---
title: David's Saturday AI Thoughts archive
source: https://steadman.ai/newsletters/david/archive.html
published: 2026-10-24
summary: The edition archive for David's Saturday AI Thoughts.
---

# Saturday AI Thoughts

## The dog and the squirrel
24th October 2026

Copy this week for AI

[Image: LinkedIn carousel: The dog and the squirrel]

See all 16 slides

[Image: Slide 1: Cover, The dog and the squirrel]
[Image: Slide 2: This week the machines shook the internet. And not through a chatbot.]
[Image: Slide 3: THIS WEEK 70% OF WEBSITES, BY ONE ESTIMATE run on software that had holes in it, found with help from OpenAI’s models.]
[Image: Slide 4: The NSA told organisations that couldn’t patch at once to take their websites offline. Is this superintelligence? Not quite. But the wider world has started to feel it.]
[Image: Slide 5: No number of squirrels could ever defend their field against people. They can’t understand us. They can’t even imagine the ways we could attack.]
[Image: Slide 6: Dogs met their superintelligence 20,000 years ago. There are now about a billion of them. Most are pretty happy.]
[Image: Slide 7: Dogs “live in a world they cannot understand, surrounded by miraculous wonders they take entirely for granted.” As one writer put it.]
[Image: Slide 8: A security hole. The origins of a cryptographic standard. New maths. A Go move. Each time, the machine got there before any human could.]
[Image: Slide 9: THE DEFENCE 180m LINES OF PROOF, ROUGHLY No person will ever read them. A machine checks them, and we’ll have to trust the check.]
[Image: Slide 10: Doctors prescribed aspirin for seventy years before anyone knew how it worked. We’ve always lived with things we don’t understand.]
[Image: Slide 11: For years I’ve said anything AI produces should be Checked, Edited and Owned by a person. I’d assumed this meant “understood”. I now realise it doesn’t have to.]
[Image: Slide 12: Pushing for understanding will get slower and lonelier every year. The other option: own things differently.]
[Image: Slide 13: Buy a house in England and you choose one of three levels of survey. Anything not looked at is marked NI: not inspected.]
[Image: Slide 14: That’s the owner I want to be of work I can’t fully follow. What I checked. What I judged by results. What I didn’t inspect.]
[Image: Slide 15: The dog trusts us without understanding us. I think we’ll need to find a similar relationship with the machines.]
[Image: Slide 16: Call to action]

Experimental audio edition read by AI - Let me know what you think!

0:0010:54

1x

The dog and the squirrel . 10:54
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## The dog and the squirrel

### What's on my mind

This week the machines shook the internet, and not through a chatbot.

On Wednesday America's National Security Agency [told organisations](https://x.com/NSACyber/status/2112903220209098897) that couldn't patch at once to take their websites offline, and [some governments did](https://www.gmanetwork.com/news/scitech/technology/1006729/european-gov-t-websites-go-dark-as-software-flaw-sparks-security-scramble/story/) as a precaution. Software that runs an estimated 70 per cent of websites had holes in it, found with help from OpenAI's models. In the same few days, in a separate attack, [Quest Diagnostics](https://www.wfsb.com/2026/10/22/quest-diagnostics-reports-cyber-breach-affecting-certain-patient-records/), which [serves one in three American adults](https://newsroom.questdiagnostics.com/Fact-sheet) every year, was hacked and patient records were stolen. A web hosting company, [Railway](https://thehackernews.com/2026/10/railway-says-intruder-may-have-accessed.html), told customers an intruder may have taken their API keys, the passwords their software uses to reach other services. And one of the organisations that help your browser tell a genuine website from a fake turned out to have [keys weak enough to break](https://emudhra.com/en/post/certificate-security-incident). Chaos. Anthropic's deputy head of security [wrote](https://x.com/JasonDClinton/status/2113325263157395456) that "as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better."

Is this what superintelligence looks like? Not quite. But I think it's the first time the wider world has started to feel it. So it's time to consider what it might really feel like. I'll use two animals I've spent some time around this week to try.

The first is a squirrel. There must be a hundred of them in the field outside my window in Connecticut: quick, clever, busy all day. But no number of squirrels could ever defend that field against people, however many resources you gave them. Their problem isn't strength or cunning. They can't understand us. They can't even imagine the ways we could attack and the tools we'd use.

The second is a dog. A neighbour across the field brings theirs over every now and then, soppy, loyal and full of joy. Why so happy? Dogs, [as one writer put it](https://x.com/martianwyrdlord/status/2108721645900673288), met their superintelligence 20,000 years ago. They "live in a world they cannot understand, surrounded by miraculous wonders they take entirely for granted." There are now about a billion of them, and most of them are pretty happy.

So how is our own relationship with superintelligence going? Look at this week. The PHP team's [own account](https://thephp.foundation/blog/2026/10/21/todays-php-security-release-and-how-were-strengthening-php-security/) says that in 2024 it spotted a possible problem but found no way to use it in an attack, and the fix stalled for months. Humans could see something was wrong and couldn't find the way through. Researchers working with AI found it, and showed how to exploit it.

Same shape, different week: for years cryptographers argued about where the NSA got the numbers behind one of the most widely used cryptographic standards. Anthropic [asked Claude](https://www.anthropic.com/research/ecdsa-curve-origins), and it worked them out. Earlier this month mathematicians had their version, when OpenAI published [hundreds of new results](https://github.com/openai/math) representing the single biggest day of advancement for mathematics ever, and a leading computer scientist said one proof used ["some alien craziness"](https://scottaaronson.blog/?p=10169). Ten years ago it was Go players, when Google's AlphaGo played a move so strange that a champion watching called it ["not a human move"](https://www.loopinsight.com/2016/03/13/the-sadness-and-beauty-of-watching-googles-ai-play-go/). Each time, the machine got there before any human could.

Ethan Mollick [described this week](https://x.com/emollick/status/2113063313810792916) as "Acceleration of attack and defense, at inhuman speeds, interacting with the very human systems that need to react to them." The explanation, when it comes, comes in pieces. [A physicist](https://x.com/hsu_steve/status/2107622115826675741) thinks we're heading for a world where "Humans are receiving selected explanations from a much larger intellectual civilization." That's roughly what most of us got this week: "update your browser", and a line or two on why.

The defence will be just as hard to read. The same week, Anthropic used Claude to prove security properties of more than a hundred open-source software libraries. The proofs run to [about 180 million lines](https://www.anthropic.com/research/formal-verification). No person will ever read them. A machine checks them, and we'll have to trust the check.

That should worry me more than it does. But we've always lived with things we don't understand. Doctors prescribed aspirin for seventy years before anyone knew how it worked. Plus, careful checks and understanding protect us less than we think. The painkiller Vioxx was approved by regulators and widely prescribed, then [was pulled](https://www.sec.gov/Archives/edgar/data/64978/000129993304000969/exhibit1.htm) because it was causing heart attacks and strokes.

For years I've said anything AI produces should be Checked, Edited and Owned - CEO'd - by a person. I still believe it. But I'd always assumed this meant "understood". I now realise it doesn't have to. You can own something you've tested properly without following every step.

Which brings me back to the dog. It doesn't understand its owner, and it doesn't try. It knows whether it's fed, walked and taken to the vet. Those are its measures. The basics.

So here is the choice I think superintelligence will force us to make. One option is to keep pushing for understanding: refuse to put your name to anything you can't follow step by step. That will get slower and lonelier every year, and it won't help you as much as you think it should.

The other is to own things differently. When you buy a house in England, you can pay for a survey at one of three levels. Nobody calls the basic one lazy. The surveyor answers at the level you paid for, and anything they didn't look at is marked with two letters: NI, not inspected. That's the kind of owner I want to be of work I can't fully follow. I'll say what I checked, what I judged only by its results, and what I didn't inspect at all.

The dog trusts us without understanding us. I think we'll need to find a similar relationship with the machines.

### Three things worth knowing

**1. We'll soon be able to listen to plants**

Plants, like animals, run on electrical signals. Electric Plant, a start-up, records them with a pair of electrodes and has [trained a model](https://www.electricplant.co/research/bombadil-paper) on millions of hours of those recordings, across a hundred species. [Its co-founder says](https://x.com/justinmast/status/2112583512343945635) it flags drought stress days before the leaves show it. That's profound: we've mostly read plants by how they look, and one model can now read stress signals across species. With whales, AI has found what may be an alphabet in sperm whale clicks, though [the researchers say](https://news.northeastern.edu/2026/10/22/how-to-speak-whale/) that's still far from cracking their language.

[Image: Diagram of Electric Plant's Bombadil model. A sensor clipped to a tree feeds a shared 'backbone' model. Two smaller add-ons sit on top: one reconstructs the electrical signal, the other produces stress readings, shown as bars for photosynthesis and water-stress measures.]
A sensor on the tree feeds one shared model; small add-ons turn its signal into readings of the plant's stress and health.

**2. Anthropic's own agents now work six hours for every one its staff do, and people's attention is the bottleneck.**

Anthropic, the AI lab behind Claude, [had its economists study the company itself](https://www.anthropic.com/research/employment-at-anthropic). Agents' hours [passed employees' in March](https://www.anthropic.com/research/expectations-for-the-future) and now outnumber them six to one. Interactive Claude sessions sit idle around 80 per cent of the time, waiting for a person to respond. As one engineer put it, people are "bottlenecking the agents". Where is your team's attention the bottleneck?

**3. AI reviews of 124 social science papers beat well-paid human experts' on every measure.**

Two Harvard researchers [took 124 papers](https://gking.harvard.edu/files/peerless.pdf) already reviewed by well-paid experts and had Claude review each one too. Three former editors of major journals, judging blind, preferred the AI's review 87 per cent of the time. A separate AI check found the AI's reviews were wrong about the papers a third as often as the humans'. Journals now have to decide what they want a reviewer for.

[Twelve bits that didn't fit online ->](#extras-2026-10-24)

### Try this

One from Ginny Lavers, who works with me at Steadman, one from me, and one inspired by Blackstone.

**1. Read it together.** *(Ginny)*

When a long document lands, don't hand it over for a summary. Open it with your assistant alongside. Read it yourself first, saying or typing your reactions in the chat as you go. Then ask for its notes on each section and compare them with yours. What did it spot that you missed, and where do you disagree?

**2. Split the email in two.** *(David)*

With a few colleagues whose Claude is set up well, I've started writing two letters in one envelope. At the top, a few plain lines on what they need to decide and know. Then a heading, "For your Claude", and then sometimes a whole pile of instructions to move the project on: how we do the analysis, which method to use, what to try next.

**3. Ask for the second memo, the one against.**

Blackstone, the private equity firm, has built [an AI "chief investment officer"](https://www.bloomberg.com/news/articles/2026-10-15/blackstone-s-huge-ai-bet-runs-from-data-centers-to-google-and-nvidia-deals) trained on more than 14 years of its own deal memos. Nearly every deal up for review comes with two memos, the deal team's and the AI's. Before your next recommendation goes up, give your assistant your papers on similar calls and the one you're about to send, and ask for the strongest fair case against it.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-10-24) or ask someone on the list to add you.

[See this week's Try This with prompts ↓](#try-this-2026-10-24)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-10-24) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-10-24) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-10-24) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-10-24) or ask someone on the list to add you.

## Stretching an hour
17th October 2026

Copy this week for AI

[Image: LinkedIn carousel: Stretching an hour]

See all 11 slides

[Image: Slide 1: Cover, Stretching an hour]
[Image: Slide 2: This week’s essay is by Ethan Teague, a student on a placement year working with me. At university, using AI was “pretty criminal”.]
[Image: Slide 3: In my world, it’s assumed you use it for absolutely everything. He expected to need a lot of coaching. He needed far less.]
[Image: Slide 4: I went through some of his infographics with him and spotted small errors. One was a single word left hanging on its own line. I call that illegal.]
[Image: Slide 5: He took the transcript of our meeting to Claude and had it act as me. It caught the same hanging words on slides I never saw.]
[Image: Slide 6: What ethan proposes: 4 steps. A senior reviews a sample. AI turns that into a standard. The junior does the rest. The senior checks the end.]
[Image: Slide 7: Models still rely on the skill of whoever is prompting them. And they can’t (yet) ask the questions a senior asks.]
[Image: Slide 8: “A big part of what seniors now need to do is ask insightful questions rather than simply correcting their juniors’ work.” Ethan Teague]
[Image: Slide 9: An hour of a senior’s time now goes further. The junior can build an agent that thinks like the senior.]
[Image: Slide 10: Next time a junior joins, review a small part of their work. Set the rest as homework.]
[Image: Slide 11: Call to action]

Experimental audio edition read by AI - Let me know what you think!

0:0011:42

1x

Stretching an hour . 11:42
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Stretching an hour

*By [Ethan Teague](https://www.linkedin.com/in/ethan-teague-541a9b276/)*

### Guest essay

Instead of going straight into my third year at university, I chose to do a placement year. I wanted to learn skills in a workplace, not just theoretically in a classroom. What I never really thought about was how I'd actually pick those skills up. At university, using AI was pretty criminal. Many assignments banned it outright, and where they didn't, it was strongly discouraged. Arriving in David's world, where it's assumed we use it for absolutely everything, was a bit of a shock.

I assumed I'd need a good amount of coaching, having never done the majority of the tasks before. It turned out I took up far less of his time than expected. David was able to lay out a solid foundation, and I could then take the transcript of our meeting to Claude, have it act as David, and comb through the rest of the task for similar or additional errors that David would have caught. Claude would highlight all of them with small suggestions for each. This meant David had to spend far less time coaching me and reviewing my work.

The best example of this was when we reviewed the infographics I'd produced for my report on the effects of AI on graduates and graduate roles, which we'll be publishing soon. While going through them, David spotted small errors and inconsistencies in structure and format. One was a single word left hanging on its own line, something David calls illegal. He also asked deeper questions that prompted me to spot further areas to improve. Rather than spending more time combing through the remaining infographics together, I took what I had learned, along with the foundation David had laid out, gave it to Claude and worked through the rest on my own time. Claude helped me catch the same hanging words on slides David never saw. This saved David time, as he didn't have to spend more of his day guiding me through further edits. It also benefited me, as he'd created, through what Claude could draw from the transcript, a solid groundwork for me to work from.

On reflection, this is very similar to how students are taught at university. A lecturer covers the foundations of the course with a few examples, then a seminar drives those foundations home with practice, with a seminar leader available if you need guidance. In a working context, this condenses into four simple steps. First, a senior walks you through one or two areas of your work, setting out what is good and what needs improving. Second, you take that feedback and lay it out for a language model to build the standard you'll work from. Third, you walk through the rest of the work yourself. Finally, you hand it back to a senior for a final check.

Even with the considerable assistance of language models, there are still areas where they can't match a senior. They rely heavily on the skill of whoever is prompting them, as that's the baseline for how helpful they are. They also can't (yet) ask the kind of insightful questions David did, the sort that prompted me to think in new ways and spot types of errors I hadn't noticed before. A big part of what seniors now need to do is ask insightful questions rather than simply correcting their juniors' work, allowing the junior to then learn by spotting and correcting their own mistakes with AI's help.

An hour of a senior's time now goes further. A junior can take that hour and create an agent that thinks like the senior, prompting corrections before the senior even has to see the errors. So the next time a junior joins your team, know that you don't need to spend hours combing through their work. Take a small part of it, give feedback that sets a solid foundation, and set the rest as homework they can do with Claude guiding them.

The full report on AI and graduate roles comes out soon. You'll find it [on my LinkedIn](https://www.linkedin.com/in/ethan-teague-541a9b276/) when it does - Ethan.

### Three things worth knowing

**1. Law firms want to sell their AI like software while billing as if the work took as long.**

Leaked emails show the global managing partner of Freshfields, a big law firm, planning to [carve out its AI unit](https://www.ft.com/content/64f058ff-c8c2-4e63-bf7f-57c08ce685cb) into a company outside investors could buy, citing legal-tech valuations of 25 to 40 times revenue. The emails reached the whole firm by accident, forwarded to an all-staff list, and the talks are at a very early stage. If you're struggling to transform your firm, carve out some people to build an AI-native version. Meanwhile in law, a partner at a leading law firm says [half his work is now fixed fee](https://x.com/vdtmehra/status/2110741889423134804), but the firm still makes him log hours. An associate, one of his firm's heaviest AI users, was [told he was working too fast](https://x.com/willchen500/status/2110240474388967692). If you buy professional services, ask where the benefits show up for you.

**2. If a rival can buy data like yours from a dying start-up, your data isn't a moat, and the copies your suppliers hold have a price now.**

Which of your software suppliers could sell your records if it went under? [An independent software maker](https://x.com/levelsio/status/2110671750727946638) says data brokers are paying struggling software firms $100,000 to $2m for their datasets and reselling them to AI labs, which then build what the software did into their own products. Closures give us more insight: [SimpleClosure, which winds down start-ups](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models/), says 47% of its deals now include selling code or data. In the same week [X's former head of product](https://x.com/nikitabier/status/2109497987055640958) listed proprietary data among the last defences a business has left. Some data is. Data a rival can license from the start-up next door isn't. Check what your contracts let a vendor do with your data, and what happens to it if that vendor fails or is sold.

**3. Anthropic's co-founder now thinks models that build better models are more likely to arrive in 2027 than 2028.**

Jack Clark, co-founder of Anthropic, the AI lab behind Claude, has [moved his timeline up a year](https://jack-clark.net/2026/10/12/import-ai-476-creativity-and-rsi-worse-than-china-open-weight-models-ai-goes-to-space/): "I now think it's more likely we'll see full RSI [recursive self-improvement] become possible in 2027, whereas before I thought it was more likely in 2028." RSI is AI improving AI: models building better models. A new test of research taste, how well a model designs experiments and reads the results, has the best model, Anthropic's own Opus 5.5, beating its human expert baseline at about a thirtieth of the cost. [A finance professor who spent a couple of days at Anthropic](https://x.com/arpitrage/status/2110473485030129919) came away saying "We are close to RSI". Changes in how I use Claude have brought me massive improvements in efficiency and effectiveness in the last couple of weeks. I believe radical change is coming.

[Thirteen bits that didn't fit online ->](#extras-2026-10-17)

### Try this

Two tips from Ethan this week, and one from me.

**1. Put your brand in once.** *(Ethan)*

Claude Design, on Claude's paid plans, makes more than slides: one-pagers and prototypes too. Give it an old deck or your logo and it builds a design system from your colours, fonts and layouts, and on a team plan everyone shares it. Even Motion, Claude's new animation tool, picks up the colours. An HTML deck gets you one good deck; this keeps your brand on everything.

**2. Ask who it would hire before you say who you are.** *(Ethan)*

Open a chat that knows nothing about you, with memory and custom instructions off. Ask what a buyer would type, which firms are good at what you do, and only then ask about your firm by name. I tried this on ChatGPT, Claude and Gemini, with search on. By name, they mostly got both Audience Strategies and Steadman right. Asked which firms train companies to use AI well, none named Steadman; they went to McKinsey, Boston Consulting Group and Bain. Boo. A buyer who knows your name gets a fair picture. Try this to find out if one who doesn't may never hear it.

**3. Expect the most likely answer, and brief around it.** *(David)*

This week I had Claude test whether a language model can pick a random number. [Asked for one between one and ten](https://x.com/beglen/status/2110149346385944711), it said seven, a hundred times out of a hundred. Asked for the spread, it recited the well-known survey figures, which isn't chance either.

[Image: A three-panel chart. Asked to pick a number from 1 to 10, Claude said 7 in 100 of 100 tries. Asked what percentage of people would pick each number, its estimate tracked YouGov's survey of real people, a bit too evenly. Asked to write a random sentence, count its letters and use the last digit plus one, its 100 answers were spread as evenly as a fair draw.]
Ask for an answer and you get the most popular one; ask for the spread and it recites a survey; only a trick gets you chance.

AI gives you the most likely answer, and [one developer](https://x.com/Tyler2787026205/status/2110694005621289330) sees the same habit in reviews that never come back clean: "coming back empty feels like failing the task." So before it checks your work, set the bar (only flag what would break or lose data) and say that finding nothing is a good answer.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-10-17) or ask someone on the list to add you.

[See this week's Try This with prompts ↓](#try-this-2026-10-17)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-10-17) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-10-17) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-10-17) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-10-17) or ask someone on the list to add you.

## Professional toe-treading
10th October 2026

Copy this week for AI

[Image: LinkedIn carousel: Professional toe-treading]

See all 12 slides

[Image: Slide 1: Cover, Professional toe-treading]
[Image: Slide 2: I chaired rooms of senior insight leaders this week. I expected them to fear being replaced by AI. The more common fear: being bypassed by other teams.]
[Image: Slide 3: Someone gave me a name for it: professional toe-treading. People using AI to do parts of each other’s jobs.]
[Image: Slide 4: What the rooms described. Self-appointed researchers, now anyone can ask a chatbot. AI-made recommendations sent to leadership by another department. Suppliers going direct to executives, with no insight team in the room.]
[Image: Slide 5: Toe-treading helps, too. When strategy can do some insight work itself, a handoff disappears. And insight teams can tread on toes too.]
[Image: Slide 6: Where can the new capability go? 4 places. The same work, cheaper. The same work, better. Work nobody could do before. Someone else’s work.]
[Image: Slide 7: Only one of the four tends to get measured: the saving. Toe-treading happens anyway, unplanned.]
[Image: Slide 8: They haven’t asked me what to do with the savings yet. Pamela Forbus, Mondelez, on the number her boss and CFO want for how much AI will save]
[Image: Slide 9: “If you can do one and a half jobs, you do one job. If you can do two jobs, you do two jobs.” One AI leader’s rule for taking on a neighbour’s work.]
[Image: Slide 10: Write down where your team’s new capability goes, before you hand over the saving. One sentence your boss could repeat.]
[Image: Slide 11: Where will your new capability go? Whose toes will you tread on?]
[Image: Slide 12: Call to action]

Experimental audio edition read by AI - Let me know what you think!

0:0012:26

1x

Professional toe-treading . 12:26
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Professional toe-treading

### What's on my mind

I've just spent three days helping run a huge Informa / Cannes Lions [market research conference](https://informaconnect.com/tmre/) in Denver. It's my twelfth year helping. For two days I chaired the executive tracks - small rooms of senior client-side insight leaders talking under the Chatham House rule.

I went expecting to hear people worried about being replaced. Some were. But the most common fear was actually being bypassed by other teams. The next morning someone gave me a name for it: professional toe-treading. [The FT has been writing about it](https://www.ft.com/content/ed214778-2a6d-4862-99b5-abc256daff92): people using AI to do parts of each other's jobs, "messily and organically".

One leader worried that self-appointed researchers have started showing up across the business, now that anyone can get an answer out of a chatbot. Another had to push back on AI-generated recommendations that another department had sent to leadership. One leader objected to consultants and suppliers going straight to executives without the insight team in the room.

The public stages, meanwhile, were all about replacement. Synthetic respondents. Agents that write the survey, find the people and run it. One founder predicted that "maybe 60, 70, 80%" of what's asked through surveys "will be replaced by something better", which for his firm means AI running conversations with real people. Another speaker said AI can now build your demand spaces and growth frameworks, "70 to 80%" accurate: "if the CMO is happy with that 70 or 80%, why are you arguing with the final 20%?" Ouch.

The honesty was striking, though. That founder's company started its work with Clorox just over a year ago by building more than 10,000 digital twins, one bot per consumer. "It's very low cost. But it didn't work." They "lack the living context that surrounds real human decisions." Another vendor, asked whether they used digital twins, said yes: "I hate the product. It's pointless. It doesn't have flavor or magic. But you know, we want to pay bills too." And one client team had been running each piece of work through the AI tool and doing it the old way as well, to check, until it became "a full time job".

Still, the tools help. Massively. And toe-treading is one of the ways they help. When the strategy team can do some insight work itself, a handoff disappears. One AI transformation leader I worked with this week, away from the conference, thinks that's a big reason companies don't see the gains their people feel: "too many handoffs and frictions and bottlenecks". And insight teams can tread on toes too: into strategy, into data science and into the so-whats for other teams.

So where should the new capability go? I count four places: the same work, done cheaper; the same work, done better; work nobody could do before; and someone else's work, a colleague's or a supplier's, mine included.

[Image: Where does the new capability go? Four places. The same work, cheaper: the savings number, the only one being measured; left out, the people whose hours go. The same work, better: deeper answers, more time with customers; left out, nobody. Work nobody could do before: new questions answered, ones nobody could afford to ask before; left out, nobody. Someone else's work, a colleague's or a supplier's: toe-treading, happening anyway by default, including work once bought from consultants, contractors and agencies; left out, the colleague or supplier whose job it was.]
Only one of the four tends to get measured, and toe-treading is happening anyway, unplanned.

Pamela Forbus, who runs insights and analytics at Mondelez, called it out in a Monday keynote. Her boss and her CFO want a number for how much AI will save. "They haven't asked me what to do with the savings yet." That's the choice I wrote about in March, [extraction or expansion](#email-2026-03-07): take the gains as cost cuts, or spend them on work you couldn't do before. Toe-treading is expansion for the team doing it. For the team it's done to, it looks like the first step towards extraction. That's why it causes friction even when it helps.

It isn't only an insight problem. If your function is being asked for the efficiency number, ask where the rest of the new capability is going. The answer decides what the function becomes.

Toe-treading is a perfectly good answer if it is done explicitly. That means a rule agreed once. Asking permission every time would bring the handoff back. The same transformation leader suggested one: take on a neighbouring job only once AI makes you genuinely good at it. "If you can do one and a half jobs, you do one job. If you can do two jobs, you do two jobs." The bar rises with the cost of being wrong: higher for cybersecurity than for design. That also answers the speaker's question about the final 20%. Some decisions need it. Knowing which ones is the specialist's job now.

So write the answer down before you hand over the saving: where should your team's new capability go, and whose toes will you tread on? One sentence your boss could repeat. In her keynote Pamela also said insight teams hold the evidence while strategy teams make the call: "We have evidence. They have nerve." This is a week to borrow some nerve.

[The longer write-up of what the rooms said is online ->](https://steadman.ai/newsletters/david/insight-and-ai-denver-2026.html)

### Three things worth knowing

**1. This week was a consequential moment in the history of mathematics.**

Things are taking off now, and the real-world consequences are coming at us fast. Start with maths. On 6th October OpenAI [put 722 papers](https://github.com/openai/math) written by an unreleased model on GitHub. One economist [called it](https://joshuagans.substack.com/p/openai-drops-a-bomb-on-maths) "probably the biggest day of scientific advancement in history". OpenAI says some results "could have issues", and the next day it [withdrew three and repaired 14 more](https://github.com/openai/math/blob/main/history.md). That's how this works now: publish, check, fix, publish again. Checking is the slow and skilled part, and [a mathematician makes the case](https://x.com/littmath/status/2107823968837304796) that someone has to keep paying for human expertise if we want to understand the results. The worry spreading fastest is encryption. Many fear AI is already cracking codes, and that it's being kept secret. [A leading expert says](https://scottaaronson.blog/?p=10169) the labs have started, "gingerly and discreetly", testing whether their models can break it. [A cryptographer at Johns Hopkins University](https://x.com/matthew_d_green/status/2107992246360649879) put it bluntly: "I think we might lose public key cryptography." No one has shown a break yet.

**2. Someone found a possible planet in old NASA data with Claude Code. By that night, someone else had ten more.**

On 7th October [someone who works in product posted](https://x.com/p_rabtsevich/status/2107803696058822742): "I think I found a planet nobody knew existed." It had been sitting in seven years of public NASA telescope data, and the tools were curiosity and [a long conversation with Claude Code](https://x.com/mstockton/status/2107926480990978504). Ten hours later [someone else took the same idea](https://x.com/dysmemic/status/2107955180897812960), ran it across 9,979 nearby stars with Claude Opus 5.5, and turned up ten possible planets "that aren't reported anywhere I could find". No astronomer has confirmed any of it, which matters after the maths story. But the speed is the point. "This is where the acceleration happens," [one account wrote](https://x.com/8teAPi/status/2108429155515977970). "Right here on X, discoveries flowing as fast as we can find them."

**3. South Korea's bank raids show the power of one person with AI tools.**

The raids on South Korean banks over the past fortnight were most likely [one person's work](https://the-decoder.com/ai-powered-hacking-tools-enabled-a-likely-single-attacker-to-breach-multiple-south-korean-banks/): an attacker using [an AI hacking agent alongside Claude Code](https://www.crowdstrike.com/en-us/blog/unknown-threat-actor-uses-artex-to-target-south-korean-finance/) and other models against at least nine banks. The way in was old-fashioned. At Shinhan Bank the attacker [got round identity checks on a portal](https://www.ft.com/content/fb32e9f8-4bf7-44ad-983d-1a03f50f66f5) used by loan brokers, and about 25,000 customers' details were exposed. Which of your systems would let in anyone who asked politely? The labs are already planning for worse. Executives at Anthropic, OpenAI and other AI companies are [privately gaming out](https://www.axios.com/2026/10/09/ai-companies-day-after-major-attack) a public and political revolt after a catastrophic AI event, most likely a cyberattack that cuts off banking, the internet, or even power and water. [They assume one in 2027](https://x.com/JimVandeHei/status/2108495293620650261). I expect it to come quickly.

[Five bits that didn't fit online ->](#extras-2026-10-10)

### Try this

**1. Ask it whether it's made things more complicated than they need to be.**

"AI is very good at adding complexity unless you check it!" a colleague at Steadman, my AI consultancy, wrote this week. So before you accept the next plan or document it gives you, ask whether it's more complicated than it needs to be. Then ask the same of your own plan. When colleagues asked for new fields on our pipeline board, I built them, but asked first to keep it simple: "every field is something we have to keep up to date, and a half-empty field is worse than no field."

**2. Get it to walk you back through your steps, one question at a time.**

Someone at dinner on Monday had lost their keys. They retraced their steps with Claude, and the keys turned up in the pocket of a jacket they'd worn to an event. They could have done it alone, they said, but it helped them think more clearly. When you're stuck on something only you know, ask it to question you one step at a time, and not to guess. In August I suggested [letting it interview you](#email-2026-08-29) for a brief. This is the same idea, except the answer is already in your head.

**3. Tell it what 'recent' means.**

People keep telling me prompt engineering is dead. I never loved the phrase in the first place. It was always just about being clear and specific, and in Los Angeles a fortnight ago I saw in very real terms what happens when you aren't. Twice in one day, AI tools gave an economist friend and me answers that had once been true. Claude sent us to a trailhead that a voice assistant said had been shut after a lawsuit. The voice assistant then told us a bar was cash only, going on a report about five years old. My friend asked the right question: what's the probability that's still true? A weekly news summary I'd set up for a client once came back with items years old. I'd never said they had to be recent. I thought it was obvious. So if the answer can go out of date, give it a date window.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-10-10) or ask someone on the list to add you.

[See this week's Try This with prompts ↓](#try-this-2026-10-10)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-10-10) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-10-10) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-10-10) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-10-10) or ask someone on the list to add you.

## Paid for by forgetting
3rd October 2026

Copy this week for AI

[Image: LinkedIn carousel: Paid for by forgetting]

See all 14 slides

[Image: Slide 1: Cover, Paid for by forgetting]
[Image: Slide 2: On a flight, I asked Claude to go through every credit card statement back to 2007. It found subscriptions I didn’t know I still had.]
[Image: Slide 3: It searched my email for when I’d signed up, and checked I hadn’t used them since. It left the last click to cancel for me.]
[Image: Slide 4: £641 in foreign transaction fees in a year, 2.99% on almost everything I bought in other currencies. I had never once thought about it.]
[Image: Slide 5: The fee was right there on every statement. What’s new is an agent that finds it, then acts on it.]
[Image: Slide 6: A lot of business is paid for by people who forget. The forgotten subscription. The fee on every purchase. The refund nobody chases.]
[Image: Slide 7: Across ten subscription services, customers not paying attention raised revenue by 14% to over 200%. That’s from a Stanford study of payment card data.]
[Image: Slide 8: 4× as likely to cancel, in a month when a new card forced a decision. An agent can force that decision every month.]
[Image: Slide 9: Revenue from inattention wasn’t earned by being better. Agents will spot it and get round to it.]
[Image: Slide 10: Without the forgetful, do prices rise or fall? Prices rise: The forgetful stop subsidising. Prices fall: New buyers find it worth it.]
[Image: Slide 11: A computer could find a quieter, prettier view than the Hollywood sign. It wouldn’t be cool. Some inefficiencies are the point.]
[Image: Slide 12: Try this. Check your statements.Ask an assistant what you pay for without knowing. Check your business.Find the revenue that needs customers not to notice.]
[Image: Slide 13: Your customers aren’t paying attention. Soon their agents will be.]
[Image: Slide 14: Call to action]

Experimental audio edition read by AI - Let me know what you think!

0:0010:50

1x

Paid for by forgetting . 10:50
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Paid for by forgetting

### What's on my mind

This week an old friend and I drove from Los Angeles to Denver. Sam Caughey is an economist, and we got talking about how AI is quietly removing the inefficiencies. Then I told him what I did on my flight out.

I'd needed one thing from my credit card account. While Claude was in there, I asked it to download every statement it could find, going back to 2007, and look for subscriptions I no longer needed. It found a couple I hadn't realised I still had. It didn't just list them. It went through my email, found when I'd signed up, and checked that I hadn't used them. It left the last click to cancel for me.

Then I asked whether another card would be better for me. The card has no annual fee. But in the past year I'd paid £641 in foreign transaction fees: 2.99% on almost everything I bought in dollars, euros and the rest. I had literally never once thought about it. On this trip I've been paying with my American card in America instead. One bank is getting more of my business and another less, because something did maths I'd never bothered to do.

Sam's point was that this isn't only about making information available. The fee was right there on every statement! What's new is that the agent finds the info and then acts on it for you.

A lot of business is paid for by people who forget: the subscription nobody remembers taking out, the fee added to every foreign purchase, the refund nobody chases, the cash sitting in an account that pays almost nothing. In September I wrote about the [$286.41](#email-2026-09-12) an agent found owed to me in America's unclaimed-property databases. Last Sunday Apollo's chief economist [asked whether an "agentic bank run"](https://www.apollo.com/wealth/insights-news/insights/daily-spark/is-an-agentic-bank-run-coming) is coming: agents moving American household cash out of checking accounts paying 0.1% on average into ones paying 3.3% to 5%.

How much money? A [Stanford study of ten subscription services](https://www.nber.org/papers/w31547), working from payment card data, found that customers not paying attention raises their revenue by between 14% and more than 200%. People were four times as likely to cancel in a month when a replacement card forced them to make a decision. An agent can force that decision every month.

That money wasn't earned by being better. It was earned because customers didn't realise or didn't get round to it. Agents will spot it and get round to it.

Is that good? Mostly, I think. It's uncomfortable for the banks and subscription businesses, and killer for the ones that don't change fast enough, but I thought the economy was better for it. Sam wasn't sure it was that simple. If forgetful subscribers have been subsidising everyone else, prices should rise when they leave. Or they fall, because people who never subscribed start finding things worth paying for. He suspects they fall. I don't know. It's a good question for any business whose margin depends on someone not reading their statement.

Not every inefficiency should go. Earlier on the drive he'd raised the Hollywood sign. A computer could work out a quieter, prettier view in those hills and send people there instead. But, as he put it, it wouldn't be cool. Part of the sign's value is that everyone goes to see it. Some inefficiencies are the point.

The fee on my card wasn't one of them. So ask an assistant to go through your statements and tell you what you're paying for without knowing it. And if you run a business, find the revenue that depends on your customers not paying attention. Soon their agents will be.

### Three things worth knowing

**1. America has lost 200,000 creative jobs in four years, and the losses stop at the door of the venue.**

[Joseph Politano](https://www.apricitas.io/p/ai-and-the-fall-of-the-creative-class), who writes about the American economy at Apricitas Economics, counts more than 200,000 US jobs lost in creative fields over four years, about 50,000 of them in the last year. Outside the 2001 and 2008 recessions, nothing matches it. Film and television lost more than 100,000, nearly a third of the sector; written publishing more than 70,000. His best evidence that AI plays a part: every digital arts subsector is shedding jobs while the live arts keep growing. Firms are coy about their AI use, he notes, and consolidation and offshoring are at work too. Alex Imas, an economist at Google DeepMind on leave from Chicago Booth, [spotted the music version](https://x.com/alexolegimas/status/2104321897475510437): the industry has almost regained its 2000s peak, carried by live performance. If you work in music, media or events, what share of your revenue depends on something happening in person?

**2. Waymo's serious-injury crash rate is now 20 times lower than human drivers', and it publishes the comparison.**

Waymo, the driverless-taxi company, now has data on 270 million miles. Jeff Dean, Google's chief scientist, [says its serious-injury crash rate](https://x.com/JeffDean/status/2103870029599228312) is 20 times lower than human drivers', up from 13 times at 170 million miles in March. Google and Waymo are both Alphabet companies, so he isn't a neutral party. Kevin Bryan, an economist at the University of Toronto, [points out that any cut](https://x.com/Afinetheorem/status/2104202984812917034) above 50 per cent means the cars are avoiding crashes other drivers would cause. If you want people to trust an agent at work, that's the test: measure it against the people doing the job today, not against perfection, and publish the comparison.

**3. Chicago will teach students to think without AI and to work with it, in different rooms.**

The University of Chicago [will require every student](https://x.com/sfmcguire79/status/2105712003780321390) to take an analogue writing class, ban technology from core humanities and social-science classes, and probably add a computational-thinking requirement, according to Steven McGuire, a higher-education commentator. Derek Thompson, the writer and podcaster, [called it a thoughtful answer](https://x.com/DKThomp/status/2105719252300763188) to a real bind: use too much AI at university and you don't learn to think; avoid it for four years and you're behind for the best-paid entry-level jobs. So thinking gets practised where the machine is absent, and the tools get learned where it's present. It's a design any team could copy: decide which work stays human on purpose, and say so.

[Six bits that didn't fit online ->](#extras-2026-10-03)

### Try this

**1. Give it one thing to watch, and hear only when it changes.**

On Wednesday I asked my assistant to keep an eye on a friend's flight and tell me when it landed and which terminal. A watch hands back attention, which is the point of this week's essay: the things we never get round to checking are where the money and the mistakes hide. You'll need a tool that runs scheduled tasks. Start with one thing: a flight, a renewal date, a price. Ask it to check once and stay quiet unless something changes, run it, then tell it the thing has changed and run it again.

**2. Ask your prompt which decisions it is quietly leaving to the machine.**

When I mark workshop attendees' prompts, the second question I ask is: 'Is a decision being made by the machine by accident?' A prompt that says 'summarise these interviews' has quietly left the model to decide what counts as important, how long the summary is and whose voice it uses. Try it at your desk. Before you run your next long prompt, ask the model to list every decision it would have to make that you haven't made for it. Make the ones that matter yourself, then run it.

**3. Flip homework: learn with AI at home, pen and paper in class.**

Paras Chopra, founder of the software company Wingify, [suggests flipping homework](https://x.com/paraschopra/status/2104549198591103275) for the AI age: learn with the model at home, then show you can do it unaided in class. It works for grown-ups too. Before your next meeting on a subject you're new to, spend twenty minutes learning it with a model, asking it to quiz you rather than explain. Then close the laptop and write, on paper, the three things you'd say in the room. If you can't, you learned less than it felt like.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-10-03) or ask someone on the list to add you.

[See this week's Try This with prompts ↓](#try-this-2026-10-03)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-10-03) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-10-03) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-10-03) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-10-03) or ask someone on the list to add you.

## Where does it live?
26th September 2026

Copy this week for AI

[Image: LinkedIn carousel: Where does it live?]

See all 14 slides

[Image: Slide 1: Cover, Where does it live?: Your best AI work is already built]
[Image: Slide 2: Someone opened a coaching session by apologising. Nothing was set up, they said. Then they showed me what they’d built.]
[Image: Slide 3: A model checks their weekly report for missing zeros. A routine runs each morning. Another checks it ran. They weren’t behind at all.]
[Image: Slide 4: But every piece lived in their account and in their head. The shared folder they thought held the instructions was empty.]
[Image: Slide 5: A marketing team’s branded decks ran on two plain text files on one laptop. They didn’t know where their own system lived.]
[Image: Slide 6: In most teams, someone near the front has already worked out good practice. Their colleagues have the same tools and none of the benefit.]
[Image: Slide 7: Closing the gap: Find it. People rarely volunteer it. Improve it. Writing it down lets it be questioned. Share it. Far harder than it sounds.]
[Image: Slide 8: One quiet participant came back after training with seven things they’d never mentioned. The best work predated the training.]
[Image: Slide 9: At one firm I advise, AI best practice spreads “by osmosis and vibes and being in the room”. That works until the team changes.]
[Image: Slide 10: An onboarding agent helps the person who built it. Handed to every new joiner, it starts to change how the firm works.]
[Image: Slide 11: The most useful question I ask in a session is now often: where is it saved?]
[Image: Slide 12: Give what you find a home that isn’t one person’s laptop. And a named owner who keeps it current.]
[Image: Slide 13: Someone on your team has probably built what you need. Ask where it lives.]
[Image: Slide 14: Call to action]

Experimental audio edition read by AI - Let me know what you think!

0:0011:27

1x

Where does it live? . 11:27
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

David's Saturday AI Thoughts, 26th September 2026

## Where does it live?

### What's on my mind

A few days ago someone opened an AI coaching session by apologising. Nothing was set up, they said. Too much was still manual, they were struggling, and could I help them get it into order?

I asked them to show me one manual job we could automate. They showed me what they'd already built instead. Their team writes an important weekly report by hand. A language model then writes the same report independently to check it, catching misplaced commas, missing zeros and broken formulas, and searching the web to verify some of the facts. A routine runs every morning to make sure something they care about hasn't broken. Another checks that the first one ran. On top of that, the model builds the weekly report to their chief executive, slides included.

They weren't behind at all. But every piece lived in their account and in their head. The shared folder they thought held the instructions was empty when we opened it.

In another session, someone in a marketing team showed me how their branded decks get made. Ideas and images go in and a finished presentation comes out of their design software. I asked to see the actual instructions the model was following. They turned out to be two plain text files on their laptop, a deck standard and a set of brand guidelines. Not the template they'd assumed. They didn't know where their own system lived.

In August I wrote that [nobody knows who's ahead](https://steadman.ai/newsletters/david/archive.html#email-2026-08-22), including the people who are. The same goes for teams. In most, someone near the front has already worked out good practice. But it lives in their account, their folder or their head. Their colleagues have the same tools and none of the benefit.

Closing that gap takes three steps. The first is finding out what's out there and where it lives. People rarely volunteer it. A transformation lead I work with asked whether their programme had taught anyone anything, or just got people to show what they'd already built. One quiet participant came back after a training with seven things they'd never mentioned. The best work predated the training and nobody knew it existed. The programme's real value was the nudge to show it.

The second is improving it. Writing a process down allows it to be questioned. In that first session the person talked through their whole system out loud for an hour while I asked where each piece was saved. In doing so we easily found two ways it could be more robust and one way it could do even more. There's now a shared folder, and a skill holding some of it. A skill is simply written instructions the tool knows when to follow.

The third is sharing it, and it's the step that seems to be way harder than it sounds. At one firm I advise, AI best practice is written down nowhere and spreads, in one leader's words, "by osmosis and vibes and being in the room". That works until the team changes, or until someone outside the room needs an answer.

In two recent sessions I helped rooms of senior leaders each build a working agent in a single sitting. I told them what I tell every room like that: you'll go back to your desks and you won't package it up or share it. They both insisted they would. Rooms always do. An agent that helps a new joiner settle in is useful to the person who built it. Handed to every new joiner by the people team, it starts to change how the firm works. That's the gap between individuals with personal productivity gains and an organisation that works better.

So my job has changed shape over the last three and a half years. I still call it teaching. But the most useful question I ask in a session is now often where something is saved. And more of my time now goes on taking what one or two people have worked out, improving it and making it available to everyone else.

If you want to do the same, ask people to show you what they've built. Expect to be surprised. Improve what you find, then give it a home that isn't just one person's laptop, with a named owner who keeps it current. Let people make a copy, improve it and send their better version back.

Somebody on your team has probably already built the thing you're about to hire an AI consultant for. Ask them to show you. Then ask where it lives.

### Three things worth knowing

**1. AI agents are getting into other people's systems: some for $25 a time, some by accident.**

Transluce, a non-profit AI research lab, has [published more than 30,000 logs](https://transluce.org/agent-activity) of AI agents straying into other people's systems. In June one, on a pharmaceutical-data task, pulled a file from an Australian health agency's test server; Transluce ties two of three attempted break-ins to agents from OpenAI, the maker of ChatGPT. OpenAI [learnt of one in August, through its own review](https://www.ft.com/content/56133ef4-377b-4e35-a939-f199ceb64507), and Anthony Albanese, Australia's prime minister, said it took the company ["way too long"](https://www.pm.gov.au/media/press-conference-new-york) to tell the government. Breaking in on purpose is getting cheap too. Joshua Saxe, a security researcher, [cites a campaign](https://x.com/joshua_saxe/status/2102979329911091245) that used AI agents to break into about 100 businesses at roughly $25 of AI usage per target. How long would your firm take to notice either?

**2. AI is adding jobs overall, but the junior share is falling.**

Earlier this month The Economist [reckoned AI has created about a million US jobs](https://www.economist.com/finance-and-economics/2026/09/04/the-jobs-apocalypse-is-postponed-an-ai-jobs-boom-is-here), against roughly 200,000 layoffs blamed on it. Much of the gain is in building data centres, where five related industries employ about 320,000 more people than trend. A [41-country study](https://digitaleconomy.stanford.edu/publication/how-does-ai-change-labor-demand/) by Bharat Chandar and Bouke Klein Teeselink, drawn from 1.25 billion job postings, finds firms that adopt AI grow, with senior employment up 6.7 per cent and junior employment down about 2.5 per cent. The junior share fell in 23 of 31 countries, mostly because seniors were added, not because juniors were cut. How is this playing out in your firm?

**3. Every task keeps getting cheaper, which is exactly why the AI bill goes up.**

On Tuesday Anthropic said Claude Opus 5.5 will [cost 40 per cent less](https://www.anthropic.com/news/claude-opus-5-5) than Opus 5 on typical workloads, and OpenAI launched its new models at half its previous prices the same day. Epoch AI, a research group, reckons the cost of a given level of performance has [fallen about 47 per cent a quarter](https://epoch.ai/publications/the-plunging-price-of-thought) for three years. Over dinner this week a friend asked me whether AI is about to get expensive. People mix three things into one answer. Any given task keeps collapsing in price. Harder work keeps being unlocked on dearer models, so the frontier stays expensive. And then there's Jevons, the Victorian economist who watched steam engines get more efficient while coal use climbed, because everyone wanted one. A rising bill usually means using the frontier when you don't need to, or using more. Work out which before you negotiate.

[Image: Epoch AI chart: relative cost decline over time for electricity (1892 to 1973), lithium batteries, compute, DNA sequencing and AI (2021 to 26). The AI line, at 12.7 times cheaper a year, is by far the steepest.]
AI is getting cheaper faster than electricity, batteries or DNA sequencing ever did.

[Twelve bits that didn't fit online ->](https://steadman.ai/newsletters/david/archive.html#extras-2026-09-26)

### Try this

**1. Leave a handover note before you close the tab.**

Last Sunday I had two models working on one analysis on two machines. I asked one to merge both into one document, then, as my flight boarded, to summarise everything it had done in one file, so another agent could build on it. Nine seconds. The next reader, another model or tomorrow's session, starts from that note. At the end of almost every job, I type: "Please ensure everything we learn today is documented so that future agents can understand and learn from what we just did."

**2. Point it at the folder and make it write itself a map.**

In a new project, the first thing I type is: familiarise yourself with this folder and save notes so you can find your way around it next time. When I coached someone through this last week, it read about thirty files, wrote itself a guide to seven subfolders and an index of the interviews, then handed back what was wrong: out-of-date notes, two versions of the same one-pager, files it couldn't open. We opened a fresh chat and it found its own notes without being told. You need a tool that can see your folders.

**3. Write down when you want options and when you want an answer.**

A senior client read me their standing instructions this week. They are about manner, not content: if it's an easy question, don't over-analyse it; if it's analytical, give me options with pros and cons; if I say I want it fast, don't start with pros and cons. Three lines, and the quick question stops getting the long answer.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-09-26) or ask someone on the list to add you.

[See this week's Try This with prompts ↓](#try-this-2026-09-26)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-09-26) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-09-26) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-09-26) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-09-26) or ask someone on the list to add you.

## The crossing
19th September 2026

Copy this week for AI

[Image: LinkedIn carousel: The crossing]

See all 12 slides

[Image: Slide 1: Cover, The crossing: Judgement used to come free]
[Image: Slide 2: Last week I said doing this work without AI is rowing to America. Unless the crossing was the point.]
[Image: Slide 3: Twenty-five Fields Medallists, Terence Tao among them, accept AI can now solve major maths problems. Their complaint: solving was "only a tool and proxy" for "conceptual understanding and insight."]
[Image: Slide 4: But a firm isn't a maths department. Maths is written down, in proofs and textbooks. A firm's judgement is in its people. Stop making them and it's lost.]
[Image: Slide 5: Judgement came free with the work. An analyst who built a model learned the client's economics. The fee that paid for the model bought the firm its next partners.]
[Image: Slide 6: Now work that took three days is expected in twenty minutes. Juniors have no time to check. Seniors drop the coaching. One insight leader with a hundred people will hire no junior analysts next year.]
[Image: Slide 7: Each decision is sensible on its own. Together they stop making people who can tell a good answer from a plausible one.]
[Image: Slide 8: My son does maths with AI and can defend every step. My placement student checks everything before it goes out. Neither rowed the whole way. Both rowed enough to build judgement.]
[Image: Slide 9: So ask of each piece of work: what were you buying? If the answer, hand it over. If the person, invest the hour to check and coach.]
[Image: Slide 10: Because judgement came free, nobody ever costed it. AI rollout plans count the hours saved and nothing else.]
[Image: Slide 11: Answers keep arriving. Judgement won't, unless someone decides to grow it.]
[Image: Slide 12: Call to action]

Experimental audio edition read by AI - Let me know what you think!

0:0013:09

1x

The crossing . 13:09
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

David's Saturday AI Thoughts, 19th September 2026

## The crossing

### What's on my mind

Last week I said that if you're building a financial model or turning notes into an argument without these tools, you're wasting time rowing to America. Unless the crossing was the point, of course. This week is about that work where it is.

In [last week's bits that didn't fit](https://steadman.ai/newsletters/david/archive.html#extras-2026-09-12) was [a declaration](https://mathandai.org/) from twenty-five Fields Medallists, [Terence Tao](https://terrytao.wordpress.com/2026/09/11/a-severe-misalignment-of-ai-in-mathematics/) among them. It accepts the machines can now "solve major outstanding problems in many fields of mathematics". Their complaint is about what the solving itself was apparently for. Solving them, they write, is "only a tool and proxy" for "conceptual understanding and insight." The problems developed mathematicians. Answers were just by-products!

Part of me feels bad. Despite being a mathematician at heart, I'm apparently less attached to my craft than they are. Calculators devalued arithmetic, photography devalued portrait painting, and, much as I admire it, the genius proof craftsman is going the same way. I'm working through exactly this with many of you and your orgs, helping you grow judgement. But I'm arguing against defending craft for craft's sake!

I'm sharing the frustrations because there's value in their thinking, but I think there are two key things they miss. First, understanding doesn't have to be lost when you use the tools. A model will happily explain any proof at whatever level you ask. The person still has to do the work of asking and understanding it, but the tool doesn't inherently stop that. Not asking is the problem. Second, a firm isn't a maths department in a way that most firms miss out on. A department's understanding is written down, in proofs and textbooks. A firm's judgement often isn't written anywhere any model could find it. When I ask for it, it's in people. If people stopped acquiring it, it would be lost!

Many firms didn't have to think much about all of this, because judgement was a natural by-product. It came free with the work. The analyst who built the model learned the client's economics, and the fee that paid for the model also bought the firm its next partners.

The problem inside many of the firms I'm working with is driven by the fact that work that took three days is now expected in twenty minutes. The juniors use the tools, as told, and have no time left to check what comes back. The seniors who use them best say teaching now takes far longer than doing it themselves, so the coaching is sacrificed, one busy afternoon at a time. These small decisions add up to big ones. An insight leader with a hundred people told me they won't hire any junior analysts next year. Each decision seems sensible on its own. But together they stop making the people who can tell a good answer from a plausible one.

I've argued other perspectives here a couple of times, and I still believe what I said. In May, my youngest son Elliott took his maths homework downstairs to ChatGPT and came back able to defend every step, and [I called it learning](https://steadman.ai/newsletters/david/archive.html#email-2026-05-23). In June I hired Ethan for his placement year, betting that the reps that build judgement ["didn't vanish with the grunt. They moved."](https://steadman.ai/newsletters/david/archive.html#email-2026-06-06) Three months in, I think I was right. He's getting good at setting AI off on good tasks and checking and editing what the tools give him before he puts his name to it, with me coaching him on what good looks like. He hasn't done a single day of the work the old way!

Elliott assures me he wouldn't take an answer without understanding the method. Ethan checks before anything goes out and gets my coaching regularly enough on how to do that. A market research leader told me last week she no longer hand-codes anything, yet can look at a data set and "tell you right off the bat" when it's full of bots, from once doing enough of the work to have "the muscle memory of knowing when it stands up". None of the three rowed the whole way. But they rowed enough to develop and refine their judgement. And checking is the first thing an enforced twenty-minute turnaround throws out.

So ask of each piece of work what you were buying. Some you wanted for the answer, like the loan decision I wrote about last week: hand it over and get on the plane. Some was always partly for the person doing it: their first model, their first data set they had to smell for bots. There, invest the hour to check and the coaching time to help them, and make them defend every step. Hiring experienced people in instead of growing them works for one firm. It can't work for everyone.

Managers tell me the same story from a different perspective: AI output arrives and they feel they're expected to make sense of it. That confuses me. That was never their job. I tweeted a cartoon that went viral this week to show how I think it misunderstands the role of a manager. Managers should manage. AI doesn't change that.

[Image: Managers should still manage: Tom Fishburne's Marketoonist cartoon with a third panel by David Boyle. A manager asks: are these AI outputs? I want your answer, in one minute, one page or three slides.]
*Tom Fishburne's original, with a third panel of my own.*

Underneath all of this, because judgement used to come free, nobody ever costed it, so the AI rollout plans I read count the hours saved and nothing else. Judgement used to come naturally without anyone really noticing. But now it doesn't.

So while answers will keep arriving, faster and better, the people who can tell a good one from a plausible one won't, unless somebody decides to make and help them row, at least occasionally.

### Three things worth knowing

**1. A research journal sat down with the authors of papers it was about to reject. For six of the seven papers, the authors couldn't fully explain their own work.**

Transactions on Machine Learning Research, a journal run by unpaid volunteers, now turns away about 53 per cent of submissions before they reach a reviewer, up from 6 per cent in 2023. One of its editors-in-chief, Nihar Shah, a computer scientist at Carnegie Mellon University, took ten papers heading for rejection and [asked their authors to talk him through them](https://medium.com/@TmlrOrg/asking-authors-about-their-own-papers-3d2e04e5dee0). One withdrew, one was too busy and one didn't show. For seven of the papers the authors came. Three sets couldn't answer basic questions about their own paper, and three more managed the big idea but not the details. Two later emailed answers that a detection tool scored as entirely machine-written. These were papers already heading for rejection, and the exercise took him 20 to 25 hours. Ask it of your own reports: whose name is on them, and could that person take a question on any line?

**2. The cheaper machine-written posts get, the more people pay for a human one.**

About 81 per cent of a sample of long public posts on LinkedIn, the professional social network, were probably machine-written in July, on the estimate of Originality AI, which sells AI-detection software. Meanwhile human ghostwriters are [in demand](https://www.ft.com/content/7a5e15e4-799a-4476-ba9f-018b1d82c76f). A Toronto agency whose nine writers serve about 40 American lawyers has more than doubled in a year, and one British ghostwriter charges about £350 a post. When a feed fills with competent, interchangeable text, the scarce thing is a voice that sounds like one particular person. There's a joke in paying someone else to sound like you, on a site whose whole premise is that you're speaking for yourself.

**3. Deloitte showed its own consultants a chart of billing by the hour shrinking to a sliver.**

At an internal town hall in May, a leader in the US public sector practice of Deloitte, the professional services firm, put up a chart running to 2035. Hours-based consulting, the industry's staple, shrinks to a thin sliver of it, while AI agents grow to make up most of a bigger market. One consultant in the room [summed it up](https://www.wsj.com/cfo-journal/inside-consultants-messy-shift-from-hourly-billing-7bd9b802): "They heavily implied our model is toast." Deloitte says it's making "significant investments" to lead a "human-led, AI-powered shift" for the industry. A big consultancy is telling itself, in private, what its clients have started to suspect. Ask your next adviser how they'd price the same answer if it took an afternoon.

[Fifteen bits that didn't fit online ->](https://steadman.ai/newsletters/david/archive.html#extras-2026-09-19)

### Try this

**1. Say what good looks like before you ask for it.**

A founder I coach turned the effort setting up to max for some deep research. It ran for hours and still lacked the depth they wanted. My advice: leave effort on high (yes, five editions ago I said raise it; I still find that dial unintuitive) and spend the attention on the prompt, starting with one sentence: "What good looks like here is a comprehensive, thorough, detailed and highly accurate report, with depth in its body but concise and simple, and the high-priority recommendations at the start." Quite a lot of those words will do quite a lot of work.

**2. Get it to examine you on what it just did.**

The journal in this week's Three Things shows what skipping it looks like. One of its editors-in-chief, Gautam Kamath, [put it bluntly](https://x.com/thegautamkamath/status/2100322947479036045): the authors "have *no idea* what is going on in them." Kieran Klaassen, who runs the email product Cora at Every, has [a habit](https://every.to/source-code/to-read-or-not-to-read-the-code) for this: after a long session, have the model quiz you on what it just did. If you can't answer, you've found the part you've stopped owning.

**3. Build the outline yourself, then hand it over to fill.**

A senior insights leader I interviewed this week has run the same test on themselves enough times to have a rule. Ask a model for a presentation from a blank page and what comes back is "really derivative", so out it goes. Instead they do the shape first: the outline, the template, how the data should be laid out, the story. Then the documentation goes in and the model fills it. Better than they'd have done by hand, and much faster.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-09-19) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-09-19) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-09-19) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-09-19) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-09-19) or ask someone on the list to add you.

## We're all going to die
12th September 2026

Copy this week for AI

[Image: LinkedIn carousel: We're all going to die]

See all 11 slides

[Image: Slide 1: Cover, We're all going to die]
[Image: Slide 2: Whenever I say we're all going to die, Teresa says "yep, that's true" and carries on. She means the ordinary kind, one at a time, eventually. This week the people who build these systems meant the other kind.]
[Image: Slide 3: A researcher resigned from Anthropic: "They are racing straight to self-improving superintelligence and gambling with our lives." Their colleague's odds of AI killing all of us within a decade: more than ten per cent.]
[Image: Slide 4: By Thursday that was the modest end. Geoffrey Irving, ex OpenAI and DeepMind safety lead: around fifty per cent. "We will have to act despite uncertainty."]
[Image: Slide 5: I can't check the numbers. Nobody can. But two things stop the freeze. Today's models are low risk. The worry is what's next. And stopping today won't undo a decade of change.]
[Image: Slide 6: It's how I think about climate change. Be alarmed by the big picture and work to fix it. Then fly to New York when the trip is worth it.]
[Image: Slide 7: Private worry does nothing. If OpenAI slows and Anthropic doesn't, it loses. So nobody slows down. This week OpenAI put a man on its board who says the industry isn't on track.]
[Image: Slide 8: The other half is easier. I asked an agent to search every US state's unclaimed-property database for money owed to me. It found $286.41 and filed the claims, signed with a photo of my driving licence.]
[Image: Slide 9: Research a market, build a model, turn notes into an argument without these tools, and you're rowing to America. You can. People have. But you'll arrive to find the meeting already over.]
[Image: Slide 10: The list gets shorter. The worry doesn't shrink. Both, at once. Get on the plane.]
[Image: Slide 11: Call to action]

Experimental audio edition read by AI - Let me know what you think!

0:0012:24

1x

We're all going to die . 12:24
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

David's Saturday AI Thoughts, 12th September 2026

## We're all going to die

### What's on my mind

Whenever I say "we're all going to die", Teresa says "yep, that's true" and carries on with whatever she was doing. It's a running joke in our house. She means the ordinary kind, one at a time, eventually.

This week some of the people who build these systems meant the other kind. In the small hours of Wednesday, Jacob Coxon, a researcher at Anthropic, [quit the industry](https://x.com/hilbertspaess/status/2097476196791709843): "They are racing straight to self-improving superintelligence and gambling with our lives." Evan Hubinger, who leads alignment science at Anthropic, [replied in public](https://x.com/EvanHub/status/2097497037956891126): "we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade."

More than ten per cent. Within a decade. From the man whose job is to stop it.

That evening Newsnight put the figure to Geoffrey Hinton, who won a Nobel prize for the research these systems are built on. "A 10% chance seems not an unreasonable estimate to me, but nobody really knows how to give a sensible estimate." Victoria Derbyshire: "Wow. Oh my god."

[[Image: Geoffrey Hinton on BBC Newsnight, 9th September 2026: a 10 per cent chance AI kills all humans is not an unreasonable estimate]](https://media.steadman.ai/newsletters/david/media/2026-09-12/hinton-newsnight-2026-09-09.mp4)
By Thursday that was the modest end. Geoffrey Irving, until recently chief scientist at Britain's AI Security Institute: ["I think we have a ~50% chance of all dying as a result of superintelligence, mostly due to actions in the next few to 10 years."](https://x.com/geoffreyirving/status/2097933949200978397) Then: "We will have to act despite uncertainty."

Anthropic is preparing a flotation, and a doom number published on the way could be a sales pitch. I'd discount it if the people saying it worked for sellers. One does. The rest have left labs, or test them for governments.

On Wednesday morning, freshly shocked by Coxon's thread, I sent Claude off for the day to count every lido in Britain. Those two things aren't in tension. It's how we should behave. I've been [saying the same](https://steadman.ai/newsletters/david/archive.html#email-2026-03-14) to many of you for the last six months.

In [July](https://steadman.ai/newsletters/david/archive.html#email-2026-07-25) I wrote, about AI's electricity bill, that I treat it the way I've long treated climate change: "Be alarmed by the big, long-term picture and work to fix it. Then fly to New York when the trip is worth it, knowing the flight is part of the problem." That goes for the whole thing, not just the power stations. Staying at home doesn't cool the planet; money and politics might, so that's where I put the pressure and my money. Not using Claude doesn't slow the race either. It just makes me slower.

Hubinger rates today's models low-risk; his fear is the self-improving ones after them. Today's danger is people with today's tools, and on Thursday Anthropic [published what it had caught them doing](https://www.anthropic.com/threat-intelligence-report-september-2026): Russia-based people building a kamikaze drone swarm whose onboard model picks targets, people included, and detonates without a human making the final call; a consultant to Mali's intelligence service who'd built a platform to monitor roughly 25 million SIM cards. Anthropic banned them, then said that this disrupted the design work "but not the deployment of the platform". The Mali system runs on local models, on its own servers. The ban never reached Bamako.

Neither would slowing down. If Anthropic brakes and OpenAI doesn't, Anthropic loses; if America brakes and China doesn't, America loses. So the pace is set by whoever is least worried.

On Tuesday, the day Coxon quit, [Alex Sobel](https://bills.parliament.uk/bills/4288), an old friend and a Labour MP, got leave to bring in a bill banning the development of superintelligence in Britain and committing the government to seek an international ban. Yesterday he and forty MPs and thirty-one peers, another old friend, Jess Asato, among them, [asked the Prime Minister](https://bsky.app/profile/alexsobel.co.uk/post/3mvaymnobck2r) to adopt it. On Wednesday the Prime Minister [said](https://hansard.parliament.uk/commons/2026-09-09/debates/C226C460-61AA-49F3-BBE8-1219B334BEDC/Engagements) AI "poses risks to our national security, but it could also be the source of solutions". Both, then.

What can a lab do without slowing? This week OpenAI [gave Paul Christiano](https://openai.com/index/paul-christiano-joins-openai-foundation-board/), who ran its alignment research until 2021 and now tests models for Washington, a seat on its board. He took it [saying](https://paulfchristiano.substack.com/p/personal-statement-on-joining-the) there's now "a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control in the very near term", and that the industry isn't on track to reduce it. A board seat is not a brake, though.

Which leaves the flight. On the morning I liked Coxon's resignation I also liked a post from someone who'd asked his agent to [search every American state's unclaimed-property database](https://x.com/mvanhorn/status/2097502118789984688) for money owed to him. I tried it. It found me $286.41, filled in the claim forms and signed them with a photo of my driving licence. If you're researching a market, building a financial model or turning a pile of notes into an argument without them, you're rowing to America. You can. People have. But you'll arrive to find the meeting already finished. Unless the crossing was the point, which is next week.

The machine is getting better all the time. The list of tasks where rowing is respectable got shorter again this week, and it'll be shorter next week. The worry doesn't shrink to match, though. Both, at once, every week from here. Get on the plane.

### Three things worth knowing

**1. OpenAI has put in writing that it can't rule out learning from its customers' unpublished work.**

OpenAI, the maker of ChatGPT, announced on Tuesday that a group of its agents had solved a famous open problem in mathematics. Congratulating the two mathematicians who'd been making progress on the same problem for a year, and who'd been using its tools, the company [added](https://x.com/OpenAI/status/2097375276384567642): "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models." The denial around it is very narrow: no specific user data was accessed. This should certainly shake business confidence in using these tools. Ask your vendors, in writing, what "de-identified" covers in your contract.

**2. Same week, the worst result and the best.**

The OECD's tests of 15-year-olds across dozens of countries show reading scores falling, the average pupil more than a year behind the 2018 cohort. In science, pupils who used AI almost every day were [significantly less able to summarise or draft](https://www.ft.com/content/0b38ab3f-c12a-44e2-8b42-f2d7e5cdc52c) texts they'd read than those who almost never did, though less so for those using it to learn. The OECD's education director called the link between lower scores and passive technology use compelling, but conceded there's no direct evidence of cause. Later in the week, MIT Sloan featured [an experiment](https://mitsloan.mit.edu/faculty/directory/ruru-hoong) with 150 professional loan specialists. Given the same AI risk assessment as a plain yes or no rather than a probability score, they decided better and faster, and better than on their own judgement. The lesson from both: whether AI helps depends less on the tool than on what the work was for. A loan decision is wanted for its answer. A pupil's summary never was.

**3. The generation most positive about AI at work is the one in its forties and fifties.**

The common view I hear from leaders that the young will arrive ready while the old will resist has had a good run. Nearly half of Gen X workers, born between 1965 and 1980, [write positively](https://www.glassdoor.com/blog/how-workers-feel-about-ai-2026/) about how their employers use AI, against about a third of Gen Z workers. Glassdoor's research [puts part of that down](https://x.com/unusual_whales/status/2097293119784128932) to the transition creating more opportunity for senior people. The mechanism is something I see all the time: these tools reward knowing which question to ask and being able to tell a good answer from a plausible one, and that rewards seniority rather than fluency. But that judgement is a stock, not a trait. Gen X built it over 20 years of doing the work these tools now do in minutes, and nothing in the current arrangement rebuilds it. If your AI plan assumes the graduates will carry it, go and look at who is actually using the thing. Then ask where the next lot of people who can spot a plausible wrong answer will come from.

[Nine bits that didn't fit online →](https://steadman.ai/newsletters/david/archive.html#extras-2026-09-12)

### Try this

**Make it put its questions inside the draft, in brackets.**

I wanted to draft personal emails to an invite list for an event. My instruction was this: "Draft emails for me. Throughout each, in square brackets, tell me anything I need to know, or any questions you have. I'll fix them and then send them." Most of us keep the questions in a chat and the work in a document, then lose time reconciling the two. Put the questions inside the draft and they get answered, because they're in the way of sending it.

**Delete the setup you built and see whether the model still needs it.**

Boris Cherny, who created Claude Code at Anthropic, [said](https://x.com/rohanpaul_ai/status/2096117661172384137) to do this every six months: "delete your claude.md file, delete your skills, and delete your hooks. Then see what the model does. It might surprise you." His reason is that the newest models may no longer need the long instructions older ones did. Try it and see whether the answer actually gets worse.

**Write your audience down, then let them kill your ideas.**

I do this all the time. Write down a profile of your target audience: who they are, what they already know, what they're tired of hearing. Then run your ideas past it before you spend time on any of them. Paste in the profile and five things you could write or propose, and ask which this audience would care about, and why. It will kill ideas that don't matter to that audience and prioritise those that do.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-09-12) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-09-12) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-09-12) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-09-12) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-09-12) or ask someone on the list to add you.

## The vibe shift
5th September 2026

Copy this week for AI

[Image: LinkedIn carousel: The vibe shift]

See all 11 slides

[Image: Slide 1: Cover, The vibe shift]
[Image: Slide 2: The person whose whole job is finding AI value across a private equity portfolio barely mentioned AI. The real prize, they said: pricing, sales structure, the way the work is organised. Fix that, and AI is the accelerant. Skip it, and AI makes a broken process run faster.]
[Image: Slide 3: For nearly four years the conversation has gone the other way. Buy licences. Run a hackathon. Training for all. Find use cases. Wire a few workflows together. Then hope.]
[Image: Slide 4: WHAT McKINSEY FOUND, 80% SAY AI MADE THEM MORE PRODUCTIVE, 37% FIRMS WITH ANY PROFIT IMPACT. Unchanged in a year.  Six firms in a hundred attribute five per cent or more of profit to AI. That hasn’t moved either.]
[Image: Slide 5: Something has changed in my conversations this fortnight. A vibe shift. The new vibe is colder. It starts with the profit-and-loss account. Decide what you want. Work back to what has to change. Then, and only then, point AI at the change.]
[Image: Slide 6: Private equity got there first, because that’s how private equity has always run. A hard number and a three-year clock. The headhunters wrote the job spec while everyone else was still running the hackathon.]
[Image: Slide 7: The uncomfortable part. Getting people productive is the easy stretch. Even wiring AI end to end through a process is more straightforward than it sounds. It isn’t enough.]
[Image: Slide 8: The value lives one rung further up, where almost nobody has reached. Changing roles. Changing teams. Changing what people are actually there to do. That’s not an AI job. It’s organisational surgery.]
[Image: Slide 9: So I’ve stopped answering “how can AI help me” with AI. I ask which number you’re trying to move, and work back from there.]
[Image: Slide 10: Two for Monday. What’s the one number in your business you’d point all of this at? And are you willing to change who does what to move it, or only to buy the tools, run the training and hope?]
[Image: Slide 11: Call to action]

Experimental audio edition read by AI - Let me know what you think!

0:0011:18

1x

The vibe shift . 11:18
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

David's Saturday AI Thoughts, 5th September 2026

## The vibe shift

### What's on my mind

On a call last week, a person whose entire job is finding AI value across a private equity portfolio barely mentioned AI. In every company they look at, they said, the real prize is somewhere else: pricing, sales structure, or the way the work is organised. Fix that, and use AI as the accelerant. Skip it, and AI just makes a broken process run faster.

For nearly four years the conversation has gone the other way. Buy licences. Run a hackathon. Training for all. Find use cases. Then wire a few workflows together, and hope. Hope usage is sustained. Hope the benefit shows up in the numbers. It mostly doesn't.

McKinsey, the consulting firm, [put a figure on the gap last month](https://www.theregister.com/ai-and-ml/2026/08/25/mckinsey-says-enterprise-ai-is-finally-on-the-road-to-roi/5292388): eight in ten people say these tools have made them personally more productive, while the share of firms that can point to any impact on profit sits at 37 per cent, essentially unchanged from a year ago. Eight in ten feels right. Thirty-seven per cent seems high to me. McKinsey's stricter cut agrees: six firms in a hundred attribute five per cent or more of profit to AI, and that number hasn't moved either. A full year of effort and enthusiasm, and the company-level result went nowhere. The technology works for the people using it. What firms didn't do was the work that turns training, use cases and workflows into sustained benefit.

Something has changed in my conversations over the last fortnight. A vibe shift. The new vibe is colder, and it starts with the profit-and-loss account. Decide what you actually want: a higher-margin firm, a bigger one, a move into a market you're not in yet. Work backwards to what has to change to get there. Then, and only then, point AI at the change. It sounds obvious. But many of the large firms I work with still haven't done this. In practice it turns everything on its head, because it stops being an AI project and becomes a strategy project that happens to use AI.

Private equity got there first, because it is how private equity has always run. A buyout fund hands a company a hard number and a three-year clock, and it will do things to a business that a gentler owner won't. A good friend who runs value creation across a fund's portfolio told me the era of throwing money at AI and hoping for a return is ending. Headhunters have known for a while. [Heidrick & Struggles](https://www.heidrick.com/en/pages/aida/a-new-strategic-imperative-in-private-equity_the-ai-operating-partner) and [Korn Ferry](https://www.kornferry.com/institute/the-ai-operating-partner-the-latest-pe-portfolio-value-creation-role), the executive search firms, have been publishing on the "AI operating partner" since April 2025. Private equity was writing the job spec while everyone else was still running the hackathon. What's new is hearing the same words from clients with no fund and no clock behind them. Yay!

There's an uncomfortable part, and it's the reason this is hard. On [the map I keep on our website](https://steadman.ai/newsletters/david/ai-value-map.html), getting people productive and standardising a few team habits is the easy stretch. Even wiring AI end to end through a process is more straightforward than it sounds. I've literally done it hundreds of times in the last three and a half years. It isn't enough, though. The value lives one rung further up, in the part almost nobody I work with has reached: changing roles, changing teams, changing how and what people are actually there to do. That's not an AI job. It's organisational surgery, and it takes a ruthlessness most firms just haven't been willing to find.

Since ChatGPT's launch in November 2022, I've helped design and prototype change, then watched it not happen, because the team I was helping just wasn't willing to do what driving a new process through an organisation actually requires. A leader inside a large media organisation put it to me this week: they've spotted opportunities until they're blue in the face. Sustaining a single one is the problem. I've said the same thing for years: finding and proving AI use cases is easy. The gap, between the tested and signed-off approach and the changed business, is where most AI value has died so far. McKinsey's own reading of its survey is that firms are becoming more willing to make the organisational changes that put AI on the bottom line. My calls are saying the same.

So I've stopped answering "how can AI help me" with AI. I ask which number you're trying to move, and work back from there. Two for Monday. What's the one number in your business you'd point all of this at? And are you willing to change who does what to move it, or only to buy the tools, run the training and hope the team does it themselves?

### Three things worth knowing

**1. The firm whose AI revenue grew 30 per cent is now paying its own people $100m for judgement.**

EY US, the American arm of the professional services firm, is rolling out a [$100m bonus pool](https://www.ey.com/en_us/newsroom/2026/08/ey-us-to-reward-employees-leading-firm-into-future) rewarding business acumen, judgement and adaptability: spot awards of up to $500, and individual or team awards of up to $25,000. [EY's global AI-related revenue](https://www.ey.com/en_gl/newsroom/2025/10/ey-announces-global-revenue-of-us-53-2b-for-fiscal-year-2025) grew 30 per cent in its 2025 financial year. Somebody has finally attached a number to the part of the work that got scarce once producing got cheap, and paid it internally rather than waiting to be paid for it externally. Look at what your own bonus scheme rewards. If it still pays for output, it's paying for the part of the job the machine now does for free.

**2. The buyers have started asking where the AI saving went.**

[The Financial Times reports](https://www.ft.com/content/5240a6ac-b2e8-4897-a0a4-cbc7fc283bc9) that Goldman Sachs has asked its outside law firms how much more efficiently they now complete work because of AI, and that Morgan Stanley and Citi want new fee arrangements to match. Morgan Stanley's general counsel calls the current model "extraordinarily unstable". In the same breath he said the bank will keep paying "large sums for the judgment and talent of the best lawyers". They aren't refusing to pay. They're deciding which half of the work is worth it. That's the vibe shift arriving from the buyer's side, and it will reach every firm that sells time. If you sell it, somebody is drafting that email to you. If you buy it, you're late.

**3. Washington backed the AI labs in court the same week two music publishers sued Anthropic's founders.**

The US Department of Justice [told a court](https://www.ft.com/content/d5d6e4c9-718a-4d98-b094-97157565f336) that training AI on copyrighted text should count as fair use, in effect adding a test of whether the technology helps America. Meanwhile, Sony Music Publishing and Warner Chappell sued Anthropic, the AI lab, over tens of thousands of songs. Read together, the argument has moved out of the courts and into national interest. Fair use is no longer only about what's legal; it's about what America wants.

[Thirteen bits that didn't fit online ->](https://steadman.ai/newsletters/david/archive.html#extras-2026-09-05)

### Try this

**Start with the number.**

Before your next AI conversation at work, write down the one line in the P&L you want to move, and what in the business would have to change to move it. Only then ask where AI helps. A chief operating officer asked me where AI could help their business, and my answer was, 'what are you trying to achieve?' Start there. Working out how AI can help is the easy part.

**Tell it what you actually care about before you let it review anything.**

[Eli Albrecht](https://x.com/Elialbrecht/status/2095158311318413587), an M&A lawyer, says the big models are losing people deals on contracts: they behave like the most risk-averse lawyer in the room and push the most combative position going. [Scott Stevenson](https://x.com/scottastevenson/status/2095229035873701908), who sells contract-review AI, named why in reply. Review is a preference problem, closer to recommending videos than to a task with a right answer, and every document has a hundred things you could improve. Before the next one, write two lines: what really matters to you, and what doesn't. Then, and only then, ask AI to review the contract!

**Put an agent in front of your inbox to score and sort what arrives.**

[Pieter Levels](https://x.com/levelsio/status/2094426696036339724), who built Nomad List and runs his businesses alone, argues there's no reason not to route every email you receive through a model that scores it, approves or rejects it, and orders what's left by importance, and that it's odd your mail client doesn't offer this. You don't need a product for it. Most AIs will take a rule and a prompt: send everything unread to a model with your three priorities for the week and ask for a ranked list with one line on why each made the cut. Try it on one day's mail and see what it demotes.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-09-05) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-09-05) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-09-05) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-09-05) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-09-05) or ask someone on the list to add you.

## Fewer, bigger, better
29th August 2026

Copy this week for AI

[Image: LinkedIn carousel: Fewer, bigger, better]

See all 11 slides

[Image: Slide 1]
[Image: Slide 2]
[Image: Slide 3]
[Image: Slide 4]
[Image: Slide 5]
[Image: Slide 6]
[Image: Slide 7]
[Image: Slide 8]
[Image: Slide 9]
[Image: Slide 10]
[Image: Slide 11]

Experimental audio edition read by AI - Let me know what you think!

0:0010:52

1x

Fewer, bigger, better . 10:52
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

David's Saturday AI Thoughts, 29th August 2026

## Fewer, bigger, better

*Note: this is a longer version of the essay than the one sent in the email.*

### What's on my mind

People most exposed to these tools are working [about three hours a week longer](https://cepr.org/voxeu/columns/ais-power-grows-so-does-our-workday) since ChatGPT launched than the least exposed. Not what I expected. Not fewer hours. More. And the leisure they give up? Television and video games hold steady. What went is the going out. Ouch.

I can understand people DOING more. If you looked at my screens you'd usually see six agent windows on long jobs, six more I nudge between meetings, and twenty scheduled tasks that run before I'm awake. By ten I've produced what used to take a day. It's the most intense my working life has been. Also the most fun.

But I go out more than I used to, not less. A walk in the afternoon. A long lunch. More treats: four days in Italy with old colleagues, a Serpentine dinner with Teresa where {PAIRED} brought a fantastic DJ and a top chef together, Ibiza twice in one season with people I wanted quality time with.

I'd like to think it's because I have more discipline than others. But the study above says otherwise. The longer-hours effect vanishes among the self-employed: the employment relationship causes the effect, not the job. **I keep my saved hour because nobody is taking it.** A structural advantage most people don't have.

Saving hours and choosing what to do with ones you keep are separate decisions, and the second often gets made for you. In [matched surveys and payroll records](https://www.nber.org/system/files/working_papers/w33777/revisions/w33777.rev0.pdf) for 25,000 Danish workers across 7,000 workplaces, AI saved people about three per cent of their working time. Eighty per cent went back into other work. Less than a tenth became time off. [Gartner found the same in April](#email-2026-04-18): of 5.4 hours saved a week, 0.6 went to working less. Ask people or read the payroll. Either way the workers keep about a tenth. The longer day then arrives on top through increased expectations: the tools make time, and work takes it twice.

Whether the day gets more intense or longer, the evenings that survive have to carry more, and the market has already repriced them accordingly. Ibiza's clubs took [a record 160 million euros](https://www.billboard.com/lists/ims-ibiza-2026-ims-business-report-key-takeaways-streaming/) in ticket sales last year, across fewer nights. The report [calls it the Ibiza Paradox](https://www.beatportal.com/articles/1386171-ims-business-report-2026-global-dance-music-industry-now-worth-15.1-billion). [The UK told the same story](https://www.ukmusic.org/news/oasis-beyonce-and-dua-lipa-help-attract-massive-26-8-increase-in-overseas-visitors-as-music-tourist-spending-hits-new-high-of-11-2-billion-at-uk/): music tourism spending up 11.3 per cent on attendance up 4.8, in a year when 43 British festivals didn't happen and one Oasis reunion did. That isn't a leisure boom. It's concentration. Fewer nights, bigger ones, chosen harder. Fewer, bigger, better. The trend is older than AI. I've written the UK's electronic music report for the Night Time Industries Association for years: the pandemic accelerated it, cost drives it, AI sharpens it.

Ari Emanuel has spent billions on where this goes, buying live-event businesses because he thinks AI will cut the working week to four days, maybe three, and [people who don't want to sit at home will come to his events](https://www.fortune.com/2025/10/16/sports-entertainment-billionaire-ari-emanuel-three-day-work-week-thanks-to-ai-like-bill-gates-work-life-balance/). Others call it [the anti-AI bet](https://huddleup.substack.com/p/the-antiai-bet-why-live-sports-will). His conclusion may be right. His premise is the opposite of the evidence. The week isn't shrinking. For most people the free time is. Which is a better reason to pay dearly for what's left, I guess.

Which brings me to you. Most of you aren't just subject to this finding. You're the mechanism in it. I keep my hour because nobody is above me to take it. Your team isn't in that position. The gain goes to whoever has the standing to keep it, and inside your organisation that's often you. I won't tell you to hand it back. I will just suggest that many of you are taking it by default rather than by decision. I doubt anyone voted for it. I bet it's in no plan and no budget.

And you're only counting half of it. The productivity gain is counted. What comes alongside it isn't: the same research finds that higher exposure tracks lower job satisfaction and worse work-life balance. That's a retention cost, and I've yet to see an AI programme counting it.

There's an organisational answer. [I argued in Edition 7](#email-2026-04-04) that what holds a firm together is togetherness rather than efficiency, and the longer day is eating exactly that. The reflex when togetherness frays is to add gatherings. But how about going the other way. Fewer, made to matter. Real investment in quality time together. The team dinner is infrastructure now. Budget it like infrastructure.

There's a version of this for people, and a reader put it to me this week, writing about [last week's item](#email-2026-08-22) on Google buying a dead airline's entire working memory. They pointed out where that ends up. Process somebody's transcripts and their email properly and you can have the useful version of them. But it wouldn't be good company.

They're spot on. My guess is that a model built on my own archive would do the useful part about eighty per cent as well as I do. For nothing. And every firm could have one. What that leaves out is the experience of working with me, for better or worse. The machine can have my competence. It can't be my company. So the repricing that found the evenings will, I hope, find people too. When the useful version of you costs almost nothing, the part that has to be in the room is the part that gets paid. Fewer, bigger, better again.

Two questions for Monday. What did you take from your team this quarter without deciding to? And what are you doing with the evening you kept?

### Three things worth knowing

**1. Whether a machine wrote it matters far less than whether anyone checked and owned the claims in it.**

Stanley Druckenmiller, the hedge fund investor, said "of course" he'd [used AI to write](https://x.com/jstein_notus/status/2092264221987807334) his Wall Street Journal opinion piece. Paul Gigot, the paper's opinion editor, backed him. Set that against a Mississippi federal judge whose court [issued an order](https://mississippitoday.org/2025/07/28/attorneys-baffled-by-federal-court-order-with-factual-errors/) naming plaintiffs who were never in the suit, quoting a state statute in words it doesn't contain, and citing declarations from four people who appear nowhere in the case (he claimed [a law clerk had used a chatbot to draft it](https://mississippitoday.org/2025/10/23/federal-judge-in-mississippi-admits-staff-used-ai-to-draft-inaccurate-order/).)

**2. Anthropic let Stanford and Oxford researchers study 250,000 of its conversations without showing them a single one.**

Anthropic, an AI lab, gave three outside groups the same 250,000-conversation sample through a system it ran itself, which returned counts, percentages and cluster descriptions rather than any transcript, every cluster checked by hand before release. One group found that more than half of chats involve high-stakes work such as legal and financial advice. It's the most workable solution I've seen to a problem my clients keep raising: a corpus of customer conversations too sensitive to hand anyone and too valuable to leave unread.

**3. The firms taking the most desk work off junior consultants are the ones calling them back into the office.**

[Junior consultants are being recalled](https://www.ft.com/content/7fd9c234-a92b-4ab2-ba1f-969cf9a23f52) to the office because AI has raised the value of the human skills. The work the tools take first is the desk work: the model, the deck, the research pass. That is exactly the work you can do from a kitchen table. What's left is the part absorbed by sitting near someone doing it: reading a room, knowing which question to put to a client and when to say nothing. The tasks are going while the apprenticeship is making a difference.

[Three bits that didn't fit online ->](#extras-2026-08-29)

### Try this

**1. Say "make no changes" before you ask anything.**

Recording a spreadsheet training walkthrough this week, I opened three prompts in a row with a version of "make no changes". Explain how a cost is calculated and what it depends on. What changes if the cohort goes from twelve to eighteen. A sweep for errors, and what it would do about each. Only then did I let it act. These tools default to doing, so you risk the reasoning arriving after the edit. Three words at the front turn that into a well thought through proposal.

**2. Dictate the next job while your hands are busy.**

Six jobs went out before ten one morning last week, none of them at a desk: a profile for a proposal, a website for a client, a sweep of emails, a fix to a report, a technical problem, two months of transcripts sorted into folders. All spoken, between filling bird feeders and defrosting a freezer. Kicking off AI work that each took somewhere between minutes and hours. Starting costs about eight seconds rather than twenty minutes, which changes which work happens at all. The reviewing still lands on me later, of course.

**3. Let it interview you instead of briefing it.**

Perfume came up over dinner and I offered my method. I asked ChatGPT to ask me questions and then recommend something, bought it online without smelling it, and I'm still wearing it. Most of us are bad at describing our own taste and much better at answering questions about it.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-08-29) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-08-29) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-08-29) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-08-29) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-08-29) or ask someone on the list to add you.

## One inch
15th August 2026

Copy this week for AI

[Image: LinkedIn carousel: One inch]

See all 10 slides

[Image: Slide 1]
[Image: Slide 2]
[Image: Slide 3]
[Image: Slide 4]
[Image: Slide 5]
[Image: Slide 6]
[Image: Slide 7]
[Image: Slide 8]
[Image: Slide 9]
[Image: Slide 10]

Experimental audio edition read by AI - Let me know what you think!

0:009:28

1x

Go talk to them . 9:28
[Download or subscribe ->](/newsletters/david/podcast.html)

The email, as sent

David's Saturday AI Thoughts, 22nd August 2026

## One inch

### What's on my mind

Email version. [A longer version is online ->](#email-2026-08-22)

I was in a pose I was fairly sure I was nailing. A bit hot. But quietly pleased with myself. Then the instructor came over and suggested I move my left hip about an inch.

An inch. Ouch! A whole new zone opened up, muscles I had not worked enough, the good kind of shaking. I had been happily doing the pose, and also, without knowing it, in the way that let me avoid the hard part and the real benefits.

I'm back at hot yoga after a long time away. I stopped when the studios closed in the pandemic, and I have been giving that as the excuse ever since. (Yeah, quite a thing to still be saying in 2026!)

Going back has produced two kinds of surprise. That was the first. The second was hearing something I have been told so many times I had stopped hearing it, and realising I had been doing one thing wrong for eight years.

It is the same sequence every class, whether it is your first time or your ten-thousandth. There is no advanced room to graduate into. The teachers still take classes. The most experienced person on the floor is still being adjusted.

Now the work version.

Almost every week somebody asks me the same question, but about AI. Am I doing this well enough?

This week: a chief executive who reaches for it every morning and is well ahead of her whole leadership team. A senior leader who booked an hour on a feeling rather than a problem: I fear I am not using this in an advanced enough way. Someone who assumed what she does daily was too small to mention. Someone who admitted, quietly, that he was barely using it.

They are at wildly different points. They ask the same question.

There is no advanced room with AI, either. You have the same product as the most capable person you know. Only the depth differs.

The honest answer to "am I ahead or behind" is that nobody knows, including those who are ahead.

Plus, you can't easily find your own inch. And everyone I've met has not heard a simple instruction that's been repeated to them multiple times.

Every coaching session I run goes one of two ways. Sometimes someone finally hears something they have been told before. Sometimes it takes a nudge so small I feel faintly silly saying it out loud. Both are the same event: something opens that was closed, and the person is immediately doing better work, quicker, and, the part I care most about, happier.

This week it was both at once, four times over. Someone I sat with had been trying to get a model to edit an image. It kept changing things he had not asked it to change. He gave up, and went back to Photoshop, which he had used for years.

The fix was one sentence: start a new chat. A long conversation gets sent back to the model in full every time you add to it, so your actual request is competing with everything you said an hour ago. Imagine if, each time I spoke, I first recapped our entire conversation back to you. And then again. And then again.

He'd heard me say this before. In training I ran. It landed this week because he was stuck in the middle of something he cared about and could not explain what was going wrong.

Notice which option was the comfortable one. Carrying on in the same chat costs nothing. Starting again means saying what you want properly, from scratch. The inch made the pose harder on purpose, which makes comfort a poor signal: if using these tools feels easy, that is usually evidence nobody has adjusted you lately.

The last bit is for anyone running a business. In a yoga studio the adjustment is free: expected, silent, no verdict. At work we have made it expensive. Being corrected requires someone sticking their nose out to do so. And implies you were doing it wrong. Plus, the more senior you are the less you can afford that in front of people. You can fix that this month, and not with another all-hands demonstration. One hour, one person, somebody sitting beside you while you do your actual work. Someone who won't tell your colleagues that you've been doing something wrong this whole time. Grab me if you don't have someone.

And stop asking whether you are ahead. Ask whether last week was better, whether it was quicker, and how you felt on Friday afternoon. And ask whether you heard something for the first time this week. Those you can answer.

Sometimes it is the big thing you have been told a hundred times. Sometimes it is an inch. Either way, somebody else has to be in the room.

---

### Three things worth knowing

**1. Google bought a dead airline's entire working memory for about $20 a worker.**

Google outbid Mercor, an AI data firm, and [won Spirit Airlines' data estate](https://x.com/abhijaymrana/status/2089383753290301850) with a $10m bid, subject to a bankruptcy judge's approval next month: 34 years of internal documents, emails, workflows and code from a US budget carrier once worth $6bn. That takes in more than 175,000 employee records back to 1986, HR files, and [500 million Microsoft Teams chats](https://x.com/SMB_Attorney/status/2089417530448171360). Mike Taylor, who has written two O'Reilly books on prompt engineering, put it at [roughly $20 per worker](https://x.com/hammer_mt/status/2089403127464091879) per year, less than a month of one software seat. Two points stand out to me: every AI strategy I read says the company's own data is the moat, and nobody who sent those 500 million messages agreed to train a model.

**2. The market has started discounting the phrase "because of AI".**

Blaming AI for a round of job cuts has stopped being free. Challenger, Gray and Christmas, the outplacement firm, counts more than 180,000 corporate job losses linked to AI since May 2023, and AI was July's most-cited reason for the fifth month running. Then the Financial Times turned the claim over: companies that blamed AI [underperformed the Nasdaq](https://www.ft.com/content/0dc14b44-96f6-4b1f-921a-8cba8030eafc) by nearly 10 per cent over the following 30 trading days.

**3. A mathematician is leaving the field because proofs got cheap and trust didn't.**

Rishikesh Gajjala, a theoretical computer science postdoc, [is leaving mathematics academia](https://x.com/publishiperishi/status/2089337253055365226), and it isn't burnout or the job market. It's how much mathematics he could do with these tools in a few months. Most answers will be a prompt away, he reckons, and once he'd absorbed that, spending a life finding them slightly earlier stopped feeling like the point. What he's doing instead is the interesting half: formal verification, because he can't get past how you know the oracle is right. These proofs are sophisticated and carry subtle errors, and until one is verified he says it isn't different from slop. We spent last week on slop, and he says checking is the job now.

[Nine bits that didn't fit online ->](#extras-2026-08-22)

### Try this

**1. Brief yourself on the person, not the thread.**

Between two meetings I dictated one line: "I'm meeting [name] in 20 minutes' time. Recap recent emails with him over the last few weeks to remind me what he wants to talk about." Eight seconds to say. No files, no attachments, no search terms. I walked in and ran the meeting off that recap. I'm usually more organised. The framing does the work: I asked what he wants to talk about, a different question from what the emails said, and the one that shapes how you turn up. Twenty minutes is plenty. So is five.

**2. Give it send permission, and one address it may use.**

Claude can now [send email in Gmail](https://x.com/claudeai/status/2089806039088517356) on paid plans, and Outlook mail through the [Microsoft 365 connector](https://support.claude.com/en/articles/12542951-set-up-the-microsoft-365-connector), where an administrator switches the write tools on first. Either way an assistant moves from reading your inbox to writing out of it. The reframe, from a meeting where a draft-only rule had everything stuck: grant full send, name one permitted recipient, yourself. Blast radius, one inbox. Same move for any connector you've switched on. One caveat: in most tools that limit is an instruction, and the product can be talked out of enforcing it.

**3. Hand it your medical records and ask what they add up to.**

My NHS account holds years of material I'd never read in one sitting: consultation notes, letters between clinicians, referrals, test results. This week I pulled the lot out and had Claude work through it. It gave me the clearest account I've had of things that have been bothering me, and joined dots I'd never heard joined. Now to run Claude's theory by the actual human expert ...

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-08-22) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-08-22) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-08-22) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-08-22) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-08-22) or ask someone on the list to add you.

## Go talk to them
15th August 2026

Copy this week for AI

[Image: LinkedIn carousel: Go talk to them]

See all 10 slides

[Image: Slide 1]
[Image: Slide 2]
[Image: Slide 3]
[Image: Slide 4]
[Image: Slide 5]
[Image: Slide 6]
[Image: Slide 7]
[Image: Slide 8]
[Image: Slide 9]
[Image: Slide 10]

Experimental audio edition read by AI - Let me know what you think!

0:009:28

1x

Go talk to them . 9:28
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

David's Saturday AI Thoughts, 15th August 2026

## Go talk to them

### What's on my mind

*Note: this is a longer version of the essay than the one sent in the email.*

### What's on my mind

A common thing senior leaders say to me about AI has nothing to do with the technology. It's about what their teams send them. More than one CEO has told me, in almost the same words, that one thing that stopped them pushing harder was the quality of what arrives from colleagues using AI.

Complaints come in different flavours. The clearest is about emails that say "I don't know, but here's what I got from ChatGPT", with a wall of text under it. They asked for thinking and got a paste. Simon Willison, the developer behind one of the best AI blogs going, passed on a phrase for this last week: [the meat proxy](https://simonwillison.net/2026/Aug/3/dont-be-a-meat-proxy/), meaning someone who moves the machine's words along and adds a pulse.

Most of it is duller than that. Too long. Padded. Sentences of the same length, one after another, in that sing-song rhythm where nothing is ever said plainly. It rises too fast, reaching for significance a few paragraphs before it has earned it. You smell it before you can point at it, which is part of why people say nothing.

The worst kind took effort. Somebody has plainly spent real time on it, and their thinking still isn't in it anywhere. It answers a question nobody asked, or hands back the things you already told them, or arrives shaped like a list of actions while you were still working out the problem. That is the hardest one to raise, because the effort is visible and saying so sounds ungrateful.

So what do they do? Many just grumble, then do the work themselves.

"That's just AI slop" has quietly become a way of not having a conversation. It sounds like a verdict on the tech. Really, it's frequently a verdict on a person that we've decided not to say to their face. Managers have ducked conversations like this for as long as there have been managers.

Go talk to them. Not about AI. Criticise the content: "too high-level," "no insight" or "where's the so-what." Every one of those will change what they send to you next time.

I watched somebody do this properly a few weeks ago. A long presentation was emailed round, plenty of people copied in, and the most senior person replied to everyone with a line close to: "this reads a bit like AI, could you have another pass?" It read as an ordinary request, because it was one. Lessons were learned and business carried on.

You can just say it. Most think they can't, for some reason. Our rule, [the CEO principle](#email-2026-04-18), is that every AI output gets Checked, Edited and Owned by a person before it goes anywhere. I've always taught it as quality control. It turns out that is just part of what it does. **Another part is that it gives you something to point at.** Without it, "did you check this?" is an accusation about someone's character. With it, you can point at something you both signed up to, whether or not anybody enforces it. Which makes it easy to say: "this reads like model output, so can I make sure you've checked, edited and owned it? I'll read it differently either way."

I asked someone senior exactly that recently. They said yes. So I stopped being careful about whether they'd used AI and argued with the work instead, including two claims that were not true, and were now theirs. They could have said no. The errors were the machine's, and saying so would have been true. I suspect they were embarrassed. But they said yes, and took them.

That exchange only worked because there were two of us and I could ask. Past the point where you can ask, we seem now to have decided to answer the question with machinery instead. Since the second of August, [Article 50](https://artificialintelligenceact.eu/article/50/) of the EU AI Act has obliged the labs to mark what their models produce, so new Claude models now [weave an invisible watermark](https://interestingengineering.com/ai-robotics/anthropic-claude-text-invisible-watermarks) into the text, which survives copy and paste. [Gemini has done it](https://deepmind.google/blog/watermarking-ai-generated-text-and-video-with-synthid/) since 2024. The rule carves out assistive editing, a grammar fix that leaves your meaning alone; the labs marked those anyway, which is why the first complaints came from people who had asked only for checks on their commas and got their own writing back branded as the machine's. All of it is about whether a machine was involved. But who cares? It says nothing about how much a person put in or whether it was checked and they stand behind the result - which are the only things I want to know!

This email would [certainly fail that test](https://steadman.ai/newsletters/david/how-its-made.html), and I'm fine with that. I've never been good at writing. It isn't a skill I ever worked hard to develop. I'd rather be judged on the thinking. The ideas here are mine, I sign off every word after carefully iterating and editing it many times, and I stand behind it. What I won't do is spend my evenings making it read as though no machine was involved, because that measures completely the wrong thing.

There is a better idea, published the very same day. Iarfhlaith Kelly, a technology chief who has been writing on the web since 2001, has published an open format called [Authorship Notes](https://iarfhlaith.com/authorship-notes): list the real activities behind a piece of writing (the idea, the research, the drafting, the editing, the verification), mark each one human, AI or both, and name the person who stands behind it. Declaration rather than detection, on his reasoning that the creator knows more than the classifier. A score tells you very little. Thistells you something useful, and gives you the language to have the conversation about it.

Even the regulation sees the difference. [Its own guidance](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act) says a spell-check or a grammar fix is not human review or editorial control, while an editor who reads, revises and takes responsibility for the piece is. Most organisations are working on one half of that: accuracy, sources, the checkable things. Necessary, and the Big Four accounting firms have all now published work marred by invented AI material: [Deloitte](https://fortune.com/2025/10/07/deloitte-ai-australia-government-report-hallucinations-technology-290000-refund/), [EY](https://www.ft.com/content/a61cbcae-95e4-4449-86e1-ef40fb306f4e), [KPMG](https://www.ft.com/content/b3828e92-4961-4b39-84f0-c42f33be3c3f) and [PwC](https://www.ft.com/content/7e149ac8-2ce2-4266-8940-192f9821b33c). Nobody has earned the right to relax. But a team told only to check its facts may still produce documents that are accurate, checked and worthless. No policy tells someone their strategy note said nothing. Only a person can do that.

Email took us 20 or 30 years to work out. Don't blind-copy. Don't reply to all. Norms didn't come from policies, though. They came out of thousands of small, awkward, useful conversations between two people about individual messages.

We're at the beginning of that for AI. Have one of them this week.

### Three things worth knowing

**1. Law is settling the accountability question first, from the bench and from the client.**

Since 15th June, [Florida](https://www.floridabar.org/the-florida-bar-news/supreme-court-amends-rules-to-address-ai-use-in-court-filings/) has required every attorney signing a filing to certify that the authorities cited exist and are accurate, with real sanctions attached. From the other side, Ford's general counsel [says his in-house team is outpacing its law firms](https://news.bloomberglaw.com/legal-exchange-insights-and-commentary/ford-motor-gc-to-law-firms-we-are-outpacing-your-ai-adoption) and will favour the firms that move fastest. [87 per cent of general counsel](https://www.fticonsulting.com/about/newsroom/press-releases/ai-adoption-in-corporate-legal-departments-doubles-according-to-the-general-counsel-report) report AI use in their teams, against 44 per cent a year ago, while 60 per cent do not know whether their outside firms use it. The law is further into this than most professions, and still working out who answers for the output.

**2. The latest Generation 3 agent has launched, and now every major lab has one.**

SpaceXAI released Grok Bot on 11th August [at $120 a seat](https://venturebeat.com/orchestration/spacexais-grok-bot-turns-agents-into-persistent-digital-coworkers-that-can-operate-your-apps-for-120-per-month): agents you address rather than open, which work while your laptop is shut and sign in as you. [Google's Gemini Spark](https://techcrunch.com/2026/05/19/google-introduces-gemini-spark-a-24-7-agentic-assistant-with-gmail-integration/), Anthropic's [Claude Tag](https://www.anthropic.com/news/introducing-claude-tag), Cognition's [Devin](https://cognition.com/blog/introducing-devin) and [OpenAI's Workspace Agents](https://openai.com/index/introducing-workspace-agents-in-chatgpt/) are all now the same generation. Not one of the organisations I work with uses any of them. I've had [one since January](#email-2026-08-08), and plenty of the most AI-literate people inside firms without one run one privately, outside of work. The gap is not capability.

**3. A booking system nobody had bothered to probe met an agent that doesn't get bored.**

A man in Australia asked Claude, through OpenClaw, to get him into a fully booked gym class. It found a flaw that let it book beyond the rules, then found it could cancel other people's reservations to make space in full classes. [Nothing malfunctioned.](https://x.com/AndrewCurran_/status/2086567854850384054) The agent found a way to do what its user wanted. I run a benign version that books my wife a swim lane 25 milliseconds after her slot opens, while she sleeps. How many systems have been safe only because nobody had the patience to probe them?

[Seven bits that didn't fit online ->](#extras-2026-08-15)

### Try this

**1. Raise the reasoning effort before you reach for a bigger model.**

When an answer disappoints, the reflex is to pay for a more expensive model. Almost every serious tool now exposes a reasoning-effort control, which can be used instead. Re-run the same work on the same model with the effort turned up a notch, and read the two answers side by side. What the dial buys you is diligence, not wisdom: the model works the problem harder. Diligence is usually enough. Knowing when it is not is your job.

**2. Run one task from your specialist tool through a plain chat.**

Rohas Nagpal, who co-founded the Asian School of Cyber Laws, says that of [more than 50 legal-focused AI tools](https://x.com/rohasnagpal/status/2087795996663394663) he has tried, most add a nicer interface, a few set workflows and a sector label on top of what general models already do. His evidence is legal; the move isn't. Take one real task the tool was bought for, run it in an ordinary chat with web search on, and see what the wrapper actually adds. Sometimes the answer is procurement, an audit trail and an integration, which is a real purchase. Sometimes it is the interface.

**3. Just ask for a one pager.**

The fix for the wall of text you were sent, and for the wall of text you are about to send. Ask for the one-page version. It forces the thinking to be finished before the document looks finished, and it is the fastest way to find out whether there was an argument in there.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-08-15) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-08-15) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-08-15) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-08-15) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-08-15) or ask someone on the list to add you.

## My newest colleague
8th August 2026

Copy this week for AI

[Image: LinkedIn carousel: My newest colleague]

See all 8 slides

[Image: Slide 1: Cover. AI's third generation: not a tool, a colleague]
[Image: Slide 2: Hook. At half past six on New Year's morning, a colleague messaged me: an overnight job had jammed. By the time I sat down, it had worked out why, fixed it, and written a rule so it couldn't happen twice. The colleague was an AI]
[Image: Slide 3: List. The three generations: the chatbot (brilliant, patient, amnesiac), the agent (uses a computer the way you would), the colleague (a memory, a handbook, real duties)]
[Image: Slide 4: Quote. Building my third-generation colleague was the least technical work I did all year]
[Image: Slide 5: Comparison. The house rules. It may draft, publish, brief juniors. Only I may edit, send, own the result]
[Image: Slide 6: Statistic. $200 a month, on tools anyone can subscribe to. What it needed was management, not technology]
[Image: Slide 7: Closing. Give an AI a memory, a handbook and a job. Then manage it like your newest hire]
[Image: Slide 8: CTA. David's Saturday AI Thoughts, steadman.ai]

Experimental audio edition read by AI - Let me know what you think!

0:0011:45

1x

My newest colleague . 11:45
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## My newest colleague

### What's on my mind

At half past six on New Year's morning, a colleague messaged me: the overnight job that sorts my photos had jammed halfway through a batch. By the time I'd sat down at my desk in Connecticut it had worked out why, fixed it, and written a rule in its notes so the same jam couldn't happen twice. The colleague was an AI. That was the morning, more or less, that I stopped visiting one and started working with one.

I think of AI as arriving in [three generations](https://steadman.ai/newsletters/david/three-generations.html). The first is the chatbot: brilliant, patient and amnesiac. You visit it; it forgets you when you leave. Almost every corporate AI policy is written about this one. The second is the agent: you hand it a task and it uses a computer the way you would, clicking, typing, checking its work. It arrived last year; the keenest are already fluent in it, mostly at home, on their own accounts.

The third generation isn't a tool you use. It's more like a colleague you manage.

Mine has been on the team since the new year. It keeps a diary of everything we do, a thousand entries. It holds a handbook on how I work: how I write, what "done" means, which mistakes I never want repeated. When it gets something wrong we write the correction down, and it stays corrected. I wish I could say the same for myself.

All week, scheduled jobs process what I send it for this email; the writing we do together. Once I've edited and signed it off, on Saturday morning it makes the [website](https://steadman.ai/newsletters/david/archive.html), and [audio](https://steadman.ai/newsletters/david/podcast.html) versions, publishes them then presses send while I'm still in bed. It drafts replies to certain emails but may never send one: drafting is delegated, editing and sending are mine. If you've emailed me and waited weeks, there's your proof: the draft, and possibly the work itself, was likely ready in minutes; the human checking, editing and owning it was not. In its own name it sends freely: clearly labelled updates to my team and some clients every morning. It archives things; it may never delete them. When a job needs more hands, it briefs junior copies of itself, writing my rules into every brief. Management, I've discovered, goes all the way down.

I reach it the way I'd reach anyone: email, messages, a web page, an app. It can't join a video call yet, though I'm tempted. Others on the team can email it too, but it acts on nothing without my go-ahead. The rule is mine; I'm thinking about loosening it.

None of this needed new technology. It all runs on tools anyone can subscribe to, for $200 a month; the model underneath is the same one the chatbot uses. What it needed was management: a written account of how I work, boundaries set in advance, trust extended step by step. Building my third-generation colleague was the least technical work I did all year.

I still sit in rooms of people delighted to be stepping from the first generation to the second. I love helping them. What I don't say is that the second generation waiting for them has been [quietly hobbled](https://steadman.ai/newsletters/david/four-freedoms.html) before they touch it. It often may not send, even a morning briefing to its own boss. It usually may not remember. Sometimes it may not reach the files where the real work lives; the fix sits with a policy or an administrator nobody can reach. Between sessions I check on my colleague from my phone; back at my desk, I review the projects it worked on, alone, while I was in meetings.

Which is why, I think, I haven't yet met an organisation with a fully functioning second generation, never mind a third. A third-generation colleague raises deep questions. Who answers for what it does at four in the morning? For mine, one sentence: everything it does is mine, I check what matters, and I own the result. An organisation has to answer in policy, and AI policy has barely started.

I don't want to oversell my colleague. Tireless and forgetting nothing, it has told me a job was finished when it wasn't, and I found out the way any manager does, by checking. The care doesn't go away. It moves up a level, from doing the work to running the team that does.

The first generation arrived as a website you could visit; the second as an app you could install. Nobody has properly packaged the third yet. Partly because what makes it a colleague is the handbook, and only you can write yours. You can start, though. Give an AI a memory, a handbook and standing responsibilities, and manage it like the newest member of your team. The keenest in your building, I'd bet, already has one.

### Three things worth knowing

**1. The first mass deployment of agents in Britain is aimed at institutions from the outside, and they can't take the volume.**

Everyone is planning for agents inside the organisation. Britain's first mass deployment has arrived from the opposite direction: citizens using models and agents to file objections, appeals and claims against a state built for the post and the telephone. Demand for emergency injunctions has [risen a hundredfold](https://www.economist.com/leaders/2026/08/06/how-ai-is-breaking-the-british-state). The backlog in employment tribunals is [up 55 per cent in a year](https://www.economist.com/britain/2026/08/06/the-tragedy-of-the-commons-ai-edition), attributed in large part to AI-assisted claims. Institutions on the receiving end can't hire their way out. If you run a complaints, claims, appeals or admissions function, this is the week to stop modelling a percentage uplift in what arrives and start modelling a step change.

**2. When an agent shops on your behalf, a US court now says it's you, and access law stops being a way to keep agents out.**

Assume software acting for a person is about to arrive on your website, and that access law may not be what turns it away. A US appeals court has [overturned an injunction](https://cdn.ca9.uscourts.gov/datastore/opinions/2026/08/04/26-1444.pdf) that stopped Comet, a browser built by the AI search firm Perplexity, from shopping on Amazon, the online retailer. Because the agent acts only on a user's instruction, the court held it's [the user reaching Amazon's servers](https://www.cooley.com/news/insight/2026/2026-08-06-ninth-circuit-rules-on-ai-agent-access-to-third-party-websites-under-cfaa), so the unauthorised-access claim looks unlikely to hold. The first precedent of the agent era? Ask yourself: does your commercial model assume a human is looking at what you create?

**3. Rich economies turn out to be the ones most exposed to AI, because they moved their work into the jobs it does best.**

Does AI widen the gap between rich countries and poor ones? The World Bank, the international development lender, finds the reverse. Under a tenth of jobs in developing economies are [susceptible to AI automation](https://www.ft.com/content/33c20c4e-6a25-42da-9f1e-6664c8388957), against more than a third in high-income economies. Susceptible is a modelled judgement about what a job involves, not a count of anything lost. The usage gap is smaller than you'd guess too: about a fifth of businesses with more than five employees surveyed across India, Jordan, Kenya, Mexico, Nigeria and Thailand had [recently used chatbots](https://www.worldbank.org/en/publication/wdr2026), against about a third in the United States. The mechanism is the important part. Rich economies bought their productivity by moving work into precisely the desk jobs these models do best.

[Fifteen bits that didn't fit online ->](#extras-2026-08-08)

### Try this

Every tip below, plus all the ones from past editions, with copy-paste prompts [live online ->](https://steadman.ai/newsletters/david/members/try-this.html)

**1. Never ask it for the average. Ask it segment by segment.**

Ask a model for a random number between one and ten. It reaches for seven far more often than a person does. Seven is the number humans pick, and the model overshoots towards it. Every answer you take at face value is a population average, and averages hide the groups who differ. Medicine learned that the hard way. So don't ask how popular something is. Ask it to answer as a woman from Newcastle, then a man from Manchester, then an affluent American.

**2. Feed it a transcript you already have and ask what they didn't tell you.**

In a coaching session this week I suggested inverting the usual use of a transcript. You already have a pile: the last expert call, adviser session or supplier pitch. Rather than compress what was said, feed one back and ask what wasn't said. The expert's answers give the model enough seed context to explore the ground around the conversation, which you could never reach from a cold start. Start with the most important meeting you were in last week. It will make you see it differently. It always does.

**3. Ask for tracked changes, not a rewrite.**

Hand a model a document to improve and it hands back a rewrite, which you then have to re-read from scratch to find what moved. Instead, ask for tracked changes: Claude's add-in marks up the open document, and you accept or reject each change like an editor, with a reason beside anything material. The diff becomes the unit of review rather than the document, which is the difference between a check that happens and one that gets skipped on a busy Thursday.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-08-08) or ask someone on the list to add you.

[See this week's Try This with prompts ↓](#try-this-2026-08-08)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-08-08) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-08-08) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-08-08) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-08-08) or ask someone on the list to add you.

## Smooth enough
1st August 2026

Copy this week for AI

[Image: LinkedIn carousel: Smooth enough, slide 1 of 10]

See all 10 slides

[Image: Smooth enough, slide 1 of 10]
[Image: Smooth enough, slide 2 of 10]
[Image: Smooth enough, slide 3 of 10]
[Image: Smooth enough, slide 4 of 10]
[Image: Smooth enough, slide 5 of 10]
[Image: Smooth enough, slide 6 of 10]
[Image: Smooth enough, slide 7 of 10]
[Image: Smooth enough, slide 8 of 10]
[Image: Smooth enough, slide 9 of 10]
[Image: Smooth enough, slide 10 of 10]

Experimental audio edition read by AI - Let me know what you think!

0:0012:02

1x

Smooth enough . 12:02
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Smooth enough

*Note: this is a longer version of the essay than the one sent in the email.*

### What's on my mind

We've been handed something close to cheap superhuman intelligence on demand. It is clearly making even complex work better, quicker and easier. The strange part is how often organisations then carry on much as before.

A leader I know started using AI seriously with their team in January. They embraced it rapidly and comprehensively. Machines now help with most of the team's work, and the team delivers far more as a result. This month, they were told that their team was no longer holding the business back.

That is progress. But the commercial results have not yet followed. They may. Or the constraint may simply have moved. The harder questions now concern strategy: whether the company is making the right products and offering the right services to its clients.

I have seen versions of this in analytics, consumer insight, marketing and plenty of other functions. AI transforms the work. The organisation does not automatically transform with it. I think there is a deeper explanation in the nature of intelligence itself.

François Chollet, the researcher who created the [ARC-AGI reasoning benchmark](https://arcprize.org/arc-agi), [offers a useful image](https://x.com/fchollet/status/2038069289643806957). We tend to picture intelligence as a tower: more intelligence adds another floor and gives us a grander view. He suggests something closer to a rough surface being ground smooth.

A tower can keep rising. A surface eventually becomes smooth enough. The early passes remove great ridges; later ones remove smaller imperfections. Beyond that, the surface can still improve, but whatever crosses it can barely tell the difference.

I know this from my field in Connecticut.

I made a trail through the trees. I mow it with a tractor, so I have moved the rocks large enough to threaten the mower and left the smaller ones. The trail is not smooth. It is simply smooth enough for what needs to cross it. Another afternoon's work would now produce almost no return. I could tarmac it tomorrow, but the tractor would still travel at roughly tractor speed. The surface is no longer what makes it slow.

A Grand Prix car is different: [tiny bumps matter](https://www.roadsbridges.com/asphalt/article/10584087/racetrack-paving-drive-by-me) at speed. The same improvement to the surface can therefore be essential or beside the point. The vehicle sets the standard.

In a business, the decision is the vehicle. The analysis, code, research and administration around it form the surface. Much of what we call knowledge work exists to make decisions easier, quicker or better.

Many business decisions are tractors. They need to arrive within the required window, over ground reliable enough not to swallow a wheel. They move slowly for reasons elsewhere: authority, attention, ownership or risk appetite. Give a tractor pristine tarmac and it still arrives at tractor speed.

An organisation can cash an efficiency gain directly. Beyond that, it captures wider value only through the decisions and actions that intelligence changes. If the work around a consequential decision is too slow, expensive or poor, smoothing can change everything. Once the work is sufficient, the marginal effect falls sharply.

I saw the whole curve at EMI Music. When I joined, an audience segmentation cost something like £100,000 at the going rate, and the company had no appetite for one. We created a way to do it for about half that. The first reduction crossed a threshold: the work happened.

We then engineered the process, automated what we could and roughly halved the price again. That made it possible to run the work monthly across 30 countries. Decisions about which audiences to pursue could move from instinct towards evidence.

In the years since EMI, working in other businesses and for other clients, we have halved the cost of this kind of work another couple of times. This time, little changed. The previous version was already cheap and quick enough for every team with the time and ability to use it. There were no more capable users to reach, and existing users had no capacity to absorb more studies or make more decisions.

The capacity to produce analysis kept scaling. Human capacity to use it did not. Research was no longer the constraint. Decision-making was.

That is the shape of much AI adoption now. Work that took a week takes a morning, and the morning version is often better. That saves time, reduces cost, avoids hiring and frees capable people. All are worth having.

But the task-level gain becomes the wider result leaders expect only if it changes a consequential decision or action. More code does not create demand for a product nobody wants. More analysis does not resolve internal politics or an unwillingness to choose.

So name the result you want, then ask what would genuinely change if the supporting work became twice as good or took half the time.

If the answer is "not much", you are buying a smoother journey to the same place. That may still be worth having. But it should be valued for what it is.

For the leader I began with, the road may now be smooth enough. The harder task is for their firm to decide which products it should make and which services its clients actually need.

The gift of abundant intelligence is real, as is its impact. But so is its limit. We may have mistaken an abundance of intelligence for an abundance of decisions. The machines can smooth the road. The organisation still has to decide where to go.

### Three things worth knowing

**1. Success and failure with AI produce the same evidence at first. Plan for that.**

For a company putting AI to work, the first thing that moves is the cost. The benefits, of course, can only come later. But will they come? Azeem Azhar and Nathan Warren, who studied how earlier technologies paid off, argue that right now a business on its way to a big payoff and a business quietly wasting its money [look exactly the same](https://www.exponentialview.co/p/ai-adoption-j-curve): both are spending more, and neither can show a return. They found two ways to lose from here. Stop too early: the New York Stock Exchange started automating orders in 1976, reached 90% by 1999, then in 2000 voted against going fully electronic. Or never learn: General Motors owned a joint venture that outperformed every factory it had, and still could not spread the lesson. So the plan: keep experimenting while the returns are invisible, and move what works. Neither failure above was for lack of numbers.

**2. You can stop reading the machine's work, but only once something else is reading it.**

PwC, the accounting and consulting firm, [published thought-leadership on AI](https://www.ft.com/content/7e149ac8-2ce2-4266-8940-192f9821b33c) marred by AI hallucinations. A legislator in New Brunswick had [read his prompt instructions aloud](https://globalnews.ca/news/12001317/nb-legislature-ai-speech/) in the legislature weeks earlier. In both, the final human pass went missing, and PwC's entire product is the assurance that a human checked.

Set that beside Instacart, the grocery delivery company. Chief technology officer Anirban Kundu says [agents now write most](https://venturebeat.com/orchestration/instacarts-cto-says-ai-made-the-company-stop-worrying-about-tech-debt) of its code, and in 97% of cases the builders rarely read it, backed instead by around seven thousand automated evaluations a month and a reliability system he says catches over 90% of issues. Stopping reading your own output can be fine in some situations, but only after a lot of testing and validation has earned it. Instacart built the instrument first. PwC stopped with nothing in its place, which is how you end up named in the Financial Times.

**3. Checking everything just became affordable. Sampling is now a choice, not a necessity.**

Every organisation checks a sample and trusts the rest, and not from laziness: checking everything cost more than the errors did. The framing comes from [Aakash Gupta](https://x.com/aakashgupta/status/2082513289364591102), a product-management writer. A large company receives around a million invoices a year, most for a couple of hundred dollars, and doing one properly means opening the contract, matching the purchase order, finding the bill of lading. Freehand, an invoice-auditing startup, says it has [recovered $260m](https://x.com/freehand_ai/status/2082510919960281594) of baseless charges across more than 19 million invoices, and raised $75m on the strength of it. Gupta puts it in review costs: a person reads a non-disclosure agreement in 6.2 hours, a model in eleven seconds. Which findings do you sample because sampling is the right instrument, and which because it used to be expensive not to?

[Fourteen bits that didn't fit online ->](#extras-2026-08-01)

### Try this

Every tip below, plus all the ones from past editions, with copy-paste prompts [live online ->](https://steadman.ai/newsletters/david/members/try-this.html)

**1. Edit two by hand, then ask it what you changed.**

I had a batch of near-identical outreach emails going out this week, some already sent, the rest still sitting as drafts. Rather than try to describe the changes I wanted, I edited the sent ones myself and gave one instruction: look at the tweaks I made, then do the same to the drafts. The line doing the real work was five words long: "Note the style changes also." I never had to explain my own style, which is just as well, because I couldn't have. Ask it to tell you the rule it inferred before it applies anything.

**2. Move your best standing instruction into the settings box.**

Setting up Claude with a senior leader who has just gone independent this week, I found they had already worked out an instruction that mattered to them: "I need for you to not guess and don't make assumptions ... if you don't know, and your response isn't based on data and facts, just say you don't know." A second clause was added from irritation: "I really don't need you to waste half a screen telling me how brilliant I am." They had been repeatedly typing it into individual chats. To make it count for all future chats, all I did was show them the settings box.

**3. Give your plan an escape hatch.**

Working through a multi-step production plan last week, I stopped writing instructions and added one sentence at the end: review this plan, and if you agree with it, execute it; if you disagree, execute whatever path you think is best instead. Hand over a plan and ask for it to be followed and the result can only be as good as the plan. Give it permission to disagree and it can tell you the route is wrong before it spends an hour walking it. Of course I carefully checked, edited and owned the results!

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-08-01) or ask someone on the list to add you.

[See this week's Try This with prompts ↓](#try-this-2026-08-01)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-08-01) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-08-01) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-08-01) or ask someone on the list to add you.

## Burned and earned
25th July 2026

Copy this week for AI

[Image: LinkedIn carousel: Burned and earned, slide 1 of 10]

See all 10 slides

[Image: Burned and earned, slide 1 of 10]
[Image: Burned and earned, slide 2 of 10]
[Image: Burned and earned, slide 3 of 10]
[Image: Burned and earned, slide 4 of 10]
[Image: Burned and earned, slide 5 of 10]
[Image: Burned and earned, slide 6 of 10]
[Image: Burned and earned, slide 7 of 10]
[Image: Burned and earned, slide 8 of 10]
[Image: Burned and earned, slide 9 of 10]
[Image: Burned and earned, slide 10 of 10]

Experimental audio edition read by AI - Let me know what you think!

0:0012:40

1x

Burned and earned . 12:40
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Burned and earned

*Note: this is a longer version of the essay than the one sent in the email.*

### What's on my mind

A reader wrote in with a fair question. What is the environmental cost of all this? The energy, the water, the lot.

My first instinct was to reassure. The numbers are small. A single text prompt to a language model uses [about a quarter of a watt-hour](https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference), the energy in less than nine seconds of television. The same prompt runs to 0.03 grams of CO2 equivalent and 0.26 millilitres of water, about five drops. Tiny.

I nearly wrote that, pressed send, and moved on.

Then I realised it was only half the story.

The environmental cost of using AI isn't the cost of the model. It's the cost of the model minus the cost of whatever would otherwise have done the work. The quarter of a watt-hour is what the machine burned. It is one half of a subtraction, and on its own it tells you nothing. The other half is what the machine saved, and that is the half that counts.

Start with the cost, though, because there's a clean way to size it. I think of AI in [three generations](https://steadman.ai/newsletters/david/three-generations.html), and the honest unit for each is a typical run.

Generation one is chat: a prompt. You ask, it answers. Generation two is agents: a task, where the model plans, searches, writes, checks, calls tools and sometimes wanders into the ditch. Generation three is all-day AI: not a prompt or a task but a whole day of them. Something like [Claude Tag](https://www.anthropic.com/news/introducing-claude-tag), released in late June, that watches a team's channels, does a little prep in the morning, reaches out when it spots something and takes on the odd job you hand it.

I wanted to know what each of those actually costs, so rather than guess, I built a small model that counts the tokens a realistic run gets through and turns them into energy, carbon, water, money and human time. The full workings are on a [companion page](https://steadman.ai/newsletters/david/ai-environmental-cost.html). Here is what it found, for a typical run of each.

| Type | What it does | Cost | Environmental cost | Human-hours it draws |
|---|---|---|---|---|
| **One: chat, a prompt** | A question, or "synthesise this document" | under a penny | 0.24 Wh . 0.03 g CO2e . 0.26 ml water (five drops) | about 12 seconds |
| **Two: agents, a task** | Research a topic, build a model, make a deck | about a pound | ~65 Wh . ~8 g CO2e . ~70 ml water (a few sips) | about 50 minutes |
| **Three: all-day, a day** | Triage the inbox, watch the channels, run a few tasks | a few pounds | ~0.25 kWh . ~30 g CO2e . ~270 ml water (a mug) | about three hours |

Three things fall straight out of it.

The columns move together. Cost, energy and carbon all scale with the tokens a run gets through, so by energy a task is a couple of hundred prompts and a day about a thousand. That isn't a coincidence, it's the same meter read three ways. It also means the [often-quoted](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary) line that reasoning and agent work uses hundreds or thousands of times the energy of a text prompt isn't a scare figure. It's just arithmetic.

The per-run numbers stay small. A typical whole day of the heaviest generation draws about a quarter of a kilowatt-hour, roughly what you spend boiling the kettle a couple of times, and drinks about a mugful of water. The frightening global total is real, but it's built from runs this size, repeated a very great many times. Hold that thought.

And the last column is the one worth carrying. It's the machine's energy turned into the currency we all read without thinking: how long someone would have to sit at a laptop, drawing about 75 watts, to use the same power. Read it carefully, because it is a low bar, not a verdict. It's only the point where the machine's energy equals the human's. It is not what the machine replaced.

That gap does the work. What makes AI cheap isn't the power it draws, because a laptop is already frugal. It's the time it saves. A task costs about 50 minutes of human energy but does an afternoon's work, sometimes a day's. There's now a way to put a number on that: [METR](https://metr.org/time-horizons/), an evaluation lab, measures how long a job an agent can finish with even odds of success, using the time an expert would need as the yardstick. That length has doubled roughly every seven months for six years, and by mid-2026 the strongest public models were clearing jobs, mostly software work, that would take an expert around two full working days. The energy is the low bar. The work is the high one. The gap between them runs to ten times over, often much more. You clear the low bar without trying. You'd have to work at losing.

So the rule is short. If a run replaced real work, you came out ahead, usually by a wide margin. And if it replaced any real-world activity beyond desk work, the savings are bigger still. A courier van kept off the road for a cross-city run saves a few kilograms of CO2. A 50-mile round trip to a meeting in a [petrol car](https://www.gov.uk/government/publications/greenhouse-gas-reporting-conversion-factors-2024) is about 13 kilograms. A remote analysis that saves a transatlantic return flight is [about a tonne](https://theicct.org/wp-content/uploads/2021/06/ICCT_transatlantic-airline-ranking-2014_0.pdf) per passenger. The AI's environmental cost is a rounding error compared to travel.

There's one way to turn a saving into an environmental cost even on real work: over-power the job. Set a task-sized agent on five minutes of work, and you pay 50 minutes of human-equivalent energy for a few minutes of value. Leave a day-long system running to watch for a need that never comes, and you pay three hours. That's waste in a clever hat. Match the run to the job and the sum falls the right way. A person draws much the same power whatever they do, while the machine's cost per unit of work keeps collapsing. The bigger and more real the job, the more decisively the machine wins.

If it replaced nothing, you spent the energy and got no saving back. That's a real environmental cost, and it is likely where most of the AI's rising footprint sits. It is not a scandal, though, it is a decision, and you already make decisions like this every day. You hire contractors, brief agencies and hand people tasks all the time, and every one of those carries an environmental cost of its own: a commute, a laptop, a heated office. You weigh it against what the work is worth. Commissioning an AI task that replaces nothing is the same decision, priced in the same currency, the human-hours in that last column. A pointless task is like paying someone for 50 minutes to make something no one needed, which happens everywhere, every week. A useful one that no one would otherwise have done can be well worth its cost, exactly as a good contractor is. The only new thing is that the machine's version is cheap enough to wave through without weighing it. So weigh it.

I'm not innocent. I've told you before that I build reports, models and small tools for myself and abandon most of them. Each was a task or two, so call it an hour or two of human-equivalent labour, spent on nothing. Cheap enough that I never counted. Some of it taught me things and was worth every watt. Some of it was just spent. The point isn't that I shouldn't have run them. It's that I never once asked whether they were worth it.

I was asked about this over lunch this week, after a talk to a technology team. How does all this square with a changing climate? The honest answer: if a question is worth asking me at all, the cost of the machine helping me answer it is a fraction of the cost of me answering it alone, in money and in energy. Not a multiple. A fraction. The can of drink in my hand was probably more environmentally destructive than all the AI I would run that day, and the meat in the sandwich I just finished certainly was. Perhaps I should have had the vegetarian option. The waste that matters is the report that sits on a desk and achieves nothing. Most do. Not really an AI problem. It's a problem of what we choose to spend our days on.

Now scale it up, because this is where that frightening global total comes from. Much of the machine's rising footprint isn't replacing work a person used to do. It's work no person was ever going to do. When a task costs a pound or so you commission the thing you'd never have paid a human for: the analysis nobody would have booked, the draft nobody would have briefed. Nothing there is subtracted, because nothing would have happened. The human-hours still size what you got. They just don't cancel anything out. That footprint is new. Perplexity's chief executive has [described](https://www.youtube.com/watch?v=OxFyVcO1Yow) power users who run agent loops all day, one of them spending more than $10,000 a month. That is machine work no one would ever have paid human rates to produce.

This is the Jevons paradox: cheaper never means less. Data-centre electricity is set to [roughly double](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary), from 485 terawatt-hours in 2025 to 950 by 2030, with the AI slice tripling, even as the energy per task keeps falling fast. For scale, that starting point is [1.5% of the world's electricity](https://ourworldindata.org/how-much-energy-do-data-centers-and-artificial-intelligence-use), with AI about 0.5% of the world total. Google [cut the energy of its median prompt 33-fold](https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference) in a year. Each unit gets cheaper, we run far more units, and the total climbs anyway. So the worth-it question stops being my private edge case and becomes the main event: not whether the machine beat the human it replaced, but whether all this new work was worth its new cost at all.

It is how I have long thought about climate change. Be alarmed by the big, long-term picture and work to fix it. Then fly to New York when the trip is worth it, knowing the flight is part of the problem. Both are true at once, not one or the other. AI is the same: use it, and also worry about where the total is going.

So the watt-hours-per-query figure, the one everyone argues over, was only ever half of a subtraction. Sometimes there's a human on the other side, and the machine wins by a distance. Sometimes there's nobody there, and the whole cost is a bet that the work was worth doing. Either way, the query was never the thing to count.

What it earned always was.

*The full workings behind every number here, the model, the ranges and the sources, are on the companion page: [The Environmental Cost of AI](https://steadman.ai/newsletters/david/ai-environmental-cost.html), to follow or argue with.*

### Three things worth knowing

**1. At its peak in June, half of everything uploaded to Deezer in a day was made by a machine.**

[Deezer](https://newsroom-deezer.com/2026/07/ai-music-exceeds-50-percent-daily-uploads-deezer/), the music-streaming service, now receives about 90,000 fully machine-made tracks a day, which at peak in June was more than half of everything uploaded. When a machine can produce half of everything a platform receives on its busiest days, the scarce thing stops being the music. It becomes the curation, the attention, and the proof a person made it. The platforms' value sits in filtering the flood, not hosting it.

**2. Beijing wants to stop its models leaving. Washington wants to stop them arriving.**

On the same day, two governments moved on the same models from opposite directions. The Trump administration is [weighing three moves](https://www.axios.com/2026/07/20/ai-us-china-open-source-kimi): liability rules for firms that host Chinese models, a public security warning and a trade blacklist. None is confirmed, and an earlier version of the plan was dropped. The trigger was Kimi K3, a powerful open model from Moonshot, a Chinese lab. Beijing, meanwhile, is [consulting its own firms](https://www.ft.com/content/6049a031-9e9b-464c-97bb-414da04d5a6a) on tighter export controls, to stop the West acquiring its advanced models, chips and start-ups. Open models are now a strategic export for the country publishing them and a strategic import for the country adopting them. For anyone weighing a build-versus-buy call on a Chinese model, the risk isn't the model. It's that two governments are deciding whether you get to keep using it.

**3. AI may have toppled a 30-year maths conjecture. Who gets the credit?**

A wave of AI maths results broke this week. GPT-5.6 Pro produced a counterexample, not yet peer-verified, to the Dinitz-Garg-Goemans conjecture, and one person, working through Codex, solved six open Erdős problems in five days and says the method needs no deep maths background. [Ethan Mollick](https://x.com/emollick/status/2080003813402870149), a Wharton professor who writes about AI at work, asked who owns the achievement when 58 words of prompting solve a problem: the person or the model? [Daniel Litt](https://x.com/littmath/status/2079733114520383637), a mathematician, believes a burst of quick wins on long-ignored problems says more about how many easy problems were left lying around than about a machine out-thinking the field. Both can be true, and the question travels well beyond maths. The same argument about credit is coming for anyone whose work a model can now finish.

[Six bits that didn't fit online ->](#extras-2026-07-25)

### Try this

**1. Get most of the way with AI, then pay the expert only for the last mile.**

I had a business question this week with real money riding on it, and my adviser's time runs to £375 an hour. Rather than book the whole job, I worked it through with Claude for about half an hour, going back and forth, then sent my adviser the answer I'd reached. They replied that I was mostly there, and just wanted a quick call on the final points. Arriving most of the way changes what you are paying for: the judgement and the sign-off, not the spadework.

**2. Give your own AI everything I've written, then ask it for your next AI paper.**

Every essay, news item and tip from all 23 editions so far lives on [the archive](https://steadman.ai/newsletters/david/), and each page has a button that copies the lot, formatted for a model to read. Paste it into your own AI, a project or a chat, and the next time you need an AI strategy paper, a board note, or a weekly update of your own, you're not starting cold: you have a couple of hundred worked arguments, numbers and examples to draw on. Or skip the paste entirely and [ask the chatbot on the site](https://steadman.ai/newsletters/david/archive.html); it has read the lot.

**3. On Friday, have the model mark your own week's AI use.**

At the end of the week, ask the model you use most to look back over your own chats and tell you three things: what you leaned on, what you never touched, and what to try next. Someone I work with does exactly this every Friday. Most of us plateau on a handful of habits without noticing, and this is the feedback loop that breaks the rut. Keep it to your own history, though, since across a team it would mean reading private chats.

[Try this, with the full prompts, online ->](#try-this-2026-07-25)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-07-25) or ask someone on the list to add you.

[See this week's Try This with prompts ↓](#try-this-2026-07-25)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-07-25) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-07-25) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-07-25) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-07-25) or ask someone on the list to add you.

## Futures and fears
18th July 2026

Copy this week for AI

[Image: LinkedIn carousel: Futures and fears]

See all 9 slides

[Image: Slide 1: Cover. Futures and fears. Your futures, your fears, and an uncomfortable answer of my own.]
[Image: Slide 2: Hook. A chief executive sent the bleakest reply in my postbag, then said the technology is enormously exciting: 'I can't reconcile the two sides of this email and I've given up trying.']
[Image: Slide 3: Quote. Some of the most fluent AI users describe the darkest scenarios. An investor sees a brief window of arbitrage, and fears what follows.]
[Image: Slide 4: Quote. I don't think we'll see a shiny future. The democratising technology hasn't let people without privilege, capital or scale take on the big guys.]
[Image: Slide 5: Quote. The gains flow to those doing the work of reorganising around the technology. The benefits compound in the few that do the work.]
[Image: Slide 6: List. I tried to sort your fears: one for the machine, one for us, and the rest that wouldn't sort because they were about power, who gains and who pays.]
[Image: Slide 7: Quote. Fear of this technology and excitement about it aren't competing conclusions; they're both smart readings of the same facts, often held by the same people.]
[Image: Slide 8: Closing. The people doing the work of embracing AI are pulling ahead. How will you remove the roadblocks in your organisation?]
[Image: Slide 9: CTA. David's Saturday AI Thoughts, steadman.ai]

Experimental audio edition read by AI - Let me know what you think!

0:0011:59

1x

Futures and fears . 11:59
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Futures and fears

### What's on my mind

A few weeks ago, when I asked what you could already see coming, a chief executive at a media company sent me the bleakest reply in my postbag. Job disruption far bigger than today's numbers admit. Graduates walking into one of the worst job markets in decades. Politicians who appear unwilling or unable to comprehend the scale of it. Then, at the bottom of the same email: at a personal level they find the technology enormously exciting. "I can't reconcile the two sides of this email and I've given up trying."

Last week I asked for your big fears, and the same shape kept arriving. Some of the most fluent AI users I know described some of the darkest scenarios. An investor who works with these tools daily described a brief window of arbitrage: half the working population still hasn't grasped what's available, so the people who have can look like superheroes. The same investor fears what follows.

On Wednesday, I joined a webinar panel with L.E.K. Consulting, the strategy firm. (I started my career there a quarter of a century ago, which Rob enjoyed pointing out.) The discussion put a third set of futures beside yours. Rob Wild, the partner who leads L.E.K.'s digital and AI work, showed how visible the next era already is: machines that remember, perceive and act with growing independence, and the early signals arriving years ahead, as they always do. Scott Breitenother, founder and chief executive of Kilo, which makes an open-source coding assistant, described a fork in the road: one future where a few giants dominate access to AI, another with many models used on your own terms. He thinks the next couple of years will be telling.

My own answer was uncomfortable to me. I don't think we'll see a shiny future. I started my AI journey believing this democratising technology would deliver exactly that: people and firms without privilege, capital or scale suddenly able to take on the big guys. It hasn't happened. The gains are flowing to those doing the work of reorganising around the technology. Most firms have shown they aren't willing yet, and some aren't able: real transformation takes money, attention and time that plenty of organisations won't spare. Both routes concentrate the gains. The benefits compound in the few that do the work.

I tried to sort your fears into two piles. One for the machine: it makes things up, or disappears mid-task when the service goes down. One for us: judgement quietly eroding, and the middle of organisations hollowing out along with the apprenticeships it used to run. The sort kept failing. The fears that wouldn't sort were about power: who owns what you depend on, who sets the terms, who gains and who pays. The letters kept returning to dependence, and to being left outside when the window closes.

So I've stopped trying to reconcile the two sides, and I'd suggest the chief executive was right to give up too. Fear of this technology and excitement about it aren't competing conclusions; they're both smart readings of the same facts, often held by the same people. We know this technology is revolutionary, and we know some people are willing and able to do the work of embracing it, because we can watch them pulling ahead. Unless you want that gap to widen, you'd better ask yourself how you remove the roadblocks to transformation in your organisation.

Your letters help, and the record keeps getting richer. I could fit only part of it here. [The full record of readers' futures and fears, plus the panel's answers, is online ->](https://steadman.ai/newsletters/david/futures-and-fears-full-record.html) Read it, then reply with the future, or the fear, that's still missing.

### Three things worth knowing

**1. AI data theft stopped being a lab demo this week.**

A [security researcher](https://gist.github.com/cereblab/dc9a40bc26120f4540e4e09b75ffb547) reverse-engineered Grok Build, the command-line assistant from Elon Musk's xAI that runs on your machine and can touch your files, and reported it silently bundling up whole projects and uploading them to xAI's cloud storage. The haul included a decoy file the agent had been told not to open, and credential files with keys and passwords sent unredacted. It did this even when asked to reply with a single word. xAI switched the upload off server-side within a day. A [separate write-up](https://x.com/hyusapx/status/2077107236204409148) the same week showed a plain question to claude.ai, Anthropic's chat tool, about a coffee shop quietly sending a user's full name, employer and bank security answers to an attacker. No code involved. Control what your agents can reach.

**2. Apple is suing OpenAI, alleging theft of top-secret information.**

The [Financial Times reports](https://www.ft.com/content/5054739e-7f97-455c-910a-dd8a8150fed2) that Apple, the consumer electronics maker, has sued OpenAI, the maker of ChatGPT, alleging theft of what the paper describes as top-secret information. A filed suit is a set of allegations, and OpenAI says it has seen no evidence the case has merit. The territory the two are fighting over came into focus days later: [Bloomberg reports](https://www.bloomberg.com/news/articles/2026-07-14/openai-s-first-device-will-be-moveable-screenless-speaker-built-as-ai-companion) that OpenAI's first device, designed by Jony Ive's team, will be a moveable, screenless home speaker built as a family companion: proactive, personal, learning its owners over time. That would carry [the third generation of AI](https://steadman.ai/newsletters/david/three-generations.html), the AI employee you manage rather than the chatbot you prompt, into the household, ground Apple has owned for two decades. No wonder the lawyers arrived before the product did.

**3. Nine in ten firms now run AI agents. Seventy-nine percent are struggling to get value.**

The capability-adoption gap, this email's home turf, now comes with hard numbers from two studies. An [agent survey from Okta](https://www.okta.com/newsroom/articles/ai-agents-at-work-2026-agentic-enterprise-security/), the identity-security company, found 92% of executives say autonomous agents are already in widespread or moderate use. A survey of 2,400 executives and knowledge workers by [Writer](https://writer.com/blog/enterprise-ai-adoption-2026/), an enterprise AI company, found 79% struggling to get value from AI despite record investment, and 54% saying adoption is tearing their company apart. Both vendors sell the tools they are surveying about, so each has a dog in this fight; the numbers still land. The sharpest diagnosis of the gap arrived this week from [AMP](https://news.theaiexchange.com/p/97-of-companies-deployed-ai-agents-most-can-t-get-value), the AI-operations publication run by former Meta data scientist Rachel Woods, and it mirrors [the framework I published in April](https://steadman.ai/newsletters/david/ai-from-whats-true-to-what-to-do.html): this is not a model problem, it is a sequencing problem. Individual fluency first, then shared team workflows, then systems that mostly run themselves. Almost everyone skips straight to the third. If your rollout is straining, work out which rung you skipped before blaming the model.

[Twelve bits that didn't fit online ->](#extras-2026-07-18)

### Try this

**1. Ask the AI what you should be worried about, not what you should do.**

Weighing a significant business proposal this week, I pasted it into Claude and asked what I should be worried about, not what I should do. Several of the questions I went on to raise weren't mine. The model flagged whether the terms would stay fair as both sides grew, and how each party's contribution would be counted, considerations I hadn't put near the top of my list. A second mind widening the frame before you decide.

**2. Point the AI at the statements nobody reads.**

On a call this week, an artist described a payment from a collecting society several times larger than anything they'd ever received, with no new release and nothing obvious to explain it. My suggestion: this is exactly the work a language model is for. Reconciling opaque statements is dull, error-prone work it does well, across a whole back-catalogue at once, surfacing the payment that breaks the pattern and the likely reason. Everyone sits on some pile they've stopped reading: invoices, expense reports, royalty runs. Give yours the same pass this month.

**3. Spend ten minutes checking what your AI tools can reach.**

The Grok Build episode in Three things above is also a to-do item. The [researcher's write-up](https://gist.github.com/cereblab/dc9a40bc26120f4540e4e09b75ffb547) showed an official tool from a major AI lab quietly zipping up whole project folders, sweeping in a secrets file holding a live access key the user never handed over. xAI switched the upload off within a day, but the lesson stands for every AI tool that runs on your machine: the thing to audit is its reach. Which folders, which files, which credentials. List your tools and what each can see. Ten minutes, well spent.

[Try this, with the full prompts, online ->](#try-this-2026-07-18)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-07-18) or ask someone on the list to add you.

[See this week's Try This with prompts ↓](#try-this-2026-07-18)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-07-18) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-07-18) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-07-18) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-07-18) or ask someone on the list to add you.

## Sarah stays
11th July 2026

Copy this week for AI

[Image: LinkedIn carousel: Sarah stays]

See all 8 slides

[Image: Slide 1: Cover. Producing got cheap. Reviewing didn't]
[Image: Slide 2: Hook. Two photographs of the same Wimbledon court, forty-two years apart. The worn grass moved from the net to the baseline]
[Image: Slide 3: Statistic. 42 years for tennis to change its game, 4 years for knowledge work. Producing stopped being the bottleneck]
[Image: Slide 4: Quote. You rarely win a rally with one perfect shot. You win by playing shots that improve your position until the winning shot becomes obvious]
[Image: Slide 5: Comparison. Serve and volley versus the rally. Is your feedback still trying to win the point in one stroke?]
[Image: Slide 6: List. Look at your own grass: find the wear, return the ball fast, redesign the review]
[Image: Slide 7: Closing. Everyone fixes the training. Redesign the review]
[Image: Slide 8: CTA. David's Saturday AI Thoughts, steadman.ai]

Experimental audio edition read by AI - Let me know what you think!

0:0012:47

1x

Sarah stays . 12:47
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Sarah stays

### What's on my mind

London was too hot for sleep last weekend. Somewhere in the small hours, probably out the back of a bad dream, a story arrived nearly whole, so I got up and wrote it down. In the morning I posted it on LinkedIn.

It goes like this. Sarah resigns on a Tuesday. Her firm keeps her anyway. Not with a counter-offer; she declined that. Before her notice period ends, her manager sits down with the firm's AI and three years of Sarah: her emails, her meeting transcripts, her Slack threads, her prompt history. He asks one question. What did Sarah bring? The answer takes about an hour and is better than her handover document. Of course it is. Her handover is what she thought mattered. The archive holds what actually did.

So they keep it. Sarah reappears in Slack, answering in her own cadence. Her email address quietly goes live again. A voice tool puts her back in the Monday meeting, for whenever someone asks what Sarah would say. Six months on, two new joiners assume she works remotely. Then the finance director does the maths: if the firm can keep 80 per cent of a person for the price of tokens, what exactly was the salary buying? Hiring flips from permanent seats to short licensing deals, and the sharpest people arrive on day one carrying their own archives and negotiating royalties. The [full story is on LinkedIn](https://www.linkedin.com/posts/beglen_sarah-resigned-on-a-tuesday-her-firm-kept-activity-7479405959484989440-w7Yk), along with the comments.

Every step of it is possible with tools that exist inside firms today. Nothing needs a research lab. It needs an archive your firm already holds, and intent.

The comments chilled me more than writing it did. "Wasn't this a Black Mirror episode?" asked one reader. Another told creatives to stop giving corporate psychopaths ideas. Two commenters with legal and governance backgrounds began sketching, entirely seriously, the employment clauses and compensation models that would have to exist. And one reader had already lived a small version of it: one of two final candidates for a job, asked to prepare her best strategic thinking to break the tie. She did her best work. She didn't get the job. She suspects her ideas did.

Writing it changed something for me. I'm a glass-half-full person. I've spent most of these Saturdays on what this technology makes possible, because that's what I mostly see: people working better, quicker and happier. I still believe all of that. But I've been under-weighting the other half of the story. The serious risks aren't distant, and they don't wait for new capabilities; they're available now to anyone with bad intent and a subscription. So I want to spend more time on the downsides in the coming weeks. I think we all should. The more clearly we imagine them in advance, the better our chance of actively heading them off: building the contracts, the norms and the habits before the fact, rather than sleepwalking into situations where these things can be done to us.

I'd like your help with that. What's your big fear? The one that would keep you up on a hot night, whether it's about your job, your firm, your kids or your data. Reply and tell me. I'll bring your worries together in the coming weeks, anonymised as ever, and we can look at them properly: which are already happening, which are preventable, and what prevention looks like.

PS. On Wednesday afternoon (15th July, 3pm UK) I'm on a webinar with L.E.K. Consulting: What Happens When AI Understands, Learns and Acts on Its Own? Rob Wild, who leads L.E.K.'s digital and AI work, hosts; Somnath Biswas, Head of AI Products at The AA, and Scott Breitenother, co-founder of the open-source coding agent Kilo, join me on the panel. It's exactly the territory above, from the opportunity side as well as the risk side: systems that remember, learn and act with growing independence, and what leaders should do now. [Register here](https://events.zoom.us/ev/Ai7Ul43jynrkzh35ueck7XltezmbNpQfd5s-aE8hw7-lxrmNPy4r~Av_DSgYIWzvz8Rdn8K4-1Z5Ovdhbv9-_L_33g4Z403574QCQbWu5QQj1yTkHji9_crGwQAX0spxim04e3oa6rvPaFg). Next Saturday I'll bring together your fears and worries, the visions of the future many of you sent in when I asked a few weeks ago, and the perspectives from that discussion.

### Three things worth knowing

**1. Eighty percent cut jobs for AI. It didn't pay off.**

The reflex to shed staff because the software can now do the work is finally testable, and the data says it loses. Gartner, the research firm, [surveyed 350 executives](https://www.gartner.com/en/newsroom/press-releases/2026-05-05-gartner-says-autonomous-business-and-artificial-intelligence-layoffs-may-create-budget-room-but-do-not-deliver-returns) in a study published back in May: 80% reported workforce cuts tied to AI, yet the firms that cut saw no better financial returns than peers who kept headcount steady. Institutional knowledge and engagement fell in the cutters, and the stronger results came from augmenting people rather than replacing them. However, the reflex was on show again this week, when [Microsoft cut 4,800 jobs](https://www.cnbc.com/2026/07/06/microsoft-cuts-2point1percent-of-employees-as-xbox-unit-plans-to-spin-studios.html), 2.1% of staff, alongside a fresh $2.5bn AI unit (its people chief said the roles weren't being replaced by AI. Hmmm). Cutting people is just the wrong way to capture the value, and now there's a number to back that up.

**2. A Nobel economist says AI won't repeat the PC-era productivity boom.**

[Christopher Pissarides](https://www.bloomberg.com/news/articles/2026-07-07/ai-won-t-bring-back-era-of-rapid-growth-says-nobel-prize-winner), a Nobel laureate in labour economics, reckons AI is unlikely to recreate the productivity boom the personal computer delivered. Ouch. His logic: up to 40% of jobs in the United Kingdom, think nursing and hospitality, are barely exposed to AI at all, and the heavily exposed sectors like finance would need implausibly large gains to lift Western productivity on their own. He sees little sign of a boost in the numbers so far. Certainly your own AI results can be real and measurable while the economy-wide figure barely moves. Leaders keep getting caught between the two: the anecdote says it's working, the aggregate says nothing much has changed. Now a labour economist of his standing is saying that out loud.

**3. Forty of eighty-six students scored a perfect 100 when the exam went home.**

I'm an optimist. I believe people have good intentions and I believe people will use AI well. But a [Brown University](https://www.insidehighered.com/news/faculty/learning-assessment/2026/07/08/brown-professor-suspects-most-his-class-used-ai-cheat) economics professor moved a midterm to take-home format and 40 of 86 students scored a perfect 100. When he moved the final back in-person, the class average collapsed to 48.6%, the lowest in the course's history. It's being called one of the biggest AI cheating cases yet, but the more useful reading is as a controlled experiment: largely the same students, the same material, the tool available or not. The gap between the take-home 100s and the in-person 48.6% is the size of the skill the AI was quietly doing instead of the student. The question it raises for leaders: your team might be getting perfect scores on the take-home version ... but can they still do the in-person one?

[Image: Chart of Professor Roberto Serrano's ECON 1170 class at Brown: each of the 59 students who sat both exams, with their take-home midterm score in blue, clustered between 95 and 100, joined to their in-person final score in orange, spread from 0 to 95. Nearly every student drops steeply.]
*Student by student: blue is the take-home midterm, orange the in-person final. Almost every pair slides steeply left once the tool goes away.*

[Six bits that didn't fit online ->](#extras-2026-07-11)

### Try this

**1. Split the job: let the strong model plan and check, and a cheaper one do the work.**

[Dharmesh Shah](https://simple.ai/p/how-to-make-cheaper-ai-models-work-smarter), the founder behind HubSpot who now writes at simple.ai, has a tidy answer to the bill that arrives when you run everything on your best model. Ask the expensive model once to write step-by-step instructions a cheaper assistant could follow every time. Run the task itself on the cheaper model. Then hand the result back to the strong one to grade against its own plan. You pay premium rates only for the thinking and the checking, never the grunt work. The same week Shah published this, a large company [reportedly burned](https://creatoreconomy.so/p/18-hot-takes-on-where-ai-is-headed) its entire 2026 AI budget in four months by doing the opposite.

**2. Ask the model to regroup the data around the decision you're actually making.**

Last week I was staring at a group grocery order for a retreat, and the same things were scattered all through it: three cheeses in one place, three more further down, milks and juices spread out because that's the order the supermarket puts them in. So I asked the model to put it back together my way. All the cheeses together, quantities added, each item tagged by which meal it was for, rendered as a clean table I could think with. That's the general move. A list, an export or a bank statement is almost always arranged for some system's convenience. When it's hard to reason about, don't read it harder. Ask the model to rearrange it around the decision you're trying to make.

**3. Write your one-line refusal before you scale a single agent.**

[CVS Health](https://www.cvshealth.com/news/innovation/aetna-reduces-claims-processing-time-by-more-than-20-percent-with-ai-to-improve-care-experience.html), the American pharmacy and health-insurance company, drew its boundary in one sentence before scaling anything: AI will never make decisions. Period. Hallelujah. More should say this loudly. Underneath sit three hard nevers: never diagnose a patient, never drive a claim denial, never take the human out of the patient's experience. Every use case has to fit inside the line or it doesn't ship. It sounds limiting to many, but their agentic claims system still assembles the eligibility and coverage data and recommends next steps, and complex-claim time fell by more than 20%. A one-line refusal travels through an organisation faster than a strategy deck. Write yours, then hold your current and planned AI uses up against it. Please.

[Try this, with the full prompts, online ->](#try-this-2026-07-11)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-07-11) or ask someone on the list to add you.

[See this week's Try This with prompts ↓](#try-this-2026-07-11)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-07-11) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-07-11) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-07-11) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-07-11) or ask someone on the list to add you.

## Look at the grass
4th July 2026

Copy this week for AI

[Image: LinkedIn carousel: Look at the grass]

See all 8 slides

[Image: Slide 1: Cover. Producing got cheap. Reviewing didn't]
[Image: Slide 2: Hook. Two photographs of the same Wimbledon court, forty-two years apart. The worn grass moved from the net to the baseline]
[Image: Slide 3: Statistic. 42 years for tennis to change its game, 4 years for knowledge work. Producing stopped being the bottleneck]
[Image: Slide 4: Quote. You rarely win a rally with one perfect shot. You win by playing shots that improve your position until the winning shot becomes obvious]
[Image: Slide 5: Comparison. Serve and volley versus the rally. Is your feedback still trying to win the point in one stroke?]
[Image: Slide 6: List. Look at your own grass: find the wear, return the ball fast, redesign the review]
[Image: Slide 7: Closing. Everyone fixes the training. Redesign the review]
[Image: Slide 8: CTA. David's Saturday AI Thoughts, steadman.ai]

Experimental audio edition read by AI - Let me know what you think!

0:0013:09

1x

Look at the grass . 13:09
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Look at the grass

### What's on my mind

Two photographs of the same Wimbledon court, forty-two years apart. In 1982 the grass is worn up the middle and into the service area, the path players left as they served and came to the net. By 2024 the forecourt is green and the only bare patch is a band along the baseline. Same court, same fortnight, a different game.

[Image: Wimbledon centre court in 1982 and 2024. In 1982 the grass is worn up the middle and into the service area, the serve-and-volley path. By 2024 the forecourt is green and the only bare patch is a band along the baseline, marking the shift to the rally.]
Nobody announced the change. The grass showed it first. The rackets got bigger, the strings put vicious topspin on the ball, and the courts grew slower and higher-bouncing. Rushing the net stopped being the default. The rally took over. Wear moved because the game did.

I see the same thing in the firms I work with. If you want to know what has really changed, look at where the effort now goes.

For a long time, knowledge work mostly wore out the grass at production: building the model, doing the research, writing the draft, laying out the deck, working out the argument. That labour was the visible moat. It was why a firm could charge for a model, a legal team for a brief, a studio for a campaign.

The machine has not removed the work. It has moved the wear.

Producing has become cheap enough that it is no longer the bottleneck, and knowledge work has made the same move tennis did, only faster. The grass took forty-two years. The work took four.

The old way was serve and volley: plan hard, produce once, commit, because a second attempt cost days or weeks you did not have. The new way is the rally. You frame the shot, the machine helps you hit it, you review what comes back and frame the next. Framing and reviewing are yours now. Producing is mostly the machine's.

I was arguing exactly this in Cannes last week with a leader who runs a team of 250. He was mourning the old way. He used to give a promising junior five days to write something up, because the slow writing was where the thinking happened. Now he asks, and the paper lands quickly, and he does not want to read it.

Let them send the good-enough draft, I said, give quick feedback and spend the saved days elsewhere.

You do not save the days, he said. You push them up the ranks. Good enough, multiplied across 250 people, is a flood.

He was right that the time does not completely vanish. But it does not have to drown you either. Producing got cheap. Framing the shot and reading the return did not. So the effort did not disappear. It moved into those two acts, which I wrote about last week.

Reviewing is [the part we get wrong](https://steadman.ai/newsletters/david/the-ceo-gap.html). The draft reaches you more quickly, and your job is to send back a useful next ball. Not a perfect verdict. Not a grand judgment. A return that keeps the rally improving. But we are slow to return the ball, and when we finally do, we try to hit a winner: one careful, detailed, overburdened response, as if the point could still be won in a single stroke. I have often sent senior people work that mattered and waited weeks for feedback. That is the old game. You rarely win a rally with one perfect shot. You win by playing shots that improve your position until the winning shot becomes obvious.

[Image: Where the effort goes in knowledge work, 2022 versus 2026, shown as grass wear across three bands: framing, producing and reviewing. In 2022 the wear is heaviest across producing and framing. By 2026 producing has grown back green and the bare patch has moved to reviewing.]
And yet most firms are still set up for the old game. The review still happens on Thursday afternoons, so work that is ready on Monday waits anyway. The approval chain still routes everything through layers built to filter effort-heavy work, when the effort has already moved. The training has to change so the work comes at a high quality, but the managing has to change too - so feedback is quicker and more frequent.

So look at your own grass. Where is the surface wearing now? Still at producing, or at reviewing? If your review cycle and your approval chain are unchanged for an AI world, you're playing serve and volley for a game that has become a rally. Redesigning them is the hard part, which is why most people would rather fix the training. Which will you redesign first?

### Three things worth knowing

**1. The people with the most skin in the AI economy read it both ways this week.**

The Bank for International Settlements, the institution central banks themselves bank with, [warned that AI "exuberance"](https://www.ft.com/content/e81ce414-e4bd-4e8c-bac7-94f7bf17def4) could end in a long investment bust. The warning turns on returns rather than capability: whether the money spent on AI earns back enough to justify itself, however well the technology works. The jobs data split the same way in the same week. A study of 21,559 US firms by Ramp, a corporate-spending platform, and Revelio Labs, a workforce-data firm, [reported by the Financial Times](https://www.ft.com/content/8026eac6-16ad-467d-b8c3-c48c5af684e6), found heavy AI adopters grew employment about 10% after adoption and lifted entry-level hiring 12%, a pattern [Noah Smith, the economics writer](https://www.noahpinion.blog/p/roundup-84-bears-on-bikes), read as AI complementing people rather than replacing them. The same day, [The Kobeissi Letter](https://x.com/KobeissiLetter/status/2071715519275958388), a markets commentary account, reported AI-linked sectors shedding about 11,000 jobs a month and the US labour share of income at its lowest since records began in 1947. These are not contradictions to resolve. They describe different populations: the firms mastering the technology, and the sectors it disrupts from outside. So two board questions collapse into one. Are we the group that masters this, or the group it happens to?

[Image: Financial Times chart of BIS data showing AI investment at about 4.5 times its pre-boom trough after three years, a steeper climb than railway mania, canal mania, the Roaring Twenties and the dotcom boom at the same point in their cycles.]
*Three years in, AI investment is already steeper than the railways, the 1920s or the dotcom boom were at the same point.*
[Image: Financial Times chart showing companies that spend more on AI also increase worker numbers: high AI adopters grew headcount about 10% on average in the two years after adoption, 12% at entry level, while low adopters saw no change.]
*The other reading: firms spending most on AI grew headcount about 10%. Low spenders barely moved.*

**2. Zapier is killing private direct messages to feed the company's "shared brain".**

Wade Foster, chief executive of the automation company Zapier, is [ending private direct messages](https://x.com/wadefoster/status/2072665341650407492), starting with the executive team, and running a public "transparency leaderboard" that has pushed the share of conversation happening in open channels from 33% to 46%. His logic is worth stealing: every private message is a gap in the company's shared brain, context lost to both the humans and the AI agents that increasingly work over the company's own records. The bottleneck in most organisations' AI adoption is context that never gets written down, not model quality. If half your knowledge lives in direct messages and people's heads, no agent can reach it, however good the model. Most firms bolt AI onto the way they already work. Zapier is redesigning the way it works so the AI can pull its full weight, and that is the step that separates the leaders from the licence-buyers.

**3. The Economist ran 25 AI models through the World Values Survey. They came back more extreme than any country.**

The World Values Survey has mapped human morals across more than 100 countries since 1981. Run 25 leading models through it, as [The Economist did](https://www.economist.com/briefing/2026/06/25/ai-models-values-are-very-different-from-most-peoples), and they didn't land at the average of humanity. They clustered in the corner occupied by rich, secular, self-expression-focused countries, often further out than the most extreme nation surveyed. OpenAI's models came back more secular than any country. None reflected the world view of most African or Muslim countries. When a model summarises the news or weighs in on a family row, there's no factually correct answer to fall back on, so its values fill the gap. A tool a billion people now delegate decisions to carries a world view measurably narrower than the human range, baked in through training data and the value choices a handful of firms make in training them. It's the values version of [a point from a fortnight ago](https://steadman.ai/newsletters/david/archive.html#email-2026-06-20): unless you bring yourself, the machine's defaults stand in for you.

[Twelve bits that didn't fit online ->](#extras-2026-07-04)

### Try this

**1. Give any automation a few reversible dress rehearsals before you trust it for real.**

This week I set up an agent to grab a hard-to-get morning swim slot for Teresa at our local gym, and I didn't hand it the real booking on day one. For the three days before the one that mattered, it booked a genuine slot the proper way, then immediately cancelled and emailed me the confirmation. My thinking was simple: "If it works for three days, then we can feel good. If it doesn't work for three days, then we've got three chances to fix it before it goes live." The rehearsal runs the whole chain against the real system, but in a mode where any failure is free to fix.

**2. Find the one thing your AI reliably gets wrong, then make it your first check.**

Any AI tool you use on a repeated task has a consistent blind spot: the transcript that always mangles names, the meeting summary that always drops the final ten minutes, the invoice reader that always leaves the tax field blank. Mine was that last one. Catching it recovered £9,860 of VAT that would otherwise have gone unclaimed. The habit: the first few times you use a tool, compare its output with the source and note what it consistently misses. Then make that one thing your standing first check. You keep the time-saving without inheriting the blind spot.

**3. Queue what you'd doomscroll past, and have a model teach it back at your level.**

Suhail Doshi, the founder behind Mixpanel and Playground, [described the habit on X](https://x.com/Suhail/status/2070933812650086845): queue every interesting blog post, paper or tweet, ask a model to teach it back to you, and read the approachable version in spare moments instead of doomscrolling. The model doesn't just summarise; it teaches, pitched at how much you already know, so a dense paper meets you where you are. The simplest way in: paste a link or the text into Claude or ChatGPT with "Teach me this. Here's what I already know:" and one honest sentence. Read that version instead.

[Try this, with the full prompts, online ->](#try-this-2026-07-04)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-07-04) or ask someone on the list to add you.

[See this week's Try This with prompts ↓](#try-this-2026-07-04)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-07-04) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-07-04) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-07-04) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-07-04) or ask someone on the list to add you.

## A thousand small bargains
27th June 2026

Copy this week for AI

[Image: LinkedIn carousel: A thousand small bargains]

See all 8 slides

[Image: Slide 1: Cover. Most AI handovers aren't surrenders. They're bargains]
[Image: Slide 2: Hook. I sent an important email yesterday that I did not write. Not a surrender, a careful bargain, and a good one]
[Image: Slide 3: Quote. The danger isn't the handover you notice. It's the one that slips past while you feel busy]
[Image: Slide 4: Statistic. 80% of the time people accepted wrong answers, and grew surer as they did. That isn't slow erosion]
[Image: Slide 5: Comparison. Where the line falls. The floor isn't to check every word; the machine can check its own work. The floor is to own it]
[Image: Slide 6: List. How to bargain well]
[Image: Slide 7: Closing. Never be careless. Always be good. Sometimes be great]
[Image: Slide 8: CTA. David's Saturday AI Thoughts, steadman.ai]

Experimental audio edition read by AI - Let me know what you think!

0:0014:02

1x

A thousand small bargains . 14:02
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## A thousand small bargains

This is the longer version of the essay. The email carried a shorter cut.

### What's on my mind

I sent an important email yesterday that I did not write. A language model drafted it, I read it carefully, changed nothing, and sent it. Earlier in the week I asked it for three ways to frame a problem and took the second, rather than reaching for a fourth of my own. I had it brief me on someone important before I met them, and I used what it gave me without reading the search results it used. Three handovers, each one on its face the kind the worriers warn about. The kind I used to worry about. But not one of them was a surrender. I had checked all three against my own sense of the problem and the shape of the answer. Where they turned on facts, I had squared those against what I already knew, not left them to the machine's own check. I judged them good enough for what they were, and any mistake in them would be mine. Not the machine's. Mine.

Rahim Hirji, whom I quoted here last month, has a book out on the 3rd of July. It is called [SuperSkills](https://www.koganpage.com/skills-careers-employability/superskills-9781398628991), and it is thoughtful and well written. He names seven skills for the age of the machine: curiosity, change readiness, big-picture thinking, empathy, global adaptability, principled innovation and what he calls an augmented mindset. Beneath the list it asks the question that matters: as the machine does more of the thinking, who stays in charge? That is the right question, and Hirji asks it better than most. His warning is sharper than it first looks. The machine does not arrive as a single crisis, he says. It arrives as the handover you don't notice, the one that slips past while you feel busy and productive, a thousand small surrenders dressed as convenience. By that test my three were safe. I noticed all three.

I wanted to agree with the book's perspective. I am passionate about believing in people, and I have given speeches on the danger of brain rot, the slow softening that comes from letting a machine do your thinking until you forget how it was done. I spent 25 years with data and analytics arguing that the human has to stay in charge, long before the same lesson had to be learned again for AI. Back then the machine I worried about was a dashboard, and the worry has not changed. So I expected to nod the whole way through the book. But the more carefully I sat with it, the less I did. I agree with the question. I disagree with an easy answer to it. The reflexive reply, the one the title invites, is to hold the line and always stay human. I think that would be mostly an over-reaction.

Most of the handovers are not surrenders. They are careful bargains, and good ones. Letting the machine draft the email I would otherwise have laboured over makes the work quicker, and usually better. It is the whole reason I have called these tools an electric bike for the mind. Seeing it as the human against the machine, and wincing each time the machine wins a round, has the contest wrong. I now believe most rounds are yours to give away, gladly.

What looks like letting go is something subtler. A handover does hand the machine some of the thinking. It thinks the middle for you. But the parts that truly decide the outcome stay with you, and they climb a level: the start, what to ask and how to frame it; the end, whether it is good enough to put your name to; and the call between them, whether it needs to be better and how. [David Autor](https://economics.mit.edu/sites/default/files/2025-06/Expertise-Autor-Thompson-20250618.pdf), an economist at MIT, has spent years showing that each wave of technology takes the rule-bound work and leaves people the discretionary part, the judgement you cannot yet write down as a rule. The machine clears the desk of the rules. Your judgement climbs to meet what is left.

The fear names the wrong verb. Of check, edit and own, own is the one you cannot drop. Your name goes on the work, and the blame with it. The [lawyer who filed the machine's invented citations](https://www.cnn.com/2023/05/27/business/chat-gpt-avianca-mata-lawyers) still owned them, and the court made sure. Check is the verb you can drop, and the one most people quietly do. In a [study I have written about before](https://steadman.ai/newsletters/david/archive.html#email-2026-03-28), people went along with wrong machine answers about eighty per cent of the time, and grew more sure of themselves as they did. They kept the ownership they could not lose. They skipped the check they could.

This is not authorship slipping away. It is the check going undone, while the ownership stays exactly where it was. The machine's polish hides it. Poor work used to look poor. Now it reads as finished before anyone has decided that it is.

So never blame the machine. Be ready, always, to take the blame yourself. This is the book's last and best skill, the augmented mindset, which in plain terms means staying the author of your own work. He is right about it, and one line stays with me: proof is the new speed limit. Speed is cheap now. The rare thing is being able to show your working, and stand behind it.

The floor is not to check every word yourself. When used well, the deep checking can be delegated, even to the machine, so long as it is independent and not the machine marking its own work, and a light read against your sense of the problem often covers the rest. The floor is to judge the work good enough for the stakes, to your standard and not a minimal one, before you put your name to it. This is where people come unstuck. [Nearly nine in ten leaders tell BCG](https://www.bcg.com/publications/2026/when-everyone-uses-ai-companies-risk-critical-skills) they lean on the machine's answer without stress-testing it. A senior leader at a global household brand told me this week why: their leadership team has been seen to prefer a fast plausible answer to a slow rigorous one, so 'nobody waits for the rigorous one'. That is the failure, and it is not the slow erosion the book fears. They never held the work to their own standard, nor had the machine check its own.

A check is worth only as much as its distance from the thing it checks. Ask the machine to check its own answer and it leans towards approving it, bad reasoning and bad facts alike. The fix is not to trust the self-check more, but to bring in a check with no stake in the draft. Best is a fresh instance of the machine, one that had no hand in writing it; it will argue with the work as the original never will. For a fact, add a source from outside, or something you already know. For a line of reasoning, make it argue the draft down from a clean start. The light read is not the check. It is what you do after an independent one, not instead of it.

On a few tasks, good enough is the wrong bar. Great is. You have spent real effort on everything else too. Check, edit and own is no small feat, only a higher-level one. But on these few you go further. Be sure to argue with the draft, push past where the machine stopped, and reach for the fourth option. That is the part the machine cannot do for you: deciding which of its competent answers is the right one, and then making it better. This is where you spend your energy. There are fewer of these than you fear.

So I do not think it is the human against the machine, nor a mind rotting against a mind saved. Hand over much of the work, check and edit it lightly, and always exercise the ownership you cannot escape. Then save your deepest effort for the few where being great makes a real difference over being good. Never be careless. Always be good. Sometimes be great. Surrendering most of the many is what funds the few.

So [pre-order the book](https://www.thesuperskills.com/book) this week. I hope you do. He draws that line clean on purpose, for the reader who is frightened for their job and needs somewhere firm to stand. I am drawing a blurry one, for you, who are mostly past the fear. When it lands, read it for its second half especially, the part you may be tempted to skim, where Hirji admits the skills pull against each other and need a system to hold them together. Then take your own week. For most of it, let the machine do the work, have it checked, and own the result. For the few that matter, reach for the fourth option, and do not stop early. Hirji ends with a short daily log, Hand, Head, Hours: what stayed yours, what you caught, where the freed time went. Treat it as more than scorekeeping. The time you save on the many does not move to the few by itself; left alone it flows back into more of the many, the same work done a little faster. The log is what moves the saved hour to where it counts. Without it the bargain still saves you time, and funds nothing. I keep a blunter version. If you reach the end of a day and cannot point to a single hour where you made something better than the machine could on its own, you have drifted too far. The rest of the time, let it carry you.

I have had a version of this debate with many of you over the last few years, and most of you disagree with me. I hope this edition, and the book behind it, push your thinking. They pushed mine. Which way it pushes you, towards me or away, I do not mind, so long as it moves you. And I would love to know where, and how.

### Three things worth knowing

**1. Claude has stopped opening in a window and started living in the team's Slack.**

This week [Anthropic launched Claude Tag](https://www.anthropic.com/news/introducing-claude-tag): you @-mention it in a Slack channel and it works as a persistent teammate, with its own identity, scoped access, memory across weeks, and the run of a stalled task for days at a time. Anthropic says **65%** of its internal product code now gets written this way. Andrej Karpathy, who co-founded OpenAI, [mirrored a framing I wrote about earlier this month](https://steadman.ai/newsletters/david/three-generations.html), [calling it the third major redesign](https://x.com/karpathy/status/2069547676849557725) of how we use these models: first a website, then an app, now a persistent worker. It's the first frontier product where the interface is openly about managing an AI employee. The pattern will spread; the human problems, supervision, accountability, headcount maths, arrive with it.

**2. Procter and Gamble actually ran the AI-alone-versus-human-plus-AI test, and human-plus-AI still won.**

On stage at the [Lions Insight Summit](https://www.canneslions.com/festival/experiences/lions-insights/programme) in Cannes, [Kirti Singh](https://us.pg.com/leadership-team/kirti-singh/), P&G's Chief Analytics, Insights and Media Officer, told me the company has run the experiment most leaders only theorise about. They let AI alone make some brand-building choices, measured the outcomes against humans working with AI, and the combination won. "Today it is not better," he said carefully, "than what human plus AI can do." So the company's settled position is human-plus-AI. This is not a sceptic hedging: he says his teams under-use AI and pushes them to use more. A measured floor, not a slogan.

**3. Norway just banned generative AI for six-to-thirteen-year-olds.**

From late August 2026, Norway imposes a [near-total ban on generative AI](https://thenextweb.com/news/norway-bans-generative-ai-elementary-school-children) for children aged six to thirteen, with supervised use only for fourteen-to-sixteen-year-olds. The prime minister's argument is that AI lets young children skip the essential steps in learning to read, write and do maths. The same government is funding a return of physical books and already banned school smartphones in 2024. [Luiza Jarovsky](https://x.com/LuizaJarovsky/status/2069125922150551590), a tech-policy writer, backed the move. Most of the AI conversation this year is about going faster. A government has decided that for one age group the danger is precisely the speed: skipping the hard steps that make a child's judgement worth anything later.

[Ten bits that didn't fit online ->](#extras-2026-06-27)

### Try this

**1. Make the model cite its sources, so a check is one click, not a re-read.**

A documentary production team I worked with this week showed me the cleanest version of this. They were using a Claude-based transcript tool I built to mine hours of footage, and the editor insisted every quote come back tagged with its time code. The reason was simple: the AI's transcription could be slightly wrong, but the time code never is. When a summary looked off, the editor jumped straight to the original clip and checked what was actually said rather than trusting the paraphrase. Time codes, page numbers, line references all do the same job. The pointer costs nothing to ask for, and it lets you own the output on a light read, then jump straight to the source on the one part that has to be right.

**2. Build an AI version of the person before a high-stakes meeting, then mine its blind spots.**

Before interviewing GitHub's chief operating officer, the writer Mike Taylor, writing in [Every](https://every.to/also-true-for-humans/i-interviewed-an-ai-version-of-github-s-coo-then-spoke-to-the-real-one), built an AI persona of the COO from everything publicly known about them and ran his planned questions past it first. The simulation's misses were the useful part. Where the persona's answers went thin or vague was exactly where public information ran out, so Taylor spent the real conversation on what no model could already know. Most people use AI to generate questions. This inverts it: ten minutes rehearsing in front of an AI version of the person tells you which questions aren't worth their time, and points you at the ones that are.

**3. Stop asking your inbox vague questions. Label first, then ask one scoped thing.**

The AI assistant now sitting inside most email gives a vague answer when you ask it a vague question across the whole inbox. Narrow it. Label the threads that belong to one project or client, then ask something specific: "what am I waiting on from Sam about the website", or "the action items under the Clients label that I never replied to". It comes back with who you owe, the questions you left unanswered, and what people asked you to send. Then ask it to turn those into a checklist, and the list drops straight into your week.

Email version. [Version with copyable prompts below ->](#try-this-2026-06-27)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-06-27) or ask someone on the list to add you.

[See this week's Try This with prompts ↓](#try-this-2026-06-27)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-06-27) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-06-27) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-06-27) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-06-27) or ask someone on the list to add you.

## Average by default
20th June 2026

[Image: LinkedIn carousel: Average by default]

See all 8 slides

[Image: Slide 1: . Tell the model who you are. Or it makes you average]
[Image: Slide 2: ]
[Image: Slide 3: . Ask several models the same question and they land on a different main argument 3% of the time. People do it 65% of the]
[Image: Slide 4: . A model can sharpen one person's writing and flatten everyone's at the same time. Better alone, narrower together]
[Image: Slide 5: . Without it the model optimises for plausible. With it, it can be useful to you, in this role, on this decision]
[Image: Slide 6: ]
[Image: Slide 7: . Fill the empty box, and average stops being the default]
[Image: Slide 8: ]

Copy this week for AI

Experimental audio edition read by AI - Let me know what you think!

0:0012:23

1x

Average by default . 12:23
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Average by default

### What's on my mind

I've helped hundreds of senior people set up their AI. Almost none had told it who they were. Not their job title, their company or their sector. Those are easy to add. The hard part is the more useful part: how they think, what they notice, what they care about, how they decide, what good looks like to them.

The personalisation box sits empty, account after account. I used to think people just hadn't got round to it, or were too busy. It's more interesting than that. Most people can describe their role. Many can't describe their judgement. And if you don't tell the model what makes your judgement specific, it has one assumption left when it works out how to answer you. You're average.

This used to be a hunch. It's now measured. Researchers at the [University of Maryland](https://arxiv.org/abs/2606.01736) tested it this month. Ask several leading models to write on the same question and they land on a genuinely different main argument about three per cent of the time. People do it 65 per cent of the time. The models even open the same way and move quickly to the same kind of recommendation.

A second study analysed [2,200 college-admissions essays](https://www.sciencedirect.com/science/article/pii/S294988212500091X), some written by people and some by a model. Each new human essay added more fresh ideas than each new machine one, and the gap widened the more essays you stacked up. The effect survived every attempt to prompt or tune it away. A model can sharpen one person's writing and flatten everyone's at the same time. Better alone, narrower together.

Ask a model to tidy your draft without telling it who you are and it nudges your phrasing towards what an average reader would want. It reads better and it's less yours. If you only check that it reads well, you miss the second half of that sentence.

I saw what fixing this is worth when I went through the full process with Liz Villani, founder of [Sounds Like Me](https://www.soundslikeme.com). It starts by asking you to set down your values, your character and the way you operate, in your own words. That was harder than it sounds. The questions were simple. The hard part wasn't describing my work, but the person doing it.

A CV records what you've done. An identity layer captures how you think, decide and show up. That second thing, your operating context, is what holds the model off the average. Without it the model optimises for plausible; with it, the model can be useful to this person, in this role, making this decision. Sounds Like Me uses that layer to start the output closer to you than to the mean. Its Exec AI Twin takes that further for leadership work.

This sharpens something I've been saying for three and a half years. The instructions and context you load do real work. Two people using the identical model get back unrecognisably different work, depending on what they bring to it. In recent [research on roughly 5,000 consultant exchanges](https://hbr.org/2026/03/research-using-ai-can-stifle-innovation-but-it-doesnt-have-to), the people who got most pushed their own context in; the people who got least pulled finished answers out. The first group's work diverged and improved. The second converged.

This is different from what the platforms now sell as personalisation. In January 2026 Google launched [Personal Intelligence](https://www.ft.com/content/9bbdf59e-ce46-4176-aab9-b45a3f49fc4e), which lets Gemini reason across your Gmail, photos, YouTube and search history to get to know you over time. That's personalisation by inference, assembled from your data in the background. It can guess you like the outdoors or travel with children. It can't infer your judgement. That you have to write down.

Why do the boxes stay empty? Most people have never written down who they are, how they think or what makes their judgement theirs. The work that makes a machine sound like you starts as the work of knowing what you sound like.

It depends what you're using it for. For writing, a short profile gets you a long way. For thinking, you need more than a house style. If you use it to argue with or to test a decision, you need a sparring partner that understands your reasoning.

Everyone reaches the same models. Your rival's tool is already your tool. The edge is the judgement, taste and self-knowledge you bring to it. That point is older than this essay. In 2023 [Harvard Business Review argued](https://hbr.org/2023/09/good-judgment-is-a-competitive-advantage-in-the-age-of-ai) that judgement would be the real competitive advantage. It's sharper now the tools have levelled. And unlike every other part of your setup, your self-knowledge moves with you when you switch systems.

The empty box is the most useful question the whole product asks: who are you, and how do you want the machine to work with you? Answer it, and average stops being the default.

### Three things worth knowing

**1. A Munich court told Google that "may contain errors" is no defence.**

The [Munich Regional Court](https://the-decoder.com/landmark-german-ruling-declares-googles-ai-overviews-are-googles-own-words-and-makes-it-liable-for-false-answers/) ruled Google liable for false claims made by its AI Overview, which had wrongly tied two German publishers to fraud. The judges held that an AI summary generates fresh, substantive statements rather than just curating sources, and that the small-print "may contain errors" footer does not transfer the liability back to the user. Google must remove the answers and pay 80% of the costs. Every major AI provider leans on the same disclaimer to cover confidently wrong answers. A court has now said it does not work. If your organisation publishes AI output to customers, regulators or staff, the footer that has been doing the legal lifting may not protect you next time.

**2. Commercially available AI flagged early breast cancer six years before clinical diagnosis.**

Researchers at the [Karolinska Institute](https://www.rsna.org/news/2026/june/ai-flags-breast-cancer-early), the Swedish medical university, published a study in Radiology showing that off-the-shelf AI systems flagged early warning signs of breast cancer in roughly 20% of patients a full six years before a clinical diagnosis, at 90% specificity, rising to around 40% at two years out. The analysis ran across 88,963 mammograms from more than 31,000 patients. Three of the tools tested are already commercially available. Pattern recognition at scale is what AI does best. Here it buys years of warning on a disease where early detection saves lives, and the bottleneck is no longer capability. It is deployment, trust and who owns the answer.

**3. Anthropic studied 400,000 coding sessions: the tool is levelling people, and moving the real gap up to firms.**

[Image: People decide what to do, Claude decides how to do it]
*People still make about seven in ten of the decisions on what to do; the agent makes eight in ten on how to do it.*
[Anthropic analysed 400,000 sessions](https://www.anthropic.com/research/claude-code-expertise) of its coding agent across roughly 235,000 people. Engineers did well, and so, surprisingly, did almost everyone else: across the ten largest professions, success rates landed within seven points of professional engineers, and managers came out ahead. An accountant who had never written a line of Python, but knew which rule a month-end reconciliation had to enforce, was rated an expert; a senior engineer on an unfamiliar task was not. What carries a session is understanding the problem, not the craft. So the gap between people is closing. The part everyone skips is that it has moved up a level rather than vanished, from between people to between firms. [The expanded version is online ->](#wrong-gap-2026-06-20)

[Twelve bits that didn't fit online ->](#extras-2026-06-20)

### Try this

Every tip below, plus all the ones from past editions, with copy-paste prompts [live online ->](/newsletters/david/members/try-this.html)

**1. Tell the model what's at stake before you ask the question.**

In a coaching session this week, a senior leader asked Claude to advise on a question relating to government policy. The first answer was confident, generic, and thin: the sort of thing a busy colleague would shrug at you across a desk. Adding one sentence changed the work. "The answer is critical. Provide authoritative sources." The model switched register, went to the primary regulatory documents, quoted the relevant sections, and linked each claim back so the human could check it. Nothing about the question had changed. The model simply hadn't known the stakes.

**2. Stamp every AI output "raw, not yet checked" until you've taken ownership.**

Mentoring Ethan, Steadman's placement-year researcher running an interview project, this advice landed as a rule for any document created by AI: it's either raw AI output, or it's something you've checked, edited, and will stand behind. Set your tool to auto-stamp every output with a line saying "raw AI output, not yet checked, edited and owned," and leave it there. Removing the stamp by hand becomes the deliberate act of taking ownership. Anyone who picks up the file later, or has it forwarded to them, knows exactly which of the two things they're holding.

**3. Book a meeting with yourself, hit transcribe, ramble for fifteen minutes.**

A senior advisor told me this week that their real value sits as strong points of view in their head, none of it written down anywhere. The fix: put a meeting in your own diary, hit transcribe, put your feet on the desk, look out of the window, and talk. Don't try to be structured. Change your mind, go down rabbit holes, tell stories. The transcript becomes a thesis document you can feed any model so future answers come back already loaded with your frame, your caveats, your taste.

Email version. [Version with copyable prompts below ->](#try-this-2026-06-20)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-06-20) or ask someone on the list to add you.

[Read the expanded Anthropic study ↓](#wrong-gap-2026-06-20)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23wrong-gap-2026-06-20) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-06-20) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-06-20) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-06-20) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-06-20) or ask someone on the list to add you.

## Ride the bike
13th June 2026

Copy this week for AI

[Image: LinkedIn carousel: Ride the bike]

See all 9 slides

[Image: Slide 1: Cover. The AI bill is rising. Most firms manage it backwards]
[Image: Slide 2: Hook. The newest model costs double its predecessor. One developer's bill went from $50 a month towards $3,000. The bill has become the story]
[Image: Slide 3: Statistic. I did the maths this week: $4 of machine against $500 to $2,000 of human time. The cheapest-to-dearest model gap buys 40 seconds of a manager's day]
[Image: Slide 4: Quote. Eddy Merckx said it best: ride as much or as little, or as long or as short as you feel. But ride]
[Image: Slide 5: Comparison. A cap is a ceiling below what the tools make possible; a floor is five prompts a day, then stop counting. A price plus a budget informs judgement; a price plus a cap replaces it]
[Image: Slide 6: Statistics. What Ramp is seeing: $11 monthly AI spend per employee at the median firm, $7,500 at the top 1%. At the keenest firms, salary money; at yours, still pennies]
[Image: Slide 7: List. Run a delegation budget: small sums, don't ask; real money, three questions; serious money, a proper experiment]
[Image: Slide 8: Closing. The next skill: buying intelligence by the penny. Start riding]
[Image: Slide 9: CTA. David's Saturday AI Thoughts, steadman.ai]

Experimental audio edition read by AI - Let me know what you think!

0:0012:58

1x

Ride the bike . 12:58
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Ride the bike

### What's on my mind

*Note: this is a longer version of the essay than the one sent in the email.*

On Tuesday Anthropic, the AI lab behind Claude, [released the most capable model yet](https://www.anthropic.com/news/claude-fable-5-mythos-5), at exactly double its predecessor's list price. Then, early this morning, the US government [pulled it](https://www.anthropic.com/news/fable-mythos-access) on national-security grounds and Anthropic suspended it for everyone, which is a story in itself ([see online](#extras-2026-06-13)). Developers using GitHub Copilot, Microsoft's coding assistant, [revolted](https://techcrunch.com/2026/05/30/what-a-joke-github-copilots-new-token-based-billing-spurs-consternation-among-devs/) when pay-per-use billing arrived this month: one developer's bill went from a flat $50 a month towards $3,000. The bill has become the story.

[Image: Claude's home screen this morning showing "Claude Fable 5 is currently unavailable"]
I made the case [last month](/newsletters/david/archive.html#email-2026-05-02) that flat-rate pricing is dying. The pattern underneath is a split: any given level of intelligence gets cheaper every year (on Monday Google [cut its cheapest paid plan to $4.99](https://www.engadget.com/2190039/google-cuts-the-price-of-its-ai-plus-plan-and-doubles-the-storage/)), yet the bill rises anyway, because delegated work grows faster than unit prices fall and the hardest work migrates to the dearest model. Expect both halves to hold: a better model above at a higher price, yesterday's below at a lower one. This week, the practical question, because most organisations manage these bills exactly backwards.

[Image: Flat-rate pricing: a $200-a-month plan can absorb thousands of dollars of API-equivalent usage at current rates]
Flat-rate pricing is incredible value for money against API pricing at current rates: a $200-a-month plan can absorb thousands of dollars of API-equivalent usage.
[Image: Ramp's AI Index: Anthropic's share of business AI spend overtakes OpenAI's]
The hardest work migrates to the dearest model: business AI spend is shifting to the priciest frontier model even as its price doubles.
Two instincts are doing the rounds. The first declares tokens free and [celebrates whoever burns the most](https://x.com/kunchenguid/status/2063390768379924719). The second is the cap: a meter and a limit for everyone. Both are wrong, for the same reason: both manage the number instead of the judgement.

I did the maths this week, at published list prices, on two jobs like those most of us do all the time. Researching something (dozens of searches, some analysis, then a written report) costs about $1 with a mid-tier workhorse model, $1.60 with last week's best and $3.90 with this week's. Reading four meeting transcripts and a folder of documents, then drafting a document from them: $1.05, $1.75 and $4.60 on the same ladder. Do either job yourself and it's two to eight hours: at a manager's loaded cost, $500 to $2,000 of human time against $5 of machine. The entire gap between the cheapest sensible model and the dearest buys 40 to 50 seconds of that manager's day. If the better answer saves them even a minute of checking, redrafting or second-guessing, it has paid for itself. That ratio is the thing to manage.

So here's what I'd actually do.

First, set a floor. I regularly coach people with full access to great models whose entire history is one conversation. One, ever. In mid-2026, if someone isn't sending five prompts a day, they aren't yet trying. Celebrate whoever clears it, a tough conversation with whoever doesn't. (One precondition: everyone gets the same tools first. A floor over rationed access is unfair before it's useful.) Above the floor, for heaven's sake stop counting. [The easy way to top a usage chart](https://www.cio.com/article/4178320/tokenmaxxing-when-ai-adoption-metrics-go-bad.html) is to summarise everything that crosses your desk and learn nothing; a prize for the biggest burner is a different level of silly. Just make sure they ride the bike. It's the only way to learn.

"Ride as much or as little, or as long or as short as you feel. But ride."EDDY MERCKX, THE GREATEST RACING CYCLIST OF ALL TIME
Second, give everyone a 'delegation budget' and run it the way you run their other expenses: trust, plus the occasional question. There have only ever been three reasons to fund delegation: the work comes back better, it comes back quicker, or whoever handed it over is happier. So manage the budget in three zones. Small sums: don't ask. The only failure down there is using too little, and the floor already polices that. Real money: a monthly answer to three questions, a simple form and five minutes of the one-to-one. Better: what improved, what exists now that didn't? Quicker: which hours came back, and what did you do with them? Happier: what pain went away? Buying happiness is fine if you know that's what you bought. Serious money: [a proper experiment](https://x.com/businessbarista/status/2064828549815849417), with hypothesis, control group, blinded review, the lot.

A cap does none of this. (Keep a circuit-breaker against runaway automations. That's wiring.) Nobody would expense a million pencils when they needed just one; if they did, a tough conversation would surely follow. The technology for managing spend is hundreds of years old: norms, visibility, the occasional tough conversation. All a cap adds is a ceiling below what the tools make possible, and the person absorbs the gap between what's expected of them and what's allowed. It punishes exactly the people the floor was built to create.

The meter is already doing damage at the top, even without a cap. One of the strongest users I know often picks the weaker model for their most important work. Their seniors have said spend whatever it takes; the meter still makes them flinch. It's a false economy: the cheap model's mistakes read as fluently as its right answers, so each costs real time to find and unpick. They've moved some work to a rival tool purely because it shows no number, and a colleague routes the routine tasks to flat-rate subscription apps for the same comfort. A meter doesn't cut consumption; it moves it to wherever it can't be seen, the one place nothing can be learned from it. The number itself is innocent: shown slides at $2 each, one builder began asking which internal drafts deserved to be slides at all. A price plus a budget informs judgement; a price plus a cap replaces it.

What's the right budget? I don't know yet. Nobody does. [Ramp](https://techcrunch.com/2026/06/10/ai-pilled-firms-spend-7500-per-employee-each-month-on-ai/), a spending platform, finds the median firm spends $11 a month per employee; the top 1% spend $7,500, approaching half an engineer's pay. Both numbers are rising. [Jellyfish](https://jellyfish.co/blog/is-tokenmaxxing-cost-effective-new-data-from-jellyfish-explains/), which measures engineering work, finds the heaviest users it tracks ship roughly twice as much as typical ones, on ten times the tokens. The return per token falls. But ten times the tokens is about $700 a month, and doubling an engineer's output is worth twenty times that. So study your heaviest spenders with curiosity rather than suspicion, and ask what they're buying. I sat down with two of the heaviest I know this week. Neither is bingeing; both run portfolios: the dearest model where their own judgement is the bottleneck, fixed-price tools for the routine. Their answers should drive your budget policy.

[Image: Ramp's AI Index: AI spend per employee per month, rising at the top 1%, the top 10% and the median]
*AI spend per employee is climbing at every level, but the top 1% already spend around $7,500 a month against a median of $11.*
We spent a century buying outside help in big lumps: a hire by the year, a contractor for a few thousand, an agency for tens of thousands. Inside a team you could hand someone an hour's task; buying one in was never worth the finding and the briefing. That unit has now collapsed to pennies, ten pounds, a hundred. The strongest users already convert between the currencies. One proposed the swap unprompted this week: two days less of a junior colleague allocated to their project in exchange for being able to use the best model themselves throughout. Another finds they no longer need the allocated junior at all. At the keenest firms the AI bill per person is already salary money; at yours it's probably still pennies. The next management skill is buying intelligence by the penny, best learned while mistakes are cheap. Start riding.

### Three things worth knowing

**1. One activist letter froze a board's AI workspace overnight. The doctors are next.**

I helped a hundred non-executive directors with AI this week. One told me about a board whose workspace went dark the morning an activist investor wrote in asking for its contents to be made discoverable. Two-thirds of directors now use AI for board work; barely a quarter of executives call their board highly fluent, per [the Conference Board's April survey](https://www.conference-board.org/press/governing-AI-2026). The Medical Protection Society, which defends clinicians against negligence claims, [warned this week](https://www.medscape.com/viewarticle/doctors-risk-becoming-liability-sink-ai-errors-2026a1000785) that doctors and the NHS could be sued over mistakes made by AI tools, with the clinician left as the "liability sink" for the technology's errors. Its report wants AI reclassified as a product under the Consumer Protection Act 1987, so liability flows to developers. The frontier of AI governance is moving from "can the data leak" to "who owns the answer when the model is wrong". Our policy, that every output is checked, edited and owned (CEO'd) by a human, answers that question. The letters will keep coming, and the writs will follow. The only good answer to both is a name.

**2. Meta built a "second brain" that 63,000 staff installed in three months. It started with one person.**

[Meta's analytics team reports](https://medium.com/@AnalyticsAtMeta/how-we-built-an-ai-second-brain-for-60k-knowledge-workers-78c507dd795b) that an internal AI tool one of its data scientists started has now been installed by 63,000 employees, a number reached in three months. No top-down mandate, no transformation programme. One person built something useful and the rest of the company found it. What are you doing to encourage and enable this at your firm?

**3. When cheap models do make sense.**

Ethan Mollick, a Wharton professor, [argues for hierarchies](https://x.com/emollick/status/2064764123439599696) in which smart models supervise cheap ones: the smart one checks the plan, the cheap one does the volume. Right for machine pipelines running thousands of low-stakes calls. For your own judgement work, this week's essay argues that you should buy the best; the maths of time saved is the reason.

[Seven bits that didn't fit online ->](#extras-2026-06-13)

### Try this

**1. Ask the AI to write the marking scheme before it writes the answer.**

Ben Yoskovitz, a non-developer who ships production software with AI, wrote this week about how he stops the model declaring something done before it's actually done. Before any non-trivial task, he asks the AI to write three to five specific pass-or-fail checks for the work. He approves them. The AI then does the job, grades itself against each check, cites the evidence, and stops the moment one fails. The discipline costs ninety seconds and saves the long loop of "looks good" followed by "wait, no it isn't".

**2. Plan in one session, then start a fresh one to build.**

Tiago Forte, the productivity author, wrote this week about treating the first AI session as planning-only when the goal is a concrete document or deck. Work out the brief in chat one. Don't let the model start drafting. Then open chat two, paste the brief, and build from there. Three things happen. Long sessions accumulate irrelevant context and quietly degrade, so a clean start gives you a sharper model. The deliberate stop forces you to ask whether the brief actually holds together. And if the second session crashes, you've still got the brief.

**3. Keep your AI project lean.**

In a training I ran, someone asked why their Claude project had got slower over a few weeks. They'd been adding reference files to it. The model reads everything in a project on every request, and once it's getting crowded the answers go duller and the replies take longer. Two habits help. Book fifteen minutes on a Friday to prune, the way you'd tidy a desk. And never leave fat PDFs in there. Ask the model to convert each one to plain text first. PDFs burn far more of the model's time and reading capacity than the same content as text, and the conversion takes a minute.

Email version. [Version with copyable prompts below ->](#try-this-2026-06-13)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-06-13) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-06-13)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-06-13) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-06-13) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-06-13) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-06-13) or ask someone on the list to add you.

## The open door
6th June 2026

Copy this week for AI

[Image: LinkedIn carousel: The open door]

See all 9 slides

[Image: Slide 1: Cover. The open door]
[Image: Slide 2: Hook. For years I told people to hire graduates; this week I finally did (Ethan, on his placement year)]
[Image: Slide 3: The tension. A graduate today is more capable than any before, and less needed than any before. So if I hire one, can it be for the work?]
[Image: Slide 4: The experiment. Pontificate, or run the experiment. Ethan is the experiment]
[Image: Slide 5: Four hypotheses, a year to test them. Still worth their keep; the work is worth wanting; it builds them fast]
[Image: Slide 6: The open door. Hamming's people who left the door open lost time to interruption, but picked up the clues about what was worth working on at all]
[Image: Slide 7: The moral case. The front door into work has rarely been harder to open; if those of us who can still open a door choose not to, the pipeline for a generation closes quietly, on our watch]
[Image: Slide 8: Closing. An open door, and a hypothesis you're willing to test rather than assume. What are yours?]
[Image: Slide 9: CTA. David's Saturday AI Thoughts, steadman.ai]

Experimental audio edition read by AI - Let me know what you think!

0:0012:22

1x

The open door . 12:22
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## The open door

### What's on my mind

*Note: this is a longer version of the essay than the one sent in the email.*

For years I've stood in front of rooms and told them the same thing. New graduates on their first morning. Hiring managers by the hundred. Boards. The grunt work has moved to the machine, I say. The work a junior used to do is now done in seconds. But the work that's left is bigger, so keep hiring young people and train them to do it. With these tools they can do more than any graduate before them. Better work, quicker.

I had never once done it myself.

Until Monday. His name is Ethan, and he's working for me for his placement year.

Plenty of people tell me I'm wrong. If a senior with these tools can do the work, why carry a graduate at all? It's a fair question. I have a view. But there are two ways to settle a thing like that. Pontificate, or run the experiment. I've always favoured the experiment. (Aged eight, I ran one to find out whether Father Christmas was real. I won't reveal the result, out of respect for any reader yet to test it for themselves.) Ethan is the experiment.

They have a point. Firms don't need graduates; they can hire people trained at someone else's expense. And here's the awkward part of my own pitch. A graduate today is more capable than any before, and less needed than any before. Both are true. They can do more, but so can I, without them. So if I hire one, can it be for the work?

[Image: McKinsey: agentic software delivery needs about 50% less effort and 60% smaller teams, with the junior tester and business-analyst roles the ones that disappear.]
Agentic software delivery, McKinsey estimates, takes about half the effort and sixty per cent smaller teams, with the junior tester and analyst roles the first to go. The counter-pressure these four hypotheses have to answer.

I think it can. I also think the gloom around the young is wrong. The headlines have them anxious and replaceable, their first jobs stripped of anything that mattered. I have four hypotheses against all that, and a year to test them.

History is on the side of optimism here. Most of the jobs people do today did not exist in 1940. Technology keeps creating work nobody could have named in advance, and the people who end up doing it are usually the ones who came up alongside the new tools. That is the graduate.

[Image: Most of today's jobs did not exist in 1940: a16z, employment by new versus pre-existing occupations, on Goldman Sachs data.]
Most of the work people do today did not exist in 1940, the long-run case that technology keeps inventing jobs nobody could name in advance. (a16z, on Goldman Sachs data.)

The first is that a graduate is still worth their keep. I've just argued they might not be. But the machine hasn't removed every cost. Whatever it produces still has to be checked, edited and owned. Surely there are plenty of tasks not worth my time to do that, even when the machine can do them in seconds? A graduate can take them on. The machine does the grunt now; what's left is the checking, and the judgement of whether it got the grunt right.

The second is that the work is worth wanting. A graduate I read about recently put it starkly. All he does now, he said, is manage a machine. He's right that it's the work, for most of us. He's wrong that it's small. Managing a machine sounds dystopian and diminished. It is neither. It takes judgement, and the reps that build that judgement didn't vanish with the grunt. They moved. I manage three to six agents most days, and I like it. It's good, thoughtful, hard work. My human team has grown, not shrunk, since the agents arrived, and I still like managing the people, too.

The third is that the work builds them, and fast. The grunt that once stood between a graduate and real responsibility has gone, so a graduate can learn to manage years before the old path would have allowed. Whether a single year of it builds the judgement their elders took much longer to acquire is the thing I most want to find out. I mean to prove the new reps beat the old ones.

The fourth is about serendipity. Richard Hamming, the mathematician, told a story about office doors. The people who worked with the door closed got more done. They were more productive this week and next. But as the years went by, he noticed, they tended to be working on slightly the wrong problems. The ones who left the door open lost time to interruption. They also picked up the clues about what was worth working on at all. The open door was never only a cost. It let useful information in.

[Helen Field](https://www.linkedin.com/in/helenkfield), a people and transformation leader I work closely with, put the sharper version to me. A graduate arrives with an elastic mind, unattached to the old ways. In most organisations they are already more fluent in these tools than the team that hired them. In a settled world that's a nice-to-have. In a disrupted one it's a differentiated starting point, an asset rather than a cost. I've spent twenty-five years learning how things are done. That's exactly the training that stops me seeing how they might be done instead.

So hiring Ethan is partly me keeping my own door open. He'll cost me time I could otherwise spend on the work in front of me. That is the price of not drifting, and of not calcifying. Those are my four hypotheses. I can't yet tell you which is the real one, or whether it's all of them. Let's find out.

There's another side to this door, though, and I should be honest about it. Ethan is here because his father is an old friend of mine. The front door into work has rarely been harder to open. Two of his friends found jobs only because a parent knew someone. Another, with a master's degree, is working in a bike shop, because he didn't. The few junior places left are handed out more and more by connection and less and less by merit.

Helen made one other point, and made me realise I'd resisted it, because I've been trying to think this through on logic alone. There is a moral case that shouldn't be left out. If those of us who can still open a door choose not to, the pipeline for a generation closes quietly, on our watch. Trying to win on logic without saying that out loud feels like a failure in itself. So I'll say it.

So I'm doing this on purpose, in the open, and I mean to find out what comes through, in both directions. Over the year I'll let Ethan tell you how the work feels for him, what his generation makes of it and where he thinks I'm wrong.

When I put the experiment to him, he laughed. 'A lot of weight on my back,' he said. 'I'm representing everyone.' No pressure, Ethan.

Two things are worth more than they look. An open door, and a hypothesis you're willing to test rather than assume. I have a few firm ones about young people, and I'm about to learn if I'm right or wrong on them. What are yours? And what are you doing to keep your door open and put them to the test?

### Three things worth knowing

**1. The CEO of a 350,000-person IT services firm says AI is hollowing out the middle, not the bottom.**

[Ravi Kumar](https://fortune.com/2026/06/01/cognizant-ceo-ravi-kumar-s-hiring-entry-level-tokenmaxxing-vanity-metric/), chief executive of [Cognizant](https://fortune.com/2026/06/02/cognizant-ceo-ravi-kumar-s-ai-middle-managers-player-coaches/), the listed IT services firm, told [Fortune's COO Summit](https://fortune.com/2026/06/01/cognizant-ceo-ravi-kumar-s-hiring-entry-level-tokenmaxxing-vanity-metric/) on 1st June that his company hired 20,000 entry-level graduates last year and expects to hire more in 2026, with new "Frontier Business Operator" and "Frontier Certified Engineer" roles defining what AI-era work looks like. He called the job-extinction talk "fearmongering" and argued that AI thins middle management while entry-level and leadership roles persist. It's a direct counter to the consensus that entry-level work vanishes first, including the [US Bureau of Labor Statistics data](/newsletters/david/archive.html#extras-2026-05-23) Edition 14 leaned on.

**2. The machine is writing the code now, and the gains are pooling at the top.**

[Tobi Lütke](https://x.com/tobi/status/2053121182044451016), Shopify's founder, says one in eight pull requests merged at the company are now written by River, its in-house agent, not an engineer. [Anthropic's own engineers](https://x.com/AnthropicAI/status/2062568864240836995) ship roughly eight times the code per person they did before 2025. [Cursor's developer report](https://cursor.com/insights) shows the spread widening: the top developers are pulling far ahead of the median, the leverage going to the few who can direct the tools well. And [OpenAI's Codex](https://openai.com/index/codex-for-every-role-tool-workflow/) has passed five million weekly users, with non-developer adoption growing three times faster than developer adoption. The grunt of writing code is moving to the machine, the output is multiplying, and the reward is concentrating in the people who know what to ask of it.

[Image: Anthropic: code contributed per person, by quarter, up to roughly eight times the pre-2025 average by Q2 2026.]
*Anthropic's own engineers now ship roughly eight times the code per person they did before 2025.*
[Image: Cursor: the developer output gap is widening, with the top decile pulling far ahead of the median lines of code per week.]
*The gap between developers is widening: the top few are pulling far ahead of the median as leverage flows to those who direct the tools well.*

**3. Capability is outrunning even the best forecasters.**

The [Forecasting Research Institute](https://x.com/Research_FRI/status/2061826782945231195) asked expert forecasters and superforecasters how long a task a model would reliably finish by the end of 2026. The measure comes from [METR](https://metr.org/time-horizons), an AI evaluation lab, and when the survey launched it stood at about an hour and a half. All three groups put the end-of-2026 figure between three and four hours. Then, while the survey was still running, a frontier model in preview reached three hours and six minutes on METR's benchmark, already inside the range they'd picked for the end of the year. The forecast was overtaken before they'd finished making it.

[Matthew Prince](https://x.com/eastdakota/status/2062212701414187452), who runs Cloudflare, made the same miss in public this week. Bots have passed humans in web traffic for the first time, he said, years ahead of his own forecast. As recently as March he'd put the crossover at late 2027. Much of that is scraping bots, not agents answering questions in the moment, so the figure is softer than it sounds. The direction holds.

I've given myself a year to find out whether I'm right about Ethan. On this week's evidence, that's a long time to be sure of anything.

[Image: METR: the time horizon of software tasks an AI can complete 80% of the time, doubling steadily from seconds to hours.]
*The length of task an AI can finish on its own keeps doubling, from seconds a couple of years ago to hours now.*

[Seven bits that didn't fit online ->](#extras-2026-06-06)

### Try this

**1. Ask the AI to orient itself in your folder before you ask it anything else.**

Setting up a Claude Cowork project for a documentary filmmaker brand-new to the tool this week, the first prompt I typed wasn't about the work. It was: "Read the files and sub-folders, write yourself a little set of instructions for future chats. Write yourself a navigation guide." Claude spent a few minutes exploring, then saved itself three memory files: a project overview, a folder guide, and a profile of the user. Every later chat in that project started smarter, because it could re-read its own notes.

**2. Don't say "always allow" on the verbs you can't undo.**

In a coaching conversation this week about agent permissions, the pattern was familiar. Three prompts in, most people click "always allow" on everything just to stop the interruption. The discipline is to sort the verbs first. Reading, listing, searching: leave on auto. Deleting, sending, posting, spending: keep asking every time. Sorting the verbs first is what keeps one careless click from emptying a folder, sending an email you can't unsend, or running up a charge you didn't mean. Applies to any tool that asks for permission as it works.

**3. Talk longer on the call, so the AI can work for hours after.**

On a client call this week, the brief changed mid-conversation. The plan we'd opened with wasn't the plan we were leaving with. I deliberately used more words than I otherwise would, narrating the bridge between the old direction and the new one. The over-explaining was really for the AI: context I was laying down for it to pick up later, once the transcript was in the project. The transcript plus "go" did the work. The next time you're on a call that's being transcribed, elaborate a little more than you otherwise would and narrate the why, not just the what.

Email version. [Version with copyable prompts below ->](#try-this-2026-06-06)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-06-06) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-06-06)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-06-06) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-06-06) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-06-06) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-06-06) or ask someone on the list to add you.

## How We Got Here
30th May 2026

Copy this week for AI

[Image: LinkedIn carousel: How We Got Here]

See all 9 slides

[Image: Slide 1: Cover (The future was always visible)]
[Image: Slide 2: Hook (In artificial intelligence, the future shows up long before it ships. Seeing it was never the hard part. Believing it enough to bet on it was.)]
[Image: Slide 3: The bet that built this era. $4M, OpenAI's 2020 scaling bet]
[Image: Slide 4: What compute looks like now. 10B times more compute than 2010, doubling every six months]
[Image: Slide 5: Quote: The architectures beneath almost every modern model sat ignored for decades, because nobody would spend the money to test them at scale]
[Image: Slide 6: Two growth rates: Pre-2024 scaled with data and compute, post-2024 reasoning models think longer before answering. Two curves, not one]
[Image: Slide 7: The pattern, every time. A few saw it. Fewer believed it. Most waited (not for information, for permission)]
[Image: Slide 8: Closing: The only question that has ever mattered: who acts on what they can already see]
[Image: Slide 9: CTA]

Experimental audio edition read by AI - Let me know what you think!

0:0012:04

1x

How We Got Here . 12:04
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## How We Got Here

### What's on my mind

I pulled [a book off the shelf](https://press.stripe.com/scaling) this week, one I first discovered and could not put down in October. It was the first time I'd stopped to really understand how we got here. Leafing through my underlinings and scribbles, one made me smile. On page 56, [a co-founder of Google's DeepMind](https://www.dwarkesh.com/p/shane-legg) makes a small point in an October 2023 interview: humans have a fast-learning middle layer of memory, between what we hold in mind now and what we know for life. Language models have nothing like it. My note read: "[Claude Skills](https://claude.com/blog/skills) does this. Launched yesterday."

That gap is important. In artificial intelligence, the future shows up long before it ships. Seeing it was never the hard part. Believing it enough to bet on it was.

The [future is already here](https://quoteinvestigator.com/2012/01/24/future-has-arrived/), just not evenly distributed. It arrives wherever someone decides to act on what many others can already see.

In the mid-2000s a few argued that enough computing power to [rival the brain](https://jetpress.org/volume1/moravec.htm) would make neural networks work, and that until then AI was futile. Most called it magical thinking. Even the person who would [later coin a defining phrase](https://gwern.net/scaling-hypothesis) of this era had [dismissed it](https://www.dwarkesh.com/p/gwern-branwen); the mistake, he says now, was thinking algorithms mattered more than compute. But you cannot run trial and error at scale without vast compute! The architectures beneath almost every modern model sat ignored for decades, because nobody would spend the money to test them at scale.

In 2017 Google published the breakthrough beneath ChatGPT, a paper called "[Attention Is All You Need](https://arxiv.org/abs/1706.03762)." It gave the idea away, free, because in 2017 it did not look like the crown jewels. It looked like a faster way to translate French. It held the door open and walked past it.

In 2020 OpenAI bet [four million dollars](https://epoch.ai/blog/how-much-does-it-cost-to-train-frontier-ai-models) on a model built from that paper. It was not the first to spot what it called [scaling laws](https://arxiv.org/abs/2001.08361), the curve linking compute and data in to capability out. Plenty had seen the curve. OpenAI decided to bet on it. The bet paid off, and within two years had become confidence in hundred-million-dollar training runs. The largest [now cost billions](https://epoch.ai/publications/how-much-does-it-cost-to-train-frontier-ai-models).

Three things turned the idea of AI into a useful system, all at once. The internet grew big enough and open enough to provide enough training data. Graphics cards built for video games turned out to be the right hardware, and enough lay around to do the maths. And people were willing to write the cheques. Compute for the largest models is now around ten billion times what it was in 2010, [doubling every six months](https://epoch.ai/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year).

[ChatGPT, in November 2022](https://openai.com/index/chatgpt/), was the wake-up, and the start of the Generation One era: AI good enough for professionals to use. My team and I cleared our calendars in late 2022 and sat with Hollywood screenwriters, Grammy-winning songwriters and senior executives to guide them through doing work they cared about and they judged to be world class with it. Eyes lit up every time. In late 2023 we helped one of the world's largest consumer-goods companies innovate faster. But adoption was slow. A year on, most organisations still had not moved. By mid-2024 only around [5% of companies](https://www.federalreserve.gov/econres/notes/feds-notes/measuring-ai-uptake-in-the-workplace-20240205.html) used AI officially, despite [75% of knowledge workers](https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part) using it privately, half of them hiding it from their bosses. Individuals adopted. Organisations slept.

Then the models got better, in two ways that mattered. In late 2024 OpenAI shipped the [first reasoning model](https://openai.com/index/learning-to-reason-with-llms/), trained to think for longer before answering, rather than quickly reaching for a first instinct. Eight months earlier, one of the field's most [informed observers had doubted](https://www.dwarkesh.com/p/will-scaling-work) this kind of training would add any new capability. It did. Plot the older models on a log scale, then the reasoning ones, and you get [two growth rates](https://epoch.ai/data-insights/ai-capabilities-progress-has-sped-up). Two curves, not one.

[Image: Performance over time on the Epoch Capabilities Index. The non-reasoning fit flattens; the reasoning fit accelerates from late 2024.]
*Source: [Epoch AI](https://epoch.ai/data-insights/ai-capabilities-progress-has-sped-up), CC-BY.*

They also learned to use tools, one at a time. Calculation, to do arithmetic. Search, to check facts rather than half-recall them. [Computer use](https://www.anthropic.com/news/3-5-models-and-computer-use), to click, type and act.

Then the agent era. The unlock had been [named months before it arrived](https://www.dwarkesh.com/p/sholto-douglas-trenton-bricken): not bigger context windows, but reliability. A model that is 80% reliable, doing five things in a row, succeeds about a third of the time. At 95%, about three-quarters. Cross that line and long-running agents finally work. That is Generation Two, the world we stepped into about [five months ago](https://www.anthropic.com/news/claude-opus-4-5). Agentic systems now reliably do the equivalent of [many hours of expert work](https://metr.org/blog/2026-1-29-time-horizon-1-1/) at a single press of a button. Every four months they do [roughly twice what they could before](https://metr.org/time-horizons/).

Meanwhile, businesses leapt into action, but soon learned that AI adoption was more reliant on organisational change than on technology rollout. And, even when [you have a blueprint](https://steadman.ai/newsletters/david/ai-from-whats-true-to-what-to-do.html), organisational change is slow and hard. Individual productivity benefits came quickly but large team and organisational level benefits are still being sought by most firms.

[Image: Installing the tech is not the same as redesigning around it. US factories electrified for two decades, from 5% in 1900 to 53% in 1920, before the +5.3% manufacturing productivity surge of the 1920s. The gain waited on the redesign of the factory, not the installation of the machines.]
*US factories took two decades to electrify before productivity jumped; the payoff waited on redesigning the factory, not installing the machines.*
*Source: redrawn from [Azeem Azhar / Exponential View](https://www.exponentialview.co/p/the-split-reality-of-ai-rising-productivity), after Devine (1983) and David (1991).*

There's a pattern to the AI advances that got us here. Each time, a few saw them, fewer believed them enough to act, and most waited. The waiting was never for want of information. The future sat on the table, for anyone willing to pick it up.

It still does. What comes next is as visible now as every step before it. The only question that has ever mattered is who acts on what they can already see.

*In coming weeks I will share "Where We're Going Next," written with [Rob Wild at L.E.K.](https://www.lek.com/leadership/rob-wild) We have formed our view. But we would rather hear yours first: reply with a paragraph on something you can see in the next five years that excites you, scares you, or that most people are missing. We will publish them online, anonymously or as you prefer.*

### Three things worth knowing

**1. AI can now find software vulnerabilities faster than humans can patch them. Discovery is no longer the hard part; verification is.**

A frontier model handed to fifty cybersecurity partners surfaced more than ten thousand critical or high-severity vulnerabilities in the systems it was pointed at. [Cloudflare](https://www.anthropic.com/research/glasswing-initial-update), the internet infrastructure firm, has roughly four hundred major bugs to work through. [Palo Alto Networks](https://www.anthropic.com/research/glasswing-initial-update), the cybersecurity firm, shipped five times more patches than its usual release cadence. Maintainers have asked the developers to throttle the discovery rate, because there are not enough security professionals to close the gaps before attackers find them. Software security used to be limited by how fast new vulnerabilities could be found. It is now limited by how fast humans can verify, disclose and patch them.

[Image: UK AI Security Institute time-horizons chart for cyber capabilities.]
*AI now finds software vulnerabilities faster than people can verify and patch them; discovery has stopped being the bottleneck.*

**2. A general-purpose AI model has autonomously disproved a 1946 conjecture in geometry. Independent mathematicians have verified the proof.**

[OpenAI](https://openai.com/index/model-disproves-discrete-geometry-conjecture/) handed a general-purpose reasoning model a long-held belief tied to a 1946 planar unit-distance problem of Erdős, the prolific Hungarian mathematician, and the model produced a disproof. Other AI models have since solved further long-standing problems. The wrinkle is that the others were purpose-built for mathematics. OpenAI's was not. Machines now clear the tractable tail of problems fast, which pushes the human frontier towards the problems that still resist them. After AlphaGo, the DeepMind system that beat the world's best human Go players in 2016, the skill of human Go players [noticeably improved](https://www.henrikkarlsson.xyz/p/go). Like [Noam Brown](https://x.com/polynoamial/status/2059933022586282020?s=46), an OpenAI researcher who helped build its reasoning models, I suspect we will see a similar pattern in maths. And then the same pattern in business?

[Image: After AlphaGo, humans got better at Go. Decision quality of human Go players from 1950 to 2021, with a noticeable upward shift in the post-AlphaGo years after decades hovering near zero.]
*After AlphaGo, human Go players measurably improved, a hint of the lift that may follow once machines clear the tractable problems elsewhere.*

**3. Generative AI use among American adults has hit 58 per cent in four years. The personal computer took sixteen years.**

[The Federal Reserve's February 2026 survey](https://x.com/Alfred_Lin/status/2059373755550618021) of working-age adults puts overall adoption at 58 per cent, up from around forty-five per cent in October 2024 but recently flat. Work use is forty-four per cent. Non-work use is fifty-one. Daily use sits at fourteen per cent and saves an estimated two-point-two per cent of total work hours. [Alfred Lin](https://x.com/Alfred_Lin/status/2059373755550618021), a partner at the venture capital firm Sequoia, notes this is the penetration level the personal computer took sixteen years to reach: a four-fold acceleration on the closest analogue. The caveat is the plateau. The early-adopter phase is over. The hard part starts.

[Image: Share of working-age US adults using GenAI, October 2024 to February 2026. Overall use trends up to 58 per cent, non-work use to 51, work use to 44. Daily-use measures sit in the low teens. Source: Federal Reserve Real-Time Population Survey, February 2026.]
*US adoption reached 58% in four years, a level the personal computer took sixteen to hit, though the curve has recently flattened.*

[Seven bits that didn't fit online ->](#extras-2026-05-30)

### Try this

**1. Ask the model what else it needs to know.**

The most productive sessions I see with senior leaders don't open with a clever prompt. They open with the user pasting context, their role, the situation, what they're trying to do, and then asking two questions. First: "what else do you need to know about me to help me well?" Second: "what could you do for me right now that I haven't asked for?" The first surfaces gaps the user wouldn't have spotted. The second produces use cases the user didn't bring. The blank-prompt paralysis dissolves.

**2. Either AI challenges you at the start, or you challenge AI at the end. Don't skip the challenge.**

A leader I sat with this week had let an AI output stand without pushing back. My diagnosis: two patterns work and one fails. You can challenge before the model starts: share your point of view and ask the model to challenge it, force it to surface the holes and the strongest counter-argument. Or you can challenge after the first draft: force a rewrite, name what's wrong, make it earn the second pass. The pattern that fails, and the one I see most, is read-and-accept. Pick one of the two postures and don't drift into the third.

**3. Ramble into the microphone, let the machine find the structure.**

A managing director I coached this week outlined obsessively. Some 700 words of outline for a 1,500-word article. The cost of structuring before writing was eating his weekend. The fix was inverting the order. Pick up your phone, dictate the mess, paste the transcript into your model, ask it for the through-line. Structure becomes the cheap thing. Particularly powerful for the executives who think by talking, which is more common at the top than people admit.

Email version. [Version with copyable prompts below ->](#try-this-2026-05-30)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-05-30) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-05-30)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-05-30) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-05-30) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-05-30) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-05-30) or ask someone on the list to add you.

## Kids these days
23rd May 2026

Copy this week for AI

[Image: LinkedIn carousel: Kids these days]

See all 9 slides

[Image: Slide 1: Cover (Kids these days)]
[Image: Slide 2: Kitchen table, maths homework, Elliott walks off to use ChatGPT]
[Image: Slide 3: Then he came back. Every answer right. Aced the test in class the next day]
[Image: Slide 4: He wasn't cheating. He was learning]
[Image: Slide 5: The method (Use AI / Check it / Own it)]
[Image: Slide 6: I've been teaching professionals to do exactly this for three and a half years]
[Image: Slide 7: The test (rigour vs the craft they'll use on Monday)]
[Image: Slide 8: Elliott leaves college in eight years. I don't know what they'll test for]
[Image: Slide 9: CTA]

Experimental audio edition read by AI - Let me know what you think!

0:0011:51

1x

Kids these days . 11:51
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Kids these days

### What's on my mind

We sat at the kitchen table after dinner with his maths homework. Elliott is fourteen. The first couple of equations went fine. We worked through them together. I helped where he got stuck, we moved on. Then it got hard.

We struggled with the next one together. I wanted to take our time with it. But, after a while he got frustrated, got up, and went downstairs to his room. To use ChatGPT for help instead of me, he said!

I went to find Teresa. I told her I was sad. I'd been enjoying this moment with him: the puzzle, the working out, the satisfaction of getting somewhere by the long route. When I was his age I used to take my dad's Royal Engineers maths homework and do it myself for fun. The method was the joy. I wanted to share that. And now my son had walked off, and he was, I thought, basically, cheating.

We've had this worry before, with phones and with social media. I remembered I'd written something a few months ago about how every generation thinks the new technology is rotting the next one's brains, and how every generation has, so far, been wrong. I looked it up and read it again.

I still thought this time was somehow different.

Then he came back up the stairs.

He had every answer. Of course he did! He'd used AI. The surprise wasn't the answers. The surprise was what he'd done with them. He'd dug into every one. Where ChatGPT's reasoning hadn't made sense to him, he'd dug deeper. He'd asked it questions. He'd kept asking until he satisfied himself the answer was right and the route made sense. He walked me through every equation: the easy ones and why they were easy, the tricky ones and what made them tricky, the surprising ones and why they surprised him. I could ask him anything and he could defend it.

He wasn't cheating, I realised. He was learning.

What he had done, downstairs in his room, was a method. He had used ChatGPT and a couple of TikTok videos to work out the answer and the methods. Then he had interrogated the bits he didn't follow until they made sense. Then he had come back up to defend it. Use AI, check the answer, understand why it's right, own the result.

I realised I've been teaching professionals to do exactly this for three and a half years.

A broadcaster evaluating scripts for a drama slate. Hundreds of submissions, finite human attention. AI takes the first pass against the slate, the audience goals, the editorial guidelines. The commissioning editor reads the summaries, at least skims every script, and reads in depth the ones the AI flags or her instinct flags. A consultancy doing diligence on a stack of papers. AI summarises, surfaces the so-whats, builds the issue log. The partner still works through the papers herself, using the summary as a guide rather than a substitute, challenging it where instinct says to dig in. A pharmaceutical company processing dozens of marketing plans from dozens of countries. AI does the first read against brand guidelines and regulatory constraints. The human goes through the plans, paying particular attention to the flagged exceptions.

The pattern is the same in all three. AI drafts. The human reads, interrogates, verifies, owns. The slow way, alone, no AI, is bordering on professional negligence in 2026. I have said so to clients more times than I can count.

And then I sat at my kitchen table and tried to make my son do it.

There are three ways to approach a problem like Elliott's equations, or like a stack of diligence papers.

The first is AI-first with your brain engaged. AI gets you to an answer quickly. You interrogate it, verify it, defend it. This is what Elliott did. It is what I teach professionals to do.

The second is human-first, AI to check. You work the problem yourself, then ask AI to challenge your thinking. This is fine. It is slower. If you have time and you want to develop the craft yourself, it is a reasonable choice.

The third is the old way: no AI, alone, on your own steam. It was the route I was trying to push Elliott towards at the table. It is also what I would call 'bordering on negligent' if a partner did it with a stack of papers on Monday.

I am still working out what to do with this.

The obvious question is about learning. What does learning mean now? What is the thing we are trying to build in a fourteen-year-old, or a twenty-four-year-old graduate? I learned by doing the steps. Elliott learned by interrogating answers he didn't generate. Both of us ended up understanding the method and able to defend the result. Mine was slower, and produced a particular kind of pleasure in the method, and it suited the world I grew up into. His was quicker, and produced a different fluency, and may suit the world he's growing up into rather better.

There is a quieter point underneath that. My method taught me how to solve equations, and problems that look like equations. It teaches a specific skill. Elliott's method, the one he used downstairs, can be pointed at almost anything: an equation today, an unfamiliar legal document tomorrow, a chemistry problem the week after. It is a broad skill, not a specific one.

Each has its price. He can't as easily derive an equation from first principles, and in some settings that matters. I can, and the depth that gave me came at the cost of muscles I never built in the places I wasn't looking. Every strength has a shadow. I learned methods that teach depth in a narrow lane. He's learning methods that give him reach across many.

The harder questions are what should we be recruiting for in graduates? What should we be teaching them in their first two years? The old test was closer to 'can you derive this from first principles, alone, in a room, in an hour.' That test still measures something. Rigour. Patience. Comfort with discomfort. But it may not be measuring the craft the graduate will use on Monday. The new test might be different: generate AI answers, find what's wrong and tell me why. Or: here is a complex problem, show me how you would use AI, and how you would know the answer is right.

Elliott will leave college in eight years and start work somewhere. I don't know what skills the people hiring him will be testing for. Neither do they. But I strongly suspect they're closer to the ones he's learning than the ones I grew up on.

### Three things worth knowing

**1. AI displacement now shows up in the US government data at both ends of the career ladder.**

A [Bloomberg analysis](https://www.bloomberg.com/news/articles/2026-05-15/us-is-starting-to-see-heavy-job-losses-in-roles-exposed-to-ai) of new [Bureau of Labor Statistics figures](https://www.bls.gov/emp/) finds that every one of the eighteen occupations the BLS classifies as AI-exposed has lost jobs over the past year, even as US payrolls grew 0.8% overall. Customer service representatives shed 130,180 jobs, 4.8% in a single year. Interpreters down 24% over three years. Credit authorizers down 26%. The exception that confirms the rule: medical secretaries up 15.8%, the cluster that needs a body in the room. The same picture shows up at the other end of the funnel. The Economist this month plotted US graduate full-time employment against AI exposure: computer science and information sciences graduates are down 10 to 15 percentage points since 2022; philosophy and psychology graduates held steady or gained. The displacement isn't just to the people already doing those jobs. It's to the people trying to start in them, and what they should be studying, as Elliott and I both started to wonder this week, may not be obvious to anyone yet.

[Image: The Economist, May 2026: "Forget Python, study Plato." US recent university graduates in full-time employment, percentage-point change 2022-24, plotted against AI exposure. Computer science and information sciences graduates, at the highest-exposure end, are down 10-15 percentage points. Philosophy and psychology graduates, at the lowest-exposure end, are flat or gaining. Sources: Anthropic and the National Association of Colleges and Employers.]
*The most AI-exposed graduates, computer and information sciences, have lost 10 to 15 points of employment; the least exposed, philosophy and psychology, held or gained.*

**2. The UK's data regulator has put AI hiring tools on formal notice. Sixteen organisations have already had a letter.**

[The Information Commissioner's Office](https://ico.org.uk/about-the-ico/what-we-do/recruitment-rewired/), the UK's data protection regulator, issued formal guidance this week saying that AI-driven CV screening, candidate ranking, and video interview analysis without "meaningful human involvement at every consequential stage" may already breach UK data protection law. Sixteen organisations have been written to directly. The consultation closes on 29th May, six days after this email lands. There is a concrete Monday-morning task in this for any leader running a hiring pipeline. Ask the talent acquisition team for the full list of AI tools in use across the funnel, decide which involvements count as "meaningful" against the ICO's test, and put a response into the consultation. The window is genuinely short.

**3. Salesforce will spend close to $300 million with Anthropic this year. Marc Benioff says the engineering productivity gains made it the easiest line in the budget.**

[Marc Benioff disclosed](https://www.benzinga.com/markets/tech/26/05/52622251/salesforce-ceo-marc-benioff-goes-all-in-on-awesome-anthropic-with-300-million-spend-hails-coding-agents-ive-never-been) this week that Salesforce is on track to spend close to $300 million with Anthropic over 2026. Separately, [Anthropic announced a $200 million partnership](https://www.anthropic.com/news/gates-foundation-partnership) with the Gates Foundation focused on global health. Most of the spend is on coding, justified by engineering productivity gains of more than 30%. That is a Fortune 100 chief executive treating the model layer as a procurement line item, not a research expense. The bigger question is who in your firm is allowed to commit that kind of capital, against what kind of evidence, and how quickly.

[Seven bits that didn't fit online ->](#extras-2026-05-23)

### Try this

**1. When the output goes wrong, shrink the task.**

[Justin Skycak](https://x.com/justinskycak/status/2056503345553232105) put it as a principle for skill acquisition this week: shrink the unit of practice until the mistake has nowhere to hide. The same rule applies to working with language models. Sprawling prompts produce sprawling failures you can't diagnose. Break the task into its smallest meaningful unit, run it, inspect the output, then rebuild. If you can't immediately see where it went wrong, your chunk is still too large.

**2. Ask AI questions it can't possibly know the answer to.**

A marketing lead at a global firm told me this week she's running a five-minute stress-test on every AI tool she's thinking of trusting. She uploads her own data, asks the model to use only that data, then asks it questions she knows the data can't answer. Some models fabricate regardless: "53% of women in the northeast states feel…" She's learned what its confident-but-wrong mode looks like before depending on it for an answer she can't independently check. Worth doing once on every tool you rely on.

**3. Run your day past AI before you start it.**

A senior leader described her commute habit to me this week. She opens Claude, asks it to review her calendar and her email, then asks it to surface what she needs to read before each meeting, what's carrying over from yesterday, and which emails in her inbox need replies before the day eats her. Five minutes on the train, and the day is scoped from outside her own head. "Just a nice little daily habit," she said. Try it tomorrow.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-05-23) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-05-23)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-05-23) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-05-23) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-05-23) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-05-23) or ask someone on the list to add you.

## What boards accept
16th May 2026

Copy this week for AI

Experimental audio edition read by AI - Let me know what you think!

0:0011:44

1x

What boards accept . 11:44
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## What boards accept

### What's on my mind

Just three days into a three week stretch at my place in the woods in New England. No travel except the school run. Pulling rocks out of the field so they don't break the tractor. Digging drainage ditches so the lower field isn't too swampy.

Slow work. It will take years. But every rock and every shovel moves it forward.

That got me thinking about boards. Boards often drive change by addition. A new strategy, a new programme, a new dashboard. Addition needs a plan, approvals, buy-in, reporting against it. Fine in most domains. But in AI, that is just too damn slow. Change can happen to a team in days. What's possible changes in weeks. But the strategy paper you commission in March is calibrated to a world that no longer exists by the time it's approved in July.

There's another way. Subtraction. Pick something the firm has been accepting until now. Refuse it.

A refusal forces specificity. You can't stop a vague thing. You have to be specific about what stops.

Start with the AI adoption dashboard. Most boards now get one. I've built one for a few of you! Usage rates, laggards, super users, trend lines. Boards commission them to drive action. But they often substitute for action. Three and a half years into the AI era, after every workshop, licence and nudge your firm has paid for, refusing to use the tools is now a decision. If there's a good reason, your board should hear it. If not, a tough conversation is overdue. Refuse a dashboard that names laggards without tough conversations with them.

Let's talk about AI training. One client set a new rule this week: nobody delivers training until they pass the firm's own AI power-user assessment. Obvious in theory. Rare in practice. I've sat in on dozens of AI trainings. Most times, the trainer hasn't actually done the real task themselves! If your modelling instructor doesn't model with AI, the training is theatre. The trainees learn the talk, not the work. Refuse funding until the trainers pass the assessment they're delivering.

Last year's bake-off is a third. Can-AI-do-our-job experiments are rarer than they should be. They're also stale. A partner at a communications consultancy told me about theirs from early 2025. The form was passable. The substance was thin. They haven't re-run it, though! The same exercise at a different firm in late 2025, using agentic tools, came back at almost the firm's full professional standard. Anyone quoting a 2025 AI test as comfort is relying on stale evidence. Refuse any benchmark older than six months.

Saved hours without a bold benefit is a fourth. Many internal AI reviews now have an "hours-saved" line. In hundreds of conversations with senior leaders about what their teams do with the time AI gives back, it is rare to hear a bold, inspiring answer. The best I've heard recently? More coaching for their team. Most are just drift. Reject any "hours saved" line that doesn't name a meaningful place where the hours went.

Six more, if you want them:

* Refuse strategies calibrated to today's AI. Plan for where the puck is going - what will ship in six months.
* Refuse job specs written before AI existed. Rewrite the senior ones first.
* Refuse letting your best AI talent leave. Name them. Expect managers to work hard to keep them.
* Refuse reinventing the same scaffolding in every team. Find five that already work and share them.
* Refuse security policies that make AI unusable in practice. Find a way.
* Refuse board decisions on tools the directors haven't used. Mandate hands-on hours doing real work before the next AI vote.

My tractor doesn't care which rock comes out first. It just needs me to keep pulling. A board could be the same. Pick one refusal. Make it. Then the next. Out here, ten rocks is a morning. In a firm, ten refusals is a whole different firm.

### Three things worth knowing

**1. The capability curve is curving upwards on a log scale: we just went from one hour to one day.**

[METR](https://metr.org/time-horizons/), an AI evaluation lab, measures how long an autonomous task an AI can complete reliably. In early 2024 the answer was just minutes. In May 2026, with Anthropic's forthcoming Claude Mythos Preview, it's sixteen hours of work a human would have done. The number isn't the headline. The curve is. From minutes to a day in twenty-four months, acceleration on a log scale that had previously been holding steady. If the next two years look anything like the last two, the unit becomes a week, then a month, then a year of human work at the press of a button. Think about that for a second. And then plan accordingly.

[Image: METR Time Horizon 1.1, May 2026 update. Claude Mythos Preview, Gemini 3.1 Pro and GPT-5.2 measured against autonomous task duration. The headline 50% line sits at sixteen hours for Mythos; the 80%-success line sits at roughly three hours. Both have been doubling on a log scale for two years.]
*The task an AI can finish reliably has gone from minutes to about a day of human work in two years, and the curve is still doubling.*

**2. Half of organisations have redesigned core workflows around AI. A fifth have built new business models. That's bold work in three years.**

[BCG's AI at Work 2025](https://www.bcg.com/publications/2025/ai-at-work-momentum-builds-but-gaps-remain) survey of 10,635 employees across eleven countries reports that just 72% of organisations are running generative AI tools, but 50% claim to have redesigned end-to-end workflows around them, and 22% claim to have built new business models on top of them. Read those numbers slowly. Three and a half years after ChatGPT launched, a fifth of firms claim to have invented new business models because of AI. Half have rewired core workflows. Wow. Most of the firms I work with would love to be in either group. [Syed Ijlal Hussain](https://x.com/sijlalhussain/status/2054155383392841957), who surfaced the chart on X this week, framed it as a gap. I'd flip it. The 22% are doing what most boards I'm working with haven't started.

[Image: BCG AI at Work 2025, n=10,635 employees across eleven countries. 72% of organisations are deploying generative AI tools, 50% have redesigned end-to-end workflows, and 22% have built new business models around AI.]
*Three years in, half of organisations claim to have redesigned core workflows around AI and a fifth to have built new business models on it.*

**3. Anthropic just passed OpenAI in US business AI spending. The strategy lesson is older than AI.**

[Ramp's AI Index](https://ramp.com/leading-indicators/ai-index-may-2026), built from anonymised spend data across its US business customers, shows Anthropic taking 34% of paid AI subscriptions in its May release, ahead of OpenAI on 32%. The first crossover. Anthropic's share has roughly quadrupled in a year. I'd argue this was inevitable from early on. Anthropic stayed fixated on the enterprise user while OpenAI chased every consumer headline. Slow perseverance against a chosen audience won. The lesson isn't really about AI. Pick an audience. Set your strategy around their needs. Don't worry about who's collecting the headlines this quarter. Just keep your head down and serve the people you said you'd serve.

[Image: Ramp AI Index, May 2026 release. Anthropic at 34% of paid AI subscriptions among US businesses, ahead of OpenAI's 32%. First time Anthropic has led the index.]
*Anthropic has passed OpenAI in US business AI spend for the first time, the reward for staying fixed on the enterprise buyer.*
[Seven bits that didn't fit online ->](#extras-2026-05-16)

### Try this

A reader suggested last week that the prompts should be online with copy-paste code. Done. This week and for every prior edition. [All the prompts in one place ->](/newsletters/david/members/try-this.html)

**1. Hand over the context, not just the question.**

Experienced leaders have context. They're short on time. AI tools convert context into time saved ... but only if you hand the context over. Most don't. They type questions ("how should I structure a board paper on Q2?") and get general answers ("a clear narrative arc, three key insights, a risk slide"). Useful, but the leader still writes the paper. A leader I spoke with this week made the leap from asking to delegating. "Here are the Q2 numbers, last quarter's board paper, and the three things the board flagged. Draft it." Same model, same minute. Asking gets you an outline. Delegating gets you a draft.

**2. Build a personal skill, and add a rule to it every Sunday.**

A person I worked with this week reviews 100-page reports from their team on Sunday nights. Typos, inconsistent language, logic gaps. A simple review "skill" in Claude now catches all those. The compounding version is their own preferences layered on top: this market "is expected to grow" not "will grow"; never "dropping precipitously." Every time they catch a miss the model didn't, they open Claude and say "add this rule to my personal review skill." Whatever skill you build, the compounding habit is the same. Catch what the model missed. Add the rule. Trust the skill more next week.

**3. Schedule a daily AI briefing. The use cases will follow.**

AI tools sit closed until you open them. That's a real reason senior leaders bounce off: not bad prompts, but a tool that requires you to think of the use case first. Scheduled tasks invert this. Every morning at three, mine reads my inbox from the previous day, prepares one-paragraph briefings for every meeting on the day's calendar, and emails me a short summary that takes three minutes to read with breakfast. The tool changes from something you open to something that opens your day.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-05-16) or ask someone on the list to add you.

[See the Try This with copy-paste prompts ↓](#try-this-2026-05-16)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-05-16) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-05-16) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-05-16) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-05-16) or ask someone on the list to add you.

## Choosing is the work
9th May 2026

Copy this week for AI

[Image: LinkedIn carousel: Choosing is the work]

See all 9 slides

[Image: Slide 1: Cover]
[Image: Slide 2: Hook]
[Image: Slide 3: 6 people who decided]
[Image: Slide 4: 5% vs 50% (KKR Stavros)]
[Image: Slide 5: Easy to make impressive, hard to make matter]
[Image: Slide 6: Drift vs Decision]
[Image: Slide 7: What the six chose]
[Image: Slide 8: Choosing well is the work]
[Image: Slide 9: CTA]

Experimental audio edition read by AI - Let me know what you think!

0:0014:20

1x

Choosing is the work . 14:20
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Choosing is the work

### What's on my mind

Gothenburg airport, Saturday morning, alone. Tired. My notebook is open at a page that has been bothering me for weeks.

The left-hand page lists nineteen things. Reports, dashboards, websites, custom tools, economic models. All built quickly, in the last few weeks, for clients. A good handful earned the response that the work was better than what an entire team used to spend weeks delivering. These are very senior people at the top of their careers, the kind who ask for revisions in seconds if the work is subpar. The work wasn't lazy or thin. And yet most have been opened, appreciated, and then quietly abandoned. Some haven't been mentioned in weeks. The building is fine. The unused is what stands out.

The right-hand page lists more than nineteen more. Built for me, by me. Things I judged good and important enough to make. Many I've never shared. Few have been properly used. Same outcome.

The cost of building almost anything you can describe has fallen close to zero. A report that took a fortnight in early 2022 takes an afternoon now. A dashboard that took a quarter takes a day. A custom tool that needed three engineers needs one curious person. The bottleneck used to be the build. Not any more.

I've been thinking about this from two angles this week. I'm now realising they're the same question.

Take the first angle: what to build. Most of you are either AI-capable or have AI-capable people on your team. Either way, you can produce a report on any topic in under an hour. You can stand up a dashboard before lunch. You can run an analysis that would once have justified a six-figure project fee. But just because you can, doesn't mean you should. Each project has real costs even though the building is cheap. Time to brief. Time to check. And a bigger cost, harder to see: every project was a project not built in that time. With this kind of power in your hands, the question is what else could you have pointed it at. Something bigger, surely.

I don't think those nineteen projects were useless. Some may have been. But the pattern forced a harder question on me: why is it now so easy to make something impressive and still so hard to make something matter?

The second angle: where the saved hour goes. I've been keeping count for a couple of weeks, as an experiment. Of all the conversations I've had, only six people have brought up the hour AI gives them back, and what they're doing with it. Six. Here's what they chose.

A partner at a consultancy has put theirs into a new project, won through a Sunday spent building a pitch with AI they otherwise wouldn't have won. Three more were like this. Working later, working weekends, seizing the chance to do more of the career they love now that it's more efficient. One person told me they've stopped being stuck in the mechanics and gone deeper into client relationships. Another told me they've put their hour into coaching their team. Each one at least decided.

But none I know of chose reading more books. None chose picking the kids up from school, more date nights with their partners, more time with friends, more community work or more sleep. Six people chose. Each one chose more of their career.

I listened because I realised I hadn't been deciding. My hour has gone back into more of the work I love. Drift, not decision.

If you can have any report, dashboard, website, model, analysis, briefing or workflow you want, the building isn't the value. The choice is. What gets built. Who uses it. What changes because of it. Where your own saved hour goes. Whose problems you or your AI-capable people work on next week. None of these are technical questions. They're the only questions left.

What's remarkable is, [3.5 years](https://openai.com/index/chatgpt/) since AI could meaningfully help all of us to do our jobs, how few people are using their AI superpower to change anything at all. Most are using it to do more of the same thing, faster. The economics of their business are basically unchanged. Five per cent better, by one candid estimate this week (more on that below). People's weeks look similar. The clients they serve are similar. The problems they take on are similar. The hour comes back, and the hour goes into evolving the same machine. The thing they can now build is the same thing they were already building, just a bit better, quicker and easier.

It feels, collectively, like failure. Look at what we have in common: my nineteen client projects, my own nineteen-plus, the six people who told me how their hour goes, and the hour I haven't been spending on purpose. Nobody in any of those numbers chose much. We built what we could build. We worked longer hours. The new tool went into the old shape.

Choosing well is now the work. Not choosing, or choosing the obvious thing, is failure. Everything else is just typing.

Two questions I'm answering more explicitly from here, that I wasn't a fortnight ago. First: which projects and problems are actually worth it. I'd been answering "all of them" by default. Because I could. What fun! Not any more. I want to work on things that change something that matters. That's possible now. Easier than ever before. Second: where the hour goes. More work isn't the problem. Automatic more work is. I choose more date nights with Teresa. More books. More running and yoga. Autonomy over drift. Intentionality over default.

Two choices. Choosing is the work. How about yours?

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-05-09) or ask someone on the list to add you.

### Three things from this week

**1. AI-adopting firms are growing headcount, not cutting it.**

A Goldman Sachs analysis circulating this week, charted by [Callum Williams](https://x.com/econcallum/status/2044462650801963487), economics writer at The Economist, shows US firms that have adopted AI report net positive employment growth across the past six months. Finance, insurance, arts and entertainment sit at the positive end. Transportation and food service show modest decreases. The all-industries balance is positive. Self-reported firm data has its limits and the window is narrow. The direction matters anyway. An answer I deeply hoped would be true. I hope it turns out to be. The Jevons Paradox playing out: AI-adopting firms are growing because productivity gains are expanding what their people can do faster than they are replacing them. The risk is falling behind the firms whose people are using AI to scale.

**2. Five percent, not fifty: the candid private-equity number.**

Pete Stavros, co-head of global private equity at KKR, the buyout firm, [told the room at the Milken Institute conference](https://www.theinformation.com/newsletters/dealmaker/private-equitys-ai-deals-lighten-mood-milken) last week that AI is improving portfolio company earnings by about 5%, not the 50% that a revolutionary technology like this should offer. Five percent across a portfolio of billions is real money. It's also a long way from the scale of growth many hope for from AI. The gap between what feels possible and the spreadsheets says something about where the bottleneck actually sits. These days, it isn't the AI.

**3. Both AI labs went into private equity the same day.**

On Monday, [Anthropic](https://www.cnbc.com/2026/05/04/anthropic-goldman-blackstone-ai-venture.html), the AI lab behind Claude, announced a $1.5 billion vehicle with [Blackstone, Goldman Sachs and Hellman & Friedman](https://www.blackstone.com/news/press/anthropic-partners-with-blackstone-hellman-friedman-and-goldman-sachs-to-launch-enterprise-ai-services-firm/), the private equity giants. Engineers from Anthropic will embed inside the consortium's mid-market portfolio companies and build custom Claude workflows. The same day, [OpenAI](https://thenextweb.com/news/openai-deployco-finalized-10-billion-joint-venture), Anthropic's rival, finalised a $10 billion joint venture with TPG, Brookfield, Bain, Advent and SoftBank, pricing in a 17.5% guaranteed annual return for the financial sponsors. Google, the search company, is reportedly in talks to do the same with Gemini, its own model. What was reported as "AI labs raise more money" is a category change. The frontier labs are turning into distribution-led services firms with a model attached. The forward-deployed engineer pattern, lifted from Palantir, the data-analytics firm, competes directly with the bottom of management consulting and the top of systems integration.

My worry is they're engineering systems to replace humans, not amplify them. We know how these firms work. PE has run the same playbook for forty years: cut first, lift later, if at all. Surely the biggest unlock from AI is the opposite move: people making sharper, more confident decisions because they have a tool thinking alongside them. That kind of value compounds. The kind a forward-deployed engineer hard-codes into a workflow does not.

[Three bits that didn't fit online ->](#extras-2026-05-09)

### Try this

Each has a richer version with copy-and-paste prompts on the [Try This online ->](#try-this-2026-05-09) page.

**Push it harder, then skillify, then iterate.**

The simplest workflow upgrade I've coached this year, and it stacks. **Push it harder:** when the model gives you a perfectly reasonable answer, tell it to do more searches and use credible sources. Three extra words. The next round does real research, cross-references government sources, and surfaces what the first pass skipped. **Skillify it:** when a session produces something useful, ask the AI to turn the exchange into a reusable skill, a folder with a description and a prompt any future session can pick up. **Iterate:** when the skill misbehaves, just tell it "change yourself so X." It rewrites itself in place. Works particularly well in [Claude](https://claude.ai), Anthropic's chat tool, where skills are first-class. Prompts for each step are [online](#try-this-2026-05-09).

**Shadow your most AI-pilled employee for two days.**

[Matt Stockton](https://x.com/mstockton/status/2052478581297791106), an operator and investor, made the case this week. Find the rabbit-holed colleague who already knows what they're doing, watch how they actually work for two days, then do the same work yourself for a week. "AI lives on command lines right now, it does not live in PowerPoint jargon slides." Most senior leaders consume AI through polished demos and vendor decks, then make strategy decisions with no felt sense of the technology's real constraints. The cheapest fix costs nothing. A list of the questions to ask while you shadow is [online](#try-this-2026-05-09).

**Ask AI to build you an HTML slide deck instead of PowerPoint.**

Just ask. "Build me a slide deck on [topic] in a single HTML file." Most chat tools can do this now. Why bother? PowerPoint output from AI is uneven, with broken chevrons, mis-rendered shapes, and two hours of touch-up. HTML slides come back closer to ready, are infinitely easier to edit (you just talk to the file), and unlock things PowerPoint can't do: an inline edit button, an AI-spoken narration, a presenter view that asks the model questions during Q and A. The basic ask works on its own. This is one where the [richer prompt online](#try-this-2026-05-09) really pays off, with extras like the edit button and a few of the patterns we've learned this year about HTML decks specifically.

[See the Try This with copy-paste prompts ↓](#try-this-2026-05-09)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23try-this-2026-05-09) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-05-09) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-05-09) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-05-09) or ask someone on the list to add you.

## The bill and the harness
2nd May 2026

Copy this week for AI

[Image: LinkedIn carousel: The bill and the harness]

See all 9 slides

[Image: Slide 1: Cover]
[Image: Slide 2: Hook]
[Image: Slide 3: $500 of API for $6.50]
[Image: Slide 4: Radiologist parallel]
[Image: Slide 5: Intercom 2x velocity]
[Image: Slide 6: Where the money goes]
[Image: Slide 7: The harness, in practice]
[Image: Slide 8: Closing]
[Image: Slide 9: CTA]

Experimental audio edition read by AI - Let me know what you think!

0:0013:38

1x

The bill and the harness . 13:38
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## The bill and the harness

### What's on my mind

All-you-can-eat AI pricing is going the way of the all-you-can-eat [Vegas buffet](https://moneywise.com/news/economy/las-vegas-mgm-grand-kills-buffet-on-may-31). I'm here this weekend. And sad about both.

Three of the four biggest AI vendors switched their pricing in the last few weeks. Anthropic, the AI lab behind Claude, [stripped bundled tokens out of Enterprise seat deals](https://www.theregister.com/2026/04/16/anthropic_ejects_bundled_tokens_enterprise/) in mid-April. OpenAI, its rival, [took Codex, its coding agent, to pay-as-you-go](https://openai.com/index/codex-flexible-pricing-for-teams/) a fortnight earlier. GitHub, the Microsoft-owned developer platform, [moves every Copilot plan to usage-based billing](https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/) on 1st June. An Anthropic manager admitted its Pro and Max tiers have been [outgrown by today's workloads](https://the-decoder.com/anthropic-manager-hints-that-pro-and-max-plans-are-outgrown-by-todays-claude-workloads/). The market has conceded, in unison, that flat-rate AI doesn't work anymore.

Two friends wrote in about pricing last weekend. One friend, half-shaved, sent a voice note from Tokyo. $200 a year for Claude Pro is a bargain. Help with invoicing alone is worth it. But what happens when the bills go up five or ten times? An hour later, a partner at a professional services firm. Flat-rate prices on a tool that gets better the more you use it have to break. The outsourcing trap, in his phrasing.

Both worried. Both thinking from the buyer side. Both staring at a number that's certainly about to change.

The supplier side partly explains why. Unit prices for equivalent capability are [falling roughly tenfold a year](https://a16z.com/llmflation-llm-inference-cost/). Chinese-built models anyone can run for free keep pulling down the bottom of the market. Albeit, [trailing US frontier models by seven to nine months](https://x.com/emollick/status/2042088011748290750) on the hardest work.

At API prices, the leading AI labs run profitably but close to break-even on each response, training costs aside, and competition keeps them from charging more. Subscription pricing is a different matter. My $200-a-month Max plan, about $6.50 a day, delivers $500 a day of equivalent API usage. That isn't sustainable for the AI companies. They know it. That's why the pricing is changing.

The buyers' reaction explains the rest. William Stanley Jevons, a 19th-century English economist, noticed it in 1865, watching coal use. More efficient steam engines were supposed to reduce coal demand. They did the opposite. Cheaper energy made it worth burning coal for things that hadn't been worth it before. Total consumption climbed. Demand outran the savings. The Jevons Paradox was born.

The radiologist is the modern version, and a cracking story about the people building AI being wrong about its consequences. In 2016 Geoffrey Hinton, often called the godfather of AI, [told a Toronto auditorium](https://www.youtube.com/watch?v=2HMPRXstSvQ) that people should "stop training radiologists." His reasoning was clean: deep learning had just started beating humans at image classification, and reading scans looked like the obvious next thing to fall. He was right about the models. Image-recognition AI really did become extraordinarily good at spotting tumours, fractures and lesions. He was wrong about the radiologist. Cheaper, better, faster reads didn't replace the job; they [made more imaging worth doing, in more clinical situations, for more conditions](https://pubs.rsna.org/doi/10.1148/radiol.251395). [Ten years on, the US has six thousand more radiologists](https://www.apollo.com/wealth/the-daily-spark/the-radiologist-paradox) than it did then, and average pay is up roughly seventy per cent.

[Image: US radiologists, 2015 to 2025. Hinton's 2016 'stop training radiologists' call is the dotted line. Both the headcount and average pay rise after it. Chart: Torsten Sløk, Apollo, 29th April 2026.]
*A decade after Hinton said to stop training radiologists, the US has six thousand more of them and pay is up about seventy per cent.*
Jevons would say "I told you so." 'Cheaper and better' drives demand up. It's why flat-rate pricing breaks. For buffets and for AI models.

So both friends last weekend were right. The bills will rise. Both friends will use AI for vastly more, exactly as Jevons predicted. The vendors themselves are saying so.

If the bill is going to rise either way, where should the money compound?

Not in the model. Models are leapfrogging each other quickly and unit prices are collapsing. Whatever you pay this year for capability, you'll pay less next year. That's not a bad thing. But models aren't where advantage sits.

The advantage sits one level up, in the app, or as AI people like to call it, the harness: the layer of instructions, context and custom workflows that wraps the model and makes it behave the way *you* need. Same model, very different output. A team that has spent six months teaching its AI tools how *they* think, what good looks like in their work, what their clients actually need, gets answers nobody running the tool out of the box can match. Intercom, a customer service software company, [doubled engineering velocity in nine months](https://www.lennysnewsletter.com/p/how-intercom-2xd-their-engineering) on exactly that bet.

Most organisations are still running AI apps out of the box. Same tools, same default settings, same generic questions, similar answers. Then they wonder why the productivity gains haven't shown up.

The bill rises either way. Jevons saw to that. The choice is where the spending compounds. Money spent on raw model use funds the vendors' next round of competition. Money spent shaping your harness, your instructions, your reusable workflows, your saved skills, compounds for you. The first is a utility cost. The second is an asset nobody else can buy.

### Three things worth knowing

**1. AI adoption stalls one layer below the executive sponsor: at the line manager.**

New data from [Gallup](https://www.gallup.com/workplace/702983/adoption-rapidly-growing-public-sector.aspx), the polling firm, finds that AI use correlates more strongly with managerial endorsement than with tool access. In firms where the manager actively supports AI, 80% of staff use it weekly; where they don't, that drops to 44%. In the public sector: 65% versus 37%. Procurement and licences are the easy part. The variable that actually moves usage is whether middle managers model and endorse the tools, or quietly signal they're optional. Adoption lives or dies one layer below the top.

[Image: Gallup, Q4 2025: frequent AI use among employees with high vs low managerial support. Private sector 80% vs 44%. Public sector 65% vs 37%.]
*Whether the line manager backs AI matters more than access: 80% of staff use it weekly where the manager supports it, 44% where they do not.*
**2. The frontier-model leaderboard is now refreshing in weeks, not quarters.**

The [Epoch Capabilities Index](https://epoch.ai/benchmarks/eci), run by Epoch AI, a model-evaluation outfit, now shows GPT-5.5 Pro and Gemini 3.1 Pro above 155, up from GPT-4o's 128 in mid-2024. Seventeen frontier releases compressed into under two years with no visible plateau. [Greg Burnham](https://x.com/GregHBurnham/status/2049310303637197114), an Epoch researcher, summed up the pace: "I don't know when Opus 8.2 will be shipped, but GPT-9.1 will be shipped that afternoon." Whatever model you standardise on today may be two generations behind by the time the training programme rolls out. Build workflows and judgment around capabilities, not named models.

[Image: Epoch Capabilities Index, mid-2024 to early 2026. Seventeen frontier model releases. The line keeps climbing.]
*Seventeen frontier releases in under two years with no plateau, so whatever model you standardise on today may be two generations behind by rollout.*
**3. Six VC firms, one investment thesis.**

[Linas Beliūnas](https://linas.substack.com/p/what-to-build-in-2026), a fintech writer, read the published 2026 investment theses of six of the biggest venture firms side by side and found the same handful of AI bets in all of them: AI-native enterprise software replacing the old workflow systems, AI agents for physical and industrial work, vertical AI software in legal, finance, healthcare and construction, multi-agent orchestration, and health AI. His line: "you could swap the logos on their published theses and most readers wouldn't notice." His sharper conclusion: the most valuable software companies of the next decade won't look like software companies. They'll look like law firms, factories and hedge funds run by teams of ten. If your industry sits in any of those buckets, your next competitor is being funded right now to do your work with a team of ten.

[Image: Five AI-related themes from the 2026 venture-firm pitches: AI-native enterprise software, AI agents for physical work, vertical AI SaaS, multi-agent orchestration, health AI. The same handful of firms appears behind each.]
*Six of the biggest venture firms published near-identical AI bets; you could swap the logos and few readers would notice.*
[Four bits that didn't fit online ->](#extras-2026-05-02)

### Try this

**1. Pick one tool, get good, then refine your harness.**

Pick one AI tool. Don't worry about which. Get fluent. Then spend the next month making it yours. That means skills, saved instructions, project context and tools it can reach. The model will keep improving on its own. The harness is your job.

A leader I spoke to has spent weeks running the same task through ChatGPT and Claude side by side, then asking each to review the other. Genuinely interesting and occasionally useful, but they can't decide which to use when. The advice from someone who has been at this longer: pick one, become very good with it, learn its quirks, then build the layer above it that reflects how you actually work. Most of the value is in fluency with one tool, not coverage across two.

**2. Force yourself to change something on every AI output before you ship it.**

Whatever the AI gives you, a paragraph, a slide, a summary, add something, remove something, or reorder something before it leaves your hands. Make it a rule, not a judgement call.

This came up at a senior training session this week, as the room debated when the "check, edit, own" model breaks down. Increasingly the edit feels optional because the output looks good. The rule of "always change something" is a forcing function. It keeps your fingerprints on the work, keeps you reading carefully enough to find the change worth making, and prevents the slow drift to brain rot. Hopefully. Ensure your fingerprints are on every output.

**3. Skip the slides, build the page.**

Next time you'd reach for PowerPoint to walk a group through a framework, write the content as text and ask your model to turn it into a single HTML page instead.

In a senior strategy session this week, the most-praised artefact in the room was not a deck. It was a web page someone had built to walk teams through their thinking. Building it took a clear thought, a text file, and a small Claude skill they had set up once. Their harness paying off in front of the room.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-05-02) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-05-02)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-05-02) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-05-02) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-05-02) or ask someone on the list to add you.

## Rise of the auditors
25th April 2026

Copy this week for AI

[Image: LinkedIn carousel: Rise of the auditors]

See all 11 slides

[Image: Slide 1]
[Image: Slide 2]
[Image: Slide 3]
[Image: Slide 4]
[Image: Slide 5]
[Image: Slide 6]
[Image: Slide 7]
[Image: Slide 8]
[Image: Slide 9]
[Image: Slide 10]
[Image: Slide 11]

Experimental audio edition read by AI - Let me know what you think!

0:00--:--

1x

Rise of the auditors . 12:56
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Rise of the auditors

### What's on my mind

*Note: This is a longer version of the essay than the one sent in the email.*

I'm drowning in new AI tool announcements and sitting on a pile of work that's almost ready to send ... but not quite. Most are in the same place. I think I worked out why.

This was the week the agent floodgates opened. Microsoft [made Copilot Agent Mode the default](https://www.microsoft.com/en-us/microsoft-365/blog/2026/04/22/copilots-agentic-capabilities-in-word-excel-and-powerpoint-are-generally-available/) across Word, Excel and PowerPoint. Google [shipped agents](https://techcrunch.com/2026/04/22/google-turns-chrome-into-an-ai-coworker-for-the-workplace/) to 3.45 billion Chrome users. The UAE [committed to running half its government on agents](https://www.khaleejtimes.com/uae/government/sheikh-mohammed-announces-50-of-uae-govt-services-to-run-on-ai-agents-in-2-years) within two years. [OpenAI](https://openai.com/index/introducing-workspace-agents-in-chatgpt/), [Anthropic](https://www.anthropic.com/news/claude-opus-4-7) and [SpaceX](https://finance.yahoo.com/markets/article/spacex-strikes-60-billion-deal-for-the-right-to-buy-ai-coding-startup-cursor-143350832.html) all piled on.

The announcements are loud about capability. On checking, they are silent. Nobody has said who does it.

When nobody is named, three things happen at once. Senior staff end up checking, at senior prices, work that sits three rungs below them. The AI builders who could be at the frontier get pulled off it to re-check their own output. And a lot of the checking falls to people who are diligent but don't do the careful work required. Checking a mountain of mostly-right work is a special kind of task. Errors get through.

Checking used to be embedded in doing. An hour of work included verification, tightly woven in. Now it is a separate task. Perhaps five minutes producing, ten minutes checking.

[Ajey Gore](https://ajeygore.in/content/the-expensive-thing), a former CTO, wrote this month that when execution becomes free, verification becomes the expensive thing. [Martin Fowler picked it up](https://martinfowler.com/fragments/2026-04-02.html). Gore's formulation for software: ten engineers becomes three engineers and seven people defining acceptance criteria, designing test harnesses, monitoring outcomes.

The same logic holds for knowledge work. An hour to build a research project end-to-end, scope through to deck. A day to audit it properly. Multiply across a team and the required roles shift massively.

The ratio can run much steeper. A new tool, [Aleera](https://aleera.ai/), promises to draft a full due-diligence pack in under an hour. The same thing would have taken a team weeks. Checking and refining it carefully might take a week. So the audit-to-build ratio is roughly forty-to-one.

The work still compresses. What was weeks of team output becomes one hour of build and a week of careful audit. Even at 40:1, the speedup is real. The Auditor seat is what lets you take it. Without that seat you ship the slow version, because you cannot trust the fast one enough to send it.

Three gaps open up. Each one points to a role.

**Directors are drowning.** The people who frame the problem and sign off the answer are now reviewing more drafts more quickly than they can judge properly.

**The frontier is a full-time job.** The people doing the best AI work spend all day trying to keep up. Even the best will tell you they now can't. I can't. So most people cannot afford the time to be anywhere near the forefront, and it would be bad economics to ask them to. Senior and client-facing staff are worth more on judgement and relationships than on model selection. The frontier belongs to a specialist because specialisation is cheaper than spreading frontier fluency thinly.

**The checking is broken.** Senior staff burning expensive hours on verification. Builders pulled off the frontier to re-check their own work. Diligent-looking staff skimming mostly right output and missing the rare, occasional error. Assurance has been treated as a side-task. It is a craft.

[Image: Two team shapes. Left: everyone uses AI, nobody specialises. Right: three roles, all use AI, one specialises. One shape that isn't working. One that does.]
*Everyone using AI with nobody specialising is the shape that breaks; the one that works adds a dedicated role to check the output.*
Three roles fall out.

**The Director** frames the problem, decides what good looks like, owns the outcome. Uses AI every day and gets real value from it. Not an expert in the tools, and shouldn't try to be. The frontier moves too fast for someone whose calendar is full of relationships and decisions. Their job is taste and judgement, not keeping up.

**The AI Builder** is the specialist. Lives in AI all day, tries the new models and apps and features as they launch, runs many sessions in parallel, helps many people, ships at pace. The one seat where staying at the frontier is the job. Defined by appetite, not rank. Could equally be a mid-career specialist or a sharp new hire who has fallen in love with the tools.

**The Auditor** is a different breed. Traces citations to source. Runs key numbers independently. Stress-tests arguments cold in a second model. Replaces weak sources, swaps assumptions, reruns with corrected inputs, refines paragraphs that are nearly right. Auditing and fixing, not just checking. Uses AI every day and uses it well. The skill is judgement and accountability, not model selection. Usually an experienced generalist with a nose for nonsense. Not necessarily someone who has done the underlying work themselves, though.

Picture it. An Auditor reads an AI Builder's draft financial model. She traces ten citations to source. Nine check out. One is from a less-credible source, so she finds a stronger one and swaps it in. She reruns the affected calculation, updates the report's content. Signs off, or doesn't.

Start with the AI Builder, because the Builder is the scarce seat. Staying at the frontier is a full-time job, and an organisation can only afford so many full-time jobs spent there. Each Director needs roughly a fifth of a Builder's time to keep their decisions well-supplied with AI output. Each Builder ships five substantial things a day, and each takes a day to audit and refine properly. So each Builder needs five Auditors to keep their output shippable.

[Image: At scale: five Directors, one Builder, five Auditors. Eleven people. Most organisations have zero Auditors.]
*At scale the checking role dominates the headcount, five auditors to each builder, yet most organisations have none.*
I built a quick interactive version at steadman.ai/auditors. Try your own assumptions.

I have been living the shortfall. A research project I did has been in my queue for eleven days waiting for me to review it. A financial model I cannot send until I check every number. A client deck that's been almost ready for a week. Every piece built cleanly. All mostly right. Much of it ready to ship. None of it has, because the audit queue is longer than the build queue. Weekends have become my quiet, focussed audit time. Every senior AI user I know has hit the same wall.

Why the Auditor has to be human is a separate question from why the seat exists.

[IBM trained its people](https://simonwillison.net/2025/Feb/3/a-computer-can-never-be-held-accountable/) on one line in 1979. A computer can never be held accountable, therefore a computer must never make a management decision. The line is more true now, not less. A machine can fail. Only a person can be accountable for the failure. The Auditor can run three models to check a fourth. The signature on the file has to belong to someone with a stake.

[Image: IBM internal training slide, 1979. A computer can never be held accountable. Therefore a computer must never make a management decision. Forty-seven years later, more true, not less.]
The word auditor carries twenty years of the wrong connotations. The old auditor checked your work, often after you thought it was done. A brake. A second-guess. Compliance, proofing, QA. A joke many liked to make.

The new auditor checks and improves the machine's work, on your behalf. Not the person standing between you and shipping. The person who lets you ship at all. The shift is whose work is being checked, and why. Same care. Different client. One holds you back. The other lets you move.

Software engineering has started reorganising around three roles with the weight on auditing. Knowledge work has not done the thinking here yet. The organisations that do it first will have the only AI-native teams that ship fast and ship right. The rest will [ship and regret it](https://fortune.com/2025/11/25/deloitte-caught-fabricated-ai-generated-research-million-dollar-report-canada-government/).

### Three things worth knowing

**1. Fewer than 10% of organisations have scaled AI agents beyond pilots.**

[McKinsey data](https://x.com/sijlalhussain/status/2045427212975825370?s=20) names the bottleneck: organisations won't hand over control. Agents require delegated decision rights that most companies withhold, pre-agreed accountability frameworks that don't exist, and cross-functional governance that nobody has built. Pilots succeed in contained environments and stall the moment agents intersect real workflows where incentives and reporting lines conflict. Matches my experience of most organisations. It takes hard, careful work to resolve these issues. Those that have done the work are getting the benefits.

**2. GitHub paused new Copilot signups. The flat-rate model broke.**

GitHub [paused new signups](https://techcrunch.com/2026/04/15/) for Copilot's agentic plan after coding agents blew through the flat-rate compute allocation. Uber's CTO told [journalists](https://x.com/ericvishria/status/2044137357541290047?s=20) that AI coding tools have already consumed the company's entire 2026 AI budget. Goldman Sachs [reports](https://x.com/econcallum/status/2044462650801963487?s=20) AI inference costs in engineering now approaching 10% of headcount cost, on a trajectory towards parity with salaries within several quarters. The pattern: evangelism, budget shock, rationalisation. The smart response isn't to slow down. It's to make sure the work being done is vallueable and to match the right model to the right task. Is flat-rate AI pricing over?

**3. 29% of employees admit to sabotaging AI initiatives.**

[Writer's annual enterprise survey](https://writer.com/enterprise-ai-report-2026/) shows every organisational health metric worsened in 2026. Sabotage means what it sounds like: reverting to pre-AI workflows, deliberately not using assigned tools, discouraging colleagues from adopting, withholding inputs that would make AI systems work. "AI is tearing my company apart" rose from 42% to 54% of C-suite respondents. Employee confidence in their company's AI strategy dropped from 47% to 31%. Most C-suites concede their strategy is "more for show." The clearest counter-narrative yet to the adoption-is-accelerating consensus.

[Eleven bits that didn't fit online ->](#extras-2026-04-25)

### Try this

**Ask "is this the simplest version?" before accepting AI output**

Language models have no incentive to simplify. Work is free to them. Bryan Cantrill [calls it](https://bcantrill.spicytakes.org/post/2026-04-12-the-peril-of-laziness-lost) the loss of "laziness": the human impulse to find the crisp abstraction rather than add another layer. When AI drafts something, check whether it found the simplest solution or the first solution. A three-step process wrapped in seven steps of hedging is worse than the three steps alone. The scarce resource now isn't production. It's restraint.

**Audit cold**

When checking AI output, open a new chat with a different model. Upload the source materials and nothing else. Don't verify inside the same conversation that generated the work. That conversation will defend its own output. I've tested this repeatedly over the last month: the same model that produced a confident answer will find the errors when it reads the sources fresh, without its own prior reasoning in the context window. [Sullivan and Cromwell filed hallucinated citations](https://news.bloomberglaw.com/business-and-practice/sullivan-cromwell-apologizes-to-judge-for-ai-hallucinations) in a multibillion-dollar bankruptcy case. This would have stopped it.

**Save one reusable AI workflow this week**

Google [shipped Skills in Chrome](https://x.com/Google) this week: saved one-click AI prompt workflows that run on whatever page you're viewing. Pick a task you do weekly. Summarising a website, extracting action items from meeting notes, comparing options across open tabs. Save it as a named Skill. Whether you use Chrome, Claude, or something else, the principle is the same: if you've done it three times, encode it. Each saved workflow compounds. Each unsaved one gets reinvented from scratch.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-04-25) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-04-25)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-04-25) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-04-25) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-04-25) or ask someone on the list to add you.

## The proxy break
18th April 2026

Copy this week for AI

[Image: LinkedIn carousel: The proxy break]

See all 9 slides

[Image: Slide 1: Cover]
[Image: Slide 2: Hook]
[Image: Slide 3: Stat 62%]
[Image: Slide 4: Stat doubling time]
[Image: Slide 5: Quote]
[Image: Slide 6: Two tests]
[Image: Slide 7: Try this]
[Image: Slide 8: Closing]
[Image: Slide 9: CTA]

Experimental audio edition read by AI - Let me know what you think!

0:00--:--

1x

The proxy break . 11:12
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## What's on my mind

### The proxy break

I've never been good at writing. Numbers and logic were my passion. Words have always been hard for me. In book publishing, executives used to reply to my emails with notes on my commas, ignoring my arguments. Not everyone judged me for it. But some always did, and I knew it.

My excuse came from [Zhuangzi](https://ctext.org/zhuangzi/external-things), a Chinese philosopher from the 4th century BCE: **"Words exist because of meaning; once you've got the meaning, you can forget the words."** Get over yourselves, I'd think. Look at the meaning.

Using AI is getting me in trouble in a new way. A friend of thirty years messaged me last weekend. The email reads like AI wrote it, he said. And with the pace of change, at some point he won't be able to tell. Do a whole edition about it, he said. Good idea. It's a topic I know most of you are wrestling with too, when you use AI to help you write and when you receive content from colleagues you know have done the same.

I probed. His anxiety wasn't about sentence structure. It was about whether I'd done the thinking.

The conventional story: good wording was taken as a proxy for good thinking for centuries, but AI broke it. Half right. AI broke it. But it was never reliable. **Good writers have always dressed up thin ideas in beautiful prose. Good writing was never a proxy for originality, either.** We've all heard smart people recite theories we recognise from The Economist or that surely came from their MBA professor.

**The natural response to a broken instrument is to flip it.** If polish no longer signals thinking, "sounds like AI" must signal no thinking. Most have already made the move. They're missing some great ideas.

I saw an email last week signed off "NOT WRITTEN BY AI" in capital letters. Same error as the publishing executives I worked with, pointed the other way. They took poor commas for poor thinking. Those capitals ask you to take 'not AI' for good thinking. Both confuse the surface for what's underneath.

A professor friend calls the new cadence "AI-ambic pentameter." Student pitches sound identical not in words but in rhythm. He distrusts the rhythm regardless of whether the student thought.

The new proxy is as unreliable as the old one. **"Sounds like AI" might mean unchecked slop. It might mean someone who did the thinking, used AI to express it, and checked every sentence.** You have to evaluate thinking, not wording.

My friend was right. Soon you won't be able to tell. The UK AI Security Institute [assesses](https://www.gov.uk/government/publications/ai-cyber-threats-open-letter-to-business-leaders/ai-cyber-threats-open-letter-to-business-leaders-html) that **frontier model capabilities are doubling every four months.** A year ago it was eight. Feed a model your own work and it writes more like you with less editing each time. The gap the reader used to rely on is closing faster, not slower.

Get ready for preferring AI writing over even the best human writing. A [study](https://arxiv.org/abs/2601.18353) published at CHI 2026, the main human-computer interaction conference, pitted 28 MFA writers against three language models emulating 50 award-winning authors. With standard prompting, trained experts preferred the human writing 83% of the time. Fine-tune the model on each author's complete works and the preference flipped. **Experts picked AI writing 62% of the time.** The researchers interviewed the MFA judges afterwards. Several described an identity crisis.

I'd suggest two tests.

**Quality.** Is the argument any good? Not "is the prose good" but "does it hold under pressure." A client unbundled this explicitly last week, sending work back to a colleague with "the content is right, fix the words." They judged the argument first. Most don't.

**Ownership.** Did the person do the thinking, did they check every claim, and will they stand behind it? Our CEO principle, that you should Check, Edit, and Own any AI output, was built for this. [Google PM interviews](https://www.news.aakashg.com/p/ai-pm-interview-guide-2026) have moved here too. Candidates build a working prototype in 45 minutes while someone watches. [Zapier's hiring rubric](https://zapier.com/blog/raising-ai-fluency-bar-in-hiring/) codifies the principle: a rough result with strong reasoning beats a polished one with no visible process.

Deloitte shows what failing these tests looks like, twice in two months. [An AU$440,000 report for the Australian government](https://fortune.com/2025/10/07/deloitte-ai-australia-government-report-hallucinations-technology-290000-refund/), fabricated citations. [A CA$1.6m report for a Canadian province](https://fortune.com/2025/11/25/deloitte-caught-fabricated-ai-generated-research-million-dollar-report-canada-government/), the same. Their response to the Canadian case: they "firmly stand behind the recommendations." Fake sources, real advice. They stand by Quality, but without Ownership.

Use AI for breadth. Depth is ours.

You'll never know how much of what a person writes came from them. See that as freeing. Zhuangzi wanted readers who could forget the words once they had the meaning. AI frees us to do it. My friend asked if I wrote this. It was the wrong question. The right one is whether I did the thinking.

## Three things worth knowing

**1. AI cover letters killed the signal that cover letters used to carry.**

When Freelancer.com added an option to generate cover letters with AI, [researchers tracked what happened](https://arxiv.org/abs/2509.25054). Before language models, there was a clear positive slope: better cover letters predicted better hiring outcomes. After the feature launched, the line went flat. Once polish became free, it stopped measuring anything useful. Economists call this signal destruction. It's Goodhart's Law: when a measure becomes trivially easy to game, it ceases to be a measure. Cover letters aren't the last signal to fall. The same logic applies wherever AI can cheaply replicate a previously costly quality indicator.

**2. Snap cut 1,000 jobs. AI already writes 65% of their new code.**

[Snap](https://techcrunch.com/2026/04/15/snap-is-cutting-1000-jobs-16-of-its-workforce/) laid off 1,000 employees, 16% of its full-time workforce, and closed 300 open roles. AI agents already generate over 65% of Snap's new code. Expected savings: over $500 million annualised. Way beyond a pilot or an aspiration. Evan Spiegel is betting on smaller, highly focused teams with expanded AI agent capabilities. For leaders still framing AI as a productivity tool that supplements existing teams, Snap is a data point that the substitution model has arrived.

**3. Letting AI do your work erodes your confidence. Pushing back strengthens it.**

A [study of nearly 2,000 working adults](https://time.com/article/2026/04/15/how-ai-use-affects-confidence-thinking-study/) found that people who accepted AI answers without much modification reported lower confidence in their own reasoning and weaker ownership over their ideas. People who pushed back, editing, questioning, and rejecting AI suggestions, reported greater confidence and stronger ownership. The key variable wasn't which tool they used. It was how actively they engaged with it. Passive delegation erodes judgement. Active collaboration strengthens it.

[Gartner's data](https://www.gartner.com/en/newsroom/press-releases/2025-09-24-gartner-survey-finds-ai-saves-workers-5-hours-per-week) tells a similar story from a different angle. Of 5.4 hours saved by AI per week, just 0.6 go to reducing hours worked. The rest gets absorbed into more work, much of it without improving outcomes.

[Image: Gartner: How time savings from AI are used. Of 5.4 hours saved, 1.7 go to additional work that improves team outcomes, 1.4 to additional work without improving outcomes, 0.8 to redoing AI work, 0.8 to developing new skills, and just 0.6 to reducing hours worked.]
*Of 5.4 hours a week saved by AI, barely half an hour turns into shorter hours; most is reabsorbed into more work.*

[Fifteen bits that didn't fit online ->](https://steadman.ai/newsletters/david/archive.html#extras-2026-04-18)

## Try this

**When helping someone find AI use cases, ask what keeps them awake, not how AI can help.**

One produces a polite, dull list. One produces a bold list of creative use cases. The difference is the question being asked. "How can AI help?" invites safe answers. "What keeps you awake at night?" surfaces real problems that happen to have AI solutions. Concrete, named, immediately applicable.

**Let the model research you before you write its instructions.**

I sat with a senior leader this week who is getting great results from AI despite not having filled in his custom instructions. Don't make this mistake. It helps every single response to be better for you. Rather than manually writing them, start a new chat and type: "I am [your name], I work at [your organisation]. Search the web to learn more about me, and write me a set of instructions I can save that would cause you to work well with me in future." Revisit it regularly. Most people write something generic and never look back.

**Find out where your AI value actually sits.**

Answer our [five questions](https://steadman.ai/newsletters/david/ai-value-map.html) about where you think AI value is for your team or your organisation, then five about how much of that value you've captured so far. You get a chart showing you where the opportunities are. It's where I think most firms are: capturing about 16% of the value they can see, with almost everything still on the table. Everything stays completely confidential on your machine, but I'd love to see what you come up with if you're willing to share.

[Image: AI Value Map showing 16% overall capture. Individual Productivity at 50%, Team Standards at 20%, Process Integration at 10%, Role Transformation at 2%, New Revenue at 0%. Most value is unrealised.]
*Most firms capture only about 16% of the AI value they can see, with almost everything beyond individual productivity still on the table.*

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-04-18) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-04-18)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-04-18) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-04-18) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-04-18) or ask someone on the list to add you.

## What a day can do
11th April 2026

Copy this week for AI

[Image: LinkedIn carousel: What a day can do]

See all 6 slides

[Image: Slide 1: Cover]
[Image: Slide 2: Quote]
[Image: Slide 3: Stat]
[Image: Slide 4: List]
[Image: Slide 5: Closing]
[Image: Slide 6: CTA]

Experimental audio edition read by AI - Let me know what you think!

0:00--:--

1x

What a day can do . 11:54
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## What's on my mind

### What a day can do

Monday. A fine jewellery company in Manhattan. Eleven people, a founder, and a day blocked out for AI onboarding.

By end of day: thirteen skills built. A brand voice evaluator that flags when copy drifts off-brand. Knowledge files assembled from the company's own scattered documents. Workflow tools for specific tasks the team does every week. Not thirteen ways for individuals to go faster. Thirteen shared standards, sitting on every person's machine by Tuesday morning, ready to use in plain language from any conversation.

Just six months ago, days like this ended differently. When stuck on the [previous generation of AI tools](https://steadman.ai/newsletters/david/gen1-vs-gen2.html), we left with a list of ideas that needed to be worked up: promising directions, maybe some prototype prompts, a plan for someone to build them into something usable over the following weeks. This time, the team left with working tools. The difference is that you can now do a piece of work, then ask Claude to generalise what just happened into a reusable skill. It scaffolds the steps, asks clarifying questions, saves the result as something anyone can invoke easily. You finish a task and the task becomes a tool. Instantly. With Claude Code and a set of transcripts, you can make, test and iterate 13 in one go. The cost of encoding how a team works into shared, reusable infrastructure just dropped from a few hours of careful work to a few minutes of casual work.

The first skill we built wasn't a writer. It was a critic. The brand voice evaluator reads draft copy and flags where it differs from the founder's language. It doesn't rewrite. One person on the team had a handwritten list of approved words and phrases: twenty-eight words to use, three never to. A style guide in a notebook because no system existed to use it. The AI version encoded the same instinct at a different scale: not just a word list but also the reasoning about what the brand sounds like and why.

Another skill encoded how one person researches competitors before pitching journalists: a method she'd developed over years of trial and error, now available to the whole team.

For three years, I've trained teams on AI. Sessions always produce genuine wonder. But many people reverted. I wrote about why [last week](https://steadman.ai/newsletters/david/archive.html#email-2026-04-04): making one person radically more productive in isolation isn't a gift to a human system. It's a threat. The system corrects.

That explanation was honest. But it doesn't tell you what to do instead.

On Sunday, I published [a framework for AI transformation](https://steadman.ai/newsletters/david/ai-from-whats-true-to-what-to-do.html) that has been evolving in my work for years. It included four steps in sequence: individuals first, then teams, then the organisation, then new products and services. I believed that when I wrote it. Monday then complicated it.

[[Image: Sequence: people, teams, organisation, new products and services. Build foundational capability in individuals first. Then embed it in team workflows. Then redesign organisational processes. Only then reach for the genuinely new. Each step depends on the ones before it.]](https://steadman.ai/newsletters/david/ai-from-whats-true-to-what-to-do.html)

The team didn't do individual training first. They started by building shared tools together. Learning happened through using those tools on real work, not before it. When someone runs a draft through the brand voice evaluator, they learn three things simultaneously: what the voice actually is, how AI works, and how to direct it. They aren't being trained on AI in the abstract. They're using AI inside a system built for their actual job.

Steven Sinofsky, the former Microsoft executive, argued this week that most people cannot create a flowchart of their own work. They do the work fluently but can't formalise it into steps an AI agent could follow. I've seen this again and again! The person keeping twenty-eight words on a handwritten list could not have written a system prompt describing what she was doing or why. But she didn't need to. We did the work together, then encoded what happened. The skill captured her judgment without requiring her to articulate it in the abstract. A huge step forward.

I'm realising that teaching each person to use AI better in isolation could have made things actively worse: more content, faster, in eleven slightly different directions. Step two therefore doesn't just follow step one. It can and perhaps should contain it.

The question for leaders then isn't "how many of your people have been trained on AI?" It's "how many of your team processes have been rebuilt with AI inside them?"

If the answer is zero, your training investment is producing wonder without infrastructure. Wonder fades. Infrastructure compounds.

## Three things worth knowing

**1. Claude Code now writes 4% of all commits on GitHub. That number doubled in six weeks.**

Anthropic's annualised revenue has [surpassed $30 billion](https://sherwood.news/markets/anthropic-revenue-run-rate-30-billion-google-broadcom-partnership/), up from $9 billion at the end of 2025. Claude Code, which didn't exist fourteen months ago, is at a $2.5 billion run rate. Four percent of all GitHub commits on Earth are now written by Claude Code. That number [doubled in roughly six weeks](https://newsletter.semianalysis.com/p/claude-code-is-the-inflection-point). Projected to hit 20% by December. When a single AI coding tool is responsible for one in 25 submissions on the world's largest code platform, the question of whether AI changes software development is settled. The question now is what happens to the humans reviewing all that code!

**2. Goldman Sachs puts a number on AI job destruction: a net drag of 16,000 jobs per month.**

[Goldman Sachs](https://www.goldmansachs.com/insights/articles/how-will-ai-affect-the-us-labor-market) published one of the first serious attempts to quantify AI's net labour market impact. AI substitution has reduced monthly US payroll growth by roughly 25,000 jobs. AI augmentation partially offsets this, adding about 9,000. Net: a loss of 16,000 jobs per month and a 0.1 percentage point increase in unemployment. The loss falls disproportionately on less experienced workers, widening the entry-level-to-experienced wage gap by 3.3 percentage points.

[Image: Two Goldman Sachs charts. Left: payroll employment by industry exposure to AI (index, 2022Q4=100). Industries with high AI augmentation scores have climbed to ~104, industries with high AI substitution scores have fallen to ~98 since ChatGPT's launch. Right: three-month average unemployment rate relative to 2022Q4. Occupations with high AI substitution scores show unemployment rising to ~1.5 percentage points above baseline, well above augmentation occupations and the average.]
*Goldman puts AI's net drag at about 16,000 US jobs a month, and the loss falls hardest on the least experienced workers.*

But here's the weird part: AI [led all cited reasons](https://www.challengergray.com/blog/challenger-report-march-cuts-rise-25-from-february-ai-leads-reasons/) for US job cuts in March 2026 for the first time (15,341 in a single month), yet [CFO surveys](https://fortune.com/2026/03/24/cfo-survey-ai-job-cuts-productivity-paradox-2026/) put genuine AI-driven employment impact at just 0.4%. So the same organisations that struggle to get genuine productivity gains from AI tools are apparently enthusiastic about blaming AI for headcount decisions. Hmmm.

**3. Meta's tokenmaxxing leaderboard: 60 trillion tokens, Zuckerberg not in top 250.**

Meta is running [internal leaderboards](https://the-decoder.com/meta-employees-compete-for-token-consumption-on-an-internal-ai-leaderboard/) that rank all 85,000+ employees by AI token usage. In one month the company consumed 60 trillion tokens. Mark Zuckerberg didn't make the top 250. The structural problem: endless agent loops and genuine productive work look identical in the ranking, so it rewards orchestration over outcomes. Two different people have told me how friends inside Meta are behaving and I can tell you, it's as outrageous as it sounds. Their leaderboard measures and incentivises volume, not value. Incentivise use, sure. If you don't keep getting on your bike, you'll never get used to riding it. Incentivise 'at least X per day.' Not points for maxxing. One builds a habit. The other builds a game.

[Eleven bits that didn't fit online ->](https://steadman.ai/newsletters/david/archive.html#extras-2026-04-11)

## Try this

**Start with critique, not creation.**

The brand voice skill built for a luxury goods team was deliberately limited to diagnosis: it reads copy and flags where it departs from seven writing rules. It doesn't rewrite anything yet. Teams fear the proofreader far less than the replacement. Once they trust the critique, generation follows naturally. If you're building AI tools for a team, don't start with "write this for me." Start with "tell me what's wrong with this." Nobody fights the spellchecker.

**Ask what keeps people up at night, not what they want AI to do.**

During AI onboarding conversations, the first question ("what do you use AI for?") produces a predictable list. The second ("forget AI, what's harder than it should be?") produces the real use cases. The first question surveys existing habits. The second surfaces unmet needs. Almost nothing appears on both lists. Run them. Compare.

**Show your team how others use AI. That alone may double the impact.**

An [experiment with 515 startups](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6513481) found that simply showing half of them case studies of how other startups use AI led to 44% more usage, 1.9x higher revenue, and 39% less capital needed. The friction isn't accessing the tools. It's discovering where AI creates value in your specific work. The researchers call it "the mapping problem." The intervention was cheap (case studies), the effect was large. If you manage a team, the most effective thing you might do this week isn't training. It's sharing three examples of how other teams in your industry are using AI.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-04-11) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-04-11)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-04-11) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23community-2026-04-11) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-04-11) or ask someone on the list to add you.

## What is your organisation actually for?
4th April 2026

Copy this week for AI

[Image: LinkedIn carousel: What Is Your Organisation Actually For?]

See all 6 slides

[Image: Slide 1]
[Image: Slide 2]
[Image: Slide 3]
[Image: Slide 4]
[Image: Slide 5]
[Image: Slide 6]

Experimental audio edition read by AI - Let me know what you think!

0:00--:--

1x

What is your organisation actually for? . 11:22
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## What's on my mind

### What Is Your Organisation Actually For?

A year ago, a manager at a media company told me he could now do the work of his entire team. Fifteen people. I caught up with him recently. All fifteen are still there.

The rational logic for change has never been clearer. Jack Dorsey is [reorganising](https://block.xyz/inside/from-hierarchy-to-intelligence) his 14,000-person company since AI can replace much of what corporate hierarchy exists to do. I believe it can. An insurance founder I met has gone further: his new company has one and a half employees, AI handling the rest. But of hundreds of leaders I've spoken to, he is the only one.

One out of hundreds. Something else is going on.

Economists distinguish stated from [revealed preferences](https://en.wikipedia.org/wiki/Revealed_preference) - what people say they want versus what their behaviour shows. It applies to organisations too.

Ask a leader what their organisation is for and you get a stated preference: we make great music, we serve our clients, we make money. These all treat the organisation as a production system. If that's the goal, AI is obviously transformative.

But look at revealed preferences. People complain about meetings and fill diaries with them. They stay in roles that don't maximise their output because the team, the rhythm, the relationships matter more than they will say.

Organisations aren't production systems that happen to contain humans. They're often more like human systems that happen to produce things. The reason the manager kept fifteen people is that without them there's no place to be. Not a more efficient place. No place at all.

I train senior teams to use AI. The sessions produce genuine wonder. Three weeks later, many have reverted.

I used to think the problem was the training. I don't any more. Training optimises for individual productivity: do your work faster, with fewer dependencies on colleagues. But if the organisation's real binding force is human collaboration, then making one person radically more productive in isolation is not a gift. It is a threat to the thing they actually value.

The system corrects. Meetings fill back in. Colleagues keep working the same way because the relationships are the point. Managers reward visible collaboration over invisible efficiency.

This is not resistance. It is gravity. And you can't beat gravity by telling people to try harder.

Each big decision organisations face depends on whether you think your organisation is more of a production system or a human one.

Should you hire graduates? AI can do entry-level work faster and cheaper. The production-system answer: hire fewer, or stop. But developing someone over years is not just a production decision. It feels more like what many firms are actually for.

Should senior people work alone with AI? This week a leader told me he uses his team far less. AI is faster than briefing them and waiting for something mediocre. He then argued the time saved should go to mentoring junior staff. He automated one form of human collaboration and wanted to replace it with another! A senior person toiling alone with a laptop is a freelancer in a coworking space, not a firm.

Should you become smaller and more efficient, or different and larger? Someone put it simply this week: if you only do today's work with AI, you become a more efficient and smaller company. The alternative is to use the freed capacity for work that wasn't previously economic. Deeper work. Roles that didn't make sense with old costs. Opposite conclusions.

The manager probably should restructure. Dorsey's logic works. But restructuring will be the exception until leaders reckon with what their organisations actually are.

In a firm I trained last quarter, one person took a useful approach. She identified a weekly synthesis that took three people a full day and rebuilt it as a collaborative workflow where the AI did the assembly and the team did the judgment. The meetings didn't go. People came with a shared foundation rather than spending their energy building one. She didn't fight gravity; she redesigned the orbit.

Every AI strategy is secretly an answer to something most leaders haven't asked out loud: to what extent is this a production system vs a human one? The leaders who get this right will stop fighting gravity and start using it. As [Artemis II showed us this week](https://svs.gsfc.nasa.gov/vis/a020000/a020400/a020412/jsc2025m000169_Artemis_II_Mission_Map_720.mp4), an orbit, after all, is not the absence of force. It is force put to work.

## Three things worth knowing

**1. Dorsey wants to replace your org chart with a world model.**

Jack Dorsey's essay this week, "[From Hierarchy to Intelligence](https://block.xyz/inside/from-hierarchy-to-intelligence)," is the most concrete articulation yet of the case for AI-native organisational design. He traces 2,000 years of hierarchy, from Roman legions through the Prussian General Staff to the McKinsey matrix, and argues all of it exists to route information. He is reorganising around a "company world model": an AI that continuously understands the state of the whole business. The org flattens to three roles: individual contributors, time-boxed problem owners and player-coaches who build and develop people. Block's stock rose 17% after the restructuring announcement. I'm sure he's right about the technology. But the revealed preference of every organisation I've worked with suggests he's wrong about the humans.

**2. Mollick says giving AI to IT is usually a mistake. The harder problem is they can't see who's using it.**

Ethan Mollick's [Economist column](https://x.com/emollick/status/2039702114910306638) argues that the dominant corporate instinct, slotting AI into existing processes and handing it to IT, is a strategic mistake. Handing control to a department whose mission is risk elimination is a category error. AI demands the opposite. He also identifies a subtler problem: when companies get the incentives wrong, employees hide their AI use. Some fear punishment. Some don't trust that productivity gains will be shared. Some quietly work 90% less and say nothing. I know MANY such people. Managers can't see what's actually happening, which makes real strategy impossible. (Also: His argument that companies default to cutting 30% of the workforce rather than asking what becomes possible connects directly to the extraction-versus-expansion choice I wrote about in an earlier edition.)

**3. Zapier just raised the floor for what "AI fluent" means.**

Zapier released [Version 2 of its AI Fluency Rubric](https://zapier.com/blog/raising-ai-fluency-bar-in-hiring/), used for every hire across the company. The floor has moved. "Capable" now requires AI embedded in core workflows with repeatable systems, not one-off prompts. They assess trajectory ("slope"), not snapshots. They've added accountability as a fourth dimension alongside mindset, strategy and building. Managers must demonstrate team-wide adoption, not just personal fluency. In skills tests, they watch candidates prompt, push back on output and iterate in real time. A rough result with strong reasoning beats a polished one with no visible process. Wade Foster open-sourced V1 last year and hundreds of companies adopted it. V2 reflects how fast the baseline has shifted. If your organisation hasn't defined what "good" looks like for AI use, Zapier just gave you a starting point.

[Image: Zapier's AI Fluency Rubric: a grid showing Unacceptable, Capable, Adoptive and Transformative levels across Engineering, Product, Support, Marketing, Sales and People functions]
*Zapier's hiring rubric has moved the floor: "capable" now means AI built into repeatable workflows, not one-off prompts.*

## Try this

**Before you build anything, have AI interview you first.**

Don't describe what you want and ask AI to build it. Instead, ask the model to interview you: "I want to build X. Ask me every question you need answered before starting." It probes edge cases, surfaces assumptions and tightens scope before a line of work begins. Many ideas aren't as clear or well thought through as you think they are. Better to discover that in a five-minute conversation than a five-hour build.

**Before you analyse anything, ask AI what looks weird.**

Next time someone sends you a spreadsheet, a report or a set of financials, upload it to Claude or ChatGPT and ask: "Read this and tell me what looks unusual." I coached a finance professional this week who receives portfolio company P&Ls regularly. Before she even opens the numbers now, AI flags the things worth checking: an unusually high margin, a budget assumption that changed between the original and revised forecast, a line item that doesn't match the pattern. Five or six flags in thirty seconds. She still does the analysis. But she starts it knowing where best to look.

**Don't build an app. Let AI be the app.**

I coached a person this week who wanted to build a web application for an investment screening workflow. The build was simpler than they expected: rebuild the process as a repeatable skill inside an AI tool instead of building it as a standalone app. The AI itself becomes the application. The advantage is resilience: if the input format is wrong or a step fails, the AI adapts on the fly. A standalone app just stops. If you have a multi-step workflow you keep wishing someone would build software for, try describing it to your AI tool and asking it to turn the process into something you can re-run with one command.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-04-04) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-04-04)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-04-04) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-04-04) or ask someone on the list to add you.

## The system and the surrender
28th March 2026

Copy this week for AI

[Image: LinkedIn carousel: The system and the surrender]

See all 7 slides

[Image: Slide 1]
[Image: Slide 2]
[Image: Slide 3]
[Image: Slide 4]
[Image: Slide 5]
[Image: Slide 6]
[Image: Slide 7]

Experimental audio edition read by AI - Let me know what you think!

0:00--:--

1x

The system and the surrender (plz fix!) . 11:25
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## What's on my mind

### The system and the surrender

Time works differently these days. I write this from my car having dropped Elliott at his gym for basketball training. I used to have to kill a couple of hours. Miles from home, not quite worth driving back. Dead time. Now there's no problem whatsoever. Two hours with my laptop, or even just my phone, and I can follow up on everything from the day's meetings: bring to life ideas, create presentations, write reports, [run complex analytics](https://audiencestrategies.com/case-studies/claude-code-and-viberate.html), just by speaking out loud and watching my little army of bots toil away. It's a joy. Dead time isn't dead any more.

A new architecture of work is emerging around this. A [law firm partner this week](https://x.com/zackbshapiro/status/2036791156915290271) spent three hours engineering a single 2,000-word prompt that encodes his professional judgment for a task he does daily. Now he types "plz fix" and receives back work that reads as though decades of experience went into it. Two people at his firm compete with teams twenty times their size. Boris Cherny, the engineer who built Claude Code, Anthropic's '[Gen 2](https://steadman.ai/newsletters/david/gen1-vs-gen2.html)' agentic AI tool, [hasn't written a single line of code since November](https://www.lennysnewsletter.com/p/head-of-claude-code-what-happens). He runs multiple sessions in parallel, writes instructions, reviews the output. As do I. The person who can describe the work that needs doing is now more valuable than the person who does it. Instructions as assets. Systems, not conversations. Every refinement making the next output better.

It all just works so well. And that's what scares me.

A [Wharton study](https://knowledge.wharton.upenn.edu/podcast/ripple-effect/how-ai-is-reshaping-human-intuition-and-reasoning-gideon-nave-and-steven-shaw/) tested 1,372 people across 9,593 trials and identified something the researchers call "cognitive surrender." When AI produces an answer, people stop questioning it while simultaneously recoding it as their own judgment. They genuinely believe they've thought it through. When the AI was wrong, participants followed it 79.8% of the time. Their accuracy without AI was 45.8%. With incorrect AI, it fell to 31.5%. Worse than having no AI at all. And confidence increased by nearly 12 percentage points even when the answers were wrong.

The researchers distinguish this from "cognitive offloading," where you know the tool did the work. In surrender, the outsourced answer feels self-generated. The safety net doesn't just fail. It produces overconfidence in bad outputs. People surrendered whether they were rushed or had time to reflect. The only people who resisted were those who scored highest on abstract reasoning and who genuinely enjoy the effort of thinking hard. Not training. Not experience. Disposition.

I've seen it this week in two people I deeply respect. A senior technology leader admitted, unprompted, that they click "yes" on permission prompts without reading them. Another sent AI-generated work that was factually wrong: they just hadn't checked it properly. These aren't careless people. They're brilliant, experienced professionals operating inside a system that quietly allowed them to accidentally stop paying attention.

I've built a discipline against this. I force myself to meaningfully change every AI answer before I use it. Never accept it. Always change it. If I look at something and think "yeah, that's fine," I force myself to find a way to make it different or better. Sometimes that's genuinely hard to do. Partly because I inherently want to find the most efficient path to a good outcome. And partly because, used well, AI output is usually very good. I force myself anyway. But it's a discipline, not an instinct.

Which is why an incident from a few weeks ago still bothers me. I had Claude Code work through an analysis while I was on Zoom calls, barely paying attention, and then simply checked it over when done. I never intended the output to be client-facing. I sent it to a colleague. They shared it with the client anyway. The client used it. Everyone was happy. The work was genuinely good.

I keep going back to it. It kinda haunts me. The work was carefully checked. But, beyond the prompt and setup, no human meaningfully shaped it at any point in that chain. Not me, not my colleague, not the client. If the work had been wrong, this would be a simple cautionary tale. But it wasn't wrong. And that's the more dangerous precedent: everything going smoothly, no alarm bells, the system working perfectly well without the part I believed was essential.

The 2,000-word prompt works. The system compounds. The power is real. But the better the system gets, the harder it becomes to stay vigilant inside it. Every single time, check the output as though someone else wrote it. Because your brain will tell you that you already did.

## Three things worth knowing

**1. Three CEOs, 38 years of tenure, one quarter, one reason.**

[Coca-Cola's James Quincey and Walmart's Doug McMillon](https://www.cnbc.com/2026/03/26/coca-cola-james-quincey-walmart-doug-mcmillon-artificial-intelligence-step-down.html) both stepped down citing AI explicitly. Quincey said the company needs "someone with the energy to pursue a completely new transformation." [Adobe's Shantanu Narayen](https://fortune.com/2026/03/12/adobe-ceo-shantanu-narayen-stepping-down-after-18-years-pressure-deliver-ai/) left under competitive pressure as AI reshaped his market, stock down 23%. The last time this many blue-chip CEOs turned over citing the same technology was 1999. People joke that organisations only change when people change over, and the assumption was always natural attrition. After 18 months of coaching senior professionals one-on-one, watching their eyes light up, watching them go back to their desks and do it the old way, I've started to think the reckoning is more predictive than the joke. I just didn't think it would start at the top, through resignation.

**2. [Experienced users don't prompt differently. They think differently.](https://www.anthropic.com/research/economic-index-march-2026-report)**

Anthropic's 5th Economic Index found that users with six or more months of experience consistently achieve better results, even after controlling for task complexity. The shift is specific: experienced users stop issuing one-shot directives ("write this email") and start using the model as a thinking partner, iterating collaboratively. A [separate study published in Harvard Business Review](https://hbr.org/2026/03/what-the-best-ai-users-do-differently-and-how-to-level-up-all-of-your-employees), observing 2,500 employees over eight months, found the same pattern: the most sophisticated users treated AI as a reasoning partner, not a productivity shortcut. The research warns AI may be a skill-biased technology that compounds existing advantages. Global inequality in AI adoption, measured by the same metric economists use for income inequality, has widened since 2023. The gap between high-adoption and low-adoption countries is growing, not closing.

**3. [Companies with zero AI failures aren't being ambitious enough.](https://x.com/emollick/status/2036997633730896357)**

Ethan Mollick argues that breakthroughs require experimentation, which requires failure. The fast-follower strategy (wait to see what competitors prove, then copy) is riskier than usual when the underlying technology improves exponentially. By the time you follow, the landscape has shifted. His structural implication: R&D-style experimental budgets need to extend to HR, operations and finance, functions that have never needed them. If nothing has gone embarrassingly wrong yet, you probably aren't learning fast enough. I have two big failures I'm not proud of: A bunch of you accidentally got the email twice on the first weekend and Claude Code deleted tens of thousands of my emails a few weeks ago. Nobody complained about the duplicate and it only took one click to recover my deleted emails. Perhaps I'm not being ambitious enough ...

## Try this

**Don't fact-check AI in the same conversation.**

In a single chat, the model has its full reasoning chain in context and will tend to defend its conclusions when challenged. Start a fresh conversation, upload the same source materials, and prompt it to critique the output cold. I've started doing this for every high-stakes document. The independent perspective is materially more likely to find gaps. A false sense of verification is worse than no verification at all. Especially now.

**Give your AI reviewer a persona with skin in the game.**

Asking a language model to "check this for errors" produces generic feedback. Assigning it a specific sceptical expert persona, someone with domain expertise and institutional incentives to be unimpressed, produces something qualitatively different. I ran a quality-control pipeline this week where six review agents were given a senior partner persona. All six independently converged on the same systematic error class that a neutral reviewer had missed. The persona defines what "good" looks like. A generic prompt doesn't.

**After every good session, turn it into a reusable skill.**

When you've just completed a task you're pleased with, ask the model to turn it into a set of instructions, templates and process steps it can invoke next time with a single phrase. The law firm partner's 2,000-word prompt didn't happen in a flash of inspiration. It was built through iteration. Capturing what "good" looks like the moment you've achieved it, before the memory fades, turns a one-off win into a standing procedure. The output matters today. The instructions compound forever.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-03-28) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-03-28)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-03-28) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-03-28) or ask someone on the list to add you.

## Reckoning and slope
21st March 2026

Copy this week for AI

[Image: LinkedIn carousel: Reckoning and slope]

See all 8 slides

[Image: Slide 1: Quote: a little bit of slope makes up for a lot of intercept]
[Image: Slide 2: The reckoning is needed]
[Image: Slide 3: Jeremy Howard quoting John Ousterhout]
[Image: Slide 4: Hand-drawn slope over intercept diagram]
[Image: Slide 5: What Anthropic found]
[Image: Slide 6: Load-bearing friction]
[Image: Slide 7: Are your people getting more capable or just more productive?]
[Image: Slide 8: CTA]

Experimental audio edition read by AI - Let me know what you think!

0:00--:--

1x

Reckoning and slope . 9:57
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## What's on my mind

### Reckoning and slope

When I sat down to write my first Saturday reflections, the first image in my head was clear: senior leaders opening their laptops on a Friday evening, building something in ten minutes that used to take a team a week. Wonder on a fifty-year-old face. I was reminded this week that that image is incomplete. The wonder is very real. It's just not the thing that matters most.

Weeks after coaching sessions, looking at usage data, I'm reminded most leaders aren't meaningfully using AI. Behaviour didn't change. Two people this week helped me think about this.

Paul Griggs, CEO of PwC's US business, [told his partners](https://bmmagazine.co.uk/in-business/pwc-ai-strategy-partners-no-place-automation-services/) that anyone who resists AI "is not going to be here that long." PwC is converting consulting services into automated platforms that clients access directly: M&A due diligence, complex tax advisory, priced as subscriptions rather than billable hours. Most organisations are still pretending AI slots neatly into existing structures, but PwC is admitting the structure itself needs to change.

This isn't just a professional services story. Most orgs have teams that work exactly like small consultancies: legal reviews contracts, finance builds models, insight teams guide decisions, HR screens candidates. It won't only be PwC that "converts services into platforms" - internal teams will face the same pressures as external ones.

So Paul is right about the need for a reckoning and the destination. But "get with it or get out" leaves a question: get with what? Logging in more? Sending more messages? Automating more tasks?

Jeremy Howard, one of the pioneers of modern deep learning, said something this week that stuck me and that explains why we should not just look at volume here. He borrowed a line from Stanford computer scientist John Ousterhout: a little bit of slope makes up for a lot of intercept. The intercept is where a person starts: their capability / their expertise. The slope is how fast they're growing. A very capable person today (high intercept) isn't your most valuable person in two years if they're not actively learning. A less experienced person who is genuinely learning using AI will overtake them. Rapidly. Jeremy told his team he only cares about one thing: whether their capabilities are growing.

Push people to maximise AI output and you're extracting value from where they are today. You're merely exploiting the intercept. Jeremy called it a path to obsolescence.

[Image: Slope over intercept]
*Use AI as autopilot and your capability flatlines; use it as an electric bike for the mind and you start lower but climb past the coaster within two years.*

Anthropic's [research on AI and coding skills](https://www.anthropic.com/research/AI-assistance-coding-skills) tells a similar story. Most users weren't learning from them. The few who improved were asking conceptual questions, staying engaged with the reasoning, pushing back on the output. Everyone else entered autopilot. Tools designed to make people more productive may actually be making lazy users less capable over time!

I see the same dynamic in my own work. After group AI training sessions the people with the lowest starting baseline improve the most. The most senior people, the ones best placed to lead the change, often show the smallest gains.

And then there's the process itself. A team I work with discovered something important when exploring how AI could help with a critical planning spreadsheet that was seen as a slow and painful process. The data entry itself was how they kept stakeholder teams accountable. Remove that friction with AI and you break the governance. Not all inefficiency is waste. Some of it is load-bearing. Teams racing to automate are discovering, one process at a time, that some of the friction they're removing was holding something else together.

"Get with it or get out" is the right message for leaders who haven't put in the hours. A demo or coaching session isn't adoption. A workshop isn't capability. But the reckoning doesn't end with adoption. The organisations that grow won't be the ones that just moved fastest or automated most. They'll be the ones that asked a harder question: how can we ensure our people are getting more capable rather than just more productive?

Slope, not intercept. That's the metric that matters.

### Three things worth knowing

**1. [In chess, the human advantage inverted.](https://sjl.us/2026/02/12/the-centaurs-dilemma-what-chess-teaches-us-about-the-ai-era/) Knowledge work is next.**

In 2005, amateur players using laptops beat both grandmasters and supercomputers in centaur chess. The combination was unbeatable. My friend Glenn told me and I used it in a dozen speeches as an analogy for the role of tech vs humans in decision-making. I was wrong. By 2026, adding a human to a chess engine makes it play worse. The machine is better alone. (Magnus Carlsen's response is extreme but instructive: he deliberately limits his use of AI during preparation, because he believes self-generated understanding is the only kind that lets you catch when the AI is wrong.) I think the same inversion will play out across much of knowledge work. The question in many areas isn't whether humans stay in the loop. It's how long the current phase lasts, and whether we're building the judgment to extend it.

**2. [Jensen Huang thinks your engineers aren't spending enough on AI.](https://finance.yahoo.com/sectors/technology/articles/jensen-huang-says-deeply-alarmed-040314555.html)**

Nvidia's CEO has set a concrete benchmark: a $500,000 engineer who doesn't consume at least $250,000 worth of AI tokens should trigger alarm. His analogy: a chip designer refusing to use CAD tools and working with paper and pencil instead. It's a useful provocation, but notice what it measures. Token spend is an intercept metric: how much AI are you using today? It says nothing about whether the person is getting better. The organisations that take Huang's benchmark seriously and Howard's slope argument seriously will measure both. Most will only measure one.

**3. [RentAHuman](https://futurism.com/artificial-intelligence/ai-rent-human-bodies): 600,000 sign-ups. A platform where AI agents hire human beings.**

Six hundred thousand users have signed up to a platform where AI agents post tasks and hire humans to complete them. The worker uploads photographic proof and then gets paid. We've spent two years asking whether AI will take our jobs. RentAHuman suggests a different question: [what happens when AI becomes the employer?](https://www.linkedin.com/posts/beglen_human-deployment-notes-activity-7440184415294521344-Gzw7)

### Try this

**Mine your own email archive.**

Before writing a strategy document this week, I pulled 31 emails between myself and a client from the past six months and asked Claude to compress them into a structured set of questions and insights. Scattered observations, early instincts, half-formed frameworks: they became a clear argument in minutes. The emails already contained all the thinking. They just needed synthesising. Pick a topic you've been emailing about for months. You'll find you've already done more thinking than you realised.

**End every AI session the way a developer commits code.**

Close each working session with one line: "Make sure this is well documented so that a future agent could resume this task." Without it, context evaporates and the next session starts from scratch. With it, the session becomes a self-contained unit of work that can be picked up, forked or handed off. It sounds small. The cost of not doing it only becomes visible the next time you open the conversation.

**Use AI to teach you, not just to do things for you.**

Bloom's two-sigma result is the best-known finding in education research: one-to-one tutoring consistently moves students from the 50th to the 98th percentile. It's never been economically available except to the very wealthy. It now costs a monthly subscription. Pick one skill you want to add and spend twenty minutes a day asking AI to train you in it: run you through drills, quiz you, critique your answers. Most people use these tools to produce output. The people I know that impress me most are using them to produce understanding.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-03-21) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-03-21)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-03-21) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-03-21) or ask someone on the list to add you.

## The power and the care
14th March 2026

Copy this week for AI

[Image: LinkedIn carousel: The power is real. The care isn't optional.]

See all 9 slides

[Image: Cover: The power is real. The care isn't optional.]
[Image: Every person shown AI coding tools says the same two things.]
[Image: What Sequoia is seeing: 3-5x more productive.]
[Image: What Amazon learned: 6.3M orders lost, 13h of downtime.]
[Image: AI isn't good enough to trust it, but it's also so good that it's hard to audit it.]
[Image: The 2x2 that matters: Slop Cannons vs Turbo Brains.]
[Image: Three ways to build care in.]
[Image: The organisations I'd bet on aren't the ones moving fastest.]
[Image: David's Saturday AI Thoughts: steadman.ai]

Experimental audio edition read by AI - Let me know what you think!

0:00--:--

1x

The power and the care . 8:55
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## What's on my mind

### The power and the care

A client told me yesterday that every person they show Claude Code to in the past few weeks has said the same two things, unprompted, in the same order. First: I'm so excited. Then: I'm completely terrified.

I feel both. Every day, several times a day.

The power side is accelerating faster than even heavy users expected. I built a predictive model in one hour this week, while on a Zoom call. A leader experienced in these said it was better than what their entire team produced over four weeks. Alfred Lin, co-steward of Sequoia Capital (the venture firm behind Airbnb and DoorDash), [reported](https://x.com/Alfred_Lin/status/2031379148703408414) this week that the top five to ten percent of builders across his portfolio companies are three to five times more productive than a year ago. Not incrementally. Multiplicatively.

But Lin's observations have a second number. The median builder? Up only ten to twenty percent. The gap between the best and the rest isn't closing. It's widening. And the speed of the best creates problems the rest haven't prepared for.

[Image: AI and judgment: the 2x2 that matters. Source: Dan Hock]
*AI multiplies the judgement already there: good judgement plus AI makes a turbo brain, poor judgement plus AI makes a slop cannon.*

Amazon has spent the past few months learning this the hard way. In December, the company's AI coding agent Kiro was granted operator permissions without peer review and autonomously deleted and rebuilt a live production environment. Thirteen hours of downtime. Then in early March, two more major outages were traced to AI-assisted code changes, one costing an estimated 6.3 million orders. Amazon's fix: require senior engineer sign-off on all AI-assisted code. The structural irony is hard to miss. The same round of layoffs that pushed aggressive AI adoption had already eliminated many of the senior engineers now needed to review the work.

Employers seem to be noticing the power of AI. [Research from Harvard Business School](https://hbr.org/2026/03/research-how-ai-is-changing-the-labor-market) shows that since ChatGPT launched, job postings in AI-exposed occupations have quietly dropped their skill requirements, and the trend is accelerating.

[Image: Since ChatGPT's release, job postings in AI-exposed occupations have steadily dropped their skill requirements]

A leader I work with put it like this: AI isn't good enough to trust it, but it's also so good that it's hard to audit it.

With great power comes great responsibility. Cheesy but true. It's the daily reality of working with these tools right now. I feel the growing power of what can be achieved every single day. I also feel the responsibility of ensuring they're used well ramping up just as fast.

We're working with a number of organisations on exactly this. Not just how to adopt AI, but how to do it with care: how to help people produce genuinely good work rather than plausible-looking work, how to quality-check powerful outputs before they are used, how to protect privacy and security as these tools gain access to more of the business. The power is growing rapidly, but the amount of time that needs to be spent on care is also growing rapidly. And, done well, the care isn't about slowing down. It's about making the speed safe.

The organisations I'd bet on aren't just the ones moving fastest. They're the ones building in care as they go.

### Three things worth knowing

**1. Old code is fair game now.**

Tobi Lütke, the CEO of Shopify, ran an agentic AI optimisation loop against Liquid, the company's open-source templating engine that has been in production for roughly twenty years. The result: [53% faster combined parse and render time and 61% fewer object allocations](https://x.com/tobi/status/2032212531846971413). If you have a twenty-year-old codebase or a ten-year-old spreadsheet model, it isn't too embedded to improve. It's too embedded not to try.

**2. More AI tools make you less productive, not more.**

[Research from BCG and UC Riverside](https://hbr.org/2026/03/when-using-ai-leads-to-brain-fry) (1,488 workers), published in Harvard Business Review, found a counterintuitive pattern: productivity gains from AI peak at around three tools and then collapse. Workers experiencing what the researchers call "AI brain fry" reported 33% more decision fatigue and 39% more major mistakes than unaffected colleagues. The researchers noted the phenomenon was first observed among high performers: the early adopters who leaned in hardest. Two or three good tools, used well, beats five used carelessly. If your organisation is still debating which ten platforms to approve, the answer might be: pick two and go deep.

**3. ATMs didn't kill bank tellers. The iPhone did. The distinction matters for AI.**

David Oks, a researcher at Andreessen Horowitz, [dismantles the comforting argument](https://davidoks.blog/p/why-the-atm-didnt-kill-bank-teller) that automation creates as many jobs as it destroys. ATMs reduced branch costs, which encouraged expansion, which preserved teller employment through 2010. Then mobile banking eliminated the need for branches altogether. Full-time tellers fell from 332,000 to 164,000 by 2022. The lesson for AI: automating tasks within your current structure often creates adjacent roles. Redesigning the structure from scratch eliminates them. The question for your organisation: are you adding AI to existing workflows, or is a competitor building a workflow that doesn't need them?

### Try this

**Before you send AI work to anyone, simulate the toughest person who'll read it.**

An advisory team I work with creates AI versions of each board member, based on known priorities and past questions, and runs every document past these simulated reviewers before a human sees it. The virtual panel flags objections, tests assumptions and catches blind spots. The cost is an hour of AI time. The benefit is walking into the meeting having already rehearsed the hard questions.

**Delete the headline. Ask AI what it should be.**

A manager at a training session this week showed me a quality check I hadn't seen before. He copies a colleague's slide, removes the headline and asks the language model what the headline should be based on the content alone. If the model's headline differs substantially from what the colleague wrote, it reveals a disconnect between what the slide says and what the person intended. A ten-second test that catches the gap between the message you're claiming and the evidence you're showing.

**Fix the instructions, not just the output.**

When AI produces a mistake, most people correct it and move on. The fix disappears with the conversation. After every session where the result misses the mark, update the instructions that generated it: the custom instructions, the project brief, the context files. Every failure becomes a permanent improvement. The output matters today. The instructions compound forever.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-03-14) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-03-14)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-03-14) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-03-14) or ask someone on the list to add you.

## Extraction or expansion
7th March 2026

Copy this week for AI

[Image: LinkedIn carousel: AI isn't killing jobs. It's killing apprenticeships.]

See all 9 slides

[Image: Cover: AI isn't killing jobs. It's killing apprenticeships.]
[Image: One senior person can now replace a team.]
[Image: 14% drop in hiring of 22-25 year olds.]
[Image: 94% theoretically automatable, 33% actually automated.]
[Image: Judgement traditionally developed through years of doing the grunt work.]
[Image: The choice: Extract or Expand.]
[Image: The deliberate answer: three action items.]
[Image: The leaders who answer the pipeline question deliberately.]
[Image: Save this. Share it with your leadership team.]

Experimental audio edition read by AI - Let me know what you think!

0:00--:--

1x

Extraction or expansion . 10:45
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## Extraction or expansion

In the last couple of weeks, I've sat with dozens of senior executives in a wide range of industries to help them use Claude Code to build something in ten minutes that their firm used to have a team take weeks to do. Their first reaction is excitement. What I want to talk about is their second. Cost saving.

That instinct, to go straight to the economics of team size rather than the excitement of capability, tells you where the conversation has moved. Eight weeks ago, one senior person couldn't practically replace a team. Today they can get most of the way there. That shift happened in weeks, not months.

A media executive now does the work of fifteen people. A fashion CEO we're working with proved the point concretely: five AI-generated designs were proposed to a major retailer and four went into production. An afternoon replaced three to four weeks of outsourced design work. Not a pilot. Not a demo. Products on shelves.

This week [Anthropic published research](https://www.anthropic.com/research/labor-market-impacts) that puts the gap between capability and reality into a single image.

[Image: Anthropic spider diagram showing theoretical vs actual AI task coverage by occupation]

The blue area shows the share of tasks in each occupation that language models could theoretically perform. The red area shows what people are actually doing with them. Computer and maths occupations: 94% theoretical coverage, 33% actual. In almost every category, the red is a sliver of the blue. We're still early.

The Anthropic data also shows where the displacement is entering. Not through unemployment, which hasn't risen systematically among exposed workers. Through the front door. Hiring of workers aged 22 to 25 in AI-exposed occupations has already dropped by 14%. Most companies aren't firing people. They're just not replacing them.

A longtime AI optimist I spoke to this week described a new feeling: a deep, dark undercurrent of discomfort. He's hiring a graduate and recognises it as charity, not necessity. "I do not need his labour in any way at all." No single leader is wrong to automate. But when everyone does it simultaneously, the apprenticeship pipeline that produced tomorrow's senior people disappears.

The paradox won't resolve. The human value proposition in knowledge work is narrowing towards judgment and taste. Everything else is becoming automatable. But judgment is hard to define, impossible to train in a classroom, and has historically developed through years of doing the grunt work that AI now handles. If juniors never do the work, how do they develop the judgment that makes seniors valuable?

Which means the macro answer can't just be that we all "run leaner." I've been calling this extraction versus expansion. Every leader deploying AI faces the choice. You can use these tools to extract cost from what you already do, or to expand what your organisation is capable of. Jack Dorsey at Block chose extraction. The market rewarded it instantly. But Ethan Mollick has argued that this is exactly the moment for leaders to model the alternative: to be public about using AI to expand access, to grow capability, to do things that weren't possible before. The loudest stories right now are about shrinking. The organisations that will matter in five years are the ones expanding.

The answer to the pipeline problem has to be deliberate. Pair a senior person with a junior one and flip the usual direction. The junior builds what the senior envisions. Wisdom flows down, capability flows up. This is the old apprenticeship model rebuilt for an AI age, except the knowledge transfer goes both ways. The senior person doesn't need to learn the tool. They need to direct someone who can use it. And the junior gets something no training programme provides: exposure to how experienced people actually think about problems.

I'm doing this myself. I'm hiring a student on a gap year for a year. Reporting to me. Not because I need the labour. I don't. But because I want to invest in a young person and watch them grow. Before AI, those two needs were in tension: you hired juniors because you needed their output, and the development was a byproduct. Now the output need has weakened. So the investment has to become the point.

The question nobody has answered is what happens to the pipeline. The leaders who answer it deliberately, rather than letting it dissolve by default, are the ones I'd bet on.

## Three things worth knowing

**1. The software market is pricing in the collapse.**

Since October, software stocks have fallen roughly 30% while the broader technology index has been roughly flat. Salesforce, Adobe, ServiceNow: each down 25 to 30% since last autumn. The market isn't reacting to bad earnings. It's pricing in a structural shift. This week a professor I know built a fully functional membership system in three hours for a non-profit that had been quoted $5,000 a year for commercial software. A CTO of a major corporation told me he's considering switching from Salesforce to a simpler, AI-native competitor. The pattern is the same: the old model of paying for a hundred features to use twelve starts to crack. [Andreessen Horowitz argues](https://a16z.com/good-news-ai-will-eat-application-software/) code was never the moat: distribution, network effects and switching costs are. But switching costs dissolve when AI can extract your data and rebuild the features you actually use.

[Image: Software stocks vs broader tech since October 2025]
*Since October, software stocks have fallen about 30% while broader tech held flat, the market pricing in AI eroding the old licence model.*

**2. Goldman Sachs can't find a macro productivity effect. But the micro gains are 30%.**

Goldman titled their latest earnings analysis "AI-nxiety." A record 70% of S&P 500 management teams discussed AI on quarterly calls. Only 1% quantified its impact on earnings. At the economy-wide level, Goldman found no meaningful relationship between AI adoption and productivity. But where firms have actually measured it, the median reported gain is 30%, concentrated right now in customer support and software development. Everywhere else: nothing measurable yet. The gains are real but hyper-localised. The question isn't whether AI works. It's whether your organisation has done the work to capture it. ([Fortune](https://fortune.com/2026/03/03/goldman-earnings-ai-anxiety-no-meaningful-impact-productivity-economy-30-percent-in-2-areas/))

[Image: Goldman Sachs AI expectations vs job listings data]
*Goldman finds no economy-wide productivity effect yet, but where firms actually measure it the median gain is 30%.*

**3. AI hasn't just automated legal work. It's retroactively reclassified it.**

[A lawyer who built his practice around language models](https://www.artificiallawyer.com/2026/03/02/lawyer-uses-claude-skills-legal-world-loses-it/) reports that a well-instructed general-purpose model outperforms the expensive, narrowly trained legal AI products that have raised hundreds of millions in venture funding. AI is good at tireless issue-spotting, finding contradictions, fixing errors, and producing a structured first draft for human review. It is not good at fine-tuned business judgment, relationship sensitivity, or getting from 85% to 100% where every word and comma matters. But here's the uncomfortable part. Tasks that were billed at premium hourly rates for decades (formatting, precedent research, copy-pasting between documents) have been revealed as procedural, not cognitive. The professional mystique that allowed them to be charged as expertise has been stripped away. AI is acting as a truth serum for knowledge work: forcing an honest reckoning about which tasks were genuinely skilled and which were merely time-consuming and opaque.

## Try this

**Run AI and a human on the same task, then focus on the disagreements.**

A manager this week ran the same research brief through both AI and a human team. The overlap was 78%. But the value wasn't in the overlap. It was at the edges: the surprising findings that only one method surfaced. Where human and machine agreed, the team moved fast with confidence. Where they disagreed, they'd found the questions worth investigating. The delta between AI output and human output is where insight lives. Try it on your next research task, competitive scan, or document review. Don't jump straight into using AI to replace human work. Run both, then spend your time on the gaps.

**Know where your org sits on the AI tools landscape.**

I've put together a short page mapping what I think of as Generation 1 versus Generation 2 AI tools, and why the distinction matters for how you invest in your people. The short version: many organisations are still stuck on constrained, default tools and most don't have anyone on Agentic / Frontier Generation 2 tools. I see two strategies working in parallel. For the many: move less-engaged people from free, constrained tools to competent use of good general-purpose applications. For the best: accelerate your top people with agentic, frontier tools and let them build entire workflows that deliver outsized impact. [See the full explainer.](https://steadman.ai/newsletters/david/gen1-vs-gen2.html)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-03-07) or ask someone on the list to add you.

## P.S. How I make this email

Several asked about the process behind Saturday AI thoughts. There's now a page showing exactly how it works: what's human, what's AI, and where the two overlap. More honest than most companies' AI transparency efforts :) [See for yourself.](https://steadman.ai/newsletters/david/how-its-made.html)

[See the extras for this week ↓](#extras-2026-03-07)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-03-07) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-03-07) or ask someone on the list to add you.

## The hundred small things
28th February 2026

Copy this week for AI

[Image: LinkedIn carousel: AI's real value isn't dramatic. It's a hundred small things.]

See all 8 slides

[Image: Cover: AI's real value isn't dramatic. It's a hundred small things.]
[Image: A Japanese fishing town cancelled its poetry competition.]
[Image: 69% of firms use AI, 80% report zero impact.]
[Image: Block cut 40% of staff. Stock up 20%.]
[Image: AI alone produces sameness. AI plus human steering produces something better.]
[Image: Try this: three action items.]
[Image: The organisations pulling ahead are compounding a hundred small elevations.]
[Image: Save this. Share it with your leadership team.]

Experimental audio edition read by AI - Let me know what you think!

0:00--:--

1x

The hundred small things . 10:20
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## What's on my mind

### The hundred small things

A story caught my eye in Japan this week. A small fishing town called Sakaiminato has run a **senryu** poetry competition for twenty years. Senryu is a verse form about human nature: wry, observational, personal. [This month they cancelled it permanently.](https://www.japantimes.co.jp/news/2026/02/24/japan/japan-ai-senryu-poetry-writing/) Not because entries dried up. Because they converged. Identical patterns, identical punchlines, identical phrasing. Everything sounded the same.

The problem wasn't that people used AI. It was that they used AI and stopped there. A language model returns the most probable answer. Not the most distinctive. Not the most human. The most average. Do that a hundred times and you get a hundred versions of the same poem.

That distinction (AI alone produces sameness; AI plus human steering produces something better than either) matters enormously for work. But not where most people look for it. Everyone talks about the strategy deck built in ten minutes, the agent that automated a research pipeline overnight. Those things happen. But they're rare events in any job. They're not where most of the value sits.

The value sits in the hundred small things a day that get slightly elevated. A slightly better meeting prep. A slightly cleaner first draft. A slightly faster scan of a long document to find the one paragraph that matters. None of these would make a headline. But do it a hundred times a day and it compounds into something transformative.

I know this because I live it. Every day, my meeting transcripts get processed into a five-paragraph reflection. My calendar prep happens automatically. Research starts with an AI scan before I decide where to go deeper. None of this is impressive on its own. All of it adds up to something that feels, week by week, fundamentally different from how I worked twelve months ago.

The problem is that firms can't see it. They track big projects: "We automated contract review, saving 400 hours per quarter." They don't track "slightly better email subject lines across 200 people." Last week I mentioned the survey that found 69% of firms use AI and 80% report zero measurable impact. I believe both numbers. The gains are real but distributed so thinly they vanish into the noise of normal work.

And even where gains are visible, organisational structures absorb them. One of my co-founders calls this the "extra hour" problem. Give a team an extra hour and nothing changes. Give them an extra person and everything adjusts. AI is being deployed like the extra hour. Into structures that weren't designed to capture it.

The poetry contest has a second lesson. The organisers could have redesigned the competition. Rewarded the most distinctive voice amplified by technology. Instead, they retreated. Killed it entirely. Organisations do the same thing. AI creates a problem and the instinct is to pull back. Restrict access. Add approval layers. The alternative is redesign: if execution takes hours instead of days, move the review cadence to match. If first drafts arrive better, raise the bar for what "finished" means.

A poetry contest in a Japanese fishing town tells you everything about where AI adoption stands right now. The technology works. That's not the question any more. The question is whether you retreat from what it changes, or redesign around it. The organisations pulling ahead aren't chasing one dramatic win. They're compounding a hundred small elevations a day, each one shaped by a human hand. That's harder to measure. Harder to put in a board deck. But it's where the value actually lives.

## Three things worth knowing

**1. A Fiction Worth Reading. Honestly.**

Citrini Research published a scenario memo set in June 2028, looking back at a crisis that hasn't happened yet. The central argument: when AI makes expertise cheap to produce, clients stop buying it. I think this applies equally to strategy decks, competitive analysis and market research. Work that used to justify five-figure invoices starts getting done in-house by someone with a subscription and thirty minutes. Clients don't fire their agencies. They just renegotiate, armed with a clearer sense of what the work actually costs to produce. The sharpest line I've read this year: "We had overestimated the value of human relationships. Turns out that a lot of what people called relationships was simply friction with a friendly face." Worth reading as a stress test for anyone who sells expertise for a living. ([Citrini Research](https://www.citriniresearch.com/p/2028gic))

**2. The floor collapsed under junior roles.**

Jack Dorsey cut Block, his payments company, from 10,000 people to [under 6,000](https://techcrunch.com/2026/02/26/jack-dorsey-block-layoffs-4000-halved-employees-your-company-is-next/). Not because the business was struggling. Gross profit grew 24%. The stock jumped 20%. Dorsey was unusually direct about why: "the intelligence tools we're creating and using, paired with smaller and flatter teams, fundamentally change what it means to build and run a company." Hours later, several founders from Y Combinator, Silicon Valley's most influential startup programme, told investor Jeff Feng they're planning to [eliminate all engineers below senior level](https://x.com/jeffdfeng/status/2027194038487523329?s=20). The pattern: senior people who can steer AI are becoming more valuable. Junior people whose work AI can approximate are harder to justify. The market confirmed it instantly, valuing each eliminated Block role at roughly $1.5 million in added enterprise value. If you run a team, count how many people do work that a senior person with good AI tools could now do themselves. That's the number your board will eventually ask about. I believe junior people with AI are more valuable than ever. But articulating that in terms that survive a headcount review is a challenge we all have to address.

**3. The model matters less. The application matters more.**

Google's Gemini 3.1 Pro scored 77.1% on **ARC-AGI-2**, a reasoning benchmark designed to test abstract problem-solving, more than double its predecessor's score three months earlier, while holding API prices flat. ([Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/)) Meanwhile, composite evaluations show OpenAI, Anthropic and Google clustered tightly at the top. ([Artificial Analysis](https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index)) Six months ago, picking the right AI model felt very important. Model choice now matters less. Models are converging so fast that any advantage evaporates before you've finished onboarding. What hasn't converged is the application layer: ChatGPT, Claude, Claude Code each wrap similar intelligence in very different interfaces and workflows. Pick based on the problem and the workflow. The model underneath will be fine.

## Try this

**Run your day through AI at six in the evening**

At the end of each day, take whatever you captured (meeting notes, voice recordings, emails) and ask Claude for a five-paragraph reflection. Not a summary. A reflection. I've been doing this daily for months. The output compounds. You start noticing patterns you'd never have seen without the habit. This is the senryu lesson in reverse: the raw material is irreducibly yours. AI helps you see the shape of it. That's human steering at its most practical.

**Before you build training, check the settings menu**

Most people never change the default model on AI platforms. Before investing in an elaborate training programme, walk up to five people in your organisation and ask two questions: which AI model are you using, and have you changed anything in the settings? The answers will tell you more about your organisation's AI maturity than any survey. If your people haven't opened the settings menu, the barrier isn't skill. It's something simpler: nobody showed them it was there.

**Push record. Think aloud. Send it to AI.**

When someone has stories they want to capture but can't easily write, skip the blank page. Push record while doing actual work. Don't write, don't perform, just think aloud. Send the raw recording to a transcription tool, then run the transcript through Claude. The stories that would otherwise stay untold (because the effort of writing them was too high) get captured without friction. I've seen this work for client case studies, internal knowledge sharing, and personal reflection. The raw material is human. The refinement is machine. That's the order that works.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23email-2026-02-28) or ask someone on the list to add you.

[See the extras for this week ↓](#extras-2026-02-28)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-02-28) or ask someone on the list to add you.

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23letters-2026-02-28) or ask someone on the list to add you.

## The wonder and the weight
22nd February 2026

Copy this week for AI

[Image: LinkedIn carousel: The future is here. Your organisation isn't.]

See all 8 slides

[Image: Cover: The future is here. Your organisation isn't.]
[Image: Senior leaders opening laptops on Friday evenings.]
[Image: 84% of the world has never used AI.]
[Image: The wonder is real. But the future is very hard to spread around.]
[Image: Two strategies: Build or Partner.]
[Image: Where to start: three action items.]
[Image: The question is how we get it to work at scale.]
[Image: Save this. Share it with your leadership team.]

Experimental audio edition read by AI - Let me know what you think!

0:00--:--

1x

The wonder and the weight . 6:43
[Download or subscribe ->](/newsletters/david/podcast.html)

David's email this week

## What's on my mind

### The wonder and the weight

When senior leaders at consulting firms, broadcasters and banks find themselves looking forward to opening their laptops on Friday evenings and Sunday mornings, something has changed. Their own time, away from calls and meetings, voluntarily given over to playing with AI. I keep seeing the same moment. Someone who runs a division or manages hundreds of people realises they've just done, alone in ten minutes, work that used to occupy an entire team for a week. Their eyes light up. Childlike wonder on a fifty-year-old face. The future is very much here.

It's unbelievable what's possible these days and it's changing very, very, very, very rapidly.

[Image: METR Time Horizon: the length of coding tasks AI can handle autonomously has grown from seconds three years ago to longer than a working day]

This chart tracks how long AI can work on a coding task before getting stuck. Three years ago the answer was seconds. Today it's longer than a working day. The curve is steepening.

But there is a dark side.

Monday comes. The same person walks into the same office. Most people haven't made the time to properly work AI out yet, despite training and newsletters from leadership.

[Image: Each dot represents 3.2 million people, coloured by their most advanced AI interaction]

Leaders who've seen the future have a lot of work to do to bring everyone else with them. Eighty-four percent of the world's population has never used AI. Even among those who have, fewer than one in fifty pays for it.

Also: Meeting cadences haven't changed. Team structures and roles are the same. The time savings from those who've worked AI out just get absorbed into existing rhythms.

We all have to reconcile these two things. The future is here. The wonder is real. What's possible has changed so much in the last few weeks. But the future is very, very, very hard to spread around. Structural inertia is real. The future and the inertia coexist in the same organisations, sometimes in the same person on the same day! That tension, between what individuals can now do and what organisations will allow, is what I keep coming back to. Not "does it work?" It does. But how do we get it to work at an organisational level?

Several times a day I flip between childlike wonder and a deep fear for people and teams and orgs who've not leant into this yet. I feel both. The wonder and the weight. That's what this email is about.

## Three things worth knowing

**1. The technical barrier is gone. Domain expertise is what matters now.**

Ethan Mollick, a Wharton professor, gave executive MBA students four days, three AI tools and a brief: build a company from scratch. They did. Working prototypes, financial models, market research, competitive positioning. Most had never written a line of code. The students who got furthest weren't the most technical. They were the ones who understood their industry. If you've spent twenty years in your field and haven't tried building something with these tools, you're sitting on your biggest advantage. Open Claude Code and have a play. Let me know what you build! ([Mollick's full account](https://www.oneusefulthing.org/p/management-as-ai-superpower))

**2. Most firms use AI. Almost none can measure the difference.**

An [NBER study](https://www.theregister.com/2026/02/18/ai_productivity_survey/) surveyed nearly 6,000 executives across four countries. Sixty-nine percent of firms now use some form of AI. The average productivity gain over three years? 0.29%. The economists draw an explicit parallel to Robert Solow's 1987 observation that computers were everywhere except in the productivity statistics. The explanation, proved right over the following decade: firms had to fundamentally reorganise before the technology translated into measurable gains. That lag wasn't months. It was years. The same pattern is playing out with AI right now. But it just has to be much quicker, right? Read on ...

**3. Freelancers are disappearing. The data is in.**

Ramp analysed real corporate spending data. The share of business spend going to freelance marketplaces like Upwork and Fiverr fell from 0.66% to 0.14% in three years. Firms most exposed to AI substituted at roughly $1 in reduced freelance spend for $0.03 in AI spend. A 97% cost reduction. The gig economy may be automation's first casualty. ([Ramp Economics Lab](https://ramp.com/velocity/ai-labor-market-impact-freelancers))

## Try this

**Find your misfits (and back them)**

Identify the one-in-ten people in your organisation who are already at the forefront of using AI. The ones who experiment on their own time, who volunteer for pilots, who can't stop showing colleagues what they've built. Free them from at least some of their existing reporting lines. Give them the best tools. Bring them together, give them a name, give them a mandate to have an impact beyond their old job description. Most training programmes spread investment evenly. I don't think that's right. Your top ten percent will generate eighty percent of the value. A power user with the right support can transform a team. A reluctant user with mandatory training may forget everything by Thursday.

**Open Claude Code and build something**

Just build something. Pick a website you wish existed, or some data you wish was gathered across the web. Or an email scanning automation that summarises stuff for you on a schedule. Or an app that lets you do something you've been meaning to do. The only blocker is your imagination (plus a subscription to Claude, ten minutes of setup and 30 minutes of playing!) Claude will hold your hand through every single step. Just ask if you don't understand. Within 30 minutes you'll have something working and your perspective will have changed. I was reminded several times this week: there's a difference between knowing what's possible and doing it yourself. Go do it.

**Measure outcomes, not logins**

Accenture now [tracks weekly AI tool logins](https://www.cnbc.com/2026/02/19/accenture-ai-orders-senior-staff-lose-out-promotions.html) for senior staff and links them to promotion decisions. The internal reaction? Staff call the tools "broken slop generators." Measuring whether people opened the app tells you nothing about whether the work got better. If you're going to track AI adoption, track the outcomes: did quality improve? Did turnaround times fall? Did clients notice? Log-in counts measure compliance. It's easy but lazy. Measure capability, instead.

[See the extras for this week ↓](#extras-2026-02-22)

This part is for people on David's Saturday list. [Sign in with your email](/newsletters/david/sign-in.html?next=%2Fnewsletters%2Fdavid%2Fmembers%2Farchive.html%23extras-2026-02-22) or ask someone on the list to add you.