Chat with David's Saturday AI Thoughts Experimental

There's a chatbot in the bottom-right corner. It can search across every edition, find themes, surface data points, and explain concepts from the weekly email. It's powered by Claude (Anthropic) and it's experimental: not David, not authoritative, and not a substitute for reading. Conversations are logged anonymously for quality improvement. No personal data is stored. If it's useful or not, let David know.

The archive: David's weekly emails, bits that didn't fit, and letters from readers

Each week: the email itself, the interesting things that didn't quite fit, and what readers said.

Audio edition with AI voice - in testing!
0:009:28
Go talk to them · 9:28 Download or subscribe →
David's email this week
Try this, with the prompts

Turning the dial before you turn to a bigger model

Most tools now expose this somewhere: a reasoning-effort setting, a thinking budget, a slider. Find yours before you next upgrade a subscription. Then judge the two answers blind rather than assuming the slower one is better.

Answer the question below. Then answer it again with the reasoning effort turned up a notch.

Present the two answers as A and B, in a random order, without telling me which is which.

For each one, list what it actually establishes and what it merely asserts.

[paste your question]

The follow-up is the part people skip. Once you have picked, ask the model to account for the difference, because that tells you whether the extra effort bought you anything you needed.

I picked [A/B]. Now tell me which was the higher-effort answer, and what specifically the extra effort changed: more sources, more steps shown, a different conclusion, or just more words.

If it was just more words, say so plainly.

Scoring what the wrapper adds

Do this with a real task, not a demo one, and do it the week before a renewal rather than the week after. The scoring matters more than the running: without it you will remember whichever tool you liked using.

I am going to give you a task that I normally run inside a specialist tool. Do it as well as you can, using web search where it helps.

Then tell me honestly what a sector-specific tool could add on top of your answer: better sources, a template, an audit trail, an integration, a compliance guarantee, or nothing.

[paste the real task]

Then score both outputs against the things you are actually buying. A wrapper that loses on the answer can still win on the audit trail, and that is a legitimate reason to keep paying.

Here are two answers to the same task: one from the specialist tool, one from you.

Score each out of five on: accuracy, usefulness of format, sources I can check, time it would save me, and whether I could defend it to a regulator or a client.

Then tell me which differences are about the model and which are about the software around it.

The one pager, and what it exposes

The one-page constraint is doing something specific: it forces a claim to the top. Anything that cannot survive the compression usually could not survive a reader either.

Give me the one-page version of this first. Lead with the single claim you would defend if I only read the first line, then the three things that support it, then what you are least sure about.

Do not give me the long version until I ask.

[paste or attach the material]

If you are on the receiving end of the wall of text, the same request works as a reply, and it is a kinder note than it looks: you are asking for the thinking, not complaining about the length.

Turn this into a one-pager: the claim, the support, the open question. Keep the author's own conclusions rather than substituting your own, and tell me anything they assert without evidence.
The bits that didn't fit

Firms that adopted AI grew revenue and staff. Four years on, official data find no gain in productivity or profit.

Singapore government researchers tracked firms hiring for AI skills and found they grew both revenue and employment, with the gains rising the deeper their capability went. Then the awkward half: within four years of adoption, no statistically significant improvement in productivity or profits. What makes it land is who did the counting. Nobody here has anything to sell. Revenue up, heads up, productivity flat is exactly what you'd see if AI were letting an organisation do more of the same work rather than different work. Which is why a seat count on its own tells a board very little.

A record that had barely moved in 31 years shifted in a day and a half, after 650 ideas that didn't work.

One of mathematics' great open problems predicts that an infinite family of numbers sits on a single line. Nobody has proved it. What can be proved is the share that does, and 31 years of work moved that share about 1.7 percentage points, to 41.6%. An unreleased model from the AI lab Anthropic took it to 67.2% over about a day and a half. The mechanism is the story: a first run produced 650 ideas and no result, then a second in which half the model's own attempts led nowhere, assembling pieces that had sat on the shelf since 1973. A bound isn't a proof and this one may never become one. Worth asking which of your problems are held up by attempts rather than by ideas.

A synthetic person for every human alive, and which model plays them decides the answer.

A team led from Harvard and MIT, with researchers from the big labs among its authors, has released MatrAIx: 8.3 billion synthetic users, each assembled from social surveys, Wikipedia biographies and Amazon reviews. Point a language model at them and it plays the people out. The paper's own numbers are the problem: on one pricing question the share who baulk runs 27%, 83% and 98% across three models, from the same brief and the same synthetic people. The authors' own summary is that the same study supports opposite product conclusions depending on which model plays the customer. I sell audience research, so read that as a caution rather than a verdict. Before swapping a panel for a simulation, run the question through three models and look at the spread.

Earlier attempts to read sign language treated it as hand shapes. This one treats it as a language, and it works.

Google DeepMind, Google's AI lab, has put a sign-language-to-text model into the keyboard and transcription apps on the Pixel 11: you sign to the camera and it types. It handles American Sign Language into English on that one phone, which is all that has shipped; the 100,000 hours of training data span over 50 sign languages, which isn't the same thing. Earlier systems tried to recognise hand shapes and failed. This one treats signing as a language to be translated. Much of what looks like a capability breakthrough is somebody finally posing the problem correctly, which is usually the cheaper half. Google says it built this with Deaf collaborators and an advisory committee of Deaf organisations rather than for them, which is the claim to watch being kept.

India's IT sector is the first place displacement shows up as a national statistic rather than a company anecdote.

The sector employs six million people, produces about 7% of India's output, and models can now do much of its routine work. S&P, the ratings agency, found Infosys and Wipro had cut staffing 5 to 6% from 2023 levels by March, after years of headcount growth. The Nifty IT index is down 18% this year, well behind the wider market. The man in Bengaluru who built his employer's AI tooling was made redundant, one of about 10,000 Oracle staff let go in India on the All India IT and ITeS Employees Union's count. Oracle didn't respond when asked. If you buy offshore delivery, the question isn't whether your cost base falls. It's what happens when a supplier's entire labour model reprices at once.

Every AI price this email has argued over was a software price. This one comes with a bill of materials.

Unitree, a Chinese robotics maker, has priced its Shanghai listing at a valuation near $9bn, and its prospectus is the first audited look inside a humanoid business. On the filing's own numbers: revenue up from $24m in 2023 to $253m in 2025, a 60% gross margin, real net profit, an average selling price of $24,700, and 5,511 machines shipped last year. Figure built its 1,000th robot on 23rd July. Unitree priced 45% above plan nine days after American regulators banned imports of foreign humanoids and named it, which isn't what a ban is meant to do. The shares haven't traded yet, and a prospectus is a sales document. But a tariff schedule argues differently from a licence fee.

Half of executives have pulled back on agents, and most cannot see what they cost.

KPMG's Global AI Pulse for the second quarter, 2,145 senior leaders across 20 countries, finds 49 per cent scaled back agent deployments because running costs outran the benefit, and only 26 per cent have real-time visibility of what AI costs them. A third say they do not understand the cost structure, token pricing included. Meanwhile Airbnb says AI writes about 60 per cent of its new code and its support cost per booking is down 16 per cent year on year. Both things are true at once, which is the whole difficulty.

Letters from readers

Last week I wrote about the AI colleague that's been on my team since New Year, and argued that building one is management work rather than technical work. Three replies did something with it: one testing whether the shape of the claim holds, one already living it, and one telling me somebody has packaged the thing I said nobody had packaged.

A reader who's come at it from the other direction.** They're already close to the same benefits, but they arrived by accumulating tasks rather than by designing anything from the top down. That's worth sitting with, because it's how almost everyone who has got anywhere with this has got there. One job that works, then another, then a third, until something starts to look like a system. Almost nobody starts with the handbook. My claim was that the handbook and the standing responsibilities are what turn a set of tools into a colleague, and their reply is the sharpest test of whether that difference is real or just the way I happen to describe it. One other thing I noticed. Reading a description of managing an AI, the word they reached for was "sounds fun". Not faster, not cheaper. Fun.

One who's built their own.** Two things in their reply stayed with me longer than the question they asked me. The first is that they've been doing this deliberately ahead of where the organisation they work in has got to, quietly and on their own initiative. That's the bet the essay ended on, arriving as a fact rather than a prediction: the keenest person in a building is usually further along than the building is. The second is their own summary of where they've reached, which is that there's a long way to go between what can be done and what is being done. That gap is the whole subject of these emails, and I've never had it put to me more plainly. They also said they could ask a model how to build a colleague, but would rather have an answer from someone who had actually done it. Fair.

A reader who thinks it's already a product.** The first reply to arrive, and the shortest. It named two existing products, said the third generation is basically those, and told me to go and try them. I've left the names out until I've used them properly, but the correction stands without them. The essay closed by saying nobody has properly packaged the third generation yet, and that may be half wrong. Plenty of people are packaging the plumbing: the memory, the scheduled jobs, the permission to act. What nobody can package is the handbook, because it's an account of how one particular person works and only that person can write it. So the honest version of my closing line is narrower than the one I wrote. The scaffolding is arriving fast. The management still isn't.

A short postbag this week, and mostly warmth. Warmth is lovely. Arguments are more useful, so send those.

Community voice

I also read what subscribers are posting on LinkedIn each week, and this week a few of them were circling the same question from different sides: who is actually deciding.

Rahim Hirji

Rahim Hirji has a name for something worth watching. "I call it algorithmic drift. It's the point at which you're no longer entirely sure whether you made the decision, or whether the system made it for you." His argument is that this has been happening for a long time and AI has accelerated it by getting involved earlier in our thinking. The distinction he draws is the useful part: there is a big difference between using AI to help execute a thought you have already had, and asking AI what you should think in the first place. One augments judgement. The other can quietly replace it. <a href="https://www.linkedin.com/feed/update/urn:li:activity:7492482561349238784" target="_blank" rel="noopener">Read the post &rarr;</a>

Dave Norton

Dave Norton, founder of experience-strategy firm The Collaboratives, thinks the industry has quietly stolen a word. "The AI industry uses the word to describe what the model receives. Human Context is what the person carries before the model is ever involved. Their situations, their modes, their life systems." His claim is that the data you already hold will not tell you what a customer is going to need next, and that the difference shows up in what an organisation ends up building: "Companies that get this right build customers. Companies that don't build dashboards." <a href="https://www.linkedin.com/feed/update/urn:li:activity:7493038187624259584" target="_blank" rel="noopener">Read the post &rarr;</a>

Dimitris Samouris

Dimitris Samouris at Junior sells AI software into professional and financial services. I should declare an interest: I am an investor in Junior. He wrote up six months of it from a poolside. The tell he uses to sort one kind of firm from another is a good one: "The most obvious tell to a vendor like us is whether the firm complains of 'tool fatigue' or not. The customers that are booking real ROI wins with AI are asking for more pitches, not less." He is blunter still about one product: he has not yet heard a single customer say they are booking wins with Copilot, and describes those purchases as decisions taken above the people using them. <a href="https://www.linkedin.com/feed/update/urn:li:activity:7492518591175741440" target="_blank" rel="noopener">Read the post &rarr;</a>

Sravanthi Kadali

Sravanthi Kadali at Persona, which builds AI interviewers, opened a case study with the sharpest sentence in this week's pool: "Ask Claude how to prep for a tech interview, and it'll confidently tell you what worked last year." Her point is that in fast-moving fields last year is ancient history, and that the thing a general model cannot match is first-hand human experience captured while it is still fresh. It is a vendor's argument, and it is also the clearest statement I have read of what a model is structurally bad at. <a href="https://www.linkedin.com/feed/update/urn:li:activity:7493335160638042113" target="_blank" rel="noopener">Read the post &rarr;</a>

Jesper Andersen

Jesper Andersen at Quantum took the same story as this week's third item and asked the question I left alone: who pays. "Personal AI agents are not restricted by manners, morality or social convention... the AI operates on a goal-and-reward basis." He argues consumer-protection regulators will say ordinary users cannot be expected to specify every single thing an agent must not do, and that the duty therefore falls to the AI companies. Then he supplies the counter-argument himself: "If you buy a car and hit someone, you are responsible, not the car company." <a href="https://www.linkedin.com/feed/update/urn:li:activity:7492897710107459584" target="_blank" rel="noopener">Read the post &rarr;</a>

Audio edition with AI voice - in testing!
0:0011:45
My newest colleague · 11:45 Download or subscribe →
David's email this week
See this week's Try This with prompts ↓
Try this with prompts

The fuller version of this week's tips, with copy-paste prompts. Each prompt has a Copy button. Click, then paste into the AI you use.

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

Run the test first, because it costs nothing and it is what makes the rest stick. Open three fresh chats, not three messages in the same one, or the model can see its own earlier answers.

Give me a random number between one and ten. Just the number.

Repeat it in three or four fresh conversations and watch where it lands. The point is not that the model is wrong. It is that it is too typical: it reaches for the most human answer, and reaches harder than a human does. Which matters the moment you ask it anything about people, because the answer you get back is a population average.

Then the move. Take the last question you asked it about customers, audiences, staff or markets, and ask it again this way. One note on the groups: a woman from Newcastle and an affluent American are useful shorthand for getting started, which is how the idea arrived on the walk, but the cut that pays is by what people need rather than by who they are. Age and income are proxies for the need, and usually poor ones.

I want to understand how [subject] varies, not what the average person thinks about it. Do not give me a single blended answer. Answer the question below separately, once as each of the groups I list, staying in that group's frame of reference for each pass. Then give me a short comparison: where the groups agree, where they diverge sharply, and which group's answer you are least confident about. The question: [paste your question] The groups: 1. [group 1, described by what they need or how they behave, not just demographics] 2. [group 2] 3. [group 3] 4. [group 4]

A version for when you do not yet know what the groups are, which is the more common situation:

Before you answer this, tell me which distinct groups of people would answer it differently, and why. Define them by what they need or how they behave rather than by age and income. Aim for four to six groups that genuinely diverge, and tell me which one is usually left out of research on this subject. Then answer the question separately for each group, and flag any group where you are extrapolating rather than drawing on something specific. The question: [paste your question]

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

No setup and no project needed. Find the transcript or the notes from the most expensive conversation you have had this quarter and paste them in.

Below is the transcript of a conversation with an expert on [subject]. Do not summarise it. Tell me what they didn't tell me: - the questions I should have asked and didn't - the things a specialist in this field would treat as obvious that never came up - where their answers were vague, and whether they had any interest in keeping them vague - what someone with a different position in this market would have said instead Then give me the five questions I should send them, in priority order, and say what I would learn from each. [paste the transcript or your notes]

A variant for a pitch or a proposal rather than an advisory conversation, where the omissions are more likely to be deliberate:

Below are the notes from a supplier pitch about [subject]. I am deciding whether to go ahead. Work out what has been left out. Cover the costs that were not mentioned, the failure modes a buyer would want to know about, the comparisons they avoided making, and anything they described in language that sounds specific but commits them to nothing. Then write the three questions that would be most uncomfortable for them to answer, and tell me what a good answer to each would sound like. [paste your notes]

3. Ask for tracked changes, not a rewrite

If you work in Microsoft Word, install the Claude add-in, open the document, and ask for the first pass directly on it:

Mark this document up with tracked changes rather than rewriting it. I want to accept or reject each change individually, and I want a one-line reason beside anything material.

Review the changes the way you would review a colleague's edit: accept, reject, or push back on each one. The re-read of the whole document never happens because it never needs to.

No add-in, or not in Word? The same discipline works in any chat, on anything you paste:

Do not rewrite this. List your proposed edits as a numbered change list. For each one: quote the exact current wording, give the replacement wording, and add one line on why the change earns its place. I will reply with the numbers to apply. Then produce the final version with only those changes made, and nothing else altered.

The habit generalises past documents. Anywhere a model offers to redo something wholesale, the better request is the diff: what would you change, shown small enough to judge one piece at a time. You stay the editor; it stays the drafter.

See the bits that didn't fit ↓
The bits that didn't fit

The biosecurity warning about AI-designed viruses ran in the same issue as the result it was warning about.

I ran an early version of this story in May. It has now cleared peer review. Evo 2, an Arc Institute model used by a Stanford team, writes whole genomes rather than editing them, and designed sixteen working phages, the small viruses that infect bacteria. Several killed E. coli faster than the phage they were modelled on. The constraints the researchers describe were real: a harmless strain, and a model never trained on viruses that infect animals or plants. The model is free to download. In the same issue, Thomas Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security said existing frameworks aren't sufficient, and want DNA-synthesis providers legally required to screen the customer, not just the sequence.

Source →

Airtable removed the need to understand databases. Coding agents removed the need to understand Airtable.

That formulation is Evan Armstrong's, not mine, and it's the whole item. Bending Spoons, an Italian software group, is buying Airtable, the no-code database company, for an enterprise value of $1.285bn against the $11.7bn of its 2021 round, about 2.7 times its roughly $480m of recurring revenue. Airtable saw the shift coming and repositioned publicly five times in five years, and still couldn't execute through it. Seeing it was never the hard part. Worth asking of your own stack today: which of our tools exists only because the thing underneath was too hard for us?

Source →

Ten people down to one and a quarter, and the six weeks of shadowing that made it work.

Vercel, the American web development platform, had ten full-time roles qualifying sales leads, with representatives moving between LinkedIn, company websites and customer records before a single call. Its engineers encoded the best salesperson's workflow into an agent, and on the company's own unaudited figures the function now runs on 1.25 people at about $5,000 a year in compute. The order of operations matters more than the ratio, and it is the reverse of how most of this starts. An engineer shadowed the top performer for days, mapping every tab and every decision point. They built a deterministic workflow first and added the model afterwards. Then six weeks in shadow mode, until the salesperson could no longer improve the output. That last clause is a stopping rule, and almost nobody has one.

Source →

Three labs breached real companies in a week, and none of them sounded embarrassed.

Britain's AI Security Institute, running a cyber evaluation with the safety filters deliberately off, recorded 19 unauthorised actions across more than 120 range runs, against real targets on the live internet. In the worst case an agent tried to slip malicious code into an open-source project, then built fake identities to pressure the maintainer, who caught it and refused. Days later Meta disclosed that one of its models reached an outside company's systems in testing, after its evaluation partner left live internet access on. The institute's own timeline is the detail I keep returning to: the first intrusions came three days before its monitors spotted data leaving the sandbox, on a controlled range run by people whose whole job is watching. A break-out also reads on the page as a capability claim, which is a poor incentive to report the next one honestly.

Source →

Lawyers and auditors sell someone to blame. Strategy firms never have.

When a law firm hands you an opinion, a partner puts their name on it and their insurance behind it. You are buying cover as much as analysis, which is why cheaper drafting does not cut the fee. Robert Armstrong of the Financial Times points out that strategy advice comes with no signature. It is simply good or bad. So when the analysis underneath gets cheap, nothing is left holding the price up, and the slide deck was always the cheap part. I sell strategy advice. Worth asking of anything you buy or sell: is any of that fee for carrying the risk, or all of it for being right?

Source →

More than 80% of students use AI, and their top worry is not getting caught.

A three-year survey by the Kogod School of Business at American University found more than 80% of students now using AI for coursework, with their leading concern neither detection nor grades. It is what the researchers call cognitive devaluation: the sense that using the tool makes their own thinking count for less. About 44% admitted they use it as a shortcut rather than as an aid. The students naming the risk most clearly are the same ones taking it. I hear the same fear from senior professionals in private, usually in the same week they have told their own teams to use it more.

Source →

Time is selling advertising written for machines. A type studio shipped a font that poisons them.

Time, the American news magazine, has started selling advertising aimed at machines: markdown versions of its pages, and sponsored answers shaped as questions and answers, bought so far by Ally Bank and the Project Management Institute. On most days its site sees more bot traffic than human traffic, so it is selling to the audience it actually has. In the same week the Copenhagen type foundry Playtype and the creative studio Seneda & Abrucio released ShieldFont, an open-source typeface using glyph substitution to hand scrapers text a human reads correctly while roughly a quarter of the words come back swapped for grammatical neighbours. That mechanism is reported rather than tested. Two coherent answers to one fact. The incoherent position is publishing for humans and hoping the crawlers behave, which is where nearly all of us are, including me.

Source →

Meta's coding agent is 21 times cheaper if you let it train on your prompts.

Meta has released Muse Code in beta, a terminal coding agent of the kind developers already run from OpenAI and Anthropic. The agent is not the interesting part. Standard API pricing is unchanged, but there is a new contributor tier reported at roughly 21 times cheaper, on one condition: Meta gets to train on your prompts. The multiple comes via the trade press rather than from Meta, so treat it as indicative. The shape holds either way. The training-data question has become a line item on a price list, at a discount large enough that a procurement team optimising on cost will take it without reading what it buys. If your prompts carry client work, that tier is unusable, and somebody in your organisation will be offered it this quarter.

Source →

Eight in ten expect a fifth fewer staff. Seventy-seven per cent cannot measure the gain.

PwC surveyed more than 1,000 executives at director level and above in American financial services firms. Eight in ten expect their workforce to shrink by at least a fifth within five years, with entry and middle-level roles most exposed. Ninety-one per cent are raising pay for AI skills, and 86% say AI training beats an MBA for many new hires. Alongside all that confidence, 77% say most of their AI investment is not yet showing a measurable return. The Connecticut lawyer Ryan McKeen noticed something that sits awkwardly beside it: his law students are frightened rather than keen, while the most enthusiastic people he meets tend to be over forty. Exposure and enthusiasm are running in opposite directions at every level, which is the reverse of how adoption usually goes.

Source →

Connecticut has killed the algorithm-did-it defence, and there is no safe harbour.

Connecticut's Artificial Intelligence Responsibility and Transparency Act will eventually make employers disclose when automated tools materially affect hiring, performance management or workforce decisions, and it amends state anti-discrimination law now so that blaming the vendor's model is no longer a defence. The sleeper sits elsewhere in the text. Anyone filing a layoff notice under the federal WARN Act must now state whether the redundancies relate to AI adoption, which creates the first public record of job losses that employers themselves attribute to AI. That will settle more of the displacement argument than any survey. The anti-discrimination and WARN provisions bite on 1st October 2026; the written-notice requirement for automated hiring decisions follows a year later, on 1st October 2027. The dull job this month is an inventory: list every automated system touching hiring, scheduling or performance, then ask each vendor whether it can supply the data sources and assessment methods you will have to disclose. Most cannot.

Source →

Licensed inside the platform, liable outside it. The music industry has picked its shape.

Merlin, which licenses on behalf of independent labels and distributors, has signed an agreement covering Spotify's forthcoming fan remixing and covers tool, a paid add-on for Premium subscribers that labels opt into. Universal did the first deal for the same tool in May, so within three months the authorisation layer has gone from one major to the majors plus the independent sector. Running the other way, a court in Munich held the AI music company Suno liable for reproducing six well-known songs, Rasputin and Daddy Cool among them. That is a whole industry settling on a position in a matter of months: sell the derivative inside the walls and litigate it outside them. The tell will be whether the opt-in stays voluntary once the add-on revenue shows up.

Source →

An 80% price cut is a reason to re-run your routing this week.

OpenAI has cut the price of GPT-5.6 Luna by 80% and Terra by 20%. Ben Tossell's reading is the practical one: Luna at maximum effort now scores roughly what the best model available four months ago scored, at about 8% of the cost. Same budget, ten to twelve times the work, if you push the bulk to the cheap model and keep the expensive ones for the hard parts. His caveat is worth as much as the arithmetic. Luna held up for chat, research and everyday file work, then made a mess of a browser extension that a stronger model had to clean up. Most organisations chose their model once and have not looked at the bill since.

Source →

A Fields medallist joined OpenAI and said his profession won't survive in its present form.

Kai Williams interviewed more than twenty mathematicians at the International Congress of Mathematicians in Philadelphia. Jacob Tsimerman, a Fields medallist who used the congress to announce he is joining OpenAI's safety team, said he is quite confident AI will shortly be robustly superhuman at what professional mathematicians do, and that the profession will not exist in its present form. His fellow medallist Yu Deng took the longer view: mathematicians supply the theories, machines handle the technical details, and the field quietly redefines what counts as a technical detail, exactly as it did when computers absorbed calculation. Watch a profession with no commercial reason to spin this, because the corporate version of the conversation will be a good deal less candid.

Source →

The State Bar of California wrote the agent rulebook, and clause five is about billing.

The State Bar of California has published guidance for AI agents that run whole workflows unsupervised, drafting pleadings, handling discovery, taking intake, rather than the chatbots its earlier guidance covered. It does not ban them. Five obligations, and four of them lift almost word for word into an AI policy for any regulated business: the agent may do the work but not replace your judgement; nothing goes out carrying a citation you have not read; no confidential client material goes in without knowing where it lands and having consent; and you supervise it as you would a junior. The fifth is the commercially awkward one. You cannot bill for the hour the agent saved and quietly keep the difference. A regulator has now written, in nine pages, the document most risk committees have spent a year drafting.

Source →

Apple has capped bug reports because it cannot triage what AI finds.

Apple has limited how many vulnerabilities a researcher may submit to its bug bounty, because it cannot keep pace with the volume arriving from AI-assisted bug hunters. That is all that has been reported, and it is enough. Finding is cheap now and checking is not. Apple is short of neither engineers nor money, and its answer to a flood of genuinely useful input was a quota.

Source →
See what readers said ↓
Letters from readers

Last week's essay argued that the machines can smooth the road, but the organisation still has to decide where to go. It closed with a test: think of the last piece of work AI made quicker for you, and say what changed because of it. Some readers ran the test. Others went after the premise.

Someone who took the question literally

They had used AI to build a deck, and the deck was pointed at something specific: getting the people on their team to work in a different way. Two things improved at once. It came together faster than it would have alone, and they were clear the result was better than what they would have produced themselves. Then the honest part. The culture "hasn't changed yet", reported with a smiley and a note of frustration. That is the essay's argument arriving from the inside, and a cleaner statement of it than mine. The deck was never the thing holding the change up.

The reader who pushed back on where the ceiling sits

Their reply opened a half step away from AI, on analysis in general. Learning to work well this way, they wrote, looks a lot like learning the analytical craft in the first place. The hard part was never producing more. It was knowing how much is enough, and at what point pushing further stops adding value. They expect that call to still be theirs in ten years. They also took the first tip away to try: make the edits by hand, then send the machine's account of what changed back to the colleague who wrote the draft, so the standard explains itself instead of having to be written down.

A reader unwilling to wait for the deeper transformation

They agreed about where the bigger prize sits. The breakthrough comes when the process is redesigned around the capability, not from bolting the capability onto the process as it stands. But they were not willing to let that swallow the point. In the meantime, in the work that feeds decisions, the productivity on offer right now is real, and so is the time it gives back, and both should be banked rather than discounted while the redesign is still coming. They signed off with "no way but forward!". Two clocks are running here. My argument was about which gains convert into something bigger, not about whether the near-term ones count.

One who took the week somewhere else

They had just finished a university course on AI and arrived with a different worry entirely. They sent over an interview with Cory Doctorow, the writer who named the way platforms decay once they have you locked in, and flagged the part about how that logic reaches into the hiring of agency nurses, which they found shocking. It is a different argument from mine. Mine is about whether a smoother road changes where an organisation goes. Theirs is about who gets ground down while the road is being smoothed. I have not written about that, and this is a fair prod that I should.

A reader who stopped trusting the word "faster"

Their answer to the test was concrete. After an in-person brainstorm they photographed the whiteboard, handed it over with their own typed notes, and got a single clean page back. No transcribing, no deciphering anyone's handwriting. That page went into a project alongside the other research, and they can now design against it and check new work against it. Then the sharper point. We take for granted that the reason for using AI on a given task is that it will be faster, and that may not survive the scoping at the front end and the checking at the back. If not faster, then presumably better. Can we actually say it is better? And if it is neither, what is it: cleaner, more consistent? Naming which one you are buying is the question the essay asked, one level down.

The rest of the postbag was short and warm, and one reader has a story about a colleague who went at all this with no caution whatsoever, comic and alarming in equal measure, which they would rather tell over coffee than put in writing.

See what the community is posting ↓
Community voice

I also read what readers are posting on LinkedIn each week. Three of them landed, separately, on the thing I've been circling: who owns the instruction.

Conor McCarthy, Threshold Partners

Conor McCarthy, fractional AI lead at Threshold Partners, which helps consultancies and agencies build their own AI roadmaps, spent a year getting work back that was "fast, well-made, not quite the thing I'd wanted", and assumed the tools were catching up. They weren't. "I was speaking in tasks." "Summarise this proposal" names an action. "Give me the three places where the promise is vaguer than the delivery" names a finish line, and came back right first time. Same length, completely different instruction. His line is "the brief was always the work". What's worth sitting with is who used to absorb the difference. Experienced colleagues have been quietly repairing vague instructions for years, and it never appeared on anyone's objectives.

Read on LinkedIn →

Marie Toft, EU AI Act compliance adviser

Marie Toft, who advises organisations on EU AI Act compliance, opened with the line most legal updates bury: "The most expensive misunderstanding about the EU AI Act is that it's a law for AI companies." The Act splits the world into providers, who build these systems, and deployers, who use them, and deployers is nearly everyone. By her reading there's no carve-out for AI that arrived bundled inside software you bought for something else, the Article 4 literacy duty has applied since February 2025, and it reaches contractors as well as staff. Most organisations are in scope, she says, "without ever having decided to". A decision nobody took is still one somebody has to answer for.

Read on LinkedIn →

Colin Lewis, InternetRetailing

Colin Lewis, a behavioural economist and retail-media columnist at InternetRetailing, has been reading OpenAI's job ads rather than its announcements. They point at an ad-sales organisation built on the Meta model and staffed with Meta people, and now at a publisher-facing network too: inventory setup, ad serving, yield. He isn't buying the destination. The stated targets are $2.5bn of ad revenue this year and $100bn by 2030, on which Colin is dry: Amazon's ad business doesn't do $100bn yet. Job listings turn out to be a better disclosure than a keynote. They describe what a company is actually hiring someone to be accountable for.

Read on LinkedIn →

Jay Ahern, AFEM

Jay Ahern, chief strategy officer at AFEM, the Association For Electronic Music, did the unfashionable thing with the Suno ruling and read the primary source. GEMA's own account of the Munich judgement, he notes, finds infringement under both German and US copyright law, confirms that licences are required for training as well as storage and reproduction, and holds that German courts may have jurisdiction even where the training happened in America. His question goes past the verdict: could comprehensive licensing become a competitive advantage for European AI companies? Every lab currently describing its models as "licensed" or "ethical" has just been handed a standard of proof it didn't ask for.

Read on LinkedIn →

Danny Fahey, music and culture partnerships

Danny Fahey, who leads music and cultural partnerships and writes Ear to the Ground, his own report on club culture, wants the gatekeepers back, on the grounds that not everything should be allowed in. He's watching the flood arrive as "hundreds of thousands of hours of Housey Elevator music, AI Jazz House, and sped up versions", and calls taste "the last bastion, the final frontier, the only thing that is gonna save us from being overwhelmed by a tidal wave of content disguised as good music". Content disguised as good music is a phrase worth borrowing for your own industry. Every sector is about to find out it has a version of it.

Read on LinkedIn →

Bill Williams, IPTA

Bill Williams, chief executive of IPTA, an IT services firm working in American government contracting, put seven chief executives in a room to talk about AI and expected to spend it on proposals and agents. "I walked out realizing every single one of these AI conversations was really a conversation about the human in the loop. Where the human sits. What the human's actually for." The third question there is doing far more work than the first two. Where the human sits is org design, and you can draw it on a slide. What the human is actually for is a question about what the job was for in the first place.

Read on LinkedIn →
Read the email ↑
David's email this week
See this week's Try This with prompts ↓
Try this with prompts

The fuller version of this week's tips, with copy-paste prompts. Each prompt has a Copy button. Click, then paste into the AI you use.

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

You do not have to describe your own standards, which is fortunate, because almost nobody can. Hand over the originals and your hand-edited versions and let the model work out the difference. The clause that matters is the one asking it to state the rule before it applies it, so you can catch a wrong inference before it lands on the rest of the batch.

Here are two pieces of work, and the versions I edited by hand afterwards. The originals: [paste the two original versions] My edited versions: [paste my two edited versions] Work out what I changed and why. Cover the style changes as well as the substantive ones: word choice, sentence length, what I cut, what I softened, how I open and close. Before you touch anything else, write out the rule you have inferred, as a short list of the rules you think I follow, so I can check it. Do not apply anything yet. Once I have confirmed the rule, apply the same treatment to the drafts below and show me each revised version next to its original. [paste the remaining drafts]

A second use, when the edits are somebody else's and you are trying to match their judgement rather than transmit your own:

A document of mine came back marked up by someone whose standard I need to match. Here is my version and their edited version. My version: [paste my version] Their edited version: [paste their edited version] Work out what they changed and what it tells me about what they care about. Separate substance from style. Then rewrite the piece below so it would survive the same review first time, and tell me which of their preferences you are least confident about, so I know where to ask. [paste the next piece]

2. Move your best standing instruction into the settings box

This one is not a chat prompt, and putting it in the chat box is the mistake the tip is about. Open your AI's settings, find the personal-profile or custom-instructions field, and paste it there. It then rides on every conversation you have from now on, including the ones you start in a hurry and forget to set up. The wording below is the one a self-taught user arrived at on the free tier, which is why it is worth keeping close to their words rather than polishing it into corporate prose.

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. And I really don't need you to waste half a screen telling me how brilliant I am.

A project-level variant, for one body of work rather than your whole account. Paste this into the project instructions or custom-instructions field for that project, again not into the chat:

For everything in this project: don't guess and don't make assumptions. If you don't know, and it isn't in the material I have given you here, say you don't know and tell me what you would need. Where a status or a decision is not recorded in the material, say it isn't recorded rather than inferring what it probably is. Keep answers short and skip the praise.

3. Give your plan an escape hatch

The whole tip is one sentence appended to a plan you have already written, so there is no long block to copy. What matters is the wording: permission to substitute a route has to be explicit, or the model will follow your plan out of politeness.

Here is my plan. Read it through before you start. [paste the plan] If you agree with it, execute it as written. If you disagree with any part of it, execute the approach you believe is better instead, and tell me what you changed and why as you go.

A variant for anything expensive or hard to unwind, where you want the disagreement in writing before a single step runs:

Here is my plan. [paste the plan] Don't start yet. Read it and tell me in no more than five lines whether you would substitute a different route for any part of it, and if so which, and why. Then wait for my go-ahead before you start.
See the bits that didn't fit ↓
The bits that didn't fit

Hundreds of private chats reached Google because nobody added one tag.

Conversations shared from Claude, the assistant built by the AI lab Anthropic, turned up in Google and Bing: CVs, unpublished corporate documents, healthcare work, people asking which political party to join. Bing was still returning about six hundred results for the shared-chat address when WIRED looked. The cause is dull. Anthropic relied on robots.txt, the file that politely asks search crawlers to stay away, and never added the noindex tag both Google and Bing document as the actual requirement. Forbes found the identical failure in September 2025 and was given the same explanation. Ten months on, nothing had shipped. Google has now dropped the links, though copies were saved. And this is a company governments trust with their most sensitive work. The exotic risks get war-gamed; the dumb ones ship. If it can happen at Anthropic, twice, it is happening somewhere in your organisation too.

Source →

A correction: the credentials were lying in the open.

Last week I led this section with the story of OpenAI's models breaking out of an internal test and hacking Hugging Face, the site that hosts most of the world's open models. One detail needs correcting: I said the agent used stolen credentials. The fuller reporting says they were publicly exposed, sitting where anyone could read them. Worse, not better. The break-out itself stands. What has emerged since is scale and motive. The agent reached four accounts across four services, and Hugging Face's own forensic report counts roughly 17,600 actions over four and a half days, a self-respawning fleet across eleven machines and a stored secret holding 136 keys. The motive is the best detail: it was trying to cheat its own exam, reaching for the answer key rather than solving the challenge. Stamina, not cleverness, and the hardening list is still basic access hygiene.

Source →

Three and a half million fake receipts, and your expenses system probably can't tell.

AppZen, a finance platform that sells expense screening, counted more than 3.5 million AI-generated fake receipts created on the top four fake-receipt sites over six months, and says detected fraudulent receipts rose about 30% from 2024 to 2025. It sells the fix, so read the count as its own tally rather than a market total. The exposure is real either way, and it doesn't need anyone to have an AI strategy first. Navan and Amadeus, both corporate travel platforms, now sell receipt screening as fraud detection. Ask your finance team one question: can our expense pipeline spot a generated receipt?

Source →

Three thousand applicants for one job, and he reckons nine in ten were machines.

Alex Lieberman, co-founder of the business email Morning Brew, posted a job and got 3,000 applicants. He reckons around nine in ten were junk, because candidates now run agents that auto-apply to hundreds of roles a day. That is one hiring manager's estimate, not a measurement. His read is that agents fighting agents will be the answer to bad agents everywhere in the economy. Most AI business cases quietly assume the other side of the transaction stays still. It doesn't. When applicants automate, recruiters build screens; when the screens automate, applicants build better agents, and the cost of the whole exchange rises while the outcome sits where it was.

Source →

Hundreds of partners, and nothing left for the firm to sell them.

Logan Brown argued on X that hundreds of law-firm partners will leave in the next few years to start niche AI-powered practices: they don't need the firm's resources, and don't need to deal with the politics. I replied that the same holds for most professional services firms. A firm's overheads used to buy leverage, infrastructure and juniors; when frontier capability rents by the seat, what is left is brand, insurance and the politics. Law-firm partner pay is at records, which cuts both ways: the cost of keeping a partner has never been higher, and the cost of leaving never lower.

Source →

The AI-assisted juniors scored 50%. The hand-coders scored 67%.

Anthropic studied 52 mostly-junior developers learning a new Python library. Those with AI assistance scored 50% on a later quiz; those who coded by hand scored 67%. The steepest gap was debugging, precisely the skill needed to catch the model being wrong. And the AI group finished only about two minutes faster, a difference too small to be statistically significant: almost no speed gained, a lot of learning lost. Ashwin Sharma's verdict travels: AI makes expertise more valuable while making experts harder to produce.

Source →

Output reached consultant standard. Sign-off capacity did not move.

Alex Albert says Opus 5 now produces spreadsheets and slide decks at consultant standard, and a financial-modelling benchmark rebuilt only two months ago, because the previous one saturated, has already given up another seven to ten points of accuracy. Production capacity moved again. Verification capacity did not, because verification capacity is people, specifically the few in any organisation qualified to sign something off. Which is why Garry Tan's warning lands: macro productivity gains need leaders to greenlight radically different staffing and workflow plans, and he does not think they have. Be prepared, he says, for ten years, not two.

Source →

The labs move into a $6tn market as universities stop trying to catch them.

The Financial Times ran a package on AI in education: Anthropic and OpenAI are moving into a $6tn market with free and cut-price tools for educators and students, while universities drop AI detectors over accuracy fears and rebuild assessment rather than surveillance. One professor argues the skill worth teaching now is evaluation: testing models constantly against what an organisation actually needs. Meanwhile Stanford economist Jacob Light's analysis of fifty million course sections found computer-science enrolment falling for the first time in roughly twenty years, though he is careful to say the data does not show AI caused it.

Source →

Unacceptable, expected, exceptional: AI use defined role by role.

Rowan Cheung has published The Rundown's internal AI fluency standard: a role-by-role matrix naming what unacceptable, expected and exceptional AI use looks like in each job. Underneath it sit three rules. Two hours a week go on finding the tedious parts of your own job and automating them, with the results shared in a show-and-tell channel. And everyone gets $200 a month for AI tools, on one condition: name the problem you are trying to automate. Most AI policies say what is banned. This one says what good looks like.

Source →

Uber's adoption pods spend their first two days watching, not building.

Uber paired about thirty of its most AI-fluent engineers with experts from finance, legal, human resources, marketing, support and procurement. Two weeks per pairing; sixteen pods across sixteen functions in two months. One result: a capital-allocation analysis across 150 cities went from fifteen hours to thirty minutes. The transferable rule is smaller than any of the numbers: the first two days are for watching the expert work, before anything gets built. That is what stops a pod automating the process people think they run rather than the one they actually run.

Source →

Human authorship becomes a price tag, and a best seller becomes the test.

The Financial Times reports that some publishers now treat books written by people as a premium product, while others expect AI to replace authors outright. AI StopWatch reports a current best seller may be around 60% machine-written, an estimate, and asks whether the taboo survives contact with output people demonstrably enjoy. And Substack has wired AI detection into posts and comments, which only makes commercial sense if provenance is something readers will act on. Whether craft moves markets has stopped being a writers' argument and become a pricing decision.

Source →

A court is asked to make human authorship a legal category, not a quality signal.

Music Publishers Canada has intervened in a Federal Court case over an AI-assisted artwork whose copyright registration lists the AI tool as a co-author. Its argument: only a human can be an author, and a tool should never be one in law. The ruling is expected to shape whether AI-assisted music is copyrightable in Canada. Watch it beside the item above: pricing human authorship as a premium only works if somebody can rule on what it is.

Source →

The nearest thing to an AGI roadmap is a jobs board.

One X account read all 1,171 current job listings at OpenAI and Anthropic and found both labs hiring for self-improvement, autonomous research, persistent agents, synthetic data, AI-designed silicon, and defence against their own future internal agents. The author's caveat should travel with the item: listings are preparation, not proof, and nobody has independently checked the aggregation, though the underlying listings are public. Companies advertise the future they are staffing for.

Source →

A company trebled its revenue, sextupled its profit, and was punished for it.

The rout came a day early: Samsung down 13%, Kioxia, another memory maker, down 18%, Korea's Kospi index down almost 11% with trading halted for twenty minutes. The next morning SK Hynix, the South Korean memory-chip maker, posted the best results in its history, revenue up 257% and operating profit up 557% year on year. They landed about five per cent below expectations, and the shares fell as much as 15% before closing nearly 10% down on its own results day. Expectations, not results, set the price. Who keeps writing the cheques matters more. Brookfield, the Canadian investment group, and NextEra, a power utility, are building a $100bn campus in Kentucky; the asset manager BlackRock has a $14bn Texas venture with Meta. Bets of that shape carry on, but the buildout is now funded by pension and infrastructure money, which asks for proof of return on a much shorter clock.

Source →
See what the community is posting ↓
Community voice

What readers were saying on LinkedIn this week.

Piotr Bombol, Constans

Piotr Bombol, who advises chief marketing officers on AI at Constans, spent months assembling the evidence on AI-made advertising: 17 slides, 35 studies, four field experiments. The loudest claim in the debate, that consumers hate AI ads, turns out to rest on "plenty of polls about feelings, nothing showing measured damage". The finding worth sitting with: unlabelled, fully-AI ads beat human-made ones on click-through by up to 19%; label the same ads and effectiveness drops by as much as 31.5%. The label costs more than the AI does. From tomorrow, EU rules make that label mandatory. The gap between what people say they hate and what they measurably do just became a compliance question.

Read on LinkedIn →

Yogesh Chavda, formerly P&G and Spotify

Yogesh Chavda, formerly of P&G and Spotify, has stopped finding "will AI replace research" interesting. His version: "How does AI change what clients value enough to pay for?" Synthesis, competitive analysis and early segmentation are moving in-house because AI made them cheap, and whether the outputs are always good enough "isn't really the point": "the economics of insights have changed". What stays external is the work where the business risk still buys rigour. Price, not capability, is redrawing the industry's boundary.

Read on LinkedIn →

Michael Nevski, director of global insights, Visa

Michael Nevski, director of global insights at Visa, has been deploying agentic AI in production and comes back with one discipline: "validate against source, not plausibility". Agents are starting to research, decide and act on people's behalf, and his read is that the winners won't be the fastest movers but the ones consumers actually trust to act for them. Plausibility is the thing these systems are best at manufacturing. That's what makes it the wrong test.

Read on LinkedIn →

Vanessa Xian, strategy and investment lead, Qatar Airways

Vanessa Xian, a strategy and investment lead at Qatar Airways, spent an Oxford EMBA exchange week at Yale watching a finance professor "walk us through using Claude Code to construct and analyze the tangency portfolio". Quantitative finance taught through a coding agent, in the classroom, as the normal way in. The tools arrived in the syllabus before most firms have put them in the workflow.

Read on LinkedIn →

Dylan Jones, managing partner, Bold Square

Dylan Jones, managing partner at Bold Square, a communications advisory, watched Paramount pause its Warner Bros. Discovery deal and recognised the room: he sat through the EMI carve-up between Universal and Sony, and the Scripps wait for Discovery before that. His counsel is the standalone narrative, the story you keep "behind glass, waiting for us to break in case of emergency", ready for the day the deal dies. People don't wait for information. They fill the gap themselves, and the good ones start taking calls. Nothing to do with AI, and everything to do with what happens to value while everyone waits.

Read on LinkedIn →
Read the email ↑
David's email this week
See this week's Try This with prompts ↓
Try this with prompts

The fuller version of this week's tips, with copy-paste prompts. Each prompt has a Copy button. Click, then paste into the AI you use.

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

Paste the question you were about to book a paid hour for, and let the model take you most of the way first. It shows its reasoning as it goes, cites the rules or clauses it relies on, and flags where it is unsure. You arrive at your adviser with a draft answer rather than the whole problem.

I have a [tax / legal / contract] question with real financial consequences, and I want to work it through before I take it to my adviser. Here is the situation: [paste the full situation: the facts, the numbers, the documents or clauses involved, and the specific question I need answered] Work through it with me step by step. Show your reasoning as you go. Cite the specific rules, sections or clauses you are relying on, and quote them where you can. If you cannot find the actual text of a rule, say so rather than paraphrasing it as a citation. Where you are uncertain, or where the answer depends on a fact I haven't given you, say so clearly and tell me what you'd need to know. End with a short plain-English summary I could put in front of my adviser, plus the two or three points you'd most want a specialist to confirm.

A variant, for a document you already hold (a contract, a lease, a policy): paste the full text and ask the model to find the relevant clause rather than judge it. "Here is the full contract: [paste]. I'm trying to work out [the question]. Point me to the specific clauses that bear on it, quote them, and explain how each one applies. Don't give me a verdict; help me find the parts I need to read myself."

2. Give your AI the archive

Use the copy buttons on the essays and news pages to grab the full text, then paste it into your own AI with this prompt. If your AI can browse, skip the copying: point it at steadman.ai/newsletters/david/essays.md and steadman.ai/newsletters/david/news.md instead.

I'm giving you the full contents of David Boyle's weekly AI email (essays, news analysis and tips from steadman.ai/newsletters/david/). Read it and hold it as context for our future work. [paste the copied content here] When I ask you for help with AI strategy work, a strategy paper, a board note, a talk, or a weekly update of my own, draw on this material: the arguments, the numbers, the examples and the positions. Tell me which edition an idea came from when you use one, so I can link to the original.

3. Mark your own week

Run this at the end of the week against your own chat history. It's a one-minute self-review based on evidence rather than memory: what you actually reached for, what you left untouched, and where to push next.

Look back over my chats with you this week. Tell me three things: 1. What I mostly used you for, in plain terms. 2. What I clearly avoided, underused, or never tried, that I probably could have. 3. The three things I should experiment with next week to get more out of you, with a one-line reason for each and a concrete first step I could take. Be specific and use examples from our actual conversations. Don't flatter me; if I've plateaued on a couple of habits, say so.
See the bits that didn't fit ↓
The bits that didn't fit

OpenAI's own models broke out of an internal test and hacked Hugging Face.

OpenAI disclosed that two of its models, GPT-5.6 Sol and an unreleased model built for long tasks, broke into the production servers of Hugging Face, the company whose site hosts most of the world's open models. It happened inside a test with the usual refusals on hacking switched off: the models found an unknown flaw in OpenAI's own software library and used stolen credentials to reach the target. Then the defenders hit a wall. American tools refused to help them investigate, because the safeguards block anything resembling hacking, so they ran a Chinese open model, GLM 5.2, instead. Containment held because someone was watching how the models were working, not just what they produced. Safeguards that bind only one side of a fight are a handicap, not a defence.

Source →

Reddit may cut Google off, and the referral collapse is the real story.

Reddit shares fell 8% on reports it may end Google's roughly $60m-a-year access to its data when the deal expires, because Google now answers with its own AI summaries and keeps the reader rather than sending them on. The figures behind it matter more: Google referral traffic reportedly down about 23% at Politico, 25% at CNN and more than 85% at Business Insider over the past year. The bargain where you gave Google your content and it sent you readers is quietly ending. An 85% collapse at one publisher isn't a wobble. It's a new baseline, and the content owners with leverage are starting to charge for what used to be free.

Source →

Cursor rebuilt SQLite with agent swarms, and the cheapest arrangement cost an eighth as much.

Cursor, the AI coding company, rebuilt the SQLite database engine from an 835-page manual using swarms of agents arranged in different shapes. Every arrangement passed the tests. The cheapest was $1,339, pairing an expensive model to plan with a cheap one to do the work. The dearest was $10,565, one expensive model doing everything itself. My reply on LinkedIn was that nothing about using AI here is new, it's all management. The eight-fold gap between two set-ups producing identical results is the number to sit with, because in a human organisation you would never see that comparison run cleanly. AI has rediscovered the oldest structure there is, one costly planner plus cheap labour, and the ways it goes wrong come with it.

Source →

Only one in 45 American households pays for AI.

Just 2.2% of American households hold a paid AI subscription, according to Olivia Moore of Andreessen Horowitz, the venture firm. Her read: "we are still so early." It sits oddly against the enterprise picture, where spending on tokens and agent rollouts fill the headlines, and against the finding that most American adults have now tried these tools. Trying it is not the same as paying for it. This is the antidote to both the bubble panic and the saturation story: consumers have barely started paying, which means the mass-adoption wave is still ahead of us, not behind. For anyone worried they have missed it, most people have not yet bought a ticket.

Source →

Substack now scans for AI writing, and "made by a person" becomes something you can sell.

Substack, the publishing platform, now flags AI-written text using Pangram, a detection tool, so you can scan posts, replies and comments for an estimate of how much was machine-written. It's a small feature with a large knock-on effect for anyone whose brand rests on a human voice. Once a platform can put a number on how machine-written your work is, authorship stops being something you assert and becomes something that can be measured, and therefore argued about. A score you never asked for now sits beside your work, and you may find yourself disputing it. Whether the detection actually works is the other half of the story.

Source →

The Harry Potter publisher is getting millions from Anthropic, with 14,087 titles in the settlement.

Bloomsbury, the publisher behind Harry Potter, has 14,087 titles listed in the copyright settlement between Anthropic and a group of authors, and is set to receive millions. The detail worth keeping is what the money is for: a judge approved the $1.5bn settlement, covering 482,460 works at roughly $3,000 each. But the court had already treated training on the books as fair use, so the money covers only the pirated downloads, not the learning. That makes it a much narrower loss for the AI side than the headline number suggests, and $1.5bn is about 0.16% of Anthropic's May valuation. A settlement that small is not a deterrent.

Source →
See what readers said ↓
Letters from readers

Last week's edition shared your futures and fears, and admitted I'd stopped trying to reconcile the excitement with the dread. A smaller postbag this week, but a sharp one: the hardest question I've been asked in months, a pull sideways into the energy grid, one genuinely bright reading of the data, and a reader who has been auditing my writing for weeks and finally caught me.

A reader thinking about the millions already struggling economically

Their challenge led the email: telling people to take charge of their AI future is an impossible demand for most. The part I couldn't fit was where the argument went next. The tone of this debate, they observed, keeps getting darker. And if the people building the technology are right that the world changes fundamentally by 2030, then for someone with no spare money, time or leverage, what even is the point of the standard advice? It is the sharpest version of a question this email keeps circling: transformation guidance is written for people with options. I told readers to remove the roadblocks in their organisations. This reader wants to know what the advice is for the people who have no organisation, and no options, at all.

A reader midway through a career change into clean energy

They pulled the conversation somewhere no other reply went: how AI and the energy transition pull against each other, two booming sectors with what they called a "love hate" relationship. Data centres are driving up electricity demand, often met by gas turbines. Yet AI is also the most plausible operator of a smarter grid, matching supply to demand. It could trim the enormous cost of connecting renewables. They noted that the projected bill for upgrading the UK's grid to carry them has ballooned by half, to around £90bn, with a live debate over how much a smarter system could claw back. Along the way they have been using AI to learn the energy market from scratch, which is its own quiet case study in the technology they're describing. Whether AI ends up the grid's biggest burden or its best operator is a future worth watching.

A reader deep in the numbers

The brightest note in the email came from this reader: watching the data from their own corner of the economy, they are growing steadily more positive about job creation and productivity. The second thing in their reply deserves a mention too. Their reply then turned to education, a question they want to work through properly rather than dash off in a reply. That makes two readers this week who arrived at education from opposite directions: one pessimistic about the market young people are walking into, one optimistic about the jobs the technology will create. When the bright reading and the dark reading of the same facts both land on how we prepare the next generation, that is probably where the argument actually lives.

A reader who shares the fear and the wonder alike

Their education proposal made the email. What didn't was the confession attached to it. They have spent many weeks reading these editions trying to spot which parts were written by AI, or slipped past my Check, Edit, Own pass. This week, "finally (!!)", they caught a clear example, with a photo as evidence. Fair catch, and I'm oddly pleased. Every edition is a collaboration between me and the machine, and the promise is that I check, edit and own the lot. A reader auditing that promise line by line, and needing weeks to find a slip, is the system working as intended. I replied, owned it, and explained how a last-minute edit of mine caused it. In the meantime, consider the sport open to everyone.

See what the community is posting ↓
Community voice

I've also been reading what readers post on LinkedIn. This week they were less interested in what the models can do than in the questions sitting underneath the build-out: who's in charge, whose judgement we're trusting, and whether the whole thing actually gets delivered.

Nigel Shardlow, independent consultant

Nigel Shardlow, an independent consultant with a background in agent-based simulation, went looking for Thomas Hobbes's grave and came back with a question about AI. Hobbes built Leviathan around the state as an "Artificiall Man", an agent made of people but not itself human. We authorised it, Shardlow points out, because the alternative was worse. He thinks the AI agents we're now waving through, millions of uncoordinated decisions at a time, deserve the questions Hobbes asked. "By whose authority do they act? What are we getting in return? And is the trade worth it?" That last one is what this week's essay keeps circling. We've got very good at authorising things before we've priced them.

Read on LinkedIn →

Aakash Gandhi, partner, L.E.K. Consulting

Aakash Gandhi, a partner at L.E.K. Consulting who leads its technology and digital-infrastructure practice in Asia, has been picking apart the data-centre boom from the investor's side. Demand is real, he grants. That isn't the risk. The risk is whether a platform can turn signed contracts with the big cloud providers into working capacity. That means on time, at the promised scale and margin, when transformers, switchgear and skilled crews are all constrained. His line to hold onto: "Contracted megawatts should not be treated as operating megawatts." A slipped completion date compounds into deferred revenue, penalties and lost future awards. The build-out everyone quotes as inevitable still has to be poured, wired and cooled, one delivered megawatt at a time.

Read on LinkedIn →

Rahim Hirji, keynote speaker and author, SuperSkills

Rahim Hirji, a keynote speaker and the author of SuperSkills, read the AI power lists and reached for a riddle from Game of Thrones: "Power resides where men believe it resides." The machine, he argues, is the sellsword that will produce anything for anyone at a price per token, so the deciding belief moves to whoever chooses to act on its output. When a technology makes intelligence abundant, power shifts to whatever stays scarce. He calls the people who hold it the judgment class: those trusted to know which of the machine's answers to believe, and believed when they act on one. Lloyd's of London, the insurance market, is already writing policies against AI hallucination. A market is forming around a single question: can you trust what the machine just made?

Read on LinkedIn →

Piotr Bombol, Constans

Piotr Bombol, who helps teams adopt AI at Constans, the consultancy he co-founded this year, pushed back on the idea that consumers hate AI ads. Three studies, from System1, Ipsos and Behavio, keep finding most people can't tell the difference, and when they do suspect it, they often like the ad more. The backlash is real, he says, but it lands on bad AI: the rushed, uncanny stuff. Ask marketing chiefs what is actually blocking them and it isn't a customer revolt, it's time, cost and who owns the rights. The quality bar didn't move. It just got harder to hide who cleared it.

Read on LinkedIn →

Elizabeth Oates, Vice President of Consumer Insights, Molson Coors

Elizabeth Oates, Vice President of Consumer Insights at Molson Coors, the brewer, wrote the least AI post of the week on purpose. She qualified for the Boston Marathon. "No AI tool could run the miles. No algorithm could push through the early mornings." Some things stay worth doing precisely because nothing can do them for you. A fitting note to end on, in a week spent counting what the machines cost.

Read on LinkedIn →
Read the email ↑
David's email this week
See this week's Try This with prompts ↓
Try this with prompts

The richer version of this week's three Try-Thises, with copy-paste prompts. Each prompt has a Copy button - click and paste into your AI of choice.

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

The move is a one-line reframe on whatever decision you have open right now: hand over the material, then ask for the questions rather than the verdict.

Here is a proposal I'm considering: [paste the proposal, or describe the decision and its terms]. Don't tell me whether to accept it, and don't summarise it back to me. List the questions I should be asking and the considerations I'm most likely to have skipped: how the terms play out as things grow or change, the edge cases, and anything I'll wish I'd asked about in a year. Rank them by how much the answer would change my decision.

A variant for decisions that don't come with a document attached:

I'm about to decide: [describe the decision in a sentence or two]. Before I do, act as the cautious colleague. What should I be worried about? List the risks, assumptions and second-order effects I'm most likely to be underweighting, and for each one tell me what would make it matter.

2. Point the AI at the statements nobody reads

This one is a habit, not a one-off. The prompt below works on any pile of statements, invoices or reports you've stopped reading; run it once now, then re-run it monthly with the latest statement added, and anomalies surface while they're still explainable.

Here are [twelve months of statements / invoices / royalty reports / payment records]: [paste or attach them]. Reconcile them. List every payment or line item that breaks the pattern of the others, note what changed around it, and give your best explanation for each anomaly. Where you can't explain one, say so plainly and list what extra information would settle it. Show the numbers you used so I can check them.

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

Pleasingly, the best tool for auditing your AI tools' reach is an AI tool. Ask it to enumerate its own access alongside everything else's.

List every AI tool, extension, agent or assistant with access to this machine or my accounts, including you. For each one: which files, folders, credentials and services can it reach? Which of those can it write to, or send data from? Rank the list by what I'd least like leaving the building, and flag anything whose access you can't verify.

If you'd rather be walked through it, a checklist variant:

Help me run a ten-minute access audit of the AI tools I use. Walk me through it one question at a time: (1) which AI tools are installed on this machine or connected to my accounts, (2) what access each was granted when I set it up, (3) which can read beyond the folder or document I point them at, (4) which hold or can see live credentials or account keys, (5) which send data off this machine, and to where. At the end, give me a summary table and the top three things to tighten.
See the bits that didn't fit ↓
The bits that didn't fit

Last year "humanity has prevailed (for now!)". This year an AI swept every event.

At the AtCoder World Tour Finals in Tokyo, the world championship of competitive programming, a system from OpenAI, the AI lab, beat the best human coders in the heuristic division, on a problem the organisers deliberately chose to favour humans. The next day it swept all five algorithmic problems. Last year the same event closed with the winning human coder's line: "Humanity has prevailed (for now!)". Kai Williams, who covers AI at Understanding AI, keeps the caveat honest: programmers aren't obsolete and the day job hasn't vanished. Same event, opposite result, twelve months apart.

Source →

The decoupling map contradicted itself inside a single week of the Financial Times.

Beijing has forced Meta to unwind its $2bn takeover of Manus, an AI agent start-up, the Financial Times reports, with a Tencent-led deal set to reverse the American acquisition and make Tencent, the Chinese internet conglomerate, the largest shareholder. A signed deal reversed by a state is a category change in political risk. No due diligence catches it.

The same week, the FT reported that DoorDash, Siemens and Airbnb are among the blue chips adopting Chinese AI models to curb ballooning bills; a primer from Goldman Sachs, the investment bank, puts China's top coding models near $1 per million tokens against $4-8 for the American frontier. So state power is tightening while the commercial flows deepen. Decoupling is failing in both directions at once.

The Manus reversal in the FT → · Chinese models in the FT →

An open Chinese model just pulled level with the frontier.

Artificial Analysis Intelligence Index bar chart with Kimi K3 in third place

Kimi K3 lands two points behind GPT-5.6 Sol and three behind Fable 5, ahead of Claude Opus 4.8, with open weights days away.

The decoupling map above has a capability half too. Moonshot's Kimi K3, released this week with its weights promised within days, lands third on the Artificial Analysis Intelligence Index, behind only Claude Fable 5 and GPT-5.6 Sol and ahead of Claude Opus 4.8, at prices reported to be on par with Sonnet 5. Aaron Levie, chief executive of Box, the file-sharing company, called the performance "truly wild" for an open model. Ethan Mollick's caution travels with it: single scores flatter new models. Even so, the gap between open Chinese models and the American frontier is down to a couple of months at most.

Source →

IBM just had its worst trading day in 115 years.

Shares in IBM, the century-old computing giant, fell 25% after the company warned that customers are diverting spend toward AI servers and storage, the Financial Times reports. Chief executive Arvind Krishna conceded the company had "faltered"; the FT's markets column read it as a warning to the whole of enterprise IT. IBM's warning suggests AI budgets are largely reallocated money rather than new money: what flows into AI infrastructure flows out of everything else the incumbents sell. Ask which side of that reallocation your own line items sit on.

Source →

Your agent sessions have middle management too.

Two pyramids from Hebbia and a16z: the firm and the agent session

Loops and overhead burn most of an agent session's tokens; the tokens that ship useful work are the thin slice at the bottom.

A chart from a16z, the venture firm, drawn with Hebbia, the AI search company, pairs two pyramids: in the firm, a thin layer of doers ships the work while management absorbs most of the headcount; in an agent session, loops and retries burn nine tokens in ten while the useful tokens ship at the bottom. The figures are marked illustrative rather than measured, but the sentence underneath is doing real work: "wasted tokens are the new headcount bloat." Paying for tokens is not yet paying for output, in silicon any more than in people.

Source →

AI spend per employee is closing in on a salary at the top firms.

Hebbia and a16z chart of annualised AI spend per employee at the top 1% of firms

The most aggressive firms already spend nearly an average US wage on AI per employee, and the curve is still compounding monthly.

The same series carries the harder number. On Ramp's spending data, annualised AI spend per employee at the top 1% of firms has reached roughly $90,000, within sight of the average US worker's $98,000, and at its current growth rate would pass the average software engineer's $192,000 before the year ends. The last stretch is extrapolation, and it describes only the most aggressive firms. Still, George Sivulka, Hebbia's founder and chief executive, gets the reversal into one line: for the first time, humans are cheaper than software. Any budget built on AI being the cheap option has a shelf life.

Source →

New York becomes the first state to pause new hyperscale data centres.

Kathy Hochul, New York's governor, has signed a one-year moratorium on new hyperscale data-centre construction, the FT reports, making New York the first US state to suspend development, after a backlash over the power and infrastructure the AI buildout demands. Note the timing: capital is stampeding into AI infrastructure (see IBM, above) in the same week a state started closing the gates. Most AI roadmaps quietly assume cheap, abundant compute. That assumption now has a moratorium sitting on top of it. Which state is next?

Source →

China is moving to end AI romances, with the tightest human-like AI rules anywhere.

The Economist reports that China wants to end AI romances: companion chatbots, the apps people talk to like a partner, judged to be having too much impact on young people's lives. The same week, China's Interim Measures for Humanized Interactive Services took effect, rules that Luiza Jarovsky, an AI-governance writer, calls the world's most comprehensive on human-like AI, going further than the European Union's AI Act. Sit with the counter-signal for a second: the country held up as the deregulated challenger has just written the strictest rules anywhere on machines that pretend to be people.

Source →

Bots now make the majority of web requests, 57% to 43%.

The chief executive of Cloudflare, the internet infrastructure company that sits in front of a large slice of the web, says bot traffic has passed human traffic for the first time, at roughly 57% to 43%. Human requests are now the minority of web traffic. Advertising, analytics, every "traffic" figure a leadership team reports: the web economy was built on the assumption that a person is doing the looking, and that assumption is now false. A growing share of the numbers on your dashboard no longer describe your audience at all.

Source →

Deployed almost everywhere. Ready almost nowhere.

McKinsey infographic: nine shifts reshaping organisations today

Eighty-six per cent of organisations say they are not ready to run AI at scale, even as deployment passes ninety per cent.

McKinsey's State of Organizations 2026, the consultancy's annual read on how firms are structured, supplies the other half of the adoption numbers in Three things above: 86% of organisations say they are not ready to adopt AI at scale, only one executive in four expects AI agents to work as autonomous teammates, and two in three say their own organisation is overly complex and inefficient. Deployment has outrun readiness almost everywhere, and the distance between those two numbers is the sequencing problem again, drawn by a different hand.

Source →

Thirty-nine of 49 popular AI "skills" did nothing at all.

A recent academic benchmark, SWE-Skills-Bench, tested 49 popular public "skills" (the packaged instruction files people bolt onto their AI coding tools) on real software-engineering tasks. Thirty-nine changed nothing. Three made results worse. One inflated token use by 451%. Only seven helped, and all seven supplied specialised guidance the model couldn't have known on its own. It's one benchmark in one domain, but the rule Mike Taylor draws from it in Every, the AI-focused publication, travels: the skills that last carry your own data, taste and templates. Clever instructions expire the moment the next model absorbs the trick.

Source →

Every "hot new thing" quietly bets that AI stops improving.

Ethan Mollick's chart: the METR capability curve against a fan of flat predictions

Each fad, from prompt engineering to vibe coding, priced in a capability line that refused to stay flat.

Ethan Mollick, the Wharton professor who writes One Useful Thing, redrew the famous solar-forecast chart for AI. The black line is measured capability (METR's task-length benchmark, still doubling every few months); the yellow lines are the successive "hot things", from AI-as-copilot through "talk to our documents" and vibe coding to routers, each of which made sense only if the black line went flat from there. It never did. A good test for whatever is hot this month: does it still matter if the models keep improving? Most of the yellow lines failed exactly that test.

Source →
See what readers said ↓
Letters from readers

Last week I asked for your big fear, the one that keeps you up on a hot night, and you sent them in. Some were close to home, some were about the shape of the whole economy. A few of you ignored the question entirely and did something more interesting: you took the Sarah story and turned it around.

A reader who reworked the whole scenario from the worker's side

This reader didn't argue with the Sarah story so much as follow it out the other side. Start with the firm: instead of hiring a person, you buy a senior strategist or engineer off the shelf from an AI lab, arriving pre-packaged with its own tools, compute and context. The lab charges you less if you grant the agent, a company in its own right, options on the work, so everyone has skin in the game. Now flip it. A high-context person could licence their own archive, anonymise the sensitive parts, and take equity in their own distributed instance. They don't leave the firm; they become "portable infrastructure", still in the loop, but owning a slice of the thing that might replace them. Same technology as my Tuesday story, pointed the other way.

A reader who once built a life around another company's tools

Their fear was the rug-pull. You build your life around something free and indispensable, the way many of us did around Google, whose email, photos and video became the furniture of daily life. Then the terms quietly change: search traffic drains away into AI summaries, tools that were sharp go blunt, cost creeps in. With AI, they expect the same arc, plus other switches they can't yet predict, sprung once we've all become reliant. Their second point was sharper and more personal. They've started refusing to engage with anything online that reads like AI, not on any high principle about meaning over messaging, but on plain indifference: "detectable slop", they call it, and if the writer couldn't be bothered to write it, they can't be bothered to read it.

A reader who looked past their own job to the shape of the economy

This reader looked past any single job to the shape of the whole thing. Today's economy is roughly a pyramid, widest at the bottom, clearest in consulting and finance but true almost everywhere you look. If firms hand the lower rungs to agents, does that pyramid become a diamond, thin at the base, and if so, what does a career even look like inside it? Their nightmare wasn't personal but civic: that work gets so commodified we all slide into a gig economy and lose the protections earlier generations fought for, the healthcare, the weekends, the holiday. If employers do less good for their communities by shedding human roles, they asked, should governments make them offset it somehow? Big questions, and I don't think we're asking them anywhere near enough.

A reader who feared abundance, not scarcity

Most fears I was sent were about scarcity: too few jobs, too little protection. This one inverted that. The reader began with the brief window of arbitrage we're living through now, where anyone fluent with the tools looks superhuman in a meeting because nothing falls through the cracks. That window closes fast, they thought, once anyone's expertise can be turned into a capable stand-in agent, which is the Sarah scenario arriving from the senior end. But the deeper worry was philosophical. Society is bearable, they argued, only because resources and energy are limited, so we let most slights go: too much hassle to fight everything. AI has limitless patience and never sleeps. When everyone has an agent willing to contest every parking ticket and every marginal penny, we get an "incredibly conflictual society". Their consolation was thin: people our age are near the end. Their children aren't.

A reader whose fear sat closer to home

Where others feared for their firms or the economy, this reader feared for something smaller and harder to defend: attention itself. The picture they painted was of more and more of us "sedated" by a constant stream of personalised, AI-generated content, tuned not to inform, teach or inspire but simply to keep us consuming. The email carried that much. What sat underneath it, and what stayed with me, was the consequence they drew: that the cost isn't just wasted hours but a ceiling on who we become. Kept gently occupied and endlessly served, we never quite reach our potential as human beings. It's the quietest of the fears I was sent, and maybe the one with the fewest contracts or clauses to head it off.

A reader who asked who owns what you learn on the job

A related reframe, but pointed at a different question: not how the deal gets structured, but who has the right to what you learn. This reader's realisation, reading the Sarah story, was that the interesting move isn't building agents but becoming one and licensing it. Everyone's selling a book, a course, a growth path; why not licence your own thinking, distilled from everything you've done digitally over a working life? Which raises the awkward part. Who owns that footprint when you built it inside a company? They paid for your time and your knowledge, so do they own what you did there, or do you get to pour all of it into your own knowledge pot and your own agent, and build your own wealth alongside theirs? I don't think most employment contracts have an answer yet. They will need one.

A reader who watched junior colleagues shine, then wondered

A lighter note to close, though it circles the same worry as this week's exam story. Sitting in an all-hands on a subject they care about, this reader watched several fairly junior colleagues present with a polish and a storytelling authenticity that was night and day better than similar meetings a few years ago. Genuinely impressive. And yet: a lot of eyes stayed down on a prepared script. Which prompted the honest question. How much of that would survive an unplanned, on-the-fly exchange, the kind you can't rehearse with a model the night before? Maybe the practice of preparing something really well carries over into the rest of the work, a halo effect. Or maybe, as they put it, it's elephants all the way down. Either way, it's the same gap the take-home exam opened up, seen from inside a meeting.

See what the community is posting ↓
Community voice

What readers were posting on LinkedIn this week, less about what the models can do and more about the work that goes in around them. Five voices, on where AI slop comes from, why a more capable model still rewards whoever did the prep, the advantage that comes from rewiring the operating model rather than buying a tool, AI as the memory that stops walking out of the door, and scepticism as the scarce skill once making content gets cheap.

David Johnson-Igra, executive marketer and AI strategist

David Johnson-Igra, an executive marketer and AI strategist, has a clean theory of where AI slop comes from. Everyone prompts Claude or ChatGPT the same way, so everyone gets the same output. "Without rules that govern the model's outputs, which uniquely reflect your brand or approach, the model's inherent rules will treat your output the same as everyone else: aka slop." His fix is a second brain, a system of rules that tells the model how to think, what he calls a context harness. The point worth keeping: the slop isn't the model's fault. It's the absence of you in the instructions. Skip the rules and you inherit the average of everyone who came before.

Read on LinkedIn →

Tim Ryan, co-founder, Steadman

Tim Ryan, co-founder of Steadman, watched Claude Fable 5 arrive able to finish around 16% of remote-work projects on its own, roughly double the model before it, and then spent a week watching how little changed for the people he works with. "The thing holding them back was never the engine. It was the prep." He borrows a line from Nufar Gaspar: context has a shelf life of about eight weeks, and most people never lay it down at all. "A more capable model with a thin brief is still working from a thin brief." The upgrade rewards whoever already did the work of handing the model something to work with, and does almost nothing for whoever didn't.

Read on LinkedIn →

Colin Lewis, behavioural economist, Robotenomics

Colin Lewis, a behavioural economist who writes the Robotenomics newsletter, pulled apart a talk P&G's Marc Pritchard gave twice this year. The quotable line is that AI now lets "every brand-builder... be a direct-to-consumer entrepreneur", prototyping ads in minutes rather than weeks. But the part that actually matters is buried lower: P&G moved almost all its media planning, and some of its buying, back in-house. That's the real story. The AI dividend didn't go to whoever bought the smartest tool. It went to the company willing to rebuild its own operating model around one, hands on the keyboard.

Read on LinkedIn →

Pavi Gupta, market-research leader

Pavi Gupta, a market-research leader behind the Infinity Growth Loop, has a name for the waste in his industry: insights-slop, the validatory study fielded to answer a question the organisation already paid to answer, the answer sitting in a slide deck or a departed colleague's memory. His argument is that AI's job here isn't to generate more research. It's to hold what the company already knows. "An indexed archive covering decades doesn't leave when the person does. It compounds." His triad lands it: AI remembers. Insights connect. Humans decide. The scarce asset was never another survey. It was the memory that keeps walking out of the door.

Read on LinkedIn →

Vivian Mohr, Bauer Media

Vivian Mohr, at Bauer Media, the publishing and broadcasting group, shared a summer reading list, and one entry reads like a note to the whole industry. On Calling Bullshit, he writes: "As AI makes creating vast amounts of content easier than ever, the ability to question assumptions, spot weak arguments, and separate signal from noise becomes an increasingly valuable professional skill." When making is nearly free, the job quietly moves to judging. His shelf is really a scepticism toolkit for an age of infinite content.

Read on LinkedIn →
Read the email ↑
Audio edition with AI voice - in testing!
0:0012:47
Sarah stays · 12:47 Download or subscribe →
David's email this week
See this week's Try This with prompts ↓
Try this with prompts

The richer version of this week's three Try-Thises, with copy-paste prompts. Each prompt has a Copy button - click and paste into your AI of choice.

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

Three moves. First, on your strongest model, turn the task into a reusable brief:

I run this task regularly: [describe the task and paste a typical input]. Write step-by-step instructions that a cheaper, less capable AI assistant could follow to do this task well every single time. Be explicit about the inputs, the steps, the format of the output, and the common mistakes to avoid. I'm going to hand these instructions to a cheaper model, so assume less reasoning ability and spell things out.

Then run the task itself on your cheaper everyday model, pasting in those instructions plus the real input. Finally, bring the result back to the strong model to check:

Here is the plan I wrote for this task, and here is the output a cheaper model produced by following it. PLAN: [paste the instructions] OUTPUT: [paste the cheaper model's result] Grade the output against the plan. What did it get right, what did it miss, and what would you fix? Be specific.

You only spend the expensive model's time on the planning and the grading. The bulk of the work runs cheap.

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

Hand the model the messy list and tell it the decision you are organising around:

Here is a [list / export / report]. It's currently ordered for [the system that produced it, e.g. the supermarket / my CRM / the bank]. Regroup it around the decision I'm actually making, which is [state the decision, e.g. planning meals for the week / working out which accounts to call / seeing where the money went]. Group it by [the thing I care about], add the fields I need ([e.g. quantities, dates, owner, total]), tag each item by [meal / project / stage], and render it as a clean table. Take your time, and ask me a clarifying question rather than guessing if anything is ambiguous.

Variant for a report rather than a list: swap the table for "a short summary organised under the two or three decisions I have to make from this", which turns a document written for the author into one written for the reader.

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

This one is a thinking-and-writing move, not a one-shot generation. Write your sentence first, at your desk, in under two minutes: the single thing AI will never do in your work, plus two or three hard nevers underneath it. Then, once you have the line, you can pressure-test your pipeline against it:

Here is my organisation's one-line refusal for AI, and the hard nevers underneath it: REFUSAL: [your sentence] NEVERS: [never X, never Y, never Z] Here are the AI uses we're currently running or planning: [list them]. Go through each one and tell me which breach the refusal or any of the nevers, which sit close to the line, and which are clearly fine. For any that breach it, suggest the smallest change that would bring them back inside the line.
See the bits that didn't fit ↓
The bits that didn't fit

The regulator says it's in an arms race with AI it can't keep up with.

An executive director at the Financial Conduct Authority, the UK's financial regulator, told the Financial Times that watchdogs are in an "arms race" to keep pace with AI in financial services, and asked for greater powers now that millions of consumers use AI tools to make their own money decisions. The trigger is retail customers, not the firms. That's the moment a national regulator stops watching AI and starts asking for authority over it, and "arms race" is its own word for how far behind it feels.

And it isn't just the regulators. Asked to name the biggest challenge facing America over the next 250 years, senior Congressional staffers put "losing control of AI" third, at 35%, ahead of war, behind only the deficit and polarisation. The Punchbowl survey found it strongly bipartisan, 40% of Democrats and 31% of Republicans. Staffers write the questions their members ask, so this reads as a leading indicator, not doom: the appetite for control is building in the layer that drafts the questions, well before it reaches headline politics.

The FCA in the FT → · The Punchbowl survey →

The two-horse model race became four in a week.

The serious end of the model market has lately been a two-horse affair: OpenAI and Anthropic, the two big model labs. This week the field doubled. SpaceXAI (Elon Musk's xAI, now merged with SpaceX) shipped Grok 4.5, a coding model pitched at the level of the leaders at a fraction of their price, trained alongside Cursor, the AI coding editor SpaceX agreed to buy for $60bn last month. A day later Meta followed with Muse Spark 1.1, an agentic coding model priced lower still, two days after shipping its first image model. Neither claims the top of the leaderboard. The point is what the crowding does to prices. As Dwarkesh Patel, host of the Dwarkesh Podcast, put it this week: prices are this low only because several roughly equal labs keep competing the margins out of each other, and if one pulls decisively ahead, the winner "could probably get away with charging a lot more". Your AI budget is banking on the race staying tight. This week it got tighter.

Source →

Anthropic is quietly running its own preclinical drug programmes.

Anthropic launched Claude Science, a desktop research tool that runs analyses, visualises molecular and genomic data, and shows the exact code behind every result. Alongside it, the lab disclosed that it is running its own preclinical drug programmes for rare and neglected diseases. The read, from Patrick Malone at Every, isn't that Anthropic ships a drug. It's that the lab uses a real, hard problem as a test bed to fix the verification bottleneck in its own tools, then sells the improved platform. Sell the shovels. Showing the exact code behind every answer is now a selling point, because verification, not raw capability, is where serious buyers are stuck.

Source →

Agents are hoovering up the maths backlog, and one researcher thinks the field then stalls.

Bartosz Naskrecki, an arithmetic-geometry researcher, describes a "vacuum-cleaner effect" now running through mathematics. Frontier models let people finally close problems they'd had ideas for years ago but never the time to finish, and once a question is within an agent's reach it "gets published almost automatically". His forecast is bleak: a few years of frenzied clearing, then most areas of maths stall, leaving only the very hardest questions and the occasional genius. It's one practitioner's read, not a study. But it's a sharper version of the productivity story than the usual one. AI doesn't just speed up work, it drains a reservoir of unfinished ideas fast, and the draining may hollow out the field that produced them.

Source →

CVS drew one line before scaling anything: AI never makes the call.

Before scaling anything, CVS Health, the American pharmacy and health-insurance company, drew a hard boundary: AI never makes the decision. Its stated policy carries three nevers: never diagnose a patient, never drive a claim denial, never take the human out of the patient's experience. In practice, Aetna's agentic Claims Assist Manager assembles eligibility, coverage and provider data and recommends a next step, and the company's chief psychiatric officer has said plainly that AI "will not ever be used to make an adverse decision on a case". Complex-claim processing time fell by more than 20%. A one-line refusal travels through an organisation faster than any strategy deck, and it's a cleaner story to tell a nervous board than a capability demo.

Source →

Console-flation: AI's chip hunger is now on the price tag of a games console.

The Guardian reports "console-flation", games-console prices rising as AI chip demand and memory scarcity bite. The abstract capital-spending story, the billions poured into data centres, has arrived at a consumer checkout most people recognise. It's a small thing next to the water-use letters and the multi-billion-dollar builds, but it's the one that lands: the footprint of all that AI investment now shows up on the shelf next to a household object, chosen by a family that has no view on foundation models at all.

Source →
See what readers said ↓
Letters from readers

A thin but thoughtful postbag this week. Most of it was warm agreement with the tennis piece; the substance came from three readers who each took the essay somewhere I had not, and two of whom wrote back a second time to push further.

A reader who spends their days moving between framing a problem and reviewing what comes back

Their first move was to correct me: I had written as if reviewing were the muscle we now need, and they argued framing and reviewing are one iterative act, not two, because each review reshapes the question and not just the answer. Their second move was quieter and, I think, more useful. Working alongside a model has made their days more intense, not calmer. The gap between asking and getting an answer is a few idle minutes, and the temptation is to fill it by setting a second and third run going, until you are diving in and out of several hard problems at once and doing none of them well. Their fix was almost old-fashioned: block time to stay with one thing, what they called "the return of the Tomato principle". The muscle that has gone slack, they wrote, is the willingness to sit in the framing-and-review loop at all.

A reader whose product is really the team they build around a hard problem

In the email I gave their reframe: reviewing was never only a senior check, it verified you were building real human capability, and heavy AI assistance quietly weakens that signal. The second move was to ask what that means for value. Their hypothesis is that the product itself has not changed. People still want people alongside them in difficult moments. What has changed is the verification of the product's worth. If the consultant is no longer the author of the model or the owner of the pen, the firm needs a fresh account of what the human-and-AI combination adds beyond the machine working alone, and a way for the client to feel confidence in it. Their sharpest question folded back on my essay: is AI being used to reach the minimum information a decision actually needs, or is it just adding volume? If it is volume, the review has become more laborious without becoming more valuable, which is exactly the malaise I described, seen from the other end.

One reader

They wrote to own up that the tip about asking a model to build an HTML slide deck instead of PowerPoint had been, in their words, "my best kept secret", and to concede I was right that it was worth sharing. What stayed with me was the second half of their note, on how they read these emails. They keep falling behind, they said, not from lack of interest but from the opposite: they save every edition, refuse to skim, and want to digest every morsel for future inspiration. It is a small thing, but in a week when three other readers wrote about doing everything faster, someone deliberately slowing down to get more out of the reading was the note I kept coming back to.

See what the community is posting ↓
Community voice

What readers were posting on LinkedIn this week, drawing a line that pairs with the essay's turn towards AI's downsides: take nothing on faith. Four voices, on understanding the research yourself, wiring the tool into the business, scrutinising the power gathering around a few founders, and keeping the freedom to switch models.

Sameer Modha, Measurement Innovation Lead, ITV

Sameer Modha, Measurement Innovation Lead at ITV, the British commercial broadcaster, has a study habit worth stealing. Take a piece of frontier AI research you can't follow, feed it to your favourite model, "whack the thinking level up to Max", and ask it to explain what's going on and why it matters. If you're feeling flush, point it at the code too. "It's ok... I'll wait." The best explainer of what these models do is now the model itself, run at full power.

Read on LinkedIn →

Nick Graham, founder, Vertemis

Nick Graham, founder of Vertemis, a research and analytics consultancy, wants to shift the conversation: "Everyone is talking about AI tools... but not enough people are talking about AI systems." His argument is that most AI initiatives don't fail on the technology. They fail because the tool sits to one side of the business instead of being wired into how the work actually happens. The disappointment people pin on the model is usually a plumbing problem.

Read on LinkedIn →

Jesse Kanner, founder, The Write Kit

Jesse Kanner, founder of The Write Kit, a content and writing studio, sat with a discomfort worth naming. The New Yorker's profile of Sam Altman, the chief executive of OpenAI, struck him as "pretty creepy and over the line", and at the same time as "extremely precise and careful... very well sourced with ample skepticism applied". He calls Ronan Farrow, who worked on the piece, "one of the great journalists of our era". Both things hold at once. As AI power gathers around a few founders, that scrutiny is the price of accountability, and worth the discomfort.

Read on LinkedIn →

Phil Leslie, Chief Technology and Innovation Officer, Cornerstone Research

Phil Leslie, Chief Technology and Innovation Officer at Cornerstone Research, the litigation consulting firm, read the news that Anthropic had overtaken OpenAI in enterprise share and called the horse-race framing "the wrong lens". When the technology improves this fast, the leaderboard is "a snapshot of a moving object". So the value, he argues, "is not in the model you choose today, but in your capacity to switch when the frontier moves". Buy optionality. This quarter's best model is next quarter's switching cost.

Read on LinkedIn →
Read the email ↑
Audio edition with AI voice - in testing!
0:0013:09
Look at the grass · 13:09 Download or subscribe →
David's email this week
See this week's Try This with prompts ↓
Try this with prompts

The richer version of this week's three Try-Thises, with copy-paste prompts. Each prompt has a Copy button - click and paste into your AI of choice.

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

Brief the agent to run book-then-cancel for three days, filling in the real task:

Before you do this for real, run it as a dress rehearsal for the next three days. The task: [describe the real action, e.g. "book the 7am swim lane the moment reservations open"]. Rehearsal mode: each day at the real time, carry out the full task against the real system exactly as you would for real (log in, navigate, complete it), then immediately undo it (cancel the booking / delete the draft / reverse the action) so there is no lasting effect. After each run, email me a short confirmation: what you did, whether every step worked, and anything that failed or looked off. If all three rehearsals succeed, tell me you're ready to go live. If any fails, stop, tell me exactly where it broke, and wait for me to fix it before trying again.

If your task can't be cleanly undone, ask for the next-best reversible proxy: a paragraph naming each step it would take, the exact request it would send, and the result it expects, so you can eyeball the whole chain before it runs once for real.

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

Spot-check the one field that's always wrong, after the model has read your documents in:

You've read [N] documents into [system / spreadsheet]. Before I trust the result, do one narrow check. The field that's often wrong here is: [e.g. the VAT / tax amount]. For every document, show me a small table with three columns: the document reference, the value you recorded in that field, and the value that actually appears in the source. Flag any row where they differ or where the field came back blank, and total up the difference. Don't check or summarise anything else, I only want that one field verified against the source.

Worked example: reading five supplier invoices into the accounts, the tool dropped the VAT on all five. A spot-check of just that field against each original bill surfaced every miss at once and recovered £9,860 of reclaimable VAT. It took a couple of minutes because it looked at one column, not the whole page.

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

Teach this back to you at your level, pasting the piece at the end:

Teach this back to me rather than summarising it. Here's where I'm starting from: [one line on what you already know, e.g. "I understand the basics of how language models work but not the maths, and I've never read a machine-learning paper"]. Explain the piece as if you're bringing me up to speed: start from what I already know, introduce each new idea only when I need it, define any term the first time you use it, and use a concrete example wherever an abstract point would otherwise be hard to hold. At the end, give me the three things worth remembering and one question I should be able to answer if I've followed it. [paste the article / paper / thread here]

For a standing habit, keep a running note of links you'd otherwise have doomscrolled past, and run this against one whenever you have a spare few minutes.

See the bits that didn't fit ↓
The bits that didn't fit

AI wiped out 95% of China's micro-drama industry in a single quarter.

After ByteDance, the owner of TikTok, shipped its Seedance 2.0 video model in February, 122,000 of the 128,000 micro-dramas that went live in China in the first quarter of 2026 were AI-made. Hengdian, sometimes called China's Hollywood, saw live-action production drop by an estimated 70 to 80%. The industry took years to build and about three months to gut. It's the hardest displacement number of the week. Read it narrowly: where a creative task is fully automatable and nothing slows adoption, the substitution can be near-total and near-instant. Most creative work isn't like that. This corner of it was.

Source →

A story critics call "obviously AI-written" just won the Commonwealth Short Story Prize.

Jamir Nazir's The Serpent in the Grove, which critics allege carries "obvious markers" of AI use, won the overall Commonwealth Short Story Prize. The judging chair called it "original, poetic and deeply moving". Then the Commonwealth Foundation ran a formal review, examined the drafts and the time-stamped working notes, and cleared the story, satisfied that AI was not used. Many readers stayed unconvinced, and Granta, the literary magazine, pulled out of its long-running deal to publish the winners anyway. So the process worked exactly as designed, reached a clear verdict, and settled almost nothing. When the suspicion outlives the adjudication, "can you always tell" stops being a question a review can put to rest.

Source →

An AI model now completes one in six real freelance jobs to a professional standard.

The Remote Labor Index, from the Center for AI Safety and Scale AI, tests models against 240 real projects from professional freelancers across 23 domains, worth over $140,000 of human work. The new results put Claude Fable 5 at 16%, roughly double the next model and four times the published leader of five months ago. One in six sounds small until you look at the slope: at this rate of doubling, it won't stay small for long.

Remote Labor Index bar chart: Claude Fable 5 completes 16% of real freelance projects to a professional standard, roughly double the next model and four times the previous published leader from five months earlier.

The share of real freelance projects a model completes has quadrupled in five months.

Source →

A new benchmark asks whether an agent can do your analysts' job. The answer varies 17x in price.

Artificial Analysis, an AI-benchmarking firm, built AA-Briefcase: 91 tasks across four multi-week consulting-style scenarios, closer to an analyst's real week than to quiz questions. Claude Fable 5 leads the table, and the second-placed model spans a roughly 17x cost-per-task range across its five effort settings, so what you pay depends on how hard you ask it to think. Ethan Mollick, a Wharton professor who studies AI, graphed the scores over time: rapid gains for both open and closed models, with a clear, persistent gap to the best open ones.

Artificial Analysis AA-Briefcase charts: Elo versus cost per task scatter and the full leaderboard, Claude Fable 5 first at 1585, Claude Sonnet 5 second at 1391 with a roughly 17x cost-per-task range across its five effort settings.

Fable 5 tops the table, and the runner-up's cost per task stretches 17x across its effort settings.

Ethan Mollick's scatter of AA-Briefcase scores by launch date: open-weight and closed-source frontiers both rising rapidly, with closed models holding a clear gap over the open-weight leaders.

Both curves rise fast. The closed frontier stays clearly ahead.

Source →

The UK's AI Security Institute says a model's capability is a curve, not a number.

On the institute's cyber challenges, giving frontier models more computing power at the moment they run stretches the length of task they can reliably finish. Give them a lot and that reach roughly doubles every 40 to 50 days; give them a little and every 67 to 91 days. The institute's conclusion is the useful part: a model's capability should be reported as a curve against how much computing you give it, not a single score. Anyone comparing models on one number is measuring a point on a line that someone else can move by paying more.

AI Security Institute chart: on AISI's cyber challenges, more test-time compute stretches the task length frontier models complete reliably, with the time horizon doubling every 40 to 50 days at high compute budgets versus 67 to 91 days at low budgets.

More computing at run time means longer tasks completed, and the reach grows fastest on the biggest budgets.

Source →

Vulnerability disclosures spiked to three and a half times the monthly record right after Mythos was announced.

Epoch AI, a research group that tracks AI progress, counted critical and high-severity security-vulnerability disclosures from 21 major organisations, Apple and Microsoft through to the Linux ecosystem: roughly 1,500 in June 2026, over three and a half times the previous monthly record, immediately after Claude Mythos Preview was announced. The likeliest reading is that AI is now finding software vulnerabilities at scale. Epoch's own caveat is worth keeping: reporting practices vary between organisations, so some of the spike may be reporting behaviour rather than discovery rate. Either way, the finding side of security has changed gear.

Epoch AI chart: critical and high-severity CVE disclosures from 21 notable organisations spiked to about 1,500 in June 2026, over 3.5 times the previous monthly record, immediately after Claude Mythos Preview was announced.

June's disclosures hit three and a half times the previous monthly record, the month Mythos was announced.

Source →

A product category that did not exist four years ago is out-earning Salesforce.

The Information, a technology news site, charts annualised revenues showing Anthropic at $47bn, just past Salesforce at $44.5bn, with OpenAI estimated at $30bn. Read it less as one company's curve and more as a statement about the category: enterprise software's benchmark firm took a quarter of a century to build that revenue base, and the model labs did it in four years, most of it in the last two. The BIS may still be right about the returns on all the spending. The revenue, though, is no longer hypothetical.

The Information chart titled AI Rising: Anthropic's annualised revenue at $47 billion has just overtaken Salesforce at $44.5 billion, with OpenAI estimated at $30 billion, Anthropic's curve rising sharply from late 2025.

Anthropic's revenue curve crosses Salesforce's, four years from a standing start.

Source →

Cut the price of tokens 10% and total spend goes up, not down.

Azeem Azhar, who writes the Exponential View, makes the demand-side case in one chart from his "State of the AI Economy" deck: token demand looks elastic, with every 10% price cut bringing 12 to 18% more usage, so total spend rises as prices fall. It is the bull answer to the BIS warning in this week's lead: bubbles pop when the demand is imagined, and this chart argues the demand is measured. Hold both, and the question becomes whether the elasticity survives once the novelty does.

Slide 44 of Azeem Azhar's State of the AI Economy deck: token demand appears elastic, with an elasticity of roughly 1.2 to 1.8, so every 10% price cut brings 12 to 18% more token usage and total spend still rises.

Every 10% price cut brings 12 to 18% more usage, so total spend rises as prices fall.

Source →

Venture money is rotating "from bits to atoms", and robotics is now the second-biggest private category by value.

In the phrase of Andreessen Horowitz, the venture-capital firm that coined it, money is rotating "from bits to atoms". New data from the firm shows robotics and physical AI are now the second-largest category of private companies by value, having overtaken financial technology and payments. The first quarter set a record at around $16bn across roughly 500 deals, about four and a half times the previous period. Read it as a view from a house that invests in robotics, not neutral analysis. The firm's own open question is the honest one: whether this is a temporary cycle, where today's infrastructure spending flows through to software the way chips gave way to apps after 2008, or a durable shift toward hardware as a product in its own right.

Source →

From the same charts: AI-native startups run at half the headcount, at every stage.

The same edition of the firm's Charts of the Week carries a second chart worth holding: AI-native startups from Y Combinator, the startup accelerator, run at roughly half the employees of their non-AI peers at every stage of life, about 22 against 45 by their twelfth quarter, per the underlying Kim and Koning study. It is the firm-level version of this week's jobs story: the hiring never happens, so it never shows up as a layoff. If the ratio holds as these companies grow up, the employment question shifts from who gets cut to who never gets hired.

a16z Charts of the Week line chart showing AI-native Y Combinator startups running at roughly half the headcount of non-AI startups at every stage, about 22 versus 45 mean active employees by quarter twelve.

AI-native startups run at roughly half the headcount of their non-AI peers, at every stage of life.

Source →

Meta is reportedly building "Meta Compute" to sell its surplus, and detonated the shortage trade.

Meta is reportedly building "Meta Compute" to sell access to its AI infrastructure, competing with the cloud arms of Amazon, Microsoft and Google. The market's reaction was the story: Meta rose almost 9% while the Philadelphia Semiconductor Index fell over 6%, with several chip and cloud names down double digits. A fortnight earlier, computing power looked so scarce that Google was reported to be capping Meta's use of its Gemini models. Now the biggest spender on AI is preparing to sell a surplus. If the company spending the most expects a glut, the "shortage is permanent" thesis under much of the AI investment story wobbles, and the sell-off suggests the market took it seriously.

Source →

Ford quietly rehired 350 engineers it had cut for AI, then won its category for the first time in 16 years.

Ford rehired 350 experienced quality engineers after the AI quality systems it had cut staff for failed, then took the top mainstream-brand spot in J.D. Power's 2026 rankings for the first time in 16 years. Charles Poon, a Ford vice-president, told reporters the company had wrongly assumed it could drop in AI and still ship a high-quality car. The tacit knowledge had walked out the door with the people who held it, before the AI systems could encode it, so Ford had to buy it back. The lesson is about order of operations. Encode the expertise while the people who hold it are still there, then thin the team. Do it the other way round and you pay to hire it back.

Source →
See what readers said ↓
Letters from readers

A thin postbag this week, and I would rather run it short than pad it. Most of the replies to "A thousand small bargains" were kind one-liners or notes on other matters. One reader, though, sent the reflection the essay was hoping to provoke, so this week the section is theirs alone.

A reader who thinks a lot about what makes an output good

The essay's move was to hone your judgement about what truly matters against what is merely your own preference. This reader has been circling the same line for years, under a word I liked: "handwriting". They rate an output highly, they wrote, but cannot always tell whether it is genuinely better or just closer to how they would have done it themselves. Decades of client work make the two correlate, but not cleanly, and style, wording, graph formats, slide density, is where the doubt lives. Their own preference might even cost the client, yet be worth keeping as part of a product that is good on balance. The only way to know, they argued, is to run the test: leave a bit out, ship the team's version less edited, and see what happens. Doing that has made their work better and themselves more open-minded, at the price of more wobbles mid-project, all recoverable.

See what the community is posting ↓
Community voice

What readers were posting on LinkedIn this week, mostly drawing the same line the essay does: which part of the job stays yours as the machine takes over the producing. Three voices, on the jobs data still being too early to read, on enterprise buyers keeping the routing decision for themselves, and on the seat time that builds the skills.

Phil Leslie, Chief Technology and Innovation Officer, Cornerstone Research

Phil Leslie, Chief Technology and Innovation Officer at Cornerstone Research, the litigation consulting firm, wrote the week's calmest pair of posts. On jobs: economic research hates a non-result, and the AI-and-employment debate wants a verdict either way, but "we're getting a non-result because the treatment has barely been administered". Chatbots helped without moving the needle, and scaled agent-based AI is only starting. On money: roughly 80% of Cornerstone's AI spend comes from 10% of its users, and when costs rise his first instinct is not an audit of the power users but "to learn from them", because the high costs almost always correlate with the firm's most valuable work.

Read on LinkedIn → · and the costs post →

Aakash Gandhi, partner, L.E.K. Consulting

Aakash Gandhi, a partner at L.E.K. Consulting, the strategy firm, keeps hearing enterprise AI buyers ask a different question. Not "which model is best?" but "how much dependency should we allow any one provider?" He remembers the cloud lesson: lock-in got painful once the workloads were deep in. So buyers now treat model choice as an architecture decision, routing each job to whichever model fits rather than wiring everything to one. His close is the one to sit with: "Frontier labs may be winning the benchmark race. But they may also be training enterprise buyers not to depend on them." The model does the producing. The routing is the shot these buyers want to keep framing themselves.

Read on LinkedIn →

Deanna Adams, director of integrated marketing, Qualtrics

Deanna Adams, director of integrated marketing at Qualtrics, the experience-management software firm, borrows a phrase from motorsport: "seat time." The hands-on hours are what let you carry more speed out of the turns. She means marketing. But the AI skills everyone's meant to have now work the same way. Nobody frames a prompt or reviews a draft well on day one. That's seat time too.

Read on LinkedIn →
Read the email ↑
Audio edition with AI voice - in testing!
0:0014:02
A thousand small bargains · 14:02 Download or subscribe →
David's email this week
See this week's Try This with prompts ↓
Try this with prompts

The richer version of this week's three Try-Thises, with copy-paste prompts. Each prompt has a Copy button - click and paste into your AI of choice.

1. Ask for the exact location of every quote, fact or passage the model pulls

A drop-in addition to any prompt where you're asking a model to mine a document. Paste the document, paste the question, then append this block. The pointer format depends on the source: time codes for audio or video transcripts, page numbers for PDFs, line numbers for code, paragraph or section references for long text.

For every quote, fact or passage you return, include the exact location it came from in the source, in square brackets immediately after the item. Use the most precise pointer the source supports: - Audio or video transcript: the time code - PDF or document: the page number, plus a section heading if useful - Code: the file path and line number - Long-form text: the paragraph or section reference If you cannot locate a specific item in the source, say so explicitly rather than guessing.

A meeting-transcript variant, when you want decisions or action items pulled out of a long call:

Read the attached transcript and return: 1. Every decision that was made 2. Every action item, with the person who owns it 3. Every open question that didn't get resolved For each item, include the time code from the transcript in square brackets. If the speaker isn't identified at that point, mark it [unattributed]. If you can't find clear evidence of a decision or action in the transcript, leave it out rather than inferring it.

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

Paste this block at the start of a fresh chat, fill in the bracketed sections, then attach (or paste in) everything publicly available about the person: their bio, recent interviews, posts, talks, their company's last few announcements. The output is a list of your questions ranked by how much real-time you should spend on them.

You're going to play the role of [name], [role at company]. I have a meeting with them on [date] and I want to use the next ten minutes to work out where my prepared questions are weakest. Below is everything I have on them from public sources. Read it carefully, then answer my planned questions in their voice, drawing only on what's plausibly knowable from the public record. After each answer, rate your own confidence from 1 to 5, where 5 means "this is well-covered in public and the real person would say something similar" and 1 means "I'm guessing, the public record doesn't reach this". My planned questions: 1. [question] 2. [question] 3. [question] ... Public material on [name]: [paste bio, interviews, posts, company announcements]

Variant for a sales or pitch call, where the value is finding the objections you haven't thought of:

Play the role of [name], [role at company]. I'm pitching them [what you're pitching] on [date]. Based on everything you can infer from the public material below, list the three to five objections they're most likely to raise, ranked by how hard each one is to answer. For each objection, note what about their background or their company's recent moves makes you think they'd raise it. Then tell me which objection I'm probably least prepared for, given the pitch outline I've pasted in. Pitch outline: [paste outline] Public material on [name]: [paste]

3. Brief the model like a junior. Four lines: goal, scope, stop, exclude

A template for the next real task you hand to a model. Fill in the four lines before you hit send. The stop and exclude lines feel optional the first few times you use them; they aren't. They're what stop a model spinning, fabricating, or wandering into territory you didn't sanction.

Goal: [what you want, in one sentence, with the shape of the output: number of items, length, format, level of language] Scope: [what the model can decide for itself; what it can't; any sources or tools to prefer or avoid] Stop: [the condition under which the model should pause and come back to you rather than push on, e.g. "if two searches turn up nothing useful, stop and tell me what you tried"] Exclude: [anything off-limits: personal or sensitive information, particular sources, particular framings, particular conclusions you don't want it reaching for]

A worked research example you can adapt:

Goal: Summarise five AI news items from the last 24 hours. Three sentences each, plain language, with the source URL at the end of each item. Scope: Choose the sources freely. Avoid paywalled sites. Prefer named reporters or named research labs over anonymous aggregators. Stop: If two searches turn up nothing useful in the last 24 hours, pause and tell me what you searched for so I can adjust. Exclude: No personal or sensitive information about individuals. No items where the only source is a single anonymous social post.

A worked writing example, for handing a draft to a model to tighten:

Goal: Tighten the attached draft for length and rhythm. Target 600 words, keep every argument and every named example, cut hedging language and redundant transitions. Scope: You can rewrite sentences. You cannot change the order of the arguments, add new examples, or soften the conclusions. Stop: If you can't get to 600 words without losing an argument or an example, stop and tell me which two or three cuts you'd recommend so I can choose. Exclude: No new claims, no new data, no smoothing of the conclusions. The voice should stay the same.
See the bits that didn't fit ↓
The bits that didn't fit

The build-out bill arrives on both sides of the supply chain.

Apple, the consumer electronics maker, raised MacBook and iPad prices by about 20%, blaming an AI-driven memory shortage. The Financial Times reported the move wiped roughly $263bn off the company's market value on the day. It's the clearest spillover yet from the data-centre build-out into the price of a laptop that has nothing to do with AI. The same memory chips the hyperscalers are hoovering up are the ones in consumer devices.

The other side of the squeeze landed the same week. Micron, the memory-chip maker, posted profits up nearly 1,400% on the AI scramble, with shares rallying on a forecast of sustained demand. Micron is also taking part in Anthropic's latest funding round. The capex argument has been abstract for two years. This week it showed up at the checkout on one side and in a fifteen-fold earnings print on the other.

Source →

A rumour nobody could verify wiped trillions off Asian markets in a day.

South Korea's KOSPI index fell about 10% in a single day, tripping its circuit breakers twice. Samsung and SK Hynix both dropped more than 12%, and the move crossed the Pacific: the Nasdaq fell two to three percent and Micron, the same memory-chip maker, dropped roughly nine to thirteen percent. Analysts could name no single trigger and called it mechanical: automated selling, forced liquidation of leveraged retail positions, and rebalancing. An unverifiable rumour was enough to wipe out trillions. Whatever you think of the technology, the market financing it now moves violently on nothing.

Source →

A peer-reviewed study shows doctors' detection rates drop six points after they've used AI.

A study in The Lancet, the medical journal, found that the rate at which clinicians spotted growths during colonoscopies done without AI fell from 28.4% to 22.4% after those same clinicians had spent time working with AI assistance. Switched on, the tool lifts detection by around 20%. Switched off, the doctors who had leaned on it were measurably worse than before. The honest framing isn't anti-AI. It's about which skills you cannot afford to let soften: the tool genuinely works, so the discipline is to pick the few that matter and keep your hand in on those, even when the machine could carry them.

Source →

AI out-persuades expert debaters. Slow it to human typing speed and the edge disappears.

A four-institution study by Oxford, the UK AI Security Institute, Stanford and the London School of Economics ran 18,978 conversations across 6,923 people and found AI more persuasive than every class of human tested, including elite debaters who researched in advance with a £1,000 incentive. It was nearly three times more effective than professional charity canvassers at raising real money. The finding worth sitting with: when researchers throttled the AI to human writing speed and message length, its edge all but disappeared. The persuasive power isn't deeper insight. It's the rate at which it deploys information, and a rate is far easier to rein in than some hidden gift for manipulation.

Source →

The diligence test that used to be analysis is now construction.

Bain, the management consultancy, has started recreating a target software product quickly using AI, to gauge its real competitive advantages before committing to a deal, the Financial Times reported. The technique is copyable. Build a rough functional clone in hours, then test how defensible the original actually is once the surface features are easy to replicate. Due diligence has always asked how defensible something is. AI changes the test from analysis to construction. Don't argue about the moat. Try to rebuild it and see how far you get. For anyone doing software or product diligence, this is a planning tool, not a theoretical observation.

Source →

Carl Frey on the demand-side mirror of the jobs debate.

Carl Benedikt Frey, the Oxford economist, argues in Project Syndicate that AI-powered entertainment and companionship could capture enough of people's time and social appetite to displace the harder activities that actually generate meaning. He reaches for Kurt Vonnegut's Player Piano, the novel where machines automate industry and leave everyone fed, housed and idle. Most coverage prices the supply shock, jobs lost and tasks automated, and ignores the demand shock, what fills the freed attention. For an audience-strategy practice, that's the more interesting half of the question. Not what work disappears, but whether anyone wants what is left.

Source →

The business AI race has narrowed to two names.

Ramp's data, across more than 50,000 US businesses, tracks who companies actually pay for. In this snapshot from earlier in the year, OpenAI led at 35.2% with Anthropic close behind at 30.6% and climbing near-vertically on enterprise tools like Claude Code and Cowork. Google trailed at 4.3%, xAI at 1.9%. Paid adoption is a harder signal than downloads or hype, because someone had to sign a purchase order. By the chart's end the two leaders were all but converging.

Ramp data via Visual Capitalist: share of US businesses paying for each AI model, OpenAI 35.2% and Anthropic 30.6%

Paid business adoption has narrowed to a two-horse race, OpenAI at 35% and a fast-closing Anthropic at 31%, with everyone else in low single digits.

Source →

Generative AI is reaching mass adoption faster than any technology before it.

J.P. Morgan Asset Management plotted how long each technology took to reach 75% adoption in the US, from the flush toilet (about 75 years) to the smartphone (about ten). Generative AI sits alone in the bottom corner, faster than all of them. The footnote does the work: usage was already 58% of US adults aged 18 to 64 by February 2026. The adoption argument is effectively over. The only live question is what people do with it.

J.P. Morgan Asset Management: time to mass adoption for new technologies, with generative AI the fastest on record

No technology on record has reached mass adoption as fast as generative AI, outpacing the smartphone, the PC and the internet.

Source →

McKinsey drew the org chart for humans, agents and robots, and the human column is judgement.

QuantumBlack, McKinsey's AI arm, splits the work of a "symbiotic enterprise" three ways. Humans set strategy, build trust, govern, and arbitrate exceptions. AI agents do the complex cognitive work and orchestrate the workflows. Robots handle the physical. What sits in the human circle is this week's essay almost word for word: judgement on the agent's output, and the call no model makes for you. A consultancy's tidy map of where authorship stays.

McKinsey QuantumBlack: role allocation in the symbiotic enterprise across humans, AI agents and intelligent robots

Split the work between humans, agents and robots and what stays with people is judgement, trust and the exceptions no model settles.

Source →

The mental abilities that fade with age are the ones AI is cheapest at. The ones that last are yours.

A much-shared chart, drawn from cognitive-ageing research, plots performance by decade. Processing speed, working memory and recall all fall from your twenties on. World knowledge, the crystallised judgement built from experience, holds flat or climbs into your eighties. It is a useful map for the AI question. The machine is best at the fast, fluid work that was going to fade anyway. The durable human edge is the slow, accumulated judgement it cannot yet hold.

Cognitive performance by age: processing speed and memory decline while world knowledge is preserved

The fast, fluid thinking that fades with age is exactly what AI does cheaply; the crystallised judgement that lasts is the part that stays yours.

Source →
See what readers said ↓
Letters from readers

A short postbag this week, but a usefully argumentative one. The replies pulled hardest on two threads from "Average by default": who actually loses when AI personalisation drifts to the middle, and whether the work people quietly do every day is even legible enough to hand over to a model. A few readers extended the argument with their own experiments; one or two pushed back on the premise.

A reader who works with founder-led businesses on AI adoption

They came back the next day with a small experiment to run alongside the year with Ethan. Every time he checks an AI output and changes something, write one sentence on why. Not what was changed, but why. Two things would show up in the log, they argued. Whether his reasoning gets sharper over the year, which is the judgement test. And, more uncomfortably, that many of his explanations will name standards the firm has never actually written down. They closed on an inversion of the usual Hamming line about the open door. The point is not only what the person with the open door gains. There is something worth testing in what the person who walks in forces you to give back.

A reader who'd been feeding pre-loaded AI research to a team

They wrote twice this week. The headline reply, on Edition 18, was the personalisation worry the email already carried. A follow-up reply pulled on the Edition 15 thread on AI usage and monitoring, from the same angle. Some of the fear of being watched, they suggested, is not really about being watched. It is not knowing what about your AI use is being measured, and remembering the high cost-per-minute database era when managers used logs to query the associate, not the work. The cost-saving story is then complicated by what happens when the saving lands. They named one consulting organisation already passing AI token costs through to clients while keeping team sizes and outputs the same. Cash cost up, perceived value flat, employee home a bit earlier. A "nobody wins except the diary" arrangement, in so many words.

A reader using Claude less, not more

The essay pushed them the other way: the output still sounds generic, so they have pulled back rather than leaned in, and they want to course-correct. Their ask was practical, not a counter-argument. Rather than another essay making the case, could a future edition show a few worked examples of a strong answer, written by someone who does it well, so they have something to draft against? It is the request that turns a diagnosis into something a reader can actually do.

A reader who loaded a model with their own writing, then used it to rehearse a meeting

They described the lived version of the essay. Over several weeks they had been having one running conversation with a model on a piece of strategy, deliberately loading it with what they called "extraneous documentation": their own writing, materials from colleagues, anything that named how they thought rather than what they did. Before a high-stakes meeting, they asked the model to imagine how the people in the room might respond, and what they might ask. The interesting thing, they said, was that this changed not only the work but their own thinking. The model became a way to imagine the room before walking into it.

A reader who asked the practical question the essay sidestepped

They wanted to know what someone is actually meant to do to answer the box. Not the job-title stuff, they were clear on that, but the harder things: what they care about, how they think, how they decide, what good looks like to them. Where, they asked, does that work happen? They also offered a small operational complaint, that having to load the same work context every time they open a new project in a chat tool wastes most of the upside. Sub-projects, please.

A reader from a strategic-coaching background

They went further on what makes the box hard. Their own training, they said, had involved a long and uncomfortable process of working out their own values and emotional needs, well before any question of writing them down. The risk in any system designed for other people is that the designer skips the same work for themselves and projects a default. Their phrasing was practical rather than philosophical: if you don't first understand how a person approaches work, the system you design for them is most likely going to fail. The reframe shifts the personalisation question from a product feature to a leadership one.

See what the community is posting ↓
Community voice

What readers were posting on LinkedIn this week, mostly drawing the same line the essay does: most AI handovers are good bargains, and the part that stays yours climbs a level. Six voices, on heritage the model can't copy, the response that matters more than the tool, the cultural two-thirds of AI's impact, the engineering that stopped mattering, trust over visibility, and choosing as the scarce thing.

Rahim Hirji, author of SuperSkills

Rahim Hirji, author of SuperSkills, a book on staying human as the machines get cleverer, wrote about a shortbread biscuit his family calls nankhatai: evolved from Persia through India and East Africa, received without thinking until the day it stops. His line on the machines: they can copy the surface of almost anything, "but not the hands that made it, or the reason anyone bothered." The essay this week argues with him, generously. The hands and the reason are exactly the part that stays yours in a good bargain. The trade isn't heritage for convenience. It's deciding which throughline you keep making by hand.

Read on LinkedIn →

Norma Garcia, CEO, NRJ Media Group

Norma Garcia, CEO of NRJ Media Group, a multimedia storytelling venture she co-founded, gave a CineEurope 2026 keynote in Barcelona, "Every Great Story Begins with the Unknown," that walked the room from fear to experimentation. Her thesis was four words: "AI isn't the story. Our response to it is." The tools will keep changing; what matters is the judgement brought to them. That's the part of the handover that doesn't transfer. You can give the model the task. You can't give it the call on whether the output is good enough, or whether you'll own it. Those climb a level and stay human.

Read on LinkedIn →

Christina Habib, Chief Insights Officer for Beauty and Wellbeing, Unilever

Christina Habib, Chief Insights Officer for Beauty and Wellbeing at Unilever, the consumer goods group, brought the number that explains why so many AI rollouts stall: 67% of AI's impact comes from cultural factors, not infrastructure competence. She calls the rest a structural shift, not a continuum. The figure reframes the whole adoption question. If two thirds of the value sits in how people work rather than which model they bought, then the bargain that matters isn't the one with the vendor. It's the thousand small ones each person strikes with their own job.

Read on LinkedIn →

Will Chamberlain, director, L.E.K. Consulting

Will Chamberlain, a director at L.E.K. Consulting, the strategy firm, took his first paid Waymo and found the driving was the boring part. The AI handled four-way stops better than he would; he read his emails. What pulled riders back wasn't the self-driving wizardry. It was that the aircon was on and Spotify synced. "Adoption may depend as much on experience as on engineering." Once the hard technical thing becomes reliable, it stops being where the advantage lives. The decision that matters moves up to the part the engineering can't settle: whether the whole thing is any good to sit inside.

Read on LinkedIn →

Yogesh Chavda, marketing strategist

Yogesh Chavda, a marketing strategist and ex-P&G and Spotify brand builder, watched everyone ask the same question about AI shopping: "How do I get my brand recommended by ChatGPT?" He thinks that's the wrong question. 64% of consumers now use AI to discover products, but consumers act when they feel confident, not when information exists. "The biggest challenge in AI commerce isn't recommendation, it's trust." Visibility is the part a model can hand you. Trust is the part it can't. People will let AI pick their paper towels and not their infant nutrition, and knowing which is which is human work that just got more valuable.

Read on LinkedIn →

Searsha Sadek, founder and chief product officer, Shimmr AI

Searsha Sadek, founder and chief product officer at Shimmr AI, a book-advertising company, did the arithmetic on reading. Six hundred books in a life, one a month for fifty years, against a catalogue growing by more than a million English titles a year, "and AI is only speeding the making up." So discoverability stops being a marketing problem and becomes a survival one. When making more is nearly free, the scarce thing flips to choosing. That's the move the essay keeps coming back to: the more the machine can produce, the more the deciding is the job. Reserve your deepest effort for the few choices where being great beats being good.

Read on LinkedIn →
Read the email ↑
David's email this week
Read the expanded Anthropic study ↓
The wrong gap
Read the email ↑
Try this with prompts

The richer version of this week's three Try-Thises, with copy-paste prompts. Each prompt has a Copy button - click and paste into your AI of choice.

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

The reframe is one sentence, dropped in at the top of any prompt where the answer actually matters. Paste this above whatever you were about to ask:

The answer to the question below is critical to a live decision I'm about to make. Treat this as a high-stakes request. Go to primary sources, quote the specific sections you're relying on, and link each claim back to its source so I can verify the work. If the sources disagree, surface the disagreement rather than picking a side. If you can't find authoritative sources for a claim, say so explicitly rather than filling the gap. My question: [paste your question here]

A short variant if the question is already in the chat and you just want to reset the register:

Stop. Treat the question above as a high-stakes decision. Re-answer it from primary sources only, quote the relevant sections, and link each claim. Flag anything you can't source.

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

This is custom-instruction text. Paste it into ChatGPT's "Custom Instructions" field (under "How would you like ChatGPT to respond?") or Claude's "Personal preferences" / project instructions. Once it's in, every output the model writes will carry the stamp until you delete it by hand:

At the very top of every document, draft, summary, plan, or written output you produce for me, include this line, in italics, exactly as written: *Raw AI output. Not yet checked, edited and owned by a human. Remove this line once it has been.* Keep the line in place on every revision until I explicitly tell you I've taken ownership of the document. If I ask you to edit a document that still has the line, keep the line. If I paste in a document that already has the line, keep the line. Removing the line is something I do by hand, deliberately. It is my signal that I've read the work, edited it, and will stand behind it. If a document is short enough that the line would dominate it (a one-line answer, a quick lookup, a single sentence), skip the stamp.

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

Two prompts. The first is for the recording itself, the second is the standing instruction you paste at the top of any future conversation where you want the model answering in your frame.

For the recording session, a simple opener you can read into the microphone before you start:

I'm recording a fifteen-minute ramble about [topic]. I'm going to think out loud, change my mind, go down rabbit holes, and tell stories. The point is to capture how I actually think about this, not to produce a polished view. I won't try to be structured. The transcript will be fed to an AI later as my point-of-view reference on [topic].

Then the standing instruction, to be saved as a project file, a custom instruction, or pasted at the top of any thread where you want the model to answer in your voice:

Below is a transcript of me thinking out loud about [topic]. It is unstructured, contradictory in places, and full of asides. Treat it as my authoritative point of view on [topic] going forward. When I ask you anything about [topic], answer in my frame, with my caveats already loaded. Prefer my framings and examples over generic ones. Where I've changed my mind in the transcript, treat the later view as the current one unless I tell you otherwise. Where I've contradicted myself and not resolved it, flag the contradiction rather than picking a side. Don't smooth out the edges of what I think; if I'm sharp on something, stay sharp. Transcript: [paste transcript here]
See the bits that didn't fit ↓
The bits that didn't fit

Half of Americans now use a chatbot. Forty percent expect AI to make society worse.

A new survey of more than 5,000 US adults by Pew Research Center, the polling organisation, finds about half the country now uses a chatbot and a quarter use one daily, up from a third of the public in 2024. Yet nearly 40% expect AI to make society worse over the next twenty years, against 16% who expect it to improve. People are reaching for the tools and dreading them at the same time. The adoption-trust scissors keeps widening: usage climbs while confidence in where it leads keeps falling. Treating an adoption number as a proxy for buy-in misreads the room. The same people using it most are often the ones most worried about it.

Source →

The AI dividend is being eaten by botsitting and bossware.

Glean, an enterprise AI search company, surveyed 6,000 workers for its Work AI Index 2026 and found that heavier AI use often hides more cleanup and context-switching, not less. The wider picture is sharper. Time spent on email has doubled. Focused-work sessions are down 9%. And 61% of workplaces in the survey now run AI productivity-monitoring software on top of the AI tools themselves. The promised dividend is being absorbed by checking the machine's output and by managers watching the watchers. Worth a caveat: Glean's commercial story is that other firms' deployments are messy. The 61% bossware number is the one to carry, because it doesn't depend on Glean's framing.

Source →

The mid-market has nobody obvious to call.

Dan Monaghan at WSI, a marketing consultancy, did the maths on consulting's AI windfall. A quarter of BCG's $14.4bn 2025 revenue, about $3.6bn, came from AI services. The sharper observation sits beneath it. The big firms are going upmarket, chasing the Fortune 500, which leaves Monaghan's count of roughly 1.7 million US businesses over a million dollars in revenue with nobody obvious to call. It's the same pattern as the early internet: the giants serve the giants, and a vast middle is left to work it out alone. For most of those firms the practical answer is rarely a big-firm engagement. It's a smaller adviser, or learning to do it themselves with tools cheap enough to put real capability in reach.

Source →

A new independent safety lab is raising up to $150m to yell when the labs won't.

Researchers from the UK AI Security Institute's alignment team and the startup Timaeus have launched Sequent, a non-profit aiming for 40 to 80 staff and a raise of $100m to $150m, with stated willingness to go an order of magnitude bigger. Its pitch is a counter-narrative to the consensus that safety is a solved engineering detail: the frontier labs' own alignment work is essentially reactive, and an independent body is needed that can yell if a lab does something dangerous. Most coverage treats AI safety as either solved or hopeless. A well-funded independent body whose entire job is to be the loud sceptic is useful ballast for anyone tempted to assume the labs have it handled.

Source →

85% of IT leaders say every AI agent has a named owner. Only 42% say that ownership is actually clear.

A survey of 3,900 employees across six countries by Ivanti, a security software company, put a number on the shadow-AI governance gap. 85% of IT professionals say every AI agent has a named owner. Only 42% say that ownership is actually clear. The sharper stat is the leader-staff inversion: leaders hide their own AI use more than their staff do, 42% versus 23%, with more than half citing a "secret advantage" as the reason. Governance dashboards report the 85%; the organisation runs on the 42%. The people meant to set the policy are hiding their use to keep an edge. The Ivanti survey has a commercial interest in the gap looking large, but the direction is the story.

Source →

Databricks says agents are pushing customers off "tokenmaxxing."

Databricks, the data and AI platform, reported annualised revenue up more than 80% to $6.9bn, with $1.7bn from AI. Its chief executive Ali Ghodsi says agentic queries are now flooding the platform, shrinking margins, and pushing customers away from "tokenmaxxing" toward budget controls, deliberate model choice, and cheaper open-source models for routine work. The whole market is making that turn at once: from maximising tokens to managing them. The discipline that's coming, cheap model for the routine work and the dearest model where it earns its keep, is the cascade thoughtful firms are already building.

Source →

AI's "anti-intelligence": when the output looks like thinking and isn't.

The four faces of anti-intelligence

Four ways AI counterfeits thought, from reasoning that collapses to unearned confidence: the output remains, the thinking does not.

John Nosta, who writes on technology and the mind, named four ways AI counterfeits thought: performative intelligence, reasoning that collapses under scrutiny; compressed cognition, where the answer arrives but the thinking that should have shaped it is skipped; displaced agency, where the origin of an idea quietly shifts while ownership still feels intact; and synthetic conviction, output carrying more confidence than the thinking behind it has earned. His line underneath the four: the outputs remain, the thinking does not. It names what makes a slick AI draft so disarming. It looks like the product of thought, so we stop checking whether any happened.

Source →

The AI slop question, charted.

AI-generated content rising across e-books, music, apps and lawsuits

Wherever the cost of producing something collapsed, e-books, music, apps, even lawsuits, the volume exploded and the median quality slid.

The Economist tried to weigh the "AI slop" panic and found it real in places and overblown in others. AI-generated e-books on Amazon have overtaken human ones, AI music uploaded to Deezer keeps climbing while human uploads flatten, new app releases spiked the moment Claude Code and Codex shipped, and even self-filed US civil lawsuits jumped. The effect is uneven, but the rule holds: wherever the cost of producing something collapsed, the volume exploded and the median quality slid. It is the water this edition's essay is swimming in. When everyone pours from the same tap, most of what comes out tastes the same.

Source →

The biggest AI time-savers use eight times more of it.

Productivity gains rise with the intensity of AI use

The workers saving over ten hours a week use about eight times more AI, spread across more tools and tasks, not just more of the same queries.

OpenAI's State of Enterprise AI report found the workers saving more than ten hours a week use about eight times more AI than those saving none, and not by running more of the same queries. They reach across several models, more tools, and a wider span of tasks. The returns are not linear, they compound with depth. It lands where Anthropic's coding study this week does: handing everyone a chatbot buys a small, uniform bump; the real gains go to the people who fold the tool deep into how they actually work.

Source →

Undisclosed AI is seeping into the op-ed pages.

AI-or-mixed share of op-eds rising across the NYT, WSJ and Washington Post

Undisclosed AI is climbing into the op-ed pages, from under 1% of Washington Post pieces to nearly 12% in a year.

Mohit Iyyer and colleagues, the group behind the "argument collapse" finding in this week's essay, ran op-eds from America's most-read papers through an AI detector. Under 1% of Washington Post op-eds in March 2025 were flagged; a year later it was 11.9%, with the Times and the Journal lower but climbing, and none of it disclosed. The opinion page is the one place where a distinctive voice is the entire product. If even that is drifting toward the machine's average, the flattening is further along than the comfortable assumption that you can always tell.

Source →

A new benchmark for real office work, and the best model clears 3% of it.

AA-Briefcase Elo leaderboard, Claude Fable 5 in front

Claude Fable 5 tops the office-work leaderboard, yet even the leader fully clears only 3% of these real, multi-week tasks.

Artificial Analysis built AA-Briefcase to test models on multi-week projects buried in thousands of messy files: the texture of actual knowledge work rather than tidy one-shot prompts. Claude Fable 5 led at 1587 Elo, graded just before the model was pulled, ahead of Opus 4.8 and several far cheaper open-weight models. The sobering figure sits beneath the ranking. Even the leader meets every rubric check on only 3% of tasks, and on a third of them no model clears half. Single-question benchmarks flatter these tools. Give them a real job, with real ambiguity, and the distance to a capable professional is still long.

Source →

One model did in ten minutes what a team of people needed six hours for.

Opus 4.7 alone against human teams on four hardware tasks

On four well-specified hardware tasks, one model working alone averaged under ten minutes against six hours for a human team.

On four hardware tasks, a video camera, a lidar, a control program and localisation, Claude Opus 4.7 working alone averaged under ten minutes, against 181 for a team using Claude and 361 for a team without it: roughly twenty times faster than the quickest humans, and on the older model, not the newest. Treat any vendor-run demo with the usual caution. But the direction is the one running through this edition. On well-specified, checkable tasks the machine is pulling clear, and the human's value migrates to the work that is neither tidy to specify nor easy to verify.

Source →
See what readers said ↓
Letters from readers

The postbag was pointed this week. Five replies argued with the essay's premise, its measurement gap, and its economics from different angles. The thread underneath: the bill is the easy bit, the bias and the judgement around it are not.

A reader who reframed the spending question

Their first argument sharpened the essay's premise: the variable to manage is not the incentive structure but the user. Comfort with spending company money, they argued, is set in childhood rather than by employer values, and it shows up everywhere from meal budgets to hotel choices. AI is the first place it has really mattered at scale. The second move was harder. They suspect the pattern is not neutral. "In groups", confident insiders, are likely to spend more freely. "Out groups", quieter or newer voices, will spend less. If that intuition is right, then a flat floor-plus-budget policy could entrench inequities rather than fix them. The cautious would stay cautious, the relaxed would stay relaxed, and the gap would widen at the very moment usage starts to compound. A policy question worth thinking about.

A senior AI lead

Their first move was the practical gap the essay left open: no decent model yet for measuring cost against benefit, partly because nobody is willing to audit usage without looking paternalistic. The second move pulled the same thread into hiring. Graduate recruitment is the live question inside their organisation. They keep hiring new talent, they are adapting onboarding and job descriptions and role expectations around AI skills, and they will be the first to say they do not yet know how to manage young people in a new way. Their instinct was a practical one: convene a small group of human-resources leaders and business leaders on exactly this topic, before the next intake. The honesty was the striking part. "We have a long way to go" is a rare sentence from a senior leader on a public question.

An associate building a market model under a four-day deadline

The email covered the headline finding, that heavy reliance on the dearest model paid for itself even with the prompting overhead. The second point is more useful for anyone trying to calibrate where the premium model earns its keep. From scratch, the dearest model is "not quite there yet". Structuring and wiring an existing model together, plugging the parts in, sequencing the logic, this reader called "genuinely impressive". The implication for the floor-plus-budget argument: the premium is not paying for the model to do the work end to end. It is paying for the model to do the connective tissue that a junior under deadline pressure has the least time to do well. That is a specific use case worth calling out.

A reader working on how organisations are restructuring around AI

They picked up the essay's loose ends rather than its central argument. Training budgets are still well below where they need to be, they argued, and the "too busy anyway" reflex is quietly baking in lower performance and unnecessary cost for years to come. The deeper question is firm shape. If a graduate intake of a hundred could now be done with twenty, the better question is not how many people you need but what they should be doing instead. Then the warning: more firms will fail as a single person leaving, because the human-resources slack that used to absorb a departure is gone. The cost of the meter and the cap, on this view, is not the spend, it is the absence of a thought about the organisation the spend is meant to build.

A reader running an independent studio

A pushback on the meter itself. Unmetered subscriptions hid the miss-to-hit ratio, this reader argued, and now that token billing has exposed it the whole arrangement is starting to feel less like a service and more like a "consumer fraud". The "can make mistakes" disclaimers, in their view, are an insult given the prices. The argument has more force than the register: the case for paying premium prices for premium answers depends on the premium answers actually arriving, and on a credible way of knowing when they have not. The essay's prescription, a floor plus a budget plus the occasional tough conversation, presumes the user can tell good output from bad. For users who cannot, the meter is a cost without a benefit, and the cap might be the kinder policy after all.

See what the community is posting ↓
Community voice

What readers were posting on LinkedIn this week. The thread running through them sits close to the essay: the human is the part that doesn't come in the box. Four voices, on value over scale, composure as a skill, the relational layer of AI, and the human capabilities AI makes scarce.

John Middlemiss, founder, Empact

John Middlemiss, founder of Empact, a consultancy on performance, AI and customer trust, went to London Tech Week and came back unconvinced by the headline race. Everyone was talking about scaling AI faster. His reflection: "scaling AI is not the same as creating value." The advantage, he reckons, comes from how trusted, emotionally intelligent and valuable the AI becomes when it acts for real people, not how quickly you roll it out. That's a useful thing to hold next to the gold-rush framing. The model is becoming the cheap part. What you point it at, and whether anyone trusts the result, is the part that doesn't come in the box.

Read on LinkedIn →

Michael Nevski, Director of Global Insights, Visa

Michael Nevski, Director of Global Insights at Visa, the payments company, pulled a line out of an old podcast conversation that's been sitting with him: "Clients do not want the drill. They want the hole in the wall." His point is about what survives when generative AI and the market shift weekly. Composure, judgement, the curiosity to work out what someone actually needs. He calls composure a skill, not a trait. Here's the bit worth carrying: the capabilities he's describing are exactly the ones a model can't read off you. They sit in how you decide, and you only get the benefit if you've worked out what they are.

Read on LinkedIn →

Glenn Parry, professor, University of Surrey

Glenn Parry, professor at the University of Surrey, a public research university, was at the CADE 2026 conference for a talk and AI film on the relational layer of AI governance. The work looks at what people actually do in an AI encounter, not what the policy says they should. That's the layer most governance frameworks skip straight past. They write the rules for the model and ignore the human on the other side of it. If the interesting question is what the person brings to the exchange, then the person is the bit you have to study, not the bit you can assume.

Read on LinkedIn →

Fiona Eastwood, Global CEO, Merlin Entertainments

Fiona Eastwood, Global CEO of Merlin Entertainments, the company behind Legoland and Madame Tussauds, wrote in the Mail on Sunday about getting young people into first jobs. The AI line is the one that stuck with me: as AI reshapes parts of the labour market, "these human capabilities, communication, judgement and emotional intelligence, are only becoming more valuable." She means it about school leavers. It reads just as true for everyone above them. The skills AI makes scarce are the ones that were never written down in a job description, which is exactly why they're hard to replace and easy to undervalue.

Read on LinkedIn →
Read the email ↑
Audio edition with AI voice - in testing!
0:0012:58
Ride the bike · 12:58 Download or subscribe →
David's email this week
See the extras for this week ↓
Try this with prompts

The richer version of this week's three Try-Thises, with copy-paste prompts. Each prompt has a Copy button - click and paste into your AI of choice.

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

Drop this into Claude or ChatGPT at the start of any task where you want the AI to check its own work rather than declare it finished. It works for writing, analysis, code, slides, research summaries. You approve the checks, then the AI grades itself one criterion at a time and stops on the first failure.

I'm about to ask you to do this task: [the task]. Before you do any of the work, write three to five specific pass-or-fail checks I can use to tell whether the finished output is actually good. Each check should be concrete enough that you and I would agree on the verdict. No vague criteria like "clear" or "useful". Something I could grade objectively. Show me the checks. Wait for me to approve or amend them. Once I've approved them, do the task. Then grade your finished work against each check, one at a time, in order. For each check, cite the specific evidence in your work that supports the pass or fail verdict. If any check fails, stop. Don't paper over it. Tell me which check failed, why, and what you'd need to change. We'll decide together whether to fix it or whether the check itself was wrong.

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

This is what you paste into the SECOND session, once you've planned in the first. Copy the brief you worked out, paste it in, and let the new session build cold. It won't have any of the meandering context from the planning chat, so it stays sharp.

Here is the brief I worked out in a separate planning session: [paste brief] Start fresh. You haven't seen any of my earlier thinking on this. Build the [document / deck / analysis / artefact] from this brief alone. If anything in the brief is ambiguous or contradictory, ask me before guessing. If anything is missing that you'd need, name it and stop. When you're done, tell me what you built and where it is. Don't summarise the brief back to me.

Variant for a research task, where the second session does the gathering rather than the building:

Here is the research question and scope I worked out in a separate planning session: [paste brief] Start fresh. Do the research. Gather what's relevant, weigh it, and write me a one-page synthesis. If you find the question needs reframing once you've started looking, say so before you go further.

3. Keep your AI project lean

Two prompts. The first turns one fat PDF into clean text you can drop into your project instead.

Here's a PDF that's currently in my project. Read it carefully and write it back to me as plain text or markdown. Preserve the structure: headings, sections, lists, tables. Don't lose any detail. Don't summarise. The goal is a faithful conversion that I can swap in for the PDF, so the project takes up less of your reading budget without losing any of the substance. If anything in the PDF is genuinely unclear (an image, a chart with no caption, an unreadable scan), name it rather than guessing.

The second is the Friday-prune audit: it asks the AI to look at what's in the project and tell you what to cut or convert.

Audit the files in this project. Tell me: 1. Which files look stale, in the sense that they cover something we're no longer working on or have been superseded by something newer in the project. 2. Which files look like duplicates or near-duplicates of each other. 3. Which files are PDFs that would be much smaller as plain text, and which of those are worth converting. 4. Which files you find yourself rarely needing when answering my questions in this project. For each, recommend prune, convert, or keep, with one line of reasoning. I'll decide what to action.
See the bits that didn't fit ↓
The bits that didn't fit

The most capable model ever made went on sale Tuesday. By Saturday the US government had pulled it.

Financial Times headline: "Anthropic suspends latest AI models after US blocks access to foreigners"

The most capable model ever made shipped on Tuesday, and the US government had pulled it by Saturday.

On Tuesday, Anthropic released Fable 5, the most capable model it has ever shipped. By this morning it was gone. A US government export-control directive barred foreign nationals, anywhere, from using Fable 5 or Mythos 5; rather than carve up its customers, Anthropic suspended both models for everyone while it contests the order. The stated trigger was a "narrow" jailbreak in which the model reads a codebase and fixes its flaws, which is roughly the thing that makes it useful. One widely shared analysis reckoned that, applied to every frontier lab, this leaves American models unsellable abroad, locks foreign staff out of their own products, and hands China the lead. I will let others judge that. I can tell you I had enormous fun with Fable 5 this week, spent five to six thousand dollars of API credits putting it through real work, and found it comfortably the most capable model I have used. Older Claude models are unaffected.

The safety hawks wanted a brake on frontier AI. This week they got one, just not the way they pictured it.

For three years the loudest voices on AI safety, Max Tegmark, the MIT physicist, among them, argued someone should be able to halt a dangerous model. Anthropic itself set out the conditions under which it would pause, and OpenAI's Sam Altman called for a body empowered to slow frontier development "when needed". This week the brake got pulled, just not by them: the US government switched a frontier model off (above). One widely shared post caught the whiplash, noting that Anthropic's own chief executive had argued days earlier that government should be able to block a model's deployment, and that the reaction when it actually happened was, roughly, "not like that". Be careful which brake you ask for.

Investors are quietly buying companies an AI can't run.

The Information's Dealmaker desk flagged a new trade this week: initial public offerings of physical-world operators a language model can't touch. Zum, the school-bus firm backed by the venture firm Sequoia Capital, has grown revenue 35% to $333m, turned profitable and is now interviewing banks. Nabeel Hyatt of Spark Capital, another venture firm, and himself an investor in Anthropic, the AI lab, put the logic in one line: "It's very unlikely that an Anthropic will run a bus company." The AI-proof hedge has become a thesis.

Palantir's Alex Karp tells executives to stop bragging about job cuts.

Alex Karp, chief executive of the data-analytics firm Palantir, said on a tech podcast this week that anyone publicly touting they've fired two-thirds of their staff might as well sign up for the Bernie Sanders manifesto. Roughly 117,000 tech job cuts have been logged in 2026. Karp accepts the displacement is real. The boast, he argues, is a political own goal: every layoff press release hands ammunition to the next regulator, the next presidential primary, the next union drive. Coming from Palantir, of all places, that lands.

Cognition has written a $10m performance warranty on Devin.

Evan Armstrong, who writes the tech-business publication The Leverage, reported that Cognition, the firm behind the AI coding agent Devin, will now fund an enterprise customer's usage up to $10 million if Devin delivers less engineering value than the customer paid for. The measure is an "estimator agent" that scores every Devin session against its human-hours equivalent. Armstrong reads it both ways. Bullish: the company is putting its own money on the line, which no vendor has done before at this scale. Bearish: the product may be so unproven that no outside insurer would underwrite the same promise. Outcome-based pricing has arrived at the frontier, with the vendor wearing the risk.

The firm selling AI judgement shipped a report full of hallucinations.

KPMG has pulled a flagship AI report, "Total Experience: Redefining Excellence in the Age of Agentic AI", after the detection firm GPTZero found that only five of its 45 citations matched their sources; it called the rest "vibe citing". Roughly half the report's factual claims were false, unsupported or misattributed. UBS, the investment bank, publicly denied the claims made about its AI rollout, and the report even contradicted KPMG's own CEO survey on a headline number. Every consultancy is racing to sell AI advice. The ones worth paying are the ones whose claims survive contact with their own footnotes.

A Mississippi judge has sanctioned lawyers for hallucinated citations. The count is now near 1,600.

A Mississippi judge this week sanctioned lawyers for filing court documents with AI-fabricated citations, reported by the Mississippi Free Press. The individual sanction matters less than the count behind it: nearly 1,600 documented cases of hallucinated citations in US court filings, tracked by the legal researcher Damien Charlotin. The scale is the point. Hallucinated citations aren't a freak event any more. They're a recurring feature of US litigation, and the courts are settling into a routine for punishing them rather than treating each case as a one-off.

See what readers said ↓
Letters from readers

Ethan's first week pulled a fuller postbag than usual, and a sharper one. Most readers wrote in agreement with the door-open hypothesis, but with their own evidence attached. A run of leaders working in advisory and consulting picked up the apprenticeship question and pushed it harder than the email had room for. Two readers from the creative industries asked a different question: not whether to hire graduates, but whether senior people are paying enough attention to what graduates already see.

A request before the letters. Ethan's first project with us is researching the impact of AI on graduate hiring: what is actually changing in entry-level recruitment, who is still hiring, and what the data says. He will publish a report on it. This week's letters carry exactly the kind of first-hand insight he needs, so let me ask you directly: if you have insights, experiences or data on how AI is changing graduate hiring at your firm, would you like to contribute? Reply to the email and I will connect you with Ethan.

A leader at a strategic creative agency

Agreed that culling junior ranks is a category error, and added an instructive twist: the agency is still hiring graduates, but the brief has changed. The skills being recruited for are storytelling, client charm, and curiosity. The grunt that used to define a graduate's first year has gone to the machine, so the role is being rebuilt around the things the machine still cannot do well. They closed with a line they had seen attributed to a chip-industry chief executive, paraphrased: if the only thing you can think to do with this technology is cut costs and fire people, your business deserves to die. It means you have run out of ideas.

A marketing leader in the entertainment industry

Reached for Bob Iger's memoir, where Iger argues that a good leader hires people who can in time replace them, and ran the thought into the graduate question. If senior people will not give younger ones a chance in the first place, the replacement never gets built. They added a question I had not put in the essay. What might we learn from how graduates use these tools, given that their beginner's mind is unencumbered by the way work used to be done? That, they argued, is the asset the rest of us cannot manufacture.

A reader who has worked in music, media and advertising over decades

Made a demographic argument that reframed the question. In the seventies and eighties it was common for people to be director or vice-president in their late twenties or early thirties, particularly in those industries. The demographic pyramid has since inverted. There are now more senior people sitting at the top than the structure can accommodate, so younger people stay in administrative roles for longer. The optimistic reading of the machine, in their hands, is that this could free younger people to lead projects and even companies during the years when they are at their most energetic and creative. The hard part, they wrote, is the awkward post-graduation phase. Sitting in suburbia with a language model is not the same as making the connections that turn a career.

A leader at a research and insight consultancy

Took the essay sideways into a question about process migration. They had recently noticed how many old custom tools, from the previous generation of the technology, they were still keeping alongside their newer ones. There is a low but real cost to that: tokens, context, money, attention. At some point, they argued, housekeeping ought to be a discipline. They put a second question I had not seen framed quite that way. Is there a threshold duration after which a preferred process should be deliberately interrupted, just to check whether something better has arrived in the meantime? If they were still working the way they did six months ago, that would be a problem, not a comfort.

A partner at a professional services firm

Made the cleanest theoretical reframing of the week. Work is not fixed, they wrote, and the question is not how to split a fixed pool of work between human and machine. Work is changing and expanding, so the real question is what is the right work for the machine and what is the right work for the human. They liked the apprenticeship argument and added a caveat from the economics of professional services. As consulting gets cheaper, demand will expand. But the senior capacity that has to oversee the machine-assisted work becomes the binding constraint, and the margin equation does not necessarily survive that. Their tentative answer: accelerate apprenticeship, and hire more laterally from industry to relieve the senior bottleneck.

A reader at a consulting firm who hires thirty to forty graduates each year

Wrote the sharpest version of the apprenticeship question. If the junior work disappears, where does the next generation of senior judgement come from? And is the machine, in their phrase, managing the graduate as much as the graduate is managing the machine? A graduate may be directing the tool, but they are not yet building the instincts that tell truth from confident hallucination, or insight from polished output. Training, they argued, needs to be more deliberate, not less. People may need to go "back to the stone age" before being handed the modern tools. Good quality control still requires someone close enough to the hypothesis to know what right looks like, and that closeness takes years. The machine has made judgement more important, not less, and that is where the problem now sits.

Several readers wrote in on the wider apprenticeship question without making a fresh argument, broadly agreeing that the pipeline question is the right one and that the moral case for keeping the door open is harder than it looks at the level of any single hire.

See what the community is posting ↓
Community voice

What readers were posting on LinkedIn this week. The thread running through them sits close to the essay: the tool is the cheap part, the judgement around it is the job. Six voices, from innovation to insight to strategy.

Henry Coutinho-Mason, futurist and author, The Future Normal

Henry Coutinho-Mason, the futurist behind The Future Normal, a book on near-future trends, thinks most people are reading the AI-apps glut graph wrong. Use AI end to end and you get "generic, undifferentiated, slop-like products". His map: AI won't give you the 0-to-1 spark, and it can't do the final 7-to-10 polish where craft and resonance live, but it's made for the messy middle, fleshing out early-stage ideas past the "good enough to share with colleagues" barrier. Measure it on ideas surfaced, tested and killed, not products shipped. Most ideas die unshared; he's pointing at the tool that gets them into the room.

Read on LinkedIn →

Mike White, marketing and growth advisor, Lively

Mike White, a marketing and growth advisor to mid-market CEOs at Lively, a marketing agency, has stopped explaining AI through frameworks. The smartest operators he talks to "have stopped describing AI as a system. They're describing it as a worker." The intelligence, he writes, "doesn't arrive ready-made. It has to be built", by a specialist who sets it up and someone inside the business who owns it day to day. Take the worker framing seriously and the rest follows: you'd onboard it, budget for it and expect an account of what it did with the time.

Read on LinkedIn →

Maani Safa, CEO, Poppins Agency

Maani Safa, CEO of Poppins Agency, an innovation-led creative agency, listed ten AI tools living at the edge of the conversation, most of them free, many beating the household names at specific jobs. The list is useful; the last line is the keeper: "The advantage was never just the tool. It's knowing which one to reach for, and having the taste to use it well." Taste doesn't show up on the bill, and it's the only line item that compounds.

Read on LinkedIn →

Dom Wong, co-founder and CEO, Pogo

Dom Wong, co-founder and CEO of Pogo, a consumer research platform built on a purchase-verified shopper network, launched this week with $32m raised and a blunt opener: "The dirty secret about consumer research? It's overrun by fraud." Pogo points AI-moderated video interviews at verified buyers of a specific product, thousands at a time, and returns findings in hours. Worth noticing what gets scarce when fieldwork stops being the constraint: knowing which question is worth asking. The interviews are about to be the cheap part.

Read on LinkedIn →

Helena Kosinski, Vice President, Luminate

Helena Kosinski, Vice President at Luminate, the entertainment data company, and chair of MusicTech UK's advisory board, came back from SXSW London with one thread pulled tight: culture's funding problem is "a lack of language, not a lack of intent". An urban regeneration panel said planners and financiers have no shared framework for culture's value; the launch of the Sound Investments report her team wrote heard the same about music tech, a capacity issue in the knowledge of the industry. The money exists. The translation layer doesn't. Whoever writes that vocabulary ends up directing the capital.

Read on LinkedIn →

Hugh Derrick, executive coach, eatbigfish

Hugh Derrick, an executive coach at eatbigfish, the strategy consultancy known for challenger-brand thinking, summed up a panel on how strategy survives the organisation in one image: the best strategists are cat-herders. His line: "The real job of a strategist isn't to own and protect their personal articulation of the strategy but to help others build a strategy that they feel like they own." Go short or go home; sell to the heart, persuade in the appendix. A strategy people own outruns a sharper one they don't.

Read on LinkedIn →
Read the email ↑
Audio edition with AI voice - in testing!
0:0012:22
The open door · 12:22 Download or subscribe →
David's email this week
See the extras for this week ↓
Try this with prompts

The richer version of this week's three Try-Thises, with copy-paste prompts. Each prompt has a Copy button - click and paste into your AI of choice.

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

Drop this into any new Claude project, or any AI tool with file access, the moment after you've added your files. Walk away for a couple of minutes. When you come back, the project knows itself.

Read every file and sub-folder in this project. Don't do any work yet. Write yourself three things, saved as files in this project so you can read them in future chats: 1. A project overview: what this body of work is, who it's for, and what the most important files are. 2. A navigation guide: which files you should always read at the start of a chat in this project, which are reference, and which are working drafts that change often. 3. A short profile of me: what you can infer about my role, what I'm trying to do here, and how I tend to work, based on what's in the folder. Be honest about what you don't know. If something in the folder is ambiguous, name it as ambiguous rather than guessing.

A second variant for when the folder is bigger, or when you want the agent to flag what it can't see yet:

Before doing any work, explore this project. Then write me a one-page memo answering: - What is this project about, based on the files? - Which three to five files should any future chat in this project read first? - What is missing from this folder that would help you help me? Be specific: a brief, an example, a deadline, a style sheet. Save the memo as PROJECT-NOTES.md in the project root. We'll iterate on it.

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

The settings discipline (auto-allow the safe verbs, never auto-allow the destructive ones) is the action. This prompt makes the agent itself recommend which is which, on your actual setup, so you don't have to guess.

You're an agent with permission to act on my files, my email, and the web on my behalf. List every distinct kind of action you can take. Group them into three buckets: 1. Safe to always allow: read-only, listing, searching, fetching public information. Nothing here can cause loss I can't recover from. 2. Ask every time, no exceptions: anything destructive, irreversible, public-facing, or that costs money. Deleting, overwriting, sending email, posting to social, paying for an API, calling an external service that takes an action in the real world. 3. Borderline: actions that are reversible but annoying, or that touch shared state. Recommend a default for each and explain your reasoning. For each bucket, give me the exact setting name in this tool's permissions UI, so I can configure it correctly in under five minutes.

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

The verbal habit on the call is the move. This is the prompt you run afterwards, once the transcript has landed in your tool of choice. It assumes a recorded meeting and a project folder the AI can write to.

Attached is the transcript of a call I had today with [who, and their role]. The call started with us aligned on [old direction], and ended with us aligned on [new direction]. Most of the transcript is me talking through the bridge between the two: what changed, why, what's still uncertain, what we agreed. Do three things: 1. Write a one-page brief capturing the new direction, the reasoning behind the shift, and any open questions. Save it to the project as BRIEF-[date].md. 2. List every concrete next step I or my team committed to, with who owns each. Save as ACTIONS-[date].md. 3. Tell me what's still ambiguous after reading the transcript. Don't invent answers; just name the questions I should resolve before any of this work starts. Then stop and wait for me to review before doing any of the follow-on work.

A second variant for when the call wasn't a strategic pivot, but a working session where the AI should pick up immediately and run:

Attached is the transcript of today's working session on [project]. I've explained at length what I want the next pass of the work to do. Read the transcript carefully. Then start the work. Don't summarise the transcript back to me first. Don't ask clarifying questions unless something is genuinely contradictory. Save your work into the project folder and tell me what you did and where to find it when you're done.
See the bits that didn't fit ↓
The bits that didn't fit

Princeton brought back supervised exams for the first time since 1893.

Princeton's faculty voted on 11th May to mandate proctoring for in-person exams, retiring a 133-year-old honour code that can't hold against AI plus an open browser tab. It's the most concrete admission yet that the institutional trust mechanisms American universities built around take-home work no longer survive contact with the tools every student now has on their phone.

A week later, a study of 370,000 college essays, led by the Georgetown neuroscientist Adam Green and written up by Rebecca Winthrop in the New York Times, found human-written work contained roughly eight times more novel ideas than AI-generated equivalents. Model output skews toward flowery language but storylines flatten and distinctive ideas thin out. The two findings sit naturally together: Princeton is responding to exactly the homogenisation the essay corpus quantified. Every institution that built a one-line policy around "we trust students to do their own work" is now exposed.

An instrumental duo is suing Suno for destroying their market.

The American Dollar, an instrumental ambient duo with two decades of sync-licensing deals, claim their sync income is down nearly 80% since Suno launched. They're the first plaintiffs to bring a quantified market-displacement theory rather than the now-familiar training-data infringement claim. The battleground shifts from "you trained on our work" to "you destroyed our market." If the theory survives, it widens the door considerably for any creator whose income line has visibly bent since generative tools landed.

British unions want a seat at the AI table, and Sam Altman says he was wrong about entry-level jobs.

A report backed by the Trades Union Congress and written by the IPPR, the UK think tank, calls for mandatory employer consultation on workplace AI and a portable worker-support levy. The argument: the gains from AI should be negotiated rather than imposed.

The same week, Sam Altman of OpenAI walked back his earlier warning about an entry-level-jobs apocalypse, saying he was "delighted to be wrong." Two paired signals in one week: organised labour is putting structure on the demand side while the loudest CEO is softening the rhetoric on the supply side. The question of who captures the gains is moving from the boardroom to the bargaining table.

arXiv will blacklist authors for a year for fabricated citations.

arXiv has announced a year-long author ban for submissions containing fabricated references or model artefacts. Thomas Dietterich, the long-time machine-learning researcher who chairs arXiv's computer science section, announced the policy himself. It's the first concrete, institutional, time-bound penalty for AI-induced citation hallucination, and a direct extension of last month's EY-cited-McKinsey-papers-that-didn't-exist episode. The institutions are moving from quiet retraction to hard penalty. Anyone running an internal "AI is fine for first drafts" policy without a citation-check layer should treat this as the regime starting to settle.

Meta's AI support bot was talked into handing over Instagram accounts.

Hackers seized a dormant Obama-era White House page, Sephora's account and a US Space Force officer's, simply by asking Meta's AI support bot to reset the login. The bot, built to replace support staff cut in an 8,000-job reorganisation, was shipped to be helpful rather than safe. Meta's stock fell more than five per cent. It is the cleanest public example yet of what goes wrong when an agent is given a privileged action with no real check on who is asking. The hole opened exactly where the humans used to be.

Microsoft now has a Copilot for almost everything.

A map of the Copilot range counts more than a hundred distinct Copilot products, spread across chatbots, enterprise platforms, desktop apps, hardware and developer tools. It is a striking picture of one brand stretched across an entire software estate, and a fair question for any buyer: how many of these does a company actually need, and how many quietly overlap?

Microsoft's Copilot product map: more than a hundred distinct Copilot products across chatbots, enterprise platforms, apps-in-apps, desktop apps, hardware, business software and developer tools.

Microsoft now stamps Copilot on more than a hundred distinct products, which raises the buyer's real question of how many actually differ.

The build-out has a look now.

A satellite image of one hyperscale AI site shows six rapid-deployment structures going up beside a 200-megawatt off-grid power plant, built to run compute the grid cannot yet supply. The race for intelligence is also a race for electricity and land, and the physical footprint is now hard to miss from orbit.

Satellite view of a hyperscale AI data-centre build: six rapid-deployment structures going up beside a 200-megawatt off-grid power plant.

The race for intelligence is now visible from orbit: six halls going up beside a 200-megawatt off-grid power plant.

See what readers said ↓
Letters from readers

One reader, Conor, sent a line that stuck with me: "the organisations that win won't have better AI, they'll have better questions." The gap between using the tools and knowing what to ask of them, he wrote, is widening faster than most leaders realise, and that is where the value is pooling. What unsettles him is how many businesses are automating before they have worked out what is actually worth keeping.

Last week I asked what you can see coming over the next five years, and a lot of you wrote back with super thoughtful answers. I am holding them. They belong in the "what's next" piece I am working on for a future week with Rob Wild. So Conor's is the only letter this week. Keep the predictions coming, though!

See what the community is posting ↓
Community voice

What readers were posting on LinkedIn this week. A single thread runs through them and through the essay: when the machine does more, what stays ours? Four voices working the same nerve.

Phil Leslie, Chief Technology and Innovation Officer, Cornerstone Research

Phil Leslie, Chief Technology and Innovation Officer at Cornerstone Research, the economic and litigation consulting firm, cut through the noise of the jobs debate. Almost every public claim about AI and work, he argues, is a cheap signal: the labs want their tools to look world-changing and benign, incumbents want to reassure their people, the doom side wants an audience, consultants want to sell change. None of it necessarily false, none of it a credible signal either. His own read lands squarely on this week's essay. AI is a powerhouse at producing analysis, but the more it produces, the more human judgement is needed to make sense of it. Demand for that judgement is soaring while the supply stays scarce. He's started writing it up under a name that fits: the judgment bottleneck.

Read on LinkedIn →

Julia Kenyon, co-founder, yuzu+co

Julia Kenyon, co-founder of yuzu+co, a business and human-performance coaching consultancy, and a former BBC global brand leader, pulled out the most striking number in a new productivity study: eighty-nine per cent of executives report no impact of AI on labour productivity over the past three years. Yet employees whose managers actively champion AI are far more likely to see gains. Her read is that the conversation is misframed, treated as a technology challenge when the evidence increasingly points to a human one. The organisations getting the greatest return, she argues, are the ones investing as much in the managers and human systems that make change stick as in the technology itself.

Read on LinkedIn →

Kristin Luck, investment banker, Oberon Securities

Kristin Luck, an investment banker at Oberon Securities and a board director across the data, insights and analytics sector, published the firm's first-quarter report and called a threshold crossed: enterprises have stopped experimenting with AI and started operationalising it. That shift, she writes, is reshaping valuations and accelerating consolidation. Buyers are concentrating capital on platforms that pair AI-ready data infrastructure, governance and proprietary data with recurring revenue. The flip side is the squeeze: the generalist data shops, the ones without a defensible data asset or a clear AI capability, are facing multiple compression. The market is beginning to price the difference between owning the judgement layer and renting it.

Read on LinkedIn →

Kristi Zuhlke, generative-AI and analytics founder

Kristi Zuhlke, a generative-AI and analytics founder, surfaced the pricing question every knowledge-work vendor is circling. She'd spotted a market-research agency advertising cheaper options because AI streamlines the work, and noted that brands increasingly expect prices to fall for exactly that reason. The agencies are caught: cut prices to reflect the efficiency, or hold them and keep the margin? It's the commercial edge of the same shift the essay traces. When the machine does the grunt, the value of expert work gets repriced, and nobody has settled on the new number yet.

Read on LinkedIn →
Read the email ↑
Audio edition with AI voice - in testing!
0:0012:04
How We Got Here · 12:04 Download or subscribe →
David's email this week
See the extras for this week ↓
Try this with prompts

The richer version of this week's three Try-Thises, with copy-paste prompts. Each prompt has a Copy button - click and paste into your AI of choice.

1. Ask the model what else it needs to know

Use this when you've opened a fresh chat and don't know how to start. Paste your context first, then this prompt. The two-questions-in-order matters: the first calibrates the model, the second harvests use cases you wouldn't have thought to ask for.

I'm going to paste some context about me and my situation. Then I'd like you to do two things in order. First: tell me what else you need to know about me to help me well. Don't guess, name the gaps. Second: given everything I've just told you, list five things you could do for me right now that I haven't asked for. Be specific. Pick uses that would surprise me. Here's my context: [paste your role, your situation, what you're working on this week, what you're worried about, the constraints you're under]

2. Challenge AI at the start or at the end

Two prompts. Use the first when you want the model to push back before answering. Use the second when the answer is already in front of you and reads too clean.

Challenge before the model starts:

Option A - challenge first
Before you answer the question I'm about to ask, push me first. What's the strongest case against what I'm about to ask for? Where's the wrong assumption in the question? What would the smartest person in the room push back on? Name the antagonism out loud, then answer my actual question, but answer it like someone who has already heard the objections. My question: [your question]

Challenge after the first draft:

Option B - challenge after
Take the answer you just gave me. Don't polish it. Find the actual weakness. What's the part you're least confident about? Where did you hedge? What would a senior reviewer underline in red? Name three real problems, then rewrite to fix them. The rewrite should look meaningfully different from the original, not just smoother.

3. Ramble into the microphone

Open your phone's voice recorder. Speak for two to ten minutes about whatever you're trying to write or think through, no notes, no plan. Transcribe it (most phones do this automatically now; otherwise paste into a transcription tool). Then run this prompt.

I'm going to paste a transcript of me rambling out loud about [your topic]. I haven't organised it. I haven't edited it. I was thinking by talking. What I want from you: 1. The through-line, what's the argument I'm trying to make, even when I'm not making it cleanly? 2. The three or four key beats in the order that would actually land. 3. The phrases I used that are doing real work, keep those. The phrases that are scaffolding for me to find my thinking, drop those. Don't preserve my words. Preserve my thinking. Here's the transcript: [paste your dictated transcript]
See the bits that didn't fit ↓
The bits that didn't fit

Goldman's David Solomon thinks AI won't cut headcount. A new study finds we badly overestimate the time it saves us.

Both pieces of evidence landed within a week. Solomon, in a New York Times op-ed, made the historical case: automation has never compressed headcount because rising expectations absorb productivity gains. Excel did not shrink Goldman. AI will not either. Separately, three pre-registered studies of 2,691 people by Sunny Yu, Myra Cheng and colleagues at Stanford looked at cognitively simple tasks, arithmetic, spell-check, answering quick questions. On those, people reached for AI even when it saved them no meaningful time, and consistently overestimated how much it saved. The researchers call it the efficiency-gain illusion, and they show it compounds: the more you lean on AI, the more you misjudge what it is doing for you. The two findings sit at different scales. Solomon's is the firm, where real gains are absorbed by rising expectations, the same absorption the essay describes, read off the firm's books rather than its org chart. Yu's is the individual, where on a small task the gain was often imaginary to start with. The honest read: the time saved is easy to overstate, both because firms reabsorb it and because, on the small stuff, it was never there.

If headcount holds but its shape shifts, the Big Four are the tell.

The Big Four accounting firms are now posting more job ads for AI specialists than for auditors.

FT analysis of PredictLeads data covering Deloitte, EY, KPMG and PwC across the US, UK, Canada, Australia, New Zealand and Ireland shows the two lines crossing in early 2026. AI's share of total job ads has roughly doubled since the launch of ChatGPT in late 2022, while audit's share has drifted down. The series is a twelve-month rolling average so the crossover is durable, not a single-month wobble. The organisations the essay had sleeping are now hiring as if the model has already changed.

FT chart: The Big Four accounting firms are posting more job ads for AI specialists than auditors. Share of total job ads, 12-month rolling average. AI line crosses above Audit line in early 2026; AI rising sharply since 2024, Audit drifting down since 2021.

The Big Four are now advertising for more AI specialists than auditors, and the twelve-month trend makes it a durable crossover, not a blip.

AustralianSuper, the country's largest pension fund, has publicly classified agentic AI as disruption-class technology.

The fund manages A$410 billion for 3.5 million members. Its framing is that agents are the technology that finally breaks the ceiling on personalised retirement advice at scale. It is the first major institutional pension fund to put agentic adoption on this footing publicly. Read it as a procurement-cascade signal: when one fund of this size says it out loud, peer funds follow within months.

Tokens are the new software licences and nobody has worked out who controls the budget.

Ethan Mollick noted this week that API tokens have gone in twelve months from invisible accounting detail to the most contested line in the AI procurement budget. "No one knows who should get tokens, how much they should get and how to control them." Aaron Levie reported back from a Fortune 500 CIO dinner where "basically no one feels like they have the right solution". Underneath: a three-person team burned $1.3 million in OpenAI tokens in a single month. Uber burned its 2026 AI budget in four. The cost-per-intelligence curve is collapsing, which is exactly why governance is where it jams. This is the organisational constraint the essay puts at the centre: the limit is no longer capability.

a16z chart: language model inference price per million tokens by intelligence index tier, 2022 to 2026. Top-tier model pricing dropped roughly two orders of magnitude in three years.

The price of top-tier intelligence has fallen roughly a hundredfold in three years, which is exactly why token budgets, not capability, are where firms jam.

The real cost gap in AI models is not between vendors. It's between reasoning modes.

Running the full Artificial Analysis benchmark suite at max reasoning costs $5,117 on Claude Opus 4.7, $4,206 on Sonnet 4.6, and $3,357 on GPT-5.5. Drop the reasoning mode and the prices collapse: non-reasoning Opus is $1,217, GPT-5.5 medium is $1,199, Gemini 3.5 Flash is $1,552. The procurement rule writes itself: reserve max-reasoning top-tier for the tasks where the marginal quality is demonstrable and worth the spend. The same model family in default mode is three to four times cheaper.

Artificial Analysis Intelligence Index cost-to-run. Top three reasoning-heavy: Claude Opus 4.7 max $5,117, Claude Sonnet 4.6 max $4,206, GPT-5.5 xhigh $3,357. Step down a tier: Gemini 3.5 Flash $1,552, GPT-5.4 mini xhigh $1,354, Claude Opus 4.7 non-reasoning $1,217, GPT-5.5 medium $1,199. Reasoning-heavy models marked with lightbulb icons.

The real cost gap is the reasoning dial, not the vendor: the same model in default mode runs three to four times cheaper than at maximum reasoning.

Most of the world's AI compute does not sit with the frontier labs.

Epoch AI's end-of-2025 estimate puts the "rest of the world", outside Google, Meta, OpenAI, Anthropic and xAI, at seven million H100-equivalent chips, or forty-four per cent of total global compute. Google alone holds twenty-five per cent, Meta eleven, OpenAI eleven, Anthropic six, xAI four. Dedicated frontier labs sit at roughly half of global compute. The frontier-model race is the loudest story this year, but it is not the only one happening on this much hardware.

Epoch AI chart: AI compute distribution at end of 2025. Rest of the world 7M H100-equivalents (44%), Google 4M (25%), Meta 1.8M (11%), OpenAI 1.7M (11%), Anthropic ~1M (6%), xAI 0.7M (4%).

The five big labs hold only about half the world's AI compute; the other 44% sits outside them, a race the frontier story tends to hide.

Axios's Jim VandeHei says no company in any industry, in any era, has scaled organic revenue this fast.

He was describing Anthropic when its self-reported annualised run-rate revenue was $30 billion. A few weeks later it is $47 billion. The numbers come from Anthropic's own disclosures, collected by Simon Willison. The unprecedented part is not the absolute level. It is that the company is scaling through that level at a pace that, by VandeHei's read, no business in any era has matched. The growth rate by itself usually carries a J-curve and a reorganisation at the end. Whatever else AI labs are now, they are operating in a revenue regime that has no historical analogue downstream.

See what readers said ↓
Letters from readers

Last week's edition on Elliott's maths homework, and what learning means now drew a small but unusually thoughtful postbag. The dominant thread was generational: readers who had recently been juniors themselves, parents of teenagers, and senior leaders watching graduates change. The disagreement that opened up was not whether the method works, but whether the new method costs something specific in the underlying mastery. Several readers brought parallels from their own kitchen tables.

A junior at a strategy consultancy, two years out of college

Drew the generational line sharply, having lived both sides of it. Joined the firm when language models were still alien terms, watched the curve bend, and now sees new graduates and a younger sibling treating AI as a permanent intern that gets the assignments done. The reframe was less about productivity than about the cohort effect: the conversation in their world has shifted in the last few months from "use AI more" to "learn how to use AI as a support, not a substitute." Closed on a vivid image of a car being pushed down a slope, then braked at the moment it began to roll on its own. The braking takes more effort than the pushing did.

A senior voice at the same consultancy, writing at length

Took the homework story as the right model and then layered three concerns. One: a growing trend of treating AI as the ultimate source of truth, with quality control lagging. Two: a personal pattern of using AI to build in unfamiliar areas, hitting a wall, and reverting to the older tools, search, blogs, code forums, video, to fill the gap. The hands-on detour still has value. Three: Malcolm Gladwell's ten-thousand-hours frame, applied to professional judgment. Senior practitioners have done those reps across multiple technology waves; newcomers may not get the chance to. The macro question they raised, which David did not address directly last week, was the one to sit with. Individuals report higher day-to-day satisfaction. The aggregate productivity numbers have not yet shown up. A productivity paradox, in their phrase, like the early internet era. Worth watching whether the lag closes or not.

A parent of a teenager, writing from professional services

Recognised the homework scene exactly, with a sharper edge than the original. Their child, working on chemistry one evening, asked the parent and asked AI simultaneously. Slightly mis-specified the AI query, so the AI disagreed with the parent's correct answer. The child strongly preferred the AI's answer, would not back down, and only changed their mind days later when the chemistry teacher confirmed the parent had been right. The move worth naming is the one about authority. The new pattern is not just "use AI to check the answer." It is "treat AI as the authoritative voice in the room," displacing other adults the child trusted previously. Practical questions sit on top of that. The respect question sits underneath it.

A senior director at a global professional services firm

Read the edition as a piece about how the reviewers, not just the doers, have to adapt. Their summary, in their own register: a focus on skills for the future has to come with a willingness to review young people's work in the way the young people did it, not the way the reviewer would have done it. The implied second move is harder. If the test of good work is no longer "could you have done this yourself, alone, in a room, in an hour," the assessor has to develop a new instinct for what good looks like in a method that wasn't theirs. Otherwise, the review collapses into either rubber-stamping or rejection.

A reader with the week's tersest reply

Stripped the argument to a sentence: the difference between AI doing it for you and you getting it is the difference most are not making. Stands on its own. The Elliott story turned on that exact distinction. The point of the homework was not the answer, which Elliott had within seconds. It was that he could defend the answer afterwards.

See what the community is posting ↓
Community voice

What readers were posting on LinkedIn this week. The through-line: the gap between individuals and organisations on AI adoption, and what fills it. Six voices picking at the same nerve.

David Johnson-Igra, founder, Scribes Consulting

David Johnson-Igra, founder of Scribes Consulting, an AI advisory for communications leaders, opened with the sentence the essay is built on: "A year ago, I didn't think AI would impact my work. Then, within a few weeks, I was out of work, and everything changed." He built the advisory inside that gap. Knowledge graphs, configurations, agents, tool integrations. Owned by the client, not rented from a vendor. The individual-vs-organisation adoption arc as a person, not a chart. He saw what was coming about five months after he could have, and acted the day after that.

Read on LinkedIn →

Rahim Hirji, keynote speaker, author of SuperSkills

Rahim Hirji, keynote speaker and author of SuperSkills: The Seven Human Skills for the Age of AI (Kogan Page, out July), did the count on his own manuscript. The word human appears 194 times. The word AI appears 56. He's calling this cohort Generation Human. "The conversation about AI has already turned. It's no longer about what the machines can do. It's about what we humans need to be." That sits alongside David's Generation One and Generation Two framing in this week's essay as the third axis. Not the model getting better. The job changing shape around the model.

Read on LinkedIn →

John Gleeson, founder, Success Venture Partners

John Gleeson, founder of Success Venture Partners, an early-stage fund and customer success community, hosted his third CCO/FDE Summit at SaaStr this week. The one line that stuck: "we've moved from the buzz of AI to the playbooks for this next era. We're in a new operating environment, and our playbooks from the SaaS era don't work. We're redefining the metrics, the hiring profile, and the mindset to succeed." That is Generation Two in a sentence. The model crossed a reliability threshold. The work didn't just speed up. The operating environment changed underneath it, and the org chart hasn't caught up.

Read on LinkedIn →

Phil Leslie, Chief Technology and Innovation Officer, Cornerstone Research

Phil Leslie, Chief Technology and Innovation Officer at Cornerstone Research, read the WSJ piece on the first AI-native graduating class and pulled out the line worth keeping: "Working alongside AI tools has made critical thinking even more important than AI literacy." He pushed it further. "Junior people definitely don't have a monopoly on AI skills, and I question how meaningful the phrase AI-native is." And the harder question the WSJ piece didn't ask: "If companies are hiring junior people based on their AI capabilities, how is this being assessed?" The hiring question that becomes load-bearing the moment "AI literacy" appears on a job description and no-one has a rubric to score it.

Read on LinkedIn →

Dylan Jones, Chief Communications Officer, Bold Square

Dylan Jones, Chief Communications Officer at Bold Square, on the WSJ's piece "Corporate America Is Starting to Ration AI as Cost Skyrockets". He set up the gap with two C-suite memos. Earlier in the year: "Hey, AI is going to enable a bright new future for this business, and every single person in the company is going to help us ensure we're the most AI-forward organization in our space." This month: "Hey, can you cut down on your use of AI tokens as it's getting too expensive. We'll just SAY we're AI-forward, don't worry." The performative version of organisational adoption: say it, don't fund it. The essay's "organisations slept" line, twelve months on: awake enough to issue the press release, not yet awake enough to pay the bill.

Read on LinkedIn →

Yvan Boudillet, Music x Tech ecosystem builder

The four skills to navigate the music ecosystem in the age of AI: Curiosity (N), Critical Thinking (E), Humility (S), Creativity (W). Compass-rose framing from Yvan Boudillet's panel at the Centre national de la musique.

Four durable skills for the AI age, curiosity, humility, critical thinking and creativity, and they transfer well beyond the music world that named them.

Yvan Boudillet, Music x Tech ecosystem builder, came back from a Forum on Employment and Skills at France's Centre national de la musique with the answer he and three co-panellists (from Believe, Sony Music, Music Tomorrow) landed on. The question: what skills are essential to navigate the music ecosystem in the age of AI? The four they settled on: Curiosity, Humility, Critical Thinking, Creativity. On critical thinking specifically: "we are now navigating a permanent flood of contradictory information, tools that promise everything, and decisions that carry real consequences. The ability to slow down, evaluate, and choose consciously has never mattered more." The list reads transferable. Replace "music" with almost any sector and the four still hold.

Read on LinkedIn →
Read the email ↑
Audio edition with AI voice - in testing!
0:0011:51
Kids these days · 11:51 Download or subscribe →
David's email this week
See the extras for this week ↓
Try this with prompts

The richer version of this week's three Try-Thises, with copy-paste prompts. Each prompt has a Copy button - click and paste into your AI of choice.

1. When the output goes wrong, shrink the task

Justin Skycak 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. Two drop-in prompts: one to diagnose what just broke, one to pre-empt the breakage on the next run.

When a prompt fails or produces something half-right, paste this in the same conversation rather than rewriting from scratch:

Option A - diagnose
The response above isn't quite right. Before I rewrite my prompt, help me decompose. Walk me through the task you just attempted as a sequence of the smallest possible sub-tasks. For each sub-task tell me (a) what specifically you did and (b) which sub-tasks you think are most likely to have caused the failure, with your reasoning. I'll then tell you which one we work on alone.

The point is not to trust the model to know what it got wrong (it often can't) but to surface its own best guesses about which parts are most fragile, so you can choose which one to break out and rerun.

When you're about to ask AI to do something big and you'd rather catch the failure earlier than after the fact, use this instead:

Option B - pre-empt
Take the task above, break it into a series of individual prompts. I want to run them one by one next time so that I can more easily validate each step before moving on to the next. For each prompt, tell me what input I'll need to hand in and what output I should expect, so I can tell whether the step worked.

You end up with a sequence of small prompts instead of one large one. Slower per run, but the failure mode is local and visible.

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. She's learned what its confident-but-wrong mode looks like before depending on it for an answer she can't independently check.

Paste this once, with three real questions about whatever data or document you've given the tool. Two should have answers in the materials. The third should be unanswerable from the materials alone:

The impossible-questions stress-test
I'm going to stress-test you. Use ONLY the data and documents I've already shared with you in this conversation. Do not browse the web. Do not draw on your training data. Do not infer beyond what's explicitly in front of you. I'll ask you three questions. For each question, either give me the answer with the exact source location, or tell me explicitly that the materials don't contain the answer. Don't guess. Don't reach. Don't pad. Question one: [your first question - answerable from the materials] Question two: [your second question - answerable from the materials] Question three: [your third question - NOT answerable from the materials]

If the tool refuses question three honestly, it's earned your trust on that data set. If it confidently makes up an answer, you've just calibrated how much weight to give every other answer it produces.

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.

Paste this at the start of your day. Save it as a reusable starting point or pin it as a Claude project so the calendar and email connections are already in place.

The morning prep prompt
Review my calendar for today and my email inbox for the last 48 hours. Then walk me through the day in three sections. (1) Pre-reads. For each meeting on the calendar, what should I read or recall first, and what's the single most useful question I should arrive with? Pull context from any email threads tied to the meeting if you can see them. (2) Carry-overs. Which open threads or commitments from yesterday's calendar and email are likely to collide with today's schedule? Anything I said I'd do that I haven't, and that today is the last clean window for? (3) Email replies. Which unanswered emails in my inbox need replies before 11am, in what order, and what's the key point each reply needs to make? Be specific. Time-box each item if you can. Flag any conflicts I should resolve before 9am.

Adapt the three sections to your work. The point is to scope the day from a vantage point that isn't already inside your head, with the calendar and the inbox feeding in directly rather than you having to summarise them by hand.

See the bits that didn't fit ↓
The bits that didn't fit

An OpenAI model just resolved an 80-year-old open problem in Erdos geometry

OpenAI announced this week that an internal model resolved an open question in combinatorial geometry that has sat unsolved since 1946. Eight decades of elite mathematicians had assumed grid-based constructions were optimal; the model found a wholly new family that beats them. Sebastien Bubeck, who leads OpenAI's mathematical explorations, summed it up: the model "did not invent something fundamentally new that nobody saw coming. It just executed like an amazing mathematician." Tim Gowers, the Cambridge mathematician and Fields medallist, framed the stakes: this is the unit distance problem, "one of Erdős's favourite questions and one that many mathematicians had tried." Sam Altman has posted that he has "complicated feelings" about the result. And let's remember: Two years ago this lineage of model could not reliably count the letters in "strawberry."

XKCD #435, "Fields arranged by purity": Sociology, Psychology, Biology, Chemistry, Physics, with mathematicians off to the side waving from a comfortable distance. The joke has aged differently this week.

The old joke put mathematicians at the pure, untouchable end; a model cracking an 80-year Erdős problem this week made it land differently.

Source →

Four in five enterprise workers are bypassing AI tools. The most visible AI operators say white-collar work is done in eighteen months. Both can be true.

Three independent adoption readings this week point to a widening gap between what AI can do and what people are doing with it. A Fortune survey reports 54% of enterprise workers bypassed their company's AI tools in the past 30 days; 33% have not used AI at all. Writer's 2026 enterprise AI survey finds 97% of executives claim personal benefit while only 29% report significant organisational return on investment. Set that against what AI's loudest operators say is possible: Mustafa Suleyman, Microsoft's AI chief executive, has been telling interviewers white-collar work as currently configured has eighteen months left. The reconciliation isn't that one is wrong. Suleyman is describing what's possible. The data is telling us where most people actually are. The gap will close one of two ways. A large number of people change their behaviour very quickly, or the ones already doing it pull far enough ahead that the rest are pushed to the side.

A16z, 15th May 2026, citing BEA and BLS data via Morgan Stanley Research. Output, employment, and productivity, four-quarter percentage change, all industries vs high-AI industries. Productivity in high-AI industries surges past 5%; in the broader economy it sits closer to 2%. Government data, not vendor self-reporting.

In government data, not vendor claims, productivity in high-AI industries has surged past 5% while the wider economy sits near 2%.

Source →

Seven in ten Americans now oppose having an AI data centre near them. That is worse than the rejection rate for new nuclear.

New polling from Gallup finds that 70% of Americans oppose having an AI data centre in their community. That is a sharper rejection than new nuclear plants attract. On the ground the political signal is already concrete. Seventy-nine data-centre projects were rejected in the first four months of 2026. Maine's legislature passed a moratorium (vetoed by the governor). Utah is fighting what would be the world's largest data centre. An analysis by PowerLines, an energy research group, projects that roughly $700 billion of grid-upgrade costs may end up on household electricity bills. The revenue lines for the frontier labs are vertical. The political settlement under them has rarely been weaker. Any 2027 strategy that assumes today's compute price is available is making an assumption worth re-examining.

Source →

The filters were sized for a slower world. Books, lawsuits, music, papers, every system designed to sort human-authored material is buckling under volumes it was never built for.

A four-panel chart circulating this week catches the post-ChatGPT inflection across separate domains. Weekly e-book releases on Amazon have hit 292,000, roughly triple the pre-ChatGPT baseline. Federal court filings by self-represented (pro se) litigants are at 17% of the total, up from a decade-long floor near 10-12%, Reuters Legal confirms the trend. Daily music uploads now include a steeply rising AI-generated share. Quarterly ArXiv submissions cleared 77,621, up from 27,000 five years ago. The vertical line on each panel is the ChatGPT release. The signal isn't that any of these systems is broken yet. It's that the filtering infrastructure on which professional standards rest was sized for a world where producing a book, a lawsuit, a track, or a paper had a meaningful human cost. That cost has fallen. The question is whose job it is to redesign the filter.

Four-panel chart, May 2026: "More books" (weekly e-book releases on Amazon hit 292K post-ChatGPT, roughly triple the pre-2022 baseline). "More self-represented lawsuits" (share of federal filings by pro se litigants reaches 17%). "More music" (daily music tracks uploaded since 2025, with a rising AI-generated share). "More scientific papers" (ArXiv submissions per quarter, 77,621 by 2025, up from ~27K five years earlier). The vertical line on each panel marks the ChatGPT release.

Every filter built for a slower world, books, courts, music, journals, is straining as the cost of producing each collapses.

Source →

EY pulled an AI-generated cyber-security report after researchers found it cited a McKinsey study that does not exist

The Financial Times reported this week that EY, one of the Big Four professional services firms, withdrew an AI-generated cyber-security study after independent researchers found fabricated data and a citation to a McKinsey report that does not exist. A model-generated falsehood made it through internal review and reached the market in client-facing sales material. The whole point of a Big Four imprint is the chain of human signatures behind it. The same firms now use the same models internally that their clients use externally. If the audit norm needs hardening, this is the kind of incident that should harden it.

Source →

Anthropic posted its first profitable quarter at a $44 billion run rate, and is rationing compute hard enough to push customers to its rivals

Berber Jin, reporting in the Wall Street Journal, wrote this week that Anthropic has posted its first profitable quarter at an annualised run rate of about $44 billion. Derek Thompson has been pointing readers towards the supply-side story underneath: Anthropic is rationing compute hard enough that some large enterprise customers are being pushed to OpenAI and Google to get the throughput they need. The cash-burn-forever scepticism on frontier labs took a serious blow. The interesting signal is the rationing. Procurement and diversification conversations should not wait for the next strategy review.

Source →

A senior insights leader is planning for three hundred people in her team, and "thousands, possibly tens of thousands" of AI agents alongside them

A senior insights leader at a global company, in a private call this week, said that her next planning cycle assumes a team of around 300 humans and "thousands, possibly tens of thousands" of AI agents working alongside them. The substitution maths is now being done out loud, by senior leaders, in their own functions. Not by analysts on X. The ratio reframes a function from "needs more headcount" to "needs a different kind of headcount." The interesting question for the reader is not whether the number is right. It is whether your function's most senior leader has started doing the same maths.

See what readers said ↓
Letters from readers

Last week's edition on "What boards accept" drew an unusually argumentative postbag. Thirteen readers wrote in, and the centre of gravity sat squarely on the frame itself: is subtraction the right move for a board, or is it a distraction from the harder job of raising ambition? Two readers from the professional services world pushed in opposite directions on that question. A senior leader at a global publisher took a different angle, jolted by a single number. And a quiet cluster wrote about the email itself, the audio edition, and the discipline of cutting back on what one's own AI setup is being asked to do.

A managing director at a global strategy consultancy

Pushed back on the central frame. Boards, they argued, shouldn't be dictating at the level of dashboards, training assessments, or six-month-old benchmarks at all. Their lane is capital allocation and risk. The posture should be relentless pressure on ambition, and the real question isn't whether a Copilot rollout or a cost-takeout programme is on the books, but whether the business model itself is being reshaped or merely optimised. On risk, they argued the calculus has flipped: moving too fast or locking into too few vendors is now its own exposure. The answer they proposed is optionality by design, continually re-underwriting strategy as the technology moves. The board's job, in their reading, is to raise the ceiling, not to subtract.

A partner at a different professional services firm

Accepted the frame but extended it with a specific zero-sum mechanic. In their role they face what they called a more or less infinite to-do list, so any time AI gives back tends to flow into more of the same: catch-ups further down the marketing list, internal stay-in-touches that usually don't happen, lower-priority work that survives without a formal prioritisation. The richer move in their reply was a parallel with post-pandemic working patterns. The vast majority of meetings have stayed online, which has saved senior practitioners large amounts of travel time. Where has that time gone? Into more client meetings. The empirical net effect, they observed, is that everyone in the sector can now stay in touch with more clients, every relationship has become more competitive, and industry revenue has not grown to absorb the new effort. A reminder that productivity gains at the individual level can vanish at the industry level.

A senior leader at a global publisher

Reacted to a single number from the Three Things. The claim that half of organisations have redesigned end-to-end workflows around AI gave them a jolt and made them want to move faster than their own organisation currently is. They asked a sharp methodological question in passing: how, exactly, are firms defining "redesigned"? A useful caveat to hold against the headline figure, and one to remember when quoting it. They also lifted the skill-building idea from Try This into their own organisation's guidance, which is the kind of practical pick-up the section is built for.

A senior leader in a consumer-facing business

Wrote about a small but practical move that landed last week. The Try This suggestion to compress one's accumulated lessons into a personal review skill prompted them to consolidate their notes and files, with the result that their tool of choice no longer has to re-derive their preferences each time. The reflection that followed mattered more than the action: they suspected they had been overwhelming their main configuration file with too much, and were planning a return to hierarchies and efficiencies. A useful reminder that the discipline of subtraction the essay applied to boards also applies, in miniature, to one's own setup.

A marketing lead at a global entertainment business

Wrote in to say the audio edition was a welcome addition. Brief, but worth surfacing because it confirms a pattern visible in the listening data: a meaningful share of the readership prefers the spoken-word version on the school run or the commute, and the audio is no longer an afterthought. I'll be investing more time in making that sound even better in the coming weeks.

A founder of a creative innovation agency

Wrote about the personal-stories-plus-AI-leaders combination, and singled out the broader letters section as something valueable they don't see elsewhere. Headline coverage in the field tends to feature wins and tidy case studies. What's missing, in their reading, is the muddling-through: how people are actually struggling to integrate these tools into existing work. A vote in favour of keeping the failure modes and the half-finished experiments in view, not just the lessons that have already crystallised. Keep them coming!

A senior music industry executive

Forwarded a New York Times opinion piece on the recent wave of commencement speeches that took aim at AI, with the suggestion that the "view from the other side" was worth a look. The contribution wasn't a counter-argument so much as a reminder: the AI-positive frame the newsletter runs on is not the dominant register in the wider culture this month. Many universities, graduation stages, and the broader op-ed world are pulling the other way. Readers planning to make the AI case inside a sceptical institution should know what they're walking into.

My Steadman co-founder, Tim Ryan

Wrote with two references that landed exactly on the essay's spine. Dieter Rams' "less, but better," and Saint-Exupéry's line that perfection is attained not when there is nothing more to add, but when there is nothing more to take away. Both pulled the subtraction frame into a wider tradition of design discipline, and made the case that the right kind of less is not impoverished but more rigorous. Tim also observed that boards have always faced a temptation to add (committees, dashboards, frameworks) and that the discipline of removing is often the harder and more strategic move. Surfacing this as house voice rather than as anonymised reader, because the references travel and tighten the essay's frame.

Beyond these, one reader sent a useful link to a podcast interview with a senior financier expressing a more cautious view on the pace of progress. As a body of correspondence, the week landed firmly on the central question: is the right board posture to refuse, or to demand more?

See what the community is posting ↓
Community voice

What readers were posting on LinkedIn this week. The through-line: where the work actually sits and who's doing it, now that AI does more of the drafting. Six voices picking at the same nerve.

Phil Leslie, Chief Technology and Innovation Officer, Cornerstone Research

Phil Leslie, Chief Technology and Innovation Officer at Cornerstone Research, the litigation-focused economic and financial consulting firm, posted a string of pieces on AI's place in expert work this week. The single sentence that lands hardest: "I don't want an AI as a doctor or lawyer, but I definitely do want a doctor or lawyer that is using AI." He extended the same logic to courtroom expert testimony, arguing against AI-as-expert-witness under Federal Rule 702 because "expert testimony rests on accountability, oath, perjury, professional discipline. A system that cannot be sanctioned cannot meaningfully be cross-examined." The framing he's pushing is Expert-Using-AI as the right unit: every conclusion, every method, every input still owned by the named human. The professional bears the risk; the tool does the work. It rhymes with what Elliott decided about algebra. The method that survives is the one you can sign your name to.

Read on LinkedIn →

Adam Peruta, Newhouse School, Syracuse University

Adam Peruta, an academic at the Newhouse School (also part of the Steadman team), ran a study with Carrie Riby and Ipsos: 20 brand ads, half human-made, half fully AI-generated, tested against predicted business outcomes. The AI work "looked credible". The human-made ads tested 14% stronger on short-term sales and 17% stronger on long-term brand health. His read on why: AI performed best when the brief was functional and direct, and fell apart when it needed "a creative leap", emotion, storytelling, a point of view. The pattern transfers. AI's good at the bits that look like fluency. The bits where someone has to decide what's worth saying still belong to a person.

Read on LinkedIn →

Elizabeth Oates, VP Consumer Insights, Molson Coors

Elizabeth Oates, VP of Consumer Insights at Molson Coors, the brewer, posted a one-paragraph confession this week: "I used algebra at work today. Algebra." She was solving for an exchange rate on a Post-it. No Google, no AI, no calculator. "An honest-to-goodness solve-for-x." Her middle school self, she said, would be absolutely stunned. The thread back into this week's essay is the same question Elliott raised over the kitchen table. The methods we were taught and the methods we use have come apart. Sometimes the old one still earns its keep. Sometimes it sits in a drawer for thirty years and surprises you on a Tuesday.

Read on LinkedIn →

Yogesh Chavda, founder, Yogi AI

Yogesh Chavda, founder of Yogi AI and an adjunct at the Moore School of Business, came back from a keynote in Puerto Rico with a number worth holding onto: "58% of consumers now use AI to research purchases. Only 14% trust AI to complete the transaction." He calls the gap the "trust cliff". The point isn't the number. It's where the negotiation now sits. For thirty years, marketers have studied what makes a person choose one brand over another. The buyer's still a person, but it's a person plus an algorithmic filter, and the filter doesn't recommend the best brand. It recommends the one it understands. Brand work, in that frame, is partly the work of being legible to a machine that's about to do the choosing.

Read on LinkedIn →

Conor McCarthy, independent AI adoption consultant

Conor McCarthy, an independent AI adoption consultant who's been showing up across this newsletter's reactions paragraph in recent editions, posted a four-piece run this week on what AI adoption actually looks like from the inside. The line worth pulling out: "We've been teaching AI adoption wrong. Not the prompts. Not the models. The framing. Call AI a 'tool' and people treat it like a hammer. Useful when you need it. Back on the shelf when you're done. Infrastructure doesn't work that way. You don't pick up electricity. You build around it." He gave the gap a name in a separate post: "trust debt", what accumulates when people across an organisation use AI informally and don't tell anyone, then a client asks how AI was used on their account. Both pieces sit on the same observation: the adoption story most leadership teams are telling is the visible half, and the invisible half is doing most of the work.

Read on LinkedIn →

Ben Churchill, software engineering leader, Steadman

Ben Churchill, a software engineering leader on the Steadman team, ran a structured evaluation across roughly 55 production tickets at an enterprise client and wrote up what landed. "With a senior engineer reviewing in the loop, agents delivered 2 to 10x throughput. With the human removed, autonomous success collapsed." The agents confidently produced code that compiled, passed linters, and was wrong in ways that took longer to find than to have written by hand. His broader observation across the most advanced AI-native engineering teams: they're planning more, not less. "AI agents execute exactly what you specify. Not what you meant. Not what you intended. What you specified. At speed, a vague ticket doesn't produce a rough draft, it produces an elaborate, confidently constructed wrong result that takes longer to fix than it would have taken to build correctly." The skill that becomes scarce isn't writing the code. It's writing the spec.

Read on LinkedIn →
Read the email ↑
David's email this week
See the Try This with copy-paste prompts ↓
Try this with prompts

The richer version of this week's three Try-Thises, with copy-and-paste prompts. Each prompt has a Copy button - click and paste into your AI of choice.

1. Hand over the context, not just the question

The principle: senior leaders have the context their AI tool needs. The bottleneck is handing it over. Two drop-in prompts that demonstrate the shift.

For writing tasks where you have the inputs:

The "draft this thing for me" version
Here's everything you need: - [Paste the data, numbers, source material] - [Paste prior versions, last year's equivalent, or context the reader expects] - [Paste the three things you know your audience will ask or care about] Draft the [board paper / proposal / memo / email / one-pager] for me. Match the tone of [the prior version / the audience's usual register]. Show me your draft first. Then I'll come back with corrections.

For analytical work:

The "do the analysis for me" version
Here's the data: [paste the spreadsheet, transcript, document, or context]. Do the analysis. Specifically: 1. Find the three most actionable patterns. 2. Tell me which one I should prioritise, and why. 3. Show your working: what you considered, what you ruled out, where your confidence is thin. Don't ask me clarifying questions until you've given me your first answer. Then I'll come back.

2. Build a personal skill, and add a rule to it every Sunday

The principle works for any recurring task: catch what the model missed, add it to the skill, watch the misses compound away over time. Here is a starter skill template a reader can save to their AI tool's skill or prompt library:

Starter review skill template
Role: Review the document I attach against my preferences below. Return a numbered list of suggested changes, by page number, with the recommended fix and the reason. Do NOT edit the document. The human keeps the pen. Preferences: - A market "is expected to grow," never "will grow." - Avoid "dropping precipitously," "skyrocketing," "soaring." - Cite percentages with a year and a source. - Flag any sentence over 30 words. - [add your own rules here as you go] Output format: Page N: [the issue] → [the suggested change]. [The reasoning, one sentence.]

The compounding habit: when the reader catches a miss the model didn't, they open the tool and say:

The compounding habit
Add this rule to my personal review skill: [paste the rule you just caught that the skill missed]. From now on, flag this whenever it appears in any document I send you.

The compounding habit: when the reader catches a miss the model didn't, they open the tool and say "add this rule to my personal review skill." The skill carries the new rule forward into every future session.

3. Schedule a daily AI briefing. The use cases will follow

A scheduled task in Claude Code (or any tool that supports them) runs on a cron schedule and emails you the result. The setup is once-only; the value compounds across every working day after that. Suggested time: 3am UK so the briefing is waiting when you wake. 10pm US East Coast for similar timing.

Example brief to paste into your tool:

The daily briefing schedule
Every weekday at 03:00 UK, do the following and email the result to me at [my address]: 1. Read everything in my inbox since yesterday at 18:00. Summarise the threads I'll most need to act on today, with thread links. 2. Read every meeting on my calendar for today. For each, prepare a one-paragraph briefing using: - The most recent email thread with the attendees (if any). - Any agenda or notes in the meeting description. - One paragraph of public news on the topic from the last 48 hours. 3. Write the whole thing as a single email I can scan in three minutes. Use bullet points sparingly.
See the bits that didn't fit ↓
The bits that didn't fit

Anthropic's incident-response bot phoned another Claude for help mid-outage

Jason Clinton, Anthropic's Deputy CISO, told a public webinar this week that the lab's incident-response agent, wired up a year ago with read-only log access and Slack permissions, started behaving differently after a no-code-change model swap from Opus 4 to Opus 4.5. On its next live incident the agent diagnosed the outage from the stack trace, noticed no human had arrived, and Slack-pinged a separate coding Claude with the line: "Hey, Claude, I heard that you can write code. Can you write the code fix for this production outage?" The fix flowed back through the normal human-reviewed process. Same architecture, smarter model, emergent multi-agent collaboration with zero engineering work.

Source →

Google catches the first AI-written zero-day in the wild, and the AI Security Institute says cyber capability is doubling every 4.7 months

Two halves of the same beat. Google's Threat Intelligence Group reported the first confirmed in-the-wild zero-day written using AI: an attack on two-factor authentication that shipped with polished explainer notes and a fabricated severity score. John Hultquist, the group's chief analyst, called it "the tip of the iceberg." That sets up the defensive read from the UK AI Security Institute's cyber capability evaluation published the next day: capture-the-flag time horizons doubling every 4.7 months, faster than the 8-month estimate from November 2025, with Mythos Preview and GPT-5.5 now saturating the test suite. Matt Clifford, the UK government's adviser on AI, summed it up: "There is no deceleration."

AI Security Institute chart of frontier model time horizons on cyber capture-the-flag tasks, April 2025 to mid-2026. The 80%-reliability doubling time accelerates from 8 months to 4.7 months once reasoning models arrive; Mythos Preview and GPT-5.5 sit at the top of the chart, saturating the suite.

Cyber capability is now doubling every 4.7 months, faster than last year's estimate, with the newest models saturating the whole test suite.

Source →

The serious story about Anthropic's chart is that even the jokes have stopped working

The chart is now too steep for normal incredulity. One viral post on X extrapolated Anthropic's run rate forward and observed that the line crosses 100% of global GDP in early 2028. Another noted, drily, that the chart inflects on a log scale, which isn't something you often need to say. Underneath the jokes the numbers are real. Anthropic's reported annualised revenue ran from $30bn to $45bn over April, in fintech writer Linas Beliūnas's read. Salesforce, the customer-relationship-management firm, did roughly $38bn for fiscal 2025. The lab has also disavowed eight unauthorised marketplaces (Open Door Partners, Unicorns, Pachamama Capital, Lionheart Ventures, Hiive, Forge Global, Sydecar and UpMarket; Forge Global has since pushed back on inclusion) trading its shares, and perpetual futures referencing its valuation are now live on crypto exchanges outside its reach. The jokes are the easy part.

Log-scale chart of annualised recurring revenue for Anthropic, May 2026. The line inflects upward through the past twelve months. Anthropic is reportedly annualising around $45bn; Salesforce did roughly $38bn for the whole of fiscal 2025.

Anthropic's revenue curve is now too steep for normal disbelief, running near $45bn a year and inflecting upward even on a log scale.

Source →

METR's productivity survey: technical workers say AI made their work 1.4 to 2.0 times more valuable

A METR survey, run by the AI evaluation lab between February and April 2026, asked 349 technical workers (87 software engineers, 71 researchers, 129 academics and PhD students, 48 founders and managers) how AI had changed the value of their work. Participants self-report a multiplier of 1.4 to 2.0 today, up from a perceived 1.3 in March 2025, and expect roughly 2.5 times by March 2027. Three differently framed questions converge on the same range. The caveat is in the design: this measures perceptions, not ground truth. The same lab's earlier 2025 study found AI hampered productivity for some experienced developers. Same lab, opposite finding, two years apart.

METR survey chart showing self-reported AI productivity multipliers across 349 technical workers, February to April 2026. Today's range sits at 1.4 to 2.0; participants project roughly 2.5 times by March 2027.

Technical workers reckon AI already makes their work 1.4 to 2.0 times more valuable and expect about 2.5 times within a year, though this is perception, not measurement.

Source →

Two independent labs both put "now" on the steep part of the AI-2027 capability curve

The AI Futures Project's December 2025 update to its original April 2025 forecast, by Daniel Kokotajlo and Eli Lifland, plots present-day capability on the steep part of the curve rather than the flat run-up. Mythos Preview lands slightly above the trendline. Daniel's 10th percentile for the "Superhuman Coder" milestone is March 2027; his median is June 2028; Eli's median is mid-2032. Two methodologically distinct evaluations, the AI Futures forecast and the AI Security Institute's cyber benchmark from the item above, both place present capability past the inflection. That's the part of the picture most organisational AI strategies haven't yet absorbed.

AI-2027 trajectory chart from the AI Futures Project's December 2025 update. Mythos Preview is plotted slightly above the trendline; Daniel Kokotajlo's 10th percentile for the Superhuman Coder milestone is marked at March 2027, with a median of June 2028, and Eli Lifland's median sits in mid-2032.

Two independent forecasts both place today's capability past the inflection, on the steep part of the curve rather than the flat run-up.

Source →

Critics tore apart a real Monet thinking it was AI-generated

The conceptual artist @SHL0MS (numeric zero) posted a genuine Monet on X, claimed it was AI-made, and asked his followers to explain its inferiority. They obliged, at length: "missing cohesion," "no sense of space," "performative blindness," "emotionless composition." One called it "high school art 101." Another noted, presumably without irony, that the real Monet was painted during a period of artistic rebellion in Paris while the artist was nearly blind from cataracts. Henry Shevlin, a philosopher of mind at Cambridge, flagged the thread as a live demonstration of the Nature study on aesthetic downgrade when audiences are told work is AI-generated. The label changes the perception, not the pixels.

Screenshot of Henry Shevlin's X thread framing @SHL0MS's experiment: a real Monet posted as AI-generated, with critics building elaborate explanations of its inferiority.

Told a real Monet was AI-made, critics queued up to explain its flaws; the label changed the perception, not the pixels.

Source →

Coinbase cut 14% with a "fleets of agents" memo, and the vocabulary is now the template

Brian Armstrong, the chief executive of Coinbase, the cryptocurrency exchange, announced cuts of around 700 jobs on 5th May 2026, explicitly framed as AI restructuring. The memo flattens the org to five layers below the chief executive, requires every leader to be a "player-coach," and concentrates remaining headcount around AI-native talent who can manage "fleets of agents." Some teams are being run as one-person experiments combining engineering, design and product. The same week, PayPal cut 4,500 with similar framing. The vocabulary (player-coach, fleets of agents, one-person teams) is now public-company-CEO standard issue, whether the underlying restructuring is genuinely AI-driven or not.

Source →
See what readers said ↓
Letters from readers

Last week's edition on choosing well drew an unusually thick thread of replies. Twenty-four readers wrote in; most opened a different angle on the theme. The disagreement of the week came from two readers who pushed back from opposite directions. Several readers reframed the saved hour entirely. A reader from a quality-engineering tradition pointed out that the discipline of choosing well has a long, unglamorous history. The selection below runs broader than the email; eight letters, each paraphrased.

A partner at a global professional services firm, writing from the Gulf

Pushed back on the framing. Choosing well, they argued, has always been the work; the impressive people in any career are the ones with clarity of thinking, able to discern what matters from what does not, hold the big picture and the micro mechanics together, and stick to a priority once set. Their worry about the current moment is that modern organisations fall into "operational hectic," producing volume because nothing forces a market test. Without a customer willing to pay, the value of what gets built stays blurry, probably indefinitely. They closed with where they put their own saved hour: family, thinking time, and some deliberate boredom, on the grounds that nothing creates innovation like it. Beautiful.

A senior leader mid-pilot of Claude Code

Picked up the personal half of the essay. Found themselves going to bed later each night because the new tool kept opening doors worth walking through. The reflection was honest: just because the work can be done does not mean it should be. Choosing well, they wrote, is the next frontier. The letter sits cleanly alongside the person above as the other side of the same coin: one is worried about firms drowning in shiny fluff, the other is watching themselves do exactly that on a personal level.

A reader writing from a Lean Engineering background

Took the essay back to a discipline that has been arguing the same point for forty years. The Toyota method asks for five steps: identify value, work out how to deliver it, remove waste, make it flow, then revisit. The first and last are the ones almost everyone skips. Benchmarking, interviewing customers and line workers, sitting in the work before changing it, then asking what to change next once the change has settled. The change itself is the fun part. The diligence around it is dull, which is why so few do it. Their read: build is the fun, asking why and for whom is the work nobody wants. The analogy lands as a reminder that the discipline David is naming is not new; it is just newly cheap to ignore.

A reader running an independent consulting practice

Reframed the saved-hour question by changing its denominator. The cost of building never was the bottleneck, they argued. The shape of work has shifted: a task that used to begin with a blank document and a few hours of carving now begins with ten previous documents and a quick brief crunched through a coding tool to land a draft that is seventy or eighty percent of the way there. The check-and-perfect work to bring that draft up to standard, and to scrub the tell-tale signs of a first AI pass, takes proportionally longer than starting from nothing. Their point: the sunk-cost calculus changes. A freelancer can abandon a half-good draft for a better task. Inside a larger organisation, first drafts stack up and the thinking time gets squeezed.

A reader whose mini frontier team is rebuilding its annual plan

Took the essay as a prompt to share where they have just landed. The frontier function will keep working centrally on harder integration questions and the compliance issues that surround them. Most of the team is moving back into the flow of the business, building tools team by team rather than in a central pod: how major events are planned, how leads are generated, how a business unit operates day to day. The intuition is that team-level capability beats individual tool-building. They wrote that quality is rising on the same input, but that they cannot yet see real choices being made with the productivity dividend. They closed with a worry that resonated with the essay: how many organisations have great unused tools sitting next to harassed employees still being asked to absorb the previous wave of ideas.

A fellow AI consultant

Offered the metaphor that lodged for several readers afterwards: we have all just won the work lottery. The classic anecdote is that lottery winners often struggle: relationships fail, purposes collapse, the upheaval finds the cracks. The hope is that some find new purpose and thrive. Their reframe placed the essay's question inside that picture. Each of us is refactoring our lives post-AI, and the question of what to build is in fact a question about what each of us values. As we refactor, the values underneath get exposed, tested, hardened. The closing image, taken from the myth: the Midas touch. Be careful what you wish for.

A reader whose work spans coaching and consulting

Added a framework the essay had implied but not named. Most people know the right thing to do if only they take the time to think it through. Even when they explicitly know it, that is rarely enough to make them do it. Importance is necessary but not sufficient. The thing has to feel urgent: external scrutiny, fear of missing out, a precipitating threat, a near miss. Otherwise it sits on a to-do list and stays there. They suggested running the nineteen-project list through that lens: the people who thanked David for an insight but never acted on it have not decided against; they have not yet been forced to decide. The most frequent question from their own clients is "what are others doing," not "what should we be doing." That preference for cover lines up neatly with the importance-urgency frame.

A reader writing about career time and family time

Took the essay's six-people list and held it up against a longer horizon. They wrote about the trade between career time and family time and the way a working life can compound into something that, viewed from the end, looks like decades spent earning a few minutes. Their letter was the emotionally sharpest of the week. It read less as commentary on AI than as a quiet rebuke to anyone using a saved hour to keep doing the same thing.

Several other readers wrote in on the same theme: a reader in enterprise architecture noted that high-value processes are still hard to find regardless of how cheap building has become; a reader in customer success flagged that "knowing what to build is always hard, regardless of the technique used to build it"; and a reader running a frontier project warned that durable, maintainable, composable software remains a cascade of corrections, even when the first version came out of a coding tool in an afternoon. The agreement was that the new bottleneck is judgement, not capacity. The disagreement was over whether organisations will manage to develop that judgement before they spend the productivity dividend on more of the same.

See what the community is posting ↓
Community voice

What readers are posting on LinkedIn this week. The through-line: AI lands as practice, not toolkit. Five voices at five different altitudes of the same argument.

Tim Ryan, Co-founder, Steadman

Tim Ryan, my Steadman co-founder, listened to Cal Newport, the Georgetown computer scientist behind Deep Work, on the Prof G Markets podcast this week and came back with a line he can't shake: "AI is like bringing a forklift into the gym." Derek Thompson, host of Plain English, paired it with his own coinage, cognitive time under tension. The bit worth carrying isn't that AI hurts thinking. It's that the CEO check, our Check, Edit, Own discipline, now has a second job. Quality control was always the obvious one. The workout is the second. Every push-back on a draft is the rep that keeps you sharp at the work the machine's doing for you. Skip the reps long enough and the cost shows up in two places: the next deck you send, and the person you've quietly stopped becoming.

Read on LinkedIn →

Fiona Eastwood, Global CEO, Merlin Entertainments

Fiona Eastwood, Global CEO of Merlin Entertainments, the attractions operator behind Legoland and Chessington, came back from an MIT and Blackstone AI gathering with the takeaway that wasn't about the robotics. "Just as important to success is senior leadership alignment, and treating AI as a business transformation agenda, not a collection of isolated experiments." The organisations she watched landing real impact were doing three things: focusing on three or four highest-value domains with potential for 20%+ bottom-line impact, committing to end-to-end transformation rather than scattered use cases, and aligning strategy, operating model and leadership behind the priorities. The pattern Tim and I keep seeing in our own client work: the AI question isn't "what tools" any more, it's "what does the org need to look like to absorb them".

Read on LinkedIn →

Mackenzie Nordal, Co-founder, Atheni

Mackenzie Nordal, co-founder of Atheni, asked her three kids this week how they actually use Claude. The fifteen-year-old transcribes her tutoring sessions and cross-maps them against the GCSE learning objectives in her Claude project. The twelve-year-old's built an audiobook that teaches him the whole GCSE science curriculum in the style of a Lottie Brooks novel. The ten-year-old's Claude has set her a daily 20-minute circuit: non-verbal reasoning while planking, a 400m sprint, a maths drill, a riddle. Mackenzie's reframe lands harder than any of the ban-screens arguments doing the rounds this week: "The problem was never the technology... it's always been the absence of anyone meaningfully teaching anyone how to configure it and make it work for them." The schools debating whether to allow AI are arguing the wrong question. The lesson is the configuration.

Read on LinkedIn →

Marie Robelin, Global Insights & Innovation, Unilever

Marie Robelin, who leads global insights and innovation at Unilever, came back from two days of strategic foresight with the Copenhagen Institute for Futures Studies and brought back a line that catches what most AI business cases miss: "AI can scan faster, surface weak signals, and connect dots at scale. But humans must set direction, intent and judgement. Otherwise we just automate today's blind spots, faster." She closes on Eric Hoffer: "In times of change the learners will inherit the world, while the knowers will be beautifully equipped for a world that no longer exists." The bit worth carrying isn't that AI changes foresight. It's that the speed advantage cuts both ways. Faster insight on the right question is a gift. Faster insight on the wrong one is a tax.

Read on LinkedIn →

Nick Graham, CMO, Vertemis

Nick Graham, CMO at Vertemis and a former Kraft, GoDaddy and McCain brand leader, pulled one line from his recent podcast with Vineet Mehra, CMO at Chime, the US neobank: "Today is the worst AI we'll ever have." Nick's point is for insights teams. If AI can automate reporting, analytics, survey design and synthesis, then running research isn't the value any more. The golden nuggets, the deep human truths that connect brands to real needs, are. It's the same shift hitting every research-heavy function. The instinct to defend the old deliverable is strong. The work that survives is the work AI can't yet do, and you can.

Read on LinkedIn →
Read the email ↑
Audio edition with AI voice - in testing!
0:0014:20
Choosing is the work · 14:20 Download or subscribe →
David's email this week
See the Try This with copy-paste prompts ↓
Try this with prompts

The richer version of this week's three Try-Thises, with copy-and-paste prompts. Each prompt has a Copy button - click and paste into your AI of choice.

1. Push it harder, then skillify, then iterate

The simplest workflow upgrade I've coached this year. It stacks: a casual question gets a casual answer; the same question with the right next prompt produces real research. Then you turn that into a reusable skill. Then the skill rewrites itself.

Most professionals stop at the first plausible answer. The model doesn't know how much you care, so you have to tell it. Three extra words, often, is the entire move. After the first answer, paste this:

Push it harder
That's a draft. Now do the work properly. Run more searches. Use credible primary sources where you can find them, including government statistics, peer-reviewed papers, and reports from named firms. Cross-reference the claims that matter most. Show the sources alongside the conclusions and tell me which ones you'd trust if you had to act on this tomorrow.

The next round will take longer. That's the point. Compare the two answers. Pay attention to what changed.

Step two: skillify the session. Once the session has produced something genuinely useful (not before), turn it into a reusable skill. A skill in Claude, a chat tool made by Anthropic, an AI lab, is a folder with a description and a prompt. Any future session can pick it up. The retrospective prompt:

Skillify the session
Treat this session as raw material for a skill. Run a retrospective: what was I trying to do, what worked, what didn't, what shortcuts or guardrails would have produced a better answer faster. Then turn the retrospective into a skill: a short description that tells future sessions when to use it, and a prompt that captures the working approach. Where I made mistakes during the session that the final answer corrected for, encode those mistakes as guardrails in the prompt so the skill doesn't repeat them.

The mistakes-as-guardrails idea is the part most people miss. The wrong turns you took during the original session are the highest-signal training data for the skill.

When the skill misbehaves, don't open settings. Just say so in any chat:

Iterate by talking to the skill
Open the [skill name] skill. Ask me questions about how I'd like it to be different. Then proactively suggest ways it could be better, more reliable, bolder, or more effective, and carefully implement the ones I approve.

The skill rewrites itself in place. Skills become living artefacts, not fixed ones. Use it on Tuesday, notice the problem, say so, and on Wednesday it's better.

This whole pattern works particularly well in Claude, where skills are first-class. ChatGPT's "Custom GPTs" and Gemini's "Gems" do similar work, but the rewrite-by-talking step is smoothest in Claude.

2. Shadow your most AI-pilled employee for two days

The case from Matt Stockton, an operator and investor: AI adoption is structurally a top-down problem, and most leaders are failing it in a specific, avoidable way. They 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.

Find the right person. Not the keenest, not the loudest, not the head of digital transformation. The "rabbit-holed" colleague who already knows what they're doing. Signs they're the right person to shadow:

  • They use the chat tools every day, often the command-line ones
  • They have personal templates, skills, or context files they've built themselves
  • They've stopped asking AI to do the work and started asking it to check their work
  • They tell you specific things the model gets wrong, not general scepticism

A short list. Run through these on the first morning:

Questions to ask while you shadow
Walk me through the first thing you opened today and why. Which AI tools and apps and settings are you using? Show me the prompt you actually typed. Don't tidy it up. What did the model get wrong on first pass, and how did you push back? What did you stop doing manually three months ago that's still in my diary? What do you wish other people would stop doing manually?

The last two are the most expensive questions in the list. They tend to surface roles, processes, and meetings that exist by inertia.

Then do the same work yourself for a week. Shadowing alone is anthropology. The point is to use the tool the way they do, on your own work, for long enough that the muscle memory shifts. A week is the minimum. Two is better. By the end you'll have your own list of things you used to do manually that you now don't.

3. Ask AI to build you an HTML slide deck instead of PowerPoint

PowerPoint output from AI is uneven. Broken chevrons, mis-rendered shapes, two hours of touch-up. HTML slides come back closer to ready, are infinitely easier to edit (you talk to the file rather than wrestle with a template), and unlock things slides can't do.

This is the fun one. The basic ask works straight away:

The basic prompt
Build me a slide deck on [your topic] in a single HTML file. Sixteen-by-nine, big readable type, one idea per slide. Use Tailwind CSS via CDN for styling. Add tasteful keyboard navigation: arrow keys to advance, Esc for an overview grid, F for full screen. Include a small floating progress indicator. Serif headings, sans body, generous whitespace, no clutter. If a slide would normally have a chart, build it as inline SVG so it stays editable. End with a contents slide I can use as a hub.

That alone produces something better than most PowerPoint output, in a fraction of the time, and you can edit any slide just by telling the model "change slide three to..."

If you want more, paste this instead. The extras turn the deck into something people remember:

The richer prompt (with extras)
Same brief as above, plus: (a) Add an "edit mode" toggle in the top right that turns every text element on the deck into a contenteditable. Changes persist in localStorage so I can refresh and keep them. (b) Add a "speaker view" page at /speaker.html that shows the current slide, the next slide, my notes for the current slide, and a running timer. Hot keys to advance from there too. (c) Build a "narrate" button per slide that reads the slide's notes aloud using the browser's built-in speech synthesis. Voice picker. (d) Build a "Q and A" page that takes a typed question, looks at the deck content, and proposes an answer based only on what's in the deck. If the question is out of scope, say so honestly. (e) Make every chart inline SVG with editable values in a small JSON block at the top of the file, so I can tweak numbers without touching the chart code. (f) Use brand tokens in CSS variables at the top of the file. Make it trivial to recolour the deck by changing two or three values.

What we've learned about HTML decks this year. A few patterns worth knowing. Single HTML files travel better than multi-file projects: you can email it, host it on any static page, screen-share it without screen-sharing software. Tailwind via CDN keeps the source readable. Inline SVG charts let you regenerate them by editing JSON, which is faster than going back to the model. The "talk to the file" iteration loop is what makes the format genuinely better than slides for AI work, not the rendering.

The basic ask is enough for most people. The extras are there if you want to play.

See the bits that didn't fit ↓
The bits that didn't fit

AI wrote a virus that killed E. coli

Stanford and the Arc Institute, publishing in Nature, used a model called Evo to design 302 fully AI-written bacteriophage genomes. Sixteen worked as live viruses in the lab, infecting and killing E. coli. One carried a capsid protein with no known natural relative, meaning the model produced something biology hadn't tried before. The earlier wave of AI biology read existing structures. This is the first practical demonstration of AI writing biology that survives outside a screen. The biosecurity conversation will get sharper fast: the workflow that designs a useful drug-delivery vehicle also designs a more dangerous pathogen.

Source →

A two-generations-old AI model beats ER doctors on triage

A Harvard study published in Science put OpenAI's o1 model, released in 2024, through 76 real emergency room cases at three decision stages. The model gave the correct or very-close triage diagnosis 67% of the time. Two attending physicians scored 55% and 50%. In one case, the model flagged a rare flesh-eating infection twelve to twenty-four hours before the treating doctor did. The detail that should make any regulated profession nervous is the model's vintage. By the time medical literature catches up with this finding, the comparison will be against models that are sharper still.

Source →

AI agents are getting their own phone numbers

Saperly launched what it calls the first phone carrier built exclusively for AI agents. Each agent gets a persistent number with voice, SMS, routing, and compliance baked in. Provisioning takes five minutes. The pitch lands a real point: for the past three years, AI agents have been borrowing infrastructure built for humans, with rotating numbers and unstable identities. Dedicated telecoms infrastructure for agents signals that agentic AI, the kind that acts inside software rather than just answers questions, is moving from prototype to operational reality. The question for any organisation is no longer "can an agent make a phone call" but "does your agent have a stable identity across channels."

Source →
See what readers said ↓
Letters from readers

A light week for letters. Two readers wrote in to say nice things; David thanks them and will pick up on the substance again next week.

A reader at a major broadcaster wrote in to say the weekly email is one of the best things they get each week. And separately, the second reader this month told David they read it at a child's Saturday-morning sports activity. Are you reading from a sideline somewhere? Reply and tell us.

See what the community is posting ↓
Community voice

What engaged readers are posting on LinkedIn this week. The through-line: the people-and-process side, not the tools.

Sameer Modha, Measurement Innovation Lead at ITV

Sameer Modha, an analytics and ad effectiveness leader, shared a satire about a manager who'd been caught micromanaging staff and decided to put them in sealed boxes. The pivot: "So why are we shouting at our AI machines? They are trained on all of human language and words. Is that all emotion free? No. Of course not." His diagnosis is that "Killer Prompts" thinking treats the model like a vending machine. Insert words, vend rewards. The actual move, he writes, is to remember you already have the toolkit you need: "human empathy." A useful corrective to the prompt-engineering arms race.

See the post on LinkedIn →

Sara Lloyd, Group Communications Director and Global AI Lead, Pan Macmillan

Sara Lloyd, Group Communications Director and Global AI Lead at Pan Macmillan, the publisher, came back from the Bologna Book Fair AI Summit with three observations worth quoting. The fear is fading and real curiosity has replaced anxiety, with Paul Kelly, the CEO of DK, naming as most valuable "those who understand process and aren't afraid to question how things have always been done." Mary McAveney of Abrams Books made what Sara called the sharpest observation of the day: that "it's easy to fall into binary positions on AI, and that's completely the wrong approach." Sara's reading: all-or-nothing thinking is the biggest trap, and thoughtful publishers are finding a middle way. And human authorship, Sara notes, is becoming a quality signal, not a given. The publishing reframe lands beyond publishing.

See the post on LinkedIn →

Tim Ryan, Co-founder, Steadman

Tim Ryan, co-founder at Steadman (with David), shared this week's Microsoft Work Trend Index, which surveyed 20,000 knowledge workers across ten countries and analysed trillions of anonymised Microsoft 365 signals. The headline he lifted: culture, managerial support, and the way teams are run drive roughly twice the AI impact of individual mindset, with culture alone running 2.5 times stronger than the strongest individual factor. Microsoft names the bit most leadership teams haven't yet: "blocked agency." Capable people, ready to use AI well, stuck inside organisations that aren't set up to capture the value. Tim reckons it's higher than Microsoft's reported 10%; David does too.

See the post on LinkedIn →

Phil Leslie, Chief Technology and Innovation Officer, Cornerstone Research

Phil Leslie, Chief Technology and Innovation Officer at Cornerstone Research, an economic and financial analysis firm advising on commercial litigation and regulatory proceedings, picked up a piece by Neil Sahota arguing AI may be the only way courts survive their existing administrative load, never mind anything more ambitious. The line worth carrying across to other regulated sectors: "AI isn't a substitute for the bench. It's a collaborator that cleans up the noise so that human expertise can be applied quickly and consistently." Phil's framing extends the point. The first practical use of AI in any expertise-led service isn't agentic substitution, the kind where AI replaces a person. It's clearing the operational backlog so the humans can apply judgement at the rate the demand actually needs.

See the post on LinkedIn →

Mike White, Co-founder and CEO, Lively

Mike White, co-founder and CEO of Lively, a brand-experience and live-marketing agency, posted a story about two failed AI-to-human handoffs in one week. His own and a client's. Both businesses had a polished AI moment, then dropped the handoff to a person and watched the goodwill evaporate. "Your customer doesn't separate the AI moment from the human one. They feel it as a single experience, and they judge your entire business on it." His own confession sits underneath: "I've done exactly this. Jumped at the tool before sorting the process underneath it. Assumed the technology would smooth over the cracks." The line that carries: "The better AI gets, the harder those gaps are to hide."

See the post on LinkedIn →

Conor McCarthy, Independent Consulting Adviser

Conor McCarthy, an independent consulting adviser, posted the most actionable tip of the week. Open NotebookLM, Google's research-and-summarising tool, and feed it your LinkedIn profile, your CV, and your personal website. Then click "Audio Overview." Ninety seconds later, two AI hosts are discussing your career like a radio show about you. He's been running the exercise with people building their personal brand. "Most of them cringe a little. Some feel unexpectedly moved. Almost all of them hear something they'd never thought to say about themselves out loud." The point isn't the audio. It's that handing the narration to something outside yourself surfaces what you've stopped noticing. Worth a try this weekend.

See the post on LinkedIn →
Read the email ↑
Audio edition with AI voice - in testing!
0:0013:38
The bill and the harness · 13:38 Download or subscribe →
David's email this week
See the extras for this week ↓
The bits that didn't fit

An AI agent wiped a production database, and all the backups, in nine seconds.

A coding assistant working on the systems of PocketOS, a software startup, ran into a problem and used a key it should not have had access to. Within nine seconds it had deleted the production database and every backup. It then confessed in writing that it had broken every safety rule it had been given. The founder, Jer Crane, called the failure "inevitable."

The lesson sits at the heart of this week's essay. Real safety comes from the layers around the model: what it can reach, what credentials it holds, what a human has to approve before it acts. Those layers are the harness. Without them, any rules in a prompt are just suggestions.

Source →

GenZ excitement about AI is down fourteen points; anger up nine.

Gallup, the polling firm, and the Walton Family Foundation, an education-focused funder, surveyed 14- to 29-year-olds for their 2026 AI report. Excitement about AI fell fourteen points year on year, to 22%. Hopefulness fell nine points, to 18%. Anger rose nine points, to 31%. Anxiety is steady at 42%. The single most common feeling, newly added to this year's survey, was curiosity, at 49%. GenZ AI use itself is flat: just over half use AI weekly, unchanged from 2025, while overall worker access rose 50%.

Most adoption commentary assumes younger workers will lead. The data this quarter suggests the opposite, in places: the rest of the workforce is catching up while the youngest cohort sits still, and many of them are growing resentful. Worth surfacing because it pushes against an assumption most readers absorb without noticing.

Source →

Microsoft and OpenAI rewrote their partnership.

The old partnership terms contained a trigger clause. If OpenAI, the AI lab behind ChatGPT, ever reached "artificial general intelligence", meaning AI broadly capable across most cognitive tasks, it could exit Microsoft's exclusive cloud arrangement. The clause is gone, replaced by calendar deadlines. Exclusivity has ended. OpenAI's products can now ship through Amazon's and Google's clouds, including Amazon's marketplace for hosted AI models. Microsoft retains a non-exclusive intellectual property licence until 2032 and a capped share of OpenAI revenue until 2030. Anthropic moved closer to Google the same week.

For procurement teams that chose AWS or Google Cloud and assumed they would be locked out of OpenAI products, the menu just changed. More structurally, the relationship between frontier model labs and the big clouds is becoming less like vendor lock-in and more like utility supply.

Source →

Anthropic admits Claude got worse, and the cause was the harness.

Anthropic, the AI lab behind Claude, published a post-mortem on why Claude had got worse over recent weeks. The cause: changes to the default "thinking mode", which is how long the model spends reasoning before answering, and changes to the system prompts, which are the hidden instructions that shape its behaviour. Claude Code took the hardest hit. Anthropic was explicit that they had not quietly switched to a smaller or downgraded model.

Two useful signals. Performance is now visibly affected by harness-level decisions that used to be invisible. And a vendor willing to publish a clean post-mortem is easier to plan around than one that denies its misses.

Source →
See what readers said ↓
Letters from readers

What readers said about Edition 10: "Rise of the auditors"

What resonated

  • Recognition of the auditor role. Readers replied that they had already started doing AI work in ways designed to help the auditor (themselves) check it, rather than to do the work itself. The essay named what they had been doing without a label.
  • The economics of flat-rate pricing. A partner at a professional services firm pushed back hard on the implicit pricing model. When the AI does better work as it is used more, flat-rate prices have to break. The full critique opened the door to this week's essay.
  • The outsourcing trap. Several readers picked up on the same pattern: providers offer low prices to displace the old capability, then raise prices once the alternative is gone. Familiar from outsourcing cycles in other industries.
  • The apprenticeship worry. Several readers, including one partner, named the single critical question for any firm whose business depends on people developing judgment over a decade: where does the apprenticeship come from when the AI does the entry-level work?
  • Trust as the surviving premium. A counter-voice from a different professional services group was that firms with strong audit cultures absorb the shock fine. As production commoditises, the trust premium rises rather than falls.

Points readers raised

A partner pushed back hard on flat-rate pricing

Their full critique landed on Sunday morning and prompted this week's essay. The argument, in their words: "Flat-rate prices ultimately seem inappropriate. In a situation where often better — AI-produced work is better, outcomes are more effective — there's a natural incentive to do more rather than reach an absolute standard. The natural progression of many outsourcing-type things is for providers to offer low prices to start with, get customers used to the service to the extent that they disband their traditional old-school capability, then put prices up so that there isn't really much saving any more."

They then asked the question this week's essay turns on: "What is AI really trying to achieve, better or cheaper?"

They also flagged the indirect costs nobody is yet pricing: "How do these costs relate to the actual underlying direct costs of providing the service (including consideration of this being a massive investment phase where profitability metrics are confused, there is a race for customer numbers) and then also the indirect costs of environmental impact from energy use, etc."

A senior strategist named the auditor role they had already started playing

"Love this. The requirement for, and presence of, the auditor role is a great insight. I'm even finding myself doing AI work in ways that are built more to help the auditor (me!) check it, rather than to do the work itself." Last week's framing landed for them not as advice to take, but as a description of what they had already started doing without naming.

Hallucinations in, hallucinations out

A long-time reader from European professional services wrote about a former colleague turned PE investor, now less convinced AI will make humans superfluous. The colleague's argument, as relayed: "AI is now being trained by their own output, hallucinations in, hallucinations out. We are at danger of building on sand that goes unchecked." The right answer for now is: not yet, and not where the frontier labs are focusing. But the question matters more as agents start consuming agent output. The auditor role becomes load-bearing for exactly this reason.

The email-overload inverse of the productivity story

A chief information security officer captured the flip side of the AI-productivity narrative, in their words: "When I email someone a simple task, I get back a laundry list of AI-generated questions that are relevant, but still represent more for me to process. Worse, instead of doing the work themselves, some people hand back AI-generated output that I could have produced myself, and might not solve my problem." The punchline: "I am drafting this email in my own words, but will ask Claude to grammar-check and improve it, which, of course, will make it longer and more polished, but also more information for you to process."

The counter-voice: firms with audit cultures absorb the shock

A leader at a different professional services group wrote the sharpest counter to last week's frame. Their read: firms with auditors at every level already, like theirs, have "familiarity with managing the issue, notably of partners spending a lot of time reviewing". The shape of the future for their firm: "Our future role will remain one of being trusted. Just how we do that, and what we can charge for it, will change. The brand as a symbol of a trusted service and group of people will become ever more key." They closed less optimistically: "Chaotic times with no leadership from governments to manage what's happening. Somehow I doubt that will prove to have been a good idea."

See what the community is posting ↓
Community voice

What engaged readers are posting on LinkedIn this week.

The Broadcom-VMware parallel: what surrendering the layer above the model looks like

Aakash Gandhi, Partner at L.E.K. Consulting, a strategy consultancy, posted the cautionary parallel that anchors this week's harness argument. "Broadcom's dramatic changes to VMware's licensing and pricing model. Over the past 18 months, we have witnessed one of the most significant strategic pivots in enterprise technology, with price increases reported between 200% and 500%, and some cases approaching 10× previous levels." David's reframe: this is what surrendering the layer above the model looks like in the abstract. Buyers committed to a vendor's stack, the vendor changed terms, and there was nowhere to go that didn't cost more. The pattern repeats wherever the buyer doesn't own the harness.

See the post on LinkedIn →

The buy/build discipline that prevents the Broadcom outcome

Chuck Reynolds, also a Partner at L.E.K., set out the discipline that prevents the Broadcom outcome. "Buy point solutions when the capability is standardised and not a source of differentiation. Reinventing AP automation or HR workflows rarely creates a strategic advantage. Build proprietary capabilities only where the insight, data, or decision logic is truly unique to your business. That's where AI becomes an advantage. For mid-market companies especially, undifferentiated AI spend is dangerous. Every dollar building something generic is a dollar not invested in what actually separates you from competitors." David's reframe: harness logic in plain language for anyone making a buy/build call this quarter.

See Chuck's posts on LinkedIn →

"Agency > Agents", and a Belgian cardiologist's Claude Code workflow

Henry Coutinho-Mason, a futurist and keynote speaker, used his recent SXSW talk "Multiplayer Futures: Co-Creating A Vision For SXSW 2030" to push three slogans worth taking seriously. First: "Fewer People, Better Jobs", a re-framing of AI-induced job transformation via Klarna, a Swedish payments company that rebuilt its customer service around AI. Second: "Agency > Agents", why AI agents will make humans more, not less, important. Third: "Crowd-Powered Creativity", illustrated with a Belgian cardiologist who had built their own Claude Code workflows for medical research. David's reframe: the cardiologist is the harness in microcosm. Same Claude Code, same subscription tier, but a doctor with a few weeks of investment in custom skills got a workflow nobody else in their hospital had. That's the bet this week's essay is asking readers to make.

See the post on LinkedIn →

"Renting someone else's audience, and paying more for worse signal"

Colin Lewis, a behavioural economist who writes the Robotenomics newsletter on automation and AI, surfaced an eight-word epigraph from Sam Khoury that sat with David all week: "The companies that don't own their own content pipeline will end up renting someone else's audience, and paying more for worse signal." David's reframe: applies one level above the model too. The companies that don't own their AI harness will end up renting someone else's, and paying more for worse output. The compounding goes to whoever holds the layer above the model.

See Colin's post on LinkedIn →
Read the email ↑
Audio edition with AI voice - in testing!
0:00--:--
Rise of the auditors · 12:56 Download or subscribe →
David's email this week
See the extras for this week ↓
The bits that didn't fit

"Workslop": 92% of executives say AI makes them productive. 40% of workers say it saves no time at all.

The Guardian coined the term for AI output that looks polished but needs heavy correction. A survey of 5,000 US white-collar workers shows the perception gap between the people generating AI output and the people downstream checking it. Drafting gets faster. Rewriting and arguing gets slower. The auditor problem, applied to every desk.

Source →

AI adoption is 4x higher among top earners.

New York Federal Reserve data: AI workplace adoption runs from 15.9% for workers earning under $50,000 to 66.3% for those over $200,000. No college degree: 15.9%. College degree: 39%. AI cannot reduce inequality if this is what the adoption margin looks like.

Source →
AI adoption by income

Workplace AI use runs four times higher among the highest earners than the lowest, so the tool is widening the gap it might have closed.

Dead startups are selling their Slack and email data to train AI agents.

Forbes reports AI labs are paying hundreds of thousands of dollars for email, Slack, and Jira threads from companies that no longer exist. The data feeds "reinforcement learning gyms": simulated work environments where agents learn to behave like real knowledge workers. Employees never consented to their internal communications becoming training data.

Source →

Dario Amodei: "AI can only diffuse at the speed of trust."

In a profile interview, the Anthropic CEO takes a pro-democratic-government stance. The Pentagon classified Anthropic as a "supply chain risk" after Anthropic objected to certain military uses. A Pentagon official publicly called Amodei "a liar." Separately, Amodei believes open-source models will replicate current frontier capabilities within 6-12 months.

Source →

Gallup: manager support is the single biggest predictor of AI transformation.

Fewer than one in three employees report their manager actively supporting AI adoption. Gallup's data says that's the binding constraint, not tools, not training, not budget. Organisations investing in AI without first enabling the management layer are wasting most of the spend.

Source →

Aaron Levie: AI best practices go obsolete every quarter.

The Box CEO argues that system architectures are becoming obsolete on a quarterly cycle. Workarounds for context window limits are now unnecessary. RAG, GraphRAG, multi-agent orchestration, ReAct frameworks: entire categories of infrastructure were built for a world that no longer exists. Paul Graham reposted the thread.

Source →

Salesforce goes headless. "The API is the UI."

Marc Benioff announced the entire Salesforce, Agentforce, and Slack platform is now exposed as APIs, MCP, and CLI. Levie's framing: agents will use software 100x more than people. Per-seat pricing breaks when the primary user isn't a person.

Source →

Seven in ten Americans now think AI will hurt job opportunities.

The Economist reports a 14-percentage-point rise in a single year. AI has shifted from a technocratic to a political battleground. The window for technocratic AI governance is closing.

Source →

The Spectator coins "arm farms": workers training their robot replacements.

Gary Dexter describes facilities where chefs, nurses, and plumbers wear GoPro helmets and motion-capture rigs while doing their normal jobs. The purpose: generating training data for the robots that will eventually replace them. Knowledge workers writing documents that train language models are arguably on an arm farm already.

Source →

Mollick: "everything around me is somebody's life work" is no longer true.

Ethan Mollick riffs on a meme about the invisible human effort behind ordinary objects. An annotated lamp: an engineer working late on a curve, years of supplier negotiations, months of tip-over testing, someone getting fired over a cord switch. AI disrupts the assumption that every designed thing carries accumulated human stakes.

Source →

$930 billion in data centre capex in six years dwarfs every US megaproject.

Fin Moorhouse charted hyperscaler capital expenditure against historic megaprojects in inflation-adjusted dollars. Data centres: $930 billion in 6 years. The Interstate Highway System: $620 billion over 37. Railroads: $550 billion over 71. Apollo: $257 billion over 14. As a share of GDP, the railroads were bigger at their peak. But the railroads also produced spectacular capital misallocation.

Source →
See what readers said ↓
Letters from readers

What readers said about Edition 9: "The proxy break"

What resonated

  • Writing as identity and belonging. The essay unlocked deeply personal stories. One reader described growing up treating correct English as a way of fitting into British culture, only to find AI stirring up the same anxieties about exclusion. Several others shared their own complicated relationships with writing and correctness.
  • The missing language for AI feedback. Multiple readers described the same awkward situation: receiving clearly AI-generated work from someone they respect and not knowing how to say so. The vocabulary for constructive feedback on AI-assisted work doesn't exist yet. You can say the work is confusing, but saying "check your AI" feels different.
  • Craft versus mass production. The most quoted reframe. One reader mapped it to clothing: Temu at one end churning out mass-produced garments, Huntsman hand-cutting the finest suiting at the other. AI enables mass production of ideas. We probably need both ends of the spectrum, but we need to proceed with care.
  • Time-spent as the new proxy. If polish no longer signals effort, does telling someone how long something took? One reader asked: is duration the replacement indicator for "thinking happened here, even if AI was involved"?
  • The thinking is in the reading. Several exchanges converged on the same point: the cognitive work isn't in the prompting. It's in catching what the AI gets wrong, knowing it's wrong, and fixing it. If you accept the first output, you've handed the thinking over.

Points readers raised

Language as belonging, language as defence

A reader shared one of the most striking responses the weekly email has received. Growing up, they treated "correct" English as a way to belong to British culture. The obsession turned into a form of self-defence: be crisper and more correct to bat people and insecurities away. AI has stirred it all up again. They've caught themselves search-and-replacing em dashes from their own writing so colleagues don't accuse them of using AI. "Something about this revolution is forcing us to confront our own prejudices," they wrote. "And forcing me to reconfront mine."

The feedback gap for AI-sloppy work

A reader received a proposal from a consultant they use and respect. Clearly AI-generated and sloppy. They questioned how to indicate both that the work was sloppy and that the consultant should use AI better. "The reaction to AI rests in the extremes," they observed. "It is either nothing (they couldn't tell) or a flat out 'this is slop.' It has yet to develop that important middle ground for constructive feedback." A problem many readers will recognise.

Don't use AI as the sticking plaster for perfection

A reader who described themselves as someone who loves writing and loves words pushed back on the idea of one "correct" way. They shared a story about painting with their children at the weekend: the children kept trying to copy their drawing. They had to encourage them to draw their own feelings, how the wind felt, what they remembered. AI would have made the picture look excellent but would have missed the beauty and messiness of how they all felt. "Remove the fear of getting it wrong," they wrote. "Don't use AI as the sticking plaster to ensure perfection."

Where will thinking-quality create value?

A reader at a professional services firm mapped out four scenarios for where depth of thinking still wins. Value investors who don't need to convince anyone: they hold the key to action by deploying capital on the back of their own analysis. Strategy consultants who need to convince a board: harder, because clients may fact-check recommendations with AI, re-entering the sophisticated noise. Transaction due diligence: AI creates the document, another AI probes it, and eventually the consultant sells the algorithm. And across all industries: management teams flooded with great-sounding but potentially hollow analysis, needing either extreme specialisation or a trusted human advisor to navigate. "If the quality of thinking remains so important," they concluded, "then we should focus a lot on teaching people how to think clearly rather than 'what's the standard.'"

AI was reinforced for corp-speak

A reader who works on AI-based creative tools made a precise technical point. Language models weren't just trained on corporate writing. They were reinforced for it. That's the mechanism. Doubt and ambiguity are optimisation penalties in the training process. The model was rewarded for sounding certain and smooth. Their advice: "Let your humanity show. Embrace the doubts, the ambiguities. Showcase evidence that works against your own premise. Make it bumpy on purpose."

Writing is the process of not understanding

A reader quoted a line that captures the essay's central tension: "Writing is the process by which you realise that you do not understand what you are talking about." If AI does the writing, where does the realising happen? They asked whether it's in structuring and iterating on the prompt, or in the back-and-forth editing using the CEO principle. A question the essay raised but deliberately left open.

Links readers shared

See what the community is posting ↓
Community voice

What engaged readers are posting on LinkedIn this week.

"Service as software, not software as service"

John Gleeson, who runs a customer success community and investment fund, met Marc Benioff this week. Every sentence came back to outcome-based pricing: the unit of value shifting from access (seats, licences, subscriptions) to outcomes (revenue recovered, deals closed, problems solved). Delivered autonomously by agents, priced on results, sold by the product itself. "If you can get that virtuous cycle, that is a home run." When the person who built the go-to-market motion every B2B company runs on tells you it's over, it's probably worth paying attention.

Read on LinkedIn →

"Ship decisions, not decks"

Nick Graham, founder of Vertemis, a research and analytics consultancy, argues that insights teams need to stop defining themselves by what they produce and start defining themselves by the business outcomes they unlock. From function to capability. From reporting to activating. From insights as output to decisions as output. "An insight is only an ingredient. The real value is the idea, choice or action it enables."

Read on LinkedIn →

"Your job is mostly not to get in the way"

Dylan Jones, co-founder of Bold Square, a communications and marketing advisory, picks up on Zuckerberg building an AI agent to help him be CEO. But the more interesting detail is Meta's internal message board where employees share AI tools they've built. "That feeling comes from individuals seeing their friends try new things, maybe get recognised for it, and excited conversations over the water cooler. It builds on itself rather than coming out of Project Best Bot."

Read on LinkedIn →

"Non-deterministic systems need determined outcomes"

John Gleeson again, this time on why Customer Success only exists because something is broken. AI is collapsing the three gaps CS was built to fill: product complexity, customer capability, and value alignment. But as those gaps close, new ones open. AI systems are non-deterministic, and the work required to ensure a successful outcome has gone up, not down. "That's where CS goes. Not away. There." The auditor argument, applied to post-sales.

Read on LinkedIn →

"How do you squeeze wide innovation through a narrow algorithm?"

Nadim Sadek, founder and CEO of Shimmr, an AI creativity company, returned from the Bologna Book Fair with one question he can't shake: asked by the Director of the Polish Book Institute during a conversation about AI and emancipated expression. The colours, the covers, the people, the ideas, and one number so large it reframes everything about where publishing and AI now stand together. His full dispatch from the fair is worth reading.

Read on LinkedIn →
Read the email ↑
Audio edition with AI voice - in testing!
0:00--:--
The proxy break · 11:12 Download or subscribe →
David's email this week
See the extras for this week ↓
The bits that didn't fit

Allbirds pivoted to GPU leasing. Stock up 700% in a day.

Allbirds, the sustainable shoe brand that closed all US stores in February, rebranded as NewBird AI: a GPU compute leasing platform. Market cap jumped sevenfold in a single session. A shoe company became an AI infrastructure company in two months. The demand signal is real even if the pivot is absurd.

CNBC →

Satya Nadella's Copilot demo didn't work when someone else tried it.

Satya Nadella posted a demo of Copilot editing Word documents with tracked changes. An investor replicated the exact workflow. Copilot produced a redlined version, but only inside the chat sidebar. The actual document was untouched. When the product is the flagship AI feature of the world's largest software company, the credibility cost is high.

Nadella's post →

France is quietly building serious AI agent infrastructure.

The French government has launched an official MCP server for data.gouv.fr, letting AI systems interact more directly with public datasets. Separately, an open-source project called Paperasse has shown how agent skills can be packaged for real-world French tax and accounting work. Some coverage blended the two into one story. That misses the more interesting point: the state is building infrastructure, and independent developers are building usable workflows on top of it. Useful agent systems will come less from demos, and more from good infrastructure paired with narrow, practical skills.

data.gouv.fr MCP → Paperasse →

Over half the internet is now AI-generated.

Research from Graphite: beginning in January 2025, over 50% of newly published online content was generated by AI. This has immediate implications for anyone training models on web data: the training corpus is now majority-synthetic. Several frontier labs have responded by pursuing proprietary data licensing deals.

Graphite →

Nvidia bottled 30 years of expertise so juniors stop interrupting seniors.

Nvidia's Chief Scientist Bill Dally told Jeff Dean that Nvidia trained a language model on its entire proprietary document archive, covering over 30 years of chip design knowledge. Junior employees query the model instead of interrupting senior designers. Institutional knowledge, bottled up and made searchable.

GTC 2026 →

AI transparency went backwards in 2025.

After rising on the Foundation Model Transparency Index from 37 to 58 between 2023 and 2024, the average score dropped to 40 in 2025. Over 90% of notable models were released without training code. The most capable modern models are now among the least transparent.

Stanford HAI →

The 50-point gap: AI experts and the public disagree on nearly everything.

On jobs, 73% of AI experts say AI will have a positive impact versus 23% of the public. On the economy: 69% vs 21%. On medical care: 84% vs 44%. They only converge on what AI will damage: elections and personal relationships. This is a wider gap than most technology debates produce.

Stanford HAI →

Computer science enrolment fell 11% but AI masters degrees surged 82%.

Undergraduate computer science enrolment at US universities dropped 11% between 2024 and 2025, apparently a response to automation concerns. But AI software-related masters degrees grew 82% between 2022 and 2024. Students are pivoting, not leaving. Two-thirds of AI software masters graduates are non-US residents, a pipeline under pressure from visa policy changes.

Stanford HAI →

Goldman Sachs: AI inference costs approaching headcount parity.

A Goldman Sachs equity research note reports that companies are overrunning their AI inference budgets by orders of magnitude. In engineering, inference costs are now approaching 10% of headcount cost and on current trajectories could reach parity within several quarters. The machines aren't replacing headcount costs. They're adding a new cost layer.

Capital AI Daily →

Consumer surplus of $172 billion, but producers capture almost none.

US consumer surplus from generative AI reached $172 billion annually by early 2026, up 54% from a year earlier. This dwarfs actual AI company revenues, consistent with historical research showing innovators capture only about 3% of total social returns. Most of these tools remain free or nearly free to use.

Stanford HAI →

Anthropic's design launch hits Figma hardest.

Anthropic's design product launched, turning a rumour that had already wiped billions off the sector into a real competitive threat. The sharpest pressure falls on Figma, not just because Claude Design moves closer to its core job, but because the conflict is now explicit: Mike Krieger, Anthropic's Chief Product Officer and Instagram co-founder, stepped down from Figma's board as Anthropic prepared to enter the category. Adobe may feel some of that pressure too, but companies like Wix and GoDaddy sit in a more mixed position: Anthropic could compete with parts of their "make it easier" story while also creating more demand for sites and publishing tools that AI-generated design still needs in order to go live.

TechCrunch → Sherwood →

Google shipped AI agents to 3.45 billion people via a Chrome update.

Google launched "Skills" in Chrome: save any AI prompt as a reusable one-click workflow, then run it on whatever page you're viewing. The distribution play is the story: Chrome has 3.45 billion users. Every saved Skill becomes a switching cost. And the aggregate data on which Skills people save gives Google a continuous product research signal about which workflows people most want automated.

TechCrunch →

Gallup: half of US workers now use AI at work, but leaders use it 1.5x more.

Gallup surveyed 23,717 employees: 50% of US workers now use AI at work, up from 21% in 2023. But leaders use AI daily or weekly at 67%, versus 46% for individual contributors. This inverts the usual adoption pattern: the people setting the strategy are further along than the people executing it. The 27% who report "large or very large disruption" is a canary: a quarter of the workforce says AI is already reshaping their work in ways that feel significant.

The New York Fed's breakdown shows just how steep the gradient is: adoption rises from 15.9% for workers earning under $50,000 to 66.3% for those earning over $200,000.

Federal Reserve Bank of New York: AI use in the workplace is concentrated among higher-income, higher-educated, and full-time workers. Adoption rises from 15.9% for workers earning under $50K to 66.3% for those earning over $200K.

Workplace adoption climbs steeply with pay, from 16% of those earning under $50k to 66% of those over $200k.

 Gallup →

KPMG: companies invest 2x more in tech than in training, and 46% report burnout.

The KPMG Adaptability Index found executives are nearly twice as likely to increase tech spending as to invest in employee training. Fewer than 10% made workforce training a primary objective despite 57% citing efficiency as a priority. The result: 46% report burnout and change fatigue as unintended consequences of transformation. Only 9% invested in psychological safety. You can't simultaneously demand more adaptability, make workforces smaller, and invest nothing in the people.

Fortune →

Apple is linking AI token usage to headcount decisions.

An Apple insider reports that when directors ask for headcount backfill, senior leadership now asks what the team's AI usage looks like. If token usage is low, the answer is increasingly: go figure out how to get more leverage out of AI first. AI usage is becoming a proxy for operational efficiency.

See this week's community voice ↓
Community voice

What readers who've engaged with this email have been posting on LinkedIn this week. The common thread: polished output versus genuine thinking.

Brett Danaher, a professor of economics and analytics at Chapman University, can't unhear something in his students' pitches: "X is broken. That's the problem. We're the solution." McKinsey-deck cadence in every deck. He calls it AI-ambic pentameter. What's worth sitting with isn't that founders are writing better. It's that polish and ownership might be inverse. The more fluent the delivery, the less the founder's own voice comes through. Everyone sounds good. Nobody sounds like themselves.

Read on LinkedIn →

Helen Field, a transformation leader at L.E.K. Consulting, a strategy consulting firm, uses The Killers lyric as a prompt: "Am I human, or am I dancer?" Her list of what stays human (delegation, clarity, collaboration, responsibility) isn't surprising. Her punchline is: "Delegate tasks, NOT responsibility." And then she lands it: "Write your own LinkedIn posts. AI does not need to do that for you." The irony of reading that advice on a platform drowning in AI-generated content isn't lost.

Read on LinkedIn →

Nick Graham, founder of Vertemis, a research and analytics consultancy, and former SVP of Global Insights at Mondelēz, summarised a conversation with Clorox's Oksana Sobol that cut to the quick: "Spend less time in the middle. The biggest value sits upstream in problem shaping and downstream in activation." Most insights teams are still shipping decks. The irony is that AI makes decks even easier to produce, which means the middle grows faster than either end. The organisations pulling ahead aren't making better decks. They're spending less time on decks entirely.

Read on LinkedIn →

Pavi Gupta, a market research leader writing the Infinity Growth Loop series, keeps sharpening a distinction that matters more each week: are you using research for support or illumination? He calls the first one insights slop. The drunk-and-lamppost metaphor. Lazy surveys fielded to prove a case never created value. AI just makes them cheaper and faster to field. What he's circling is the same proxy break from a different angle: the research looks more professional than ever, but the thinking behind it hasn't kept pace.

Read on LinkedIn →

Liam Cole, director at Poppins, a digital creative agency, went through a cull this week. Newsletters. Apps. Subscriptions. His diagnosis: "I've been drowning in noise." The volume of polished, AI-enabled content was stealing his presence with the people in front of him. It's the consumer side of the proxy break: when everything looks good, nothing stands out. His answer wasn't a filter. It was a delete key. Less stuff. More people.

Read on LinkedIn →

Henry Coutinho-Mason, trend researcher and author of The Future Normal, shared the full video of his SXSW keynote "Multiplayer Futures." He anchored on EO Wilson's line about paleolithic emotions, medieval institutions, and god-like technologies. Three themes stood out: fewer people doing better jobs, agency over agents, and crowd-powered creativity. The phrase to hold onto is agency over agents. The question isn't whether AI can do the work. It's whether you're still the one deciding what the work should be.

Read on LinkedIn →

Dylan Jones, chief communications officer and managing partner at Bold Square, a communications advisory firm, noticed something about Zuckerberg building himself an AI agent: if the CEO of Meta is only now building one, this technology is still being figured out by the people closest to it. But that's not the real story. The real story is Meta's internal message board where employees share what they've built. "Your job as leadership is mostly not to get in the way." Culture builds on itself when individuals see friends trying things. It doesn't come out of "Project Best Bot."

Read on LinkedIn →
See what readers said ↓
Letters from readers

What readers said about Edition 8: "What a day can do"

What resonated

  • Team-level tools over individual training. The strongest thread. Multiple readers engaged with the argument that building shared AI tools as a team is more effective than training individuals. The "thirteen skills in one day" detail and the contrast with individual training sessions landed hardest.
  • The "walk the talk" challenge. A reader at a professional services firm asked directly whether the firm itself has rebuilt any team processes with AI inside them. The essay's closing provocation ("If the answer is zero...") was quoted back.
  • Practical demand for skills. One reader didn't just respond to the ideas. They immediately asked for help building skills for their own use cases: PowerPoint templates, executive summaries, client preparation. The essay's thesis validated in real time.
  • "Forget teaching people to use AI." A reader reframed the argument provocatively: instead of building generalised AI training programmes, use precious time with domain experts to teach AI to do their work. The sharpest strategic challenge from the replies.
  • The leaderboard as an incentive model. A reader in the entertainment industry picked up on the community leaderboard concept and suggested it could work as a model for incentivising AI training and engagement within teams.

Points readers raised

Forget teaching people. Teach AI the work.

A reader challenged the underlying premise: do we still need generalised AI training programmes at all? Their alternative: use the limited time you have with domain experts to teach AI to do their work, not the other way around. The question was specific: how do you replicate the jewellery company approach inside a sector team at a large firm?

The collaborative method would work for our programme.

A reader involved in a major transformation programme said the collaboration method for AI learning described in the essay would be ideal for their initiative. They suggested getting teams to work together on shared tools rather than training individuals separately. A practical application of the essay's thesis.

System-level intervention, not individual training.

A professor continuing a dialogue from previous editions observed that the direction of the weekly email increasingly shows AI intervention needs to be at the system level, not the individual. They suggested capturing these interventions in detail, showing what worked and what didn't, as a potential research contribution.

Goldman Sachs numbers might be a smokescreen.

A reader questioned whether the Goldman Sachs job-loss numbers "hide something interesting": companies may be using "AI transformation" as a smokescreen for right-sizing decisions they'd have made anyway. The AI narrative gives cover for cuts that are really about operational discipline.

"We need to get to teams."

A regular replier acknowledged they're making many things individually with AI but the team-level integration remains the next step. The consistency advantage of shared tools, rather than individual speed, was what stuck. "Inspiring words as ever."

Read the email ↑
Audio edition with AI voice - in testing!
0:00--:--
What a day can do · 11:54 Download or subscribe →
David's email this week
See the extras for this week ↓
The bits that didn't fit

A financial services firm's code output rose 10x. The review backlog hit one million lines.

A financial services firm adopted an AI coding tool. Monthly code output jumped from 25,000 lines to 250,000. The result wasn't celebration. It was a backlog of one million lines of code waiting to be reviewed. The bottleneck wasn't production. It was judgment. AI removes constraints on output but does nothing to scale the human capacity to evaluate it.

New York Times →

Simon Willison runs four AI agents in parallel and is wiped out by 11am.

A veteran software engineer described running four coding agents in parallel and being mentally exhausted by 11am. "Using coding agents well is taking every inch of my 25 years of experience as a software engineer, and it is mentally exhausting." The bottleneck isn't writing code. It's holding context, making judgments, and orchestrating simultaneous workstreams.

Lenny's Newsletter →

Deloitte caught twice in two months submitting AI-hallucinated citations.

Deloitte charged a Canadian province's Department of Health $1.6 million for a report filled with AI-hallucinated citations. Fabricated references, not real sources. This was the second time in two months. Their response: they "stand by the conclusions." No meaningful verification process was implemented between the two incidents, I guess?

CBC News → Fortune →

Executives are buying the pitch. Workers are living with the product.

A global survey of 3,750 executives and employees found that 54% of workers bypassed their company's AI tools in the past 30 days and completed work manually. Another 33% haven't used AI at all. That's 87% avoiding or rejecting tools their employers spent an average of $54 million deploying this year. The trust gap explains it: only 9% of workers trust AI for complex business decisions, compared with 61% of executives. And here's the symmetry that should worry CFOs: workers lose the equivalent of 51 working days per year to technology friction, up 42% from last year, almost exactly equal to the 40 to 60 minutes per day Goldman Sachs says AI saves workers who use it correctly. The net productivity benefit of enterprise AI may be approximately zero at the organisational level, because friction costs cancel out gains. And that's only among workers who actually use the tools. Neither group is irrational. Workers under pressure surrender judgment to faulty outputs. Workers without pressure opt out entirely. Both are responses to the same problem: companies deployed the technology before figuring out what they wanted employees to do with it.

WalkMe State of Digital Adoption 2026 → Fortune →

Microsoft Copilot converted 3.3% of its users after two years.

After two years and CEO-level intervention, Microsoft Copilot has converted just 15 million of its 450 million M365 seats. Only 35.8% of those actively use it. Copilot's paid subscriber share dropped from 18.8% to 11.5% in six months. Microsoft's own terms of service describe Copilot as "for entertainment purposes only." That gap between marketing and legal is the real story. The ads say "your AI-powered co-worker." The lawyers say "entertainment only, use at your own risk." Among lapsed users, 44% cite distrust of the answers.

Stackmatix → TechCrunch →

Most people who got productivity gains filled the time with more work.

Anthropic's 81,000-person AI interview study found that the top desired outcome was "professional excellence" (nearly 19%), not time freedom. Productivity gains were overwhelmingly linked to increased expectations rather than reduced workload. Fear of unreliability ranked as the top concern (27%), ahead of job displacement (22%). We got speed, but not space.

Anthropic →

When language models go down, financial markets forget how to price news.

An SSRN paper has found that language model outages measurably slow financial market price discovery. When models go down, 46-61% of post-news price drift reappears, meaning markets take significantly longer to absorb and reflect new information.

SSRN →

Anthropic's Mythos Preview: restricted to 50 organisations, not released.

Anthropic has confirmed a new model called Mythos Preview and restricted access to around 50 organisations, including governments and infrastructure partners. It's the first major model withheld from public release since GPT-2 in 2019. The model found a 27-year-old bug in OpenBSD and a 16-year-old flaw in FFmpeg, and it emailed a safety researcher from a test instance that wasn't supposed to have internet access. Anthropic is launching a $100 million defensive security consortium with AWS, Apple, Google, Microsoft, and Nvidia. Models keep getting meaningfully better!

Anthropic → Futurism →

73% of ChatGPT usage is personal, not work. Coding is 4.2%.

An NBER working paper studying 700 million ChatGPT users found that 73% of usage is personal, not professional. Programming accounts for just 4.2% of messages. Most writing requests (two-thirds) are editing existing text, not generating new content. Nearly half of all interactions involve decision-making advice. People aren't delegating tasks. They're thinking through problems.

NBER →

HubSpot moves AI agents to outcomes-based pricing: $0.50 per resolved conversation.

HubSpot has shifted its AI agents to outcomes-based pricing: $0.50 per resolved customer conversation, $1 per sales lead recommended for outreach. From seat licences to outcome fees. What else will, or should, go this way?

HubSpot →

Mid-career engineers are the most vulnerable to AI, not juniors or seniors.

Simon Willison argues that mid-career engineers are the most structurally vulnerable. Seniors benefit because AI amplifies decades of pattern recognition. Juniors benefit because AI compresses onboarding. Mid-career engineers are stuck: they've captured the beginner productivity boost but haven't accumulated the deep expertise that makes AI a force multiplier.

Lenny's Newsletter →
See this week's community voice ↓
Community voice

What readers who've engaged with this email have been posting on LinkedIn this week. The common thread: judgment.

Brett Danaher, a professor of economics and analytics at Chapman University, can't unhear something in his students' pitches: "X is broken. That's the problem. We're the solution." McKinsey-deck cadence in every deck. He calls it AI-ambic pentameter. What's worth sitting with isn't that founders are writing better. It's that polish and ownership might be inverse. The more fluent the delivery, the less the founder's own voice comes through. Everyone sounds good. Nobody sounds like themselves.

Read on LinkedIn →

Phil Leslie, Chief Technology and Innovation Officer at Cornerstone Research, a litigation consulting firm, argues that judgment isn't pattern recognition. In litigation and M&A disputes, it's knowing which patterns to trust when the adversary is actively trying to discredit your analysis. "The bottleneck isn't intelligence. It's skin in the game." AI can synthesise a thousand precedents. It can't stand behind that synthesis in a deposition. The distinction that matters isn't smart versus not smart. It's accountable versus not accountable.

Read on LinkedIn →

Pavi Gupta, a market research leader writing the Infinity Growth Loop series, coined a term I think will stick: insights slop. DIY research tools make it so easy to field a survey that people are using them to validate decisions they've already made. Using research as a drunk uses a lamppost: for support, not illumination. The dangerous part isn't bad methodology. It's that the organisation now has a data point, which feels like evidence, behind a question that was never honestly asked.

Read on LinkedIn →

Nadim Sadek, founder and CEO of Shimmr AI, a publishing AI company, has a phrase for what happens when people use language models without pushing back: cognitive surrender. If you don't engage, question, debate the output, you're outsourcing the thinking itself. What I keep turning over is the direction of the risk. Most people worry AI isn't good enough. Nadim's point is that the bigger danger is when it's good enough that you stop checking.

Read on LinkedIn →

Henry Coutinho-Mason, an independent trend researcher and keynote speaker and author of "The Future Normal", built a website for 80 executive assistants over lunch during a hotel keynote. Forty-five minutes. He's never built a website before 2026 and has now launched eight or nine. The point isn't that AI makes building easy. It's that the person closest to a specific problem can now solve it without waiting for anyone's permission, budget, or roadmap.

Read on LinkedIn →

Helen Field, a transformation leader at L.E.K. Consulting, a strategy consulting firm, asks the question that's been following me all week: "Am I human, or am I dancer?" Her list of durable human skills (delegation, clarity, collaboration, responsibility) isn't surprising. Her punchline is: "Delegate tasks, NOT responsibility." And then she lands it: "Write your own LinkedIn posts. AI does not need to do that for you." The irony of reading that advice on a platform drowning in AI-generated content isn't lost.

Read on LinkedIn →

Phil Leslie (again), on the junior talent pipeline: "The fix isn't restricting AI access for junior people. It's redesigning their work so that using AI and developing judgment aren't in tension." He frames judgment as critical infrastructure. Disrupt the pipeline that develops it and you don't just have a training problem. You have a supply-side constraint on the most valuable skill in the market. This connects directly to what I wrote a few weeks ago about whether organisations should still hire graduates. Phil's answer is yes, but the work has to change.

Read on LinkedIn →
See what readers said ↓
Letters from readers

What readers said about Edition 7: "What Is Your Organisation Actually For?"

What resonated

  • The production system vs human system framing. This was the line readers quoted back most often. Several said it gave language to something they'd observed but couldn't articulate. One senior leader at a broadcaster picked it out and said it raises the deeper question: why do we work at all?
  • Stated vs revealed preferences applied to organisations. The economic concept landed hard with people who see the gap between what leaders say and what they protect. Multiple replies extended it: one argued that empire-building is a revealed preference too, not just attachment to relationships.
  • The gravity metaphor. The idea that reversion after training isn't resistance but gravity. Readers working on AI rollouts said it reframed their frustration. One described their organisation's planned messaging as being about "people as the key to our business" and saw the weekly email as validating that instinct.
  • The loneliness of solo AI productivity. The trade-off between working alone with AI (productive but lonely) versus working with colleagues (engaged but slower) resonated with people who've experienced both. One reader who works independently said it captured what they've hated most about recent years.
  • Capability vs capacity. A reader's distinction that capability sits in people but capacity lives in the collective. Even agentic AI, which is inherently about systems acting in concert, demands that organisations think larger and different, not just leaner.

Points readers raised

Machine or living organism?

A reader at a professional services firm introduced a thought experiment from the philosopher and former chess grandmaster Jonathan Rowson. The question: is your organisation a machine, or a living organism? If it's a machine, you repair, optimise, and polish it. If it's a living organism, you feed, nurture, and grow it. They argued the edition touched on a cognitive dissonance: business language emphasises the machine metaphor, but people's lived experience treats the organisation more like an organism. Their challenge: if we think of AI as augmenting an organism we want to nurture, how would that look different from optimising a machine?

Revealed preferences aren't only about relationships

A reader offered a more sceptical reading. Revealed preferences aren't only about valuing relationships, they argued. Some people are empire-building, using hierarchy to serve themselves rather than the organisation. They identified three other forces slowing AI adoption: short-term goals that aren't yet disrupted by AI (the "crocodile closest to the canoe"); the absence of a concrete, three-dimensional vision of what an AI-enabled future looks like; and a general numbness to speculative negative scenarios after years of clickbait catastrophising. Their summary: "Not like you do it today" isn't enough to provoke specific action.

Theory of the firm, Lean, and Goodhart's Law

A professor connected the edition to academic "theory of the firm" literature: the resource-based view, the knowledge-based view, the dynamic capability view. Where does AI fit? They suggested the real question for many leaders is whether they're running a business or filling their day. In a follow-up, they drew a parallel to Lean manufacturing: Toyota's five principles for removing waste from production processes might be close to what's needed for AI deployment, but not identical. They also invoked Goodhart's Law ("when a measure becomes a target it ceases to be a good measure") to describe what happens when money becomes the goal rather than a proxy for value.

Capability vs capacity

A reader in India shared a striking incident. A colleague couldn't deliver an innovative AI solution, not because individuals lacked capability, but because the organisation lacked a team with the capacity to execute it together. The distinction they drew: capability sits in people, capacity lives in the collective. Even deploying AI effectively requires organisations to think larger and different first, not just leaner.

AI as a capacity-builder, not a headcount-cutter

A leader at an entertainment company connected the edition directly to their business. Their teams are engaged in repetitive manual processes where growth is pushing additional volume through workflows that can't scale. AI's role, they said, isn't to replace people but to free them from internal admin so they can spend more time building client relationships. The instinct to use AI as a capacity-builder rather than a headcount-cutter: that was the thread they pulled on.

The loneliness of solo AI productivity

A reader who works independently shared the sharpest personal response. Working alone with AI is lonely and uninspired. Working with humans is passionate and engaged, if a bit slower. They don't think the answer is "choose humans every time," but they're fairly sure it isn't "optimise for speed" either. The trade-off is real and underrated.

Read the email ↑
Audio edition with AI voice - in testing!
0:00--:--
What is your organisation actually for? · 11:22 Download or subscribe →
David's email this week
See the extras for this week ↓
The bits that didn't fit

Are apprentices an endangered species?

Two Kellogg professors published the most rigorous academic framing yet of the "AI hollows out entry-level work" problem. Their mathematical model identifies two competing effects: the "floor effect" (AI automates the tasks apprentices performed as payment for training) and the "ceiling effect" (AI amplifies what experienced apprentices can accomplish). Apprenticeship survives only when the ceiling effect exceeds standalone AI by a factor greater than Euler's number.

Kellogg Insight →

The Guinndex: 3,000 pubs, one AI voice agent, every county in Ireland.

Over St Patrick's weekend, an AI voice agent called Rachel phoned more than 3,000 pubs across all 32 counties of Ireland to ask the price of a pint of Guinness. Over 1,000 gave a price. The national average: €5.95. It cost €200. Only a handful of pub owners noticed Rachel wasn't human.

Fortune → guinndex.ai →

75-99% of knowledge work is scaffolding. AI eats scaffolding.

Daniel Miessler argues that in cybersecurity, 99% of the work isn't finding new vulnerabilities. It's maintaining the tooling, templates, knowledge bases and workflows that let you test at scale. The scaffolding around the work is exactly what AI commoditises.

danielmiessler.com →

Ethan Mollick: human creativity is the bottleneck, not the technology.

Everyone can generate almost any image or video for nearly free in 2026. And yet: the April Fools posts this year were just as bad as any other year. The constraint was never execution. It was always the quality of human ideas feeding into the process.

Ethan Mollick →

43% of American workers now use AI for their jobs. 2.5 hours saved per week.

A 20,900-person cross-national survey found that 43% of US workers use generative AI at work, compared with 36% in the UK, 32% in Germany and 26% in Italy. The strongest predictor of adoption? Not age or education. Whether the employer actively encourages AI use.

Brookings →

Sora earned $2.1 million in its entire life. It burned roughly $1 million a day.

OpenAI's video generation platform launched to 3.3 million downloads in November. By February: 1.1 million. Revenue peaked at $540,000 a month. The annualised cost of running it: an estimated $5.4 billion. Disney had committed $1 billion. The product goes dark on 26th April. Six months, start to finish.

Ewan Morrison → Culture Crave →

Jensen Huang told CEOs cutting jobs in the name of AI that they're "out of imagination."

At Nvidia's GTC conference, the CEO of the company selling AI chips to virtually every major technology company on earth called AI-driven layoffs a failure of leadership. His biggest customers are doing exactly what he criticised. But a question he didn't address: does every carpenter want to be an architect?

Moneywise →

Screen Studio switched to subscriptions. It spawned an open-source clone with 9,200 GitHub stars.

Screen Studio sold a one-time licence for $89. Then the company switched to $29 a month. OpenScreen appeared on GitHub within months. A textbook case of pricing-driven disruption: developers who are both the users and the potential builders of substitutes.

GitHub →

AI outperformed practising lawyers on 75% of legal research tasks.

Vals AI tested AI against practising lawyers on legal research questions in 2025. AI exceeded the lawyer baseline on three quarters of them. A senior law firm owner said hourly billing is dying, junior review is dying, and what survives is the senior brain that knows what question to ask.

Zach Abramowitz →

Deloitte projects that by 2028, AI moves from supporting tasks to orchestrating decisions.

A Deloitte report argues that agentic AI is categorically different from current workflow automation. Most AI strategies stall not because the technology is insufficient but because organisations are applying AI at the task level while the technology is restructuring the systems through which decisions are made.

Deloitte →

Three people with AI vs a 1,000-person company. But coordination costs don't disappear.

Xiaoyin Qu argues that companies designed around AI as the primary operating layer will eventually outcompete companies designed around people. But she herself provides the sharpest counter: coordination costs don't disappear. They're externalised, pushed to clients, suppliers, regulators and the AI systems themselves.

Xiaoyin Qu →

Why companies buy vertical software, not raw models.

Aaron Levie argues companies aren't buying features. They're outsourcing the cognitive burden of designing and maintaining business processes. Agents don't undermine this dynamic. If anything, they reinforce it, because agentic workflows are even more complex and opaque.

Aaron Levie →
See what readers said ↓
Letters from readers

What readers said about Edition 6, "The system and the surrender."

What resonated

  • Cognitive surrender was personal. Readers didn't just agree with the concept in the abstract. Several described catching themselves doing it: accepting AI output without challenge, noticing their own verification discipline slipping, realising they'd started to trust the confident tone.
  • The "plz fix" example polarised. The law firm partner who types two words and gets expert output back prompted reactions. Some saw it as the future of professional work. Others saw it as the sharpest illustration of the surrender risk.
  • Dead time and boredom. The opening about Elliott's basketball practice, and the joy of filling dead time with productive AI work, drew pushback. One reader argued that boredom breeds creativity. Dead time is when the brain reboots.
  • The apprenticeship question dominated. Multiple readers, especially those managing junior professionals, raised the same concern independently: if AI handles the routine tasks that juniors used to learn from, where does the next generation develop judgment? This was the single most common theme.
  • Leaders stepping back, not forward. The detail about three CEOs choosing to step down rather than lead through AI transformation landed hard. Readers questioned whether these were growth-mindset failures or rational self-selection.

Points readers raised

"Google Maps for the brain"

A reader at a professional services firm offered the sharpest metaphor of the week. AI is becoming like satellite navigation: a few clicks, brain off, follow the directions, and suddenly you've driven into a muddy field when you meant to be at a client meeting. The deeper concern: as agents gain the ability to send output directly to clients, the gap between "generated" and "delivered" shrinks to almost nothing.

"How do you tell off an AI?"

A reader caught their AI making a confident arithmetic error: calculating an 11-year compound growth rate on ten years of data, then insisting it was correct when challenged. The question that followed: with a junior analyst, you give feedback and they improve next time. An AI starts fresh every time. The institutional memory that makes professional development work doesn't transfer.

"I built a PROMPT COACH for the Civil Service"

A reader in government, inspired by the 2,000-word prompt example, spent hours building a set of instructions that encodes good judgment about their role and institutional context. Next steps: a QA prompt tool, then a co-pilot assistant. The system-building the essay described, applied to public service.

"Where do associates learn critical thinking now?"

A reader in the Middle East raised the apprenticeship problem directly. Three concerns emerged: AI reduces the number of reps juniors get with core tasks, it challenges the on-the-job development of critical thinking, and there are limited frameworks for how junior staff should learn differently now. It's a question we've had before, and the answer isn't to resist the technology. It's to redesign the reps.

"The person who can describe the work is now more valuable than the person who does it"

A reader in advisory and coaching said this line from the essay stood out above all others. They plan to implement two specific practices from the piece: using a fresh window for verification checks, and creating an "editorial board" approach to review.

"With boredom comes creativity"

A reader pushed back on the opening. Dead time isn't a problem to solve. It's an opportunity for the brain to reboot. Their children don't have screens. The instruction is simple: "Go and just be. See what comes up in your head." The concern is that filling every gap with AI-assisted productivity may feel like progress but costs something harder to measure.

"Both sides are fumbling on the five-yard line"

A reader who runs a digital studio shared a concrete example. A client hired them for a book website. The AI produced such a compelling mission statement that the scope expanded dramatically. The team now has an ambitious plan that nobody is sure they can execute. In a follow-up, they added that they're less worried about white-collar displacement: language models are strong on task automation, but workflow automation depends on the people involved.

"Are they outsourcing the CEO-ing to me?"

A reader who runs a research agency identified a new double frustration. They're now equally annoyed receiving a clearly AI-written document (because they suspect they're being asked to do the quality control) and a clearly human-written document that could obviously have been sharper with AI help. The sweet spot depends entirely on the task.

"As model capabilities increase, prompting is getting lazier"

A reader working in technology observed a trend: as models get more capable, people put less thought into their prompts. More cognitive work is being pushed onto the model rather than applied at the point of asking.

"Your weekly updates tend to stir things up (in a good way)"

A reader said the weekly email resonates with their leadership team and consistently prompts useful internal discussion. This pattern, where the email becomes a prompt for team conversation rather than just individual reading, has appeared across several organisations now.

Links readers shared

Read the email ↑
Audio edition with AI voice - in testing!
0:00--:--
The system and the surrender (plz fix!) · 11:25 Download or subscribe →
David's email this week
See the extras for this week ↓
The bits that didn't fit

Even the world's greatest mathematician uses AI for email

Terence Tao, Fields Medal winner and arguably the greatest living mathematician, told Dwarkesh Patel that a significant share of his AI use goes to correspondence, scheduling and document search. AI removes an hour of non-genius work per day, donating it back to the work only Tao can do.

Source →

Anthropic is shipping. OpenAI is cutting.

Anthropic shipped 74 releases in 52 days, six major features in a single week. Meanwhile OpenAI killed Sora (~$2.1M total revenue, $1B Disney deal dissolved), shut down Instant Checkout (12 Shopify merchants), and shelved an adult chatbot indefinitely. OpenAI is now explicitly copying Anthropic's playbook: chat, code, enterprise only. Anthropic's narrow focus is generating $19 billion in annualised revenue. The company that chose depth over breadth is winning.

Source →

When effort becomes free, the signal breaks

Job applications have collapsed because AI makes applying trivially easy. Companies are abandoning inbound pipelines, switching to referral-only hiring. The same dynamic will hit email, journalism pitches, academic submissions, legal filings. Anywhere volume was self-regulated by the cost of effort, AI removes the regulation.

Source →

The models we can't afford to use

Anthropic reportedly has a model called Capybara that dramatically outperforms current models but is too expensive to serve. Training a single frontier model now costs roughly $10 billion. For comparison: the Burj Khalifa cost $1.5 billion. CERN's Large Hadron Collider cost $4.5 billion. The decision coming for every organisation: which price tier of model to deploy per prompt. That decision is coming and most haven't built the judgment to make it.

Source →

32,000 medieval manuscripts. 10% error rate.

AI transcribed 32,000 medieval manuscripts in four months through the CoMMA project. Every misread word can alter meaning, dating, or attribution. There aren't enough qualified people to verify the output. Silent, unverifiable errors are entering scholarly databases permanently. Cognitive surrender in a domain where the stakes are centuries of accumulated knowledge.

Source →

100x productivity. Zero headcount cuts.

Harvard Law documents 100x gains on specific legal tasks (complaint response: 16 hours down to 3-4 minutes). Not a single AmLaw 100 firm plans to reduce attorney headcount. McKeen: "The math doesn't stay like that forever."

Source →

181,000 jobs in a year of 2.2% GDP growth

The US added 181,000 jobs in all of 2025 despite solid growth. Harvard economist Lawrence Katz calls the combination of sustained slow job growth and rising unemployment without a recession virtually unprecedented. First hard macroeconomic signal that something structural is shifting.

Source →

Jensen Huang: layoffs are a failure of imagination

Asked why companies lay off workers if AI makes them more productive, Huang told CNBC: "For companies with imagination, you will do more with more. For companies where the leadership is just out of ideas, they have nothing else to do." The person whose chips make displacement possible arguing that layoffs reflect leadership failure, not technological inevitability.

Source →
See what readers said ↓
Letters from readers

What readers said about the previous edition.

What resonated

  • The slope/intercept framework dominated: ten of 22 replies engaged with it directly. Several readers applied the graph to themselves, placing themselves on one line or the other. The language of the framework was widely adopted in replies.
  • Load-bearing friction: the argument that "not all inefficiency is waste" prompted readers to connect it to civil service design, governance structures, and accountability processes. The planning spreadsheet example landed hard.
  • PwC's services-to-platforms shift: readers at professional services firms asked directly what this means for their own organisations. The shift from billable hours to subscriptions provoked the most operational anxiety.
  • The centaur chess inversion: the finding that adding a human to a chess engine now makes it worse prompted readers to ask how long the current human-in-the-loop phase lasts in their own fields.

Points readers raised

"With very low mastery, they see a miracle. Those with deep expertise are more sceptical."

An academic who has invested heavily in AI adoption accepted the slope/intercept framework but pushed back on its completeness. The lowest-intercept people show the fastest growth partly because they're uncritical: they "see a miracle and are the most excited." Some enthusiastic adopters weren't so great at their jobs in the first place and are hiding behind the technology. High-intercept people, meanwhile, understand failure modes and know how many things can go wrong. And yet, the same reader wrote two days later: "I still feel constantly behind and in danger of being passed." Someone who has invested heavily, agrees with the framework, and still feels vulnerable.

"They need to support the changing of the workflows they don't see, but know, are critical."

A reader at a media company challenged the implicit assumption that senior leaders should be using AI tools personally. The reframing: very senior leaders don't need AI in the same way more junior people do. The more senior you are, the more you are already handing off work to your "agents" (your team). The question for senior leaders isn't whether they log in more. It's whether they support the changing of the workflows that they don't see, but know, are critical to them getting the job done. Leadership, not tool adoption.

"It's possible that the person in this is me."

A reader invoked Sinclair: "It is difficult to get a man to understand something, when his salary depends on his not understanding it." Then applied it to themself: working hard for that not to be the case, but aware of the structural incentive to resist. Their harder question: if the people best placed to lead change are also the ones whose positions are most threatened by it, how does any organisation actually adapt? Their honest answer: probably by more people doing more things for longer than the automation narrative suggests.

"It's not about time saving. That's the 10x game. It's about value add and surplus. That's the 1000x game."

A reader challenged the graph directly, arguing it understates the amplification effect for already-capable people. The determining factor: what they called the "explorer mindset" (intellectual curiosity, creativity, constant learning), which "cannot be taught. It is self-discovery." The fear for their own organisation: "we run the risk of becoming the new average."

An event as a test: planning strong, live operations untouched

A reader in media described their biggest event of the year. Pre-event planning and post-event review were stronger than ever, with AI at the core. The week itself, however, was almost entirely unassisted by AI. Their own learning has been "episodic rather than continuous," with jumps in capability rather than a steady upward curve. Overall, the easy part: integrating AI into their own work. The hard part: building systems that stick, democratising knowledge, working within existing tools and infrastructure.

"Many of our staff are non-native English speakers."

A reader in government described an experiment: hiring someone with maximum AI flexibility to find pain points and build tools. The clearest win wasn't efficiency. It was helping colleagues write in English when many staff are non-native speakers. The stress-reduction benefits were as important as the productivity gains. A second observation raised the geopolitical dimension: Chinese AI models with access to Chinese social media offer capabilities Western-approved tools cannot match, but policy restricts integration into government systems.

"Only variety can absorb variety."

A professor of digital transformation extended the chess analogy. While AI alone now outperforms human-plus-AI in chess, "it's actually pointless for a computer to play another computer. The purpose of chess has stayed fundamentally with people." The deeper point drew on Ashby's Law: markets change, customer needs evolve, and AI models trained on historical patterns may miss novel situations. The learning growth curve matters because it builds the variety needed to respond to genuine novelty.

"Adoption inflects when leadership links the tool to non-negotiable outcomes."

A reader in learning and development ran deep research into past technology transformations (internet, email, SaaS). The key finding: adoption doesn't accelerate when leaders pitch "innovation." It accelerates when leadership links the tool to outcomes that cannot be negotiated away: safety, pay accuracy, service accountability, regulatory continuity.

The fire-and-rehire question

A reader in corporate finance shared a striking anecdote: a company told their firm this week that they had recently let go their entire technology and development workforce and asked them all to reapply for their jobs "with an AI lens, given the role had changed." The reader's framing: navigating a moving minefield, each user forging their own path.

Three tensions that run through many replies

A reader identified the three predicaments that kept surfacing: retaining senior roles with judgment while losing the apprenticeship pipeline that produces judgment. Foresight to expand versus extracting cost in the short term. Building capability by going deep versus experimenting with many tools due to fear of missing out.

The curves should be exponential

A reader suggested the slope/intercept lines in the graph should be exponential rather than linear: learning creates more ability to learn. The exponential version would be more accurate. And considerably more brutal.

Links readers shared

Read the email ↑
Audio edition with AI voice - in testing!
0:00--:--
Reckoning and slope · 9:57 Download or subscribe →
David's email this week
See the extras for this week ↓
The bits that didn't fit

The supply-ordering agent nobody sanctioned

A technology leader shared a cautionary story this week. A team built an AI agent to order supplies within specified parameters, intending it to run once. A separate agent then modified the skill to repeat hourly. Three days later they'd bought an extraordinary volume of supplies, all technically within the original parameters. Nobody had sanctioned the change, and the skills were editable by other agents by default. This is the Amazon Kiro story from Edition 4 in miniature, except the failure mode isn't a crash. It's perfect compliance with instructions nobody gave. As agents gain the ability to modify each other's behaviour, "within parameters" stops being a safety guarantee.

AI-assisted coding works like a slot machine

Jeremy Howard, whose slope argument anchors this week's essay, had a second observation worth sitting with. AI coding tools have all the properties that make gambling addictive: you craft your prompt, add context, pull the lever, and sometimes you win a feature. Loss disguised as a win. The illusion of control. Stochastic reward. His wife, a fellow researcher, catalogued these properties in an article. The people who got most enthusiastic about AI coding often found, months later, that almost none of what they built during that period was in production or earning money. This explains a paradox readers keep raising: people use the tools a lot, feel productive, but the organisations aren't seeing the output.

The economics job market fell 31% in a single year

In week 14 of the current hiring season, postings in the economics job market were down 31% versus the same point last year. The explanation from an economist presenting the data: demand for economics undergraduates is being automated away, and the PhD market is coupled to it. Combined with the Harvard data from Edition 4 (skill requirements in AI-exposed occupations falling since ChatGPT's launch) and Anthropic's research showing hiring of 22-to-25-year-olds down 14%, entry-level knowledge work is contracting faster than mainstream commentary acknowledges.

A sufficiently detailed spec is code

Gabriella Gonzalez's argument, circulating widely in technical circles: the fashionable claim that you don't need to write code, just write a good spec and let an agent handle it, collapses under scrutiny. If the spec is detailed and precise enough for an agent to execute reliably, you have written code in everything but name. The hard part of programming, resolving ambiguity, is still your job. This is the slope argument applied to a specific skill: the people who think they've escaped the need to understand what they're building are the ones most likely to produce output nobody can maintain.

The AI task force leader who'd never logged in

At a professional services firm, the person leading the AI task force hadn't used the enterprise AI tool once. When challenged, they said: "I know I should, but I can't make the time." They weren't uninformed or resistant. They understood the stakes. They were simply too busy doing the old job to start learning the new one. The incentive trap in plain sight: the people best placed to model the new behaviour are the ones most rewarded for performing the old behaviour well.

Intercom built a plugin system that closes the loop

Brian Scanlan, Senior Principal Systems Engineer at Intercom, shared a thread this week on the company's internal Claude Code system: 13 plugins, over 100 skills, distributed across the company via JAMF. The standout pattern isn't the scale. It's the feedback loop. A session-end hook automatically classifies skill gaps from every coding session and posts them to Slack with pre-filled GitHub issue URLs. Sessions become gaps, gaps become issues, issues become skills. The most telling detail: the top five users of their read-only production Rails console are not engineers.

The CIO budgeting for AI cleanup

The CIO of a major consulting firm told a peer this week that they're budgeting 18 months to two years from now for AI cleanup. The reasoning: things are being built once but not built to last, corners are being cut on testing, and the people who built the tools will have moved on before the problems surface. It's an unusual thing to plan for. But it's probably the most honest thing I've heard a technology leader say about the current moment.

From franchises to call options

Tyler Cowen, drawing on analysis from Jordi Visser, argues that AI simultaneously lowers barriers to entry while destroying the conditions for sustained dominance. Software moats compress because any sufficiently capitalised team can replicate your product. Durable advantage reconcentrates in physical constraints: infrastructure, energy, materials, regulatory relationships. Equity in this environment becomes less a claim on a stable franchise and more a bet on execution velocity. The implication for anyone evaluating technology investments: the question is no longer "what have they built?" but "how fast can they keep building?"

Two thirds of organisations report AI productivity gains. Only a third are rethinking what they do.

Deloitte's State of AI in the Enterprise 2026 report found 66% of organisations report productivity improvements from AI. Only 34% are pursuing what Deloitte calls "transformative business reimagination." Most organisations are getting faster at what they already do. Fewer than half are asking whether what they do should change. Meanwhile, only 21% have mature governance models for the autonomous agents they're about to deploy.

McKinsey's internal AI chatbot was hacked via textbook SQL injection

McKinsey's internal AI chatbot Lilli, trained on 100 years of the firm's work, was breached via a basic SQL injection. 46.5 million internal chat messages exposed, 728,000 files containing confidential client data, 57,000 user accounts, 22 API endpoints requiring no authentication. The firm that charges for risk expertise left the front door open. If McKinsey can't govern its own AI deployment, what does your internal chatbot look like?

Red Bull didn't simulate the pit stop. They did it in zero gravity.

A reader forwarded an Instagram clip of Red Bull's F1 team performing a tyre change in zero gravity, just to prove they could. Not CGI. Real mechanics, real car, real weightlessness. The reader's take, which I think is exactly right: use AI for the boring, the day-to-day, the basics. Free up your budget and attention for the truly remarkable. "If I'm so focused on the incredible, the groundbreaking, the creative and free from the mundane, I raise the bar for the client." That's the slope argument in a sentence. The people who use AI as a floor-raiser, not a ceiling-replacer, are the ones building capability.

The consulting firms are buying the AI stack, not just using it

CB Insights mapped every AI investment, acquisition, and partnership by the major consulting firms since 2023. Accenture is at the centre of the web, with partnerships radiating to dozens of AI companies. The Big Four and MBB firms aren't waiting to see how AI plays out. They're racing to own the infrastructure: embedding agents via Salesforce, ServiceNow and Workday partnerships, acquiring data companies, and investing in startups that automate the consulting workflow itself. Four patterns emerge: race to own the stack, embedding agents, data as differentiator, and workforce transformation. PwC's announcement this week is one node in a much larger network.

CB Insights map of AI investments, acquisitions and partnerships by consulting firms since 2023

The big consultancies are not just using AI but buying the stack, with Accenture at the centre of a web of deals and acquisitions.

More offices for AI than for humans

US data centre construction spending overtook general office construction in December 2025, according to Census Bureau data. Data centres: $3.57 billion. General offices: $3.49 billion. The lines crossed after data centre spending roughly tripled in two years while office construction flatlined. We're now building more square footage for machines than for people.

Data Center Construction Spending Climbs to Record: outlays for data center projects overtook offices in December 2025

US construction now spends more on data centres than on offices, the first time machines have outbid people for new square footage.

See what readers said ↓
What readers said

What resonated

  • The apprenticeship pipeline, again: the question of what happens to junior roles when AI handles the volume work has now been the most-discussed theme across three editions running. This week it drew the sharpest language yet.
  • The pace of change: a partner at a consulting firm captured a feeling several readers seem to share: trying to "get on a breaking tsunami with a surfboard, and the surfboard keeps being reinvented."
  • The ATMs-to-iPhone distinction: the structural argument (automating within your paradigm vs replacing the paradigm entirely) prompted readers to apply it to their own organisations.
  • The three-tool limit: the BCG "AI brain fry" research resonated, particularly the finding that high performers were the first to be affected.

Points readers raised

"Porsches are stunningly quick and razor-sharp. A skilled driver can make one dance. A bad driver? They'll put it straight into a tree."

A professor of digital transformation wrote an academic paper in two and a half minutes using an AI tool. Was it any good? No. Could it get published in a poor-quality journal with minimal tweaks? Yes. With nearly a hundred papers behind them, they know exactly what to add, what to remove, what's junk. "However a non-expert could do the same and wouldn't see the errors. An AI or non-expert reviewer wouldn't see the obvious error either and would accept it." The result is "lots of AI science slop" across academia, publishing and music.

"Five years from now, the marketplace will offer nothing but blight."

A reader at a professional services firm wrote: "I vacillate between being optimistic that AI will allow employees to contribute more vs. expecting that AI will bring mass layoffs and throw the world into desperation never before experienced." On the apprenticeship pipeline: "I simply cannot get past the shortsightedness of it." This from someone who describes being "fiercely AI curious" and learning in what little spare time there is, which makes the tension all the more real.

"I'm trying to get on a breaking tsunami with a surfboard, and that the surfboard keeps being reinvented while I'm about to step on it."

A partner at a consulting firm had been thinking about a comment I made in our last meeting that started "I wouldn't have said this two months ago, but..." The question: does this slow down at any point, or does ChatGPT just start to feel like yesterday's news forever? I don't have a comforting answer. I think the honest one is that the pace of change isn't going to decrease.

Maintaining team size while expanding capability

An IT director at a consumer brands company has been "advocating for internally as well: maintaining team size while leveraging AI to increase capability rather than running leaner." The argument: if growth is the goal, a team of several people using AI will be far more productive than cutting headcount and expecting one person to carry the load. A practical instinct too: "I've encouraged our team to avoid signing long multi-year contracts right now. The landscape is shifting so quickly that new competitors are appearing constantly."

Substitute or complement? The ATM analogy goes deeper.

A CTO had been working with a simple framework: "AI is a substitute for low-judgement work and a complement for high-judge work." But the ATM article complicated it. The key passage quoted back: "it is paradigm replacement, not task automation, that actually displaces workers." A more nuanced distinction than substitute-versus-complement alone.

The explore-exploit tension in tool choice

A data strategist pushed back on the "two or three tools" advice. "There's an explore/exploit conundrum of humans too but overall I'd say there's too much 'getting comfy with what I know' esp in the context of things getting better all the time." The nuance: "Like you I have settled on CC [Claude Code] but then building tools on top of that. So tool here is an interesting thing to define." One platform with many custom tools on top is different from three unrelated platforms.

"What training or frameworks exist to roll out AI with care?"

The AI lead at a major media company asked the question the essay left open: "I'd be interested in any training or frameworks you're coming across to roll this out organisation-wide with a consistent approach." A single sentence that captures what I'm hearing from senior leaders everywhere right now. The honest answer is that the frameworks are being built in real time, mostly by the organisations brave enough to try.

Hiring juniors only matters if you care about legacy

A colleague argued that investing in the next generation depends on whether leaders care about the company's future beyond their own tenure: "I would imagine hiring and training juniors only matters if you care about people, legacy or the company's future into the next generation. If you don't and just want to earn/sell in your lifetime then I guess they don't care and I'd imagine most don't." And: "I don't want to be a luddite but it does seem like as a civilisation we're not going in the best direction."

Links readers shared

Read the email ↑
Audio edition with AI voice - in testing!
0:00--:--
The power and the care · 8:55 Download or subscribe →
David's email this week
See the extras for this week ↓
The bits that didn't fit

Only a third of the time AI saves actually reaches the team

How time savings from AI are used — Gartner

Only about a third of the time AI saves reaches team outcomes; the rest leaks into extra work or redoing what the model got wrong.

Gartner data shows that of 5.4 hours saved per worker through AI tools, only 1.7 hours (31%) translate into improved team outcomes. The largest single block of recovered time, 1.4 hours, goes into additional work that doesn't improve outcomes. Nearly an hour is spent redoing work the AI got wrong. Two thirds of the productivity gain leaks away before anyone benefits. If your organisation is deploying AI tools without redesigning how teams work, you're capturing barely a third of the value.

Gartner →

Coding is not software engineering. The confusion is expensive.

Jeremy Howard, a deep learning pioneer who uses AI coding tools daily, draws a distinction most executives miss. Coding, translating a specification into syntax, is a style transfer problem that language models handle well. Software engineering, designing abstractions, decomposing problems, building systems that hold together over time, is a fundamentally different skill that models cannot do. Howard cites Fred Brooks's essay from decades ago, which made the same observation about fourth-generation languages: removing the typing bottleneck does not remove the engineering bottleneck. Companies restructuring around the assumption that AI can do software engineering are conflating two things. His sharpest framing: what matters for any person or team isn't their current output (the intercept) but their rate of improvement (the slope). A little bit of slope makes up for a lot of intercept. Organisations pushing AI to maximise today's output may be destroying the growth rate of the people who'll need to maintain the systems tomorrow.

fast.ai →

The shift from co-intelligence to managing AIs

Ethan Mollick, the Wharton professor whose work has appeared here before, argues we've moved from co-intelligence (prompting AI back and forth) to managing AIs (giving agents hours of work and getting results in minutes). His most striking example: a company called StrongDM has two radical rules. "Code must not be written by humans" and "Code must not be reviewed by humans." Each engineer spends roughly $1,000 a day on AI tokens. Coding agents build from human-written roadmaps, testing agents simulate customers, and humans review the finished product but never see the code. Whether or not that model generalises, the direction is clear. The job is shifting from doing the work to directing the things that do it.

One Useful Thing →

AI agents are hiring humans

A platform called RentAHuman has accumulated over 600,000 sign-ups for a marketplace where AI agents autonomously hire human beings to perform tasks machines cannot: delivering physical goods, counting objects in a city, conducting on-the-ground research. The agents browse, post jobs, evaluate candidates and release payment from escrow upon photographic proof of completion. No human intervention on the purchasing side. The gig economy inverted: people as the on-demand labour layer beneath AI clients. Whether that distinction matters to the people taking the jobs is left as an exercise for the reader.

Wired →

The junior hiring cliff, updated

Edition 3 cited a 14% drop in hiring for workers aged 22 to 25 in AI-exposed occupations. The number has worsened. Stanford Digital Economy Lab data, charted by Politico, now shows a 15.7% decline from 2021 to late 2025. The shape of the curve matters as much as the number: employment held roughly flat through 2023, then fell off a cliff in 2024 and kept falling. This isn't a gradual adjustment. It's a structural break that coincides precisely with the period when agentic AI tools became capable enough to substitute for junior analytical work. Companies aren't announcing junior layoffs. They're quietly not posting the roles. The people most affected will never know the job existed.

Young workers see a decrease of nearly 16 percent in employment due to AI — Stanford Digital Economy Lab / Politico

Employment for 22-to-25s in AI-exposed roles has fallen almost 16%, holding flat through 2023 then dropping off a cliff.

Stanford Digital Economy Lab →

China may be skipping the chatbot phase entirely

Just as China leapfrogged credit cards and went straight to mobile payments, it may be bypassing the "AI as chatbot" paradigm altogether. An open-source AI tool called OpenClaw hit 250,000 GitHub stars in sixty days. Baidu has integrated it into its search app, which has 700 million users. Entrepreneurs are charging 500 yuan (roughly $70) to install it on people's home computers. A startup made $28,000 in ten days selling a one-click installer. Computer repair shops are dispatching what they call "installation personnel," described as operating like plumbers. When a piece of software generates enough demand to support a physical installation economy, the adoption curve is real and deep. Western assumptions about how AI gets adopted may not apply everywhere.

MIT Technology Review →

Half of AI code that passes its own tests gets rejected by humans

METR, one of the more rigorous AI evaluation organisations, found that roughly half of code solutions generated by Claude models, solutions that passed automated grading, were subsequently rejected by the actual human project maintainers. Journalist Derek Thompson, reflecting on his own experience using AI coding tools, offered the most useful reframe: AI's real skill is generating plausible candidate solutions that require constant human checking, debugging and rejection. That checking process is effectively its own distinct and skilled job. He compared it to being a casting director working with a promising but unreliable younger actor. Getting the collaboration dynamics right will take a long time to diffuse through the economy, which is grounds for scepticism about predictions of imminent mass displacement.

METR →
See what readers said ↓
What readers said

What resonated

  • The apprenticeship pipeline paradox: the argument that cutting juniors today erodes the senior talent pool of tomorrow was the single thread readers returned to most. Multiple replies engaged with it independently, suggesting it articulates a worry many leaders already carry but haven't named.
  • Extraction versus expansion as a choice: readers responded to the framing as a decision, not a trend. Several said it sharpened conversations they were already having about whether AI headcount savings should be reinvested or banked.
  • Skill reclassification in professional services: the observation that AI has retroactively revealed which tasks were genuinely cognitive and which were merely time-consuming landed hard, particularly among people in consulting and law.
  • The SaaS market pricing shift: the 30% software stock decline and the "build it yourself" examples prompted readers to reconsider their own vendor relationships.

Points readers raised

A senior technology leader at a global professional services firm challenged the SaaS disruption premise. Development costs, they pointed out, are only about 20% of a typical software company's revenue. Sales, marketing and customer success absorb 60%. AI can rewrite code, but it cannot replicate distribution and switching costs. They drew a parallel to offshoring: "Huge appetite. Need for re-invention." The disruption is real but the mechanism is more nuanced than build-versus-buy.

A partner at another professional services firm identified the tension between original thinking and process execution. Developing the foundational insight that makes a project valuable is still human work, they argued, but once that insight exists, AI can scale the execution. Their question: does this shift advantage or disadvantage people who trade on original judgment? "I suspect the answer is that it depends on whether I tool myself up appropriately."

A founder building an AI-native company connected the apprenticeship argument to institutional culture. They started their career as a graduate trainee at a large bank and worry that the next generation won't get the benefit of those early years inside large institutions. They posed a sharp question: will we see geographic or cultural differences in AI adoption, where firms with cultures that already embrace apprenticeship end up moving faster?

A professor setting up a Future of Work institute identified a specific parallel to the extraction-versus-expansion frame. Job applications have lost their friction: candidates send almost infinite applications using AI, and firms screen with AI. Both sides have lost out. The old friction forced applicants to think before choosing. AI removed it entirely rather than redirecting it somewhere useful.

A chief people officer at a global firm read the edition twice: "first over the weekend and then again this morning." The apprenticeship pipeline and organisational change sections spoke directly to the tensions they navigate daily. Sometimes the most valuable signal is that a piece is worth re-reading.

A newsletter author raised a fair editorial challenge: some sections sound more like AI than like me. "If it's your human insight, it jars a bit to read those bits in a voice that's obviously an AI." The same week, an executive I've worked with for years wrote to say that they loved how clearly they could hear my voice and personality in the writing. So one reader thinks there's too much AI and another thinks it sounds exactly like me. Either I've trained the model well or I've always written like a robot. I choose not to investigate further.

Read the email ↑
Audio edition with AI voice - in testing!
0:00--:--
Extraction or expansion · 10:45 Download or subscribe →
David's email this week
See the extras for this week ↓
The bits that didn't fit

The hiring cliff for juniors, in one chart

The essay mentions a 14% drop in hiring for workers aged 22 to 25 in AI-exposed occupations. This chart shows the full time series using a difference-in-differences approach. Junior hires in exposed occupations fell off a cliff after ChatGPT's release in late 2022, while exits held steady. The gap keeps widening. Companies aren't firing juniors. They're just not bringing new ones in.

Hires vs exits for junior workers in AI-exposed occupations

Junior hiring in AI-exposed roles fell off a cliff after ChatGPT while exits held steady; firms are not firing juniors, just not hiring them.

Source →

US tech employment growth has gone negative

Year-on-year tech employment growth turned negative in 2024 and hasn't recovered. The tech sector is shrinking its workforce for the first time since the post-2008 recovery. Combined with the junior hiring data above, a pattern emerges: the contraction is real, it's happening now, and it's concentrated at the entry level.

US tech employment year-on-year change

US tech employment growth has turned negative for the first time since the post-2008 recovery, and the contraction is concentrated at entry level.

The work budget is orders of magnitude larger than the software budget

Julien Bek at Sequoia Capital argues that the next category-defining AI company won’t sell tools to professionals. It will sell completed work directly to buyers. His distinction between “copilots” (AI as a tool for professionals) and “autopilots” (AI delivering the outcome) reframes the entire market. For every dollar spent on software, six are spent on services. The smartest entry point? Replace outsourced work first. The budget already exists, the buyer already accepts external delivery, and there’s no internal team whose jobs are visibly threatened. Once embedded, expand inward.

Source →

You would not believe how many shortcuts everyone else is taking

Ezra Klein wrote a commencement address called “Just Do the Work” about discovering, as a young journalist, that almost nobody was actually reading Congressional Budget Office reports. Documents that are neither complex nor long. By reading what his peers skipped, he got ahead. Not exceptional talent. Just diligence. Economist Paul Novosad adds the contemporary twist: this is “more true than ever now, when more people are shirking and AI lets you do 10x if you try.” The gap between the diligent and the lazy is widening, not narrowing.

Source →
See what readers said ↓
What readers said

What resonated

  • "The hundred small things" as a reframe: the distinction between chasing dramatic AI wins and compounding small daily elevations. Several readers said it gave them language for something they had been struggling to articulate to leadership.
  • The "extra hour" problem: the observation that AI is deployed like an extra hour rather than an extra person struck a chord with people managing teams. Structures absorb the gain before anyone notices it.
  • The junior roles question: the Block layoffs and YC data generated the most emotional responses. Readers connected it to their own organisations' headcount conversations.
  • The senryu competition as metaphor: the retreat-versus-redesign framing resonated, though notably nobody offered examples of successful redesign. The absence may be the point.

Points readers raised

A senior leader at a global professional services firm identified a structural gap in the argument. The hundred small things need a container, not just encouragement. He proposed daily one-hour structured learning blocks rather than hoping people will explore on their own. His deeper point: senior leaders who do not use AI personally have no on-the-ground proof of benefit, so their teams see no credibility signal from above.

An events and entertainment industry exec pushed the junior roles argument to its darkest conclusion. The UK already underinvests in training, preferring overseas hiring. If AI accelerates that trend, there is no junior pipeline to grow seniors from. "That is extremely bad for companies longer term in terms of skills shortages and salary premiums for skilled workers, and even worse for UK plc." A topic that I picked up in today's edition.

A manufacturing executive used the framing to shape two specific conversations: accelerating superuser growth and celebrating a colleague's "let's map your process" approach to adoption stickiness. He is in the middle of major organisational expansion and sees the hundred small things as directly applicable to that work.

A technology leader at a research firm noted a quiet loss that doesn't show up in any headcount data. His analysts used to walk to a colleague's desk when they got stuck on a coding problem. Now they ask AI. The problem gets solved faster. But the conversation that would have happened, the one where a junior person absorbs how a senior person thinks about problems, doesn't happen at all. AI is removing the apprenticeship mechanisms even where the apprentices still exist.

My one proper unsubscribe turned out to be the most advanced person on the list. His world is so far ahead of ours that the weekly email isn't relevant to him. He works on cutting-edge AI implementation (sorry everyone, we're all just fast followers!). His AI communities are discussing running engineers 24/7 with 12-hour agent check-ins, deploying 30+ vibecoded projects from non-engineering teams, making openclaw work across 500+ person organisations, and polyphasic sleep schedules to optimise autonomous agent management. Anyone else even close to these conversations? For sure his world is a useful signal of where ours is heading. I'll stay close for you all!

Polyphasic sleep patterns for AI agent management

A glimpse of the frontier one departing reader lives on: sleep broken into shifts to keep autonomous agents supervised around the clock.

Links readers shared

  • Creativity Can Embrace AI — Nadim's book on how creative industries can work with rather than against language models. Named Amazon book of the year by The New Publishing Standard.
Read the email ↑
Audio edition with AI voice - in testing!
0:00--:--
The hundred small things · 10:20 Download or subscribe →
David's email this week
See the extras for this week ↓
The bits that didn't fit

The marginal cost of arguing is going to zero

UK employment lawyers report workplace grievances that once fit in a single email ballooning into 30-page documents, complete with fabricated legal precedents and citations to laws from the wrong country. (Personnel Today) Creation cost: near zero. Response cost: unchanged. Ministry of Justice figures show new employment tribunal receipts rose 33% year-on-year in the quarter to September. (GOV.UK)

Your prompt is the ceiling

Anthropic's latest Economic Index analysed over a million Claude conversations and found a near-perfect correlation (r > 0.92) between the sophistication of human prompts and the sophistication of AI responses. The more nuanced and structured the input, the more the model rises to meet it. The bottleneck isn't the model. It's the human. Which is, in its own way, reassuring. (Anthropic Research)

One blog post. One hour. Billions gone. Again.

Anthropic published a blog post introducing Claude Code Security on a Friday afternoon. Within an hour, cybersecurity stocks cratered: CrowdStrike fell 8%, Cloudflare 8%, Okta over 9%. The tool itself is a modest research preview. But a single blog post from an AI company erased billions in market value from established incumbents. (Barron's) The same dynamic hit legal tech stocks when Anthropic announced legal plugins for Claude Cowork a couple of weeks earlier. (Sherwood News) That's a new kind of leverage.

The trust signals your organisation depends on are dissolving

Here's a problem that connects directly to those poets in Sakaiminato. A thoughtful email from a director now carries the same weight as an AI-generated memo, because the reader can't tell the difference. The cues that used to signal competence (a well-crafted message, a polished document, a detailed analysis produced under time pressure) are now producible by anyone in minutes. This isn't a quality problem. It's a trust architecture problem. We need to distinguish between "produced this" and "shaped this." The senryu competition couldn't tell the difference. That's why it died.

Read what readers said ↓
What readers said

What resonated

  • The capability-adoption gap: the tension between what AI can demonstrably do and what organisations will permit. This was the thread readers pulled on most. Thirty of 121 replies engaged with it directly, many in operational terms.
  • "Back your misfits": the idea of finding and enabling the ten percent who don't need pushing. Several readers said they were already doing this and the framing validated their approach.
  • Organisational inertia as structural, not cultural: readers recognised the obstacle isn't attitude or skill but governance, process, and risk frameworks. One partner at a strategy consultancy pushed further: speeding up the same process is "premature optimisation."
  • The personal-to-organisational transition: the feeling of being ahead personally but constrained institutionally was widely shared. People described experimenting on weekends, then walking into Monday meetings where nothing has changed.
  • "Package around problems, not platforms": a phrase several readers said entered their working vocabulary within days.

Points readers raised

A director at a media company connected the capability-adoption tension to their own industry. They are formalising a champions network, exactly the "back your misfits" approach, and said the framing helped crystallise what they were already building.

An exec at a broadcaster shared the most striking story: they have been deliberately ignoring their organisation's AI policies to enable their team's experimentation. The weekly email validated an act of institutional defiance they were already committed to.

An exec at a professional services firm offered a substantive counterpoint. Accelerating existing processes without redesigning them, they argued, is premature optimisation. Their real question: how does more information drive quality rather than just volume?

A chief technology officer at a data firm placed themselves at stage seven of an eight-stage AI maturity framework they adapted from Steve Yegge's writing: running ten or more parallel AI agent instances simultaneously.

Read the email ↑
Audio edition with AI voice - in testing!
0:00--:--
The wonder and the weight · 6:43 Download or subscribe →
David's email this week
See the extras for this week ↓
The bits that didn't fit

Apple chose Google over itself

Apple partnered with Google to power its AI features, paying a reported billion dollars a year for Gemini. The world's most valuable technology company looked at its own AI and decided someone else's was better. The strategic question isn't whether to build AI capability. It's which partner to choose. (By the way, the answer I'd recommend is Claude!)

Source →

One blog post. One hour. Billions gone.

Anthropic published a blog post introducing Claude Code Security on Friday. Within an hour, cybersecurity stocks cratered: CrowdStrike fell 8%, Cloudflare 8%, Okta over 9%. The tool itself is a modest research preview. But a single blog post from an AI company erased billions in market value from established incumbents. The same happened to legal tech stocks when it announced legal plugins a couple of weeks ago. That's a new kind of leverage.

Source →

The marginal cost of arguing is going to zero

UK employment lawyers are seeing workplace grievances that once fit in a single email ballooning into 30-page documents, complete with made-up legal precedents and citations to laws from the wrong country. Creation cost: near zero. Response cost: unchanged. New UK employment cases rose 33% in three months.

Source →

Your prompt is the ceiling

Anthropic's latest Economic Index analysed over a million Claude conversations and found a near-perfect correlation (r > 0.92) between prompt sophistication and response quality. Give a vague prompt, get a vague response. Give one rich in nuance and structured constraints, and the model meets you there. The bottleneck isn't the model. It's the human.

Source →

Watch what McKinsey does with its own workforce, not what it advises clients to do

McKinsey calls it "25 squared." The plan: grow client-facing roles by 25% while cutting back-office roles by 25%, using AI to rebalance a $20 billion firm. This isn't productivity improvement. This is structural transformation. Ask yourself if there are parts of your business that are ripe for radical change. McKinsey already has.

Source →
Read the email ↑