David's Saturday AI Thoughts

Rise of the auditors

Sat, 25 Apr 2026 09:00:00 +0100

AI-native teams need three roles: Director, Builder, Auditor. Execution is cheap, verification is expensive. Most organisations have zero Auditors and are shipping nothing because nobody is named to check.

What happened this week

<10% of organisations scale agents beyond pilots (McKinsey)
GitHub paused Copilot signups / Uber blew 2026 AI budget / Goldman inference costs approaching headcount parity
29% of employees sabotaging AI initiatives (Writer survey)

What to try

Ask is this the simplest version (Cantrill laziness)
Audit cold: different model, fresh context
Save one reusable AI workflow (Chrome Skills)

Read the full edition with all links and sources

The proxy break

Sat, 18 Apr 2026 09:00:00 +0100

AI broke the old proxy (good writing = good thinking) but the new proxy ('sounds like AI' = no thinking) is equally unreliable. A friend's challenge prompted a deeper question: evaluate thinking, not wording. Two tests proposed: Quality (does the argument hold under pressure?) and Ownership (CEO principle). Fine-tuned models now preferred over human writing 62% of the time.

What happened this week

AI cover letters killed the signal: Freelancer.com study shows Goodhart's Law in action, better letters no longer predict better hires
Snap cut 1,000 jobs (16% workforce), AI writes 65% of new code, $500M annualised savings: substitution model has arrived
Passive AI delegation erodes confidence, pushing back strengthens it: 2,000-person study. Gartner: of 5.4hr saved, only 0.6hr reduces working time

What to try

Ask what keeps people awake at night, not how AI can help: surfaces real problems with AI solutions
Let the model research you before writing custom instructions: web search + self-portrait generates better instructions than manual writing
Find where your AI value sits: AI Value Map interactive tool, five questions on value allocation then five on capture

Read the full edition with all links and sources

What a day can do

Sat, 11 Apr 2026 09:00:00 +0100

Team-level AI infrastructure can precede and contain individual training. The cost of encoding how a team works into shared reusable tools just dropped from hours to minutes with Gen 2 tools (Claude Code + transcripts). A small jewellery company built thirteen shared skills in a day. Step two doesn't just follow step one, it can contain it.

What happened this week

Claude Code now writes 4% of all GitHub commits, doubled in six weeks; Anthropic run rate $30B (up from $9B at end of 2025), Claude Code alone $2.5B; projected 20% of commits by December
Goldman Sachs quantified AI's net labour market drag: -25k jobs substituted + 9k augmented = 16k net monthly loss; entry-level-to-experienced wage gap widened 3.3pp. But CFO surveys put genuine AI impact at just 0.4%, while Challenger reports 15,341 job cuts blamed on AI in March alone
Meta's internal tokenmaxxing leaderboard: 85k+ employees, 60T tokens in one month, Zuckerberg not in top 250. Rewards orchestration over outcomes. Incentivise use yes, incentivise maxxing no

What to try

Start with critique, not creation. Brand voice evaluator was diagnosis-only; teams fear proofreaders less than replacements. Nobody fights the spellchecker
Ask what keeps people up at night, not what they want AI to do. First question surveys existing habits; second surfaces unmet needs. Almost nothing appears on both lists
Show your team how others use AI. 515-startup field experiment: case studies alone led to 44% more AI usage, 1.9x revenue, 39% less capital needed ('the mapping problem')

Read the full edition with all links and sources

What is your organisation actually for?

Sat, 04 Apr 2026 09:00:00 +0100

Organisations say they're production systems but behave like human systems. The revealed preference is togetherness, not efficiency. AI adoption reverts because training optimises for individual productivity while the real binding force is human collaboration.

What happened this week

Dorsey wants to replace org charts with AI world models (Block restructuring)
Mollick says de-weirding AI is a mistake; hidden AI use is the harder problem
Zapier raised the AI fluency hiring bar: slope not snapshot, accountability added

What to try

Have AI interview you before building anything
Ask AI what looks weird before analysing data
Let AI be the app: build skills not standalone software

Read the full edition with all links and sources

The system and the surrender (plz fix!)

Sat, 28 Mar 2026 09:00:00 +0000

A Wharton study of 1,372 people identified 'cognitive surrender': when AI produces an answer, people stop questioning it while recoding it as their own judgment. Accuracy drops from 45.8% alone to 31.5% with incorrect AI. The better the system gets, the harder it becomes to stay vigilant inside it.

What happened this week

Three CEOs (Coca-Cola Quincey, Walmart McMillon, Adobe Narayen) stepped down in one quarter citing AI transformation pressure; 38 years of tenure in one turnover, most since 1999
Anthropic 5th Economic Index + HBR 2,500-employee study: experienced users (6+ months) treat AI as thinking partner, not productivity shortcut. AI may be skill-biased tech that compounds existing advantages
Ethan Mollick: companies with zero AI failures aren't being ambitious enough. R&D-style experimental budgets need to reach HR, operations, finance

What to try

Don't fact-check AI in the same conversation: model defends its own chain. Start fresh, upload source materials cold for critique
Give your AI reviewer a persona with skin in the game: six senior-partner personas converged on the same systematic error a neutral reviewer missed
After every good session, turn it into a reusable skill: capture what 'good' looks like the moment you've achieved it, before memory fades

Read the full edition with all links and sources

Reckoning and slope

Sat, 21 Mar 2026 09:00:00 +0000

The gap between AI wonder and behaviour change. PwC's CEO says get with it or get out, but the real question is get with what. Jeremy Howard's slope-over-intercept frame: capability growth matters more than current output. Anthropic's research shows most coding tool users enter autopilot. Not all inefficiency is waste: some friction is load-bearing.

What happened this week

Centaur chess inverted: in 2005 amateurs with laptops beat grandmasters, by 2026 adding a human to an engine makes it worse. Carlsen deliberately limits AI prep to maintain self-generated understanding
Jensen Huang's benchmark: $500K engineer should consume $250K in AI tokens. But token spend is an intercept metric that says nothing about slope
RentAHuman: 600,000 sign-ups to a platform where AI agents hire humans for physical tasks. Inverts the usual displacement narrative

What to try

Mine your own email archive: pull months of correspondence on a topic, ask AI to synthesise the intellectual arc
End every AI session the way a developer commits code: one line documenting context for the next session
Use AI to teach you, not just to do things for you: Bloom's two-sigma tutoring now costs a subscription

Read the full edition with all links and sources

The power and the care

Sat, 14 Mar 2026 09:00:00 +0000

The dual experience of AI acceleration: excitement and terror, said in that order. Top builders 3-5x more productive, median only 10-20%. The gap widens. Amazon outages show what happens when power outpaces care. Skill requirements dropping in AI-exposed jobs. The organisations that will matter are those building in care alongside speed.

What happened this week

Shopify CEO ran AI optimisation against 20-year-old Liquid engine: 53% faster parse/render, 61% fewer allocations. Legacy code isn't too embedded to improve, it's too embedded not to try
BCG/UC Riverside research (1,488 workers): AI productivity peaks at 3 tools then collapses. 'AI brain fry' causes 33% more decision fatigue, 39% more major mistakes. High performers hit first
ATMs didn't kill bank tellers, the iPhone did. Automating tasks within current structure creates adjacent roles. Redesigning the structure from scratch eliminates them (David Oks, a16z)

What to try

Simulate the toughest reader: create AI versions of board members from known priorities and past questions, run every document past them before a human sees it
Delete the headline, ask AI what it should be: if the model's headline differs from what the colleague wrote, it reveals a disconnect between message and evidence
Fix the instructions, not just the output: update custom instructions, project briefs, context files after every miss. Output matters today, instructions compound forever

Read the full edition with all links and sources

Extraction or expansion

Sat, 07 Mar 2026 09:00:00 +0000

Leaders deploying AI face a binary choice: extraction (cut costs from existing operations) or expansion (grow what the organisation is capable of). The apprenticeship pipeline paradox: if juniors never do the grunt work AI now handles, how do they develop the judgment that makes seniors valuable? Hiring as deliberate investment, not necessity.

What happened this week

Software stocks down ~30% since October while broader tech flat (Salesforce, Adobe, ServiceNow each down 25-30%). Market pricing in structural shift as AI enables smaller tools to replace enterprise suites
Goldman Sachs 'AI-nxiety': 70% of S&P 500 discussed AI on earnings calls, only 1% quantified impact. Median reported gain 30%, concentrated in customer support and software development only
AI retroactively reclassified legal work: general-purpose Claude outperforms expensive legal AI products. Tasks billed at premium rates for decades revealed as procedural, not cognitive

What to try

Run AI and human on same task, focus on disagreements: 78% overlap in one test, but value was at the edges where only one method surfaced findings
Map your org on Gen 1 vs Gen 2 AI tools landscape: most orgs stuck on constrained defaults, need parallel strategies for the many (move to competent use) and the best (accelerate with agentic tools)

Read the full edition with all links and sources

The hundred small things

Sat, 28 Feb 2026 09:00:00 +0000

AI value sits not in dramatic one-off wins but in a hundred small daily elevations (better meeting prep, cleaner drafts, faster document scanning) that compound into transformation. Firms can't see it because they track big projects, not distributed micro-gains. The Japanese senryu poetry cancellation illustrates the risk: AI alone produces sameness, AI plus human steering produces something better.

What happened this week

Citrini Research fiction set in 2028: when AI makes expertise cheap, clients stop buying it. 'A lot of what people called relationships was simply friction with a friendly face'
Block cut from 10,000 to under 6,000 people while growing gross profit 24%. Y Combinator founders planning to eliminate all engineers below senior level. Market valued each eliminated Block role at ~$1.5M enterprise value
Google Gemini 3.1 Pro scored 77.1% on ARC-AGI-2 (double predecessor in 3 months). Model choice matters less; application layer and workflow matter more

What to try

Run your day through AI at 6pm: daily five-paragraph reflection from meeting notes, voice recordings, emails. Pattern recognition compounds over weeks
Before building training, check the settings menu: ask five people which model they use and whether they've changed settings. Answers reveal maturity better than any survey
Push record, think aloud, send to AI: skip the blank page for case studies, knowledge sharing, personal reflection

Read the full edition with all links and sources

The wonder and the weight

Sat, 21 Feb 2026 09:00:00 +0000

The tension between individual excitement about AI (senior leaders working weekends, doing a week's work in ten minutes) and organisational inertia (unchanged meeting cadences, team structures, adoption rates). 84% of the world has never used AI.

What happened this week

Technical barrier gone, domain expertise matters now: Wharton MBA students built companies in four days with AI tools, non-technical students went furthest
Apple chose Google over itself for AI features, paying a billion a year for Gemini: build vs buy resolved in favour of partnerships
McKinsey '25 squared': grow client-facing 25%, cut back-office 25% using AI to rebalance a $20B firm

What to try

Find your misfits: identify the one-in-ten AI enthusiasts, free them from reporting lines, give them mandate beyond old job description
Change the cadence: if AI makes execution faster, review more often (daily not every two days). Review cycle is now the bottleneck
Package around problems not platforms: 'this is your sales coach' drives adoption, 'look at our new AI tool' doesn't

Read the full edition with all links and sources