David's Saturday AI Thoughts

Look at the grass

WHAT HAPPENED THIS WEEK * The people with the most skin in the AI economy read this week's data both ways at once.: The Bank for International Settlements warned that AI 'exuberance' could end in a long investment bust, a warning about returns (whether the money earns back enough) rather than capability. The same week a Ramp and Revelio Labs study of 21,559 US firms found heavy AI adopters grew employment about 10% and lifted entry-level hiring 12%, which Noah Smith read as AI complementing people; on the same day The Kobeissi Letter reported AI-linked sectors shedding about 11,000 jobs a month and the US labour share of income at its lowest since records began in 1947. Not contradictions to resolve but different populations: the firms mastering the technology and the sectors it disrupts from outside. Two board questions collapse into one: are we the group that masters this, or the group it happens to? * Zapier is killing private direct messages to feed the company's 'shared brain'.: Wade Foster, chief executive of the automation company Zapier, is ending private direct messages, starting with the executive team, and running a public 'transparency leaderboard' that has pushed the share of conversation in open channels from 33% to 46%. His logic: every private message is a gap in the company's shared brain, context lost to both the humans and the AI agents that increasingly work over its records. The bottleneck in most organisations' AI adoption is context that never gets written down, not model quality. Most firms bolt AI onto the way they already work; Zapier is redesigning how it works so the AI can pull its full weight, which is the step separating the leaders from the licence-buyers. * The Economist ran 25 AI models through the World Values Survey and they came back more extreme than any country.: The World Values Survey has mapped human morals across more than 100 countries since 1981. The 25 leading models did not land at the average of humanity; they clustered in the corner occupied by rich, secular, self-expression-focused countries, often further out than the most extreme nation surveyed. OpenAI's models came back more secular than any country, and none reflected the world view of most African or Muslim countries. Where there is no factually correct answer, a model's values fill the gap, so a tool a billion people delegate decisions to carries a world view measurably narrower than the human range. It is the values version of Edition 18's point: unless you bring yourself, the machine's defaults stand in for you. WHAT TO TRY * Give any automation a few reversible dress rehearsals before you trust it for real.: David set up an agent to grab a hard-to-get morning swim slot for Teresa at their local gym, and did not hand it the real booking on day one. For the three days before the one that mattered, it booked a genuine slot the proper way, then immediately cancelled and emailed him the confirmation. His logic: if it works for three days you feel good, and if it does not you have three chances to fix it before it goes live. The rehearsal runs the whole chain against the real system, but in a mode where any failure is free to fix. * Find the one thing your AI reliably gets wrong, then make it your first check.: Any AI tool you use on a repeated task has a consistent blind spot: the transcript that always mangles names, the summary that always drops the final ten minutes, the invoice reader that always leaves the tax field blank. David's was that last one, and catching it recovered £9,860 of VAT that would otherwise have gone unclaimed. The first few times you use a tool, compare its output with the source and note what it consistently misses, then make that one thing your standing first check, so you keep the time-saving without inheriting the blind spot. * Queue what you'd doomscroll past, and have a model teach it back at your level.: Suhail Doshi, the founder behind Mixpanel and Playground, described the habit on X: queue every interesting blog post, paper or tweet, ask a model to teach it back to you, and read the approachable version in spare moments instead of doomscrolling. The model does not just summarise; it teaches, pitched at how much you already know, so a dense paper meets you where you are. The simplest way in: paste a link or the text into Claude or ChatGPT with 'Teach me this. Here's what I already know:' and one honest sentence. Read the full edition with all links and sources [https://steadman.ai/newsletters/david/#edition-2026-07-04]

A thousand small bargains

WHAT HAPPENED THIS WEEK * Claude has stopped opening in a window and started living in the team's Slack.: Anthropic launched Claude Tag: you @-mention it in a Slack channel and it works as a persistent teammate with its own identity, scoped access, memory across weeks, and the run of a stalled task for days. Anthropic says 65% of its internal product code now gets written this way, and Andrej Karpathy called it the third major redesign of how we use these models — website, then app, now persistent worker. It's the first frontier product openly about managing an AI employee, and the human problems (supervision, accountability, headcount maths) arrive with it. * Procter and Gamble actually ran the AI-alone-versus-human-plus-AI test, and human-plus-AI still won.: On stage at the Lions Insight Summit in Cannes, P&G's Chief Analytics, Insights and Media Officer Kirti Singh told David the company let AI alone make some brand-building choices, measured the outcomes against humans working with AI, and the combination won — "at this point in time, AI on its own is not better." This isn't a sceptic hedging: he says his teams under-use AI and pushes them to use more. It's a measured, evidence-backed floor for the human-plus-AI position rather than a slogan. * Norway just banned generative AI for six-to-thirteen-year-olds.: From late August 2026 Norway imposes a near-total ban on generative AI for children aged six to thirteen, with supervised use only for fourteen-to-sixteen-year-olds. The prime minister's argument is that AI lets young children skip the essential steps in learning to read, write and do maths; the same government is funding a return of physical books and banned school smartphones in 2024. Most of the year's AI conversation is about going faster — here a government has decided that for one age group the danger is precisely the speed, the skipping of the hard steps that make a child's judgement worth anything later. WHAT TO TRY * Make the model cite its sources, so a check is one click, not a re-read.: A documentary production team using a Claude-based transcript tool David built insisted every quote come back tagged with its time code: the transcription could be slightly wrong, but the time code never is. When a summary looked off, the editor jumped straight to the original clip rather than trusting the paraphrase. Time codes, page numbers and line references all do the same job — the pointer costs nothing to ask for and lets you own the output on a light read, then jump straight to the source on the one part that has to be right. * Build an AI version of the person before a high-stakes meeting, then mine its blind spots.: Before interviewing GitHub's COO, the writer Mike Taylor (writing in Every) built an AI persona of the COO from everything publicly known and ran his planned questions past it first. The simulation's misses were the useful part: where the persona went thin or vague was exactly where public information ran out, so Taylor spent the real conversation on what no model could already know. Most people use AI to generate questions; this inverts it — ten minutes rehearsing against an AI version of the person tells you which questions aren't worth their time. * Stop asking your inbox vague questions. Label first, then ask one scoped thing.: The AI assistant inside most email gives a vague answer to a vague whole-inbox question. Narrow it: label the threads belonging to one project or client, then ask something specific — "what am I waiting on from Sam about the website", or "the action items under the Clients label that I never replied to." It comes back with who you owe, the questions you left unanswered, and what people asked you to send. Then ask it to turn those into a checklist that drops straight into your week. Read the full edition with all links and sources [https://steadman.ai/newsletters/david/#edition-2026-06-27]

Average by default

WHAT HAPPENED THIS WEEK * A Munich court told Google that "may contain errors" is no defence.: The Munich Regional Court ruled Google liable for false claims in its AI Overview, which wrongly tied two German publishers to fraud, holding that an AI summary generates fresh substantive statements rather than just curating sources, and that the small-print disclaimer does not transfer liability back to the user. Google must remove the answers and pay 80% of costs. Every major provider leans on the same footer to cover confidently wrong answers; a court has now said it does not work, so any organisation publishing AI output to customers, regulators or staff can no longer assume the disclaimer protects them. * Commercially available AI flagged early breast cancer six years before clinical diagnosis.: A Karolinska Institute study in Radiology found off-the-shelf AI systems flagged early warning signs of breast cancer in roughly 20% of patients a full six years before clinical diagnosis at 90% specificity, rising to around 40% at two years out, across 88,963 mammograms from more than 31,000 patients, with three of the tools tested already commercially available. Pattern recognition at scale is what AI does best; here it buys years of warning on a disease where early detection saves lives, and the bottleneck is no longer capability but deployment, trust and who owns the answer. * Anthropic studied 400,000 coding sessions: the tool is levelling people and moving the real gap up to firms.: Across roughly 235,000 people and the ten largest professions, success rates landed within seven points of professional engineers and managers came out ahead; an accountant who had never written Python but knew which rule a month-end reconciliation had to enforce was rated an expert, while a senior engineer on an unfamiliar task was not. What carries a session is understanding the problem, not the craft, so the gap between people is closing. The part everyone skips is that it has not vanished but moved up a level, from between people to between firms. WHAT TO TRY * Tell the model what's at stake before you ask the question.: In a coaching session a senior leader asked Claude about a government-policy question and got a confident, generic, thin answer. Adding one sentence, "The answer is critical. Provide authoritative sources," changed everything: the model switched register, went to the primary regulatory documents, quoted the relevant sections and linked each claim back so the human could check it. Nothing about the question had changed; the model simply hadn't known the stakes. * Stamp every AI output "raw, not yet checked" until you've taken ownership.: Mentoring Ethan, Steadman's placement-year researcher, the rule landed for any AI-created document: it's either raw AI output or something you've checked, edited and will stand behind. Set your tool to auto-stamp every output "raw AI output, not yet checked, edited and owned," and leave it there. Removing the stamp by hand becomes the deliberate act of taking ownership, so anyone who later picks up or is forwarded the file knows exactly which of the two things they're holding. * Book a meeting with yourself, hit transcribe, ramble for fifteen minutes.: A senior advisor said their real value sits as strong points of view in their head, none of it written down. The fix: put a meeting in your own diary, hit transcribe, put your feet on the desk, look out of the window, and talk, without trying to be structured. Change your mind, go down rabbit holes, tell stories. The transcript becomes a thesis document you can feed any model so future answers come back already loaded with your frame, your caveats and your taste. Read the full edition with all links and sources [https://steadman.ai/newsletters/david/#edition-2026-06-20]

Ride the bike

WHAT HAPPENED THIS WEEK * One activist letter froze a board's AI workspace overnight. The doctors are next.: From David's week helping a hundred non-executive directors: one board's AI workspace went dark the morning an activist investor asked for its contents to be made discoverable. The Conference Board's April survey finds two-thirds of directors use AI for board work while barely a quarter of executives call their board highly fluent, and the Medical Protection Society warned that doctors and the NHS could be sued over AI tools' mistakes, with the clinician left as the 'liability sink'; it wants AI reclassified as a product under the Consumer Protection Act 1987 so liability flows to developers. The governance frontier is moving from 'can the data leak' to 'who owns the answer when the model is wrong', and the CEO'd policy (every output checked, edited and owned by a named human) is the only good answer to both the letters and the writs. * Meta built a 'second brain' that 63,000 staff installed in three months. It started with one person.: Meta's analytics team reports that an internal AI tool one of its data scientists started has been installed by 63,000 employees within three months, with no top-down mandate and no transformation programme. One person built something useful and the rest of the company found it. The standing question for readers: what are you doing to encourage and enable this at your firm? * When cheap models do make sense.: Ethan Mollick argues for hierarchies in which smart models supervise cheap ones: the smart one checks the plan, the cheap one does the volume. Right for machine pipelines running thousands of low-stakes calls; for your own judgement work, the essay's time-saved maths says buy the best. The item is the deliberate counterweight to the essay's argument, marking where it does and does not apply. WHAT TO TRY * Ask the AI to write the marking scheme before it writes the answer.: Ben Yoskovitz, a non-developer who ships production software with AI, has the model write three to five specific pass-or-fail checks before any non-trivial task, approves them, and then has the AI do the job, grade itself against each check, cite the evidence and stop the moment one fails. Ninety seconds of discipline against the long loop of 'looks good' followed by 'wait, no it isn't'. * Plan in one session, then start a fresh one to build.: Tiago Forte treats the first AI session as planning-only when the goal is a concrete document or deck: work out the brief in chat one, never let it start drafting, then open chat two, paste the brief and build. A clean start gives a sharper model, the deliberate stop tests whether the brief holds together, and a crash in session two still leaves you with the brief. * Keep your AI project lean.: Someone at a training David ran asked why their Claude project had slowed over a few weeks; they had been piling reference files into it. The model reads everything in a project on every request, so book fifteen minutes on a Friday to prune, and never leave fat PDFs in there: have the model convert each one to plain text first, since PDFs burn far more reading capacity than the same content as text. Read the full edition with all links and sources [https://steadman.ai/newsletters/david/#edition-2026-06-13]

The open door

WHAT HAPPENED THIS WEEK * The CEO of a 350,000-person IT services firm says AI is hollowing out the middle, not the bottom.: Ravi Kumar, chief executive of Cognizant, told Fortune's COO Summit on 1st June that his company hired 20,000 entry-level graduates last year and expects to hire more in 2026, with new 'Frontier Business Operator' and 'Frontier Certified Engineer' roles defining AI-era work. He called job-extinction talk 'fearmongering' and argued AI thins middle management while entry-level and leadership roles persist. It's a direct counter to the consensus that entry-level work vanishes first — including the US Bureau of Labor Statistics data Edition 14 leaned on — and a real-world data point for the essay's bet on graduates. * The machine is writing the code now, and the gains are pooling at the top.: Tobi Lütke says one in eight pull requests merged at Shopify are now written by River, its in-house agent, not an engineer. Anthropic's own engineers ship roughly eight times the code per person they did before 2025. Cursor's developer report shows the output gap widening, with top developers pulling far ahead of the median. And OpenAI's Codex has passed five million weekly users, with non-developer adoption growing three times faster than developer adoption. The grunt of writing code is moving to the machine, the output is multiplying, and the reward is concentrating in the people who know what to ask of it. * Capability is outrunning even the best forecasters.: The Forecasting Research Institute asked expert forecasters and superforecasters how long a task a model would reliably finish by the end of 2026, measured on METR's time-horizon benchmark (about an hour and a half when the survey launched). All three groups put the end-of-2026 figure between three and four hours. Then, while the survey was still running, a frontier model in preview hit three hours and six minutes — already inside the range they'd picked for year-end. The forecast was overtaken before they'd finished making it. Cloudflare's Matthew Prince made the same public miss: bots passed humans in web traffic for the first time, years ahead of his own late-2027 estimate, though much of that is scraping rather than agents, so the figure is softer than it sounds. WHAT TO TRY * Ask the AI to orient itself in your folder before you ask it anything else.: Setting up a Claude Cowork project for a documentary filmmaker new to the tool this week, David's first prompt wasn't about the work. It was: 'Read the files and sub-folders, write yourself a little set of instructions for future chats. Write yourself a navigation guide.' Claude explored, then saved itself three memory files — a project overview, a folder guide, and a profile of the user. Every later chat in that project started smarter because it could re-read its own notes. * Don't say 'always allow' on the verbs you can't undo.: In a coaching conversation about agent permissions this week, the familiar pattern: three prompts in, most people click 'always allow' on everything just to stop the interruption. The discipline is to sort the verbs first. Reading, listing, searching — leave on auto. Deleting, sending, posting, spending — keep asking every time. Sorting the verbs first is what keeps one careless click from emptying a folder, sending an email you can't unsend, or running up a charge you didn't mean. * Talk longer on the call, so the AI can work for hours after.: On a client call this week the brief changed mid-conversation. David deliberately used more words than he otherwise would, narrating the bridge between the old direction and the new one. The over-explaining was really for the AI: context he was laying down for it to pick up later, once the transcript was in the project. The transcript plus 'go' did the work. Next time you're on a call that's being transcribed, elaborate a little more and narrate the why, not just the what. Read the full edition with all links and sources [https://steadman.ai/newsletters/david/#edition-2026-06-06]