David's Saturday AI Thoughts
Each Saturday, David reflects on what feels important in the world of AI. Not the breathless hype or the doom. The practical, analytical perspective: what happened this week, what it means for people who use language models in their work, and what to try next.
A spoken-word audio edition of David's weekly email, with different voices for each section.
Testing. I'm experimenting with an AI text-to-speech version of the weekly email. Different AI voices read different sections: one for the essay, another for the news, another for tips, another for reader reactions. I'd love your feedback. Is the quality good enough to listen to each week? Reply to any Saturday edition, or email david@steadman.ai.
Episodes
Where does it live?
Someone opened a coaching session by apologising: nothing was properly set up. They then showed David a system of automated checks they had already built, kept in one account only. A marketing team's branded decks turned out to run off two plain text files on one laptop. That is this week's finding: the best AI practice in most firms already exists, hidden in someone's head or folder. David's job has quietly changed, from teaching to finding what one person built, improving it, then giving it a home that outlives a laptop.
The crossing
Twenty-five Fields Medallists complained this week that the machines can solve the problems, but the problems were set to make mathematicians. David takes the point into firms: judgement used to come free with the work, and a twenty-minute turnaround throws out the checking that built it. His son, his placement student and a market research leader show the crossing can be made without rowing the whole way, if somebody pays for the checking. Also this week: the journal whose authors couldn't explain their own papers, Deloitte's billable-hour chart, and fifteen bits that didn't fit.
We're all going to die
Two AI safety researchers put public numbers on human extinction this week, more than ten per cent from one and around fifty from the other, and David takes them seriously without letting them stop him working. On the same Wednesday morning one of those threads landed, he set an agent counting every lido in Britain, and another one found him $286.41 of unclaimed property and filled in the claim forms. That is the position he argues for: be alarmed by the big picture, back the fight worth having over slowing down, and still get good at the tools, because the list of tasks where doing it the slow way is respectable got shorter again this week. Plus three things worth knowing, including what OpenAI cannot rule out about its customers' work, three things to try, and what readers said.
The vibe shift
David names a vibe shift in how serious people talk about AI at work: from buying licences, running hackathons and hoping, to starting with the profit-and-loss account, deciding what has to change, and only then pointing AI at it. Private equity got there first. The hard part is not wiring AI through a process but changing roles, teams and what people are there to do.
Fewer, bigger, better
The jobs most exposed to these tools are working about three hours a week longer since ChatGPT launched, and the leisure that goes is the going out. That is the essay: the hour AI saves rarely comes back as an evening, the market has repriced the evenings that survive, and the person with the standing to keep the gain is usually the one reading. Also this week: a judge's order nobody checked set against an investor's owned AI column, Anthropic letting outside researchers study 250,000 conversations without reading one, and the firms taking the most desk work off junior consultants calling them back into the office.
One inch
An instructor moved David's hip about an inch in a hot yoga class and turned an easy pose very hard. That is the essay: there is no advanced room to graduate into, at yoga or with these tools, so the question four leaders asked him in one week, am I doing this well enough, cannot be answered by anyone, including the people who are ahead. Also this week: Google buying a dead airline's entire working memory for about twenty dollars a worker, the market starting to discount the phrase because of AI, and a mathematician leaving the field because proofs got cheap and trust did not.
Go talk to them
The most common complaint I hear from senior leaders about AI isn't about the technology. It's about what their own teams send them. "That's just AI slop" has quietly become a way of not having a conversation: it sounds like a verdict on the tech, and it's frequently a verdict on a person nobody will say it to. So go and talk to them, and criticise the content rather than the tool.
My newest colleague
At half past six on New Year's morning, a colleague messaged me: the overnight job sorting my photos had jammed, worked out why, fixed it, and wrote itself a rule so it wouldn't happen twice. The colleague was an AI. The third generation of this technology isn't a tool you use. It's a colleague you manage, with a memory, a handbook and standing responsibilities of its own.
Smooth enough
We have been handed something close to cheap superhuman intelligence, and the strange part is how often organisations carry on much as before. The machines can smooth the road; deciding where to go is still ours. The essay follows the whole curve, from a trail in Connecticut to the halvings at EMI, and asks what would genuinely change if the work behind your next big decision were twice as good.
Burned and earned
A reader asked me what AI costs the planet. My first instinct was to reassure: one prompt is a quarter of a watt-hour and about five drops of water. Tiny. Then I realised that was only half the story. The cost of using AI is the cost of the machine minus the cost of whatever would otherwise have done the work. Count what it earned, weigh what it burned.
Futures and fears
The bleakest reply in my postbag came from a chief executive who forecast job disruption far bigger than the numbers admit, then added at the foot of the same email that he finds the technology enormously exciting and has given up trying to reconcile the two. Fear of this technology and excitement about it aren't competing conclusions. They're both smart readings of the same facts, often held by the same person.
Sarah stays
I couldn't sleep on a hot London night, so when this arrived nearly whole I got up and wrote it down. Sarah resigns on a Tuesday. Her firm keeps her anyway, rebuilt from three years of her emails, meeting transcripts and prompt history. Her handover is what she thought mattered. The archive holds what actually did. Every step is possible today.
Look at the grass
Two photographs of the same Wimbledon court, forty-two years apart. In 1982 the grass is worn up the middle, the path of serve-and-volley. By 2024 the forecourt is green and the bare patch runs along the baseline. Nobody announced it: the game became a rally, so the wear moved. Knowledge work has made the same move in four years. The machine has not removed the work. It has moved the wear.
A thousand small bargains
I sent an important email last week that a machine wrote, read it, changed nothing, and sent it. Taking Rahim Hirji's new book SuperSkills as a generous foil, I argue most AI handovers are good bargains, not a thousand small surrenders: the machine takes the middle, the parts that decide the outcome climb a level to you. Never be careless. Always be good. Sometimes be great.
Average by default
I have helped hundreds of senior people set up their AI, and almost none had told it who they were. Not the job title, the easy part, but how they think and what good looks like to them. The personalisation box sits empty account after account. If you don't tell the model what makes your judgement specific, it has one assumption left. You're average.
Ride the bike
Anthropic's newest model lists at double its predecessor, and the bill has become the story. I did the maths: the gap between the cheapest sensible model and the dearest buys about 40 seconds of a manager's day. So manage the judgement, not the number: a five-prompt floor, delegation budgets run like expenses, and Eddy Merckx's advice. Ride as much or as little as you feel. But ride.
The open door
For years I told people to hire graduates and never did it myself. This week I did. Ethan is here on his placement year, and he is the experiment. I have four hypotheses about why a graduate is still worth it when the machine can do the work, and a year to test them. The fourth is Hamming's open door: leaving it open costs you time, but it lets the useful things in.
How We Got Here
I pulled a book off the shelf and found a note I'd written in the margin in October: "Claude Skills does this. Launched yesterday." The passage next to it said language models had no fast-learning memory layer. In AI the future shows up long before it ships. Seeing it was never the hard part. Believing it enough to bet on it was.
Kids these days
Elliott walked off the kitchen table to use ChatGPT instead of me. He came back having mastered the method I've been teaching professionals for three and a half years: use AI, check the answer, understand why it's right, own the result. And he aced the test in class the next day.
What boards accept
Boards drive change by addition: a new strategy, a new programme, a new dashboard. In AI, the loop is too damn slow. Subtraction is the alternative. Refuse what the firm has been accepting. Ten refusals worth making, and the case for picking the one you've been avoiding longest.
Choosing is the work
The cost of building anything you can describe has fallen close to zero. So choosing well — what to build, who uses it, where the saved hour goes — is now the work. Not choosing is failure. Everything else is just typing.
The bill and the harness
Flat-rate AI pricing is breaking. Bills will rise either way. The choice is whether the spending compounds for you, in the harness, or for the vendor in raw model use.
Rise of the auditors
AI-native teams need three roles: Director, Builder, Auditor. Execution is cheap, verification is expensive. Most organisations have zero Auditors.
The proxy break
AI broke the old proxy (good writing = good thinking) but the new proxy ('sounds like AI' = no thinking) is equally unreliable. Evaluate thinking, not wording.
What a day can do
Team-level AI infrastructure can precede individual training. A small jewellery company built thirteen shared skills in a day. Step two can contain step one.
What is your organisation actually for?
Organisations say they're production systems but behave like human systems. The revealed preference is togetherness, not efficiency.
The system and the surrender (plz fix!)
A Wharton study of 1,372 people identified 'cognitive surrender': when AI produces an answer, people stop questioning it. The better the system gets, the harder it becomes to stay vigilant.
Reckoning and slope
The gap between AI wonder and behaviour change. Jeremy Howard's slope-over-intercept frame: capability growth matters more than current output.
The power and the care
The dual experience of AI acceleration: excitement and terror. Top builders 3-5x more productive, median only 10-20%. The gap widens.
Extraction or expansion
Leaders deploying AI face a binary choice: extraction (cut costs) or expansion (grow capability). The apprenticeship pipeline paradox.
The hundred small things
AI value sits not in dramatic one-off wins but in a hundred small daily elevations that compound into transformation. Firms can't see it because they track big projects.
The wonder and the weight
The tension between individual excitement about AI and organisational inertia. 84% of the world has never used AI. The wonder of what's possible, and the weight of spreading change.