David's Saturday AI Thoughts

David's Saturday AI Thoughts

By David BoyleTechnologyTech News
Download on the App Store

David's Saturday AI Thoughts episodes

  • Same AI, two experiences
    What happened this week
    • The median firm spends $16 a head a month on AI, but the top one per cent spends 600 times that: A November index finds the top one per cent of firms spending a median $10,200 per employee per month, the top tenth $801 and the median firm $16, while the average cost of using these models fell 21 per cent in a month. One investor describes 'swarm riders', people who hand work to large groups of AI agents and use 1,000 to 10,000 times as much AI as a typical user. The question for readers is who in their firm works like that, and what the firm is learning from them.
    • One bank spent AI's saved hours on more deals; another firm on customers who couldn't afford it before: A VP at a boutique investment bank says every task got shorter but analysts' hours did not fall at all, because seniors put each analyst on more deals, managing directors chase more mandates and directors invent new work; one reply noted that analysts didn't get 40-hour weeks when Excel and PowerPoint arrived either. A business-to-business firm used an agent to cut onboarding each new customer from 100 hours to about 30 and spent the saving chasing customers who couldn't afford its services before. The question is whether you are being deliberate about where the savings go.
    • Have we been labelling the wrong type of writing?: David has long argued that AI output should carry a disclaimer until a human has checked, edited and owned it, and now wonders if that is the wrong way round. Ethan Mollick argued there is 'way too much concern' about checking everyday AI work for factual errors, citing data that recent models beat human experts, PhD specialists included, at finding facts. A Harvard economist replied that human writing now needs the label 'Written by a fallible human, may contain mistakes'.
    • What to try
      • Give your next meeting a second chair (David): In a client meeting a colleague was leading, Claude read the live transcript every minute and passed David notes: context the deck didn't have, the next question coming, when to chip in and the loose ends. Try it in a meeting someone else is leading by giving your AI the deck and the live transcript and asking for a few notes an hour, only when something is worth saying; Anthropic has since built a feature that does this and more.
      • Pass on the prompt, not your set-up (David): The reader in the essay handed their team the prompt that wrote their instructions, not a copy of the instructions. Paste in: 'Interview me about my work, one question at a time: what I do, how I like to work and what good looks like to me. Stop after about ten questions. Then write instructions I can paste into your settings so you work the way I do.' Put the result in your settings, then send the prompt, not your result, to one colleague.
      • Tell it the work is its own: Anthropic tested one sentence telling its newest model that the code it was working on was its own, which gave Claude permission to make more radical changes. On your next document or project, tell your AI 'This is yours now', then ask it to leave the work in a state it would be proud to hand over and to list what it changed.
      • Read the full edition with all links and sources

        13 min
      • Lazy, or leadership?
        What happened this week
        • Anthropic's models are improving about twice as fast as before, and AI is doing most of the speeding up: METR, an independent AI evaluation non-profit, spent 11 weeks examining the lab with access to its internal data and reckons it improves its models on well-specified tasks at around twice the pace of March 2024 to October 2025, with perhaps 80 per cent of the speed-up coming from AI used in AI research. The useful caveat: it may hold only for tasks a machine can grade, so which of your team's tasks have answers a machine could check?
        • A&O Shearman is paying for its AI push mostly in lawyers' hours: The global law firm will put $1.1bn into AI over five years, most of it staff time: every practice must spend at least 3 per cent of its billable hours building AI tools for its own area of law. Ethan Mollick sees two camps in every adopting firm, one treating AI as existential and one wanting it seen as a regular technology so nothing has to change; A&O has picked a side, and the question is which you have.
        • Language models found possible loopholes in German law, the way hackers find bugs in code: A law doctoral student and a co-author used language models to find regulations whose cited legal basis does not cover them, which would make them void, and published three with possible exploits. Courts absorb most such exploits through interpretation, but cheap legal reasoning at scale could put legal systems at risk, and your terms and conditions, pricing rules and promotions are written in the same kind of language.
        • What to try
          • Take the parts of my chief of staff that suit you (David): Someone who already had a chief of staff had their assistant compare it with David's and took the five pieces that filled real gaps; David shares a rough v0.1 with the prompts and caveats, to paste in with the instruction to compare with how you work and change nothing until told.
          • Give an uncertain colleague a safe place to start (Ginny Lavers): Check first what limits your AI use (company policy, a board's rules, a charity's duties, your own no-goes), set boundaries, and test in a familiar tool before handing anything new your keys; it came up with board directors, a charity trusteeship and a new personal AI agent called Instinct, and her free guides take about ten minutes to set up.
          • Rewrite your rules as the reasons behind them: A client's voice assistant followed a long list of clear rules badly; a Steadman colleague rewrote the same rules as guiding principles and within a couple of days the client said it did what they had asked for for months.
          • Read the full edition with all links and sources

            13 min
          • I'm calling it
            What happened this week
            • Graduate jobs at Britain's big employers are harder to get, because applications have flooded in: Ethan Teague's Steadman report: 141 applications per graduate vacancy against 86 in 2022/23; AI's clearest effect so far is on applying, so fix screening before cutting intake.
            • Most big firms now pay for Claude while a few staff in each now do the work: Anthropic's listing filing: over half the Fortune 500 pay at least $100,000 a year, yet under 0.1 per cent of their costs; in one huge company a tenth of staff produce nine-tenths of the gains.
            • A 95-year-old law firm is closing, and AI needn't do a partner's job to break a partnership: Weil Gotshal & Manges is closing; a few rainmakers carry a firm of fixed costs, and AI lets a star partner serve clients without a pyramid of associates.
            • What to try
              • Put the human before the loop: Before handing an assistant an analysis, write down what you would measure and what would change your mind, then compare its calls with yours; Anthropic's trial found analysts followed the agent's lean 86% of the time.
              • Mark up the draft in three colours (from Elliott, David's younger son): Highlight a PDF in three colours (expand this, I didn't understand this, keep this) and ask for a second version, and future ones, that follow the colours.
              • Make the task bigger, then push once more: Following Ethan Mollick, keep asking how to make the task more ambitious until you hit a wall.
              • Read the full edition with all links and sources

                12 min
              • The dog and the squirrel
                What happened this week
                • We'll soon be able to listen to plants: Electric Plant, a start-up, records plants' electrical signals with a pair of electrodes and has trained a model on millions of hours of recordings across a hundred species. Its co-founder says it flags drought stress days before the leaves show it. With whales, AI has found what may be an alphabet in sperm whale clicks, though the researchers say that is still far from cracking their language.
                • Anthropic's own agents now work six hours for every one its staff do, and people's attention is the bottleneck: Anthropic had its economists study the company itself. Agents' hours passed employees' in March and now outnumber them six to one, and interactive Claude sessions sit idle around 80 per cent of the time waiting for a person to respond. As one engineer put it, people are bottlenecking the agents. Where is your team's attention the bottleneck?
                • AI reviews of 124 social science papers beat well-paid human experts' on every measure: Two Harvard researchers took 124 papers already reviewed by well-paid experts and had Claude review each one too. Three former editors of major journals, judging blind, preferred the AI's review 87 per cent of the time, and a separate AI check found the AI's reviews were wrong about the papers a third as often as the humans'. Journals now have to decide what they want a reviewer for.
                • What to try
                  • Read it together (from Ginny Lavers, who works with David at Steadman): when a long document lands, don't hand it over for a summary.: Open it with your assistant alongside. Read it yourself first, saying or typing your reactions in the chat, then ask for its notes on each section and compare: what did it spot that you missed, and where do you disagree?
                  • Split the email in two (David): write two letters in one envelope.: With colleagues whose Claude is set up well, David puts a few plain lines at the top on what they need to decide and know, then a heading, For your Claude, and sometimes a whole pile of instructions to move the project on: how we do the analysis, which method to use, what to try next.
                  • Ask for the second memo, the one against.: Blackstone has built an AI chief investment officer trained on more than 14 years of its own deal memos, and nearly every deal up for review comes with two memos, the deal team's and the AI's. Before your next recommendation goes up, give your assistant your papers on similar calls and the one you are about to send, and ask for the strongest fair case against it.
                  • Read the full edition with all links and sources

                    11 min
                  • Stretching an hour
                    What happened this week
                    • Law firms want to sell their AI like software while billing as if the work took as long: Leaked emails show the global managing partner of Freshfields planning to carve out its AI unit into a company outside investors could buy, citing legal-tech valuations of 25 to 40 times revenue; the emails reached the whole firm by accident and the talks are at a very early stage. Meanwhile a partner at a leading law firm says half his work is now fixed fee but the firm still makes him log hours, and an associate who is one of the firm's heaviest AI users was told he was working too fast. If you buy professional services, ask where the benefits show up for you.
                    • If a rival can buy data like yours from a dying start-up, your data isn't a moat: An independent software maker says data brokers are paying struggling software firms $100,000 to $2m for their datasets and reselling them to AI labs, which then build what the software did into their own products. SimpleClosure, which winds down start-ups, says 47% of its deals now include selling code or data. Data a rival can license from the start-up next door isn't a defence, so check what your contracts let a vendor do with your data, and what happens to it if that vendor fails or is sold.
                    • Anthropic's co-founder now thinks models that build better models are more likely to arrive in 2027 than 2028: Jack Clark has moved his timeline for full recursive self-improvement, AI improving AI, up a year, to 2027. A new test of research taste has the best model, Anthropic's Opus 5.5, beating its human expert baseline at about a thirtieth of the cost, and a finance professor who spent a couple of days at Anthropic came away saying 'We are close to RSI'. David adds that changes in how he uses Claude have brought massive gains in efficiency and effectiveness in the last couple of weeks, and that he believes radical change is coming.
                    • What to try
                      • Put your brand in once. (Ethan Teague's tip): Claude Design, on Claude's paid plans, makes one-pagers and prototypes as well as slides. Give it an old deck or your logo and it builds a design system from your colours, fonts and layouts, which a team plan shares with everyone; even Motion, Claude's new animation tool, picks up the colours. An HTML deck gets you one good deck; a design system keeps your brand on everything.
                      • Ask who it would hire before you say who you are. (Ethan Teague's tip): In a chat that knows nothing about you (memory and custom instructions off, search on), ask what a buyer would type, which firms are good at what you do, and only then ask about your firm by name. Tried on ChatGPT, Claude and Gemini: by name they mostly got both Audience Strategies and Steadman right, but asked which firms train companies to use AI well, none named Steadman; they named McKinsey, Boston Consulting Group and Bain. A buyer who knows your name gets a fair picture; one who doesn't may never hear it.
                      • Expect the most likely answer, and brief around it. (David's tip): David had Claude test whether a language model can pick a random number: asked for one between one and ten, it said seven a hundred times out of a hundred, and asked for the spread it recited the well-known survey figures, so only a trick produced chance. One developer sees the same habit in code reviews that never come back clean ('coming back empty feels like failing the task'), so before it checks your work, set the bar (only flag what would break or lose data) and say that finding nothing is a good answer.
                      • Read the full edition with all links and sources

                        12 min
                      • Professional toe-treading
                        What happened this week
                        • This week was a consequential moment in the history of mathematics: On 6th October OpenAI put 722 papers written by an unreleased model on GitHub, and one economist called it probably the biggest day of scientific advancement in history; the next day it withdrew three and repaired 14 more, so the pattern is publish, check, fix, publish again. Checking is the slow, skilled part, which is the case a mathematician makes for continuing to pay for human expertise. The fastest-spreading worry is encryption: a leading expert says the labs have started, gingerly and discreetly, to test whether their models can break it, and a cryptographer at Johns Hopkins University says he thinks we might lose public key cryptography, though no one has shown a break yet.
                        • Someone found a possible planet in old NASA data with Claude Code. By that night, someone else had ten more: On 7th October someone who works in product posted that they had found a planet nobody knew existed, in seven years of public NASA telescope data, using curiosity and a long conversation with Claude Code. Ten hours later someone else ran the same idea across 9,979 nearby stars with Claude Opus 5.5 and turned up ten possible planets not reported anywhere they could find. No astronomer has confirmed any of it, which matters after the maths story, but the speed is the point: discoveries flowing as fast as people can find them.
                        • South Korea's bank raids show the power of one person with AI tools: The raids on South Korean banks over the past fortnight were most likely one person's work, an attacker using an AI hacking agent alongside Claude Code and other models against at least nine banks. The way in was old-fashioned: at Shinhan Bank the attacker got round identity checks on a portal used by loan brokers and about 25,000 customers' details were exposed. Executives at Anthropic, OpenAI and other labs are privately gaming out a public and political revolt after a catastrophic AI event, most likely a cyberattack that cuts off banking, the internet or even power and water, and assume one in 2027.
                        • What to try
                          • Ask it whether it's made things more complicated than they need to be.: A colleague at Steadman wrote that AI is very good at adding complexity unless you check it. Before accepting the next plan or document, ask whether it is more complicated than it needs to be, then ask the same of your own plan. When colleagues asked for new fields on the pipeline board, David built them but asked first to keep it simple: every field is something you have to keep up to date, and a half-empty field is worse than no field.
                          • Get it to walk you back through your steps, one question at a time.: Someone at dinner had lost their keys, retraced their steps with Claude and found them in the pocket of a jacket worn to an event; they could have done it alone but it helped them think more clearly. When you are stuck on something only you know, ask the model to question you one step at a time and not to guess. It is the same idea as letting it interview you for a brief, except here the answer is already in your head.
                          • Tell it what 'recent' means.: Twice in one day in Los Angeles, AI tools gave David and an economist friend answers that had once been true: Claude sent them to a trailhead that a voice assistant said had been shut after a lawsuit, and the voice assistant said a bar was cash only on a report about five years old. A weekly news summary David set up for a client once came back with items years old because he had never said they had to be recent. If the answer can go out of date, give it a date window.
                          • Read the full edition with all links and sources

                            13 min
                          • Paid for by forgetting
                            What happened this week
                            • America has lost 200,000 creative jobs in four years, and the losses stop at the door of the venue: Joseph Politano of Apricitas Economics counts more than 200,000 US jobs lost in creative fields over four years, about 50,000 of them in the last year, with film and television down more than 100,000 and written publishing down more than 70,000. His best evidence that AI plays a part is that every digital arts subsector is shedding jobs while the live arts keep growing, though he notes consolidation and offshoring are at work too. Alex Imas saw the music version: the industry has almost regained its 2000s peak, carried by live performance.
                            • Waymo's serious-injury crash rate is now 20 times lower than human drivers', and it publishes the comparison: On 270 million miles of data, Jeff Dean says Waymo's serious-injury crash rate is 20 times lower than human drivers', up from 13 times at 170 million miles in March, though Google and Waymo are both Alphabet companies. Kevin Bryan points out that any cut above 50 per cent means the cars are avoiding crashes other drivers would cause. The test David draws for trusting an agent at work is to measure it against the people doing the job today, not against perfection, and publish the comparison.
                            • Chicago will teach students to think without AI and to work with it, in different rooms: According to Steven McGuire, the University of Chicago will require every student to take an analogue writing class, ban technology from core humanities and social-science classes, and probably add a computational-thinking requirement. Derek Thompson called it a thoughtful answer to a real bind: use too much AI and you do not learn to think, avoid it for four years and you are behind for the best-paid entry-level jobs. David reads it as a design any team could copy: decide which work stays human on purpose, and say so.
                            • What to try
                              • Give it one thing to watch, and hear only when it changes.: David asked his assistant to keep an eye on a friend's flight and tell him when it landed and which terminal. A watch hands back attention, which is the point of the essay: the things we never get round to checking are where the money and the mistakes hide. Start with one thing, such as a flight, a renewal date or a price, ask it to check once and stay quiet unless something changes, then tell it the thing has changed and run it again.
                              • Ask your prompt which decisions it is quietly leaving to the machine.: When David marks workshop attendees' prompts, his second question is whether a decision is being made by the machine by accident. A prompt that says 'summarise these interviews' leaves the model to decide what counts as important, how long the summary is and whose voice it uses. Before a long prompt, ask the model to list every decision it would have to make that you have not made for it, then make the ones that matter yourself.
                              • Flip homework: learn with AI beforehand, then do it unaided on paper.: Paras Chopra, founder of Wingify, suggests flipping homework for the AI age: learn with the model at home, then show you can do it unaided in class. Before a meeting on a subject you are new to, spend twenty minutes learning it with a model that quizzes you rather than explains, then close the laptop and write on paper the three things you would say in the room. If you cannot, you learned less than it felt like.
                              • Read the full edition with all links and sources

                                11 min
                              • Where does it live?
                                What happened this week
                                • AI agents are getting into other people's systems: some for $25 a time, some by accident: Transluce has published more than 30,000 logs of AI agents straying into other people's systems, including one that pulled a file from an Australian health agency's test server; Anthony Albanese said OpenAI took "way too long" to tell the government it knew. A security researcher separately cites a campaign that used AI agents to break into about 100 businesses at roughly $25 of AI usage per target.
                                • AI is adding jobs overall, but the junior share is falling: The Economist reckons AI has created about a million US jobs against roughly 200,000 layoffs blamed on it, much of the gain in data-centre construction. A 41-country study of 1.25 billion job postings finds firms that adopt AI grow senior employment 6.7 per cent while junior employment falls about 2.5 per cent, with the junior share falling in 23 of 31 countries mostly because seniors were added, not because juniors were cut.
                                • Every task keeps getting cheaper, which is exactly why the AI bill goes up: Anthropic's Opus 5.5 costs 40 per cent less than Opus 5, and Epoch AI reckons the cost of a given level of AI performance has fallen about 47 per cent a quarter for three years. David separates three things people mix into one answer: any given task keeps collapsing in price, harder work keeps opening up on dearer frontier models, and Jevons's paradox means a rising bill usually signals using the frontier when you don't need to, or simply using more.
                                • What to try
                                  • Leave a handover note before you close the tab.: With two models working on one analysis across two machines, David asked one to merge both into a single document and, as his flight boarded, summarise everything it had done in one file so another agent could build on it. He now ends almost every job asking that everything learned be documented so future agents can understand and continue it.
                                  • Point it at the folder and make it write itself a map.: In a new project, the first instruction is to familiarise itself with the folder and save notes for next time. Coaching someone through this, the model read about thirty files, wrote itself a guide to seven subfolders and an index of interviews, then flagged what was wrong: out-of-date notes, two versions of the same one-pager, files it couldn't open.
                                  • Write down when you want options and when you want an answer.: A senior client's standing instructions are about manner, not content: don't over-analyse an easy question; give options with pros and cons for an analytical one; skip the pros and cons entirely if told to move fast. Three lines stop the quick question getting the long answer.
                                  • Read the full edition with all links and sources

                                    12 min
                                  • The crossing
                                    What happened this week
                                    • A research journal sat down with the authors of papers it was about to reject. For six of the seven papers, the authors couldn't fully explain their own work.: Transactions on Machine Learning Research now rejects about 53 per cent of submissions before a reviewer sees them, up from 6 per cent in 2023. An editor, Nihar Shah, asked ten authors heading for rejection to talk him through their own papers; three of the seven who showed up couldn't answer basic questions, three more managed the big idea but not the details, and two later sent answers a detection tool scored as entirely machine-written. It matters because the same question, whose name is on a report and could they take a question on any line, applies just as well inside a firm.
                                    • The cheaper machine-written posts get, the more people pay for a human one.: About 81 per cent of a sample of long LinkedIn posts were probably machine-written in July, on Originality AI's estimate. Meanwhile a Toronto ghostwriting agency serving about 40 American lawyers has more than doubled in a year, and one British ghostwriter charges roughly £350 a post. It matters because once a feed fills with competent, interchangeable text, the scarce thing becomes a voice that sounds like one particular person.
                                    • Deloitte showed its own consultants a chart of billing by the hour shrinking to a sliver.: At a May town hall, a Deloitte leader put up a chart running to 2035 in which hours-based consulting shrinks to a thin sliver of a bigger market dominated by AI agents. One consultant in the room summed it up: "They heavily implied our model is toast." It matters because a big consultancy is telling itself in private what its own clients have started to suspect, and the question worth asking any adviser is how they would price the same answer if it took an afternoon.
                                    • What to try
                                      • Say what good looks like before you ask for it.: A founder David coaches turned the effort setting on deep research up to max and still didn't get the depth wanted after hours of running. His advice was to leave effort on high and spend the attention on the prompt instead, opening with one sentence stating exactly what good looks like: comprehensive, thorough, detailed and accurate, but concise, with the high-priority recommendations first.
                                      • Get it to examine you on what it just did.: Kieran Klaassen, who runs the email product Cora at Every, has the model quiz him on what it just did after a long session; if he can't answer, he's found the part he has stopped owning. It's the practical antidote to the journal's finding that authors increasingly can't explain their own AI-assisted papers.
                                      • Build the outline yourself, then hand it over to fill.: A senior insights leader David interviewed found that asking a model for a presentation from a blank page produces something "really derivative", so she does the shape herself first, the outline, the template, how the data should be laid out, the story, and only then lets the model fill it in from the documentation. The result is better than she would have done by hand, and much faster.
                                      • Read the full edition with all links and sources

                                        14 min
                                      • We're all going to die
                                        What happened this week
                                        • OpenAI has put in writing that it can't rule out learning from its customers' unpublished work.: OpenAI announced on Tuesday that a group of its agents had solved a famous open problem in mathematics. Congratulating the two mathematicians who had been working on the same problem for a year and who had been using its tools, the company added: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models." The denial around it is very narrow, covering only access to specific user data. It matters because David expects it to knock business confidence in using these tools, and his instruction is practical: ask your vendors, in writing, what "de-identified" covers in your contract.
                                        • Same week, the worst result and the best, and the variable was the shape of the answer.: The OECD's PISA tests of 15-year-olds show reading scores falling, with the average pupil more than a year behind the 2018 cohort; in science, pupils who used AI almost every day were significantly less able to summarise or draft texts they had read than those who almost never did, though less so for those using it to learn. Andreas Schleicher, the OECD's education director, called the link between lower scores and passive technology use compelling but conceded there is no direct evidence of cause. Later in the week MIT Sloan featured an experiment by Ruru Hoong with 150 professional loan specialists: given the same AI risk assessment as a plain yes or no rather than a probability score, they decided better and faster, and better than on their own judgement. It matters because the lesson from both is the same, that whether AI helps depends less on the tool than on how people use it.
                                        • The generation most positive about AI at work is the one in its forties and fifties.: Nearly half of Gen X workers, born between 1965 and 1980, write positively about how their employers use AI, against about a third of Gen Z workers. Glassdoor's research puts part of that down to the transition creating more opportunity for senior people. David's mechanism goes further: these tools reward knowing which question to ask and being able to tell a good answer from a plausible one, and that is seniority rather than fluency. It matters because the consensus that the young arrive ready and the old resist has had a good run, and if your AI plan assumes the graduates will carry it, you should go and look at who is actually using the thing.
                                        • What to try
                                          • Make it put its questions inside the draft, in square brackets, rather than keeping them in a separate chat.: David wanted emails to go to an invite list for a client event, and instructed: "Draft emails for me. Throughout each, in square brackets, tell me anything I need to know, or any questions you have. I'll fix and then send." Most people keep the questions in a chat and the work in a document, then lose time reconciling the two; questions inside the draft get answered because they are in the way of sending it.
                                          • Delete the setup you built and see whether the model still needs it.: Boris Cherny, who created Claude Code at Anthropic, told a Y Combinator audience to do this every six months: "delete your claude.md file, delete your skills, and delete your hooks. Then see what the model does. It might surprise you." His reason is that the newest models may no longer need the long instructions older ones did. Most readers do not have those files, so the version for them is to turn off the standing-instructions box in their chat tool's settings for one real task and see whether the answer actually gets worse.
                                          • Write your reader down, then let them kill your ideas.: David does this all the time: write a profile of your target audience, who they are, what they already know and what they are tired of hearing, then paste in the profile alongside five things you could write or propose and ask which this audience would care about, and why. It kills the ideas that do not matter to that audience and prioritises those that do, before you have spent time on any of them.
                                          • Read the full edition with all links and sources

                                            13 min

                                          About David's Saturday AI Thoughts

                                          From the publisher's feed

                                          Each Saturday, David Boyle 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.