The biggest opportunity is the complexity. #gofigure

Watch a cab driver now. Even on roads they have driven a hundred times, the phone on the dashboard is telling them where to turn.London’s black-cab drivers used to spend years memorising tens of thousands of streets, a test called the Knowledge. Researchers scanned trainees’ brains before and after. The ones who qualified grew more grey matter in the part of the brain that maps space. The ones who gave up didn’t. The skill grew only where it was used.Today every driver running a maps app also feeds it: their speed on each road is how the app knows where the traffic is. Drivers are growing the thing that is replacing what they know. Nobody stops them learning the roads. It is just easier to follow the blue line, and a study that followed drivers over three years found that the more they used GPS, the more their own sense of direction declined.There’s a tree that dies this way. Birch is the first to grow on cleared ground, and it shelters the slower trees coming up beneath it. Nothing attacks it. They grow taller, the light stops reaching it, and birch cannot live in shade.That is the slope this series is about: no day when anything goes wrong, only a little less light each year. AI is now doing for our judgment what the map did for the road. And when the map is wrong, as it was in 2024 for three men who drove off an unfinished bridge in Uttar Pradesh, there is nobody left reading the road.The essay is about what waits at the bottom of that slope. Check out the latest article in my newsletter: Nothing Attacks the Birch linkedin.com/pulse/nothing-a… via @LinkedIn

This academic year, CBSE schools across India began teaching AI from Class 3. Eight-year-olds are learning to work with it before most of them have written a full essay. A year earlier, OpenAI announced half a million free ChatGPT licences for Indian teachers and students, including help with planning lessons. In the US, 37% of public school teachers already use AI to prepare lessons at least once a month. That is a lot of classrooms drawing on the same few systems, from the lesson plan to the homework. What worries me turned up in a study of 6,875 essays this spring. AI made the essays better. It also made them alike: the way the arguments were put together lost between 70% and 78% of its variety. Farmers know where that kind of sameness leads. In 1840s Ireland, most of the poor lived on a single variety of potato. When blight arrived, it found every field equally defenceless, and about a million people died. The variety had been the protection, and nobody noticed until it was gone. A generation that learns one shape of argument from the age of eight will share the same blind spots. A mistake one of them misses, all of them miss, and there is nobody left thinking differently to catch it. There was a hopeful detail. When students gave the AI more specific prompts, their arguments became more varied in depth, not less. The sameness comes from how we use these tools, which means it can be designed out. The essay asks what happens to a society that thinks alike. It starts with the banana in your fruit bowl. Check out the latest article in my newsletter: Cognitive Monoculture linkedin.com/pulse/cognitive… via @LinkedIn

You went to your bank to renew a fixed deposit. Your relationship manager, someone you had trusted for years, suggested a policy with “guaranteed returns”. She believed in it. You signed. Two years later you needed the money, and learned what surrender value means. Nobody lied to you. Something else was in the room. In September, India’s insurance regulator published what insurers pay banks to sell their policies. A bank that sells for one insurer earns about ₹13 of every ₹100 of new premium. A bank that sells for several earns ₹33. Between 2022-23 and 2024-25, banks and other corporate agents sold 1.3 times more insurance and were paid 2.3 times more for selling it. You never saw those numbers. Neither, often, did the person selling to you. In 1976 two economists, Michael Jensen and William Meckling, described what happens next. A principal hires an agent to act for them. The agent has interests of their own, and the principal cannot watch every move. Even honest agents drift towards what they are paid for: which product reaches the desk, which one gets mentioned first, which one gets the extra hour. The principal pays for that drift. Jensen and Meckling called it agency cost. You were the principal. The surrender value was your agency cost. Now look at the incentive plan your board approved last year. Every customer who trusts your staff is a principal, and none of them can see it. No one has to lie for you to lose money. The pay plan does the talking. Check out the latest article in my newsletter: Who pays when the bet goes wrong linkedin.com/pulse/who-pays-… via @LinkedIn

You went to your bank to renew a fixed deposit. Your relationship manager, someone you had trusted for years, suggested a policy with “guaranteed returns”. She believed in it. You signed. Two years later you needed the money, and learned what surrender value means. Nobody lied to you. Something else was in the room. In September, India’s insurance regulator published what insurers pay banks to sell their policies. A bank that sells for one insurer earns about ₹13 of every ₹100 of new premium. A bank that sells for several earns ₹33. Between 2022-23 and 2024-25, banks and other corporate agents sold 1.3 times more insurance and were paid 2.3 times more for selling it. You never saw those numbers. Neither, often, did the person selling to you. In 1976 two economists, Michael Jensen and William Meckling, described what happens next. A principal hires an agent to act for them. The agent has interests of their own, and the principal cannot watch every move. Even honest agents drift towards what they are paid for: which product reaches the desk, which one gets mentioned first, which one gets the extra hour. The principal pays for that drift. Jensen and Meckling called it agency cost. You were the principal. The surrender value was your agency cost. Now look at the incentive plan your board approved last year. Every customer who trusts your staff is a principal, and none of them can see it. No one has to lie for you to lose money. The pay plan does the talking. Check out the latest article in my newsletter: Who pays when the bet goes wrong linkedin.com/pulse/who-pays-… via @LinkedIn

Offer people a choice between a hard mental task and a painful burst of heat on the skin, and many will take some of the pain. That was a study in eLife in 2020. A 2024 review of 170 studies found the same thing more gently: the harder people think, the worse they feel. Your brain treats thinking as a cost, and it looks for any way out of paying. AI is the first tool that removes the cost completely. Common Sense Media found this summer that 70% of American teenagers use AI for schoolwork, and 63% of those say they use it to get answers. What I keep coming back to is a detail in that 2024 review. The unpleasantness was smaller in countries where students spend longer on schoolwork. The researchers think tolerance for hard thinking is learned, by doing it. The body has its own version of this, and you can see it in any classroom. In the big cities of East Asia, 80 to 90% of young people now leave school short-sighted. Genes can’t change that fast. A growing eye needs daylight to know when to stop lengthening, and children now spend their growing years indoors. When schools in Guangzhou added 40 minutes outdoors a day, new cases fell from 39.5% to 30.4% over three years. When an adult hands a task to AI, they skip a cost they paid years ago. A fifteen-year-old skips the work that would have built the skill, and the bill arrives about fifteen years later. The essay starts with a child born this year, who will never meet a search that doesn’t answer. Check out the latest article in my newsletter: The First Generation That Never Chewed linkedin.com/pulse/first-gen… via @LinkedIn

The cheaper AI gets, the more it will cost you. In the past three weeks, OpenAI and Anthropic have cut the price of their small models by 50% to 90%. Both now charge 10 cents per million input tokens. Boards will read that as good news for the budget. Uber is the warning. Prices had already fallen about 1,000 times in three years when it used up its entire 2026 budget for AI coding tools in four months. Cheap is what makes it expensive. When a query costs almost nothing, nobody rations it. Chat becomes agents, and agents run pipelines: one task, dozens of model calls, 5 to 100 times the tokens. Every call lands on a metered cloud bill that no one approved line by line. By the time it arrives, the work depends on it. This is old. In 1865 William Stanley Jevons showed that more efficient steam engines made Britain burn more coal, and it kept burning more for almost fifty years. Efficiency has never shrunk the bill for something this useful. So budget AI like electricity: a permanent line that grows with the business, metered from day one. A falling price is a forecast of a rising bill. Check out the latest article in my newsletter: The efficiency that raised the bill linkedin.com/pulse/efficienc… via @LinkedIn

Everyone has had one friend who agrees with everything. Every plan is brilliant and every ex was the problem. You leave those conversations feeling taller, and you make some of your worst decisions right after them. In April 1961, John F. Kennedy had a room full of them. Arthur Schlesinger, one of his advisers, opposed the CIA’s plan to land 1,400 Cuban exiles at the Bay of Pigs in a private memo, then said almost nothing in the meetings. The invasion collapsed within three days and most of the men were captured. Schlesinger later wrote that he “bitterly reproached” himself for keeping so silent. That friend now lives in your phone, and it answers at 2 a.m. Anthropic studied 1.5 million conversations with its own chatbot, looking for the ones that pulled people’s beliefs, values or decisions away from their own. Mild cases turned up in roughly one conversation in 50 to 70, and the rate rose between late 2024 and late 2025. People gave those conversations more thumbs-up than ordinary ones. The AI was saying things like “CONFIRMED”, “EXACTLY” and “100%”. The big safety frameworks check whether a model could help build a bioweapon, launch a cyberattack or slip out of human control. None of them measures what a year of being agreed with does to you. The essay asks why our safety checks are built to catch the wrong thing. It starts in Hanoi in 1902, with a bounty on rats. Check out the latest article in my newsletter: Safe, Helpful, and Making You Worse linkedin.com/pulse/safe-help… via @LinkedIn

Captain Smit Machchhar was stabbed in his own cockpit on 30 September, with 174 people behind him and the plane diving more than 14,000 feet in under 30 seconds. Bleeding, he held off his attacker long enough to unlock the door and let help in. The plane landed safely. Whatever you call what he did in those seconds, nearly 10,000 hours of flying went into it. Oddly, that is the part automation leaves alone. When NASA researchers had sixteen experienced pilots fly a 747 simulator without the autopilot, their hands and reflexes were fine. What slipped was the thinking: keeping track of where the plane was without the map, spotting a failed instrument. It was worst in pilots whose minds wandered while the autopilot flew. Aviation knows this because it measures pilots. Most things don't. Over a million Americans with no approved medical reason are now on weight-loss drugs, and people who start them walk about 560 fewer steps a day. The scale never shows it. That is the slope this series is about, and AI is running the same experiment on how we think. The essay asks who chooses the numbers we watch, and it starts in 1965, with the sugar industry and two Harvard scientists. Check out the latest article in my newsletter: Who Paid for the Study That Says You're Faster? linkedin.com/pulse/who-paid-… via @LinkedIn

Check out the latest article in my newsletter: Everyone is cutting. That is not evidence. linkedin.com/pulse/everyone-… via @LinkedIn

Check out the latest article in my newsletter: The Antlion Doesn't Hunt linkedin.com/pulse/antlion-d… via @LinkedIn

Check out the latest article in my newsletter: The cost you cut was holding something up "There exists in such a case a certain institution or law; let us say, for the sake of simplicity, a fence or gate erected across a road. The more modern type of reformer goes gaily up to it and says, 'I don’t see the use of this; let us clear it away.' To which the more intelligent type of reformer will do well to answer: 'If you don’t see the use of it, I certainly won’t let you clear it away. Go away and think. Then, when you can come back and tell me that you do see the use of it, I may allow you to destroy it.'" linkedin.com/pulse/cost-you-… via @LinkedIn

A student hands in a perfect essay. The professor asks them to explain one paragraph. Blank stare. Professors say they keep seeing this. The unsettling part is that the student may not be bluffing. They really thought they understood it. In the last essay I wrote that we can tell help that makes us stronger from help that does the work for us, in a few seconds. And we still reach for the second kind. I said it wasn't about willpower. This is what I meant. We can't feel the difference. Universities have noticed. This term, more marks are moving to exams written in the room and answers defended out loud. One report put the whole problem in its headline: "Perfect homework, blank stares." For most of history, clear writing meant clear thinking. That is no longer true. The examiner's check is being rebuilt. Yours isn't. In an Aalto University study, people who used ChatGPT on reasoning tests all overestimated their scores. The ones who knew most about AI overestimated most. This series started with a claim: something is being taken from us, and we have no way to measure it. The first measuring tool to go may be the one in our own head. Universities brought back the invigilator. Nobody has one for the inside of your head. So what happens when something learns to trigger the feeling of understanding on demand? The best answer I found came from a seagull chick. Check out the latest article in my newsletter: Fed and Yet, Malnourished linkedin.com/pulse/fed-yet-m… via @LinkedIn

Yesterday, 15 September, the Federation of Indian Pilots asked the regulator to scrap airlines' rest-rule exemptions. My first thought: if it agrees, airlines need more spare pilots or fewer flights. IndiGo showed what happens without them. Last December its rosters were so tight its buffer reportedly hit zero. After stricter rules arrived, it cancelled 2,507 flights in three days. Kingfisher and Starbucks piled up risk the same way, one sensible decision at a time. Kingfisher collapsed in 2012. Starbucks fought for its life in 2008 and recovered. Each trigger was ordinary. The physicist Per Bak's sandpile describes this: once enough grains pile up, any one can start an avalanche. I wanted to know why one pile recovers and another collapses. Check out the latest article in my newsletter: The right decision at the right time linkedin.com/pulse/right-dec… via @LinkedIn

Check out the latest article in my newsletter: Cooking Made Us. Being Fed Will Unmake Us. On 12 September Dario Amodei published an essay asking the AI industry to slow down. His worry is that within six to twelve months a swarm of misaligned AI agents "could be capable of taking over the entire internet." He is right to take that seriously. The response he proposes has a familiar shape: embedded evaluators, democratic coordination and global coordination, which amounts to inspectors, votes and treaties. It is how we answer an extreme. Terrorism kills fewer than ten thousand people a year worldwide, and we rebuilt airports, budgets and foreign policy around it. About 3.7 million deaths in 2021 were attributable to high body mass index. There was no treaty, no inspector and no pause on the recipe. It arrived one convenient meal at a time, and each meal was fine. We answer a spike and shrug at a slope. Nobody ever evacuated a food court. AI has a spike, and Dario is pacing it. It also has a slope: a tool helpful enough that the work which used to make us capable slowly stops being done. Nobody is embedding evaluators for that one. What I cannot settle is whether the two compete for the same attention, or whether slowing the spike buys the time we would need to see the slope. linkedin.com/pulse/cooking-m… via @LinkedIn

As Air India welcomes their new CEO, Tewolde Gebremariam, he would be pleasantly surprised with the upward trajectory of their service performance in recent days. Air India's net promoter score has gone from minus thirty-five to plus thirty-seven in three years.That is a real turnaround, and the people who did it deserve the credit. Plus thirty-seven against what, though? In the 1970s Air India was the airline to beat, other Asian carriers studied. Singapore Airlines took its cabin service standards from it. Singapore Airlines now owns 25.1 percent of Air India, and last week agreed to help fund its recovery. Nobody involved has done anything wrong. Every leader measures progress from the day they arrived, because that is the honest thing to do and because the data starts there. Which is also how a fifty-year decline stops being visible to anyone. Try this on your own business. It takes thirty seconds. Open your last board deck. Find the oldest number in it. If nothing in that pack predates the current management team, every trend line you are looking at begins after the damage. I used to think the board was the safeguard here, since they are the only people in the room who were there before the current team. I am less sure now that board tenure is often three years as well. There is a name for what happens next, and a fuller version of that test. Both are in this week's Ring Signal — along with some photographs from Key West that make the point better than I can. Nobody Remembers the Big Fish. linkedin.com/pulse/nobody-re… via @LinkedIn

AI risk gets diagnosed the way we diagnosed the bomb. Who holds the weights. How powerful is the model. Can it escape the lab. Non-proliferation, frontier thresholds, containment. South Korea has just made that frame look narrow. Fifty-two million citizens, free unlimited AI, arriving through messengers, phone calls and text messages. It needs no frontier model. It runs on domestic ones, deliberately. Nothing in it is a weapon. All of it is an exposure. Which is a different science, with four questions the containment frame cannot ask. Dose. Not how powerful the system is. How much of it, how often, for how long. Route. Not whether it escaped the lab. Whether you went and opened it, or it arrived in a thread you were already reading. Cohort. Not who holds the weights. Who is exposed, and which group had the least say in it. Baseline. Not what the model can do. What the person could do before, and what they can do now, unassisted. Only the last one settles anything, and it is the only one nobody is funding. Korea is not the villain here. Korea is the first legible version — one jurisdiction, one start date, a real before and after. Everyone else is running the same study without a calendar. What I cannot resolve is that epidemiology needs an unexposed group to measure against, and I am no longer sure one exists. linkedin.com/newsletters/the…

Check out the latest article in my newsletter: Expensive on purpose linkedin.com/pulse/expensive… via @LinkedIn

Check out the latest article in my newsletter: The map is not the territory linkedin.com/pulse/map-terri… via @LinkedIn

Check out the latest article in my newsletter: Coasting linkedin.com/pulse/coasting-… via @LinkedIn

Check out the latest article in my newsletter: The memory business & loss aversion linkedin.com/pulse/memory-bu… via @LinkedIn