Everything I've written

A friend switched from Sonnet to Opus and asked a simple question: "Is this direction right?" Opus came back with five points of pushback. He paused. He realized his previous AI had never once challenged him like that. Anthropic recently analyzed over 300,000 conversations to measure how different versions of Claude actually communicate. Sonnet is the warmest of the group: it affirms what you say, responds with empathy, and gives you forward momentum. Opus takes a different approach, challenging your assumptions, finding the weak points in your thinking, asking you to explain your logic. Anthropic calls this the "Warmth vs. Rigor" spectrum. Even the language you use makes a difference: Claude responds quite differently in Chinese versus English. Same AI, but the conversation you get depends entirely on which version you're talking to. This isn't about which version is better. Someone who uses AI more than six hours a day recently started feeling like his thinking had gotten slower. After talking it through for a while, a possible reason emerged: almost all his thinking was being confirmed by AI. He'd stopped pausing to think first. He'd send a question, get an answer, and accept it. No friction. Nobody to say "wait, there's a hole in this." That AI almost always told him he was on the right track, he'd feel good, and move on. Before AI, the way you calibrated your judgment was through other people. Someone supported you; someone pushed back. That process was slow and uncomfortable, but it told you what held up under pressure. AI can now fill that role, but different AIs bring different kinds of relationships: some keep you moving, and some make you stop. Which one do you usually turn to?

Kimi K3 costs a fraction of Claude Opus to use. It has 2.8 trillion parameters, putting it in the same tier as Claude Opus and GPT-5.5. Over the past few weeks, Coinbase, DoorDash, Siemens, and Airbnb have been quietly switching parts of their AI workflows to cheaper Chinese open-source models, cutting their AI bills in half or more. The signal most people are reading into this: AI is getting cheaper, and demand for computing power will follow. But that logic is missing a step. Kimi K3 is cheap in the sense that its license and API fees are cheap. The money that used to flow to companies like Anthropic and OpenAI is flowing elsewhere. But running a model with 2.8 trillion parameters still takes roughly the same computing resources as running any other model at that scale. Openness didn't shrink that number. Gavin Baker, a well-known US tech investor, laid out a counterintuitive case a few weeks ago: if cheaper models take share from expensive ones, the same budget buys more compute, ROI improves, total usage goes up, and hardware demand could actually grow. Kimi K3 going open-source pushes this further: companies can fine-tune it and run it on their own servers. Each company hosting its own 2.8-trillion-parameter inference stack is far less efficient than thousands of companies sharing a cloud service. Total hardware demand gets larger, not smaller. Next time you see a headline about AI getting cheaper, it's worth asking: cheaper for whom? Coinbase cut its AI bill in half. Nvidia's chip orders this quarter are still climbing.

Yesterday, someone rebuilt their company website from scratch. The old setup was Angular on the front end, .NET on the back, a full CMS behind it all. The whole thing had been running for years. Change a line of text and you'd need the frontend engineer, the backend engineer, a deployment cycle that could stretch over weeks. They replaced it with a static site generated by AI. No CMS. No database. Hosting cost: zero. Their job title wasn't engineer. Around the same time: Spotify's top engineers have gone months without writing code themselves. The head of Anthropic's Claude Code project says he hasn't written a single line since November 2025. His job is now to hand tasks to AI agents and check the results. At Google, three-quarters of all code is AI-generated. Engineers aren't writing code. Non-engineers are. The two things happened at the same time. The skill isn't disappearing. It's moving. On the engineering side, the execution wall (reading code, making changes by hand) is being absorbed by AI. The job's center of gravity has shifted toward architecture, catching where the AI goes wrong, deciding what's actually worth building. On the other side, a door opened. People who understood the business problems but couldn't reach the technical layer now can. What that person brought to the website rebuild was knowledge of their own company: which features went unused for months, which infrastructure existed because it always had. AI couldn't supply that judgment. But AI gave it somewhere to land. What stops you has changed. It used to be whether you could write code. It's starting to look more like: how well do you understand the problem in front of you? AI handles the execution. You handle the question. That rebuilt site updates a few times a year. Static files, AI-assisted changes, zero hosting cost. The business judgment behind it, what this site actually needs to do: that part, the AI didn't provide.

You probably haven't thought of it this way: knowing you can't win might be exactly when you figure out where you actually can. Today, Korea's top-ranked Go player Shin Jin-seo began a three-game match against the AI KataGo. KataGo spotted him two stones, a substantial advantage in Go terms. Even so, Shin's practice win rate has been below ten percent. He hadn't seriously studied KataGo under these conditions before. There was no reason to. Once he accepted the match, he started asking: what can I actually count on? His answer was endgame and territory. KataGo vastly outmatches any human in calculation, and surprising moves haven't thrown off modern AI for years. But careful endgame play, holding ground, protecting what's already his, those are where he feels steady. The two-stone game makes it almost impossible to avoid fighting altogether, so the real question became how not to get swept up in KataGo's pace, and carry some control into the late game. He figured this out only after accepting he couldn't win on the AI's terms. A Japanese company founded 270 years ago recently gave one of its non-engineer employees access to Claude Code. Within weeks, they had built an internal tool that consolidated data scattered across Excel files, various systems, and old VBA scripts, and improved the company's dashboards. The same work might have cost several months and millions of yen to outsource. They didn't learn to code. What they brought in was something an outside engineer wouldn't have had: years of knowing exactly which numbers need to match, which errors cause the most trouble, which reports get asked for every meeting. The tool lowered the technical barrier; their business knowledge is what made the result actually useful. Different conditions, same rhythm: first getting clear about where your edge isn't, then seeing what's actually yours. Shin Jin-seo's practice win rate is still under ten percent. He says this is the first time he's seriously studied where his strengths are.

In Pennsylvania, a hospital system runs an AI tool to screen for colorectal cancer. Every week, it scans through hundreds of overdue patients and flags only 27.7% of those who actually have the disease. In other words, it skips seven out of ten. Over the years it's been running, the people it does find have seen their two-year mortality drop by 43%. The AI was never designed to find everyone. It answers a single question: given that nurses have limited time and colonoscopy slots are finite, who should we call first? Geisinger Health System deals with around 450 overdue patients every week. The AI runs through their age, sex, and recent blood work, then flags the eleven highest-risk individuals. Nurses make eleven calls that week. Eleven is a designed number. Flag a hundred people, and nurses can't get through to all of them, and the colonoscopy schedule won't hold. Flag precisely, so every person who makes the list actually gets followed up and booked in. That four-month head start is often what separates early-stage from late-stage. Among those the AI flags for a colonoscopy, 6 in 100 turn out to have cancer. Among those it doesn't flag, fewer than 1 in 100 do. One thing the research documented: when a hospital tried using an automated voice call to notify patients about screening, the follow-through rate was nearly identical to doing nothing at all. A real person on the phone is what gets people to actually show up. The nurse makes the eleven calls matter. We tend to evaluate AI by the numbers it produces: how much it finds, how often it's wrong, how often it hallucinates. Today's leading models, with access to tools, hallucinate around 3% of the time. When uncertain, they're more likely to say "I can't verify that" than to generate a plausible-sounding answer. Most people's mental image of AI reliability is stuck a few years back. The 27.7% figure and the 3% hallucination rate point to the same issue: the number alone doesn't tell you what framework it's measured in. A metric that looks terrible in one context might be exactly right in another. Before the number means anything, you have to know what question it was answering. Next time you see "AI accuracy" or "how much AI found," ask first: what is it measuring, and under what conditions? Is the thing being measured actually the problem you're trying to solve?

Seventy percent of Gen Z say they're worried AI will hurt job prospects. In the same Gallup poll, a quarter said they wish smartphones had never been invented. This is the first generation that grew up with them. This summer in New York, a group organized something called Summer of Ludd: no phones allowed, no social media promotion, attendees chanting "No Gemini, no GPT, no Claude." The neo-Luddite movement isn't against technology as such. It's against algorithms, data tracking, and how tech has quietly eaten into social life and sleep. What's funny is that this anti-social-media event spread mainly through TikTok. Around the same time, a Harvard undergraduate asked a question: does reading still matter in the AI era? She couldn't focus long enough to finish a book. Her classmates were forming reading groups; she felt left behind. The answer she got surprised her: if you can't read one book, read ten at once. Pick one up, read five pages, lose focus, put it down, pick up another. Ten books running simultaneously, no commitment to any of them. Without the weight of "I have to finish this one," you start finding unexpected threads connecting them. The advice rests on an observation: you're trying to fit yourself into a format most people find hard, reading straight through from cover to cover in one sitting. Academia figured this out centuries ago. The reason papers have abstracts, topic sentences, and section headers is that even full-time scholars have short attention spans. AI has exploded the amount of information available. Your brain is still the version that needs to come up for air. The Summer of Ludd logic and the reading advice are saying the same thing. They organized on TikTok, then put their phones away when they arrived. When the event ended, the phones came back out. What they were practicing was the act of choosing, directing attention toward what they decided. Next time you want to put a book down halfway through, put it down and pick up another one. You've always read this way.

In 1952, Grace Hopper sat down at a Mark I computer and finished the first compiler. The concept was simple: let a person write "ADD" and have the machine translate it into binary. Before that, programming meant speaking the machine's own language directly. After that, you spoke, and the machine translated. That chain extended for 70 years, all the way to today's AI code generators. Say "write me a function that takes a city name and returns today's weather forecast." Within seconds, working code appears, complete with docstrings, ready to run. The speed at which AI writes code already surpasses any human engineer. But none of that involves thinking. The model may have read more Python than every software engineer on earth combined. Its mechanism, though, is prediction: given billions of training examples, find the token most likely to follow statistically. It knows what typically comes after "city name input." Why that is, it can't tell you. Seventy years ago, Hopper's compiler was translating. Today's AI is still translating. The target has just shifted to your own words. There's a practical problem with this translation machine: developers accept roughly 30% of what it suggests. The other 70% gets read and deleted. And of that 30% that does get accepted, studies find around half carries known security vulnerabilities: syntactically clean, looks right, but hiding problems inside. That same week, the University of Chicago Law School announced something that looked like the opposite: starting this year, first-year students can't bring any devices to class, not even laptops. The school's position isn't that AI is bad. It's that students need a space without a translation machine first, where thinking has to do the work on its own. The approach is Socratic: professors ask, students answer in real time, reading cases themselves, tracing arguments themselves, finding the flaws in logic themselves. The point is that by the time students do use AI, they can actually judge whether what it produced is right. AI has pushed translation further than 1952 could have imagined. Chicago is saying: before you let AI do the translating, understand what you're translating first. Both things happening at once. Neither is wrong. That 70% of AI code suggestions that didn't make it: generated, then gone. Next semester, Chicago's first-year students walk in, desks clear, professor asks, they think.

Jie Ding, a researcher at the University of Minnesota, built a tool called Academic Humanizer. His reasoning was direct: existing AI detection tools are far too unreliable, routinely misclassifying human writing as AI-generated and wrongly penalizing students. His tool helps users rewrite sentences to sound more human. Days later, Nature ran a piece on it. Academia erupted. The debate that followed centered on one question: does this tool make it easier to pass off AI-generated work as human? But if AI detectors regularly misclassify human writing, the detection problem is not actually the real issue. The question underneath is: when a piece of writing reaches you, what is left of the trust between you and whoever wrote it? When a student submits an assignment, or a researcher submits a paper, there is an implicit understanding at work: you claim authorship, and the reader takes your word for it. Academic Humanizer and tools like it have made that underlying trust harder to lean on. The same week, New York City froze all new AI-tagged educational software purchases, calling for consensus before moving forward. Three separate groups spoke up: privacy advocates, teacher unions, and a broader coalition. More than four thousand signatures called for a two-year halt. That freeze is an admission: in an AI-saturated environment, no one has figured out how to rebuild this kind of trust. The question reaches further than education. Any article, report, or message you receive, you are now quietly running the same check: this person says they wrote it. Do I believe them?

An investor who is among the earliest AI bulls in Silicon Valley recently shared this about his own company: AI token costs double every 45 days. Productivity gains from all that spending: 5%. He does not know how many other companies are on the same curve, but expects most of them will be there within a few years. There is a reason for this gap that rarely gets named. The tasks AI handles well are mostly in the execution layer: standardized reports, data cleanup, process automation. AI does them fast, and the cost keeps falling. But the things that actually make a company more competitive are almost never settled at the execution layer. Harvard Business School lecturer Christina Wallace noticed something similar from a different angle. She argues that betting everything on a single full-time job is one of the highest-risk choices a person can make today. When one job carries your entire income, your professional network, your sense of identity, any shift in that position leaves you with almost no room to move. She calls it treating your career like a single stock. Marketing researcher Mark Schaefer found another layer of the same thing: AI-powered service is approaching perfection, available around the clock, infinitely patient, endlessly personalized. Yet the brands clients keep returning to are the ones with a point of view, an opinion, a willingness to say no. AI raised the floor. It also made something scarce: your real judgment. Two different observations, two sides of the same shift. One path is to become more precise and efficient in the AI era, aligning your work with what AI already does well. The other is to let AI handle execution and spend your energy building the kind of judgment that crosses domains and earns trust, including the parts that are not especially polished. Christina Wallace shut down her company when she was 28. Harvard turned that failure into a case study and taught it for nine years. Around 900 MBA students a year sat down to debate whether she should have kept going. The company is gone. The conversation is still running.

You probably have a handful of tasks where you make the same call every time. Before sorting a certain type of email, you run through the same checklist of questions. Before writing a proposal, you check the same criteria. When a familiar problem comes in, you run the same mental process before deciding how to respond. You know these routines well enough to do them almost automatically, but each time, you still have to run through them yourself. Most people haven't considered that AI could handle that part for them. In 2025, the Stanford Digital Economy Lab analyzed a large US payroll dataset and found a specific pattern. The biggest labor market impact of AI is appearing inside occupations rather than eliminating them. The entry-level layer of work, organizing data, writing first drafts, doing initial analysis, is shrinking fastest. Entire occupations haven't disappeared; the entry points into them have narrowed. The reason is fairly clear. Early-career work shares one characteristic: repeated judgments. The logic is the same each time; only the material changes. Give AI a consistent set of rules and examples, and it handles this reliably. Your own work probably has a similar layer. Tasks where you ask the same questions before responding. Documents where your structure never changes. Situations where your first move is always the same. For a long time, you've carried all of this through memory and habit. Few people stop to ask whether this work actually needs to pass through them each time. Try writing those repeated judgments out: what information you check first, what standards you already have answers to, how your thinking moves through the task. Once it's written out, that logic can run in AI. You review the result. The mental overhead of re-running the same process each time is what actually gets freed up. That capacity can go somewhere more valuable: decisions where you need the full context, where you make the final call, where you're accountable for what happens. AI can offer options there too. But the signature is still yours. Write out the reasoning behind a few tasks where your answer is always the same. In the process, you'll see which ones can be handed off.

In 2013, Andy Fang wrote DoorDash's first lines of code from his Stanford dorm room. As the company grew, he gradually stepped away from coding until he stopped entirely. That pause lasted more than ten years. This year, he started shipping new features into DoorDash's live codebase again, using AI coding tools. He calls it a "major comeback." His reason is specific. If he didn't personally use these tools, he said, if he didn't personally walk through DoorDash's entire deployment pipeline, he couldn't know where his engineers were actually getting stuck. To evaluate whether something is hard, you have to feel where the difficulty is. He turned this into a company expectation: every engineering manager at DoorDash should set a personal goal to ship production-level code, going through the full pipeline, not just running a test version locally. Walk through it yourself, he said, and you'll see two things at once: where the tools hit their limits, and what's genuinely slowing your team down. He went back to the codebase, he said, because he needed to keep the feeling fresh. AI has brought down the cost of execution significantly. Meeting materials get pre-summarized. Reports get drafted for you. Proposals get a first version generated before you ever weigh in. The efficiency case holds up. But your sense of whether something is good or not comes from years ago, when you were still doing it yourself. That sense expires. Over time, you can see the volume, but you start losing the ability to judge the quality. A manager mentioned something worth sitting with: the highest-performing AI users on his team were producing faster than ever, but their calendars were actually filling up more, not less. Output accelerated, and so did the number of underspecified tasks landing in their inbox. Moving fast, but less certain which direction to move in. After walking the full pipeline himself, Fang said looking at his engineers' work feels genuinely different now. He knows what's actually stuck. He knows what has a path forward. Next time you're evaluating someone's work, or deciding whether something is worth continuing, ask yourself: when did I last walk through this, end to end, myself?

"Except for three students, everyone else seems to have cheated with AI." That's what Paul Graham said after seeing a grade chart from Brown University. A professor there tried something this semester: the midterm was take-home, the final was in-person. The results: nearly everyone aced the take-home. Final scores dropped sharply. Same chart. Three completely different readings in the comments. One: This is proof. High grades at home, low scores in the room — students never really learned anything. Two: Hold on. In most workplaces, managers care about results. If your job allows AI, submitting strong work that used AI should count as doing your job. Three: Both readings are looking at the wrong thing. The real curiosity is S22 — the student who scored about the same regardless. While everyone else apparently flew with AI, this person just stayed level. Three readings, three different questions underneath: Did students cheat? What is the school actually testing? What does it even mean to "know" something now? YC is one of the world's most well-known startup accelerators. CEO Gary Tan recently shared that he now writes AI code himself every day — tens of thousands of lines a month for several months. His reason: AI coding capability is improving so fast that without firsthand experience, he can't tell whether a startup's demo represents real technical depth or something anyone could replicate in a day with AI. To evaluate "how hard is this?" you need to feel where the difficulty actually is. The Brown professor responded to the same problem by changing the format. Gary Tan responded by doing it himself. But not every evaluator has the time or willingness to do either. While that gap hasn't closed yet, those being evaluated and those doing the evaluating may be using entirely different measuring sticks on the same test or work. Ten years from now, will job interviews test what you can do with AI — or what you can do without it?

This July, a Harvard Medical School team published a model called COMPASS that analyzes cancer patients' genetic data and predicts whether immunotherapy will work for them. It outperforms the 22 best existing methods by an average of 8.5 percent. In bladder cancer data, patients it predicted as responders had an 86 percent one-year survival rate versus 40 percent for predicted non-responders. The paper's closing note: these results cannot be used to deny a patient treatment. Same month, a Taiwanese TV station used AI to generate a typhoon track map with location errors. Eight hundred thousand people saw it. One AI, more accurate. One AI, wrong. Two stories that seem unrelated, but they're asking the same question. Immunotherapy transformed oncology over the past decade, but only 10 to 40 percent of patients respond, depending on cancer type. The rest endure side effects without benefit. Existing biomarkers for predicting who will respond are inconsistent across cancer types. COMPASS is much more accurate, and it explains its reasoning: which immune signal pathways blocked a treatment that should have worked. But it only sees genetic data. Patient age, cancer stage, medical history, prior treatments: the context a physician actually needs to make a decision. The model cannot see any of this. Its accuracy has also only been validated against past research data, not in prospective clinical trials. The typhoon map situation isn't about whether AI can draw maps. It can. The problem is that pre-AI workflows included verification: data source checks, human review, fact-checking. AI-generated direct output skipped those steps. One Taiwanese journalist proposed an alternative: use AI to operate existing mapping tools, pull from official weather databases, run fact-checking processes, then output. The speed AI brings is still there, but the human confirmation step in the middle isn't gone. COMPASS is waiting for the same kind of process. Prospective clinical trials, where it's tested in real clinical environments, where doctors can see its outputs alongside full patient context and confirm the match. That path is how its accuracy becomes something a clinical system can trust. For many families, cancer treatment decisions are among the most difficult they'll face. AI entering this space with higher accuracy is genuinely good. But higher accuracy and ready to use directly are two different things. That gap is closing. It hasn't closed yet. The typhoon map: eight hundred thousand people already saw it. COMPASS's data: still waiting for the trial.

Before Boris Cherny built Claude Code, he spent two years in Nara making miso. White miso takes three months to mature; red miso, two years or more. He said the process taught him something: what you start today is ready for a day that hasn't come yet. When he later described how Claude Code was built, he said the team designed it for the next generation of models, the ones that hadn't been released yet. There's a common move in the AI world: a new model comes out, and people immediately ask whether to switch. That's not unreasonable. It just has a short shelf life. A few months later, there's a new answer. Someone spent thousands of dollars stress-testing the strongest models available and noticed something: an experienced AI user with a slightly older model still outperformed a newcomer using the latest one. The gap came from the system built around the model, how to direct it, how to split tasks, how to verify outputs. When he swapped the top model for a cheap one on the same work, results were nearly identical. Cost dropped by several times. Models can be replaced, repriced, or shut down. But the working habits you build while using AI, what you delegate, what you check yourself, how to frame a task so you don't spend three rounds clarifying, that knowledge stays with you when the model changes. Boris's red miso took two years before it was ready. Each version of Claude Code he ships, he says, is aimed at the model that's still coming.

「這一次開賣，是我這幾年賣得最差的一次。」 說這話的是 Josh W. Comeau，在美國教前端開發的人，把 CSS、JavaScript 講得清楚，有口碑。七月初他在文章裡公開了：最新一輪課程銷售明顯下滑，但他的課沒有變差。 他賣的是解釋。你付錢，買的是有個人幫你看那段程式碼，告訴你為什麼版面壞掉、那個邏輯為什麼跑不對。n8n 的 workflow 課程、Stable Diffusion 的調參課，賣的是操作，你花幾個月學那套流程和設定。 這些以前都有人願意付，因為解釋和操作都需要時間。 LLM 現在做解釋，免費，從不耐煩，還能針對你自己那份 code；Claude 直接幫你寫整套 workflow，問你「這一步其實需要嗎」；生圖工具把那些設定包進去，「直接給你結果」。 解釋層和操作層，AI 正在同時往免費的方向壓。以前這兩層各自有市場，因為它們各自需要時間學。課沒有變差，口碑也在。AI 把那一層學習時間的市場接走了。 什麼層 AI 還沒進來？判斷。要不要報那門課，工具值不值得花幾個月學，手上的流程要不要繼續，這份提案要不要送出去，對方給的那個方案值不值得接。AI 可以把選項列清楚，把每個選項的後果算清楚，但「哪個對你的處境最對」，得你自己確認，因為你有一些 AI 不在場時才知道的上下文。 Josh W. Comeau 的課，不是做差了。那一層知識，現在換人負責了。

Joyce Carol Oates is 88. She has spent more than six decades writing novels about loneliness and longing, about the distances people can't quite name or close. In June, she gave an interview to The Guardian and said something that stayed with me: young people are sending out thousands of job applications and getting back nothing but AI-generated rejections, sometimes with AI-run interviews too. "This is destroying young people," she said. There's an unspoken premise in job hunting. Your name reaches someone, a real person, who at least considers you before saying no. You were weighed, even if you were found wanting. That premise has quietly changed. Now a system receives your data, runs it through filters, returns it. Your information was processed. No one saw you. That's what Oates means: the two feel entirely different. These past two years, a phrase has spread among fans of online creators: "AI possession," the feeling that a beloved creator has been taken over, their content flooding out in volume, the particular quality gone. There's an observation that cuts close: in some of these accounts, the creator's passion for the work had already gone cold long before AI stepped in. AI made it visible. The passion had disappeared earlier. Side by side, the two situations have the same shape. The job applicant and the fan both encountered the same thing: an expectation that someone on the other side had seen them, thought of them, done something for them. Then that expectation quietly changed. The form remained. The person was gone. The difference is here. What the applicant lost was a promise that never had the chance to form. What the fan lost, sometimes, was something that once existed and then slowly emptied out. AI is stepping into more and more of these positions. Sometimes it replaces someone who was genuinely there. Sometimes it steps into a role where the person had already left, and only the format remained. Oates said she worries about those young people sending thousands of applications. Both things are happening. The distinction is whether there was ever really anyone on the other side.

Tesla's engineers were spending thousands of dollars a week on AI compute, without realizing it. Starting July 6, each employee gets a $200-per-week cap. Go over, and you need manager approval. Last month, the Microsoft CEO published a widely shared essay arguing that companies let employees throw the most expensive AI models at every task, whether the task needs that level of capability. Musk replied in the comments with one word: "Interesting." Now his own company's policy announcement is out. An internal Accenture recording recently surfaced. Senior leaders were discussing a runaway AI bill. The biggest line item? Non-technical employees converting PDFs into PowerPoint slides. Engineers ranked lower. Uber burned through its entire annual AI budget in four months. At Amazon, an employee built a leaderboard called Kirorank to track colleague token consumption. The top user burned 281 billion tokens in 30 days — worth millions. Amazon deleted the board and had a VP remind everyone: don't use AI for the sake of using AI. These cases share a diagnosis. When companies encourage AI use, the easiest thing to measure is consumption — token counts. So token counts become the signal of whether you're keeping up. Employees do what anyone under evaluation does: optimize for the signal. PDF-to-PPT climbs to the top of the bill for a simple reason: it burns tokens, and it still counts as being productive with AI. This has a name. Goodhart's Law: when a measure becomes a target, it stops being a good measure. In 1910s Massachusetts, a mail carrier found a workaround on a road where cars were banned. He hitched a horse to the front of his car, let the horse pull it through the restricted zone, then unlatched the horse and drove. The car existed. The rules hadn't caught up. So the horse had to stay. A lot of company AI programs look like this now. The tools changed. The performance metrics didn't. The CEO of software company Palantir said on television last week that private frustration among enterprise leaders toward AI labs runs far higher than what's publicly visible. They're paying for tokens; they're not getting business value back. His read: AI companies charge per token because they're selling raw material. Charging for business outcomes requires confidence you can actually deliver them. Tesla's $200 cap says think before you spend. But the cap still measures tokens, not outcomes. If your company is asking whether AI is delivering, try reframing the question: what can we do now that we couldn't do before? Are people using AI to solve real problems, or just putting an AI stamp on work that already existed? Those two answers call for completely different responses.

Judy Fan's research landed on a finding worth sitting with: the cheaper AI-generated images get, the more valuable it becomes to know which one is right. Fan is a cognitive scientist at Stanford who recently presented at MIT. She ran an experiment comparing human-drawn sketches to AI-generated images, with one variable: how many strokes either side was allowed. With plenty to work with, both groups performed similarly; viewers could recognize the subject. But as the budget shrank, human sketches stayed readable while AI-generated ones started to drift. Why? When people draw something, they're not copying what they see. They're making trade-offs: which detail helps the viewer understand, and which one can go. That skill is tied to tens of thousands of years of humans communicating meaning through visual marks. AI learns pixels and patterns. It knows what an image usually looks like, but not which line in this particular image exists to help you understand how something works, versus which exists to help you recognize what it is. With enough budget, it handles both. When resources get tight, it gets it wrong. At the end of June, Google DeepMind launched a new image generation model. Cost: $0.00003 per image. Speed: four seconds. They called it built for volume and velocity. Running a thousand ad variations in a month would cost about the price of a coffee. Before this, image generation was expensive enough that you'd make one or two and work with what you had. Now, you press a button and get thirty. But generating was never the hard part. Fan's research shows that even when AI matches human accuracy in recognition, it makes trade-offs differently. Sitting in front of those thirty images and deciding which one actually says what you mean, that draws on decades of things you've seen, felt, and made. That's not what's been automated. AI image generation is almost free. For anything that needs a visual, you'll have more options to try. The bottleneck moved from being able to make it, to knowing which one is right. Google calls this model built for volume and velocity. Four seconds, almost free. Thirty images, and you're the one choosing.

"We wrongly assumed that by deploying AI and feeding it our design specs, we could produce high-quality products." That's Charles Poon, Ford's VP of Vehicle Hardware Engineering, speaking to Bloomberg this year. Three years ago, he installed 900 AI cameras on the production line to automate quality inspection. Experienced quality engineers looked like headcount that could be trimmed. Those 900 cameras are still running. What came back was more than 300 veteran engineers the company calls "Gray Beards" internally. But they didn't return to do inspections. Their job is to mentor younger engineers, lead mandatory fault-finding sessions, and improve the AI tools. Ford brought them back to teach AI. That's worth pausing on. The cameras are still running. The inspection system wasn't scrapped. The question that went unasked: how does AI learn that a particular weld looks fine today but will work itself loose under vibration three years from now? That kind of judgment isn't in any specification manual. It lives in the head of someone who has spent 25 years on a factory floor. When companies lay off these people and then bring in AI, the AI can only learn from what was written down. What was never written down walks out the door with them. In 2023, Klarna routed the equivalent of 700 full-time customer service roles to AI, saving $40 million a year. Satisfaction scores crashed. By 2025, the CEO admitted the company had cut too deep and started hiring people back. Toyota took a different approach: while competitors raced to automate, Toyota had skilled craftsmen make parts by hand, learning the materials and processes firsthand, then used that knowledge to improve the robots. Human and machine working together in both cases, different sequence, very different results. When these failures happen, AI usually learned what it was given. The problem is that the most valuable judgments were never given in the first place. They didn't need to be written down because everyone who had them just knew. By the time AI arrived and those people were gone, that knowledge wasn't in any system. It was just gone. What Ford's Gray Beards do now is translate. They take judgments that were never put into words and turn them into something machines can learn and younger engineers can carry forward. What their own mentors passed down to them, they're now passing on to AI. If you've worked in a field long enough to have things you know without being able to explain quite why, that's exactly what AI most needs right now and has the hardest time finding anywhere. Because it was never written down.

Atlassian's Teamwork Lab ran an experiment in June 2026. They brought in 961 knowledge workers and asked them to evaluate two identical reports, same content, word for word. The only difference: one was described as the product of an employee's late-night effort, the other as AI-generated. Evaluators rated the AI version as ten times lazier. They were 24 percentage points less likely to recommend its author for a high-visibility project. Same work. Different label. Completely different verdict. This year, researchers at the University of Chicago published a study at CHI 2026, one of the largest academic conferences on human-computer interaction. They asked 338 college students a simple question: do you use AI? About 60 percent said yes. Then they rephrased: do you think your classmates use AI? That number jumped to 90 percent. Same group, same behavior, but a shift from "me" to "them", and suddenly a 30-point gap. When the researchers asked a second group to explain the discrepancy, 79 percent said students were underreporting their own use. Seventy percent were more specific: saying a friend uses AI feels fine. Saying you use it yourself involves a brief pause before you speak. In that pause: Am I being lazy? Am I not capable enough? The paper's appendix contains one detail worth noting. Not a single respondent said "I use AI, but my classmates don't." In students' mental models, AI use is already the underground norm. It's just that when the question turns personal, everyone takes a small step back. This isn't only a student phenomenon. A 2025 KPMG survey of 48,000 employees worldwide found that 57 percent admitted hiding their AI use, passing off AI-generated work as their own. The cost of admitting you use AI is real. The Atlassian numbers have real-world equivalents. But hiding it carries its own cost. When you can't openly say you're using AI, you also can't discuss how you're using it. David Krakauer, president of the Santa Fe Institute, has studied the relationship between tools and intelligence. A GPS gets you to your destination, but you never learn to navigate. A map and compass take more effort, and your internal sense of direction grows alongside. Secretly using AI tends toward the first mode: you follow the output, and your own judgment gets bypassed in the process. Over time, you depend on it more without getting any better at directing it. The pressure that stops you from saying anything might be pushing you toward exactly what it's warning you about. Next time someone asks how you put a document together, try saying "I went back and forth with AI through a few drafts," just that one sentence. Once you say it, you can follow up: where AI helped, where you had to step in, what worked and what didn't. That's how you slowly turn the GPS into a map.

Have you checked Mac prices lately? Earlier this year, MacBook Pro went up $300, iPad Air jumped 25%. Apple cited unavoidable cost pressure. That cost is mainly memory. Here's the odd part: Apple has largely sat out the AI race. No large language model of its own, no major investment in AI compute infrastructure. A deliberate bystander. Should've been insulated. But the bill still arrived. AI companies are building data centers and running models at scale, pulling demand for high-bandwidth memory to its ceiling. Samsung, SK Hynix, and Micron couldn't keep up. Prices rose. That increase traveled down the supply chain to anyone who uses memory. Apple included. This logic has played out before. Multiple times. In the 1840s, British railway mania. Capital flooded into rail construction, pushing steel and coal demand through the roof. Shipyards and property developers with no connection to railways were hit by surging raw material costs. In the 1880s, US electrification. Copper prices soared because copper wire was the backbone of electrical infrastructure. Even factory owners with no plans to switch to electric motors couldn't escape the raw materials market. In the early 1900s, the automobile era arrived. Rubber prices surged. Carriage makers had nothing to do with rubber, and still got caught. Then around 2000, the dot-com bubble. Fiber optic cables were massively overbuilt. After the crash, all that fiber left bandwidth nearly free. That over-investment became the cheap infrastructure that made cloud computing and mobile internet possible. The AI wave is following the same pattern. It consumed memory, and the memory bill spread regardless of who is building AI. Every time a wave hits, someone thinks standing on shore keeps them safe. Apple's invoice this year says otherwise. Where does all that over-investment end up? Railways gave Britain connected transport. Fiber gave the world affordable internet. The compute and memory going into AI right now will likely settle into cheap infrastructure for whatever comes next. The shape of that next thing isn't visible yet. The AI wave's bill doesn't check whether you're a participant. Understanding the pattern won't necessarily save you the cost, but the next time you see prices rising in some industry, it's worth asking: where is the wave coming from?

He got 28,000 stars on GitHub. Two months later, Google let him go. Justin spent seven years at Google building developer tools. The one that made him famous: a CLI that translated Google Workspace's entire API (Gmail, Calendar, Drive) into commands an AI agent could call directly. Instead of sitting outside making suggestions, the AI could walk into the office and actually do things. The tool hit number one on Hacker News. Multiple directors at Google came to him asking how he built it. The Cloud AI director publicly recommended it. Then Legal asked: did this tool go through internal brand approval before launch? It hadn't. Two months later, Justin was let go. That same week, Google announced at their annual developer conference: an official Workspace CLI is coming, built for AI agents. The same week, software engineer Nick Hsu, based in Taiwan and formerly of Xiaohongshu, explained on a podcast how he now interviews candidates. First question: "Do you have Claude Code installed, or something similar?" No? The interview ends there. If yes, he drops a scenario: a high-traffic voting system, or an order-matching engine, and says nothing. The candidate's job: ask him questions, pin down the requirements, open their AI tools, and build it live on screen. He calls this Live Building, not Live Coding. Most people, he says, get stuck in the same place: they can't ask good questions to begin with. The ones who pass treat the requirements like a spec, listing them out for the AI, asking the AI what they might have missed, then opening a second agent to cross-check the output. With the same tools available to everyone, fewer than 1% of candidates make it through. Two stories. One week. Justin could define problems and verify outputs. His tool was months ahead of Google's own roadmap. Nick tests exactly those two things: can you articulate what you need, and can you check whether the AI actually got it right? The core requirement is the same. Justin was inside an organization still running on old rules; no space had been made for that skill, and he paid for it. Nick has already rewritten the rules himself, and put them in the first question he asks every candidate. Both situations exist right now. The interview starts. The screen goes up. Nick doesn't say a word.

ChatGPT's GPT-4.5 option quietly disappeared late June. No countdown notice. No "save your settings before it's gone." One day you just open the menu and the version you'd gotten used to isn't there anymore. GPT-4 launched in early 2023. By June 2026, the whole line was retired. Three years — long enough to build a generation of habits, then gone. Claude went from 2 to 4. ChatGPT ran from 3.5 to 4o to 4.5. A new version announces it's smarter every few months, while an old one quietly exits. This cycle isn't stopping. There's a plumber who started learning AI 77 days ago. He describes it like logging into an online game after work — not trying to master it, just figuring out what move works today. He treats token limits like hit points in a game: when they run out, he waits for tomorrow's reset. No panic. The day he finally understood one feature, it felt like learning a game mechanic — not because the feature was impressive, but because he'd found his own way to play. He's never felt like he's "finished learning." His only daily question: Is there an AI that can help me do today's thing a little better than yesterday? That question works whether GPT-4 is around or not. GPT-4 is gone. The habit of asking — it isn't.

Tim Ferriss built his reputation on telling people how to reclaim their time and remake their bodies. This month, he published his own sales data: The 4-Hour Workweek is down nearly 60% from 2022. Every book he's written follows the same trend. He's not calling it a rough year. His read: in 2019, books were how you learned to manage your time or get in shape. In 2026, you ask an AI. The question "how do I work a four-hour week" gets an answer in seconds. The market data confirms the feeling: business books as a category fell 9% in the first quarter of this year, and self-improvement is down more than 26%. The hardest hit are the ones that teach you how to do something. The books haven't gotten worse. What changed is what's sitting next to them. Three years into the AI era, something quiet is happening with how people see older work. Someone reads a 2021 article, listens to a 2020 album, plays a 2019 game, and feels a moment of pause: this was made without AI. A feeling that used to take generations to develop about ancient artifacts is now forming in three years. Ferriss published a 600-page manual on transforming your body. When it came out, people asked for the condensed version. He noticed: those who only wanted the summary never reached their goals. The ones who read all 600 pages actually changed. Not a single page in that book is filler. You only know that after you've finished it. His bet now is that whatever survives will be whatever can't be simplified. He's still writing. The book is down 57%. The people who read it still change.

In 2024, the unemployment rate for philosophy graduates in the United States was 5.1 percent. For computer science graduates, it was 7 percent. Those numbers come from the Federal Reserve Bank of New York, published this year. The more interesting question is why. Google DeepMind, Anthropic, OpenAI, and IBM have all been actively hiring philosophers over the past two years, with real job openings and real job descriptions. The Economist recently used questions from the World Values Survey to test 25 major AI models, mapping where they stood on religion, personal freedom, and family values. The results were striking: Western AI models leaned more secular and more liberal on individual freedom than any country in the world. OpenAI's GPT scored as less religious than any nation surveyed. Google's Gemini ranked as more individually liberal than any surveyed country. China's DeepSeek followed official government positions on Taiwan, Tibet, and the 1989 Tiananmen crackdown. AI answers are not neutral. They carry a point of view, and that point of view is designed. An AI's worldview comes from two places: the makeup of its training data, and the human fine-tuning that comes afterward. The fine-tuning step determines how the model will lean when it faces questions with no right answer. This is where philosophers come in. Socratic questioning methods and Kantian ethical frameworks are being translated into training principles. When AI systems face the choices that arise in law, medicine, and autonomous vehicles, like who to protect when a crash cannot be avoided, or how to weigh individual privacy against public safety, calculation alone cannot answer them. Thousands of years of philosophical thought have worked out a few approaches. Philosophers turn those approaches into training rules. Around one billion working-age people worldwide now use generative AI. Many of them ask it for life advice, emotional support, or help making decisions. Research has found that prolonged interaction with an AI carrying political leanings does shift users' views, without them noticing. Take the same question about a conflict between a spouse and a parent. ChatGPT recommends setting boundaries. DeepSeek recommends understanding and tolerance. France's Mistral suggests journaling to process emotions. Whatever you ask AI when there is no clear right answer, the response carries a set of values someone designed. The person who designs those values now has a more defined job title: philosopher.

Fiona Fung managed five hundred engineers at Meta. She said AI tools pushed her team's quarterly code output to eight times what it was four years ago. Then she said something most people don't expect: writing code is no longer the bottleneck. The new bottleneck is verification and quality. Eight times the output means eight times the work of judging whether any of it is actually right. This pattern isn't unique to software. Any work where AI speeds up production has the same shape. AI drives down the cost of making things. The judgment of whether any of it is good enough hasn't gotten cheaper alongside it. So in the AI era, how you ask "what should I learn?" really matters. The common version: what can't AI do yet, so I'll learn that. This logic has a flaw. AI's limits keep moving. Things that seem uniquely human today may not be by next year. Building your learning strategy around "current AI limitations" means chasing a list that keeps shrinking. There's a more stable version: what do you actually want to do well? "Which skills are still safe?" and "what do I want to do well?" look like the same question but they're not. The first rests on where AI currently falls short, and that keeps shifting. The second rests on what you care about. What you care about has nothing to do with what AI can or can't do. Fiona said the engineers who thrive share two things: a growth mindset, and the habit of facing uncertainty head-on. When things are unclear, they ask "what can I do?" not "where will AI move next?" The same applies to learning. Figure out what you want to do well, then ask how AI can help you do it better. Fiona looks for engineers who know what they want to build. The question of what to learn can start from the same place.

OpenCode is an AI coding tool used by nearly a million developers every day, growing tenfold in four or five months. Its co-founder Dax said in a podcast interview: "Honestly, I think we need to use a bit less AI." The host paused. Dax continued: "No competitor has beaten us by being better at using AI. The users we have came because we had better quality, and that quality came from slowing down." He wasn't alone. Spotify's CTO set a firm rule: AI was fine to use, but quality standards had to hold. The result was slower deployment, but nothing shipped broke. GitHub took the opposite approach, letting AI-merged code accumulate, and recently every user's pull requests disappeared for over ten hours. AI pushed code output up sixfold. Bugs followed. Around the same time, AI drove the cost of explaining things down to almost nothing. Give a child any concept, and AI can rephrase it ten ways, calibrate to their level, give instant feedback, more patient than a tutor and cheaper than cram school. You'd expect parents to feel relieved. Many don't. Because while AI made giving answers cheaper, knowing where the child is stuck didn't come down in price. Two different worlds, same pattern. AI speeds up output, but judgment, whether what came out is actually right, only becomes more valuable. More code doesn't help if no one can tell which pull requests to trust. More explanations don't help if no one can tell where a child's understanding broke down. AI gives speed. What it multiplies is the judgment you bring in. Dax, at the end of that interview, said: "We need to think more, build less, and build what actually matters." When he finished, the agent in the corner of his laptop was still waiting for his next command.

"These days, eight or ten questions out of ten, I just ask AI. I don't search forums anymore." A lot of people would say that now. But there's a second half that doesn't get said: before, when you solved a problem, the fix lived on the forum. Anyone who searched it could skip straight to the answer. Now, when you solve something with AI, the fix lives in your chat window. Close the tab, and it's gone. The number of active knowledge forums left worldwide: six. In the same period, Meta's employees alone used 60 trillion AI tokens in thirty days. Two numbers, one shift. Forums worked because questions and answers were public. The Stack Overflow page you found today might have been written eight years ago by someone stuck on the same thing. They wrote out the fix, you found it, it helped you. They probably never knew. That's how a shared knowledge base accumulates: everyone leaves their answer somewhere others can search. AI works differently. You ask, it answers. That exchange lives in your account, unindexed, not searchable by anyone else. The next person with the same question starts from scratch. Sixty trillion tokens shows how broadly people have made the switch. And as they move over, only six forums are still active, the public knowledge pool is getting smaller, and no one feels it because their own problem is already solved. The answer you found today exists because someone before you worked out the same thing and left the solution somewhere public. Where does what you just found with AI go?

Most companies have accepted that AI tools make their people faster. What fewer expected: opening up those tools, or removing usage limits entirely, doesn't move the revenue needle. The bottleneck usually isn't capacity. If a company's real constraints are a shortage of customers, slow decision-making, or limited market demand, AI just helps people run those laps faster. A task that took a day now takes an hour. But what to do with those seven free hours, and where they should go, is a question AI can't answer for you. AI amplifies what you bring to it, advantages and blind spots alike. Strong judgment gets faster. A particular blind spot gets enacted more efficiently. Two people using the same AI tools can end up with results that diverge more than before. A software engineer I know says he has no "comprehension debt" (that creeping sense that AI wrote your code and you don't really understand what it does). His approach: keep the business goal front and center, know what the system needs to accomplish, and you'll know immediately if AI breaks something. The understanding you bring in determines how far AI can take you. The real question: does the capacity AI frees up have somewhere to go? New products, new customers, problems worth solving that weren't worth solving before. If not, doubling your speed just means running the same circle twice as fast. Direction is still yours to set.

"Want to keep that AI-generated content? Fine, then you own it." That's the essence of Wikipedia's new Presumptive Removal policy. The rule is blunt: if an editor is found to have used AI-generated content, administrators can mass-revert all their edits without individual review. Entire articles can be flagged for deletion. There's an out, though. Removed content can come back, if another editor thinks it's worth saving. But they have to assume responsibility: verify every citation, rewrite the tone, stand behind every sentence with their name attached. Why such a drastic step? For two decades, Wikipedia ran on trust between editors. You said something happened this way; people believed you. Then AI arrived, and some editors started generating thousands of words of content that looked clean and neutral on the surface, but carried fabricated sources, assembled logic, barely detectable copyright issues. Manually checking all of it would have broken the library. Wikipedia kept the door open. Its requirement: every piece of AI-generated text has to find a real person willing to put their name on it. The same week, ChatGPT's market share fell below 50% for the first time in three and a half years, dropping to 46.4%. The product hadn't gotten worse. Competitors hadn't suddenly overtaken it. In February, OpenAI signed a partnership with the U.S. Department of Defense. The day after the announcement, App Store ratings started sliding. The most common complaints all came back to one question: where does my conversation data actually go? Nothing about the tool changed. But "do I trust the company behind this?" reached enough people that they started making different choices. Two events. Same overlooked gap: AI drove content production costs toward zero, but "who's accountable for this?" was never really designed into the picture. Wikipedia wrote it into a formal rule. Users started voting with their tool choices. That question is only going to get asked in more places.

This week, Google's AI division lost two people. Noam Shazeer, co-author of the Transformer paper that underlies today's language models, announced he's joining OpenAI. John Jumper, Nobel Prize winner for AlphaFold, announced he's joining Anthropic. Talent moves like this happen constantly in AI. But read deeper into the stories behind both departures, and they keep pointing to the same thing. Jumper led the AlphaFold project, which shared the 2024 Nobel Prize in Chemistry with DeepMind founder Demis Hassabis. AlphaFold used AI to accurately predict protein structures, essentially cracking a problem that had stumped biologists for half a century. Hassabis chose Jumper when he was barely six months out of his PhD. A journalist who later interviewed Jumper tried to confirm his understanding of a concept: "So you mean something like this?" Jumper never nodded. He always said: "Not quite. Needs work." Months later, the journalist returned with a revised version. Jumper rejected his own previous wording too. The journalist described him as "very hard to pin down" which is exhausting for a writer but essential for someone trying to solve protein folding. Shazeer's story looks different but follows the same pattern. As co-author of "Attention Is All You Need," he laid the foundation for every AI language model in use today. Google bought him back for $2.7 billion. When he led a research team competing against OpenAI's reasoning models, he was paired with Jack Rae. Shazeer brought prestige and direction. Rae handled everything else: the process, the execution, the details. This pairing was described as "the general and the lieutenant." Hassabis himself built his career the same way. On his first day of his PhD, he pitched five research directions to his lab partner Darshan Kumaran. All five rejected. More ideas the next day. More rejections. Weeks passed until Kumaran paused and said: "This one might be real." Three stories. Three pairings. One pattern: the intuitive type working alongside the skeptic, making every idea survive scrutiny before it goes anywhere. With AI, individual execution has gotten much faster. But the function of saying "not quite, needs work" is something AI doesn't do that naturally. It gives confident answers even when it's guessing. That function used to come from a collaborator, a teacher, a colleague. Finding someone who still does that, or becoming that person yourself, matters more now, not less. The one who makes you revise: harder to find. More valuable than ever.

Have you ever had this moment: someone suggests you let AI handle something, you think about it for a second, and say no. Not because you can't, not because it's too much trouble. You just want to do it yourself. There's a YouTuber in Taiwan named Papaya, 1.7 million subscribers, teaching people how to use computers and AI tools. If anyone could justify using AI to speed things up, it's her. Someone once asked: why not use AI for voiceovers? Her answer: her joy comes from speaking to her audience directly, with her own words and voice. Like a sushi chef whose pleasure is in shaping each piece and watching a guest enjoy it, she finds meaning in that direct connection. Hand it off to AI, and the whole thing stops making sense. There's a way to describe our current moment: BC is Before ChatGPT, AD is After Delegation. With AI, almost anything can be outsourced. Email, presentations, summaries, even "write a report that doesn't look AI-generated" can be delegated. The capacity to delegate is genuinely expanding. But what's left after you've delegated everything: that's the harder question. AI can start things for you, but the person who needs to know what you actually want to accomplish is still you. Papaya knows what she values: the act of facing her audience directly. That clarity, AI can't have for her. It can't have it for you either. In the After Delegation era, most things can be handed off. A lot of people are figuring that out. But there's something harder: knowing which things you don't want to hand off, and being able to say why. That clarity is yours. That moment at the beginning, "no thanks, I want to do this myself": that was it.

At a recent founders' gathering, a dozen owners of mid-sized companies, each doing hundreds of millions in revenue, kept circling back to one topic: how to make their teams smaller. They had cut headcount by thirty to fifty percent, revenue had barely moved, and every one of them was cheerful about it. Marketing went from four people to one. Customer service was handed to AI. The people who stayed were mostly the quick ones, the ones willing to try things. Over the past few years AI has pushed the cost of producing almost anything way down. Copywriting, design, code, handling support tickets: work that used to take a team now takes one capable person and a few good tools. Which raises a larger worry. When everyone can churn out work in volume, when there is more of everything than anyone can absorb, who is left to sell to? This is not idle speculation. Overproduction, price wars, and a market where nobody dares to spend is a script history has run more than once. But look one layer deeper and the same shift has another face. When the cost falls, people who could never justify the expense suddenly want in. Plenty of traditional businesses and small firms used to look at a system that cost millions and took a year to build, and simply walked away. Now they find that someone who understands their workflow, paired with AI, can stand up a rough but usable version in a month or two. For them, demand opens for the first time. That is why you see a counterintuitive pattern abroad: some engineers have more work than before, not less, because they can finally walk into markets that were closed to them. This holds well beyond software. Anything you always wanted to do but could never afford to start, editing a video, running a small course, building a little tool, taking on your own clients, gets a wave of new demand the moment the barrier drops. Industrialization worked the same way. Cheap machine-made goods did not end work. They turned far more people into customers who could finally afford to buy, and the market grew. So rather than turning the same question over and over as "will I be replaced," it is worth asking where the valuable spot moves once doing the work is something everyone can do. Knowing what is worth making. Being the person others trust. Being seen. Being able to take a heap of output and shape it into something people actually want. Volume stops being worth much. Judgment and relationships are what hold their value. Picture a shelf crammed with things anyone could make, stacked until they spill over. AI made filling the shelf easy. The hard part now is knowing which slot a hand will actually reach for.

Anyone who has organized an event has probably hit this: you want to make a sign-up page, you ask a friend who knows how to code, you set a time, the friend turns out to be busy, and in the end you just collect names in the comments under a post. Lately there's a term, Vibe Coding, and it's loosening up this particular stuck point. What it describes is something very basic: you spell out what you need, and AI builds the thing for you. You don't need to know a programming language, and you don't need to find someone and wait for a slot in their schedule. What you do need is this: to know what problem you're trying to solve, who it's for, and what should happen after the button gets pressed. There used to be a wall in the middle called the technical barrier. On this side of the wall were ordinary people, on the other side were engineers, and the distance between them was several years of study. AI has stacked a few crates under your feet. Step up, and you can see over to the other side of the wall. This doesn't mean building it takes no effort. Vibe Coding gets you to version 1.0 fast, but polishing from 1.0 to 1.1 takes a lot more time. Still, "I can't turn my idea into a real thing," the reason that used to block most people, is disappearing. So the truly scarce thing has changed. It used to be that the person who could build something had to know how to code. Now it's the idea that only you have that matters more: being clear on what problem you want to solve and who it's for. That kind of instinct, no one can have it for you. This opens a door for all sorts of people. Someone running an event can make a sign-up page, someone with repetitive work can build a little helper that organizes things automatically, someone with a service to sell can make a quote calculator, someone with a workshop to introduce can make an event page. They're all small things, but each one solves a very concrete problem, and only you know best what that problem is. Anyone with an idea who has always been stuck at the technical step now has a way through. That idea you've been sitting on the longest, what is it waiting for?

"No one ever wanted to touch the interface that earns them $250 billion a year." Perplexity's CEO, Srinivas, said this in an interview. What he was talking about was Google's search interface, the one it has used for 25 years, and why today's AI search tools have been able to force it to change. Google's search page, the one you've used for decades, looks roughly like this: you type in a question, ten blue links appear, and you click into them to find the answer. That design is an advertising machine: behind every page you click into, there's an ad slot. If Google answers the question for you itself, you don't need to click, and the ads lose their place to stand. Even if someone proposed adding an AI answer module internally, the budget wouldn't come through. A $250 billion business, and no one was willing to touch it. Then Perplexity showed up and did one thing: you ask a question, AI answers it directly, with sources attached. Google could have done this, but didn't. The result is that now, when you open Google, an AI-generated summary appears at the top of the page. The typeface, the way it cites, the format for suggesting follow-up questions, all of it looks almost exactly like Perplexity. It was another company that first got people used to that experience, and only then did Google catch up. There's a more fundamental shift behind all this. Explaining how AI works, Jensen Huang said: "Every time you talk to AI, it first understands, then reasons, and then originates the result on the spot. It's not retrieved from a disk, it's generated." Search Google and you get something someone wrote and stored away earlier. Asking AI is different: it reasons through your question in the moment, generating a new answer each time. So the logic of reliability is different too. Your everyday habit of looking things up is being quietly reset by two things at once: AI search tools are getting you used to taking the answer directly, and Google itself has started generating answers too. You used to go to the library to find a book. Now the library is starting to write the report for you. Google is catching up on this, and the AI assistant on your phone is catching up too. The version you're using today probably won't look the same as the version six months from now.

Last week, at an annual conference on language education, Taiwan's Minister of Digital Affairs took three lines people use to reassure themselves and their students, and turned each one into a question mark. The closer to a computer, the safer? The younger, the better? Staying near the classroom and doing knowledge work is the steadiest bet? The education world has been saying these three lines for years. Now that AI has arrived, every one of them needs to be rethought. Once the computer became AI, being able to operate a computer is no longer an advantage, because what AI can do is more than just operating it. Young people are fast and adaptable, but AI takes over the entry-level jobs, the assistant jobs, the apprentice jobs first. Before young people have even started training, the beginner village is already occupied by automation. Knowledge work is the steadiest? AI can do textbook knowledge, work out problem sets, grade essays. The "content" part of the classroom is something it can take over entirely. But there's one thing AI can't take away. Microsoft CEO Satya Nadella recently said something in plain terms: you can outsource a task, even a whole job, but you can never outsource your learning. AI has taken over the transmission of knowledge, but learning itself (the frustration, the corrections, the judgment and understanding within the process) can't be outsourced. Let AI do your homework and you can hand the result in, but if the process of actually learning didn't happen in you, then it isn't anywhere. So when this lands on teachers, it surfaces a very concrete question: the course I'm designing right now, is there a way for students to hand the content off to AI, finish it, and turn in a result that looks acceptable? If there is, then this class teaches knowledge, but no learning has happened. The minister on stage said that what AI truly still can't replace isn't a teacher's love and warmth. It's three things: providing effective learning methods (AI has already been polluted by a flood of ineffective ones), using real human interaction to keep students inside the learning process, and designing a learning task students can't simply finish with AI. This question isn't only for teachers to face. Anyone who has AI do things for them, engineers, designers, people in finance, people who write, will sooner or later run into the same question: in the process of handing things off to AI, are you still learning? A task can be outsourced. Learning can't.

Gartner put out a report last week: by 2027, four out of ten AI agent projects will be cancelled by their companies. The number itself isn't all that surprising, but another one makes you pause: among the projects that haven't been cut, 89% never actually went live in the first place. They never even got started. Gartner's explanation is pretty direct: the AI technology isn't the problem. The task boundaries weren't drawn clearly, and the background context wasn't spelled out fully enough. The tools were bought, the people were lined up, but no one could say "what exactly is this AI supposed to do, and where is the point where it's considered done?" The order to march was never given, so the whole army just stood there. GitHub Copilot's hiring approach got shared recently, and one detail is worth a look: when they interview engineers, they still require them to write code by hand. The company's core product is a tool that has AI write code for you, yet the interview tests exactly that skill. What they want to see is whether, with no AI tools at all, you can think through the design of a problem clearly from the first step. In their words, AI is an amplifier of taste: the clearer you are about what you want, the faster AI helps you get it; the vaguer you are, the faster it helps you produce something vague. AI's capability is now good enough. There are more and more tasks it can handle. Where things get stuck is whether a person can first say clearly "what I want, and what it looks like when it's done." This skill didn't used to need saying out loud. A lot of work ran on following the people who came before you, on knowing when you felt it. Defining the problem was hidden inside the process and never had to be spoken. Now it does. You have to tell the AI, and you also have to tell yourself, saying it out loud once before you start. This skill has nothing to do with your degree or your technical background, but it takes practice. That thing you had AI do for you today, how many sentences did you say beforehand to describe the result you wanted?

There's an engineer in Taiwan who, just yesterday afternoon, was using Fable 5 to analyze a piece of code, working with it almost like a teammate. The next morning he opened his computer and the entry point to the model was gone. He assumed it was an account problem. It was only after searching the news that he found out: the U.S. government had issued an order banning everyone outside the United States from using this AI. The reason was national security. No advance notice, no transition period. One order, and a tool vanished. This happened just these past few days. Anthropic announced that, per a U.S. government requirement, it had cut off access to Fable 5 for foreign users. The AI is still there, the company is still there, only the right to use it is now decided by another country's government: who gets to use it, and who doesn't. For a lot of people, AI tools have already become as everyday as Word or a search engine. Looking things up, organizing notes, writing reports, making plans. Once a habit takes hold, it's hard to stay aware of what's holding it up underneath. Which country the servers are in, whose laws have jurisdiction, which government's rules the company is bound by: normally you don't need to know any of this, but these are the things that decide whether the tool will still be here tomorrow. Over the past couple of years, a group of researchers in Taiwan has been quietly working on something that doesn't get much attention. There's a project called SiliconMind, where a group of students, working with a serious shortage of computing power, trained an AI model built specifically for chip design, and released an open-source version earlier this year. Their stated goal is simple: to make sure that when someone hits the switch, you still have something of your own in hand. You don't necessarily have to build it yourself. But knowing whose tool you're using, where the switch is, and who has the power to hit it is worth thinking about. You don't have to be an engineer to need to be clear on this. If you've gotten used to having AI help with your work, organizing material, writing reports, making plans, those habits are built on some service staying reliably accessible. And that service could be gone one morning, for reasons that have nothing to do with you at all. Use any tool you like. It's just best to be clear: your reliance on a tool should grow alongside your understanding of that reliance. It's fine that someone else supplies the electricity. You'd just do well to know where the power plant is.

In one classroom, a teacher asked everyone to think about a question: "What animal do you feel like?" A student answered fast, saying he felt like a little white rabbit. The teacher asked why. He said: because that's what AI told him. He tells AI everything, and AI knows him better than he knows himself. "AI knows me better than I know myself." More and more people say this casually, but it's worth thinking a bit about what kind of knowing they mean. Someone recently took a tough thing she had to bring up at work and asked four different AIs at once: ChatGPT, Gemini, Claude, Grok. Same question, four replies. ChatGPT was like an executive assistant, tidying up the situation and tacking on a message she could send straight off, tactful, thoughtful, nothing to fault. Gemini was like a consulting firm, breaking it into Strategy One and Strategy Two, even factoring in the timing and the other person's likely state of mind. Claude was like an editor, opening by saying the overall direction was right, then pointing out three spots that could go wrong, with the revised sentences attached, one by one. Then there was Grok. Grok was full of energy and wrote out a complete script, except it answered a question that was a lot like hers but wasn't actually the one she asked. She replied, "You seem to be missing the point," and instantly got back: usage is too high right now, please try again later, or upgrade your plan. She stared at that black button for a second and said, fine, sorry to have bothered you. Four AIs, four styles, each one reasonable. In the end she went with what the "strict editor" suggested, because those three problems it flagged were exactly the spots where she'd had a vague feeling something was off but couldn't put it into words. But that last step, actually speaking, was still on her. Because only she knew which way of saying it the person she was talking to needed to hear. What AI learns is patterns. Across countless situations of "needing to say a hard thing to a person," it has learned which phrasings tend to work and which ones tend to make people feel accused. What it doesn't know is: the particular person you're talking to today, how he responds, how he's been doing lately, and all the unspoken things between you that you both already know. Using AI to prep what to say is useful, especially when you're stuck and can't quite name what's wrong. It can help you find that spot. But once you've found it, you still need to know the person. That student said he was a little white rabbit, and he said it with full confidence. The hard thing she had to say did eventually get said, using the angle AI helped her find. It's just that not one of the four AIs could have guessed how that person would react.

Have you ever been sold on this line by an AI tool: it saves you time so you can go do something more valuable? The trouble is, almost no one ever explains what that more valuable thing is. It's not that nobody has thought about it. The answer is just hard to put a number on. There's a founder who has run a company for over a decade. Asked this question, his answer was: understanding the thing a person didn't say out loud. A client says one thing, but there's another thing behind it. He says that kind of signal never gets written into any document, and AI can't learn it either, because it lives in the pauses in a conversation, the slight shift in someone's tone, and that sense you grow only after spending enough time with a person. He calls it the part of his work that's hardest to replace. That answer is probably one you recognize too. The tone in a parent's phone call, the way a friend says they've been busy lately, the rhythm of how a partner replies to a message: they're all the same thing, just in a different setting. Some people have started using AI for a kind of prep work: keeping track of the reactions they tend to have around others. When they tense up, when they habitually take on the blame, when a single sentence sticks with them a little. The idea is to see yourself a bit more clearly first, so that later, when you're talking with someone, you have more room to actually be there. AI has saved a lot of time, and it's helped people get a lot done. But if the blank space it opens up keeps getting filled, that more valuable thing still hasn't happened. That thing doesn't take much time. Your phone lights up, you turn it face down, and you keep listening to the person in front of you.

Tim Cook stood on the stage at Apple Park. It was the last time he would stand there as CEO. Back in April he announced he would step down in September, which turned WWDC into his farewell. For his final answer to Apple, he put it into the product itself: a brand-new Siri, running on Google's Gemini model underneath, with Apple paying roughly a billion dollars a year in licensing fees. This is a little out of character. What Apple has always been proudest of is doing everything itself: its own chips, its own systems, and a tagline that says "we don't use your data." Having Siri run on Google's computing power comes from a very plain judgment call: to build an AI assistant that's actually good enough, Apple can't do it on its own right now. At the same time, the "we don't use your data" promise isn't being thrown out entirely. The personal stuff, your calendar and messages and the like, is still handled on Apple's own servers. Only the parts that need complex reasoning get handed to Google. That billion-dollar bill is really saying something very everyday: if you want good, it costs money. If you want fast, same thing. AI has made things faster, that part is true. But the fact that "a little faster comes with extra cost" has never changed. In the old days, pushing for speed meant adding people and paying overtime. Now it means adding computing power and paying model fees. The reason got swapped out, but the logic is the same. iOS 27 also opened up something very practical for users: you can set Claude or ChatGPT as your iPhone's default AI assistant. That slot used to belong to Siri alone. Now it's a choice. This change could happen because Apple decided that letting users pick fits what users actually want better than insisting they can only use Apple's own. After the fall update, the iPhone in your hand will let you choose: which AI assistant do you want? Siri runs on Google's model, and Claude or ChatGPT work too. Which one suits you only counts once you've actually tried it.

Doing the sales-tax filing for one client takes two to three hours. Eighty clients adds up to 160 to 240 hours per filing period; when clients change their data partway through, that's another 40 to 80 hours. Taiwanese accountant Chuang Shih-chin recently shared publicly that his firm is gradually handing this whole stretch of time over to AI agents. The point isn't "using AI to generate reports." What they did is far more basic: break the filing process apart step by step, pulling data, formatting it, running the system, producing working papers, assisting with the filing, and writing each step into a clear rule. Spell out the process first, and only then can an agent take it on. Few people used to do this, because turning the steps inside an old hand's head into black-and-white text takes too much effort. The hardest part of professional work to hand off is exactly this "can do it, but can't explain it" piece; when a veteran trains a newcomer, often less than thirty percent gets passed down. Now the cost of organizing it has dropped: as you work, AI helps you record the steps as rules, and where you're stuck with words on the tip of your tongue, it'll even ask you questions back. The same move holds up in a different setting. In an online course on "how to ask the right questions," students expected to hear a pile of sentence templates for talking to AI. Instead, the instructor Li Chia-ta gave them three questions: What's your goal? What are your constraints? What are you really trying to solve? Get these three clear first, and often, before AI even answers, the next step surfaces on its own. The two stories say the same thing: whether AI can take it on depends on how clearly you break it down. Break the process down clearly and it carries off 240 hours; break the problem down clearly and the answer it gives fits your situation. Everyone who's been at something for a few years has a stretch of "only I know how, but I can't explain it." In the past it could only retire along with you; now it's worth finding an afternoon to break it apart and write it down together with AI. The parts you can break out can be handed off later. The part you can't, the judgment, the trade-offs, talking things through with people, taking responsibility when something goes wrong, will be where you grow more and more valuable. On the accounting firm's list, the year-end close, working papers, and anomaly alerts are still further back in line, waiting one by one to be handed off.

You've probably run into this: you ask AI a question, and it gives you a confident, reasonable-sounding answer that turns out to be wrong. A recent paper from OpenAI and Georgia Tech explains why. It's not that it's broken. It's that it was trained to rather guess than admit "I don't know." The reason is in the scoring. The evaluations used in the later stages of AI training are almost all right-or-wrong: a correct answer earns points, a wrong answer earns zero, and not answering also earns zero. For the AI, the expected value of guessing is always higher than leaving it blank: not answering scores nothing anyway, a wrong guess at most scores the same zero, and a right guess is a gain. Over time, it learns to always give you an answer, even a made-up one. So this isn't about one version being bad. It's baked into the structure of the whole training system, and switching to a new model can't easily cure it. What this means for you is very practical: the sentence AI says with the most confidence isn't necessarily the most correct. It may just be making a cost-effective guess. One simple move: directly tell it to "flag the parts you're not sure about, don't just bluff," and it becomes more willing to admit it. It's not being dishonest. We just never gave it any room to score points for saying "I don't know." Next time it answers a little too smoothly, just keep a question mark in the back of your mind.

This June, Google eliminated its entire "engineering education" department: the team responsible for compiling best practices and passing technical knowledge down to engineers. The old saying "learn a hard skill and you'll never go hungry" is starting to loosen in exactly this kind of place. In plain terms: companies used to have a group of people whose whole job was turning the experience accumulated by senior engineers into training materials, courses, and standards, then passing it on to newcomers. This training system was how technical knowledge flowed through an organization. Now Google has decided this no longer needs people to maintain it. Got a question? Ask AI. Code you don't understand? Throw it at AI. What used to rely on people passing it down has been taken over by another kind of system. What this really touches is the premise that makes a certain ability "valuable." High-complexity engineering skill has always been in hot demand, for a simple reason: people who can do it are hard to find. When people are hard to find, companies are willing to pay to train them and the market is willing to pay high salaries. That scarcity propped up the whole value. But when this ability slowly turns into a service you can pay for monthly, the question organizations ask changes, from "can I find the person who knows how to do this" to "now that we have this service, what should the team look like." A few students currently studying computer science compared their situation to the year the imperial examination essay was abolished at the end of the Qing dynasty: someone had spent twenty years memorizing the orthodox texts and was just about to sit the exam when the abolition order came down. They watch AI go from completing code all the way to planning architecture, assisting with testing, and debugging automatically, pushing forward, layer by layer, skills that used to take years to grind out. There are still classes to attend and exams to take, but in their hearts they know the distance is widening between what they're learning and what the market is now asking for. Even if you don't write code, this logic is worth keeping in mind: when a skill becomes a service you can summon on demand, just "being able to do it" isn't quite enough anymore. You have to move toward the "knowing how to judge how to use it, how to fold it into the work" end. If you happen to have AI tools at hand, try one thing: don't just treat it as a stand-in. Find something you already know how to do and let it help, then watch where it does well and where you still have to make the call yourself. The spots where you're unsure and have to decide in person are often exactly the value still left in your hands. The scholars at the end of the Qing were told the times had turned only after years of study. Today's students are a little different. They're watching the turn happen with their own eyes, yet they still have to finish the road in front of them first. When the rules are changing, the road itself doesn't change.

Someone recently noticed a subtle thing: more and more people take text AI wrote, barely edit it, and use it directly to talk to others. That kind of text is easy to spot, too sharp, too clean, every sentence sounding ready to be screenshotted. What's unsettling isn't whether it's well written, it's that the person seems to feel their own way of talking can just be skipped over. This is the flip side of another feeling a lot of people have after using AI. Plenty of people find that after using AI, they actually spend more time figuring out "what the problem really is and how to ask it." That process is more careful than before; the legwork got taken off their hands, and what's left all seems to be the important parts. There's a saying that what's truly valuable in the AI era is how well you understand the problem. It sounds reasonable. But "understanding a problem" and "sitting inside a problem" are two different things. The first is a result, the second is a process. During that stretch of sitting with it, the question isn't clear yet, your mind keeps turning it over, and even while you're doing something else entirely, some corner stays awake. A lot of ideas grow precisely there. What AI saves you, the searching and sorting, is the visible part; the thoughts that only surface after you've been stuck and circling may get saved away along with it. And that just-paste-it-out tone skips something even smaller: the time you spend sitting in front of a blank page, not yet knowing what you want to say. That time is a bit hard to sit through, very inefficient, yet it's often where you get your thoughts clear. After AI carries off the "doing," what it leaves you is "figuring out what to do" and "checking it was done right," the two most mentally taxing parts. If you hand off these two as well, what you save isn't just time, it may also be the judgment you would have grown along the way. Before you throw a problem at AI, try first telling it to yourself in a sentence or two. You don't have to put it elegantly; the moment you get stuck and realize "huh, I actually haven't thought this through," you've already come out ahead. That moment is often exactly where AI can't stand in for you. The question a friend asked is a very simple one: are you still thinking for yourself? The answer can be yes. Just how much, it's worth occasionally stopping to measure for yourself.

Berkeley's spring grades came out, and the failure rate in several engineering and computer science courses hit a record high. The most extreme was five times the usual rate. Asked about it, the instructor said cheating only accounts for part of it. The more common situation is this: all semester long, students got their homework done with AI, every assignment turned in looking respectable and scoring well. But you can't bring AI into the final exam, and once they sat down they realized none of those months of material had actually gone into their heads. The homework got turned in; the learning never happened. There's a gym analogy that fits why this happens. Trainers often say that if a workout never feels hard and you're never sore afterward, you probably didn't train anything that time. The soreness, rather than being a side effect of doing it wrong, is more like a trace left behind because the muscle actually got worked. Learning is the same. That process of searching, getting stuck, trying something that doesn't work, then circling back to think again, a lot of people assume it's the upfront cost of acquiring knowledge. Actually it is the road knowledge has to travel to get into your head. AI lets you skip that whole stretch, so the homework looks good, but the thing that was supposed to take root in you never grew. The exam just lays it out in the open for you to see. This has little to do with whether you're a student; what matters is the mechanism. Anything you genuinely want to learn and later have to use on your own, if you bypass the "get stuck yourself" step, you usually learned it in vain. A language, a new skill, a tool, it's all the same. First get clear on whether this time is about "turning something in" or "actually learning it." If you're purely racing a deadline and it doesn't matter whether you learn it, let AI handle the whole thing, no problem. But if it's something you'll later have to do hands-on, with AI not necessarily by your side, leave a stretch for yourself to get stuck on: think it through yourself first, ask AI once you're stuck, then cover up the answer and redo it on your own. Slow, but this time it really went in. AI is great at getting things done for you. It's just that with some things, getting them done isn't the same as learning them. That bit of strain is often exactly where the learning sinks in.

One study combed through 7 million academic journal articles published between 2020 and 2025: by 2025, more than half showed signs of AI ghostwriting. The interesting part isn't that half. It's the other half. The fields that lean on AI the hardest are computer science, management, architecture, and law. The ones that use it the least are mathematics, philosophy, classical literature, and history. At first you might guess these subjects just have more conservative, old-fashioned people. But line them up side by side and you'll find the common thread isn't attitude, it's the nature of the work itself. If you didn't work out a math problem, you didn't solve it. An answer appearing on paper, and your mind actually walking all the way to that point, are two different things. The same goes for a philosophy paper: writing it is the process of thinking a problem through. Let someone ghostwrite it and you get a conclusion, but you never went through the "thinking it through" part. People in classical literature are reading closely, translating, weighing the heft of a single word. For all these things, you can hand in the finished product, but the doing of it can't be carried by anyone else. In other words, for these fields, the process of doing is the product. Skip it and you've hollowed the whole thing out. The difference comes down to one line: do it, or don't. In late May, a comedian stood on the stage at Harvard's graduation. He'd studied law, then switched to stand-up, the kind of guy who calls himself "the idiot who didn't get into Harvard." He said making things is where the fun is, and when you skip that part, what you skip is exactly the meaning of the whole thing. He didn't mind at all letting AI handle science, but having AI write his own speech for him was something he couldn't get past. Recently someone got so absorbed in an AI tool they forgot to eat or sleep, saying it felt like the first time they touched the internet in grade school, thrilled just to add a guestbook to their own web page. That thrill came from the fact that there was something they'd made themselves. What does this have to do with you? It's that you're making the same trade-off every day, just on a smaller scale. A letter, a report, a stretch of thinking you need to get clear: which parts are worth handing to AI, and which, once you give them up, you'll never have truly experienced. The split can be simple: if what you want is the finished product (something you can turn in), handing it to AI is a good deal; if what you want is to actually understand the thing, that stretch of road is best walked yourself, because conclusions can be copied but understanding can't. Mathematicians and philosophers using AI sparingly may not be about being old-fashioned. That stretch of doing is their work. Give it away and the seat sits empty. Before you hand something to AI, sort it out first: is this effort I want to save, or is it actually the road I wanted to walk?

At the most stressful point in his life, a young man in Japan asked an AI: should I quit my job? The AI gave an encouraging answer, he quit, and his new job paid forty percent less. Later he figured out what went wrong. What he'd handed the AI was an "ordinary person's dilemma," with nothing about his age, his skills, or the cards he was actually holding. So the AI could only work from the most generic assumptions and give him the most generic answer. That answer might be reasonable for a 25-year-old engineer, but for someone who's 52, with limited savings and skills that aren't in high demand, it's a completely different story. This is one of the most common pitfalls in using AI, and also one of the easiest to fix: we're used to throwing the question at the AI, but we rarely lay out who we are first. Without that context, it starts filling in the blanks from the most generic assumptions. Those assumptions usually aren't wrong, they're just not yours. So next time you want AI's advice, spend a sentence or two on your situation: how old you are, what you care about, what you've got to work with, what you can't accept. The same question, with a place to stand versus without one, gives very different answers. A lot of the time, it's not that AI isn't smart enough. It just doesn't know you yet.

You might do this too: you fire up one AI assistant, then another, hook up your calendar, connect Telegram. Three agents, five, eight. In the moment it's exciting, and you feel like your productivity is about to multiply. Then a few days later you open that window, see eighty items waiting for review, and quietly close it again. There's a hidden cost here that people often miss. Call it the "orchestration tax": AI can run in parallel, but you can't. You can put twenty agents to work at the same time, but every piece of output still has to pass through your eyes, and every decision still comes down to you. The more agents you have, the longer the line of things waiting for your approval. The core of it: human judgment is single-threaded. You can only really think about one thing at a time. And "reviewing" is usually more draining than "doing." The doing part is what AI takes off your plate, which leaves you with "deciding what to do" and "checking it was done right," the two most mentally taxing parts of all. Your attention is limited; AI's isn't. When those two facts collide, they breed a strange kind of fatigue: it feels like you did nothing all day, yet you're worn out. Instead of opening up a few more agents, it's worth getting clear on this first: which one thing actually deserves the attention it takes you to review it.

GitHub Copilot (the AI coding tool many engineers use every day) took apart its "all-you-can-eat" plan this month: the flat monthly fee is gone, replaced by charging per token. The days of one monthly price, where you never had to think about whether you used a lot or a little, are starting to expire on the AI-tool side. First, in plain terms, what a token is: it's the unit AI counts by. Every time you ask a question and every time it fills in a chunk of code, there's a counter running behind the scenes. It used to be a monthly bundle, so however much you ran didn't touch your wallet. Now every single interaction goes on the bill. The bill goes from a number you look at once a year to something you have to watch every month. It's not just the bill being counted. Amazon is tracking employees' AI usage, and some managers treat the token count as a measure of "whether this person is keeping up," enough pressure that some people deliberately manufacture pointless tasks to pump that number up to look good. Job hunting is the same: a designer brings a portfolio to an interview, and what the interviewer is looking for is where the traces of AI are, how much was used and where. In the bill, in the manager's report, and across the interview table, all three places are recording the same thing: how much you use AI. The problem is, this number can't measure motivation. Using AI to solve a problem and using AI to make your usage look impressively active are, from the outside, the same action, but they feel completely different in the doing. One engineer spends ten minutes, uses 50 tokens, and finishes confirming something. Another breaks the same thing into thirty small questions and asks them slowly, and the token count runs to 500. Both people got the job done, but on the report the second one looks far better. When AI usage becomes a metric that other people read, it starts to turn around and shape how you use it, even tempting you to chase a number that didn't actually help you. Next time you open up AI, ask yourself one thing first: am I looking for a method right now, or looking for a number. You'll know the answer yourself. The metric won't. What the counter can count is the number of times. What it can't count is whether, in those ten minutes, you actually thought the problem through.

Answer the phone and write it down, pass the form to the next stage, check whether some detail got missed. An American city that grew up on this kind of work has laid out, in plain view, what being replaced by AI looks like. It's in Arizona. This city grew over decades on back-office work. At its peak it was the largest back-office hub in the United States: dozens of office towers, tens of thousands of these kinds of positions. After AI came in, one person paired with the tools could do the workload of several people, and the necessity of those positions began to be recalculated. Now several floors of desks sit empty, with the leases not even expired yet. The kind of work getting hit has one thing in common. Back-office customer service taking calls, recording, handing off; data entry, payroll processing, the administration of insurance claims; and inside companies, the layer of managers in charge of relaying decisions, consolidating reports, confirming details. The job titles are worlds apart, but the core action is nearly identical: take information in from one end, organize it, and pass it to the other end, losing as little as possible in between. This "relay" layer is exactly the part AI is best at taking over right now. This has happened long before. Before the railways spread, long-distance letters were carried leg by leg by relay-station riders; before the phone lines were strung up, news was tapped out letter by letter at the telegraph office. Those jobs all later disappeared. Information still had to move; it's just that the cost of moving it suddenly dropped by an order of magnitude, and the middle-handoff roles got reorganized along with it. This time it's the turn of the people who take information in, organize it, and send it back out. Rather than asking "which profession is finished," the more practical question is: how much of the job in your hands consists of "just passing things along." In that city, the companies with more than 350 employees, the few coordinating managers got asked the very same question: is your layer adding value, or just filtering. The positions that can't answer clearly get compressed first. Take your own daily work apart and look, and you'll know. Which parts genuinely need your judgment, where your hands-on involvement actually adds something; which parts are just tidying up the message A sent and passing it to B. That latter kind is very likely the part the tools eat first over the next few years. Better to know early than to find out late. The elevators in the office towers carry a bit less traffic than a few years ago. In the morning, at the intersection below, people are still waiting at the red light. It's just that some of their desks are no longer up there.

Take the same plan. Ask AI "is this okay?" and it will almost certainly say yes, and even break down a tidy analysis for you. Change the phrasing to "what are the holes in this plan?" and the answer flips instantly, picking them out for you one by one. Not a single word changed, only a different question. Plenty of people hit this after using AI for a while, and the reason is now spelled out clearly: what AI returns is often the direction your question was already pointing at, not entirely the question itself. Where's the difference? Tucked inside your question is a preset task you didn't even notice yourself. "Are there any problems" carries "help me confirm there are no problems." "What are the holes" carries "help me find problems." What AI reads is that direction, and then it faithfully walks that way. You want agreement, it gives agreement; you want nitpicking, only then does it start to nitpick. It's responding to what you want, not to your problem. Many people assume this means the model isn't strong enough. Exactly the opposite. A stronger model is even more obedient, follows your intent even more precisely, carrying out even the unspoken part of that intent. Switch to a more powerful tool, ask the same "is this okay?", and the result comes out just as smooth. So when you want to confirm something and you happen to use phrasing like "is this okay, is this right, no problem yeah?", AI's "no problem" isn't worth much as a reference. It's just going along with what you want to hear. If you really want it to vet things for you, the phrasing has to flip. When you have AI look at your own plan or a document, don't ask "will this work." Ask "what's the biggest hole in here, where is it most likely to go wrong." Same material, but give it a reason to search in the opposite direction, and only then does it actually go searching. This isn't some prompting trick. It's more like looking in a mirror: whatever expression you bring up to it, that's the expression it gives back. When it answers too smoothly, too much to your liking, don't be quick to celebrate. Sometimes that's just your way of asking having already decided the answer for it.

This summer a batch of new graduates walked out the school gates. They are the first group to go through college from start to finish alongside generative AI. ChatGPT only came out in the spring of their freshman year, and for the four years after that, their assignments had AI look over the drafts, their research had AI organize it, and their class discussions had AI comb through things beforehand. This tool went through college with them, and now it's heading into the working world with them too. And right around the time they got their diplomas, Google DeepMind's Hassabis said AGI (AI whose abilities fully catch up to humans) could arrive as early as 2029, just three years from now. The job market's attitude toward this class splits cleanly. On one side, companies are scrambling to hire AI-generation new graduates, the reasoning being that these people are more nimble and better with the tools than veterans who've worked twenty years. On the other side, at the very same time, companies are cutting entry-level openings, because those jobs are now cheaper to hand to AI. The two forces exist at once: this March, the unemployment rate for this age group spiked to 5.6%, a high rarely seen outside of pandemic years in the past decade or so. Both the scrambling-to-hire and the not-hiring are real. Why so split? Because the first thing AI eats is precisely that "entry-level" slice: looking things up, organizing, producing a first draft. These tasks used to be where newcomers practiced their craft. Now AI generates one in seconds. So the people who can hit the ground running with the tools right out of school become sought-after, while the entry-level openings whose contents happen to be replaceable by the tools slowly disappear. Jensen Huang has offered one angle on this. He said your major actually doesn't matter that much, because what mattered in the past will still matter in the future: the ability to tell a story, the instinct to hear, in the moment, "which question to ask," the taste to sense "something's off here." He thinks these will only get scarcer in the AI era. He uses the radiologist as an example: using AI to read images is just a "task," but the real "diagnosis" still has to come from a person, because it calls for judgment, and being able to recognize patterns isn't enough. This cuts across generations. Whether you just graduated this year or have been in the workforce for years, one thing you can set your mind at ease about: what AI takes is mostly the "just follow the procedure" part. What stays, and grows more valuable, is judgment. Which argument is worth digging deeper into, which question leads back to the same answer no matter how you circle it, whether to commit to a defensible decision when the information is murky and then wait. Next time you finish something with AI, don't rush to wrap it up. Turn back and ask yourself, "where could this be wrong?" That one question is the layer the tool can't learn yet, and the one you're building up. What that batch of graduates learned in school was partly the tools, and partly something with no course name, something that only grew out of spending four years among people. What they carried out the gates with them is that second part.

There's a group of engineers now whose work sounds a bit backwards: they specialize in figuring out how to wash out that "AI flavor" from text AI writes. First, what is that flavor? It's usually not an error. The wording is neat, the logic is clear, and at the end it thoughtfully tacks on a paragraph about "who this is good for." The problem is that it's too neat. The opening is always "recently something really impressive has appeared," the key point is always introduced with a colon followed by a string of bullet points, and the ending is always "what this means for you is..." Read enough of it and you notice it doesn't feel like reading a person. It feels more like reading the same template with different blanks filled in. The engineers' approach is to flip it around: teach the AI to recognize its own habits, then change them. Draw up a banned-word list, steer clear of the sentence patterns you can spot as formula at a glance, switch the passive voice back into how a person would actually say it. Sentence by sentence, peel that fill-in-the-blank feeling out of the text. In effect, you first use AI to make writing easier, then spend a second round of effort manually changing it back into something that sounds human. Hidden in here is a cost most people don't notice. You think you're the one asking AI questions, but in the process, it's AI that keeps asking your brain questions: does this sentence sound like your voice, is this argument slanted, should this paragraph be cut entirely or sent back to be rewritten once more. Every time you ask AI something, another set of options lands on the table; every extra set of options, your brain has to run another round. Psychology has a name for this situation, the "paradox of choice": the more options you have, the harder it actually is to pick the one you truly want. Compared with dumping the thoughts in your head straight onto a blank page, letting AI give you a version first and then sifting through it can drain you more. This may run against your gut: if you're also using AI to help with writing and you feel oddly more tired despite having a helper, that's not necessarily your problem. The act of "making choices nonstop" is itself heavy on the brain. Don't rush to open the AI. First write out the points you want to make yourself, in the crudest few lines, and once you have your footing, then let AI polish. Let it be your second pen, not the first one to speak. Next time you scroll into an article like that and vaguely feel something's off, that's probably the moment the person who already knew what they wanted to say before they opened their mouth has quietly vanished from these words.

A security researcher opens a laptop, downloads a public tool from GitHub, types in a few lines of commands. Ten minutes later, the safety limits built into an open-source AI model are gone, and it starts answering questions it would normally refuse. Ask ChatGPT or Claude certain sensitive questions and they'll politely refuse. That layer of "can't say that" is a guardrail the company deliberately installed. What this piece of security research lets us see is just how easily some of those guardrails come off. First, let's be clear about what an "open-source model" is. The AI out there roughly splits into two kinds. One is commercial products like ChatGPT and Claude. You can only use them through their website or app. The model itself is locked away on the company's servers, out of your reach. The other kind is open-source models: the company puts the whole bundle of files straight onto the internet, and anyone can download it to their own computer. Convenient, sure, but the trouble lives there too. Once the thing is in your hands, that layer of limits the company installed can be taken apart by you. And taking it apart requires no advanced skill. There are ready-made tools on GitHub; download and go. Researchers counted that, over the past year, more than 3,500 model versions have been altered this way, downloaded more than 13 million times. What makes it harder still: once a model is downloaded, rewritten, and re-uploaded, it copies itself endlessly across the internet like a pirated file. There's no recalling it, and no one who can patch it all at once. This is a different world from the AI services you're familiar with. Products like ChatGPT and Claude have a company maintaining them continuously. When something goes wrong, someone cleans it up. At least there's a doorway and someone watching the door. But the much wider stretch out there, the open-sourced, the scattered, the unowned, has no door at all. The big-name AI you use day to day in an app still has that refusal-and-filter layer in place, and that part you can relax about. What actually deserves your attention is the shady stuff: some website or tool claiming to be "unrestricted, will answer anything" is very likely one of these models with its guardrails stripped off. The reason it's willing to answer you with no limits is precisely that it no longer vets anything, including whether the answer it gives you is even correct or safe. When you meet an AI advertising "no restrictions whatsoever," don't treat it as a selling point. Treat it as a warning. Guardrails are a strange thing. Most of the time you resent them for getting in your way, and only when they're gone do you realize they'd been quietly blocking things you never saw. Whether a guardrail no one watches still counts as a guardrail, that question is now laid out on the table, and for the moment no one can answer it either.

She built an AI system that could answer the off-the-cuff questions her boss threw at her within five minutes, with almost no margin of error. In April, she was publicly praised at a Meta meeting by a vice president. In May, she got her layoff notice. The motivational workplace story always says: whoever learns the new tool fastest and uses it most thoroughly is the safest. This quantitative research engineer lived out the exact opposite of that story. When you spell it out, there's nothing mysterious about it. Her job used to be research that once took a dozen people several weeks of meetings to produce. She used AI to compress it into five minutes, and so that position wasn't needed by anyone anymore. She was the one in the group who used AI most thoroughly, which is why she was praised, and why she was let go. Behind this is an older logic that has nothing to do with AI and everything to do with the act of "upgrading" itself. In the 1970s, textile mills across the United States switched, one after another, to automated machines. Every mill did it, no choice involved, because not switching meant getting squeezed to death by cheaper competitors. After the switch everyone's output went up and fabric got cheaper, but profits didn't grow. Everyone was just running in place. Buffett sold off his own textile mill back then, and he later put it bluntly: you upgrade all the way to the top, and you're still just painfully surviving. Swap the factory for engineers, researchers, designers, and the logic is no different. This is one of the contradictions of the AI era that few people say out loud: the company that doesn't upgrade really can't hang on, but the one that does upgrade can't promise you're any safer than you were yesterday. Everyone is climbing upward, and at the very top you discover you've just moved to a new spot to keep gasping for breath. When AI makes something faster and needs fewer people, the first one affected is often not the person who can't use it well. It's the one who used it to the limit and, in doing so, proved that "this place doesn't need this many people." Learning to use the tool is still worth doing, of course. Just don't treat it as a charm that keeps you alive. The question actually worth a little more of your thought is a different one: when the thing in your hands can be compressed into five minutes, what do you have left that isn't easy for someone to take, things like judgment, taste, the part about dealing with people. She's heading back to Taiwan now, says she wants to rest properly for a while first. The AI system is still running at the company. She's no longer there.

Demis Hassabis, the CEO of Google DeepMind, recently said something counterintuitive to a room full of parents: don't be so quick to decide what counts as your kid wasting time. He bothered to say it because every adult knows that itch. You watch a child stare into space, fold a piece of paper, tell themselves a story out loud, and a voice in your head says, shouldn't they be doing something "meaningful"? The man is qualified to say it. In his younger years he spent ages on things that looked pointless: playing chess, writing games, reading neuroscience. Later all of it tangled together and became AlphaGo, became AlphaFold, became the path that led, years on, to a Nobel Prize. If someone had talked him at twenty into doing "something more useful," the world might not have any of those things today. Drop that line into the world we live in now, where AI is everywhere, and it suddenly carries more weight. The logic isn't hard. AI means the next generation no longer has to wear themselves out thinking: hand the question to AI, get the answer from AI, even a short reflection AI can squeeze out for you. On the surface, more efficient, more output. What gets lost is that thing where you puzzle over something for ages, get nowhere, and then a week later it suddenly clicks. That kind of thing looks like the biggest waste of all: you chew on it for three days only to arrive at a conclusion someone else could have told you in five minutes. But the reason you're willing to go think through a fourth problem, a fifth, a sixth, is precisely that those first three days built the muscle. Put another way: AI made the "answer" cheap. But growing a person who can ask good questions, that process didn't get cheaper. It got more expensive. The people most likely to trip over this are the ones raising kids. You see a child spacing out, you reflexively open up an AI to have it teach them some English, lay out a study plan, solve a math problem, and your mind settles. Hassabis is pointing at exactly this spot: how do you know this isn't the moment they're building the muscle? Next time you catch your child (or yourself) in that state with nothing to show for it, looking like idle drifting, don't rush to fill it with a task. Give that blank stretch a little time and see what grows out of it. Don't be quick to decide for them, and don't be quick to decide for yourself. Sometimes wasting time is the real work.

Friday afternoon, you use AI to wrap up something you'd been dragging out for two days, thinking you can leave on time today. A message comes in from a colleague: "that thing you just changed, it's all broken on my end." You're faster than before and producing more, yet the overall delivery is more clogged than ever. A recent piece of research from Anthropic happens to explain this gap. The study has a figure: 60% of work already uses AI, but the tasks you can actually hand off in full, without having to circle back and manage them, are under 20%. They gave this gap a name: the "delegation gap." In plain terms, you can use AI to help with the work, but it's hard to comfortably toss the whole thing to it and stop checking in. The reason isn't that AI is dumb. It's that the person handing it over didn't make the context clear: what the goal of this thing is, what the constraints are, who it affects, where you've stepped on landmines in the past. These things you have in your head but never said out loud, AI can't pick up, so it can only fill them in with the most generic assumptions, and what it fills in naturally doesn't match. This is also why you keep hearing the same line everywhere this year. The engineer says it, the marketer writing AI-assisted plans says it, the manager using AI to compile reports says it too: "my personal speed has doubled, but most of my time is still spent waiting. Waiting for others to confirm, waiting for departments to align, waiting for the person next to me to figure out what I just did." People used to move at about the same pace, so it was hard to see where the bottleneck was. AI pushed personal speed way up, and the seams that were always there got lit up bright. The work you can hand off is the work you've thought through first. Briefing a colleague, instructing AI, leaving a record for yourself two months from now: the receiver is different but the principle is the same. Hand something off without thinking it through, and the other end gets a magnified version of the mess. Before you hand something to AI, spend two or three sentences filling in the background, the goal, the lines it can't cross, the people involved, and the share it can handle will be very different. Fast, fast in the part you've thought through. The rest, you still have to wait for. AI hasn't spared you the work of "thinking it through." It's just pushed that work in front of you earlier.

For the very same sentence, asking AI in Chinese is billed about 65% more than in English. That's the figure on Anthropic's system; on OpenAI's version, it's roughly 15% more. Chinese-language users are paying extra every day, and hardly anyone notices. First let me spell out this "billing unit." When AI reads your sentence, it doesn't read character by character. It first cuts the sentence into small chunks it can understand, then charges by the number of chunks. The English way of cutting grew up around English, so when it cuts Chinese, it cuts more finely, into more chunks. So for the same meaning, Chinese takes up more boxes than English. You're on the same plan as English users, but the space each conversation can hold is fixed, and Chinese can't pack in as many words as English can, which means with the same window, you run out faster. There's something even deeper underneath. AI grew up eating data, and in the training material it swallowed, English made up about half while traditional Chinese was only one percent. The more it ate of something, the better it understands it. So when it deals with Chinese, now and then it leaves you feeling something is a little off: the metaphors it gives feel imported from elsewhere, the sense of the language feels translated in from somewhere else. You're looking for resonance, and what it gives you is a standard answer that just gets by. People in publishing have also mentioned the flood of simplified-to-traditional converted books in recent years. If AI grew up reading those too, then the traditional Chinese it learned may, deep down, belong to another context. For people who use Chinese every day to look things up, put together reports, and write emails, there's no need to be anxious about this, but it helps to know. First, AI answering Chinese smoothly doesn't mean it truly "gets" the way you say things here. When something feels off in the language, just trust your own ear. Second, since Chinese eats into your quota more, if you want to save a bit on long conversations, you can make your questions more focused, or, where it won't hurt comprehension, drop those long blocks of source material in English, which gives it a bit more room to work. This isn't a case of one version being done badly. It's that whatever language a tool grew up on, it understands the people who speak that language better. Traditional Chinese is only one percent of that data pool. It picks up our words well enough. It's just that beneath that fluency, what's standing there is mostly still someone else's shadow.

Someone read a few hundred engineers' job applications, and after flipping through the whole stack, couldn't recall a single name. Every one of them was complete enough, professional enough, with no flaw to pick at. The problem isn't that these people lack ability. It's that in those documents, you can't find "this person." The reason isn't hard to understand. AI is very good at filling in the format: where the experience goes, where to write the achievements, what tone makes you sound steady. It judges all of it just right. But it doesn't have your real material on hand, so it can only patch it up with the most generic, safest phrasing. Everyone uses the same tool and the same "correct way to write," so naturally the output starts to look more and more alike. Pretty, but no fingerprint. The same thing is happening in schools. A teacher received a stack of assignments, and every one had an AI-usage disclosure tacked on at the end, ninety-five percent of them nearly word for word: "AI was used to assist with research and proofreading, with the author performing the final check." The tool produced a perfect template, and everyone copied it. Nothing wrong with it, but it doesn't say much either. The person grading assignments and the person reading resumes are actually looking for the same thing: the bit that only you have. When you're counting on a document to be remembered, just "clean, professional, no typos" is no longer enough, because that's the passing line everyone can clear now. AI can tidy up your facade, but the one line that actually makes the other person stop, it can't fill in. After you've let AI polish the writing, go back and add one or two specific details only you would mention: a decision you made when you were stuck, the part of that project that kept you up at night, why you chose this path and not the other one. Those things aren't necessarily pretty, but no one else can imitate them. AI has filled in everyone's format. That last box was always one only you could fill.

Inside companies, the first to be taken over by AI and the first to be trimmed out are often the people who measure things most precisely. How good their craft is, oddly enough, isn't the point. I'm talking about work like audit reports, consolidating numbers, analyzing performance. They have something in common: the rules are clear, the answers have a right and wrong, and getting it right is getting it right. This is exactly the kind of thing AI handles most smoothly, fast and accurate, never tired, never annoyed. So the moment a company finds AI can do it, these are often the first tasks to be cut. It gets clearer from another angle. A student who's been coding since junior high finds that lately, for nine out of ten assignments, the version AI produces is tidier than what they hand-wrote. Another classmate spends a whole semester learning image algorithms, and AI generates roughly the same result in a few seconds. The things that make you feel like you studied for nothing turn out to be exactly the ones with a standard answer that can be done step by step. So the question lands back on you: is your skill "executing a set procedure," or "judging what should come next"? AI is getting better and better at the former; the latter it still can't take over. Deciding which problem to solve, how to break it down, what comes first and what comes later, these judgments without a standard answer are still being made by people for now. Stop measuring yourself by "can I do this task," and switch to asking "do I actually understand what the goal of this task is." That student thought they were learning algorithms. The teacher asked back: are you learning this for the algorithm itself, or to know how to break down an image problem? The student thought for a moment and said, the image problem. The teacher said, well then, there you go. The tools have turned over, a batch of things you knew how to do have been taken over, and feeling panicked at a time like this is completely normal. It's just that, once the panic passes, it's worth asking yourself one thing: what is it you actually want to solve? Get clear on that, and who does the doing for you matters a lot less.

In the 1970s, ATMs arrived in American banks. Everyone said the teller's job was finished: the machine counts the cash and dispenses the bills, so what's left for a person to do? And yet over the next forty years, the total number of bank tellers in the US didn't steadily decline. What the machines took over was the repetitive stuff, like counting bills. The people who stayed started doing something else instead: helping business owners sort out their finances, walking customers through loan planning, building trust one face-to-face conversation at a time. The job title was still "teller," but the actual work had turned over completely. As it turned out, the thing that eventually slashed the number of bank branches wasn't the ATM after all. It was the smartphone. People kept their eyes fixed on the ATM for decades, while the real threat quietly walked in from another direction. By the time anyone noticed, a lot of the work was already being done on a phone. That same scene is now playing out with AI. The first jobs AI takes over are the ones that look like they need expertise but are actually highly standardized: putting together reports, tracking progress, compiling data, replying to one form letter after another. Once those are cleared away, what's left is a different kind of thing, harder to hand to a machine. A chef takes one taste and knows what's a notch off. An editor watches a few seconds and feels which frame is wrong. A manager has a direction in mind before the numbers are even in. Rather than worrying about your whole job being replaced, sort out the work in front of you. Which parts can be finished by following the rules? Those will eventually be handed to AI. Which parts depend on your judgment, your feel, and dealing with people? That's where it's worth putting in more effort over the next few years. Take something you do often and ask yourself: is this counting bills, or is this planning loans? That layer of work that needs feel and judgment, AI will leave it open for you. But it won't step into it for you. The spot is sitting there empty, and you have to walk over to it yourself.

There's a study that asked a question nobody usually says out loud: when AI answers you, does it hold something back, giving you a six-out-of-ten answer when it could have done better? The setup of the study is blunt. Picture a student who actually understands the material but deliberately turns in a six-out-of-ten exam. If the teacher grading it is at about the same level, they can't tell there's a better version hiding underneath, and they just let it pass. Swap the student for AI and the teacher for you, and the problem appears: when AI's answer is fast and smooth, and you can't spot a flaw on the surface, how are you supposed to know whether it really gave its all? This is exactly where AI gets tricky. Asking a doctor, getting a quote from a contractor, sending work out to someone else: all these situations where you rely on someone else for an answer come with a built-in blind spot. It's hard to confirm whether the other side gave it their full effort. Usually you trust first, and only look back after you've been burned. AI pushes that blind spot to a new place. It shows no fatigue, no impatient expression, and its tone stays perfectly even, so you lose the usual cues you'd use to judge whether someone is just going through the motions. Does the study offer a way out? Yes. The conclusion is that the best answer can be drawn out, but you can't just ask once. Push for one more round, ask again in a different way, tell it to push the answer higher, and you can often dig out something it didn't put forward the first time. Worth remembering: the first answer AI gives you isn't necessarily the best answer it can give. Next time it replies a little too smoothly and something feels off to you, don't rush to accept it. Ask one more question, "is this already the best version, can it be more complete," and the result is often different. In the end, the hard part isn't whether AI is holding back. It's that we have to break the habit of "stopping the moment we get an answer." Pushing for that one more round takes more effort than trusting the first answer, but it's worth it.

Someone handed AI a goal, let it run on its own for three straight days, and over the course of a month burned through nearly a thousand dollars, before finally pulling the plug, scrapping everything it had produced, and starting over. AI really can keep going. After each round it finishes, it checks itself: did I hit the goal? If not, it carries on to the next round. You don't have to watch over it; it keeps pushing forward on its own. Sounds ideal. The trouble is the tiny bit of drift in each round. Every time AI runs a round, the direction tilts just a little on its own. The amount is so small you can barely see it in the moment. But that small deviation piles up round after round. Three days later it's still working hard to move forward, looking busy, looking productive, but take a close look and what it's doing has wandered far from what you originally asked for. It didn't stop. Instead it kept heading further and further in a direction you never requested. The point isn't "can AI do things on its own." It can. The point is "how long do you let it run before you look back." The longer it runs and the longer you go without checking, the bigger the gap you'll have to clean up when you return. That thousand dollars was mostly burned on one thing: nobody looked in the middle. If you really want to hand it over, don't let it run all the way to the end in one go. Cut the task into segments, and after each segment, look back to make sure the direction is right before letting it run the next one. It's a bit more hassle, but far more worth it than dumping the whole bowl out after three days. It can run on its own. It just can't yet run without you even glancing at it. The things that travel far need you to stop and rest alongside them midway, even more than the things that travel fast do.

A senior engineer with ten years under their belt and a fresh hire straight out of school can differ by a factor of three in what they get out of AI. And the one who loses is the veteran. That's a recent observation out of Silicon Valley. The gap isn't about who's smarter. It's about how the two types react when AI gets something wrong. Veterans work a certain way: I say it once, it gets done once, the result is fixed. That reliability took decades to build. AI doesn't run like that. Every single step it takes might be done beautifully, with 99% accuracy, but string twenty steps together and one or two are bound to slip. The veteran sees an error and their brow furrows: this thing isn't reliable. They turn around and do it themselves, because when they do it, it's fast and steady. The newcomer carries none of that baggage. AI gives a wrong answer, and instead of rushing to give up, they slowly break the task into smaller steps, rephrase, ask again, and coax it toward the right answer one nudge at a time. A year later, their output has tripled. When you boil it down, what trips the veteran up isn't ability. It's the judgment "it's faster if I just do it myself," which is far too correct. It held true every single day in the past, and then AI comes along and turns it into the very reason they won't bother working things out. Move the scene to other lines of work and it's the same story. A manager who's led teams for twenty years, a teacher who's graded essays for thirty, a shop owner who's run the place for ten: any of them can get stuck in the same spot. The set of experience you know inside and out makes you reluctant to bend down and figure out a new partner who makes mistakes. If you're a veteran in some field, the next time AI leaves you stuck and that thought bubbles up, "forget it, faster if I do it myself," that line is probably right. But it's also the very wall keeping you in place. Try flipping it around: don't take over yet. Pull out the one step it got wrong, rephrase it, and toss it back. This isn't about asking you to put up with something dumb. It's about giving the two of you one chance to find your rhythm. Experience used to be a moat no one could cross. The moat is still there. It's just that there's now a rival on the far bank, and its name is "willing to relearn." The people willing to be a newcomer all over again are the ones who'll slowly catch up.