AI is taking your ability to think. You should probably do something about it.
2026 May
Originally a thread on X — the longer version is below.
This one is a bit of a brain dump. It started as a late-night voice memo to a friend who heard me out and said: "you should actually write this down". So here it is. Cleaned up, but not sanded down.
I'm writing it because last week PayPal announced layoffs, and one of the people inside that number is a dear friend of mine. Another close friend lost their seat at Visa earlier this year and hasn't found work since. I carry both of them in my mind and in my prayers, and they were not the first this year. A few years ago we were a comfortable bunch: early-career, killing it in FAANG and FAANG-adjacent seats, the kind of jobs that commanded great salaries in your early 20s and felt like a moat you were lucky to be on the right side of. The moat, it turns out, was rented.
The terrain is changing. And the uncomfortable part is I think it's changing for the better. Not for everyone — but for thinkers. This piece is my attempt to explain why, and what I believe is the only real way to survive (and maybe win) in this new AI world.
1. Why we need to re-learn thinking
1.1 The failure of the default loop
Here is the loop a lot of us have quietly settled into.
You paste in a spec, or an error message, or a half-formed question. The model hands back something that works like magic. The symptom vanishes. You ship. And somewhere inside that loop, the messy, productive struggle between the problem and the solution — the part that used to be the actual job — just stopped happening.
Here's the first-principles problem, and the annoying thing is that it's just arithmetic. Skills compound. A = P(1 + r)^t. Your capability tomorrow is your capability today, multiplied by some small daily growth rate, raised to the power of every day you put in the reps. That equation is why a senior engineer is worth more than a junior one: it's a decade of r stacked up. But the equation has a terrifying corollary. If r goes negative — if you shed a sliver of capability each week instead of adding one — the same exponential runs in reverse. Cognitive debt compounds exactly the way cognitive equity does. Just in the wrong direction. AI keeps getting smarter while you (no offense) might be getting dumber.
The reason I decided to write this was an X post I saw the other day, by Addy Osmani, a genius in the field and the author of many great software engineering books:
Don't outsource the learning. Right now, it's too easy to let AI write the code while you skip the learning. The bug gets fixed but your mental model doesn't move. It might get worse over time. We are silently trading future capability for present speed.
— Addy Osmani
The full piece lives on his blog: Don't Outsource the Learning.
The piece is quite alarming. The setup is that exact loop from above: paste the error, accept the fix, ship, repeat. The symptom dies. Your understanding flatlines. Osmani calls the bigger version of this cognitive surrender — the moment the model's verdict evicts your own. He is very much not an anti-AI guy yelling at a cloud. He says he's shipped more with these tools in the last year than in all the years before it. The tools are not the villain. The problem is that they are tuned for closing the task and that's it.
Other research seems to converge at the same uncomfortable spot. MIT strapped EEG caps on people writing essays and watched brain connectivity drop with every extra layer of AI help, until 83% of the AI group could not quote a single line of the thing they had supposedly just written. They named the effect cognitive debt, and the name is the whole warning: you borrow speed today, and you repay it, with interest, in thinking.
There's another one I want to point you to. It is by Vlad Feinberg, who works on Gemini at Google DeepMind, and it is called How to Land a Frontier Lab Job. On the surface it has nothing to do with anything I have said so far. Read it anyway, because it is a love letter to grind.
Feinberg's claim is that the people who land the hardest technical jobs on the planet tend to share three traits. Intent: they picked a problem space that actually matters instead of drifting. Mathematical maturity: the generalized problem-solving muscle that lets you wrestle an ambiguous question to the ground. And grit: they survived the soul-crushing, proof-based classes that quietly break most people. Notice that not one of those three is "good at prompting." Not one of them is a thing you can download. He describes his own college weekends with a kind of fond horror — quadruple-shot iced coffees on a Saturday morning, problem sets stacked on problem sets, the campus party audible through the library window and very deliberately not for him.
He says to use AI for what you already know how to do, only — but aggressively so. That is the entire discipline in one breath. AI is a forklift, not a gym membership. It is phenomenal for moving weight you have already learned how to lift, and it builds you exactly zero muscle the moment you let it lift the things you have not.
Put the two articles side by side and you basically get the thesis of this whole piece. Osmani tells you what quietly rots when you outsource the thinking. Feinberg shows you what compounds when you refuse to.
Therefore, we need a new solution: we need to re-learn how to think, on purpose, against the grain of tools designed to do it for us.
1.2 The requirements of the new solution
For about a decade, the actual moat was knowing things. The engineer who had the framework memorized. The analyst who knew the comps cold. The person in the meeting who simply had the fact. Information was scarce and expensive, so the people who hoarded it won.
Today, the situation has completely changed. Every fact is free, instant, and conversational. Knowledge is no longer scarce. In fact, it's the most abundant commodity on Earth, priced at roughly zero.
We are transitioning from a regime where knowing things was the moat, to a regime where knowing how to think is the only moat left — and the bridge is drawing up.
So what would a real solution have to do? Three things. One, it has to keep you using AI. Going Amish is not an option and pretending otherwise is a fantasy we frankly cannot afford. Two, it has to deliberately rebuild the friction the tools have sanded away, because friction was where the learning lived. Three, it can't be a product you buy; it has to be a posture you adopt, something you run inside the exact same tools you already have open.
Here's what it looks like.
2. Four features of the independent thinker
2.1 Hypothesis before prompt
Before you ask the model anything, write two or three sentences on what you think the problem is.
From first principles, this is just the scientific method, and the scientific method has one non-negotiable rule: you cannot test a hypothesis you never formed. If the model's answer is the first hypothesis in the room, it is also the last one because there's nothing left to weigh it against. The analogy here is an open-book exam. Same book for everyone. The student who studied, and the student who didn't, get wildly different scores, because the book only helps the person that already has somewhere to put the information. Forming your hypothesis first is studying for the open-book exam. Not studying is… well, not studying.
2.2 Explanation before code
When you're in unfamiliar territory, your first prompt should never be "write it." It should be "explain how this works, what the alternatives are, and what the tradeoffs are."
There's a real principle underneath this. Neurons that fire together wire together — Hebb's rule, the closest thing learning has to a law of physics. The wiring happens during the explanation. It does not happen during the copy-paste. Asking for the code first is ordering the answer key before you've seen the test; asking for the explanation first is hiring a tutor.
2.3 The re-derive check
Every so often, take something the model wrote for you and rebuild it from scratch, alone, no AI in the room. This is the calibration check, and here's the "wait, actually" part.
Anthropic ran a randomized trial on exactly this. Engineers learned a brand-new Python library, half with AI and half without. Both groups finished the tasks at the same speed. On the surface, identical. Then came the comprehension quiz. The manual group scored 67%. The AI group scored 50%. But the real finding was inside the AI group: the engineers who used AI to ask conceptual questions scored above 65%, while the ones who just copy-pasted the generated code scored under 40%. Same tool. The re-derive check is just stepping on the scale. It's the only way to know what you actually know.
2.4 Falling in love with knowing
This is the one that actually matters, so I'll say it plainly: fall in love with learning, and fall in love with knowing.
From first principles, motivation is the only sustainable engine here. Discipline runs out — it's a battery, and it drains. Curiosity is a generator; it makes more of itself. If learning stays a tax, you will spend the rest of your career finding clever, AI-shaped ways to dodge it. But if learning becomes an appetite, you'll happily overpay for it, and the overpayment — the knowing you picked up that nobody asked you for — is the moat. There has genuinely never been a better moment in human history to be a person who simply loves to know things. Infinite knowledge, one prompt away, and almost nobody bothering to keep any of it. The shelves are full and the store is empty.
3. The unconstrained TAM
Here is where I think the consensus badly mis-sizes this.
Most people file "think for yourself" under self-improvement. A nice-to-have. A productivity tip somewhere between journaling and buying a standing desk. That is a category error, and it's an expensive one.
Think about what AI actually did to the labor market. It raised the floor and the ceiling. The floor — producing a working, plausible answer — is now standing room only, crowded with every single person who can type a prompt. That's a commodity, and it is being priced like one: junior developer employment is down roughly 20% since 2022, per Stack Overflow's own data. But the ceiling — the person who can tell when a confident answer is confidently wrong, who can steer the tool instead of being steered, who actually owns the architecture when it breaks — that space is wide open. Empty, even. So the question isn't whether AI takes jobs.
…it's whose.
Worried it's coming for yours? Fair. But ask the better questions. Who debugs the AI's code when it falls over at 2am? Who catches the hallucination before it ships to a customer? Who does the company keep when it's cutting 20% of the building? It is not the fastest prompter. It is the clearest thinker — probably, every single time, hopefully.
Therefore, the TAM for thinking is not some slice of the job market. Every seat AI clears off the crowded floor adds value to the empty ceiling. The market for people who can genuinely reason is the one market that runs inverse to the layoffs.
4. Other solutions
Let me speedrun the alternatives people could bring up.
"Just get really, really good at prompting." Sure, prompting matters today. But prompt-craft is a skill the model makers are actively, deliberately working to make unnecessary. Every release needs less of it than the last. Building a career on prompting is building on a sandbar that the tide is engineered to take — unless you're getting paid to teach it.
"AI will just get good enough that none of this matters." Maybe — and if that day comes, notice the asymmetry: the person who kept thinking loses nothing, and the person who didn't loses everything. That's a free option. Take the free option.
"Pick a recession-proof, AI-proof field." There isn't one. PayPal is profitable and still moved to cut around a fifth of its people, citing AI. Profitability stopped protecting headcount.
"Wait it out." No. The compounding clock from Section 1 does not pause while you wait. Every week you stall, r stays negative and the debt keeps stacking. There is no waiting it out. There is only the compounding, and it is already running, whether you've opened your notebook or not.
5. Conclusion
So here is the whole thesis in one breath. The terrain stopped paying people for knowing things, because knowing things became free, and it started paying enormously for an asset it cannot download, fine-tune, or lay off: a creative mind that works. My friends from PayPal and Visa are going to be just fine, because they are thinkers and life learners.
If you do exactly one thing after reading this, make it tomorrow morning: before you open the chat, form your own answer first. Carry a notebook, and actually write in it. Re-derive something by hand. Fall in love with knowing. Go long your own brain.
It is the trade of the decade, and the window is open right now.
Further reading: Addy Osmani — Don't Outsource the Learning. Vlad Feinberg — How to Land a Frontier Lab Job.
Original thread: @pantsyd on X