Learning to Fly

A big company measured its best people after a couple of months of AI, and they had gotten worse.

The story comes from a short video by @stay_human25, the account behind the book Stay Human. Leaders at a company well along in AI adoption ran an experiment on their own experts. After a couple of months of working with AI, their top people were about 20 percent less accurate at catching errors. “Everyone asks how juniors are gonna build expertise,” the narrator says, “but fewer people are asking how fast senior people lose it. And it turns out like, quite fast.”

That is one company, told secondhand, and we will treat it that way. But it landed on a question this blog has been circling for fourteen months without quite naming it. Because the same fourteen months have taught us more, faster, than any stretch of working life either of us can remember. Seventy-three posts. Six distinct phases. Lessons that would normally take a decade of technology cycles to collect, collected between one Labor Day and the next Halloween.

So which is it? Is AI making us learn faster than anyone in history, or making our experts forget? This post argues it is both, and that the difference between the two has a name.

Tom Petty put the whole thing in a chorus in 1991: learning to fly, but I ain't got wings. And the line after it is the one that matters for the rest of this post: coming down is the hardest thing.

Experience, stripped down, is a loop: you try something, it works or fails, you notice which, and you adjust. A career is the number of times you get around that loop, multiplied by how much you notice each time. The reason a thirty-year engineer is worth more than a three-year one is not that they read more. It is that they went around the loop ten times as often.

AI multiplies the number of trips. Reading this blog back from the first post to the last, the trips are visible as phases, and each phase starts where the previous one ran into a wall:

• Learning the tool (fall 2025). Context is the job; memory has to live outside the model. 50 First Dates, The Shawshank Method.
• Learning about ourselves (late 2025). Speed does not return time; it resets expectations. The Event Horizon, The Co-Pilot Paradox.
• Learning the economics (winter). Subsidized prices, AI bolted into every SaaS tool, and being replaced by a platform vendor ourselves. Software Is Dead.
• Learning the human cost (spring 2026). Addiction, re-entry, organizations nobody can explain. Flowers for Algernon, Be Kind, Rewind, I've Created a Monster.
• Building discipline (summer). Director, coder and QA; specs a machine can check; watching beliefs drift over a long conversation. Under My Thumb, The Arc.
• Learning what we can no longer see (fall 2026). Official-looking codes nobody checks, dementia signals polished away, dashboards that go red because they are working. Sharp Dressed Man, The Great Pretender, I Can See Clearly Now.

The direction matters more than any single stop. It runs from how do I use this, to what is it doing to me, to what is it doing to my organization and my market, and finally to what is it doing to my ability to tell what is true. A normal technology wave takes a decade or more to walk a person down that whole path. This one did it in a little over a year.

And most of it is not widely shared. Plenty of people have read about AI. Far fewer have come back from a week of vacation to fifty times the output they can absorb, or watched a model's stated beliefs move from 18 to 35 percent over one long conversation, or run a director-coder-QA loop across two vendors' models in production. These are lessons you only get by living at the speed.

One flight log is an anecdote. So we went looking for a second, and found one in Steve, who posts on TikTok as @sd141414 under a bio that reads “30 years shipping production systems. Judgment is the edge.” He published 190 videos between December 2025 and this week. We pulled the full list, transcribed a systematic sample of every sixth video, and counted themes across all 190 captions.

The numbers tell the story before the transcripts do. In December, three captions in four were about vibe coding, and the videos were prompting tips: give it the files, spell out the edge cases, use plan mode, tell the model what it must not do. He was teaching beginners to hold the AI's hand. By fall the vibe-coding tag was gone, agents and loops had taken its place, and the voice had turned personal. In October he wrote that AI built the math in his imaging tool and “is better at math than I am,” and that line-by-line code review “does not survive AI-speed delivery.”

That reads like a move from caution to full embrace, and it is. But look closer and the hand-holding never stopped. It moved. In December he held the AI's hand through prompts. In October the AI holds the keyboard and he holds the tests, the specs, the decision records and the accountability. Responsibility language doubled in the last period. The bio line about judgment never changed.

What stopped us was how closely his route matched ours. Different platform, different audience, different projects, and the same lessons in the same order, usually within a few weeks of each other. Step through the legs:

The one real difference is tone. Steve almost never writes about the emotional cost. This blog wrote Flowers for Algernon and Turn! Turn! Turn!. He writes as the instructor; we tend to write as the witness. Two different cockpits, same weather.

The @stay_human25 video points at a piece of brain anatomy that explains both arcs. It is called the salience network: a set of regions, mainly the anterior insula and the anterior cingulate cortex, that tend to activate together. Its job is to flag what matters right now and to switch you from autopilot into deliberate attention when something does not fit.

Two details make it relevant here. First, it reacts most strongly to surprise: the gap between what you expected and what you got. In Wolfram Schultz's classic studies, dopamine neurons stopped firing at a reward once the reward was fully predicted, and fired at the cue instead. A fully predictable thing stops being interesting. Second, the anterior insula is where the brain reads the state of the body, so its flags often arrive as a feeling before they arrive as a thought. That is the expert's “something's off” that comes before they can say what.

Now look at the flight logs again. In leg one, the surprises were bad output: the hallucination, the model that redesigned your app. As the tools got better, those errors stopped happening, and so they stopped being surprising, and so they stopped being noticed. Steve recorded the moment precisely. He cut his standard prompt set from 212 instructions to 61, because “the ones I cut told the agent to do things it now already does on its own.” Attention moved up a level, to the next place surprises were still happening: the decisions, then the loop, then our own understanding, then the people around us, then the things we could no longer see at all.

Read that way, both arcs are the same thing: a record of the salience network climbing the stack, one layer at a time, as each layer below it went quiet.

That is the good version. It is how expertise is supposed to work: you stop noticing what you have mastered so you have attention left for what you have not. The bad version needs a dog.

Ivan Pavlov won his Nobel Prize in 1904 for work on digestion, and found conditioning almost by accident: his dogs began salivating before the food arrived, at the sound of the assistant's footsteps. Pair a neutral signal with something that matters, often enough, and the signal alone produces the response. Popular retellings use a bell; his lab more often used metronomes and buzzers. The principle is the same.

Pavlov also documented the reverse, and it is the part that matters for experts. Ring the bell again and again with no food, and the drooling fades. He called it extinction. The dog has not forgotten the bell. It has learned that the bell no longer means anything.

An expert's gut feeling is conditioning. Years of pairing certain textures, an odd number, a clumsy phrase, a step that does not follow, with “this turned out to be wrong” make those textures ring a bell. AI threatens that in two ways. The obvious one: fluent output removes the textures, so the bell never rings even when the error is there. That is the mechanism in The Great Pretender, where polish erases the signal families relied on. The quieter one is extinction. If the “something's off” feeling keeps firing at AI output that turns out to be fine, the feeling learns to stop firing. Then the real error arrives and nothing rings.

We want to be careful here. Nobody, as far as we know, has measured conditioned extinction in people reviewing AI work. The nearest established research is on automation complacency: people monitor less carefully once a system seems reliable. The extinction account is our hypothesis for why a 20 percent drop could arrive in a couple of months instead of a couple of years. It fits the timescale. It has not been tested.

Steve logged a version of it from the other side. In August he wrote that he “kept waiting for a reaction to how fast I have been shipping” and none came: “they don't have the same baseline that I do.” That is conditioning in both directions at once. He was conditioned to expect applause for speed, and the expectation extinguished. His stakeholders were never conditioned to the old pace at all, so the new one was not surprising enough to notice. Salience depends on the baseline. People without the old baseline cannot feel the change.

Pavlov offers one hopeful footnote. Extinguished responses come back after a rest, which he called spontaneous recovery. Extinction suppresses an alarm; it does not delete it. That suggests the loss is reversible, if you give the alarm a reason to ring again.

Here is the finding we did not expect. Working on different problems, both logs arrived at the same countermeasure, and it is exactly what the brain science says should work.

Steve's most recent video, built on Addy Osmani's essay about critical thinking when agents do the work, gives it as a two-step habit: “Write down what you expect before you prompt. Have the agent explain its own diff afterward. The gap between those two is where your judgment still lives.” In April he gave the gentler version: ask the AI “what should I consider?” instead of “what should I do?”

This blog's versions look different on the surface. In I Can See Clearly Now the rule was to write down, before go-live, which number will spike and which should fall. Every recent post carries an honest counter-case section, written before we decide we are right. The director-coder-QA loop writes the acceptance criteria before the code exists.

Underneath, they are one move: commit to a prediction before you see the output. Fluent AI output does not create surprise on its own, because it always looks right. A written prediction gives the output something to disagree with. The gap between what you expected and what you got is the signal the salience network runs on, and you have just manufactured it on purpose. It is the bell, with food behind it again.

The habit is cheap and almost nobody does it, because it adds a step to a workflow whose whole appeal is that steps disappeared. That is the trade. The friction used to be free, built into doing the work by hand. Now you have to pay for it deliberately, or lose what it bought.

In 2009 Daniel Kahneman, who spent his career documenting the failures of intuition, and Gary Klein, who spent his documenting its successes, wrote a joint paper to settle when expert gut feeling can be trusted. Their answer had two conditions. The environment has to be regular enough to learn, and the expert needs plenty of practice with quick, clear feedback. Firefighters and chess players qualify. Stock pickers and long-range political forecasters mostly do not.

AI breaks the first condition. The environment is no longer regular; it changes every few months. Steve put a number on one piece of it: “Prompt engineering had about an 18-month shelf life.” And a sharper version: “Stack knowledge fades in six months. Default behavior doesn't.” This blog has its own expired lessons. Half of The Shawshank Method is about repeating constraints so a model cannot ignore them, which today's models need far less.

So the acceleration has a catch. AI speeds up how fast you build conditioned intuition, and it speeds up how fast that intuition goes stale, by about the same amount. You learn the sky faster, and the sky changes faster. If experience were the whole story, that would be a treadmill. Run faster, stay in place.

Experience is not the whole story, and the part that is not experience has a name. Psychologists at the Max Planck Institute in Berlin, led by Paul Baltes, spent years trying to measure wisdom as something distinct from intelligence or knowledge. Their definition starts with two basics, a deep store of facts and a deep store of know-how, and then adds three things that sit above them:

• Context. Understanding that the right answer depends on the situation and the stage of life. Steve: “There is no universally ‘correct’ architecture. There is only architecture that matches context.”
• Values. Recognizing that different people reasonably want different things. Every “honest counter-case” section on this blog starts from that.
• Uncertainty. Knowing the limits of what you know and acting well anyway. Steve: “Judgment is the ability to stay open longer than feels natural.”

Notice what those three have in common. None of them is a conditioned response. They are the ability to look at your own conditioned responses and ask whether they still apply. Intuition is the bell. Wisdom is knowing when the bell has stopped meaning food.

That is why the 212-to-61 cut is the most instructive thing in Steve's whole feed. The 151 prompts he deleted were conditioned lessons that had expired. The 61 that survived, in his words, “carry a decision, or a rule from a real failure.” The tactics expired. The judgment about which tactics to keep did not.

AI compresses experience. It does not compress wisdom. You get more trips around the loop than anyone has ever had, but only reflection turns the trips into something that lasts, and reflection takes as long as it always did.

That also answers why these lessons are so unevenly spread. The line usually credited to William Gibson says the future is already here, it is just not evenly distributed. The unevenness is not only about who adopted early. Even among early adopters, the people who come out with the full lesson are the ones who kept some friction: who still predicted before they looked, still wrote things down, still did some of the repetitions themselves. The rest got the speed and the 20 percent drop.

1. Two pilots is not a sample. This is the strong objection. Both logs come from senior people with long systems careers, and both of us make a living teaching others how to build with AI. The arc we describe might be what seniority does with any new tool, not what AI does to everyone. A junior developer or an accountant might walk a completely different route. We think the match is striking because the two arcs were built from different work. We do not think it is proof.

2. The scariest number is the weakest. The 20 percent drop comes from one unnamed company, told secondhand in a short video. The extinction mechanism is our hypothesis. If someone runs a proper study and finds expert error-detection holds steady with AI, the second half of this post falls over. We would rather know.

3. Falling views might mean the later lessons are worth less, not that they are rarer. Steve's median plays dropped by about two-thirds as his videos moved from tips to reflection. We read that as evidence that audiences reward the early, conditioned lessons more than the hard-won ones. It could also mean the later videos are simply less useful, or that recent videos have not had time to collect views. The data cannot tell those apart.

4. “Wisdom” can be a story seniors tell to stay relevant. If the answer to every disruption is that experience still matters, that is convenient for the experienced. We made the same worry in Turn! Turn! Turn! and it applies harder here. The real test is whether someone starting a career now, who gets the speed without the decades, can build the reflective layer on purpose. Wax On, Wax Off argued that has to be designed in. Nobody has shown yet that it works.

Pilots keep logbooks for a reason that has nothing to do with paperwork. A log forces you to write down what happened while it is still fresh, and it lets you see, years later, which of your habits came from a real incident and which came from habit. That is the practice this whole post points at.

• Predict before you prompt. One sentence: what you expect the output to do. Compare afterward. The disagreements are where your judgment lives.
• Make the agent explain its work, then check the explanation against your prediction, not against the code. Explanations that surprise you are the bell ringing.
• Keep one manual repetition a week. Debug one thing by hand, review one diff line by line, write one paragraph without help. Not as nostalgia; as calibration for the alarm.
• Put an expiry date on your lessons. Every quarter, go through your prompt set, your checklists, your rules of thumb. Delete what the tools now handle. Keep what carries a decision or a real failure.
• Write the log. Decision records, a team wiki, a blog. Anything that captures why, not just what. It is the only way the experience leaves your head and reaches the people who are not flying at your speed yet.

That last one is why this blog exists, it turns out. We did not set out to keep a flight log. We set out to write about AI. Reading all seventy-three posts in order, a log is what we have: the record of one fast apprenticeship, with the expired lessons left in so you can see them expire.

The flying got easy. The coming down, the noticing, the writing it down, the deciding what to keep, is still the hardest thing, and still the part that makes you a pilot.

This post sits with Wax On, Wax Off on how judgment gets built, The Great Pretender on signals that polish erases, I Can See Clearly Now on writing the prediction down before go-live, and Turn! Turn! Turn! on what the people living through this are feeling.