The cap
A developer I know hit his monthly token ceiling on the 11th. Not the soft warning email. The hard cap, the one finance set after the last invoice landed and someone senior asked what exactly we were getting for the money. For the rest of the month he had to write code the way he did in 2023.
He couldn't. Not wouldn't. Couldn't. He'd shipped for eighteen months straight, his velocity numbers were the envy of the team, and somewhere in those eighteen months he had quietly stopped being able to do the job without the machine doing it for him. The cap didn't slow him down. It exposed him.
There's a term going around for this, and people either laugh at it or go very quiet: AI psychosis. It's what happens when engineers, and frankly everyone adjacent to engineering, spend two years outsourcing their thinking so completely to LLMs that they forget how to do their actual jobs.
How we got here
This wasn't a gradual slide. BigAI arrived fast and arrived everywhere. Design, architecture, code, tests, automation, debugging. There wasn't a discipline it left alone. For a while it felt like acceleration. Ship faster, think less, produce more. The tools were cheap, the output was good enough, and the dashboards looked fantastic.
Then the token costs went through the roof.
Companies are scaling back. Usage policies are tightening. The free lunch is over, and what's underneath it is worth looking at directly.
I wrote in January that when AI availability got constrained, teams would discover their generated systems fail first because the understanding was never there. That was a prediction. This is the invoice.
Two failure modes
There are two, and I worry about the second far more than the first.
The first is skill atrophy. Developers, PMs, designers, QA engineers who knew their craft and slowly stopped exercising it. They handed more and more to the model and the muscle faded. That's recoverable. Humbling, uncomfortable, but recoverable. You do the work by hand again until it comes back.
The second is harder, because there was nothing to recover to. These are the people who used AI to produce output they never understood in the first place. Developers shipping code they couldn't explain. QA engineers whose automation passed review but who couldn't tell you what it tested. POs presenting ideas assembled by a model, fielding questions with a confidence that had no floor under it. The output looked right. The understanding wasn't there. Nobody checked.
What makes this worth writing about isn't that it happened. It's the scale. This isn't a handful of people cutting corners. It's a pattern across disciplines, organisations and levels of seniority. AI democratised the appearance of competence faster than anyone built the real thing.
What it actually looks like
The people who've been through the full arc and come out the other side describe it consistently enough to be worth laying out, because a lot of people are living it right now without naming it.
The entry point is a dopamine hook. The slot-machine quality of firing off a prompt and coming back to see whether it did something impressive. One sufficiently good output, something clever enough that you think I might have done it that way myself, and you stop reading the rest so carefully. Then, because the models were slow, you parallelise. Multiple agents at once, multiple work trees open, four or five things in flight so you're never waiting. The metric shifts. How big a diff can I land today? How much ground can I cover?
At the peak, some stop reviewing the code at all. Not slowed down. Stopped. There's no time to read it. We have to keep going.
The signs it's gone wrong are physical. Waking in the night to fire off another prompt. The laptop in the bedroom. Not being able to sit through thirty minutes without doing more. The burnout that eventually forces a stop is not gentle.
I'm not writing this from the outside. I've had the laptop in the bedroom.
The honest account from the people who recovered is not that they quit AI. It's that they slowed down, went single-threaded, and became outcome-focused instead of output-focused. The question stopped being how much code can I generate and became is the thing I'm building any good.
So where does that leave us
AI is still a force multiplier. But we're entering an era where that has to mean something deliberate. Not the first tool you reach for. A considered one. I expect to see companies stand up dedicated teams around AI usage and acceleration, whose entire job is to think carefully about where and how it gets applied. A considered layer sitting on top of the chaos.
Meanwhile the day-to-day teams will work inside tight token budgets. A couple of dozen dollars a month per developer sounds reasonable until you realise that's one badly-built prompt away from hitting the ceiling before the sprint closes. Scarcity forces discipline. When every token costs something visible, people think before they paste.
What comes out of that pressure is two camps, with the middle ground priced out of existence.
The first camp uses AI for the simple stuff. Boilerplate, repetitive mapping, scaffolding they could write in twenty minutes but knock out in two. Low cognitive overhead, output that's trivial to verify precisely because they already understand it. Acceleration for the obvious.
The second saves it for the genuinely hard problems. The moment a developer stares at something and thinks I don't even know where to start. They reach for AI not to skip the thinking but to get a foothold. A suggested approach, an unfamiliar pattern, a starting point they then pull apart and actually understand. A guide into the unknown, not a ghost-writer for the known.
Both are legitimate. Both require you to engage with what comes back.
What's no longer viable is the middle. The sprawling, multi-turn, token-heavy sessions that used to carry people through moderately complex work they only half-understood. That's where the budget runs out. A tight allowance won't get you from confused to done on a real feature through prompting alone. The people who leaned on AI hardest for that middle tier, the work that wasn't simple enough to verify in a glance and wasn't hard enough to justify real learning, are the ones most exposed right now.
One more warning sign
This one is less about how you use AI and more about what you've started to believe.
If you find yourself nodding along every time Jensen Huang tells a room that software engineers are finished, or Sam Altman suggests most knowledge work gets automated inside the year, stop and ask what it costs them if you believe it. Nvidia sells more chips when developers panic-adopt AI infrastructure. OpenAI raises at a higher valuation when the narrative is inevitability. These aren't neutral observers sharing an inconvenient truth. They're businessmen steering a market, and the product needs you to believe resistance is futile and your skills are already gone.
Then look at their job listings. They're hiring engineers by the hundred. That contradiction should land.
It usually doesn't when you're deep in it. That's part of how you know you're deep in it.
Recovery
For skill atrophy, the plan is simple and unpleasant: do the work by hand again until it returns. For the people who never had the skill, it's harder. They have to actually learn the thing they were producing, which is a different and longer road. For organisations, recovery starts with admitting that the last two years built a competence debt cheap tokens hid, and expensive ones are now surfacing line by line.
The scarcity doesn't punish the extremes. It punishes the dependency.