Let’s talk about AI Psychosis/Brain Fry

1

Top highlight

50

1

In a recent interview with Sarah Guo on No Priors, Andrej Karpathy said something that made me stop and stare at the screen a little longer than usual.

He said he’s in “a state of psychosis trying to figure out what’s possible.” He said he hasn’t typed a line of code since December. He said he watches people on Twitter building incredible things with AI agents and feels extremely nervous that he’s not at the forefront.

Andrej Karpathy. Worried he’s falling behind.

That hit me harder than I expected, not because it sounded dramatic, but because it sounded familiar.

He was trying to describe something I’ve felt for a while now: the obsession, the mental noise, the sense that even when I step away from the tools, some part of my brain is still pacing. When his comments started spreading, a lot of people responded like they’d just been handed language for something they’d already been dealing with. That was my reaction too. Not surprise. Recognition.

Now we have a name for it. So it’s worth talking about what it actually feels like from the inside.

What AI Psychosis Means

To me, it doesn’t mean psychosis in the clinical sense. It’s a very specific mental state that shows up when you’re not just watching AI move fast, but trying to work inside that speed every day without losing your footing.

It comes from three things colliding at once.

First, the mental strain of supervising AI. A recent Harvard Business Review study found that people actively managing multiple AI systems at once reported more fatigue, more information overload, and more mental effort than people with lighter oversight responsibilities. The researchers called it “AI brain fry.” That tracks with what a lot of people, myself included, seem to be feeling. AI can absolutely remove friction from the work. But if you’re the one monitoring it, steering it, correcting it, and keeping multiple threads from drifting, the strain just comes back in a different form. Sometimes it feels like I’m not doing the work anymore. I’m managing work that’s being done around me, and somehow that can be even more tiring.

Second, the fear of falling behind becomes constant. Karpathy said he wants to be at the forefront and he’s anxious that he isn’t. A lot of people understood that immediately. I did too. This doesn’t feel like ordinary professional anxiety to me. It feels more like living in a field where a few weeks of inattention can make your mental model feel old. I tried stepping away for a while, hoping distance would calm all this down. It didn’t. Mostly it just made me feel like the ground had moved without me.

Third, the natural stopping point is gone. Before all this, there were limits built into the work. Writing took time. Debugging took time. Research took time. There was friction everywhere, and yes, it was annoying, but it also created a boundary. Now that boundary is weaker. The tools remove so much drag that it’s easy to keep going long past the point where I should have stopped. What used to feel like leverage starts turning into expectation. Then expectation turns into baseline.

That’s the state I’m talking about.

What It Actually Looks Like

When other people describe this, the details vary, but the pattern is pretty consistent. A lot of it feels familiar.

There’s the buzzing feeling. Other people describe it as a low-level hum, like part of their brain is still switched on somewhere. That rings true to a lot of people right now, including me. It’s not always panic. It’s more like too many tabs open in my head and none of them fully closed, even when I’m technically off.

Then there’s compulsive monitoring. A lot of people talk about how hard it is to leave an agent alone once it’s running. That pattern is familiar. I’ll tell myself I’ll come back later, then check five minutes later anyway. Then again during daily standup. Then at lunch. Then while driving, knowing I shouldn’t be doing it, knowing it’s bad. Once the agent is running, part of my attention runs with it.

Then context starts splitting apart. People describe the mental strain of keeping too many threads alive at once, and that feels exactly right. One agent is doing one thing, another is halfway through something else, a third is running in a different codebase, and I’m trying to hold all of it in my head. It doesn’t feel like normal multitasking. It feels like my attention getting chopped into pieces.

Then there’s the loop where I can’t stop pushing it. Other people describe the same urge to keep testing, tweaking, and seeing what else the tools can do. That part feels familiar too. One more workflow. One more setup. One more prompt structure. Even when it makes me tired, I keep going, because the pace of improvement makes it feel stupid to look away.

And underneath all of that, there’s knowledge decay anxiety. One engineer who wrote about his own AI fatigue described building a sophisticated prompt engineering workflow over two weeks in early 2025 — carefully tuned, working well. Three months later, model updates and shifting best practices made half his templates obsolete. “Those two weeks were gone. Not invested. Spent.” .

Why This Happens

This doesn’t feel like it’s just about weak boundaries or bad habits. A lot of it comes from the structure of the work itself.

The HBR researchers made a distinction that makes sense to me: **the real issue isn’t AI use by itself. It’s AI oversight.**When AI quietly handles repetitive work, the burden can actually go down. But when you’re the one supervising output, checking for errors, redirecting the process, and switching between tools, your brain stays engaged the whole time. You’re never fully at rest because the system only works if you’re still there to catch what it misses.

Then work expectations start shifting too.

That part feels especially real now. Early on, using AI heavily felt like an advantage. It felt like I was ahead of the curve, like I was finding leverage before most people understood what the tools could do. But over time that stopped feeling exceptional. It just became the new expectation. The output that once felt impressive became normal. Then normal became the minimum. That’s when the pressure changes shape.

And the fear of falling behind is not imagined. Karpathy said that if you haven’t kept up over the last thirty days, your view may already be outdated. That sounds extreme until you live inside this for a while. New models. New tools. New agent patterns. New frameworks. Constant movement. The surface area keeps expanding, and no matter how engaged I am, I still have the sense that I’m only seeing part of it.

That’s why stepping away hasn’t felt simple, at least not for many people trying to stay current. In slower fields, a break feels restorative. Here, sometimes it feels expensive. You come back and the tools are better, the workflows have changed, and the gap doesn’t feel theoretical. It feels real.

What I’ve Had To Admit To Myself

One thing that’s hard to admit is that the pressure doesn’t just come from the outside. Some of it comes from ambition, curiosity, and the desire to stay close to what’s changing.

I am genuinely fascinated by this stuff. I don’t want to just use the tools. I want to understand what they’re making possible. I want to see around the corner a little. I want to know whether the way I’ve worked for years is about to become irrelevant, or whether it still matters in a deeper way. So part of this stress is curiosity mixed with ambition mixed with fear, and that combination is hard to regulate.

I’ve also had to admit that unplugging didn’t fix it for me. I tried stepping back. I tried reducing how much AI touched my personal workflow. I thought distance might quiet the obsession.

Instead, I mostly felt behind.

When I came back, the models had improved, the workflows had changed, and the baseline had moved again. That didn’t feel calming. It felt like proof that the anxiety had something real underneath it.

And I think that’s why this is hard to talk about cleanly. Because some of the fear is irrational, sure. But not all of it is. Some of it is just what it feels like to live in a field that’s changing fast enough to make your own judgment feel unstable in real time.

What Should We Do About It

The answer probably isn’t to walk away from AI. It’s to work with it in a way that doesn’t fry the part of you that’s supposed to stay sharp.

The first thing I’m trying to get better at is limiting how many active things I supervise at once. Every time I convince myself I can track a bunch of parallel agent work without a cost, I end up proving myself wrong. My attention has limits, whether I like that or not.

I’ll also try to stop hovering. The worst mental drain, at least for me, comes from live-monitoring everything. Watching every step. Checking too often. Staying half-attached to unfinished processes all day. It feels productive in the moment, but I don’t think it is. It mostly just keeps my brain in a constant state of partial engagement.

I’m trying to protect some work that I do without AI in the loop too. Not because I think that’s morally better. Just because I can feel the difference when I haven’t used my own mind in a direct, uninterrupted way for a while. Writing, reviewing, thinking through architecture, making sense of tradeoffs myself. I still need that. Maybe more now than before.

I’ll also try to separate actual learning from ambient panic. Not every new thing deserves my attention. Not every release matters. Not every framework is worth caring about. The problem is that when everything arrives with urgency, it’s easy to lose the ability to tell signal from noise.

And honestly, sleep matters more here than people want to admit. Fatigue makes me sloppier, more suggestible, more likely to trust output I should question. A tired human paired with fast, confident model output is not a great system.

If you’re leading a team, I think this matters even more. You can’t keep raising output expectations forever just because the tools got faster. At some point you’re not creating leverage anymore. You’re creating cognitive debt.

The Part I Still Can’t Shake

Karpathy used the word psychosis because softer words didn’t quite get at it.

“Fatigue” gets part of it. “Burnout” gets part of it. “Overload” gets part of it. But none of those fully capture what it feels like to keep trying to build a stable mental model inside a field that won’t stop moving.

I don’t think we’re broken. I think we’re early.

But early isn’t free.

The people who do well in this era probably won’t just be the ones who move the fastest. I think they’ll be the ones who can stay clear-headed while everything accelerates. The ones who know when to hand something off to the agent and when to slow down and think for themselves. The ones who can use the tools without getting absorbed by them.

And I think that’s where a lot of the fear lives. Not just in whether we’ll stay relevant, but in whether the kind of judgment we’ve spent years building will still count the same way in a world where the tools can move faster than we can.

That’s the part that feels personal.

And for now, it’s still one of the few parts the AI can’t carry for us.

50

50

1