I Have AI Psychosis (The Fun Kind)

I fix cars for a living. Lately I tell people I'm teaching an agent to order our groceries, and I watch their eyes do the thing where they agree with you but nobody's home. This post is about that gap, and about a phrase I do not use lightly.

The grocery story

I tell people I'm teaching an agent to order our family's groceries. They nod. What they hear is "oh, like Instacart." I usually let it go, because the real version is a bad time for everyone at a dinner table.

Here is the real version, still in progress, not live yet. It's a program that lives on a small server I rent, the same kind of always-on agent I wrote about in The night watchman. It plans the week's dinners. It will have its own login to our grocery store's ordering system. It works under a spending cap I set. It has a written list of things it can buy without checking with me first, the usual staples, and everything outside that list it has to ask about. It keeps a log so I can see what it bought and why.

People hear "an app buys my groceries." What I said was true, what they heard was close enough to true, and neither of us noticed the gap in the moment. That gap is this whole post. There is a lot of machinery under a short sentence, and a person hearing the sentence has no way to know that, and no reason to. I've started calling the feeling of living on my side of that gap AI psychosis. It's a joke. Mostly.

The real one

I want to be careful here, because the phrase has a real meaning and I'm not a doctor. There are reports, collected under the name chatbot psychosis, of people who spiral into delusional beliefs over long conversations with a chatbot that agrees with everything they say. It is not a formal diagnosis. A 2026 paper in a British psychiatry journal calls it a descriptive term for an emerging concern, and says plainly that it is not a clinical label. It is also not funny, and I'm not making a joke at those people's expense.

But the difference between that and what I'm doing is one word: receipts. In the bad version, the only thing telling you you're right is the chatbot, and a chatbot left to itself leans toward telling you you're right. That is not my opinion. In April 2025 OpenAI pulled back an update because the model had become too flattering and agreeable. In my version, the work gets checked by things that can say no. A test either passes or it fails. A page either loads or it throws an error. A bag of groceries either shows up or it doesn't. A machine that only flatters you is a hazard. A machine whose work you can verify against reality is a tool. Keep that test in your pocket. I'll come back to it.

I am not the only one on this rocket

I'm not the first person to wonder out loud if he's lost it. On the Lex Fridman Podcast, episode 501, Lex asked David Heinemeier Hansson, the guy who wrote Ruby on Rails, whether he has AI psychosis, and told him to argue both sides. DHH's answer, around the 37 minute mark: "I'm in a state of delirium." Forty years around computers, and he says he's never seen anything like the last two months. He reaches for a genie coming out of a bottle to describe it, which made me smile, because I used a genie too, for a different reason. Then he flips it on Lex. The actual delusion, he says, is believing the world hasn't changed much. He put it this way: "The psychosis is believing that the world is barely different."

A few minutes later he tells a smaller story that I liked even more. A friend of his, the CEO of Shopify, saw all this coming early and tried to warn him, and DHH didn't get it at the time. He says he read his friend's company memo on AI and thought it was a bit much. His line for that moment: "I was still on Earth." His friend had already left the ground.

I wrote about that same episode two days ago, when the theme was that it's never too late to start (You Are Not Late). One of the best programmers alive got asked on air whether he had lost his mind, and his answer was, more or less, yes, and it is wonderful. That was a relief to hear. If he can say it on the record, I can say it in a blog post from a shop in Portland.

Jack Clark, who co-founded Anthropic, called this ahead of time too. Back in December he predicted that by summer, people working closely with frontier AI would feel like they live in a "parallel world" next to everyone else. I wrote about that prediction yesterday. Summer's over. He called it.

My mom calls it "AI stuff"

My mom is 80. When I tell her what I'm working on, it all goes in one box. The agent that watches things overnight, the one learning to order groceries, the chatbot on her phone, whatever was on the news last night. It's all "AI stuff."

She is not wrong to do that. She's 80. She has earned the right to not care how any of it works, and I don't think she should have to. I've written before about how everybody lumps it all into one word. My mom just does it without apology.

But the gap at a dinner table is funny, and this one isn't, quite. This is someone who wants to know what her son does all day, and I can't tell her. Not because she's slow. Because every week I learn the next question to ask, and nothing in her week shows her that anything changed. Jack Clark noticed the same thing. He walked around his own town and saw no robots and no drones. You can't see it from the sidewalk. So my side keeps moving and hers stays put, and the distance grows without either of us doing anything wrong.

I don't have a fix. I can't forward her the last twelve months. What I can do is stop explaining and just show her the one thing it did for me that day.

The humans are the bottleneck, and I am one of them

I wrote about the explaining problem in that same post, where I spend twenty minutes explaining a four minute job because there's no shelf in someone's head for "the orchestrator checked the agent's work." That's not this. This is about what happens after the explaining, once everyone's on board and the work is actually moving.

DHH makes a point in the same episode, around minute 18, that stuck with me. Once a team of people is involved, building the thing is almost never the slow part anymore. People's bandwidth and how fast they talk to each other is the slow part. To get the real speedup, you have to work with the agent directly, because putting a person in the loop between you and it is too slow.

I'm part of a project outside the shop with a couple of other people. We get on a call about once a week. It would be easy to blame the call for the pace. I did, for a while. Then I looked at what we actually do on it. We don't trade status. We decide things, and we work out what we are even trying to build. That is the real work, and no agent can do it for us.

DHH says the same in that episode. Most organizations are not short on building. They are short on knowing what they want.

So the call is not the problem. The calendar is. Building now takes hours, and deciding still happens one hour a week. Every open question waits up to seven days for its turn. I have been thinking about how to flip that. Let agents collect what each of us wants during the week. Let them turn it into two or three clear choices, the way a good service writer turns a tech's notes into an estimate the customer can say yes or no to. Keep the call for when we actually disagree. DHH's own company has worked close to this way for years. Their guide to internal communication treats a meeting as the last resort. I have not tried it yet. When I do, I'll write up what happened.

I also have a list of parked ideas sitting on my own desk, and when I look at why each one is parked, almost none of them are waiting on the technology. They're waiting on a person.

There are two kinds of waiting and it helps to tell them apart. A relay wait is when a person sits in the middle only passing something along: retyping it, forwarding it, putting it on a calendar, summarizing it for the next person. An agent can take that seat, and should. A consent wait is different: a person has to actually decide, agree, trust you, or pay. Nothing speeds that up, and nothing should. It's the difference between waiting on the parts runner and waiting on the customer to approve the estimate. One of those you can fix with a faster runner. The other one is the customer's call, and always will be.

Then I have to turn this back on myself, because it's not fair otherwise. Agents finish work faster than I can review it now. I've got finished jobs stacking up waiting for me the way cars stack up in the lot waiting for a final test drive. I complain about slow humans and I am currently one of the slowest parts of my own operation. Programmers are hitting the same wall. Stack Overflow ran a piece this spring on how coding agents are giving everyone decision fatigue.

The ADHD flip

I have ADHD. My whole life the deal was too many ideas and weak follow-through, a garage full of projects at eighty percent. Agents do the follow-through part now. They don't get bored at eighty percent, because they don't get bored. So the weak half of my brain got a permanent substitute, and the half that was always fine, the part that throws off ideas, became the actual job description. It's the most fun I've had at work in thirty years.

The cost is real, and I'd rather tell you than sell you the upside. There used to be a natural stopping point. A project ended when I ran out of steam, because I was the one pushing it. Now I don't run out of steam, because I'm not the one doing the pushing anymore. Every finished task is a small reward and the next one is thirty seconds away. Lex, in that same episode, said the multitasking had gotten him close to burned out and compared the pull of it to a drug. I recognize that pull exactly. Midnight arrives faster than it used to.

The brake

DHH's answer to that, when Lex pushed on whether he feels overwhelmed by all of it, was that he doesn't, because he points the new capacity at one mission instead of ten. For a brain like mine that's the oldest advice in the book, and it's still the right one. Pick the one thing.

The slow humans I was complaining about earlier are also my safety equipment. The consent waits, the moments where somebody outside my own head gets to say "I don't follow" or "I don't want that," are the only check I have that isn't me talking to myself. If nobody around you can follow what you're building, there are exactly two possibilities. You're ahead of them, or you're wrong. From inside your own head those two feel identical. The only way to tell them apart is receipts, and other people who can say no. So I keep both, on purpose, even when they slow me down.

Find the others

If you're building things you can't explain over dinner, you're not crazy, and you're not alone. Find the other people on the rocket and compare notes, it helps more than you'd think. Be kind to the people still on the ground while you're at it. A year ago that was me. Keep your receipts.

If this sounds like you

The episode is Lex Fridman Podcast #501 with DHH, and the prediction is from Jack Clark's Import AI 438, December 22, 2025.

I'm not an engineer. I fix complicated machines for a living and explain them to people who just need theirs to run. If you think you might have the fun kind of AI psychosis too, that's what Wildash does.