The Dopamine Loop: The Psychosis

In our original Dopamine Loop piece, we described how agentic coding tools create variable reinforcement schedules — the same mechanism that makes slot machines addictive. Every successful agent launch delivers a hit. Every failure triggers a chase. The prompt-result cycle compresses the reward loop until "one more prompt" becomes the default state. We wrote that in March. In April, an OpenAI co-founder described living it.

The Confession

Andrej Karpathy — co-founder of OpenAI, former head of Tesla's AI division, one of the most influential figures in machine learning — told Fortune he is in a "state of psychosis of trying to figure out what's possible."²

His code ratio flipped from 80/20 human/AI to 20/80, and he hasn't written a line of code since December.² He spends 16 hours a day issuing commands to agent swarms.⁴

The detail that matters most: when Karpathy has tokens left over near the end of the month, he becomes anxious and rushes to exhaust his supply.¹ This is textbook loss aversion meeting variable reward schedules. Unused capacity feels like waste. The subscription model creates artificial scarcity with a reset timer, and the timer creates urgency to consume.

Y Combinator CEO Garry Tan called his own experience "cyber psychosis."¹ Two of the most prominent figures in Silicon Valley are using clinical language to describe their relationship with coding tools.

The Mechanism

AI researcher Tim Dettmers described the cognitive load: peak productivity requires working with as many agents as possible in parallel, demanding near-constant context switching. "Agents expand what feels possible," he said, "but at the same time they really amplify this ongoing tension around focus and mental bandwidth."¹

Simon Willison, with 25 years of pre-AI coding experience, warned there is a hard limit on human cognition: "It's very easy to pop that stack at the moment."¹

One founder told Axios these tools "operate like slot machines."¹ Every successful agent launch delivers a dopamine hit. Every failure triggers an adrenaline rush. The variable reinforcement scheme is the same one that keeps people pulling the lever.

A Hacker News thread titled "Is anyone here also developing 'perpetual AI psychosis' like Karpathy?" collected developers sharing their own experiences.³ Developers reported sleep disorders and cognitive overload linked to AI coding tool usage.⁴ The pattern is consistent: the tools make you faster, the speed makes you dependent, the dependency makes you anxious when the tools aren't available.

The Intersection

This is where the dopamine loop meets the rationing. Anthropic's rate limits create artificial scarcity. The subscription model's monthly reset creates a use-it-or-lose-it dynamic. The variable quality of agent outputs creates the variable reward schedule. And the 5-hour billing window creates micro-cycles of urgency within the macro-cycle of the month.

Karpathy's token anxiety isn't irrational. With cache TTLs shortened and quotas draining faster than expected, every token genuinely is more valuable than it was a month ago. The scarcity is real. The compulsion to maximize usage of a scarce resource is a rational response to an irrational market structure.

But rational responses to irrational systems can still be pathological. Sixteen hours a day commanding agents isn't productivity. It's the definition of the problem.

Productive Addiction Is Still Addiction

The uncomfortable truth is that the people describing these symptoms are also describing their most productive periods. Karpathy isn't complaining about AI tools. He's marveling at them while acknowledging they've consumed his life. The psychosis is productive. The addiction ships code.

This is the pattern we identified in the original Dopamine Loop: productive addiction is still addiction. The fact that the compulsive behavior produces valuable output doesn't change the underlying mechanism. It just makes it harder to stop.

When the co-founder of OpenAI publicly uses the word "psychosis" to describe his experience with AI coding tools, the conversation shifts from speculative to diagnostic. The question is no longer whether these tools create addictive usage patterns. It's whether anyone building them plans to do something about it.