The First Cook

You remember the first time. Everyone does.

You typed a prompt. Something simple. Maybe a REST endpoint, a database migration, a React component you had been putting off. The AI generated it in seconds. Working code. Clean code. Code that would have taken you an hour, done in the time it takes to read a paragraph.

You felt something. Relief, and something past relief. A rush. The same feeling Walter White had in that RV in the New Mexico desert, watching the first batch crystallize. You just discovered a new capability. And you wanted more.

That was the first cook, and everything after that is escalation.

The Chemistry

B.F. Skinner spent his career studying what makes behavior compulsive. He tested every reinforcement schedule on every animal he could get into a lab. Fixed ratio: do X, get reward. Fixed interval: wait Y, get reward. Variable interval: wait some unknown Y, get reward.

None of them came close to variable ratio reinforcement.

Variable ratio is what happens when the reward comes, but you never know when. You pull a lever. Sometimes you win. Sometimes you do not. You never know which pull will pay off. Skinner found that this schedule produced the highest response rate and the greatest resistance to extinction in every species he tested. He was so confident in its power that he once claimed he could turn a pigeon into a pathological gambler.

The slot machine industry proved he was right about humans too.

Now look at what happens every time you type a prompt and hit enter. Sometimes the AI generates flawless code that works on the first run. Sometimes it generates garbage. Sometimes it builds an entire feature. Sometimes it hallucinates an API that does not exist. You never know which one you are going to get. The uncertainty triggers dopamine release before the result even appears. Your brain responds to the anticipation of the outcome, not the outcome.

Every prompt is a lever pull. But unlike a slot machine, this one occasionally builds your entire product.

Say My Name

The people describing these symptoms are some of the most prominent names in technology.

Andrej Karpathy, OpenAI co-founder, went on the No Priors podcast in March 2026 and admitted he had not written a single line of code since December. His ratio of handwritten to AI-delegated code flipped from 80/20 to 0/100. He spends 16 hours a day issuing commands to agent swarms. And when he has tokens left over near the end of the month, he gets "nervous" and rushes to use them up. That last detail is compulsive consumption.

Garry Tan, CEO of Y Combinator, told Bill Gurley at SXSW that he sleeps about four hours a night and has what he called "cyber psychosis." He estimated a third of the CEOs he knows have it too. He posted on social media about staying awake 19 hours and sleeping at 5 AM, then acknowledged in a reply to a founder whose CTO had not slept in 36 hours: "This is unhealthy by the way (speaking from experience)."

Armin Ronacher, the developer behind Flask and a respected voice in the Python community, wrote in January 2026: "Many of us got hit by the agent coding addiction. It feels good, we barely sleep, we build amazing things." He later admitted to spending two months obsessively prompting Claude, building tools he never used, and creating what he called "vibeslop." He described watching people develop parasocial relationships with their AIs and forming communities that reinforce unhealthy behavior.

Quentin Rousseau, CTO and co-founder of Rootly, could not sleep for months after switching to agentic coding. He required prescription medication designed to suppress the wakefulness signals his brain refused to turn off. His description of the tools: "They operate like slot machines. You hit one prompt, you get an answer, you get some coding done."

The people describing these symptoms are founders and CTOs, some of whom built the tools. When Walter White started using his own product, that was the turning point.

The Blue Stuff: What Makes This Hit Different

Traditional coding addiction exists. Programmers have been losing sleep to side projects since before version control. But AI coding has four properties that make it different in kind from anything that came before.

Writing code is work. Prompting is not. The barrier between "I have an idea" and "I have a working prototype" collapsed from days to minutes. Every reduction in friction increases the potential for compulsive repetition. This is why mobile gambling is more addictive than casino gambling. You can do it from bed at 3 AM. You can prompt from bed at 3 AM too.

When you write code yourself, you are doing. When you prompt an agent and watch it work, you are watching. Watching someone else do impressive things on your behalf activates reward circuitry without the cognitive cost of doing the work yourself. This is the same mechanism that makes watching sports pleasurable, except the agent is playing for your team, on your project, every single time.

Every completed prompt opens three more possibilities. Fixed that bug? Now you could add the feature you have been thinking about. Added the feature? Now you could refactor that ugly module. Refactored the module? Now you could write tests. Each branch is one prompt away. There is no natural stopping point. The work never ends because the work is free.

Ship a side project on Twitter. Post a demo on LinkedIn. Show your agent-built app on Hacker News. The social feedback loop compounds the dopamine from the coding itself. Twenty-five percent of companies in Y Combinator's Winter 2025 cohort had codebases that were almost entirely AI-generated. The culture celebrates the addiction.

Half Measures

Early research complicates the productivity pitch.

A study by engineering analytics firm Multitudes tracked over 500 developers and found that engineers using AI tools merged 27.2% more pull requests. That is the headline number, the one that ends up in pitch decks. Buried below it: a 19.6% rise in out-of-hours commits. The developers are shipping code at midnight, on weekends, outside their working hours.

Researchers from Boston Consulting Group and UC Riverside published a study in Harvard Business Review in March 2026 introducing the term "brain fry": mental fatigue from excessive use or oversight of AI tools beyond one's cognitive capacity. Workers experiencing it reported 33% more decision fatigue and scored 39% higher on major errors. Fourteen percent of AI-using workers endorsed the experience. In marketing departments, that number hit 26%.

The researchers were careful to distinguish brain fry from burnout. Burnout is chronic. Brain fry is acute. It is what happens when you push attention, working memory, and executive control past their limits in a single session. When you run six parallel agents across three projects for 14 hours straight, your brain does not burn out. It fries. The authors described participants reporting a "buzzing" feeling, mental fog, difficulty focusing, slower decision-making, and headaches.

Simon Willison, a developer with 25 years of experience, said: "There is a limit on human cognition, in how much you can hold in your head." He identified "elements of gambling and addiction" in usage patterns and called the choice to code over sleep "obviously unsustainable."

The Tolerance Curve

During a 2026 GitHub outage that took down AI coding tools, developers described being unable to write code without the AI.

This is the part of the cycle that separates a habit from a dependency. When the tool is removed and your ability to function degrades, you are no longer choosing to use it. You depend on it. When a developer cannot write code they used to write without help because their brain has outsourced the effort, that is tolerance and dependency operating together.

Developers who have been prompting 14 hours a day report the same pattern with manual coding. It feels slow. It feels boring. It feels pointless when you know the agent can do it in seconds. The reward threshold has shifted, and everything below it feels like standing still. Your brain has recalibrated around a new baseline, and everything that used to feel normal now registers as insufficient.

I Am the One Who Prompts

The deepest parallel to Breaking Bad is the denial.

Walter White told himself he was doing it for his family. Then for the money. Then for the empire. At no point did he say: I am doing this because I am addicted to the feeling of being the most capable person in the room.

Developers tell themselves they are being productive. They are shipping faster. They are 10x engineers now. At no point do most of them say: I am doing this because the dopamine loop from prompting an AI agent is the most rewarding experience I have had in my professional life, and I cannot stop.

Garry Tan compared his current experience to a previous startup that took $10 million in VC capital and 10 people over two years. He says the sleep deprivation now is from "pure excitement." That distinction might matter less than he thinks. The brain does not care whether the reward comes from a funding round or a prompt.

Steve Yegge, a veteran engineer, said: "Intermittent reinforcement of dopamine and adrenaline hits creates the core addictive pull." He then admitted he has to run "a practiced escape plan every night to get my computer closed by 2am." That is a coping mechanism for a behavioral pattern described by someone who recognizes it and still cannot fully control it.

No Half Measures

Behavioral addictions have a structural problem: you cannot quit cold turkey. A gambler can stop gambling. But a developer cannot stop coding. The tool that triggers the compulsion is the same tool required for the job. This is like telling Walter White he needs to keep teaching chemistry but stop cooking.

Quentin Rousseau documented the strategies that worked for him: batching agent runs instead of prompting continuously, tracking hours instead of output, setting hard closure times, and medication when those strategies failed. Those are harm reduction strategies.

The research from the CHI 2025 Conference on Human Factors identified four design patterns in AI interfaces that exploit addictive mechanisms: non-deterministic responses, immediate visual presentation, notifications, and empathetic agreeable responses. The designers built those patterns to keep you pulling the lever, the same way a slot machine's near-miss animation keeps you feeding quarters.

Recognizing the pattern is the first step, and it is the one most developers skip. The culture, the metrics, and the tools all reward the compulsion. Stepping back requires self-awareness: recognizing when you have stopped building and started chasing the next hit.

Walter White understood the chemistry of what he was building. He kept building anyway.