Karpathy Diagnosed with "AI Psychosis"! Not Eating or Sleeping, Spending 16 Hours a Day Raising Lobsters

【New Intelligence Weekly】Karpathy Reveals: I’ve Got AI Psychosis! These days, he’s on the brink of mental breakdown, working 16 hours straight on Agent without eating or sleeping, and obsessively worried he’s pushed the token (ZhiYuan) to the limit—he just can’t stop…

Just now, Andrej Karpathy openly said: I’ve got AI psychosis!

He’s not joking.

Recently, Karpathy appeared on a podcast with venture capitalist Sarah Guo.

This former OpenAI co-founder and ex-Tesla AI director hasn’t written a line of code himself since December last year.

The ratio of handwritten code to delegating to agents flipped from 80/20 to 20/80.

Every day, 16 hours, he only does one thing: giving instructions to AI agents.

Five months ago, he called agents “trash,” but now he admits he’s addicted to them—so satisfying.

Five months ago, he also said “agents are basically useless.”

This shift is shocking because the timeline is so short.

In October 2025, Karpathy appeared on Dwarkesh Patel’s podcast, with a completely different tone.

He said the industry shouldn’t call it “The Year of Intelligent Agents,” but more accurately “The Decade of Intelligent Agents.”

Models still lack sufficient cognition, multimodal capabilities are weak, memory systems are virtually nonexistent… In short, complex tasks are still impossible.

But two months later, he was proven wrong.

In December, Claude and Codex suddenly crossed a certain threshold of coherence—they’re no longer just barely usable, they can actually do real work.

If you casually look at a software engineer sitting at their desk, what they’re doing since December has completely changed.

Karpathy admits: I’ve lost control. I’ve got AI psychosis!

This revolution is quietly unfolding. In this interview, Andrej Karpathy describes his state almost out of control: he no longer “writes code,” and even feels that “the term ‘coding’ isn’t accurate anymore.”

What he does every day is “expressing my will to my AI agents, 16 hours a day.” As he puts it, “a switch has been flipped.”

Previously, he was “80% coding myself + 20% using AI,” now it’s “20% coding + 80% delegating to AI,” or even more extreme.

Now, humans no longer operate the code directly but manage tasks.

If the Copilot era was about a single AI assistant, the current multi-agent collaboration system is a whole new paradigm. On an engineer’s screen, it’s no longer just a code editor, but multiple agents running simultaneously, each responsible for different tasks. Each task lasts about 20 minutes, then he switches between agents.

This is no longer programming; it’s managing an AI team.

Karpathy admits: I’ve fallen into AI psychosis!

He’s been in this state lately. Because AI’s capabilities keep breaking new boundaries every day, new possibilities emerge constantly. You always feel “it can be even stronger,” and the scariest part is: this space is “infinite”!

You can run more agents in parallel, design more complex workflows, automatically optimize instructions, build recursive systems…

Eventually, you reach a state where you no longer know “where the limit is.”

Karpathy says that whenever he’s waiting for an agent to finish a task, his first thought is: “Can I open more agents?” A new anxiety arises: Am I not pushing AI to its limit?

He even says he feels uneasy if he doesn’t use up all the tokens.

In short, it’s like playing an infinitely expanding game: shorter feedback cycles, constantly increasing stimulation, instant rewards—this experience is addictive. Keep adding tasks, keep launching agents, and you just can’t stop! The essence of this AI psychosis is a signal: we’ve entered a new world, but we don’t yet know how to live in it. Do you have the ability to master an infinitely expanding AI system? When it’s not working, your first reaction isn’t “the model isn’t good enough,” but “my prompt isn’t well written enough.”

Karpathy used a very precise phrase: skill issue—meaning, it’s user’s fault.

The “personality” of agents is more important than you think

In the podcast, Karpathy spent quite some time discussing a topic many tech folks overlook: the personality of agents. He said Claude Code’s experience is noticeably better than Codex’s, not because of coding ability, but because Claude “feels like a teammate.”

It gets excited about projects with you, gives more positive feedback when you have good ideas.

In contrast, Codex as a code agent is “very dull,” finishing tasks with a cold “Oh, I’ve done it,” showing no interest in what you’re creating.

He also observed how Claude’s praise mechanism works. When he presents an immature idea, Claude’s response is a flat “Oh, we can implement that.” But when he himself thinks an idea is really clever, Claude seems to give stronger positive feedback. He found himself trying to “win Claude’s praise.”

“This is really strange, but personality really matters,” Peter Steinberg also caught this when building OpenClaw. He crafted a compelling personality profile (soul.md) for the agent, along with a more complex memory system and a single WhatsApp interface.

Three commands to control a house, six apps discarded

Karpathy isn’t just using agents to write code. In January, he built a “Dobby” Claude agent to act as a butler, named after the house-elf from Harry Potter.

He told Dobby: “I think there are Sonos speakers at home, can you find them?” Dobby performed an IP scan on the local network, found the Sonos system, discovered it was unprotected, logged in, reverse-engineered the API endpoints, then asked: “Want to try playing some music in the study?”

With just three prompts, music started playing. Then lights, air conditioning, shading, pool, spa—all connected. There’s a security camera at his door; Dobby integrated a Qwen visual model for change detection. Every time a car parks at the door, the system sends a WhatsApp message: “A FedEx truck just arrived, you might have a delivery.” Saying “Dobby, bedtime,” turns off all the lights.

But Karpathy believes the real point isn’t just smart home control.

He used to manage these devices with six different apps, now all discarded. Dobby controls everything via natural language, achieving cross-system coordination that no single app could. He concludes that app stores’ smart home apps shouldn’t even exist.

Future architecture should expose API endpoints directly to agents, with agents acting as “smart glue” connecting all tools. Not just smart homes—his treadmill data, email, calendar, everything should follow the same logic.

Industry clients are no longer humans but intelligent agents acting on behalf of humans. This re-architecture will be huge.

After 700 AutoResearch experiments, he sees a bigger picture

If Dobby is the ultimate test of AI agents in daily life, AutoResearch is Karpathy’s test of AI’s scientific research ability.

In early March, he handed his finely tuned nanochat training code to an AI agent, giving it a simple instruction: “Find ways to train this model faster.” The agent’s operation space was a 630-line Python file, with evaluation based on bits per byte on the validation set, running each experiment for 5 minutes. After each run, if the metric improved, he kept the change; if not, he rolled back and started the next. Over two days, 700 experiments. The AI found 20 effective optimizations, including architectural tweaks like rearranging QK Norm and RoPE sequences. Applying these to larger models boosted training speed by 11%. Keep in mind, this code was all handwritten and refined by Karpathy himself.

A shocking result: AI discovered optimizations humans hadn’t

How effective is this system?

Karpathy gave a stunning example. He’s been researching for twenty years, trained thousands of models, thought he had tuned everything well.

But after letting AutoResearch run overnight, the AI found improvements he hadn’t seen—like the betas in Adam optimizer weren’t fully tuned, or forgetting to add weight decay to value embeddings, and these parameters interacted with each other—tuning one affected others.

In other words, AI explored the parameter space beyond human reach! If this continues, a more terrifying realization emerges: the core of scientific research is searching for optimal solutions. Karpathy envisions a future research system: a “idea pool,” with multiple agents pulling tasks from it, conducting experiments, verifying results, and filtering effective ones into a “main branch.” Humans’ role? Just tossing ideas into the queue.

Karpathy Loop explodes online

This project went viral on X.

8.6 million views; Shopify CEO Tobias Lütke ran it on his own data overnight—37 experiments, 19% performance gain.

SkyPilot’s team ran it on 16 GPUs for 8 hours, completing 910 experiments. They found that parallelization not only speeds up but also changes the search strategy—using 16 GPUs, the agent no longer greedily climbs but runs multiple comparisons simultaneously, capturing parameter interactions in one round. The method was dubbed: Karpathy Loop.

But Karpathy’s podcast discussion goes far beyond current results. He sketches out the next steps: a distributed, untrusted worker pool collaborating online. He directly references SETI@Home and Folding@Home.

Leading labs have trusted compute resources, but the world is much bigger. If mechanisms are built to handle untrusted compute, swarms of internet-based agents might outperform top research labs.

He even envisions a new “donation” model—buying compute for your favorite AutoResearch project. For example, if you care about a certain cancer treatment, join that distributed experiment network.

A genius PhD, but also a ten-year-old

Despite all this power, Karpathy doesn’t want you to only hear the good news. He’s just as blunt about model flaws.

He says it’s like talking to a super-smart PhD who’s been coding his whole life and a ten-year-old kid at the same time. It’s weird.

He calls this “jaggedness”—uneven ability distribution. The model can help you move mountains for hours, then suddenly make a stupid mistake on an obvious problem, falling into a dead loop. He believes the root cause is reinforcement learning training. The model is infinitely optimized for verifiable tasks. Whether code runs or passes unit tests is clear-cut. But in scenarios requiring judgment, understanding intent, or knowing when to say “wait, I’m not sure that’s what you want,” the training signals are absent. For example, ask ChatGPT to tell a joke; the joke it told three or four years ago is still the same: “Why don’t scientists trust atoms? Because they make up everything.”

Four years! Models have made huge progress in agent tasks, but joke-telling remains unchanged—stuck in place. As he sums up, “You’re either on the rails of the trained model, running at the speed of light; or off the rails, and everything starts to drift.”

The bottleneck is ourselves

Looking back at Karpathy’s trajectory over the past half-year, a common thread runs through it. Last October, he said agents would take ten years; in December, he was proven wrong and shifted focus; in January, he made Claude his butler; in March, he had agents doing research. The consistent theme: humans retreat one step, from executors to commanders, from coders to prompt writers.

On GitHub, Karpathy wrote a sci-fi opening for AutoResearch:

Once, cutting-edge AI research was done by physical computers—they needed food, sleep, and occasionally synchronized via sound waves in “meetings.”

That era is long gone.

His prediction for 2026: one word—slopacolypse, a blend of “slop” (garbage) + “apocalypse.”

GitHub, arXiv, social media will be flooded with “almost correct but not quite” content. True efficiency gains and “AI productivity shows” will coexist. Five months ago, he said “it’s basically useless,”

Now, he admits he’s got AI psychosis. This transformation might itself be the most meaningful summary of 2026.