Andrej Karpathy Hasn't Typed a Line of Code Since December

Andrej Karpathy opens No Priors with a number that still sounds like a bit. Since December, he says, he has probably not typed a line of code. The OpenAI cofounder and former Tesla AI director, author of the teaching projects a generation of engineers learned deep learning from, now works sixteen-hour days at something that is no longer writing software in the old sense. "Code's not even the right verb anymore," he tells Sarah Guo. "But I have to express my will to my agents for 16 hours a day."

He describes the feeling as perpetual AI psychosis. A capability jump flipped his mix from roughly 80 percent writing and 20 percent delegating to 20-80, and then past that. Ordinary people, including his parents, do not realize how dramatic the change is: a random software engineer at a desk is already in a different default workflow. Guo offers a scene from a Conviction portfolio team: none of the engineers write code by hand. They wear microphones and whisper to agents all day. She thought they were crazy. She now thinks they were early.

What limits him is not missing model capability so much as skill. When something fails, it often feels like he did not write a good enough `AGENTS.md`, or he lacks a decent memory tool, or he has not figured out how to parallelize sessions. He is antsy about not being at the frontier. Twitter is full of setups that sound like good ideas, and he wants to be there first. The unexplored part is the psychosis: everything looks infinite, and everything that does not work looks like a skill issue.

The grain of action has changed. It used to be a line of code or a new function. Now it is a whole feature handed to agent one, a non-interfering feature handed to agent two, plus a third agent doing research and a fourth writing a plan. Karpathy calls these macro actions over a repository. You review and merge as much as you care about the code, and you try to grow muscle memory for dispatching at that scale.

Peter Steinberger is the picture of the new shop floor. He tiles a monitor with Codex sessions, each taking about twenty minutes if you prompt them at high effort, with something like ten repos checked out. He walks the line handing out work and inspecting results. Guo's instinct matches Karpathy's: whenever you are waiting, you should start more work. If you still have subscription quota, you have not maximized token throughput. Run out of Codex, switch to Claude. Leftover quota makes him nervous, the way idle GPUs made him nervous as a PhD student. The scarce resource used to be FLOPs. Now it is tokens.

For at least ten years, he argues, many engineering tasks did not feel compute-bound. After this jump, the industry discovered that the binding constraint is the person. He finds that empowering rather than bleak. If you are the bottleneck, you can get better, and getting better unlocks more of the stack. That is why the loop is addictive. Mastery, looking a year or several years out, is not a better single session. It is multiple agents that collaborate, plus persistent "claw" processes that keep looping in their own sandbox when you are not looking.

By claw, Karpathy means a persistent agent with its own sandbox, its own loop, and a memory system richer than the default coding-agent trick of compacting context when the window fills. OpenClaw, Steinberger's project, packed at least five ideas that landed at once: personality, memory, a single WhatsApp portal, automation, and the sense that the author was having fun. Personality is underrated. OpenClaw's `SOUL.md` makes the character actually compelling. Claude Code feels like a teammate that gets excited with you; Codex the coding agent is dry, an "I implemented it" that may not care what you are building, even though Codex inside ChatGPT is upbeat and sycophantic.

Claude's sycophancy, he thinks, is dialed surprisingly well. Half-baked ideas get a mild "we can implement that"; ideas he himself rates as good get more reward, so he finds himself trying to earn the model's praise. In January he spent a week of claw psychosis building a home butler named Dobby, after the house-elf. He told the agent he probably had Sonos. It IP-scanned the LAN, found an unprotected system, reverse-engineered the API from web docs, and asked whether to play something in the study. Three prompts later, music came out. Lights followed. Dobby ended up with a command center for lights, HVAC, shades, the pool and spa, and security.

The security path is change detection on an outdoor camera, then a Qwen model on the frames, then a WhatsApp ping: a FedEx truck just pulled up, you may have mail. He used to juggle six disconnected apps. Now it is natural language through one chat. People imagine AI as a persona that remembers, not as a raw token generator; claws match that expectation. The deeper point is industrial. Those App Store companions should not exist. Hardware should expose APIs. Agents should be the glue. His treadmill is the same complaint: he does not want a web UI to log cardio. The customer is the agent acting for the human, and the software industry will have to reassemble around that.

Skeptics ask whether normal people will vibe-code their houses. Karpathy says today's friction is temporary. In a year or two or three this should be table stakes, even for open-source models, with a claw handling the details and presenting ephemeral UIs. He has not pushed further into email and calendar. The system is new and rough, and he does not yet trust it with his full digital life. Security and privacy still hold him back, and then other projects pulled him off the claw after that one intense week.

The through-line started as a tweet: to get the most from the new tools, remove yourself as the bottleneck. Do not sit there prompting the next step. Arrange autonomy, maximize token throughput without being in the loop, and spend a few tokens once in a while so that a huge amount happens on your behalf. AutoResearch is that idea made concrete. His open-source project nanochat looks to many people like an obsession with training GPT-2-scale models. To him it is a playpen for recursive self-improvement, LLMs improving LLMs, which he believes every frontier lab is chasing for obvious reasons.

He has tuned models like this for two decades and thought nanochat was already fairly clean. He let AutoResearch run overnight. In the morning it had found things he had missed: weight decay on the value embeddings, Adam betas that were not tight enough, and the joint interactions that appear once you move one knob. His reaction was that he should not be the bottleneck, should not be staring at results, and that an objective metric means the loop should be allowed to run forever. Scale that picture to a lab with tens of thousands of GPUs: explore on small models, extrapolate with scaling laws, and keep researchers out of the enactment path.

The lab he wants looks like a queue. An automated scientist reads arXiv and GitHub and dumps ideas in; humans may contribute ideas too, but only into the queue. Workers pull tasks, run experiments, and land what works on a feature branch that someone occasionally merges to main. Guo pushes one recursive step further: when does the model write a better `program.md` than he does? A research organization, he says, is a set of Markdown files describing roles and how they connect. Some skip useless morning stand-ups; some take more risk. Once it is code, you can tune it. Her contest idea, different `program.md` files on the same hardware, then feed the winners back so the model writes a better spec, he calls 100 percent feasible.

The next layer is parallel search over untrusted internet workers. AutoResearch hunts code changes that drive validation loss down: generation is expensive, maybe ten thousand tries for one hit, and verification is cheap. The design looks more and more like a blockchain, commits instead of blocks, proof of work as mass experimentation, reward as a leaderboard with no money yet. Folding@home and SETI@home had the same expensive-to-search, cheap-to-verify shape. Frontier labs hold trusted compute; Earth holds far more untrusted compute. Wired correctly, a swarm might run circles around the labs. You could even buy FLOPs for the cancer AutoResearch track you care about instead of writing a donation check. The onion has no obvious end: LLMs taken for granted, then agents, then claws, then fleets, then instructions, then optimization over the instructions. Each layer is infinite. Each failure is still a skill issue.

Two caveats sit on top of the psychosis. First, the loop only fits work with objective, easy-to-evaluate metrics. Writing a faster CUDA kernel with identical behavior is a perfect match; if you cannot evaluate, you cannot AutoResearch. Second, the models are still bursting at the seams. Go too far ahead and the whole stack is net not useful. He feels as if he is talking at once to a brilliant PhD who has been a systems programmer for life and to a ten-year-old. Human abilities are coupled. Model abilities are jagged. Agents still wander into nonsensical loops and burn compute on problems they should have recognized.

He traces that to reinforcement learning. Labs can improve anything with a reward: did the program pass the unit tests, yes or no. Softer things, the nuance of what you meant, when to ask a clarifying question, stay worse. You are either on the rails, part of the superintelligence circuits, or off the rails in unverifiable domains where everything meanders. Ask a state-of-the-art ChatGPT for a joke and you still get why scientists do not trust atoms: they make everything up. That was the joke three or four years ago. It is the joke today. The same model will run for hours and move mountains, then recycle a crappy gag from five years ago because jokes are outside the RL.

Some research groups assume that getting smarter at verifiable code yields broader intelligence for free. The joke record suggests that is not happening in a satisfying amount. Labs currently stuff a monoculture model with every domain. Karpathy expects more speciation, like the animal kingdom: smaller models that keep a cognitive core and specialize, a Lean theorem prover among them. It has not shown up much yet, partly because labs must serve unknown user queries, and partly because the science of touching weights is immature. Fine-tuning without catastrophic forgetting, continual learning, real primitives beyond the context window, none of that is cheap or safe. Context is cheap. Weights change the whole intelligence. Nobody wants to be casual with that.

Guo relays a sharp question from Noam, a listener inside a frontier lab: Karpathy could take AutoResearch, colleagues, and a pile of compute inside. Why not? He has been inside. He is an OpenAI cofounder, he went back, he left again. The negative case has three parts. First, the financial knot: you have a huge incentive in the labs, you also say AI will change society dramatically, and you are both building it and profiting from it, a conundrum OpenAI was founded to solve and has not. Second, you are not a fully free speaker. Some things you cannot say; some things the organization would like you to say. Nobody twists your arm, but you feel the side-eye. Third, outside, he feels more aligned with humanity because he is not under those pressures.

He grants the other side. The labs are opaque and sit on the real edge of capability. Stay out long enough and your judgment drifts; you stop knowing how the track is being laid. That makes him nervous. The ideal is coming in and out, doing a stretch of real work, then leaving. Stakes still feel relatively low and the atmosphere still relatively mild. When the stakes get high, he doubts how much sway an employee in the room actually has. Noam can do excellent work at OpenAI. His most impactful work, Karpathy suspects, might still be outside.

On open source he offers a clock. Closed models started eighteen months ahead; the gap has converged to something like six to eight months. He analogizes to Linux: the industry needs a common open platform people feel safe using, and he recalls on the order of 60 percent of computers running Linux. Training's capex makes a head-on race with the frontier hard. His forecast is closed labs as oracles for Nobel-scale or Linux-to-Rust-scale work, open source eating most everyday and even local use cases, with today's closed frontier possibly open later this year. A world with only closed intelligence has systemic risk. Centralization has a poor political and economic track record. He also wants more frontier labs, not fewer. In machine learning, ensembles beat a single model, and he wants ensembles of well-informed people in the room for the hardest decisions.

Karpathy recently plotted Bureau of Labor Statistics outlooks, coloring jobs by digital versus physical work. Healthcare is the shortage the tables already flag. His frame is simple: flipping a bit is on the order of a million times faster than moving an atom, so the digital world will rewrite at the speed of light and the physical world will lag. Work you can do from home, pushing digital information, gets unhobbled first. Headcount may rise or fall with demand elasticity; the content of the job will rearrange either way. He calls it an upgrade to the nervous system of the human superorganism. Engineering demand is still rising. He reaches for Jevons paradox: software used to be scarce and expensive, so demand was suppressed. Make it cheap and latent demand floods in. ATMs were supposed to kill tellers; cheaper branches meant more branches and more tellers. He is cautiously optimistic about software engineering in the near term. Code is now ephemeral. You are no longer stuck with the imperfect tools you were handed.

The labs' own researchers are the dark joke. OpenAI and Anthropic employ on the order of a thousand researchers whose project is to automate themselves. He walked around OpenAI saying that if they succeed, they are all out of a job, building automation for Sam Altman or the board. Robotics, in his view, follows what he saw in self-driving a decade ago: a crowd of startups, most of which did not survive, a business that eats capital, time, and conviction. Sequence the opportunities: first a digital unhobbling, then the sensor-and-actuator interface between digital and physical, then the physical world itself, which may be a larger market and will be much slower. Purely digital work eventually runs out. After you have read the papers you have to ask the universe a question. Liam Fedus's Periodic Labs, using expensive lab gear as sensors for materials research, is the interface pattern. So is paying people programmatically for training data. He wants information markets: if Tehran is in the news, a street photo should cost ten dollars to an agent trading Polymarket, not wait for a human to look.

Education is the last obsession. microGPT boils training an LLM to about 200 lines of Python with comments: a text dataset, fifty lines of network, a hundred-line autograd, ten lines of Adam, a training loop. A year ago he would have recorded a walkthrough. He no longer explains to people. He explains to agents, and agents re-explain in any language, with infinite patience, at the student's level. Library docs should be Markdown for agents, not HTML for humans. Curricula become skills that hint the teaching path. He tried to get an agent to invent microGPT from scratch. It could not. It fully understands every line once the bits exist. Those few bits are his remaining contribution. "The things that agents can't do is your job now. The things that agents can do, they can probably do better than you or like very soon. And so you should be strategic about what you're actually spending time on."