Karpathy's Autoresearch Project: AI Now Conducts Its Own Research

Once upon a time, frontier AI research was conducted by meat computers. These "computers" would work between eating meals, sleeping, having other fun, and synchronizing occasionally through a ritual called "group meeting" using sound wave interconnect. That era is now behind us.

Andrej Karpathy's autoresearch project, launched in March 2026, represents the first step toward turning this vision into reality. Imagine this: you give an AI agent a small but real LLM training setup and let it experiment autonomously overnight. It modifies the code, trains for 5 minutes, checks if the results improved, keeps it if successful or discards it if not, and the cycle continues. You wake up in the morning to find a log of experiments and hopefully a better model waiting for you.

So How Does It Work?

The project is intentionally kept minimalist. There are only three files that really matter:

• prepare.py — Fixed constants, data preparation, and utility functions. This file is off limits.
• train.py — The only file the agent edits. Model architecture, optimizer, and training loop live here. The agent can play with everything.
• program.md — Baseline instructions for the agent. The file humans edit.

The system works like this: Each experiment runs with a fixed 5 minute time budget. This makes results comparable regardless of platform. The metric used is validation bits per byte (val_bpb). Lower is better.

Why Is This So Important?

Karpathy's project is not just a technical experiment, but a vision of the future of research. In traditional research processes, humans spend hours writing code, manually managing experiments, and analyzing results. Autoresearch automates this entire process.

And this system isn't just for big institutions. If you have a single NVIDIA GPU (preferably H100, but forks exist for smaller systems), you can run this experiment at home or in your office. Community forks already exist for MacOS, Windows, and AMD.

Tips for Smaller Computers

If you don't have an H100, don't worry. Karpathy has shared some recommendations for smaller systems:

1. Use less complex datasets like TinyStories.
2. Reduce vocab size (from 8192 down to 256).
3. Lower MAX_SEQ_LEN (even down to 256).
4. Reduce model DEPTH (from default 8 to 4).
5. Decrease TOTAL_BATCH_SIZE (to around 2^14).

Conclusion

Autoresearch offers an exciting perspective on the future of AI research. Perhaps in the near future, researchers will wake up to review the discoveries made by AI agents working through the night. Who knows, maybe when we reach the 10,205th generation of the code base, we'll look back nostalgically at the days when this project laid its foundations.

If you want to try this experiment yourself, check out the GitHub repo. Maybe the next big discovery will happen on your GPU in the early morning hours.