Karpathy Open Sources Autoresearch: Take a Nap, Let AI Run 100 LLM Training Automatically

According to 1M AI News Monitoring, Andrej Karpathy, founder of the AI education company Eureka Labs and co-founder of OpenAI, publicly shared the open-source project autoresearch yesterday, packaging the AI Agent's auto-tuning workflow previously used in the LLM training project nanochat into a standalone repository for developers to get hands-on experience over the weekend.

The core design of the project is "Humans write Markdown, AI writes code": humans write program.md to define the research direction, and the AI Agent autonomously modifies around 630 lines of train.py (including the full GPT model, Muon + AdamW optimizer, and training loop). Each experiment runs for a fixed 5 minutes, with the validation set byte per byte (val_bpb) as the sole metric. If it outperforms the baseline, the submission is retained; otherwise, it is discarded, and the process repeats. At this pace, approximately 12 experiments are run per hour, around 100 times in a full night. In the example chart presented by Karpathy, 83 experiments resulted in 15 retained improvements.

The project only requires a single NVIDIA GPU (tested on H100), relies on PyTorch and a few dependencies, and is open source under the MIT license. Community-driven macOS and MLX compatibility branches have already emerged.