AI agents running research on single-GPU nanochat training automatically
About autoresearch
autoresearch is a repository from Andrej Karpathy that gives an AI agent a small language model training setup and lets it run experiments on its own, typically overnight. The training code is a simplified single-GPU version of nanochat.
Each experiment follows the same loop: the agent edits train.py, trains for a fixed five minutes, evaluates and keeps or discards the change. That gives about twelve experiments an hour. The human writes instructions in program.md, while the agent only changes train.py, which keeps diffs easy to review. It is tested on NVIDIA H100 GPUs, with community forks for macOS, Windows and AMD.
One source says Karpathy followed it with agenthub, a platform for agent swarms working on one codebase. Star counts differ widely between snapshots, so check the repository.
Key features
- Agent runs training experiments unattended
- Fixed five-minute runs
- Agent edits only train.py
- Instructions in program.md
- Single NVIDIA GPU
- Community forks for macOS, Windows and AMD
Links
All content and agents from autoresearch
Frequently asked questions about autoresearch
- What is autoresearch?
- AI agents running research on single-GPU nanochat training automatically. It is listed on botshub under other agents.
- Is autoresearch free or open source?
- autoresearch is tagged as open source on botshub. Check the official site for current terms.
- What can autoresearch do?
- Key features include agent runs training experiments unattended, fixed five-minute runs, agent edits only train.py, instructions in program.md and single NVIDIA GPU.
- What are alternatives to autoresearch?
- Other other agents on botshub include data-to-paper, OpenResearch, Awesome Embodied AI Safety and GPT Researcher. See all alternatives to autoresearch.
