# karpathy/autoresearch

Canonical: https://openpick.ing/r/autoresearch
As of: 2026-11-21 20:30 UTC


autoresearch is an open source tool in AI agents with 100,376 GitHub stars (+329 in the last 7 days), not maintained, as of 21 Nov 2026.

## Answer
A small LLM training setup on one GPU where an AI agent edits the code, trains for 5 minutes, and keeps better results.

## Takeaways
- The agent only edits train.py, keeping changes small and reviewable
- Each run has a fixed 5-minute budget, so results are comparable
- You program the agent through program.md instructions, not Python

## Use cases
- Running about 100 training experiments overnight without hands-on work
- Finding the best model for your GPU within a 5-minute budget
- Learning how AI agents can run simple research loops
- Testing changes to model architecture, optimizer, or batch size quickly

## Fit for
- People with a single NVIDIA GPU who want automated training experiments
- Researchers curious about agent-driven model iteration

## Not for
- Users without an NVIDIA GPU, since other platforms are unsupported
- Projects needing distributed training or complex configs

## Setup difficulty
medium

## Quick answers
### Is autoresearch still maintained?
Not maintained. 0 commits in the last 90 days.

### How popular is autoresearch?
100,376 GitHub stars. +329 in the last 7 days. #51 all-time. As of 21 Nov 2026.

## Stats
- Category: AI agents
- 100,376 stars, as of 2026-11-21 20:30 UTC
- +329 stars in 7 days
- License: No license listed (Check the license). Not legal advice. License: https://github.com/karpathy/autoresearch
- GitHub: https://github.com/karpathy/autoresearch
- Ranks: #51 all-time

## Health
Maintenance: Not maintained
Health: 26/100 (as of 2026-11-18)
- Issues: 55 open / 61 closed
- Pull requests: 140 open / 7 merged
- Commits in the last 90 days: 0
- Latest release: none (0 releases)
Stale: no recent pushes to the repository.

Cite as: autoresearch on OpenPick, https://openpick.ing/r/autoresearch, data as of 21 Nov 2026.

AI-generated content. Content updated 2026-11-17 11:31 UTC.

See Methodology: https://openpick.ing/methodology · Disclaimer: https://openpick.ing/disclaimer