autoresearch (Karpathy's Overnight Research Agent)
activeOrigin: Andrej Karpathy, March 2026 (GitHub release)
An open-source AI agent that autonomously runs ML research experiments on a single GPU overnight. Key mechanic: `program.md` is a complete research methodology document; a coding agent reads it and executes it indefinitely — roughly 12 experiments/hour, ~100 overnight. Karpathy used it to speed up LLM training by 11%. Tobi Lütke ran 37 experiments overnight and got 19% performance gain. 66K+ GitHub stars within weeks of release.
Usage
The canonical example of 'AI as autonomous research assistant.' 'program.md' is now a pattern name (like WORKFLOW.md or SPEC.md) for agent instruction documents. Use to represent the most advanced demonstration of agentic engineering in 2026.
Mutations & variants
'program.md' — the research instruction file pattern
'Running autoresearch overnight' — shorthand for autonomous AI research
'100 experiments overnight' — the benchmark for AI-accelerated research
Cross-platform escapes
- GitHub: 66K+ stars
- Fortune, VentureBeat, Axios: covered
References
- GitHub: github.com/karpathy/autoresearch
Shelf life: Active — will remain a reference point for 2026 AI research automation
'Autoresearch' is both a product name and an emerging generic term (like 'vibe coding'). When using it, clarify which: Karpathy's specific tool, or the general concept. Shopify CEO test (37 runs, 19% gain) is the best 'CEO validates this' proof point available.