Karpathy's AutoResearch runs 700 experiments in 2 days

This is less self-improving AGI than agentic AutoML with tighter ergonomics, and that's exactly why it matters: once you can define a metric and a time box, the search loop becomes software.

• –700 experiments in 2 days is the point; Fortune reports Tobias Lutke tried the same loop on internal data overnight and got a 19% gain after 37 experiments. [Fortune](https://fortune.com/2026/03/17/andrej-karpathy-loop-autonomous-ai-agents-future/)
• –The repo is intentionally tiny: `program.md` sets the mission, `train.py` is the only editable code path, and each trial is capped at 5 minutes. [GitHub](https://github.com/karpathy/autoresearch)
• –That makes the pattern easy to port to other measurable workflows, from model tuning to data prep to CI-style optimization.
• –It is still narrow, not recursive intelligence, but the narrowness is the whole value proposition because teams can ship it today.