This page may contain stale information. Last updated: 2026-06-12
Definition
Autoresearch is the practice of automating the ML research loop: propose an idea, implement it, run an experiment, validate the result, and use learnings to choose the next experiment — without continuous human intervention.
Benchmark Ecosystem
Popular autoresearch benchmarks from andrej-karpathy’s ecosystem:
- nanochat-autoresearch — fixed-budget training to lowest BPB
- nanogpt-speedrun — fastest training to target loss
- sol-execbench — GPU kernel optimization toward hardware limits
June 2026 Milestone
recursive achieved SOTA on all three benchmarks simultaneously and open-sourced artifacts (2026-06-11-recursive-automated-ai-research). The system outperformed the entire autoresearch@home community on NanoChat.