autoresearch

Andrej Karpathy’s research-loop repo. Hit 50,000 GitHub stars in a single week at release — one of the fastest growth curves the wiki has tracked for an OSS agent project. The pattern itself is probably more valuable than the tool.

What it does

Give it a goal (“make this model better”, “find a higher-converting landing page”, “lower customer acquisition cost”) and it runs an experiment loop overnight. By morning you wake up to a tested set of improvements, with the loop having only kept the winners. Conceptually a “robot research intern that runs hundreds of experiments and only keeps the winners.”

Type and platform

  • Type: repo (research loop pattern)
  • Platforms: model-agnostic
  • Status: tracked

Why it matters in the wiki

This is the closest thing the wiki has to a primary-source example of autoresearch-evals in practice. Sibling concept to loop and the Ralph loop community pattern (still deferred for a primary source). The user has flagged the pattern as more valuable than the tool itself — worth studying the loop structure even without running the repo directly.

Hands-on notes

(empty)

Sources

See Also