Andrej Karpathy's Autoresearch framework hits 70k GitHub stars as AI begins automating its own research code

Andrej Karpathy published Autoresearch on GitHub in recent weeks, and the project has accumulated over 70,000 stars. It is a framework that guides an AI agent to automatically improve research code by making changes and testing them against quality criteria, keeping only improvements that raise metric scores.

Unlike human code review, Autoresearch performs continuous local search (hill climbing) — trying hundreds of candidate edits, measuring each against fixed benchmarks, and building on wins. Related projects like Sakana's AI Scientist and AlphaEvolve use similar strategies to search for better algorithms and entire scientific ideas at once, not just code refinements.

For ML researchers and mathematicians, this automates the trial-and-error cycle that occupies weeks of manual work. For philosophers and research communities, it raises a hard question: AI agents now write papers and improve algorithms without understanding what scientific model they are actually building — a gap philosophers of science have spent decades mapping that AI teams often overlook.