Karpathy autoresearch — Deep-Dive Reference
Karpathy's autoresearch (March 7, 2026) is a ~630-line single-file LLM training repo that hands a Claude Code agent the job of iterating on its own training code. The human edits program.md to describe research goals; the agent edits train.py to test hypotheses; every experiment runs for exactly five wall-clock minutes and reports valbpb; the agent keeps improvements and reverts regressions. This document covers the three-file architecture, the valbpb metric, the Muon optimizer, agent permission model, HPC/Slurm integration, Apple Silicon adaptation, and how autoresearch differs from conventional HPO frameworks.