#nanochat

Andrej Karpathy says his autoresearch setup reduced nanochat's Time to GPT-2 from 2.02 hours to 1.80 hours. He said the agent explored roughly 700 changes over about two days and found around 20 additive improvements, but the result should still be read as a source claim rather than an independently audited benchmark.

Andrej Karpathy has published autoresearch, a minimal repo that lets AI agents iterate on a stripped-down nanochat training loop overnight. The project turns agent evaluation into a closed-loop research workflow with fixed 5-minute runs, Git branches, and validation-loss-based selection.