Voices
Karpathy's Autonomous AI Research Agent Ran 700 Experiments Over 48 Hours, Found 20 Optimizations
Karpathy shared results from a two-day continuous autonomous agent run: the agent conducted 700 experiments optimizing the training loop of a small language model and independently discovered 20 improvements that reduced training time. Fortune covered it as confirmation that AI-for-AI-research is crossing from demo to real utility. This is a material update on his earlier autoresearch work — up from 333 experiments overnight to 700 over 48 hours, with Fortune framing it as a 'why everyone is talking about this' moment.
Source
↳ Follow the thread