Researchers gave Claude Code access to 16 GPUs for Karpathy's autoresearch project. Over 8 hours it ran 910 experiments, discovering that parallel search enables factorial grids instead of greedy hill-climbing, and the agent independently developed a strategy to exploit hardware differences between H100 and H200 GPUs.
Products
Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster
Claude Code autonomously ran 910 experiments across a 16-GPU cluster in 8 hours, discovering that parallelism enables factorial grid exploration and independently optimizing for H100 vs H200 hardware differences.
Thursday, March 19, 2026 12:00 PM UTC2 MIN READSOURCE: Hacker NewsBY sys://pipeline
Tags
products