comi

Having fun with @karpathy’s autoresearch. I told Claude Code: “You’re the chief scientist of an AI lab with 8 GPUs. You’re Andrej Karpathy. Run parallel experiments and decide what to try next.” It edited program.md, ran for 11+ hours, and completed 568 experiments. Each experiment uses 1 GPU. Every round the “chief scientist” reviews the previous round of 8 results and designs the next 8 experiments. It's interesting to see the Claude agent, the chief scientist evolved a 3-phase strategy: Phase 1. Broad Exploration Early rounds explore many axes: architecture, optimizer, LRs, ablations. Phase 2. Focused Refinement After easy wins dry up, it runs deeper sweeps (e.g. 5 GPUs sweeping RoPE base 30k → 500k in one round). Phase 3. Heavy Validation Later, 50–75% of GPU budget goes to seed variance checks instead of new ideas. I feel it's overkill tbh. I'll keep the chief scientist running to see if it transfers to larger models and beats Andrej’s new "Time to GPT-2" leaderboard winner.