Autoresearch CPU fork runs on any hardware

Karpathy's autoresearch went viral for good reason — autonomous overnight ML research is a compelling idea — but locking it to H100s meant 99% of developers couldn't touch it. This fork fixes that.

• –The original hit 34.9k stars and required Flash Attention 3 on an H100; this fork swaps that out for standard PyTorch SDPA, opening it to consumer hardware
• –The added "Folding Mode" using a local Qwen 2.5 0.5b model is a clever low-resource twist — background research agents running during idle time
• –The reported val_bpb improvement (2.29 → 2.23) mirrors real gains users saw on the original, suggesting the CPU path isn't just a demo — it can produce meaningful results overnight
• –Multiple parallel forks for macOS MLX, Windows/RTX, and CPU show a coordinated community effort to democratize access — Karpathy himself signaled openness to linking them
• –The ~5 minute training loop on any hardware is the key unlock: developers can now iterate on autonomous ML experiments without cloud GPU budgets