Autoresearch fork lands on Modal H100s
A fork of Karpathy's autoresearch ports the autonomous 5-minute training loop to Modal's serverless H100s, so experiments can run without a local GPU or CUDA setup. The author says each run costs about $0.32, cold starts are around 2 seconds, and data persists in Modal volumes.
This is less about a new training idea than about making autonomous research loops cheap, repeatable, and easy enough to run overnight. Once GPU provisioning disappears, the bottleneck shifts to prompt quality, the experiment loop, and the metric, while the tiny mutable surface, fixed 5-minute budget, persistent Modal volumes, and reported $0.32 runs with ~2-second cold starts make the workflow practical.
DISCOVERED
21d ago
2026-03-22
PUBLISHED
21d ago
2026-03-22
RELEVANCE
AUTHOR
Ready-Interest-1024