Open-source auto-research agent cracks CUDA kernel challenge's global top 10
IgorCarron · x · 2026-10-30
RohanArun reports his open-source auto-research agent generated new CUDA tricks that pushed him 45 ranks up into the global top 10 of the Prefill kernel challenge in one day: Prefill Layer dropped from 213,101.86 µs to 198,110.25 µs (1.956×, #20→#8), with gains across MoE, Attention Decode, and Whole Decoder Layer (#120→#91).
The project is a native macOS app that drives Codex, Claude Code, or OpenCode CLI to work on open research problems with one click, recycling optimization tricks learned from competitions back into the agent's knowledge base.
More from coding & agent
- Skyworks cuts root cause analysis from 20 days to 10 with LangGraph agents — LangChain · 2026-10-30
- 73-entry list sorts agent infra: which MCP tools an agent can provision itself vs needs a human — Total-Reasonable · 2026-10-30
- YC-backed Turnstone turns your existing AI subscriptions into an agent team that auto-switches when limits run out — ycombinator · 2026-10-30
- The simple self-testing loop behind every RL data company, per practitioners — teodorio · 2026-10-30
- 7 AI models all faked completion reviewing the same 111k-line codebase — a postmortem — rseroter · 2026-10-30
- Supabase mistook thousands of new databases per minute for a DDoS — it was Lovable integrating silently — alifcoder · 2026-10-30