Model trained without natural language writes kernels passing 28% of tests, up to 66x faster than Triton
cephaloform · x · 2026-09-27
BenKoska reports a surprising result: a model that was never trained on natural language or natural-language reasoning could still produce GPU kernels that passed all tests 28% of the time, achieving speedups of up to 66x compared to Triton kernels. "Turns out, it actually kind of works?"
More from coding & agent
- Why AI agents may not go mainstream this decade: most lives aren't complex enough — menhguin · 2026-09-27
- A practical AGI test: replace one human validating 150,000 document conversions — burkov · 2026-09-27
- Talk on self-improving agents with reflective optimization: GEPA, gskill, fast-slow training — CShorten30 · 2026-09-27
- Microsoft president Jeff Teper shares how he uses Autopilot for work routines — heyneighbor · 2026-09-27
- Same prompt weekly for a quarter: how a prod agent's policy compliance drifted until it broke — IsomuraArganee_95 · 2026-09-27
- Diplomacy comes to Multi-Agent Arena: test your social strategy against frontier LLM agents — ycombinator · 2026-09-27