Formalizing PTX in Lean to formally verify GPU kernels written by AI agents
gharik · x · 2026-11-04
BenKoska announced a formalization of NVIDIA's PTX instruction set in Lean, enabling formal correctness proofs of GPU kernels. The motivation: as agents increasingly write kernels no human will read line by line, machine-checkable proofs offer confidence the generated code is correct without manual review. A thread with details follows.
More from Infra
- 15-year-old's attention kernel survives 32k tokens where standard PyTorch OOMs, 11x faster — vankownlaw · 2026-11-04
- Five local models on DGX Spark: Qwen3.8-27B most reliable at 28/30, fastest models finish last — JeremyCMorgan · 2026-11-04
- Prediction: every hospital runs its own local AI model by 2028, small models get there first — MaziyarPanahi · 2026-11-04
- LangSmith BYOC goes GA on AWS, hits Beta on Azure and Google Cloud — LangChain · 2026-11-04
- SK Hynix offers $45B US chip plant in exchange for license to upgrade its China fab — demian_ai · 2026-11-04
- AI is reviving magnetic tape: Fujifilm spends $63M to boost tape output 5x — demian_ai · 2026-11-04