srush's blog verifies floating-point certificates for 152 ML operators, finds unreported bugs
ChrSzegedy · x · 2026-11-15
- Serge (srush) published blog #2, "When Rounding Breaks Machine Learning," producing floating-point correctness certificates for 152 operators to verify the whole ML stack.
- The approach surfaced issues across core ML libraries, several of which had never been reported before.
- The takeaway: rounding can silently break ML computation, and the entire stack deserves formal verification.
More from Infra
- AGI would still face a world running on decades-old infrastructure no one dares to touch — SuB8u · 2026-11-15
- SpaceX cuts AI datacenter build time to ~4 months, MiniHard done in 5 — Dr_Singularity · 2026-11-15
- 33B MiniMax-H3-Image image model runs fully on a Snapdragon phone NPU, no cloud — shirosoro · 2026-11-15
- LMDeck: a macOS menu-bar app to unload local LLMs across backends in one click — Fickle_Tradition4491 · 2026-11-15
- Kosmo and Ropedia release Physical AI dataset: 1,589 3D-scanned environments, 131 hours of egocentric video — liuziwei7 · 2026-11-15
- Retired Microsoft engineer explains what broke in the Outlook outage — nicolascraske · 2026-11-15