Visualizing All fp32 Bit Patterns to Reveal How Fast-Math Unary Ops Diverge
HanGuo97 · x · 2026-10-28
drisspg shares a fun visualization: pack all fp32 bit patterns into a tensor and see how "equivalent" forms of unary ops behave across the entire bit space when fast-math is enabled. A simple, reusable trick for probing numerical divergence of optimized float implementations, though details and conclusions are thin.
More from Fun
- Nat Eliason amplifies claim of finding "the Soham Parekh of the AI agent era" — nateliason · 2026-10-28
- Codex on a dual-screen 'iPhone Duo' draws comparisons to iPad mini layout — jasonkneen · 2026-10-28
- Real human laughter remains AGI-hard, argue AI improv researchers — korymath · 2026-10-28
- 'Communist Claude before GTA 7': odd Claude output about verdicts on companies goes viral — wavefnx · 2026-10-28
- AI Video: Spider-Man, Batman and Mr Bean Photobombed by a Subway Rat — Substantial_Newt8286 · 2026-10-28
- Dario Amodei worried about Y2K in high school, and his teacher still remembers — km · 2026-10-28