Sasha Rush: numerically verifying core ML code is easier and cheaper than you think
srush_nlp · x · 2026-11-14
Sasha Rush (srush) says that unlike his past formal verification blogs, this direction feels both useful and doable: getting most core ML code numerically verified isn't that hard, and is certainly easier and cheaper than continual learning — pointing to an underrated path for ML engineering reliability.
More from Research
- CMU Launches New Center on Mathematical Foundations of AI Safety — aran_nayebi · 2026-11-14
- 'Weight decay is a psyop' meme gets the perfect reply: Weight Agnostic Neural Networks paper — pbaylies · 2026-11-14
- MIT Press 496-page ML theory textbook 'Learning Theory from First Principles' is free — burny_tech · 2026-11-14
- Gradient vs Jacobian vs Hessian: one chart connects calculus to ML — burny_tech · 2026-11-14
- Cohere Labs talk: Harness² recursively improves agent harnesses at test time — Cohere · 2026-11-14
- Encoder-free multimodal LLMs re-learn encoder-like traits in early layers — rdesh26 · 2026-11-14