AIs far worse at fiction than code hints gradient descent generalizes policies badly
ZeroStateReflex · x · 2026-11-10
allTheYud observes AIs are much worse at fiction than code—constantly losing track of which facts depend on which, and resolving conflicts the wrong way. His sharper point: we may only still be safe because gradient descent on neural networks is such a bad way to generalize over policies.
More from AGI Musings
- JD Pressman: rigorously reverse-engineering fun might have changed AI debates — jd_pressman · 2026-11-11
- AGI governance is a children's fantasy, argues researcher JD Pressman — jd_pressman · 2026-11-11
- Thread: Confucianism and Kant were governance theories first — solved governance is a new problem — jd_pressman · 2026-11-11
- Speculative thread: technology may translate imperceivable math structures into new senses — ShanRizvi · 2026-11-11
- OpenAI and Upwork study freelance job postings as early signals of AI changing work — Exp_Mark · 2026-11-11
- As AI grows more powerful, researchers reflect on agency, happiness and meaning — prajdabre · 2026-11-11