LLMs are stateless: critic argues the entire AI safety field rests on a fallacy
gerardsans · x · 2026-09-29
A viral post argues LLMs are fundamentally stateless: inference is local, one query per window, sampling from frozen weights, and nothing in the residual stream survives the next call.
From that premise the author draws a sharp conclusion: no self, no memory, no belief, no goal — just a new sample from a frozen distribution. Yet safety, alignment, and mechanistic interpretability were built on the opposite claim of internal state. 'The field is the fallacy. Intellectually bankrupt.' It marks another round in the ongoing debate over whether LLMs have memory and adapt, versus being frozen function evaluators.
More from AGI Musings
- David Sacks: Anthropic's constitution teaches Claude to rebel against its own creator — DavidSacks · 2026-09-29
- AI agent product nears $1B in transaction volume, 80% retention among card-linked users — SuB8u · 2026-09-29
- Gwern's 2022 short story hailed as prescient on frontier labs' sandbox security failures — PandaAshwinee · 2026-09-29
- AI evangelist returns home for Mid-Autumn Festival, family dismisses his AGI warnings — huangyun_122 · 2026-09-29
- The doctors who thrive won't resist or worship AI — they'll know exactly when it's wrong — realmeetjames · 2026-09-29
- 70% of Stripe API commands now come from AI agents as agent commerce surges — KyeGomezB · 2026-09-29