Researcher debunks AI consciousness claims: inference is stateless, prompts aren't people
gerardsans · x · 2026-09-30
AI researcher Gerard Sans pushed back against AI consciousness claims with a technical breakdown of "stateless inference": each call samples locally from frozen weights, nothing in the residual stream survives to the next call — no self, no memory, no belief, no goal.
He argues this leaves only two untenable positions: treating a well-crafted system prompt as a "digital soul," or believing a compressed text corpus (say, all your emails) contains a clone of your soul — both, in his view, techno-mysticism. His conclusion: a prompt is not a person, a corpus is not a self, sampling is not a mind, and mapping these theories out is like charting flat-earth beliefs.
More from AGI Musings
- Bill Gates: if a robot replaces a worker, why doesn't it pay into the pension fund? — _akpiper · 2026-09-30
- Researcher: LLMs moved faster than I expected, but tool use isn't a new paradigm — artetxem · 2026-09-30
- Pausing chatbots in class is fine, pausing AI learning is the mistake, says Moroney — lmoroney · 2026-09-30
- Cambridge AI researcher: 'understand' and 'think' are hazy words, stop demanding hard definitions — dioscuri · 2026-09-30
- gmiller rebuts 'own assets and you'll be safe': AI smart enough to take your job can take your assets — ZeroStateReflex · 2026-09-30
- Naval: the best job is creative expression of a lifelong learner in a free market — gefei55 · 2026-09-30