No case for AI consciousness: stateless transformers can't host a self, argues engineer
gerardsans · x · 2026-09-24
Gerard Sans argues that AI consciousness in current models is purely a philosophical question, not a technical one.
- Frontier models are transformers; inference is local and stateless—each output is independently conditioned on the input against a frozen corpus, with no shared state, so a self or world model cannot exist in this architecture.
- The sense of identity in chat is a curated illusion: first-person UI, a system-prompt persona, and RLHF stacked on pre-training. Change any piece and the speaker disappears.
- His bottom line: sampling is not a mind, a distribution is not a self, and any consciousness debate should start with a definition—which neither neuroscience nor philosophy has managed.
More from AGI Musings
- Yoav Goldberg: LLM Spotted the Pattern by Analogizing It to CRISPR — yoavgo · 2026-09-24
- Nina Schick: If you believe in AI safety, go long on AI compute — NinaDSchick · 2026-09-24
- After Astra: where does value go when frontier labs ship robotics brains? — mihdalal · 2026-09-24
- New note extends 'The Economics of Recursive Self-Improvement' paper — CFGeek · 2026-09-24
- NBER conference data: US dominated genAI last year, not today — professor_ajay · 2026-09-24
- Robotics' GPT-3 moment may just be GPT-6 itself, argues investor — mihdalal · 2026-09-24