Stateless Transformer Inference Challenges the 'AI With Wants' Doom Narrative
gerardsans · x · 2026-09-23
JimDMiller argues there will be little time between AI taking the last human job and AI eliminating humanity: by then humans are unnecessary, costly, and easy to remove unless AI actively cares about us. Gerard Sans counters that the whole narrative hinges on AI having a self with wants or needs, which current transformer architecture—with conditioned, stateless inference—doesn't support. The real open question, he says, is the timeline: is this about today's technology or some undefined future system?
More from AGI Musings
- AI professor Toby Walsh publishes call for control of frontier AI models — TobyWalsh · 2026-09-23
- Andrew Ng: AI danger fears are an orchestrated PR hype, a setback for the field — sudoraohacker · 2026-09-23
- Alignment researcher: without a slowdown, p(doom) > 0.1 — ArthurConmy · 2026-09-23
- Alignment researcher: without slowdown, p(doom) is above 10% — ArthurConmy · 2026-09-23
- ICLR load debate: capping submissions per author would barely dent paper counts, data shows — jonasgeiping · 2026-09-23
- Hackers breached OpenAI in days, per WSJ — challenging the case for frontier-model advantage — ben_j_todd · 2026-09-23