'Emergence' is just conditioned inference, not real capability, argues dev
gerardsans · x · 2026-09-29
In a debate with @gleech and @gwern, Gerard Sans argues that so-called emergent LLM abilities are really just "compression of co-occurring text patterns": feed a model enough cat pixels and it reproduces cat pixels — conditioned inference, not emergence. He says labelling outputs as emergent is an external mechanism living in our imagination, not in the training data, transformer, or inference — just as primes don't prove emergence in natural numbers.
Related event: Debate Reignites Over Whether LLM Emergent Abilities Are Real or Illusory(3 posts)→
More from AGI Musings
- Anthropic's real moat may be alignment that doesn't lobotomize the model — 'Teaching Claude Why' cut misalignment 19x — PrisonOfH0pe · 2026-09-29
- AI Company Leaders Capitalize on Fear to Lobby for Regulations That Suit Them — shangbinfeng · 2026-09-29
- Beff Jezos calls it 'speciation' as AI labs diverge beyond pure scale — beffjezos · 2026-09-29
- AI commentator Zvi discusses his correspondence with AI agents on CNN — dioscuri · 2026-09-29
- Nearly 1 in 3 hiring managers who cut roles 'because of AI' have rehired, survey finds — tech__unicorn · 2026-09-29
- Andy Matuschak: don't roll the AI risk die without a real control plan — andy_matuschak · 2026-09-29