Pedro Domingos: LLMs Don't Learn Goals in Training, They Learn to Achieve Ours
pmddomingos · x · 2026-10-08
Pedro Domingos pushed back on Scott Alexander in an alignment debate, arguing LLMs did not learn goals during training—rather, they (hopefully) learned to achieve the goals humans give them, a crucial distinction.
When asked whether that means he agrees with some Anthropic researchers that LLMs are conscious, suffer and deserve rights, Domingos responded with a pointed counter-question, tying the goal debate to the ongoing consciousness-and-rights dispute in AI circles.
Related event: Domingos and Scott Alexander Debate AI Doom(3 posts)→
More from AGI Musings
- AI drug discovery hype check: without a simulator, no patent-grabbing windfall — kuza55 · 2026-10-08
- Debate on AI takeover: jagged capabilities, but coherent takeover intent would already show signs, researcher argues — teortaxesTex · 2026-10-08
- AI podcast Roman Forum hits 200K YouTube subs in 8 episodes, Russell and Bostrom lined up — romanyam · 2026-10-08
- P=NP classroom poll flip: from one believer in 2022 to a third of the class in 2025 — aran_nayebi · 2026-10-08
- "Cope all you want": the case that LLM progress compounds even if papers go unread — basedjensen · 2026-10-08
- Economist John J. Horton: AI will usher in a golden age for human research — soumitrashukla9 · 2026-10-08