Researcher argues human alignment theory could be imported to align LLMs
willcb · x · 2026-10-17
willcb argues we should compare LLMs to humans holistically instead of dismissing the premise: current LLMs differ mainly in speed, cost, breadth, and error rate — not some unbounded "intelligence." Doing so would let us import a large body of inter-human alignment theory and technology to LLM alignment, alongside (not instead of) mech interp.
Related event: Researcher argues LLMs lack "1000 IQ," urges porting human alignment theory(2 posts)→
More from AGI Musings
- Dev got 10x easier, marketing 100x harder: founders clash over AI's net effect — MarcJSchmidt · 2026-10-17
- "In a decade most deeptech will look like a grift" — and the sad agreement it draws — sachdh · 2026-10-17
- Open-source maintainer asks: what's the incentive to release anything open-source now? — MarcJSchmidt · 2026-10-17
- EU border system was down 3 of 6 trips — researcher says don't bet on sovereign tech — Afinetheorem · 2026-10-17
- Dev observation: models excel with a reference implementation, but struggle at designing something truly new — lucasmeijer · 2026-10-17
- Dev laments: too many people tinkering with custom agent setups, too few actually shipping — intellectronica · 2026-10-17