Why LLMs haven't automated more: RLHF rewards likable answers, not task completion
venturetwins · x · 2026-09-29
On the a16z podcast, @CompleteSkeptic argues LLMs are already smart enough to automate more human tasks — the bottleneck is that current models aren't built to function as reliable components inside software. RLHF rewards responses humans like, which doesn't necessarily lead to getting tasks done.
More from AGI Musings
- Many people mistakenly think AI is a database of scraped content — dreamwieber · 2026-09-29
- Commentator: OpenAI and Anthropic are both unkillable, their mog-countermog dance will outlast ASI — teortaxesTex · 2026-09-29
- What's your counterargument to "LLMs are just next-token predictors"? — thekokoricky · 2026-09-29
- Review articles are getting bland and AI-indistinguishable, scientist warns — NikoMcCarty · 2026-09-29
- "Agency Plus Curiosity Will See You Through the Intelligence Explosion," Argues Viral Post — StewartalsopIII · 2026-09-29
- Personal AI Agent Race Will Come Down to Meta, OpenAI and Google, Analyst Argues — RihardJarc · 2026-09-29