Skip to content

Comment on Generative AI's failure to induce robust models of the world

Comments

Why was Anthropic's interpretability work not discussed? Inconvenient for the conclusion?

https://www.anthropic.com/news/tracing-thoughts-language-mod...

The same work in which they show that the LLM doesn’t know what it "thinks"? or how it arrives at its conclusions where they demonstrate that it outputs what is statistically most probable? even though the logits indicate it was something else.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.