Comment on Generative AI's failure to induce robust models of the worldComments−energy1231yWhy was Anthropic's interpretability work not discussed? Inconvenient for the conclusion?https://www.anthropic.com/news/tracing-thoughts-language-mod...−lossolo1yThe 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.
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.