I understood this purely as a pragmatic notion. LLM's produce some valid stuff and some invalid stuff. It would be useful to know which is which. If there's inf

I understood this purely as a pragmatic notion. LLM's produce some valid stuff and some invalid stuff. It would be useful to know which is which. If there's information inside the machine that we can extract, but isn't currently showing up in the output, it could be helpful. It's not really necessary to answer abstractions about truth and knowledge. Just being able to reject a known-false answer would be of value.