Theoretical interpretability question.

Consider two models, A and B.

• Model A is 100% interpretable. It can predict what Bob says and what he thinks.^[1] But sometimes it makes random-looking mistakes ("noise"). You care about predicting the noise, but you don't care about understanding the mechanism behind the noise. Ignoring the noise, model A is pretty perfect, capturing something like the platonic ideal of Bob, the necessary and sufficient conditions for thinking/talking like Bob.
• Model B is not interpretable. It predicts what Bob says and what he thinks, but it can't tell us what he thinks. Unlike A, it never makes mistakes - it does predict the noise.

Question: in this setup, can you modify A to predict everything correctly? (e.g. by locating the noise predictor inside B and combining it with A) What if you had a hypercomputer or an unphysically large computer?

If the answer is "yes", then it solves ELK in situations where we have a complete enough understanding of the thing we want to elicit.

Knowing people's opinions about what the answer is would be really useful for my research.

1. ^{^}

   This example is inspired by Paul Christiano. See "inaccessible information".