If you find this interesting, you should look into Solomonoff induction. It combines Kolmogorov complexity with Bayes rule to provide a general framework for inductive inference, and naturally formalizes Occam's razor.
I'm not sure the cause was really similar. In the case of language switching, it was caused by malformed supervised training data where the prompt was translated, but the answer was kept in the original language. In the…
> I find the fact that this only looks at the activations of some specific layer l a bit interesting. Some layer l might 'think' a certain way about some input, while another later layer might have different 'thoughts'…
If you find this interesting, you should look into Solomonoff induction. It combines Kolmogorov complexity with Bayes rule to provide a general framework for inductive inference, and naturally formalizes Occam's razor.
I'm not sure the cause was really similar. In the case of language switching, it was caused by malformed supervised training data where the prompt was translated, but the answer was kept in the original language. In the…
> I find the fact that this only looks at the activations of some specific layer l a bit interesting. Some layer l might 'think' a certain way about some input, while another later layer might have different 'thoughts'…