haven’t thought about this before tho so could be way off base here

i actually think this is a probabilistic argument: after something so OOD all conditional probabilities are extremely small, so it winds up sampling something weird which then collapses the subsequent distribution onto some random part of the training corpus

agreed. mechanistically i would buy that the induction heads get overloaded but also its just so out of distribution that the model has to resort to digging for scraps

a circle in each plane? the matrices are in an o(2) x … x o(2) subgroup of o(n) right

its just a sequence-position-dependent rotation in latent space

/pol/ Pot