@eshear ultimately I think my sticking point is there is an unstated assumption here that LLMs are mesa-optimizers and are in pursuit of a goal (specifically, their pretraining objective), no part of which has been remotely demonstrated.