Yes. One point worth making though is that pre-training trains the model to produce high probability completions - in practice, most of what it's trained on is a good completion as a human would judge it, i.e. it's not factually false data, so you get "creates a bias _towards_ "correct answers" with each inference step", but the model isn't really learning what correct or incorrect means in any way a human would consciously think about it.
replydoes that lead to much inside the box thinking by models? Do they try to follow the clearest path?
replyIt depends on what you mean by “inside the box thinking”. In one sense, you are always mostly sampling high probability completions and they are going to be close to things that have been observed during training. Low probability completions will likely be low quality too. But, in a case where the context is forcing the model to generalize, I.e. interpolate in the latent space, the most probable completion may appear novel.
reply