zlacker

This paper and other similar works changed my opinion on that quite a bit. It shows that to perform text prediction, LLMs build complex internal models.

>>38893456