Comment by jyounker

One of my current models for LLMs is that they're compression algorithms. They compress a large amount of training data into a set of weights. A query is a key into that compression space. Hallucinations happen when you supply a key that corresponds to something that wasn't in the training set.