It reasons over the semantic network between tokens, in a feedforward inference pass over the 2k(ish) words or tokens of the prompt. Sometimes that reasoning is

It reasons over the semantic network between tokens, in a feedforward inference pass over the 2k(ish) words or tokens of the prompt. Sometimes that reasoning is superficial and amounts to probabilistic linear relationships, but it can go deeply abstract depending on training material, runtime/inference parameters, and context of the prompt.