Comment by cluckindan

This is it. They’re language models which predict next tokens probabilistically and a sampler picks one according to the desired ”temperature”. Any generalization outside their data set is an artifact of random sampling: happenstance and circumstance, not genuine substance.