An LLM like ChatGPT ends up with a bunch of probabilities for the next word-ish. A random number generator then picks from them, weighted by probability, and then the process loops. So while the weights are static, the actual execution is very much non-deterministic.