LLMs are a stochastic tool that generalizes patterns imprinted into the neural network during training. And then these patterns are expressed in the given context during inference. The stochastic aspect is what makes it possible for the LLM to adopt generalized training to the specific problem being solved. So, the outputs LLMs produced are not random in a sense of playing bingo because randomness is not the dominant property of the system. It is predicting the most likely next token within the given context.

I think this is a gross misunderstanding of the underlying tech. The probability distributions are fixed at training time. That is why the weights are static. When doing inference, models are selecting from a top-k of probable tokens at each autoregressive loop for your typical LLM. When you adjust top-k to 1 or simply set the “temperature” to 0, it will always give you the same result from the same input.

Something is deterministic if given the same conditions and inputs it produces the same output.

If you give an LLM the same parameters, seed, and prompt, it will produce the same output.

That’s certainly deterministic. As deterministic as any other procedural generation of content.