You said “that’s not how they work”. But that is how they work. Same prompt = same output. Throw some random data in there to jumble things around and you get a little variance. That’s the seed, and we only need to do that because LLMs are inherently deterministic.

Same reason Minecraft has a random seed for world generation, and block cipher algorithms use an initialization vector and/or feedback loop. We don’t want the same thing every time.

I did say that you’re right, because the tooling we use around the LLM itself does exactly what you’re talking about. So, in practice, you’re right.