[nearbuy]

An LLM absolutely doesn't respond the same way each time if asked the same question three times in a row, with temperature (randomness) set to zero. It responds the same way only if you start a new chat, which is a clean instance with no memory of the previous conversation. For a human, this is like if you went back in time to just before you asked the question, and asked them the same question again, in which case the person would give the same answer.

[abeyer]

> For a human, this is like if you went back in time to just before you asked the question, and asked them the same question again, in which case the person would give the same answer

Is it? Would they?

You seem to assert that there's no "temperature" in human behavior... which is a reasonable theory, but not one that's universally accepted nor likely to be provable.

[therealdrag0]

No I think they’re saying the temperature in human behavior comes from the “random” noise of inputs around us and ongoing history. But rewinding history and playing it back with the same temperature dice rolls is the only way to have the same thing a a LLM with no random inputs.

LLMs run in simulated environments where you can control randomness so you need the same for a human to compare the two. You can’t just ask a human a question multiple times as everything around them changes and conclude the human is behaving differently because they answer differently the same question. The question is not the bounds of relevant context; the entire operating environment is!

[juhanima]

And of course "temperature" is just an euphemism for the artificial randomness that is mixed in to make the output appear more magical.

[nearbuy]

The term "temperature" has been used in machine learning for a long time and came from using it as a parameter during training, analogous to physical temperature in https://en.wikipedia.org/wiki/Boltzmann_distribution.

But the relevant point is that we can reset the state of an LLM to its initial state before you asked it anything. This is a feature. You can choose to persist memory (through training, fine-tuning, databases, or context window), or you can choose to wipe it. If we could do the same for a human (eg, by going back in time), the person would behave the same way as the LLM. They wouldn't get annoyed that you asked the same question three times. They wouldn't know they've been asked before.