[justinclift]

> LLMs are non-deterministic; that is each time you run something you might get a different answer.

That's a choice. With (at least) Ollama, you can set the temperature parameter so the results will be deterministic based upon the seed value. Which you can set.

It's used for things like running test cases, but there are likely other uses too.

[daniel-payne]

You are correct about temp and seed, but for some models it still does not guarantee identical answers. Only models explicitly designed for deterministic output can do that, and I am not sure the commercially available models are built like that. I might be wrong.

The thing I was trying to say (not very well) was AI is going to make mistakes.

When we had systems based on relational databases, it was possible to recreate the conditions where the mistake was made. But in a world of LLMs and rapidly changing graph databases, this is almost impossible. So we should be recording everything so we don't make the same mistake twice.

Someone said to me yesterday, "The best doctors have only killed one person." Apparently, it's a common saying in UK healthcare. But it illustrates the point.

I currently work for a bank that is hot on AI ethics; I initially thought it was a bit of HR hype. But as I come to understand the inner workings of these systems, I am thoroughly convinced "recording how decisions are made is just as important as the decisions themselves."