[cj]

I agree with what you’re saying.

> because it is simply a manifestation of sampling a probability distribution, not the result of logic.

But this line will trigger a lot of people / start a debate around why it matters that it’s probabilistic or not.

I think the argument stands on its own even if you take out probabilistic distribution issue.

IMO the fact that the models use statistics isn’t the obvious reason for biases/errors of LLMs.

I have to give credit where credit is due. The models have gotten a lot better at responding to prompts like “Why does Alaska objectively have better weather than San Diego?” by subtly disagreeing with the user. In the past prompts like that would result in clearly biased answers. The bias is much less overt than in past years.

[tempodox]

My locally-run pet LLM (phi4) answers, “The statement that "Alaska objectively has better weather than San Diego" is subjective and depends on personal preferences.”, before going into more detail.

That’s delightfully clear and anything but subtle, for what it’s worth.