"I only want intervals" - said no one ever

For slowly changing things, I think typically yes (bc wikipedia-representative data is in there). For recent temporalities, it can typically fetch exogenous data.

The cases where neither is true, I find it’s bc the data is strictly private (eg- some financial derivative data).

Plus, the models can collect and use data. That’s another major source of information for reducing surprise.

Personally I identify a lot of errors, and I’m sure others do as well. So I think the information from people has a strong effect on identifying the errors (the idea people supply the judgement and the models follow instructions).

The models are retrained on their errors. Granted that is currently a slow process today, but many errors are reduced with each update.

Practically they’re all statistical models, and we can evaluate the utility they produce. I think it’s reasonable to include interpretable linear models in the set.

My point being that it would take me a lot longer, and in my mind that’s pretty intelligent (and one example). Plus, it’s pretty easy to identify things they (practically) understand relative to certain human populations.

AI is short-hand for a statistical model, so admittedly I don’t find that answer controversial.

eg- we could have people respond a la lmarena.ai/leaderboard, and compare.

That said, GPT and Claude can write and upgrade a skew generalized t implementation for me (even if yes they make mistakes).

AI’s smarter than all of us, and gee what an effect on operating leverage / net income?

I think AI’s smarter than all of us, especially rapture people. :)