Addition: Hell, it could probably solve 2. LLM that does sentiment analysis. Sentiment drops below a certain level, the model knows to put in an order for Dominos for a company pizza party, since this is about as much morale boasting as most employees get.

I still wonder if this is something AI will ever be able to do, or if it’s much harder than we suspect and at least several more AI winters away.

It’s basically the is/ought problem, where the theory is that one cannot get to what one “ought” to do based only on what “is”.

Like we might say “it is cold outside, therefore you ought to put on a coat”. But, should you? Maybe you enjoy the cold, or maybe you live in an area ruled by a street gang and your coat is the wrong color and you risk being shot if you wear it outside.

AI could eventually make a much more informed recommendation, with more context, but at some point it begs the question of how close the model needs to get to reality, which obviously has some limit.