exactly what robot Kai WOULD say

Seeing this all the time in our "treatment resistant depression" clinic. Patients who say that they never could fit in or get it right, and got increasingly depressed/suicidal as they moved through adolescence. The article points the arrow the wrong way. ASD traits get you an MDD diagnosis.

Ok, you two, then let's continue my thread w @shrewshrew.bsky.social bc this cuts right into anxiety and OCD. Do we know how/ where uncertainty gets tagged w negative valence? That seems like really valuable clinical target.

Which is close to what you say upthread but there's I think a subtle difference (maybe confounded). Which is interesting question bc do we know that exploit is less effortful than explore? I guess the RTs say that?

But also mental effort. cuz this was coding tasks. One could be asking it "write me a function that summarizes the output of these other 3" without having to hold those 3 in working memory. One of the worst things about coding is when it turns into basically N-back.

That's a different thing, though. The argument you were making to start, as I read it, was the idea that it's a sure-payoff action, so people are doing it bc it's easier than searching/trying to think of an answer, not bc it directly reassures them.

Also this thread really makes me think that at ACNP we should set up a giant beach arena, get some basketballs, and sponsor a marble burying contest for humans

So, logically, it is a low-variance, medium average payoff bandit. Going for long nature walks or drinking tea and coloring are high variance, often will produce nothing but lower blood pressure, but MIGHT suddenly produce "holy crud, what if we..." ideas, i.e. high variance.

But what about if the reward rate were mid? Bc isn't that the argument you're making here? GenAI doesn't produce huge payoff answers bc it can't -- it ultimately can only recombine tokens. The truly wild "nobody has ever thought this before" ideas are out of scope.

So if this is true, would we predict that given two bandits, one with lower variance on its random walk, mammals would prefer that? Because I thought the IGT data went the other way.