primbin

Maybe this is too reductive, but I think interpretability is a good thing. If matrices are actually scrutable, that's good for alignment. This line of research reads as more promising than worrying to to me.

Open to being wrong though.

If you believe in the natural or convergent abstraction hypothesis, then "human comprehensible" isn't necessarily arbitrary.

IMO, the idea of recursive generalization doesn't really seem narrow the hypothesis space enough for this specific neuron for me to feel confident drawing a definite conclusion. I think recursive generalization is probably right, but I still feel confused about this neuron.

This is great. I've been looking for ways to improve my communications and this is exactly the kind of advice I think I've been in need of.

No, there's no single "optimal play" for math, but it wouldn't surprise me if there were a set of fundamental mathematical structures which would be consistently rediscovered by a wide variety of different "math playing" algorithms (of sufficient math-playing skill).

I can't tell whether this was very well done or if I'm just very gullible

It was good! I'm surprised to hear that this was your first piece of short fiction! As a writer I'm inexperienced, but as a reader I thought the dialogue felt pretty naturalistic and well-paced. And I too liked the bat's tendency to interrupt Dwarkesh, lol

Six months ago I read this and thought, "Yeah, this is obviously true of me."

For example, a thing I'd do is ruminate for days on a difficult conversation I needed to have, in order to devise a way to deliver the information while causing the other person the least pain — but that's not always possible. If, for instance, I need to break up with someone, there's no way for me to guarantee they will not be hurt. By breaking up with someone, I am *by nature* inflicting a painful experience on them! The information itself is painful! The task is impossible, but... (read more)