I don't think you need to be ambiguity / risk averse to be worried about robustness of long-term causes. You could think that (1) the long term is extremely complex and (2) any paths to impact on such a complex system that humans right now can conceive of will be too brittle to model errors.

It's becoming increasingly apparent to me how strong an objection to longtermist interventions this comment is. I'd be very keen to see more engagement with this model.

My own current take: I hold out some hope that our ability to forecast long-term effects, at least under some contingencies within our lifetimes, will be not-terrible enough. And I'm more sympathetic to straightforward EV maximization than you are. But the probability of systematically having a positive long-term impact by choosing any given A over B seems much smaller than longtermists act as if is the case — in particular, it does seem to be in Pascal's mugging territory.

My understanding is that:

1. Spite (as a preference we might want to reduce in AIs) has just been relatively well-studied compared to other malevolent preferences. If this subfield of AI safety were more mature there might be less emphasis on spite in particular.
2. (Less confident, haven't thought that much about this:) It seems conceptually more straightforward what sorts of training environments are conducive to spite, compared to fanaticism (or fussiness or little-to-lose, for that matter).

Thanks for asking — you can read more about these two sources of s-risk in Section 3.2 of our new intro to s-risks article. (We also discuss "near miss" there, but our current best guess is that such scenarios are significantly less likely than other s-risks of comparable scale.)

I've found this super useful over the past several months—thanks!

Given that you can just keep doing better and better essentially indefinitely, and that GPT is not anywhere near the upper limit, talking about the difficulty of the task isn't super meaningful.

I don't understand this claim. Why would the difficulty of the task not be super meaningful when training to performance that isn't near the upper limit?

I am (clearly) not Tobias, but I'd expect many people familiar with EA and LW would get something new out of Ch 2, 4, 5, and 7-11. Of these, seems like the latter half of 5, 9, and 11 would be especially novel if you're already familiar with the basics of s-risks along the lines of the intro resources that CRS and CLR have published. I think the content of 7 and 10 is sufficiently crucial that it's probably worth reading even if you've checked out those older resources, despite some overlap.

Anecdote: My grad school personal statement mentioned "Concrete Problems in AI Safety" and Superintelligence, though at a fairly vague level about the risks of distributional shift or the like. I got into some pretty respectable programs. I wouldn't take this as strong evidence, of course.