This is a special post for quick takes by Wei Dai. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
[Epistemic status: butterfly speculation, not confident about this, but I think it's an idea worth taking seriously.]
Terence Tao spends his time wireheading on abstract math
So I was initially [skeptical of]/[slightly repulsed by] this framing[1], but after davidad's recent LLM-induced "awakening", I am starting to wonder that perhaps very high-g (+ high-NFC?) people tend to have a tendency towards something that is not very badly described as wirehead-y.[2]
If you think about cognition/theorizing as divided into generation+verification, then we can take verification as equivalent to receiving a reward signal. A wireheading, in this regime, is yearning for an unambiguously strong positive verification signal[3] and the two polarities of domains which could satisfy this are: (1) domains where an "objective" signal is relatively easy to get; (2) domains where it's easy to delude yourself, because your generation easily outruns and tricks your verification[4] (and maybe that of other people, who are sufficiently interested/enthused by your ideas but not smart enough to inside-view-rebut them).
You're putting Tao in the first one, and IIRC you made somewhat similar remarks about von Neumann elsewhere. Davidad might be going from (1) to (2). Newton also fits the pattern of someone who did (1) and (2) at various parts of his life, having invented calculus but also wasted a lot of his life on alchemy and astrology.
(There might be a simple explanation of this in terms of active inference, perhaps combined with the lethal reality hypothesis.)
ETA: Insofar as we want to speculate about this in terms of well-trodden constructs of modern academic psychology, the need for cognitive closure might be relevant, alongside g/IQ and need for cognition.
With the caveat that if you're aiming for it robustly/intelligently, then you're willing to do a lot of stuff that turns out to be very wrong in the short-term for the sake of getting a better verification signal in the long-term (at least in expectation or something).