If I had to guess I think it’s relevant to like, anthropic reasoning, or something.
AprilSR
Karma: 1,126
Okay, this makes sense! It’s not obvious to me exactly how ambitious 2 is, but I get why you might be skeptical.
I’m having a little trouble understanding the whole argument. It’s not obvious to me why exactly this line of reasoning doesn’t prove too much by ruling out human speech? Plenty of human phonemes are like 10ms long?
I have no idea why the duration of an individual click is supposed to be relevant. There’s like, at least 30 milliseconds between clicks (according to Claude), and usually more than that, which seems like the relevant number to me?
Oh, I meant in the category of (topological) vector spaces, which requires the quotient maps to be linear.
I think maybe part of the confusion is that, when you’re working with vector spaces in particular, subspaces and quotient spaces are the same thing.
My intuition is like… you get a topological circle by gluing the two ends of an interval together, but no subspace of the interval is homeomorphic to a circle. I’m not entirely sure that this sort of issue meaningfully impacts neural networks, but I don’t immediately see any reason why it wouldn’t?
Claim 2 sounds very likely false to me.
Surely 28K should be at least 5 points!
I learned that antipsychotic medications have unpleasant side effects that can make people unwilling to get on or stay on them. Once a brain malfunctions that badly, without treatment, it never gets fixed.
I’d consider cobenfy as an option here, maybe? Or sometimes nicotine, allegedly, though that obviously isn’t ideal.
I very much know how continuous functions work and precisely what differentiability is, and I have filed taxes, but I would probably need a refresher on tax brackets to be sure I had everything right...
Yeah I’m pretty sure it’s an idiosyncratic mental technique / human psychology observation, there isn’t technical agent foundations progress here.
Wow, this sure is a much clearer way to look at the self-pseudo-prediction/action-plan thingy than any I’ve seen laid out before.
Do we mean active LessWrong users? <10% would shock me, if you use a filter or weighting that solves the “probably there are a lot of people who look at LessWrong ever other than ‘real’ LessWrongers” aspect.
Maybe it’s less than half though. There might be a large contingent that has only read like, HPMOR and the Sequences Highlights.
Point out that most peoples’ clothes don’t really manage to signal what they intended even when they’re trying, and someone will say something like “well, it’s largely about signalling to oneself, e.g. to build confidence, so it doesn’t matter if other people get the signal”. And, like… I roll to disbelieve?
I think having a good model of when vibe people get from clothing is important and I find it plausible that there is some rationalization going on with this… but also, the self-signaling thing does seem like a large enough aspect to be the most important part, to me, even if the other-signaling aspect isn’t entirely unimportant.
I think part of this is probably that beauty is much less one-dimensional?
I agree Tessa’s explanation isn’t especially good, though it’s maybe more “incomplete” than “bogus”.
I don’t think the minimax theorem comes anywhere close to implying the existence of some sort of true optimal strategy, though, which I think becomes clear if you consider two types of chess bots. Bot A plays the same move as (something like) LeelaPieceOdds—unless that move would be moving from a game theoretically won position to a draw or loss, or from a draw to a loss, in which case it, say, randomly selects from all the moves that don’t do that, or ideally picks a move that humans are inclined to blunder against (maybe LeelaPieceOdds’s second choice, or something.)
On the other hand, Bot B immediately resigns whenever the position is a game theoretic loss, and immediately offers a draw whenever the position is a game theoretic draw. If its opponent rejects the draw offer, Bot B prefers to stay in states with the fewest opportunities for its opponent to blunder.
While both are “inexploitable”, Bot A beats humans every time, and Bot B draws them every time. (Unless chess is a game theoretic win for (WLOG) white, in which case Bot B wins as white but immediately resigns as black.) If you made Bot B a chess.com account, it very well might literally never break 1000 Elo.
So in pathological cases the non-transitivity can get pretty bad. The tops result sounds really neat but I haven’t read it yet and don’t know exactly how Bot B would fit into it—obviously “<1000 Elo” isn’t really going to be a description of Bot B that captures how it fits into such a structure very well.
I swear I don’t laugh about janus generally it’s just that the way you wrote that paragraph was really funny
My cursed proposal is to have a second language model extract the code from the original response.