Optimization at a Distance has the mental picture in it, which… is not not a definition.
johnswentworth
Karma: 59,098
Yes.
In my head I usually think of it as non-myopic in spacetime (as opposed to just time), but the version which is (somewhat) justified by the Toy Coherence Theorem is non-myopia over time.
My standard answer to the kinds of points you’re making here is in A Simple Toy Coherence Theorem, specifically this section.
Those are useful examples, thanks. I’m gonna come at this from a different angle, but first to answer this question:
Maybe put another way: under your model, can the guy in the neckbeard meme dress well?
Unsure, either “no” or “it would be hard enough that I have no idea how he’d do it”. I do not think that makes this a model of attractiveness per se, rather I think every outfit has some level of attractiveness required to pull it off (which, for the worst outfits, may be beyond the attractiveness of any actual human). As the relevant section of the post said, attractiveness and charisma are necessary conditions.
Now back to the main cruxy part.
Suppose the neckbeard you linked to looks in the mirror and, in complete self-honesty, gets the vibe that he looks awesome. Like, he imagines someone else who looks just like that, and his gut response is “that guy looks awesome”. Well, as long as he’s being honest with himself… his own reaction is in fact extremely strong evidence that lots of other people would agree that he looks awesome. It’s very unlikely that he is so unique in his tastes (or lack thereof) that nobody else shares them. There may be lots of people who hate it, but at that point it is plausibly just correct to respond “well you can’t please everyone, and anyone who stands out is gonna bother some people”.
Personally, at this point I would tell him that he’s farther than he thinks from the best he could be doing, but yes there are probably many people who will in fact get a pretty good vibe from that outfit (again, assuming the premise that his own honest read was that it looks awesome). So at that point it’s down to numbers: just how many people would think it looks awesome, and how many would be driven away? In that guy’s case, I’d guess the “awesome” camp would be in the minority, but not so rare that it would never make sense to take the tradeoff (“better to be loved by a few and hated by many, than for everyone to feel nothing” as the saying goes).
This is importantly different from the fat guy in the meme! For that guy, the number of people who think he looks awesome is going to be very close to zero, and probably even in his own head he needs to motivatedly-ignore some things about the guy in the mirror in order to think he looks good.
Main point of all that: insofar as one is honest with oneself, one’s own reaction/vibe to a mirror is damn strong evidence of other peoples’ reaction/vibe. Probably at least a large minority will feel the same reaction/vibe, even if one lacks the majority’s taste.
And for most people, whatever reaction/vibe they get will basically match the reaction/vibe most other people would get; by definition most people are in the majority. If most people have the relevant taste, then most peoples’ honest vibe from the mirror will reflect that, and in that case the neckbeard is in the minority. If most people don’t have the relevant taste, then most peoples’ honest vibe from the mirror will reflect that, and in that case it’s fashion twitter which would be in the minority.
I think the neckbeard really truly believes that their outfit is vibe-appropriate.
I buy that, but I think the cause is more like a motivated blindspot than lack of taste.
Let’s look at the classic meme:
This guy’s problem is not combining an ironic graphic tee with a visor and a fedora and a katana. The guy’s outfit is not the main problem. The guy’s taste is not the main problem. The problem is he weighs like 300 pounds. What he needs is Ozempic.
Now, what does that guy think when he looks in the mirror?
When that guy looks in the mirror, I think he’s telling himself that the outfit looks great, and he’s motivatedly-not-thinking about how overweight he is. He’s telling himself that the outfit puts him in the same bucket as a better-looking guy wearing that same outfit and facial hair.
That’s not a problem of lack of taste, that’s a problem of lack of self-honesty. He needs to be honest with himself about what he sees in the mirror.
Maybe a concrete deviation: It seems like you see fashion largely through the lens of offensive vs. nonoffensive? At least this seems to be the main failure mode you describe, when talking about how an outfit can fail. [...] Is it your intuition that people fail to be fashionable by failing to be appropriately noisy? This would explain why bravery is the main requisite virtue to you while being nearly irrelevant to me.
I wouldn’t say that’s the main lens I use in my head, but that is an accurate description of a major failure mode, plausibly the most common failure mode.
(Aside: I’d prefer the term “loud” over “noisy”, because “noisy” is easily confused with noise in the sense of signal vs noise, and signal vs noise is a very central concept here.)
By choosing colors and silhouettes and layers that make sense to me on a particular day, though, I can somewhat-reliably show my friends that I’m feeling energetic and we should do something interesting, or I’m feeling cozy and we should stay in and laugh together, and I can interpret the same from their outfits.
I want to dig in on this. You of course have much more visibility into your own experience than I do, but just going by that description, my first guess would be that there’s a different thing actually going on.
Notably, “how energetic is <person> feeling” is one of the easiest things to read off of body language and speech patterns. There’s usually a pretty clear energy level being broadcast.
My guess would be:
you sometimes choose an outfit to match your intended energy level for the day
that outfit mostly acts as a self-signal, reminding you of your intended energy level for the day, so you act more in accord with that intended energy level
your friends then notice your energy level from all the clear nonverbal cues being sent, most of which is not the clothes themselves.
In particular, one prediction this model makes is that if you dress low-energy but feel/act high energy, your friends will mostly pick up high energy vibes rather than low energy vibes. They might not even notice the clothes mismatching the vibes.
Curious to hear how well that does/doesn’t feel like it fits your experiences.
There’s a relevant selection effect: plenty of people would benefit from just trying to emulate an archetype, but these are mostly people who are good at picking up non-verbal communication, so will learn a lot by just trying an outfit, noticing how they are perceived, repeating. [...]
I disagree with this model in multiple ways. I don’t currently have a short summary of that disagreement, so I’ll hit a bunch of points which might gesture in the right direction:
I expect that even people who are relatively good at picking up non-verbal communication will either get very few bits on how they are perceived (e.g. just general positive/negative valence and strength), or (more often) hallucinate a perception which is not there. The thing you’d want to get a read on is, like, what story/role/narrative the clothing is perceived as conveying, and that’s a relatively hard thing to get from a nonverbal read.
I don’t imagine feedback from other people being the main tool anyway. Like, the core way I imagine someone not falling into the neckbeard trap is by looking in the mirror and honestly asking themselves “does that person in the mirror actually give the vibe I’m looking for?”. The hard part is to pay attention to one’s gut and give an honest answer (and different people will be differently biased in that regard).
I guess feedback from others could substitute for the “look in the mirror”, especially if one does not trust oneself. But that substitutes trusting the feedback from others, which is itself tricky. Either one needs nonverbal reads on relatively detailed things like perceived story/role/narrative, or one needs to ask for verbal feedback in a way which both elicits the relevant information (which may be hard to articulate at all) and elicits honest answers.
Insofar as one does rely on feedback from others as a main driver of fashion choices… man, that feels like a trap. Someone who iterates on that feedback is going to shoot themselves in the foot, by e.g. doing things which offend nobody and therefore just look generic, or by doing things which seem kinda vaguely nice but don’t send a coherent message at all, or ???.
When I imagine someone who “understands the relevant things but hasn’t been brave enough to use it”, a central example would be someone who can look in the mirror and accurately assess whether they’ve got the vibe they want, but has mostly optimized their wardrobe based on valence-feedback from other people, without a coherent story/role/narrative. I want to tell that person “pick a strong archetype, make the person in the mirror vibe that archetype, be honest when assessing yourself, then be brave and go rock it”.
Channeling an archetype very hard without deeply understanding that archetype makes you look like somebody who really wants to be perceived in a particular way...
I mean, yes, obviously you need to understand the archetype deeply in order to channel it real hard? Ok, I suppose the existence of neckbeards indicates that that is not obvious and I should be explicit about it. But yeah, as with most things, one can certainly shoot oneself in the foot by not understanding what one is doing.
I guess you’re imagining that the advice would be “disastrous in basically all cases” because basically everyone who tries it will execute incompetently rather than competently? I’m not sure why you think that, “basically all cases” seems way too strong here. There are, for example, probably people who already understand the relevant social signals and are capable of competent execution, but either haven’t thought through what would actually benefit them most or haven’t been brave enough to actually do what would benefit them most.
Regarding your four specific points:
I totally buy the “signals to self” value proposition.
I also totally buy the “syntactic correctness” point, though I wouldn’t frame it that way in my head. Coherence of a look is recognizable and valuable even when it’s unclear what signal it sends.
The people who get to know me better are the least important for my clothes to signal to; they have loads of other data which will largely overwhelm my clothes. Clothes are most important for signalling to the people I know least.
I think you are vastly overestimating the amount of effort which other people will put into grokking one’s style, regardless of whether their taste is similar to one’s own. Especially since the people I care most about signalling to with my clothes are exactly those who I know least.
On a meta note: I’m responding to most of your points because I find this discussion interesting. But I know a dense thread like this can be tiring, so please don’t feel too obligated to respond if you find it draining.
If you want to get better at fashion because you want to unify external perception, I expect this to fail unless you want your perception to be a stereotype (scene clothes, suit-and-tie, fedora and neckbeard).
That would be roughly my recommendation: pick a universally-recognizable archetype and channel that archetype real hard. (This generalizes beyond “coolness” to other goals.) Otherwise, you’re just sending messages which are in-practice far more ambiguous and weak than you think to an audience which is small to begin with, and probably the number of people who will actually recognize what you intended the clothes to say is approximately zero.
The second one: the AIs will be producing legible-and-verifiable breakthroughs in other fields, but those same AIs will be producing slop in the case of alignment.
This should be reasonable on priors alone: hard-to-verify problems are a systematically different class than easy-to-verify problems, so it should not be a huge surprise if generalization failure occurs across that gap.
It also makes sense from the perspective of “what’s incentivized by RL?”. In hard to verify areas, evaluators have systematic and predictable biases; outputs which play to those biases can and sometimes will outscore actually-good outputs. So we should expect an RL’d system will be actively selected to produce slop which plays to those biases (as opposed to straightforwardly good things), e.g. sycophancy.
And we already see this phenomenon on easier problems today. Today’s AI already do the most impressive things in e.g. math and programming competitions, the easiest cases for verification. The harder verification gets, the more they spit out slop. So just extrapolate what we already see to the more extreme regimes of stronger AI.
(To be clear, my median guess is that we’re about one transformers-level paradigm shift away from strong AI still, and it makes a lot less sense to extrapolate today’s AI to a different paradigm. But insofar as one expects LLMs to hit criticality, especially using RL, one should expect that they’ll produce systematically more slop in harder to verify areas.)
If I understand correctly, your position is that we lose status points when we say weird (as in a few standard deviations outside the normal range) but likely true things, and it’s useful to get the points back by being cool (=dressing well).
That is not my position.
Being cool is a specific social strategy (not synonymous with dressing well). The nature of that strategy frees one to do a lot more nonconformist-signalling things, without losing status each time, because doing nonconformist-signalling things is an expected part of people playing a “cool” social role.
If anyone does this exercise, I’d be very interested in people sharing what they thought various outfits said, to see if different people converge on similar answers. Whether-or-not-this-is-pseudoscience seems to hinge entirely on that question, and it’s very testable.
I would guess that even the “in the know” people are over-updating, because they usually are Not Measuring What They Think They Are Measuring even qualitatively. Like, the proxies are so weak that the hypothesis “this result will qualitatively generalize to <whatever they actually want to know about>” shouldn’t have been privileged in the first place, and the right thing for a human to do is ignore it completely.
My immediate critique would be step 7: insofar as people are updating today on experiments which are bullshit, that is likely to slow us down during early crunch, not speed us up. Or, worse, result in outright failure to notice fatal problems. Rather than going in with no idea what’s going on, people will go in with too-confident wrong ideas of what’s going on.
To a perfect Bayesian, a bullshit experiment would be small value, but never negative. Humans are not perfect Bayesians, and a bullshit experiment can very much be negative value to us.
+1 to this. Tying it back to coolness specifically: coolness will absolutely alienate some people. That is not just an unavoidable side effect, it’s load bearing for the game theory to work.
Trying to offend as few people as possible leaves one looking generic, or like a politician/middle manager.
As the opening of the post points out, nonconformity is a strict requirement for better-than-baseline epistemics. If one is trying to offend few people in general, then one’s epistemics are doomed to basically-baseline. So I don’t recommend that goal.
One could still argue that looking generic and inoffensive is fine, so long as one’s beliefs aren’t under lots of conformity pressure. But man, I sure would be very skeptical if someone claimed that were true of themselves.
I do not understand the statement of your current understanding, in particular point 1. Could you please state in different words and/or state it formally and/or give an example?
… did you read the section after that one?
I’d start by looking at some fictional characters or musicians for inspiration.
Personally, I ended up with my current look over many little steps, but I don’t necessarily think that’s the best way to do it.
Fixed, thanks.
This can be done in a way which doesn’t cost significant weirdness points.
The key, I think, is to include a lot of coolness in whatever you’re signalling. Being “cool” is largely about embracing countersignalling in all its glory; coolness and conformity are strongly opposed. Being cool is the main way I know of that one can be very nonnormal/nonconformist without paying lots of weirdness points and seeming low status.
Ah, yes, that is almost correct. You need redundancy over TWO distinct observables (i.e. the subset must be at least size two), not just one, but otherwise yes. With just one observable, you don’t have two branches to dangle a Λ′ off of in the graphical proof, so we can’t get Λ between two copies of Λ′.