07mk
4mo ago
Yeah thats not proof, thats a theory. There are other theories about what the LLM is doing and they are just as explanatory as yours is. You have run no experiments to isolate those alternatives and test whether or not they exist. You have run no ablation studies and no studies to attempt to isolate co-occurring variables. It is by definition a theory. Hence why I asked for proof, because I am certain you have none.
I'm not sure where the disconnect is here. This is a simple logical question, not an empirical one. There's no need to check for some physical representation of a model, because making accurate predictions requires an implicit model, QED. I've yet to see anyone suggest an alternative - certainly not in this thread - and I'm not sure what alternative explanation could exist, logically.
We can prove the child has an internal model of physics because we have 7 billion humans, including our own selves, we can extrapolate our internal abilities across a generalized set of all humans.
But we don't need 7 billion humans or even any other human than that child to conclude this. If we landed on an alien planet and observed an alien doing this, we would also know that it had an internal model of physics. If someone made a robot that could do this, we would know it as well.
07mk
4mo ago
I'm reminded of the phenomenon of, "When I look at how barbarous orcs are depicted in your work of fiction, my mind can't help but be immediately reminded of black people/other minorities, which is clearly an indication of your innate, implicit racism that you should do the hard work of inspecting and extracting."
07mk
4mo ago
So at some level, it's a question of semantics. But I think it also has real-world implications. If an LLM lacks models (or perhaps I should say "sophisticated models," then in my view (1) it's missing an important ingredient of human-level intelligence; and (2) it can't be conscious.
This clears up my confusion. I agree with you in that, the current evidence of generic LLMs is that they lack "sophisticated" models of chess, for some reasonable definition of "sophisticated." Now, whether or not that means it's missing an important ingredient of human-level intelligence or can't be conscious, I don't know, and I'm not sure how anyone can know. What seems very likely to me is that, lacking a "sophisticated" model of chess (or the world, or social life, or physics, or etc.), it's lacking an important ingredient of human-emulating or human-like intelligence, but that doesn't imply that it lacks human-level intelligence. In terms of consciousness, I think the Hard Problem remains Hard.
And yet most people would agree that it doesn't actually model the game in the sense that the computer system contains no internal representation of a chessboard.
Perhaps most people would agree with that - I might, depending on what you mean by "internal representation" - certainly I doubt that the computer would have a model that could trivially show an accurate representation of each of the 64 squares and where each of the 32 pieces sit on the board and whose turn it is. But I'd say that doesn't mean that the computer isn't modeling the game or that it doesn't have some sort of internal model of the chessboard. It's just a wrong model, one that is far wronger than any typical human would have, and one that is wrong due to bizarre mistakes that no stupid human would commit.
07mk
4mo ago
They like predictable fiscal outcomes, safe products
Is this actually true? One major point of criticism these days towards AAA games the last few years is that so many companies are trying to create the new Fortnite, pouring hundreds of millions of dollars into live service game after live service game, with the Venture Capital-like math where 1 "infinite money glitch" success pays for 9 failures.
Concord is the most prominent example (rumored to have cost $400MM, almost certainly at least $100MM - shut down in under 2 weeks due to lack of players), but there are others, like Marathon this year (also by Sony, which bought the devs Bungie for almost $4 billion, with likely an additional >$100MM spent on development - currently averaging around 10-20k players on Steam). Sony is also currently working on a live service game based on the Horizon franchise. I'm not the first person who has pointed out that Sony owns a ton of franchises known for great singleplayer games that they could have poured a "modest" sum of around $20MM-$50MM for development to generate a "modest" profit fairly reliably instead of going for moonshots. Likewise, Ubisoft owns a ton of great singleplayer franchises like Splinter Cell or Prince of Persia, which they tend to ignore in favor of trying to make the next 10MM+ selling AssCreed game with microtransactions. Maybe they finally hit rock bottom and are reversing course with remaking Black Flag.
Heck, if you gave me $20MM, I bet I could get enough devs together to make a competent remaster Bloodborne which would almost certainly make a nice, safe profit, given how much bigger Soulslikes and Fromsoft have gotten in the 11 years since the original release. Yet Sony seems to have no interest in that.
07mk
4mo ago
Refer to the 2nd part of my comment about providing evidence of the internal workings of the LLM to prove it has an internal model.
Again, the proof is in the external behavior. To be able to predict something external at a rate better than chance requires some model somewhere. We know that these LLMs don't have an external model. Therefore it must have an internal one. Much like how, say, a 5 year old who can throw a ball towards home plate certainly has some internal model of physics, as proven by the fact that he can, at a rate better than chance, throw the ball towards home plate instead of at 3rd base or straight up or just dropping it on the mound. We don't need to plant electrodes in his brain or do some fMRI studies to know this, the proof of the pudding is in the eating.
07mk
4mo ago
That is not actually true. It could merely be well trained or even overfit on a statistical distribution of chess moves such that it can proffer a valid move.
Well-trained or overfitting on a statistical distribution of chess moves is a model of chess, though. A model that's wrong (like most models), and one that's likely not very useful (like some models), but that doesn't make it not a model.
07mk
4mo ago
Is that an internal model, or just "usually a player castles after moving their knight and bishop" correlation?
The correlation would be part of the internal model.
07mk
4mo ago
The evidence (or rather; proof) is that it generates text that conforms to chess rules at a rate better than chance. To be able to predict something better than chance, it requires an implicit internal model (well, assuming there's no explicit/external model, anyway, which is the case here) of it.
07mk
4mo ago
Your logic here seems exactly backwards to what was being stated, so I'm not even sure how to interpret your question.
07mk
4mo ago
I'm not sure I understand your point, but in my view, unless the LLM outputs a textual representation of the game board for each turn, it's not actually modeling the game.
This is where I disagree. If it's outputting correct moves at a rate greater than chance, then it's certainly got an internal model of the game in there somewhere, in order to predict moves. The model is certainly wrong and, again, likely doesn't resemble an 8x8 grid with 16 pieces on each team, with each piece having a set of legal moves, etc. But rather might involve bizarre rules like "if white starts with XX, then black responds with YY" and such. But that just makes it a wrong model - which makes it similar to most models - not not a model.
07mk
4mo ago
Fortunately for the "battle of the sexes," it seems evident that plenty of women don't consider having a one-night stand to be intrinsically analogous to using a man and then throwing him away like a snotty tissue.
07mk
4mo ago
The problem is that that is literally, objectively, what LLMs are doing.
Sure, and also, we can say that what both LLMs and humans are doing is having the atoms and energy (but I repeat myself?) that make them up following the laws of physics in a way that creates physical motion. That's something that's literally, objectively true. Now, what the atoms and energy that make up the LLMs are doing can be, in aggregate, described as "next token prediction." We don't know if what is creating human cognition is something that is meaningfully analogous to "next token prediction," because the atoms and energy are aggregated in very different ways in forms of things like "neurons" and "neurotransmitters" and many many other things. But given that human cognition arises from a bunch of dumb atoms and dumb energy dumbly following a dumb algorithm that we call physics, it's evident that a bunch of dumb things following dumb rules isn't necessarily incapable of producing the equivalent of human cognition.
07mk
4mo ago
I disagree, another possible reason is that simply makes a good (but imperfect) guess as to what's likely to be the next move after a sequence of moves, based on all the chess games stored in its database.
That's not another possibility, though; that's just describing actually how the LLM works for generating the model of chess (via the training) and the chessboard (via the text input) and then using the model to generate next moves (the generated text).
07mk
4mo ago
In another thread, I echoed the idea that LLMs don't model the universe. So for example, if you play chess with an LLM, there's no model of a chessboard in the system, which is why it sometimes makes illegal moves.
I've seen this kind of notion argued in many different contexts, and I don't understand what's the disconnect. Because OF COURSE the LLM has an internal model of the chessboard in the system; that's the only reason it could possibly make moves that are correct at a rate better than chance. That model almost certainly doesn't looks like a model that any human would recognize, such as containing a grid of 8x8 with pieces each representing a team, a position, and a set of allowed moves, which is why it makes mistakes in ways that no human would. But the fact that the model of chess - or the world - would be incomprehensible to humans and isn't based on any real empirical or experienced understanding of physics or rulesets doesn't make it not a model.
07mk
4mo ago
Baby Daddy material means that the man is so attractive that you're willing to fuck the long-term negative consequences in favor of fucking him. Husband material means that you'll fuck him only because of the long-term positive consequences that follow.
07mk
4mo ago
Yes, I would wager there's a significant such bloc, though I'd also wager that the bloc of former-Dems or borderline-Dems who would be heartened by such a report to such an extent as to influence their vote positively in the Dem direction is even more significant. If not in absolute numbers, then certainly in the effect on votes. It's almost certain that such a report would cause a significant bloc of current Dem voters to peel away, but they don't have a mainstream party to go for, and I'd also wager that a very significant number of that group are concentrated either in blue states or blue enclaves of red states where the POTUS election, at least, would have minimal negative impact.
07mk
4mo ago
For a lot of people, and not necessarily fully blind tribalists, their side is better because of prior assumptions that are not in question.
I would say that this sentence is essentially self-contradictory. The "fully" can sorta save it, but even then, to whatever extent these people are only partially blind tribalists, it just doesn't touch on the actual, meaningful thing about not being a blind tribalist, which means being open to questioning such prior assumptions.
07mk
4mo ago
Well, I understand that a lot of Democrats are blind tribalists, but a lot of them still do value the idea that our side should win because it's actually better than the other side, not merely because it's our side. If we can't openly analyze the "soul of the dnc," then we can't be confident that our side actually is better.
07mk
4mo ago
Every poll on this I've seen has shown significant split between the two choices, at most maybe 70-30 one way. This has convinced me that, if this were done IRL, there's basically no way that Blue would get 50%, and I'm skeptical it'd get over 20%. If the voting is split when there are no consequences and you can choose whatever makes you feel virtuous knowing that you won't ever have to walk the talk, then in a situation of fatal consequences, there's simply no plausible way that the "don't die" button wouldn't have overwhelming victory. Given that, I don't see how I could justify adding one more body to the pile, instead of gritting my teeth and accepting the responsibility of keeping society running after it's been approximately decimated.
07mk
4mo ago
I'm in the DNC and want the party to have success in the future, the best situation is to move on entirely from anything that had to do with Biden.
Hard disagree. The best situation is to highlight it so much that every Democratic politician has no choice but to learn from it. Prove to the electorate that this party is actually better than the other one, because it actually learns from its mistakes and takes punitive actions, even against its own ego, to make sure it doesn't happen again. Losers who don't learn usually stay losers, and people usually don't like to side with losers.
07mk
4mo ago
If the Democrats were to release a report like that, the fact that the DNC would put their names on such a pathetic ego-protecting report is something every Democratic voter would find immensely valuable, for deciding how much reform the party leadership needs. Because a DNC that would produce such a report is one that is neither interested in getting things right nor in winning, and those are important characteristics for any supporter of any party to consider.
07mk
4mo ago
I'm not sure what the point of your comment here is, because all your points seem entirely orthogonal to the phenomenon I talked about. So I'll just directly answer the direct questions that were in your comment.
men will praise women for having big, natural tits
In their face? Not the best strategy unless you're already having sex.
It's probably not the best strategy, but it's absolutely a very common one, and for good reason. Men complimenting women for their great figure or other genetically-determined aspects of their physical appearance, such as their "big beautiful eyes" as part of flirting is pretty much cliche.
Or among the boys?
AND among the boys, not OR, though in all-male settings, they'll often feel more free to use crude language, such as using the phrase "tits."
Don't women also fawn about a guy in non-personality ways when among trusted female friends?
I'm not a woman, so I lack any meaningful insight into this, but I'd guess that this is probably the case.
No one's talking about rubbing anything in here. The conversation is about praising others.
Also praising makes sense in relation to stuff you did. You expended effort and achieved a positive result, that's laudable. You deserve no cookies for how your face looks or similar.
Why not? Someone having a prettier face due to luck of genetics makes things more pleasant for others around them, almost by definition. If such people receive praise that they value, that provides incentive for such people to show their faces more often than those who aren't genetically lucky, which makes the lives of those around them, including my own, better.
But even before we get into the logic of incentives, by default I'm going to praise people based on how I appraise them. Proving you can accomplish things with effort is one way of raising my appraisal of you, but also proving that you are genetically gifted in a way that makes my life more pleasant is another way. This is why, again, women praise men for things like being tall and assertive and men praise women for things like having big, natural tits. They don't care about how much effort these people put into accomplishing these things, they just care about the effect they have on themselves.
07mk
4mo ago
Same reason men would prefer a woman who's naturally beautiful over a woman who uses tons of makeup and cosmetic surgery to "fix" her looks.
I've noticed this cliche and also the mirror cliche in both sexes where men/women will tend to praise other men/women for looking good in ways that are the results of effort, not genetics. E.g. men will praise other men for successfully bulking up at the gym, whereas women will praise men for having a "great personality," and women will praise other women for doing such a bang-up job with their make-up, while men will praise women for having big, natural tits. I think there's a heavy influence of selfish interest in both sexes here, where if you can bootstrap your way into convincing the other sex (or at least bullying them at least long enough for you to escape the game) that [effort-based] rather than [genes-based] (everything is based on both, of course, and this is a matter of degree) things are greater contributors to one's attractiveness, then you individually have more control over your own destiny.
Human-emulating would be reaching conclusions and decisions through a process that is similar to how humans do in some way. The most obvious way would be that it follows some sequence of "thoughts" that a typical human could look at and honestly think, "That's similar to how I might think through this." Another way might be if it literally emulates our entire brain, possibly down to the sub-atomic particles.
Human-level would simply be if it's able to pass intelligence tests (any that you could come up with, including IQ, but also things that might involve social awareness or performance in physical tests) at a rate similar to humans. How it accomplishes this wouldn't matter; perhaps tomorrow we discover that God is real and He can be communicated with via a new antenna we developed. Then we put that antenna on a computer and tell it to ask God what to do, in order to behave as intelligently as a human, and God in His great benevolence, decides to answer accurately. That computer would have human-level intelligence, but certainly not human-emulating.
More options
Context Copy link