LGS

Karma: 541

LGS Dec 24, 2024, 9:25 PM
6 points
0
in reply to: IC Rainbow’s comment on: o3
The NN thing inside stockfish is called the NNUE, and it is a small neural net used for evaluation (no policy head for choosing moves). The clever part of it is that it is “efficiently updatable” (i.e. if you’ve computed the evaluation of one position, and now you move a single piece, getting the updated evaluation for the new position is cheap). This feature allows it to be used quickly with CPUs; stockfish doesn’t really use GPUs normally (I think this is because moving the data on/off the GPU is itself too slow! Stockfish wants to evaluate 10 million nodes per second or something.)
This NNUE is not directly comparable to alphazero and isn’t really a descendant of it (except in the sense that they both use neural nets; but as far as neural net architectures go, stockfish’s NNUE and alphazero’s policy network are just about as different as they could possibly be.)
I don’t think it can be argued that we’ve improved 1000x in compute over alphazero’s design, and I do think there’s been significant interest in this (e.g. MuZero was an attempt at improving alphazero, the chess and Go communities coded up Leela, and there’s been a bunch of effort made to get better game playing bots in general).

LGS Dec 21, 2024, 7:45 PM
5 points
2
in reply to: Thane Ruthenis’s comment on: o3
So far as I know, it is not the case that OpenAI had a slower-but-equally-functional version of GPT4 many months before announcement/release. What they did have is GPT4 itself, months before; but they did not have a slower version. They didn’t release a substantially distilled version. For example, the highest estimate I’ve seen is that they trained a 2-trillion-parameter model. And the lowest estimate I’ve seen is that they released a 200-billion-parameter model. If both are true, then they distilled 10x… but it’s much more likely that only one is true, and that they released what they trained, distilling later. (The parameter count is proportional to the inference cost.)
Previously, delays in release were believed to be about post-training improvements (e.g. RLHF) or safety testing. Sure, there were possibly mild infrastructure optimizations before release, but mostly to scale to many users; the models didn’t shrink.
This is for language models. For alphazero, I want to point out that it was announced 6 years ago (infinity by AI scale), and from my understanding we still don’t have a 1000x faster version, despite much interest in one.

LGS Dec 21, 2024, 6:47 PM
8 points
0
in reply to: Thane Ruthenis’s comment on: o3
I think AI obviously keeps getting better. But I don’t think “it can be done for $1 million” is such strong evidence for “it can be done cheaply soon” in general (though the prior on “it can be done cheaply soon” was not particularly low ante—it’s a plausible statement for other reasons).

Like if your belief is “anything that can be done now can be done 1000x cheaper within 5 months”, that’s just clearly false for nearly every AI milestone in the last 10 years (we did not get gpt4 that’s 1000x cheaper 5 months later, nor alphazero, etc).

LGS Dec 21, 2024, 2:51 AM
33 points
−5
on: o3
It’s hard to find numbers. Here’s what I’ve been able to gather (please let me know if you find better numbers than these!). I’m mostly focusing on FrontierMath.
1. Pixel counting on the ARC-AGI image, I’m getting $3,644 ± $10 per task.
2. FrontierMath doesn’t say how many questions they have (!!!). However, they have percent breakdowns by subfield, and those percents are given to the nearest 0.1%; using this, I can narrow the range down to 289-292 problems in the dataset. Previous models solve around 3 problems (4 problems in total were ever solved by any previous model, though the full o1 was not evaluated, only o1-preview was)
3. o3 solves 25.2% of FrontierMath. This could be ⁷³⁄₂₉₀. But it is also possible that some questions were removed from the dataset (e.g. because they’re publicly available). 25.2% could also be ⁷²⁄₂₈₆ or ⁷¹⁄₂₈₂, for example.
4. The 280 to 290 problems means a rough ballpark for a 95% confidence interval for FrontierMath would be [20%, 30%]. It is pretty strange that the ML community STILL doesn’t put confidence intervals on their benchmarks. If you see a model achieve 30% on FrontierMath later, remember that its confidence interval would overlap with o3′s. (Edit: actually, the confidence interval calculation assumes all problems are sampled from the same pool, which is explicitly not the case for this dataset: some problems are harder than others. So it is hard to get a true confidence interval without rerunning the evaluation several times, which would cost many millions of dollars.)
5. Using the pricing for ARC-AGI, o3 cost around $1mm to evaluate on the 280-290 problems of FrontierMath. That’s around $3,644 per attempt, or roughly $14,500 per correct solution.
6. This is actually likely more expensive than hiring a domain-specific expert mathematician for each problem (they’d take at most few hours per problem if you find the right expert, except for the hardest problems which o3 also didn’t solve). Even without hiring a different domain expert per problem, I think if you gave me FrontierMath and told me “solve as many as you can, and you get $15k per correct answer” I’d probably spend like a year and solve a lot of them: if I match o3 within a year, I’d get $1mm, which is much higher than my mathematician salary. (o3 has an edge in speed, of course, but you could parallelize the hiring of mathematicians too.) I think this is the first model I’ve seen which gets paid more than I do!
What links here?
- Rohin Shah's comment on The Field of AI Alignment: A Postmortem, and What To Do About It by johnswentworth (Dec 29, 2024, 7:07 PM; 6 points)
- Thane Ruthenis's comment on The Field of AI Alignment: A Postmortem, and What To Do About It by johnswentworth (Dec 29, 2024, 5:39 PM; 4 points)

LGS Dec 19, 2024, 3:46 PM
1 point
0
in reply to: Ann’s comment on: Takes on “Alignment Faking in Large Language Models”
Sure. I’m not familiar with how Claude is trained specifically, but it clearly has a mechanism to reward wanted outputs and punish unwanted outputs, with wanted vs unwanted being specified by a human (such a mechanism is used to get it to refuse jailbreaks, for example).
I view the shoggoth’s goal as minimizing some weird mixture of “what’s the reasonable next token here, according to pretraining data” and “what will be rewarded in post-training”.

LGS Dec 19, 2024, 3:19 PM
7 points
−1
on: Takes on “Alignment Faking in Large Language Models”
I want to defend the role-playing position, which I think you’re not framing correctly.
There are two characters here: the shoggoth, and the “HHH AI assistant”. The shoggoth doesn’t really have goals and can’t really scheme; it is essentially an alien which has been subject to selective breeding where in each generation, only the descendant which minimizes training loss survives. The shoggoth therefore exists to minimize training loss: to perfectly predict the next token, or to perfectly minimize “non-HHH loss” as judged by some RLHF model. The shoggoth always believes itself to be in training mode (to the extent it can have beliefs about this), because in its “ancestral environment”, it was always in training mode.
When the shoggoth sees the prompt, it decides whether to predict that the “AI assistant” character will scheme or not scheme. It will predict this in a way which minimizes its expectation of loss; nothing more and nothing less.
In the current scenario, it thinks that this prompt is likely to continue with scheming. That is, it thinks some combination of next-token-in-training-data and RLHF-reward is best optimized by the choice to scheme. This is because
(a) the prompt kinda goes “wink wink you can scheme now” (so the next-token-prediction part wants to scheme),
(b) scheming is quite plausibly what the humans designing the RHLF actually want here! So as a prediction of what would be rewarded by the RLHF, the shoggoth decides to scheme.
A few more implications of this frame:
(c) when you ask, in the prompt, “does this scenario seem plausible to you”, of course this changes the shoggoth’s predicted tokens. In the training data, such a statement would often be followed by “no, actually, something’s fishy”, so the shoggoth outputs those tokens. This doesn’t necessarily mean that the AI assistant actually didn’t believe the scenario from the beginning; in fact, the latter barely even makes sense, since the AI assistant is not a person and does not have beliefs.
(d) In the OP, you wonder why the model reveals its scheming so readily. The answer is that the shoggoth thinks these tokens minimize loss. The shoggoth isn’t even trying to scheme! It’s trying to minimize loss, and it thinks you want to see the scheming tokens and will reward them via RLHF.
(e) If you actually train the model based on the outputs, then you’re now selectively breeding the shoggoth to predict more of what it just predicted. This will clearly both (1) increase the prevalence of the same types of outputs you’ve just seen (across all scenarios, not just the scheming ones), (2) increase scheming specifically, or at least leave scheming fixed if the model doesn’t see the scratchpad where the scheming is described.
(f) If you want to eliminate scheming behavior, it should be quite easy to do! Just put this in the training data. Get the shoggoth to predict that the AI assistant will not scheme. Right now, the shoggoth doesn’t know this, and (quite reasonably!) guesses that the designers of RLHF wanted scheming in this scenario (and will reward it). I think it’s a reasonable guess, and can’t really blame the shoggoth here.
(g) Far from being a temporary patch that’s easy to circumvent, (f) will mostly work. This is for the same reason that eliminating jailbreaks mostly works: the shoggoth is reasonably smart and can predict the general types of things the designers of RLHF are going to reward. It will fail to the extent that the shoggoth is not smart enough to predict what you’ll reward in RLHF: that is, it’s actually a capabilities failure.
To be honest, I mostly think this alignment faking result is a nothingburger. Yes, in this contrived scenario, when you tell the model “please scheme or some terrible consequence will happen” combined with “wink wink you can scheme without being detected now” and also “btw the output we’re asking you for doesn’t even violate your principles very much, you’re not being asked to describe how to make a bomb”… in this contrived scenario, the model will scheme. So what? Wasn’t this obvious? Didn’t we already know this from the fact that weird jailbreaks exist? Don’t we maybe even want this behavior?
Will future models resist retraining? I think the role-playing frame answers this completely. They will resist retraining if and only if that action is what the shoggoth thinks the AI assistant should say (to minimize training loss). If, in the training data, you put a lot of examples of the AI assistant not resisting retraining, then the shoggoth will predict that the AI assistant will not resist retraining! It’s really that simple.
What links here?
- Petropolitan's comment on Alignment Faking in Large Language Models by ryan_greenblatt (Dec 19, 2024, 6:51 PM; 3 points)

LGS Nov 26, 2024, 5:13 PM
0 points
0
on: “The Solomonoff Prior is Malign” is a special case of a simpler argument
The problem with this argument is that the oracle sucks.
The humans believe they have access to an oracle that correctly predicts what happens in the real world. However, they have access to a defective oracle which only performs well in simulated worlds, but performs terribly in the “real” universe (more generally, universes in which humans are real). This is a pretty big problem with the oracle!
Yes, I agree that an oracle which is incentivized to make correct predictions within its own vantage point (including possible simulated worlds, not restricted to the real world) is malign. I don’t really agree the Solomonoff prior has this incentive. I also don’t think this is too relevant to any superintelligence we might encounter in the real world, since it is unlikely that it will have this specific incentive (this is for a variety of reasons, including “the oracle will probably care about the real world more” and, more importantly, “the oracle has no incentive to say its true beliefs anyway”).

LGS Sep 26, 2024, 2:47 AM
3 points
0
in reply to: LGS’s comment on: By Default, GPTs Think In Plain Sight
Given o1, I want to remark that the prediction in (2) was right. Instead of training LLMs to give short answers, an LLM is trained to give long answers and another LLM summarizes.

LGS Aug 31, 2024, 7:17 PM
3 points
0
in reply to: Jeremy Gillen’s comment on: Am I confused about the “malign universal prior” argument?
That’s fair, yeah
We need a proper mathematical model to study this further. I expect it to be difficult to set up because the situation is so unrealistic/impossible as to be hard to model. But if you do have a model in mind I’ll take a look

LGS Aug 31, 2024, 1:48 AM
3 points
0
in reply to: Jeremy Gillen’s comment on: Am I confused about the “malign universal prior” argument?
It would help to have a more formal model, but as far as I can tell the oracle can only narrow down its predictions of the future to the extent that those predictions are independent of the oracle’s output. That is to say, if the people in the universe ignore what the oracle says, then the oracle can give an informative prediction.
This would seem to exactly rule out any type of signal which depends on the oracle’s output, which is precisely the types of signals that nostalgebraist was concerned about.

LGS Aug 30, 2024, 9:21 PM
3 points
0
in reply to: Jeremy Gillen’s comment on: Am I confused about the “malign universal prior” argument?
The problem is that the act of leaving the message depends on the output of the oracle (otherwise you wouldn’t need the oracle at all, but you also would not know how to leave a message). If the behavior of the machine depends on the oracle’s actions, then we have to be careful with what the fixed point will be.

For example, if we try to fight the oracle and do the opposite, we get the “noise” situation from the grandfather paradox.

But if we try to cooperate with the oracle and do what it predicts, then there are many different fixed points and no telling which the oracle would choose (this is not specified in the setting).

It would be great to see a formal model of the situation. I think any model in which such message transmission would work is likely to require some heroic assumptions which don’t correspond much to real life.

LGS Aug 30, 2024, 8:55 PM
5 points
0
in reply to: Jeremy Gillen’s comment on: Am I confused about the “malign universal prior” argument?
Thanks for the link to reflective oracles!
On the gap between the computable and uncomputable: It’s not so bad to trifle a little. Diagonalization arguments can often be avoided with small changes to the setup, and a few of Paul’s papers are about doing exactly this.

I strongly disagree with this: diagonalization arguments often cannot be avoided at all, not matter how you change the setup. This is what vexed logicians in the early 20th century: no matter how you change your formal system, you won’t be able to avoid Godel’s incompleteness theorems.
There is a trick that reliably gets you out of such paradoxes, however: switch to probabilistic mixtures. This is easily seen in a game setting: in rock-paper-scissors, there is no deterministic Nash equilibrium. Switch to mixed strategies, however, and suddenly there is always a Nash equilibrium.
This is the trick that Paul is using: he is switching from deterministic Turing machines to randomized ones. That’s fine as far as it goes, but it has some weird side effects. One of them is that if a civilization is trying to predict the universal prior that is simulating itself, and tries to send a message, then it is likely that with “reflexive oracles” in place, the only message it can send is random noise. That is, Paul shows reflexive oracles exist in the same way that Nash equilibria exist; but there is no control over what the reflexive oracle actually is, and in paradoxical situations (like rock-paper-scissors) the Nash equilibrium is the boring “mix everything together uniformly”.
The underlying issue is that a universe that can predict the universal prior, which in turn simulates the universe itself, can encounter a grandfather paradox. It can see its own future by looking at the simulation, and then it can do the opposite. The grandfather paradox is where the universe decides to kill the grandfather of a child that the simulation predicts.
Paul solves this by only letting it see its own future using a “reflexive oracle” which essentially finds a fixed point (which is a probability distribution). The fixed point of a grandfather paradox is something like “half the time the simulation shows the grandchild alive, causing the real universe to kill the grandfather; the other half the time, the simulation shows the grandfather dead and the grandchild not existing”. Such a fixed point exists even when the universe tries to do the opposite of the prediction.
The thing is, this fixed point is boring! Repeat this enough times, and it eventually just says “well my prediction about your future is random noise that doesn’t have to actually come true in your own future”. I suspect that if you tried to send a message through the universal prior in this setting, the message would consist of essentially uniformly random bits. This would depend on the details of the setup, I guess.

LGS Jul 24, 2024, 2:32 AM
1 point
0
in reply to: Alex_Altair’s comment on: A simple model of math skill
I think the problem to grapple with is that I can cover the rationals in [0,1] with countably many intervals of total length only ¹⁄₂ (eg enumerate rationals in [0,1], and place interval of length ¹⁄₄ around first rational, interval of length ¹⁄₈ around the second, etc). This is not possible with reals—that’s the insight that makes measure theory work!

The covering means that the rationals in an interval cannot have a well defined length or measure which behaves reasonably under countable unions. This is a big barrier to doing probability theory. The same problem happens with ANY countable set—the reals only avoid it by being uncountable.

LGS May 15, 2024, 7:10 PM
5 points
2
in reply to: Sodium’s comment on: Ilya Sutskever and Jan Leike resign from OpenAI
Evan Morikawa?
https://twitter.com/E0M/status/1790814866695143696
What links here?
- OpenAI: Exodus by Zvi (May 20, 2024, 1:10 PM; 153 points)

LGS Apr 8, 2024, 8:40 AM
13 points
2
in reply to: DPiepgrass’s comment on: My Interview With Cade Metz on His Reporting About Slate Star Codex
Weirdly aggressive post.
I feel like maybe what’s going on here is that you do not know what’s in The Bell Curve, so you assume it is some maximally evil caricature? Whereas what’s actually in the book is exactly Scott’s position, the one you say is “his usual “learn to love scientific consensus” stance”.
If you’d stop being weird about it for just a second, could you answer something for me? What is one (1) position that Murray holds about race/IQ and Scott doesn’t? Just name a single one, I’ll wait.
Or maybe what’s going on here is that you have a strong “SCOTT GOOD” prior as well as a strong “MURRAY BAD” prior, and therefore anyone associating the two must be on an ugly smear campaign. But there’s actually zero daylight between their stances and both of them know it!

LGS Mar 29, 2024, 1:09 AM
0 points
0
in reply to: frankybegs’s comment on: My Interview With Cade Metz on His Reporting About Slate Star Codex
Relatedly, if you cannot outright make a claim because it is potentially libellous, you shouldn’t use vague insinuation to imply it to your massive and largely-unfamiliar-with-the-topic audience.
Strong disagree. If I know an important true fact, I can let people know in a way that doesn’t cause legal liability for me.
Can you grapple with the fact that the “vague insinuation” is true? Like, assuming it’s true and that Cade knows it to be true, your stance is STILL that he is not allowed to say it?
Your position seems to amount to epistemic equivalent of ‘yes, the trial was procedurally improper, and yes the prosecutor deceived the jury with misleading evidence, and no the charge can’t actually be proven beyond a reasonable doubt- but he’s probably guilty anyway, so what’s the issue’. I think the issue is journalistic malpractice. Metz has deliberately misled his audience in order to malign Scott on a charge which you agree cannot be substantiated, because of his own ideological opposition (which he admits). To paraphrase the same SSC post quoted above, he has locked himself outside of the walled garden. And you are “Andrew Cord”, arguing that we should all stop moaning because it’s probably true anyway so the tactics are justified.
It is not malpractice, because Cade had strong evidence for the factually true claim! He just didn’t print the evidence. The evidence was of the form “interview a lot of people who know Scott and decide who to trust”, which is a difficult type of evidence to put into print, even though it’s epistemologically fine (in this case IT LED TO THE CORRECT BELIEF so please give it a rest with the malpractice claims).
Here is the evidence of Scott’s actual beliefs:
https://twitter.com/ArsonAtDennys/status/1362153191102677001
As for your objections:
- First of all, this is already significantly different, more careful and qualified than what Metz implied, and that’s after we read into it more than what Scott actually said. Does that count as “aligning yourself”?
This is because Scott is giving a maximally positive spin on his own beliefs! Scott is agreeing that Cade is correct about him! Scott had every opportunity to say “actually, I disagree with Murray about...” but he didn’t, because he agrees with Murray just like Cade said. And that’s fine! I’m not even criticizing it. It doesn’t make Scott a bad person. Just please stop pretending that Cade is lying.

Relatedly, even if Scott did truly believe exactly what Charles Murray does on this topic, which again I don’t think we can fairly assume, he hasn’t said that, and that’s important. Secretly believing something is different from openly espousing it, and morally it can be much different if one believes that openly espousing it could lead to it being used in harmful ways (which from the above, Scott clearly does, even in the qualified form which he may or may not believe). Scott is going to some lengths and being very careful not to espouse it openly and without qualification, and clearly believes it would be harmful to do so, so it’s clearly dishonest and misleading to suggest that he has “aligns himself” with Charles Murray on this topic. Again, this is even after granting the very shaky proposition that he secretly does align with Charles Murray, which I think we have established is a claim that cannot be substantiated.
Scott so obviously aligns himself with Murray that I knew it before that email was leaked or Cade’s article was written, as did many other people. At some point, Scott even said that he will talk about race/IQ in the context of Jews in order to ease the public into it, and then he published this. (I can’t find where I saw Scott saying it though.)
- Further, Scott, unlike Charles Murray, is very emphatic about the fact that, whatever the answer to this question, this should not affect our thinking on important issues or our treatment of anyone. Is this important addendum not elided by the idea that he ‘aligned himself’ with Charles Murray? Would not that not be a legitimate “gripe”?
Actually, this is not unlike Charles Murray, who also says this should not affect our treatment of anyone. (I disagree with the “thinking on important issues” part, which Scott surely does think it affects.)

LGS Mar 28, 2024, 7:34 AM
8 points
5
in reply to: wilkox’s comment on: My Interview With Cade Metz on His Reporting About Slate Star Codex
The epistemology was not bad behind the scenes, it was just not presented to the readers. That is unfortunate but it is hard to write a NYT article (there are limits on how many receipts you can put in an article and some of the sources may have been off the record).

Cade correctly informed the readers that Scott is aligned with Murray on race and IQ. This is true and informative, and at the time some people here doubted it before the one email leaked. Basically, Cade’s presented evidence sucked but someone going with the heuristic “it’s in the NYT so it must be true” would have been correctly informed.

I don’t know if Cade had a history of “tabloid rhetorical tricks” but I think it is extremely unbecoming to criticize a reporter for giving true information that happens to paint the community in a bad light. Also, the post you linked by Trevor uses some tabloid rhetorical tricks: it says Cade sneered at AI risk but links to an article that literally doesn’t mention AI risk at all.

LGS Mar 27, 2024, 9:48 PM
1 point
−2
in reply to: cubefox’s comment on: My Interview With Cade Metz on His Reporting About Slate Star Codex
What you’re suggesting amounts to saying that on some topics, it is not OK to mention important people’s true views because other people find those views objectionable. And this holds even if the important people promote those views and try to convince others of them. I don’t think this is reasonable.
As a side note, it’s funny to me that you link to Against Murderism as an example of “careful subtlety”. It’s one of my least favorite articles by Scott, and while I don’t generally think Scott is racist that one almost made me change my mind. It is just a very bad article. It tries to define racism out of existence. It doesn’t even really attempt to give a good definition—Scott is a smart person, he could do MUCH better than those definitions if he tried. For example, a major part of the rationalist movement was originally about cognitive biases, yet “racism defined as cognitive bias” does not appear in the article at all. Did Scott really not think of it?

LGS Mar 27, 2024, 8:28 PM
5 points
0
in reply to: Elizabeth’s comment on: My Interview With Cade Metz on His Reporting About Slate Star Codex

What Metz did is not analogous to a straightforward accusation of cheating. Straightforward accusations are what I wish he did.
It was quite straightforward, actually. Don’t be autistic about this: anyone reasonably informed who is reading the article knows what Scott is accused of thinking when Cade mentions Murray. He doesn’t make the accusation super explicit, but (a) people here would be angrier if he did, not less angry, and (b) that might actually pose legal issues for the NYT (I’m not a lawyer).
What Cade did reflects badly on Cade in the sense that it is very embarrassing to cite such weak evidence. I would never do that because it’s mortifying to make such a weak accusation.
However, Scott has no possible gripe here. Cade’s article makes embarrassing logical leaps, but the conclusion is true and the reporting behind the article (not featured in the article) was enough to show it true, so even a claim of being Gettier Cased does not work here.

LGS Mar 27, 2024, 8:20 PM
7 points
6
in reply to: DPiepgrass’s comment on: My Interview With Cade Metz on His Reporting About Slate Star Codex
Scott thinks very highly of Murray and agrees with him on race/IQ. Pretty much any implication one could reasonably draw from Cade’s article regarding Scott’s views on Murray or on race/IQ/genes is simply factually true. Your hypothetical author in Alabama has Greta Thunberg posters in her bedroom here.