I believe this is standard/acceptable for presenting log-axis data, but I'm not sure. This is a graph from the Kaplan paper:
It is certainly frustrating that they don't label the x-axis. Here's a quick conversation where I asked GPT4o to explain. You are correct that a quick look at this graph (where you don't notice the log-scale) would imply (highly surprising and very strong) linear scaling trends. Scaling laws are generally very sub-linear, in particular often following a power-law. I don't think they tried to mislead about this, instead this is a domain where log-scaling axes is super common and doesn't invalidate the results in any way.
From the o1 blog post (evidence about the methodology for presenting results but not necessarily the same):
o1 greatly improves over GPT-4o on challenging reasoning benchmarks. Solid bars show pass@1 accuracy and the shaded region shows the performance of majority vote (consensus) with 64 samples.
What do people mean when they say that o1 and o3 have "opened up new scaling laws" and that inference-time compute will be really exciting?
The standard scaling law people talk about is for pretraining, shown in the Kaplan and Hoffman (Chinchilla) papers.
It was also the case that various post-training (i.e., finetuning) techniques improve performance, (though I don't think there is as clean of a scaling law, I'm unsure). See e.g., this paper which I just found via googling fine-tuning scaling laws. See also the Tülu 3 paper, Figure 4.
We have also already seen scaling law-type trends for inference compute, e.g., this paper:
The o1 blog post points out that they are observing two scaling trends: predictable scaling w.r.t. post-training (RL) compute, and predictable scaling w.r.t. inference compute:
The paragraph before this image says: "We have found that the performance of o1 consistently improves with more reinforcement learning (train-time compute) and with more time spent thinking (test-time compute). The constraints on scaling this approach differ substantially from those of LLM pretraining, and we are continuing to investigate them." That is, the left graph is about post-training compute.
Following from that graph on the left, the o1 paradigm gives us models that are better for a fixed inference compute budget (which is basically what it means to train a model for longer or train a better model of the same size by using better algorithms — the method is new but not the trend), and following from the right, performance seems to scale well with inference compute budget. I'm not sure there's sufficient public data to compare that graph on the right against other inference-compute scaling methods, but my guess is the returns are better.
What is o3 doing that you couldn't do by running o1 on more computers for longer?
I mean, if you replace "o1" in this sentence with "monkeys typing Shakespeare with ground truth verification," it's true, right? But o3 is actually a smarter mind in some sense, so it takes [presumably much] less inference compute to get similar performance. For instance, see this graph about o3-mini:
The performance-per-dollar frontier is pushed up by the o3-mini models. It would be somewhat interesting to know how much cost it would take for o1 to reach o3 performance here, but my guess is it's a huge amount and practically impossible. That is, there are some performance levels that are practically unobtainable for o1, the same way the monkeys won't actually complete Shakespeare.
o3 is very performant. More importantly, progress from o1 to o3 was only three months, which shows how fast progress will be in the new paradigm of RL on chain of thought to scale inference compute. Way faster than pretraining paradigm of new model every 1-2 years
o1 was the first large reasoning model — as we outlined in the original “Learning to Reason” blog, it’s “just” an LLM trained with RL. o3 is powered by further scaling up RL beyond o1, and the strength of the resulting model the resulting model is very, very impressive. (2/n)
The prices leaked by ARC-ARG people indicate $60/million output tokens, which is also the current o1 pricing. 33m total tokens and a cost of $2,012.
Notably, the codeforces graph with pricing puts o3 about 3x higher than o1 (tho maybe it's a secretly log scale), and the ARC-AGI graph has the cost of o3 being 10-20x that of o1-preview. Maybe this indicates it does a bunch more test-time reasoning. That's collaborated by ARC-AGI, average 55k tokens per solution[1], which seems like a ton.
I think this evidence indicates this is likely the same base model as o1, and I would be at like 65% sure, so not super confident.
edit to add because the phrasing is odd: this is the data being used for the estimate, and the estimate is 33m tokens / (100 tasks * 6 samples per task) = ~55k tokens per sample. I called this "solution" because I expect these are basically 6 independent attempts at answering the prompt, but somebody else might interpret things differently. The last column is "Time/Task (mins)".
I appreciate your point about compelling experimental evidence, and I think it's important that we're currently at a point with very little of that evidence. I still feel a lot of uncertainty here, and I expect the evidence to basically always be super murky and for interpretations to be varied/controversial, but I do feel more optimistic than before reading your comment.
You could find a way of proving to the world that your AI is aligned, which other labs can't replicate, giving you economic advantage.
I don't expect this to be a very large effect. It feels similar to an argument like "company A will be better on ESG dimensions and therefore more and customers will switch to using it". Doing a quick review of the literature on that, it seems like there's a small but notable change in consumer behavior for ESG-labeled products. In the AI space, it doesn't seem to me like any customers care about OpenAI's safety team disappearing (except a few folks in the AI safety world). In this particular case, I expect the technical argument needed to demonstrate that some family of AI systems are aligned while others are not is a really complicated argument; I expect fewer than 500 people would be able to actually verify such an argument (or the initial "scalable alignment solution"), maybe zero people. I realize this is a bit of a nit because you were just gesturing toward one of many ways it could be good to have an alignment solution.
I endorse arguing for alternative perspectives and appreciate you doing it. And I disagree with your synthesis here.
I agree it's plausible. I continue to think that defensive strategies are harder than offensive ones, except the ones that basically look like centralized control over AGI development. For example,
Provide compelling experimental evidence that standard training methods lead to misaligned power-seeking AI by default
Then what? The government steps in and stops other companies from scaling capabilities until big safety improvements have been made? That's centralization along many axes. Or maybe all the other key decision makers in AGI projects get convinced by evidence and reason and this buys you 1-3 years until open source / many other actors reach this level of capabilities.
Sharing an alignment solution involves companies handing over valuable IP to their competitors. I don't want to say it's impossible, but I have definitely gotten less optimistic about this in the last year. I think in the last year we have not seen a race to the top on safety, in any way. We have not seen much sharing of safety research that is relevant to products (or like, applied alignment research). We have instead mostly seen research without direct applications: interp, model organisms, weak-to-stong / scalable oversight (which is probably the closest to product relevance). Now sure, the stakes are way higher with AGI/ASI so there's a bigger incentive to share, but I don't want to be staking the future on these companies voluntarily giving up a bunch of secrets, which would be basically a 180 from their current strategy.
I fail to see how developing and sharing best practices for RSPs will shift the game board. Except insofar as it involves key insights on technical problems (e.g., alignment research that is critical for scaling) which hits the IP problem. I don't think we've seen a race to the top on making good RSPs, but we have definitely seen pressure to publish any RSP. Not enough pressure; the RSPs are quite weak IMO and some frontier AI developers (Meta, xAI, maybe various Chinese orgs count) have none.
I agree that it's plausible that "one good apple saves the bunch", but I don't think it's super likely if you condition on not centralization.
Do you believe that each of the 3 things you mentioned would change the game board? I think that they are like 75%, 30%, and 20% likely to meaningfully change catastrophic risk, conditional on happening.
Training as it's currently done needs to happen within a single cluster
I think that's probably wrong, or at least effectively wrong. Gemini 1.0, trained a year ago has the following info in the technical report:
TPUv4 accelerators are deployed in “SuperPods” of 4096 chips... TPU accelerators primarily communicate over the high speed inter-chip-interconnect, but at Gemini Ultra scale, we combine SuperPods in multiple datacenters using Google’s intra-cluster and inter-cluster network (Poutievski et al., 2022; Wetherall et al., 2023; yao Hong et al., 2018). Google’s network latencies and bandwidths are sufficient to support the commonly used synchronous training paradigm, exploiting model parallelism within superpods and data-parallelism across superpods.
As you note, public distributed training methods have advanced beyond basic data parallelism (though they have not been publicly shown at large model scales because nobody has really tried yet).
While writing, I realized that this sounds a bit similar to the unilateralist's curse. It's not the same, but it has parallels. I'll discuss that briefly because it's relevant to other aspects of the situation. The unilateralist's curse does not occur specifically due to multiple samplings, it occurs because different actors have different beliefs about the value/disvalue, and this variance in beliefs makes it more likely that one of those actors has a belief above the "do it" threshold. If each draw from the AGI urn had the same outcome, this would look a lot like a unilateralist's curse situation where we care about variance in the actors' beliefs. But I instead think that draws from the AGI urn are somewhat independent and the problem is just that we should incur e.g., a 5% misalignment risk as few times as we have to.
Interestingly, a similar look at variance is part of what makes the infosecurity situation much worse for multiple projects compared to centralized AGI project: variance is bad here. I expect a single government AGI project to care about and invest in security at least as much as the average AGI company. The AGI companies have some variance in their caring and investment in security, and the lower ones will be easier to steal from. If you assume these multiple projects have similar AGI capabilities (this is a bad assumption but is basically the reason to like multiple projects for Power Concentration reasons so worth assuming here; if the different projects don't have similar capabilities, power is not very balanced), you might then think that any of the companies getting their models stolen is similarly bad to the centralized project getting its models stolen (with a time lag I suppose, because the centralized project got to that level of capability faster).
If you are hacking a centralized AGI project, say you have a 50% chance of success. If you are hacking 3 different AGI projects, you have 3 different/independent 50% chances of success. They're different because these project have different security measures in place. Now sure, as indicated by one of the points in this blog post, maybe less effort goes into hacking each of the 3 projects (because you have to split your resources, and because there's less overall interest in stealing model weights), maybe that pushes each of these down to 33%. These numbers are obviously made up, and they would get to a 1 – (0.67^3) = 70% chance of success.
Unilateralist's curse is about variance in beliefs about the value of some action. The parent comment is about taking multiple independent actions that each have a risk of very bad outcomes.
Thanks for writing this, I think it's an important topic which deserves more attention. This post covers many arguments, a few of which I think are much weaker than you all state. But more importantly, I think you all are missing at least one important argument. I've been meaning to write this up, and I'll use this as my excuse.
TL;DR: More independent AGI efforts means more risky “draws” from a pool of potential good and bad AIs; since a single bad draw could be catastrophic (a key claim about offense/defense), we need fewer, more controlled projects to minimize that danger.
The argument is basically an application of the Vulnerability World Hypothesis to AI development. You capture part of this argument in the discussion of Racing, but not the whole thing. So the setup is that building any particular AGI is drawing a ball from the urn of potential AIs. Some of these AIs are aligned, some are misaligned — we probably disagree about the proportions here but that's not crucial, and note that the proportion depends on a bunch of other aspects about the world such as how good our AGI alignment research is. More AGI projects means more draws from the urn and a higher likelihood of pulling out misaligned AI systems. Importantly, I think that pulling out a misaligned AGI system is more bad than pulling out an aligned AGI system is good. I think this because I think some of the key components about the world that are offense-favored.
Key assumption/claim: human extinction and human loss of control are offense-favored — if there were similarly resourced actors trying to destroy humanity as to protect it, humanity would be destroyed. I have a bunch of intuitions for why this is true, to give some sense:
Humans are flesh bags that die easily and never come back to life. AIs will not be like this.
Humans care a lot about not dying, their friends and families not dying, etc., I expect extorting a small number of humans in order to gain control would simply work if one could successfully make the relevant threats.
Terrorists or others who seek to cause harm often succeed. There are many mass shootings. 8% of US presidents were assassinated in office. I don't actually know what the average death count per attempted terrorist is; I would intuitively guess it's between 0.5 and 10 (This Wikipedia article indicates it's ~10, but I think you should include attempts that totally fail, even though these are not typically counted). Terrorism is very heavy tailed, which I think probably means that more capable terrorists (i.e., AIs that are at least as good as human experts, AGI+) will have high fatality rates.
There are some emerging technologies that so far seem more offense-favored to me. Maybe not 1000:1, but definitely not 1:1. Bio tech and engineered pandemics seem like this; autonomous weapons seem like this.
The strategy-stealing assumption seems false to me, partially for reasons listed in the linked post. I note that the linked post includes Paul listing a bunch of convincing-to-me ways in which strategy-stealing is false and then concluding that it's basically true. The claim about offense is easier than defense is sorta just a version of the strategy stealing claim, this bullet point isn't actually another distinct argument, just an excuse to point toward previous thinking and the various arguments there.
A couple caveats: I think killing all of humanity with current tech is pretty hard; as noted however, I think this is too high a bar because probably things like extortion are sufficient for grabbing power. Also, I think there are some defensive strategies that would actually totally work at reducing the threat from misaligned AGI systems. Most of these strategies look a lot like "centralization of AGI development", e.g., destroying advanced computing infrastructure, controlling who uses advanced computing infrastructure and how they use it, a global treaty banning advanced AI development (which might be democratically controlled but has the effect of exercising central decision making).
So circling back to the urn, if you pull out an aligned AI system, and 3 months later somebody else pulls out a misaligned AI system, I don't think pulling out the aligned AI system a little in advance buys you that much. The correct strategy to this situation is to try and make the proportion of balls weighted heavily toward aligned, AND to pull out as few as you can.
More AGI development projects means more draws from the urn because there are more actors doing this and no coordinated decision process to stop. You mention that maybe government can regulate AI developers to reduce racing. This seems like it will go poorly, and in the worlds where it goes well, I think you should maybe just call them "centralization" because they involve a central decision process deciding who can train what models when with what methods. That is, extremely involved regulations seem to effectively be centralization.
Notably, this is related but not the same as the effects from racing. More AGI projects leads to racing which leads to cutting corners on safety (higher proportion of misaligned AIs in the urn), and racing leads to more draws from the urn because of fear of losing to a competitor. But even without racing, more AGI projects means more draws from the urn.
The thing I would like to happen instead is that there is a very controlled process for drawing from the urn, where each ball is carefully inspected, and if we draw aligned AIs, we use them to do AI alignment research, i.e., increase the proportion of aligned AIs in the urn. And we don't take more draws from the urn until we're really sure we're quite confident we're not going to pull out a misaligned AI. Again, this is both about reducing the risk of catastrophe each time you take a risky action, and about decreasing the number of times you have to take risky actions.
Summarizing: If you are operating in a domain where losses are very bad, you want to take less gambles. I think AGI and ASI development are such domains, and decentralized AGI development means more gambles are taken.
Noting that I spent a couple minutes pondering the quoted passage which I don't think was a good use of time (I basically would have immediately dismissed it if I knew Claude wrote it, and I only thought about it because my prior on Buck saying true things is way higher), and I would have preferred the text not have this.
I don't see anybody having mentioned it yet, but the recent paper about LLM Introspection seems pretty relevant. I would say that a model which performs very well at introspection (as defined there) would be able to effectively guess which jailbreak strategies were attempted.
There is now some work in that direction: https://forum.effectivealtruism.org/posts/47RH47AyLnHqCQRCD/soft-nationalization-how-the-us-government-will-control-ai
Some prompts I found interesting when brainstorming LLM startups
I spent a little time thinking about making an AI startup. I generally think it would be great if more people were trying to build useful companies that directly add value, rather than racing to build AGI. Here are some of the prompts I found interesting to think about, perhaps they will be useful to other people/AI agents interested in building a startup:
What are the situations where people will benefit from easy and cheap access to expert knowledge? You’re leveraging that human expert labor is hard to scale to many situations (especially when experts are rare, needs are specific, it’s awkward, it’s too expensive — including both raw cost and the cost of finding/trusting/onboarding an expert). What are all the things you occasionally pay somebody to do, but which requires them coming in person? What is a problem people know they have but they don’t seek out existing solutions (because of perceived cost, awkwardness, unsure how). e.g., dating profile feedback, outfit designer.
Solve a problem that exists due to technological development, e.g., preventing the social isolation from social media, reducing various catastrophic risks during and after intelligence explosion.
Some other problem attack surface opened up by LLMs:
Cheaply carry out simple straightforward tasks.
Analyze data at scale.
Do tasks that there was no previous market for (e.g., provided $5 of value but took an hour, and you can’t hire people for $5/hour because they don’t want to work for that little and the overhead is high). Reasons for lack of market: not enough money to be made, can’t trust somebody (not worth the time needed to grow trust, or substantial privacy concerns), communication cost too high (specify task), other overhead too high (travel, finding person), training cost too high compared to salary (imagine it took 8 years to become a barber).
Provide cheap second opinions, potentially many of them (e.g., reviewing a low-importance piece of writing).
Some other desiderata I had (for prompting LLMs):
I want to have a clear and direct story for making people's lives better or solving problems they have. So I have a slight preference for B2C over B2B, unless there's a clear story for how we're significantly helping the business in an industry that benefits people.
We don't want to be obsoleted by the predictable products coming out of AI development companies; for instance a product that just takes ChatGPT and adds a convenient voice feature is not a good idea because that niche is likely to be met by existing developers fairly soon.
We don't want to work on something that other well resourced efforts are working on. Our edge is having good ideas and creative implementations, not being able to outcompete others according to resource investment. We should play to our strengths and not try to get in a losing battle with strong existing products.
I mainly don't want to be directly competing with existing products or services, instead I want to be creating a large amount of counterfactual value by solving a problem that nobody else has solved.
The MVP should be achievable by a team of 5 working for <6 months, ideally even a very basic MVP should be achievable in just a week or two of full-time work.
I want to be realistic, we won't be able to solve everything or do everything. I want to aim for a fairly niche product, rather than solving a huge-scale problem like fixing medical care. That is, instead of a general medical chatbot, a better idea would be a first-aid tutor that can help people learn first-aid basics and refresh their knowledge later.
I want to be providing a service people are excited to receive. For instance, a sustainable living advisor isn't a great idea because if it actually got people to make more sustainable decisions, that would be annoying — people don't actually want to hear that they shouldn't fly home to their family for the holidays, even though this is one of the more important sustainability decisions they could make.
I probably want to provide a service that is not currently provided by a simple google search. For instance, a cooking assistant is pretty much just glorified google search. I want to be providing more value than that. Services which can be provided by a simple google search are likely to be filled in by existing developers.
I do not want to be pushing the frontier of AI capabilities in dangerous domains such as: synthetic biology, cybersecurity, autonomous weapons, ML engineering and AI development, manipulation and persuasion. Generally pushing the frontier of scientific capabilities may also be in this group due to its effects on the other domains, but it is unclear.
Yep, I basically agree with those being the cruxes!
On how much of the goal reasoning is happening out loud: Nowadays, I think about a lot of AI safety research as being aimed at an AI Control scenario where we are closely supervising what models are thinking about in CoT, and thus malign goal seeking must either happen in individual forward passes (and translated to look harmless most the time) or in an obfuscated way in CoT. (or from rogue deployments outside the control scheme)
Studying the intrinsic goals of the models in naturalistic settings would also be interesting, but I expect this to be hard, and also less relevant to the threat models we're worried about. I'm mainly concerned about models becoming more goal-directed with scale and someone figuring out long-horizon RL training on top of that, leading to models that are qualitatively different from current ones.
By naturalistic, I mean "from a realistic training process, even if that training is designed to create goals". Which sounds like what you said is the main threat model you're worried about? If you have the time, I would push you harder on this: what is a specific story of AI catastrophe that you are trying to study/prevent?
Some quick thoughts (only skimmed the post, writing quickly), as you asked for feedback:
It looks like the main thing you're testing is some variant of "when prompted to do goal directed behavior, how effective is the model at satisfying the goal?" That's a reasonable thing to investigate, but I'm not sure it would be near the top of the list of "empirical research on goal-directed-ness that I want to see". I'm probably mainly interested in the deceptive alignment motivation, read the rest of this comment as focusing on that.
Aside: To state it directly, I think the main reason to study goal-directedness in this lower-validity setting (of giving models goals in prompts) is that CoT-based goal-directedness might act as a precursor for in-forward-pass goal directedness (which seems far more worrying re deceptive alignment) — so we can study it earlier. So again, reasonable to study, but if you agree with me that this is the main reason for such experiments being valid, it's an important frame to have when thinking about this kind of work: artificially inducing goal directedness is a model-organism approach rather than a natural experiment.
Thinking out loud, a list for goal-directedness work I want to see might be; sub-bullets are more detailed ideas:
Are base models goal directed? Are RLHF finetuned models goal directed? (naturalistic setting)
Could look like this recent work on the consistency of model answers to values questions, but more adapted to goals you hypothesize the models to have (like how consistently do models follow a particular behavior outlined in the Model Spec)
How do RLHFed models deal with conflicting goals — do they engage in sophisticated reasoning about this or instead seem to follow simple heuristics?
To the extent these models are goal directed (including because you induce this via prompting), is anything interesting going on:
Do they goal generalize the way we would expect? Similar to this recent work, but aimed at some risk other than reward hacking, I would be particularly interested in the time horizon over which the world is affected, as this is perhaps a proxy for a model having non-myopic goals.
As mentioned, maybe CoT and prompting provide an early warning sign for forward-pass goal-directed-ness. Is this true? How much of an early warning sign?
It looks like the settings in this post are sorta a general capability eval for a model accomplishing goals. I wonder if you think they add a ton of value over existing agent benchmarks like SWE-Bench? My intuition says you would be better off on trying to focus in on a narrower question that is particularly relevant to safety, like one of those I mentioned.
Sorry if this comment was rude or mean, it's been a couple weeks and this post has no feedback even though you asked, I figured something might be better than nothing. It looks to me like your overall approach and ways of thinking about this are good!
What's the evidence that this document is real / written by Anthropic?
This sentence seems particularly concerning:
We believe the first two issues can be addressed by focusing on deterrence rather than pre-harm enforcement: instead of deciding what measures companies should take to prevent catastrophes (which are still hypothetical and where the ecosystem is still iterating to determine best practices), focus the bill on holding companies responsible for causing actual catastrophes.
Nice work, these seem like interesting and useful results!
High level question/comment which might be totally off: one benefit of having a single, large, SAE neuron space that each token gets projected into is that features don't get in each other's way, except insofar as you're imposing sparsity. Like, your "I'm inside a parenthetical" and your "I'm attempting a coup" features will both activate in the SAE hidden layer, as long as they're in the top k features (for some sparsity). But introducing switch SAEs breaks that: if these two features are in different experts, only one of them will activate in the SAE hidden layer (based on whatever your gating learned).
The obvious reply is "but look at the empirical results you fool! The switch SAEs are pretty good!" And that's fair. I weakly expect what is happening in your experiment is that similar but slightly specialized features are being learned by each expert (a testable hypothesis), and maybe you get enough of this redundancy that it's fine e.g,. the expert with "I'm inside a parenthetical" also has a "Words relevant to coups" feature and this is enough signal for coup detection in that expert.
Again, maybe this worry is totally off or I'm misunderstanding something.
Leaving Dangling Questions in your Critique is Bad Faith
Note: I’m trying to explain an argumentative move that I find annoying and sometimes make myself; this explanation isn’t very good, unfortunately.
Example
Them: This effective altruism thing seems really fraught. How can you even compare two interventions that are so different from one another?
Explanation of Example
I think the way the speaker poses the above question is not as a stepping stone for actually answering the question, it’s simply as a way to cast doubt on effective altruists. My response is basically, “wait, you’re just going to ask that question and then move on?! The answer really fucking matters! Lives are at stake! You are clearly so deeply unserious about the project of doing lots of good, such that you can pose these massively important questions and then spend less than 30 seconds trying to figure out the answer.” I think I might take these critics more seriously if they took themselves more seriously.
Description of Dangling Questions
A common move I see people make when arguing or criticizing something is to pose a question that they think the original thing has answered incorrectly or is not trying sufficiently hard to answer. But then they kinda just stop there. The implicit argument is something like “The original thing didn’t answer this question sufficiently, and answering this question sufficiently is necessary for the original thing to be right.”
But importantly, the criticisms usually don’t actually argue that — they don’t argue for some alternative answer to the original questions, if they do they usually aren’t compelling, and they also don’t really try to argue that this question is so fundamental either.
One issue with Dangling Questions is that they focus the subsequent conversation on a subtopic that may not be a crux for either party, and this probably makes the subsequent conversation less useful.
Example
Me: I think LLMs might scale to AGI.
Friend: I don’t think LLMs are actually doing planning, and that seems like a major bottleneck to them scaling to AGI.
Me: What do you mean by planning? How would you know if LLMs were doing it?
Friend: Uh…idk
Explanation of Example
I think I’m basically shifting the argumentative burden onto my friend when it falls on both of us. I don’t have a good definition of planning or a way to falsify whether LLMs can do it — and that’s a hole in my beliefs just as it is a hole in theirs. And sure, I’m somewhat interested in what they say in response, but I don’t expect them to actually give a satisfying answer here. I’m posing a question I have no intention of answering myself and implying it’s important for the overall claim of LLMs scaling to AGI (my friend said it was important for their beliefs, but I’m not sure it’s actually important for mine). That seems like a pretty epistemically lame thing to do.
Traits of “Dangling Questions”
They are used in a way that implies the target thing is wrong vis a vis the original idea, but this argument is not made convincingly.
The author makes minimal effort to answer the question with an alternative. Usually they simply pose it. The author does not seem to care very much about having the correct answer to the question.
The author usually implies that this question is particularly important for the overall thing being criticized, but does not usually make this case.
These questions share a lot in common with the paradigm criticisms discussed in Criticism Of Criticism Of Criticism, but I think they are distinct in that they can be quite narrow.
One of the main things these questions seem to do is raise the reader’s uncertainty about the core thing being criticized, similar to the Just Asking Questions phenomenon. To me, Dangling Questions seem like a more intellectual version of Just Asking Questions — much more easily disguised as a good argument.
Research on scaling laws shows that as we increase model size, training compute, and dataset size, language models get “better”. … But this is a complete misinterpretation of scaling laws. What exactly is a “better” model? Scaling laws only quantify the decrease in perplexity, that is, improvement in how well models can predict the next word in a sequence. Of course, perplexity is more or less irrelevant to end users — what matters is “emergent abilities”, that is, models’ tendency to acquire new capabilities as size increases.
Explanation of Example
The argument being implied is something like “scaling laws are only about perplexity, but perplexity is different from the metric we actually care about — how much? who knows? —, so you should ignore everything related to perplexity, also consider going on a philosophical side-quest to figure out what ‘better’ really means. We think ‘better’ is about emergent abilities, and because they’re emergent we can’t predict them so who knows if they will continue to appear as we scale up”. In this case, the authors have ventured an answer to their Dangling Question, “what is a ‘better’ model?“, they’ve said it’s one with more emergent capabilities than a previous model. This answer seems flat out wrong to me; acceptable answers include: downstream performance, self-reported usefulness to users, how much labor-time it could save when integrated in various people’s work, ability to automate 2022 job tasks, being more accurate on factual questions, and much more. I basically expect nobody to answer the question “what does it mean for one AI system to be better than another?” with “the second has more capabilities that were difficult to predict based on the performance of smaller models and seem to increase suddenly on a linear-performance, log-compute plot”.
Even given the answer “emergent abilities”, the authors fail to actually argue that we don’t have a scaling precedent for these. Again, I think the focus on emergent abilities is misdirected, so I’ll instead discuss the relationship between perplexity and downstream benchmark performance — I think this is fair game because this is a legitimate answer to the “what counts as ‘better’?” question and because of the original line “Scaling laws only quantify the decrease in perplexity, that is, improvement in how well models can predict the next word in a sequence”. The quoted thing is technically true but in this context highly misleading, because we can, in turn, draw clear relationships between perplexity and downstream benchmark performance; here are threerecentpapers which do so, here are evenmorestudiesthat relate compute directly to downstream performance on non-perplexity metrics. Note that some of these are cited in the blog post. I will also note that this seems like one example of a failure I’ve seen a few times where people conflate “scaling laws” with what I would refer to as “scaling trends” where the scaling laws refer to specific equations for predicting various metrics based on model inputs such as # parameters and amount of data to predict perplexity, whereas scaling trends are the more general phenomenon we observe that scaling up just seems to work and in somewhat predictable ways; the scaling laws are useful for the predicting, but whether we have those specific equations or not has no effect on this trend we are observing, the equations just yield a bit more precision. Yes, scaling laws relating parameters and data to perplexity or training loss do not directly give you info about downstream performance, but we seem to be making decent progress on the (imo still not totally solved) problem of relating perplexity to downstream performance, and together these mean we have somewhat predictable scaling trends for metrics that do matter.
Example
Here’s another example from that blog post where the authors don’t literally pose a question, but they are still doing the Dangling Question thing in many ways. (context is referring to these posts):
Also, like many AI boosters, he conflates benchmark performance with real-world usefulness.
Explanation of Example
(Perhaps it would be better to respond to the linked AI Snake Oil piece, but that’s a year old and lacks lots of important evidence we have now). I view the move being made here as posing the question “but are benchmarks actually useful to real world impact?“, assuming the answer is no — or poorly arguing so in the linked piece — and going on about your day. It’s obviously the case that benchmarks are not the exact same as real world usefulness, but the question of how closely they’re related isn’t some magic black box of un-solvability! If the authors of this critique want to complain about the conflation between benchmark performance and real-world usefulness, they should actually bring the receipts showing that these are not related constructs and that relying on benchmarks would lead us astray. I think when you actually try that, you get an answer like: benchmark scores seem worse than user’s reported experience and than user’s reported usefulness in real world applications, but there is certainly a positive correlation here; we can explain some of the gap via techniques like few-shot prompting that are often used for benchmarks, a small amount via dataset contamination, and probably much of this gap comes from a validity gap where benchmarks are easy to assess but unrealistic, but thankfully we have user-based evaluations like LMSYS that show a solid correlation between benchmark scores and user experience, … (if I actually wanted to make the argument the authors were, I would be spending like >5 paragraphs on it and elaborating on all of the evidences mentioned above, including talking more about real world impacts, this is actually a difficult question and the above answer is demonstrative rather than exemplar)
Caveats and Potential Solutions
There is room for questions in critiques. Perfect need not be the enemy of good when making a critique. Dangling Questions are not always made in bad faith.
Many of the people who pose Dangling Questions like this are not trying to act in bad faith. Sometimes they are just unserious about the overall question, and they don’t care much about getting to the right answer. Sometimes Dangling Questions are a response to being confused and not having tons of time to think through all the arguments, e.g., they’re a psychological response something like “a lot feels wrong about this, here are some questions that hint at what feels wrong to me, but I can’t clearly articulate it all because that’s hard and I’m not going to put in the effort”.
My guess at a mental move which could help here: when you find yourself posing a question in the context of an argument, ask whether you care about the answer, ask whether you should spend a few minutes trying to determine the answer, ask whether the answer to this question would shift your beliefs about the overall argument, ask whether the question puts undue burden on your interlocutor.
If you’re thinking quickly and aren’t hoping to construct a super solid argument, it’s fine to have Dangling Questions, but if your goal is to convince others of your position, you should try to answer your key questions, and you should justify why they matter to the overall argument.
Another example of me posing a Dangling Question in this:
What happens to OpenAI if GPT-5 or the ~5b training run isn't much better than GPT-4? Who would be willing to invest the money to continue? It seems like OpenAI either dissolves or gets acquired.
Explanation of Example
(I’m not sure equating GPT-5 with a ~5b training run is right). In the above quote, I’m arguing against The Scaling Picture by asking whether anybody will keep investing money if we see only marginal gains after the next (public) compute jump. I think I spent very little time trying to answer this question, and that was lame (though acceptable given this was a Quick Take and not trying to be a strong argument). I think for an argument around this to actually go through, I should argue: without much larger dollar investments, The Scaling Picture won’t hold; those dollar investments are unlikely conditional on GPT-5 not being much better than GPT-4. I won’t try to argue these in depth, but I do think some compelling evidence is that OpenAI is rumored to be at ~$3.5 billion annualized revenue, and this plausibly justifies considerable investment even if the GPT-5 gain over this isn’t tremendous.
I agree that repeated training will change the picture somewhat. One thing I find quite nice about the linked Epoch paper is that the range of tokens is an order of magnitude, and even though many people have ideas for getting more data (common things I hear include "use private platform data like messaging apps"), most of these don't change the picture because they don't move things more than an order of magnitude, and the scaling trends want more orders of magnitude, not merely 2x.
Repeated data is the type of thing that plausibly adds an order of magnitude or maybe more.
I sometimes want to point at a concept that I've started calling The Scaling Picture. While it's been discussed at length (e.g., here, here, here), I wanted to give a shot at writing a short version:
The picture:
We see improving AI capabilities as we scale up compute, projecting the last few years of progress in LLMs forward might give us AGI (transformative economic/political/etc. impact similar to the industrial revolution; AI that is roughly human-level or better on almost all intellectual tasks) later this decade (note: the picture is not about specific capabilities so much as the general picture).
Relevant/important downstream capabilities improve as we scale up pre-training compute (size of model and amount of data), although for some metrics there are very sublinear returns — this is the current trend. Therefore, you can expect somewhat predictable capability gains in the next few years as we scale up spending (increase compute), and develop better algorithms / efficiencies.
AI capabilities in the deep learning era are the result of three inputs: data, compute, algorithms. Keeping algorithms the same, and scaling up the others, we get better performance — that's what scaling means. We can lump progress in data and algorithms together under the banner "algorithmic progress" (i.e., how much intelligence can you get per compute) and then to some extent we can differentiate the source of progress: algorithmic progress is primarily driven by human researchers, while compute progress is primarily driven by spending more money to buy/rent GPUs. (this may change in the future). In the last few years of AI history, we have seen massive gains in both of these areas: it's estimated that the efficiency of algorithms has improved about 3x/year, and the amount of compute used has increased 4.1x/year. These are ludicrous speeds relative to most things in the world.
Edit to add: This paper seems like it might explain that breakdown better.
Edit to add: The below arguments are just supposed to be pointers toward longer argument one could make, but the one sentence version usually isn't compelling on its own.
Arguments for:
Scaling laws (mathematically predictable relationship between pretraining compute and perplexity) have held for ~12 orders of magnitude already
We are moving though ‘orders of magnitude of compute’ quickly, so lots of probability mass should be soon (this argument is more involved, following from having uncertainty over orders of magnitude of compute that might be necessary for AGI, like the approach taken here; see here for discussion)
Once you get AIs that can speed up AI progress meaningfully, progress on algorithms could go much faster, e.g., by AIs automating the role of researchers at OpenAI. You also get compounding economic returns that allow compute to grow — AIs that can be used to make a bunch of money, and that money can be put into compute. It seems plausible that you can get to that level of AI capabilities in the next few orders of magnitude, e.g., GPT-5 or GPT-6. Automated researchers are crazy.
Moore’s law has held for a long time. Edit to add: I think a reasonable breakdown for the "compute" category mentioned above is "money spent" and "FLOP purchasable per dollar". While Moore's Law is technically about the density of transistors, the thing we likely care more about is FLOP/$, which follows similar trends.
Many people at AGI companies think this picture is right, see e.g., this, this, this (can’t find an aggregation)
Arguments against:
Might run out of data. There are estimated to be 100T-1000T internet tokens, we will likely hit this level in a couple years.
Might run out of money — we’ve seen ~$100m training runs, we’re likely at $100m-1b this year, tech R&D budgets are ~30B, governments could fund $1T. One way to avoid this 'running out of money' problem is if you get AIs that speed up algorithmic progress sufficiently.
Scaling up is a non-trivial engineering problem and it might cause slow downs due to e.g., GPU failure and difficulty parallelizing across thousands of GPUs
Revenue might just not be that big and investors might decide it's not worth the high costs
OTOH, automating jobs is a big deal if you can get it working
Marginal improvements (maybe) for huge increased costs; bad ROI.
There are numerous other economics arguments against, mainly arguing that huge investments in AI will not be sustainable, see e.g., here
Maybe LLMs are missing some crucial thing
Not doing true generalisation to novel tasks in the ARC-AGI benchmark
Not able to learn on the fly — maybe long context windows or other improvements can help
For jailbreaking you are trying to learn the policy "Always imitate/generate-from a harmless assistant", here you are trying to learn "Always imitate safe human". In both, your model has some prior for outputting harmful next tokens, the context provides an update toward a higher probability of outputting harmful text (because of seeing previous examples of the assistant doing so, or because the previous generations came from an AI). And in both cases we would like some training technique that causes the model's posterior on harmful next tokens to be low.
I'm not sure there's too much else of note about this similarity, but it seemed worth noting because maybe progress on one can help with the other.
Cool! I'm not very familiar with the paper so I don't have direct feedback on the content — seems good. But I do think I would have preferred a section at the end with your commentary / critiques of the paper, also that's potentially a good place to try and connect the paper to ideas in AI safety.
It looks like the example you gave pretty explicitly is using “compute” rather than “effective compute”. The point of having the “effective” part is to take into account non compute progress, such as using more optimal N/D ratios. I think in your example, the first two models would be at the same effective compute level, based on us predicting the same performance.
That said, I haven’t seen any detailed descriptions of how Anthropic is actually measuring/calculating effective compute (iirc they link to a couple papers and the main theme is that you can use training CE loss as a predictor).
Claude 3.5 Sonnet solves 64% of problems on an internal agentic coding evaluation, compared to 38% for Claude 3 Opus. Our evaluation tests a model’s ability to understand an open source codebase and implement a pull request, such as a bug fix or new feature, given a natural language description of the desired improvement.
...
While Claude 3.5 Sonnet represents an improvement in capabilities over our previously released Opus model, it does not trigger the 4x effective compute threshold at which we will run the full evaluation protocol described in our Responsible Scaling Policy (RSP).
Hmmm, maybe the 4x effective compute threshold is too large given that you're getting near doubling of agentic task performance (on what I think is an eval with particularly good validity) but not hitting the threshold.
Or maybe at the very least you should make some falsifiable predictions that might cause you to change this threshold. e.g., "If we train a model that has downstream performance (on any of some DC evals) ≥10% higher than was predicted by our primary prediction metric, we will revisit our prediction model and evaluation threshold."
It is unknown to me whether Sonnet 3.5's performance on this agentic coding evaluation was predicted in advance at Anthropic. It seems wild to me that you can double your performance on a high validity ARA-relevant evaluation without triggering the "must evaluate" threshold; I think evaluation should probably be required in that case, and therefore, if I had written the 4x threshold, I would be reducing it. But maybe those who wrote the threshold were totally game for these sorts of capability jumps?
Can you say more about why you would want this to exist? Is it just that "do auto-interpretability well" is a close proxy for "model could be used to help with safety research"? Or are you also thinking about deception / sandbagging, or other considerations.
Nice! Do you have a sense of the total development (and run-time) cost of your solution? "Actually getting to 50% with this main idea took me about 6 days of work." I'm interested in the person-hours and API calls cost of this.
Hm, can you explain what you mean? My initial reaction is that AI oversight doesn't actually look a ton like this position of the interior where defenders must defend every conceivable attack whereas attackers need only find one successful strategy. A large chunk of why I think these are disanalogous is that getting caught is actually pretty bad for AIs — see here.
Not sure I love this analogy — moving to NYC doesn't seem like that big of a deal —, but I do think it's pretty messed up to be imposing huge social / technological / societal changes on 8 billion of your peers. I expect most of the people building AGI have not really grasped the ethical magnitude of doing this — I think I sort of have, but also I don't build AGI.
Note on something from the superalignment section of Leopold Aschenbrenner's recent blog posts:
Evaluation is easier than generation. We get some of the way “for free,” because it’s easier for us to evaluate outputs (especially for egregious misbehaviors) than it is to generate them ourselves. For example, it takes me months or years of hard work to write a paper, but only a couple hours to tell if a paper someone has written is any good (though perhaps longer to catch fraud). We’ll have teams of expert humans spend a lot of time evaluating every RLHF example, and they’ll be able to “thumbs down” a lot of misbehavior even if the AI system is somewhat smarter than them. That said, this will only take us so far (GPT-2 or even GPT-3 couldn’t detect nefarious GPT-4 reliably, even though evaluation is easier than generation!)
Disagree about papers. I don’t think it takes merely a couple hours to tell if a paper is any good. In some cases it does, but in other cases, entire fields have been led astray for years due to bad science (e.g., replication crisis in psych, where numerous papers spurred tons of follow up work on fake things; a year and dozens of papers later we still don’t know if DPO is better than PPO for frontier AI development (though perhaps this is known in labs, and my guess is some people would argue this question is answered); IIRC it took like 4-8 months for the alignment community to decide CCS was bad (this is a contentious and oversimplifying take), despite many people reading the original paper). Properly vetting a paper in the way you will want to do for automated alignment research, especially if you’re excluding fraud from your analysis, is about knowing whether the insights in the paper will be useful in the future, it’s not just checking if they use reasonable hyperparameters on their baseline comparisons.
One counterpoint: it might be fine to have some work you mistakenly think is good, as long as it’s not existential-security-critical and you have many research directions being explored in parallel. That is, because you can run tons of your AIs at once, they can explore tons of research directions and do a bunch of the follow-up work that is needed to see if an insight is important. There may not be a huge penalty for having a slightly poor training signal, as long as it can get the quality of outputs good enough.
This [how easily can you evaluate a paper] is a tough question to answer — I would expect Leopold’s thoughts here to dominated by times he has read shitty papers, rightly concluded they are shitty, and patted himself on the back for his paper-critique skills — I know I do this. But I don’t expect being able to differentiate shitty vs. (okay + good + great) is enough. At a meta level, this post is yet another claim that "evaluation is easier than generation" will be pretty useful for automating alignment — I have grumbled about this before (though can't find anything I've finished writing up), and this is yet another largely-unsubstantiated claim in that direction. There is a big difference between the claims "because evaluation is generally easier than generation, evaluating automated alignment research will be a non-zero amount easier than generating it ourselves" and "the evaluation-generation advantage will be enough to significantly change our ability to automate alignment research and is thus a meaningful input into believing in the success of an automated alignment plan"; the first is very likely true, but the second maybe not.
On another note, the line “We’ll have teams of expert humans spend a lot of time evaluating every RLHF example” seems absurd. It feels a lot like how people used to say “we will keep the AI in a nice sandboxed environment”, and now most user-facing AI products have a bunch of tools and such. It sounds like an unrealistic safety dream. This also sounds terribly inefficient — it would only work if your model is very sample-efficiently learning from few examples — which is a particular bet I’m not confident in. And my god, the opportunity cost of having your $300k engineers label a bunch of complicated data! It looks to me like what labs are doing for self play (I think my view is based on papers out of meta and GDM) is having some automated verification like code passing unit tests, and using a ton of examples. If you are going to come around saying they’re going to pivot from ~free automated grading to using top engineers for this, the burden of proof is clearly on you, and the prior isn’t so good.
AIs that do ARA will need to be operating at the fringes of human society, constantly fighting off the mitigations that humans are using to try to detect them and shut them down
Why do you think this? What is the general story you're expecting?
I think it's plausible that humanity takes a very cautious response to AI autonomy, including hunting and shutting down all autonomous AIs — but I don't think the arguments I'm considering justify more than like 70% confidence (I think I'm somewhere around 60%). Some arguments pointing toward "maybe we won't respond sensibly to ARA":
There are not known-to-me laws prohibiting autonomous AIs from existing (assuming they're otherwise following laws), in any jurisdiction.
Properly dealing with ARA is a global problem, requiring either buy-in from dozens of countries, or somebody to carry out cyber-offensive operations in foreign countries, in order to shut down ARA models. We see precedence for this kind of international action w.r.t. WMD threats like US/Israel's attacks on Iran's nuclear program, and I expect there's a lot of tit-for-tat going on in the nation state hacking world, but it's not obvious that autonomous AIs would rise to a threat level that warrants this.
It's not clear to me that the public cares about autonomous AIs existing in many domains (at least in many domains; there are some domains like dating where people have a real ick). I think if we got credible evidence that Mark Zuckerberg was a lizard or a robot, few people would stop using Facebook products as a result. Many people seem to think various tech CEOs like Elon Musk and Jeff Bezos are terrible, yet still use their products.
A lot of this seems like it depends on whether autonomous AIs actually cause any serious harm. I can definitely imagine a world with autonomous AIs running around like small companies and twitter being filled with "but show me the empirical evidence for risk, all you safety-ists have is your theoretical arguments which haven't held up, and we have tons of historical evidence of small companies not causing catastrophic harm". And indeed, I don't really expect the conceptual arguments for risk from roughly human level autonomous AIs are likely to convince enough of the public + policy makers that they need to take drastic actions to limit autonomous AIs; I definitely wouldn't be highly confident that will will respond appropriately in the absence of serious harm. If the autonomous AIs are basically minding their own business, I'm not sure there will be major effort to limit them.
I appreciate this post. Emphasizing a couple things and providing some other commentary/questions on the paper (as there doesn't seem to be a better top level post for it) (I have not read paper deeply and could be missing things):
I find the Twitter vote brigading to be annoying and slightly bad for collective epistemics. I do not think this paper was particularly good, and it did not warrant the attention it got. (The main flaws IMO are a lack of (empirical) comparison to other methods — except a brief interlude in the appendix; and lack of any benchmarking — for example testing if clamping sycophancy features affects performance on sycophancy benchmarks)
At an object level, one concerning-to-me result is that there doesn't appear to be a clean gradient in the presence of a feature over the range of activation values. You might hope that if you take the AI risk feature[1], and look at dataset examples that span its activation values (as the tool does), you would see highly activating text be very related to AI risk and low activating text be only slightly related. I think that pattern is weak — there are at least some low-activation examples that are highly related to AI risk, such as '..."It's what they're programmed to do." "Destroy all technology other than their own"' (cherrypicked by me). This is related to sensitivity, which the paper mentions is difficult to study in this context (before mentioning one cherry-picked result). I care about this because: one way to use SAEs for safety is as a classifier for malicious behavior (be checking if model activations correspond to dangerous features); this would really benefit from having a nice smooth relationship between feature activation magnitude and actual feature presence, and it pretty much needs to have high sensitivity. Given the existence of highly-feature-related samples in the bottom activation interval, I feel fairly worried that sensitivity is poor, and that it will be hard to do magnitude-based thresholds — it pretty much looks like 0 is the reasonable threshold given these results.
In the paper this is labeled with "The concept of an advanced AI system causing unintended harm or becoming uncontrollable and posing an existential threat to humanity"
I don’t have strong takes, but you asked for feedback.
It seems nontrivial that the “value proposition” of collaborating with this brain-chunk is actually net positive. E.g., if it involved giving 10% of the universe to humanity, that’s a big deal. Though I can definitely imagine where taking such a trade is good.
It would likely help to devise more clarity about why the brain-chunk provides value. Is it because humanity has managed to coordinate to get a vast majority of high performance compute under the control of a single entity and access to compute is what’s being offered? If we’re at that point, I think we probably have many better options (e.g., long term moratorium and coordinated safety projects).
Another load bearing part seems to be the brain-chunk causing the misaligned AI to become or remain somewhat humanity friendly. What are the mechanisms here? The most obvious thing to me is that AI submits jobs to the cluster along with a thorough explanation of why they will create a safe successor system, and then the brain-chunk is able to assess these plans and act as a filter, only allowing safer-seeming training runs to happen. But if we’re able to accurately assess the viability of safe AGI design plans that are proposed by a human+ level (and potentially malign) AGIs, great, we probably don’t need this complicated scheme where we let a potentially malign undergo rsi.
Again, no strong feelings, but the above do seem like weaknesses. I might have understood things you were saying.
I do wish there was more work thinking about standard trades with misaligned AIs, but perhaps this is going on privately.
I appreciate this comment, especially #3, for voicing some of why this post hasn't clicked for me.
The interesting hypotheses/questions seem to rarely have strong evidence. But I guess this is partially a selection effect where questions become less interesting by virtue of me being able to get strong evidence about them, no use dwelling on the things I'm highly confident about. Some example hypotheses that I would like to get evidence about but which seem unlikely to have strong evidence: Sam Altman is a highly deceptive individual, far more deceptive than the average startup CEO. I work better when taking X prescribed medication. I would more positively influence the far future if I worked on field building rather than technical research.
Just chiming in that I appreciate this post, and my independent impressions of reading the FSF align with Zach's conclusions: weak and unambitious.
A couple additional notes:
The thresholds feel high — 6/7 of the CCLs feel like the capabilities would be a Really Big Deal in prosaic terms, and ~4 feel like a big deal for x-risk. But you can't say whether the thresholds are "too high" without corresponding safety mitigations, which this document doesn't have. (Zach)
These also seemed pretty high to me, which is concerning given that they are "Level 1". This doesn't necessarily imply but it does hint that there won't be substantial mitigations — above the current level — required until those capability levels. My guess is that current jailbreak prevention is insufficient to mitigate substantial risk from models that are a little under the level 1 capabilities for e.g., bio.
GDP gets props for specifically indicating ML R&D + "hyperbolic growth in AI capabilities" as a source of risk.
Given the lack of commitments, it's also somewhat unclear what scope to expect this framework to eventually apply to. GDM is a large org with, presumably, multiple significant general AI capabilities projects. Especially given that "deployment" refers to external deployment, it seems like there's going to be substantial work to ensuring that all the internal AI development projects proceed safely. e.g., when/if there are ≥3 major teams and dozens of research projects working on fine-tuning highly capable models (e.g., base model just below level 1), compliance may be quite difficult. But this all depends on what the actual commitments and mechanisms turn out to be. This comes to mind after this event a few weeks ago, where it looks like a team at Microsoft released a model without following all internal guidelines, and then tried to unrelease it (but I could be confused).
Sam Altman and OpenAI have both said they are aiming for incremental releases/deployment for the primary purpose of allowing society to prepare and adapt. Opposed to, say, dropping large capabilities jumps out of the blue which surprise people.
I think "They believe incremental release is safer because it promotes societal preparation" should certainly be in the hypothesis space for the reasons behind these actions, along with scaling slowing and frog-boiling. My guess is that it is more likely than both of those reasons (they have stated it as their reasoning multiple times; I don't think scaling is hitting a wall).
This might be a dumb question(s), I'm struggling to focus today and my linear algebra is rusty.
Is the observation that 'you can do feature ablation via weight orthogonalization' a new one?
It seems to me like this (feature ablation via weight orthogonalization) is a pretty powerful tool which could be applied to any linearly represented feature. It could be useful for modulating those features, and as such is another way to do ablations to validate a feature (part of the 'how do we know we're not fooling ourselves about our results' toolkit). Does this seem right? Or does it not actually add much?
Thinking about AI training runs scaling to the $100b/1T range. It seems really hard to do this as an independent AGI company (not owned by tech giants, governments, etc.). It seems difficult to raise that much money, especially if you're not bringing in substantial revenue or it's not predicted that you'll be making a bunch of money in the near future.
What happens to OpenAI if GPT-5 or the ~5b training run isn't much better than GPT-4? Who would be willing to invest the money to continue? It seems like OpenAI either dissolves or gets acquired. Were Anthropic founders pricing in that they're likely not going to be independent by the time they hit AGI — does this still justify the existence of a separate safety-oriented org?
This is not a new idea, but I feel like I'm just now taking some of it seriously. Here's Dario talking about it recently,
I basically do agree with you. I think it’s the intellectually honest thing to say that building the big, large scale models, the core foundation model engineering, it is getting more and more expensive. And anyone who wants to build one is going to need to find some way to finance it. And you’ve named most of the ways, right? You can be a large company. You can have some kind of partnership of various kinds with a large company. Or governments would be the other source.
Now, maybe the corporate partnerships can be structured so that AGI companies are still largely independent but, idk man, the more money invested the harder that seems to make happen. Insofar as I'm allocating probability mass between 'acquired by big tech company', 'partnership with big tech company', 'government partnership', and 'government control', acquired by big tech seems most likely, but predicting the future is hard.
Um, looking at the scaling curves and seeing diminishing returns? I think this pattern is very clear for metrics like general text prediction (cross-entropy loss on large texts), less clear for standard capability benchmarks, and to-be-determined for complex tasks which may be economically valuable.
Capability benchmarks: see epoch post, the ~4th figure here
Complex tasks: See GDM dangerous capability evals (Fig 9, which indicates Ultra is not much better than Pro, despite likely being trained on >5x the compute, though training details not public)
To be clear, I'm not saying that a $100m model will be very close to a $1b model. I'm saying that the trends indicate they will be much closer than you would think if you only thought about how big a 10x difference in training compute is, without being aware of the empirical trends of diminishing returns. The empirical trends indicate this will be a relatively small difference, but we don't have nearly enough data for economically valuable tasks / complex tasks to be confident about this.
Yeah, these developments benefit close-sourced actors too. I think my wording was not precise, and I'll edit it. This argument about algorithmic improvement is an argument that we will have powerful open source models (and powerful closed-source models), not that the gap between these will necessarily shrink. I think both the gap and the absolute level of capabilities which are open-source are important facts to be modeling. And this argument is mainly about the latter.
Yeah, I think we should expect much more powerful open source AIs than we have now. I've been working on a blog post about this, maybe I'll get it out soon. Here are what seem like the dominant arguments to me:
Scaling curves show strongly diminishing returns to $ spend: A $100m model might not be that far behind a $1b model, performance wise.
There are numerous (maybe 7) actors in the open source world who are at least moderately competent and want to open source powerful models. There is a niche in the market for powerful open source models, and they hurt your closed-source competitors.
I expect there is still tons of low-hanging fruit available in LLM capabilities land. You could call this "algorithmic progress" if you want. This will decrease the compute cost necessary to get a given level of performance, thus raising the AI capability level accessible to less-resourced open-source AI projects. [edit: but not exclusively open-source projects (this will benefit closed developers too). This argument is about the absolute level of capabilities available to the public, not about the gap between open and closed source.]
The implication of ICL being implicit BI is that the model is locating concepts it already learned in its training data, so ICL is not a new form of learning that has not been seen before.
I'm not sure I follow this. Are you saying that, if ICL is BI, then a model could not learn a fundamentally new concept in context? Can some of the hypotheses not be unknown — e.g., the model's no-context priors are that it's doing wikipedia prediction (50%), chat bot roleplay (40%), or some unknown role (10%). And ICL seems like it could increase the weight on the unknown role. Meanwhile, actually figuring out how to do a good job in the previously-unknown role would require piecing together other knowledge the model has — and sufficiently strong building blocks would allow a lot of learning of new concepts.
For example, if the GPT-4 evaluator gave a weighted score of 2.0 to a summary generated by Claude 2 and a weighted score of 3.0 to its own summary for the same article, then its final normalized self-preference score for the Claude summary would be 2/(2+3)=0.4.
Should this be 3/(2+3) = 0.6? Not sure I've understood correctly.
I expect a lot more open releases this year and am committed to test their capabilities and safety guardrails rigorously.
Glad you're planning on continual testing, that seems particularly important here, where the default is every once in awhile some new report comes out with a single data point about how good some model is and people slightly freak out. Having the context of testing numerous models over time seems crucial for actually understanding the situation and being able to predict upcoming trends. Hopefully you have and will continue to find ways to reduce the effort needed to run marginal experiments, e.g., having a few clearly defined tasks you repeatedly use, reusing finetuning datasets, etc.
Slightly Aspirational AGI Safety research landscape
This is a combination of an overview of current subfields in empirical AI safety and research subfields I would like to see but which do not currently exist or are very small. I think this list is probably worse than this recent review, but making it was useful for reminding myself how big this field is.
Other capability evaluations / benchmarking (useful for knowing how smart AIs are, informing forecasts), including persona evals
Understanding normal but poorly understood things, like in context learning
Understanding weird phenomenon in deep learning, like this paper
Understand how various HHH fine-tuning techniques work
AI Control
General supervision of untrusted models (using human feedback efficiently, using weaker models for supervision, schemes for getting useful work done given various Control constraints)
Unlearning
Steganography prevention / CoT faithfulness
Censorship study (how censoring AI models affects performance; and similar things)
Model organisms of misalignment
Demonstrations of deceptive alignment and sycophancy / reward hacking
Trojans
Alignment evaluations
Capability elicitation
Scaling / scalable oversight
RLHF / RLAIF
Debate, market making, imitative generalization, etc.
Reward hacking and sycophancy (potentially overoptimization, but I’m confused by much of it)
Weak to strong generalization
General ideas: factoring cognition, iterated distillation and amplification, recursive reward modeling
Robustness
Anomaly detection
Understanding distribution shifts and generalization
User jailbreaking
Adversarial attacks / training (generally), including latent adversarial training
AI Security
Extracting info about models or their training data
Understanding optimization in multi-agent situations
Attacks vs. defenses for various problems
Unsorted / grab bag
Watermarking and AI generation detection
Honesty (model says what it believes)
Truthfulness (only say true things, aka accuracy improvement)
Uncertainty quantification / calibration
Landscape study (anticipating and responding to risks from new AI paradigms like compound AI systems or self play)
Don’t quite make the list:
Whose values? Figuring out how to aggregate preferences in RLHF. This seems like it’s almost certainly not a catastrophe-relevant safety problem, and my weak guess is that it makes other alignment properties harder to get (e.g., incentivizes sycophancy and makes jailbreaks easier by causing the model’s goals to be more context/user dependent). This work seems generally net positive to me, but it’s not relevant to catastrophic risks and thus is relatively low priority.
Fine-tuning that calls itself “alignment”. I think it’s super lame that people are co-opting language from AI safety. Some of this work may actually be useful, e.g., by finding ways to mitigate jailbreaks, but it’s mostly low quality.
I'll think more about the outer shell stuff, it's possible that my objection actually arises with the consequentialist assumption, but I'm not sure.
It's on my todo list to write a comment responding to some of the specifics of Redwood's control post.
I would be excited to read this / help with a draft.
Yes, approximately, as I believe you and I are capable of doing.
The analogy is strained due to not being able to gradient update my brain with arbitrary training data. It's pretty unclear to me if I could pull off deception like this, it seems pretty hard imo.
Seems easy enough to predict given roughly human-scientist-team level of capabilities.
One situation I'm imagining here is that we've (explicitly) trained our AI on scientific output from January to March, and we're now telling it to do June-level research. It does not know if we have labeled data for just up to March, for April, or all the way to June. I think if you took a human research team, stuck them on an island, and had them get up to speed in some subfield they're not part of, they would be unable to distinguish between research that had just been done on the mainland or is just past the frontier. This just seems pretty hard. I think this example is closer to what we'll get with AIs rather than the bar being at "human level" it's more like "what has been accomplished by the humans+AIs in the outside world, for which I've received untrustworthy information". And every time the AIs do useful work we can add this to our supervised training corpus.
> The data need not be human generated, it could be generated by other AIs, as long as we trust it.
?? This seems to be assuming a solution to the problem.
One version of this is: you have GPT-5 (potentially untrustworthy) do some research, you have humans carefully check this research using the best available tools (e.g., AI debate / other 'scalable oversight'), and then you use this research as supervised / labeled data for training GPT-6 to do good work.
Latent adversarial training seems like it could make an agent robust to a few extra categories of internal perturbation, but doesn't seem like it would help with any problem that we were focused on in this post.
Here's a specific thing I mean: we can look at activations and use these to aid in the training process, even without a full understanding of what activations mean. Take this recent paper (which I think is super confused but helps get at the thing), "We find a strong positive relationship between the LLM's attention to constraint tokens and the factual accuracy of generations". My interpretation is that they found a probably-causal link between 'the model is attending to relevant tokens' and 'the model gets the answer correct', because duh, it's like how when you give somebody a riddle you can tell whether they're on track based on what parts of the question they're focused on. While it's probably non-trivial, I expect you could use this internal activation to train models which do a better job paying attention to the tokens you want them to. Similarly, I think it's probable (but far from guaranteed) that deceptively aligned models are representing training/deployment in their activations, or that sandbagging models are representing human-level / super-human-level difficulty for tasks. If so, it it possible in principle to do all kinds of things with these activations which may help us. e.g., get the model to solve tasks it otherwise would have intentionally failed at, or giving us hints about the accuracy of a research result. I expect this is difficult implementation wise, but still a reason for hope. IMO activation engineering is off to a good start.
"leaning more on generalization" ≈ "leaning more on agency",
That's not why I meant to say. By leaning more on generalization, I mean leaning more on the data efficiency thing, and also weak-to-strong generalization ideas. I think this is different from "leaning more on agency".
The problem is shaping the goal to be a specific thing, not changing it in unpredictable or not-fully-understood ways.
I'm not so sure of this. I could imagine starting with a deceptively aligned AI whose goal is "Make paperclips unless being supervised which is defined as X, Y, and Z, in which case look good to humans". And if we could change this AI to have the goal "Make paperclips unless being supervised which is defined as X, Y, and Q, in which case look good to humans", that might be highly desirable. In particular, it seems like adversarial training here allows us to expand the definition of 'supervision', thus making it easier to elicit good work from AIs (ideally not just 'looks good').
I think that there is a lot of room for the evidence to be ambiguous and controversial
I agree that this is a major potential problem which I think is pretty likely to occur; maybe I'll think/write more about it soon. I think I brought this up in response to some of the Redwood stuff and got a response mentioning that we can play things forward through our oversight scheme, and get clearer answers to this question. I think this remains an important concern, but it's also true that we'll have the ability to actually run some outputs through to e.g., see if they trigger oversight AIs to mess up, see if a vulnerability would have been successfully inserted in production, etc. It might also end up being very expensive to do this for everything that seems like a possible coup attempt, idk.
I buy this for many environments and levels of power, but it's obviously wrong for AIs that have taking-over-the-universe in their easily-within-reach options. And that's the sort of AI we get if it can undergo self-improvement.
I don't feel like I'm talking about AIs which have "taking-over-the-universe in their easily-within-reach options". I think this is not within reach of the current employees of AGI labs, and the AIs I'm thinking of are similar to those employees in terms of capabilities, but perhaps a bit smarter, much faster, and under some really weird/strict constraints (control schemes).
Caveat that I have only skimmed this. I don't think your arguments engage with any of my cruxes or provide detailed enough argument on the key questions for me; perhaps I'm not the main audience. My reactions:
Section 4: Behavioral training is an imprecise way to specify goals - most the sub-arguments have me going "shrug, maybe". I definitely don't feel confident that any of the mentioned problems will arise in practice. I don't see why I should believe in an inner/outer shell breakdown of constraints — this section seemed quite speculative.
Section 5: Control of misaligned AIs is difficult - This is probably where I disagree the most. Most of my comments are reiterations of argument made in Redwood's recent post, because I think their arguments are stronger than the ones you make here. The main other source for my thinking here is the capability elicitation stuff (1, 2, 3).
"Misaligned goal-directed AI won’t cooperate with humans" Well, what the hell else is it going to do? Is the claim here that the AI performs well on ~all the human-checkable tasks and then reverts to being absolutely useless or sabotaging on the hard stuff? Predicting where that line is seems pretty difficult, especially given the AI cannot trust sensory inputs, does not know what time it is, does not know if other AIs are helping the humans (specifically, have leaked higher capability data that is being used for SFT). [Additionally, I think it's plausibly the correct strategy for such an AI is to just actually help the humans and try to make peaceful deals. Not very confident, but there's definitely some other solution space here.]. I think this subsection title is possible but unlikely.
"There needs to be sufficient human-generated research in the domain where we are attempting to elicit the AI’s capabilities." Well, no. You either need sufficient labeled data in the domain we're trying to elicit capabilities, or good generalization from other domains. The data need not be human generated, it could be generated by other AIs, as long as we trust it.
"It only applies to relatively easy to check research results" seems possible, and worrisome. This is the subsection I think is most right and important in section 5. I think there are still numerous directions of hope though, e.g., leaning more on generalization, latent adversarial training.
"In order for the AI to do useful research, it needs to pursue the goal (or sub-goal) of doing research. There’s no particular reason for the gradient based training to actually modify the goal representation, rather than building some additional narrow machinery on top." I agree that we don't have strong reasons to think one thing or another here, but I think goal modification is reasonably likely: humans can iterate a bunch and maybe have okay interpretability tools (assuming decent but not fundamental advances). Also, as discussed, goals might be changing a bunch during training — that's not totally asymmetric, it also gives us hope about our ability to modify AI goals.
"It may be hard to efficiently elicit capabilities which require additional (perhaps chain-of-thought) reasoning steps." I agree that capability elicitation seems non-trivial in some cases, but I'm reasonably optimistic that future work will improve the situation.
Many of the issues in this section are things that, if we're not being totally idiots, it seems we'll get substantial warning about. e.g., AIs colluding with their AI monitors. That's definitely a positive, though far from conclusive.
Section 6: Powerful misaligned AI would be bad - I think a couple of the arguments here continue to be legitimate, such as "Unclear that many goals realistically incentivise taking over the universe", but I'm overall fine accepting this section.
Going into reading this, I was hoping it would tackle some of the recent discussion around capability elicitation and Control. I think it didn't really engage with those arguments in a way that sways me much.
