Curated. This seems like an important set of considerations for alignment researchers to think about.
I agree with the main points made in the post, though I want to recognize there is some difficulty that comes with predicting which aspects will drive capability advances. I think there is value in reading papers (something that more alignment researchers should probably do) because it can give us hints at the next capability leaps. Over time, I think it can improve our intuition for what lies ahead and allows us to better predict the order of capability advances. This is how I’ve felt as I’ve been pursuing the Accelerating Alignment agenda (language model systems for accelerating alignment research). I’ve been at the forefront, reading Twitter/papers/etc to find insights into how to use language models for research and feel like I’ve been gaining a lot of intuition into where the field is going.
As you said, it's also important to remember that most of the field isn't directly aiming for AGI. Safety discussions, particularly about self-improvement and similar topics, may have inspired some individuals to consider pursuing directions useful for AGI, when they might not have otherwise. This is why some people will say things like, "AI safety has been net negative and AGI safety discussions have shortened AGI timelines". I think there is some truth to the timelines argument, but it’s not clear it has been net negative, in my opinion. There's a point at which AI Safety work must be done and investment must be made in AGI safety.
One concern I’d like to bring up as a point of discussion is that whether infohazard policies could backfire. By withholding certain insights, these policies may leave safety researchers in the dark about the field's trajectory, while capability researchers are engaged in active discussions. Some of us were aware about the AgentGPT-like models likely happening soon (though unsure about the exact date), but it seems to have blindsided a lot of people concerned about alignment. It’s possible that safety researchers could be blindsided again by rapid developments they were not privy to due to infohazard policies.
This may have been manageable when progress was slower, but now, with global attention on AI, it may lead to some infohazard policies backfiring, particularly due to alignment people not being able to react as quickly as they should. I think most of the balance favours keeping infohazard policies as is for now, but this was a thought I had earlier this week and figured I would share.
Complementary ideas to this article:
The tech tree picture is in large part a reaction to single-factor "capabilities/alignment ratio" pictures (as in FAI Research Constraints), or even two-factor pictures (as in State Space), because while they summarize the results of decision-making about research priorities, they didn't help me summarize my reasoning process.
Gotcha. What determines the "ratios" is some sort of underlying causal structure of which some aspects can be summarized by a tech tree. For thinking about the causal structure you may also like this post: https://forum.effectivealtruism.org/posts/TfRexamDYBqSwg7er/causal-diagrams-of-the-paths-to-existential-catastrophe
I previously had had a cruder model of "There's an AI capabilities fuel tank and an AI alignment fuel tank. Many research items fill up both at the same time, but in different ratios. If you fill up the capabilities tank before the alignment tank, we lose. You want to pursue strategies that cause the alignment tank to get filled up faster than the capabilities tank." (I got this from Andrew Critch during in-person converation)
I like this post for putting forth a higher resolution model, that prompts me to think a bit more specifically about what downstream effects I expect to happen. (though I think the tank model might still be kinda useful as a fast shorthand sometimes)
Could we get this curated ASAP? I think it should be pinned indefinitely. Until such time as there is a better intro, this seems like the best technical intro to the coordination game of not destroying the world we have; focused on the dynamics of the game rather than the fact of the existence of the red nodes, which is what other discussions focus on.
When you accidentally unlock the tech tree by encouraging readers to actually map out a tech tree and strategize about it
No, excellent analysis though.
I think an important point to make is that whether a point is red or yellow depends on other considerations than just capabilities vs alignment. For example the current state of hardware progress, how integrated fully autonomous robots are in our factories, society, and warfare and how common AI companions are.
Consider that case that the current (2023) deployed hardware/compute is not X-risk. That is even with the most advanced/efficient algorithms the resulting AGI could not take over (because say the von Neumann bottleneck), but would have to wait until more advanced chip factories where built and started production.
To be clear, if the current hardware/compute does not constitute X-risk, then all the nodes are arguably currently yellow. I appreciate it could be more complicated than this, if you regard it as highly likely that such an AI in the situation described before will convince people to make more sophisticated chip factories or this will happen by default anyway, then you could say the node is actually red, or some probability yellow/red.
I disagree. I think you underestimate how powerful a embodied but highly persuasive and superhumanly strategic and super fast AGI could be. I think we (2023) are already well beyond an X-risk danger threshold for deployed hardware/compute, and that the only thing holding us back from doom is that we haven't stumbled onto a sufficiently advanced algorithm.
Do you think this model can handle disagreement over what color particular nodes in the tech tree are?
With uncertainty over tech trees yeah, but I don't think there's a silver bullet. When people disagree with you, you should be swayed for normal epistemological reasons, and also for social / logistical reasons, but sometimes you should go ahead anyway.
Dual use is the wrong name for this. The dangerous research is any and all gain-of-function research that increases probability of out-of-control powerseeking AIs. I'm not sure what to call that, but certainly not "dual" use.
The LessWrong Review runs every year to select the posts that have most stood the test of time. This post is not yet eligible for review, but will be at the end of 2024. The top fifty or so posts are featured prominently on the site throughout the year.
Hopefully, the review is better than karma at judging enduring value. If we have accurate prediction markets on the review results, maybe we can have better incentives on LessWrong today. Will this post make the top fifty?
RLHF is already a counterexample. It's had an impact because of alignment's inherent dual use - is that cheating? No: RLHF is economically valuable, but Paul still put work into it for alignment before profit-seeking researchers got to it. It could have been otherwise.
I think this works as a counterexample in the case of Paul, because we assume that he is not part of the 'profit-seeking' group (regardless of whether its true, depending on what it means). However, as you pointed out earlier, alignment itself is crucial for the 'profit-seeking' case.
Let's pretend for a minute that alignment and capabilities are two separate things. As you point out, a capable but misaligned (wrt. its creators' goals) is not profit generating. This means alignment as a whole is 'profitable', regardless of whether an individual researcher is motivated by the profit or not. The converse can be said about capability. It is not necessarily the case that a capability researcher is motivated by profit (in which case maybe they should focus on alignment at some point instead).
Why are your competitors so few? If the development of superintelligent AI is going to be a hugely important event in the course of human history, why aren't more people doing research motivated by it? You might as well ask why there are only 300 alignment researchers. For whatever reasons, most AI research simply isn't that future-oriented.
'300' publicly disclosed alignment researchers.
I think quite a lot happens, or has already happened, in more secretive circles.
Do you have any particular reason to believe that, or is this a fully-general unfalsifiable belief in the existence of secretive circles devoted to arbitrary pursuits?
(Note that despite the phrasing, I recognize this to be as eminently reasonable as believing in aliens: the world is a big place, and it's very easy to form a secretive circle.)
A crux here is whether there is any node subset that autonomous superintelligent AI would converge on in the long-term (regardless of previous alignment development paths).
I’ve written about this:
No control method exists to safely contain the global feedback effects of self-sufficient learning machinery. What if this control problem turns out to be an unsolvable problem?
…
But there is an even stronger form of argument:
Not only would AGI component interactions be uncontainable; they will also necessarily converge on causing the extinction of all humans.
https://www.lesswrong.com/posts/xp6n2MG5vQkPpFEBH/the-control-problem-unsolved-or-unsolvable
This is pretty basic. But I still made a bunch of mistakes when writing this, so maybe it's worth writing. This is background to a specific case I'll put in the next post.
It's like a a tech tree
If we're looking at the big picture, then whether some piece of research is net positive or net negative isn't an inherent property of that research; it depends on how that research is situated in the research ecosystem that will eventually develop superintelligent AI.
Consider this toy game in the picture. We start at the left and can unlock technologies, with unlocks going faster the stronger our connections to prerequisites. The red and yellow technologies in the picture are superintelligent AI - pretend that as soon as one of those technologies is unlocked, the hastiest fraction of AI researchers are immediately going to start building it. Your goal is for humanity to unlock yellow technology before a red one.
This game would be trivial if everyone agreed with you. But there are many people doing research, and they have all kinds of motivations - some want as many nodes to be unlocked as possible (pure research - blue), some want to personally unlock a green node (profit - green), some want to unlock the nearest red or yellow node no matter which it is (blind haste - red), and some want the same thing as you (beneficial AI - yellow) but you have a hard time coordinating with them.
In this baseline tech tree game, it's pretty easy to play well. If you're strong, just take the shortest path to a yellow node that doesn't pass too close to any red nodes. If you're weak, identify where the dominant paradigm is likely to end up, and do research that differentially advantages yellow nodes in that future.
The tech tree is wrinkly
But of course there are lots of wrinkles not in the basic tech tree, which can be worth bearing in mind when strategizing about research.
Ultimately, which technologies are the right ones to research depends on your model of the world / how you expect the future to go. Drawing actual tech trees can be a productive exercise for strategy-building, but you might also find it less useful than other ways of strategizing.
We're usually mashing together definitions
I'd like to win the tech tree game. Let's define a "good" technology as one that would improve our chances of winning if it was unlocked for free, given the state of the game and the policies of the players, and a "bad" technology as one that would harm our chances. Dual-use research in what I think of as "the tech tree sense" is a technology that's an important prerequisite to both good and bad future technologies.
This is different from how we use "dual-use" outside of AI alignment! Dual-use nuclear technology is technology that all by itself can be used either for electricity or for bombs. This information is about the technology itself, no need to condition on the policies of players in a multiplayer game or anything like that. Any AI technology that can be used for both good and bad things sometimes gets called "dual-use" in this way.[1]
It's also different from how a lot of people use "dual-use" inside of AI alignment. Any technology that advances both alignment and capabilities is often called "dual-use."[2] This is different from the tech tree definition because "alignment" and "capabilities" are often measured locally, not rigidly connected to our chances of aligning superintelligent AI.
All these definitions are fine categories to have, but it can be a little confusing when people mash them together. I'm mostly not going to worry about it, because what's really important is the question of which technologies are helpful (or harmful) to our chances of getting aligned AI.
No definition of dual-use is a complete answer to the question of which technologies are good to research. Instead, this idea of dual-use technology should be taken as a guide to the imagination, a reminder to weigh the good things you want to use a technology for against its potential negative impacts.
The dual use you imagine depends on the research
The key questions to get your imagination kick-started are "Would other people like to use this too?" and "How do you use this in ways unrelated to its usefulness for AI alignment?" Vivid answers to both these questions can help you clarify how the research in question fits into your overall strategy.
Imagining dual use is often an exercise in trying to peer out into the next few layers of the tech tree. It's going to be idiosyncratic for each research topic - predicting the tech tree depends on specific models of technology and society, with few general rules.
Alignment research is inherently dual-use.
One of the reasons dual-use alignment research is salient right now is RLHF's broad usage. A key influence on this post is Paul Christiano's Thoughts on the impact of RLHF research.
RLHF is basically the poster child of "Other people would like to use this too." It lets you guide AI to do otherwise-hard-to-specify tasks, and surprise surprise this is economically valuable.
But RLHF is also the poster child of those claiming "Maybe it's fine for other people to want to use it." Yes, this technology has unlocked some green nodes for OpenAI that might indirectly hasten superintelligent AI. But mostly, people want to use RLHF directly as it was intended as alignment research, getting complicated human preferences into an AI.
Something like this is going to be the case for all alignment research with direct applications. The race isn't exactly "AI alignment" versus "AI capabilities" because alignment is a type of capability. An AI that's super smart but doesn't do what you want is useless (as Katja Grace points out vividly). We want to use AI alignment research to design the motivational system of future AI, but present-day AIs also have motivational systems that benefit from clever design.[3]
Inherent dual use might seem to be safe because it's directly linked to how good the research is at being alignment research. If you design a motivation system so clever it directly makes bad AI feasible, presumably it's also so clever it can be used for good AI (or else why did you research it?).
I basically buy this argument for reduced threat, but it doesn't eliminate all danger from dual use. If alignment research increases the economic value of building larger models and systems that make plans in the real world, this can make bad AI[4] more feasible before we understand how to build good AI.
Instrumental research has instrumental dual use
We don't always research the entire alignment problem at once.
Maybe you have an alignment strategy that relies on having a system that's robust to unusual inputs, and so you work on robustness. Someone else might use your work on increased robustness to get good capabilities generalization to new domains, without caring about the rest of your alignment strategy.
Compared to alignment's inherent dual use, this is more like to the sort of dual-use pointed at in Pragmatic AI Safety. If you're pursuing some research direction as an instrumental objective, the whole point of instrumental objectives is that lots of people will find them useful!
I think what gets unlocked on the tech tree is the most important question for evaluating dual use of instrumental alignment research. I'm less worried about alignment research making existing products 5% better, and more worried about alignment research helping to unlock totally new non-alignment capabilities.
Can alignment itself be used instrumentally in AI design, turning inherent dual use into indirect dual use? One example might be motivating components of an AI to work on their tasks, sort of like how AutoGPT relies on GPT's "alignment" to faithfully do the subtasks parceled out to it.
We (sometimes) won't get lost in the crowd
One possible reaction to the prospect of dual use research is that, strategically, it shouldn't matter at all. After all, there's a bunch of people already working on AI capabilities, isn't it a bit outlandish to think your alignment research is going to have a nontrivial impact on when it's not even what you're directly trying to do?
There are several interconnected reasons why alignment research isn't obviated by all the people working on capabilities.
RLHF is already a counterexample. It's had an impact because of alignment's inherent dual use - is that cheating? No: RLHF is economically valuable, but Paul still put work into it for alignment before profit-seeking researchers got to it. It could have been otherwise.
Let's actually back up and consider instrumental research first - like the example of robustness. Suppose you're considering working on robustness as a step towards later alignment research. If robustness is instrumentally useful, won't your work just get lost in the crowd?
First: Yes, absolutely, other AI researchers are definitely working on robustness for reasons other than yours. If you want to work on robustness, you had better read the literature.
Second: The field is big enough that your impact is small, but not so big that your impact is negligible, even on instrumentally useful sub-capabilities of alignment. This might sound strange given that there are 300,000 AI workers or whatever, and only 300 AI alignment workers or whatever, a tidy ratio of 1000:1. But there were only 50 or so papers on robustness in deep learning at ICML 2022. If you do a high-quality paper's worth of work, that's 1/50th of the sub-field.
Third: AI alignment suggests specific research directions that are more relevant than average to superintelligent AI. If you look at the papers on robustness in ICML, a lot of them are things like "one simple trick to make language models robust to adversarial attacks." That's important research for many present-day applications of AI, but it's not intended to be part of an AGI design. If you're doing AI alignment research because you take AGI seriously, you're going to want robustness in a different sense, and that sense is going to be more relevant both to aligning superintelligent AI and also to building it.
My guesstimate is that this drops the effective size of the "crowd" the dual use of your research is supposed to get lost in down to about 10. Even though there are hundreds of thousands of AI workers, even though a few hundreds of them are sometimes working on robustness, only an effective handful are working on the sort of robustness that advances superintelligent AI timelines.
Why are your competitors so few? If the development of superintelligent AI is going to be a hugely important event in the course of human history, why aren't more people doing research motivated by it? You might as well ask why there are only 300 alignment researchers. For whatever reasons, most AI research simply isn't that future-oriented.
This sort of future orientation and willingness to take powerful AI seriously is an advantage of the AI alignment community. It lets things like RLHF be investigated for alignment before they're used for economic gain, rather than after. It just also means that the research we find instrumentally useful will often also be selected for its import to powerful AI, and we can't plan on getting lost in the crowd.
This leads to fairly common positions on strategy
Thanks to MAIA for some of the desks this was written on. Thanks to Logan Riggs for some chats. Thanks to LTFF for some of the support.
E.g. Urbina et al., Dual use of artificial-intelligence-powered drug discovery.
E.g. See Dan Hendrycks and Thomas Woodside's Pragmatic AI Safety, particularly this section.
I'd still agree that more abstract work on "what does it even mean for an AI to be aligned with human values?" is less useful to other people, and we should try to have that research outrace the practical application of that research to motivation system design.
Not "bad" because it's built by bad people, but "bad" because its motivation system finds perverse optima - the classic example being the motivation system of reinforcement learning based on a button the human presses. When the AI is dumb, this motivation system increases the AI's economic value by directing it to do the tasks the human wants. When the AI is smart, it makes plans to take control of the button.