So I've become more pessimistic about this sort of outcome happening over the last year, and a lot of it comes down to me believing in longer timelines, which means I expect the LLM paradigm to be superseded, and the most likely options in this space that increase capabilities unfortunately correspond to more neuralese/recurrent architectures. To be clear, they're hard enough to make work that I don't expect them to outperform transformers so long as transformers can keep scaling, and even after the immense scale-up of AI that will slow down to the Moore's law trend by 2030-2031, I still expect neuralese to be at least somewhat difficult, and it's possibly difficult enough that we might survive because AI never reached the capabilities we expect.
(Yes, the scaling hypothesis in a weak form will survive, but I don't expect the strong versions of the scaling hypothesis to work. Reasons for why I believe this are available on request).
It's still very possible this happens, but I wouldn't put much weight on this for planning purposes.
I agree with a weaker version of the claim that says that the AI safety landscape is looking better than people thought 10-15 years ago, with AI control probably being the best example here. To be clear, I do think this is actually meaningful and does matter, primarily because they focus less on the limiting cases of AI competencies, but I currently am not as optimistic about some of the important properties of LLMs that are relevant for alignment surviving for newer AI designs, which means I disagree with this post.
Relevant posts on this point which argue that catching misalignment is a big help in fixing it (which is relevant to the bumpers plan):
Catching AIs red-handed by Ryan Greenblatt and Buck Shlegeris:
https://www.lesswrong.com/posts/i2nmBfCXnadeGmhzW/catching-ais-red-handed
Handling schemers if shutdown is not an option, by Buck Shlegeris:
https://www.lesswrong.com/posts/XxjScx4niRLWTfuD5/handling-schemers-if-shutdown-is-not-an-option
A really important question that I think will need to be answered is whether specification gaming/reward hacking must be in a significant sense be solved by default in order to unlock extreme capabilities.
I currently lean towards yes, due to the evidence offered by o3/Sonnet 3.7, but could easily see my mind changed, but the reason this question has such a large amount of importance is that if it were true, then we'd get tools to solve the alignment problem (modulo inner optimization issues), which means we'd be far less concerned about existential risk from AI misalignment (at least to the extent that specification gaming is a large portion of the issues with AI.
That said, I do think a lot of effort will be necessary to discover the answer to the question, because it affects a lot of what you would want to do in AI safety/AI governance if alignment tools come along with better capabilities or not.
My second concern is that “AIs solving specific technical problems that the human wants them to solve” is insufficient to avoid extinction and get to a good future—even if the AI is solving those problems at superhuman level and with the help of (1-3) superpowers.[3]
I won’t go much into details of why I have this concern, since I want to keep this post focused on technical alignment. But I’ll spend just a few paragraphs to give a hint. For more, see my What does it take to defend the world against out-of-control AGIs? and John Wentworth’s The Case Against AI Control.
Here’s an illustrative example. People right now are researching dangerous viruses, in a way that poses a risk of catastrophic lab leaks wildly out of proportion to any benefits. Why is that happening? Not because everyone is trying to stop it, but they lack the technical competence to execute. Quite the contrary! Powerful institutions like governments are not only allowing it, but are often funding it!
I think when people imagine future AI helping with this specific problem, they’re usually imagining the AI convincing people—say, government officials, or perhaps the researchers themselves—that this activity is unwise. See the problem? If AI is convincing people of things, well, powerful AIs will be able to convince people of bad and false things as well as good and true things. So if the AI is exercising its own best judgment in what to convince people of, then we’d better hope that the AI has good judgment! And it’s evidently not deriving that good judgment from human preferences. Human preferences are what got us into this mess! Right?
Remember, the “alignment generalizes farther” argument (§4 above) says that we shouldn’t worry because AI understands human preferences, and those preferences will guide its actions (via RLHF or whatever). But if we want an AI that questions human preferences rather than slavishly following them, then that argument would not be applicable! So how are we hoping to ground the AI’s motivations? It has to be something more like “ambitious value learning” or Coherent Extrapolated Volition—things that are philosophically fraught, and rather different from what people today are doing with foundation models.
I agree with the claim that existential catastrophes aren't automatically solved by aligned/controlled AI, and in particular biological issues remain a threat to human survival.
My general view on how AI helps in practice is more likely to route around first figuring out ways to solve the biology threat from a technical perspective, and then using instruction following AIs to implement a campaign of persuasion that uses the most effective way to change people's minds.
You are correct that an AI can convince people of true things as well as false things, and this is why you will need to make assumptions about the AI's corrigibility/instruction following, though you helpfully make such assumptions here.
On the first person problem, I believe that the general solution to this involves recapitulating human social instincts via lots of data on human values, and I'm perhaps more optimistic than you that a lot of the social instincts in humans don't have to be innately specified by a prior.
In many ways, I have a weird reaction to the post, because I centrally agree with the claim that corrigible/instruction following AIs aren't automatically sufficient to ensure safety, and yet I am much more optimistic than you do that mere corrigibility/instruction following is a huge way for AI to be safe, probably because I think you can actually do a lot more work to secure civilization in ways that semi-respect existing norms, but semi-respect is key here.
Yeah, I think the crux is precisely this, in which I disagree with this statement below, mostly because I think instruction following/corrigibility is both plausibly easy in my view, and also removes most of the need for value alignment.
"The AI's values must be much more aligned in order to be safe outside the text domain"
I think where I get off the train personally probably comes down to the instrumental goals leading to misaligned goals section, combined with me being more skeptical of instrumental goals leading to unbounded power-seeking.
I agree there are definitely zero-sum parts of the science loop, but my worldview is that the parts where the goals are zero sum/competitive receive less weight than the alignment attempts.
I'd say the biggest area of how I'm skeptical so far is that I think there's a real difference between the useful idea that power is useful for the science loop and the idea that the AI will seize power by any means necessary to advance it's goals.
I think instrumental convergence will look more like local power-seeking that is more related to the task at hand, and not to serve some of it's other goals, primarily due to denser feedback constraining the solution space and instrumental convergence more than humans.
That said, this is a very good post, and I'm certainly happier that this much more rigorous post was written than a lot of other takes on scheming.
Oh, now I understand.
And AIs have already been superhuman at chess for very long, yet that domain gives very little incentive for very strong instrumental convergence.
I am claiming that for practical AIs, the results of training them in the real world with goals will give them instrumental convergence, but without further incentives, will not give them so much instrumental convergence that it leads to power-seeking to disempower humans by default.
To answer the question:
So, as a speculative example, further along in the direction of o1 you could have something like MCTS help train these things to solve very difficult math problems, with the sparse feedback being given for complete formal proofs.
Similarly, playing text-based video games, with the sparse feedback given for winning.
Similarly, training CoT to reason about code, with sparse feedback given for predictions of the code output.
Etc.
You think these sorts of things just won't work well enough to be relevant?
Assuming the goals are done over say 1-10 year timescales, or maybe even just 1 year timescales with no reward-shaping/giving feedback for intermediate rewards at all, I do think that the system won't work well enough to be relevant, since it requires way too much time training, and plausibly way too much compute depending on how sparse the feedback actually is.
Other AIs relying on much denser feedback will already rule the world before that happens.
[insert standard skepticism about these sorts of generalizations when generalizing to superintelligence]
But what lesson do you think you can generalize, and why do you think you can generalize that?
Alright, I'll give 2 lessons that I do think generalize to superintelligence:
The data is a large factor in both it's capabilities and alignment, and alignment strategies should not ignore the data sources when trying to make predictions or trying to intervene on the AI for alignment purposes.
Instrumental convergence in a weak sense will likely exist, because having some ability to get more resources are useful for a lot of goals, but the extremely unconstrained versions of instrumental convergence often assumed where an AI will grab so much power as to effectively control humanity is unlikely to exist, given the constraints and feedback given to the AI.
For 1, the basic answer for why is because a lot of AI success in fields like Go and language modeling etc was jumpstarted by good data.
More importantly, I remember this post, and while I think it overstates things in stating that an LLM is just the dataset (it probably isn't now with o1), it does matter that LLMs are influenced by their data sources.
https://nonint.com/2023/06/10/the-it-in-ai-models-is-the-dataset/
For 2, the basic reason for this is that the strongest capabilities we have seen that come out of RL either require immense amounts of data on pretty narrow tasks, or non-instrumental world models.
This is because constraints prevent you from having to deal with the problem where you produce completely useless RL artifacts, and evolution got around this constraint by accepting far longer timelines and far more computation in FLOPs than the world economy can tolerate.
I don't get why you think this is true? EG, it seems like almost no insights about how to train faithful CoT would transfer to systems speaking pure neuralese. It seems to me that what little safety/alignment we have for LLMs is largely a byproduct of how language-focused they are (which gives us a weak sort of interpretability, a very useful safety resource which we are at risk of losing soon).
I think the crux is I think that the important parts of of LLMs re safety isn't their safety properties specifically, but rather the evidence they give to what alignment-relevant properties future AIs have (and note that I'm also using evidence from non-LLM sources like MCTS algorithm that was used for AlphaGo), and I also don't believe interpretability is why LLMs are mostly safe at all, but rather I think they're safe due to a combo of incapacity, not having extreme instrumental convergence, and the ability to steer them with data.\
Language is a simple example, but one that is generalizable pretty far.
It sounds like you think safety lessons from the human-imitation regime generalize beyond the human-imitation regime
Note that the primary points would apply to basically a whole lot of AI designs like MCTS for AlphaGo or a lot of other future architecture designs which don't imitate humans, barring ones which prevent you from steering it at all with data, or have very sparse feedback, which translates into weakly constraining instrumental convergence.
but we're moving away from the regime where such dense feedback is available, so I don't see what lessons transfer.
I think this is a crux, in that I don't buy o1 as progressing to a regime where we lose so much dense feedback that it's alignment relevant, because I think sparse-feedback RL will almost certainly be super-uncompetitive with every other AI architecture until well after AI automates all alignment research.
Also, AIs will still have instrumental convergence, it's just that their goals will be more local and more focused around the training task, so unless the training task rewards global power-seeking significantly, you won't get it.
Some good picks the for how to design reward functions starter pack (though I should note that their empirical support is very weak due to focusing on toy models) are Defining Corrigible and Useful Goals and Defining Monitorable and Useful Goals.
The first post focuses on how you can get a goal for AIs that allow you to shutdown the AI while having the AI be useful, and the approach to corrigibility it takes is extremely different to how human brains work, using the corrigibility transformation to get corrigible AIs.
One big caveat here is that it definitely requires the assumption that Causal Decision Theory is used but I mostly am fine with that assumption, given that humans intuitively use Causal Decision Theory and it's in the spec of the transformation rather than a background assumption.
The other big caveat is that you want the model to optimize for the reward in order for this to work, so in terms of under-sculpting vs over-sculpting, or whether an AI is driven by the reward vs driven by another goal, you want to have the AI reward-maximize and be over-sculpted (though in this case it's just appropriately sculpted via reward), which makes it incompatible with corrigibility/alignment hopes that depend on AIs not maximizing the reward, but I think this is a good property to have.
The post on defining monitorable and useful goals proposes the idea of the monitorability transformation to get AIs to not be incentivized to fool monitors generally, and I'd recommend reading that over any explanation I'd give.
These are admittedly curveballs compared to standard LW thoughts on this, but this is why I picked them for the reward functions starter pack, as they contain novel ideas to deal with some notorious problems.