Victoria Krakovna. Research scientist at DeepMind working on AI safety, and cofounder of the Future of Life Institute. Website and blog: vkrakovna.wordpress.com
Vika
Karma: 2,893
When discussing AI risks, talk about capabilities, not intelligence
Re 4, we were just discussing this paper in a reading group at DeepMind, and people were confused why it’s not on arxiv.
The issue with being informal is that it’s hard to tell whether you are right. You use words like “motivations” without defining what you mean, and this makes your statements vague enough that it’s not clear whether or how they are in tension with other claims. (E.g. what I have read so far doesn’t seems to rule out that shards can be modeled as contextually activated subagents with utility functions.)
An upside of formalism is that you can tell when it’s wrong, and thus it can help make our thinking more precise even if it makes assumptions that may not apply. I think defining your terms and making your arguments more formal should be a high priority. I’m not advocating spending hundreds of hours proving theorems, but moving in the direction of formalizing definitions and claims would be quite valuable.
It seems like a bad sign that the most clear and precise summary of shard theory claims was written by someone outside your team. I highly agree with this takeaway from that post: “Making a formalism for shard theory (even one that’s relatively toy) would probably help substantially with both communicating key ideas and also making research progress.” This work has a lot of research debt, and paying it off would really help clarify the disagreements around these topics.
Thanks Daniel, this is a great summary. I agree that internal representation of the reward function is not load-bearing for the claim. The weak form of representation that you mentioned is what I was trying to point at. I will rephrase the sentence to clarify this, e.g. something like “We assume that the agent learns a goal during the training process: some form of implicit internal representation of desired state features or concepts”.
Thanks Daniel for the detailed response (which I agree with), and thanks Alex for the helpful clarification.
I agree that the training-compatible set is not predictive for how the neural network generalizes (at least under the “strong distributional shift” assumption in this post where the test set is disjoint from the training set, which I think could be weakened in future work). The point of this post is that even though you can’t generally predict behavior in new situations based on the training-compatible set alone, you can still predict power-seeking tendencies. That’s why the title says “power-seeking can be predictive” not “training-compatible goals can be predictive”.
The hypothesis you mentioned seems compatible with the assumptions of this post. When you say “the policy develops motivations related to obvious correlates of its historical reinforcement signals”, these “motivations” seem like a kind of training-compatible goals (if defined more broadly than in this post). I would expect that a system that pursues these motivations in new situations would exhibit some power-seeking tendencies because those correlate with a lot of reinforcement signals.
I suspect a lot of the disagreement here comes from different interpretations of the “internal representations of goals” assumption, I will try to rephrase that part better.
The internal representations assumption was meant to be pretty broad, I didn’t mean that the network is explicitly representing a scalar reward function over observations or anything like that—e.g. these can be implicit representations of state features I think this would also include the kind of representations you are assuming in the maze-solving post, e.g. cheese shards / circuits.
Thanks Alex! Your original comment didn’t read as ill-intended to me, though I wish that you’d just messaged me directly. I could have easily missed your comment in this thread—I only saw it because you linked the thread in the comments on my post.
Your suggested rephrase helps to clarify how you think about the implications of the paper, but I’m looking for something shorter and more high-level to include in my talk. I’m thinking of using this summary, which is based on a sentence from the paper’s intro: “There are theoretical results showing that many decision-making algorithms have power-seeking tendencies.”
(Looking back, the sentence I used in the talk was a summary of the optimal policies paper, and then I updated the citation to point to the retargetability paper and forgot to update the summary...)
Sorry about the cite in my “paradigms of alignment” talk, I didn’t mean to misrepresent your work. I was going for a high-level one-sentence summary of the result and I did not phrase it carefully. I’m open to suggestions on how to phrase this differently when I next give this talk.
Similarly to Steven, I usually cite your power-seeking papers to support a high-level statement that “instrumental convergence is a thing” for ML audiences, and I find they are a valuable outreach tool. For example, last year I pointed David Silver to the optimal policies paper when he was proposing some alignment ideas to our team that we would expect don’t work because of instrumental convergence. (There’s a nonzero chance he would look at a NeurIPS paper and basically no chance that he would read a LW post.)
The subtleties that you discuss are important in general, but don’t seem relevant to making the basic case for instrumental convergence to ML researchers. Maybe you don’t care about optimal policies, but many RL people do, and I think these results can help them better understand why alignment is hard.
Here is my guess on how shard theory would affect the argument in this post:
In my understanding, shard theory would predict that the model learns multiple goals from the training-compatible (TC) set (e.g. including both the coin goal and the go-right goal in CoinRun), and may pursue different learned goals in different new situations. The simplifying assumption that the model pursues a randomly chosen goal from the TC set also covers this case, so this doesn’t affect the argument.
Shard theory might also imply that the training-compatible set should be larger, e.g. including goals for which the agent’s behavior is not optimal. I don’t think this affects the argument, since we just need the TC set to satisfy the condition that permuting reward values in will produce a reward vector that is still in the TC set.
So think that assuming shard theory in this post would lead to the same conclusions—would be curious if you disagree.
Great post! I especially enjoyed the intuitive visualizations for how the heavy-tailed distributions affect the degree of overoptimization of X.
As a possibly interesting connection, your set of criteria for an alignment plan can also be thought of as criteria for selecting a model specification that approximates the ideal specification well, especially trying to ensure that the approximation error is light-tailed.
David had many conversations with Bengio about alignment during his PhD, and gets a lot of credit for Bengio taking AI risk seriously
Thanks Alex for the detailed feedback! I agree that learning a goal from the training-compatible set is a strong assumption that might not hold.
This post assumes a standard RL setup and is not intended to apply to LLMs (it’s possible some version of this result may hold for fine-tuned LLMs, but that’s outside the scope of this post). I can update the post to explicitly clarify this, though I was not expecting anyone to assume that this work applies to LLMs given that the post explicitly assumes standard RL and does not mention LLMs at all.
I agree that reward functions are not the best way to refer to possible goals. This post builds on the formalism in the power-seeking paper which is based on reward functions, so it was easiest to stick with this terminology. I can talk about utility functions instead (which would be equivalent to value functions in this case) but this would complicate exposition. I think it is pretty clear in the post that I’m not talking about reinforcement functions and the training reward is not the optimization target, but I could clarify this further if needed.
I find the idea of a training-compatible goal set useful for thinking about the possible utilities that are consistent with feedback received during training. I think utility functions are still the best formalism we have to represent goals, and I don’t have a clear sense of the alternative you are proposing. I understand what kind of object a utility function is, and I don’t understand what kind of object a value shard is. What is the type signature of a shard—is it a policy, a utility function restricted to a particular context, or something else? When you are talking about a “partial encoding of a goal in the network”, what exactly do you mean by a goal?
I would be curious what predictions shard theory makes about the central claim of this post. I have a vague intuition that power-seeking would be useful for most contextual goals that the system might have, so it would still be predictive to some degree, but I don’t currently see a way to make that more precise.
I’ve read a few posts on shard theory, and it seems very promising and interesting, but I don’t really understand what its claims and predictions are. I expect I will not have a good understanding or be able to apply the insights until there is a paper that makes the definitions and claims of this theory precise and specific. (Similarly, I did not understand your power-seeking theory work until you wrote a paper about it.) If you’re looking to clarify the discourse around RL processes, I believe that writing a definitive reference on shard theory would be the most effective way to do so. I hope you take the time to write one and I really look forward to reading it.
Which definition / result are you referring to?
We expect that an aligned (blue-cloud) model would have an incentive to preserve its goals, though it would need some help from us to generalize them correctly to avoid becoming a misaligned (red-cloud) model. We talk about this in more detail in Refining the Sharp Left Turn (part 2).
Just added some more detail on this to the slides. The idea is that we have various advantages over the model during the training process: we can restart the search, examine and change beliefs and goals using interpretability techniques, choose exactly what data the model sees, etc.
[Linkpost] Some high-level thoughts on the DeepMind alignment team’s strategy
Power-seeking can be probable and predictive for trained agents
Thanks Alex for the detailed feedback! I have updated the post to fix these errors.
Curious if you have high-level thoughts about the post and whether these definitions have been useful in your work.
This post provides a maximally clear and simple explanation of a complex alignment scheme. I read the original “learning the prior” post a few times but found it hard to follow. I only understood how the imitative generalization scheme works after reading this post (the examples and diagrams and clear structure helped a lot).
I agree that a possible downside of talking about capabilities is that people might assume they are uncorrelated and we can choose not to create them. It does seem relatively easy to argue that deception capabilities arise as a side effect of building language models that are useful to humans and good at modeling the world, as we are already seeing with examples of deception / manipulation by Bing etc.
I think the people who think we can avoid building systems that are good at deception often don’t buy the idea of instrumental convergence either (e.g. Yann LeCun), so I’m not sure that arguing for correlated capabilities in terms of intelligence would have an advantage.