Your posts about the neocortex have been a plurality of the posts I've been most excited reading this year. I am super interested in the questions you're asking, and it has long driven me nuts that I don't find these questions asked often in the neuroscience literature.
But there's an aspect of these posts I've found frustrating, which is something like the ratio of "listing candidate answers" to "explaining why you think those candidate answers are promising, relative to nearby alternatives."
Interestingly, I also have this gripe when reading Friston and Hawkins. And I feel like I also have this gripe about my own reasoning, when I think about this stuff—it feels phenomenologically like the only way I know how to generate hypotheses in this domain is by inducing a particular sort of temporary overconfidence.
I don't feel incentivized to do this nearly as much in other domains, and I'm not sure what's going on. My lead hypothesis is that in neuroscience, data is so abundant, and theories/frameworks so relatively scarce, that it's unusually helpful to ignore lots of things—e.g. via the "take as given x, y, z, and p" motion—in order to make conceptual progress. And maybe there's just so much available data here that it would be terribly sisiphean to try to justify all the things one takes as given when forming or presenting intuitions about underlying frameworks. (Indeed, my lead hypothesis for why so many neuroscientists seem to employ strategies like, "contribute to the 'understanding road systems' project by spending their career measuring the angles of stop-sign poles relative to the road," is that they feel it's professionally irresponsible, or something, to theorize about underlying frameworks without first trying to concretely falsify a sisiphean-rock-sized mountain of assumptions).
Still, I think some amount of this motion is clearly necessary to avoid accidentally deluding yourself, and the references in your posts make me think you do at least some of it already. So I guess I just want to politely—and super gratefully, I'm really glad you write these posts regardless! If trying to do this would turn you into a stop sign person, don't do it!—suggest that explicating these more might make it easier for readers to understand and come to share your intuitions.
I have more proto-questions about your model than I have time to flesh them out well enough to describe, but here are some that currently feel top-of-mind:
I feel confused about why, given your model of the situation, the researchers were surprised that this phenomenon occurred, and seem to think it was a novel finding that it will inevitably occur given the three conditions described. Above, you mentioned the hypothesis that maybe they just "weren't very familiar with AI." Looking at the author list, and at their publications (1, 2, 3, 4, 5, 6, 7, 8), this seems implausible to me. While most of the eight co-authors are neuroscientists by training, three have CS degrees (one of whom is Demis Hassabis), and all but one have co-authored previous ML papers. It's hard for me to imagine their surprise was due simply to them lacking basic knowledge about RL?
And this OpenAI paper (whose authors I think you would describe as familiar with ML), which the summary of Wang et al. on the DeepMind website describes as "closely related work," and which appears to me to describe a very similar setup, describes their result in similar terms:
We structure the agent as a recurrent neural network, which receives past rewards, actions, and termination flags as inputs in addition to the normally received observations. Furthermore, its internal state is preserved across episodes, so that it has the capacity to perform learning in its own hidden activations. The learned agent thus also acts as the learning algorithm, and can adapt to the task at hand when deployed.
The OpenAI authors also seem to me to think they can gather evidence about the structure of the algorithm simply by looking at its behavior. Given a similar series of experiments (mostly bandit tasks, but also a maze solver), they conclude:
the dynamics of the recurrent network come to implement a learning algorithm entirely separate from the one used to train the network weights... the procedure the recurrent network implements is itself a full-fledged reinforcement learning algorithm, which negotiates the exploration-exploitation tradeoff and improves the agent’s policy based on reward outcomes... this learned RL procedure can differ starkly from the algorithm used to train the network’s weights.
They then run an experiment designed specifically to distinguish whether meta-RL was giving rise to a model-free system, or “a model-based system which learns an internal model of the environment and evaluates the value of actions at the time of decision-making through look-ahead planning,” and suggest the evidence implies the latter. This sounds like a description of search to me—do you think I'm confused?
I get the impression from your comments that you think it's naive to describe this result as "learning algorithms spontaneously emerge." You describe the lack of LW/AF pushback against that description as "a community-wide failure," and mention updating as a result toward thinking AF members “automatically believe anything written in a post without checking it.”
But my impression is that OpenAI describes their similar result in basically the same way. Do you think my impression is wrong? Or e.g. that their description is also misleading?
I've been feeling very confused lately about how people talk about "search," and have started joking that I'm a search panpsychist. Lots of interesting phenomenon look like piles of thermostats when viewed from the wrong angle, and I worry the conventional lens is deceptively narrow.
That said, when I condition on (what I understand to be) the conventional understanding, it's difficult for me to imagine how e.g. the maze-solver described in the OpenAI paper reliably and quickly locates the exit to new mazes, without doing something reasonably describable as searching for them.
And it seems to me that Wang et al. should be taken as evidence that "learning algorithms producing other search-performing learning algorithms" is convergently useful/likely to be a common feature of future systems, even if you don't think that's what happened in their paper, assuming you assign some credence to their hypothesis that this is what's going on in PFC, and to the hypothesis that search occurs in PFC.
If the primary difference between the DeepMind and OpenAI meta-RL architecture and the PFC/DA architecture is scale, then I think there's reasonable reason to suspect that something much like mesa-optimization will emerge in future meta-RL systems, even if it hasn't yet. That is, I interpret this result as evidence for the hypothesis that highly competent general-ish learners might tend to exhibit this feature, since (among other reasons) it increased my credence that it is already exhibited by the only existing member of that reference class.
Upthread, Evan mentions agreeing that this result is "not new evidence in favor of mesa-optimization." But he also mentions that Risks from Learned Optimization references these two papers, describing them as "the closest to producing mesa-optimizers of any existing machine learning research." I feel confused about how to reconcile these two claims. I didn't realize these papers were mentioned in Risks from Learned Optimization, but if I had, I think I would have been even more inclined to post this/try to ensure people knew about the results, since my (perhaps naive, perhaps not understanding ways this is disanalogous) prior is that the closest existing example to this problem might provide evidence about its nature or likelihood.
I appreciate you writing this, Rohin. I don’t work in ML, or do safety research, and it’s certainly possible I misunderstand how this meta-RL architecture works, or that I misunderstand what’s normal.
That said, I feel confused by a number of your arguments, so I'm working on a reply. Before I post it, I'd be grateful if you could help me make sure I understand your objections, so as to avoid accidentally publishing a long post in response to a position nobody holds.
I currently understand you to be making four main claims:
Does this summary feel like it reasonably characterizes your objection?
I agree, in the case of evolution/humans. In the text above, I meant to highlight what seemed to me like a relative lack of catastrophic *within-mind* inner alignment failures, e.g. due to conflicts between PFC and DA. Death of the organism feels to me like a reasonable way to operationalize "catastrophic" in these cases, but I can imagine other reasonable ways.
As I understand it, your point about the distinction between "mesa" and "steered" is chiefly that in the latter case, the inner layer is continually receiving reward signal from the outer layer, which in effect heavily restricts the space of possible algorithms the outer layer might give rise to. Does that seem like a decent paraphrase?
One of the aspects of Wang et al.'s paper that most interested me was that the inner layer in their meta-RL model kept learning even once reward signal from the outer layer had ceased. It seems reasonable to me to hypothesize that in fact what's going on between PFC and DA is something closer to "subcortex-supervised learning," where PFCs input signals are quite regularly "labeled" by a DA-supervisor. But it doesn't feel intuitively obvious to me that the portion of PFC input which might be labeled in this way is high—e.g., I feel confused about what portion of the concepts currently active in my working memory while writing this paragraph might be labeled by DA—nor that it much restricts the space of possible algorithms that might arise in PFC.
I mean, it could both be the case that there exists catastrophic inner alignment failure between humans and evolution, and also that humans don't regularly experience catastrophic inner alignment failures internally.
In practice I do suspect humans regularly experience internal (within-brain) inner alignment failures, but given that suspicion I feel surprised by how functional humans manage to be. That is, I notice expecting that regular inner alignment failures would cause far more mayhem than I observe, which makes me wonder whether brains are implementing some sort of alignment-relevant tech.
The thing I meant by "catastrophic" is "leading to the death of the organism." I'm suspicious that mesa-optimization is common in humans, although I don't feel confident of that. I can imagine it being the case that many examples of e.g. addiction, goodharting, OCD, and even just everyday "personal misalignment"-type problems of the sort IFS/IDC/multi-agent models of mind sometimes help with, are caused by phenomenon which might reasonably be described as inner alignment failures (although I can also imagine them being caused by more mundane processes).
But I think these things don't kill people very often? People do sometimes choose to die because of beliefs. And anorexia sometimes kills people, which currently feels to me like the most straightforward candidate example I've considered.
Things could be a lot worse. For example, it could be the case that mind-architectures that give rise to mesa-optimization simply aren't viable—that it always kills them. Or e.g. that it always leads to the organism optimizing for a set of goals which is unrecognizably different from the base objective. I don't think you see these things, and I'm interested in figuring out how evolution prevented them.