LeCun’s “A Path Towards Autonomous Machine Intelligence” has an unsolved technical alignment problem

Nice, I was actually just thinking that someone needed to respond to LeCun's proposal.

That said, I think you may have gotten some of the details wrong. I don't think the intrinsic cost module gets raw sensory data as input, but instead it gets input from the latent variables of the world model as well as the self-supervised perception module. This complicates some of the safety problems you suggest.

[-]Steven Byrnes2y*Ω330

Thanks. That did actually occur to me, but I left it out because I wasn’t sure and didn’t want to go on an exhausting chase down every possible interpretation of the paper.

Anyway, if the input to the Prosociality Score Model is a set of latent variables rather than a set of pixels then:

My OP claim that there are two adversarial out-of-distribution generalization problems (in the absence of some creative solution not in the paper) is still true.
One of those two problems (OOD generalization of the Prosociality Score Model) might get less bad, although I don’t see why it would go away altogether.
…But only if the labels are correct, and the labeling problem is potentially much harder now, because the latent variables include inscrutable information about “how the AI is thinking about / conceptualizing the things that it’s seeing / doing”. I think. And if they do, then how are the humans supposed to label them as good or bad? Like, if the AI notices someone feeling physically good but psychologically distressed, we want to label it as low-energy when the AI is thinking about the former aspect and high-energy if the AI is thinking about the latter aspect, I imagine. And then we start getting into nasty neural net interpretability challenges.

Also, aren't the latent variables changing as we go, thanks to self-supervised learning? But the Intrinsic Cost Module is supposed to be immutable. I'm confused about how this is supposed to work.

[-]Roman Leventov2y20

But only if the labels are correct, and the labeling problem is potentially much harder now, because the latent variables include inscrutable information about “how the AI is thinking about / conceptualizing the things that it’s seeing / doing”. I think. And if they do, then how are the humans supposed to label them as good or bad? Like, if the AI notices someone feeling physically good but psychologically distressed, we want to label it as low-energy when the AI is thinking about the former aspect and high-energy if the AI is thinking about the latter aspect, I imagine. And then we start getting into nasty neural net interpretability challenges.

I believe the idea was (though this could be a false memory of my interpretation because I cannot find it in the paper now) that separate NNs would be trained to interpret world model representations (states) as human-understandable text. At the very least, this could make the brute size of the interpretability task much smaller because interpreting the internal activations and circuits of the world model or actor NNs producing these representations might not matter. This is related to Bengio's idea of separating knowledge/models from the "inference" network, where the latter could be much larger. This idea is, of course, embodied in GFlowNets, and indeed it seems to me that they could be used in place of the "world model" in the APTAMI (although LeCun doesn't like H-JEPA to be probabilistic).

However, I don't know how to bootstrap the training of these interpretability modules for representations alongside the critic, but this doesn't seem like an insurmountable engineering problem.

Also, aren't the latent variables changing as we go, thanks to self-supervised learning? But the Intrinsic Cost Module is supposed to be immutable. I'm confused about how this is supposed to work.

To me, "mutability" / online trainability of various parts of the architecture, or AI as a whole seems more like a spectrum than a binary distinction, and in fact not a very important one. You can always think of "immutable" modules as just "learning very slowly, on the level of AI engineering/training iterations", a.k.a. "model selection".

So, the "immutability" of the intrinsic cost just highlights that these things should be inferred "very slowly, with governance and caution", but you can basically configure the learning rate of other parts of the architecture, such as the world model, to also be very low, so that all the surrounding infrastructure: trainable critic modules, the "altruism controller" that I proposed, interpretability modules, alignment processes and protocols, and, finally, the human understanding (i.e., "theories of AI mind" in human heads) all can "track" the changes in the world model and its representations sufficiently closely so that they don't "fall out of alignment" with them.

Actually, the fact that NNs in human brains en masse should track at least some changes in AI's world models "in time" and not fall out of alignment with them practically means that the rate of learning in the world models on the high levels of the JEPA hierarchy should be so slow that they de facto could be "immutable" on the level of a particular AI "version"/iteration, and only "learned"/updated between versions, as well as the Intrinsic Cost module. Moreover, it might need to be much slower than even the typical technology release cadence (such as "a new version of AI every 6 months"), e.g., not changing much on any 5-year interval, which seems to me the minimum realistic time in which humans en masse could adjust to a new kind of alien mind around them.

[-]Roman Leventov2y10

Update: I wrote a big article "Aligning an H-JEPA agent via training on the outputs of an LLM-based "exemplary actor" in which I develop the thinking behind the comment above (but also update it significantly).

[-]Seth Herd2y83

I agree with everything you've said here.

I think LeCun has probably ignored the alignment problem (while saying he's addressing it) not because of an intellectual oversight, but for emotional reasons.

LeCun's dismissal of alignment as a serious issue seems to me to be a polarization issue. I'm guessing he dislikes the whole topic because he feels that the people who are pushing it might slow progress in AI.

I think his beliefs are sincere, but poorly thought through. I think they're poorly thought through because he's developed and ugh field around arguments for AI risk.

I don't know exactly what LeCun's feelings are, nor am I certain this is the largest part of his reasons for taking his hardline anti-risk stance. I think it's important, though, because LeCun is the biggest name and loudest voice on that side of the discussion, and he is helping form the nucleus of a camp that is polarized against the AI safety issue. Engaging in the discussion in ways that avoid polarization is probably an important consideration.

To that end, if LeCun could actually be engaged in this discussion, that would be a big step toward preventing polarization.

[-]Ben Amitay2yΩ252

Didn't read the original paper yet, but from what you describe, I don't understand how the remaining technical problem is not basically the whole of the alignment problem. My understanding of what you say is that he is vague about the values we want to give the agent - and not knowing how to specify human values is kind of the point (that, and inner alignment - which I don't see addressed either).

[-]Steven Byrnes2yΩ241

I don't understand how the remaining technical problem is not basically the whole of the alignment problem

Yes. I don’t think the paper constitutes any progress on the alignment problem. (No surprise, since it talks about the problem for only a couple sentences.)

Hmm, maybe you’re confused that the title refers to “an unsolved technical alignment problem” instead of “the unsolved technical alignment problem”? Well, I didn’t mean it that way. I think that solving technical alignment entails solving a different (albeit related) technical problem for each different possible way to build / train AGI. The paper is (perhaps) a possible way to build / train AGI, and therefore it has an alignment problem. That’s all I meant there.

[-]Ben Amitay2yΩ110

Yes, I think that was it; and that I did not (and still don't) understand what about that possible AGI architecture is non-trivial and has a non-trivial implementations for alignment, even if not ones that make it easier. It seem like not only the same problems carefully hidden, but the same flavor of the same problems on plain sight.

[-]Steven Byrnes2yΩ461

I think of my specialty as mostly “trying to solve the alignment problem for model-based RL”. (LeCun’s paper is an example of model-based RL.) I think that’s a somewhat different activity than, say, “trying to solve the alignment problem for LLMs”. Like, I read plenty of alignmentforum posts on the latter topic, and I mostly don’t find them very relevant to my work. (There are exceptions.) E.g. the waluigi effect is not something that seems at all relevant to my work, but it’s extremely relevant to the LLM-alignment crowd. Conversely, for example, here’s a random recent post I wrote that I believe would be utterly useless to anyone focused on trying to solve the alignment problem for LLMs.

A big difference is that I feel entitled to assume that there’s a data structure labeled “world-model”, and there’s a different data-structure labeled “value function” (a.k.a. “critic”). Maybe each of those data structures is individually a big mess of a trillion uninterpretable floating-point numbers. But it still matters that there are two data structures, and we know where each lives in memory, and we know what role each is playing, how it’s updated, etc. That changes the kinds of detailed interventions that one might consider doing. [There could be more than two data structures, that’s just an example.]

[-]Ben Amitay2yΩ110

I see. I didn't fully adapt to the fact that not all alignment is about RL.

Beside the point: I think those labels on the data structures are very confusing. Both the actor and the critic are very likely to have so specialized world models (projected from the labeled world model) and planning abilities. The values of the actor need not be the same as the output of the critic. And things value-related and planning-related may easily leak into the world model if you don't actively try to prevent it. So I suspect that we should ignore the labels and focus on architecture and training methods.

[-]Steven Byrnes2yΩ220

Sure, we can take some particular model-based RL algorithm (MuZero, APTAMI, the human brain algorithm, whatever), but instead of “the reward function” we call it “function #5829”, and instead of “the value function” we call it “function #6241”, etc. If you insist that I use those terms, then I would still be perfectly capable of describing step-by-step why this algorithm would try to kill us. That would be pretty annoying though. I would rather use the normal terms.

I’m not quite sure what you’re talking about (“projected from the labeled world model”??), but I guess it’s off-topic here unless it specifically applies to APTAMI.

FWIW the problems addressed in this post involve the model-based RL system trying to kill us via using its model-based RL capabilities in the way we normally expect—where the planner plans, and the critic criticizes, and the world-model models the world, etc., and the result is that the system makes and executes a plan to kill us. I consider that the obvious, central type of alignment failure mode for model-based RL, and it remains an unsolved problem.

In addition, one might ask if there are other alignment failure modes too. E.g. people sometimes bring up more exotic things like the “mesa-optimizer” thing where the world-model is secretly harboring a full-fledged planning agent, or whatever. As it happens, I think those more exotic failure modes can be effectively mitigated, and are also quite unlikely to happen in the first place, in the particular context of model-based RL systems. But that depends a lot on how the model-based RL system in question is supposed to work, in detail, and I’m not sure I want to get into that topic here, it’s kinda off-topic. I talk about it a bit in the intro here.

[-]Ben Amitay2y10

Sorry for the off-topicness. I will not consider it rude if you stop reading here and reply with "just shut up" - but I do think that it is important:

A) I do agree that the first problem to address should probably be misalignment of the rewards to our values, and that some of the proposed problems are not likely in practice - including some versions of the planning-inside-worldmodel example.

B) I do not think that planning inside the critic or evaluating inside the actor are an example of that, because the functions that those two models are optimized to approximate reference each other explicitly in their definitions. It doesn't mean that the critic is likely to one day kill us, just that we should take it into account when we try to I do understand what is going on.

C) Specifically, it implies 2 additional non-exotic alignment failures:

The critic itself did not converge to be a good approximation of the value function.
The actor did not converge to be a thing that maximize the output of the critic, and it maximize something else instead.

[-]Roman Leventov2yΩ340

My comments on this post winded up into a whole separate post: "H-JEPA might be technically alignable in a modified form".

[-]PeterMcCluskey2yΩ120

Why assume LeCun would use only supervised learning to create the IC module?

If I were trying to make this model work, I'd use mainly self-supervised learning that's aimed at getting the module to predict what a typical human would feel. (I'd also pray for a highly multipolar scenario if I were making this module immutable when deployed.)

[-]Steven Byrnes2yΩ220

If I were trying to make this model work, I'd use mainly self-supervised learning that's aimed at getting the module to predict what a typical human would feel.

I don’t follow. Can you explain in more detail? “Self-supervised learning” means training a model to predict some function / subset of the input data from a different function / subset of the input data, right? What’s the input data here, and what is the prediction target?

[-]PeterMcCluskey2yΩ120

I haven't thought this out very carefully. I'm imagining a transformer trained both to predict text, and to predict the next frame of video.

Train it on all available videos that show realistic human body language.

Then ask the transformer to rate on a numeric scale how positively or negatively a human would feel in any particular situation.

This does not seem sufficient for a safe result, but implies that LeCun is less nutty than your model of him suggests.

[-]Steven Byrnes2y*Ω220

Then ask the transformer to rate on a numeric scale how positively or negatively a human would feel in any particular situation.

I'm still confused. Here you're describing what you're hoping will happen at inference time. I'm asking how it's trained, such that that happens. If you have a next-frame video predictor, you can't ask it how a human would feel. You can't ask it anything at all - except "what might be the next frame of thus-and-such video?". Right?

I wonder if you've gotten thrown off by chatGPT etc. Those are NOT trained by SSL, and therefore NOT indicative of how SSL-trained models behave. They're pre-trained by SSL, but then they're fine-tuned by supervised learning, RLHF, etc. The grizzled old LLM people will tell you about the behavior of pure-SSL models, which everyone used before like a year ago. They're quite different. You cannot just ask them a question and expect them to spit out an answer. You have to prompt them in more elaborate ways.

(On a different topic, self-supervised pre-training before supervised fine-tuning is almost always better than supervised learning from random initialization, as far as I understand. Presumably if someone were following the OP protocol, which involves a supervised learning step, then they would follow all the modern best practices for supervised learning, and “start from a self-supervised-pretrained model” is part of those best practices.)

[-]red75prime2yΩ230

If you have a next-frame video predictor, you can't ask it how a human would feel. You can't ask it anything at all - except "what might be the next frame of thus-and-such video?". Right?

Not exactly. You can extract embeddings from a video predictor (activations of the next-to-last layer may do, or you can use techniques, which enhance semantic information captured in the embeddings). And then use supervised learning to train a simple classifier from an embedding to human feelings on a modest number of video/feelings pairs.

[-]Steven Byrnes2yΩ220

I think that’s what I said in the last paragraph of the comment you’re responding to:

(On a different topic, self-supervised pre-training before supervised fine-tuning is almost always better than supervised learning from random initialization, as far as I understand. Presumably if someone were following the OP protocol, which involves a supervised learning step, then they would follow all the modern best practices for supervised learning, and “start from a self-supervised-pretrained model” is part of those best practices.)

Maybe that’s what PeterMcCluskey was asking about this whole time—I found his comments upthread to be pretty confusing. But anyway, if that’s what we’ve been talking about all along, then yeah, sure. I don’t think my OP implied that we would do supervised learning from random initialization. I just said “use supervised learning to train an ML model”. I was assuming that people would follow all the best practices for supervised learning—self-supervised pretraining, data augmentation, you name it. This is all well-known stuff—this step is not where the hard unsolved technical problems are. I’m open to changing the wording if you think the current version is unclear.

[-]Review Bot1y*10

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?

[-][deactivated]2yΩ010

Now, it doesn’t immediately follow that the AI will actually want to start buying chair-straps and heroin, for a similar reason as why I personally am not trying to get heroin right now.

This seems important to me. What is the intrinsic cost in a human brain like mine or yours? Why don’t humans have an alignment problem (e.g. if you radically enhanced human intelligence, you wouldn’t produce a paperclip maximiser)?

Maybe the view of alignment pessimists is that the paradigmatic human brain’s intrinsic cost is intractably complex. I don’t know. I would like more clarity on this point.

[-]Steven Byrnes2yΩ240

The “similar reason as why I personally am not trying to get heroin right now” is “Example 2” here (including the footnote), or a bit more detail in Section 9.5 here. I don’t think that involves an idiosyncratic anti-heroin intrinsic cost function.

The question “What is the intrinsic cost in a human brain” is a topic in which I have a strong personal interest. See Section 2 here and links therein. “Why don’t humans have an alignment problem” is sorta painting the target around the arrow I think? Anyway, if you radically enhanced human intelligence and let those super-humans invent every possible technology, I’m not too sure what you would get (assuming they don’t blow each other to smithereens). Maybe that’s OK though? Hard to say. Our distant ancestors would think that we have awfully weird lifestyles and might strenuously object to it, if they could have a say.

Maybe the view of alignment pessimists is that the paradigmatic human brain’s intrinsic cost is intractably complex.

Speaking for myself, I think the human brain’s intrinsic-cost-like-thing is probably hundreds of lines of pseudocode, or maybe low thousands, certainly not millions. (And the part that’s relevant for AGI is just a fraction of that.) Unfortunately, I also think nobody knows what those lines are. I would feel better if they did. That wouldn’t be enough to make me “optimistic” overall, but it would certainly be a step in the right direction. (Other things can go wrong too.)

[-][deactivated]2yΩ010

...I think the human brain’s intrinsic-cost-like-thing is probably hundreds of lines of pseudocode, or maybe low thousands, certainly not millions. (And the part that’s relevant for AGI is just a fraction of that.) Unfortunately, I also think nobody knows what those lines are. I would feel better if they did.

So, the human brain's pseudo-intrinsic cost is not intractably complex, on your view, but difficult to extract.

[-]Steven Byrnes2yΩ220

I would say “the human brain’s intrinsic-cost-like-thing is difficult to figure out”. I’m not sure what you mean by “…difficult to extract”. Extract from what?

[-][deactivated]2yΩ010

Extract from the brain into, say, weights in an artificial neural network, lines of code, a natural language "constitution", or something of that nature.

[-]Steven Byrnes2yΩ220

“Extract from the brain” how? A human brain has like 100 billion neurons and 100 trillion synapses, and they’re generally very difficult to measure, right? (I do think certain neuroscience experiments would be helpful.) Or do you mean something else?

[-][deactivated]2yΩ010

I meant "extract" more figuratively than literally. For example, GPT-4 seems to have acquired some ability to do moral reasoning in accordance with human values. This is one way to (very indirectly) "extract" information from the human brain.

[-]Steven Byrnes2yΩ230

GPT-4 is different from APTAMI. I'm not aware of any method that starts with movies of humans, or human-created internet text, or whatever, and then does some kind of ML, and winds up with a plausible human brain intrinsic cost function. If you have an idea for how that could work, then I'm skeptical, but you should tell me anyway. :)

[-]MichaelStJules2y*-40

For starters, suppose the AI straps lots of humans into beds, giving them endless morphine and heroin IV drips, and the humans get into such a state of delirium that they repeatedly praise and thank the AI for continuing to keep the heroin drip turned on.
This dystopian situation would be, to the AI, absolute ecstasy—much like the heroin to those poor humans.

This seems to require some pretty important normative claims that seem controversial in EA and the rationality community. Based on your description, it seems like the humans come to approve of this (desire/prefer this) more than they approved of their lives before (or we could imagine similar scenarios where this is the case), and gain more pleasure from it, and you could have their approval (by assumption) outweigh the violation of the preferences for this not to happen. So, if you're a welfarist consequentialist and a hedonist or a desire/preference theorist, and unless an individual's future preferences count much less, this just seems better for those humans than what normal life has been like lately.

Some ways out seem to be:

Maybe certain preference-affecting views, or discounting future preferences, or antifrustrationism (basically negative preference utilitarianism), or something in these directions
Counting preferences more or less based on their specific contents, e.g. wanting to take heroin is a preference that counts less in your calculus
Non-hedonist and non-preferential/desire-based welfare (possibly in addition to hedonistic and preferential/desire-based welfare), e.g. objective goods/bads
Non-welfarist consequentialist values, i.e. valuing outcomes for reasons other than how they matter to individuals' welfare
Non-consequentialism, e.g. constraints on violating preferences or consent, or not getting affirmative consent
Actually, whatever preferences they had before and were violated were stronger than the ones they have now and are satisfied
Actually, we should think much bigger; maybe we should optimize with artificial consciousness instead and do that a lot more.

If it's something like 1 or 5, it should instead (or also?) model what the humans already want, and try to get that to happen.

[-]Steven Byrnes2y40

Let’s please not have an object-level debate of hedonic utilitarianism (or whatever) here. It’s very off-topic.

I personally am strongly against strapping humans into chairs on heroin drips. If anyone reading this is in favor of the heroin-drip thing (which seems unlikely to me but what do I know?), then bad news for them is that there’s no strong reason to expect that from APTAMI either.

My discussion there in OP was not a prediction of what APTAMI will lead to, it was an illustrative example of one of a near-infinite number of weird edge cases in which its classifiers would give undesired results. In particular, I think hedonic utilitarians would also sorely regret turning on an APTAMI system as described. There is no moral philosophy that anyone subscribes to, such that they would have a strong reason to believe that turning on APTAMI would not be a catastrophic mistake by their own lights.

[-]MichaelStJules2y*10

I think your scenario only illustrates a problem with outer alignment (picking the right objective function), and I think it's possible to state an objective, that if it could be implemented sufficiently accurately and we could guarantee the AI followed it (inner alignment), would not result in a dystopia like this. If you think the model would do well at inner alignment if we fixed the problems with outer alignment, then it seems like a very promising direction and this would be worth pointing out and emphasizing.

I think the right direction is modelling how humans now (at the time before taking the action), without coercion or manipulation, would judge the future outcome if properly informed about its contents, especially how humans and other moral patients are affected (including what happened along the way, e.g. violations of consent and killing). I don't think you need the coherency of coherent extrapolated volition, because you can still capture people finding this future substantially worse than some possible futures, including the ones where the AI doesn't intervene, by some important lights, and just set a difference-making ambiguity averse objective. Or, maybe require it not to do substantially worse by any important light, if that's feasible: we would allow the model to flexibly represent the lights by which humans judge outcomes where doing badly in one means doing badly overall. Then it would be incentivized to focus on acts that seem robustly better to humans.

I think an AI that actually followed such an objective properly would not, by the lights of whichever humans whose judgements it's predicting, increase the risk of dystopia through its actions (although building it may be s-risky, in case of risks of minimization). Maybe it would cause us to slow moral progress and lock in the status quo, though. If the AI is smart enough, it can understand "how humans now (at the time before taking the action), without coercion or manipulation, would judge the future outcome if properly informed about its contents", but it can still be hard to point the AI at that even if it did understand it.

Another approach I can imagine is to split up the rewards into periods, discount them temporally, check for approval and disapproval signals in each period, and make it very costly relative to anything else to miss one approval or receive a disapproval. I describe this more here and here. As JBlack pointed out in the comments of the second post, there's incentive to hack the signal. However, as long as attempts to do so are risky enough by the lights of the AI and the AI is sufficiently averse to losing approval or getting disapproval and the risk of either is high enough, it wouldn't do it. And of course, there's still the problem of inner alignment; maybe it doesn't even end up caring about approval/disapproval in the way our objective says it should out of distribution.

[-]Steven Byrnes2y*30

I’m really making two claims in this post:

First, that the vague proposal for Intrinsic Cost in APTAMI will almost definitely lead to AIs that want to kill us.
Second, that nobody has a better proposal for Intrinsic Cost that’s good enough that we would have a strong reason to believe that the AI won’t want to kill us.

Somewhere in between those two claims is a question of whether it’s possible to edit the APTAMI proposal so that it’s less bad—even if it doesn’t meet the higher bar of “strong reason to believe it won’t won’t want to kill us”. My answer is “absolutely yes”. The APTAMI proposal is so bad that I find it quite easy to think of ways to make it less bad. The thing you mentioned (i.e. that “when perceiving joy in nearby humans” is a poorly-thought-through phrase and we can do better) is indeed one example.

My main research topic is (more-or-less) how to write an Innate Cost module, sometimes directly and sometimes indirectly. I don’t currently have any plan that passes the bar of “strong reason to believe the AI won’t want to kill us”. I do have proposals that seem to pass the much lower bar of “it seems at least possible that the AI won’t want to kill us”—see here for a self-contained example. The inner alignment problem is definitely relevant.

[-]Roman Leventov2y10

I believe Steven didn't imply that a significant number of people would approve or want such a future -- indeed, the opposite, hence he called the scenario "dystopian".

He basically meant that optimising surface signals of pleasure does not automatically lead to behaviours and plans congruent with reasonable ethics, so the surface elements of alignment suggested by LeCun in the paper are clearly insufficient.

[-]MichaelStJules2y1-5

I think many EAs/rationalists shouldn't find this to be worse for humans than life today on the views they apparently endorse, because each human looks better off under standard approaches to intrapersonal aggregation: they get more pleasure, less suffering, more preference satisfaction (or we can imagine some kind of manipulation to achieve this), but at the cost of some important frustrated preferences.

[-]rotatingpaguro2y30

EA/rationalists seems to me the community who gives more conscious attention to the problem of following your officially stated preferences down a cliff, a.k.a. Goodhart.

[-]MichaelStJules2y10

FWIW, I'm not sure if that's true relative to the average person, and I'd guess non-consequentialist philosophers are more averse to biting bullets than the average EA and maybe rationalist.

[-]rotatingpaguro2y10

I suspect that you read "conscious attention to the problem" in a different light than what I mean. To clarify:

Average person won't go down a cliff on literal instructions, unless Moloch? Yes.

Average person will identify and understand such problem? No.

EA bites a bullet and does something weird? Yes.

EA bites a bullet because YAY BULLETS COOL? No.

^{^}

I note that these two out-of-distribution problems correspond respectively with [aspects of] what I call “outer alignment” and “inner alignment” in this post. “Outer alignment” (in this context) is the question “Is the Intrinsic Cost module returning high vs low intrinsic energy outputs in a way that tracks the extent to which the agent is doing things that it was intended to do?” And “inner alignment” (in this context) is the question “When the AI imagines some possible future plan, does the plan seem appealing / unappealing to the AI in a way that actually tracks its expected future Intrinsic Cost?” By the way, as discussed at that link, it’s not necessarily the case that the best way to get good AI behavior is to separately solve both these out-of-distribution generalization problems; for example, if we have sufficient neural network interpretability of the trained critic, then we get to slice through both layers, bridging directly from the design intentions to the AI’s motivations, without relying on the Intrinsic Energy function being perfect.

^{^}

Maybe you’re thinking: “OK, we’ll design the AI such that, if something seems like an out-of-the-box idea, then the AI doesn’t want to do it. The AI wants to stay in the box!” Or in conventional ML terms, if we’re worried about out-of-distribution problems, then we can just put in a penalty term that makes the AI want to stay in-distribution. I do actually think there’s a kernel of a promising research direction here, but I don’t know how to flesh it out into a plausible plan. In particular, the most obvious approaches along these lines would have an unintended side-effect of crippling the AI’s ability to learn new things, make new connections, do R&D, etc. Further discussion in Section 14.4 here.

LESSWRONG
LW

LESSWRONG
LW

140

LeCun’s “A Path Towards Autonomous Machine Intelligence” has an unsolved technical alignment problem

140

Ω 42

140

Ω 42

Summary

1. Background: the paper’s descriptions of the “Intrinsic Cost module”

2. As described in the paper, several components of the AI’s Intrinsic Cost are directly opposed to AI controllability and prosociality

3. As described in the paper, the components of the AI’s Intrinsic Cost that are supposed to motivate intrinsic kindness, are unlikely to actually work

4. Conclusion

Epilogue: We need to do better than a cursory treatment of this technical problem, even in the context of a very-early-stage speculative vision paper