This post states the problem of gradient hacking. It is valuable in that this problem is far from obvious, and if plausible, very dangerous. On the other hand, the presentation doesn’t go into enough details, and so leaves gradient hacking open to attacks and confusion. Thus instead of just reviewing this post, I would like to clarify certain points, while interweaving my critics about the way gradient hacking was initially stated, and explaining why I consider this problem so important.

(Caveat: I’m not pretending that any of my objections are unknown to Evan. We’re working together on searching for experimental evidence of gradient hacking, and this review is heavily influenced by this collaboration.)

What is gradient hacking?

Evan defines it as:

the phenomenon wherein a deceptively aligned mesa-optimizer might be able to purposefully act in ways which cause gradient descent to update it in a particular way.

So we have a trained mesa-optimizer that is deceptively aligned (its mesa-objective differs from the base-objective it is trained on). This deceptive system finds a way to influence the gradient generated during training, and thus can basically ensure it never becomes inner aligned (among other things).

Before checking on how exactly this could be possible, we should think a bit more about what this implies.

If the deceptive model has arbitrary access to the gradient, then from that point on, the base-objective has only minimal influence on the training. Some influence remains because we probably want to test the end result, and thus it should work well with the base-objective. But that's pretty much the only constraint left.

It could also pretty much deals with deception detectors because it can make itself not detectable:

That is, it could hack its own gradient to produce a new model which will still exhibit the same deceptive behavior, but without doing any of the detectable deceptive cognition. Perhaps the new model exhibits the deceptive behavior not because it's trying to act deceptive, but just because it's a hard-coded part of its algorithm, or because it has some bad proxy, or any other defect that might be more difficult to detect than outright deception.

To say it pithy: if gradient hacking happens, we’re fucked.

How could it happen, though? Well, that’s actually two distinct questions: how could a model gradient hack, and how could training create a model which gradient hacks. The post mostly focuses on the first one.

How could a model gradient hack?

The first example comes from a quoted footnote of Risks from Learned Optimization:

Thus, in the context of a local optimization process, a deceptive mesa-optimizer might try to “hack” its own gradient (by, for example, making itself more brittle in the case where its objective gets changed) to ensure that the base optimizer adjusts it in such a way that leaves its mesa-objective untouched.

This implies that the model is able to check for changes in itself and in its objective, such that changing this objective would make it work worse. An analogue version would be for the model to underperform except if it’s changed in the direction it likes, to push the gradient towards there.

Still, it might feel a bit weird to think about a model that can influence its training without somewhat hacking the computer it is in. After all, the training is external, right? One fundamental but easy to miss insight from the post is that the model can influence the loss surface, and thus the gradient (which comes from this surface) 

First, note that since gradient descent is a local optimization process, it doesn't have to somehow change what the global minima might be, or even what the local minima are—it just has to create a ridge in the loss landscape that will push the gradient in the direction it wants.

So a model could gradient hacks because its output and how it computes its output influence the local loss surface around it, which means it could in theory make the gradient go its way. I feel like this is crucial, and it’s slightly problematic that the post doesn’t put more emphasis on it.

Also note that this is hardly solving the question. Let’s say the model is such that if its mesa-objective is changed, it performs badly. Then we probably have a form of gradient hacking. But what tells us that the change at this step will not slightly undermine the very mechanism that protects the mesa-objective? Unwrapping this thought, it’s harder to conceptualize how a model could gradient hack forever. It seems to involve some sort of fixed point or self correcting mechanism.

That complexity should probably make us update slightly towards the improbability of “sustainable” gradient hacking -- still not enough to dismiss it though. Understanding the requirements for such sustainable gradient hacking looks like the most important research question on this topic for me (which makes sense, I’m working on it after all…).

How could training create a model which gradient hacks?

What about this second question? Well, another issue with the post comes from the lack of clarification around the second question: how could training create a model which gradient hacks? I’m obviously not asking for a full-fledged answer, as this is still research in progress. But there are easy to make mistakes that I believe Evan could have debunked in the post.

One such mistake is to get confused about how the model could “decide” to gradient hack. After all, if the computation is deterministic, it’s not like the model has a decision to make, right? This question can be dissolved in a similar way to free will (since it’s basically about free will of neural networks): (spoiler for the solution to free will)

The less philosophical approach to this problem is to notice that the appearance of gradient hacking would probably come from the training stumbling on a gradient hacker. There is no need to worry about how the model could turn itself into a gradient hacker; the issue is about how training could bump into a gradient hacker.

Another mistake would be to argue that what is called gradient hacking here is indistinguishable in terms of training behavior from some model that doesn’t gradient hack but happens to have this gradient. This is missing a big part of Evan’s viewpoint: that looking inside the model is fundamental for alignment. Here the question is whether we can catch the model doing some form of recognizable gradient hacking, because that allows us to infer more about what it might do (and that it will not be inner aligned). Whereas if we can be relatively sure that a model is not gradient hacking, there are less things to worry about, at least in terms of deceptive alignment.

This is almost obvious for people that talked a lot with Evan on these issues, or who are intimate with his work; but I feel that it’s only implicit in this post, which might brew confusion. A new but very observant reader could infer it from the solution (discussed below) using transparency tools, sure. I still think the post would be stronger if that possible misinterpretation was debunked explicitly.

As for further research, investigating the appearance of gradient hacking through common ML training procedure is crucial. But it probably requires a better understanding of the mechanisms we’re looking for, and thus a convincing answer (at least informal) to the first question.

Possible solution

Evan proposes a solution for gradient hacking in one case:

So, that's the problem—how do we solve it? Well, there's a simple solution in the case above, which is just to run your transparency tools throughout training—not just at the end—so that you can catch the deception as soon as it happens.

This is hardly revolutionary, and doesn’t say how to actually catch deception (although most of Evan research is focused on that). Yet I don’t think much more can be done without understanding in more detail the mechanisms of gradient hacking, and how they appear through training. That is, without solving some significant part of the two previous questions.

Wrapping-up

All in all, I think this post is very important, because the alignment issue it presents is novel, counterintuitive, and hard to dismiss without more research, once one thinks about it for long enough. That being said, this post is definitely a short presentation of an idea, written quickly to draw attention to this issue. Gradient hacking warrants a longer post that digs deeper into the details of the problem,

The original footnote provides one example of this, which is for the model to check if its objective satisfies some criterion, and fail hard if it doesn't. Now, if the model gets to the point where it's actually just failing because of this, then gradient descent will probably just remove that check—but the trick is never to actually get there. By having such a check in the first place, the model makes it so that gradient descent won't actually change its objective, since any change to its objective (keeping all the other parameters fixed, which is what gradient descent does since it computes partial derivatives) would lead to such a failure.

I don't think this argument works. After the agent has made that commitment, it needs to set some threshold for the amount of goal shift that will cause it to fail hard. But until the agent hits that threshold, the gradient will continue to point in the direction of that threshold. And with a non-infinitesimal learning rate, you'll eventually cross that threshold, and the agent will respond by failing hard.

A possible counterargument: the agent's ability to detect and enforce that threshold is not discrete, but also continuous, and so approaching the threshold will incur a penalty. But if that's the case, then the gradients will point in the direction of removing the penalty by reducing the agent's determination to fail upon detecting goal shift.

The way that this might still work is if modifications to this type of high-level commitment are harder to "detect" in partial derivatives than modifications to the underlying goals - e.g. if it's hard to update away from the commitment without reducing the agent's competence in other ways. And this seems kinda plausible, because high-level thought narrows down the space of outcomes sharply. But this is even more speculative.

If your model is deceptive, though, then it might know all of that

Could you please describe your intuition behind how the model could know the meta-optimizer is going to perform checks on deceptive behavior?

adamshimi says almost everything I wanted to say in my review, so I am very glad he made the points he did, and I would love for both his review and the top level post to be included in the book. 

The key thing I want to emphasize a bit more is that I think the post as given is very abstract, and I have personally gotten a lot of value out of trying to think of more concrete scenarios where gradient hacking can occur. 

I think one of the weakest aspects of the post is that it starts with the assumption that an AI system has already given rise to an inner-optimizer that is now taking advantage of gradient hacking. I think while this is definitely a sufficient assumption, I don't think it's a necessary assumption and my current models suggest that we should find this behavior without the need for inner optimizers. This also makes me somewhat more optimistic about studying it. 

My thinking about this is still pretty fuzzy and in its early stages, but the reasoning goes as follows: 

• If we assume the lottery-ticket hypothesis of neural networks, we initialize our network with a large number of possible models of the world. In a sufficiently large networks, some of those models will be accurate models of not necessarily the world, but the training process of the very system that is currently being trained. This is pretty likely given that SGD isn't very complicated and it doesn't seem very hard to build a model of how it works.
• From an evolutionary perspective, we are going to be selecting for networks that get positively rewarded by the gradient descent learning algorithm. Some of the networks that have an accurate model of the training process will stumble upon the strategy of failing hard if SGD would reward any other competing network, creating a small ridge in the reward landscape that results in it itself getting most of the reward (This is currently very metaphorical and I feel fuzzy on whether this conceptualization makes sense). This strategy seems more complicated, so is less likely to randomly exist in a network, but it is very strongly selected for, since at least from an evolutionary perspective it appears like it would give the network a substantive advantage. 
• By default, luckily, this will create something I might want to call a "benign gradient hacker" that might deteriorate the performance of the system, but not obviously give rise to anything like a full inner optimizer. It seems that this strategy is simple enough that you don't actually need anything close to a consequentialist optimizer to run into it, and instead it seems more synonymous to cancer, in that it's a way to hijack the natural selection mechanism of a system from the inside to get more resources, and like cancer seems more likely to just hurt the performance of the overall system, instead of taking systematic control over it.

This makes me think that the first paragraph of the post seems somewhat wrong to me when it says: 

"Gradient hacking" is a term I've been using recently to describe the phenomenon wherein a deceptively aligned mesa-optimizer might be able to purposefully act in ways which cause gradient descent to update it in a particular way.

I think gradient hacking should refer to something somewhat broader that also captures situations like the above where you don't have a deceptively aligned mesa-optimizer, but still have dynamics where you select for networks that adversarially use knowledge about the SGD algorithm for competitive advantage. Though it's plausible that Evan intends the term "deceptively aligned mesa-optimizer" to refer to something broader that would also capture the scenario above.

-----

Separately, as an elaboration, I have gotten a lot of mileage out of generalizing the idea of gradient hacking in this post, to the more general idea that if you have a very simple training process whose output can often easily be predicted and controlled, you will run into similar problems. It seems valuable to try to generalize the theory proposed here to other training mechanisms and study more which training mechanisms are more easily hacked like this, and which one are not. 

I'd be interested in work that further sketches out a scenario in which this could occur. Some particular details that would be interesting:

• Does the agent need to have the ability to change the weights in the neural net that it is implemented in? If so, how does it get that ability?
• How does the agent deal with the fact that the outer optimization algorithm will force it to explore? Or does that not matter?
• How do we deal with the fact that gradients mostly don't account for counterfactual behavior? Things like "Fail hard if the objective is changed" don't work as well if the gradients can't tell that changing the objective leads to failure. That said, maybe gradients can tell that that happens, depending on the setup.
• Why can't the gradient descent get rid of the computation that decides to perform gradient hacking, or repurpose it for something more useful?

I think this post describes an extremely important problem and research directions, and I hope a lot more research and thought goes into this!

ETA: Unless this problem is resolved, I don't see how any AI alignment approach that involves using future ML—that looks like contemporary ML but at an arbitrarily large scale—could be safe.

If the inner optimizer only affects the world by passing predictions to the outer model, the most obvious trick is to assign artificially negative outcomes to states you want to avoid (e g. which the inner optimizer can predict would update the outer model in bad ways) which then never get checked by virtue of being too scary to try. What are the other obvious hacks?

I guess if you sufficiently control the predictions, you can just throw off the pretense and send the outer model the prediction that giving you a channel to the outside would be a great idea.

I am still pretty unconvinced that there is a corruption mechanism that wouldn’t be removed more quickly by SGD than the mesaobjective would be reverted. Are there more recent write ups that shed more light on this?

Specifically, I can’t tell whether this assumes the corruption mechanism has access to a perfect model of its own weights via observation (eg hacking) or via somehow the weights referring to themselves. This is important because if “the mesaobjective weights” are referred to via observation, then SGD will not compute a gradient wrt them (since they’re not part of the differentiation graph). Thus the only “culprit” for bad performance is the corruption mechanism.

Separately, there seems like there might be a chicken and egg thing here? There’s no compelling reason for performance-maximizing SGD to create a gradient-hacking scheme to protect a mesaobjective that fails on the data. If this is a sufficiently crafty agent that it’s considering long-term consequences of its objective being modified, this still doesn’t explain how the agent used its action space to induce a specific change to its weights that implements a specific multi-input function. I really struggle to see this; do you have toy examples where this could occur? (The answer presumably can’t be hacking, because then we’re fucked for independent reasons.)

Note I’m also assuming gradient hacking is distinct from “avoid states that might change your mesaobjective”, which seems much more likely to happen.