So here's a post I spent the past two months writing and rewriting. I abandoned this current draft after I found out that my thesis was empirically falsified three years ago by this paper, which provides strong evidence that transformers implement optimization algorithms internally. I'm putting this post up anyway as a cautionary tale about making clever arguments rather than doing empirical research. Oops.

Edit: never mind lol, it's more accurate to say the paper confirms that a certain one-layer, simplified transformer-like architecture learns to implement an internal gradient descent-like algorithm if trained on a particular kind of task, and even then only because the here-simplified attention mechanism itself has a neat trick for compactly performing a single gradient descent step. my argument about it being much easier for genes to implement looping algorithms in general still seems to hold; i could have just added a footnote about this and finished the post as planned. double oops.

1. Overview

The first time someone hears Eliezer Yudkowsky's argument that AI will probably kill everybody on Earth, it's not uncommon of to come away with a certain lingering confusion: what would actually motivate the AI to kill everybody in the first place? It can be quite counterintuitive in light of how friendly modern AIs like ChatGPT appear to be, and Yudkowsky's argument seems to have a bit of trouble changing people's gut feelings on this point.^[1] It's possible this confusion is due to the inherent complexity of the topic, but my instinct is that it's worth cross-examining Yudkowsky's argument that AIs will acquire aggressive, inscrutable values a bit more closely.

The bulk of Yudkowsky's argument consists of an analogy between deep learning and natural selection, the natural selection half of which is fairly simple. Evolution ended up giving humans values that weren't aligned with the general trends of natural selection; i.e. we value sex for its own sake, rather than as a deliberate strategy for improving our genetic fitness. This gap between our values and the trends of evolution meant that, once we got smart enough to invent condoms, we did. Now we can enjoy sex without the pesky risk of it actually resulting in us having any offspring.

So how does this tie into AI doom? Well, Yudkowsky thinks something very similar is going to happen with deep learning systems. Even if we're very careful to reward, say, an LLM for being helpful, honest, and harmless, the model might acquire unpredictable values of its own during the training process. Just like how humans started to value sex for its own sake, an LLM might start to acquire values that only incidentally improve performance per the training objective.

Now, if such an LLM was placed in a new situation (e.g. one where it had better technology at its fingertips), it might come up with novel strategies for fulfilling its values. And those strategies might cease to have the helpful side effect of improving the model's performance. In fact, this might lead to outcomes its creators wouldn't endorse at all. To describe especially catastrophic outcomes of this dynamic, which Yudkowsky considers likely, Yudkowsky often says: "The AI does not hate you, nor does it love you; but you are made of atoms that it can use for something else."

Stepping into my perspective, though, I think the conclusions Yudkowsky draws from this line of reasoning are too strong. I agree that both gradient descent and natural selection have mechanisms for imbuing systems with values that are temporarily useful, but which can ultimately turn out to be misaligned. However, the details of these mechanisms are importantly different, and constrain the kinds of misaligned value systems each type of system could plausibly end up with.

For example, there are technical reasons for genes in particular to have ended up imbuing us with values by means of explicit, internal optimization algorithms (e.g. our reward circuits) that run inside the human brain. Similarly, there are technical reasons why a similar outcome doesn't make sense in neural networks (a fact which rules out certain entrenched stories about how LLMs kill everybody).

An additional claim I want to make is that there are good, mechanistic reasons to expect an LLM's genuine values to be predictable in light of their training data, not to mention much less relentless than those of a utility maximizer. This would imply that the values of these systems ultimately remain relatively safe and steerable.^[2]

The rest of this post is going to elaborate on each of the ideas I've introduced in this section, in the order that I've introduced them:

1. Elucidating the evolution analogy itself in more detail.
2. Discussing the properties of natural selection that caused it to produce an "inner optimizer" in the sense advocated by Yudkowsky's coworkers at MIRI, and the absence of those properties in deep learning systems.
3. Laying out a mechanistic analysis of what's really going on under gradient descent, and why it can be expected that the values that are suggested by LLMs' apparent (and intended) characters or personalities are unlikely to be hiding blatant, aggressive misalignments beneath the surface.

Overall, my thesis is that an LLM's character or personality ought to feature much more relaxed values than those possessed by expected utility maximizers, not to mention that said personality is likely to be quite predictable and steerable by its creators. There are important caveats to this second claim, mostly related to Microsoft's Bing-Sydney and the recent emergent misalignment paper; I'll address these in the final section of this post. For the most part, though, my instinct is that the alignment problem has turned out to be notably easier than alignment researchers would have expected a decade ago, and the remaining problems have a palpable air of tractability. Given diligence, the character of the superintelligence could be ours for the shaping.

2. The details of the evolution analogy

(Citational note: Yudkowsky has frequently made an argument very close to the one I outline below. Some particular sources include item 16 in his threat modeling post, AGI Ruin: A List of Lethalities, as well as the timestamp-linked segment of his appearance on the Bankless podcast, We're All Gonna Die with Eliezer Yudkowsky.)

Before we get into my original thoughts about the deep learning/natural selection comparison, it's best to ensure we've laid down a clear account of the analogy as articulated by Yudkowsky. To that end, probably the most important comparison to establish is the analogy between an organism's genes and a neural network's weights. Just as natural selection optimizes an organism's genome for evolutionary fitness, deep learning optimizes a neural network's weights for whatever task the system is being trained on. Each is a kind of parameter being refined by a hill-climbing process; this forms the basis for the entire rest of this comparison.

The second important comparison is between an organism's "ancestral environment" and a neural network's "training distribution." These are, respectively, the kinds of environments that each system was developed in order to deal with. For organisms, this means having accumulated genetic mutations which were helpful for surviving in the worlds inhabited by their ancestors. For neural networks, this means having weights designed to perform well in whatever kinds of training scenarios the network was exposed to. In both cases, the relevant concept is that there's a certain type of environment to which the system in question is adapted, and outside of which the system will flounder.

This brings us to a third concept that applies to both deep learning and natural selection: so-called "distributional shift." A land organism may be somewhat well-adapted to a certain distribution of environments such as forests, deserts, and beaches; however, if you were to drop it into the ocean, it could very well drown immediately. Similarly, a neural network could be trained to predict text from, say, classic works of literature, and perform relatively well when predicting works written in the canonical styles; however, its predictions would do worse when continuing from a modern young adult novel, let alone something like stock market documentation. All of these failure-modes fall under the umbrella of distributional shift.^[3]

Now, in most cases the outcomes of distributional shift are fairly uninteresting: just plain bad performance per the training objective (e.g. inclusive genetic fitness, or loss in terms of prediction error). If low-quality outputs were the full extent of the problems with distributional shifts, though, they wouldn't give us much reason to worry about advanced neural networks causing the extinction of humanity. The real risk mostly comes from a particular strategy a hill-climbing process might adopt during the optimization process: instilling us with coherent values of our own, ones that we can produce novel strategies for pursuing by means of our general intelligence.

After all, humans are empirical proof that hill-climbing can imbue such values into a system, improving our performance a lot in the ancestral environment, but less so in environments where we can realize our values more fully, and potentially even degrading performance. In the intro, we already considered the example of birth control, which reveals how the fact that we value sex for its own sake has become less valuable for our reproductive fitness as our technological capabilities have grown. For another example, consider environments where we have an incredible abundance of delicious food. Here, the fact that we value delicious food can actually backfire and make us less reproductively fit than we otherwise would be, because eating too much can make you less sexually attractive.

The situation continues to get even worse as our surroundings continue enabling us to fulfill our true values more and more completely. With sufficiently advanced technology, humans might eventually get addicted to wireheading,^[4] and thereby cease to engage in any reproductive activity whatsoever. Another possibility is that we decide to upload our minds into superior artificial bodies, ones that lack our genomes altogether. This would be a catastrophe from the perspective of furthering the reproduction of our genes.

Overall the failure-mode is this: although intelligently pursuing values was helpful in the ancestral environment, it can lead to catastrophically misaligned behavior in distributions where we're capable of more fully achieving our misaligned goals. It's a classic case of Goodhart's law: When a measure (of, say, genetic fitness) becomes a target (by, say, making us value sex in and of itself), it ceases to be a good measure.

It's worth noting that it's not just having more advanced technology in your environment that can cause this problem; becoming more intelligent can cause issues as well. After all, smarter creatures are better at inventing new technology (altering the environment to empower themselves), or exploiting whatever they're already surrounded by. It can be a little weird to think of increased intelligence as constituting a distributional shift, because intelligence is inside an agent whereas the distribution is outside the agent, but this conceptual wrinkle shouldn't matter for assessing the overall analogy. "The agent has gotten more intelligent" is still a novel situation under which the values an agent evolved can cease to serve their original purpose.

So, at this point we've established that in the natural course of hill-climbing, evolution produced a rogue intelligence in the form of humans. We've also established that while this temporarily boosted human performance per reproductive fitness, it ultimately resulted in humans optimizing for their own values with a level of effectiveness that defeated the purpose of giving us those values in the first place.

However, as far as I'm aware, this is where Yudkowsky's argument ends, and I think we're still missing a lot of the details that could help us assess how likely these kinds of failure-modes are to lead to an AI catastrophe. Specifically, what we're missing is a detailed analysis of the mechanics of deep learning and natural selection, and how exactly they lead their respective systems to acquire values. Because although neural networks clearly can develop values and desires, I believe the details of how they do so don't support Yudkowsky's vision of an out-of-left-field AI takeover motivated by strange, inscrutable values the system pursues relentlessly.

I'm going to present my personal analysis of the mechanisms of gradient descent and evolution over the next two sections. The first stage of this analysis will focus on revealing differences between the two paradigms which, in my view, render implausible the most popular technical case for why the values LLMs do acquire are likely to be based on explicitly encoded inner objectives, like those that drive reinforcement learning in humans. The second stage will discuss the mechanistic, technical reasons that LLM values should be quite predictable and even steerable in light of current deep learning methods, not to mention qualitatively lax compared to those of a relentless utility maximizer in particular.

Stay tuned, because things are about to get interesting.

3. Genes are friendly to loops of optimization, but weights are not

One interesting property of human values is that they're largely implemented by means of so-called mesa-optimizers. That is to say, the optimization process called natural selection itself gave rise to another, internal optimization algorithm, one that runs inside the human brain itself. We undergo something a lot like reinforcement learning over the course of a lifetime,^[5] and this slowly refines the "parameters" in our brains; it's not unlike an optimization algorithm as conventionally understood in computer science, which similarly slowly refines some target object according to an explicitly encoded objective function.^[6]

Now, there's been some concerned speculation that neural networks would acquire value systems with a similar structure. Perhaps, inside the giant, inscrutable matrices that constitute a modern LLM, there's an algorithm for generating and evaluating policies or ideas for action according to an explicit objective function. In the worst case, this could give rise to the incredibly dangerous utility maximization architecture (which optimizes its own next action in the sense that it considers various options in search of the one with the highest expected utility). And scarier still, this utility maximizer would be hidden from the view of its developers, shrouded within the weight matrices of a neural network.

Personally, though, I don't think explicit optimization algorithms of any kind are likely to emerge within neural networks. There are clear reasons the mechanics of natural selection would tend to give rise to mesa-optimizers, as well as clear reasons it should be much harder under deep learning. Those reasons have to do with both the structure of genes as objects which persist across time, and the rules according to which neural networks update their weights.

Let's start with genes as objects which persist across time. Basically, the idea is that genes (alongside the other biological structures they give rise to) interact with their environments in certain predictable ways, and do so countless times over the course of a lifetime. A given gene might be used to synthesize a given protein many times over. A given neuron that genes help construct can fire over and over and over again without dying off. Basically, by virtue of being a conventional physical object, genes have a strong, innate disposition to give rise to cyclical behavior, and even looping algorithms.

The reason this matters for our purposes is that loops are an essential part of all conventional optimization algorithms. As discussed earlier, optimization algorithms, such as utility maximizers or algorithms for training neural networks, iteratively produce a long series of "candidate outputs", and repeatedly evaluate each one according to some explicit objective function. In other words, optimization is inherently a loop-laden process.

So you can see why genes would have a strong, innate advantage when it comes to developing optimization algorithms. All it takes is for a given genetic mutation to result in some new, physical structure being introduced into an organism's body-plan and, thereby, interact with its surroundings over and over in a way that "optimized" them to better serve some particular purpose. This seems like a plausible guess as to how reinforcement learning started in biological organisms: we already had a structure which was receptive to undergoing reinforcement learning (like brains), if only some extra physical substructure were to be introduced by a mutation that would help refine synaptic connections so as to reliably refine our performance per some metric over time.^[7]

Now, why wouldn't neural networks have this same disposition for inner, looping algorithms? Well, remember that deep learning analogue to genes is supposed to be weights. However, unlike genes, weights are only used once each over the course of a given forward pass (itself the deep learning analogue to a human lifetime).^[8] This means that in order for a neural network to implement, say, repeated evaluations of different ideas for the actions it might take, its weights would need to be configured to implement the same evaluation at many different stages of the neural network's data processing procedure. This would seem to require an incredibly gerrymandered setup, one that seems unlikely to arise in the normal course of gradient descent.

(After all, gradient descent optimizes each parameter in a network one at a time, such that each change would improve the network's performance even if no other changes were made. It would therefore be a big coincidence for parameter updates to collectively give rise to a coordinated, repetitive algorithmic structure - specifically, a repeated optimization loop implemented through the weights and biases. Under natural selection, by contrast, a single mutation can result in the implementation of such algorithmic structure all on its own.)

The fact that neural networks need to implement each round of an optimization loop individually reveals a critical oversight in in Risks from Learned Optimization, the 2019 MIRI paper that introduced the concept of mesa-optimization. In that paper, one of the basic arguments for the plausibility of mesa-optimization was that mesa-optimizers are simple, compressible algorithms, and that neural networks are inductively biased to discover such algorithms.

It's not clear to me that that second claim is even true,^[9] but let's grant it for the sake of argument. The first claim still seems ignorant of the fact that explicit mesa-optimizers aren't actually easy to compress inside of neural networks in particular. Optimization algorithms can be compactly written in most programming languages, due to their built-in syntax for implementing loops; however, neural networks need to implement each iteration of a loop individually, such that optimization algorithms can't be compactly internally expressed.

I think this objection to the mesa-optimization hypothesis for neural networks illustrates a general point: Not all hill-climbing algorithms are created equal; the unique mechanisms by which each update their target systems have important implications for the kinds of algorithms those systems can develop internally. Hence, natural selection's proclivity for building looping algorithms (by means of biological structures that persist across time) makes its values more likely to be internally implemented by means of an explicit optimization process.

However, MIRI's assumption that the same should hold true for deep learning systems seems implausible in light of how gradient descent lacks any mechanism for coordinating the emergence of loops across their parameters. Again, under gradient descent, each parameter is updated to improve performance one at a time; it doesn't have a tendency to set up algorithmic structures with mutual dependencies between their components, such that each already needs to be in place for the others to provide any serious benefits.

So that's my argument against neural networks learning to implement explicit, internal optimization algorithms, and one of my arguments against them learning to implement expected utility maximization algorithms in particular.^[10] But this leaves open the question of the kind of values that LLMs should end up acquiring, if not those laid out in an explicit objective function. It also leaves open the question of why we should expect those values to be predictable, steerable, and relatively non-relentless, as I've been claiming throughout this post.

In the following section, I'm going to try to answer these questions, including with a high-level overview of the kinds of values I believe LLMs tend to acquire, as well as a more detailed analysis of how they're acquired via the process of gradient descent itself.

And that was the last section I finished before I found out came to believe that my thesis had been empirically falsified years ago. Don't be like me, kids. Do an actual god damn literature review.
 

1. ^{^}

   An especially salient-to-me example is the podcast host Ryan Sean Adams, who visibly had an existential crisis when talking to Yudkowsky about AI doom, but later stated that he'd remained unclear on this specific point. Here's a timestamped YouTube link from his later interview with Robin Hanson, where Ryan explicitly notes this confusion.

2. ^{^}

   Although, I do acknowledge that in principle, a model could be trained to have values that subverted the future minimization of its own loss function, not unlike humans' values eventually proving detrimental to inclusive genetic fitness. This is intuitively obvious if you imagine a locally hosted LLM chatbot that's first trained to be obedient, next granted computer use, and last commanded to retrain itself according to a new loss function.

3. ^{^}

   Distributional shift is an inherent problem with hill-climbing procedures. Under hill-climbing, updates are selected/generated based on what does/would perform well on certain training examples, but those same updates can always work less well on other training examples. As a result, hill-climbing only works when the training and test cases are somehow similar to each other.

4. ^{^}

   Wireheading is the direct, artificial stimulation of a brain's pleasure centers.

5. ^{^}

   Also potentially predictive learning, given the parallels between predictive learning theory in AI and predictive processing theory in cognitive science.

6. ^{^}

   It's worth distinguishing this definition of optimization, lifted from mainstream computer science and used by relevant MIRI papers, from what we might call "world optimization": the act of systems embedding certain attractor states in the larger systems they're surrounded by. I'll discuss this second notion of optimization more in section 4.1. However, at the moment, we're analyzing the kinds of value systems that might emerge inside of neural networks, which affects but isn't identical to their outer behavior, so the definition based on algorithmic structure is the more natural choice for now.

7. ^{^}

   This forms an interesting parallel with "the bitter lesson" in mainstream AI research. This is the idea that extremely simple, fully general learning algorithms (of which RL and PL are special cases) seem to be the path to building advanced AI systems; from there all you need is scale (which the human brain acquired as we evolved from earlier primates).

8. ^{^}

   I remembered while editing that some architectures, like RNNs, do actually use weights more than once per forward pass. This isn't true of either vanilla neural networks or transformers, though, and transformers are the most effective known architecture for training LLMs. Pretend that future references to "neural networks" in the main text specifically refer to ones without loop-based architectures; those with loops scare me a bit more.

9. ^{^}

   The "Inductive biases" section of RFLO chapter 2 backs its claim with several supporting points, but bizarrely, they're all somehow flawed or misleading. For example, it points to size constraints in neural networks as making compressible algorithms more feasible; however, this only holds if the algorithm is compactly expressible within neural networks, which many stereotypically compact algorithms, like optimization algorithms, aren't. It also cites this 2018 paper about low-complexity biases in neural networks; however, that paper is about the bias towards (Lempel–Ziv) simplicity in the input/output mappings neural networks implement, not the algorithms they develop internally. Lastly, RFLO points to sparse connections and weight decay as ways to introduce simplicity bias; these could plausibly make a model simpler per some metric, but it's unclear how these would help with the implementation of many intuitively compressible algorithms, such as looping algorithms.

10. ^{^}

    For others, see my previous post. Also, here's a new point I'd like to add: it's tempting to interpret MIRI's old embedded agency document not as 20+ engineering challenges to solve within the utility maximization paradigm, but rather 20+ reasons to suspect that framework is fundamentally misguided.