Can you get AGI from a Transformer?
I want to share my thoughts about the calculations that Transformers (such as GPT-3) do, and the calculations that I think are required for general intelligence, and how well they line up, and what I think GPT-3 is doing under the hood, and why I think an arbitrary transformer-based GPT-N might be incapable of doing certain tasks are seemingly essential for a system to qualify as an AGI.
Epistemic status: Very low confidence, to the point that I almost decided to delete this without posting it. I think some of my opinions here are very unpopular, and I would love any feedback or discussion.
Before we get into it, I want to make a couple background claims. The point here is basically to argue that the question “Can you get general intelligence by sufficiently scaling up a Transformer?” is worth asking, and does not have an answer of “Obviously yes, duh!!!” You can skip this part if you already agree with me on that.
Background Claim 1: There are types of information processing that cannot be cast in the form of Deep Neural Net (DNN)-type calculations (= matrix multiplications, ReLUs, etc.), except with an exorbitant performance penalty.
(Update: After discussion in the comments section here, I should have framed this section differently: I should have said “For any given specific learning algorithm / architecture framework (e.g. “GPT-like transformer architectures”), there are types of information processing that cannot be performed within that specific framework, except with an exorbitant performance penalty”. I should never have mentioned “DNN-type calculations (=matrix multiplications, ReLUs, etc.)”, because that’s awfully vague (what exactly does the “etc.” mean?), and anyway it’s irrelevant to this post. Thanks gwern.)
By “information processing” I mean anything from sorting algorithms to data compression, random access memories, hash tables, whatever.
Let’s take Monte Carlo Tree Search (MCTS) as an example. AlphaZero does MCTS because DeepMind engineers explicitly programmed it to do MCTS—not because a generic RNN or other deep learning system spontaneously discovered, during gradient descent, that MCTS is a good idea.
Now, in principle, DNNs are universal function approximators, and more to the point, RNNs are Turing complete. So an RNN can emulate any other algorithm, including MCTS. But that doesn’t mean it can emulate it efficiently!
Let’s say we take a generic (PyTorch default) RNN, and train it such that it is incentivized to discover and start using MCTS. Assuming that the gradient flows converge to MCTS (a big “if”!), I believe (low confidence) that its only method for actually executing the MCTS involves:
Taking a certain highly parallelized algorithm running on a GPU (matrix multiplications, ReLUs, etc.)...
...and using it to emulate a Turing-machine-type serial computer…
...and using that to emulate a different highly parallelizable algorithm!
This is absurdly inefficient when compared to MCTS written by a DeepMind engineer and compiled to run directly on bare hardware with appropriate parallelization. Like, maybe, factor-of-a-million inefficient—this is not the kind of inefficiency where you can just shrug it off and wait a year or two for Moore’s law to care of it.
MCTS is just one example. Again, you can open up your algorithms textbook and find thousands of ways to process information. What fraction of these can be implemented reasonably well in the form of DNN-type matrix multiplications / ReLUs / etc.? I expect <<100%. If any such type of information processing is essential for AGI, then we should expect that we won’t get AGI in a pure DNN.
(We could still get it in a DNN-plus-other-stuff, e.g. DNN-plus-MCTS, DNN-plus-random-access-memory, etc.)
Background Claim 2: We can’t just brush aside Background Claim 1 by appealing to the “Bitter Lesson”
Rich Sutton’s Bitter Lesson says that “general methods that leverage computation” have tended to outperform methods that “leverage...human knowledge of the domain”. I strongly agree with this. As I’ve argued—most recently here—I think that the neocortex (or pallium in birds and lizards) uses a “general method that leverages computation” to understand the world, act, reason, etc. And I think we could absolutely get AGI using that method. By contrast—again in agreement with Rich Sutton—I think it’s much less likely that we would get AGI from, say, hand-coded rules about proper reasoning, knowledge graphs, etc. Although who knows, I guess.
Now, read it carefully: Rich Sutton’s claim is that the best way to solve any given problem will be a “general method that leverages computation”. The claim is not the reverse, i.e., if you have any “general method that leverages computation”, it will solve every problem, if only you leverage enough computation!
A fully-connected feedforward neural net is a “general method that leverages computation”, but it can’t do what a GAN or MuZero or Transformer does, no matter how much computation you leverage. It still has to be an appropriate method, with appropriate learning algorithm, parametrization, dataset, etc.
So, just like a fully-connected feedforward neural net is not the right “general method” for NLP, I think it’s at least open to question whether the more general paradigm of matrix-multiplications-and-ReLUs (possibly attached to other simple components like MCTS, RAM, etc.) is definitely the right “general method” for general intelligence. Maybe it is, maybe it isn’t! It’s an open question.
(And by the way, I freely acknowledge that this paradigm can do lots of amazing things, including things that I probably wouldn’t have expected it to be able to do, if I didn’t already know.)
OK, so far, all this is just to plant the general idea that DNNs do a particular type of information processing, and that there’s no guarantee that it’s the right type of information processing for AGI, merely a suggestion from the fact that it is a form of information processing that can do lots of amazing things that seem to be related to AGI.
Generative-model-based information processing: Background and motivation
Moving on, now I’ll switch gears and talk more specifically about a certain type of information processing that I think might be critical for AGI, and after that I’ll talk about how that relates to the information processing in DNNs (and more specifically, Transformers like GPT-3).
Predicting inputs with generative models (a.k.a. “analysis by synthesis” or “probabilistic programming”)
I’ll introduce the idea of analysis-by-synthesis via an unusually simple example: Towards the First Adversarially Robust NN Model on MNIST. MNIST is a famous collection of low-resolution images of handwritten digits, and the task is to recognize the digit from the image pixels. The traditional deep learning method for solving MNIST is: You do gradient descent to find a ConvNet model whose input is pixel values and whose output is a probability distribution for which digit it is. In this paper here, however, they go the other direction: They build a generative model for pictures representing the digit “0”, a different generative model for the digit “1”, etc. Then, when presented with a new digit image, they do ten calculations in parallel: How compatible is this image with the “0” generative model? How compatible is it with the “1” generative model? … They then pick the hypothesis that best fits the data.
Suggestively, their classification method winds up unusually human-like—it makes similar mistakes and has similar confusions as humans (assuming the paper is correct), unlike those funny examples in the adversarial perturbation literature.
As in this example, analysis-by-synthesis is any algorithm that searches through a collection / space of generative models for a model that best fits the input data (or better yet, with time-varying inputs, a model that predicts the input data that will arrive next). I’ve seen two other analysis-by-synthesis digit-recognition papers: Josh Tenenbaum & coworkers using classic probabilistic programming; and Dileep George & coworkers using a brain-inspired approach (and see also a nice series of blog posts introducing the latter).
The latter captures, I think, an essential aspect of the neocortex, in that there’s a whole society of generative models. Some of these generative models are at a very low level, and make predictions (with attached confidence levels) that certain sensory inputs will or won’t be active. But more often the generative models make predictions that other generative models will or won’t be active. (These predictions can all be functions of time, or of other parameters—I’m leaving aside lots of details). To oversimplify, imagine that if the “bird” generative model is active, it also sends out a low-confidence prediction that the “is flying” generative model is also active, and that the “noun” generative model is active, and so on. This leads to not only hierarchical generative models, but also other forms of compositionality. For example, there’s a “purple jar” generative model that more-or-less just predicts that the “purple” generative model and the “jar” generative model are active simultaneously. This works because the “purple” and “jar” models make (by-and-large) compatible predictions—they agree with each other on some predictions, and in other cases they’re making predictions about different things. By contrast, there is no “stationary dancing” composite model—those two models make contradictory predictions. When two mutually-incompatible generative models are active simultaneously, they kinda fight each other for dominance. The algorithm underlying that “fight” is at least vaguely analogous to message-passing in a probabilistic graphical model. See more of my thoughts at Gary Marcus vs Cortical Uniformity and Predictive coding = RL + SL + Bayes + MPC.
What’s so great about generative-model-based processing?
Well, you might say, OK, analysis-by-synthesis is an interesting approach. But is it essential for AGI? Or is it just one of many ways to do things? I think it’s at least plausibly essential. Consider some of its features...
Feature 1: Better and better results with a longer search
Example: You look at a picture of a camouflaged animal, hidden so well that you can’t find it at first. But after staring for a minute, it snaps into place, and you recognize it.
Example: Your code has a syntax error on line 12. You stare at the line for a few seconds, and then finally see the problem.
Since we’re doing a search over a space of generative models, if we don’t find a good match immediately, we can notice our confusion and keep searching. The space of generative models is astronomically large; you can always search more, given more time.
Feature 2: Generative models are simpler & easier to learn than discriminative models
Example: You are asked to prove a simple theorem for homework. You spend a few minutes thinking through different possible lines of attack, and eventually see an approach that will work.
Example: You are asked to invent a microscope with higher resolution and speed than what exists today. You think through a wide array of many different possible configurations of lenses, lasers, filters, polarizers, galvos, sample-holders, etc. Eventually you come up with a configuration that meets all requirements.
In both these cases, the forward / generative direction has a (relatively) simple structure: If you differentiate both sides, what happens? By contrast, the reverse / discriminative direction is almost arbitrarily complex: At the end, you proved the theorem; what was the first step? A model that could immediately answer the endless variety of such reverse-direction question seems like it would be hopelessly complex to learn!
In fact, probably the only way that the reverse model could be learned is by first developing a good generative model, running it through in lots of configurations, and training a reverse model on that. And something like that is indeed entirely possible! In humans, there’s a kind of memorization / chunking that allows the system to “cache” the results of runs through the generative models, so that the corresponding reverse-lookup can be very fast in the future. Actually, this is not so far from how AlphaZero works too—in that case, the MCTS self-play functions as the generative model.
Feature 3: A generative-model-based approach is more sample-efficient and out-of-distribution-generalizable
This is related to Feature 2. Since generative models are simpler (less information content) than reverse / discriminative models, they can be learned more quickly. Empirically, I think this is borne out by those three digit-recognition papers I cited above, and by the fact that humans can do zero-shot transfer learning in all kinds of situations. (Of course, humans need to spend years cultivating a healthy and diverse society of generative models; the sample-efficiency I’m talking about is after that.)
Another way to think about it: things in the world are, in reality, generated by highly structured hierarchical-and-compositional-generative-model-type processes. If we want a good inductive bias, then we need to structure our analysis in a parallel way: we need to go looking for those kinds of generative models until we find ones that explain the data. Better inductive bias means better out-of-distribution generalization and fewer samples required.
Example: You read about a new concept you hadn’t heard of—Quantum Capacitance. You find it confusing at first, but after reading a couple descriptions and thinking about it, you find a good mental model of it—a certain way to visualize it, to relate it to other known concepts, etc. Having internalized that concept, you can then go use it and build on it in the future.
In this example, you built a good generative model of Quantum Capacitance despite little or even no concrete examples of it. How? For one thing, you’re not building it from scratch, you’re snapping together a few pieces of generative models you’ve already learned, and your textbook helps you figure out which ones, by describing the concept using analogies, figurative spatial language, etc. For another thing, you are trying to slot this concept into a dense web of previously-established physics-related generative models, each of which will protest loudly if they see something that contradicts them. In other words, there are only so many ways to create this concept in a way that doesn’t contradict the things about physics you have already come to believe. Third, you can test any number of proposed generative model against the same known examples—sorta vaguely like off-policy replay learning or something. So anyway, with all these ingredients, it’s a highly constrained problem, and so you can can find the right model with few if any concrete examples.
Feature 4: Foresight, counterfactuals, deliberation, etc.
A society of generative models is a powerful thing...
It allows foresighted plans towards goals—in fact, this capability involves almost the same information-processing steps as analysis-by-synthesis understanding of sensory inputs (see my post here, or Planning by Probabilistic Inference, or Logical Inductor Decision Theory, etc.)
It enables counterfactual reasoning—see my brief discussion here.
It enables system-2-type deliberation—e.g. you can learn a Generative Model X that invokes a Generative Model Y and then invokes either Generative Model Z or Z’ depending on some feature of Y. Then Z or Z’ can invoke yet different generative models, and so on. That kind of process enables us to create an ad hoc crappy serial computer … and we call that “deliberation”! See also Kaj’s post.
I could go on, but you get the idea.
So, at the end of the day, is the generative-model-based approach essential for superintelligent AGI?
I mean, it kind of feels that way to me, but maybe I’m anthropomorphizing.
Back to Transformers and GPT-N
OK, you’ve read all this way, now how does this relate to GPT-3? Will GPT-N be an AGI or not??
Now, when the brain does its probabilistic generative model inference thing, it does it by a complicated, decentralized process. There’s a feedforward pass that activates some of the most promising generative models. Then the models do some message-passing back and forth. Contradictory models fight it out. New models get awakened. Redundant models go away, etc. Multi-step models, extended through time, walk through their steps.
So here’s my hypothesis: this whole complicated decentralized asynchronous probabilistic-generative-model-inference process is more-or-less exactly what the GPT-3 Transformer learns to approximate.
As a special case: I feel pretty comfortable imagining that a single trained Transformer layer can approximate a single step of message-passing in a probabilistic graphical model inference algorithm.
Now, I haven’t thought this through in any amount of detail, but so far this hypothesis seems at least plausible to me. For example, the Transformer has residual connections that allow later iterations to create small increments in the activations, which is what often happens with later stages of message-passing. It has a self-attention mechanism that seems well-suited to representing the sparsely-connected structure of a probabilistic graphical model. It has the positional information that it needs to figure out how far along the various multi-step generative models have progressed. GPT-2 has 12 layers and GPT-3 has 96, enough for quite a lot of sequential processing, which is important to capture this asynchronous process as it unfolds in time.
I like this hypothesis a lot because it’s consistent with my strong belief that GPT-3 has captured human-like concepts and can search through them and manipulate them and combine them in a very human-like way.
So, does that mean GPT-3 is a human-like intelligence? No!! There’s a big difference!
If you have an algorithm designed from the ground up to build and search through generative models and do probabilistic generative model inference, it will do that, and always exactly that, and only exactly that.
If you have a Transformer, it can imitate that kind of probabilistic generative model inference, but will do so only within the range covered by its training data, e.g. a restricted time-window etc. And that same Transformer architecture can also do loads of other very different types of calculations too, that have nothing to do with probabilistic generative model inference—which means that the Transformer will have different inductive biases, and worse sample efficiency.
Let me offer a more specific list of my hypothesized fundamental deficiencies in Transformers as AGIs.
First, for the reason mentioned above, I think the sample efficiency is bound to be dramatically worse for training a Transformer versus training a real generative-model-centric system. And this makes it difficult or impossible for it to learn or create concepts that humans are not already using.
For example, I am confident that GPT-N, after reading tons of text where people use some random concept (Quantum Capacitance, say), can gradient-descent its way to properly using that concept, and to properly integrating it into its (implicit) world-model. But if GPT-N had never heard of Quantum Capacitance, and saw one good explanation in its training data, I’m pretty skeptical that it would be able to use the concept properly. And I’m even more skeptical that it could invent that concept from scratch.
To clarify, I’m talking here about sample-inefficiency in modifying the weights to form a more sophisticated permanent understanding of the world. By contrast, I think there is little doubt that it has the ability to alter its activations in a quick and flexible and sample-efficient way in response to thought-provoking input text. But that only takes you so far! That cannot take you more than a couple steps of inferential distance away from the span of concepts frequently used by humans in the training data.
Second, the finite number of Transformer layers puts a ceiling on the quality of the generative-model-search process, the time spent deliberating, etc. As I mentioned above, humans can stretch their capabilities by thinking a little bit longer and harder. However, if you have a Transformer that more-or-less simulates the first 100 (or whatever) milliseconds of the neocortex’s generative-model-search process, then that’s all you can ever get.
Third, because the Transformer is a kind of information processing imitating a different kind of information processing, I generally expect edge cases where the imitation breaks down, leading to weird inductive biases, crazy out-of-distribution behavior, etc. I’m not too sure about this one though.
Maybe there are other things too.
Then of course there are also the obvious things like GPT-3 being disembodied, not having a reward signal, not having visual or spatial input channels, etc. These are all plausibly very important, but I don’t think want to emphasize them too much, because I don’t think they’re fundamental architectural limitations, I think they’re more likely just things that OpenAI hasn’t gotten around to doing yet.
These three deficiencies I’m hypothesizing seem like pretty serious roadblocks to AGI—especially the one where I claim that it can’t form new concepts and add them permanently to its world-model.
That said, it’s entirely possible that these deficiencies don’t matter, or can be worked around. (Or that I’m just wrong.)
But for the moment, I continue to consider it very possible that Transformers specifically, and DNN-type processing more generally (matrix multiplications, ReLUs, etc.), for all their amazing powers, will eventually be surpassed in AGI-type capabilities by a different kind of information processing, more like probabilistic programming and message-passing, and also more like the neocortex (but, just like DNNs, still based on relatively simple, general principles, and still requiring an awful lot of compute).
I recognize that this kind of probabilistic programming stuff is not as “hot” as DNNs right now, but it’s not neglected either; it’s a pretty active area of CS research, moving forward each year.
But I dunno. As I mentioned, I’m not confident about any of this, and I am very interested in discussion and feedback here. :-)