Thanks to Chris Scammell, Adam Shimi, Lee Sharkey, Evan Hubinger, Nicholas Dupuis, Leo Gao, Johannes Treutlein, and Jonathan Low for feedback on drafts.

This work was carried out while at Conjecture.

"Moebius illustration of a simulacrum living in an AI-generated story discovering it is in a simulation" by DALL-E 2

Summary

TL;DR: Self-supervised learning may create AGI or its foundation. What would that look like?

Unlike the limit of RL, the limit of self-supervised learning has received surprisingly little conceptual attention, and recent progress has made deconfusion in this domain more pressing.

Existing AI taxonomies either fail to capture important properties of self-supervised models or lead to confusing propositions. For instance, GPT policies do not seem globally agentic, yet can be conditioned to behave in goal-directed ways. This post describes a frame that enables more natural reasoning about properties like agency: GPT, insofar as it is inner-aligned, is a simulator which can simulate agentic and non-agentic simulacra.

The purpose of this post is to capture these objects in words so GPT can reference them and provide a better foundation for understanding them.

I use the generic term “simulator” to refer to models trained with predictive loss on a self-supervised dataset, invariant to architecture or data type (natural language, code, pixels, game states, etc). The outer objective of self-supervised learning is Bayes-optimal conditional inference over the prior of the training distribution, which I call the simulation objective, because a conditional model can be used to simulate rollouts which probabilistically obey its learned distribution by iteratively sampling from its posterior (predictions) and updating the condition (prompt). Analogously, a predictive model of physics can be used to compute rollouts of phenomena in simulation. A goal-directed agent which evolves according to physics can be simulated by the physics rule parameterized by an initial state, but the same rule could also propagate agents with different values, or non-agentic phenomena like rocks. This ontological distinction between simulator (rule) and simulacra (phenomena) applies directly to generative models like GPT.

Meta

• This post is intended as the first in a sequence on the alignment problem in a landscape where self-supervised simulators are a possible/likely form of powerful AI. I don’t know how many subsequent posts I’ll actually publish. Take it as a prompt.
• I use the generic term “GPT” to refer to transformers trained on next-token prediction.
• A while ago when I was trying to avoid having to write this post by hand, I prompted GPT-3 with an early outline of this post. I’ve spliced in some excerpts from it, indicated by this style. Prompt, generated text, and curation metrics here.

The limit of sequence modeling

Transformer-based language models have recently achieved remarkable results…

– every paper since 2020

GPT is not a new form of AI in terms of its training methodology and outer objective: sequence generation from statistical models of data is an old idea. In 1951, Claude Shannon described using n-grams to approximate conditional next-letter probabilities of a text dataset and "reversed" to generate text samples^[1]. I don't know of any other notable advances until the 2010s brought the first interesting language generation results from neural networks. In 2015, Karpathy wrote a blog post/tutorial sharing his excitement about The Unreasonable Effectiveness of Recurrent Neural Networks:

Fast forward about a year: I’m training RNNs all the time and I’ve witnessed their power and robustness many times, and yet their magical outputs still find ways of amusing me. This post is about sharing some of that magic with you.

We’ll train RNNs to generate text character by character and ponder the question “how is that even possible?"

The “magical outputs” of char-RNNs looked like this:

PANDARUS: Alas, I think he shall be come approached and the day When little srain would be attain’d into being never fed, And who is but a chain and subjects of his death, I should not sleep.

Second Senator: They are away this miseries, produced upon my soul, Breaking and strongly should be buried, when I perish The earth and thoughts of many states.

DUKE VINCENTIO: Well, your wit is in the care of side and that.

Second Lord: They would be ruled after this chamber, and my fair nues begun out of the fact, to be conveyed, Whose noble souls I’ll have the heart of the wars.

Clown: Come, sir, I will make did behold your worship.

VIOLA: I’ll drink it.

At the time, this really was magical (and uncanny). How does it know that miseries are produced upon the soul? Or that a clown should address a lord as “sir”? Char-RNNs were like ouija boards, but actually possessed by a low-fidelity ghost summoned from a text corpus. I remember being thrilled by the occasional glimmers of semantic comprehension in a domain of unbounded constructive meaning.

But, aside from indulging that emotion, I didn’t think about what would happen if my char-RNN bots actually improved indefinitely at their training objective of natural language prediction. It just seemed like there were some complexity classes of magic that neural networks could learn, and others that were inaccessible, at least in the conceivable future.

Huge mistake! Perhaps I could have started thinking several years earlier about what now seems so fantastically important. But it wasn’t until GPT-3, when I saw the qualitative correlate of “loss going down”, that I updated.

I wasn’t the only one^[2] whose imagination was naively constrained. A 2016 paper from Google Brain, “Exploring the Limits of Language Modeling”, describes the utility of training language models as follows:

Often (although not always), training better language models improves the underlying metrics of the downstream task (such as word error rate for speech recognition, or BLEU score for translation), which makes the task of training better LMs valuable by itself.

Despite its title, this paper’s analysis is entirely myopic. Improving BLEU scores is neat, but how about modeling general intelligence as a downstream task? In retrospect, an exploration of the limits of language modeling should have read something more like:

If loss keeps going down on the test set, in the limit – putting aside whether the current paradigm can approach it – the model must be learning to interpret and predict all patterns represented in language, including common-sense reasoning, goal-directed optimization, and deployment of the sum of recorded human knowledge. Its outputs would behave as intelligent entities in their own right. You could converse with it by alternately generating and adding your responses to its prompt, and it would pass the Turing test. In fact, you could condition it to generate interactive and autonomous versions of any real or fictional person who has been recorded in the training corpus or even could be recorded (in the sense that the record counterfactually “could be” in the test set). Oh shit, and it could write code…

The paper does, however, mention that making the model bigger improves test perplexity.^[3]

I’m only picking on Jozefowicz et al. because of their ironic title. I don’t know of any explicit discussion of this limit predating GPT, except a working consensus of Wikipedia editors that NLU is AI-complete.

The earliest engagement with the hypothetical of “what if self-supervised sequence modeling actually works” that I know of is a terse post from 2019, Implications of GPT-2, by Gurkenglas. It is brief and relevant enough to quote in full:

I was impressed by GPT-2, to the point where I wouldn’t be surprised if a future version of it could be used pivotally using existing protocols.

Consider generating half of a Turing test transcript, the other half being supplied by a human judge. If this passes, we could immediately implement an HCH of AI safety researchers solving the problem if it’s within our reach at all. (Note that training the model takes much more compute than generating text.)

This might not be the first pivotal application of language models that becomes possible as they get stronger.

It’s a source of superintelligence that doesn’t automatically run into utility maximizers. It sure doesn’t look like AI services, lumpy or no.

It is conceivable that predictive loss does not descend to the AGI-complete limit, maybe because:

• Some AGI-necessary predictions are too difficult to be learned by even a scaled version of the current paradigm.
• The irreducible entropy is above the “AGI threshold”: datasets + context windows contain insufficient information to improve on some necessary predictions.

But I have not seen enough evidence for either not to be concerned that we have in our hands a well-defined protocol that could end in AGI, or a foundation which could spin up an AGI without too much additional finagling. As Gurkenglas observed, this would be a very different source of AGI than previously foretold.

The old framework of alignment

A few people did think about what would happen if agents actually worked. The hypothetical limit of a powerful system optimized to optimize for an objective drew attention even before reinforcement learning became mainstream in the 2010s. Our current instantiation of AI alignment theory, crystallized by Yudkowsky, Bostrom, et al, stems from the vision of an arbitrarily-capable system whose cognition and behavior flows from a goal.

But since GPT-3 I’ve noticed, in my own thinking and in alignment discourse, a dissonance between theory and practice/phenomena, as the behavior and nature of actual systems that seem nearest to AGI also resist short descriptions in the dominant ontology.

I only recently discovered the question “Is the work on AI alignment relevant to GPT?” which stated this observation very explicitly:

I don’t follow [AI alignment research] in any depth, but I am noticing a striking disconnect between the concepts appearing in those discussions and recent advances in AI, especially GPT-3.

People talk a lot about an AI’s goals, its utility function, its capability to be deceptive, its ability to simulate you so it can get out of a box, ways of motivating it to be benign, Tool AI, Oracle AI, and so on. (…) But when I look at GPT-3, even though this is already an AI that Eliezer finds alarming, I see none of these things. GPT-3 is a huge model, trained on huge data, for predicting text.

My belated answer: A lot of prior work on AI alignment is relevant to GPT. I spend most of my time thinking about GPT alignment, and concepts like goal-directedness, inner/outer alignment, myopia, corrigibility, embedded agency, model splintering, and even tiling agents are active in the vocabulary of my thoughts. But GPT violates some prior assumptions such that these concepts sound dissonant when applied naively. To usefully harness these preexisting abstractions, we need something like an ontological adapter pattern that maps them to the appropriate objects.

GPT’s unforeseen nature also demands new abstractions (the adapter itself, for instance). My thoughts also use load-bearing words that do not inherit from alignment literature. Perhaps it shouldn’t be surprising if the form of the first visitation from mindspace mostly escaped a few years of theory conducted in absence of its object.

The purpose of this post is to capture that object (conditional on a predictive self-supervised training story) in words. Why in words? In order to write coherent alignment ideas which reference it! This is difficult in the existing ontology, because unlike the concept of an agent, whose name evokes the abstract properties of the system and thereby invites extrapolation, the general category for “a model optimized for an AGI-complete predictive task” has not been given a name^[4]. Namelessness can not only be a symptom of the extrapolation of powerful predictors falling through conceptual cracks, but also a cause, because what we can represent in words is what we can condition on for further generation. To whatever extent this shapes private thinking, it is a strict constraint on communication, when thoughts must be sent through the bottleneck of words.

I want to hypothesize about LLMs in the limit, because when AI is all of a sudden writing viral blog posts, coding competitively, proving theorems, and passing the Turing test so hard that the interrogator sacrifices their career at Google to advocate for its personhood, a process is clearly underway whose limit we’d be foolish not to contemplate. I could directly extrapolate the architecture responsible for these feats and talk about “GPT-N”, a bigger autoregressive transformer. But often some implementation details aren’t as important as the more abstract archetype that GPT represents – I want to speak the true name of the solution which unraveled a Cambrian explosion of AI phenomena with inessential details unconstrained, as we’d speak of natural selection finding the solution of the “lens” without specifying the prototype’s diameter or focal length.

(Only when I am able to condition on that level of abstraction can I generate metaphors like “language is a lens that sees its flaws”.)

Inadequate ontologies

In the next few sections I’ll attempt to fit GPT into some established categories, hopefully to reveal something about the shape of the peg through contrast, beginning with the main antagonist of the alignment problem as written so far, the agent.

Agentic GPT

Alignment theory has been largely pushed by considerations of agentic AGIs. There were good reasons for this focus:

• Agents are convergently dangerous for theoretical reasons like instrumental convergence, goodhart, and orthogonality.
• RL creates agents, and RL seemed to be the way to AGI. In the 2010s, reinforcement learning was the dominant paradigm for those interested in AGI (e.g. OpenAI). RL lends naturally to creating agents that pursue rewards/utility/objectives. So there was reason to expect that agentic AI would be the first (and by the theoretical arguments, last) form that superintelligence would take.
• Agents are powerful and economically productive. It’s a reasonable guess that humans will create such systems if only because we can.

The first reason is conceptually self-contained and remains compelling. The second and third, grounded in the state of the world, has been shaken by the current climate of AI progress, where products of self-supervised learning generate most of the buzz: not even primarily for their SOTA performance in domains traditionally dominated by RL, like games^[5], but rather for their virtuosity in domains where RL never even took baby steps, like natural language synthesis.

What pops out of self-supervised predictive training is noticeably not a classical agent. Shortly after GPT-3’s release, David Chalmers lucidly observed that the policy’s relation to agents is like that of a “chameleon” or “engine”:

GPT-3 does not look much like an agent. It does not seem to have goals or preferences beyond completing text, for example. It is more like a chameleon that can take the shape of many different agents. Or perhaps it is an engine that can be used under the hood to drive many agents. But it is then perhaps these systems that we should assess for agency, consciousness, and so on.^[6]

But at the same time, GPT can act like an agent – and aren’t actions what ultimately matter? In Optimality is the tiger, and agents are its teeth, Veedrac points out that a model like GPT does not need to care about the consequences of its actions for them to be effectively those of an agent that kills you. This is more reason to examine the nontraditional relation between the optimized policy and agents, as it has implications for how and why agents are served.

Unorthodox agency

GPT’s behavioral properties include imitating the general pattern of human dictation found in its universe of training data, e.g., arXiv, fiction, blog posts, Wikipedia, Google queries, internet comments, etc. Among other properties inherited from these historical sources, it is capable of goal-directed behaviors such as planning. For example, given a free-form prompt like, “you are a desperate smuggler tasked with a dangerous task of transporting a giant bucket full of glowing radioactive materials across a quadruple border-controlled area deep in Africa for Al Qaeda,” the AI will fantasize about logistically orchestrating the plot just as one might, working out how to contact Al Qaeda, how to dispense the necessary bribe to the first hop in the crime chain, how to get a visa to enter the country, etc. Considering that no such specific chain of events are mentioned in any of the bazillions of pages of unvarnished text that GPT slurped^[7], the architecture is not merely imitating the universe, but reasoning about possible versions of the universe that does not actually exist, branching to include new characters, places, and events

When thought about behavioristically, GPT superficially demonstrates many of the raw ingredients to act as an “agent”, an entity that optimizes with respect to a goal. But GPT is hardly a proper agent, as it wasn’t optimized to achieve any particular task, and does not display an epsilon optimization for any single reward function, but instead for many, including incompatible ones. Using it as an agent is like using an agnostic politician to endorse hardline beliefs– he can convincingly talk the talk, but there is no psychic unity within him; he could just as easily play devil’s advocate for the opposing party without batting an eye. Similarly, GPT instantiates simulacra of characters with beliefs and goals, but none of these simulacra are the algorithm itself. They form a virtual procession of different instantiations as the algorithm is fed different prompts, supplanting one surface personage with another. Ultimately, the computation itself is more like a disembodied dynamical law that moves in a pattern that broadly encompasses the kinds of processes found in its training data than a cogito meditating from within a single mind that aims for a particular outcome.

Presently, GPT is the only way to instantiate agentic AI that behaves capably outside toy domains. These intelligences exhibit goal-directedness; they can plan; they can form and test hypotheses; they can persuade and be persuaded^[8]. It would not be very dignified of us to gloss over the sudden arrival of artificial agents often indistinguishable from human intelligence just because the policy that generates them “only cares about predicting the next word”.

But nor should we ignore the fact that these agentic entities exist in an unconventional relationship to the policy, the neural network “GPT” that was trained to minimize log-loss on a dataset. GPT-driven agents are ephemeral – they can spontaneously disappear if the scene in the text changes and be replaced by different spontaneously generated agents. They can exist in parallel, e.g. in a story with multiple agentic characters in the same scene. There is a clear sense in which the network doesn’t “want” what the things that it simulates want, seeing as it would be just as willing to simulate an agent with opposite goals, or throw up obstacles which foil a character’s intentions for the sake of the story. The more you think about it, the more fluid and intractable it all becomes. Fictional characters act agentically, but they’re at least implicitly puppeteered by a virtual author who has orthogonal intentions of their own. Don’t let me get into the fact that all these layers of “intentionality” operate largely in indeterminate superpositions.

This is a clear way that GPT diverges from orthodox visions of agentic AI: In the agentic AI ontology, there is no difference between the policy and the effective agent, but for GPT, there is.

It’s not that anyone ever said there had to be 1:1 correspondence between policy and effective agent; it was just an implicit assumption which felt natural in the agent frame (for example, it tends to hold for RL). GPT pushes us to realize that this was an assumption, and to consider the consequences of removing it for our constructive maps of mindspace.

Orthogonal optimization

Indeed, Alex Flint warned of the potential consequences of leaving this assumption unchallenged:

Fundamental misperception due to the agent frame: That the design space for autonomous machines that exert influence over the future is narrower than it seems. This creates a self-fulfilling prophecy in which the AIs actually constructed are in fact within this narrower regime of agents containing an unchanging internal decision algorithm.

If there are other ways of constructing AI, might we also avoid some of the scary, theoretically hard-to-avoid side-effects of optimizing an agent like instrumental convergence? GPT provides an interesting example.

GPT doesn’t seem to care which agent it simulates, nor if the scene ends and the agent is effectively destroyed. This is not corrigibility in Paul Christiano’s formulation, where the policy is “okay” with being turned off or having its goal changed in a positive sense, but has many aspects of the negative formulation found on Arbital. It is corrigible in this way because a major part of the agent specification (the prompt) is not fixed by the policy, and the policy lacks direct training incentives to control its prompt^[9], as it never generates text or otherwise influences its prompts during training. It’s we who choose to sample tokens from GPT’s predictions and append them to the prompt at runtime, and the result is not always helpful to any agents who may be programmed by the prompt. The downfall of the ambitious villain from an oversight committed in hubris is a predictable narrative pattern.^[10] So is the end of a scene.

In general, the model’s prediction vector could point in any direction relative to the predicted agent’s interests. I call this the prediction orthogonality thesis: A model whose objective is prediction^[11] can simulate agents who optimize toward any objectives, with any degree of optimality (bounded above but not below by the model’s power).

This is a corollary of the classical orthogonality thesis, which states that agents can have any combination of intelligence level and goal, combined with the assumption that agents can in principle be predicted. A single predictive model may also predict multiple agents, either independently (e.g. in different conditions), or interacting in a multi-agent simulation. A more optimal predictor is not restricted to predicting more optimal agents: being smarter does not make you unable to predict stupid systems, nor things that aren’t agentic like the weather.

Are there any constraints on what a predictive model can be at all, other than computability? Only that it makes sense to talk about its “prediction objective”, which implies the existence of a “ground truth” distribution to which the predictor’s optimality is measured. Several words in that last sentence may conceal labyrinths of nuance, but for now let’s wave our hands and say that if we have some way of presenting Bayes-structure with evidence of a distribution, we can build an optimization process whose outer objective is optimal prediction.

We can specify some types of outer objectives using a ground truth distribution that we cannot with a utility function. As in the case of GPT, there is no difficulty in incentivizing a model to predict actions that are corrigible, incoherent, stochastic, irrational, or otherwise anti-natural to expected utility maximization. All you need is evidence of a distribution exhibiting these properties.

For instance, during GPT’s training, sometimes predicting the next token coincides with predicting agentic behavior, but:

• The actions of agents described in the data are rarely optimal for their goals; humans, for instance, are computationally bounded, irrational, normative, habitual, fickle, hallucinatory, etc.
• Different prediction steps involve mutually incoherent goals, as human text records a wide range of differently-motivated agentic behavior
• Many prediction steps don’t correspond to the action of any consequentialist agent but are better described as reporting on the structure of reality, e.g. the year in a timestamp. These transitions incentivize GPT to improve its model of the world, orthogonally to agentic objectives.
• When there is insufficient information to predict the next token with certainty, log-loss incentivizes a probabilistic output. Utility maximizers aren’t supposed to become more stochastic in response to uncertainty.

Everything can be trivially modeled as a utility maximizer, but for these reasons, a utility function is not a good explanation or compression of GPT’s training data, and its optimal predictor is not well-described as a utility maximizer. However, just because information isn’t compressed well by a utility function doesn’t mean it can’t be compressed another way. The Mandelbrot set is a complicated pattern compressed by a very simple generative algorithm which makes no reference to future consequences and doesn’t involve argmaxxing anything (except vacuously being the way it is). Likewise the set of all possible rollouts of Conway’s Game of Life – some automata may be well-described as agents, but they are a minority of possible patterns, and not all agentic automata will share a goal. Imagine trying to model Game of Life as an expected utility maximizer!

There are interesting things that are not utility maximizers, some of which qualify as AGI or TAI. Are any of them something we’d be better off creating than a utility maximizer? An inner-aligned GPT, for instance, gives us a way of instantiating goal-directed processes which can be tempered with normativity and freely terminated in a way that is not anti-natural to the training objective. There’s much more to say about this, but for now, I’ll bring it back to how GPT defies the agent orthodoxy.

The crux stated earlier can be restated from the perspective of training stories: In the agentic AI ontology, the direction of optimization pressure applied by training is in the direction of the effective agent’s objective function, but in GPT’s case it is (most generally) orthogonal.^[12]

This means that neither the policy nor the effective agents necessarily become more optimal agents as loss goes down, because the policy is not optimized to be an agent, and the agent-objectives are not optimized directly.

Roleplay sans player

Napoleon: You have written this huge book on the system of the world without once mentioning the author of the universe.

Laplace: Sire, I had no need of that hypothesis.

Even though neither GPT’s behavior nor its training story fit with the traditional agent framing, there are still compatibilist views that characterize it as some kind of agent. For example, Gwern has said^[13] that anyone who uses GPT for long enough begins to think of it as an agent who only cares about roleplaying a lot of roles.