Nick Bostrom’s simulation argument asks whether we might be living in a computer simulation. The standard version goes: if any civilization ever runs a large number of ancestor simulations, then the number of simulated minds vastly outnumbers the number of real ones, so you’re probably simulated. Most people who hear this either dismiss it as sci-fi or get briefly unsettled and then go back to their day.
I want to propose a different inversion, one that’s more technically grounded and, I think, more unsettling. Not “are we inside a simulation?” That framing is too vague to be useful. Instead: what if the universe is not a simulation but a model, and we are its inference outputs?
Not outputs in the sense of “we are rendered pixels.” Outputs in the sense that we are what the model produces when it’s prompted. The universe is the weights. The physical laws are the architecture. The initial conditions are the system prompt. And everything that happens is inference.
I know how that sounds. Stay with me, because the version of this argument that’s actually worth taking seriously is not the metaphysics. It’s what the analogy reveals about the structure of intelligence, and about the assumptions we’re making when we build models that might one day do the same thing to their own inhabitants.
The Technical Version of the Question
Let me be precise about the analogy, because hand-wavy versions of this argument are useless.
A large language model is a function that maps a sequence of tokens to a probability distribution over the next token. When you run it, you sample from that distribution, append the result, and repeat. The output is a trajectory through token space, shaped by the weights and the prompt.
Now consider a physical universe. It’s a function that maps a state to a probability distribution over the next state (in quantum mechanics, this is literally true; in classical mechanics, it’s deterministic, which is a special case). When you “run” it, you sample from that distribution, append the result to the history, and repeat. The output is a trajectory through state space, shaped by the laws and the initial conditions.
The structural similarity is not superficial. Both are autoregressive generative processes. Both produce trajectories from a fixed set of transition rules. Both have a “prompt” (initial conditions) and a “sampling procedure” (quantum measurement, or classical determinism as a degenerate case). The main difference is that the universe’s state space is unimaginably larger than a token vocabulary, and its transition function is fixed by physics rather than learned from data.
This is not a new observation. It’s roughly the framing of digital physics, and it has a long history. But the AI framing adds something the physics framing misses: the notion of a prompt and a context window. And that’s where it gets interesting.
If the universe is a model, what’s the context window? How far back does the prompt reach? Is the initial condition the Big Bang, or is there a sequence of prompts before it? Is there a system prompt that shapes the whole trajectory, or is the whole thing emergent from a simpler set of rules?
We don’t know. But the question is not as idle as it sounds, because it tells us something about what kind of universe we’d expect to find ourselves in, conditional on the hypothesis being true.
The Sacred Cow: “The Simulation Argument Is About Probability”
Here’s the belief I want to dismantle, and it’s the one that keeps the simulation argument stuck in philosophy departments.
The claim goes like this: the simulation argument is a probability argument. Either we go extinct before running simulations, or we choose not to, or we’re probably simulated. The argument is about the ratio of simulated to real minds. The interesting question is which branch we’re in.
This framing is wrong, or at least incomplete, and it’s wrong because it treats the simulation hypothesis as a claim about counts rather than a claim about structure.
The reversed version is not about counts. It’s about the observation that the universe, whatever it is, has the formal structure of a generative model. That’s not a probability claim. It’s a structural claim. It’s true regardless of whether anyone is running us, in the same way that the observation “a river has the structure of a flow” is true regardless of whether anyone designed the river.
The interesting question is not “are we simulated?” It’s “given that the universe has this structure, what follows about the nature of intelligence, computation, and agency?”
Three things follow, and they’re the meat of this piece.
- If the universe is a model, then “physical law” is a description of the transition function, not a cause of events.
This is a subtle but important shift. In the standard view, physical laws cause events. In the model view, physical laws are the transition function, and events are samples from it. The laws don’t push the universe forward. They are the forward function. This is closer to how a language model works: the weights don’t “cause” the next token in any causal sense. They define the distribution from which the next token is sampled.
- If the universe is a model, then “initial conditions” are a prompt, and prompts shape outputs in ways that are not fully determined by the prompt.
Anyone who has worked with LLMs knows this: the same prompt, run twice with different sampling, produces different outputs. The prompt constrains the output, but it doesn’t determine it. If the universe is like this, then the initial conditions of the Big Bang constrain the trajectory of the universe, but they don’t determine it. There is genuine indeterminism, not because physics is random, but because the model samples.
- If the universe is a model, then “intelligence” is not a special property of certain configurations. It’s a property of certain trajectories.
This is the one that cuts deepest. In the standard view, intelligence is something that certain physical systems have. In the model view, intelligence is something that certain trajectories do. A trajectory is intelligent if it exhibits certain functional properties (learning, prediction, adaptation), regardless of what substrate it’s running on. This is not a new idea, but the model framing makes it structural rather than metaphorical.
Put together, these three implications suggest that the question “are we simulated?” is the wrong question. The right question is “what kind of model would produce a universe with these properties, and what does that tell us about the space of possible minds?”
The Anthropic Prompt
Here’s where it gets genuinely strange.
If the universe is a model, and we are its outputs, then the fact that we exist is evidence about the prompt. This is the anthropic principle, but in the model framing it has a sharper form: the prompt must be such that outputs like us are probable.
This is a constraint on the initial conditions. Not every set of initial conditions would produce a universe with observers in it. The ones that do are the ones we should expect to find ourselves in, conditional on being observers.
So far, so standard. But the AI framing adds a twist. In LLMs, the prompt is not just the initial context. It’s also the system prompt: the instructions that shape the model’s behavior across all contexts. If the universe has a system prompt, what would it look like?
We can’t know, but we can ask what properties a system prompt would need to have to produce a universe with the properties we observe: locality, causality, conservation laws, a low-entropy past, the existence of stable complex structures, and the emergence of observers.
Some of these properties look like constraints on the sampling procedure. For example, the arrow of time could be a property of the sampling procedure rather than the transition function: if the model only samples forward from a low-entropy initial state, you get an arrow of time even if the underlying dynamics are time-symmetric.
Others look like priors in the weights. The existence of stable complex structures could be a consequence of the model being trained (or designed) to produce them, in the same way that a language model trained on text produces coherent text because the training distribution rewards coherence.
We can’t verify any of this. But the framing is useful, because it suggests that the properties of the universe are not arbitrary. They are the kinds of properties you’d expect from a model with a specific architecture, trained on a specific distribution, sampled with a specific procedure. That’s a testable-ish hypothesis, at least in the sense that it makes predictions about which properties should be universal and which should be contingent.
What This Means for AI Builders (Yes, Really)
I know this sounds like philosophy. Let me bring it back to something actionable.
If the universe is a model, then the models we’re building are small versions of the same kind of thing. This is not a metaphor. It’s a structural claim. The universe and a language model are both autoregressive generative processes. The differences are in scale, architecture, and substrate, not in kind.
This has three implications for anyone building AI:
- The “alignment problem” is not unique to AI.
If the universe is a model, then the universe has an alignment problem: how do you ensure that a generative process produces outputs that are consistent with the intended goals? Physics solves this with conservation laws and symmetries. Biology solves it with evolution and homeostasis. AI is trying to solve it with RLHF and constitutional AI. The problem is structural, not specific to our technology.
- The “interpretability problem” is not unique to AI.
If the universe is a model, then the universe has an interpretability problem: how do you understand why a generative process produced a particular output? Physics solves this with reductionism and effective field theory. Neuroscience solves it with systems-level analysis. AI is trying to solve it with mechanistic interpretability and probing. Again, structural.
- The “emergence problem” is not unique to AI.
If the universe is a model, then the universe has an emergence problem: how do complex properties arise from simple rules? Physics solves this with statistical mechanics and phase transitions. Biology solves it with evolution and development. AI is trying to solve it with scaling laws and emergent capabilities. Structural again.
This is not an argument that AI is the same as physics. It’s an argument that the problems are the same kind of problems, and that progress in one domain might inform progress in another. The alignment problem in AI might learn something from how physics maintains stability. The interpretability problem might learn something from how biology maintains homeostasis. The emergence problem might learn something from how complex systems self-organize.
This is a new lens, and I think it’s a useful one. Not because it solves anything, but because it reframes the problems in a way that makes cross-domain analogies available.
What To Actually Do
Three moves, concrete enough to start on Monday.
- Treat your models as generative processes, not as functions.
Most AI engineering treats a model as a function: given input, produce output. The model framing suggests treating it as a process: given a prompt, produce a trajectory. The difference matters, because trajectories have properties that functions don’t. They have path dependence, they have emergent behavior, they have attractors. If you’re building models, ask what trajectory your model is producing, not just what output it’s returning.
- Study the transition function, not just the outputs.
If the universe is a model, then the interesting thing about it is the transition function, not the specific outputs. The same is true of your AI models. The outputs are samples. The transition function (the weights, the architecture, the sampling procedure) is what determines the distribution of outputs. If you want to understand your model, study the distribution, not the samples.
A rough starting point:def study_transition_function(model, prompts, n_samples=100):
trajectories = []
for prompt in prompts:
samples = [model.generate(prompt) for _ in range(n_samples)]
trajectories.append({
"prompt": prompt,
"samples": samples,
"distribution": summarize_distribution(samples),
"attractors": find_attractors(samples),
"divergence": measure_divergence(samples),
})
return trajectories
The samples are the outputs. The distribution is the transition function’s fingerprint. The attractors are the emergent structures. The divergence is the model’s indeterminism. All four are more informative than any single output.
- Look for structural analogies across domains.
If the universe and a language model are the same kind of thing, then the problems of one might be the problems of the other. When you hit a hard problem in AI, ask whether there’s an analogous problem in physics, biology, or cognitive science. The solution might already exist, in a different vocabulary.
The Mental Model
Here’s the lens I want you to keep.
The simulation argument, reversed, is not a claim about whether we’re simulated. It’s a claim about the structure of reality. The universe has the form of a generative model. The physical laws are the transition function. The initial conditions are the prompt. The trajectory of events is the inference. And we are the outputs.
This is not a new metaphysics. It’s a new framing, and the framing is useful because it makes the problems of AI look like the problems of physics and biology, rather than like unique problems that require unique solutions. The alignment problem, the interpretability problem, and the emergence problem are not new. They are the oldest problems in the universe, appearing in a new substrate.
The interesting question is not whether we’re simulated. It’s whether the framing changes how we build. If the universe is a model, then the models we build are small versions of the same thing. And the problems they face are the problems the universe has already solved, in ways we don’t fully understand.
The most productive thing we can do is study those solutions. Not because they’ll tell us whether we’re simulated. Because they’ll tell us how to build better models. And that’s a question we can actually answer.
What’s your read? If you’ve worked with generative models, does the “universe as model” framing change how you think about them? And if you haven’t, does it change how you think about the universe? Drop it in the comments. I’m genuinely curious whether this framing is useful or just a fun philosophical parlor trick.
Top comments (0)