A serious line of research is betting that the next real capability jump in AI will not come from a larger language model. It will come from a world model — a system that learns a predictive, manipulable model of how an environment evolves under actions, so it can simulate futures, plan against them, and reason about what would happen if it intervened. The bet is that planning, causal reasoning, and grounding require a model of the world's dynamics, and that predicting the next token of text, at any scale, does not build one. I think this is the most credible version of the "maybe the next leap isn't scale" thesis. I also think it is a bet, not a result, and the intellectually honest move is to hold both of those at once.
Let me be precise about what a world model is, because the phrase gets used loosely.
What a world model actually is
A world model is an internal model that predicts the consequences of actions and the evolution of a state. Formally, it approximates the environment's transition dynamics: given the current state and an action, it predicts the next state, and usually the reward. Once you have that object, three things become possible that are otherwise not. You can simulate — roll the model forward without touching the real environment. You can plan — search over sequences of actions by imagining their outcomes and picking the best. And you can reason counterfactually — feed the model a different action and read off a different predicted future, which is the "what if I had done X instead" operation.
This is not a new idea, which is a point in its favor: it has a research lineage, not a pitch deck. It is the core of model-based reinforcement learning, going back to Sutton's Dyna architecture around 1990, where an agent learns a model of its environment and uses imagined experience from that model to improve its policy. The modern instantiations are concrete and have real results. Ha and Schmidhuber's 2018 paper, titled simply "World Models," trained an agent that could learn to act partly inside its own learned "dream" of an environment. DeepMind's MuZero learned a model that predicts only the quantities needed for planning — reward, value, and policy — and used it with Monte Carlo tree search to master Go, chess, shogi, and Atari without being told the rules. Hafner's Dreamer line learns a compact latent model of the world from pixels and trains behaviors purely by rolling that model out "in imagination"; DreamerV3 obtained diamonds in Minecraft from scratch, a long-horizon task, with a single fixed set of hyperparameters.
Hold onto the MuZero detail, because it defeats the easy dismissal that world models are hand-built structure fighting a losing battle against scale. MuZero was not given a model of chess. It learned one — and specifically learned to predict the abstract quantities relevant to winning, not the pixels of the board. A world model is a learned object. The question is not "structure versus learning." It is "which learned object do you scale."
What an LLM structurally lacks
An LLM is trained to predict the next token of text. Its internal representations are rich, compositional, and genuinely surprising — this is not the tired "it's just autocomplete" line, which is lazy and wrong. But the quantity it is optimized to predict is a symbol in a corpus, and that has consequences that do not dissolve with more parameters.
I've argued the deep version of this separately: a joint distribution over text captures the correlational structure of the world humans describe, but correlation is not causation and fluency is not truth — prediction is not understanding, and no volume of observational text uniquely fixes the causal mechanism behind it. A world model is interesting precisely because it targets the layer text cannot reach. A system that models dynamics can, in principle, reason about interventions and their consequences, because intervention is the native operation of a transition model: change the action, get a different rollout. The causal question "what happens if I do X" is, for a world model, just a forward pass with X plugged in. For a text model, it is a lookup of what people have written about doing X — which works exactly until you reach an intervention nobody wrote down.
Grounding is the second thing. An LLM's only signal is textual: does this continuation look like the training distribution? A world model's signal is predictive success against a world that pushes back. Its notion of "understanding" is cashed out as "predicts the next state correctly," and a wrong model gets falsified by the environment failing to do what it forecast. That is a different and, I'd argue, stronger footing for the word understanding — not prediction of text, but prediction of a world.
Then planning. Autoregressive generation produces a token, conditions on it, produces the next — a single forward trajectory with no search over alternatives and no lookahead. You can bolt planning-like behavior on top with prompting tricks, tool loops, and tree-of-thought scaffolds, and those help. But the base model has no built-in simulator to plan against. A world model is, definitionally, that simulator.
The strongest form of the argument
Yann LeCun has made the most developed case, most fully in his 2022 position paper "A Path Towards Autonomous Machine Intelligence." Represent it accurately: it is a research proposal, largely not yet delivered at frontier generality, and LeCun himself frames it that way. The argument still has teeth.
His first move is a critique of autoregression. If a model generates a long sequence one token at a time, and each token carries some independent probability of stepping onto a wrong path, the probability of staying on a correct trajectory decays roughly like (1 − e) raised to the sequence length. Even with a small per-step error rate, that is exponential decay: at e = 0.01 over 200 steps you are near (0.99)^200 ≈ 0.13, and the failures compound rather than self-correct. This is a stylized argument — real errors aren't independent, and models do recover from some mistakes — so treat the exact number as illustrative, not a benchmark. What it names is a genuine structural pressure that longer horizons make worse.
His second move is about what to predict. Generative models that try to predict every detail of a future observation — every pixel of the next video frame — waste enormous capacity modeling what is inherently unpredictable and mostly irrelevant. His proposed alternative, the Joint Embedding Predictive Architecture (JEPA), predicts in a learned representation space instead of the raw observation space: encode the present and the future, and predict the future's abstract representation from the present's, discarding the unpredictable low-level detail. The image and video versions, I-JEPA and V-JEPA, are the concrete instantiations, and a 2025 action-conditioned successor was demonstrated on some robot planning tasks. Notice this is the same instinct MuZero acted on: predict the abstract quantities that matter for decisions, not the full observation. The bet is that abstraction, not fidelity, is what a world model needs.
Put the pieces together and you get the shape of the wager: a system built around a learned world model, predicting at the right level of abstraction, planning by search over imagined rollouts, could reason about consequences and stay grounded in a way that a bigger next-token predictor structurally cannot. If AI is going to become a genuine new instrument of reasoning rather than a very fluent librarian — a real new organon in Bacon's sense, an instrument that extends what we can figure out — then a manipulable model of the world's dynamics is a far more plausible engine for that than a larger model of the world's text.
Now the honest part: this is a bet, and a hard one
I would be doing exactly what this platform exists to avoid if I sold you the thesis without its counterweights. Here they are, at full strength.
First, LLMs keep beating their skeptics. The list of capabilities confidently declared impossible for scaled next-token prediction — arithmetic, multi-step reasoning, in-context learning, code, theory-of-mind-style tasks — that then showed up as models got larger is long and humbling. "Scale won't get you X" has a poor track record. Some of what looks structurally missing today may simply be under-elicited, or may arrive with the next order of magnitude, or with the reasoning-training paradigms now maturing. Betting against scale has been a reliable way to be wrong.
Second, world models face their own unsolved problems, and they are not small. What representation should the model predict in? The JEPA answer is a hypothesis, not a theorem, and choosing the abstraction is most of the difficulty. How do you learn the model accurately over long horizons without it drifting into nonsense the moment you roll it out past what it has seen? The sim-to-real gap: a model good enough to plan against in imagination but subtly wrong will confidently plan toward outcomes the real world won't deliver, and small model errors compound over a rollout exactly the way token errors do — the failure mode LeCun uses against LLMs reappears inside his own architecture. The abstraction gap: it is genuinely unclear how to get from world models of physical dynamics, where the field's wins live, to world models of the abstract, linguistic, socially constructed domains where LLMs are strongest. Predicting where a stacked block falls is not obviously on a continuum with predicting the consequence of a legal argument.
Third, there is a tension with the Bitter Lesson worth stating plainly. Rich Sutton's 2019 essay argues that general methods leveraging computation beat approaches that build in human knowledge. A world model is, in one light, more inductive structure — the kind of thing the Bitter Lesson warns will be overtaken. The rebuttal, which I find persuasive but not decisive, is that the strong world-model systems don't hand-code structure; they learn the model and search over it, which is precisely the learning-plus-search recipe Sutton celebrates. MuZero is a Bitter Lesson success story, not a violation. This remains a live argument, and reasonable people read the essay both ways.
Merge, most likely, rather than replace
My own forecast — labeled as forecast, not fact — is that the framing of "world model versus LLM" will look quaint in a few years, and the interesting systems will be hybrids. A large model provides the priors, the abstraction, and the vast compressed knowledge of how the human-described world tends to go; a world-model component provides the action-conditioned rollout, the search, the counterfactual query, and the grounding loop against an environment. There are early signs of this convergence: reasoning models that search over their own generated trajectories are groping toward planning; agent architectures that maintain and update an explicit environment state are groping toward a world model; robotics stacks are stitching language models to learned dynamics models. The two research cultures are leaking into each other.
What I am confident about is narrower, and more useful, than a prediction of which architecture wins. The capabilities we most want next — reliable planning, reasoning about interventions, staying tied to a world instead of to a corpus — are exactly the capabilities that a model of dynamics supplies and a model of text does not. Whether that model of dynamics arrives as a separate module, as an emergent property of a sufficiently trained large model, or as something we don't have a name for yet is open. That it has to arrive, in some form, for the next leap to happen is the part of the bet I would actually put money on.
The scaling era taught us that predicting the next token, done at enough scale, buys a staggering amount. The open question is whether the thing it cannot buy is bought by predicting the next state — and that is a question about the world, which means, fittingly, that only a system that touches the world can settle it.