World Models vs. LLMs

The fight everyone frames as “language versus the physical world” is the wrong axis. LLMs already build world models. They build them as a side effect, ungrounded, and nothing in the objective forces those models to be consistent. That is the whole argument, and most of the takes miss it.

LLMs already do this

“An LLM is just a text predictor with no model of the world” is empirically false. Train a GPT on nothing but legal Othello move sequences, with zero information about the board or the rules, and a representation of the board state shows up inside the network. Intervene on those neurons and the model’s move predictions change accordingly, so it is actually using the representation, not memorizing a lookup table (Li et al., “Emergent World Representations”).

It scales past toy games. Gurnee and Tegmark found Llama-2 carries linear representations of space and time, with identifiable “space neurons” and “time neurons,” and the fidelity climbs with model size. A world model is in there. The question was never whether, it is what kind.

The world model is a bad one

Here is the result that should end the “scaling gives you understanding for free” argument. Vafa, Kleinberg, Mullainathan, and Rambachan trained transformers on 4.7 billion tokens of turn-by-turn taxi rides in Manhattan. By every surface metric the models nailed it: the predicted next turn was valid close to 100% of the time, and the internal state seemed to encode the current location.

Then they reconstructed the map the model had implicitly built. It was an imagined New York full of streets that do not exist, flyovers crossing over other streets, intersections wired together at impossible angles. Close just 1% of the real streets and accuracy collapsed from near-100% to 67%. The model was a pile of locally correct heuristics that never once had to add up to a single coherent Manhattan. It is the Emergent and Collective Intelligence story with the catch made explicit: emergence from aggregating heuristics is real, but the aggregate is under no obligation to be globally true.

The cause is the objective, not the architecture. Next-token prediction rewards being plausible at each step. It never penalizes the whole from cohering. That is exactly the production failure mode anyone shipping on LLMs already knows: confident, locally fluent, globally wrong.

What the world-model camp is actually buying

Yann LeCun calls autoregression “doomed” for reasoning because errors compound, each prediction inheriting the divergence of the last over a long chain. He left Meta after twelve years and raised around a billion dollars to build the alternative. His JEPA predicts in an abstract latent space instead of reconstructing raw output: mask part of the input, predict the features of the missing part, not the pixels or tokens.

Fei-Fei Li is making the same bet from the perception side. World Labs shipped Marble in November 2025, generating persistent 3D spaces from a sentence or an image, and raised at a reported $5B valuation chasing what she calls Spatial Intelligence, the frontier beyond language. DeepMind’s Genie 3 generates interactive, physics-consistent 3D environments in real time at 24 fps.

Three different products, one shared move: make world state the prediction target instead of tokens. When the thing you are scored on is the state of the world, the objective forces the consistency that next-token prediction lets you skip.

The real question

It is not language versus reality. Both camps are building world models. The split is where the consistency constraint lives, and whether you buy coherence through architecture or through scale.

The optimistic read is that enough scale plus multimodality closes the gap on its own, and the Manhattan flyovers just mean the model was too small. That is the Neural Networks and Brain Scaling bet: capability tracks size, keep going. The pessimistic read, and the one I find harder to argue against, is that the objective is the bug. No amount of scaling fixes a metric that never asked for global coherence in the first place. You will get a sharper bag of heuristics, not a map.

A billion dollars of LeCun’s funding says the answer is architecture. The cheapest way to find out is to watch whether the next generation of pure LLMs can be closed out of 1% of Manhattan and still find their way home.

Try it

Coherence probe (1-2 hours, Python + OpenRouter). Build a small synthetic grid: a 5x5 city of named intersections with explicit adjacency rules. Generate sequences of moves — some valid, some with one impossible transition buried in the middle (a step that skips a block or reverses illegally). Prompt several models via the OpenRouter API with each sequence and ask for the next valid move. Score whether they flag the impossible step or continue confidently. Compare a large model against a small one. The gap, or lack of one, is the world model working, or not.

Othello-GPT mini (a weekend, PyTorch). Train a small GPT on sequences of legal Othello moves — no board representation, no rules, just move tokens. The dataset is easy to generate programmatically. Install transformer-lens and extract layer activations, then fit a linear probe to predict the actual board state. If probe accuracy climbs well above chance, you have replicated the Li et al. result from scratch. The interesting moment is watching which layer the board state crystallizes in — usually not the first, not the last.

Sources

Vafa et al., “Evaluating the World Model Implicit in a Generative Model” (NeurIPS 2024) — the NYC taxi map study.
Li et al., “Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task” — Othello-GPT.
Gurnee and Tegmark, “Language Models Represent Space and Time”.
Why Fei-Fei Li, Yann LeCun and DeepMind are all betting on world models, and how their bets differ.