World Models and the Causal Frontier
Where Next-Token Compression Ends and Causal Understanding Begins
The training dataset fed into Kenneth Li, David Bau, and Martin Wattenberg's 2023 experiment at Harvard and MIT consisted strictly of bare move strings like E4, D3, C3, C4, F5. Each string recorded a sixty-move game of Othello played on an eight-by-eight grid. Crucially, the small Transformer model—Othello-GPT—was never shown a picture of a game board. It was never told that square E4 sits adjacent to E5, and it was never given the rule that outflanked discs flip from black to white. The network was trained solely to predict the next legal move coordinate from the preceding string of letters and numbers.
Critics of statistical artificial intelligence had long argued that a model trained only on surface tokens is a stochastic parrot. In this view, next-token predictors merely stitch together shallow surface correlations without constructing any internal model of reality. Eighty years earlier, in his 1943 Cambridge monograph The Nature of Explanation, Scottish psychologist Kenneth Craik had proposed the operational test for genuine internal understanding. Craik argued that an organism carrying a small-scale internal model of external reality and of its own possible actions can test counterfactual alternatives in its head before acting in the physical world.
When Li and his colleagues trained non-linear diagnostic probes on the internal hidden activations of Othello-GPT, they found a striking result. To compress and predict legal moves accurately, the neural network had spontaneously constructed an accurate eight-by-eight internal map tracking which squares were black, white, or empty after every move.
To prove this internal board map was not a passive correlation, the researchers performed a causal surgical intervention. Mid-game, they manually reached into the network's hidden layers, flipped the internal neural activation for square C4 from white to black, and let the forward pass continue. Immediately, the model abandoned its previous strategy and output legal moves corresponding to the artificially altered board state. Sufficiently tight compression of surface observations had forced the network to discover the hidden state geometry that generated the data.
Yet whether next-token compression builds true causal laws or brittle local heuristics remains the defining open question at the outer boundary of artificial intelligence. Computer scientist Judea Pearl formalized this boundary in his 2018 book The Book of Why. Pearl's Ladder of Causation separates statistical pattern matching from scientific reasoning across three distinct mathematical rungs:
- * Rung 1 — Association ($P(y \mid x)$, Seeing): Observing passive statistical correlations in historical data, such as noticing that falling barometer needles correlate with rainstorms. Passive observation alone cannot distinguish which variable drives which.
- * Rung 2 — Intervention ($P(y \mid \text{do}(x))$, Doing): Predicting what happens when an agent physically reaches into a system and forces a variable to change. A causal model recognizes that manually taping a barometer needle higher will not clear the storm clouds.
- * Rung 3 — Counterfactuals (Imagining): Simulating alternative universes that contradict observed history, such as asking whether a patient would have survived had a physician doubled the antibiotic dose yesterday. Counterfactual reasoning requires an invariant structural model of the underlying mechanism.
When Harvard researchers trained a variant of the model on astronomical orbit coordinates generated by Newton's laws and then asked it to predict a new force law outside the training data, the network predicted the planetary positions accurately yet failed to reconstruct the inverse-square vector field.