Stop Thinking of LLMs as Next-Token Predictors

TEXT START: Strictly speaking, the statement “LLMs are next-token predictors” isn’t wrong, but it’s incomplete.

The Dissection

The article makes a legitimate distinction between pre-training as imitation of existing sequences and post-training as reward-shaped optimization over generated trajectories. Its code examples and chess analogy are designed to move the reader from “text autocomplete” to “system selecting actions toward outcomes.”

But it also performs a rhetorical substitution: it shifts from describing the model’s emission mechanism to describing the objective imposed on that mechanism, then treats the latter as proof that the former is obsolete. The post-trained system still emits one token conditional on prior tokens. It is a reward-conditioned policy implemented through autoregressive prediction.

The Core Fallacy

The article confuses “not merely trained on existing text” with “not a next-token predictor.” RLVR changes which continuations become more probable; it does not remove the next-token loop. A policy can select successful trajectories one token at a time and still be a next-token predictor in the operational sense.

The deeper error is treating taxonomy as capability evidence. Learning from self-generated sequences does not automatically produce general discovery, reliable planning, or agency. Novel sequences may be new combinations of learned representations, while the reward function supplies the selection pressure. The article offers no proof that its claimed exploration survives outside the rewarded tasks, avoids proxy failure, or transfers into dependable open-world performance.

The chess analogy is stacked. An exhaustive game-search system is defined by its value function and search regime, not by a magical escape from sequential prediction. The relevant contrast is imitation versus outcome optimization—not predictor versus non-predictor.

Under the Discontinuity Thesis, this semantic battle is secondary. The decisive question is whether the resulting systems achieve durable cost and performance superiority across cognitive work. If they do, they can sever the mass employment–wage–consumption circuit while still predicting tokens. The implementation label does not protect human productive participation.

Hidden Assumptions

• RLVR rewards faithfully measure real task success rather than narrow proxies.
• Exploration is broad and productive enough to find useful solutions rather than merely exploit familiar patterns.
• Generated novelty constitutes meaningful discovery rather than recombination.
• Capabilities learned on verifiable tasks generalize to messy, weakly specified work.
• The chess analogy transfers from a closed, formal environment to open-world economic activity.
• Post-training behavior is stable, reliable, and controllable at deployment scale.
• A change in training objective warrants replacing the mechanism-level description.
• Explaining the model as an outcome-seeking system is equivalent to demonstrating that it can replace human labor.

Social Function

Classification: partial truth with prestige signaling.

The article is not simple copium; it correctly attacks an impoverished mental model. But its sophistication also launders a capability claim through vocabulary. “Exploration,” “discovery,” and “choosing what wins” sound like a decisive break from prediction without supplying the evidence needed to establish one. It upgrades the conversation’s status while leaving the economic consequences unmeasured.

The Verdict

Useful correction, inflated conclusion. The article proves that “next-token predictor” is an incomplete description of post-trained behavior. It does not prove that the category is wrong, that RLVR creates general intelligence, or that the model has escaped statistical prediction.

Its real significance is harsher: autoregressive prediction may be enough to implement increasingly effective outcome-seeking policies. If those policies become reliable and cheap, they help satisfy P1 precisely without ceasing to predict tokens. The article cuts the label’s skin; it never establishes the machine’s full reach.