“Next-token predictor” is the wrong mental model for LLMs
Whether large language models are “just next-token predictors” is contested by people who argue that this phrase is both technically accurate and deeply misleading. Commenters note that transformers are trained to predict the next token in text, but post-training techniques like reinforcement learning, tool use, and agentic frameworks turn that basic mechanism into a goal-seeking system whose behavior can’t be understood purely as copying training data. Others counter that emphasizing anything beyond next-token prediction risks inflating LLM capabilities, and that recognizing their probabilistic, mechanical nature is essential for judging both their limits and their potential risks.
Accuracy of the “next‑token predictor” label
- Many argue LLMs are literally next‑token probability models: they take a token sequence and output a distribution over the next token.
- Others say this is technically true but misleadingly reductive, akin to calling a human a “bag of chemicals” or a jet “just a rotating machine.”
- Some see “next‑token predictor” and “stochastic parrot” as slogans used to dismiss LLMs’ capabilities and risks.
Pre‑training vs RLHF/RLVR and post‑training
- Several commenters distinguish sharply between:
- Pre‑training: classic supervised next‑token prediction against ground‑truth text.
- Post‑training with RLHF/RLVR: no “correct next token”; instead, whole sequences are scored by reward models, and parameters are updated to increase reward.
- One camp: after heavy RL, the model is better viewed as a policy or “goal‑seeking / outcome‑steering system,” not a pure next‑token predictor.
- Opposing camp: despite changed objectives, the architecture and interface still reduce to next‑token distributions, so the label remains apt.
Prediction vs decision / optimization
- Dispute over the word “predict”: some insist prediction requires checking against an external truth; RL is instead optimization over reward.
- Others argue choosing the token that maximizes expected reward is still a form of prediction (of outcomes or rewards), just with a different target.
- Chess analogies are heavily debated: is an engine “predicting the next move” or “choosing a move that wins”?
Capabilities, limits, and emergence
- One side: scaling next‑token prediction + compression of text patterns naturally yields general abilities, including reasoning‑like behavior.
- Other side: models still fail often, can’t reliably code or reason, and do not “understand” correctness; they optimize patterns, not truth.
- Emergent properties are noted, but some warn that doesn’t imply unbounded or AGI‑like capabilities.
Alternative mental models
- Proposed metaphors include: “next‑embedding / next‑meaning predictor,” “pattern matcher,” “policy,” “outcome steerer,” “informed dice,” “agentic system built around a predictor.”
- Consensus trend: multiple abstractions are useful; “next‑token predictor” is valid but incomplete, especially for modern agentic and RL‑tuned systems.
Meta: politics, hype, and writing style
- Some see attempts to move beyond “next‑token predictor” as hype to inflate AI’s status; others see trivialization as a way to avoid engaging with risks.
- A number of commenters complain the article reads like “AI‑style” prose and debate whether AI‑assisted writing undermines credibility.