“Next-token predictor” is the wrong mental model for LLMs

The article argues that viewing large language models (LLMs) merely as "next-token predictors" is an inaccurate mental model. While LLMs do emit tokens sequentially, this description only captures the mechanism's shape, not its encoded function. The author suggests that this sequential token emission can encode diverse functionalities, such as simulating a helpful assistant or representing knowledge gained through exploration, highlighting that the underlying mechanism can support various complex behaviors beyond simple prediction.