The "next token predictor" is wrong: So what is LLM doing?
LLM's real ability is not to predict the next token, but to build a compressible model of the internal world; the "next-token" metaphor itself is the culprit responsible for the infinite amplification of AGI's anxiety.
The metaphor of "the next token predictor" is leading a generation of engineers into a cognitive dead end.
Last year, I was doing a technical review at an AI startup. The founder took a 200-page PRD and asked the model one by one: "Can this code be generated? Can that logic reason?" I interrupted him: You are asking the wrong question. The real question is not whether the model can predict the next token, but what internal representations it compresses your needs into in the 80-layer residual stream. He was stunned-this was the collective cognitive bias of the entire industry.
The training goal of Transformer is indeed next-token prediction. But equating goals with mechanisms is like saying that "the essence of humans is that carbon-based organisms pursue genetic replication." This is technically correct and has zero explanatory power. Ilya Sutskever said long ago: Compression is understanding. When the model compresses "the apple fell" into a potential expression of Newton's laws, rather than memorizing 10,000 sentences "the apple fell", it is already building a model of the world. Hinton is more blunt: "The big model is not predicting the next word, it is predicting the state of the world that will produce the word."
The opposite will say: The training goal is to minimize cross-entropy, and no "understanding" modules are explicitly constructed. This may sound like an iron law, but it is actually a typical "goal is mechanism" fallacy. AlphaGo's goal is to win chess, but what it learns is not "how to win", but "how to evaluate the value of a situation"-a compressible model of the Go world. Similarly, next-token prediction is just a shortcut to forcing models to learn compression; the by-product of compression is the world model. You cannot say that the hammer is essentially a "predictor of nailing action" just because it is used to drive nails.
This reminds me of cognitive psychology in the 1950s. At that time, the mainstream compared the brain to a "telephone switch", later it became a "digital computer", and now it is a "neural network". Every metaphor change reshapes the research direction of a generation. Switch metaphors lead people to "connectionism", computer metaphors lead people to "symbolism", and the "next-token predictor" metaphor is causing countless engineers to spend their energy on prompt engineering, context stuffing, and token billing optimization-essentially optimizing input-output mappings rather than understanding internal representations.
There is a unicorn that does code generation, and engineers are listening to "How to make the model output more compliant with the specification TypeScript" every day. They treat the model as a probability automaton, pile few-shots, write system prompts, and connect linters to the bottom. Until someone used a sparse self-encoder (SAE) to pull out an eigenvector in a layer in the model-corresponding to the concept of "type safety." Directly strengthen this direction by 0.3 standard deviation in the residual stream, and the type error rate of the generated code is directly halved. No prompt changes, no RAG additions, no fine-tuning. This is the power of the world model perspective: instead of patching at the input-output level, surgery is performed in the representation space.
The cost of this cognitive bias is not just engineering efficiency. It created a language accident for AGI anxiety. When you think of the model as the "next token predictor", every benchmark increase is like approaching "real intelligence"-because the more accurate the prediction is, the closer it seems to be to "understanding." But put it another way: the model is just building an increasingly high-fidelity model of the world, and the benchmark is just a slice projected into token space. Anxiety comes from misreading "projection quality" as "model essence". It's like watching shadow plays arguing about whether shadows have souls instead of looking at the hands that control them.
Who did this matter reshape? The programmer who complains to ChatGPT every day,"It's nonsense again", the product manager who forwards "GPT-6 IQ 150" on the circle of friends, and the philosopher who argues on Twitter that "LLM will never understand." They share a false metaphor and are caught in the same language game. Tools reshape users only if the user first has the correct mental model. Use LLM as a "builder of a compressible world model" and you will start to pay attention to: Which direction in the residual stream corresponds to "causal reasoning"? How to interfere with representations rather than prompt? What kind of data can make the world model more consistent? These issues are the true path to the next generation of AI.
I make a bet: in three years, the word "next-token predictor" will exit the mainstream technology vocabulary like "electronic brain." Instead, instead of cooler marketing terms, there is a boring but precise description of the "compressible world model." By then, AGI anxiety won't go away, but it will at least turn into an engineering issue-not a language accident.