This is a brief article that argues convincingly that the 'Next-token predictor' picture is the wrong mental model for LLMs. Consider, writes Garrin McGoldrick, a chess-playing LLM. "Given a new position, it predicts the move a grandmaster would most likely play next. That is a next-move predictor." By contrast, "Given a position, it chooses the move that leads to the highest probability of winning. Unlike the first system, it is not trained only on games that grandmasters already played. It also learns from games generated by its own exploration." Calling this a next-move predictor, he says, "would be strange". Both systems will give you a 'next move'. But one of them is basing its next move on more than just statistical probability from past events. And that's what contemporary LLMs do today. Via Doug Belshaw.
Today: Total: [Direct link] [Share]
