Next-token prediction

The single objective LLMs are trained on: given all tokens so far, output a probability for every possible next token. Chat, translation, coding and reasoning are all this one operation in a loop.