During training, certain tokens are more likely to lead to a lower loss function value, which is how you "win" the game of LLM output.