LLMs are statistical next-word predictors, wrapped in a stochastic next-word generator. The next-word predictions output by the model itself are based on the p

LLMs are statistical next-word predictors, wrapped in a stochastic next-word generator. The next-word predictions output by the model itself are based on the patterns and statistics of the training set. The next word that is actually generated is typically a random selection per a top-k or top-p sampling scheme.