“Large Language Models (LLMs) compute a specific conditional probability to predict the next token based on all previous tokens in the context window.”