Large Language Models

An LLM predicts the next word, one token at a time, given everything before it. Trained on most of the public internet, it has internalized grammar, facts, code patterns, and writing styles.

Models read text as tokens — fragments roughly 3-4 characters long. 'unbelievable' might be 3 tokens. This is why models 'count letters' poorly: they don't see letters, they see tokens.

When unsure, an LLM still produces fluent, confident-sounding text. It's predicting plausible words, not retrieving facts. Always verify anything that matters — names, numbers, citations.