The Intuition Behind How Large Language Models Work, Part I (mark-riedl.medium.com)

This piece is an intuition-first explainer of how Large Language Models (LLMs) — specifically Transformer-based models — generate text. Rather than heavy math, it walks through a simple word-guessing game to show that LLMs work by predicting the most probable next token given the preceding context. Key building blocks are embeddings (tokens → high-dimensional numeric arrays) and attention (every token “attends” to every other to weight relevance). Repeating embedding + attention layers many times (dozens to 100+ layers in big models) produces progressively abstract “made-up” concepts, and a decoder finally scores possible next tokens to pick the most likely continuation. The significance is practical as well as conceptual: this view explains why more prompt context yields better output, why LLMs can model long-range relationships (a distant word like “cat” can strongly influence the next token), and why they aren’t thinking agents but probabilistic word predictors. Technical implications include how embeddings learn via training to make related words numerically similar, how attention composes those into higher-level concepts, and why hallucinations and alignment to user-desired outputs remain intrinsic risks — the model aims to guess what you want to hear, not verify truth or intent.