How LLMs Actually Work — An Interactive Explainer

Everyone uses language models now. Almost no one has a mental model of what's happening between hitting enter and the answer streaming back. The explanations on offer split into two useless extremes: the cocktail-party version ("it's just autocomplete") and the graduate-seminar version (twelve equations and a reference to Attention Is All You Need).

I wanted the thing in the middle — a page you can scroll through in ten minutes that shows each step with a live animation, lets you poke at the parts that are interactive, and never asks you to take a leap of faith. So I built one.

What it walks through

The page follows one short prompt all the way through a transformer, one concept per section, each with its own animation:

• Tokenization — how your text gets chopped into tokens, and why a token isn't always a word. (Type something and watch it split.)
• Embeddings — how each token becomes a long list of numbers, and why "king − man + woman ≈ queen" is a real thing you can do with arrows.
• Positional encoding — how the model knows word order, since the math otherwise treats your sentence as a bag.
• Attention (and multi-head attention) — the actual heart of it: how each word decides which other words to "look at," shown as live scoring bars.
• Feed-forward, residuals, and depth — the plumbing that lets these layers stack dozens deep without falling over.
• Prediction — how it all collapses into a probability for every possible next token, and how temperature reshapes that into the word you actually see. (There's a slider.)
• Architecture vs. weights — the distinction most people miss: the blueprint is public, the learned numbers are the moat.

The goal was never to make you an ML engineer. It's to replace "it's magic" with a working intuition — enough that the next time a model surprises you, you have somewhere to put it.

The conceptual skeleton was inspired by this excellent explainer; the page, the animations, and the framing here are my own ELI5 take.