July 29, 2026
Welcome to the Prompt District
TokenTown: A visual way to understand how LLMs work
This AI city demo wowed some people — and left others saying “what am I even looking at?”
TLDR: TokenTown is a visual website that turns how chatbots generate text into a little city you can watch in your browser. People loved the idea, but the comments quickly turned into a fight over whether it’s a brilliant explainer for beginners or just a stylish, confusing mess.
A new browser toy called TokenTown tries to explain how chatbot brains work by turning the whole process into a tiny moving city. Words get chopped up, loaded onto trucks, driven through factory-like neighborhoods, and eventually turned into the next word on screen. It’s part lesson, part animated theme park ride, and the creator says the point is to help people see why prompts cost money, why word limits matter, and why chatbots sometimes slice words in weird places.
But the real show was in the comments, where the crowd instantly split into two very online camps: “neat idea” and “absolutely not helping.” One user loved that you can paste in text and watch how it gets broken apart, saying it finally makes the “why is this so expensive?” question easier to grasp. Another came in with the digital equivalent of flipping the table: “this doesnt help at all.” Ouch.
Then came the educator drama. Some commenters argued the project expects too much background knowledge, basically saying, “Cool city, but where’s the map for beginners?” Others wanted it to go full 3Blue1Brown style — smoother, clearer, more magical. And the spiciest jab accused the explanations of sounding like generic AI-generated mush, complete with a roast that the copy felt like something tossed over the fence by Claude. So yes: the city is cute, but the comment section turned it into a zoning dispute.
Key Points
- •TokenTown visualizes a transformer language model as an isometric city where each district represents a processing stage in token generation.
- •The tool computes multiple transformer components live in the browser, including tokenization, embeddings, positional encoding, attention, KV cache behavior, LayerNorm, feed-forward layers, and sampling.
- •The model is intentionally scaled down to 12 dimensions, 2 attention heads, 2 to 12 layers, and a vocabulary of a few hundred words.
- •Some elements are deliberately simulated rather than fully faithful: weights are random and untrained, logits are blended with a bigram prior, and attention is modified with sharpening and biasing.
- •The interface includes paced explanations, replay and step-through controls, camera and label options, and navigation tools for exploring the visualization.