Ten demos · nothing pre-computed

The AI Iceberg

Above the waterline sits everything that gets called artificial intelligence this year; below it sits the sixty years of work that made it possible, and the layers get stranger and more fundamental the deeper you go. Each demo here is one layer of that iceberg, ported from a terminal program into a page you can poke at. Every algorithm actually runs in your browser, from scratch, in plain JavaScript: the models train live in your tab, so nothing on these pages is a recording.

WATERLINE what people mean by "AI" this year CLASSICAL AI — search and logic MACHINE LEARNING — learning from data NEURAL NETWORKS — learned representations DEEP LEARNING — depth and scale GENERATIVE & AGENTIC — models that act
Layer 01

Classical AI

No data and nothing learned. Intelligence as exhaustive search and hand-written rules, which is what the field meant by the word for its first thirty years.
01

Unbeatable Tic-Tac-Toe

Full minimax over the game tree, with a toggle that shows how many board states it evaluated for its last move.

runs: recursive search, in-tab 02

Plant Doctor

A miniature MYCIN. Twelve hand-written rules, a symptom cache that never asks twice, and a trace of exactly which rule died on which answer.

runs: forward-chaining inference
Layer 02

Machine Learning

The system writes its own rules by reading examples. It generalises past what anyone typed in, and in exchange you can no longer read what it learned.
03

Spam Classifier

TF-IDF with bigrams and multinomial naive Bayes, both from scratch. Type a message and watch the verdict move on every keystroke.

trains on page load 04

Messy Data Cleaner

Two hundred rows of realistic corruption, cleaned step by step with a reason for each step's position in the order. Canvas chart, no libraries.

seeded generator, reproducible 05

Churn Predictor

Gradient boosting on 3,000 customers with a hidden risk formula the model has to rediscover. Ends with the five people to call today.

300 stumps, trained live
Layer 03

Neural Networks

Stack simple units, let the chain rule assign blame backwards, and the middle layer invents its own features. This is where learned representation begins.
06

Neural Net From Scratch

The 2-4-1 XOR network with hand-written backprop. Watch the loss sit on its plateau and the decision boundary morph out of noise.

live loss curve + heatmap
Layer 04

Deep Learning

The same idea with more layers and far more parameters, which turns out not to be a detail. Depth is what let networks stop needing hand-designed features.
07

Draw a Digit

A four-layer network trains on 400 real handwritten digits while the page loads, then reads whatever you scrawl on the canvas.

17k parameters, trained in-tab
Layer 05

Generative AI

Predict the next token, append it, repeat. The loop is sixty years old; only the size of the context and the richness of the representation changed.
08

Markov Text Generator

An order-2 chain that sees two words of context, next to Claude answering the same five-word prompt. One click apart.

lookup table vs frontier model
Layer 06

Agentic AI

Give the model functions it can ask for and a loop to ask in, and it stops answering questions and starts doing things. That is the whole architecture.
09

Mini Agent

A model with no clock, no calculator and no memory, four local tools, and a five-step loop. Every tool call is printed as it happens.

tool-calling loop, no framework 10

MEOW, a Persona Agent

The same loop with the tools removed. What is left is a system prompt, editable live on the page, so you can watch character change without touching a weight.

persona = one editable paragraph

Every page is one standalone HTML file. All CSS and JavaScript inline, no libraries, no CDNs, no build step, no analytics, and no browser storage of any kind.

Nothing here is pre-computed. The classifiers, the boosted trees and both neural networks train from random initialisation in your tab each time you open the page. Three pages call the Anthropic API, and say so on the page; the other seven make no network request at all.