Everything the world currently calls "AI" sits on top of seventy years of older ideas, most of which still quietly run the software you use every day. This is that stack, dug through one layer at a time: hand-written logic at the surface, then models fitted from data, then networks, then depth, then generation, then agency at the bottom. Every demo here implements its algorithm from scratch and runs it live in your browser — nothing is precomputed, nothing is faked, and if you refresh the page it all happens again.
Search and logic. A human writes the rules; the machine follows them exhaustively and can explain every step.
Minimax searches every possible future of the game before each move. You cannot win — the best you can do is draw, and the page shows you how many board states that certainty costs.
recursive game-tree search · backtracking open →A 1980s expert system: twelve hand-written IF/THEN rules, asked one question at a time. It never asks the same thing twice, and it shows you exactly which rule died on which answer.
forward chaining · lazy evaluation · full trace open →the waterline. Below this point nobody writes the rules any more — the machine derives them from data, and gets steadily better at things it cannot explain.
Fit a model to examples. Still the most commercially valuable layer of the whole iceberg.
TF-IDF and naive Bayes, built from nothing, trained on 24 messages at page load. Type anything and watch the verdict and the per-word evidence update on every keystroke.
tf-idf with bigrams · multinomial naive Bayes open →200 deliberately corrupted sales rows — duplicates, sentinels, three date formats, "TORONTO" and "toronto". Clean it step by step and see why the order of operations decides whether the numbers lie.
seeded corruption · pipeline · canvas chart open →Gradient boosting over 3,000 customers, trained live. The data comes from a risk formula the model never sees; the test is whether its importance bars rediscover the drivers hidden in there.
boosted stumps · AUC from scratch open →Stacked layers of weighted sums, trained by the chain rule. Flexible, and permanently unable to explain themselves.
The same mathematics, stacked deeper and fed more data. "Deep" has never meant anything more exotic than that.
Predict the next token, then do it again. The whole family, from lookup tables to frontier models.
A model placed inside a loop, with tools. The difference between answering and acting.
Every page is a single HTML file with its CSS and JavaScript inline — no frameworks, no CDN, no bundler, no dependencies of any kind. Each algorithm is implemented from scratch in plain JavaScript and executes in your browser while you watch: the classifier really is fitted on page load, the boosted trees really are grown one stump at a time, the network really does backpropagate. Where a Python library has no browser equivalent, the page says so in a dashed box and names what it substituted and what that costs. The two pages that talk to a language model call the Anthropic API directly and handle failure by saying so rather than by inventing an answer. Nothing is stored on your device — refresh and every demo starts from zero.