★ SUMMER INTO AI 2026 ★

The final week's theme is build on yesterday — so we went back to our very first competition, Spring Into AI, and rebuilt our Method-of-Loci memory app from the ground up. Assign what you want to remember to a glowing beacon in an explorable 3-D signal station (real Three.js scene, not a flat photo with hotspots), then tune in: a real broadcast voice — the Keeper — walks every beacon in order, with spatial audio tied to each beacon's actual position, or a numbers-station cipher mode you decode yourself. Download the broadcast as one real audio file, or share a link — anyone who opens it hears the exact same transmission (the feature our original Spring build had to cut). Same idea, one competition and a lot of practice later.

Also built on yesterday — but on Week 2 this time. Back then I made ~1,000,000 GPU particles, each its own independent agent with no central control, hold one fixed shape: the flag. This week I let them hold anything you commit to the union. Type a word — 1776, WE THE PEOPLE, your own name — and a half-million free points assemble it out of nothing (your text rendered to a field of targets they spring toward, all on your GPU). Drag across them to tear it apart, and with no one giving the order, it re-forms on its own. A union isn't the symbol — it's an equilibrium a half-million free parts return to, around whatever the people decide to hold together. e pluribus unum. ★

Built on a competitor's yesterday this time — a tribute-extension of Adam's Ride, Revere, Ride!, his tight little Paul Revere lane-runner. His redcoats spawn at random; mine learn you. Behind the arcade ride sits a from-scratch learning model (no libraries) that watches which lanes you favor and how you move between them, then places each redcoat where it predicts you'll be — so the British get better at catching you the longer you ride. And you can watch it think: the Crown's Mind panel shows its live guess of your next lane, and its running read-rate on you. Hit Freeze the Crown and the whole thing collapses back to blind random spawns — the toggle that proves the intelligence is real, not scripted. Warn every town, dodge the redcoats, beat the dawn. The one thing arcade games never have: an opponent that learns. 🏇

The government took $7,009,973,667,049 last year and never mailed you a receipt — so I built one. Type what you paid in federal tax and watch it get itemized to the dollar: how much of your money went to Social Security, to Medicare, to the military — and how much went just to pay interest on money the government already borrowed (more than the entire military; 7× what you paid for NASA, foreign aid, and every national park combined). Then the part they'd rather you didn't add up: for every $1 you paid, Washington spent $1.34 — the other 34¢ borrowed in your name. The grand total is pulled live, in your browser, from the U.S. Treasury's own public API (real FY2025 actuals, cross-checked to the dollar); the receipt prints as an art-directed punk broadside with a generative seal and barcode. Nothing sent anywhere. Punk isn't dead. It's compiling. 🧾

Everyone else this week asked an AI about the government's data. This is the opposite: a neural network hand-written from scratch — the forward pass, the backprop, the optimizer, no libraries — that learns from the data itself, live in your browser, on the U.S. Dept. of Education's real numbers for what graduates of every college actually earn. Watch it teach itself to predict earnings from a school's cost, size, selectivity, and type — the glowing web of neurons rewiring, light pulsing through the wires, its guesses snapping onto the truth (R² ≈ 0.70, held-out), the math gradient-checked so you know it isn't faked. Then ask it yourself: "is this degree worth it?" Unlike the black boxes the government and the AI labs keep, you can see every neuron and every weight as it thinks. Punk isn't dead. It's compiling.

A living 3-D Earth built entirely from federal data the government collects every minute and files where no one looks. Spin the globe and turn on the living layers: ~9,000 river gauges reporting their flow live from the USGS, the white-nose bat plague, the current drought belt, endangered-species habitat, every earthquake of the last month (the Ring of Fire alight), active wildfires, volcanoes, a century of hurricane tracks, and the live aurora from NOAA space-weather satellites. Real-time day/night Earth with city lights; a ⏳ time-machine to scrub 1900→today; fly to your home for your local readout; a one-click cinematic tour. Every point is real, live, public data — the planet as the instruments actually see it. For Natasha. 🌎

The government counted a catastrophe county by county — and never built you a way to watch it. So I did. White-nose syndrome is a cold-loving fungus that has killed millions of hibernating bats since a single cave near Albany, NY in the winter of 2006. This is the federal occurrence record made visible: press play and watch the white spread out of New York, winter by winter, 2006 → 2019, each glowing dot a real county the winter it was hit. Watch it leap 1,300 miles to Washington State in 2015–16, then take Texas, the Plains, California. Pick your state to see when it reached you; three bat species are already down more than 90%. Real USGS / USFWS data (555 counties, 13 winters), baked into one file. Punk isn't dead. It's compiling. — for the wildlife rehabbers on the night shift. 🦇

The graphics finale, rendered on your GPU. A million points of light — each its own sovereign agent following only its own pull, no conductor, no central control — together hold one shape: the waving flag of the union. Drag across them and they scatter into chaos, independence pulling every which way; let go, and with no one giving the order, they find their way back into one. Real-time GPU particle simulation (a million agents, additive HDR bloom, a flag that waves in 3-D). Out of many, one — and the one survives being torn apart. A union isn't held together from the top; it's an equilibrium a half-million free parts return to on their own.

Independence as real cryptography. What makes a union legitimate with no king to stamp it? Consent — proven by math. All thirteen colonies generate their own key in your browser, no server (genuine ECDSA P-256, Web Crypto) and sign their ratification into one tamper-evident chain. Then forge one: change a single colony's signed words and that seal — and every seal after it — turns red, rejected by every honest node with no central authority to ask. Become a sovereign yourself: generate a key, sign your own declaration, and watch the proof die the instant you alter a character. Independence as self-sovereignty — answer to no one, prove it to anyone.

Thirteen colonies, each a reinforcement-learning agent hand-written from scratch (no libraries, no LLM, no cloud), starting from zero knowledge. Every round each one chooses: stay loyal and pay the Crown's tax, or revolt. Revolt alone — crushed; revolt together — trade free and the Crown can't hold. Nobody tells them to rebel. They learn it, and a revolution cascades down the seaboard. Slide Tax, Retaliation, and Trade and the learned outcome flips between a loyalist coast and 1776. Independence as an equilibrium that has to be discovered — freeze the learning and the revolution never comes.

A real GPT, hand-written from scratch in one file — attention, backpropagation, and the optimizer, no libraries, no API. It starts as pure noise and teaches itself the voice of the American founding live, while you watch: the loss falls, the gibberish resolves into 1776 cadence, and you see the attention learn. Independence as self-origination — an intelligence with no pretrained ancestors and no cloud. (Backprop verified by a built-in gradient check.)

A live starling flock with no leader — order emerges from the bottom up. Slide the dial from every bird free to one Crown, then strike a node: cut a free bird and the flock routes around it; cut the Crown and the whole sky falls. Independence isn't a symbol here — it's the architecture that survives the strike.

A self-playing anthem with no conductor. Six voices, each on its own independent loop (4–16 beats) that never quite realign — built from scratch with the Web Audio API. Mute any voice, even all but one, and the rest play on. Out of many independent loops, one song. Shareable by seed.

A real neural net, hand-written in one file (no libraries), learns to fly — either from watching you pilot, or by teaching itself from scratch (neuroevolution, no teacher at all). Then it flies a course it's never seen. Watch the brain fire live. Independence = the moment it no longer needs you.

A systems puzzle — cut the Crown's supply lines so independence cascades, then secure each free state before it's re-garrisoned. Freedom is the win condition.

An 8-bit territory battle — repel the redcoat invasion and liberate the USA.

An 8-bit anthem maker — paint a tune on the sequencer, hear it sing, share it.

Chat with an 8-bit Franklin, Washington, Jefferson, or Abigail Adams — a real AI answers back.