A demonstration · September 30, 2026
AI is here
to stay.
Everything on this page — the glass and its light, the network learning in front of you, the words that gather from particles — was designed, written and tested by an AI in one session, from a seventy-word request. All of it is running live in your browser, right now.
01 / It rendersMove your pointer. The glass turns to follow it.
—starting
The material studio
The glass is a formula.
Change it.
Every pixel above is computed from a few numbers: how much the glass bends light, how differently it bends red and blue, how the body is twisted, how fast it turns. Move them and the sculpture is recomputed sixty times a second.
02 / It learns
A network that knew nothing
a few seconds ago.
On the left, a picture. On the right, a small neural network's attempt to paint it. The network starts from random numbers every time this page opens and trains here, in your browser, on a background thread: it is shown a hundred and twenty-eight pixels at a time and adjusts its 6,595 parameters to be less wrong. Nothing is downloaded and nothing is sent. Draw your own picture and it will learn that instead.
Starting…
- Steps
- 0
- Loss
- —
- Fidelity
- —
- Speed
- —
- Elapsed
- 0 s
- Parameters
- 6,595
Teach it something else
Draw, then press Teach it this.
Whatever you draw becomes the target. Simple shapes in strong colours learn fastest; fine detail takes the network longer, which is worth watching.
03 / It listens
Type a word.
Thousands of points of light are looking for somewhere to be. Give them a word and they will spell it; move through them and they scatter and return.
— particles, simulated one by one, sixty times a second.
AI IS HERE · TO STAY · A TOOL · IN YOUR HANDS
04 / It tunes itself
It is measuring
your device.
A page like this cannot assume a fast computer. So it watches its own frame times and decides: how many pixels the glass is drawn at, how many particles to simulate, how much of each frame the network may use. These are its decisions for the device you are holding, as it makes them.
- Waiting for the page script…
05 / How it was made
One request.
One session.
The numbers below are not estimates. The page reads its own source files when it is built and counts them; the commit is the one that produced what you are looking at. This page has been open for 0 seconds.
- 70
- words in the requestOne message from the studio's founder. No brief, no mockup, no code.
- under 60
- minutes from request to liveDesign, shaders, the network, the particles, the copy, the tests, the deployment. Timed from the message to the runner's first check of the live page.
- 9
- source filesglass.ts · index.ts · learn-worker.ts · learn.ts · particles.ts · sound.ts · tuner.ts · ai.astro · ai.css
- 768
- lines of code and copy48 KB before compression; 131 of the lines are GPU shader code.
- 6,595
- parameters in the network34 inputs (position + 16 Fourier features), two hidden layers of 64, 3 outputs. Trained from random on every visit.
- up to 40,000
- particlesThe count follows your device's frame time; start with 22,000 on a computer, 9,000 on a phone.
- 2adda84
- commitBuilt 2026-10-09 18:26 UTC. The page reads its own files at build time to produce these numbers.
- The request arrives: seventy words.
- A plan: what would be spectacular and what would actually work, in this codebase, with its gates.
- The engines, written in one pass: the glass (a signed-distance sculpture, refraction per colour channel, a mirrored floor, one shader), the network (forward and backward passes by hand over typed arrays, Adam, a worker), the particles, the sound, the tuner.
- The page: copy, layout, the material studio, the drawing pad, accessibility, reduced motion. The social image rendered with the site's own fonts.
- Twenty-five type errors fixed. First build passes the repository's gates.
- First run in a headless browser: the network reaches 28 dB on the glass identity in nine seconds; a drawn stroke, 44 dB.
- Screenshots at two widths; the overlaps they showed fixed; rebuilt.
- Documentation written. Pushed. Vercel builds it. A runner waits for the deployment, then checks and measures the live page.
06 / What it means
A tool.
Not an author.
This page did not decide to exist. A person asked for it, in seventy words, and a person reviewed it and chose to publish it. What the tool did was everything in between: it held the whole design in mind at once, wrote the shader and the training loop and the copy in the same hour, tested itself, and fixed what it found. A studio of specialists could make something like this. It would take them weeks, and this page would still be waiting.
That is the point, and it is a practical one. The question for a business is no longer whether this kind of capability is real. It is what you would build with it, and who you would trust to hold the tool. Glass City Intelligence is a Toledo studio that has been building with it since the beginning: websites, software and the systems behind them, delivered with the judgement a tool does not have.
Runs on WebGL2, Web Workers, Web Audio and the 2D canvas. The demonstrations send nothing anywhere; the site's ordinary analytics apply, as described in the privacy page. If your browser cannot run part of it, that part says so and the rest carries on.