Raise the
machine.
Behind this page is a real neural network — 2.3 million learned numbers that can tell a dog from a cat. You are not going to read about it. You are going to train it, map its mind, deceive it, race it, and correct it. Seven experiments. Everything runs here, in your browser.
Teach it, from nothing, in seconds.
Bolted onto the machine’s eyes is a brand-new decision-maker that has never seen an animal. It will learn dogs-versus-cats from 200 photographs while you watch — and sit an exam of 40 photos it is never allowed to study.
decision weights randomised. i currently answer by coin flip. show me the photographs.
The machine’s eyes — dozens of pattern-finding layers — stayed frozen; they already know edges, fur, and shapes. You trained only the final decision: a few hundred numbers mapping “what the eyes report” to dog-or-cat. Borrowing trained eyes and teaching a new decision is called transfer learning, and it is how most real AI products are actually built.
The space where it thinks.
Just before deciding, the machine has compressed every photograph into a single point in an abstract space — dogs gathering in one region, cats in another. This is that space, flattened so a human can walk it. Drag to pan. Scroll to dive.
to me, every photograph is a location. things that are alike are simply things that are near.
“Learning” is exactly this: finding a squashing of 67,500 pixel numbers down to a point, such that kinds gather. Photos stranded between the two clouds are the ones the machine hesitates on. Every modern AI — chatbots included — organises its knowledge as geometry like this.
Interrogate it. Then betray it.
Register the photograph the next two experiments will operate on. Move your cursor across it — the readout shows the raw numbers the machine actually receives.
Waiting for the model…
Fool it with pixels you cannot see.
The same backwards trace that reads its mind can be weaponised. If we know which pixels push the decision, we can nudge each one — too faintly for any human eye — until the machine flips its verdict entirely.
warning. you are about to use my own gradients against me. i will not notice.
Watch the verdict, not the photo — the photo will not visibly change.
The photograph on your screen did not change in any way your eye can measure — and the machine went from near-certain one way to near-certain the other. Machine sight and human sight are different senses that happen to agree most of the time. The same trick has fooled real vision systems reading stop signs.
You, versus it.
Five rounds. A photograph appears almost totally blurred, then sharpens. The machine locks a guess at every blur level; answer whenever you dare. The earlier you answer, the more it’s worth.
i do not see outlines the way you do. i see texture statistics. blur is kinder to me than to you.
Press start when you’re ready.
Teach it your things.
Dogs and cats were only the demonstration. The machine’s frozen eyes can describe almost anything — so hand it two objects from your desk. A mug versus a phone. Your left hand versus your right.
six photographs of each will do. i have seen a million objects before yours. vary the angle.
Take a large pre-trained model, freeze it, teach a thin new layer on a handful of your own examples. You didn’t need 20,000 photos, because the hard part — learning to see — was already done.
Teach it to say “I don’t know.”
Show this machine television static and it will declare, with total confidence, that it is looking at a dog. Not because it is stupid — because nobody built the option to refuse. You are the builder now. Build it.
i have exactly two words. give me a third and i will use it.
“I don’t know” was never a missing miracle — it was a missing decision by a person. You made that decision, generated the training data for it, and shipped the correction in ten seconds. When you evaluate any AI product, ask the question you just answered: what did its builders decide it should do with the unexpected?
You didn’t watch a machine learn.
You made one.
In one sitting you trained a neural network, mapped the space it thinks in, traced a decision backwards through its layers, forged an invisible disguise that fooled it, raced it on raw perception, taught it objects from your own desk, and gave it the humility its makers left out. That is not a metaphor for how modern AI is built. It is, at small scale, exactly how modern AI is built.