What is AI, really?

Draw a number with a finger or a mouse. Nobody ever told this page what a seven looks like — it has only ever seen numbers other people wrote down and labelled.

Draw a digit, 0–9

OR PICK

All it actually sees

Sixty-four numbers. The loops and strokes are gone.

Its guess

Draw something to begin

The five it most resembles

800 real handwritten digits, written by 43 people for the UCI optical-recognition dataset. On that dataset’s own held-back images this method gets 98% right — on a finger-drawn squiggle, considerably less, because you do not write like those 43 people.

The problem it solves

Try writing down what a 7 is

Every rule you can state, a real person’s handwriting breaks. This is not a hard problem to describe — it is a problem that cannot be described.

Rules a person typed

a horizontal stroke on top✗ so has a 5, and a 2

…plus a diagonal going down-left✗ many people write it straight

…and no closed loop✗ so has a 1, and a 4

…and no crossbar✗ half of Europe crosses their 7

Each patch breaks something that worked before. The rulebook never converges.

Patterns learned from examples

Forty of the 800. Nobody described a single one — a person just wrote down which digit it was, and the machine measures its way to the rest.

How it works

Three steps, no understanding required

Turn it into numbersA real 7 from the training data. Sixty-four brightness values are the whole input.

Compare to the pastDistance is measured against all 800 examples a person already labelled.

Let them voteThe majority label becomes the answer. That guess is the whole “intelligence”.

See it break

It cannot tell you it doesn’t know

Go back up and draw a letter, a star, or a scribble. It will still answer with a digit, and it will look just as sure as before.

There is no “none of the above” anywhere in the machine. It was built to return its closest match, so a closest match is always what you get — a habit worth remembering when a much larger model answers a question it has no business answering.

So what

Photo search, face unlock, fraud alerts — all rhyme with the same idea: compare it to labelled examples, and let the closest ones win.