Try it · nothing is uploaded, it all happens in your browser
Every file has a shape.
A file is just a long row of numbers. Draw those numbers as heights and you get a landscape: hills, valleys, ridges, noise. Some files turn out to be a smooth curve wearing a disguise. Others are a mess with no pattern at all. Drop one in and see which yours is.
Waiting for a file.
Your image
What the shape alone paints, before the corrections are added back
Music that was never recorded
Here is the same idea with your ears instead of your eyes. The notes below are read
straight off the surface of z = x·y: the bar number is x, the voice is
y, and the height of the surface at that point picks the note. The bass line comes from
z = x/y, the same relationship running the other way. Nothing is stored. The
audio does not exist until you press play, and it is gone again afterwards.
What this does and does not show. It shows that when music is defined by a rule, the rule is the only thing that has to exist: a few dozen bytes standing in for a few megabytes of samples, rendered on demand. It does not show that a recording of a real performance can be shrunk to that size, because a recording is not a rule; it is the sound of a room, and the benchmark measures what happens when you try. The synthesizer that turns the rule into sound is shared code, the same way a codec sits on both ends of an MP3 and is not counted in the file's size.
What you are looking at
The blue line is your file
Each byte becomes a height between the floor and the ceiling. Reading left to right is reading the file from beginning to end. Music looks like waves. Photographs look like rolling terrain. A zip file looks like static.
The orange line is a guess
A handful of simple shapes are tried against your file (a flat line, a slope, a curve, a wave) and the closest one is drawn. That shape can be written down in a few dozen characters, which is the entire point.
The green band is the difference
Wherever the guess missed, the exact difference is kept. Shape plus difference rebuilds your file perfectly, byte for byte. Nothing is thrown away, ever. The only question is how much of the work the shape did.
Why bother, in everyday terms
A recipe is shorter than a cooked meal. If I send you the recipe, you can make the meal yourself, and I never had to ship the food. A circle is another example: you could list a million points along its edge, or you could write down one number, the radius, and redraw it whenever you like. Same circle, wildly different amount of storage.
That is the whole idea. Where data follows a pattern, keep the pattern instead of the data, and rebuild it when someone actually asks. A smooth sensor reading, a tone, a gradient, a fractal: those are recipes wearing a disguise, and the recipe is tiny.
And where it does not work, it says so. A photograph is not a simple curve, and a zip file is deliberately patternless. For those, the shape explains almost nothing and ordinary compression does better. That is not a flaw being admitted quietly; it is the honest boundary, and the benchmark page measures exactly where it sits on real files, including ones where this approach loses outright.
The one trick underneath
If you can guess what comes next, you do not have to store it. You only have to store how wrong your guess was. A good guess leaves small corrections, small corrections squeeze down to almost nothing, and the rule that made the guess is a few dozen characters. That is the mechanism in one breath, and the technical write-up takes it apart properly.