Manifold Hypothesis
The claim that real data occupies a tiny, smooth, low-dimensional subset of the space used to store it.
Here's the evidence, and it's countable. A 100 × 100 greyscale image is a point in 10,000-dimensional space, giving roughly possible images. Sample one uniformly at random and you get static — every time, for as long as you care to try. So images that look like anything are a vanishingly small part of that space.
They're also smooth: rotate a head two degrees and you're still in the set of faces. A tiny, smooth, curved subset is a manifold, and its intrinsic dimension is how many numbers genuinely vary — perhaps 30 for a portrait, against the 10,000 you're storing.
Two consequences do real work. It's why compression to a short vector loses so little. And it's why a model handed an off-manifold input returns a confident, meaningless answer, which is the mechanism behind adversarial examples.