zoenkol

Somewhere inside the space of all possible pixel arrangements is every photograph that has been taken, every image that will be made, and an incomprehensible number of scenes that no mind will ever have reason to imagine.

Begin with an ordinary digital image. Under the familiar 24-bit RGB convention, each pixel has three colour channels—red, green, and blue—and each channel can take one of 256 integer values. That gives a pixel 256³, or 16,777,216, possible colours. This is not a rule of nature; modern formats may use different colour depths, alpha channels, or high-dynamic-range representations. It is simply a clean set of constraints for the thought experiment.

A picture as a number

Fix the width and height of the image. Once the number of pixels and the number of values available to each pixel are fixed, the number of possible images is also fixed. For an image containing W × H pixels, the count is:

The number is finite, but “finite” does almost no practical work here. It is so large that enumerating the images one by one is not merely inconvenient. It is beyond any plausible physical computation. The overwhelming majority would look like unstructured noise. A tiny and difficult-to-define fraction would contain recognisable objects, readable text, diagrams, faces, landscapes, or works of art.

Still, each possible image can be placed somewhere in an ordering. Read the pixels from left to right, encode their channel values as bits, and the complete image becomes one very large binary integer. Increment that integer and another image appears. Continue long enough—in the purely mathematical sense—and every image of that exact size and colour format will occur once.

Possible is not the same as present

This is where the language becomes philosophical. It is tempting to say that all of these images already “exist” in a Platonic space. Mathematics supports the more modest claim that every valid image corresponds to a possible configuration, and that every configuration can be assigned a location in a finite set. It does not follow that the images are stored somewhere, waiting to be retrieved, or that they exist physically before a machine renders them.

The distinction matters. A coordinate can be defined without anyone visiting it. A sequence can be possible without being instantiated. The space is real as a mathematical object under the chosen rules; the status of its unseen members as “existing things” is a metaphysical interpretation rather than a computational fact.

Now let the images move

A movie adds a time dimension, but it does not change the logic. Fix the resolution, colour depth, frame rate, and duration. A two-hour film at 24 frames per second contains 172,800 frames. If each frame is a 1920 × 1080, 24-bit RGB image, the number of possible silent films is:

Inside that set are every two-hour arrangement of those frames: static noise, blank screens, almost identical films differing in one pixel for one frame, every physically accurate recording the format could represent, every impossible visual effect, every coherent story, and every incoherent sequence between them.

The set only remains finite because the constraints remain fixed. Remove the limit on duration, resolution, colour precision, or audio length, and the question changes. “All possible movies” without a bounded format is not the same finite collection.

AI does not search the ocean

This vast configuration space gives us a useful metaphor for generative AI, but the metaphor can easily become misleading. A modern image or video model does not iterate through possible outputs and inspect them until it finds one that matches a prompt. No useful system could perform that search. The space is too large, and there is no simple test that identifies the single correct image because prompts usually admit many acceptable answers.

Instead, the model learns statistical structure from examples. Natural images are not arbitrary pixel arrays. Nearby pixels are correlated. Edges continue. Lighting changes in patterned ways. Faces have recurring geometry. Language associates words with visual features and relationships. Video adds regularities of motion, persistence, causality, and viewpoint. These constraints make the meaningful parts of the full space highly non-uniform.

AI is not opening a catalogue of every possible movie. It is learning where human-like movies are likely to be.

Different model families implement this in different ways. A diffusion model begins with noise and repeatedly transforms it toward an image or sequence that is probable under its learned patterns and compatible with the prompt. An autoregressive model predicts successive tokens or patches from what has already been generated. In both cases, the system uses a learned probability distribution rather than exhaustive enumeration.

“Navigation” is therefore a metaphor for guided sampling. The prompt, the training data, the architecture, the model parameters, and the random seed constrain which regions can be reached easily. They do not provide a literal map with a named coordinate for every possible film. Nor is a particular model guaranteed to be able to produce every member of the full pixel space. Its reachable outputs may form only a minute, structured subset.

The real achievement is compression

Random selection from the complete space would almost always produce noise. A useful generator succeeds because it has absorbed enough regularity to avoid almost all of that space. In this sense, intelligence looks less like raw exploration and more like compression: the discovery of rules, correlations, and abstractions that make some configurations vastly more plausible than others.

A prompt such as “a violinist playing beneath a red moon” does not identify one pre-existing bitmap. It imposes constraints. The model then constructs one of many configurations that satisfy those constraints according to its learned distribution. The result may be new as an exact arrangement of pixels even though its possibility was already contained in the mathematics of the format and its structure is indebted to patterns learned from existing work.

That leaves a useful tension. The space of outputs is finite, but effectively untraversable. Every bounded digital artefact is representable as a number, but its meaning is not contained in that number alone. AI can produce remarkable points in the space, but only because human culture, physical regularities, and prior examples have given the space shape.

Perhaps the mystery is not that every possible image can be numbered. The mystery is that, from an ocean dominated by noise, minds and machines can repeatedly find configurations that mean something.