Compute Is the Aperture

A telescope is not made of glass. A telescope is made of aperture — the diameter of the opening that lets light in. This is what compute is, for AI.

A telescope is not made of glass. A telescope is made of aperture — the diameter of the opening that lets light in. Galileo’s first telescope, in 1610, had an aperture of about 1.5 inches. He used it to see four moons orbiting Jupiter. The James Webb Space Telescope, four hundred years later, has an aperture of 21 feet. It uses that aperture to see galaxies that formed three hundred million years after the Big Bang.

The galaxies were always there. The moons were always there. The difference between Galileo and Webb is not the universe. The difference is how much light each instrument can gather.

This is what compute is, for AI.

People talk about compute as if it were a kind of engine — bigger engine, faster car. That metaphor is wrong, and the wrongness is the whole story. Compute isn’t an engine. Compute is an aperture. It determines how much of the underlying world the system can resolve. Make the aperture small, and the system sees only the brightest features. Make it bigger, and dimmer features start to appear. Make it bigger again, and structures show up that nobody knew were there.

In 1922, the astronomer Lord Rayleigh wrote down the equation for this. Resolving power scales linearly with aperture. Want to see something twice as faint? Double the diameter of your mirror. There is no shortcut. The equation does not care about your cleverness.

In 2022, three researchers from DeepMind published a paper called “Training Compute-Optimal Large Language Models.” It’s known now as the Chinchilla paper. They worked out, with a lot of careful experiments, how the loss of a language model scales with the size of the model and the amount of data fed into it. The headline result became famous in the AI industry: for any given budget of compute, there’s an optimal balance between model size and data, and most of the models being trained at the time were getting it wrong.

But the deeper result — the one nobody outside the field talked about — is that the loss curve is predictable. Once you know how much compute you have, you can predict, in advance, what the model will be able to do. The aperture determines what you can resolve.

Same equation, different domain. Rayleigh’s law for telescopes. Chinchilla’s law for language models. Both of them say: the world is full of structure, but you can only see the structure your aperture allows.

This explains a lot about why AI keeps “surprising” people.

The ability to write coherent paragraphs was always latent in the structure of language. It didn’t appear in 2022 because somebody invented it. It appeared because, by 2022, training runs had become big enough — apertures wide enough — that the structure became visible. The same is true for the ability to do basic reasoning, to write code, to handle multiple languages, to follow instructions across a conversation. None of these were built. They were resolved, the way a galaxy is resolved when your mirror gets big enough.

This also explains the things that are still missing. Long-horizon planning. Genuine novelty. The kind of judgment that knows when to stop. These are not impossible — but they require apertures we haven’t built yet. The structure may be there in the data. We can’t see it, because we don’t have the lens.

The implication for anyone trying to predict what AI will do next is uncomfortable.

Most prediction in the AI industry is in the form of “the model can’t do X, therefore it never will.” This is exactly the prediction someone in 1900 would have made about seeing the surface of Mars. It can’t be done with current telescopes. Therefore, it can’t be done. They were wrong because they confused the limit of the instrument with the limit of the world.

The opposite mistake is also common. “The model can do X, therefore it understands X.” This is the prediction someone in 1610 would have made when they first saw Jupiter’s moons — aha, I have understood the heavens. They had not. They had merely opened the aperture wide enough to see something the naked eye couldn’t. Resolution is not understanding. It’s just resolution.

The honest version of where AI is, right now, is this: we have built the largest aperture in the history of intelligence. It is showing us things that were always present in human language, never before resolvable. Some of those things are wonderful. Some are alarming. Most of them are still slightly out of focus, because we are at the edge of what current apertures allow.

Whether anything more is possible — whether there are structures in language that no aperture, however large, will ever resolve — is the open question of the field.

For now, what we have is a telescope. We are pointing it at the sky of meaning, and discovering, planet by planet, what was always already there.

Part of the Aperture series at Nous — six essays on the physics, history, and structure of the AI industry. The argument that resolution is not understanding pairs with Hexis Asks, Logos Guesses in the Logocachexia series.