The Universe Is the Better Teacher

There is a theorem that sounds like a promise. A neural network with enough width can approximate any continuous function to whatever precision you like. Cybenko proved it for sigmoids in 1989, Hornik made it general two years later, and ever since we have repeated it as a kind of charm: the network can learn anything. It is true, and it is also a trap, because it answers a question nobody asked while hiding the one that matters. The theorem tells you the network can match a function. It does not tell you which function you handed it to match.

So ask that question instead. When we train a model, what is the function on the other side of the equals sign?

The function we are actually fitting#

Look at where the data comes from and the answer is uncomfortable. The text was written by people. The labels were drawn by people. The captions, the preferences, the thumbs-up and thumbs-down that steer a model after pre-training, all of it is human judgment, recorded and replayed. When we minimize a loss against that data, the target function the universal approximator is reaching for is not “truth” or “intelligence” in the abstract. It is us. The function is human intelligence, sampled and frozen into a dataset, and the network’s whole job is to come as close to that function as it can.

That reframing should change how the theorem feels. The universal approximator is universal in its means and entirely bounded in its target. It can take any shape, but we keep pouring it into the same mold. We have built the most flexible function-fitters in history and aimed every one of them at a single function: the one that runs in a human head.

A student cannot pass its teacher by copying#

Here is the thing that should bother anyone who wants more than a clever mimic. Approximation has a ceiling, and the ceiling is the target. A function fit to human judgment converges toward human judgment; in the limit of infinite data and a perfect optimizer it reaches human judgment and stops. That is not a flaw in the network. It is the definition of the task. You cannot exceed a function by approximating it. The best possible student of a teacher is a copy of the teacher, and a copy does not surpass.

This is why imitation alone was never going to clear the human bar. Every place the data is human, the optimum is human. We can make the copy faster, cheaper, tireless, able to hold more at once than any person, and those are real and enormous advantages. But they are advantages of scale and speed applied to a fixed function, not a better function. Run a human-level judgment a million times a second and you have a very fast human-level judgment. The shape of what it knows is still the shape we gave it.

So the interesting question is not “how do we approximate better.” We are already good at that. The question is what to approximate instead. If you want a student that surpasses its teacher, you do not give it a better copy of the same teacher. You give it a better teacher.

What is more intelligent than us#

Name something that is demonstrably smarter than human cognition and is not itself a human artifact, and the list is short. It has exactly one entry.

The universe.

Consider what it does without effort. It folds proteins into their ground state in microseconds, a search we throw datacenters at and still get wrong. It routes light along the path of least time before any equation is written down. It holds a galaxy together and steers an electron through a slit with the same small set of laws, and those laws are consistent, compositional, and exact in a way no human theory has ever fully matched. The universe is not intelligent in the way a person is, with intentions and a voice. It is intelligent in a deeper sense: it computes the right answer to staggeringly hard problems, everywhere, continuously, and it never contradicts itself. Its laws are the function we have been trying to glimpse through the keyhole of human description.

And crucially, that function is not bounded by us. Human physics is our best approximation of the universe’s laws; the laws themselves are the ground truth that our physics is the student of. When we fit a model to human knowledge, we fit it to the approximation. The universe is the thing the approximation approximates. It sits a level below, where the real answers are kept.

Pointing the approximator at the world#

This is where the universal approximation theorem stops being a trap and becomes a plan. The network’s flexibility was never the problem. The target was. So change the target. Train against the universe instead of against the transcript of what people have said about it.

We already know what that looks like when it works, because the systems that genuinely exceeded human skill did exactly this. The game-playing engines that went past every human did not learn from human games; they learned from the rules, played against themselves, and let the structure of the game, a tiny closed universe with exact laws, be the teacher. The model for protein structure surpassed decades of human guessing because its teacher was physical plausibility, not the literature. In each case the supervisory signal came from a law, not from a person, and the moment that happened the human ceiling lifted, because the function being approximated was no longer human.

Generalize the pattern and a research program falls out of it. Build models whose loss is grounded in physical law rather than human annotation. Let them predict the world and be corrected by the world: by simulation where the equations are known, by experiment where they are not, by self-play against a reality that does not negotiate. Use human knowledge as the scaffold, the way we use a coarse initialization, and then let the universe do the teaching that takes the student past the scaffold. The approximator stays the same. We just finally point it at the deeper function.

None of this asks the network to be less than human first. It asks us to stop defining its target as human. The universality was always there, waiting for a worthier function to fit. The universe is offering one.

Perhaps artificial intelligence’s greatest limitation has been our stubborn fixation on the human brain as the pinnacle of intelligence. The universe itself, governed by elegant and powerful laws, demonstrates intelligence far beyond human cognition. These fundamental laws, which shape galaxies and guide quantum particles, represent a deeper form of intelligence that we have largely ignored in our pursuit of AI.

← all writeups