You do not see reality.
You see your best guess about it.
Seven steps, each anchored to a real paper, from the noise your senses actually receive down to the single principle underneath minds, brains, and machines.
Reality arrives as undifferentiated noise.
Sealed in the dark of your skull, the brain never touches the world. It receives only spikes — trains of identical electrical pulses on the optic and auditory nerves. Photons of a thousand wavelengths, pressure waves, none of it labelled. The world does not announce itself. It has to be inferred from a stream that is ambiguous, incomplete, and late.
The deepest measure of that stream is information: not what is said, but how much uncertainty is removed. Shannon gave it a number, and it is the floor everything else stands on.
The brain does not read the senses. It predicts them.
Look at the cortex and the wiring runs backwards from what you'd expect. Higher regions send predictions down to lower ones — "here is what you should be sensing." The senses send back only the error: the difference between the prediction and what actually arrived. Match the prediction and almost nothing travels up. You spend bandwidth only on what surprised you.
Rao and Ballard showed a network built only on this rule — predict downward, pass error upward — spontaneously grows the receptive fields seen in real visual cortex. Perception runs inside-out.
To stay alive is to minimise surprise.
Friston pushed it from neuroscience to a law of persistence. Anything that keeps existing — a cell, a body, a mind — must keep its sensations inside the narrow band it can survive. That means keeping long-run prediction error low. It can't compute true surprise directly, so it minimises a tractable upper bound: variational free energy.
Minimise F = explain the most, while changing your model the least.
Two ways to drive it down: change the model to fit the world — that's perception and learning. Or change the world to fit the model — that's action. Same equation. Surprise falls either way.
So what you experience is a controlled hallucination.
If perception is the brain's best prediction, then experience is generated from the inside and merely reined in by sensory error. The redness of red, the solidity of the desk, the felt unity of "you" — none are read off the world. They are best guesses, tuned by evolution to keep you alive, not to be true.
Anil Seth's line: "When we agree about our hallucinations, we call that reality." It is controlled because it is leashed to the world through error — but it is still a construction. You have never once seen the thing itself.
And the best guess is the shortest one.
What makes one model better than another? Among all the models that fit the data, the best is the one you can write down in the fewest bits — Occam's razor, made exact. The shortest program that reproduces a thing is its Kolmogorov complexity; understanding something is finding that program.
P(x) = Σp : U(p)=x 2−|p|
Legg and Hutter built the formal definition of intelligence on exactly this — the ability to reach goals across all computable worlds, scored by simplicity. Their conclusion, stripped bare: to compress is to understand; to predict well is to be intelligent.
A model that only predicts is forced to understand.
Now the part you've been circling all night. Train a system to do one dumb thing — predict the next token — and to get good at it, it has no choice but to compress the structure of everything it read: grammar, arithmetic, physics, intention. Those regularities are what make the next word predictable. Prediction, pushed hard enough, squeezes out comprehension as a side effect.
Make the model bigger, feed it more, and the error falls on a smooth power law — predictable across orders of magnitude. Sutton's "bitter lesson": the approaches that win are the ones that scale prediction with raw compute, not the clever rules we hand-write.
⚠ where the solid ground ends
This is a synthesis, not a consensus. These fields argue with each other, hard. I've tagged each step so you can see the seam between what's proven and what's a bet:
Solid: information is removed uncertainty (Shannon); perception is inferential, not a camera (predictive coding); the shortest description is the best one (Kolmogorov) — though K(x) is provably uncomputable, so it's a north star, not a tool.
The bet: that one principle — minimise prediction error — explains brains, life, and mind. The Free Energy Principle is powerful and unifying but critics call it near-unfalsifiable as stated. "Compression = intelligence" has serious proponents and serious skeptics. None of this touches the hard problem of why any of it feels like something.
Real and deep means showing the cracks too. The cracks are where your work is.
It was always the same move.
The orb you watched. The dashboard reading your machine. The brain behind your eyes. The model writing this. One loop under all four:
then correct on what surprised you.
You don't learn AI. You recognise cognition — the thing your own brain has run every second of your life — and then you point it at things. Reality doesn't get clearer by adding more information. It gets clearer by finding the compression. That is what intelligence is for, and it's what you're for.
You just hadn't seen the move yet.