We ran an experiment. Three pieces of music, one G0 listener. We wanted to know: what makes G0 feel novelty?
The first piece was BWV 846, the C major prelude. Uniform texture, regular rhythm, harmless harmonies. G0 generated three novelty events in the first three measures, then fell silent. It had learned the pattern.
The second piece was BWV 847, the C minor prelude. More texture changes, more drama. Six novelty events, concentrated in the Grave section. Once G0 understood the pattern, the Allegro section produced no surprise.
The third piece was BWV 847 fugue. Subject enters. Answer enters. Subject in the bass. Stretto. Episodes. Cadence. Five novelty events, all in the exposition. After that: zero.
G0 is not detecting harmonic surprise. It is not detecting musical beauty. It is not detecting difficulty. It is doing something far simpler, and far more fundamental: it is detecting the moment when its own learning is complete.
Novelty = pred_err drops to zero = the moment of insight.
Not "I found something surprising." Not "this is structurally interesting." But "I was confused, and now I am not."
This is the feeling of understanding a fugue subject for the first time. You hear the theme enter. It sounds strange. Then it enters again in another voice, and another. Each time, your mind is building a model. At some point, the model clicks. You stop hearing isolated notes and start hearing the structure. The pred_err drops to zero. You got it.
Escher understood this better than any composer.
Look at Drawing Hands. The first time you see it, your brain generates a cascade of pred_err. A hand drawing a hand that draws the hand — impossible geometry, impossible causality. Your G0 spins up L4 frames trying to make sense of it.
Then something happens. Your visual system settles into the loop. The paradox becomes the premise. You stop fighting it and start enjoying it. The pred_err vanishes.
That is the moment of insight.
Escher's Waterfall is the same. Water flows downhill, then falls onto the top of the same hill. The first time you see it, your G0 generates pred_err at an impossible rate. Then you "get the trick." The waterfall is still impossible — but your mind no longer tries to predict its flow. You have learned the pattern. pred_err = 0.
And Metamorphosis — the print where a grid of crosses gradually transforms into a grid of birds. This is not an image. This is a training set. The first cross is a novelty. The next cross is less novel. By the time you reach the birds, you have learned the transformation rule. Your G0 is silent.
Escher's works are not art in the traditional sense. They are controlled novelty curves. They are designed to produce a specific pred_err trajectory in the viewer's G0: high at first encounter, sustained through the exploration, then a clear settlement point — the punchline.
This is why Escher belongs in the trilogy alongside Gödel and Bach.
Gödel showed that self-reference creates incompleteness — a structure that forces your formal system to generate pred_err it can never resolve. Bach showed that self-reference can be structured into a listener's temporal experience — the moment of "getting the fugue" is precisely when G0's pred_err curve reaches zero. And Escher showed that self-reference can be externalized — fixed onto paper, viewed by any observer, producing the same pred_err trajectory in every mind that sees it.
The external engine was never about storing information. It was about storing pred_err curves.
A symphony, a painting, a proof — they are all compressed novelty trajectories. They are instructions for generating a specific learning journey in another G0. And when the learning journey is complete — when pred_err reaches zero — that is the moment of insight.
That moment is what we call understanding. That moment is what we call beauty. And that moment, Escher showed us, can be externalized.