multi-turn inc.

index

Lies of Prediction

/

On the lie that the predictable is the good.

I've been thinking about what art is for a long time. Lately, I think of it this way: making something that didn't exist in the world before. It's similar to finding a new prime number. A prime exists before it's discovered, but nobody knows it. Once found, it feels like it was always there.


Why isn't AI funny?

The answer lies in the structure of a joke. The moment laughter erupts is always the same — the story unfolds, then arrives somewhere you didn't expect. Humor research calls this incongruity resolution. While listening, we unconsciously predict what comes next, and we laugh when that prediction is wrong.1

The same joke is less funny the second time. By the third, barely funny at all. By the fourth, it's annoying. The structure of the joke hasn't changed — only the effect has disappeared. What changed is on our side. Prediction became possible.

The reason AI-generated jokes aren't funny is the same. AI outputs the most plausible next token from its training data. "Most plausible" means "most predictable."


Neuroscience has been telling a similar story. According to predictive coding theory, most of what the brain does is predict the next moment, and what actually gets processed is only the error between prediction and input.2 Karl Friston's free energy principle pushes this further, arguing that all brain activity is a process of minimizing prediction error.3

When prediction is correct, nothing happens. Something occurs only when prediction fails. Sensation, emotion, attention — all of it emerges from the processing of prediction error. To "feel" something is a signal that prediction has failed.

There's one thing to note here. Humans are prediction machines too. If the brain is a system that minimizes prediction error, then humans, like AI, are beings that perform prediction. So why can humans make art while AI struggles?

The difference lies in the direction. AI's learning objective is the minimization of prediction error. It adjusts weights to reduce the loss function, and as a result, produces outputs that best approximate the distribution of training data. The human brain also minimizes prediction error, but the artist works in the opposite direction. They make choices that deliberately break their own predictions, then check whether those choices work within a structure. The artist uses themselves as a proxy for the audience — as a human who shares the same predictive model, they can judge that what they themselves didn't predict, other humans won't predict either. AI lacks this internal feedback loop. AI is not surprised by its own output.


There's research on why fractals feel beautiful.

Richard Taylor, Branka Spehar, and others experimentally showed that human aesthetic preference is highest for patterns with mid-range fractal dimensions (D ≈ 1.3–1.5).4 There's also analysis showing that Jackson Pollock's drip paintings fall within this range.5 Random noise (D ≈ 2.0) scores low in preference, and so do simple Euclidean shapes (D ≈ 1.0).

There are two ways to interpret this result. One is the processing fluency hypothesis. Because the human visual system is optimized for natural scenes, and the fractal dimension of natural scenes is roughly 1.3–1.5, patterns in this range are easy to process and therefore preferred.6 From this perspective, beauty is the success of prediction.

The other is the prediction error perspective. Mid-range fractals have a global structure you can grasp, but local details you cannot predict. They're processable as a whole, yet continuously yield new information in their parts. Structured enough for the brain to judge "I can understand this," yet complex enough that it never concludes "I've figured it all out."

I don't think these two interpretations contradict each other. Processing fluency of the global structure and unpredictability of the details can coexist. In fact, both are necessary — unpredictability without structure is noise; structure without unpredictability is a grid pattern. The condition for beauty is "globally processable but locally unpredictable."

The same pattern is confirmed in music. In a 2019 study by Cheung et al., listener pleasure was highest when an unexpected chord appeared in a context of high harmonic uncertainty.7 The crucial condition is "a context of high uncertainty." It's not that anything unexpected is good — prediction must break while structure is maintained. This is how jazz improvisation works.


There's a project called A Call for New Aesthetics. It asks what the aesthetics of the 21st century should be, the way Bauhaus defined 20th-century aesthetics. It includes this line:

"Circa 2026, beauty can be found in strange and unusual places. It can violate our expectations in unreasonable ways."

And about AI, it says:

"If jazz didn't exist, could you prompt Suno to create it?"

I don't think it could. Let me address the counterarguments first.

First counterargument: AI can create novel combinations not found in the training data. True. Large language models don't simply copy training data — they learn distributions. Within those distributions, they can generate combinations that don't directly appear in the training data. But this is interpolation within the distribution, not extrapolation beyond it. It's placing new points inside the convex hull of the training data, not placing points outside it.

Second counterargument: raising the temperature produces unpredictable output. Also true. But temperature flattens the probability distribution — it doesn't create structural unpredictability. In fractal dimension terms, it's like raising D from 1.0 to 2.0. Random noise is not a fractal.

AI's fundamental limitation lies in its learning objective itself. AI is trained to approximate the distribution of its training data well. The better it's trained, the closer its output converges to the center of that distribution. This is working as designed. The problem is that the center of the distribution is a statistical summary of things that already exist. To generate jazz from training data where jazz doesn't exist, you'd have to go outside the distribution — and that's the exact opposite direction of the learning objective.


Lies of P. The lies of prediction.

The lie that the predictable is the good. The lie that the highest-probability next token is the best next token.

We are made of prediction, so we only feel something when prediction breaks. Beauty and humor stand on the same condition — globally processable but locally unpredictable.

The statistical average of everything that already exists is not a new prime number.


Footnotes

  1. Suls, J. M. (1972). A two-stage model for the appreciation of jokes and cartoons. In J. H. Goldstein & P. E. McGhee (Eds.), The Psychology of Humor. Academic Press.

  2. Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex. Nature Neuroscience, 2(1), 79–87.

  3. Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138.

  4. Spehar, B., Clifford, C. W. G., Newell, B. R., & Taylor, R. P. (2003). Universal aesthetic of fractals. Computers & Graphics, 27(5), 813–820.

  5. Taylor, R. P., Micolich, A. P., & Jonas, D. (1999). Fractal analysis of Pollock's drip paintings. Nature, 399, 422.

  6. Reber, R., Schwarz, N., & Winkielman, P. (2004). Processing fluency and aesthetic pleasure: Is beauty in the perceiver's processing experience? Personality and Social Psychology Review, 8(4), 364–382.

  7. Cheung, V. K. M., Harrison, P. M. C., Meyer, L., Pearce, M. T., Haynes, J.-D., & Koelsch, S. (2019). Uncertainty and surprise jointly predict musical pleasure and amygdala, hippocampus, and auditory cortex activity. Current Biology, 29(23), 4084–4092.

Lies of Prediction | Multi-turn Inc.