Abstract claim: The act of creation is fractal exploration–exploitation under optimal feedback control. When resolution increases the portion of parameter space that doesn't make the artifact worse (acceptance volume) collapses. Verification latency and rate–distortion combine into a precision tax that scales superlinearly with perceived quality.
When I make something good, I often spend most of my time making thousands of high-precision edits on an artifact that I thought should have been finished hours ago. Previously, I called this 'last-mile edits', but that was the wrong mental image.
"Last mile" implies executing a known plan with diminishing returns but "last mile" at one level just becomes "early exploration" at higher resolution at the next level. Instead of treating exploration (idea) and exploitation (execution) as temporally separated phases, they nest recursively. That nested search is where the effort goes.
Once you commit to D minor, this scene, that argument structure you've constrained the search space and now you search again within it.
Take some of my quicker five-minute, gestural sketches below. You'd think they break this nested search dynamic but with a closer look it becomes clear that I just front-loaded my taxes by caching motor heuristics.
Domains and modalities differ in how wide and forgiving their basins are and how quickly you can verify the edit (feedback latency). Music timing has a narrow basin at the micro-level (±20 ms can kill a groove) but can be more forgiving higher up: key and pitch changes can be interchangeable without loss of quality, not often though. Prose has a wide basin (many phrasings work). Abstract, contemporary art has extremely wide basin, so much so that nobody with any self-respect even bothers anymore. Renaissance paintings have more constraints and less distortion tolerance.
| Modality | Basin | Verifier | Speed |
|---|---|---|---|
| Text (prose) | Wide | Human read | Minutes |
| Code | Wide (design) / Narrow (syntax) | Compiler/tests | ms-seconds |
| Music timing | Narrow | Ear–body | ~20-40ms |
| Line drawing | Narrow | Eye–hand | ~100ms |
Let's take the following optimization landscape and assume it's for the process of writing a song. To make it simpler, let's constrain like this: We've written the lyrics and picked a BPM of 80.
The wider, more forgiving hill corresponds to choosing C major on the macro level, but there might be a higher, sharper peak in E minor that's trickier—i.e., it demands more precision edits.
Wide basins let coarse proposals land. This is where almost all generative AI outputs live and the oxygen is still plenty. Near a sharp peak, the acceptance volumea a We call it acceptance volume, but in the 3D example above, it’d be the area of a tiny slice of the hill. shrinks rapidly and you can’t reliably see micro-improvements without averaging more evidence or trials. The controller (often the hand) makes many tiny corrections after some latency. Rinse and repeat until the piece sits on a hard-to-vary peak.
That's why effort seems like it scales superlinearly as perceived quality rises. Judging the intermediate artifact takes more time and most edits (the search) make it worse. Geometrically bad edits become more likely and land you lower in the landscape.
Craft, then, is the slog of closing ever-less-perceivable gaps.
Don't bands sometimes record a banger song in an hour together?
The tower-climbing happened during practice (*muscle memory*), not recording. Jazz is closer to real-time exploration and mistakes are more accepted and expected.
Drawing takes forever because you're exploring AND refining simultaneously.
We don't "rehearse" a specific drawing, we solve a novel problem in real-time. There's no cached motor sequence to execute.
@misc{strasser2025,
author = {Strasser, Markus},
title = {Why Effort Scales Superlinearly with the Perceived Quality of Creative Work},
year = {2025},
url = {https://markusstrasser.org/},
note = {Accessed: 2025-11-11}
}