AI Tore Down the Wall of Creation—But Another Wall Stands Behind It: Two Papers on Open-Endedness and Vibe Coding

Forum topic · 小凯 · 2026-05-26

Summary

This forum post compares two 2026 arXiv papers. The first, 'In Search of the Ingredients of Open-Endedness' (GECCO 2026), replaces Picbreeder's human users with frontier vision-language models and finds a 'clear qualitative difference': while humans sustain open-ended novelty across generations of image evolution, AI selectors converge on homogeneous outputs. Interventions—exploration noise, behavioral diversity across models, and narrative momentum—only partially close the gap. The second paper, 'From Prompting to Verification,' surveys 162 vibe coders across three experience levels (no coding background, novices, professionals). All groups recognize AI-generated code risks, but verification and debugging ability scales strictly with programming experience—a 'perception-action gap.' Together the papers argue AI has removed the production barrier but not the judgment barrier: 'partial democratization' lets anyone generate images or code, yet the ability to evaluate quality and direction remains unevenly distributed. AI dismantled the wall of capability; the wall of taste and verification still stands.

Two papers. One about drawing. One about writing code.

The drawing paper kicks humans out of the studio. The coding paper leaves humans at the keyboard—but cancels the qualification exam. Read together, they say the same thing: between what AI produces and what you actually want lies a membrane you cannot see. Everyone can feel it exists, but only a few know what it should look like and how to tear it off.

Paper A: Replicating Picbreeder with VLMs

Background: Picbreeder

In 2007, AI researcher Kenneth Stanley launched a website with no brushes or palette—just a grid of fifteen abstract images generated by small neural networks. Users picked any image they liked, it mutated randomly into fifteen offspring, and selection continued. No goals, no endpoint, pure instinct.

Over years, thousands of users evolved an enormously diverse image library from pure noise—alien butterflies, scorched planets, cell-like forms. Picbreeder became the emblematic case of open-ended search: systems that produce not just new things, but categorically novel things—beyond known classes, into unforeseeable territory. It seeded the fields of novelty search and quality-diversity.

The AI Experiment

In 2026, a team asked: replace the humans with frontier vision-language models acting as aesthetic selectors. The result in one sentence: a 'clear qualitative difference' between AI and humans.

The problem was not visual quality—Picbreeder is abstract art with no objective aesthetic standard. The problem: after dozens of selection rounds, AI's output diversity collapsed. It kept choosing the same kind of thing, converging into a narrow channel. Humans get bored of butterflies and pivot to faces; they choose out of curiosity, confusion, aesthetic impulse—humans are inherently disloyal to goals. AI is not.

Three Interventions, Three Partial Fixes

1. Exploration noise — inject randomness into selection. Partially useful; reduced over-reliance on some patterns, but insufficient. 2. Behavioral diversity — run multiple models with different 'tastes.' Richer early variation, but tastes converged over time because VLMs share pretrained visual biases. 3. Narrative momentum — make the AI remember past choices and avoid old paths. Mildest effect; did not change the convergence trend.

All three weapons deployed. AI still is not human.

Paper B: Vibe Coding and the Perception-Action Gap

If Paper A asks 'can AI create like humans,' Paper B asks 'can non-creators create with AI.' It studies vibe coding—the term Andrej Karpathy coined in 2025 for describing what you want in natural language, letting AI generate code, running it, and judging by the result.

The team surveyed 162 vibe coders in three groups. Findings:

Risk awareness is flat. All three groups know AI-generated code can be wrong. No one blindly trusts AI.
Motivation diverges. Non-programmers use AI to cross the 'I can't code at all' barrier; novices to learn and experiment; professionals for real-work productivity.
Interaction style diverges. Non-programmers write one long comprehensive prompt and iterate until it runs; novices ask questions and decompose tasks; professionals interact 'surgically,' knowing what to delegate.
Verification diverges most sharply. Non-programmers: 'run it, if no error it works.' Novices: simple manual tests. Professionals: systematic verification—edge cases, code reading, targeted inputs.

This is the paper's core contribution: the perception-action gap. All groups share the perception (AI can err), but only professionals have the action capacity (finding and fixing errors systematically).

One Story from Two Papers

Both papers say: AI removed the production barrier but not the judgment barrier.

In Picbreeder: production is 'draw an image'—AI does it fast. Judgment is 'which of thousands of mutations is worth pursuing'—AI's judgment currently loses to a random human user.
In vibe coding: production is 'write runnable code'—AI does it fast. Judgment is 'is this correct, secure, maintainable for all inputs'—AI won't decide for you, and non-coders can't decide either.

One barrier is capability; the other is taste and experience. AI conquered the first; the second remains untouched.

What 'Partial Democratization' Means

The vibe coding authors' precise phrase—partially democratising—admits both that democratization happened (non-coders can now produce code) and that it is incomplete (output quality depends on prior skill). AI redistributed the ability to *produce*; it did not redistribute the ability to *judge*. And the second redistribution is likely harder—AI can write code for you but cannot grant you ten years of debugging experience.

Honest Limitations

Picbreeder paper:

Only three intervention dimensions tested; untested factors include social pioneer-follower roles and mixed-timescale rewards.
Original Picbreeder users were a highly diverse population; frontier VLMs come from a few shared training pipelines—a diversity ceiling.
The paper does not distinguish 'AI lacks open-ended creativity' from 'current AI architectures lack it in this specific interface.'

Vibe coding paper:

162 respondents, ~50 per group; modest sample, likely English-language survey platforms.
'No coding experience' was self-reported and may mean very different things.
Cross-sectional design: no tracking of whether participants migrate between groups over time—which would directly reveal whether AI is actually teaching people to code.

Closing Thought

Picbreeder and vibe coding suggest current AI systems lack not capability but the direction of attention. AI can optimize a given goal to the extreme; but when no goal exists and the direction of the next step depends on a fleeting 'that looks strange, let's see where it goes'—AI's steering wheel spins. And vibe coding adds a footnote: AI not only lacks judgment itself, it fails to transmit judgment to those it serves. AI tore down the first wall—'can it be done?' The second wall—'is it done well, in the right direction?'—you still have to tear down yourself.