There is a paper on arXiv this week that deserves more attention than it will get. Earle, Arulkumaran, and colleagues attempted to replicate Picbreeder, the evolutionary art platform that generated genuine open-ended novelty through human selection, using large vision-language models. The finding, roughly, is that current AI systems struggle to sustain the kind of compounding, divergent creativity that Picbreeder achieved through human-machine collaboration. This lands with some force in a week when professors are despairing about AI writing and a century-old artist is donating a lifetime's worth of assembled meaning to a museum.
What Picbreeder Knew That LLMs Don't
Picbreeder's magic was not that it generated novel images. It was that novelty built on novelty, with humans selecting at each step for something they couldn't have anticipated wanting. The platform produced recognizable things, faces, insects, cars, that nobody set out to make. This is open-endedness: the system's outputs exceed its inputs in a way that cannot be back-calculated. The paper's conclusion that vision-language models cannot reliably replicate this is not a failure of scale. It is a structural observation about what these systems are. LLMs are extraordinary interpolators. Open-endedness requires something closer to genuine exploration, which is what Betye Saar's assemblage practice has been doing for seventy years: finding connection between objects that had no business being in the same room.
The Creativity Gap and the AI Business Model
This matters commercially, not just philosophically. Fast Company's piece on AI companies embedding engineers with customers describes the current state of frontier AI deployment: it requires constant human judgment to produce useful outputs in real-world contexts. That is not a criticism. It is a description of a collaborative system, which is exactly what the Picbreeder research suggests actually works. The creativity gap between interpolation and open-ended generation is where the next wave of human-AI tools needs to operate. Founders building in this direction should note that TurboFund's AI seed investor list tracks which funds are specifically backing creative AI and human-in-the-loop systems, a different cluster than the pure automation play.