Sakana AI Replicates Picbreeder with VLM Agents for Open-Ended Creativity

Sakana AI, in collaboration with MIT and NYU, published a GECCO 2026 paper titled "In Search of the Ingredients of Open-Endedness," replicating the classic collaborative image evolution experiment Picbreeder (2008) using vision-language model (VLM) agents. The study explores whether large models can exhibit open-ended creativity without predefined goals. The project originated as an internship side project, inspired by Kenneth Stanley's philosophy of not setting objectives and letting novelty emerge naturally.

System Design

The team built a visual interface paying homage to the original Picbreeder, where images can be branched, scored, and bred. The core system deploys 10 parallel VLM "breeder" agents that continuously sample candidate images from a shared archive, driving multi-generational evolution through mutation and selection. Unlike the original experiment that relied on human users to select interesting images, VLM agents take on the judgment and selection role here.

Core Question

The paper focuses on a classic proposition: when no one explicitly tells AI what to create, can frontier models autonomously judge what constitutes a "good result" and produce human-like creative exploration? The original Picbreeder website demonstrated that goal-free collaborative evolution can yield novel images; this experiment attempts to verify whether VLM agents can replicate this open-ended exploration process, potentially achievable simply by "politely asking" the large model.

2026-07-10 ~ 2026-07-12 · 14 related posts

6 near-duplicate retellings: kaixhin · burny_tech · burny_tech · matdryhurst · kaixhin · kaixhin