VLMs Replicate PicBreeder Research
burny_tech · x · 2026-07-11
This thread centers on SakanaAI's replication of PicBreeder, asking the core question: Can VLMs exhibit human-like creativity and open-ended exploration?
Background
- PicBreeder is a classic experiment by Kenneth Stanley et al.: users don't set specific goals but continuously select and evolve images based purely on whether they are "interesting."
- This method embodies open-endedness—not setting predefined goals but allowing novel results to emerge naturally through long-term exploration.
Approach of this work
- Published as a collaboration between MIT and NYU, titled In Search of the Ingredients of Open-Endedness: Replicating Picbreeder with Large Vision-Language Models.
- Uses VLM agents to explore a shared archive, select images, drive evolution, publish work, and evaluate other agents' creations.
- Key point: No target images are provided, and no criteria for "progress" are defined.
Conclusions
- The results expose both the possibilities and limitations of achieving open-ended discovery with AI.
- The discussion focuses not just on "generating good images," but on whether AI can progressively approximate novel discoveries through constraints and contingencies without clear goals, much like humans.
Related event: Sakana AI Replicates Picbreeder with VLM Agents for Open-Ended Creativity(14 posts)→
More from Research
- Stanford Team Introduces Gigatoken, the World's Fastest Tokenizer — StanfordAILab · 2026-07-22
- Tabul AI launches Metal TreeSHAP to speed up Shapley values on Apple silicon — Scobleizer · 2026-07-22
- Reddit points to OpenAI’s ChatGPT Ads page — EcstaticAsparagus509 · 2026-07-22
- Open-source runtime lets each repo define its own AI code reviewer — ibabufrik · 2026-07-22
- DeepSWE: A New Benchmark for Evaluating AI Coding Agents on Real GitHub Issues — pmz · 2026-07-22
- A Rust space-economy sim runs hundreds of autonomous ships, built with Claude — kalcode · 2026-07-22