Users praise Sakana AI's reproduction of Picbreeder with non-interactive VLM agents for highlighting key insights on preserving novelty in optimization and sparking interesting experiments.
Based on 5 visible X reactions from 11 accounts; directional sample.
Ask a question below.
Published answers will appear here.
Really interesting discussion! Agree that bootstrapping better-factorized representations is the major promise of this work. We didn't end up having room for it here, but many discussions were had about how we might probe the factorized-ness of various AI/human-generated CPPNs more systematically (and I unsecretly hope someone might scoop the raw genome data we've published to perform exactly these kinds of analyses). (E.g. perturb the weights then project the perturbed images down to a CLIP-type space and see to what extent the image's semantics have changed.) But to test beyond the base case we'd of course need to turn this collaborative sandbox into one for breeding the breeder models themselves! 💸 Interesting design challenge to imagine the minimal version of this. Also very glad to see someone poking around the breeding/viewing tools!
@MLStreetTalk @SakanaAILabs @kenneth0stanley the key insight people miss: when you optimize for a specific goal, you kill the very novelty that could unlock the real breakthrough. sakana's work just proves discovery is a distribution problem, not a gradient ascent one.
@togelius that actually sounds like a super interesting experiment.
@MLStreetTalk @SakanaAILabs @kenneth0stanley Thank you!
The AI agents produced less conceptual diversity than humans.
Users praise Sakana AI's reproduction of Picbreeder with non-interactive VLM agents for highlighting key insights on preserving novelty in optimization and sparking interesting experiments.
Based on 5 visible X reactions from 11 accounts; directional sample.
Ask a question below.
Published answers will appear here.