Mapping Aesthetic DNA
Lately I’ve been reflecting on how inspiration moves through the world - not just visually, but through deeper historical and cultural threads. Why does brutalist architecture evoke a similar stripped-down and industrial feeling as certain techno tracks? Or surrealist art as avant-garde fashion?This past spring, Michelle and I built a prototype called Prism at the Exa/Anthropic hackathon (where we placed second!) to explore these cross-domain connections. We experimented with surfacing "aesthetic DNA" to reveal serendipitous links across art, fashion, design, and research.
Beyond Visual Similarity
CLIP embeddings are powerful for first-order visual and semantic clustering, but they often miss the deeper conceptual bridges that connect ideas across domains. By augmenting CLIP embeddings with LLM-generated world knowledge, we were able to reveal deeper similarities rooted in shared cultural, aesthetic, and scientific lineages.
This allowed us to traverse embedding spaces in new ways, finding connections like: Pollock paintings ↔ Tokyo streetwear ↔ Vaporwave.
The Evolution of Taste
As we enter a new paradigm - one where AI isn’t just pushing more content at us, it’s pulling threads together - and models coalesce our inputs (text, images, queries, behavior) into one fluid context, it’s worth asking: how might taste evolve as a system?
What if taste isn’t just about personal preference, but a reflection of deeper cultural patterns we haven’t yet named? How might recommendation systems help people explore meaning rather than just more content?