youtube.com web signal

Runway details LanceDB-backed multimodal curation at Reverie

TL;DR

  • Runway's Ethan Rosenthal argues quality is an 'orthogonal improvement axis to quantity,' meaning better training data does not require less data.
  • The team's curation stack combines multimodal embeddings, semantic deduplication and large-scale similarity search, with LanceDB as the storage and query engine.
  • Rosenthal traces multimodal data through preparation, deduplication, decontamination, pretraining and low-latency search in a 25-minute Reverie 2026 talk.

Ethan Rosenthal, a principal research engineer at Runway, used a 25-minute slot at Reverie 2026 to argue that generative-video labs should treat data curation as an engineering problem on par with scaling compute. His framing, laid out on the LanceDB session page, is that "better training data doesn't have to mean less data" and that quality is an "orthogonal improvement axis to quantity."

The stack he described leans on multimodal embeddings, semantic deduplication and large-scale similarity search, with LanceDB serving as both the storage layer and the query engine. Rosenthal walks the audience through the pipeline step by step: preparation, deduplication, decontamination, pretraining and low-latency search, together with the production engineering problems the team hit along the way.

Reverie is a one-day summit hosted by LanceDB on November 5, 2026 at the Contemporary Jewish Museum in San Francisco, pitched around the data systems behind world models, generative video, physical AI and multimodal search. Two of the researchers we track shared the recording on YouTube in the days after it went up.

Shared on Bluesky by 2 AI experts