Shofo

Shofo is a data provider that supplies proprietary multimodal training data to AI labs. The company, a San Francisco startup from Y Combinator's W26 batch, says it holds petabytes and millions of hours of footage, organized into datasets that are used to train the next generation of multimodal models. The catalog covers five categories. Computer use data records expert work in tools such as CAD and Photoshop, with full screen, keyboard and mouse input logs. Talking head footage comes with transcripts and can be filtered by attributes such as objects or hands covering faces. Speech and audio data includes longform speech, podcasts, interviews and conversations, with single speaker and dual speaker material and full transcripts. Video game data captures expert gameplay with input logs. Expert work data shows physical labor such as electrician, surgical and carpentry work in close up video. Buyers can browse a self serve platform called RecipeBook, which offers filters for quality, duration, frame rate, aspect ratio and clip type, along with options to find clips with summaries or transcripts and to order selected clips. Larger requests go through the sales team, and a sample can be requested by contact. This is a service for research teams and companies building models, not a consumer tool. The vendor does not publish prices for its datasets or for the self serve platform.