What Multimodal Models Actually Learn From Video
Multimodal models are trained on video, audio, and text — but what does "multimodal" actually mean under the hood? This episode digs into late fusion versus early fusion architectures, why video is the largest slice of Emu3.5's 13-trillion-token diet, and a phenomenon called "vision laziness" that explains why rich visual data can sit unused. Then it gets to the question a listener named Daniel asked: if a model learns from video, can you retrieve that knowledge just by typing at it? The answer is yes — but only for some kinds of knowledge, and the asymmetry is stranger than you'd expect…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.