He makes a very good point here about llms or lrms not being good at taking video as inputs, and i think he's signaling his intent to help change that. He's pointing at an open problem. This isn't a carefully structured blog post, it's just a tweet.