Next was an incredible talk by Phillip Isola on why large models are all converging to similar representations at MIT. I absolutely love this work, and while I think the convergence here is on the world as represented by the Internet rather than the given hypothesis, this approach should fundamentally reshape how companies and researchers value training more models and why phenomena transfer across models (see Kolter's talk above, for example). Highly recommend
https://www.youtube.com/watch?v=V7AyriUcXZQ (5/7)
#AI