Phillip Isola

ML researcher Phillip Isola has recently spoken about a Platonic Representation hypothesis, which argues that neural networks trained on different tasks (e.g. classification, colorization), modalities (e.g. vision, language), and architectures (e.g. ResNets, Transformers) still converge toward similar internal representations.

Peter Bauman, Hito Steyerl on How It Uses Us

With ideas like Phillip Isola and colleagues’ Platonic representation hypothesis, there's always this question—and I think there are ways of getting at the question—of how much of the representation is truly, philosophically, almost pure, and how much of it is just the way that the data works:

Ryan Murdock, Ryan Murdock on Hacking AI