Comment by ACCount37

Today, a lot of it is just integrated VLAs. End to end backprop covers a multitude of sins - a single integrated model stack in cross-attention beats "an ensemble of specialized models".