I think that strong claims about the limits of large end-to-end trained models require a formal argument or decisive empirical support, because they go beyond what is scientifically known

→ View original post on X — @plinz,