Brendan Long comments on The surprising parameter efficiency of vision models