[hashta]

I’m not sure AF3’s performance would hold up if it hadn’t been trained on data from AF2 which itself bakes in a lot of inductive bias like equivariance