Great link, the one part of LeCun's argument I don't understand (or maybe I disagree with it) is that in Tweet 9 he minimizes the effect of architectural bias mentioned in Tweet 6 by saying that modern DL models have "generic architectures". But his own Tweet (#11) apparently demonstrates a problem with that reasoning; modern DL architectures are bad at generalizing from rare categories, which would seem like an architectural issue that causes significant bias, especially in the sense of racial bias. I think other people may have touched on something similar by pointing that bias might be inherent in the metrics researchers optimize for when designing new models.