Well, just had Google Bard summarize this for me, and it doesn’t sound terrible. I like the tiered organization and limits it imposes (mostly). I don’t mind the reporting requirements for training data. But if they’re going to put requirements on the “high risk” category of AI to be transparent and explainable then those kinds of systems just might not exist for a good long while if they’re going to incorporate neural networks. Unless explainability and transparency mean that you can explain how they’re trained and the structure of the network. Otherwise the model weights will never really be explainable in a way that matters.