“Andrei Karpathy predicts that in 10 years, the fundamental AI training paradigm will still involve giant neural networks trained with forward/backward passes and gradient descent.”