There is a pretty intriguing/troubling observation at the end (cf. https://news.ycombinator.com/item?id=14736254 https://news.ycombinator.com/item?id=14736254):

There is a pretty intriguing/troubling observation at the end (cf. https://news.ycombinator.com/item?id=14736254 https://news.ycombinator.com/item?id=14736254): If our brains are doing gradient descent, and we can determine the signals used to indicate the gradient, then in the future (with good enough neural prostheses) we could do backprop through an AI back into the brain!