See the first six slides here: https://media.neurips.cc/Conferences/NIPS2018/Slides/Olukotu...

This is not my phrasing but as someone that's deployed several models in production that have replaced existing conventionally written and maintained areas of code I believe it.

The significance is the ability for machine learning to displace traditional software development is small but growing and there's no real practical limit.

Thanks, I've seen Karpathy's core argument before and to be honest, I don't buy it. Throwing everything at neural networks and forgetting all domain understanding (sometimes developed over centuries) just doesn't feel right. They can be good at certain use cases, certainly, but the eating argument is going too far and just seems like lazy justification of his own research area.