Above should not be taken out of context as hubris. There are brute force improvements that cannot be denied. The paradigm called deep learning – uncountable data + many layers of uncountable simple computations – is moving us monotonically closer to AI. Domain specific short-cuts (e.g. human programmed computers) will always have some advantage, but this margin will reduce.
The presentation describes numerous experiments that are showing the efficacy of artificial neural network based inference. These include significant improvements in speech recognition, image labelling, translation and some success in more challenging areas such as analogies.