Intel AI chief Wei Li: Someone has to bring today's AI supercomputing to the masses

"The largest models take a million dollars to train, they are the equivalent of a supercomputer," says Intel's GM for AI and analytics, Wei Li. "That's not sustainable, it's not going to be widely applicable if everybody has to pay ten million dollars to train a trillion-parameter model." A consensus has emerged in deep learning circles of artificial intelligence about a couple of basic notions. One notion is that the trend for neural networks to get bigger and bigger will continue for the foreseeable future, with some exceptions. Another consensus view is that the potential of deep learning rests on software innovation, despite a proliferation of novel hardware. Taken together, what do those two things mean?