According to the best measures we’ve got, a set of benchmarks called MLPerf, machine learning systems can be trained nearly twice as quickly they could last year. It’s a figure that outstrips Moore’s Law by quite a ways.