Yeah, but the juiciest tasks are still far from solved. The amount of tasks where people are willing to accept low accuracy answers is not that high. It is maybe true for some text processing pipelines, but all the user facing use cases require good performance.