Not true. Take a look at the paper and benchmarks. The point of the thing is that it does well on a number of NLP tasks by being an expensive autocomplete. As people demonstrate in this thread, it still has significant flaws.