If you read the description, they started with 35,000x, then tripled the number of cores to get 68,000x. If your triple the cores and wind up with less than twice the performance, your scaling isn't very good.
Try to do it yourself (say take a merge sort implementation and try to make it scale), it'll be roughly this much after a few hours of work.
Now imagine the compiler is doing this for you.
You can have your python parts source compatible with the new language.
Of course they are far off from that.
But not that far off if your python code is typical ML framework consumer code.