I disagree. Parallelization, coordination shouldn't be model-level functionality. Further, how do you know if or when we'll get orders of magnitude increases in context sizes. Or, do you mean that the labs will just nail the perfect plug-and-play harness? That's fair, but why wait.