Completion fast enough to be faster than a human's reaction time has been feasible for many years now. Furthermore there are completion systems that let you type discontiguous parts of your target word, essentially giving you the best of both worlds.
To truly be fast enough not to break the flow of thought->text, it has to (for me) be ~2 tokens ahead of where the cursor is, not completing the current one. And even then I doubt it'd work completely, because it requires rapid context switching to check the completion and pick the right one.
