An order of magnitude speedup for GPT-like models? Not sure why this story isn't bigger news..