The article touches a bit on this in terms of "proto-personas" and suggests an alternative method (assumption maps). To me, there's a point that you can build products from assumptions. You'd be trading risk (less validation) for speed and perhaps focus, which isn't inherently wrong. But most importantly you shouldn't pretend it's unbiased, user-driven, or that there aren't blindspots. And you shouldn't use personas as the tool of choice in that case. The distinction is rarely made though and problematic.