The essence of my point was that flat out not estimating isn't an option.
Sometimes it is better to start developing and then tell stakeholders that it will take a really long time when you run into problems. A lot of stakeholders have problems with estimates of the form "it could take 2 days or 3 weeks". It would be nice if the industri starts giving probability about estimates especially when you know that there are two clear possibillties where one is much, much larger. They might want to invest 2 days into trying to solve the problem with the understanding that it might take much more time and if that happens they can change their mind and skip the feature instead, "only" wasting two days of investigation.
I don't remember if it's "Mythical Man Month" or Thomas deMarco where I'd read it. They're better at explaining it than I.