Have you actually developed a large React app? Because you might be surprised at how fast React really is in practice.
This is (more or less) true now. It wasn't when React first showed up. Angular is also only just as fast when you actively code your apps in the same way as React (basically using a prop/state system but in an Angular way). This isn't also the method generally championed in Angular (1.x anyhow).
Ember was always pretty fast but post-React it changed a lot of how it did it's rendering... influenced by React!
So claiming that React apps are no faster is a bit disingenuous as the reason for that is because the other frameworks saw what React was doing and agreed with it.
You can use Angular 1.x in exactly the same way you use React by only ever using 1-way binding and constructing your entire application out of directives scoped to elements. The only thing Angular 1.x doesn't let you do is use JSX, but then again, you could just use plain old javascript and imperatively build up your DOM elements (which is what JSX does under the hood anyways) if you really don't want to have .html files.
Step 1 would be slow if it created it in a real dom but as it's not (that's why it's called a "Virtual DOM") and is in fact just an object representation of the dom it's actually rather fast.
After all - it's just an object tree, and that's quite fast to parse.
The diff part is probably the slowest part in the React chain actually - because it actually has to compare to the other (virtual) representation of the DOM. Deep comparisons are just... well they'll never be trivial.
But no, it's really not extra thinking, it's not error prone, and it's basically due to it not really being what you're describing it as.
Let's say that you have a page containing a billion little images. On every update, your React-enabled code needs to create a virtual DOM containing all those images. This takes time, regardless of whether the virtual DOM is fast. Hence my point.
And even if it may be only a few milliseconds for a hundred images, it will take seconds or even minutes for millions of images. And even if only 1 of the images changed. This is of course not acceptable.
Most programmers are already happy if it works for 100 images. However, imagine that the inventor of quicksort invented bubblesort instead, stating it is fast enough for 100 elements. From a CS point of view, this situation is just ridiculous.
1. Create full new version
2. Run a diff with the old version
3. Patch
Because of 1, it is bound to be slow (in a computational complexity sense). Based on the popularity of React and its wide acceptance, I wonder if people learn computational complexity theory anymore these days?
PS: yes I know there are tricks to make 1 faster, but these are really hacks (require extra thinking and are error-prone) and deviate from the original "React" philosophy (they are basically the old style of doing things), so these don't count.