[Jensson]

> There are other benefits, pair programming reduces a whole category of simple bugs/typos to basically 0

No it doesn't, bugs happens since there is much acceptance for bugs, programmers develops just enough awareness of them to get the bug count to maximum acceptable levels. If you take two programmers who both learned to get there on their own and have them oversee each others work, you reduce the bugs temporarily, until they adjust their attentiveness and conscientiousness downwards to account for it and the bugs rise again until they are at barely acceptable levels.

You see temporary improvements like this happen everywhere. Researchers love it because it makes it trivial to publish papers, but these results usually doesn't scale, and even when they do scale the effect is significantly smaller than originally advertised.

[WindyCityBrew]

(I know you run out of people in real world teams pretty quickly but) rotating pairs regularly helps with this, with a maximum of two-week pairings to avoid this and other issues.