I get collaborative work. But pair programming for me is a net negative: two people working together outputting less than one person working alone. I don’t even consider it beneficial in the longer term (training, as you put it) given the degraded performance.
The very act of programming is buiding a house of cards in your mind and turning it into code before (or while) it collapses due to our limited brain capacity. Keeping two brains synchronized in the process just seems… too much overhead.
Like the OP, I’m just not interested in finding out if I’m wrong or not. I don’t even want to give it the benefit of the doubt lest it takes hold despite being a bad idea, like many other bad ideas that we now have to live with.
Pair programming works for me in short bursts: when I'm stuck at some problem I would call a colleague, share my screen, and then live-code with me as the driver and him watching me and producing ideas. I find it very effective.
And all other purposes (like training juniors) is not for immediate benefit.