My impression is that he often tries to 'steelman' the other side by purposely putting forth more simple-minded positions or questions than he actually holds. At least in AI territory it often seemed that way, noticeable when a guest mispeaks or mislabels a theory and Fridman corrects them, suddenly knowing the exact scientific term or implications.
I don't think he's necessarily doing a great job with this approach, but I'm guessing that there's some attempt of method behind it, rather than him being negligently oblivious to the subject at hand.
edit: grammar