RSI almost certainly sees diminishing returns in its effectiveness. If you model the feedback loop as feeding into various AI aspects that have power-law scaling, you get an initial bump and then less and less over time (yes, even with the feedback loop). Which means others can catch up.
(And the RSI methods will leak to others via all sorts of channels.)
I think the second answer applies to your first point. I don't really get this: there are multiple loops to RSI, this is well-established. You can believe in RSI in one loop (pre-training) and think that knowledge transfer from that isn't very good when it comes to say, the hardware loop. I can suggest better and better pre-training architecture, but it's hard for me to design a brand new chip. Why is this not believing in RSI?
