[ACCount37]

It's closer in principle to training a chess AI by making it play against copies of itself.

Humans and AIs play the same chess, on the same board, under the same rules. Doesn't stop modern AIs from crushing humans at it. In a narrow closed domain, vastly superhuman capability isn't just attainable - it's expected.

There are issues with going to less narrow, more open domains with it. But evolution producing increasingly advanced intelligence is a bit of an existence proof for broad open domain RSI being possible - there is a serious argument that human intelligence has evolved in a "self-play" pressure cooker, via semi-adversarial optimization against other humans. If a continuous improvement path like this exists in non-intelligent space, intelligent design has "copy that" at its floor.

We already know that AIs can come up with Move 37, or use branches of human mathematics against each other in novel ways, or come up with human-unnatural solutions like LLMs solving ARC-AGI-3 puzzles with A* pathfinding and a constraint solver instead of humanlike spatial reasoning. Matter of whether this can be pushed further. And by "whether", I mean "whether it happens in the next few decades", not "whether it's possible".