It works by "replaying its past discovery attempts" while "testing thousands of alternative strategies cheaply", then "deploying the better strategy in the next round" and "it improves the exploration policy, not the underlying model weights." ...so RSI is prompt engineering lol