While recursion is definitely the (long) way to go, I’m afraid that in short term auto-post-training will mostly bring new height of reward hacking.

There are no pure objective, have to know what you’re doing, build an understanding: that’s an actual AGI problem.