What is the current state of using a model(s) to generate coding challenges for another model or itself?
Can a coding agent generate challenges it can barely not solve? Of course it can generate problems it cannot solve by simply asking itself to prove P != NP, but can it generate problems just out of reach that could potentially be bootstrapped to?
This seems like a much more scalable loop than trying to get engineers to design manual challenges.
It’s almost like getting an agent to identify the bounds of its abilities via self reflection, and then pushing those bounds via standard methods (RL, SFT, etc). Perhaps even doing weight editing of itself, if possible.
It reminded me of the now old school GAN methods. Perhaps we’ll see the spirit of GANs make a comeback.