Hacker News new | past | comments | ask | show | jobs | submit
> a deeply introspective model that knows to take a step back and call in for aid

This would mean training/RL a model to know it's getting nowhere and call for help.

But afaik, the only way a model would be able to do that would be review of its own historical output.

And if that's what's being reviewed anyway, doesn't it make more sense to just layer a separate specialized "anti-loop/stuck" model on top to judge the output trends?