The Problem Is That Tasks Generate Could Be Not Rounded in the Real World?

The idea that tasks generate could be completely not rounded in the real world. We'll try to find disconfirming evidence now. What am I at the problem? I think it's probably right. That's also one of my questions, which is why people think of the opposite thing. And so you can imagine even things like process level replay, which takes level sample from a distribution and then replays the ones with high learning potential. But we obviously all this work on curricula, but you could do curricula over the sampled worlds for an offline algorithm.