Very interesting. I would guess that to learn in the presence of spoilers, you’d need not only a good model of how you think, but also a way of updating the way you think according to the model’s recommendations. And I’d guess this is easiest in domains where your object-level thinking is deliberate rather than intuitive, which would explain why the flashcard task would be hardest for you.
When I read about a new math concept, I eventually get the sense that my understanding of it is “fake”, and I get “real” understanding by playing with the concept and getting surprised by its behavior. I assumed the surprise was essential for real understanding, but maybe it’s sufficient to track which thoughts are “real” vs. “fake” and replace the latter with the former.
Yeah, essentially what I’m arguing (poorly) in the OP is that “surprise is necessary for real understanding” is … well, not exactly wrong, but creating some cargo-cult behaviors which are not strictly necessary.
Taking the RL vs imitation idea from the OP, it’s like people have a concept of RL as an important type of learning, but it’s making them think that they have to stub their toe at some point, otherwise they’ll never learn to avoid furniture properly. You don’t necessarily have to personally stub your toe. Trying problems is a pretty good way to test how “fake” your understanding is if you’re unsure. But there are other ways to notice this, and other remedies. It seems to me that people put trial-and-error learning up on a bit of a pedestal when thinking about things like this.
Very interesting. I would guess that to learn in the presence of spoilers, you’d need not only a good model of how you think, but also a way of updating the way you think according to the model’s recommendations. And I’d guess this is easiest in domains where your object-level thinking is deliberate rather than intuitive, which would explain why the flashcard task would be hardest for you.
When I read about a new math concept, I eventually get the sense that my understanding of it is “fake”, and I get “real” understanding by playing with the concept and getting surprised by its behavior. I assumed the surprise was essential for real understanding, but maybe it’s sufficient to track which thoughts are “real” vs. “fake” and replace the latter with the former.
Yeah, essentially what I’m arguing (poorly) in the OP is that “surprise is necessary for real understanding” is … well, not exactly wrong, but creating some cargo-cult behaviors which are not strictly necessary.
Taking the RL vs imitation idea from the OP, it’s like people have a concept of RL as an important type of learning, but it’s making them think that they have to stub their toe at some point, otherwise they’ll never learn to avoid furniture properly. You don’t necessarily have to personally stub your toe. Trying problems is a pretty good way to test how “fake” your understanding is if you’re unsure. But there are other ways to notice this, and other remedies. It seems to me that people put trial-and-error learning up on a bit of a pedestal when thinking about things like this.