I read an article some time ago (I don't remember where), about a class for rock climbing where they used exactly the opposite approach. They would start new climbers off trying to climb a relatively simple wall.

After the students had a little time falling and struggling, the instructors would go back and explain the basics to a group who were much more focused on learning not to make the mistakes they had just experienced.

I think we can all agree that random practice without comparison against a standard or correction by a more experienced mentor won't be very effective.

On the other hand, the only way to learn how to attack a problem from scratch is to do just that. It sounds like the OP has a strategy that makes sense for his learning style. It also happens to be one that has served me well in my career.

You are making a distinction between research and production code. I thought the OP made it clear he was talking about the former.