Dwarkesh Patel: uncertainty

Transcript context

…Yeah. If you are trying to self learn and there is a resource that is a close approximation of the syllabus you want. Should you just think “Hey, I don't know why I need this chapter. I'm just going to go through it.” or should you use your own judgment there? This is a pretty classic issue for learning in general. You have this problem where to bootstrap yourself in a domain you have to outsource the question of what is necessary to know. You might know, for instance, that you really want to build a model that can generate images given descriptions, like Midjourney, but you don't even know what you need to study to do that. So you pick up some textbooks on machine learning. You're outsourcing the answer to this question to the author. What is necessary to know to build things? Maybe you can find a book that's actually labeled “What you need to know to make an image generating model” But even then, you're outsourcing the answer to the author. You can take that answer as a start and treat it as tentative and revise it iteratively. And as you become more skilled you can lean less on it. And you probably should. I think a very common mistake that people make is to feel that they need to do the thing the right way and that is exhaustive and completionist. If they fail because they find themselves bored or unmotivated because the material doesn't actually seem to relate to what they want to know, but they're just going on faith that, “Well, if I follow what the author says, everything will be good.” Anyway, they find themselves having trouble for that reason, and then they just stop. This is bad. They would be better off just skipping around according to their interest and continuing. One other thing I'll say about this is that the role that these syllabi play is as a scaffold. This is a term of art from learning science, but it relates to the thing we're familiar with. If you want to get higher up a building, you may not be able to climb it yourself, but you can build some scaffolding around it and then suddenly you can reach that top shelf or the top of that building. The scaffolding is ubiquitous in education. We give you simpler versions of questions first, that's a kind of scaffolding. We partially work the answer first, that's a kind of scaffolding. We give you worked examples first, where we might ask you to predict the next step of the work example. That's also a kind of scaffolding. Where the metaphor breaks down is that once you become more capable, we try to remove the scaffolding. It's called fading. The idea is that once you have solved a lot of calculus problems, you don't need half of it worked out and you're just filling in one of the blanks anymore. And in fact, doing that would not be as effective a learning experience. If I'm studying something in computer science, which is a domain that I know really well, I don't need those syllabi, not in the same way for most subjects, and I think that's mostly just because the amount of cognitive demand that's placed on me by the subject is just much lower than it is for other subjects. So much of it is familiar already that I can deploy my own planning more effectively as I go.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.