Evidence receipt / prediction
Published · transcript-backedAndy Matuschak: prediction
12 Jul 2023 Dwarkesh Podcast Andy Matuschak — The reason most learning tools fail
“You're reading through the newspaper and you see that people have finally figured out how to do X and you're like, “Wait a minute, that means if I combine it with this other thing, like we'd be able to do Y!”
Source trail
Everything needed to verify it.
- Speaker
- Andy Matuschak
- Attribution
- Verified speaker
- Claim type
- prediction
- Recorded
- 12 Jul 2023
- Publisher
- Dwarkesh Podcast
Transcript context
…Yeah, right, right. On a separate question, do LLMs make memorization more or less valuable? There's a case you can make for both. But on net, is it more important to have more Anki cards in your deck now that GPT-4 is out? Maybe this is a good time to talk about what memorization is or what it's for. We could use that word to refer to the practice of learning more trivia. For instance, a thing that I and some people I know have done is, we’ve gone through a book called Cell Biology by the Numbers, which says all of these things like, how big exactly is a nucleotide? Like how much volume does it take up? It's kind of helpful occasionally to know that it's about a nanoliter. And that can help you model things. So you can just commit all of those things to memory, right? That's one kind of memorization. And we could talk about how LLMs affect that. But I just want to make the case that so much of what you do and experience day to day is memory bound, or is memory influenced in important ways. For instance, your ability to understand a difficult argument, even in the course of a text, is memory bound. Some of that's working memory. But your ability to understand an argument that has many steps in it, more steps than you can keep in your working memory, depends on your ability to think of some of those steps in terms of some stuff that you already know, so that you can kind of reduce it or abstract it. Likewise in creative work, there's a bunch of studies trying to catalog case studies of how it is that people have flashes of insight. It's a little hard to talk about that but one of the things that's a pretty consistent source of insight for people is noticing a surprising connection or a surprising contradiction. It probably feels pretty familiar, right? You're reading through the newspaper and you see that people have finally figured out how to do X and you're like, “Wait a minute, that means if I combine it with this other thing, like we'd be able to do Y! ” or something like that. Now that's only possible if the other thing is in your memory. If you have to think to look up the other thing, the newspaper wouldn't seem so salient to you. Early on in my time in Khan Academy I learned a whole lot of details about the education market in a very thorough way using memory systems. This let me be in high level executive kinds of conversations where we're trying to figure out strategy stuff and somebody would propose a particular direction and I could say things like, “Well the total budget spending for instructional materials is this and that market is growing by this percent per year and 10% of students in the US are in this place” and so on and so forth. Basically I could evaluate ideas on the fly in a way that others couldn’t. Anyway, this and other things are just part of my rant about how people in general under-appreciate the role that memory has in our lives. So just to come back to the question, explicit memorization or explicit making sure that you can recall the thing reliably. We can test it against these things. the role that memory has in our lives. So just to come back to the question, explicit memorization or explicit making sure that you can recall the thing reliably. We can test it against these things. So for the case of the creative instinct, for instance, noticing the contradiction, noticing the connection, I imagine that we will have future notebooks that will do some of this noticing with us and that will decrease our need to be able to rely on our own sense of salience or something like that. But I guess I don't know how much. My own experience coming up with weird ideas that feel very new is that it feels very personal, it feels very [unclear]. I often haven't been able to describe, textually, the constituents of the thing very clearly. There's just kind of a feeling that something in this general direction is connected with something in that general direction, or there's a tension. That makes me a little hesitant. LLMs depend on our ability to externalize things and to make them legible. Back to the learning point about the role of memory. If what you're trying to do is to understand something pretty difficult, your ability to understand that thing is still absolutely going to be bound on your memory of the constituent material.…
Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.