Approach L: You can treat learning as a “lookup table” with lots of chunking and abstraction layers. You can store lots of knowledge, solution-patterns, concepts, and abstractions, in a notetaking software or notebook. Instead of memorizing things, you look them up in your own indexes. Manage your knowledge as a multi-level system of memory, where lots of stuff is in cache and not much is in storage or RAM. Write more, think about less at once.
Weakly associated with: treating knowledge as a less-structured graph, fast iteration, feedback loops, practical applications, pattern-matching, IDEs, StackOverflow, TVTropes, and big lists.
Prediction L: In principle, someone with low ability to memorize things, but with a high ability to chunk and take/index notes, could produce work as high-quality as somebody with high short-term/working memory, given more time. The protagonist of Memento could eventually create Ramanujan-level work in math, if he spent his whole life doing so and had good-but-still-human-level systems for abstraction and note-taking.
Approach S: Learning is an inherent struggle to carve intuitions into your neuron patterns. This is hard by default for the things we often care about (AI safety research, writing, technical problem-solving, maths, coding). Flow is of utmost importance, meaning anything you use remotely often should already be memorized. Learning only occurs through deliberate practice, and concept absorption must always happen on a subconscious level. Any idea worth thinking about (in an EMH-sense) is going to be huge and tightly-wound, so you need to have lots of working memory to think about it at all. Any abstractions for the ideas are either unhelpful, or must be created for yourself on a subconscious level. Write less, think about more at once.
Weakly associated with: treating knowledge as a hierarchical tree, system-1 “grokking”, subconscious insight, 10,000 hours, Anki, and keeping the whole mental “stack of cards” loaded at once.
Prediction S: Prediction L is wrong (except in a trivial theoretic-bound-slowness form of running an entire brain/brute-force-theorem algorithm by hand). The guy from Memento could not create Ramanujan-level work, with anything less than 50-100x as much working-time as Ramanujan.
The core of the disagreement, as I see it: Is mental intuition (via lots of struggle-ful practice and/or lucky genetics in brain wiring) replaceable with savvy knowledge management and time, given only a modest (<10x) difference in working time?
In this thread:
Theoretical arguments about which approach is better.
Formal models of both approaches (especially
Circumstances when one approach is better than the other.
Algorithms to hedge across / decide between the approaches for a given situation, etc.
Ways this can be empirically tested (from statistics we should google, to personal exercises that can be timed (like DNB)).)? In what scenarios? Can the split be changed? How? By how much?
Lower-level productivity tips building on the ideas collected and described here (knowledge storage as multi-level cache, aggressive abstraction of ideas, Anki card designs that take advantage of these ideas...).
Things to not get sidetracked by:
Practice vs intuition. In the model above, practice is the unlucky person’s way to mold their brain-structure for intuition. The question is how much practice we can circumvent, while still producing the quality output of somebody who has the intuition we lack.
System 1 vs System 2.
What an ideal reasoner or AI would do, given no cognitive constraints. The whole question is how bad certain human cognitive constraints are, in practice, and which ones can be circumvented or must be fought head-on.
Parenting styles.
Nutrition, sleep, exercise, or brain augmentation. This isn’t a brain hardware thread, it’s a brain software thread.
Which knowledge “should” or “shouldn’t” be memorized. An ideal reasoner would have everything it’s learned memorized and grokked and available, after reading it once, all at once. The trick is whether it’s practical to replace memorization with lookups, in not-quite-trivial cases. (E.g. most contact details are trivially “lookups”, some basic rationality ideas only work as “memorized/intuitive”.)
Domain expertise vs generalism. You could do either Approach on either thing. (“Generalism” is kind of its own domain, albeit a high-leverage one.)
In my personal experience, the intuition/grokking approach works better for mathematics, but the making-notes approach works better for programming.
This is not a fair comparison, because in case of programming, the notes were usually already made by someone else (on Stack Exchange), so I can use them without spending my time to write them. For deep understanding, I must usually do the work myself, because most people who write programming tutorials actually do not have the deep understanding. (Seems to me that most tutorials are written by enthusiastic beginners.)
I have a strong personal preference for the intuition/grokking approach, so I am not doing the optimal thing despite knowing what it is. But I have to admit that this is often a waste of time in programming. Before you spend the proverbial 10 000 hours learning some framework, it is already obsolete. Or if it becomes popular, sometimes the next version keeps the similarities, but changes how it works under the hood, so ironically the deeper knowledge that should be long-term sometimes becomes obsolete first.
In math, the stuff you learn usually remains true for a long time, so building intuitions pays off.
This was helpful in understanding this problem, thank you