hex

Part of your success in life will depend on your ability to master new concepts quickly. I think this is especially true if you are the kind of person who doesn't thrive on relationships... You may need to push harder on raw knowledge. Ideally you'd have both, but that's not always possible. The problem is that the way most people approach learning, and the way we're taught in school, is extremely inefficient. The first issue is that most people learn passively. They see a concept again and again without hard practice, and end up mistaking familiarity for deep knowledge. You may have noticed this: someone seems familiar with a concept, but when you dig deeper, their knowledge amounts to almost nothing. To some extent, we fool ourselves when we say "Yes, I know this" Take me and LLMs. When someone thinks I know a lot about them, I point out that it's mostly familiarity, and my actual knowledge is quite limited. This is one of the reasons I moved, a while ago, from just reading the latest papers to implementing things. In this area, I'll consider my knowledge satisfactory when I can build a small LLM from scratch, add reasoning to a pretrained LLM, and run local models on different hardware efficiently. That's my goal. Maybe it's not the optimal one, but it's what I want to master for good over the coming months. Coming back to the problem with how most people learn, something I'm noticing right now, while preparing for an exam, is that retrieval is too organized. This morning I was reviewing material on tool design and MCP, and I thought I was acing everything. Then I realized I was just pattern matching (when I saw X, I fetched X) I'm not sure that's clear, but the point is that when you face a new problem or use knowledge in the real world, questions don't arrive labeled. You have to figure out what to use in a messier, more confusing setting. Evaluation is another issue. How do I know whether I'm actually mastering a concept? There are things you can do to retain more of what you learn, such as spaced repetition. It works, but it runs into the same matching problem. Having a skilled human tutor might also work, but that has its own problem: it's expensive. Are you familiar with Bloom's two-sigma problem? I'm starting to think LLMs could solve it. But learning with LLMs has its issues too. What I notice is that I tend to let the model explain instead of quizzing me, which slides back into passive learning rather than real understanding. The solution might be practice where you type answers to unlabeled, situational questions that the model generates from your own study material. These would be scheduled per idea, then interleaved across easily confused ideas, graded on recognition and execution separately, with worked examples shown only after a real attempt. I think a system built around this would produce a bigger improvement in first-attempt accuracy on new, unlabeled problems than normal study alone.