2024-04-05
- I continue to think about how LLMs can be the foundation for the next-generation of Anki-like tools for learning. I have been running some loose experiments in this area and wanted to document the findings. The current set of experiments is around using LLMs to generate compelling cards, to remove some of the toil associated with Anki usage. The results so far have been promising, but not earth-shatteringly good. I have found LLMs to be effective at generating solid term/definition cards for key terms from highlights of source material. And had middling success with generating “combination/relational” cards (cards that ask about how concepts relate to each other). Overall, I’m optimistic about the potential for LLMs to generate “knowledge webs/graphs” and associated flashcards. There are still very large questions, of course. The biggest one, that I haven’t even touched bc it feels so huge, is around the interface. Obviously, we are coming at this from a background of using Anki to
review flashcards. And Anki does do a solid job in this respect, by being so narrow/focused in its review interface (other parts of the Anki interface are more frustrating ofc). I could/probably should start out by simply recreating the Anki review interface. But that intimidates me and I’ve been avoiding it.
Date
April 5, 2024