It's striking how rapidly I can generate a large number of spaced repetition prompts using clozes when writing prose notes like notes.andymatuschak.org/The_cognitive_.
I notice that I have to write carefully (eg to avoid "giving away" deletions), and the prompts are lower-quality. But fast!
Conversation
I realized that even if I write these kinds of prompts carefully enough that they're effortful to review, they often end up containing so much specific contextual information that I would never spontaneously 'bump into' the ideas outside a review session.
1
2
Yeah, for sure. I'm pretty excited about this direction:
2
3
Have you personally experimented with gpt3 prompt generation? If not it might be worth putting a few days into it. You might be surprised what's possible and there's more than one creative way to use gpt3 as a tool. Quick to try and big if successful.
1
2
Yeah, I have. I didn't have much luck at all, but has gotten substantially better results out of a more specific question-generation model:
1
2
Can you share how you structured your gpt3 prompts? There are a few ways to do it and it may matter a lot.
1
2
Excerpt from my (scant, alas) notes:
1
This can work a lot better by seperating out the task of asking questions from the task of answering them. Do you have more pairing examples anywhere I can easily copy and paste? I can use those to write up an example for you of what I mean.
1
Sorry, no.
1
Mmmm. Okay. My intuition says I can make it work in like a hour or two including the dataset generation... I'll try to whip something up for you in the next week or two!
1
2