* (1,2): Creating good cards should be as easy as a Google search.
I made a prototype with Browserflow that I've fallen in love with: davidbieber.com/snippets/2021-, and we can push this even further with ML
Conversation
* (3) The UX of SR systems should lead users to drop bad cards early and often; let's make keeping bad cards hard
2
5
* (4,5): SR systems should encourage studying when you're most motivated; they shouldn't rely on you studying even on a vacation, or discourage extra studying when you complete your day's review. This requires changes to both algorithm and UX.
4
1
4
* (6) An extra app like Anki is extra friction; brings SR to your email inbox, but you still need to click through to review. Great progress, but we can do even better.
1
3
I want to share these ideas with folks thinking about spaced repetition. So, here's some unsolicited @-ing; I hope you don't mind:
@hermannebbinghaus
3
9
Broadly agreed! Though alas, I believe #1 and #2 are deep research projects: my experiments with cloze-driven extracts like the ones you describe have been pretty unsuccessful in the long-term.
4
1
I've been happy with one-press cloze for several months now. I wonder what was different in our experiences.
One possibility: cards live directly in roam and so are always in an editable mode, so clean-up friction - like adding missing context - isn't too high.
1
1
Another: the cards you create require more synthesis / a higher level of abstraction that cloze extraction allows for
1
1
Yeah, that's possible. I don't think the long-term problems became particularly glaring until 6-12 months, so you may not be feeling it yet.
The central problem for me with verbatim clozes is that I find myself not caring about the wall of text I'm being confronted with.
1
3
The tension is: use a short source range—which later feels contextless and less emotionally meaningful; or include more context, yielding a burdensome prompt not very focused on what I actually want to reinforce.
1
2
The verbatim phrasing is very rarely what I actually want to reinforce. Sometimes it works out! But the majority of the time, not.
I find verbatim clozes have much worse issues than hand-authored clozes, but problems with the latter too:
Silver lining: informal experiments suggest that verbatim clozes are much better as an input model for NLP systems generating focused Q/A style prompts, vs. just highlighting some text.
1
1
Also separately: visual clozes seem pretty reliably good! A lot more to anchor on, I think.
1
2
Show replies
Is it missing context that causes the mismatch between the verbatim and what you want to reinforce? E.g. something that could be fixed by NLP
1
1
Or is it deeper in that the thing you want to reinforce is usually not directly put to paper by the author? E.g. you want to reinforce an interaction with your own prior knowledge
1
Show replies