Comment by TheDong
1 day ago
> graded by an LLM
This seems impossible to me. In anki, there's "hard", "good", and "easy" which are all for "I got this right".
For my usage, "hard" is "I got it right, but I was only like 60% sure", "good" is "I had to actively think", and "easy" is "effortlessly correct, no real thought required".
There's no way for an AI to tell if my identical input is the result of a 50/50 guess, or a little thought, or effortless recall. "delay to answer" also isn't a good approximation, I have a habit of alt-tabbing and chatting with a friend on random cards of any difficulty.
I find distinguishing those levels of easy for totally identical answers ends up making SRS more effective, and AI just can't know my inner thoughts. Maybe once we have brain implants.
Yes, this is also something I have been thinking about, can an LLM really know how well I know something. There is the issue with the grading with again, hard, good and easy that I can cut myself some slack and say "I knew that" even when I didn't(and I have a strong memory of having done this myself). And there is the possibility of bullshitting the LLM and just all you know about the subject rather than the exact definition of the flashcard. I'm leaning towards any knowledge rather than specifying that the exact answer should be graded. Whats your take?
Bullshitting the AI maliciously doesn't matter, if you don't want to study effectively, you won't study effectively, and that's not a problem for the app.
> any knowledge rather than specifying that the exact answer should be graded
I don't understand what you mean. The important thing is to feed back into the SRS algorithm "How much does this card need to be studied", and if you mean "any knowledge means we can study it less often", then I doubt the SRS will be able to be effective.
What are you suggesting to feed back into SRS? How will you ensure cards the user knows very well quickly get pushed way back (so the user isn't overwhelmed with a boring slog), and cards they only sorta know bubble up more quickly to start to cement the knowledge?
My understanding of what is important to feedback into the SRS is, was I able to retrieve the memory, and does the representation in my memory align with what I recalled.
As an example Term: "What is the capital of France and how many inhabitants does it have?" Correct definition: "Paris, which has 2 000 000 inhabitants."
For me there is a difference in not having the answer at all, which falls into "again". But what about if I'm able to retrieve that Paris is the capital, but I remember that the population is 1 500 000. This is where the gray zone begins
2 replies →
One way it could grade you automatically is by the speed of flipping the card (or entering the correct answer). If it took less than a second to confirm then evidently it was easy.
But conversely, if I alt-tabbed to chat with a friend, or paused studying because the person sitting next to me asked a question, or I took a sip from my coffee mug, that doesn't mean it's hard necessarily. Even though all of those take at least as much time as answering a hard card un-interrupted would.
The AI cannot read my mind, there is no approximation that will work reasonably accurately here for "how confident was I in my answer", unless I input that myself.
It should definitely be added as a variable within the calculation, but the current FSRS predicts how likely you are to access the memory (if it's sufficiently available which is defined by its retrieval strength) and speed of retrieval isn't really a factor in this version. The different grades are more to define how well all parts of the memory is retrieved.
Not to say that how quickly you can access it doesn't play a role in real life.
Whenever I try to use anki I can't figure what those four buttons actually mean, so I end up with 40 cards that I still can't recall and then the thing happily drops another 10 on top and I just delete the deck or the app. Haven't learned the thing I was trying to learn with it ever.
Either I don't understand the algorithm or it doesn't understand me.
The four buttons is apparently a contentious topic in the community. It's gotten more serious because in FSRS misusing "hard" to mean "I didn't get it, but I felt close" is really bad and throws off the algorithm.
I like the design suggestions proposed at [1] and [2] for this particular problem. [2] in particular gives tooltips which are supposed to guide you toward exactly what the buttons mean:
- Again: "My answer was completely incorrect"
- Hard: "My answer was correct, but I hesitated a lot"
- Good: "My answer was correct, and I hesitated a little"
- Easy: "My answer was correct, and I didn't hesitate"
That said, you can also just reduce it to a two-button system: only ever use Again and Good. There is some evidence this works better, especially with FSRS which is doing enough machine-learning behind the scenes anyway that it doesn't need the extra signal from Hard vs. Good vs. Easy.
[1]: https://forums.ankiweb.net/t/how-to-prevent-users-from-misus... [2]: https://forums.ankiweb.net/t/how-to-prevent-users-from-misus...
My tip is to map the 1-4 difficulties as "wrong, or <60% confidence", "60-80% confidence, thought required", "90%+ confidence, thought required", and "90%+ confidence, no serious thought".
Depending on what you're learning, you might vary those. For language learning, that works well imo.
Also, make sure to switch to FSRS. The old algorithm defaulted to "again" resetting a card to 0, while "again" in FSRS does show it again, but doesn't reset it back to being effectively new.
If you want to understand the algorithm, read this: https://github.com/open-spaced-repetition/fsrs4anki/wiki/ABC...
Maybe the manual will help you: https://docs.ankiweb.net/studying.html#answer-buttons