I don’t think your 99% desired retention is the issue with your stats. Even at 99%, your intervals would grown somewhat. The problem your stats can’t overcome is your 1d max interval.
Are you for or against sub 1d intervals (with a hard cap) 
Either you’re talking about an idea that is indistinguishable from learning/relearning steps – or I have no idea what you mean by that.
He’s talking about learning steps, but controlled by FSRS.
Expertium said that FSRS is looking to take learning steps into account in the future. So I suggest the possibility of FSRS starting to schedule subday intervals as well, as it only schedules >1d intervals only.
I think normal (starting) learning steps should be kept. Just that FSRS does not ALWAYS have to begin >1d where necessary.
There is concern that this would lead to inconvenient reviewing times, so I suggested a hard cap, which already is a feature present in Anki. @voczi concurs.
I don’t see any purpose in that. Is that somehow related to Expertium’s original set of suggestions in this thread? Or are we off the rails?
This is a suggestion made on my part based on a previous conversation.
And the reason I think this is great is for the same reason we are letting FSRS schedule reviews >1d. I don’t see what could prevent FSRS schedule <1d intervals as well, that’s all. This would solve the headache of having multiple learning steps (or lack thereof). And makes learning more tailored for harder cards.
I agree that it would be good, but it seems like there is too much technical debt. On a purely theoretical level, I agree with you and disagree with Dae.
Actually, I think that both removing <1d intervals completely or letting FSRS control them are better than the current situation.
1 Like
Do you mean by technical debt that you and Sherlock haven’t got around to putting this into consideration and that you guys are busy with other things or that this would require a complete overhaul of FSRS I see this just as an extrapolation of FSRS scheduling, just sub-1d downwards (but then again, I don’t know the intricacies behind all of this 
)
Neither. This is an issue with Anki and how cards in the “learning” phase are handled very differently compared to cards in the “review” phase. So if you want this to be done, you should talk to Dae, rather than LMSherlock.
1 Like
I’ll consider this if FSRS could get the actual intervals between subday delays, but even so, with new cards in particular, you have very little info to work with so don’t think FSRS can improve the delays much.
This is why I don’t really support doing away with the original learning steps (set by the user). You would want to keep it as it is for the reason you mentioned. I imagine this:
–>The user has set the learning steps 10m and 1h.
Depending on how he fares with these learning steps, FSRS can either
- graduate this >=1d (as normal-this is the current state of FSRS)
- or make a new <1d interval depending on the difficulty of the card
- or it would be promoted to the next day if the set interval is <1d but more than the cap.
@Expertium said in fact he is for automizing the entire thing altogether, if I understood correctly. I think this is a good middle ground hybrid solution.
You want sub-day intervals after the steps? Is it even beneficial given that most people get >1d initial intervals (in fact we had to explain it in FAQs that this is normal).
You want sub-day intervals after the steps? Is it even beneficial given that most people get >1d initial intervals (in fact we had to explain it in FAQs that this is normal).
That is for FSRS to decide. Its job should be to give you the best interval within its framework. If it is optimal to have an interval within 6 hours then so be it. It just means that you have a very hard card that you are bound to keep forgetting. If you don’t want it up until a certain time of the day (perhaps it’s your sleep time or you like to take a break) you can always make the hard cap tell FSRS off.
Just like I said: We have let FSRS decide >1d intervals, so why not make it <1d as well?
I think there isn’t much room for breakthrough out-of-the-world improvement anymore (I hope I am wrong), so there is room left for fine-tuning the little things that LMSherlock and Expertium have been doing. And this is one department I hope to see this improvement in.
1 Like
So you do a review after a manually set learning step, and then a same-day review scheduled by FSRS? Ok, sorry, but that’s a terrible middle ground. Hard no. It’s confusing, and it’s a waste of time. Just let FSRS do everything and remove user-defined learning steps entirely.
I think there isn’t much room for breakthrough out-of-the-world improvement anymore (I hope I am wrong)
Unless this is implemented, yes, there isn’t much left to improve.
1 Like
Hmm…yeah I am all for automation: That would be even better. I just don’t know how that could happen so I suggested this.
And yeah factoring time of the day into scheduling is also good. Anki already documents performance by the hour and maybe that could be factored somehow into the FSRS scheduling.
What is dae’s opinion on all of this. What is the hold up
That gives me hope at least that one day we could see these changes in the foreseeable future 
“Retrievability (Descending)” is imo essential. It should probably be the default sort. When cards come due, you want to be seeing them again right when they dip under that retrievability threshold you’ve set. Making freshly due cards wait while you slog through your backlog is just causing more cards to be forgotten than is necessary.
The source of a lot of error in machine learning is just noise, and it’s not a bad thing. You don’t want to go chasing away all the errors and end up overfitting the data. That would make the algorithm worse. Fitting parameters to each card as if they were their own deck would be way overfitting the data.
Ooph, really don’t like this idea. My cards vary quite a bit on how much time I take on each, and it has nothing to do with difficulty. Some cards just have a little more to read on them, some cards have longer answers, etc. I do cards while I’m sitting at a red light, and I allow myself to look up at the light and check if it’s green, which takes time. That would show up as the card being more difficult.
I’m fine with the 2 button system, but keep it based on subjective intuition in the moment, not on something programmable.
1 Like
Relevant article in my blog. I agree that time should not be used as a proxy for an answer.
1 Like