Deciding what to do for everyone based on what some folks complain about isn’t always the best practice. If those complaints tended to be completely valid, and suggested an overall issue with a 90% default DR (or an issue with the algorithm), that would be one thing.
But most of the generic intervals-are-too-long complaints I’ve seen are about (a) the user needing to adjust to how much better/more efficient FSRS is at scheduling – i.e. intervals can be longer, but maintain retention – or (b) revealing some other issue in the user’s collection – which needs more investigation (sometimes also resulting in bug fixes, which benefit everyone), and might be partially masked by a higher default DR.
I have my DR set to 95% for 236 days now. The workload is surprisingly low, even though 95% is supposed to be on the high-end. So in my opinion, it’s managable and one can remember more, so it would be a good change.
(though I’m not sure if the bump from .90 to .95 is going to make that big of a difference)
Or maybe change the default parameters ? I don’t know, do people that complain about long intervals have way lower than expected R ?
It might be, but if it’s the case, it might also be a problem of “average R” + “RMSE” that is making a 90% becoming a 82-85%, which is not that much big of a deal since it’s still quite under control.
At the same time, I’d still say I’m in a favor to start at higher DR for people like me. I have the feeling (but that need to be verified by data) that when a card is failed, stability can be in general crushed quite by a lot, and take time to recover from it.
But I think 90% is quite a happy medium point where it’s easy to move up or down. So I’d stick to 90%
I don’t think we can find parameters that result in shorter intervals without sacrificing accuracy. For the record, I do think we can find default parameters that are better than the current ones, but the intervals would be even longer.
It’s better to just raise default DR.
I gave my vote to keeping it at 90%, mostly because there are complaints of too long ivls, but there are also complaints of unweildy review load. A lot of people quit Anki because they can not keep up with their reviews or catch up with backlog.
I can’t see people quitting Anki because of “too long ivls” but I can see people quitting Anki because of “too much work everyday”.
Only if we could do A/B testing and actually figure out what will improve user satisfaction/retention
I would follow this by adding my estimation that, while the complaint ‘my intervals tooooo big’ is an immediate reaction, we run the risk of trading that for the downstream complaint of ‘my workload tooooo big’ down the line. Which might, in a few ways, be more troublesome to solve (and bring with it more user frustration, given the cumulative effect burden).
‘Mah interval too big’ is an understandable inuitive concern, but we sort of know it’s a subjective and (fairly) unreliable reaction, at least on its own as a metric.
My interval too big = easy enough to solve.
My workload has ballooned (to the point that I’m seeking help) = likely to bring with it more accumulated friction, confusion, irritation/desperation, etc.
Not that you can’t solve it by rescheduling on change, I suppose, or whatever else, but I think you run the risk of bringing along for the ride inherited user frustration by virtue of the accumulated friction that got them to that point.
Separately, at least personally, I see a lot of this complaint often centered around “my exam XXX days.” Which I sort of see as being a related, but distinct, problem in some ways. Which might be solved by, if I remember correctly, conversations I’ve seen regarding FSRS/Anki being able to optimize knowledge centered on a specific date/timeframe, yeah? Couple birds, one big stone, etc.
Which might be solved by, if I remember correctly, conversations I’ve seen regarding FSRS/Anki being able to optimize knowledge centered on a specific date/timeframe
I believe this is just a factor of not reading the manual and understanding what DR is and how it works with the scheduler.
I believe these complaints are by students who believe Anki is simply exam preparation and not long term recall. %90 is fine and if someone complains they can just be guided to the docs and they can adjust it to their liking.
Should the DR be upped, we’ll possibly find complaints of intervals too short start to appear, can’t please everyone.
Lastly I believe ton is a bit hyperbole, a few dozen threads on Reddit is certainly no ton.
Because Anki is not an exam preparation tool and I do not think Reddit is anywhere representative of the userbase, that would require a significant effort to poll and gather data.
Should the DR be upped, we’ll possibly find complaints of intervals too short start to appear, can’t please everyone.
Considering that I almost never see complaints about intervals being too short, hard disagree on this.
To put it another way: right now 95% of the time people complain about intervals being too long and 5% they complain about intervals being too short. I’m trying to get 50/50. Unless the default DR is 97+, we won’t see complaints about intervals being too short outnumbering complaints about intervals being too long.
If users are scared of long the intervals, they won’t use Anki in the first place
But the hurdle that stopped me for two years is a little strange: The brain behind Anki - whatever it is that decides whether I am about to forget/should review a card - has way too much faith in my memory.
The only way I started seeing the benefit of it was to make a custom study deck and study my next due cards, 100 or so at a time, usually much sooner than Anki intended on showing them to me by itself.
What good is it that long intervals are “optimal” if they make people quit before they discover any benefits of Anki?
I agree with you that we should think more about people actually liking the app rather than some statistics a normal homo sapien wouldn’t care about. But I’m not sure you’re on the right track.
The manual mentions many people study a lot of cards initially and end up having a backlog and “how to clear my backlog” seems like a huge issue in the community.
Not everyone is consistent with Anki and even a few days of taking an off can build up unwieldy amount of backlog. Especially among the med students of my country, I’ve seen this opinion that “Anki bad because too much work”. So, I’m looking at things from a different perspective.
The question now is, how can we be sure we’re actually optimising the app for users? It would be useful to have some way of collecting data (A/B testing, making a temporary change to test the waters) but we seem to have none.
I have experienced both ends of the spectrum; that is, Anki has become a feast or famine, deluge or dearth tool for language learning. For context, I already have DR at 95%. But I also add vocabulary and additional cloze cards one chapter at a time. After doing so, I have a glut of new cards that overwhelm me for a couple weeks, which is tiring considering Hebrew is challenging to learn with its myriad of minute rules. The tendency is to “Good” everybody to knock the workload down, which inevitably drives me into the desert sands of review. Because Hebrew requires consistent use and review to master, Anki is simply not working well for me.
Honestly, I once again am considering my efforts using Anki wasted, and physical index cards more productive.
That user is not a great one for your to cherry-pick. Is there any reason to think that it was FSRS intervals that they were put off by for the past 2 years? It seems unlikely. FSRS is probably a solution for them, not a problem.
Is there any reason to think that it was FSRS intervals that they were put off by for the past
He said “for two years”, so almost certainly not FSRS. But remember, FSRS with default parameters and 90% DR makes people complain. Sure, he could increase DR and optimize parameters, but that requires sticking around long enough and not immediately quitting. If FSRS was the default, I guarantee you he would quit.
As I said above, if long intervals make people quit, they won’t get to the “increase DR and optimize parameters” part. Or to any other part for that matter.
As for cherry picking, I can write you a list of post IDs of people who complain about intervals being too long or too short. The latter barely ever happens. I have 57 IDs to be exact, and I would be shocked if more than 5 of them are about intervals being short.
EDIT: I have 57 IDs of posts that are purely about interval lengths. I have 122 more IDs if I include posts that are about interval lengths and also about desired retention/optimization/helper add-on/whatnot.
I’m not sure why that would be better, if it’s still going to be the same number of complaints. We’ll just need to come up with a wider variety of answers. That sounds like a change that will make it harder on helpers, while benefitting a net 0-share of users.
50/50 means that it’s well balanced. Intervals are too long for half of users and too short for the other half.
Obviously we can’t satisfy everyone, but we can at least make sure that people aren’t disproportionately dissatisfied with one particular thing.
