Hacker News Reader

That growing backlog is a big issue with this type of scheduling. I think it's one of those almost "unexpected" side-effects of spaced repetition. I mean, if you did the math up front, you should know generally how busy you're going to be over the next few weeks based on how many cards you're reviewing per day, how many new ones are coming, and how you're scoring with each card, but it's not really something people do, nor does the UI show you that.

I wrote my own spaced repetition app (see my profile) and one thing I experimented with is an "ETA" function which goes through your deck and tells you how long before you learn everything "well enough", where "well enough" just means the intervals between all cards is at least some fixed threshold, like one month. I was surprised that even with a few hundred cards and a fixed number of cards reviewed per day, it would still take a few months.

I haven't added the ETA feature yet, but if I do, I think I would need to make sure to show the user how adding new cards affects that target date, and also how reviewing more cards a day or fewer cards a day affects it. I think there's a way to design a nice UI so users feel like they are in control of that end date so they can choose if it's worth it to add new cards or not.

I think a spaced repetition app should also make "falling behind" not feel catastrophic. I think memorizing a smaller set of facts consistently is better than trying to memorize a larger set and then giving up, and an app's UI can probably help to that end.

> I think memorizing a smaller set of facts consistently is better

In general, existing SRS does a very bad job of supporting priority and prerequisites between items. There's no way of saying "learn A, B, C, then learn X, Y, Z once you know the former well enough." This limits the domains where SRS can be applied most effectively to ones with very little structure, such as raw language vocabulary.

> In general, existing SRS does a very bad job of supporting priority and prerequisites between items.

This is a feature I've wanted to implement for a while, but I think it would be hard from a user's perspective to set up the dependencies. Specifically, I mean marking a card unlearnable until another card (or cards) is learned. It's something that I think requires a good UI so you can see at a glance which cards depend on which other cards. You'd go in and essentially make a skill tree of all the cards. It might be something more useful for more serious deck creators.

My dream is being able to create a deck where you have both vocab and then sentences that use the vocab. You'd make sentences that use the vocab dependent on the vocab being learned first, so that over time you'd gradually unlock more material that uses your base knowledge.

Where this feature would be especially useful too is idioms and collocations ("words best used together"), not just full sentences. It's something that's not often addressed in formal language learning but a big part of native fluency. And yes, it's a lot of data. Though it could be made manageable by e.g. importing it from external knowledge graphs/mindmaps and the like.

In my experience, the brain collects background statistics on collocations through exposure to native media. It's an aspect of language learning I'm not convinced SRS is suited to or really appropriate for.

>My dream is being able to create a deck where you have both vocab and then sentences that use the vocab. You'd make sentences that use the vocab dependent on the vocab being learned first, so that over time you'd gradually unlock more material that uses your base knowledge.

Pretty unnecessary in my experience. I only did around 2k sentences and that was for the express purpose of internalizing grammar, and found that 99 times out of 100 it was pretty natural to just learn the new vocab that came with the sentence. There would be the occasional one that would take a few times before I got it, but it was pretty insignificant problem.

After I smashed out all the grammar I just did purely vocab after that and it worked like an absolute charm.

> My dream is being able to create a deck where you have both vocab and then sentences that use the vocab. You'd make sentences that use the vocab dependent on the vocab being learned first, so that over time you'd gradually unlock more material that uses your base knowledge.

You can also go the other way. Have a huge database of sentences, and pick example sentences for new cards which only use vocabulary which you already know (besides the new word which you want to learn through that card, of course). I do that for my Japanese SRS app, and it works decently well from the feedback I get from my users.

> Source: input 35 new vocab a day into Anki while learning Japanese a decade ago and grew my vocab deck to 18000+ cards. Reviews would take several hours a day, and at one point I stopped adding new words to try and wait for the daily review load to go down. When it never did despite this, I just deleted the entire thing and say fuck it.

Yeah, but that isn't really inherent to the concept of SRS; it's just an unfortunate consequence of Anki's geriatric SRS algorithm.

For reference, for my Japanese SRS app (which is using my own custom SRS algorithm) I have a user who currently maintains ~18k known words, with a consistent ~23 new cards per day for months, and they don't take that long with only something like ~20 minutes spent per day reviewing.

Yeah, I've got 9000 reviews waiting for me in my old Chinese Anki deck, at some point you always fall off the wagon. I've changed my mind several times on SRS going back and forth, and my conclusion is that, for language, I think it's better to split your time with 90% on consuming meaningful input and 10% SRS, because learning in context with real content that you're interested in allows you to encode your memories much more efficiently. The problem with pairing available SRS systems with consuming real content is that they don't take into account "natural" repetitions, e.g. when you encountered a word in a movie outside the SRS, so they tend to underestimate how well you know a word. But also since "natural" repetitions are baked into the SRS model via parameters that fit the "average", it also tends to overestimate your knowledge of a lot of other items. Combined, you just have a very inefficient schedule that takes too much of your time for too little gain

A prioritized queue of stuff to learn may be a better solution than staring at "over 9000" things to review. Such a queue would mean you simply bite off as much as you can chew every day and don't perpetually fall behind, assuming the priority queue has a small degree of randomness and is recomputed daily based on the decay rate of all modeled facts. Do you agree?

It doesn't help.

The rate at which you can add new items is the inverse of the review period. What you're effectively doing is increasing the review interval, but you will still run into the fact that the number of items you can remove from the queue is finite. Eventually you will have only interesting items on the end of the queue equal to the number of items you can review. You're now back at square one: you don't have the time to add any new items.

The only way to deal with this is to review items a finite number of times and hope for the best.

When you're working on clearing away a backlog, your review periods should get bigger because Anki will give you "credit" for recalling these items longer. So it seems like reviews should be manageable as long as you're not forgetting too much. And Anki will always prioritize the items you're at highest risk of forgetting.

It still doesn't work for the reason OP said. Your time between reviews reaches infinity regardless.

I think it can help to keep things less overwhelming.

But what I'm trying now is to ruthlessly suspend any card that I get wrong even once after the learning phase (i.e. the day after). I've noticed that some words stick much easier than others, there might be an 80/20 kind of distribution where 20% of the words are responsible for 80% of the time spent. In that case it might be better to add a bunch of words every day and then suspend the ones that don't stick very easily. The idea being that they'll stick more easily at a later point, at least this has been my experience with Chinese.

Actually I'm going to dig through my Anki database and see if my hunch is right.

Edit: it's more like 25% of the cards are responsible for close to 45% of reps. But 25% of the cards are responsible for 58% of review time, and then it's 36/70. That indicates to me that dropping the something like 30% hardest cards would save you a disproportionate amount of time. Of course there's some nuance here where I should probably do this per card type, otherwise you'd just drop the hardest card types.

I did the AJATT method way back in the day when it was relatively new. So heavily leaned on Anki, whilst also consuming loads of natural Japanese through music, shows, games, conversation etc.

In my experience doing both in tandem was the magic combination. Anki without native media is mildly pointless. Native media without Anki is super inefficient. Used in equal measure you find your access and understanding of native media continually ratchets up day by day every 3000 cards I would notice a very significant jump in my abilities when consuming native media and partaking in conversation.

Nothing else has really come close for me out of all the other things I've tried.

That's essentially what I tried with Chinese, and I could sort of keep up with it before I had a kid but then I just can't find the time/energy to keep the backlog down AND watch new content, so for several years I stopped learning completely.

I also want to minimize time spent in Anki because it's just not fun, you know? I know common wisdom says it _should_ be painful, that means you're learning, but I'm not so sure. Me and everyone I know learned English without much pain or Spaced Repetition Systems after all. It's a much easier language to learn than Japanese/Chinese though, but it makes me feel it shouldn't have to be this difficult.

Not quite. The review intervals increase exponentially, so if you stop adding new cards the workload quickly goes to single digits per day in a few months.

35 a day is excessive for sure. Even with foreign language stuff which is the lowest time to solve per card, I limit myself to 10 new cards/day. Math problems 2 new/day. Daily workload ends up around 50-100 cards, which is very manageable. You do need to be consistent though.

35 a day was my magic number. The method to the madness was that I'd allow for only having to hit it at least 80% of the time. In a given month if I managed it every day it meant 1000 cards a month. Every 3 months I leveled up pretty hard.

In my first year studying Japanese I first did Remembering The Kanji, which was a slow to get through and I didn't manage that pace really due to several periods of starting / stopping. After I got through it though I knew bugger all Japanese functionally speaking but the language became extremely accessible due to having a solid command of recognizing Kanji and being able to look them up easily when needing to read stuff, so I decided to front load grammar after that. At the time the JLPT was about 4 months away and I wouldn't have even passed JLPT5 (well, JLPT4 back then) at that time. I found a really amazing Anki deck that I don't know if you can find anymore called Kanji Oddysee 2001. It contained 2000 sentences which covered up to I would say around about JLPT2 grammar as well as teaching vocab + readings for the 2000~ Joyou Kanji (hence the name). The cards also had audio as well. So given 4 months and 2000 cards with 2 card types I smashed out 4000 cards and sat the JLPT2 after just under a year of study. I think I needed 60% to pass at the time and got like 55%? So, not a pass but I took it more as an impossible challenge than anything, and was stunningly happy with the result.

After that I had a vocab of around a couple thousand words, a decent command of grammar, and some limited ability to converse. I just turned my efforts towards building a vocab deck at 35 words a day. I managed to add another 10k vocab in that second year after which point conversation was fluid and I was reading novels (still needed a dictionary for that, but it was huge at the time). I got in some Japanese room mates too and we had loads of parties.

Hands down one of the fondest times of my life.