I did 98,000 Anki reviews. Anki is already dead | Hacker News Reader

I did 98,000 Anki reviews. Anki is already dead

miguelconner.substack.com

66 pointsdothereading91 comments

I think that the author was using Anki incorrectly, and that led them to the spurious conclusion that "Anki is dead". I also have attempted to use Anki this way- using someone else's deck to try to force myself to learn something new. But that doesn't work, because it is just memorizing random symbols, as they noted in the article ("The enemy is the static card"). For example in maths learning, memorizing arbitrary terms, symbols, etc is useless. However, once I am introduced to a concept for the first time, then I add it to my Anki deck so I can make sure I remember it. The key is the context, and writing terms / definitions etc that speak specifically to me. I still need to work out different variations of the concept to understand it, and that's not something that Anki can help with.

I haven't used Anki for language learning, but I imagine that if I did, it would be to add some new vocabulary I had just learned from a book, conversation, film, etc. I don't think it would help me learn a language from zero though- that would require practicing it.

In summary, Anki is great for reinforcing something you've just learned, but you can't reinforce your way into the context that is necessary to truly understand something.

Using someone elses deck is such a siren song, and I honestly believe it's detrimental to properly using Anki. Making and curating your own cards is an essential step in learning the new concepts.

Yes, thou the main problem stated in the article ("memorizing the rectangles") is very real. Now, I don't think you need much magic / LLM stuff to overcome that problem:

1. Instead of the usual "1 new word, 1 card" have 2 - 3 new words on each card, and have each new word in 3 - 4 different cards (ideally with different inflection, meaning, nuance, etc) 2. Review cards fast and very few times each. Like 5 - 8 times / card (max) instead of usual 15 - 20. 3. Don't punish yourself, keep moving on even if you just half-remember. First familiarize and then internalize language patterns instead of just memorize words.

Review intentionally, totally concentrated on the task. 10 mins / day well done >>> 30 mins / day mindlessly.

Outcome: more fun, more effective learning, no memorizing rectangles.

Look, if you want to learn a new language, move to a country and get a boyfriend/girlfriend. slaps hands my job is done here

Forty years ago, my French teacher said the easiest way to learn a language is to go live there for a few months. He didn't mention the boy/girlfriend part.

> that led them to the spurious conclusion that "Anki is dead"

It didn't. They wrote "Anki is dead" because it brings clicks.

Yeah. Anki (and flashcards in general) are great for helping you remember something _you_ learned (from a book, video, class, etc.). Not for transferring knowledge someone else learned.

Writing my own cards as I'm learning is the only way I've found it effective.

First learn the material. Then use SRS to schedule the practice of effortfully recalling the material just before you would have forgotten it.

I think too many people use SRS to learn the material instead.

As I recall, the creator of Supermemo had a list of 20 suggestions or so, in which he urged people to first comprehend the information; then to learn it; then to memorize it; and then to rehearse it (SRS) so as not to forget it.

I can't agree with this when it comes to language learning specifically. I've used all sorts of tools to learn the language of my in-law side of the family: books, audio lessons, videos, tutoring, apps, etc. I've never made as much progress as I have with Anki. My language skills have improved in leaps and bounds so much with Anki that I now rarely bother with anything else, other than for sentence mining to create more Anki cards (e.g. from grammar books or apps) or just for a bit of variety to make the process more enjoyable. I actually find classes/tutoring (traditionally seen as the "best" method) frustrating because of how slowly I learn with them compared to Anki. In a 1 hour lesson we might cover a handful of concepts and words/phrases that I will almost certainly have forgotten most of by the time the next lesson comes around. In that same hour I can create and review a ton of Anki cards and I'll remember most of them.

With tools like Google and Microsoft's neural TTS and Anki's AwesomeTTS add-on my cards have audio that is so realistic that I am also constantly exposed to near-native listening. I do 3-way cards (Writing only -> English, Audio only -> English, and English -> Other language) so I'm actually getting a reasonable simulation of real life practice (reading, listening, speaking) on an individual sentence basis. My process is: (1) find a high quality sentence from a book / app / website / ChatGPT (with verification from a native speaker); preferably one that is fairly simple apart from a single word or verb conjugation that I haven't learned yet, in keeping with the i+1 rule, (2) create an Anki card for that sentence using my own custom note templates, (3) add audio with AwesomeTTS. Creating a card like this takes me perhaps 10-20 seconds as its mostly just copy-pasting and clicking a few buttons.

Of course to become truly fluent you need practice. But when I practice I'm already able to follow the gist of conversations and I can stumble my way through speaking in most situations: I've got a huge head start thanks to all the latent vocabulary and grammar that my brain knows thanks to Anki, instead of having to constantly look blankly at the other person while I pull out Google Translate.

This sound great. I'd love to see a 1 minute video of this in action.

Me too!

So I definitely agree that this is 100% the best way to use Anki, that's why I wrote the line about "Writing cards that trigger memories of experiences I had in the real world always produced better cards."

I couldn't give you a percentage, but I made most of my own cards, including all of those 2000+ kanji cards. There's lots of debate in the language learning community about vocab cards or sentence cards, and generally the ideal is the sentence cards, as it provides the context that helps you use is naturally (as opposed to literal translations from your native language).

> I still need to work out different variations of the concept to understand it, and that's not something that Anki can help with.

But imagine if it could!

I agree that sentences are generally superior to vocab. Vocab cards are extremely problematic once you go beyond beginner level, because words can (and often do) have multiple translations in both directions. This could be because there are multiple words meaning the same thing, or because a word has multiple meanings.

For example if the English prompt is "watermelon" - are you supposed to recall the Italian word cocomero, anguria, or melone d'aqua (all of which mean watermelon)? If the English prompt is "bank" - is that a place you deposit money, a river bank, to bank (turn) a plane, or to bank (count) on something happening? You end up having to build in messy hacks like giving clues in the prompt as to which translation is intended (which means you memorise the clue instead of the word) or having cards for bank(1), bank(2), bank(3), and bank(4) which becomes very tedious for recall. Sentences mitigate these problems somewhat.

I now only use vocab cards for object nouns where there's only one important translation, and mainly because I can put pictures on these cards so that I'm learning from e.g. the concept of an orange instead of the English word for orange (which saves you the step of mentally translating when you aren't yet fluent with the word).

I'm so tempted to try improving my language skills with Anki, both for my native language and my daily use language. But the commitment feels so daunting- I've barely missed a day in my reviews for the last two years, and only have 28,000 reviews total. I'm very impressed by your 98,000!

I guess the best way to start is just to create a new deck in it with one card and then go from there. I already have a daily review habit, which is the most important part.

Yeah I was definitely doing too many for it to be fun.

I think that when you have a really high level, Anki is actually even better than when you are at a lower level, because you already have the intuition for the language in general and you are just adding one small component. At lower levels you are making more assumptions about how the word will be used and that can lead you astray.

So I'd say try it!

I think people underrate the utility of a premade deck when constructed well and used appropriately.

A use case I've found is if you can find a deck that corresponds to a book you're reading.

I found a deck for the Rust book and it's structured such that you can see cards about things in the order you read about them. You simply read the book as usual, learning from your reading and entering code into a terminal as instructed, and then test you understanding with the cards.

When you end up reviewing older cards, you end up getting the benefits of putting them in long term memory but you also get the opportunity to make more connections as you revisit concepts which has its own benefits for deepening understanding.

I've found this makes reading the book 10x more effective. I get so much more out of it.

This all depends on having a source from which you're learning and the deck is just for testing understanding.

But yes anytime you're using Anki to learn/understand instead of to remember, you're likely misusing it. Anki is a tool for memory.

That sounds like an awesome experience. Can you please share the links to the deck and the book?

When anyone mentions the Rust book, they're talking about the recommended place to start in the Rust community: https://doc.rust-lang.org/book/

To find a shared deck, I usually go here: https://ankiweb.net/shared/decks

Search for "rust".

You'll find two decks with 550-560 cards.

The older one was the original and whoever created it did most of the work and should be blessed by the heavens.

The newer one took the older one, and replaced the screenshots of code with markdown equivalents so they could be rendered by Anki while saving memory. You can see this in the difference in the number of images between the two decks. This is the one I'd download and use.

Thanks a lot!

Somewhat related: you might dig Andy Matuschak's Quantum Country project, which aims to incorporate a text's flash cards inside the text itself.

https://quantum.country/

Later edit: I should have said "Andy Matuschak and Michael Nielson's".

Wow, I've been looking for this. Thanks for sharing!

I love Andy Matuschak! His podcast with Dwarkesh was so enlightening and his blog is great as well. He's one of those people whose work I go back and read every couple of months and I always learn something new

> This will bring about personalized tutoring like in Spock’s school in Star Trek (which always seemed amazing to me even as a kid).

* Confused where in the original series Spock goes to school.

* Watches the video and sees 2009 "Star Trek"

* "as a kid"...

* feels old

> The enemy is the static card. It always has the same front, formatting, and font. After enough reps I would latch on to little cues that are irrelevant to the meaning of the card, meaning I would skip the very important step of thinking deeply about the content. A fairly common occurrence was that a word in the sentence would remind me about the meaning of the sentence, giving away the answer to the target word, which is quite different from piecing together the meaning of the target word in a brand new sentence.

Interestingly, that's the "trick" behind a lot of the seemingly magic skill of geo guessers. The best players have played so much, that they now "see" things that a regular person wouldn't even consider to look for, like the camera quality, what year the car was from, and so they narrow down the possible countries by those aspects, before even looking at the "picture".

> Interestingly, that's the "trick" behind a lot of the seemingly magic skill of geo guessers. The best players have played so much, that they now "see" things that a regular person wouldn't even consider to look for, like the camera quality, what year the car was from, and so they narrow down the possible countries by those aspects, before even looking at the "picture".

Slightly OT, but this happens constantly with ML classifiers on any highly multi-dimensional problem. At first it seems like magic, and then someone digs into the principal components of the prediction, and finds a mixture of a few highly specific factors that -- in the worst case -- is an artifact of the dataset itself (image blur or color bias, for example).

Also common is that the predictive factors aren't pathological -- they're just "boring" -- and therefore the performance of the model is dismissed by the practitioner ("oh, I'd have thought of that, since it's only using a few common traits that are well-understood.")

The one time I tried Anki u realised I started recalling the answers based on the shape of the words lay out on the card without necessarily reading them.

I used Anki pretty heavily during ~8 years of law school and it was a game-changer. Two things that helped me the most (and that I think apply to any kind of flashcards, digital or paper):

- The twenty rules of formulating knowledge (https://www.supermemo.com/en/blog/twenty-rules-of-formulatin...) ; old but really solid advice on how to actually write cards that stick.

- The book Fluent Forever; it’s meant for languages, but the general principles carry over to learning basically anything.

Wow the twenty rules article should be required reading for anyone downloading Anki.

Hallucinations in LLMs when learning is dangerous; IF you have some background, you can usually tell with LLMs go off the rails, but It would be unfortunate for you to commit to memory an incorrect fact at such a vulnerable time. It will be difficult to "uncommit it" at that point.

I don't think the LLM value prop here is to build lots of cards. It's just to interact with the model conversaitonally. If the model is wrong in this one translation, I'm not going to exactly "commit it to memory", I'm just going to keep on carrying on. I don't know that the LLM mistakes are any particularly worse than the many and sundry other mistakes I'm already continuously making as a language learner anyhow.

If I could speak a foreign language as well as a 8B parameter LLM, hallucinations and all, I'd be immensely ahead of where I am now. It's not like second languages aren't themselves often broken in somewhat similar ways.

I've used anki before. Most of the decks you get are randomly downloaded from the anki website. I'm not sure that an LLM hallucination is that much more likely than a typo or error in a random free deck I downloaded that someone compiled for me.

I use Anki to study all kinds of subjects, and more than half of the value is processing what I learned (usually with pencil and paper) into good anki cards ("atomic", as per Michael Nielsen's definition), including the insights that I had when studying the subject (like "what's the comparison with X that I used to understand Y?")

I'm not sure if it is efficient, mind you, but I suppose it's effective because I can recall information later when relevant, and I believe that like exercising just being able to stick to a study routine ends up being more important than picking the best routine

I had the exact same experience. The more work I did up front on a card, the better I was able to remember. When I was learning the kanji cards, I would take the time to draw them out in a special notebook while thinking carefully about the different components, and it really helped with retention.

I did not do this with many cards though, hoping that they would eventually stick.

I think in general the more you engage with the thing you are doing, the better you remember. Even when reading or listening to a lecture or whatever. Maybe what I'm proposing here is that by making it dynamic you create a system where deeper engagement is necessary.

Always good to see more activity in the language learning space. I wish you luck now that you've moved past Anki.

A piece of feedback: one of the common issues I've found with AI-generated questions & answers based on an article is that they will often hone in on testing values. It looks like incontextlearning.com often suffers from this same issue (over half my comprehension questions were "how many"-style). I can easily answer these types of questions even if I don't know what the content is about.

Thanks for the advice! I'll try modifying the prompt a bit and see if I can pull out more engaging questions.

I don't see the dichotomy, both tools seem rather complementary to me.

For example I use LLMs to generate cards for me, and Anki's algorithm to make them stick.

Similarly a LLM plugin could easily present a fresh sentence each time you review a particular vocab

I was thinking the same thing. Currently all my cards try to all follow the same pattern, making it necessary to really remember at based on differences between them. (For example, I would never ask "What happened in 1914?" but instead just "1914" and have "Start of WWI" on the back, as a reversible card). LLMs could take such cards and provide the user with an endless stream of questions: What year did WWI begin? What happened in 1914? Which event took place in 1914? And so on. For vocabulary this might of course be even better when the word/phrase itself will always be contextualized in practice.

I agree that they could be complementary, but I think there's a not-yet-made tool that goes even beyond this, where you are interacting with an LLM that has an Anki-like backend of some kind, keeping track not only the number of mistakes but of what kind of mistakes you made and when, so that it can later bringing up the card in a more natural way.

I totally agree with the 4 rules listed here (especially #1), but in line with other comments I don't see how LLMs make flashcards for language learning obsolete. If anything, after ~15 months of semi-casual Japanese learning, with flashcards as a core part of my routine, I'm even more convinced of the power of brute force memorization / spaced repetition. I often have moments where I'm presented with a word I can't quite remember, and 90% of the time my "gut instinct" on the pronunciation is correct (I'm usually wrong when I second guess myself). It's a spooky feeling being able to summon the correct pronunciation for a word from deep within my brain when I've consciously told myself "I don't really remember this one".

Also agree that learning a bunch of vocab / flashcards is only a small part of language learning, and it should only be a small % of your studies. After ~80k review items in ~15 months, I'm at the point where I simply spend 20-30 mins. before bed knocking out 100-140 review items (both recognition and recall). Now it's just part of my daily routine and I genuinely still enjoy it. I invested a lot of initial time with flashcards while working through Genki I+II just to lay some solid foundations, more like 200-300 review items a day (knowing some Chinese was my main headstart with this).

+1 for Skritter. I was an early user way back in 2011 for Chinese, and after setting language learning aside for many years it was a joy to boot it back up. I much prefer the Kana/Kanji interface for writing practice than Anki. Admittedly it is a bit pricey, though. I'd also highly recommend Satori reader as a reading/listening practice app. Aside from that, plenty of manga, conversational Japanese 'podcasts' on YouTube, anime, and video games helps. Next step is to find a tutor at some point....

One thing that I've always found weird among the Japanese language learning community is the emphasis on drilling/learning individual Kanji. I don't really recall this being a "thing" in Chinese (presumably b/c you're on the hook for 4-5+k Hanzi rather than 2-3k Kanji :D). Truly bizarre to me to drill individual Kanji out-of-context rather than just learn a couple of common words that use them.

For the past couple of years, I've used LLMs for helping me creating flashcards -- and loved it. Initially, I built a very simple desktop tool that would send the LLMs output to a Google spreadsheet, from where Flashcards (an iOS app, [1]) would import them as decks.

After a year or so, I wrapped it up in a Telegram bot, so I could just send a new card or phrase to it, and it would take care of translation, pronunciation, and usage examples. As I'm usually learning several languages in parallel, I also added the automatic language detection (from preselected languages).

Then I randomly found out that FSRS is better than SM2 (which Flashcards uses), and found an Elixir library that implements that. So, I decided to drop both spreadsheets and the Flashcards app, and implement the review process right in my bot.

This is how I ended up with https://lexicorn.ai -- just got my first paying subscriber (yay, I guess). The site doesn't say it, but it's 5 euros/month.

[1] https://apps.apple.com/de/app/flashcards-deluxe/id307840670

I'm working on applying LLMs thoughtfully to flashcard-based language learning. The original iteration of the app didn't use AI at all and was more focused on the study interaction, but after some experimentation, I found LLMs worked well for a few key features:

* Generating a variety of example sentences

* Determining if a response was semantically similar to the target answer (my app requires providing an answer, similar to WaniKani), and explaining why or why not

* Explaining context and usage of a word

* Suggesting vocabulary words about highly specific interests (as long as its within the knowledge window of the model)

I've used it for over 60,000 reviews so far and it's working pretty well for my use case.

It can be tried out for free and is currently available on the web and iOS. It supports 8 languages and within each pair of languages, so if you are learning or speak English, Korean, Spanish, Japanese, Italian, French, German, or Chinese you can try it out for free. If it doesn't support your language and you're interested please feel free to get in touch.

https://get.meanings.app

Appropriate use of Anki isn't so simple, it's a skill in itself. Here's my strategy specifically to learn vocabulary. TL = target language, NL = native language.

1- front: image+subtitle (in TL) back: word in TL.

2- fill-in-the-blank phrases for the word, fully in TL, translation shows up after completion (you HAVE to use the function where you actually type it out)

3- front: word in TL back: translation in NL with an image, also the inverse, but the image is always in the back. Making it a different picture as the one for 1 is essential

So each new word would generate me about 6 to 8 new cards. At a fast enough rate of card creation, you won't run into the problem of memorizing each card because you will be creating like 100 cards in a day. The "fatal mistake" (to quote the author) of this article is underestimating how much this process of card creation and organization aids in learning. Creating your own study material IS studying in itself.

That is, assuming the strategy being compared to LLMs here is the correct one of actually studying the language and creating your own Anki deck while you study the material, instead of the incorrect strategy of downloading a deck

Anki helps the best when you are at the beginning of the language journey. With a vocabulary of 0 words you won't be able to read anything at all, and there's nothing to talk to an LLM about yet.

Those 'useless' static cards are extremely efficient for learning the 1st 2000-3000 words, which is key to start reading. After about 4000 there's little sense in using SRS anymore, and then I'd rather spend more time with an actual book, but getting there with anki felt like using a cheat code compared to how I learnt my first foreign language. It's not exciting, it's pure toil, but it does work.

And when it comes to the next stage, I can't imagine how random llm-generated texts are better than, say, graded readers or real books. Most people would likely find it more interesting to spend an hour or two a day following an exciting story and characters they care about, and it's (based on a sample of one) way easier to memorise all of those new words when there's an emotional connection for each one (just how we form associations between words and experiences while growing up).

As for the app itself- I have tried it with my native language, and at the advanced level it produced a sterile and slightly unnatural text with a complexity of a typical fiction. If someone could read this, then I don't see why they would bother. At the beginner level the app generated a couple of news stories which, though simple grammatically, had a vocab that I would never have recommended to a novice. Local news of a "a firefighter saved a kitten stuck on a tree" variety are much more useful for that kind of learning, and you get this from any free newspaper.

LLMs are extremely useful for learning foreign languages, but I feel like this isn't the way to go

Alas, the mote around LLM integration is practically non-existent so I'd think that productization around this would be next to impossible.

Anki is already extremely extendable so I would think that with a not too much work deep LLM integration could be implemented in Anki. Like, instead of showing static content for a card, have Anki call an LLM to create the daily iteration of a given prompt.

I'm excited to see the next crop of real LLM powered language apps.

I could see it coming from 2 directions (since I think most people agree language learning and SRS go together):

- anki on steroids - a dynamic way for me to do SRS that feels more natural - a way to do natural language interaction with the LLM (chat, voice, etc.), with SRS as an added feature or integrated more subtly

My ideal flash card for vocabulary (foreign language) focuses on one word per card but always uses that word in a sentence since words never have one single meaning. Like a single sentence or two with the word in bold that you must define.

I think it'd be cool to then use that "{word} {part of speech} {example sentence}" info to generate more example sentences with an LLM.

That way you can grind your problem words with real sentences and cement it quickly, otherwise this process happens too slowly through regular reading.

Very good article but something bothers me:

> Language learners chase something called i+1 material

I really dislike the traditional language teaching method. It didn't work for me. It's too abstract, boring, and you end up memorizing stuff the wrong way. And usually later you need to un-learn/re-learn things properly.

What worked for me (and fellow struggling students I taught) was normal text about topics I find interesting. Like boats? Pick boating magazines, books, and documentaries/movies (turn subtitles on). From WTF to "I know some of those words" to is that a pattern? And only then go for the actual rules of the language. This way you are engaged and learn real-world things.

And for dull learning, it's better to spend time in i-1. Miyagi-stile practice repetition of the basic things to the point you can't fail even if you are tired. Then move to games like finding rhymes or tongue-twisters. [Ironically, AFAIK this is the Japanese way to learn calligraphy, Judo, etc]

Same applies to many other disciplines. YMMV

In general I agree with this, and there is definitely something a bit insidious about optimizing the entire process so that you are on a perfectly calibrated i+1 treadmill.

But, for some languages with different alphabets and roots, it really is practically impossible to get any meaning from a magazine article in your target language at a beginner level. So getting out of this beginner phase as quickly as possible is really appealing. And if you can find text at this level that is interesting then that really helps.

I really like your comment about i-1 learning, I'd never thought of it like that.

Weird, someone posted something here talking about an API that would track individual words and then feed them into an LLM; I wrote a response, but it seems the comment was deleted? At any rate, hope you all don't mind if I post my comment here, as I think it's relevant and potentially interesting.

8<----

I think I've already built what you want:

https://api-dev.laleolanguage.com/v1/docs

I started working on this system before LLMs were a thing, but its purpose was specifically to address the problem described in this article -- "flashcard blindness", I've heard it called.

The idea was to solve this, instead of with an LLM, but with a giant corpus of native input. The algorithm tracks all the "language building blocks" separately, assigns each of them a difficulty and a study value, and then calculates the total difficulty and total study value of each selection in your corpus. Using that you can find material to read (or listen to, but I haven't gotten that far yet) that balances difficulty and impact on your learning. This way you're actually reading new material, rather than "memorizing rectangles".

There's a public beta for Biblical Greek [1]; I learned Koine Greek entirely through my own system. But I initially developed it for myself for Mandarin; and it's got experimental ports to Korean and Japanese (all three of which are not yet public).

But yes, this could definitely be integrated with an LLM:

1. Using the API, the LLM could ask for the top 40 words to learn or review

2. The LLM could then generate something using something from those words

3. The LLM could send the generated content to the API, to have it graded for difficulty. If the overall difficulty was too high, it could rephrase things to make them simpler (or perhaps even rephrase things to make them more complex, if the difficulty were too low).

4. The LLM could then show the content to the user, and log that the user had seen it.

The API isn't public yet, but if you're interested in trying it out, drop me a line:

contact@laleolanguage.com

[1] https://www.laleolanguage.com

>All this set me up well to understand that the future of learning would change when in late 2022 the LLM kicked down the door, tracked mud across the carpet, ate everything in the fridge, and demanded more snacks.

Thanks for the laugh, I like your writing style but to echo others, I think you went a little extreme on the Anki.

I found Anki way too heavy and the available decks mostly wrong way around for actually learning vocabulary (that is they were from foreign language to native, instead of the other way around) and switching the direction was way too cumbersome.

I have not yet found a really good tool for learning Mandarin, except for classes and actually talking with people and doing the hard work of writing the characters again and again, for which I rarely have energy or patience.

One thing I did notice in a course was, that writing an article about a topic helped a lot. It needs to be something where you use the same new vocabulary many times. But the problem with that is, that it makes my hands and wrist hurt after a couple of writings.

I was quite successful learning Mandarin by using Anki in the way that is typically discouraged, i.e., with a downloaded deck (SpoonFedChinese). I have very little time per day to devote to learning so building my own cars isn't an option, but with this and a couple of last-minute classes I went from HSK2 to HSK3.

I also find Du Chinese and The Chairman's Bao quite useful, although indeed for the writing, nothing seems to substitute actually writing. Right now I can read much more Mandarin than I can write.

For spaced repetition practice for writing, try Skritter.

Definitely by far the best way to learn to write characters. It is very expensive (compared to most apps, anyway) but its essentially obligatory to spend the money if you are serious about learning the language. When I'm learning Mandarin my favoured method is to work my way through the HSK decks on Skritter, plus a custom deck where I add my own words. Then I make sure to put at least 1 example sentence in Anki for each word that I've learned in Skritter, to reinforce it and place it in a useful context. Skritter itself rather helpfully gives an example sentence for most words. This often then leads me to adding more words to my custom Skritter deck when the example sentence contains other words I don't know yet. In this way I expand my vocabulary beyond just HSK.

> Anki’s fatal flaw is that you think you are learning content, but you’re actually just memorizing rectangles.

I totally believe this. It's 100% in line with my observations about Duolingo. I know people who put in a lot of time into Duolingo and learned nothing. I gave it a try and despite the fact that I put maybe 50 hours into Italian (estimate) I learned nothing. I could get all the cards right, but I didn't learn any Italian (Over the years I've learned 3 languages in addition to my native language, so I know it's not my problem.) Eventually I realized that I did learn something but it wasn't the language. Somehow I just knew the right answers.

I use Anki with mostly pre-made decks (correcting mistakes when discovered) and only a couple of my own ones. This works for me, as building a deck from scratch, especially for something you don't know anything about yet, is more effort than I could ever find enough energy for. I only wish there were more curated high-quality decks for things, and I'd even happily pay for them now, but alas.

Which is to say, to each their own. I find it strange that people often insist that one way of learning or using a piece of software is objectively better than all other ways.

I'm working on an app similar to Anki, where I can make my own cards for spaced repetition, organize them by folders and tags, and can run different queries to get groups to practice from. I'm curious what features people wish Anki had. I didn't love it when I tried it.

I want to use my app mainly for language learning, but as a demo, I also have some geography cards that zoom in on a country on the world map, for the front side.

Spaced repetition is incredibly inefficient. Despite exponential guarantees, there's information and incredibly complex concepts that are with me forever, after learning about it once.

Yet somehow a German word is impossible to remember.

LLMs suck, because their goal is not to improve learning. Same way Duolingo sucks, the goal is to optimize global metrics over a massive userbase, not optimize individual metrics, where each individual has its own context.

True with all the apps. Hated this about working with Tinder.

All the conversations, KPIs are "What gets the most users, makes them stay on the app longest, paying the most money". 4 years and not a single conversation about how to improve dating.

Using Anki myself for language learning and can definitely recognize the pitfall of unrelated cues triggering the recall, but so far it's been working quite well. I'm mainly using it to practice output though, so the front side of the card is in English.

On semi-related note, currently making on a language app (gengengo.com) if anyone wants to check it out.

In my limited experience with static cards you really do just end up learning the "appearance" of a card, which is why I've been more successful with dynamic cards, like say a card showing a concordance randomly selected from a large corpus every review, showing the word I want to learn in various contexts.

Just in case someone has no idea what Anki is like me:

Anki (US: /ˈɑːŋki/, UK: /ˈæŋki/; Japanese: [aŋki]) is a free and open-source flashcard program. It uses techniques from cognitive science such as active recall testing and spaced repetition to aid the user in memorization.[4][5] The name comes from the Japanese word for "memorization" (暗記).[6]

The SM-2 algorithm, created for SuperMemo in the late 1980s, has historically formed the basis of the spaced repetition methods employed in the program. Anki's implementation of the algorithm has been modified to allow priorities on cards and to show flashcards in order of their urgency. Anki 23.10+ also has a native implementation of the Free Spaced Repetition Scheduler (FSRS) algorithm, which allows for more optimal spacing of card repetitions.[7]

Anki is content-agnostic, and the cards are presented using HTML and may include text, images, sounds, videos,[8] and LaTeX equations. The decks of cards, along with the user's statistics, are stored in the open SQLite format.

https://en.wikipedia.org/wiki/Anki_(software)

Anki is dead because, after LLMs were introduced, people lost the incentive to learn at all.

Frankly, who wants to learn something the old way when you can ask an LLM to do it for you?

All this reminds me of my home country when the Ponzi schemes were at their highest point (2008). There were small and medium cities, were a sizeable percent of the population actually stopped working at all! Why? Because the ponzi schemes were promising up to 300% returns within 6 months! And the Ponzi schemes were actually delivering (so much that the banks actually pressured the government to intervene). So people put all the money in the ponzis and simply wait a bit to cash in part of their returns. No one opened their small grocery stores, nor small pubs, nor did anyone want to do chores or manual labor. Until everything collapsed.

This is why I built Manabi Reader (currently only for Japanese)

https://reader.manabi.io

It helps you discover reading material suited to your current knowledge. It’s better to acquire new words within colorful contexts, and then use flashcards to review them after learning the material. (It has its own flashcard companion app or you can use Anki.)

Soon I am working on making the activity of reading words in native texts also count as reviewing those words in current and future flashcards, using FSRS. So that you can spend more time reading and not see it as detracting from catching up on your flashcard review workload. And because the reader tracks every word and kanji you come across, it can start to find and suggest the most effective passages to revisit or read for the first time from your personal corpus it currently accumulates from what you load in.

But next I am working on adding manga and video to enhance the fun of it, as OP mentions being important too.

I’ve recently managed to go full-time on this project and hope to bring it to more languages and platforms before long.

> only for Japanese

and only for Apple devices...

Correct

For Android there's Jidoujisho https://github.com/arianneorpilla/jidoujisho.

It's essentially Ttsu reader + Yomitan (née Yomichan) + Ankidroid integration packed into one app. You have to find your own reading material though, but there's plenty of that go around. Jidoujisho has some OCR support for reading manga too, although I never found those tools to be working well enough in any app.

For websites on Android (e.g. news or Aozora Bunko) the same can be achieved with Firefox + Yomitan - https://addons.mozilla.org/en-US/firefox/addon/yomitan/ - as Yomitan has an Ankidroid integration of its own. Or any other browser with extension support, whether desktop or mobile, desktop versions integrate with desktop Anki.

In theory, Yomitan supports other languages as well, although making it work with Finnish took me quite some effort a year ago (maybe it's easier now).

Jidoujisho seems like a great asset for Android

This basically matches my experience with Anki exactly. It's too easy to learn something that's on the front of the card other than the piece of information that you are trying to learn. I've also been thinking about the same solution.

Facts are required for critical thinking. I wonder if the writer eventually transitioned from fact-memorization flashcards, to concept-based questions (how does it make sense that....?).

No, to be honest I haven't had so much practice with those kinds of cards; I've mostly used it for vocabulary and geography.

How does it change things?

After single-fact memorization, concept-driven cards force retrieval of multiple related ideas and apply them to real-world critical thinking.

Examples:

- Cloze: "Swelling of the optic disc is called [c1::papilledema]"

Concept-based:

- Front: "Papilledema (optic disk swelling) - use the pathophysiology to explain what situations you would see this in generally."

- Back: "Anything causing central retinal vein drainage ↓ (e.g., high ICP or central retinal vein occlusion)."

- Extra field: """ Anything causing central retinal vein drainage ↓ (e.g., high ICP or central retinal vein occlusion) → back-up of blood into the capillaries of the optic nerve/disk → hydrostatic pressure ↑ → transudate of fluid leaking across the capillaries → optic nerve/disk swelling (optic disk swelling = papilledema).

Recall that the optic nerve (and thus the optic disk) drains into the central retinal vein, which ultimately drains into the cerebral circulation. """

It looks like you can’t do much on incontextlearning.com without logging in. It would be good to have more explanation of what it does and maybe a demo before taking that step.

Ah sorry about that, I made this video a while ago but I forgot to link it on the home page. https://www.youtube.com/watch?v=ntcWvoZYLec&t=10s&ab_channel...

Thanks for pointing it out!

Can we get a non-Google, username and password login option?

Anki isn’t dead, it’s just LLMs might be a better tool in many ways. Not everything must be all or nothing.

I would go further to say that anki just gets an update where you can now have more dynamic context surrounding the words on cards. Just bold the target word. The same font is also something that can be easily fixed.

I’m confused from your wording, are you saying this is an update they have done or an update they should do?

Update they should do

Ah ok cool

The Japanese immersion language learning rabbit hole goes very very deep, and it super interesting. I'm writing some info about it here because I just don't want it to be lost to the sands of the digital times:

It all started way back when with this guy named Khatzumoto who did his own guerilla academic research, for lack for a better word - and proving it by getting fluent in 18 months. The theories were based in comprehensible input (from Stephen Krashen) and nonstop immersion, which do carry weight and are great ways - if not the only way to learn ANY language. He created AJATT, All Japanese All The Time, and basically created a mini cult that legitimately got people to fluency. It might sound extreme but on the other end I heard of someone who did duolingo (100% slop product btw) for 6 months and didn't know how to say "Thank You" in Japanese. Khatzumoto's ideas were manic, strange, but sometimes truly brilliant. I've never found a blog quite like it ever since - writings that an LLM really can't emulate. The original blog 404'd but it's been revived by a community member here - https://alljapanesealltheti.me/index.html I go back to them every now and then when I want some crazy motivation.

Nobody knows what has happened to Khatzumoto, he basically just dropped off the face of the internet - I wonder if he's doing alright.

> fluent in 18 months

That's a very interesting definition of fluent.

It’s not a definition

If you only surround yourself with a single language for 18 months (literally fully immersive), you can learn any language fluently without even trying. Unless you never leave your apartment.

Can != Will though. I knew someone who lived in Tokyo for 10 years and never learned any Japanese. They worked for Yahoo.

I don't know if the same people are involved, but this is a more comprehensive and up-to-date website built around the same concepts that I came across recently, and where I would suggest anyone starting out go first.

https://tatsumoto.neocities.org

Would've tried, if it didn't require a google account to connect

I get the feeling that an LLM wrote this article as well.

If it did then I’m very impressed because the demanding more snacks thing was fantastic

It could be a combination of some manual tweaks to a generated article.

I strongly agree with the author's claim that Anki incentivizes you toward recognizing the target sentence, but not with his main claim that it's dead. For context, I've been using Anki about 2 years and ~180k reviews.

Grinding the same thing over and over works well for some things but overtrains for others. I've noticed that I'm overtrained on the default sans-serif Japanese fonts, for example, which causes slight but noticeable pain when I switch to a different font - especially a more traditional serif one which tends to have a lot more visual information. Even if you know all the stroke order stuff, suddenly switching to a new font to read hundreds or thousands of characters is a lot of extra input for your visual cortex to process. It bothers me enough that I'm switching things around to match the JLPT exam typefaces as closely as possible so I don't feel it dragging me down under the pressure of a ticking clock and high stakes.

On the other hand I think the Anki engine is more than flexible enough to accommodate dynamic cards, and it still works great for multimedia and the like. I'm leery of handing too much of this over to LLMs both for the hallucination risk the author mentions, and the reliability of the audio generation - pitch and rhythm are bad on LLM-generated audio in English and I suspect they'd be even less reliable in Japanese. A few sentences sound fine but when you start doing it on longer texts or across thousands of sentences it's gonna creep in, and for language learning toxic input is deadly.

IME the best thing once you get past the beginner stage is to read as much as you can. The biggest day-to-day gripe I have with Anki (but which is not actually the fault of the software) is you're learning words and/or sentence patterns in isolation, so the work you do in understanding a sentence is thrown away immediately. You might spend 3 minutes pondering the intricacies of 'if I don't leave the house by 7am, I'll be late for my train', but then the next sentence is 'I savored the sound of my enemies' screams while watching their ship sink.' In contrast to the i+1 competency issue, you're also missing a +1 contextual issue - the little exercise you did to understand the first sentence doesn't carry you forward into the next one, so you're starting from zero every time. Even in the exam situation, you're tested on 4 or 5 questions related to each written or audio subject matter which can help you maintain focus even where there are gaps in your understanding. I think the constant fragmentation of attention by individual sentence cards is bad for concentration and retention. Although Anki allows for sequences of cards to be presented together by the deck editor, this context gets broken up both by your daily card limit and by the fact that review order has little or nothing to do with the initial order of presentation.

By contrast, when you're reading (even if it's just paragraphs or pages) new vocabulary items or grammatical structures tend to repeat and reinforce each other, you can draw contextual inferences about the meaning of unfamiliar words rather than just being given a meaning-token in isolation, and the new material is embedded in some sort of narrative or argumentative context. I've noticed that new vocabulary I pick up from reading beds in far faster than that from isolated word or sentence cards, where everything is isolated and the only context is 'that day's Anki session'. It's like the difference between enjoying a movie and and watching an endless succession of closeups from different scenes. You could work out the story from the jumbled sequence of closeups, but it would take a lot longer and not really be enjoyable or coherent-feeling.

In my view the best changes Anki could make in this regard is to get away from the deck as the main data structure and have an intermediate structure of sets, essentially sub-decks that have something in common and can reinforce each other. It would help a lot if there were better tagging options than just ctrl+1/2/3/4//5/6/7/8 too. I use a few for noting cards I am having particular difficulty with (yellow or red) and a few others for categories, but there just aren't enough of them. I would love to able to pull tag sets of just school stuff one day, just work stuff another, just transit cards another day and so on.

Reply on news.ycombinator.com