I think Anki will remain a superior alternative to an LLM for any *fixed* data set one has to learn, such as some medical or legal topics. These *are* memory-dependent. In contrast to Anki, I would not trust an LLM currently to guarantee you have 1) reviewed every single question and 2) keep reliable and consistent statistics on your success and failure rate for each question in the data set. For open-ended topics like a language, an LLM might well prove to better. If you haven’t already, you might want to check out NotebookLM.
I love Andy's work! My friend and I were thinking a lot about Quantum Country and his essays while designing inContext. Same for NotebookLM.
I guess the app I'm describing here is not just an LLM, but rather that it uses an LLM with some non-LLM backend that keeps track of what you know/don't know and where you're struggling. Even for medical and legal topics because of the memorizing rectangles thing!
I think this is a great idea. Even when I just wanted to learn scales on the piano for an exam, I wanted flashcard app to randomly show me: A minor left hand arpeggio 2 octaves. And it was a complete pain to type them all into a flashcard phone app.
With the idea you're suggesting, you can be as vague or specific as you like. So in my car, I could have done a table in excel then pasted it into the prompt window.
Had a go at vibe coding it like you suggested. Claude Code knocked it up for me in an evening!
Ooo nice, for sure you'd want the back of the card to remain the same, but I think there's value in having variety in the front. For example:
> Quick - your meeting starts in 30 seconds, what's your Google password for account X?
> Your phone died, you're on a friend's laptop, need your Google password
> *Show an actual Google login screen mockup*
At the worst, at least you aren't memorizing the card. And at best you are making your knowledge a bit more robust by having to recall it in different "scenarios".
This is super interesting. I'd rephrase the title though via Gramsci: 'The old anki is dying; the new anki has not yet been born'.
I personally think someone needs to build the new anki - I've seen attempts but not ones that I think really hit the nail on the head - making connections between info!
I think the dying/not-yet-born position is probably a more defensible one haha. I agree with the connections, and I think knowledge graphs are the perfect tool for this.
Agreed. Honestly, I’d like to build a tool like that (this is what I’m currently building, which includes a knowledge graph of sorts: https://app.chalklearning.io/tools/curriculum-graph) - It’s a tool for elementary/ primary teachers though.
Great text. LLMs are revolutinary indeed. I tried chatgpt to improve my german and I thought it was great, but of course it lacked a structure that would keep it engaging. This inContext you mentioned solve this. Thank you very much!
happy I found this, I actually went through a similar journal for Mandarin recently. originally I was super focused on flashcards, doing review every day, growing vocabulary, etc. then, I realized I wasn't spending nearly as much time inputting/outputting *actual language* as I was just doing flashcards. I then pivoted to focus primarily on comprehensible input as a way to build the "foundation" of my practice.
I also ended up building a tool to help with this, it's basically like an AI penpal for Mandarin, with the goal being to produce N+1 input while also encouraging the user to produce output & engage with the AI. https://koucai.chat
tried out your tool, think it's a really cool concept! reminds me of standardized tests with the comprehension questions lol. I'm pretty curious how you tune the LLM to produce N+1 content, it's certainly an issue I've worked around a lot as well
As a tutor it is pretty clear when a student doesn't really understand a concept and you can push them on it - in theory LLMs should be able to do this.
In practice they don't seem aggressive enough about it - even when you put them in an agent loop.
When I have tried this myself with chatgpt it basically only works well when I have existing broad level knowledge.
I think this is where having good evaluations and good data is important. Models that have been trained or have access to good data will do much better. For language, we have pretty clear distinctions between levels with CEFR levels, which can help us keep the level of text bounded between A2 and B2, for example. This already works well as shown in the Tarzan to Tolkien study I linked.
I think this also holds for math or physics problems. If an LLM app has a bank of questions it can access, then serving up an i+1 question could be as simple as pulling up a question RAG style, or pulling one up and reasoning through a modification. Of course as you mention with your broad level knowledge comment, this is easier for middle/high school level math but would be much harder for graduate level style math questions. I do think though with good data the ceiling will be pretty high.
Agree and disagree. I too think Anki has some shortfalls, but I'm not sure I'd go as far to say it is dead (or will ever die, the runway is simply too large).
FWIW, I'd genuinely love to get your thoughts on Shaeda.io. I've been working on it (not vibe-coded) for coming up to a year now and hoping to release soon.
The inContext looks nice and promising (although it will be complicated to set up, I would imagine). One small 'concern' I'd have though is that in the screenshot provided there's a lot of text to read: would there be multiple questions/flashcards per text?
But yes on the whole I think flashcards are very underutilised, and with a bit of creativity can be 10x more productive (even though they're already incredibly productive)
First things first: thank you for introducing me to Anki. I have been studying Russian on my own for some time now for many reasons: I have always dreamed of visiting Central Asia, where Russian is a lingua franca. Russian culture itself is very fascinating, not to mention the literature; and finally, I have ADHD, so learning new things is simply one of the many ways—and certainly one of the healthiest—to get my daily dose of dopamine.
But the point is that maybe Anki is already dead, but it is relevant to me. Furthermore, there's also the issue of accessibility: perhaps more functional, LLM-based tools exist (or will exist soon), but if access to such tools is paid or somehow limited, an app like Anki continues and will continue to be more USEFUL, even if perhaps less functional.
Of course, for a professional like you, the difference might be significant. But in a world where even knowledge is commodified, sharing still has the power to resurrect the dead.
Great post. Very informative and thought-provoking. A tiny example of this from my own life that I’m eager to change. I just used the good NameShark app to learn my 140 students’ names before school starts. I receive the pictures taken for their ID cards. My memory connection is tied not only to their mostly-constant face, but a hat, shirt, barrette, hair part, grin, glasses that is missing/different when I walk in the classroom. Surely this is the last time that I’ll have to learn from the static images. Looking forward to dynamic images that are never the same each time—changing clothes, churning accessories, modifying expressions.
Thanks!! So for now it's very crudely implemented but you would definitely want supplemental infra. In this specific context, the idea would be to store the vocab word that the learner failed to understand as well as the original sentence. Then, when a certain amount of time has passed, the LLM would be given this sentence and word and asked to create a new sentence (or story) that uses this word similarly.
So it's more of a context engineering problem and building that supplemental infra. With this sentence I guess I meant that in many contexts, the LLM would somehow write to this supplemental infra what the learner doesn't know, so it can be surfaced later.
Just tried your tool and love it. I’d love to connect and trade notes about this. I spent my twenties thinking about books taxonomy applied to online developer bootcamps and have recently been re exposed to language learning.
All the problems you listed with static anki cards could be solved with a anki add-on that connects to some LLM API on the backend. I don't see how LLM signals the death of anki.
This is an arrogant take. Anki still has a rabid user base and i don't see it going anywhere anytime soon. Maybe in 10+ years.
Arrogant? Relax bro
I think Anki will remain a superior alternative to an LLM for any *fixed* data set one has to learn, such as some medical or legal topics. These *are* memory-dependent. In contrast to Anki, I would not trust an LLM currently to guarantee you have 1) reviewed every single question and 2) keep reliable and consistent statistics on your success and failure rate for each question in the data set. For open-ended topics like a language, an LLM might well prove to better. If you haven’t already, you might want to check out NotebookLM.
Just remembered one more thing: If you haven’t looked into Andy Matuschak’s work, you may appreciate it.
I love Andy's work! My friend and I were thinking a lot about Quantum Country and his essays while designing inContext. Same for NotebookLM.
I guess the app I'm describing here is not just an LLM, but rather that it uses an LLM with some non-LLM backend that keeps track of what you know/don't know and where you're struggling. Even for medical and legal topics because of the memorizing rectangles thing!
I think this is a great idea. Even when I just wanted to learn scales on the piano for an exam, I wanted flashcard app to randomly show me: A minor left hand arpeggio 2 octaves. And it was a complete pain to type them all into a flashcard phone app.
With the idea you're suggesting, you can be as vague or specific as you like. So in my car, I could have done a table in excel then pasted it into the prompt window.
Had a go at vibe coding it like you suggested. Claude Code knocked it up for me in an evening!
https://claude-flashcards-frontend.vercel.app/
My *case*. Not my car.
> I challenge you to think of a static card that would not be improved by making it a dynamic card.
A card made to remember a password?
Ooo nice, for sure you'd want the back of the card to remain the same, but I think there's value in having variety in the front. For example:
> Quick - your meeting starts in 30 seconds, what's your Google password for account X?
> Your phone died, you're on a friend's laptop, need your Google password
> *Show an actual Google login screen mockup*
At the worst, at least you aren't memorizing the card. And at best you are making your knowledge a bit more robust by having to recall it in different "scenarios".
This is super interesting. I'd rephrase the title though via Gramsci: 'The old anki is dying; the new anki has not yet been born'.
I personally think someone needs to build the new anki - I've seen attempts but not ones that I think really hit the nail on the head - making connections between info!
I think the dying/not-yet-born position is probably a more defensible one haha. I agree with the connections, and I think knowledge graphs are the perfect tool for this.
Agreed. Honestly, I’d like to build a tool like that (this is what I’m currently building, which includes a knowledge graph of sorts: https://app.chalklearning.io/tools/curriculum-graph) - It’s a tool for elementary/ primary teachers though.
Great text. LLMs are revolutinary indeed. I tried chatgpt to improve my german and I thought it was great, but of course it lacked a structure that would keep it engaging. This inContext you mentioned solve this. Thank you very much!
happy I found this, I actually went through a similar journal for Mandarin recently. originally I was super focused on flashcards, doing review every day, growing vocabulary, etc. then, I realized I wasn't spending nearly as much time inputting/outputting *actual language* as I was just doing flashcards. I then pivoted to focus primarily on comprehensible input as a way to build the "foundation" of my practice.
I also ended up building a tool to help with this, it's basically like an AI penpal for Mandarin, with the goal being to produce N+1 input while also encouraging the user to produce output & engage with the AI. https://koucai.chat
tried out your tool, think it's a really cool concept! reminds me of standardized tests with the comprehension questions lol. I'm pretty curious how you tune the LLM to produce N+1 content, it's certainly an issue I've worked around a lot as well
How do you reliably generate i+1 content?
As a tutor it is pretty clear when a student doesn't really understand a concept and you can push them on it - in theory LLMs should be able to do this.
In practice they don't seem aggressive enough about it - even when you put them in an agent loop.
When I have tried this myself with chatgpt it basically only works well when I have existing broad level knowledge.
I think this is where having good evaluations and good data is important. Models that have been trained or have access to good data will do much better. For language, we have pretty clear distinctions between levels with CEFR levels, which can help us keep the level of text bounded between A2 and B2, for example. This already works well as shown in the Tarzan to Tolkien study I linked.
I think this also holds for math or physics problems. If an LLM app has a bank of questions it can access, then serving up an i+1 question could be as simple as pulling up a question RAG style, or pulling one up and reasoning through a modification. Of course as you mention with your broad level knowledge comment, this is easier for middle/high school level math but would be much harder for graduate level style math questions. I do think though with good data the ceiling will be pretty high.
Agree and disagree. I too think Anki has some shortfalls, but I'm not sure I'd go as far to say it is dead (or will ever die, the runway is simply too large).
FWIW, I'd genuinely love to get your thoughts on Shaeda.io. I've been working on it (not vibe-coded) for coming up to a year now and hoping to release soon.
The inContext looks nice and promising (although it will be complicated to set up, I would imagine). One small 'concern' I'd have though is that in the screenshot provided there's a lot of text to read: would there be multiple questions/flashcards per text?
But yes on the whole I think flashcards are very underutilised, and with a bit of creativity can be 10x more productive (even though they're already incredibly productive)
First things first: thank you for introducing me to Anki. I have been studying Russian on my own for some time now for many reasons: I have always dreamed of visiting Central Asia, where Russian is a lingua franca. Russian culture itself is very fascinating, not to mention the literature; and finally, I have ADHD, so learning new things is simply one of the many ways—and certainly one of the healthiest—to get my daily dose of dopamine.
But the point is that maybe Anki is already dead, but it is relevant to me. Furthermore, there's also the issue of accessibility: perhaps more functional, LLM-based tools exist (or will exist soon), but if access to such tools is paid or somehow limited, an app like Anki continues and will continue to be more USEFUL, even if perhaps less functional.
Of course, for a professional like you, the difference might be significant. But in a world where even knowledge is commodified, sharing still has the power to resurrect the dead.
Great post. Very informative and thought-provoking. A tiny example of this from my own life that I’m eager to change. I just used the good NameShark app to learn my 140 students’ names before school starts. I receive the pictures taken for their ID cards. My memory connection is tied not only to their mostly-constant face, but a hat, shirt, barrette, hair part, grin, glasses that is missing/different when I walk in the classroom. Surely this is the last time that I’ll have to learn from the static images. Looking forward to dynamic images that are never the same each time—changing clothes, churning accessories, modifying expressions.
I've never heard of NameShark, but this example perfectly illustrates the memorizing rectangles thing. Thanks for sharing!
“LLMs can track what you know, store it, surface it when you're about to forget it, and wrap it in fresh context every time”
Extremely interested to hear how you implemented this. As far as i know LLMs can not do this without supplemental infra.
(loved this post by the way!)
Thanks!! So for now it's very crudely implemented but you would definitely want supplemental infra. In this specific context, the idea would be to store the vocab word that the learner failed to understand as well as the original sentence. Then, when a certain amount of time has passed, the LLM would be given this sentence and word and asked to create a new sentence (or story) that uses this word similarly.
So it's more of a context engineering problem and building that supplemental infra. With this sentence I guess I meant that in many contexts, the LLM would somehow write to this supplemental infra what the learner doesn't know, so it can be surfaced later.
Just tried your tool and love it. I’d love to connect and trade notes about this. I spent my twenties thinking about books taxonomy applied to online developer bootcamps and have recently been re exposed to language learning.
dm'd you!
Fyi bloom's two sigma is a large overestimate although tutoring is still more effective: Was the famous 'The 2 Sigma Problem' study correct?
https://www.linkedin.com/pulse/famous-2-sigma-problem-study-correct-vassili-philippov-avjme?utm_source=share&utm_medium=member_android&utm_campaign=share_via
Interesting idea. I built a CLI/MCP server because I was feeling the pain of knowledge decay. https://github.com/aisflat439/anki-cli
I like this idea that I could try to have the cards be more dynamic. Just studying the concepts... Huh. Anyway. Thanks for the article
Who is the artist of the Ibiza painting?
I took the picture and asked an AI model to recreate it in the style of hiroshi yoshida, a japanese ukiyo-e artist.
All the problems you listed with static anki cards could be solved with a anki add-on that connects to some LLM API on the backend. I don't see how LLM signals the death of anki.