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
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.
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.
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 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.
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.
When I saw a great Show HN post yesterday about a language learning with real-world content, I started to sweat a bit; the headline matched what I had been working on with a friend! But it turns out that we went in totally different directions, and I figured it was as good a time as any to share.
Here's our interpretation of the Duo-style experience with real-world content: the inContext notebook. I like to think of it as lowering the activation energy of reading native-level content in a foreign language. The user reads a short passage that's roughly at their level and then they answer content questions and review vocab from the passage. As for what the topic of the passage is... you choose! Provide a link or text and bam.
In my experience with language learning, one of the most annoying phases is when you know some stuff but you still have to keep grinding the textbook because all the real content is too difficult. Hopefully this makes that period a bit more interesting, and gets the learner there faster.
Some notes on usage after testing for a few weeks:
- if you are not sure where to find articles, try Google Discover.
- this technically works in translation as well but for best results use a source written in the target language.
Technical things:
- after quite a bit of testing and iterating, we found Claude 3.5 Sonnet works best at naturalness of text, quality of story, question quality, etc.
- in the future we'd like to try use a fine tuned or trained model but so far we don't have anything that works better than this.