Note on why this person is taking an unusual route: from https://blog.kevinzwu.com/cyborg-learning/ , they are a "second generation Chinese immigrant" and "heritage speaker"; that is, they live outside China, can speak the language because they learned it from their parents, but cannot read it.
Edit addendum: https://blog.kevinzwu.com/chinese-cursed-logographic-dags/ is a fun read. I've been using the imaginatively titled "kanji study" app, which uses the same Outlier database mentioned which has the graph based etymology.
There's an additional level of chaos when learning the "same" characters as kanji rather than hanzi.
I liked the 10% @@@ example, demonstrated their point pretty well.
Also for anyone who speaks or is currently learning Chinese... I've been working on a multiplayer CJK word game that shares a similar efficient brute force style of learning to the author's approach (although presented via gameplay instead of tooling). Every turn you get a random character and must type in a word that contains the char in ANY position. If you like fast paced word games it might be up your alley: https://danobang.com/?game_lang=cmn
> I would end up copy-pasting interesting words into the dictionary window to pull up the word entry. SLOW!
> I would then click on the component characters to open their nested dictionary entries. SLOW!
> If I needed to remember the stroke order, I would scroll down for the static display. SLOW!
So, all of these are included in Anki-xiehanzi(https://github.com/krmanik/Anki-xiehanzi). Free open source software like Anki & xiehanzi can save you from using all those tokens.
It is also allowing me to see all relevant associations easily when revealing the card in built in SRS, you add cards to SRS as you browse, so they are related to what you already know / currently exploring.
Mind you, all data visible is collected from different reputable available sources. When you click "explain" there's a clearly marked LLM explanation, but my explanation generation pipeline pushed all generated explanations through 5 different models including all top Chinese-first for verification, and on average it took a few iterations back and forth to iron out any information that could potentially mislead the learner.
> I opened Claude Code and started rambling into my mic. It wrote thousands of lines of questionably efficient JavaScript. I didn't read a single one.
Hm. I always knew voice mode was a thing, but I have never tried it. What's people's experience with it?
Being able to correct my words is a good thing. Hell, I did it ~3 times when writing this comment. I can't do that when I'm rambling. I'll trip, or CC will think I'm finished.
I maintain an Anki deck for my chinese learning. Following the HSK books, I add new words to my deck with the character on front side and pinyin + definition + audio (from the CD and sliced using Audacity) on back side.
WIP (need more work in multi-hanzi words), but won't stay in the same 5 words for more than a day. it has been working well for me
the most interesting thing was GPT helped with the sentences and simplified words meaning and bing translate provided the audios
the goal is get the ~2000 words you need to be proficient in 1 year, 5 words a day plus refresh old words, also it keep track of your progress against the year, no streaks
I very much agreed to the "build your own tools for learning" route they took. I made my own site to practice Japanese grammar as it was hard to get feedback on that. tokidoki.meertens.dev -> it's open source so anyone can adapt it to their needs as well!
I can speak Chinese fluently but I need to improve my reading hugely. This sounds like exactly my usecase. I would actually be willing to pay for this (though less certainly for a subscription, preferably Pleco-like one-time fee - even if larger).
Interesting process. I wonder if he considered doing this with Anki. That would have given him a good SRS algo for free and Anki cards are also HTML+CSS+JS. I probably wouldn't try to put LLM calls onto my cards though
> I decided to go against the grain of the near-universal advice to "learn to read by reading".
...Why? That advice is universal for a reason. The side adventure with Claude Code strikes me as a distraction from the fact that there is a hard thing you want to do but are avoiding because it's hard.
39 comments
Edit addendum: https://blog.kevinzwu.com/chinese-cursed-logographic-dags/ is a fun read. I've been using the imaginatively titled "kanji study" app, which uses the same Outlier database mentioned which has the graph based etymology.
There's an additional level of chaos when learning the "same" characters as kanji rather than hanzi.
Also for anyone who speaks or is currently learning Chinese... I've been working on a multiplayer CJK word game that shares a similar efficient brute force style of learning to the author's approach (although presented via gameplay instead of tooling). Every turn you get a random character and must type in a word that contains the char in ANY position. If you like fast paced word games it might be up your alley: https://danobang.com/?game_lang=cmn
I remember learning about your game in a mandarin forum. Thanks for making it, it's very cool!
> I would end up copy-pasting interesting words into the dictionary window to pull up the word entry. SLOW!
> I would then click on the component characters to open their nested dictionary entries. SLOW!
> If I needed to remember the stroke order, I would scroll down for the static display. SLOW!
So, all of these are included in Anki-xiehanzi(https://github.com/krmanik/Anki-xiehanzi). Free open source software like Anki & xiehanzi can save you from using all those tokens.
It is also allowing me to see all relevant associations easily when revealing the card in built in SRS, you add cards to SRS as you browse, so they are related to what you already know / currently exploring.
Mind you, all data visible is collected from different reputable available sources. When you click "explain" there's a clearly marked LLM explanation, but my explanation generation pipeline pushed all generated explanations through 5 different models including all top Chinese-first for verification, and on average it took a few iterations back and forth to iron out any information that could potentially mislead the learner.
You can actually see thousands of words I typed just working on that pipeline here https://hanzirama.com/making-of
> I opened Claude Code and started rambling into my mic. It wrote thousands of lines of questionably efficient JavaScript. I didn't read a single one.
Hm. I always knew voice mode was a thing, but I have never tried it. What's people's experience with it?
Being able to correct my words is a good thing. Hell, I did it ~3 times when writing this comment. I can't do that when I'm rambling. I'll trip, or CC will think I'm finished.
> A guy on a forum had hired a calligrapher to write three thousand characters in ballpoint pen
A shame that this amazing resource is not linked.
https://dondeng.com
WIP (need more work in multi-hanzi words), but won't stay in the same 5 words for more than a day. it has been working well for me
the most interesting thing was GPT helped with the sentences and simplified words meaning and bing translate provided the audios
the goal is get the ~2000 words you need to be proficient in 1 year, 5 words a day plus refresh old words, also it keep track of your progress against the year, no streaks
https://wq-landing.netlify.app/
> I decided to go against the grain of the near-universal advice to "learn to read by reading".
...Why? That advice is universal for a reason. The side adventure with Claude Code strikes me as a distraction from the fact that there is a hard thing you want to do but are avoiding because it's hard.