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.
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:
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