You should try working on language learning tools again. The models now are pretty good especially since you don't always need the frontier model to generate good stories. I would love for this project to get solid traction so I could distill my own models that are better at meeting my story creation requirements.
I've been working on a Spanish language learning tool that helps users learn with comprehensible input. It uses LLMs to generate targeted stories for the users and while the user is extensively reading they select any words they don't understand. Those words then get added to a space repetition deck and are used in future stories when it's time to review the words again.
The project has been really fun to work on because of the fun systems I've had to think about. I had to figure out the optimal way to store the user's knowledge of the words. For example you don't need to store the singular and plural of a word in Spanish or even every verb tense, you should probably store the lemma and track those modifiers instead. It's also been a fun challenge to tell the LLM the user's Spanish knowledge without specifically sending every lemma/word the user knows.
I have seen some similar projects out there but a lot of don't seem to focus on creating the perfect story for the user and instead have them choose a CEFR level (A1,B2 etc). Which I think defeats the whole point of using the LLM. With computers we have been able to track a user's knowledge granulary but now we can do the exact same thing but for creating.
I'm really excited to see how far I can take this project. I wanna continue to polish it, but also there is so many details I can continue to add the make the ideal language learning reading app. I just launched the beta last Tuesday, so if you are learning Spanish, I would love if you tried it out and gave feedback.
I’m working on a language learning tool where it uses LLMs to generate stories at your ideal level. The idea is the user would be provided stories that are 95% comprehensible, with the other 5% being a mix of brand new words or words you are still learning. As you read the story you click on words that you still don’t fully understand. I am only working on Spanish right now, since I want to optimize for each language. It’s been fun designing my databases, coming up with calculation ideas, designing story validation, creating an estimation system of a user’s knowledge for onboarding. I know there is some debate about LLMs in language learning and I don’t think they should be trusted to explain grammar but if you validate it’s output it can be such a great tool to learn at your perfect level.
That sounds really interesting. I have a similar project, albiet pretty small. I want to generate comprehsnible input stories for the user say with 98% known words and 2% unknown words. Instead of rewriting stories though, I thought of having compiled list of books with say a book's top 1000 common but unique words, then you can add it to your desk and have those be generated in stories. That way once you complete the deck, it will be a lot easier to read your target book. I was looking into using numPy for that, not sure if you are using Python but it might be worth looking into.
You should try working on language learning tools again. The models now are pretty good especially since you don't always need the frontier model to generate good stories. I would love for this project to get solid traction so I could distill my own models that are better at meeting my story creation requirements.