Hacker News Reader

Comment by trane_project | Hacker News Reader

I've been thinking of using GPT or similar LLMs to extract flashcards to use with my spaced repetition project (https://github.com/trane-project/trane/). As in you give it a book and it creates the flashcards for you and the dependencies between the lessons.

I played around with chatgpt and it worked pretty well. I have a lot of other things in my plate to get around first (including starting a math curriculum) but it's definitely an exciting direction.

I think LLMs and AI are not anywhere near actual intelligence (chatgpt can spout a lot of good sounding nonsense ATM), but the semantic analysis they can do is by itself very useful.

View on HN

I've seen a number of projects around using GPT to generate curriculum and also flashcards in the past three months, I think this is one of the most popular ones: https://autolearnify.com

It's a very good idea in theory but takes almost as much work to verify that the flashcards and curriculum that it generates is accurate and not a hallucinogenic nightmare.

The biggest danger is that the target audience are not experts in the desired subject domain, so they have no way of sanity checking the generated curriculum.

When I played with it, I made it output a JSON file so that it would be easier to handle the output. And I specifically gave it the text to use. It did a pretty good job, but I ran into output size limits.

I agree that using the training data would probably generate more garbage. But it's the semantic analysis part that I think it's useful. In general, I think VCs and OpenAI are overhyping it by calling it "intelligent" and obscuring the very good use cases of the technology. AFAIK, no one involved has explained how a statistical model running on a Turing machine magically develops agency and awareness, which are requirements for actual intelligence (under my definition, at least).

Curious as to where you came to this impression?

I'd say the most popular applications are Knowt in the US for now, and Saveall.ai + Revision.ai (my company) in the UK, all been around with BERT/T5 etc long before this GPT trend.

The flashcard accuracy varies wildly amongst current solutions, that's for sure.

Every now and again I come across a real gem on here. Love trane and the idea behind it, can't wait to play with it after work.

Thanks. I am working on simplifying how new material is added (for simple cases, editing a single JSON file and running a build command will be enough to build all the exercises). That should make it easier to create more stuff for it, which so far has been my bottleneck.

For now, this is the most general way to create the exercises: https://trane-project.github.io/generated_courses/knowledge_.... The JSON file thingy is just a script that automates creating these files given the specification, so they will be interchangeable.

Reply on news.ycombinator.com