Welcome to my devblog for Flashbang, an LLM-powered, spaced repetition-first learning platform for online courses.
I built Flashbang for nights & weekends season 4, which is rapidly coming to a close.
This Substack will be a way for me to post about app updates and also share my thought processes as I design the software.
In this approach I drew inspiration from Gordon Brander’s project Subconscious—a note-taking protocol/app that came with a Substack where Gordon would write thoughtfully about his process. It’s a nice way to do it, I think.
For now Flashbang is pretty simple. You watch a free online course, do a little interactive research and review, then study the material long term with GPT-4-generated spaced repetition flashcards.
But I hope it serves as a foundation to experiment with new ways of learning and studying in a world of abundant machine intelligence.
If you buy a subscription to this Substack, that will encourage me to keep working on Flashbang and seeing where I can take it.
If I can at least make enough to break even on my GPT-4 API bills, I’ll keep going.
Flashbang’s short-term roadmap: LUI, dopamine, and synthesis
Once this demo day sprint wraps up, I’ll start tackling some of the backlog.
Top on the list of next steps is designing a language use interface to replace the peek-a-boo method we’re using now.
Currently, the SRS cards quiz the user by asking them to imagine the correct answer, then clicking ‘show’ to find out if they were correct.
While it’s great to picture the whole answer manually in your head, it’d be even better to express the answer to the computer so the instruction could adapt itself to it.
So one of my first goals post-n&w is to convert this show/hide button into a text input to the language model—which the folks at Full Stack Deep Learning have been calling a Language User Interface (LUI).
Something like this:
You jot down what you remember of the answer in whatever fragments you can.
Then, when you submit, the LLM is prompted with the question, the right answer, and your attempted answer. It would then be instructed to give you a grade based on how much of your answer aligned with the actual answer.
What would be really cool is if the LLM could also output a suggested grade out of the buttons available for the user. For example, suggesting the grade of “Barely remembered” below.
Another side effect of this could be saving these user submissions to document progress in learning over time.
I have high hopes for the Dashboard section (flashbang.school/dash when you’re logged in) as a way to produce some cool progress metrics in this way. Imagine being able to show your answers for months of recall practice, where you typed out the answer each time. You’d see undeniable improvements.
Other things I want to explore next include:
designing a similar LUI system for the active review section on the lesson page. They could be concatenated at the end and analyzed for the learner at the lesson finish, too. The more personalized interactions, the better.
There are of course many UI dopamine dispensers that go into an effective edtech tool. Flashbang needs its Duolingo-izing if it is to succeed in retaining people throughout an entire learning adventure. We’ll need some silly recurring characters.
After working through an entire module, I’d love it if users could perform a “synthesis” exercise with all the atoms from each playlist. This could be a cool way to cross-reference different sources and get creative—i.e., the similarity search that powers the Q&A chat could select atoms from across each course and mix them together to come up with new interactions.
I’d like to swap out the current supermemo library we’re using for FSRS, a really exciting new algorithm that trains a machine learning model on your own personalized memory data.
For now, though, I look forward to recording my video for demo day and heading to SF for the IRL event next weekend. It has been 7 years since I was last in San Francisco.
Next time I think I’m going to do a postmortem on nights & weekends and talk about what I learned, which was an insanely large amount—buildspace is really the wave of the future and I had an incredible time.
If there’s anything about the app you’d like to talk about, please do leave it in the comments. I love to talk about this stuff.
Joe