Show HN: Prepform – AI and spaced-repetition to optimize learning
prepform.comHi, I'm Eric and I'm the founder and lead developer of Prepform.
A high-quality education helped me pursue my interests and achieve my goals. I started Prepform so students of all backgrounds have access to the same kind of education.
I grew up in Southern California, surrounded by dozens of SAT prep programs, and I swear I must have gone to all of them. Different programs followed different styles and techniques, but the strategy they shared was to create a study plan and review mistakes.
A study plan is taking a diagnostic test,
setting a target score,
creating a study schedule,
identifying mistakes, and finally
reviewing those mistakes.
I wanted to take this structure and optimize it with machine learning, while accounting for elements of human learning and memory. I'm a big fan of SuperMemo, a memorization technique developed by Piotr Wozniak, where you review material just as you're about to forget it. Cognitive psychology tells us human forgetting follows a pattern, but Piotr quantified this behavior to identify the precise moment forgetting happens.
The goal was to build on his research with AI and tailor it to not only test prep but to the individual student, and make it the engine of the study plan.
The result is Blended Prep, which guides students to internalize knowledge rather than memorize material, and gives them the best chance to ace their next exam.
I'm so excited to share this with the HN community, and would love to know what you think. You can try it out at https://prepform.com. Thanks for reading. The concept sounds interesting. My kids use DuoLingo and Khan Academy for learning and their methods seem effective. Would you consider yourself a “DuoLingo for exams”, or is your learning philosophy different enough that that’s not accurate? Some feedback on the site: 1) I’d suggest replace “Remember everything” with something less absolute. I don’t that’s a realistic goal and had me doubting the rest of what I was reading. Humans forget sometimes and that’s OK. 2) A lot of words were underlined. I thought they were links. Good luck with your business! Thanks! My learning philosophy is that the bulk of learning happens when you review material. I want students to see the topics they missed, and answer 3 questions about their mistakes:
"What did I do wrong?",
"What should I have done instead?", and
"What is the clue in this question that I missed?". This "mistake log" will use spaced-repetition to show their mistake just as they are about to forget it. I think this is a huge missed opportunity for companies like Duolingo and Khan Academy. My goal is to guide students through a study plan, so they know what to study and for how long. I appreciate the feedback! Hey I've been playing with your app, I'm very excited about it! But, it seems like the site is pretty buggy, having just tried it on a few different browsers. Specifically, the test review doesn't show relevant information that was available while taking the exam, (i.e. the constants are missing, not even whitespace where they should be). I could share screenshots of what I'm seeing with you if you like. Also, it keeps saying I have 20 tasks for today- I click and it has me take another exam, and afterward, I still have 20 tasks to do. That said I'm getting a lot out of taking these exams and I'm very excited about this project! Hey I'm really glad to hear that! I'll look in to why the tasks aren't updating, but I'm having trouble recreating the missing constants in test review. I'd really appreciate it if you can send me a screenshot of what you're seeing! Again, thanks so much for checking it out. Actually, no need for the screenshots. I found the bugs. Thanks again! https://news.ycombinator.com/item?id=28309645
https://westurner.github.io/hnlog/#comment-28309645 : Re: Phonemic awareness and Phonological awareness, > What are some of the more evidence-based (?) (early literacy,) reading curricula? OTOH: LETRS, Heggerty, PAL: > Which traversals of a curriculum graph are optimal or sufficient? > You can add https://schema.org/about and https://schema.org/educationalAlignment Linked Data to your [#OER] curriculum resources to increase discoverability, reusability. https://news.ycombinator.com/item?id=24527589
https://westurner.github.io/hnlog/#comment-24527589 : >> A bottom-up (topologically sorted) computer science curriculum (a depth-first traversal of a Thing graph) ontology would be a great teaching resource. >> One could start with e.g. "Outline of Computer Science", add concept dependency edges, and then topologically (and alphabetically or chronologically) sort. >> https://en.wikipedia.org/wiki/Outline_of_computer_science >> There are many potential starting points and traversals toward specialization for such a curriculum graph of schema:Things/skos:Concepts with URIs. > https://westurner.github.io/hnlog/ ... Ctrl-F "interview", "curriculum" OpenBadges as Blockcerts for Q12 competencies Some feedback: 1. Good that you can try it out without signing in. 2. Bad that the screen where you select the things is so complicated. 3. Bad that the maximum # questions is 0, even though I've chosen options that have questions (see screenshot) I see it happens when I check the "High" difficulty. I can't tell the difference between radio buttons and checkboxes (which should be square). Agreed. It's nice to have fresh eyes looking at the site. Thanks! Thanks for taking the time for the screenshot and highlighting! I'll look into the bug. It would be nice to be able to see pricing too. I'll make it clearer that it's free! The business model is B2B, and the B2C part is to show off the tech. the website and the idea look great overall.
But compared to Kahn academy I miss that each test (set of question) is not focused on one single new concept. When using Kahn academy, the sequencing is basically 1- watch video about 1 new concept you do not already know
2- try to do a bunch of question about that new concept to make sure you master it
3- if not go back to step #1 but if you master it (progress to the next concept) It feels more like a video game where you unlock level sequentially and it is more fun. Thanks! That's a great point. I hadn't thought about how guiding students through a curriculum could also be a way to gamify the experience. I agree that a more structured path, at least through the first pass, would be more helpful for students. I appreciate the feedback! I love spaced repetition apps (I’ve used Anki and Zorbi in the past). This looks pretty cool! What’s the purpose of AI here? Thanks! The purpose of AI here is to create a student model to personalize learning, increase students' motivation, and increase their speed of learning. It’s hard for students to know what and for how long to study, and when they’ve mastered something. By analyzing a student’s performance, we estimate their current knowledge, and use it to determine factors such as the best repeat frequency, the type of question, and the difficulty of the question. For example, studies show that students stop studying when they don’t fully know the material, and they don’t study material they think they’ve already learned. By guiding them through a study plan and using spaced-repetition, we can improve the speed of learning and reduce the effects of forgetting. Our model accounts for forgetting, guessing, the order of answers, and a student’s baseline knowledge to predict their current knowledge. While we are currently training the model for the NY Regents exam, we have tested it on EdNet, the largest publicly-available education dataset available, and it has the highest predictive accuracy among competing models, with an AUC of 0.7892. What is the role of AI here? The role of AI here is to create a student model to personalize learning, increase students' motivation, and increase their speed of learning. It’s hard for students to know what and for how long to study, and when they’ve mastered something. By analyzing a student’s performance, we estimate their current knowledge, and use it to determine factors such as the best repeat frequency, the type of question, and the difficulty of the question. For example, studies show that students stop studying when they don’t fully know the material, and they don’t study material they think they’ve already learned. By guiding them through a study plan and using spaced-repetition, we can improve the speed of learning and reduce the effects of forgetting. Our model accounts for forgetting, guessing, the order of answers, and a student’s baseline knowledge to predict their current knowledge. While we are currently training the model for the NY Regents exam, we have tested it on EdNet, the largest publicly-available education dataset available, and it has the highest predictive accuracy among competing models, with an AUC of 0.7892.
