Hi HN!
This is a tool for learners-in-a-hurry, built to make it easier to engage with learning. It’s motivated by the Testing Effect, which shows studying by reading is not as good as using quizzes/flashcards (grade difference ~20%).
Most classes are still reading or watching dominant. This changes that: Put in your reading-heavy classes topics as text, and get interactive-heavy quizzes. It isn’t suitable for medical, coding or other action-based domains presently; but it hits analytical topics like Psychology, Business & Neuroscience on the head.
Technically, the architecture uses transformer models, a mix of masked language models(answer-selection), finetuned text-to-text models (question phrasing/paraphrasing - you can find some on huggingface), and autoregressive models(generate pairings, topic relation analysis via TF etc). We also rely on RAG, lecture data, and postprocessing rules-based systems.
With stable progress over the last few years, we just added a feature to save in Anki-compatible format, and have ensured you can now trial 3 quiz generations for free – and the code HN10 will unlock 7 more if you hit that this month.
Overgenerality and factual accuracy (recall) are our greatest challenges - we found you really have to be strategic with using on-the-fly augmented data to hit both in the sweet spot. Let me know if you'd like to talk about that in the comments.
1 comment
[ 0.24 ms ] story [ 10.0 ms ] threadMost classes are still reading or watching dominant. This changes that: Put in your reading-heavy classes topics as text, and get interactive-heavy quizzes. It isn’t suitable for medical, coding or other action-based domains presently; but it hits analytical topics like Psychology, Business & Neuroscience on the head.
Technically, the architecture uses transformer models, a mix of masked language models(answer-selection), finetuned text-to-text models (question phrasing/paraphrasing - you can find some on huggingface), and autoregressive models(generate pairings, topic relation analysis via TF etc). We also rely on RAG, lecture data, and postprocessing rules-based systems.
With stable progress over the last few years, we just added a feature to save in Anki-compatible format, and have ensured you can now trial 3 quiz generations for free – and the code HN10 will unlock 7 more if you hit that this month.
Overgenerality and factual accuracy (recall) are our greatest challenges - we found you really have to be strategic with using on-the-fly augmented data to hit both in the sweet spot. Let me know if you'd like to talk about that in the comments.