Hacker News new | past | comments | ask | show | jobs | submit login
Show HN: Using GPT-3 style models to Remember Everything (saveall.ai)
11 points by petros786 on March 3, 2021 | hide | past | favorite | 5 comments



Hi HN,

We generally forget 95%+ of what we learn [1]. At Save All though we believe it doesn’t need to be like this and our mission is to help people remember everything.

After telling us what you want to remember, we use GPT-3 like technologies to generate multiple choice quizzes that test you in the right way at the right time to get the information into your long-term memory and keep it there.

The product uses spaced repetition [2] and has some similarities with anki and the main following differences:

- We will ask you about your cards in many different ways rather than only 1 way (this is better for memory)

- Question types are designed so that we can verify whether you got it right or not without you needing to tell us

- This in-built “verifiability” allows us to make remembering everything feel a lot more like a game and less like a chore

- It’s easier to create cards, just type in statements you want to remember and press enter, no need to create both the front & back of cards

Try it out at https://www.saveall.ai and watch this video to learn more: https://www.youtube.com/watch?v=nTYVRfJXKVs

We’re just getting started and are here for feedback so please also comment below or email us (petros@saveall.ai) with any thoughts on how we can improve the product

[1] https://en.wikipedia.org/wiki/Forgetting_curve

[2] https://en.wikipedia.org/wiki/Spaced_repetition


This is really cool, but also really slow - I twice got the "a script on this page is unresponsive" popup on Firefox. That said, still very exciting.

E: Config, if you need it to debug - Firefox 85 on MacOS Big Sur 11.0.1, MBP 15" 2018 w/ i7 & 16GB of RAM.


Thanks for your comments and the config details, sorry about it being slow! we will investigate...

appears to be fine at moment on other browser / OS combinations


Very interesting, would likely be useful for general knowledge.

However, when you are learning a very narrow field of study, the app does not create great suggestions. A complicated example, but see the following:

"Choroidemia patients have a normal electroretinogram"

presents: MRI, Melee, Moon, DNA.

Of course, I don't expect it to come up with something amazing for something that GPT-3 or other language models aren't trained on.


Thank you for your comment, it's a great point, the suggestions won't be as good for narrow fields. Over time we believe we can make them good for all fields but at the moment it works less well for narrow fields.

Note though that even with poor suggestions there is still a large positive impact on memory from answering the questions, so it still works even for questions where the suggestions are not so good (its just not quite as efficient).




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: