▲ 204 points
back
203 comments
Wozniak estimates that you actually get a maximum of~300,000 based on his decades of data, so it seems the simplified model misses some effect and so converges to the wrong total: https://supermemo.guru/wiki/How_much_knowledge_can_human_bra...
Wozniak is incredible, it saddens me that SuperMemo will probably drift into obscurity due to being closed source and the chances of seeing its source code drift away day by day, it took a few years to get anything beyond IE support after that died.
I wish to see a FOSS incremental reading application in the very least, such an underrated concept.
Wow, never heard of SuperMemo and this Wozniak. What an interesting and awesome rabbit hole
As a colleague of mine used to phrase it, "yeah, I could learn that new thing but then I'd have to give up an episode of Gilligan's Island" ...
But this estimate is for "free learning"?
I have been wondering about this in the context of being ready for work in the age of LLM‘s. What nobody can deny is that they memorize information at a superhuman level, so it might reduce the value of having done that myself. On the other hand, “couldn’t you just google that“has been an erroneous retort to the value of space repetition for decades, during which I’ve gotten a lot of value out of doing. has been a erroneous retort to the value of space repetition for decades, during which I’ve gotten a lot of value out of doing it.
In the same way that notation is a tool of thought [1] actually knowing facts paves routes[2] of possibility and impossibility (instead of pausing everything to research or drowning in too many paths).
It's akin to seeing 37*55 and knowing it won't be a 20 digit number, greatly reducing your search space. Imagine feeding someone without knowing what they like (how do you choose something?) vs. "allergic to dairy"?
[1] https://www.jsoftware.com/papers/tot.htm [2] https://alexalejandre.com/programming/forge-deeper-thought/
It seems pretty easy to illuminate this question once you consider concrete examples.
For example, neither google nor an LLM replaces the utility of building your own vocabulary so that you can express yourself in a foreign language when you're in convo with someone.
A software-based spaced repetition system probably could have made someone appear borderline superhuman in the era before Google. Google cleared a lot of this low-hanging fruit, and then it became "merely" the best way to cram facts into your noggin so you can become an expert at whatever you want, which is still extremely valuable. Your doctor probably used SRS to get through medical school, and now look at how much money they're making. For most professions, nothing -> Google was probably a much bigger jump than Google -> LLMs, so I expect that to still largely be the case.
TL;DR: SRS is, perhaps counterintuitively, most useful for those of us seeking true mastery over something.
> during which I’ve gotten a lot of value out of doing. has been a erroneous retort to the value of space repetition for decades, during which I’ve gotten a lot of value out of doing it.
I imagine spaced repetition is not actually a cause of repeating yourself. That's just caused by being too lazy to proofread before you post.
LLMs memorize information at a superhuman level the same way an 8 tb harddisk does.
Silly.
What you remember becomes part of your thinking and reasoning. So no, why would you google something if it doesn’t come to mind as something relevant.
FSRS is a really clever iteration on Anki's SM2 algorithm. I've been playing around with various spaced repetition implementations lately - there's a great collection of open source libraries at github.com/open-spaced-repetition. I ended up using their typescript bindings for a vocabulary builder I'm working on (deft.so) and it's been surprisingly straightforward to implement.
I found FSRS is not great for languages because it optimized for the wrong thing, ie. how likely you will recall a word after X days without encountering that word.
There are 2 big problems:
1. If you are studying seriously, you will encounter many words outside your SRS practice, so the FSRS prediction fails flat.
2. More importantly, language learning is not about recalling words.
The algorithm I use for my app has a different set of objectives, it's <30 lines of straightforward code, and it works better (I tested it for months to be sure about it)
I'll open source it soon
I used the same binding for a kanji learning app (https://shodoku.app).
I had some problems with their documentation though. Following their examples was pretty straightforward, but I always had trouble e.g. what is the difference between State.Learning and State.Relearning, or what does fuzzy mean in FSRS parameters, or reps on a Card. You have to either read the algorithm explainer or the Anki source code to understand that.
Besides LineByLine (https://linebyline.app), does anyone have tips on memorizing long texts using spaced repetition?
SuperMemo 19 (2024 - Desktop/Windows) has incremental reading support
https://super-memory.com/help/read.htm
Disclaimer: AnkiDroid maintainer
I don't how well it works in practice for long texts, but for song lyrics I've used this Anki plugin: https://ankiweb.net/shared/info/2084557901
It takes a text, and creates a bunch of cloze deletion cards, each presenting you with a number of lines, and asking you to recall the next line.
I guess it matters how long it is and if you need to know it exactly, and how long you need to retain it. I had to memorize 50 to 100 word phrases and I chained them together with index cards. The first sentence is the cue for the second, the second for the third, etc. Carried them around in my pocket in order and started from first a few times a day until I could do the whole thing. But I only needed to know it exactly for a few weeks usually.
I've gotten tremendous mileage out of exporting book highlights to BufferApp.com and setting them to release only once per day or so. Shuffled, they pop up days, weeks, months after I've put down the book, re-igniting its themes as they spark discussion.
I also made a theoretical calculation like this a couple of years ago [1]. I didn't answer the question "Can you memorize an infinite number of facts?" but rather the question "If you add a constant number of cards to Anki each day, what does your daily review load look like in the limit?"
We've recently added a simulator in the deck options of the new scheduler.
If you don't want to upgrade to FSRS
* Export a colpkg
* File -> Switch Profile -> New
* Import
* Deck Options -> Enable FSRS & Optimize
* FSRS Simulator (experimental)
I wrote a kanji learning app around Anki’s Flexable Spaced Repetition (FSRS) algorithm https://shodoku.app
My design principal is that I don’t need to be limited to few words, a sentence, and maybe a picture for each flash card. Instead each flashcard can be the all the relevant parts of the whole dictionary related to the kanji that you are learning.
Another design principal (where I deviate from Wankikani) is that drawing the kanji helps remembering it. So my app has two sided cards, one for drawing and one for reading. Both reading and writing practice include every word in the dictionary + example sentences that contains that kanji.
My third design principal is that I can learn vocabulary at the same time as I learn the meaning and the writing of the kanji (in fact, I like to learn a few vocab words instead of memorizing the different readings). So you can bookmark vocab words you want to learn along the kanji, and the next time the SRS system picks the kanji up for review, these words pop up at the top, to help you recall.
Note that the app is still in development, so use at your own risk, although I am actively using it for my kanji-studies. I still have a couple of features missing before I can call it done enough.
> I still have a couple of features missing before I can call it done enough.
I've checked your website and HN profile but there's no email to contact you, so here are some simple suggestions: when you enter a deck, I would suggest you add 2 shortcuts or buttons: one to randomly jump from the present card to another one of the same deck, and the other to randomly jump to another deck in the review set
This is because when faced with a sequence, the brain learns the order the element in the set, which may affect recall.
Being able to randomize the set helps prevent that, and give better results.
Random bit of UX feedback. When on https://shodoku.app/decks/add the + symbol means to add the deck, but in the interface even knowing that I accidentally pressed it 3 or 4 times to try and fold back the kanji after having clicked "show".
> Anki’s Flexable Spaced Repetition (FSRS)
Sorry to nitpick: Free Spaced Repetition Scheduler
Discussed at the time:
Spaced repetition can allow for infinite recall - https://news.ycombinator.com/item?id=32381206 - Aug 2022 (169 comments)
I made a well-received native iOS/macOS app for mining Japanese web and ebook content (Mokuro manga up next) into Anki or my own companion flashcard app: https://reader.manabi.io
I'm now working on flashcard auto-reviews while reading, so that if you read words/kanji that you have cards for, it'll automatically mark those cards as reviewed rather than ask you to recall them shortly after you just had to recall during your reading.
I hope this helps people replace more SRS time with actual reading (or listening)
I've done the whole spaced repetition and Anki thing, realized that if it's boring, you won't learn as well, and you won't stick with it in the long term, once you miss a week and suddenly you have 863 reps to get through to catch up. Instead, read stuff you're interested, apply it in your work, learn mostly Just In Time for when you need it. Learn by doing.
Strongly disagree. The worst thing that’s happened to educational theory is the idea that everything needs to be fun and exciting to be worth doing. Some stuff is worth doing because it works, even if it’s a little boring. SRS is one of those things.
> once you miss a week and suddenly you have 863 reps to get through to catch up.
You only have to "catch up" ASAP if you want to minimize the risk of forgetting any card at all, and then having to memorize it again. Otherwise, just working through the backlog at your preferred pace is the best strategy; you'll still be recalling most of the cards and the system will push repetitions for those that you do recall further out in time.
Perhaps the Anki devs should add some kind of special workflow in the UX that shows up when the user has a huge backlog to go through due to missed reviews, and tells them this. Because you're right that just seeing a bare number (863 cards to go through!) can be quite scary.
Spaced repetition is the cherry on the cake: learn by practicing, but create your own cards along the way, and when reviewing just focus on making a recall effort, not nailing everything.
For language, 10-20 mins / day is great, and if you are putting enough effort, 6-8 reps is enough to learn a card with 2 new concepts at ~15secs / rep, which really makes a difference long term.
I made a tool[1] with that approach in mind and tailor-made for language acquisition patterns. The key part is that creating cards (audio included) must be super-easy and fast, but they need to come from you nevertheless.
I think the mistake here is simply to many new cards a day. If you miss a week and have 863 reps, you definitely are going way too fast. You shouldn’t have to spend more than 5-20 minutes daily going over your review, and if you miss a week you shouldn’t need more than two or three half hour sessions to catch up.
I think there are diminishing returns as well by having to many daily new cards, as there is only so much you can commit into memory every day. So you will probably end up having terrible recall as well (which further adds to your reps).
I’m using Anki to learn Japanese. I actually go a step further in taking it slow as I’m still on my N5 vocab (first 1000 words) after 9 months of learning. I think it will be another 9 months before I can read stuff I‘m interested in in Japanese (which will probably be Go books).
Am I missing something here? Because I fail to understand how something like this is susceptible of a priori proof or disproof. Seems pretty obviously an empirical matter.
I merged spaced repetition with knowledge management app (), and I must say it was the biggest personal improvement for me in terms of organising and remembering stuff.
As a product though I struggle to make it work. I assume people mostly afraid of their notes and cards stored in the cloud vs local storage, which is understandable.
Logseq does it already and is well-established. Probably other solutions like Obsidian can do the same. It's a tough market to be in.
Seems nice. The page is so irritating on iOS that it must dominate your revenue losses though, not fear from cloud. Like portrait screen 4 colums of 2 character wide lines of texts. All video autostarts on fullscreen when gets into view, ehh.
Have to ask, have you ever tried this?
The redundant nature of mass scale internet content operates on spaced repetition I think. If you doom scroll HN for example, over time you’ll know more about random things. WSB might be a better example. I didn’t realize how many stock ticker prices and movements over the last few years are now embedded in my mind, without deliberately trying.
This is the 1000th time you’ll have read the same thing. The hivemind naturally regurgitates the same information at a predictable cadence. We’re due for a few Rust posts on HN, for example.
If you stay tapped in, you will always be learning from this organic body.
> The redundant nature of mass scale internet content operates on spaced repetition I think
Not really. Depends on the definition of spaced repetition, but you typically want increasing intervals (normally exponentially) for it to be effective.
See 9.1: https://pcl.sitehost.iu.edu/rgoldsto/courses/dunloskyimprovi... [PDF]
Not all spaced repetition is equally effective.
Repeated exposure (e.g. re-reading a book) does something, but active recall is much more efficient at solidifying memories.
Anyone had any experience using spaced repetition for reinforcing math learning?
I used it in my abstract algebra
classes
I basically only put two things in there: "Prove X" and my ordinary problems. Then I'd just do that. Those cards usually took at least a minute apiece, but I got really good at abstract algebra, at least when compared to similarly skilled peers. Mind you, I have always been good (perfect SAT score) but not great (never scouted for Putnam) at math.
Contrary to what you might think, memorizing definitions is less useful (though not totally useless) than memorizing the proofs of theorems which rely upon those definitions.
I would almost certainly go back to doing this if I ever want for a master's degree with a math class again. The idea crosses my mind every now and then to do it with linear algebra, which counterintuitively isn't actually a prerequisite for or all that similar to abstract algebra.
Checkout mathacademy.com
Not affiliated but 3 months in and I feel my math foundations are finally solidifying into something that will allow me to build into areas previously out of reach
I don't want to rain on anyone's parade, but: of all the really brilliant people I've ever met, interacted with or learned from, I haven't heard of anyone that used techniques like this. Like, the way they learn things and become brilliant is the exact same way every brilliant person has done it for thousands of years: they read books, they discuss and debate with other brilliant people, they study their subjects and work hard. Like, true intelligence is very rarely about rote memorization of facts, it's about making new connections, being creative, and working really hard. There are no shortcuts, you have to put in the work. Aristotle, Leibniz, Einstein or whatever brilliant person you can think of didn't become who they are using cue-cards.
Spaced repetition always seemed like those schemes to get you fit or slim in 30 days that never work. There is exactly one way to get physically healthy, and it's super-unfun: diet and exercise. Same thing with your mind, you have to exercise it and feed it appropriately for months and years. Spend the time you would spend on spaced repetition reading books or watching lectures and doing exercises instead.
Fine but what to use in practice. Tried Anki and found its algorithm bit infuriating. SuperMemo is better but thoroughly proprietary.
Hilbert's memory palace.
Spaced repetition is an excellent technique, but it shouldn't be the only tool in your learning toolbox.
I found that spaced repetition really works as advertised, with software like Anki. However, it is not really suited at learning large amount of facts and concepts. If it took an hour to review your deck, it's not a good use of your time unless the cards are really high value.
For example, if I want to learn electronics, my time would be better spent on fiddling with circuits and learning through mistakes of implementation. The nice thing is that concepts are interrelated to each other, not discrete and decontextualized that usually happened with utilizing spaced repetition. You get much better value from learning through experimentation and retain it better.
Don't get me wrong, spaced repetition still apply. However, if you can learn something that will last a lifetime, you don't have to repeat it for a long time if ever.
This is putting the cart before the horse. No matter how you slice and dice your equations, everyone intuitively understands you can't really store infinite data in your brain. Trying to claim otherwise is simply academic sophistry. No matter how hard you write your proof, it won't make information theory, basic physics, etc go away.