Learn Exponentially
saveall.ai
saveall.ai
A hacker is just someone who has practiced learning independently and has become exceptionally good at it. The reason people say you can't teach the hacker mindset is because without the underlying drive, there's nothing you can tell anyone. It's like when teachers who lament students don't care what they say so long as they get the right grade, it's because those students are optimizing for approval in a system because that's sufficent for their limited purposes. The more you profess to them, the more you reinforce that learning is passive submission to authority. If you want to make hackers, start with necessity, and technique will emerge as the artifact of navigating constraints. If you want to make people smart, challenge them instead of just telling them things. Hackers aren't defined by knowing more, they're defined by having physically done more. Spaced repetition as it's usually presented optimizes for outcomes in an approval environment that produces people who have been rewarded for cheating themselves out of knowledge and expereince.
I would say, want to learn physics? Build mechanisms or make radios. Number theory? Break cryptosystems. Astronomy and geometry? Sail at night. Lead? Ride horses. Fluid dynamics? Tune engines. Statistics? Write a spam filter. Speak a language? Tell their jokes, etc. Imo, most education is set around meaningless but scalable exercises of professed skills instead of meaningful exercises that are more powerful, but don't scale. We've optimized for scale at the expense of quality. It's the solution to an inferior problem.
So sure, learn spaced repetition, but really, find something and practice it for more joy and better results instead.
— Antoine de Saint-Exupéry, Citadelle, 1948
The naive approach definitely leads to rote learning of factual trivia, but proper prompts can definitely foster curiosity and understanding. In fact it's often times the process of creating new prompts that reveal a gap in my understanding.
This Article from Andy Matuschak is a very thorough introduction to the art of prompt writing: https://andymatuschak.org/prompts/
Of course it doesn't mean that practical experience is obsolete, quite the opposite. But spaced repetition works great in conjunction with practice. I'd go so far that it's more effective to do both, than just practical experience.
Breaking it down into such small bits makes it immediately obvious if I haven't understood something, because I either wouldn't understand the answer for one of my own cards, or wouldn't even know how to create an answer. If done properly, it's much harder for something to "slip through the cracks" so to speak.
When I first started properly integrating this into my day to day life I also noticed that I was interested in remembering many more mundane things than I would've expected. Simple stuff that I really didn't have a place to write them down for, but that also weren't really important enough to take time out of my day to actively memorize it. There's a place for that now. It's been really interesting to see how broadly applicable spaced repetition is and that it's not just useful to learn for college exams.
I think I subconsciously avoid new projects to learn something programming-related because of how often they end up unbounded in time due to one of the fundamental attributes of writing software: unknown unknowns. The kinds of sticking points in a project that could take you minutes to figure out, or days, or even weeks.
I have limited free time and I want only so much of it to be _more_ programming. (I’m a professional software engineer) If I can’t be certain I can timebox a new pet project with _some_ degree of successful outcome, I think I’m low key anxious enough to avoid starting it.
Does anyone else struggle with this? How do folks overcome it and work on new things without them taking over?
If I want to do something like move a light fixture in my house, it seems like it should be fairly straightforward. Choose where I want the new one to be. Create hole. Move wires. Connect fixture.
But it only seems that simple because I'm ignorant. I have to install a new junction box, which requires having a joist nearby, or adding a cross-support between two. Am I going to need to move any insulation? What insulation do I even have? Is the existing wiring long enough, or in the right place, or am I going to have to replace it? And I need to pin the wires down every so often? What? How? Where? Where do I even find that info? And never mind patching the hole I'm leaving behind. It's not just eyeballing a drywall patch. I can't match the old texture of the ceiling, even if I thought it was plain. My patch is really noticeable. And on and on.
Unknown unknowns take what I thought was going to be a little afternoon project and turn it into asking all my friends for recommendations of people who aren't booked 6 months out and are willing to handle a tiny little project cleaning up after my ignorance.
Taking on projects before I even know what I don't know has gotten me into more trouble than just about anything else.
I've never nearly burnt my house down taking notes from a book and hiring a professional to do the actual work!
Also learning a new (natural) language sometimes requires hammering words or gramatic rules into memory, and having a good teacher can be much faster than learning on your own by reading texts.
In general:
A priori knowledge only gets you to the starting line. Experience carries the rest. And you can only get that experience by doing yourself, not second hand.
Ever.
Even reading a great O'Reilly book being sure to complete and understand the examples isn't enough. Without that immediate practical application, it's no more educational than any other form of entertainment, and much drier.
Do you have thoughts on the cost or how to optimize that type of learning?
There are undoubtedly many areas in which memorization is useful. I tend to use memorization as a second-order tool, in the sense that it is only useful to memorize once I’ve learned that memorization would be necessary.
I memorize combinations to locks I unlock frequently. I memorize names of items I sell in my shop so I don’t have to look them up over and over again.
In school I often memorized equations just long enough to get by. The few that are still with me are not those I used most frequently; they are the equations I understood at a visceral level. Obviously this means I am more conversant in Newtonian happenings than quantum concerns, so maybe there is a place for memorization. Or perhaps I lack sufficient experience in the quantum to really feel the laws that govern the smallest realms.
Either way the article paints a dull picture of learning. What of the feeling in the minds and hands of those future carpenters swinging their first hammer blows? What of the deep learning of the pianist that happens only after the transition from the first concerto as audience to the latest as featured virtuoso?
An exponential increase in the type of “learning” furthered by spaced repetition might be useful to some. I still prefer the linear road to understanding.
For me (in medicine), it is the exact opposite. Memorization comes first, then you start to actually understand and learn things. Everything is so intermingled and there is no "learn this first to understand that". In addition, you can't start practicing things before memorizing them.
I feel like it's the same thing with the mnemonics. It's useful at first, but not very practical or efficient. However, as you use it more, you actually learn and stop using the mnemonic.
Our working memories have a capacity of 4. This means that we basically can't understand something if it requires more than 4 pieces of New knowledge to understand. To understand more complicated things we need to move some of the knowledge into our long-term memories.
We wrote an article on this topic here that i'd love to know what you think of
This comes off as condescending and dismissive. It's a poor way to treat people who have taken the time to engage with your content, especially if engagement is what you want, which appears to be the case given your other replies on this topic.
Take the time to respond to them directly rather than pointing them towards more content you've written, even if it means repeating ideas you've written elsewhere.
This approach has a number of benefits:
1. It has the effect of presenting what you've read elsewhere inline (most readers won't click that link)
2. It gives you an opportunity to revisit and refine your own thinking, and
3. It forces you to think carefully about the criticisms levied
And most importantly, it reciprocates the effort they've put into reading your post and responding to you so that you don't come off like a jerk.
This is how I learn basically everything. That and practical application, so I'd learn cooking by cooking but I'd learn about information theory by just reading for hours at a time and falling down one hole after another. All this makes me wonder, where do you get the "you have 4 working memory slots" thing from? And how would you actually go about forcing things into long term memory?
What the heck am I supposed to take away from this?
This is a half ass theory, not evidence.
Where's a shred of evidence, on the time scales here, these "units of information" are retained (under either method). Are they even relevant compared to a _skill_ like reading that enables quicker information ingestion across an entire life o and is applicable across a wider range of problems than _individual information units you read_??
The core concept that spaced repetition increases rapidly in effectiveness over time is called the Spacing Effect. There are many many studies that have investigated and proved it
Evidence? What's the evidence this is a linear growth process? How many years? At what age? What populations? There is no rigor here at all.
Then it says for the "best case" that "adults" takes years to get twice as effective at reading.
If you don't believe this then it means you think it takes adults less than a year to get twice as effective? If effectiveness doubles each year then that would mean by the time a 25 year old becomes 65 that they're 1 million times as effective at reading than when they were 25. Does that sound reasonable to you?
For example, is using a text search on an e-ink device better than knowing "I wrote this down in the first 1/3rd of the notebook on the page with a coffee stain". Maybe? Maybe not. It's a trade-off. A physical notebook and an e-notebook both offer different ways of indexing, searching, and remembering where things are. One is not clearly superior to the other in this respect.
These e-ink devices have left a clear advantage they could claim unrealized.
I want an e-ink notepad where I can turn my notes into spaced repetition. I want to take hand written notes, proven to improve memory, and then I want to blot out portions of the page and have those blotted out portions be presented to me by a spaced repetition algorithm to help me remember my own hand written notes.
I'd pay a lot for such a dedicated device. Getting hand written notes and images into Anki or other spaced repetition programs sucks. I'd love to just be able to write or draw, with my own hand, and easily integrate with a spaced repetition algorithm. This is valuable enough that I'd happily use a dedicated device just for this purpose.
With the right plugins maybe Anki wouldn't be too bad, still not as seamless as a dedicated device though. And I think this idea is important enough to warrant a dedicated device. Imagine how helpful it could be to have all students trained to use such a device to memorize their own notes. Memorization alone isn't enough, but it can serve as a foundation to build knowledge on top of. Many student lack any sort of foundation, and efficient memorization would do a lot for them.
One problem I can immediately think of with a dedicated device is "vendor lock-in". Especially when you are going to spend a long time creating decks or practicing them, you wouldn't want to lose your progress or cards because some company stopped supporting their proprietary format.
This is actually the only reason I won't ever consider using most of these "Next Anki" or "Anki but better" SRS apps that pop up every now and then.
This article talks about how the reps needed to learn one piece of information, reduces exponentially over time. You might need 1 rep per day at the beginning, but only 1 rep per 100 days after a month. This basically means that if you have a lifespan of 100 years, spaced repetition means you only need around 10-20 reps to remember each piece of knowledge for the rest of your life.
But learning N items will still take 10*N reps. It scales linearly. A far cry from exponential
Consider a uniform repetition approach that repeats each item once per day. The rate of our uniform repetition system scales linearly with respect to time. If one item takes 1 second to review, then I will lose 1 second per day. Therefore, N items causes me to lose N seconds per day. Even in the case where I don’t add items to my list, I lose N seconds per day.
In contrast, as you stated, the amount of time we lose per item per day reduces exponentially with spaced repetition, however. If we don’t add any items to our list, the review time asymptotically approaches 0 seconds/day.
Unrelated to all of this, I agree the title may be a touch embellished. :)
The first rep gets you 1 extra day of knowledge. The next rep gets you 2 extra days. The next one gets you 5 extra days.
So each rep is getting you more and more knowledge even though the time required to execute the rep remains the same. This is exponential growth.
If you only have 1 card then for most days you won't have any reviews to do so it doesn't work. But if you add cards regularly then over the long-run every day you'll be getting more and more "extra days of knowledge" by doing the same amount of reviews each day.
It's like someone giving you $10 and then trying to argue that they haven't actually given you $10 because just a few minutes ago you had a $10 bill in your pocket but you spent it on something else.
The difficult part is creating a deck and crafting the answers and questions. Because usually this is a time-consuming process. So if it was a community-effort then it would be a win-win.
This is probably not an original idea and if anyone knows already where to find such decks, that would be cool.
This is something I've wanted to do with my app (Fresh Cards). I ended up defining a simple text file format for the flashcards[1] to help make it easier to share and import cards. You could post flashcards as simple text that someone could drag and drop into the app to import. (Formats like Anki's .apkg file are great, but they don't make it easy to peruse the cards without importing into Anki.)
What's missing in all of this, though, is an actual community where you could search and browse the decks and collaborate to create new ones. Though, if you simply use text files, you could host a deck on github, for instance, and allow people to create pull requests to improve it. I think there's room for creating nicer user experiences to surface decks and encourage sharing, however. (Imagine, for instance, a social media-like feed where you could see new flashcards being created and you could search by tag for your target language.) Anyway, I think this area is ripe for exploration, but the user experience has to be done right to encourage collaboration and sharing of decks.
I’ve used Anki over the years to great effect with second language learning and am fully bought into the paradigm, but I do find Anki quite clunky as a piece of software. Fresh Cards seems like it’s designed specifically in response to that.
Does this use a similar repetition algorithm to Anki? Are there any obvious limitations versus Anki? What sort of UX differences would a long-term Anki user need to get comfortable with?
Would also be keen to hear from anyone here that’s used this.
I also included my attempt at converting Anki's algorithm to javascript so you can compare and contrast. I'm open to any feedback on improving the algo that I do use!
In terms of limitations, probably the biggest one is that Anki lets you create cards using a template system, which Fresh Cards doesn't have yet. In Anki you can create one note and then from that have multiple flashcards auto-generated using several templates. You can also do something similar with cloze deletions. One of my goals is to get to 100% Anki compatibility over the next few months, so I'll be adding a template system and similar cloze-deletion system.
In terms of UX, the app actually has multiple review styles. When I first built the app, a lot of users actually wanted to re-review cards, but the app would strictly schedule them, so you didn't have a way to do that. (Anki lets you "review ahead", but for the most part tries to keep you on the schedule.)
To satisfy both users who wanted to review whatever they wanted, whenever they wanted and users who wanted to stick to a spaced repetition system, I designed the UX to act as a "playlist". You just hit play and it'll start reviewing the cards in the order that you see. (You can sort and filter based on spaced repetition algorithm, or use several other modes.)
The app is pretty much designed to just pick up and go without a lot of instruction, so it should be easy to get used to.
For language learning, there are flash card generators that make this a simple one-click process. I think these strike the right balance of simplicity & flexibility/personalization. Of course, a tool like this relies on a free database that you can map concepts onto. But I could see this sort of working with wikipedia or some documentation.
Rarely with these SRS services do you see actual high-quality decks that outdo public Anki decks, which is a shame because it would be a great way to add value.
Refold, a company focused on language learning, does exactly this [0], and having tried their JP1K deck for Japanese for a while, I can say without any hesitation that it was shocking just how high quality everything was.
It had the full works: Japanese audio, kanji, furigana, multiple definitions, a custom background, etc. And I wouldn't be surprised if there were even more changes since the last time I tried it.
I recall there being something similar to this for medical programs, but overall I'd say that this approach sadly isn't something that a lot of people are focusing on.
So how would a different community tool provide better content? What tactics could be used to increase quality?
(2) The main problem we see with Anki public decks is that the user themselves decides whether they got the question "right" or not when reviewing the cards. This lack of a "teacher" means that it is very very difficult to learn using someone else's cards.
You basically end up kind of getting something wrong but then saying it was right anyway. Do that a few times and your trust & investment in the process goes and you'll eventually you lose motivation to carry on with the deck.
Save All decks are different. WE decide whether you got it right or not, not you. This makes it much easier for you to learn using someone else's decks
If you don't want to learn (or memorize), no technique or tool will allow you to learn. Make a tool that solves this "problem", and the user will simply find other ways for running away, like not using the tool.
That said, I can see some negatives as I have read that for learning, it's generally better to construct your own flashcards.
It's gotten so popular that now theres a shared 30,000k card deck called AnKing that a lot of students use. The deck itself is free, but the curators of the deck recently launched a new paid service to automatically update the deck as new cards from the community are added: https://www.reddit.com/r/medicalschoolanki/comments/rb62m6/a...
The Anki public decks are usually too low quality to be useful unfortunately
Would very much like a version control system for Anki decks so updating cards can be done by the community and transparently. This would allow people to validate and properly maintain decks that change with time (like programming languages / regulatory requirements / operating standards etc).
Knowing things is really important, but I don't think learning makes you "smart"; it just makes you know more things.
So, sure, you should try to apply your knowledge and layer concepts on top of each other. But if you do that AND also remember a lot more of such experiences you'll learn a lot faster.
I agree. Who ever said you should try to remember that? Remember the high-level important and transferable information, don't waste time trying to remember information that won't help you elsewhere.
Edit: to the article's point spaced repetition to memorize domain specific facts is useful but it's not exponential like reading is.
I gather you've never met a zoomer in your life
I even created a little web page for the app, but I’ve mostly abandoned the idea due to a lack of interest.
For example lots of people use Save All for this exact reason https://saveall.ai/
I’ve been seeing HN submissions of various quality to extoll the virtues of SRS in an attempt to sell Anki clones or Anki for X for almost 15 years now.
For context, I was a big fan of SRS and even contributed to Anki back in the day! I was really into foreign language learning, had majored in one language and was learning another language in a separate language family.
I built, ran and put my heart into brick and mortar language immersion school for years. Over time, I realized both from my learning experiences and those of my students that SRS fell far short of extensive reading.
It's tempting to break things down to "units of information", as you put it your assumptions document. SRS is great for decontextualized information (e.g., memorizing all the capital cities in the world), but that's not really how language works or how the brain works for most learning tasks. There are higher-level things your brain picks up, such as collocations, grammar and shared cultural beliefs.
Over the short term, SRS can be useful for building a scaffold to work from, but over the long term, Extensive Reading crushes it on pretty much every metric, including raw size of passive and active vocabulary.
If you absolutely hate books and want an online resource, then I'd suggest https://www.lingq.com. It has a lot of free content and lets you import your own. Their tech/design chops are meh, but it's run by true language learning enthusiasts and the founder dogfooded it for at least half a dozen languages.
Effectiveness of spaced repetition scales really fast / exponentially whereas other learning methods don't scale like that - do you agree?
If so then over the long-run spaced repetition is always going to be extremely efficient relative to other learning methods.
Of course not. As I wrote, in the comment you’re replying to, I was a user and evangelist of SRS for years but eventually saw the reverse.
It’s a very useful tool for a certain narrow niche of memorization tasks. However, it’s an extremely inefficient learning method over the long run compared to reading and using information or skills in context.
> Effectiveness of spaced repetition scales exponentially
Given a fixed amount of review time per day, the largest deck you can maintain grows asymptotically over time, not exponentially.
If so then a fixed amount of review time per day will let you remember things for longer & longer.
I'm not surprised. That's how I learned English.
I tried adding a few cards, which seems to work fine even in the worst case scenario. It seems to work even in other languages? I haven't tested it enough to comment on how well it does though, but my guess would be as well as GPT does in the said language. This has the potential to be a huge time saver. If it could handle something like a few pages of notes and divide it up, I can see it being the new standard of SRS for many people.
As I said, I can't see myself using any non-free software for something as critical as learning.
To learn, you need to act and have the world give you feedback.
If you’re starting a company, you need to create something and get feedback on its value and iterate.
However, I’d be very interested in learning how people use SRS for remembering information they read on books/articles. Do you state new concepts as Q/A? Do you save interesting facts, or things you think might be useful in the future?
I think this second type of information is not well suited for flashcards. The article seems to imply it is, though, and I’d love to be wrong about it.
On Save All you can create cards that are just statements, no need to turn them into Q&A. So if I hear an interesting fact I usually just dump it in quickly verbatim.
When using this method if you are learning something for the first time or you come across something that peaks your interest, an exercise or question, pause and take notes, implement the solution in multiple variations, then continue the video/audio.
The one I use for desktop is https://github.com/codebicycle/videospeed
It would be great if someone could provide an example to an old dummy like me.
The effectiveness of spaced repetition scales exponentially and much faster than other learning methods. So use spaced repetition and you’ll learn a lot faster in the long-run.
Get started, practice often, that's the only way to have compounding gains at many activities. Music, hobbies, working with your hands, etc.
The long perspective is very helpful. Don't worry about improving today, but about the long trajectory.
ive done a lot of thinking on this area and have mocked up a "Big L" notation for learning curves - shameless plug: https://www.swyx.io/big-l-notation
Is that something that we know for sure?
See here for a reference for that: https://www.livescience.com/53751-brain-could-store-internet...
Or this article we wrote also talks about it: https://saveall.ai/blog/learning-is-remembering
My opinion is that true exponential learning depends on specific content in a specific order.
What we need is the dependency tree of concepts. Does such a thing exist? Curriculums are kind of a non-rigorous attempt at this.
I did want to give some feedback, though. I think this post suffers from too much hand-waving, which is what plagues most other posts about learning and spaced repetition (excluding probably just Gwern). For example, it compares flash cards and reading to just reading, citing the results of the study as a negative:
> This is however a very slow process. One study implies that in the best case it takes adults 10 years of reading 1 hour a day to get twice as effective at reading. Even if this is technically learning exponentially, the improvement rate is so slow that the process is indistinguishable from a linear one.
...
> We said in the best case it takes adults 10 years to get twice as effective reading. With spaced repetition it takes only days for your time to get twice as effective. These growth rates are completely different.
Maybe the rate of change of effectiveness of the reading is slow, but does that matter if you're accumulating knowledge from all of that reading, especially as it builds off of prior knowledge? I also don't think it's a 1:1 comparison to contrast these. It only takes days to get twice as effective at reading with spaced repetition? Or at learning? If it's the latter, I don't think that's what the earlier study measured?
The other thing that jumped out at me is the huge focus on spaced repetition and memory for learning, which are absolutely helpful, but there seems to be a lack of what constitutes memorizing versus understanding (and I'm not sure I see that in the Learning is Remembering post either). I think about other ways to build your understanding, like working through problems and applying the knowledge, that are key to learning. Much of learning physics is getting your hands dirty in the equations, and there's a big difference between knowing a formula and really understanding it in action.
It only takes days for spaced repetition to get twice as effective, not reading. It takes you years to get twice as effective at reading.
> I think about other ways to build your understanding, like working through problems and applying the knowledge, that are key to learning. Much of learning physics is getting your hands dirty in the equations, and there's a big difference between knowing a formula and really understanding it in action.
I agree, you should definitely work through problems and apply your knowledge. The post is arguing that you should do that AND spend a little bit of time doing spaced repetition, not to only do spaced repetition and nothing else.
> The post is arguing that you should do that AND spend a little bit of time doing spaced repetition, not to only do spaced repetition and nothing else.
From my perspective as a reader, I'm not sure I see either post discussing working through problems and applying your knowledge, but I do think the inclusion of that would make them both much stronger, especially since it seems like you've clearly thought about them from your comments! I do see the mention of reading + spaced repetition and the mention of memory helping math problems in the other post, but I'm not sure I see where you've explored what that active use and problem solving looks like (or how spaced repetition helps).
Again, I think this a great idea, and I don't mean to be negative for the sake of negative — you were asking for feedback on your post, and I figured I'd offer some :)
Thanks in advance.
or: https://www.supermemo.com/de/archives1990-2015/help/read
Learn slow and you won’t reach your potential. Learn fast and you might. Learn exponentially and you’ll achieve more than anyone thought you could.
THEN
...
Else, we don't know.