Piotr Wozniak as a role model for lifelong learners

Piotr Wozniak is the inventor of SuperMemo, the ancestor of all spaced repetition programs (such as Anki).
He is currently spending his time learning (and writing) as efficiently as can be, in the most rational possible way.
In my opinion, he might very well be the most knowledgeable person on earth concerning these topics :

  • memory
  • sleep
  • probably other things

I find both his lifestyle and articles inspirational.

About his lifestyle :

Some of his articles :

I wish more people knew about his articles (and about spaced repetition in general) but he doesn’t spend a lot of time (if any) trying to attract a broader audience.

So perhaps you will want to emulate his lifestyle, at least partially, or find his articles worth the time spent reading.

And as this forum is mainly about mnemonic techniques, you might be interested in another point of view : he thinks that spaced repetition and active recall must come first, and that mnemonics (including the method of loci) should only be used in some cases.

3 Likes

There’s an incremental reading extension for Anki, but I wasn’t satisfied with it after trying. SuperMemo seems to be a powerful piece of software, but unfortunately it only works for Windows.
So I’ve come to think that when someone designs a cross-platform free SuperMemo, or a better incremental reading extension for Anki, it’s not going to be a waste of time.

EDIT 2017-12-14: this add-on had actually been improved in the meantime (I had tried it in October 2016). The current version can be found here.

Supermemo seems to be cludge software. That is, it’s bits and pieces of routines put together with no overall plan and it shows. His spaced repetition algorithm is, I think, similarly afflicted, which to me indicates it’s not based on sound math or psychology.

It is also expensive for what it is.

The incremental reading part was a good idea in that it would be a good organizational system for readings. However, the last time I checked, the incremental reading part cannot use PDF’S, it also cannot use word documents. Indeed, almost all file formats are beyond it’s reach, which makes it pretty useless.

Although I have never used SuperMemo, I wouldn’t say that, because the program was originally as simple and mathematical as can be. Then it grew more complicated over years, but given Wozniak’s knowledge and experience about memory I tend to trust him on his spaced repetition algorithm. I would be surprised to learn that he gradually made the algorithm worse over time, without realizing it. This quote from him might shed light on the “simplicity vs complexity” debate : “Further improvements to the algorithm used in SuperMemo are not likely to result in further acceleration of learning. However, there is still scope for improvement for handling unusual cases such as dramatically delayed repetitions, massed presentation, handling items whose contents changed, handling semantic connections between items, etc.” (from the article about the SM15 algorithm).

You might be interested in this page : you can get version V-2 (so currently SM15) for free.

As for PDFs, word documents and so on, I think copying&pasting is not that bad. But I agree that an ideal incremental reading program should indeed be compatible with major file formats.

The software is perhaps bits and pieces of routines put together, but I’d nevertheless really like to have a free, cross-platform software that would allow to, in addition to the spaced repetition algorithm (quote) :
"- read huge volumes of texts with minimum forgetting (see: incremental learning)

  • learn with pictures, audio and/or video
  • analyze your sleep
  • optimize your sleep for learning and optimize your learning using your sleep logs (see: Sleep Chart)
  • improve your problem solving power (e.g. with concept maps)
  • boost your creativity (e.g. with neural creativity)"

Especially the incremental reading part.

So far, I think that other SRS (I only know Anki so I’m extrapolating) can only do #0 (spaced repetition algorithm), #2 (pictures, audio, video), #5 (but not with what’s called concept maps) and #6 (but not with what’s called neural creativity).

Thank you for your comments Rodent.
I did use super-memo a long time ago, and I didn’t think his algorithm for spacing worked. So, I don’t think the spacing algorithm got worse, since I didn’t think it worked to begin with.

If one looks at the evolution of his algorithm (from SM(0) up, it seems to be based on trial and error to see what works (with mainly himself as the subject). Kind of trying to repeat Ebbinghause’s work, but Ebbinghause’s work was superior since he used nonsense syllables, and not items which have meaning (which would be easier to recall).

Also, It doesn’t seem to me to be based on a mathematical analysis (math) of lets say Ebbinghause’s forgetting curve (psychology) which would result in a solution in the form of an exponential equation. This exponential equation would be the algorithm, and I don’t think it couldn’t really be ‘improved’.

.
Wozniak’s “Twenty rules of formulating knowledge” seems to be the gold standard for people who are trying to improve their memory retention.

The twenty rules are frequently mentioned on the Anki forums. The name, “Wozniak” appears on the Mnemosyne site.

SM2 is mentioned on both the Anki and Mnemosyne sites. The Anki (An) and Mnemosyne (Mn) algorithms are both based on the original “simple” SM2 algorithm.

AFAIK, if all cards are answered correctly, and on the correct day, then it’s easy to predict the interval when the next card should be presented for that specific user. Of course, the interval will vary for each user.

I think the problem arises with failed cards. Assume that a user answers a card correctly the first 10 times, and then fails for some reason. If the user then gives (say) 3 correct answers, then both An and Mn will try to get the user back to almost where she was before she failed.

On the other hand, SuperMemo’s later algos cause the user to be stuck in “low interval hell” until her next birthday, so she then changes to An or Sm.

The above quote comes from this discussion of SuperMemo, and it’s actually the most complete description of Anki’s SM2 that I’ve seen:

http://www.antimoon.com/forum/t16080.htm

I think the statisticians would say that SM is “overfitted” or “curve fitted”. In other words, SM will try to fit an exact curve through all the points, including the 3 failed points. The curve is then “extrapolated” to find the interval for the next presentation of the current card. So, if the 3 failed points cause the curve to slope downwards, then an “overfitted” SM curve will probably continue sloping downwards until several successes cause the intervals to start sloping upwards again. The situation is similar to failed attempts to predict the “bottom” of a bear market.

From the list of links in the OP, it seems to be clear that Wozniak was a perfectionist. That might well be the cause of his excessive and obsessive tinkering and tampering with the original good SM2.

BTW: Mnemosyne is not simply an SR tool, it’s a complete research project. Therefore I think the developers will deliberately try this and that in order to prove that something will work or not work.

.

Interesting thoughts.
By reading the differences between the original SM2 and the adapted SM2 Anki uses, I feel like there is actually a lot more to spaced repetition than I previously thought. For almost every difference that’s mentioned, I wonder whether it is an improvement or not, and I just haven’t the faintest idea.

Then as soon as they get an incremental reading extension working, I’ll go for it !

Which makes me wonder why there is still no incremental reading implementation other than Supermemo’s and the Anki extension (“not anywhere as powerful as Supermemo’s incremental reading” according to its developer). Years have passed, and spaced repetition software programs are now numerous. Incremental reading seems to be popular from what I’ve read, and it doesn’t seem too hard to code (at least if we copy & paste the material by ourselves).
Has anyone an idea about this ?

To me “low interval hell” means the software doesn’t work. Of course there are reasons perhaps why it may be difficult, as people point out. But, to me, even if it there were many reasons why it was actually impossible, the bottom line is still that it doesn’t work.

I find Wozniak’s list of 20 knowledge rules to be similar to the software. That is, it seems to me that ‘the rules’ are culled from anecdotal experience and in no way represent the conclusions of responsible research. His ‘rules’ are not supported by references to responsible research and therefore to me are simply made up. One quote from them is: “You should avoid such items whenever possible due to the high cost of retaining memories based on sets.” But people must ask themselves…who says? He recommends avoiding enumerations and ‘set’ knowledge and instead advocates using simply structured knowledge (one simple isolated question, requiring an equally isolate answer. I can immediately counter that with: “but that will simply fragment information into bytes of trivia, and you won’t really have any knowledge at all, you will just know a lot of trivia.” This reasoning sounds as good to me (actually better) than Wozniak’s advice. So how do we know who is correct? There is only one real answer: correctly executed and analyzed scientific studies.

@Celtic
If it’s called “low interval hell”, then there is probably something wrong with it. Notice the interesting point OldGrantonian has made : according to him, this low interval hell seems to actually be the result of Wozniak relying too heavily on sound math! I would agree that more scientific studies are needed there, as they could indicate the best ways to handle these unusual cases, if I didn’t remind myself of the fact that I don’t spend time looking for these studies.

Also, I understand what you mean when you write that the software doesn’t work. But please take into account the fact that you are influencing other readers, who may not even know about spaced repetition yet. Some might subconsciously associate SuperMemo or even SRS with “doesn’t work” or “not based on sound math or psychology”. Yet it has to be restated how powerful spaced repetition is, be it with SuperMemo or another SRS program. From an agnostic’s point of view (someone who has only been using Anki so far), I would tend to say (as Wozniak did by the way) that the difference between spaced repetition algorithms is marginal compared to the difference between spaced repetition and no spaced repetition. I think you’ll agree with me on that.

As for your skepticism about his 20 rules, I think that it is a sound one since he didn’t mention his sources (in this article). Although I do personally feel that his rules make sense, I understand that this is very subjective, given how little I know about memory, so I’m certainly not going to try to persuade you on that. I’m only going to make 2 points :

  1. seek out disconfirming evidence. If you think that Wozniak is not using enough research, try to find information that might prove the contrary.

Piotr Wozniak. Since the articles I have read or skimmed have convinced me that he knows a lot about memory and spaced repetition, to the point that I consider myself a complete novice compared to him, I think I have to trust his conclusions more than mine. This would change of course if I began studying memory extensively by myself. Even if it was proved that his rules are made up from anecdotal evidence only, I would still value them, because he has a lot more experience than I do. However, as long as I haven’t seen and examined his sources or heard that many scientists studied his claims and found them to be backed by research (I believe that climate change is human-induced even though I never studied the evidence), I won’t consider his rules as scientific knowledge, but only as interesting advice.

Here’s a bonus about the argument from authority that I’ve been using several times in this thread. I found this interesting story on the Wikipedia page Argument from authority :

An example of the use of the appeal to authority in science can be seen in 1923, when leading American zoologist Theophilus Painter declared, based on poor data and conflicting observations he had made, that humans had 24 pairs of chromosomes. From the 1920s to the 1950s, this continued to be held based on Painter's authority, despite subsequent counts totaling the correct number of 23. Even textbooks with photos showing 23 pairs incorrectly declared the number to be 24 based on the authority of the then-consensus of 24 pairs.

This seemingly established number created confirmation bias among researchers, and “most cytologists, expecting to detect Painter’s number, virtually always did so”. Painter’s “influence was so great that many scientists preferred to believe his count over the actual evidence”, to the point that “textbooks from the time carried photographs showing twenty-three pairs of chromosomes, and yet the caption would say there were twenty-four”. Scientists who obtained the accurate number modified or discarded their data to agree with Painter’s count.

Given the long thoughtful reply of Rodent, I thought it might be wise to make a few points in return. (I apologize in advance for its length).

“Wozniak relying too heavily on sound math”. I would say that there is no such thing as relying too heavily on sound math. Fitting a curve (or over-fitting it) does exactly what you want the math to do and it does it perfectly. Any curve of any complexity can be modeled. But that isn’t the point. There is something wrong with the whole approach, and if you look at SM(0), there is no evidence of curve fitting at all. So, to my mind the software is not based on sound math, but tinkering.
I definitely agree that spaced repetition is of course immensely important for long-term retention. Indeed, I would argue that it is the only method of rehearsal that makes sense for long-term retention.
I don’t worry, however, about what people might sub-consciously decide about spaced repetition and any software that implements it. We are all adults and responsible for our own decisions. Besides, I don’t think it serves people well if one doesn’t present conflicting points of view in case they take it the wrong way.
Again, I would say that spaced repetition is essential to long-term retention, but software isn’t, and there may not even be any software that’s very good at it. But anyone can, with a bit of effort, tailor the Leitner system so that it works well for them.
With regard to his 20 rules and my seeking disconfirming evidence. Frankly, I don’t have the time or that much interest in Wozniak’s work to do so because I actually don’t think he says anything of much interest. From what I’ve read of his, admittedly quite a while ago, I don’t believe anything is original with him. I would much rather take my cues from established researchers in cognitive psychology. As far as seeking disconfirming evidence, it is the obligation of the person who makes the truth claim (in this case Wozniak and his 20 rules) to provide evidence as to the truth of the claim (which I don’t think he has done); the burden does not lie with others to demonstrate its falsity.
Finally, with regard to the argument from authority. Today, of course, we have no choice but to rely on authorities given the complexity and scope of modern knowledge. So, it isn’t the appeal to authority that is an issue. What is an issue is an appeal to an improper authority. In this case, I would submit, that we do have an appeal to an improper authority: that of Wozniak.
Now, lest you think I have something against the person, I can positively say that I don’t. I do however question the truth of his claims, and the originality of his work.

I’ve done a bit (only a bit, really) of research since last time and it appears that you’re right, at least for SM-2 (I haven’t studied the latest algorithms). Here is the SM-2 algorithm :

Algorithm SM-2 used in the computer-based variant of the SuperMemo method and involving the calculation of easiness factors for particular items:
    1. Split the knowledge into smallest possible items.
    2. With all items associate an E-Factor equal to 2.5.
    3. Repeat items using the following intervals:
    I(1):=1
    I(2):=6
    for n>2: I(n):=I(n-1)*EF
    where:
    I(n) - inter-repetition interval after the n-th repetition (in days),
    EF - E-Factor of a given item
    If interval is a fraction, round it up to the nearest integer.
    4. After each repetition assess the quality of repetition response in 0-5 grade scale:
    5 - perfect response
    4 - correct response after a hesitation
    3 - correct response recalled with serious difficulty
    2 - incorrect response; where the correct one seemed easy to recall
    1 - incorrect response; the correct one remembered
    0 - complete blackout.
    5. After each repetition modify the E-Factor of the recently repeated item according to the formula:
    EF':=EF+(0.1-(5-q)*(0.08+(5-q)*0.02))
    where:
    EF' - new value of the E-Factor,
    EF - old value of the E-Factor,
    q - quality of the response in the 0-5 grade scale.
    If EF is less than 1.3 then let EF be 1.3.
    6. If the quality response was lower than 3 then start repetitions for the item from the beginning without changing the E-Factor (i.e. use intervals I(1), I(2) etc. as if the item was memorized anew).
    7. After each repetition session of a given day repeat again all items that scored below four in the quality assessment. Continue the repetitions until all of these items score at least four.</blockquote>To his credit however, the curve (x-axis : time ; y-axis : probability to remember the item) cannot easily be found a priori for an item, since item difficulties vary. So you have to take the quality of the answer into consideration. But as soon as you do that, how do you link it to the curve you're trying to find out ? (remember these are only the thoughts of someone who knows very little about spaced repetition. I would immensely appreciate experts intervening in this discussion. But why would they ? They've already written what they think elsewhere. So here is actually simply not the place to discuss the value of spaced repetition algorithms. I'm not willing to do this here, but only to introduce some people to spaced repetition, spaced repetition software, and Piotr Wozniak's lifestyle and articles).

Neither do I (except in particular cases like climate change denial or creationism). I really find your point of view interesting, and of particular value to me because it challenges mine. I was only referring to formulation. Saying that the software doesn’t work is an overstatement.

I think I understand your position. What you’ve read hasn’t convinced you, and you just cannot give everyone on the Internet their chance by spending an infinite amount of time fully studying their claims. It’s true that he didn’t provide his sources in this article. Now it’s very likely that what we both think of Wozniak and SuperMemo algorithms is at best inaccurate, but we lack time finding the truth about everything. We have to rely on probabilities to build our points of view.

Thank you for your research and response rodent.
Yes. We can’t solve all this stuff here.
One thing I do find curious is the setting of the ‘difficulty’ factor.
To me, this is an odd way of thinking about the issue since there’s no such thing as item difficulty.
Of course one could say it’s just a choice of label, but language does influence thinking and could have led to unclear solutions to the optimization issue of SRS in general.
(as a side note, when magic numbers are hard coded it’s often a worrisome sign).

I divide the whole learning process into two steps:

  1. Encoding(mnemonics)
  2. Reviewing(SRS)

Encoding,for me,is organizing,preparing and making sense of the subject I am going to learn. I use ‘mnemonics’ for this. For example,when I convert a number or word into an image for memorizing or putting into my long term memory,that is encoding for me!

And the encoding should always come first.

Then,there is the ‘reviewing’ part. Even if I encode something with ‘mnemonic’ techniques,there is no guarantee that the encoded stuff will stay in my memory for a long long time. To solve this problem,I have to review the encoded stuffs time to time. And SRS technique is the best way to go for reviewing my ‘mnemonics’/encoded materials.

For me,SRS should always come second,not first. I tried SRS without the help of Mnemonics. And doing that,I didn’t experience the optimal learning that I do now(after combining both). Mnemonics should always come first and should be used in all cases!

Of course,Piotr Wozniak is one of my superhero in the world of learning! :slight_smile: