Exponential adoption of spaced repetition
This text is part of: "History of spaced repetition" by Piotr Wozniak (June 2018)
Slow start of Algorithm SM-2
Algorithm SM-2 was first used in learning on Dec 13, 1987, and with minor tweaks survived to this day in a number of applications. SuperMemo abandoned the algorithm in 1989, however, Algorithm SM-2 keeps popping up in new applications with a frequency that must be approaching a few new developments each month. I lost count long ago. Some of the mutations contradict the principles of SuperMemo, and still take on its label. Most often, the violations include intervals measured in minutes, or halving intervals at fail grade (Leitner style). These mutations also lead to some fake news about SuperMemo. Note that fake news was one of the greatest incentives for writing this article.
When Duolingo speaks in their paper of hand-picked parameters in reference to SuperMemo, it must be a result of relying on some older texts, perhaps second-hand texts, perhaps texts written in reference to Algorithm SM-2. After all, SuperMemo was pretty adaptable as of 1989 and Algorithm SM-17 is the most adaptable specimen in existence.
Some of the blame for misinformation is mine as I stopped caring about peer review, and let the information wild on the web with insufficient mythbusting effort.
The first applications to use Algorithm SM-2 were non-commercial offshoots of SuperMemo for Atari in the 1980s. Later, minor clones of SuperMemo (e.g. for handheld computers) opted for variants of Algorithm SM-2 with various own innovations, of which many provided painful lessons on the impact of disrespect of memory in the name of cramming.
By 2001, SuperMemo World moved ahead by five major generations of the algorithm. All major software lines, incl. on-line SuperMemo and SuperMemo for Windows adopted the data driven variants of the algorithm. supermemo.net became one of the pioneering e-learning platforms, and has now evolved into supermemo.com. SuperMemo for Windows pioneered self-learning solutions such as incremental reading, sleep-and-learning optimization, or neural creativity. In the meantime, Algorithm SM-2 became an easy first-choice option for other developers.
1998: publishing and acceleration
On May 10, 1998, Algorithm SM-2 was opened to the public and published on the web here.
Mnemosyne was first to pick the tool as the offshoot of neural network MemAid created in 2003. As of 2006, Mnemosyne keeps collecting repetition history data running a mutation of Algorithm SM-2. As a free multi-platform application, Mnemosyne quickly reached a large base of users, e.g. on Linux, or those users who have Latex requirements.
Anki was born on Oct 6, 2006. It was based on Algorithm SM-2 and for nearly a decade provided the widest reach for the algorithm. It is still going strong. Anki introduced a great deal of innovations into their algorithm but refused to advance beyond its basic principles (see: criticism of SM3+).
In 2007, when we met Gary Wolf, SuperMemo looked like a sad deserted island that begged a question: if it is so good, why others don't try to copy the algorithm. Anki and Mnemosyne were little known at that time. Wolf's article in Wired in 2008 caused a nice rush for education software developers to implement a form of spaced repetition. Algorithm SM-2 seems like a low fruit to pick and its expansion accelerated. Many users of SuperMemo claim they would never find the program without Wolf's article in Wired. Krzysztof Biedalak likes to joke though that Wolf's article was indeed a breakthrough. However, it was not a break for SuperMemo. It simply opened the floodgates for the competition to rush in into the field of spaced repetition.
2008: explosion
Quizlet was written in 2005 and released in 2007. It was initially a typical cramming tool, however, by 2015, backed by venture capital, Quizlet announced a higher emphasis on long-term retention, which resulted in adopting a variant of Algorithm SM-2. By 2017, they decided to use machine learning to deploy a new algorithm that would capitalize on billions of repetition records collected. The short stint for SuperMemo at Quizlet must have given a mutation of Algorithm SM-2 an exposure to the largest user base ever. At the time, Quizlet reported reaching every second high school student in the US.
The new approach taken by Quizlet is based on a strong foundation, and can lead to a very strong tool, however, this is very disappointing to hear the motivation behind the move towards better algorithms: "Cramming is a reality for many students, and we want to help them make the best of their study time however they spend it". Algorithm SM-17 provides for more freedom to students: (1) to advance learning when in need, or (2) to delay low priority material. However, we always discourage cramming as a bad practice. It is schools that need to adapt to human brain, not the other way around. This stubborn stance on learning efficiency hurts SuperMemo, but it will never change.
That move away from a simple review schedule by Quizlet in 2017 is probably the move past the peak of popularity for the old venerable algorithm. New competitors will need to go for intelligent tools, or perhaps for licensing Algorithm SM-17. The news is good.
How many people use spaced repetition?
In mid-1991, one of my classmates tried to cheer me up. He predicted we will be successful and we will manage to sell 10-20 copies of SuperMemo. I was more optimistic. In 1993, I predicted 1 million users by 1996. In 1994, Enter, Poland, mentioned similar optimism of Marczello Georgiew:
In questionnaires received at SuperMemo World, when asked what they like most in the program, users of SuperMemo overwhelmingly indicate its effectiveness. The software may be OK, but what really counts is results in learning. How about dislikes? Users are not pleased with this or that, most often with the fact that, even in Poland, SuperMemo is always released first in English. But there is no particular turn-off that takes precedence. Definitely, nobody questions the fact that with SuperMemo, one can learn faster and never worry about forgetting. Taking this rosy picture into heart, one might wonder why has SuperMemo not yet sold in millions of copies worldwide. Marczello Georgiew, Marketing Director at SuperMemo World proposed to recall the problems Graham Bell experienced when trying to introduce his funny machine for talking over a wire, or how pessimistic the predictions of industry futurologists were about the expansion of the air-polluting mechanical horse. Then he adds confidently: It took Wozniak 10 years to turn necessity into invention, give us half this time, and we will turn his invention into a global necessity.
In my 1 million users prediction, I was off by 3 years, and had to make a distinction between short-timers and active users. The proportion of active users of spaced repetition kept dropping with wider adoption. In 2007, we estimated the reach of SuperMemo to be 5 million, of which most were freeware and partwork users. Of those 5 million, only 0.4-4.0% were active users. This might have been as few as 20,000 students.
In 2009, Gwern Branwen estimated the population of active users to be around 100,000, which seems to agree with my numbers. This does not sound too optimistic for two decades of hard work at SuperMemo World.
Let's then have a closer look at the reach of spaced repetition today. My estimates below met with a great deal of skepticism. I agree that they are based on a great deal of guesswork. However, once you are on an exponential curve of growth, even large estimate errors make little difference. You can overestimate by 200% and still catch up quickly in no time.
This is why I do not hesitate to say that the exponential growth in the adoption of spaced repetition streaks towards the big B: one billion users. Amazon's Kindle has added spaced repetition to its Flashcard option in Vocabulary Builder. Even users of SuperMemo who use Kindle may know nothing of the fact. Flashcards with books is the general idea that was to bring SuperMemo to NASDAQ, back in 1996, if we only succeeded in persuading venture capital that the idea made sense.
However, to hit a billion users we need another breakthrough. The first obvious candidate that comes to mind is Facebook, which might wire spaced repetition into the cacophony of social interaction, and make free learning transparent, i.e. where users learn without ever showing intent.
If you think Facebook and spaced repetition are incompatible worlds, consider the world of advertising. These days we all hate advertising. No matter how well it is targeted. However, the pestering party can maximize the memory effect and minimize the annoyance (i.e. retrievability) by employing spaced repetition. Even the most captivating TV advert will get on your nerves by the third exposure. Spaced review could ensure that retrievability is low and retention high.
Last but not least, spaced review may be taken on by the bad guys: the makers of fake news and worse. A publicity charlatan might pull strings behind the back of a world leader. He may shake the world in spaces. This may expose the whole world to spaced repetition to be sure we all remember.
The top of the pyramid is so bad that I will not even list it. I don't want to give bad guys any ideas.
My estimates below include a couple of points that are pretty certain. The first user in 1985, second in 1987, one million by 2000, and my laborious estimate of 5 million in 2007. Today, Duolingo claims 200,000 users. Quizlet claims even more. The growth is still showing few signs of saturation.
Figure: We expected spaced repetition to show signs of saturation long ago. However, through transmutation, it will inevitably hit a billion users at some point. Once it becomes integrated with human digital life, it will affect nearly everyone. If my estimate is right, the speed of adoption, aided by the web, is still ahead of the telephone, car, and the radio. We never thought it was possible to compete with Pokemons or Angry Birds though. The exponential regression formula in the graph is: Reach=exp((year-1984)*0.63). The red line determined by that formula crosses 1 billion just about now
Today, with almost no barrier to entry, there are many students who try and drop out after weeks or even days of use. The proportion of active users may be very low. A billion users with negligible learning is still little learning. The next step in the job is to produce a cultural paradigm shift that will add value to efficient long-term learning. We need to begin with a change to the system of schooling and to adopt the principles of free learning.
The road ahead is still very long.