I’m really curious about what’s coming next in the world of AI and spaced repetition systems. Does anyone know who’s working on cutting-edge stuff and what they’re working on? I feel like AI is going to play a huge role in this, and I’m wondering if people are already developing systems that go beyond what we have now.
I’m imagining something like AI that observes your screen or activity, maybe using Retrieval-Augmented Generation (RAG), and saves everything as vectors in a database. The AI would not only know your knowledge base but also track your learning patterns and adapt based on that to give you the ideal material to review.
Could recommender systems also play a role here? Similar to how Netflix or Spotify predict content, could an AI system recommend the next card or concept based on what you’ve learned and where you’re struggling? I don’t know much about how recommender systems work, but it seems like they could be useful for making personalized suggestions in spaced repetition systems.
Does anyone know how a solution like this would be structured technically? What would the architecture look like? Is it just about vectorization and retrieval, or are there other advanced techniques at play here, like deep learning models for knowledge tracking, or complex recommender systems that optimize the learning flow?
I don’t know much about the technical side of this, but I’m really interested in learning more if anyone has insights or ideas!
The only person I know is Jarrett. Other than that, probably not much is going on.
I think the first goal would be to understand more of neuroscience. It’s a field barely 100 years old and we still don’t know a lot of stuff. Then maybe, yeah.
If there’s anything I want happening right now, it’s a math academy like “encompassing graphs” that connect flashcards and update a card’s scheduling when another partially related card is reviewed, and maybe we can use ML techniques to train a model to do that for us.
@Expertium Do you have any “vision” for what’s going to happen years later in SRS space?
Tracking changes in memory state outside of SRS. And we don’t need Neuralink for that. For immersion based langauge learning, we need an add-on+reader application that tracks what you are reading and update memory state based on that. So if I come across the word “deliberate”, the vocab card for the same word will get updated. Given that we have JPDB/LingQ and others who combine content with SRS, this is totally possible. It’s just that no one will do it.
For language learning, maybe the AI taking the role of a teacher. Rather than answering questions through a program like Anki, it will be an interactive conversation with the AI. It will build a picture of your vocabulary, use of grammar, pronunciation, strengths and weaknesses and lead the conversation based on that. It would give you novel, purpose-built questions and sentences tailored specifically to your needs at the current point in your learning. Giving it your Anki history could allow it to build a picture of where you are now.
Detecting Interference. It is a fact that learning things in a certain order would have different results than if they were learned in a different order.
Also newly learned information absorbed could adversely affect old learned information, which is why things memorized by heart are suddenly forgotten sometimes after taking up new information.
I would like to see this interference effect and the effect of time of the day on card retention be modelled one day, perhaps in FSRS 6. One day…
Well, the mathacademy example you gave is not actually AI. It’s more like SRS with some extra work. It’s possible in Anki but requires a lot of work, and I doubt it would be as effective for math learning as it is for language learning.
Andy presents a promising vision but I don’t think it’s possible for the next few years. The cutting edge AI model is still struggling in many things and you can’t trust it being a tutor, especially when you know little about the topic.
That’s the thing about mathacademy - their approach is far better for math than Anki’s “every card is independent from every other card” approach. They adjust intervals of card A based not only on explicit reviews of card A, but also based on “implicit” reviews of related cards B, C, etc. For highly interconnected knowledge (like math) this is much more efficient than Anki.
I agree with what you said about interconnected knowledge. Indeed, knowledge graph is essential in learning in the future.
What I mean, in short, is learning math in anki-way, like mathacademy, is only helpful to a certain degree.
Here when I say anki-way, i mean treat each knowledge separately and focus on memorizing. Mathacademy has made some progress like knowledge graph, related review and test by example, but you can see that it’s still in the scope of anki-way.
I would like to underline that learning has multiple phases and only through those stages one can construct an understanding, which is a coherent one rather than flat or graph-like. this is especially the case for math since it’s abstract and emphasize application.
Actually I’ve met several people complaining about anki being not useful at all for language learning. And I think that they are using it the wrong way, namely to treat anki like a magic memorization tool instead of a learning tool. In this sense matheacademy has a good start but i think it’s not enough for math.
I think the AI solution to such problems is to generate very powerful mnemonics.
E.g. the user puts what they want to memorize into a card then auto generates and adds (powerful) images, mnemonics, explanations, interesting stories, etc. In other words converting cards that have a difficulty level of 8-10 to about 1-2.
If so, learners will not have to work hard because they are easy to memorize or there is no learning stress.
If most cards are easy they may be able to remember them without the use of the spaced repetition algorithm. (e.g. students can get a high score if they cram right before exams and they remember roughly enough in the long run, so they don’t study more.)
There are already AI products doing [basic] conversation for language learning. Your ideas would be a way for such systems to develop in the future.
If I were them rather than get an idea of your current learning point by looking at Anki history I’d develop an in-system test to assess a person. That’s more generic, and also more commercially useful as it could form a periodic progress test for those who like to measure their progress.
Yes, a test would be another way to assess the current level. Maybe both could be used. The usefulness of the anki history would depend on how the questions are formatted and how much history there is.
Regarding ongoing testing, I think one of the strengths of an AI would be that it would monitor you throughout the conversation. Exams have always been rather blunt instruments. Some people are good at them and some people are bad at them. Exam results don’t always correlate with real-world ability.
I recently read this article from the CEO of one of the big AI companies (or rather I had my phone read it to me). If what he is predicting is anywhere near accurate, within the next decade, there is going to be a revolution, not just in language learning, but in every aspect of human life. He predicts that the AIs that will start the revolution will be ‘smarter than Nobel Prize winners across most fields’ and they could be available as early as 2026. If that’s the case then language learning could be changed completely within a year or two of that date.
Here is an ai-generated summary of his long essay.
This document is an essay titled “Machines of Loving Grace: How AI Could Transform the World for the Better” by Dario Amodei, dated October 2024. The essay presents an optimistic vision of how powerful AI could radically improve various aspects of human life within 5-10 years after its development. Here’s a detailed summary of the key points:
Introduction:
Amodei, CEO of Anthropic, explains that despite his focus on AI risks, he’s optimistic about AI’s potential benefits.
He aims to sketch out a world where powerful AI has positive impacts, while acknowledging the speculative nature of such predictions.
Basic Assumptions:
Amodei defines powerful AI as a system smarter than Nobel Prize winners across most fields, capable of autonomous task completion and interfacing with the world.
He proposes a framework for thinking about AI’s impact, introducing the concept of “marginal returns to intelligence” and factors that might limit AI’s effectiveness.
Key Areas of Potential AI Impact:
a) Biology and Physical Health:
Predicts AI could compress 50-100 years of biological progress into 5-10 years.
Anticipates breakthroughs in disease prevention and treatment, genetic engineering, and lifespan extension.
Envisions the elimination of most cancers, genetic diseases, and infectious diseases.
Suggests a potential doubling of human lifespan to around 150 years.
b) Neuroscience and Mental Health:
Expects similar acceleration in neuroscience progress.
Predicts effective treatments or cures for most mental illnesses.
Anticipates improvements in everyday cognitive and emotional experiences.
c) Economic Development and Poverty:
Discusses potential for AI to help distribute health interventions globally.
Explores possibility of AI-driven economic growth in developing countries.
Addresses concerns about inequality and the “opt-out” problem.
d) Peace and Governance:
Acknowledges that AI’s impact on democracy and peace is less certain.
Proposes an “entente strategy” for democracies to gain an advantage in AI development.
Suggests AI could potentially strengthen democratic institutions and improve governance.
e) Work and Meaning:
Discusses challenges of maintaining human meaning and economic relevance in an AI-driven world.
Explores potential economic shifts and the need for new societal organization.
Conclusion:
Amodei emphasizes that realizing this positive vision will require significant effort and struggle.
He argues that the vision, while radical, aligns with fundamental human values and intuitions about fairness and cooperation.
Throughout the essay, Amodei stresses that these outcomes are not guaranteed and will require collective effort to achieve. He also acknowledges the speculative nature of his predictions and the potential for unforeseen challenges. The essay aims to provide an inspiring vision of AI’s potential benefits while encouraging thoughtful consideration of how to navigate the transition to an AI-enabled world.
This is what AnkiHub is working on at the moment. The sky is the limit, but I think the first step is to figure out what the best user experience is for basic RAG and semantic similarity search functionality in Anki.
AI puts an infinitely patient expert tutor at our fingertips, but it’s not actually obvious how best to apply this throughout all stages of the learning process.
I love the ideas of creating ad-hoc, personalized mnemonics (@Shigeyuki) and finding related cards through vector search and using that for all kinds of things like adjusting intervals (@Expertium). So many amazing possibilities!
I gave an example of implicit reviews in langauge learning. The medical school equivalent would be to update DSR values of a card when a student say, watches a lecture video. That would be a bit hard though.
I think AnkiHub can do something like math academy with current technology. I’m not sure how useful that is for medical school as cards are not as interconnected as it is in math/science. But to even do that, would need experts in fields and students can not do it themselves if they are creating their own cards.
As Anki currently does not allow building “knowledge graphs”/“encompassing graphs” I think tags can be used to identify cards that are practising something similar and then using an add-on to add manual review logs updating all similar cards upon review.
