> But drawing on his experience creating Khanmigo, an AI-powered tutor, Sal makes a compelling case that AI-powered technologies will be different. That’s because we finally have a way to give every student the kind of personalized learning, support, and guidance that’s historically been out of reach for most kids in most classrooms. As Sal puts it, “Getting every student a dedicated on-call human tutor is cost prohibitive.” AI tutors, on the other hand, aren’t.
Personalisation fails for a simple reason - people don't really want to learn.
Classes work because the teacher can inspire (or at least push), and there's peer pressure to learn (or at least keep up with the herd). I'm sure there's a dozen or so self-taught Python or Rust programmers here who will loudly refute what I'm saying, and point out that they were perfectly capable of learning something they were very interested in, but I bet a lot of them would also like to learn a foreign language or quantum physics and haven't gotten that done.
AI will fail for the same reason Youtube, DVDs, video cassettes, radio lessons, and phonographs all failed to be a revolution. If you want to learn something, and have the motivation, then reading the textbook is easily good enough (for theoretical subjects) and for less theoretical subjects you barely need a textbook, just lots of practice.
Yes, you can probably learn a little tiny bit faster with AI, if you (or someone staring over your shoulder) have the motivation to play some AI learning game rather than a more fun-optimised game that's purely about having fun, but it's a small optimisation.
> Personalisation fails for a simple reason - people don't really want to learn.
Students/people who are interested in learning a particular subject are doing just fine learning it from already available material, be it books, videos, MOOCs, etc. They may or may not use AI, but AI won't make a huge difference for these type of people.
For people who are not interested in learning, AI won't do sh*t. Because the problem is not that they don't have a tutor available to them; the problem is that they don't have motivation to learn. I am not being judgemental here: it is perfectly understandable, for example, why a poor kid may not have motivation to learn.
A few years ago, MOOCs were touted as the next best thing to democratize education and make it available to underserved populations around the world. This did not happen. Some research shows that MOOCs are mostly used by an already-educated group of people who are interested in further learning. I think learning with the help of an AI tutor would be no different: its users will not be the intended ones.
Before MOOCs, it was TV that would revolutionize education, and before that, radio was supposed to replace the need for classes/lectures.
What each of these (including the latest wave of "AI" approaches) has overlooked is that humans learn from humans, because learning is only accomplished via motivation. Creating and sustaining motivation for years is not within the scope of a program you use on a computer -- it comes from a teacher who knows how to inspire, from a classroom and home environment that reinforces the importance of learning, and from peers who also care and reinforce this culture.
I'll quote Kentaro Toyama:
"Quality primary and secondary education is a multi-year commitment whose single bottleneck is the sustained motivation of the student to climb an intellectual Everest".
I'm going to be straight forward with you: You're wrong. Not everybody can sit their butts down and go through every word in a textbook. Watching videos on Youtube helps a lot, along with other methods.
Don't believe me? Look at those school tutorials on Youtube. Look at all the math videos, science videos and other topics. And I'm not talking about people who build cool DIY stuff and explain it at the surface level. I'm talking about people going through a school curriculum. There are kids who actually watch this, believe it or not, and it's wonderful.
This counters your statement which is "people don't want to learn". That's simply untrue. Everyday we are learning something new. You may not want to learn whatever they're teaching in school, but you will want to learn other things. And I think a personal tutor, especially a cheaper one, like Khan's AI tutor, will help a lot.
> If you want to learn something, and have the motivation, then reading the textbook is easily good enough
Anecdotally, I found it dramatically easier to learn math through Khan academy than from textbooks. For one, I have terrible handwriting which makes the pen and paper part messy and hard to read, and for two, Khan's lessons were broken down into much smaller, more digestible steps.
I took discrete math at Uni and had to learn it from the same Rosen textbook that many on this thread probably did themselves. If I had the same thing in an online version analogous to Khan Academy, I think I would've had a much better time. Now obviously, some of the harder problems and proofs in that book are much harder to grade automatically the way Khan does for grade school math, but that seems like something LLMs are trending towards being able to do
Yep. When I first saw how great AI was at answering my questions, I imagined myself talking to it every day to learn everything I'd ever wanted to know.
Turns out, what I actually want to know is just that small subset of things I need to know to get by, plus maybe a little extra for variety. And I consider myself intellectually curious. I have multiple college degrees, one of which I got purely for the fun of it.
So if I can't bring myself to get AI to teach me things, I can't believe that those less intrinsically motivated will be able to either.
Although I am not completely on the 'people don't really want to learn' train, I acknowledge one of the biggest problem in education in aligning incentives. No one is able to answer an innocent question by a kid about why do they need to learn what they are learning in class, without sharing an abstract into-the-future explanation, that the kid cant relate to.
On the topic of AI tutors will change the education, it reminds of this book i recently read (https://mitpress.mit.edu/9780262546065/teaching-machines/). People have been trying to sell similar ideas, in different packaging(technology) for decades now. Video technology had the same pitch for how schools suck, teachers dont want to teach and we will revolutionise the access and personalisation to education. While videos, Sal's or YT in general, have done a great job(specially at accessibility), our education stats are continuously deteriorating.
Before the onset of LLMs we created an instant-tutoring service(https://askfilo.com/) that could very much connect students to teachers in real time. Thousands of tutoring session laters we are still learning what students want/need from a teacher, 'they want to learn' but not everything or every subject, 'they are very much capable of learning' but our assessment are built very rigidly.....learning to test well is a learning on it own. There is so much understanding we lack in terms of mass education and it is not because of lack of good content or tooling.
This is the most nonsensical statement I’ve ever heard. I don’t know how someone could actually believe this.
As someone who loves to learn, I’ve already been using AI to learn things I never had the patience for in school. I HATED classes. Learning at my own pace, in my own way with AI has been an incredible experience. It’s far from a “small optimization”.
We used to go to our IT teacher after school to help us find programming exercises because access was so hard. It's not some sort of exception either, kids would actively reach out for more learning but providing this opportunity is super hard.
Now with AI it's right there! I'm increadibly jealous to the point where OP's statements are making me actively angry.
> , but I bet a lot of them would also like to learn a foreign language or quantum physics and haven't gotten that done.
Learn a foreign language from school? How much do I know of the German/French/Spanish (it was one of them, one of the options) that I studied for years in school? None.
For some subjects, something being taught in school seems like a magnet for making a topic or subject boring and uninteresting. Like if it wasn’t taught at all people would at least not be actively biased against it and might pick it up by happenstance.
Which makes me doubt all of those “no one knows X; therefore we should teach it in school”. I don’t think that follows.
> Personalisation fails for a simple reason - people don't really want to learn.
What does (do? Lack of school English) everyone have in common? Schooling.
> If you want to learn something, and have the motivation, then reading the textbook is easily good enough (for theoretical subjects) and for less theoretical subjects you barely need a textbook, just lots of practice.
This may be true for the average student, but I don't absorb content from written text very well at all. The single biggest reason why I got A's through college was being able to listen to video lectures (YouTube, Khan Academy, and occasionally the professors themselves) at 2x speed. I absorb and retain information that way extremely easily, but put me in front of a textbook and I can't focus on it.
YouTube may not have been an education revolution for you, but it's definitely changed my life.
Python, foreign language, physics? I've got all three self taught!
I agree with you. Investing hundreds or thousands of hours of your own time into the slog of theoretical learning is statistically pretty rare. Most of the counterexamples come from people who found a way to finesse their own neural latticework into ardor for the subject, even though to the rest of mankind it's about as interesting as watching paint dry. It's very rare to meet someone who became a self made master of something they only did for the money, etc.
I totally agree, and it's not even an issue of motivation.: School gives the time, environement, motivation, support and syllabus for the children to learn. It may not be perfect for everyone, it may be improved, it is not egalitarian but it is the best way to give education to the children of a country.
AI like all other technologies that were supposed to revolutionize teaching and failed are a dead end to educate a country, even if they can be potentially useful for some people.
AI tutors can work around student engagement, mostly by making tedious theoretical memorization into small practical problems, trivias, "hand in hand" learning and active recall keeping the average student focused for more time
This is absolutely not true for kids in the range of maybe 9-13 or something. I have kids. I learn via text books. They just can't do it yet, and they are a good way ahead of most of the kids in their grade.
> Youtube, DVDs, video cassettes, radio lessons, and phonographs all failed to be a revolution
Hard disagree. Young people may never know the struggle, but using an old book to teach you how to build your first {$project} was a really crappy gamble back in the day. Oftentimes the results of your efforts were mediocre or just flat out sucked because you tripped over every gotcha possible.
Thanks to the pandemic and WFH, I was able to spend hours every week on Youtube watching master craftspeople from around the world teach me skills I've always wanted to learn. It's unbelievable how accessible learning a new skill is compared to even a decade ago.
The people for whom personalized learning works well, can do it with any technology, even books. I recently learnt a new programming language from an O'Reilly book just fine, then did some exercises in a (non-AI) online REPL/tutorial tool, then went off to code a project of my own.
On the flip side, I've absolutely never been able to pick up a new programming language from a book.
Most of the languages I've learned I started with YouTube video tutorials, but most recently when I wanted to learn Rust do you know what worked really well for me? GitHub Copilot. I started a new project, and when I didn't know what syntax to use I made a comment and examined the output. By the time the project was done I'd learned how Rust worked and could write code without help.
Before making sweeping pronouncements about what does or doesn't work for learning, it would be worth considering that people vary widely in which learning strategies work well for them. You might just be projecting your own preferred learning style out on the world.
Personalisation fails for a simple reason - people don't really want to learn.
Classes work because the teacher can inspire (or at least push), and there's peer pressure to learn (or at least keep up with the herd). I'm sure there's a dozen or so self-taught Python or Rust programmers here who will loudly refute what I'm saying, and point out that they were perfectly capable of learning something they were very interested in, but I bet a lot of them would also like to learn a foreign language or quantum physics and haven't gotten that done.
AI will fail for the same reason Youtube, DVDs, video cassettes, radio lessons, and phonographs all failed to be a revolution. If you want to learn something, and have the motivation, then reading the textbook is easily good enough (for theoretical subjects) and for less theoretical subjects you barely need a textbook, just lots of practice.
Yes, you can probably learn a little tiny bit faster with AI, if you (or someone staring over your shoulder) have the motivation to play some AI learning game rather than a more fun-optimised game that's purely about having fun, but it's a small optimisation.