Joleen Liang: I would say a hundred percent of the parents, they, they are afraid of being left behind. So they are embracing technology. So everyone is embracing AI, so that's no issue at all.

They prefer to have AI to teach because it saves money and it can give your kids better improvements. And this, there's no place to find human teachers to teach anyway. Right.

Welcome to Manifold. My guest today is Jolene Liang. She is a co-founder of Squirrel AI, which is one of the leading AI education companies in the world. Jolene, welcome to the podcast.

Joleen Liang: Hi everyone. Hi Steve. I'm so happy to be here today with you.

Steve Hsu: It's great to have you. I probably should have introduced you as Dr. Liang because you have, I believe you have a PhD and your dissertation work is actually on the use of AI in education . Am I right about that?

Joleen Liang: Yes. My PhD was focused on intelligent science and systems, so building an AI adaptive intelligent learning system for students and teachers. Yes.

Steve Hsu: Right. And so when was Squirrel founded?

Joleen Liang: Our company, school learning, was founded in 2014, which is 11 years ago. It's really, you know, far and at that time, no people understood what AI was. So when we said Squirrel AI, people thought of one. You know, AI wasn't popular at that time. Yeah, 11 years already.

Steve Hsu: Yeah. 11 years ago you were supposed to use the word machine learning, instead of AI.

Joleen Liang: Well, yeah. Machine learning, deep learning, and then some algorithms, names and some other algorithms. Yeah.

Steve Hsu: So tell us about how the company grew and for the listeners, maybe I'll just say a few words about it. Squirrels have huge penetration in China, and so there are many, many thousands of locations. Where kids come in and they use the Squirrel AI platform to learn across a variety of subjects. And education outside the school system is a huge business, not just in China, but all over, all throughout Asia and increasingly in the United States. So Squirrel is one of the leaders in the space. It's an area that I've been interested in in a long time. 'Cause I've seen them, I've been thinking about the potential of AI for education for a long time. And most interestingly, perhaps Squirrel is expanding in the United States and getting ready for an IPO. And Jolene is the CEO of the US focused part of Squirrel AI. But let's, let's start with the early days, like when you guys were a scrappy little startup and

what, what, were your dreams of what you wanted to accomplish and, and what happened during that 11 years?

Joleen Liang: Sure. I would be happy to share the story. So the reason why we found this company was because we really wanna revolutionize the whole traditional education environment and we have four co-founders and the chairman and who is the founder, Derek Ha Lee. He has been in the education industry for 27 years now, and he had two IPO educational companies in the past before Square Learning.

But his mission is to provide personalized learning for students. So when he started to do education, he knew that. There was a huge shortage for good teaching resources. Teachers and students, they learn at the same pace, which is actually wrong. So he has been looking for something that can help him to provide a platform or product or company, which provides personalized learning for the students.

So in 2014, 13, he realized, okay, AI might be the future. But at that time in China, no one was talking about AI. Even in the United States, not many people talk about AI, but he found out in a report, some research that adaptive learning, adaptive learning technology and method was one of the key methods to provide personalized learning.

And there were only a few companies that started using AI algorithms to provide, to combine together with adaptive learning. So Derek thought, okay, let's do AI plus adaptive learning together to provide, real, the true personalized learning. Because before adaptive was kind of the regular adaptive students were, divided into different levels and groups, and they were learning adaptively. That's correct. But what Derek wanted is each different student has each different learning pathway. So with AI algorithms plus, adaptive, we did it. So at the very beginning time, we spent. Roughly three years building the product. When I say product, there are two main aspects. One is of course the AI adaptive engine, which is the platform, and the other part is actually the curriculum content design.

Because in the past, curriculum content design was in. The textbook wasn't really digitalized and there was a really big range of learning objectives. So with AI and the content design, two core teams started together, but. They had a huge fight because in the curriculum team, the designers were superior teachers in China. They were really, really like 20, 30 years experienced teachers and full of experiences.

They had their own way to teach and they didn't understand what AI was. And Derek requested to break down the learning objectives from 300 to 30,000 nano level learning objectives. Those superior teachers were crazy, you know, it was very hard for them. So they had fought with the AI team because the AI algorithm needed to combine together and link those really nano levels of the learning objective, which means in this way, the students will be learning only very precisely about their loopholes, their weaknesses.

But at the end we had this platform and then the content designing team, they were convinced that this is the correct direction. So after three years of the design, we started to launch the product. That was about 2017. We had a demo.

Steve Hsu: Can I jump in and ask a few questions?

Joleen Liang: Sure, sure. Please.

Steve Hsu: Already, in three years, you have covered a lot of ground. Right? So in Asia, the sort of classical cram school often has a kind of very charismatic teacher, someone who's really good at teaching kids and knows the material very well.

And sometimes they're wearing a headset and they're able to do problems and sometimes thousands of kids are watching them maybe online or on TV even. Or they're in the, in a, like a private classroom and the person's teaching. So there, I think everybody from Asia is familiar with that model of sort of after school cram schools where kids are learning extra stuff. And so it sounded like you had a team of people who came more or less from that background, but then you had another team that was coming more from kind of, AI applied to learning or adaptive personalized, learning perspective. And second team was trying to break the subject into really tiny, you know, extremely specific things that a kid would need to learn to say fully master algebra or fully master, you know, American history or something, and it took, I guess, three years to get those two groups to kind of see eye to eye and to build the infrastructure so that you could adaptively measure the learning of the child of, on each of these little micro goals along the way. And if the kid clearly understood everything except this one micro goal, then you would, you would kind of drill them on that

micro, goal in an automated way.

Is that how we should understand your platform?

Joleen Liang: Yes, you're right. Yeah. The difference between the well from the classroom is that it was done by the teachers, whereas in our product, our platform, everything was done by AI and the teachers, the superior teachers, they just designed the contents. So they are back on the screen. People don't see them. Right.

So they don't teach at all. So you are right. Yeah. The AI. The platform diagnoses the students' weaknesses and strengths and the whole knowledge state and the goal can be set up at the beginning, but the goals will be changed dynamically based on the student's learning behavior and the progress. So the granular level of the learning objective is very important that the AI can track, correctly, diagnose.

And also more important is once the AI understands, and then the AI will recommend the students, the very precise, granular level of their learning objective, the content, the problem exercise and videos watching and everything in a really, really precise way, then that will save time for the students. Yeah.

Steve Hsu: Yes. So the thing that you guys built over those three years, I think for people who are a little bit more advanced in their thinking, who understand AI and testing . Adaptive testing.That framework is very familiar as something you would aspire to, but I don't know of very many examples where someone has actually successfully built that whole system for a specific subject, whether it's, you know, modern Chinese history, chemistry, you know, fit, you know, algebra.

So it's kind of amazing that it's a company. So the reason I was so interested in Squirrel AI when I read about you is I thought, wow, this company, if what I'm reading is correct, they've actually accomplished this goal that many education researchers, people who are more sophisticated about AI technologies, but who are interested in education, they've been thinking theoretically that this is the kind of system that could be built.

But I don't know of a lot of examples, and so that's why when I heard about Squirrel, I really wanted to talk to you about it. At what point were you guys convinced that this was working? Because obviously no one else had built something like this before you. Right? During those three years, what was the journey like?

How did you convince the teachers who came from a different tradition that, wow, this is really working. This is really actually gonna be good for the kids. Maybe you can give us some color about that.

Joleen Liang: Sure. Well, I think the key person or the key element is Derek. He's the person. He always wants to do something that people have never done before. So he dug into this industry. I think at the beginning, before he started to build this company, he found some research papers, some efficacy studies from Western countries.

As I mentioned, that was about the regular level of adaptability of learning. So he was a hundred percent sure that adaptive learning is the future, but he just wanted to make it a higher level. So he was convinced of everything. He convinced the scientists and the scientists he recruited at the beginning.

They were from us. From Newton, from Alex. So they were the data scientists and designers of those adaptive systems. So they had full belief in adaptive learning, so that's okay with the scientist. And speaking of the curriculum, content designers, those teachers, that was hard. But the good thing is all good teachers, their goals are the same.

They wanna provide personalized learning for each different student. The teachers are good, but they understand one person, one teacher can probably cover 300 students maximum. They cannot help more, but they wanna do more. They wanna contribute more. But by what? By who? By person. No. No way. So it's gonna be by an intelligent system. So they had fights, they had arguments, but the mission is the same, the direction is the same. This is more important.

Steve Hsu: So during that three years, I, I'm guessing what happens is you built the product to the point where you felt you could launch. And when you were launching, were you launching in actual physical locations, like actual actual schools where kids would come and then use your, maybe your product on a tablet?

I think that's what I saw online to get to that goal. Did, had you already more or less done enough research that if you were at a university, you would've published like a bunch of papers saying, Hey. This learning system is, actually , extremely efficient, extremely good for the kids. Like had you already more or less done that level of research when you launched?

Joleen Liang: Yes, we, we, we've done some research at that time

and when we. Started to launch. I don't think we had our research paper published yet, but we did talk to the public schools and private schools and some after school tutoring centers and tell them that, okay, this is the difficulty you are facing.

Let's try something new. Of course, we won't charge it. So they organize, say, 500 students to use our product and to test out. Of course we will give some, we gave some, reward. Rewards to teachers, to students, to schools to organize. So all the results show that the students gain better results with AI learning.

And we did have another comparison group, which is human Teachers teaching. So of course, the AI group has better results.

So before the launch, we started doing an experiment demo with AB testing and after our product launched, we started to do AI versus human teacher competition, but it's more like an effective study.

So the very first AI versus human teaching was done in October, 2017. That's just right after we launched a system a few months later. We did it in Zhengzhou, China for four, full days. And I think there were about 40 media across the world who came to interview us because we did a very first, AI versus human teaching in the South in Asia Pacific.

I think that was very, it was a very hot topic of the alpha goal competition. So people wanna know, okay, how did you do in AI education?

Steve Hsu: Yeah, so it's kind of amazing because 2017, it's already eight years ago, right? And. What's interesting is that if you talk to US educators, or even like if I talk to someone who's at say, Khan Academy, probably you're familiar with Khan Academy after LLMs came out in the last year or two. Khan Academy has really been trying to use LLMs in, in what they do.

But, but, and I think Khan Academy is great and lots of American students, including my own kids, benefited from Khan Academy when they were in school. It just seems like China was actually more advanced because, you guys already had your platform running for the last eight years, and my understanding is what you guys are doing, that kind of adaptive learning, that breaking individual subjects into micro learning goals and then being able to test for mastery of those goals and then circle back if the student doesn't have mastery. That kind of adaptive system, I just don't know of that deployed at scale anywhere except, your example.

So do you feel like you guys have been way ahead. of everybody else

for some period of time?

Joleen Liang: Yes, we, we were and we are still ahead. Why is that? Because at the beginning we had AI plus adaptive, and later, after we launched our product, we noticed that in the market, in the western countries and also, probably mAInly in Western countries at that time, like some publishers, HMA, HMH MicroHill and of course Duolingo, DreamBox, they started to deploy adaptive learning, but it matters is the level of adaptivity, how granular they break down their learning objectives, how much AI algorithm they use, how accurate it can be recommended to the students, and how. Do the students perceive accuracy and the improvements, right and efficacy. So that's the, that matters. But anyway, those companies, they started using adaptive learning already. And in China, I think after 18, there's some companies, and even in Indonesia, in India, more companies started using adaptive learning. And at that time, in 2000. Now, I think 2020-19 we started to use LLM already.

And actually when I say we started to use LLM, we didn't use it. We designed our own large model, but it wasn't LLM. LLM is only 10% in our whole engine. So actually the current engine of our product is called LAM, which is a large adaptive model. So it is the combination of AI adaptive. Plus a multimodal large model, because when students are learning, when the AI is sending the information content to students, it must be multimodal.

It cannot be just language, right? It will be the speed, right or wrong answer, the learning behavior, voice eye, gazing time and concentration or distraction, and even when they use the pen to write on the, on the screen, it can be captured as the learning solution. So different students have different learning solutions.

The AI will analyze and also when they write the article will be AI assessment, blah, blah. So it is multimodality. So we haven't seen any companies using adaptive plus large models. They are either just adaptive plus some LLM or they are just LLM based. We don't think, especially for pre-K to 12 students, it's better to let them learn on their own with LLAM rather than to just give them the answers, the feedback, the solution recommendations, because they really need to think, they really need to generate, they really need to create their own mindset, not to be given some feedback or answers.

Steve Hsu: Wow. Okay, there's so much interesting stuff in what you just said. You know, when I said I didn't know of other cases where this kind of adaptive learning was implemented, I had forgotten about Duolingo because I was not thinking of it as a general education platform. But they do teach language. I know that Duolingo founders, actually, I think their Carnegie Mellon Cs, PhD, so they, they do know, they do understand this subject quite well. But, I didn't think about them because they're not teaching algebra or history. When you say multimodal model, have you guys trained your own or are you using something like an?

What is actually doing the multimodal processing in your, on your platform?

Joleen Liang: So we train our own. That's, you know, we spend lots of money on AI and data training and everything, so we didn't outsource many things. So we only, as I mentioned, we only use 10%, like outsource from LLM, those general LLM. So, the data that our system is training is mainly. The student's learning behavior.

But this is a, yeah, this is a very big name for learning behavior. So we need to break down. So what are that data, right? As I mentioned, when a student is recommended to watch a three minute video, it will be a teaching or animation blah, blah video. So it could be the student thought, okay, I understand, I've learned it.

I wanna fast forward. So this is the data, but maybe the next student just looks at it and watches. And another student also watched at the same speed, but the eyes were moving around, which means the student was distracted. Right? And another student may be just back forward because he or she didn't understand. So this is one scenario, but with different feedback that the AI will take into account and to analyze different students.

And if we have 1 million students, there will be all different feedback. And also that's only one data, like one aspect. And also older students, they have some backgrounds, right? History data. So based on this mastery level of that learning objective and even the prerequisite learning objective if the student has mastered or not, or maybe related learning objectives that are related to the current one, if the students mastered or not, you know, everything matters.

Everything that needs to be analyzed together. And also, the speed the time if the answer is right or wrong, right? And how many times the student relearn this learning objectives, how many problems or questions the student have practiced. You know, there all this kind of data. So we, we train and our system train itself and also learn those data from students and then train and learn from other students and train.

And also that's, that's the part about the students' learning behavior. But also we do, our system does train our own contents like the tagging. We had tagging for different contents at the very beginning by humans, like, by like manually to tag, but later on we could recognize okay, the tags may, might not be correct.

Right? So the AI will correct. So that's also based on the training. So yeah.

Steve Hsu: Very interesting. So I'm getting a better picture of your platform. I think you actually just said. That when the student is working with the tablet, the tablet is actually watching the student and doing eye tracking among other things, right? And so it actually sees the attention level that the student is paying.

And so it doesn't sound like you've trained independently on your own, what we would call a multimodal large language model, which is a language model that can do like audio, video, image, language, input, output, like that, that's a separate thing. And it doesn't sound like your platform actually requires you to have that.

Joleen Liang: Well, we do actually, yes, our platform does require that.

Steve Hsu: but do you, I mean, this is not a super important point, but I'm just

Curious, are you saying you guys like Bite dance and Alibaba and you know, Deepsseek? Well, even Deepseek hasn't actually trained its full multimodal model. But, but you, you're not saying you guys have trained a state-of-the-art

multimodal LLM, you're saying you use multimodal input in your LAM.

Joleen Liang: Yes. Yes, correct. Yeah.

Steve Hsu: Not to divert the conversation, but it seems like that based on the student data you have you could evaluate. So in the West, if you got a PhD in psychology or in education, you know, there's something known as the general intelligence of an individual which is heavily correlated to how fast they learn new topics. And if you have fully adaptive, evaluation of learning of a student, you can see the rate at which they learn things. And of course certain subjects they learn faster and certain subjects they learn less fast. But there's a general correlation between their so-called general intelligence and how fast they learn, averaged over, across, averaged across many subjects.

So it seems like your platform can, if you wanted to, I know this wasn't necessarily your goal, but you actually could. Extract that overall kind of average learning rate for each student and have an evaluation of what in, in the west we call their general intelligence. And I'm curious, is that something you guys have ever thought about?

Joleen Liang: Yeah, I think it's related as we do, we do have different subjects for students to learn, but it's. Also very important that we have a parallel system, which is called MCM, which is actually similar. MCM stands for methodology, capacity, and mode of thinking. So bad in this system we had different figures on their learning abilities, like comprehensive understanding abilities, critical thinking, and you know, all those different skills and abilities. So once they finish one session of the learning. Of course the report will show the mastery level of different learning objectives and content, but also there's a page showing the MCM elements, levels.

So once the students have been learning, say, six months in the system, their learning behavior, their learning abilities, and those MCM elements can be higher of course. Yeah. But it's not only the learning speed, it's not only the, the learning of levels, but also different elements because with MCM students and the parents can understand the students better, well, knowledge is very important, but the MCM, those abilities, skills are more important when they go to work right? So some students, they're thinking, okay, I wanna be a scientist in the future, but their math is really poor. Or maybe their logical thinking is really poor, but they just really wanna be scientists in the future. Then in that way, we can see from elements from MCM elements, okay, some specific skills and methods and abilities the student needs to train.

Then we know the direction. Or maybe I wanna be a dancer in the future, then I don't need to learn. I don't. But maybe I need to train my de dimension graph, understanding, you know, some other skills. So the MCM can diagnose different elements and to train the students towards their direction in the future.

Steve Hsu: So in, in the West, if you're a Western, if you're a researcher in a field called psychometrics, psycho means the brain, the mind and metrics means measurement. If you're a researcher in psychometrics. The main observation is those MCM elements that you're talking about. If you do factor analysis, if you do sort of principle components, decomposition, even though there are many different skills, like some may have to do with math, some may have to do with visual thinking, some may have to do with language. There's actually one largest general factor within there that's 'cause all these factors are correlated. They're not perfectly correlated, but they have some core positive correlation. And psychometrics is about extracting something called the general factor from that multidimensional space of what I think you guys would call MCM. So maybe offline we can talk about whether to do an experiment, whether you can analyze your MCM variables and extract what Western Psychometricians would call the general factor of intelligence.

Joleen Liang: Okay.

Steve Hsu: Yeah, we should look at that.

Joleen Liang: Yes, definitely. We would love to learn. Yeah. yeah,

Steve Hsu: Another way to turn it around is if you say, I have a student and she's, she's been in the system for six months or a year. I have enough data that I could actually predict. Maybe I should present things to her a little bit faster in this area or a little bit slower because I can tell what her rate of absorption is, the rate at which she learns in this area.

But, but that model, you're, you're predicting something about the person from the previous data that you've seen from the person that's, that's also related to quote, measuring intelligence or measuring, general intelligence. So yeah, we should talk about that.

Joleen Liang: Sure, definitely. Yeah. That's great.

Steve Hsu: Yeah so tell us something like, I think I read about this, but I'd like to hear it more from you like the scale of your operations in China is amazing, right? So tell us how many physical locations, how many students, how many different knowledge areas you're teaching.

Joleen Liang: Okay. So far we have about 3000 locations and about 1500, self study centers. And the others are more like a point of stand, a very, very tiny, small place. So with those, self study centers, that's a very, very new and different business model. We started a few years ago, so in those centers we have supervisors like two or three supervisors, but no teachers. Of course, supervisors are human rights, but they are not teachers. So this, and there are two or three or maybe five classrooms and the student can just bring their tablets into those self study centers and they just put on a headset and they can start to learn. And this classroom has different grades, students, and they're learning different subjects.

And so our business model is the supervision with data analysis plus. The students self study in the center or at home because we think supervisors and data analysts are very, very important to help the students. So that's actually our business model. But if a parent or student says, okay, I don't wanna go to the center, I wanna stay home.

Can I do it? Of course you can bring a tablet anywhere you want, but it's better that you have a supervisor or data analyst who can talk to you online like my daughter, she's in Melbourne and she's using our tablet to learn in Melbourne on her own every day. But she has a supervisor, a data analyst, who can talk to her on WeChat anytime she wants, so that person can help her and do analysis.

Hey Carrie, what's going on? And also some predictions AI can make. But with humans, with the data analysts, when they look at that data, that data dashboard, when they talk to the students, it's more easy to understand for us, for users. So this is how it looks. And so far, in the past 11 years, we've had 43 million users accumulated. Yes. And I think another data we are really, really, you know, we think is fantastic is the learning behavior data. We've got over 2 billion learning behavior data that actually makes the system really, really accurate and to make the students can learn, to get the improvements and yeah, that's, that's about, the business model.

Steve Hsu: So the 3000 locations that you mentioned are, are those. More like a traditional, I don't wanna use the word cram school, but are those places where there are also teachers as well as your software platform? Do you have any teachers?

Joleen Liang: No, no. no. Teachers, as I mentioned, supervisors. Yeah. And they were not teachers in the past. It could be you could be the supervisor. I could be, or anyone who has had training in our system can be well, supervisor and the roles of the supervisor too. One is when the students are there, they do emotional communication interaction or maybe organizing some games and break and to, keep an eye on the student's progress. And another key role is doing data analysis. In our whole company, we call them either supervisor or we call them the data analyst.

Steve Hsu: Okay, so if, if a kid is in one of your centers or maybe working at home and. You know, they're, they're trying to use the system, but they're confused about something. They have some questions, and the video doesn't answer the question, and they're kind of looking around and they can't find the answers to their questions.

The method you, the, what you want them to do is either to talk to the supervisor at the center or go on WeChat and talk to the supervisor online, but will the supervisor actually answer a content question? Like if the student has a really actual

content question. is there a way that. That gets answered without it being The AI,

the platform answer.

Joleen Liang: Good question. Actually, there are two ways to give the students feedback. The first

method would be actually the chat bot on the system. So the students should talk to the AI chat bot first about the contents, about any, anything. If the students think, okay, it cannot be solved, and they can talk to their data analyst, and for the supervisor data analyst, they are not allowed. To give any hints to the contents and especially they are not teachers. They cannot, right. There were other occupations in the past and they didn't understand how to teach math. Of course, they cannot give the answer in the content. Yeah.

Steve Hsu: Oh, okay. But, so let me give you an example. 'cause I, I guess I, I've not, I've

not taught a lot of little kids. I guess I taught my own kids, but, but mostly I'm teaching college students, right? Or graduate students. And it's often the case that, you know, they're reading the textbook, they're doing the problem set, and they have a legitimate question, which is really not answered well by the textbook, or even what I told them in lecture or, or in the solution to the problem.

They have a, it could be even a deep question. Sometimes students have actually very deep questions that are not in the syllabus. My role as the professor is that I actually can answer all those questions. If something like that happens in your system, what does the kid do? Like, I guess they could always go to their real teacher at the school and ask them, but, how do you guys think about those situations?

Joleen Liang: Okay. First of all, actually, in our system, all the

contents are. Breaking down into a really, really nano level of the content. So it won't be a really big question. And also in the process of learning, we have different processes and different sections, so it won't be the whole bunch of questions the student doesn't understand. So they will need to learn the video, teaching video, and they will do some exercise if the student is really frustrated and cannot move to the next step. Right. There are different analysis by the AI already, so we've already predicted

and what kind of difficulties the students may face. So the AI will provide as pure as possible already.

So I would say it can solve the student's problem, but if the students do have questions, then the AI chat bot otherwise. The supervisor can help. But basically, as I mentioned, the AI can give some hints about content because it's from the AI system, right? And, the data analyst cannot really give this content or knowledge based feedback, but AI can. And also, if the student cannot. Learn well or solve the problem and just stuck there, then the AI will skip this session. Say, okay, let's move to the next level, or maybe let's move to the next component. Then that's the time the AI understands. Okay,

Jolene is not capable of learning this knowledge. I need to give her some prerequisite or related or lower level, lower difficulty level of the learning component.

It could be one point lower, it could be 10 points lower. It really depends on the analysis by the AI. Then the AI will recommend that I learn something else. Then if, oh, I can master it, I can learn it. So that's kind of analyzed with adaptivity. Because our, yeah, because our students are not college. They don't really have very big questions. But for high school students, there are some, more difficult, components. As I mentioned, our AI has already broken down to a very granular level.

Steve Hsu: I could give you an example, like, my, when my kids were learning about evolution in AP biology. This question came up, you know, I don't know how it came up, like whether one of my kids asked it or some other kid, and it was, you know, what came first, the chicken or the egg. Right. And there is actually an evolutionary theory, this sounds like a paradox, but there's actually a correct answer in evolutionary theory because organisms that laid eggs existed before there were any birds or chickens or even reptiles or dinosaurs. So the egg definitely came first. And the first chicken was born from a parent that laid an egg. So the egg actually came before the chicken. But this is actually a subtle question and it's hard to imagine the AI or the, the syllabus or the video of the kid, you know, maybe the kid wouldn't have this weird question, but. but System. It's hard to imagine a system that's fully able to deal with things like that. Right. But I wonder if you guys are thinking that the kid always has a regular school and this is enrichment beyond the regular school, but they always have the regular school as the resources. Is that important to the way you Yeah?

Okay.

Joleen Liang: It's, yeah. So the students go to regular schools and they're a very small percentage of the students. They don't go to regular schools. Why is that? Like either, they are really, really poor students. They cannot understand anything in the classroom. Then they don't go. They will just use our system to learn.

Or another is, another extreme is the students who are really, really, really. You know, you know, higher level and smart. And then they choose not to go to school or maybe skip some classes like Derek's twin boys. Derek went to the Olympic mass when he was young. And his twin boys are super geniuses. They are in the best private schools in Shanghai but they. Skip. They don't go to school anymore. They stay home to learn two things. One is using a Squirrel AI learning tablet to learn all the subjects in order to, to, to go to GaoKao, the final exam to university in a few years. But another learning they do is the Olympics. And other programming by individual special classes or teachers.

So they wanna train themselves in a special way, but not to follow the regular progress of the classroom, even though that school is already the best private school. But they are geniuses. So two extremes. Yeah,

Steve Hsu: So the, uh, a kid who's off scale can accelerate using the Squirrel platform. I'm not

sure I

understood though. Did you say the Squirrel platform actually lets you practice for Olympiad Math or Olympiad

Joleen Liang: No,

No, no, no. Not, not yet. Not yet. but we are designing olympic mass programs with the Olympic mass team now, so we don't have it yet, but we are going to have it.

Steve Hsu: One, one of the reasons when I was introduced to you and Derek, I, first, I kind of wanted to talk to Derek because I knew that he was one of the top Chinese Math Olympiad competitors when he was young. So I was interested in that angle. But I also was interested in your angle because you're bringing all this to the United States.

So, So, we ended up you and me talking, but I would love to talk to each other. other. point, One of the things I should have said to the listeners is this guy Derek, is probably in some sense himself a genius and, probably that played a big role in building the system the first three years, because I have a feeling he had a vision and a lot of insight into how to build like this kind of adaptive system.

And that probably was quite important in realizing it. In China, what is the attitude of parents towards squirrels? Because I can imagine some parents saying, my kid really likes it, and at least according to the measurements, this AI system is really more efficient for teaching.

But I can imagine another set of parents saying, I just can't believe you're gonna, I'm gonna turn over. Even though it's not the mAIn school, it's the, it's the enrichment part of the, you know, afterschool program. But even so, I'm not gonna turn that over to some software platform. I have to have, like, the best instructor in Beijing teaching my kids, you know, at the cram school. What's the, what's the kind of attitude generally about acceptance level of, of your technology in China?

Joleen Liang: Nowadays I think. I would say a hundred percent of the parents are afraid of being left behind. So they are embracing technology. They, and they understand because of the policy in 2000, the double reduction policy in 2021, and, and also AI came to everyone. So everyone is embracing AI, so that's no issue at all.

They prefer. Right to have AI to teach because it saves money and it can give your kids better improvements. And this, there's no place to find human teachers to teach anyway. Right. So yeah, the attitude is positive and. I would say 80 to 90% of the parents in China, they prefer their kids to learn something else after regular school, no matter what.

Poor students or middle level students could be really fantastic students, but they always wanna learn at their own pace. So they definitely wanna choose something. So that's good in China's market and. I think it's more important when the student, sorry. When the parents, they wanna continue, the payment to our, you know, subscription or anything.

The results and improvements from the student. So the very first step when they started to think, okay, should I buy this product? Should I purchase the pro service? That's the point. When they see the analysis report from the AI system, they think, oh my God, that understands my kid so much. And they also ask their kid, how do you feel about it? Most kids, they'll say two things. One is, it's so fun. It's like playing a game. And I have to go one by one. The second they will say, it looks like this AI understands me better than my teacher. You know, something like that.

Steve Hsu: The value the kid can tell the evaluation is accurate. Like if there's some skill that they realize they don't quite get, the AI realizes they don't quite get it, and if they can do it with a hundred percent accuracy, the AI realizes that. Right? So, the kid really feels, the AI is understanding them.

Joleen Liang: Yeah, that's the first step. And then after say one or two months, the kids either get better scores in school or they have better improvements in the system, and also they can show, well, in the past I may have not really been into study, but now I can spend two hours studying or one hour studying. I like to study. I think I am confident and motivated. Then that's another improvement. So improvements, continuous improvements in involvement will make the final decision for the parents to continue. Yeah.

Steve Hsu: Great. So you mentioned the 2021 change in government policy, which, you know, I'm not even sure I fully understand. Let me tell you what I learned about it, and you can correct me because I don't have kids in China, so I don't. I don't really know how this worked, but my understanding is that Xi Jinping or the central government thought so. We don't want widening inequality, so we don't want these rich parents to be able to hire really excellent tutors and average people can't hire these tutors. And then it changes, you know, who can get into the top colleges. So they sort of got rid of, they sort of banned tutoring above a certain price or above a certain kind of

level and try to

kind of democratize the education process. Is that fair?What actually happened on the 22nd?

Joleen Liang: Yeah, not really correct.

So the goal, yeah, the goal of that policy was, to give learning equity to everyone. So the price really went crazy at that time, and not only rich families could afford it, but it was just too high. And that was the turning point that AI was kind of into, you know, the whole industry, but not yet.

So the goal, as I mentioned, was to bring equity to everyone and also, at a time, there are too many. Traditional learning centers across China, and some teachers are good, but only a small percentage of teachers, but some teachers might not be of good quality. We actually, we tested many tutors in the afterschool centers at that time, the mastery level of their middle school math grade seven, eight is under 60%, which mistakes cannot, you know, they are not qualified. So the quality of after school tutoring wasn't good enough at the time because of course we don't have that many good teachers anyway, so. The government wants to have better results, better quality, better environment for this whole market, for this whole country. So they decided to have that policy. Of course, it was really bad, you know, effects on the whole industry, and we also faced the same situation, but we were. Actually the company encouraged this kind of policy because we didn't have human teachers to teach. And we know that AI should be the future. And as I mentioned, all the teachers are behind the screen in our system. They don't teach in front of the classroom. So we actually encouraged, we think that's the correct decision, but the thing for us was bad because we sold our system to those traditional tutoring centers and they were affected, so they had to shut down. So we had our revenue really go to zero and we had a debt, and then now we did a pivot and everything is fine now.

So we came back, but we think the policy was good. It should have that policy. It's just too strict and too quick. Just, you know, one day and yeah, that was kind of a dark time, but well, we wanna improve the whole educational environment. We have to sacrifice something and do something new. Yeah.

Steve Hsu: But just to clarify, so in 2021, the government sort of abruptly, as you said, they didn't shut down after school learning centers. Do they just shut them down? Or did they put a cap on what you could

Joleen Liang: Shut down.

Steve Hsu: They just shut them down. Oh, okay. So they just removed the license, like you couldn't, you're no longer allowed to, and so. My understanding is there's still tutors who will

go to people's houses and teach their kid, but maybe that's, is that illegal now that that's

Joleen Liang: It was illegal after the policy in 2021. And it is still illegal now.

Steve Hsu: got it. But it happens a little bit, right?

Joleen Liang: Yeah.

Who knows? As long as nobody knows.

Steve Hsu: But. This whole situation now should be great for you, right? Because if, if the goal of the government is equity, you can deliver, you know, if there's some very poor school district

in Xinjiang, you can deliver your product at a discount to them. It doesn't cost you that much to deliver

Joleen Liang: Exactly. Yeah.

Steve Hsu: I think the way to get to equity, whether it's in the US or China, is actually to make these really excellent learning technologies available to everyone.

So does it, is the government actually kind of supporting you guys now? Like did, are they on your side? Yeah. Okay.

Joleen Liang: Yeah. Actually the government, the local government in Shanghai and also the central government disappointed us the whole time. And we just had two official donations to the Tibet area and also the Qinghai province area and was organized by the central government, the mini ministry of education in China.

So Derek went down the two trips with the ministers together and to give the donation and more importantly to help the, you know, the rural area kids to learn. So that's it. The goal is to bring learning equity. And before that, we've, we've done many donations and help with rural, remote areas of the students.

They don't have teachers, maybe just one teacher teaches all the subjects. Right. So we gave our learning account to the students and they got really, really, really great improvements. Yeah.

Steve Hsu: That's great. Yeah, that's fantastic. So we're nearing an hour. It went by really fast, like

Joleen Liang: Yeah, it

Steve Hsu: Maybe we could just talk about the challenges and the opportunities for you guys in America. So, to me, I would say it sounds, it's great that Chinese families and parents. Like your system and understand its advantages. I would guess lots of Americans will not like their kid being taught by a computer, and even though maybe by showing them lots of statistics or evidence, you can prove to some of them that the system is better. A lot of them will just have some kind of psychological or emotional barrier against letting the kid learn this way. And I'm, I'm curious what you're experiencing.

Joleen Liang: Yeah. for the US market, we are still under the content localization process, so we haven't launched a product yet, so hopefully next year, but I believe we are going to face this situation. So actually in 2017 to 2020 in China, we also had the same problem. It's just now we don't have this problem, but we had before.

So the first important thing is. You have to have a human supervisor in the center to help the students, that's for sure. And secondly, you have to see, you have to show the improvements to the parents. Like I mentioned, the two, really, really great and necessary feedback from the students. Then they will think, okay, that's the future.

And I understand there are many Americans, parents, they don't want their kids to be, you know, holding their phone and tablet and laptop all the time. But, it is the trend that everyone is using the laptop to finish the homework, right? So we have to teach the whole market. I know it's gonna take time, but it's good that as long as we can show the results and before we are going to launch our product, we will also organize a few efficacy studies in the US market.

As I mentioned, we've had many in the past, over 100. Efficacy studies, but they were all in China. So we're gonna, do the same as we say, to show that, okay, the students gain better results, blah, blah, blah, blah. They love learning. They started to improve their MCM abilities. And I think that will be another evidence.

And, also when we show and promote our product to the parents, we will tell them that, okay, all the contents are designed and localized based on the US teachers. And we had great content, designing teams and some of them were award-winning teachers from different states. And one of them, our brand ambassador Alex, he's also the year the whole.

State, award-winning teacher. So they understand the students and they believe in AI, right? So when the students use AI, they don't. Just receive some text or some PP, uh, PowerPoints with some words. They receive animations, they receive a teacher teaching. So it is more like a real, you know, learning atmosphere.

So that's what we're gonna do. So yeah, I think it's gonna be a huge challenge, but, it's the future, it's the trend. Or just take some time to educate the home market.

Steve Hsu: It's great. I mean, you know, because these subjects that all kids have to learn between kindergarten and grade 12 are really the same subjects. They don't change that much. Really building an infrastructure that has value, it's gonna have value for hundreds of years. Like the, the, the content and the ability to measure learning outcomes in that space of content. It's not gonna suddenly become useless 10 years, 20 years from now. It's gonna be useful

forever because every kid has to learn chemistry. Every kid has to learn

algebra. So it's, it's worth doing right. And putting the energy into it that you guys are doing.

Joleen Liang: Yeah.

Steve Hsu: Great. I just wanna say one thing to you that actually you're a PhD, so I think you'll, you can understand this. Many Asian kids become nearsighted.

Okay. I'm very nearsighted. Okay. I think I'm like seven diopters or something. Right? Because they, as they're developing their eyes, don't get enough natural light. So it turns out our evolutionary system is anticipating receiving natural light, which includes ultraviolet light, and it's actually higher intensity than what you get indoors. And so when I was growing up, I kept wondering like, why am I so nearsighted? Is it genetic or is it environmental? And it turns out there are a lot of studies, actually a lot of 'em in Australia, which show that kids who just get exposed to enough natural light, which you would get if you went outside and played. Regularly in the sunshine do not become nearsighted. And so when my kids were born and I still wasn't sure whether this condition was mostly genetic or not, I had read these studies. So when my kids were growing up, I always made sure they went outside and got like an hour or two of natural light. All throughout their childhood and neither of them are nearsighted.

Neither of them even really need glasses, which is amazing 'cause I'm very nearsighted and my wife is also nearsighted. So, one of the things I, I think you guys could do on your platform is if you somehow. Remind the kid to go outside, and get some sunlight.

And also I would even tell my kids, do something that requires you to focus on objects which are really far away.

Because if you're reading all the time, like smart kids like to read all the time, right? And these days kids are like looking at their phones or their, or their Squirrel, AI tablet all the time. That's fine, but they need to spend at least some time focusing far away and receiving strong natural light. Then they won't become nearsighted. So actually this is something you could do to improve the lives of millions of kids if you embed something in your platform that gets them to do this every day. I think the long-term benefit is really good. So I can talk more about this offline, but I actually did this experiment with my own kids.

So my kids do don't need glasses and,

uh, yet my wife and I are

really nearsighted because we grew up at a time when nobody knew about science, right? So, yeah.

Joleen Liang: Thanks for that suggestion. And I think that's great and we really care about the students' eyes as well. So all the screens, the, the, the, the technology we use for the screen is true with the highest protection, the highest standard,

Steve Hsu: Yeah.

Joleen Liang: international standard. But I agree that they need to go out and look out and look away.

I totally agree. That's if they stay in the center, the supervisor will do that. To encourage them, but, yeah, a notice and reminding is a really great suggestion. Thank you.

Steve Hsu: Yeah, if you, every now and then, just tell them like, if you don't wanna have bad vision, do this every day. And then like, you know, I think that that's very valuable, like not very many people actually know this. I have regularly asked eye doctors this question my whole life. I've asked, I just, every time I see a new eye doctor, I ask them like, well, how much of this is genetic and how much of this is controlled by my behavior and, and actually only a few people are familiar with these studies, but I'm really glad I read these studies before when, before my kids were born because now my kids don't. So,

Joleen Liang: Yeah.

Steve Hsu: Yeah. Anyway, so, Jolene, it's been fantastic talking to you. And actually there's a bunch of scientific things maybe offline we can, uh, maybe discuss further, like about measurement of intelligence and stuff like that. I'm sure my listeners will really enjoy this interview, so thanks a lot.

Joleen Liang: It's my pleasure. Thank you. Very nice talking to you, Steve.