Hello, and welcome to Decoder! This is Hank Green, cofounder of Complexly, where we make SciShow, Crash Course, and a bunch of other educational YouTube channels. I’m also an author, a TikToker, and what you might call a poster — you might have seen my face on the internet over the years. You might also remember last year when I turned the tables on Nilay and interviewed him on his own show, because what better Decoder guest than Nilay Patel? That was a ton of fun, and it was so much fun that they’ve brought me back again. This time, I’m stepping in for Nilay to host a few Decoder episodes while he’s out on parental leave. And because I cannot interview Nilay — that would defeat the whole purpose — I instead found some really great people to bring on the show to have conversations with. Listen to Decoder, a show hosted by The Verge’s Nilay Patel about big ideas — and other problems. Subscribe here! Today, I’m talking with a very special guest — someone who I’ve known for quite a while as what you might call a colleague in the online education community: Sal Khan, the founder and CEO of Khan Academy. Sal was actually Nilay’s second guest on Decoder, back in 2020. And, well… a whole lot has changed since then. You’ve probably heard the name Khan Academy by now. The company was officially founded in 2008, but Sal actually posted his first educational videos on his YouTube account in 2006, just a year after YouTube was created. He even beat my brother John and I by about a year to his very first YouTube upload — so that’s impressive. Sal has been around for a long time. He’s seen the growth of the online video ecosystem, alongside the online education industry, up close for nearly two decades. That’s precisely why I wanted to have him back on the show: I wanted to ask Sal not just what it’s like running Khan Academy, one of the biggest and most well-funded educational nonprofits in the world, in the aftermath of the covid-19 pandemic, but also how it’s about to change due to artificial intelligence. I’ve been thinking about the intersection of AI and education pretty deeply for a few years now. And as we’re coming up on the three-year anniversary of ChatGPT’s launch, it’s something that has me equal parts fascinated and terrified, especially because it seems like it’s moving awfully quickly, and we’re only just starting to really grapple with the effects of this technology on the classroom. I think you’ll find some really surprising answers from Sal on these very hard questions, and I hope you’ll learn a lot, too. I know I did. Okay: Khan Academy CEO Sal Khan. Here we go. This interview has been lightly edited for length and clarity. Sal Khan, you are the founder and CEO of Khan Academy. Welcome back to Decoder. I am so happy that I get to talk to you, and the only thing that I’m not happy about is that I don’t get to do it for three hours because I feel like there’s so much I want to ask. I think that people probably have heard of Khan Academy because now it has been around for a very long time at different moments of its history, or they’ve interacted with it in different moments of its history. Can you just give us a look at what Khan Academy is in 2025, which might be somewhat different from a video that someone watched at some point? I think people don’t know how big you are, basically. Oh, yeah. In some ways, the true north of Khan Academy has been surprisingly consistent. But yes, the way people might perceive it has maybe changed. If you were to go back 15 to 20 years, people might associate it with math videos that some guy made for his cousin. And today, I mean, we still have videos and things like that, but our true north has always been how we can leverage technology to scale up what we think world-class education could look like. Our mission statement as a nonprofit is to provide free world-class education for anyone, anywhere, and these are ideas like personalization or mastery learning. People should be able to practice and get feedback as much as possible. So over the years, most of our resources have actually been on our software platform, which is free, available for everyone, and philanthropically supported. Today, just to give you a snapshot, we’re in 50-plus languages. I think the latest number is 180-something million registered users. I think the big push that’s very different from where we were 15 years ago is that we do a lot of formal partnerships with school districts. We started off as a very direct-to-consumer thing, which we still do, but we realized if we really want to move the dial for real students everywhere, we have to work with their school districts. We have all these efficacy studies showing how it can improve their outcomes, but it really needs to be in a classroom setting for most students for it to work. This is wild because I chose the exact opposite. I was like, “Okay, we make YouTube videos for people that help them learn, and if people want to use them, they can use them. But man, do I not want to get involved with the process of selling something to a school board or dealing with administrators.” How did you start to take that on? I mean, trust me as a person who’s dabbled in it a little bit, it’s a different business than building a tech product or making a YouTube video to actually get in there and interface with the bureaucracy a little bit. That is, I mean, I agree with you. I think that Complexly would be a more impactful organization if we did that, but we have chosen the path of personal joy to not have to do it. Yes. And I had some similar feelings, and I had other thoughtful people give me even more of those feelings. It’s like, “Are you sure you want to do this?” But the reality is, and this is what I found very promising about our structure, that our philanthropists who were donating to Khan Academy, and our board, really were nudging us in this direction. They were saying, “Look,” and the first nudge was, “Okay, you’re popular. A lot of people are using you, but how do we know it’s working?” And so we started running all these efficacy studies. Well, look, it works, and it works more than the videos. The videos are part of it, but when students practice at their own pace, and in the studies, we see if they’re even able to put in 18 hours over a whole year, which is not that much over a year, these kids are accelerating by 30 to 50 percent. There have been 50 or more studies like that. But then our board and the philanthropists said, “Okay, well, that’s nice. It works and you have a way to get to scale, and you have a lot of teachers already using you, but how do you make sure all students are able to use you?” That’s when the answer became school districts. So, we went to school districts and they said, “Well, to use systemically, you’ve got to give us support, training, integration with our rostering systems, and district-level dashboards. You have to meet all the accessibility guidelines.” That’s when we said, “Okay, if we’re serious about moving the dial at a state, national, or global level, we have to build this ground game.” So we’ve been doing it for about seven years. But, in my mind, it’s gone better than I expected. Really? I mean, it’s gone very well, but I guess that’s true of me as well. It does feel that there was a gap waiting to be filled, a little bit. What kind of organization do you think you are? You’re not really a content company. Do you think you’re a tech company? Are you an education tech company? How do you think of Khan Academy? So, the way I’ve always aspired to think of Khan Academy — I’ve always daydreamed this way, and hopefully I’m convincing other people — is to view us as a global learning institution. The same part of your brain that might think of an Oxford or a Harvard, it’s like, “Wow, those are storied institutions.” But then part of your brain says, “Yeah, but they don’t really scale. Their research scales, but their education side doesn’t really scale.” If Khan Academy could say, “Wow, it’s like that.” Hopefully, in a hundred years, people will say, “Yes, this is one of the major institutions of our world, but it scales. It’s high quality, it’s very affordable, arguably free or very close to free.” So that’s what I’ve always aspired to be. I think in real implementation, 15 years ago, we were known most for our content. Maybe we are still most known for our content, and then we actually have a pretty large software engineering team to build everything around the content, the practice, the data dashboards. Now, as AI might be able to create content in the not-too-far-off future, I think we are turning… well, we’ve always been, but I would say there’s even more weight being pulled onto how we create systems that can help raise the ceiling inside of a school, but also raise the floor outside of a school. Those systems can be software systems, AI systems, but there could also be credentials, ways to connect students with each other. We have a sister nonprofit called Schoolhouse, where there’s peer-to-peer support. So how can we build these systems so we can do high-scale, high-quality education? So I do want to talk about Schoolhouse, and I want to talk about AI. I still want to talk a little bit about the moment when you were starting to do this thing. Was there a thought in your head of, “This could be a company, or it could be a nonprofit, this could be for-profit or nonprofit”? Oh, yeah. Did you make that choice? Yeah, it was a pretty explicit choice. In the early days, I was living out here in Silicon Valley. There are some VCs who took notice back in 2007 or 2008. The first conversation was fun. They said, “Hey, I’ll write a $100,000 check right now. You quit your job. We’ll start Khan Academy as a for-profit.” But then meeting two was always like, “Okay, let’s talk about monetization, and maybe we’ll do some freemium content, or we’ll do test prep. We’ll charge for that.” And there’s nothing wrong with that. I don’t want to get too all high and mighty for anyone who does have a model like that. But I just thought about how much psychological reward I was getting from people all over the world. You get this, too, like, “Hey, thank you. That really inspired me. That really helped me.” And yeah, I did have a bit of a grandiose dream that maybe this could be — I always cite the Foundation series by Isaac Asimov — that maybe I could be something like a Hari Seldon creating the new Foundation that will keep us from entering a dark age, or maybe it’ll make today look like a dark age because everyone’s going to get educated. So I didn’t want to give up on that dream. That very chill, not at all grandiose dream of being Hari Seldon. [Laughs] Not megalomaniacal at all. Very grounded. But, you know, why not? From my point of view, the good reasons why people will often say go for-profit are access to capital, and maybe access to people because you can pay them with equity, or maybe you could, at least in theory, pay them a little bit more, even cash-wise. There’s sometimes a stereotype that has some truth in it that nonprofits may not quite be as nimble or as vast. I had never started a nonprofit before, but I always had a little bit of a chip on my shoulder, and still do. Like hey, I think we can get the best of both worlds. We can be a nonprofit, we can have this mission. We can try to build this institution, but we can also attract the best talent. Yes, we can’t give them stock because there is no stock in Khan Academy. No one owns Khan Academy; I don’t. But we can pay them well, and you can run an organization like this as nimbly as any organization anywhere, but because we’re a nonprofit, we have some advantages. We have 50 languages. We didn’t pay to translate many of those languages. People came out of the woodwork. People, especially in this time of AI, have a bias toward a little bit more trust. Some of the best people in the world don’t want to become billionaires. They’re happy if they can have an upper-middle-class lifestyle and get to work on a fun mission with other cool people. Something that matters. So we’ve been able to attract some really amazing people. Do you ever have a moment of regret that you didn’t do a for-profit and that you get to be Hari Seldon, but you don’t get to have a billion dollars? On this journey, I’ve had to raise a lot of money, where our budget now is pushing- [Laughs] Yeah, that’s because you never get any of it. Well, I’ve raised and I’ve met many billionaires, wonderful billionaires. Khan Academy would not exist without these billionaires’ significant donations. And there’s been a couple of moments where, I won’t name names, but a couple of these folks who you and I grew up reading about, and we probably fantasized in our middle-class houses or me in my lower middle-class apartment, saying, “Oh, imagine if I had that much money, I would do this and this and this.” And they’ve told me, “Hey, Sal, I envy you.” And I’m like, “You don’t really envy me. I’ve seen how you travel. You should see how... I just got upgraded to economy premium.“ But they were very genuine. I’ve heard this from at least three or four folks now. They’re like, “You’ve really found your passion and it’s really making a huge difference. I feel lucky just to be part of this journey.” That counts for a lot. And God bless them and what they do, and God bless them for donating to Khan Academy and making it possible. But I honestly wouldn’t trade places at this point. Now, could I figure out a way to spend a few more million dollars? Yeah, probably. Do you think that it could have increased your impact if you were pushing for market share, pushing for all of that stuff with an investor breathing down your neck kind of incentive? To some degree, the experiment has been run. There were several organizations that were, let’s call it, the same vintage as Khan Academy, plus or minus a few years. I would say most notably the MOOCs, the Massively Open Online Courses. I know many of the people who started them; they’re very good people. They weren’t in it for the money. They were in it for the mission, but they were convinced by people. I was trying to convince them to go nonprofit, but “No, you’re not going to get the capital. You need that for profit.” If you look at the MOOCs now, the ones that exist, they have fallen, and I don’t want to denigrate them. I think they’re still doing things of impact, but they were about “let’s democratize college education, let’s do something that’s world-changing.” Now, the ones that still exist have become “let’s do some certification post-graduate that might help some people transition into a data science type of thing,” and it’s just a smaller vision. I mean, I’m sure they’re still helping those people, but they haven’t democratized higher education, I think, the way that they could. So when I look at that, I think, “No, I think it’s priceless to still have the dream alive.” The experiment has been run, and Khan Academy is very big, hitting a lot of eyeballs, and helping a lot of people. But there’s this sort of trope that a nonprofit is going to be less nimble. It’s going to be harder to run. There’s going to be more stakeholders, more consensus building. How do you organize yourself? How is Khan Academy organized? Do you keep that as a priority, and how is that built into the structure of it? Yeah, I mean, look. When I was in business school, I had a classmate, and it was really funny because we had this case discussion, and he was talking about how narcissistic it is for people to name organizations after themselves. And his dad, I won’t name names, had named an organization after himself, a very large financial company. Everyone was laughing about it, but I remember that I was like, “Yeah, it’s super narcissistic to do that.” But then, when Khan Academy started to grow, it wasn’t called Khan Academy, and I said, “Well, maybe this could be something, but I wanted to make it a nonprofit.” I said, “Well, how do I make sure that, with my insecure side, how do I make sure that it’s hard to fire me?” They have to have a good cause, and how do I make sure that there are no shares here? I’m not like Mark Zuckerberg, who controls the voting interests of Meta. I have a vision, or I think I do. And well, my name on the door and becoming a certain quasi-mascot of the organization have helped. One, keep it focused on the vision. And frankly- And having a source of the vision, where if there’s disagreement, people know where to go to get that disagreement settled. Exactly. That’s probably not always the funnest job to have, but it’s a thing that exists. Yeah, and any executive manager or CEO has to have some of that in their job. But especially when an organization needs to make big pivots. I know we might talk about AI, and Khan Academy’s been doing a lot of pivoting there; it would’ve been very hard for someone who didn’t start the organization whose name’s not on the door to be able to make that type of a pivot. The lucky thing is, we have a great board. Many of them are major philanthropists, but some of them are people who just know a lot about education, and they push me in all the right ways. They’re always pushing, “Are we reaching the kids who need impact?” The trade-offs between investment and overstretching our budget, we think about that very seriously all the time. But they also see that not just me, but a lot of the people that we’ve brought into the organization are, I’d like to believe, pretty disciplined managers, engineers, designers, project managers, and content creators. A lot of our funders are people who are successful tech entrepreneurs. I say, pound for pound, put our large team of 350 people against any 350-person team in Silicon Valley or anywhere else, and I think you’re going to find as good or better talent and nimbler systems. And they see that. So that’s what keeps us focused that way. So 350 people. How is that organized from the board on down? There’s the board. I’m on it, but obviously, I report to the board too, and then I have more than your average direct reports. I have 14 direct reports. Man, you work hard. Well, this is the thing I learned, though. If you have the right people in the seats, the people who report to me, they’re fairly senior in their careers. So I don’t have to spend a lot of time with angst, or the “what am I going to do with my life?” type of conversation. Kristen DiCerbo is our Chief Academic Officer. She also runs the product management, design, and content teams. We have a great CTO. Our engineering team is the largest. We have what we call the external relations team, which is our philanthropy, but also all of this work. We’re partnering with school districts, and that’s a revenue source for us too, an earned revenue source. We have Khan Academy Kids, we have an internal council, we have our CFO, and all the internal functions. Then we have Schoolhouse.world, which I’m nominally the CEO of. I’m more of an executive chairman there. But the COO there nominally reports to me. So that’s kind of how we’re generally organized. But there’s really good leadership across the board here. So I get to jump around and say, “Hey, have you thought about this?” Every now and then, because people have their heads down and they’re in the forest, I can kind of surface and say, “Hey, but look,” or they’re in the trees and I’m like, “Look at the forest here. We’ve got to pivot a little bit harder.” So 14 direct reports, and that’s 14 different departments, kind of, and then Schoolhouse is a different thing? Yeah, I threw Schoolhouse in. That’s all just my universe of people that I’m talking to on a regular basis. But yeah, it’s different departments plus finance, legal, a lot of the internal stuff that you have to do in running an organization. So let’s talk a little bit about Schoolhouse. I think that people might not know that this is a part of Khan Academy, or that this is a thing that you do. I find it extremely impressive and so cool and so smart, and you’re the only one who could have pulled it off. Tell me about Schoolhouse. I’m happy to hear that. It’s funny, there’s been a little bit of controversy about Schoolhouse last week. I can tell you about that. Oh, wow. Which I think is a good controversy. It’s going to be published like five weeks from now, so it’ll be six weeks old by then. Okay. In the recent past, there’s been controversy. We should talk about that. But the idea was, I’ve always thought, “Hey, it’s great if people can learn from Khan Academy, but what about learning from each other?” And the best implementations in classrooms have always been some kids using Khan Academy, but the teacher pairing kids up, and they’re also learning from each other. When the pandemic hit, I said, “Well, now is the moment to try to do something like this at scale.” So we ran a little pilot and we said, “What if we created a website where young people” — actually, people who are learning, they don’t have to be young — “could say what they need help with, and then we could find other volunteers who can validate that they know the material?” So we needed a vetting process, but then they would tutor these people for free just out of the goodness of their hearts.” Zoom donated a bunch of licenses, and we tried it out. It’s a very utopian idea, but it worked at the scale of a few hundred people. So then we got some philanthropic funding, and we set up Schoolhouse as a separate nonprofit from Khan Academy. Honestly, the only reason to do that is just to keep the two focused on what they each needed to do. There was a little bit of fear of liability with people having real conversations on Zoom, and what might happen to Khan Academy’s liability. So we kept them separate. The name was somewhat inspired. I wrote a book back in 2011, The One World Schoolhouse, so we called it Schoolhouse.world, although we might change it in some ways in the next couple of years to bring it closer to Khan Academy. But we started doing it, and one of the immediate things we noticed was how to certify someone. How do we know that they know calculus, for example?” So we created a mechanism. Khan Academy already has assessments that are different every time, but we don’t prove that it’s you who did it. We said, “What if you take the Khan Academy assessment while it records your face, records the screen, you explain your reasoning out loud, and then Khan Academy will say whether you hit 90 percent?” You’re following a protocol, and you can’t be looking around and doing shady things. Then that video gets peer reviewed by people you don’t know to just make sure you’re not doing shady things. And then if you’ve got 90 percent on that unit, we say, “You know unit one of calculus, you can now begin your tutoring journey. And there’s some training for you there.” When we were doing that, the University of Chicago reached out and said, “Hey, everything’s up in the air with the pandemic. Could we use your certification for college admissions?” We said, “Yes, you can.” And then it was MIT, Columbia, Caltech, and now there’s a list on the website of 40-plus universities, including Yale and Brown. I mean, you can name them. They all said, “Hey, you could use Schoolhouse certifications as a way to prove your mastery.” People are like, “How are we going to do assessments in the future with AI and everything?” And Sal Khan’s already figured it out. Well, AI is going to add a whole other layer. I mean, I think we’re going to be able to do some nice simulation-based assessments and things like that. A fun example: I just met this young woman from Afghanistan three weeks ago, and obviously, she could not go to school growing up in Afghanistan. Khan Academy was her school. The Taliban take over Kabul, and her family becomes refugees in Pakistan. She’s still not in school, but she’s been learning all this time. She wants to go to MIT. She applies, and MIT is really impressed with her application, but she has no diploma, no formal transcript, no SAT scores, no AP scores, nothing. MIT asks, “Can you go on Schoolhouse and validate yourself?” And she did. She got in based on that, and I just met her. She was at this Y Combinator event for the top AI engineers in the country, out here in Silicon Valley. My wife and I took her out to a little meal, and I was like, “Yes, this is the certification vision of Schoolhouse.” It’s pretty powerful to hear stories like that. This is where some of the controversy started coming from, but I think it’s a good controversy. The same college admissions folks, they weren’t just interested in the certification, they were also interested in the kids who tutored, because you could imagine if you certified yourself in calculus and then tutored calculus and had a high rating- Who incentivizes people to actually do the tutoring, because it isn’t paid? But I can say, as a person who tutored kids when I was a kid, that’s so powerful in just knowing the material better, if not also putting that on your little kiddo resume, or whatever they make them do these days. [Laughs] Exactly. Well, if you’re one of these colleges and there’s a lot of cynicism around people gaming and et cetera, but if young Hank ran 100 sessions tutoring chemistry, and he has a 4.8 out of 5.0 rating, and there are these quotes saying, “I learned more from Hank than I did at my school,” or, “He makes me excited about chemistry.” If I were one of these universities, I’d be like, “I want Hank on my campus. That’s the kid I want.” But we went to these universities, said, “What else could be of value?” They said, “Well, this is probably too hard, but what if we could give students practice in having dialogue about hard subjects? We don’t want them to water down their passions, but we want them to be able to have constructive dialogue because everyone’s in their bubble now.” Geographic bubbles, socioeconomic bubbles, social media bubbles. So, that’s when we launched this Dialogues platform, only a couple of months ago. Most people would get anxiety, just as I’m about to describe this, but we make young people fill out surveys on tough issues — immigration, Israel-Palestine, affirmative action, and gun control. I could go down the list, the stuff that we are afraid to talk about at dinner parties, and kids fill out a survey. We pair them with kids with the opposite viewpoint, and they have a conversation. After that conversation, which usually goes about 50 minutes, they fill out a questionnaire: You don’t have to have convinced them, but we ask, “Do you feel like the other person heard you?” And “do you think they can represent your point of view? Can you represent their point of view? Is there any feedback you want to give to that other person?” It could be constructive or it can be positive feedback. You do as many sessions as you want. And then on your transcript, it just says how many sessions you participated in. It doesn’t say what your point of view was. You can highlight anything. You choose to highlight. “Hey, Hank said that Sal really opened your mind to a different viewpoint that you never took seriously before.” Whatever it is. I could put that on my transcript. And if I want to, I could share that with many of these same universities. The controversy is that there’s a New York Times op-ed that argued, “Oh, this is just another thing to fake in college admissions.” And Nate Silver tweeted, “This is affirmative action for boring kids.” All of a sudden I’m like, “No. How can it be boring? How can it be?” It’s not boring to be willing to have an open conversation about these topics, and be able to do it constructively.” And with what we’ve seen, these kids aren’t holding back. They’re holding their positions, but they’re doing it in respectful ways. All I have to say about “this is affirmative action for boring kids” is that’s a tweet. That’s such a tweet. And I’m so tired of tweets. I mean, I’m a tweeter, don’t get me wrong. I post. [Laughs] I’m not. Right here. I post. I post way too much. I want to live that Sal Khan life, get off of these postings, but I post. But man, is that a post. That’s pure “I’m going to have the take that everybody’s going to feel good about, even if it makes no goddamn sense.” Oh, man. Well, that’s fascinating and does make me sweaty. That’s not a choice that I think I would’ve made. That seems bold. It’s brave. Yeah… did you just call yourself brave? No, I’m saying the students are brave. I actually haven’t... I’ve observed some of these... Well, it’s a little brave of me, on my part too. [Laughs] It is. I think it’s brave on your part. I’m saying it’s the opposite of boring. These students are brave enough to do that. I haven’t been able to do it because if I were to get on Dialogues, the kids will say, “Oh, that’s Sal Khan, and did you know that Sal Khan has this view about the Middle East? Did you know that Sal Khan has this view on immigration?” I’m like, “Nope, not going to be constructive.” So I can’t participate in it. That’s probably ideal. I probably won’t be on there either, but I think it’s a brave decision for an organization to make. Which leads me to a Decoder question. How do you make decisions? How does Sal Khan make decisions? How does Sal Khan make decisions? I always try to remind myself and the other people who are involved in decisions what we’re trying to solve for to begin with. There’s a friend who’s a very successful venture capitalist, and I’ve made fun of him in the past, but he has this hack that always makes him look brilliant. You go about half an hour into any meeting, and he will say, “Hey, hold on a second, everyone. Let’s take a step back. What are we solving for?” And he looks brilliant, and everyone’s like, “Oh, yeah, we lost track of what we’re solving for.” It’s true. Thirty minutes into a meeting, most people are getting into the weeds, and you’re like, “No, what we’re solving for is how we get more students to participate in society.” So it’s important to have those true norths. And then yes, I try to listen to people who ideally are as close to the ground as possible on what they have to say, and then try to make a call that both makes sense intellectually and hopefully takes the path of less cynicism and more like “maybe the world can work this way; why don’t we just try?” People say gut sometimes it’s like that’s a lazy way of doing things, but your gut has a ton of neurons. Our gut, as you know, would be a fairly smart animal if it were just on its own. I don’t think the gut’s actually making the call here, but yeah. But the gut is really… We have these neural nets. We have these hundreds of billions of neurons that are doing things subconsciously all the time. Yeah, there’s a lot that we don’t know about. At some point, they’re just giving your conscious mind a memo saying, “Yeah, I think this is the right decision.” So, at the end of the day, if you have a good true north, if your gut has paid off in the past, and you’re surrounded by people who are advising you for the right reasons, and you make a call…It is also important to have folks around you who are willing to disagree and commit. Like, “Hey, once we make this call, let’s try our best. Let’s not grumble and be passive-aggressive after that.” So maybe you can walk me through that process a little bit, because I want to talk about Khan Academy starting to use AI tools. I got a call from you a couple of years ago, and I remember pacing around my basement during this phone call when you told me about you going into AI tools. I thought, “Man, I would not make this call right now. I do not know where this is going. I don’t know what the public opinion around it is going to be. I know it’s going to be complicated and weird. I don’t know where it’s headed. I don’t know how powerful these tools are.” But I also thought, “More power to him for heading straight headfirst into it.” This was very early. This was right after GPT-4 launched. How did you make that particular decision? Yeah, and I mean, I remember our call. I don’t remember exactly when that call was in the whole timeline, but we actually made the decision well before GPT-4 launched. Yeah, but this was after. Yeah, so I probably couldn’t talk to you about it then. It was probably as soon as we were out of our NDA with OpenAI, and I probably called you and said, “Hey, maybe there’s something here.” But it was the summer of 2022, so roughly three years ago. I received a call from Sam Altman and [OpenAI cofounder] Greg Brockman. They said they’re working on their new model. They wanted to show it to us because they thought maybe there’s something interesting about Khan Academy, showing positive social uses of it. I was curious. What they showed us was GPT-4, and this was well before ChatGPT. This was about six months before ChatGPT existed. ChatGPT, for those who remember, wasn’t even launched on GPT-4. It launched on GPT-3.5. It blew my mind, and honestly, even when they showed it to us, they hadn’t fully appreciated it. They ran a demo with us, and they said, “Do you want access?” I remember it was a Friday, and they gave us access, me and our chief academic officer. I was messaging with Greg, and I said, “Hey, does this work in other languages?” And he wrote, “I don’t think so.” I barely speak Bengali. I tried to speak, and it spoke back, not only to speak back to me in Bengali, but it also wrote in Bengali, which I can’t read. But then I said, “Can you translate that into English text?” I was like, “Wow, it could speak Bengali.” I took a screenshot and I sent it to the OpenAI folks, and they’re like, “Yeah, after you asked, we checked. It looks like it can speak every language.” See, this is why I would’ve said no. I’d be like, “You guys don’t even know what languages it speaks?!” There are all these emergent qualities. It’s fascinating. I have this other friend who’s a professor at Stanford, and he’s trying to understand how, apparently, you could go from one language to another as a rigid transformation in a multidimensional space. That’s why it can talk about things in Bengali that it never got trained on in Bengali. Anyway, there’s some deep linear algebra there. But this was very early on, and this was another thing I didn’t realize that we were doing. A year after that, after we had launched… well, immediately, like, that weekend, I couldn’t tell anyone about it. We had signed an NDA, but I was up all night. I asked [GPT-4], “Write the Declaration of Independence in the tone of Donald Trump, and it did it: “George III’s a loser, we’re going to have the best country.” It was actually a pretty good one.. But then, when I said, “Hey, you are an empathetic tutor” — I think the exact prompt I said is, “You are Robin Williams in Dead Poets Society, and you are going to Socratically tutor me,” because I knew this could be used for cheating, too. But it was able to do that. Even though it made errors and experienced hallucinations, there’s something real here. So, we started. Well, we got most of the Khan Academy team under the non-disclosure agreement so that we could see it. About a month into that — and this is still months before the rest of the world even knew that this was coming — our team was surfacing all the fears: the cheating, the safety, the privacy, the hallucinations, the math errors. But I was pretty convinced. I said, “Look, yes, those are real fears and errors. We need to turn those into features, but this is going to be so powerful, and it’s getting better so quickly that if we don’t really lean into this with our mindset and our mission-focused mindset, other people are going to use it. And those people are not going to care about whether it’s cheating. They’re not going to care about whether it’s good for kids. And honestly, this stuff’s going to be so scary for people that they’re going to need, hopefully, someone that they can trust. Someone who’s built up trust over time.” So that’s when we leaned in. It’s interesting that ChatGPT has emergent phenomena. The day that product came out, on November 30th, 2022, I messaged the OpenAI team. I said, “Wait, we’re under a non-disclosure agreement. We’re not supposed to announce any of this stuff until March. What did you guys just release?” And they said, “We didn’t release anything. We just put a chat interface on an old model that’s been out for like seven months, and the world exploded.” It’s interesting. Some kids in a garage or at a college campus could have done that and now have ChatGPT. No one did it. But even that was a shock, that you just put a chat interface on these existing models and it makes people think about it differently. Wild. How do you actually turn ChatGPT or any LLM or GPT model into something that is useful for a student? So you have to work in math that works. You have to try to obviate hallucinations. You have to put walls around it so that it doesn’t do things you don’t want to do or be too sort of friendly with students in whatever way. It has to look and feel like a tool. How do you actually functionally do that? What do you do to it? It’s evolving, because the models are changing, and just people’s expectations around these things are changing so fast. But some of our original thoughts, I think, are still true. There are some just basic prompts you can do. I say basic prompts, but when you want something that can robustly work for millions of folks, the prompt has to be quite careful, and you have to create all sorts of systems to test. But to say, “Look, this is going to be Socratic, use these techniques. You’re not going to give answers, but nudge students forward.” There’s a lot more to it. But that’s the gist of it. You’re going to want to have some level of transparency and oversight, especially if you’re talking about under-18 students. And transparency buys you a lot. It doesn’t just buy you safety. If a student is saying, “Hey, I want to learn how to build a bomb,” or “I want to hurt myself,” it’s important to not only have teachers or parents be able to see it, but also to have another agent that can see that and actively flag it. We avoid hallucinations from the get-go. The models themselves have gotten a lot better over the last two to three years, but you anchor it on content that you know is good, and we have a lot of content. Do you have a big context window that has a bunch of true facts in it or something? No. There are two ways that you can do it. One is, if you’re using our tutor — which we’re calling Khanmigo in Khan Academy content, like on a Khan Academy article or Khan Academy video — it has that content context built into it. Or if you’re getting help on an exercise, it knows we’ve passed it the solution to the exercise, so it won’t hallucinate, make math errors, or very infrequently make math errors in those cases. So that’s one: anchor it on things it knows. There are also these things called vector databases. You could put a lot of content in these databases, where, based on the conversation that you’re having with the AI, it can find which pieces of content are closest to this conversation and then throw those things into the context window. You don’t have to throw everything in the context window yourself. Google now has some models where you can have like a million tokens, or roughly a million characters, in the context window. So you can actually throw a lot in there. Your average book is like 40,000 words. Well, actually, it’s not even characters. A token is like two-thirds of a word. So, you can now throw a lot into context windows too, but that also has cost issues. The other guardrails, like transparency, there’s a lot that we do around the math to just double and triple check the math — just hacks like the AI calls another AI to go through reasoning that can come up with every possible way that the student might approach it, then compare the student’s answer to that. So there are things that we’ve been playing with that have improved the AI’s ability to do that. So if I’m using Khanmigo and I get the question wrong, it might’ve already predicted that wrong answer and know that I got it wrong, like how I got it wrong. And what you’ve touched on is actually one of the hardest things. The AI models are actually pretty good at math now. They weren’t good at math two and a half years ago. Right. But now they call math out? That’s actually what they do. They’re making Python calls behind the scenes to actually do the computation. The places where you still see the most errors are when the AI is evaluating the student. So let’s say the correct answer is one-third, and the student has put in 0.33, that’s not quite right. It should be 0.3 repeating. A good tutor would say, “Okay, are you sure that’s the full answer? Are you sure it’s just 0.33? How would you express that as a fraction?” So yes, now what we have going on behind the scenes is the AI, even before looking at the student’s response, is saying, “What are reasonable responses here?” and then it compares. Or when the student gets an answer, it asks, “What are ways that the student might’ve gotten that answer?” and then it compares it. This is constantly evolving as the models get better, and we’re having to do less of that. But there is still a lot of that going on behind the scenes. One thing I’ve noticed is that people always say, “Hey, you’ve helped me. You’re the reason I got a 5 on my AP exam or whatever.” And, “I think probably you had something to do with it,” is what I say to them. But what I have learned is that there’s a big component of teaching that isn’t teaching. It’s motivational. It’s like the coach in the room who you have an obligation to and who will be disappointed if you don’t do the thing that they asked you to do. It’s a human being, and it’s easier to feel an obligation to human beings. Do you think that LLMs can play a motivational role? Is that part of this? Yeah, I think over time. And they’re already playing [that role] a little. We’ve even implemented a little bit of that, just even for myself, I had to give some commencement addresses. I took my first draft, and I got an AI to give me feedback, and it made me feel good. I’m more confident now. It sure does do that. If nothing else, it will make you feel good about your writing. I could have written a pretty trashy, bad speech, and it still probably would’ve given me positive feedback. But what you’re touching on is absolutely right. A lot of folks know Khan Academy started with me tutoring family members 20 years ago. When I really think about it, yes, I was explaining certain concepts to them, but a lot of what I was doing is exactly what you’re describing. I’d be like, “Hey, where are you? How come you didn’t do the thing I told you? Hey, look, you got to be a little more confident with how you answer these questions. Let’s lean into the problem,” or whatever it might be. So, our realization with the first version of Khanmigo was that it just kind of sat there and waited to be asked, and then it would help you as you needed it. I say that’s analogous to a tutor walking into a classroom and saying, “Hey, kids, I’m here in the back. If you need me, come get me. But I’m here. I’m going to be reading a novel.” That’s not good enough. You want something that holds them accountable. This new version we’re launching, which we’re piloting in the fall, is with the AI front and center. You come, and it’ll say, “Welcome back, Hank. It’s been a couple of days. We’re falling a little bit behind our goals. Are you ready to get started on this next task? This is what your teacher wants you to work on, and once you’re done with that, I have some ideas for you to work on.” Once you go into it, it’s constantly like, “Hey, look, you got that wrong. Not a big deal. I think it’s a good idea for a review of why you got that wrong.” There might be some game mechanics. “Hey, if you review it, I’ll give you some points,” and things like that. But that’s the future. And yes, if we fast-forward a few years, we can all imagine having AI. A real human being is always going to be better, but a real human being is not always available. That can be an even better accountability holder for us. You use a phrase for this: “It’s not the best tool, but the best available tool,” or something like that, right? Yeah. Did I make up that you coined that phrase? That’s not a me phrase. Oh, okay. I mean, I’ve said stuff like if I had to pick between an amazing teacher and amazing technology, I’d pick an amazing teacher. But hopefully you can have both. I’ve also heard you say that replacing humans in education would be a disaster. I think that a lot of teachers, parents, and students would agree with you. How do you think we can avoid that fate? People are complicated and they’re expensive, and so it feels like lots of people would like for all the work to be done by things that are simpler and don’t complain. How do we avoid this fate? Well, one, I always like to point out to people the economics of our education system. Most people would argue that the teacher role has the most direct impact on the student of everything that you’re spending money on. But if you look at a lot of places, like California, which spends around $25,000 per student per year, it has about 25 to 30 students per teacher. A lot of East Coast school districts spend $30,000 to $35,000 per student. So, depending on how you account for it, there’s as much as $800,000 per class of 25 to 30 kids. They’re not paying the teacher that. The fully loaded cost of a teacher with benefits, even a senior teacher with a pension and everything, may be $200,000, if you put all of that in there. It’s usually a lot less. So a lot of the costs of education are going into layers of other stuff. Some of that stuff is needed- Yeah, buildings are part of it. But if you had to cut costs, that’s not the way to do it. Now, I think AI might help in some of the back-office stuff. You might be able to automate the registrar’s office. You might be able to automate other functions in the office of a school, and that might save money. But I think we’ve always said, even before AI, that our goal is to raise the ceiling. Hopefully, you already have access to a reasonably good classroom and a great teacher, but it’s still hard for that teacher to personalize for 25, 30, or 35 kids. We’ll give them the tools to do it, but we also want to raise the floor. I’ve talked about this young woman in Afghanistan. She did not have access to a teacher, but she was unusually motivated, and it would’ve been sad if she didn’t have something. So, that’s where we raised the floor for her. We hear other stories. There’s a young girl in a Mongolian orphanage who used Khan Academy. There are kids in rural America who don’t have a physics or a calculus class within hundreds of miles of where they live. Kids in inner cities who don’t have calculus, physics, or chemistry classes at their school or an advanced one. That’s where we raised the floor. But the ideal… I always point out that about 150, 200 years ago, when textbooks started to become a thing, a lot of teachers were afraid that those were going to replace them because teachers thought that, “Oh, I’m the source of the knowledge. Why would anyone come to me when they can read the whole textbook?” Right. There’s a book here. There’s a whole book here with exercises and everything. But now teachers can’t imagine teaching without a textbook because they’re like, “Of course they need me still, but they need their practice. It’s good to have another resource.” I think that’s going to be the same thing with AI. A teacher who refused to use a textbook 50 years ago is going to have trouble. But a teacher 50 years from now who’s refusing to use AI might not be able to be all they can be. If they’re using the technology thoughtfully — and the AI is helping them form better human connections with the students, make more engaging interactive lessons, more personalized lessons, support certain students, while they can support other students, and they can tag team, and it can act as their teaching assistant — it’s going to become invaluable for them. I’ve heard you say that the incentives of this sector, the education sector, which is the second-biggest industry in America, are strange. I think you used the word “strange,” and then the person you were talking to did not ask a deeper, further question. I was like, “Oh my God. Do I want to know all the ways that Sal Khan thinks the incentive structure of education is strange?” So, tell me, what are some of the strange incentives? Yeah, and look, this goes back to the nonprofit question. I don’t think you need to make everything in the world a nonprofit. There are certain things that the private sector does very well. I am, at my heart, a capitalist. I believe the capitalist system generally works. There are certain things you want the government to do that the private sector wouldn’t naturally do or wouldn’t have the ability to coordinate. But there are certain areas where markets aren’t working; they aren’t doing what’s efficient or what’s aligned with our values. And maybe, for political reasons, or just because the government is too slow or bureaucratic to take advantage, that’s where the nonprofit sector matters. Education — and I would say another major sector here is healthcare — is one of the areas where the beneficiary is the student, but the decision maker is typically the school district, and sometimes it’s the teacher. And the payer, in education, is the taxpayer. They’re three different entities, three different groups. You actually see a very similar thing in healthcare. Even if it was rational, we still also have this value that just as someone is bleeding and they’re dropped at the emergency room, you don’t want to say, “Hey, let’s see your insurance.” You want to treat them. Unfortunately, I think that might’ve happened sometimes, but you want to treat them first. Similarly, if a young person in our society wants to learn, I think most of us feel that it should not be based on how much money their family has when deciding whether they should get a high-quality education. So, that value system, and the fact that these three agents all have different incentives, I think, is what has led to education not always having the most engaging, the most effective outcomes. A lot of the people in the district office are very well-meaning people, but there are certain guidelines and regulations they have to follow. A salesperson from a big publisher comes and tells them a good story. They adopt it. The kids hate it, but too bad. Before we built our district offering, I couldn’t tell you how many times we would talk to a chief academic officer or superintendent of a district, and I would say, “Hey, why don’t you all use Khan Academy? Look at our efficacy studies.” And they’re like, “Oh, we believe you. My daughter swears by Khan Academy; it got my nephew through calculus, and even got me through statistics in grad school. But we have this vendor we adopted last year, and they’re on some state list that you’re not on, and we already wrote them a $5 million check, so I think we have to use them.” I’m like, “This makes no sense. Would you use those with your own child? Would you use them yourself?” And they say, “No, not really.” Oh my god, yeah. That is what I’ve identified. My solution to that problem was simply to not engage with it. But what we have both done, I think, and to our credit, is that our first customer was the student, and that’s how we got into these places. We didn’t get in from the top down; we got in from the bottom up. I’ve seen several organizations, like for-profit companies, too, get into this business that way. That’s the way to do it if you actually want to help, whereas if your customer is the superintendent, then you’re going to have a pretty different set of structures. Can I ask you a question that is a real Hank Green question here? So once upon a time, I came across this large database of videos that had been scraped to use in a sort of database that anybody could have access to for training AI. And the only YouTube channel that I found that had more videos in it than the ones that my company, Complexly, had was Khan Academy, which indicated to me that they weren’t just grabbing every video. It was videos that had good information that we can trust, and that is more valuable. Does that make sense to you that that is why we got grabbed more than the average video? I think so, yeah. I don’t know about this database and all that. But even in the early stages of Khan Academy, even as a nonprofit, there’s definitely a bias, especially in Silicon Valley, to be more of a platform than to be focused on, say, the artisanal quality of whatever you’re trying to create. I would always point out that you don’t need a thousand explanations of L’Hopital’s rule. One good one might go a long way, but maybe four or five max [explanations], with different takes on it. It’s actually sometimes been hard for me to make people believe that because it goes against this idea that “Well, that won’t scale.” I said, “No, it kind of does scale because L’Hopital’s rule isn’t going away any,” it’s not like we’re creating some news site or something. It scales the same way a textbook does, except a lot easier because making another copy is basically instantaneous. Exactly. And then we have interchangeable parts. If L’Hopital’s rule does get updated, we can just update that part. Yeah, exactly. So given that, my question is: how do you feel as a creator of content about LLMs training on the stuff that we make? Do you think that they can learn the way anybody else can learn, or is there a difference in value offered by different content that should be compensated differently? If I thought Khan Academy could get a meaningful check for this, I would love to take it. But generally speaking, in a world where let’s say I don’t have a choice — I mean, maybe I do — but let’s say there’s a world where I would rather our content be used for training. If a model in two or three years can create a stock tile video and draw diagrams that are helping people, that’s going to be net good for society, I think. So yeah, I’m generally supportive of it on the intellectual debate around whether we just view an AI as just a super smart savant that sees a lot of material and now can paint in someone’s style or write in someone’s style, or create a video in someone’s style. Or should we view it as, “No, it’s IP theft.” The courts are going to decide this. But I’ve been kind of a little bit more on the savant side because, as we know, our videos aren’t directly encoded anywhere in these models. These models, literally, are learning to create and create associations of something similar. And there have been people who could paint in the style of Leonardo da Vinci or whatever. This is a similar thing, although on steroids, but that’s my current view. What about you? Are you like, “I don’t know”? I don’t know. Well, what I’m like is, “I’d like these things to make their way through the court so that we can all know, so that we can know what the legal situation is, regardless of how I feel,” because there is sort of a law intellectual property question here. So, I want to kind of finish with this. You talked earlier in our conversation about content creation for artificial intelligence. Of course, already these things are creating content — what an LLM exports is content in a way. Maybe it’s just for you, but it’s still stuff. But are you thinking that someday there will be an ability for me to say, “I need to know about Lagrangians,” and Sal Khan just sort of pops up and starts tutoring me on Lagrangians with a piece of content that had not previously existed, and you’re doing the thing, the digital chalkboard is happening, and I’m hearing your voice. Is that the future? Are you going to never die is kind of what I’m asking? That’s my “you know, we’ll see.” [Laughs] Mr. Seldon. I mean, when you hear someone like Demis Hassabis talk about how all diseases will be cured in the next 10 years, it gives those of us who are about to be 50 a little bit of hope. But yeah, look, I think it’s only prudent that I, other content creators, and Khan Academy prepare for that reality. As best as we can guess. It’s likely to happen. It’s probably going to happen, I would guess, with really robust quality in probably three to five years. My vanity likes to think it’ll take longer. It might. We’ll see. Or it might be next month. And then our content can still be in some vault somewhere, like this is what people did before AI took over, or this was the source. It’ll hopefully still have an almost primary document value to it. Like, “Oh, Hank was like a real person and he talked about having an upset stomach that day,” and AI won’t do that kind of thing. Who knows? I think we’re going to give the AI some pretty upset stomachs. They won’t have stomachs, but something inside of them will feel- It’ll talk about its stomach, it’ll talk about what it just ate, or what Hank just ate. [Laughs] Yeah. But yeah, it’s coming, and I think we have to prepare for it, and we are building tools like the new version of Khanmigo and the teacher tools. We are having it so teachers can co-create content with the AI, including practice and administer it through the Khan Academy platform, like Khan Academy content, and then get insights back. So we’re trying to prepare for that reality. This is wild. I don’t know. My take on what I make and the way I make it is that I think it would be hard. I also think that people connect to people. I think that if people don’t know that it’s not a real person, there will be a betrayal when they find out that it isn’t. If they do know that it’s not a real person, they won’t feel the same way about it. That’s not to say it couldn’t be a useful tool, of course, but one of the hardest things to make into an AI will be a pop idol. People want a real thing, not that that’s us, but there’s an element of it. [Laughs] We’re onto certain niche audiences. Every once in a while, somebody will be like, “I just really like a smart man who teaches me things.” All right, weird way to end it, Sal. I remain very impressed by you, and I’m glad that you got to spend a little time chatting with me today about all of the crazy ways that you are having a big impact on the world. So thank you so much for taking the time. Thanks for having me, Hank. Questions or comments about this episode? Hit us up at [email protected]. We really do read every email! Decoder with Nilay Patel A podcast from The Verge about big ideas and other problems. SUBSCRIBE NOW!