The following is a conversation between Sal Khan, the Founder of Khan Academy and Author of Brave New Words: How AI Will Revolutionize Education (and Why That’s a Good Thing), and Denver Frederick, the Host of The Business of Giving.

Denver: Sal Khan has been transforming education through Khan Academy, an online platform that brings top-tier instruction to students everywhere. Starting in 2008 as a way to tutor his cousin, it has evolved into a key educational resource covering subjects from math to humanities. It is dedicated to making personalized, mastery-based learning accessible to all by integrating technology into education. Its most recent evolution is Khanmigo, which aims to offer every student a personal AI tutor. And he is with us now. Welcome back to the Business of Giving, Sal.

Sal Khan, the Founder of Khan Academy and Author of Brave New Words: How AI Will Revolutionize Education (and Why That’s a Good Thing)

Sal: Thanks for having me.

“Even out of those thousand, when you’re getting 50 letters from folks that seem somewhat transformational, it felt like something was going on.”

Denver: You know, you’ve often shared the story of how tutoring your cousin, Nadia, sparked the idea for Khan Academy. Was there a moment, Sal, when you realized that your tutoring videos had the potential to change how people learn worldwide?

Sal: Yeah, there were a bunch of moments that I guess added up to a huge realization. Even before the videos existed, I always suspected that tutoring could help. I remember I was a volunteer tutor when I was in high school, and I saw with a lot of my peers, it made a huge difference. And then with Nadia, as I’ve said many times, she went from being a student in a remedial track to getting into an advanced track. 

Word spreads in my family, free tutoring is going on, and I’m tutoring 10, 15 cousins, family, friends, and I saw that pattern over and over and over again. 

And it was only about 2 years after I started tutoring my cousins that I started making the videos. And even before that I was making the exercises for them, the software. But on the videos, my cousins famously said they liked me better on YouTube than in person. I think what they were saying was they liked being able to pause, repeat, not feel embarrassed if they had to review something from a few years ago.

But that’s when… it was on YouTube, people just randomly found it, and I was getting comments on the videos, messaging on YouTube, emails from folks. A lot of it was just “thank yous” and “this really helped,” but some of them were more intense, “This helped me pass my Algebra class,” or “I thought I was dumb, now I realize that I love math,” or “This is keeping my kids engaged and they have learning disabilities,” or “I’m able to go back to college after leaving the military.” And for me, just making these simple videos and getting messages like that from strangers around the world, I used to wake up and jump out of bed just to see what people were sending me.

Denver: Yeah.

Sal: That was the first time I said, Wow, this isn’t just a– I would’ve been excited if I was just getting “thank yous” …“this was useful.” But at that time, there were probably about a thousand people who were using the videos a week or something like that; and even out of those thousand, when you’re getting 50 letters from folks that seem somewhat transformational, it felt like something was going on.

Denver: Yeah. Really inspiring feedback, and I think that’s a lesson that often you don’t need to go big at first. It’s the depth of the response that you get from people that is the real tell as to whether you’re onto something.

Well, Khan Academy has always been about empowering learners, and the most recent example, as we mentioned, is Khanmigo. Tell us about that, Sal, and how you got started down this AI road.

Sal: Yeah, and I actually have a book coming out, Brave New Words: How AI Will Revolutionize Education (and Why That’s a Good Thing).

Denver: When is it coming out?

Sal: It’s coming out mid-May, May 14th, it’s coming out. And I detail the journey and then where we’re going with AI and Khanmigo in Brave New Words. 

But if you start with Nadia and me tutoring cousins and writing software for them, and me having tools to keep track of them and making videos, and you fast forward all the way to say, 2 years ago, and I’ll tell you in a second why I’m fast forwarding to 2 years ago, everything Khan Academy has been doing, and it’s much more than me now… we’re over 300 people at Khan Academy. It’s been trying to somehow scale that one-on-one personalized experience that I was able to do with Nadia and then my other cousins, and we were able to approximate it using on-demand video, using software, using teacher tools. But we knew it wasn’t quite the same thing. Now, it has raised the floor for a lot of students who didn’t have access to anything before, and we also like to think it raises the ceilings in classrooms because now teachers can personalize more, et cetera, et cetera. 

But there is always a holy grail in education technology, and that’s: What if the software could really emulate what a great tutor would do, and maybe for a teacher, emulate what a great teaching assistant would do.

I have no delusions that this is in any way a substitute for having a human being in mentoring you and teaching you and guiding you, but what if it could do what a tutor could do? So, that has always been in the back of our minds. I always thought that this was really just going to be the fodder for science fiction for a while.

Denver: Yeah.

Sal: But about a year and a half ago, summer of 2022, I guess it’s almost 2 years ago now, Sam Altman and Greg Brockman from OpenAI reached out. They had their next model. This was months before even ChatGPT existed, and what they were about to show me was GPT-4. 

And for those who’ve been paying attention to this, ChatGPT, when it came out, wasn’t even based on GPT-4, it was based on GPT-3.5. 

And we had early access to it. Even though it had a lot of issues– math mistakes, hallucinations, et cetera, it also had some superpowers, where it was really able to, not just pretend to be a tutor, but in many cases, it really was reasoning through it with the student and really was acting as an empathetic, approachable tutor, almost identical to the way that I would’ve messaged with Nadia back in 2004. 

And so that’s when our organization, we said, “Hey, we have to double down on this.” And you can imagine all the debates inside of Khan Academy. How do we handle safety and privacy? How do we handle that this could be used for cheating? How do we handle oversight? How do we handle that it still makes up facts every now and then? It can still make math errors. And we’ve just decided to, “Well, let’s write all those things down. Let’s turn them into features so that we can mitigate those risks. And let’s see if we can start at least showing that this could create more value than not.”

Denver: Yeah, that’s a great way to move forward because sometimes issues like that, particularly in a nonprofit organization, can paralyze a nonprofit organization because you’re never going to be able to answer them until you get out there and learn and experiment and iterate and do things of that sort.

So, I’m an 8-year-old…I have Khanmigo… I’ve just come home from school: What do I do? How does it work? Walk me through it.

Sal: So, Khanmigo, we’re constantly working on it, so there’s a lot of changes happening. 

But essentially, well, there’s a few things that could happen. One is, let’s say, that 8-year-old has traditional Khan Academy goals. Let’s say, they’re working on a math class, and they’re spending 20 minutes a day going through their math— which is what actually I would recommend, 20 minutes a day for three or four or five days a week— and when they, let’s say, go to a math video and they still have some questions, they can click on Khanmigo, and Khanmigo will pop up and they can ask questions about the video.

So, it has context of what the student is doing. If they’re doing an exercise and they say, “Hey, Khanmigo, I need help here.” Khanmigo won’t give them the answer even if the student says, “Give me the answer.” Khanmigo will say, “Hey, I’m here to help you. I’m your tutor. What do you think the next step is?” Or, “How do you want to approach this question?”

And so it really does almost the exact same thing that an ethical tutor would do, what I was doing with Nadia back in the day. And it’s not just in math, it does this in science, it does this in humanities. 

But then there’s a whole bunch of things above and beyond that tutoring, that Socratic tutoring experience… students can use Khanmigo; it can act as a guidance counselor, as a bit of a life coach. 

You know, “I’m having trouble with procrastination,” or “test-taking stress,” and it does best practices in cognitive behavioral therapy to help you reason through that. 

We’ve even partnered with Angela Duckworth from the University of Pennsylvania. She has done a lot of research on mindset and grit, and we’ve done some of her interventions that have been shown to show students improvement in focus. So, we’ve implemented that so they could walk through that with the AI.

There are activities where they can debate the AI. They take one side of an issue, AI takes the other side of the issue. 

There is an activity where they can write an essay with the AI, and the AI gives them feedback on it, but it’s not going to do the essay for them. And it’s not just a chat interface, it feels much more like you’re working with a collaborator on a Google Doc. And as you write, it highlights parts of it and says, “Hey, how do you back this up?” Or, “This isn’t really a strong thesis; how about you move it in this direction?” et cetera, et cetera. 

There are activities where you can talk to simulations of historical or literary characters. 

So, there’s a lot there, and that’s just on the student side…you mentioned 8-year-old. If you want to talk about a 28-year-old teacher, there’s…

“What we’re excited about is there are definitely aspects of Khanmigo that are going to enhance the student experience and that whole workflow, but this is the first time that we have a solution for teachers that directly helps them save time or helps them with their work. So, Khanmigo also acts as a teaching assistant.”

Denver: We’ll get to that. 

But I’m just amazed at all of that, and I just hear the student agency in all that in terms of them having ownership of their own learning paths, and it just makes you such a more active participant than passively just taking information… it’s wonderful.

Well, let’s get to the 28-year-old. How are teachers integrating Khanmigo into their lessons plans to enhance their instruction, their support?

Sal: Yeah. The first thing that I’m really excited about is historically… and Khan Academy is guilty of what I’m about to say… is when, we used to… and we still do… we go to teachers and we say, “Look at our efficacy studies; if you just get your students using Khan Academy 30 minutes a day for three, four days a week, your kids are going to accelerate pretty dramatically.” And most teachers have always agreed with that, and they want to use that. But it is one more thing for them to learn and one more thing for them to implement and somehow slip into their lesson plans, which is work. And we know teachers are already spread thin, and you have record high attrition, et cetera. 

What we’re excited about is there are definitely aspects of Khanmigo that are going to enhance the student experience and that whole workflow, but this is the first time that we have a solution for teachers that directly helps them save time or helps them with their work. So, Khanmigo also acts as a teaching assistant, and we’ve actually just launched it so it’s free for any U.S. teacher, and the computation cost we are able to pay for with philanthropy. 

The computation is much more expensive for Generative AI, especially if you’re using these frontier models like GPT-4, but this allows teachers… teachers right now spend 5, 10, 15 hours a week doing things like writing lessons, grading papers, writing progress reports, writing IEPs, writing a whole bunch of administrative tasks, or refreshing their own knowledge in something, and so Khanmigo has activities for all of the above. And we’ve already gotten feedback from some mainstream public school districts that it’s saving their teachers 5 to 10 hours a week per teacher.

So, it does all that, but it also helps that workflow with the student and the teacher. So, a teacher can talk to Khanmigo like a teaching assistant, let’s say, “Hey, what have my students been up to?” 

Khanmigo will say what the students have been up to and will also make recommendations, “Hey, based on this, how about you assign this task to these 10 kids? Because they’re ready to move on. These 10 kids have a gap year. Let’s assign this.”

And the teacher just, let’s say, says, “Okay, sounds good to me.” The assignment is provisioned. 

They can work with the AI to come up with lessons, assignments, rubrics for, let’s say, an essay and by back to school, they’re going to be able to assign the essay through Khanmigo, say, “Hey, Khanmigo, just assign it to all my students.”

Khanmigo says, “Okie dokie,” and brings it out to the students; then the students work with Khanmigo as a writing coach, and then Khanmigo reports back to the teacher, not just the final output… which if you just focus on final output, the student could cheat. They could use ChatGPT for that, they could get their sister to write it.

But now, Khanmigo will give the entire process to the teacher, “Hey, here’s the essay. Here’s my preliminary assessment of it. It took us about four hours. Here’s my analysis of the student. It’s pretty consistent with their other writing so I know, I feel confident it’s their writing. They have trouble with the thesis statement. And by the way, half your class is having trouble with thesis statements; you might want to do a lesson, and by the way, here’s a lesson that you might want to use.” 

So, it’s really like a superpower teaching assistant.

Denver: Yeah. Yeah. So the role of the educator is evolving, and I guess you really haven’t run into that much apprehension on the part of teachers saying, “Are we going to be replaced by these little chatbots?”

Sal: I think it’s pretty, you know, we do a lot of professional development with teachers in our partner districts, and there is a little bit of suspicion initially that, “Hey, AI, I’ve heard about it might replace people in certain jobs, et cetera, et cetera.” But one thing we’ve been very clear about Khan Academy from the beginning, and especially now: This is never going to replace the value of a teacher. 

And what it does do is it hopefully allows a teacher to more personalize the education for more students, and it gives the teacher time back for themselves or for their students, so they don’t have to spend as much time lesson planning, grading papers, writing exit tickets, et cetera.

Denver: Higher order work they can do. Let’s get back to your initial conversations in the summer of 2022 when you were talking about some of the ethical considerations. What are some of the more pronounced ethical considerations that you and the team have really gotten your arms around and which we should all be concerned about?

Sal: The most glaring one is cheating and that one is in some ways, I think, the easiest to solve. Khanmigo will not cheat. You know, if you say, “Tell me the answer,” it won’t do it. 

And I just mentioned on this writing workflow, not only will it not cheat, but I actually think you could use it to undermine other forms of cheating.

And in my book, Brave New Words, I have a whole new chapter on cheating, and I start with the state of cheating before ChatGPT, and it was there, and ChatGPT just put a spotlight on something that was already happening.

Denver: Yeah.

Sal: And the real answer is: make it part of the process. 

And if we had infinite resources as a society, we would have every student have a personal coach or personal tutor, and that personal tutor would also be one of the teaching assistants for the teacher, and so that tutor would tell the teacher, “Hey, we worked on this essay together; this is where they needed help, but I know it’s their work because I’m an ethical tutor, and I’ve watched the student develop this essay.”

So, cheating is one. I think we have a good handle on that.

There are issues of oversight that we are still figuring out. We err and have erred on the side of conservatism where, right now, all conversations for an under-18 user in a school setting are viewable by teachers and administrators. 

We also have a moderator that if any of the conversations go a little bit funny, it automatically notifies parents and teachers and administrators. So, that’s how we’ve handled that.

Although, the reason why I said it’s a– I actually suspect we might have overdone it a little bit. There’s a lot of false positives right now. 

My daughter got flagged the other day for …she was brainstorming with Khanmigo, and she said, “Oh, I want some really good ideas for our project. I really want to beat the other team.” I think the AI took that literally that she wants to literally beat them physically, and so she got flagged and it wouldn’t engage in the conversation. So, we’re working on some of those tensions. 

The other ones are just making sure, and this is probably one of the harder ones, well, there’s hallucinations, which is the AI is making up things, and then the math errors in tutoring, I would say.

Hallucinations. Actually, I would say the underlying models have gotten dramatically better. I actually have trouble, I was trying to make a video on how to handle hallucinations, and I couldn’t make it hallucinate. So, the models have gotten much, much better. And on top of that, we are doing work where we anchor it on Khan Academy content. So, it’s actually hard to make them make up facts now.

The math errors are also dramatically better than they were even a year ago. But what we are seeing…it’s  less in math errors, it’s more of issues where, let’s say, the student has written an answer, let’s say, they wrote 0.33. The answer is one- third. We need to make sure that the AI can say, “Well, you’re almost there. Are you sure it’s not 0.3 repeating? And if it’s 0.3 repeating, you know, well that’s the same thing as one third. Okay, great job.” 

And for the most part, the AI is quite good at that, but it is not perfect at that. So, rounding issues where the student might round differently.

So, those are the types of things that we’re working on now, but it’s pretty much on a week… every week it’s getting better.

Denver: Yeah. Yeah. No, I run into my own hallucinations, and I sometimes think that the AI just doesn’t want to admit that it doesn’t know. So, it just comes… it’s very human-like, you know; it makes up their stuff.

You know, I know another real core value of Khan Academy is equity and access. It’s been in your DNA from the very beginning. Tell us what you’re doing about that as it relates to Khanmigo.

Sal: Yeah, this has been a tough one for us because, well, even today, there are many of these models, GPT-4, and there are other models like GPT-4 out from other organizations, GPT-4 is still probably the best one out there as of right now, although there’s talk about GPT-5, and Anthropic has just released a model that’s looking pretty promising, but these models are very, very expensive to run. They’re very computation intensive. 

Anyone who’s looked at NVIDIA stock knows where those dollars are going. These chips are in demand, and it uses a lot of them, at least these models. 

And this is more, you know, we are free, world-class education, so we care about the free and the world-class. And there’s a lot of people throwing… making apps with the previous generation of models, which are cheaper to run, but they… we don’t think, are the quality bar where you can actually put them in front of teachers and students. 

So, we are anchoring on the most advanced models, but they do cost resources to run. The good news is, a year ago when we launched Khanmigo, we thought it was going to cost about $70 per year to run Khanmigo. And so it’s like, okay, we’re going to have to charge something for this because we can’t afford, you know, all the R&D we’re funding primarily with philanthropy, but the incremental costs… we would go bankrupt if we had to pay OpenAI or whoever, those types of bills.

So, the first-year pilot, we did charge school districts for that. The good news is we’ve been able to bring the price down by about half through just efficiencies on our end, efficiencies in the underlying model, and we are pretty confident that those are going to come down even more. 

And right now, actually, the biggest cost for a school district, for us, well, it’s about 50/50. Half is the computation cost. The other half is us going and doing the support, the training, the professional development to make sure that the tools are being used well. But the cost is, let’s call it, $25 to $35 per user per year now.

So, it’s getting into an approachable range, and obviously our goal, once again, is just to cover the marginal costs.

Denver: Got it

Sal:  And on top of that, we’ve just made the teacher tools free for U.S. teachers, and the way that we’ve been able to do that is through philanthropic corporate support that’s allowing us to do that. That hasn’t been broadly announced yet, but we did a soft launch a couple of days ago.

Denver: That’s great. Very cool. So, let’s think about Khanmigo, and we talked a little bit about the student and the impact it has on them. Then we talked about the teacher… you even just brought up the school district. But when you begin to gather this all together, how do you think it’s going to change education as we know it, perhaps our traditional sense of what school is?

Sal: Interesting, I think the first place you’re going to see it isn’t going to be so obvious to parents and students, but it’s going to be on the teacher prep side. 

I think if we can help teachers save 5, 10, 15 hours a week doing things like lesson planning, grading, progress reports, that’s already a huge improvement.

Staying on that same thread, you can imagine a world where teachers are going to be able to write more progress reports and give families and students more feedback more frequently, which by itself… you know, the more engaged the parents can be, that’s a major lever in improving education. And then to learn anything well, the faster you get feedback, that’s going to improve.

From a student point of view, right now, we’re seeing Khanmigo as it’s really a way to keep students engaged. You know, the time that you’re most likely to disengage is when you’ve hit a wall; something doesn’t make sense; you’ve watched the video; you’ve done the exercises; you keep getting it wrong. Now, the AI… you can have a conversation with it, and hopefully it can unblock you more often than not. 

If you’re having a conceptual difficulty, no one is there to ask the question to. I’ve used Khanmigo myself as I still make videos, and sometimes to prepare some of my conceptual questions it’s really helped me to riff with the AI and it’s like, “Oh yeah, that’s okay. That’s an interesting way to visualize that.” Or, “Yeah, that would explain why A leads to B.”

As you fast forward 3, 4, 5 years, it’s going to get better and better and better, and I think in the five-year horizon, it’s going to approach what I was able to do with Nadia. I think right now, it is able to do 10% to 15% of what I was able to do with Nadia, but in the longer run, it’s going to be pretty darn good.

Denver: And then I think you’ve said it will probably be a bit more patient than you were.

Sal: It will be infinitely patient and always– although, you know, we do have a project in– because we’ve realized that a truly good tutor is not fully patient.

Denver: Right.

Sal: They have a certain impatience to them. So, we have a project called Proactive Khanmigo, where we are making tweaks to Khanmigo so it holds you a little bit more accountable. If you said you were going to do something, “Hey, how come you didn’t do it?” Or reaches out to you, as opposed to just waiting to be asked.

Denver: And another thing I love about this too, it’s consistent with the world that these young kids are growing up in. And very often I’ve heard complaints from my cousins and others that they feel like they’re going into a way-back machine when they go into the classroom, but this is pretty much what their contemporary world is like. So, it’s consistent with that in terms of how they communicate with their friends and how they do everything else.

Sal: I’ll say it’s even a step further than that.

Denver: Yeah.

Sal: Because a lot of us older folks now, I’m approaching 50, so I’ll put myself in that category now, we look at kids who are on social media, and they’re on their phones and they look distracted and actually, it’s not just kids, we look at ourselves, and we’re overstimulated and distracted.

Denver: Yeah. Count me in.

Sal: And there’s a fear that another technological solution will drive more distraction and obviously that’s always a risk, but what’s fundamentally different from interactions with a conversational AI from, let’s say, social media or YouTube or any of these other things, is you have to be a little bit more engaged. 

You can’t just passively swipe and look at the next TikTok video and go into that haze that we’ve all felt sometimes when we’re on social media. It’s implemented. Well, it asks you questions. 

The other day I was asking it about supernovas and supernovae… I don’t know what the plural of nova is, but I said, “Why does a supernova explode? You would think it would collapse when the fusion stops.” 

And it didn’t just give me a Wikipedia entry. It said, “Hey, well, why do you think? Well, Sal, can you explain what a supernova is?”

I was like, “You know, fusion reaction stops, but once again, that was providing all that outward pressure, but now all the gravity of the star would collapse it.” 

And Khanmigo tells me, “Well, that’s a pretty good insight. Would you think it’s going to collapse quickly or slowly?” 

And I said, “Well, you know, these are massive stars, a lot of gravity. I would say it’s pretty fast.”

And then it said, “Well, have you ever seen anything collapse or fall quickly without bouncing?” 

Then I said, “Oh, so are you telling me it collapses so fast that it essentially squeezes the atoms in the core, and then there’s a rebound, and then that shockwave sends out all of the kind of outer layers of the star?” 

And then Khanmigo was like, “Exactly.”

And once again, the reason why I can recount this conversation that I had with the AI about six months ago is because it was Socratic. If it just gave me a Wikipedia article and I just read it, I probably would not have retained that information.

Denver: Yeah.

Sal: But I give that as an example of like, this is actually, it’s not this scatterbrained, distracted, passive type of thing that we’ve all done on our phones. I mean, you might do this on your phone as well, but this is actually like a good conversation with a thoughtful person, and I left that 10-minute interaction with Khanmigo not feeling like, “Oh, I just wasted time on TikTok.” I felt like, “Wow, I had a meaningful conversation.”

Denver: Sure.

Sal: I hope that this form of technology can be implemented well and can move us more in that direction.

Denver: Yeah. It could be a real paradox because what you’re explaining there is that it has a way to keep us focused, and our biggest challenge right now is we can’t stay focused.

I even realized when I went on vacation, I used to take my iPad with all my books on Kindle, and I would never finish anything. So I ended up taking a physical book because I couldn’t jump around to nine other books. And anytime you can keep somebody engaged in a conversation, as opposed to a TikTok, where you’re going here, you’re going here, you’re going here; it’s snacking over a bunch of subjects, and you don’t remember anything.

Sal: And one of my hopes there, and I write about this in the book, and we haven’t implemented it yet, we’re implementing the first layers of it, is Khanmigo acting outside of Khan Academy. 

And so, it’s just on your phone, it’s in your browser, so even when you’re on other sites, it can say, “Hey, maybe you want to go back to the task at hand?” Or maybe, “I’m not going to even let you switch tasks right now because you told me that you want to finish this by lunch, so we’re going to finish this by lunch.”

“There’s also been questions, and a lot of funders push us on this as well, rightfully so. It’s like, “Are you going to be fully dependent on philanthropy, or should you create some earned revenue streams?’”

Denver: That’s cool. That’s great. I could use that. 

How does this impact the business model for Khan Academy and maybe other similar nonprofit educational organizations? You know, I mean, you’ve got a different model here, and you’ve got for-profit organizations probably involved in this space. I’m just wondering how you look at that entire ecosystem.

Sal: This has always been a debate for Khan Academy– to what degree, I mean, there’s a Harvard Business School case on whether we should even be a not-for-profit. Most students come to the same conclusion I did, which is we should for the benefit, for the long run, and for society. 

But there’s also been questions, and a lot of funders push us on this as well, rightfully so. It’s like, “Are you going to be fully dependent on philanthropy, or should you create some earned revenue streams?”

And the answer is: we should create earned revenue streams, especially as we have things that require a marginal cost.  Even before the AI work about five, six years ago, we went to school districts, and we said, “Look, here’s all the efficacy studies. You use this in this way, it’s going to really accelerate the students.”

Almost every school district has told us, “Well, we agree with you,” and they’d always tell a story of someone in their own family and the superintendent’s family who used Khan Academy. “But if you want us to use you all-district wide, you have to give us support, training, integration with our rostering systems, district level dashboards.”

And so that’s when we said, “Look, we can’t; that’s real bespoke cost.” And then we created our district offering where we essentially just charge– Khan Academy is still free, but if a district wants this enterprise version of it, they pay a nominal amount to us so that we can send people out there, do the training, et cetera. So, that started to build a revenue stream. 

And now, what we’re doing with Khanmigo, that introduces another cost, not actually just on the marginal cost. This obviously is a place where we also want to be able to take some of that and put it into R&D… there’s so much to do, especially now with the artificial intelligence.

So, we think that the middle road for us will be: we give as many tools for free to teachers and students and parents out there in the world. That obviously will be funded primarily with philanthropy. Philanthropy also helps fund some of our R&D projects, some of our growth projects, but we do build a flywheel around districts, and we’re doing this not just in the U.S.; we’re talking to districts in Brazil and India as well, where they pay us some amount, but we try to make it as affordable as possible because we do care deeply about access… so that we can cover the marginal costs and then maybe even help subsidize a little bit of the R&D.

But we tell folks like, “I don’t own Khan Academy. We have no shareholders. All of this is going to make the product better or to serve you directly.”

“In the early days, we had a general sense of free, world-class education for anyone anywhere. We didn’t know how that would manifest itself. So, we were really trying out a ton of different things…” 

Denver: You know, I went out to your campus a couple of years ago, and I spoke to all your team about the organizational culture, which is magnificent. But I’d be curious as to how it’s changed in terms of your bringing in AI; you have new market opportunities, new forces, probably need some new skill sets on a whole bunch of different areas. I just wonder how you have tried to lead that organizational metamorphosis and the influence that you had to try to shape it in such a way that it’s going to be able to respond to what’s happening in the world.

Sal: If I were to characterize Khan Academy in the early days, let’s call it, 5, 10 years ago, well, the whole time, early days and now, we’ve always done very well getting really strong talent.

You know, there’s debates about: Can a nonprofit attract the same folks who Google can pay a lot of money to… or that Meta could pay a lot of money to? And the answer is a resounding yes.

We have some of the top folks in the world on our team, literally people that have written the book on whatever the topic is. And we pay surprisingly well for a nonprofit. 

Our cash compensation is comparable to some of these tech companies, but we don’t give stocks, which tends to be a large part of that compensation, but we’re getting some of the top folks.

I think the big difference in culture is… and a lot of this might be coming from my own evolution. In the early days, we had a general sense of free, world-class education for anyone anywhere. We didn’t know how that would manifest itself. So, we were really trying out a ton of different things. And I also was less experienced about: How do you align people? How do you make sure that they’re on the same page? And so, that was good, like a lot of startups– you’re throwing things against the wall and seeing what sticks, et cetera. 

But I would say, in the last five years, we know who we are now. We know our mechanism for action. We know what we’re strong at and what we can be confident about, and we also know what we shouldn’t try to do.

And I think we’ve gotten a lot more focused and a lot more aligned. And I think this AI work, this has only made us more focused and aligned because it was… it’s not easy to pivot an organization.

Denver: Yeah.

Sal: And frankly, if my name wasn’t on the door, it would’ve been that much harder. I see a lot of other organizations across industries that are still, I think, fairly flat-footed when it comes to fully recognizing what’s happening with artificial intelligence. But yeah, I would say, it’s more focused and aligned than we’ve ever been.

Denver: Yeah. And sometimes something like this can be a catalyst for that because everybody has to gravitate towards it.

Let me close with this, and I want to close with your book, Brave New Words. So, I’m going to ask you two things about it. Number one, tell me about the experience of writing that book because I’m always curious as to what went into it. And then give a pitch for the book. It’s May 14th, we want to sell a few copies here. So, answer both those things if you could.

Sal: This is the second book I’ve written. The first book, the One World Schoolhouse, came out back in 2011. And I remember when the publishers first came to me to write that book, I was super skeptical. I was like, “Wow, anything I have to say, I’m just going to make a YouTube video and give it away for free.”

And they’re like, “Well, you know, it can help sometimes to put…” and I was skeptical, but I said, “Well, let me just try it.” I thought this was going to be like the last paper book ever written that, you know, everything is going to be…

But it was really good for me to structure my own thoughts back then, and a lot of the alignment that I just talked about Khan Academy having as a team came from me being able to sit down and write a book about: Why does education look the way it is? How did I fall into this? And where should education go? And what should schooling of the future look like?

And so when this AI inflection point started to surface… and we were one of the first to see it, and we were under a non-disclosure agreement with OpenAI in 2022, I kept telling the OpenAI folks, I was like, “Hey, can I write a book?” 

They’re like, “You can, but you just can’t tell anyone about it just yet.”

And I’m like, “This is going to be like the biggest inflection point in any of our careers.” 

As soon as I had the green light, I started on it. 

And a lot of the motivation was, I wanted to get my own head around this. Every hour that I thought about what is possible, I realized that I was being too close-minded.

And so, it was actually a very similar process of just like, “Okay, how does this affect cheating? How does this affect teachers? How does this affect parents? How does it affect admissions? How does this affect the future of work? How does this affect economics? How does this affect what corporations or what managers need to do?

And so it was really useful to just break that down into sections and have a point of view. And now, I have a point of view about most of these things, and they’ve stayed stable for the last few months, so I think they’ll…

Denver: There you go. Well, get it. Get it out there.

Sal: And, you know, the pitch for someone, I think these are questions that either people are already having, or if they’re not, they should be having these questions. And these aren’t narrow questions about education; these affect pretty much everyone.

Obviously, if you’re a student, you have your teacher, you’re a parent– these are like hyper-relevant for you. 

But if you’re an employer and you want to know, well: What is work going to look like? How much should I be using these tools to, let’s say, write memos? How should I be vetting people going forward? What is an appropriate use of AI for recruiting or not recruiting? Is a resume even relevant anymore when you could have an AI representative for a person? You could talk to someone’s resume, and their resume could talk to your job posting. That sounds wild, but this is doable today. It’s actually already happening.

Denver: Yeah.

Sal: So, I think anyone who cares about the future of work, future of education, future of credentials, future of parenting, teaching, this is hopefully a relevant book.

Denver: Yeah, no, it sounds like it will be. And it also sounds like you had all these ideas in your mind, but they were a bit of a jumble, and the idea of how to organize them and then explain them to somebody is a great process to go through. It really is a self-discipline process to say, “Oh, this is how I can logically talk about it.” 

Brave New Words. Well, I want to thank you so much for being here today, Sal. 

Tell me about where people can learn more about Khanmigo if they should be interested, whether they’re a student, a parent, school district, teacher, whomever.

Sal: So, if they go to Khan Academy, there’s links for Khanmigo. You can also just do a search for Khanmigo, K-, H-, A-, N-, Migo. It’s a pun on Spanish with me. Some people say “Khan Amigo” but it’s “Khanmigo,” and you’ll find it.

And teachers, we’ve just made it free for you, so definitely check that out, U.S. teachers. Hopefully, we will go broader over time, but parents and students can also get access to it.

Denver: Thanks so much, Sal, and you can count on me. We’ll give a big push to your book. Looking forward to reading it.

Sal: Great. Thanks, Denver.

Denver Frederick, Host of The Business of Giving serves as a Trusted Advisor and Executive Coach to Nonprofit Leaders. His Book, The Business of Giving: New Best Practices for Nonprofit and Philanthropic Leaders in an Uncertain World, is available now on Amazon and Barnes & Noble.


Share This: