TikTok 与 AI 已彻底改变教育——这对教师、学生和家长意味着什么

Will software in some way replace teachers?

Parents want better outcomes.

What is actually improving retention of information, learning, and memory?

They're allowing a teacher to be a lot more productive and be 10 times better at their job.

Will we see more parents say, “I want more of a truly AI-directed education for my kid”?

It's teachers who are willing to pay for and use this in their everyday workflow. We need the textbook companies to grow some innovation arms fast.

I'm Justine, and this is Olivia, and welcome back to This Week in Consumer AI. This week, we're going to be talking about a very fun trend we've seen recently blowing up all over our social media feeds related to AI and education. For this topic, we wanted to bring on our subject-matter expert here at a16z on education, our colleague Zach. Zach, do you want to give a little introduction and talk about your experience in edtech?

I would love to. I think it's fun to feel invited to a podcast by colleagues. It's an interesting feeling, for sure.

Quickly, my background: I've been operating and investing in education for quite some time. I've been here for a little over 2 and a half years, thinking about what the next generation of education is going to look like with the advent of AI. Before this, I was at General Atlantic, where one of my coverage areas in the consumer internet group was education. We were investors in Quizlet and Duolingo, and we did the Chess.com deal while I was there. So it was kind of the edutainment and consumerization of education.

Before that, I actually built and sold an education technology company that was focused on delivering high school computer science education, mostly in the Northeast. We sold that to an education-focused roll-up, where I worked for a couple of years as well, rolling up different assets in mostly adult education and corporate training. You can count how many times I've said “education.”

I know we've looked at a lot in edtech over the past 2 years. I think everyone has this intuitive sense that people are using AI for homework, but even I was surprised seeing some of the data from probably the full first year of ChatGPT. The education users were even larger than everyone else who was a non-student user, which is pretty crazy. It also feels like the backlash was as severe and as immediate as all of the anti-AI checkers.

Are schools still anti-AI? How has that evolved?

It's a good question. We had the AI wave, AI detectors, and kind of the Rock 'Em Sock 'Em Robots of education. You had the Los Angeles and New York City public schools banning AI right away. I think we're super far away from that moment now, which is really strong.

I think we're in different layers of “far away,” depending on where you are in education. If you look at K-12, there are still some skeptics. Predominantly, I think 80% of districts now have a generative AI team that's looking at and procuring new technologies. So not only is there earmarked budget, but there are people who are thinking and proactively looking for it.

There's still friction around what you want to use it for and how you want to use it. Maybe there should be a teacher in the loop who has more understanding of how students are using the AI, and maybe that's a good or bad thing. But I think it's a good thing, net-net, that there will be more AI in the classroom.

In higher ed, you have Claude for Education, and you have OpenAI releasing an education platform. They're partnering and piloting with a bunch of universities, and this is a pretty horizontal platform that will be fundamental to teachers' and students' experiences. I think higher ed is leading the pack.

I think they're realizing that this is going to be a tool that people will need to know how to use in their jobs and their everyday lives. So I've been pleasantly surprised. I know 18 months doesn't feel like a long time, but in education, it is. We've taken 5 or 6 years to move to the cloud, and most people aren't even there yet in the education world.

I think it's super impressive. There are some schools that are mandating AI usage in the curriculum. I think Ohio State is one of them. I think it will continue to trickle down. People are always a bit more careful and wary about technology with younger ages, but I think we've passed the hysteria moment and are now in the pragmatic moment, which is exciting, especially for founders building in this space.

Totally. I think you alluded to this when talking about your experience in edtech, but one of the things that's both interesting and scary about the education market for outsiders is that it's not just one industry, right?

There are public schools, private schools, charter schools, homeschooling, and supplementary products that parents buy for their kids. Then there's the whole higher-ed public and private sector. There's also adult education, future learning, reskilling—all of those sorts of things—which I would imagine are probably going to be huge in the age of AI, with many jobs being automated or enhanced.

I'm curious: Have you seen the most adoption among a segment of that market, like parents buying standalone products for their kids versus school districts?

I think I was—and still am—really surprised that this is my answer: teachers. It's not actually students. It's teachers who are willing to pay for and use this in their everyday workflow.

I've seen a lot of adoption among students who are trying to use this for homework help, but they're either turning it off or using it as a way to get their problem set done and then going back to hanging out. For adult learners, retention may have been a bit better, and the completion rates or learning efficacy on Duolingo have improved with all the AI features. But I have not seen a native AI player in the adult education space.

Maybe that's just because distribution is really hard, and layering AI onto existing strong pedagogy is the right way to do it. It seems like that's the case now. But with teachers, the rapid rate of adoption is so high.

A lot of that is because 90% of the job they hate is the administrative part, which is grading, feedback, going home, building new assignments, and creating new curriculum. A lot of them have just been borrowing curriculum year over year. They've been trying to iterate on 1 unit at a time, and now they can generate a curriculum per student. I think that's extremely exciting.

Hopefully, we'll take it farther, and it won't just be generating assets that we've already seen in education, like worksheets or multiple-choice questions. The units will start to become more of an AI experience. But for now, teachers have been the wave.

I think MagicSchool has over 5 million users. I think 50% of US teachers have used their tool. Most of the most mature AI companies from a revenue standpoint have been selling to teachers bottom-up, which is tremendous. Teachers have a very small wallet to spend on tools, and this is just an outsized return for them compared with the $15 or $20 a month that they're spending on it.

This brings up another point when you think about what's driving student outcomes. There are a couple of ways to look at what is working in AI and edtech. One would be what is getting the most usage, which, to your point, is probably still ChatGPT for students, versus maybe more focused products.

Then there's what is actually improving retention of information, learning, memory, and things like that. Have you seen anything that you've been really impressed by in terms of what's working in AI and edtech?

I think the core of the question is the question, which is: What is working? What does “working” mean?

There's the investor side of what's working, which is what you want to see when you're investing in a company, and that's different from the learning-outcome side. From the investing side, I care a lot about looking at retention and engagement.

Retention is important because you have summer months, but you also have students who come on with a test that they have in 2 days. Is the experience good enough that it captures them and gets them to work proactively on the app the next time?

We look at metrics such as cohorted engagement to see how students are performing, things like that, or monthly retention—less so weekly or yearly retention, which I think could be a red herring in this market. I love looking at the weekly level: How many days in a week is a user using this, on a cohort basis, as the company grows? Is this starting to become part of their learning?

I think that's also a good proxy for how good the product is, not just how engaging it is. If a student is doing this just to get their homework done, I don't think you're going to see that level of engagement. You're going to see it taper off around certain times or certain months, or be a little spikier.

But people who are seeing their engagement flatten in terms of the number of days per week—hopefully above 4 or 5, given the school week—I think that's exciting. On the education side of what's working, this is a really, really hard question. I think it's probably true of a lot of different markets that are suffering from this, which is: What is the benchmark and eval function in education?

Luckily, education isn't like a corporate job where you have an ambiguous performance review and maybe you're doing 10% better. There are tests at the end of the year that measure student performance.

The problem with that is that you need multiple years of student testing because there are a bunch of variables. One is figuring out what's causal, what's independent, and what's dependent in terms of what's working in the school. On top of that, AI is still at the periphery of education, right? Since it's at the periphery of education, we don't actually know what its impact is. It's helping teachers create better assignments and helping students get their homework done, but it's not core to the schooling system.

I think it's really hard to know the learning outcomes. I think it has to be a mix between the metrics of the product and how good an understanding the founder has of how to build pedagogy and learning methods within the product. I think we're still a couple of years away from actually understanding what the learning goals are. There are some papers that have come out from various schools and universities that have shown test improvements with an AI-instructed course or things like that, which is exciting. But again, these are research papers and case studies; these aren't statewide or nationwide studies.

Yeah. Yeah. Okay. To get back to the point about making AI core to the educational experience, that touches on one of the viral social media accounts we've been seeing. I'm sure anyone in the education space—and definitely you, Zach—has been seeing it. I think it's literally called The Future of Education. It's an Instagram account run by a woman named Mackenzie Price, and she runs a school called Alpha School, which I believe is a charter school in a few states now. It has had these viral Instagram videos blow up by showing how her school has 2 hours of classwork a day, delivered via AI tutors—computer programs that teach kids at their level—and then they spend the rest of the day on self-directed projects or things like that. I would love to know your take as the education expert, Zach, on what you've heard about Alpha School and whether you think it will work to deploy AI like that.

Yeah, Alpha School is everything I could ask for about what's happening in education. We talk about it, we opine on it, and then we have to sit here and wait for superintendents and district heads to buy single-point AI solutions every turn of the budget. Over 3 to 4 years, we can have a narrative.

I think what Alpha School brings us is what any labs team would bring to a company, right? It's like, "Hey, the skunkworks team is going to try out a bunch of new features. They could deprecate whatever features because they're moving fast, they're not deploying a bunch of resources, and they have an outside budget to go do this." Then let's see what works. What works, we're going to take, and what doesn't work, we're going to deprecate. We're going to continue to iterate.

I think Alpha School is doing that for private schools now and hopefully for public schools later. The tuition is $40,000, and they're a super well-funded private school. The parent and student base is self-selecting into a group that wants technology in the classroom. These might seem like small facts, but all of them in amalgamation change the outcome of what a school can do and remove a bunch of friction to allow you to move fast.

I think what Alpha School is showing us is that if you turn full tilt on AI and education and bring it into the classroom, and you're fine with overspending on software and spending a bunch of money testing different software, you can figure out what's working. My takeaway here is that this is a great signal for where technology is today: if you have the privilege to go test it out and figure it out, it will have an outsized impact.

I think they ranked in the 99th percentile on a few different assessments. Their students are now in the top 1% or top 2% in the nation. It is incredible outcomes, but again, this reflects the lack of commercialization of a lot of this software. Part of that is what the integration looks like. Part of that is simply the cost of software, and part of that is software literacy at the school level.

I think Alpha School is a net extremely positive signal. I think it's an incredible way to learn. It's leaning into a lot of the old ways that we've talked about learning: self-discovery, self-learning, and very limited instruction. And it's working.

I think public schools are looking at it. Maybe superintendents are saying, "Well, I don't have hundreds of thousands of dollars in a software budget." But I think that's okay, right? We see software get cheaper. Let the Alpha Schools of the world test out what's really working, and those companies will really benefit from early partnerships with Alpha School. To me, it's paving the way for a lot of companies to see what they're doing, obviously from a very privileged standpoint where Alpha School is interesting.

Yeah. I think the controversial question there is that when you talk about superintendents not having hundreds of thousands of dollars to spend on software every year, they do spend that on teachers. So the controversial thing is, will software in some way replace teachers? Will AI replace teachers, or will we still need a human instructor to deliver the curriculum or give lessons to the majority of students?

Yeah, I think no, or it's a very long horizon away. There's still a shortage. There are still teachers who are swamped with work. We've been talking about what consumers can do with AI and what even enterprises are doing with AI, and we are multiple years behind there.

The biggest use case of AI right now is generating an answer sheet or a worksheet assignment. It's not even a unit in the classroom. What I mean by that is that it's actually a really big, nuanced difference. One of the reasons I've had a little bit of a hard time investing in these early education companies is that they're allowing a teacher to be a lot more productive and 10 times better at their job, and to feel like their burnout is a lot lower. Incredible, right? Really, really amazing. But it doesn't mean that the students are interacting with AI in a learning environment.

They're interacting with the same asset that they would otherwise, and they're having the exact same experience. Maybe they're having updated worksheets. Maybe there are better memes or better references in those worksheets because the teachers can update them weekly. But that's very different from being in an AI-enabled environment where you can take a history lesson straight from a historical figure's avatar and speak to them, have a detailed conversation, create your own world around what you think, write a creative-writing prompt and go create your own world around it, or create a game around it. I think those experiences are far away from coming.

My point here isn't to avoid answering the question, but to say that we're so far away from even AI teaching units. I think we're very far from AI teachers. What I think will happen is that the amount of active teaching will go away, because you can have really good active teaching with AI. But I don't think it's going to fully replace a human teacher.

Super interesting. I do feel like, on the opposite end of the spectrum from Alpha School, which is very private and very high-cost, there is this explosion of free, publicly available edtech content. But it looks very different from the prior best-in-class, which was the Khan Academy videos that I learned math and other things from and that I love. The open-source animation package is called Manim; it's still around today. Now we have people putting their textbook into NotebookLM and making a podcast. We've found a bunch of TikTok videos that generate Sydney Sweeney explaining this.

There's this Instagram account called Unlock Learning. Every video they post gets millions of views, and the explanations are delivered by deepfake AI celebrities teaching math and physics content, particularly for the AP or IB exams. It also has diagrams and graphics that walk you through the explanation. I guess it just makes me think: to what extent are we going to see standalone AI edtech products versus just new formats and ways to learn that are delivered across all of the platforms that we already have?

Yeah, I think it's a great question. I've watched way more educational videos in the last 4 days than I have in the last 10 years, which again is a really good signal that they're engaging. They feel like brain-rot content, but they're actually the complete opposite. They look, feel, and sound the same and are cut the same way, but they're teaching a really detailed and pretty technical topic, and they're getting better.

I remember a year ago it was Taylor Swift teaching the Taylor series: "So this is the Taylor series. It's a tool we use in math to turn scary functions into friendly polynomials." Now you have Drake in conversation with Sydney Sweeney: "If we square-root it, we get the vector's length. And that, guys and gals, is why this expression gives you the length of a 3D vector."

Talking about three-dimensional shapes, it's getting better both on the animation and graphics side and in terms of the celebrity effects. It's all getting a lot more engaging.

What I think is really interesting here, and the thing to focus on, is—I don't know if you remember, as a kid, you were like, “Oh, Zach's a visual learner,” or “Justine likes audio,” or “Olivia likes to read and marinate with her ideas.” You kind of get boxed into a type of learner, and that was considered pretty progressive education: “We know what type of learner you are, and we're going to cater to it.”

The problem is, I'm a visual learner for one thing, but I might want to listen to a podcast for another, read for another, watch a brainrot video for another—it could be so many different things—or do 100 problem sets, right? So I think what's happening now is that we're going to have a kind of factioning into a bunch of different types of learners, where, depending on the topic and your understanding of the topic, you could pick whatever modality you want. I think that's really exciting.

Depending on how serious you want to get in the topic—if you just want to be conversational around it and it's not for a test—maybe brainrot content is perfect. But if you're taking an exam that matters for your final grade, maybe you want to generate a bunch of problem sets and then listen to audio that explains your answers, right?

So I think, for me, I don't know if any of these turn into individual companies, but what's really exciting here is that the mode of learning is going to be tailored to the type of learner you are, and the burden of fitting into a particular mode is going to be removed. To me, that's the most exciting aspect for education right now, and I think we're seeing it. The engagement of these videos is insane. I mean, this is not a brainrot video—this is not Italian brainrot; this is very different.

Well, even the Veo 3 vlogs, where it's a day in the life of some famous historical figure, are incredible. It's fascinating. It's truly entertaining to watch, and I find myself actually learning things.

I think the hard part of that is that there are a few different companies doing really interesting things at the forefront of AI and education. One of them feels like what the Veo 3 thing is doing, where you can chat and have a conversation with Napoleon about history, things like that.

But in order to get school adoption, they have to wedge in with these more rudimentary tools, which are worksheet generators, student helpers, and feedback. Once you do that, it's not so easy to get people to use the other mode because it's a new form of education, versus, “Hey, do the same thing, just do it 10 times better.”

It's interesting to see that the stuff that engages us the most and gets me the most excited—even if it's being adopted and paid for by the school—has extremely low usage compared with the more rudimentary tools. There's probably going to have to be a large professional-development movement here on how to bring AI not just into your workflow, but into your classroom. I think that's a really big difference.

Yeah, it's kind of fascinating to think, too, about if you were to design the school system from the ground up, starting in the AI world, how it would be different. One of the things that has really interested me about the brainrot videos is that, for the first time, it feels like it's separating what the content is and who is delivering the content, and then optimizing both of those.

If you look at the comments on these videos—we're sure we'll show some of them—it's often people who are like, “I was a physics graduate student, and I got a 4.0 at this prestigious program, and I'm telling you guys, this is an exceptional explanation. I've never understood this content better.” You're kind of optimizing for, “Hey, this is how you explain; this is how you deliver the content in a good way. Explain it in a good way. Make it understandable.”

And then here's the brainrot celebrity: this is a familiar voice. This is her first video on the topic. Exactly. This is an interesting graphic or image to bring those things together, which we've pretty much only seen with self-directed AI products today. Traditional schools can't really separate the content from the teachers, as I understand it.

Yeah, it's interesting. I think the answer in my head is probably that the textbook companies kind of control this dynamic a lot. They are still the gatekeepers of a lot of content and what gets in and out of the classroom.

Now, I think it depends on whether they run with this or not. Sometimes they view this as cannibalizing their existing business. Sometimes they're like, “Oh, this is extending our existing content, and we have the single source of truth, and AI is really good at augmenting but not net-new creation.”

So I think what's going to happen is really dependent on where the textbook companies and the publishers spend the next 18 months thinking about net-new products. Are they going to partner with AI companies? The AI companies will get distribution, and the textbook companies will get a really strong product and technology extension of their existing content, which feels like it's slowly losing its value day by day. So it's a weird power dynamic right now, but I think that's kind of where we are.

Yeah, we need the textbook companies to grow some innovation arms fast or outsource them. Zach's on one side of the spectrum, which is, “This is how the educational system actually works,” and it's very complicated. In some ways, it's like the health care system or the tax system, where there's a bunch of annoying, outdated stuff, but you can't really change it.

And then I'm on the other end of the spectrum, where I'm like, “Blow it all up and have everyone be taught by an AI tutor all the time.” If LLMs have ingested the whole internet, they've probably learned a lot of stuff, and they might even know what to teach us that teachers don't.

Yeah, totally. And I would assume that the Zach perspective is more where we are today. But I'm wondering: are we going to see a lot of people basically reckoning with what to do with their careers, what to do with their lives, and what to do with their education in the age of AI? Will we see more parents opting for something more on the second path, like, “I want more of a truly AI-directed education for my kid?”

Parents want better outcomes. They don't like—that's the reward model for them. It's, “Are we going to invest in technology? Are we going to invest in this tutor? Are we going to invest in this course that allows my child to be better off from an education standpoint?”

I think one of the most interesting pitches I've seen—and the company was super early—was an AI reading company, which basically sent kids books, and then AI taught the kids how to read. It was mostly three- to four-year-olds, and the promise was to get them to read at a third-grade level within 3 months. It was $500 a month, and they had a bunch of parents sign up within the first 2 months.

If they can prove this outcome, I think they'll get a lot more parents. If they can't, I think parents will say, “Well, AI and technology maybe isn't the right way to do it. I'm going to go pay a tutor.”

The other side of this is that if you have the disposable income to spend $100 or $200 an hour on a private tutor, the comp to using AI is a bit harder. If the AI software is priced against that, then you have a little bit more friction on the adoption curve of AI. But if it's, “Hey, my kid's going to watch Netflix for 4 hours, or I can put them in front of an LLM that can teach them things and keep them engaged,” then that's a very different conversation.

So I think it's very socioeconomically dependent. I think it's very geo-dependent. I think it's very resource-constrained-dependent. I think there's probably a future for a few students to feel that way.

Yeah, especially because LLMs can be controlled, right? Parents could say, “Hey, I want you to explain this a certain way.” If it gets easier to control them, even for nontechnical users—which it continues to be, right?—parents could say, “I don't want you to discuss this topic,” or, “For every question, I want you to show the mathematical equation, then write the proof and explain it.”

I think all of that control is really exciting for some parents. We just have to figure out how to harness it and how to benchmark it, and I think it could be a potential future.

Okay, so looking a year from now in AI edtech, first of all, what do you think we're going to see? Is it a lot of progress? Is it a little bit of progress? And what would you like to see ideally?

Yeah. Oh, so much. A year from now, I think we'll see a ton of progress on the higher education side. I think this is the year for higher education, this coming school year. How does higher ed use it? How do the large model companies work with education companies?

The question is, are the large model companies going to be good enough? Will there need to be smaller application companies built on top of the large model companies? I think that's one thing we're going to learn a lot about.

I think in the next 12 months, we're going to move from the outside of education—the use of AI just to make things easier—to bringing it into the classroom, whether it's adding it to a discussion. A lot of the unlocks in voice and real time are really exciting, so I think the next 12 months will show us a lot.

I don't know if it will necessarily birth a massive company in education. I still think it moves.

But I think education will look fairly similar, though. I think we'll learn a lot as the market matures. What I'd love to see from a consumer perspective is the first fully AI teacher influencer. Not one of the deepfake celebrities. I know they're super entertaining, but it would be fascinating to see what it looks like if you designed the most engaging but also truly informative teacher from the ground up.

Totally. You alluded to this as well, Zach: there’s probably not one teacher who’s that for everyone. I might learn best from an animated dog, whereas you might learn best from an actual photorealistic teacher. How can we create those adaptive, AI-driven characters in real time who can deliver really personalized learning on each student's timeline?

We touched on this a bit, but I think one of the exciting things that schools like Alpha School are enabling is that you can spend as much or as little time on a subject as you need to learn it, master it, and move on. Whereas in a traditional school system, almost everyone is moving at the pace of their classroom. You're moving at the average pace of the 30 other kids. And you can't really speed it up or slow it down unless you have an external tutor like you mentioned, Zach, that's several hundred an hour.

So anyway, that's what I would like to see: more AI teacher influencers and more apps that allow personalized learning speeds.

Yeah, no, I think that is the right push. Language models are good enough to teach. They really are, right? If you want to learn something, ChatGPT is a great place to go learn something. Obviously, make sure everything they're saying is perfect, but it's a great place to learn something. It's not really an engagement product, especially when you have to learn it and there's stress involved.

I think all of the infrastructure around avatars, around voice, and everything you're describing, Justine, is going to actually unlock a lot around education. I'm sure we'll unlock a bunch of use cases around more engaging or less engaging things that need to be more engaging.

Yeah. Right.

So I think it's interesting. We're at the moment where I think LMSs are trusted as a learning partner. Now we have to build experiences around them that feel native to how you interact with language models and native to the tools that we can build with AI today and where students are engaged. The brainrot video is a great example of all of that.

Amazing. Well, thank you for educating us all on this, Zach. This was awesome.

Yeah, of course. It was great talking to you both.