Can AI Really Tutor? Reed Hastings and Mike Goldstein Debate

What are people actually missing here about either the potential or the lack of potential around AI, the casual conversation, what you see in media coverage, the kind of thing that non-experts are thinking about. What are people not seeing that they actually need to be aware of? Reid, I'll start with you. AI is getting better rapidly. Let's call it roughly doubling in quality every year. So if you think about chat GPT three years ago, it could barely do third grade math and now we're doing breakthrough post-graduate school level math. So that's just in the last three years, it's gotten roughly eight times better. That curve is likely to continue. So that means in five years, it'll be about 30 times better than it is today. And in 10 years, it'll be about a thousand times better than it is today. So grasping these exponential phenomena where it's getting twice as good every year is very hard because there's very little in our human lives that are like that. So when anyone says, well, AI can't do X or Y, just gently correct them to yet. It can't do that yet, that is correct statement. But there really is likely to be able to do everything that we care about better than us 10 years from now. Everything? Are you saying is everything like in terms of cognitive tasks, like certain kinds of work or like, I mean, how expansive of a statement is that? Everything's a big word. It's pretty expansive. Okay. Mike, talk about, you know, I do wanna get this question of, is this a question of sort of velocity and we should just be patient or are there inherent limitations? But start with, what do you think people are missing? Like when you listen to people talk about this and you move through the world, what do you think people are not seeing here about AI that they need to be more cognizant of? Well, I think, you know, Reid and I probably have a starting, a shared common starting point, which is this is sort of an irresistible force meets immovable object and Reid described the force of this constantly iterating AI improvement. The immovable object is that there's this history in education with ed tech, but with other things, that when you give the kids a new tool that better explains math or English or something like that, a lot of the kids simply don't use that tool. And so there's all these randomized control trials over the years, products like DreamBox from back in 2013, through this week of Khan Academy, an economist in Canada just ran a trial on schools in Tennessee, where the behavior you see is the kids are like, hey, here, go use this ed tech tool. Now it's AI, but it used to be something more static. And the vast majority of kids simply disengage. And so the total amount of learning gain of the kids who get access to the technology is like, you know, maybe 0.04 standard deviations, a real tiny amount better than the kids who don't get access to the technology. And so that's the immovable object is kind of this ability of the kid or the motivation of the kid has always pushed against coercive. And by the way, that extends also to pencil and paper things it's not like we should single out technology as the only thing that's unable to get a lot of kids to try hard when they don't feel like doing it. But that's where we don't yet see any doubling as best I know from AI about motivating kids who describe themselves as, I don't really feel like doing this stuff. And that's kind of the linchpin of, we see these forces of the improving AI against the motivation of the kid and he will win in five years. Yeah, and Reed, I mean, at the end of your piece, you had a piece in the Atlantic, we'll put in the producer, we'll throw it in the chat. And I think you landed like the humanness here. And so Reed, like, as you talk about this velocity and how much better AI is getting, which I think anyone who's using it, we are seeing, how does that like, is that human dynamic fundamentally an immovable object or is it gonna get better to where that, that does not become a limiting principle insofar as engaging kids is concerned? Well, I think Mike's right that, we're subject to being very excited about new technology, whether in the 1960s, that was beaming television to people's homes for educational learning or whether it was the web generally, let's just wire up the schools and like net day in the 1990s or things like DreamBox in the teens. So I'm sure that will happen. I'm sure there'll be a hype cycle. I'm sure there'll be uses of AI that don't quite deliver on what it says. And when I think about that, some of that is the grammar of schooling. So the grammar of schooling is sort of the unit structure that you're there for 45 minutes doing English and then take a little short break and 45 minutes doing math and that kind of structure. And then all 12 year olds are gonna learn the same thing on this day. And when you think about the variation in 12 year olds, I mean, it is vast. And trying to teach all 12 year olds the same thing, I think we might find is the fundamental flaw in what we have, that we have on the charter side, we've tried to do with match and Kip and others, classroom education better. And some of that is helpful and good. It's hard, but it's helpful and good. But we may be embracing a fundamental flaw, which is all 12 year olds should not be learning the same thing on a given day. And that really the breakthrough is tutoring and individualized education by which I mean, like alpha school, no classroom, no everybody's learning the same thing, which never used to be affordable. And with AI potentially becomes affordable. So I think the promise of AI is really each kid goes at their own pace and that all 12 year olds are learning different things. And therefore they're much more engaging. Why is it hard to motivate kids? Well, it might be because there's 30 of them in a classroom learning something at the same time. And a third of them already know it, a third of them are interested and a third are really bored and behind and alienated. So I think the fundamental thing may be trying to teach 30, 12 year olds the same thing. And we got to find a bunch of experiments and data around what happens if you give up that huge assumption that embodies basically all of Western education and Eastern for that matter, which is we're gonna mass teach. So let me ask you both just a question to help people ground kind of where you are. Like Mike, you mentioned a second ago, like a lot of kids are like forget AI, a lot of kids are just disengaged in analog classrooms. And I mean, we had a disengagement problem before the pandemic, the pandemic really heightened it. So when you guys think about this, what is the counterfactual that you sort of think about? Like, what should we be evaluating AI against? Because I do think AI is often in a lot of the conversations it's evaluated against like the ideal thing. Like people say, like an AI therapist, well, it's not gonna be as good as a great therapist, but most therapists are the average therapist. That's how things work. Most teachers are the average teacher, most tutors are the average tutor. So help people, ground people in like how you think about this. What is the counterfactual we should even be thinking about this against as we discuss what we should be expecting? Well, one counterfactual I think that we can consider is a human tutor. And as you and Reed know, I've done a lot of work at MATCH at our charter school, basically had this whole army of human tutors that were really carefully picked. And we were trying to understand the ceiling effect of like how much faster could a kid learn if in addition to let's say five hours of class a day, hopefully from a good teacher, they were also getting two hours a day of tutoring. And we then rolling fire, the economists replicated some of that work in Houston in turnaround schools. And my friend, Alan Safran replicated that in Chicago, sort of had this idea of high dosage, high quality tutoring. I think that's important because the counterfactual here I think would be at what point does AI equal what a mediocre human tutor does? And at what point does AI equal or exceed what a good tutor does? One of the things we found is, you know, and you've written about in your policy work is as we took these sort of little island of excellence tutoring situations of humans where the tutors were really carefully picked, then we scaled them with public policy with a lot of indifferent managers, like, I guess I got to do tutoring this year, but it'll be gone by next year, we'll do something else. And they just got any reasonable human. The total learning gains, I think, of a typical human tutor are larger than what EdTech does now or what this sort of trial of Conmigo showed yesterday or what Stanford found in the June 26 paper with this super high disengagement by kids. But it's not like it's this game changing, wow, once we introduce typical human tutoring, we make gains. So I think the counterfactual that we're all working for is what point does the AI eclipse the typical human tutor and at what point do they eclipse a good human tutor? And to do that, we probably have to learn more about what differentiates, what separates the good human tutor from the mediocre human tutor. Yeah, and all the political constraints in the system. I mean, this reminds me a lot of class size where the class size research is actually super compelling if you really lower class size, you have great teachers. But when it hit, made contact with the political space, it was like, yeah, we'll lower it a little bit because that's what we can afford. We're gonna hire whoever we can get. So you ended up with all the usual problems with teacher hiring. And so the results were very uninspiring, that fidelity problem. And that seems like that's plagued human tutoring in a very similar way. So Reid, against that, how do you think about the counterfactual and how do you think then about what we should be thinking about in terms of success and progress? I think broadly with tutoring, there's sort of two modalities. There's remedial and catching the kids up, where's the kid stuck, which is where the high dosage tutoring has been aimed at. And then there's tutoring, like we're gonna learn a new thing. There's no teacher, there's just a tutor. Okay, and the tutor you've got all day contact with. So most of the literature that you've heard about is on the former remedial catch up. And the system uses it to try to get all seventh graders at the same level so that the teacher can do a better job. Okay, and so it's aimed around making the factory model more efficient, which is good for the kids and good for the teachers. And it's not aimed at, hey, let's race ahead because then the kid's just gonna be bored in class if you let the kids race ahead. So this was for me when the big aha with DreamBox, which is districts aided up for remedial work and pull out for remediation, but let it be adaptive and the kids go ahead. No, no, it's gonna make the kid restless in class, hard to manage. And so it was in some sense, one half of the puzzle. So what I'm being interested in at Match, Mike, you had this whole floor of tutors. And so take us through that and how much it was remedial, which again, helps those kids tremendously to feel successful in class and very positive for the whole system. Or what would it have been like if there were some kids that were living amongst the tutors up on the top floor, like the whole day long, all year long, sort of the bloom theory. So three things. One, from a liability point of view, I get nervous about the kids living upstairs with the tutors. But I take your, I think we did run a version around 2014 for one year, didn't hold, wasn't really published, but it was like essentially what you're describing, the kids all day with a group of three or four and a tutor. And I've observed that also in a number of the micro schools I've visited over the last couple of years, some version of that. I think a couple of things are going on. One is we do see this bimodal distribution where there's some kids that are just pretty much wired to try hard. And if you unleash those kids, they can learn more efficiently. They're not constrained by the group. And then there's another group that often they are missing. So they're missing a piece of their motivation, which was sort of the social dynamic of the group of 20 kids. And now it's really like, I didn't like math before and now you're telling me I'm gonna do it all day like a group of three. And it's really feels like those to the grindstone type stuff. So we did see like the kids getting classes in tutoring versus the kids getting, if you will, only tutoring. The only tutor kids did make some gains in math that even eclipsed the high gains of the regular match kids for lack of a better. So it was like regular school down here, that's typical and then even a little higher, but they lost on a little bit on their reading gains compared to some match kids. We wondered if that was missing the teacher who was guiding the discussion about literature. There's only so far you can go with like the same talking about the book. And the vibes were like the kids were ready to return to, you know what, I liked it before when it was like a couple hours of tutoring and then a bunch of hours of more class 20. So I wonder about even the social dynamic, even as this becomes more viable and the AI tutors get better. So I think there's this threshold question. One is, will the AI get better at motivation? And then there's a second question of even if they do, does the experience of five, six hours a day of we're never together really in a big group or we're less together in a big group, how does that play with kids? Do they choose that? Do their parents choose that? Mike, are you a Fitbit user? Right here, buddy. Okay, so the coaching on the Fitbit, the health app, you know, has gone from irrelevant to sometimes useful. It's still a bit light and a joke, but it's been an amazing amount of improvement in the last year of the exercise motivational and very much exercises a lot about motivation, at least for people our age. In any case, so I would say that's one way to watch it as the AI learns. And then what they're doing is a lot of learning and randomized controls and seeing, you know, they try different things and see which sets, you know, actually exercise more, which they can tell from heart rate. So I would guess that AI gets quite good at motivation over the next five years, because it's broadly useful and there's no reason it can't learn for this types of humans what is motivating. I will say, obviously, that children are wired to please adults, right, as parent substitutes. And so a teacher has this big advantage. They're sort of like a parent. And so depending on the age, they are sort of, there's pleasure for the kid in pleasing the adult. So capturing that emotion, figuring out, you're right, that it's fun to do things with others. And so if you're in a class of 20, doing something engaging, that's quite a positive experience. May not happen a ton, but it's quite a positive experience. And I think if the notion of individualized education is you're alone in a dark cube for eight hours a day, that's horrible. So clearly we need to learn, if I'm right, a new grammar of schooling, which is when do you work on your own? When do you work with others? At what size? But it's a fundamental re-plumbing, which is wildly challenging for school districts. So maybe charter schools will be able to figure out that. And Silicon Schools now, which is a funding group like Charter School Growth Fund, they're working hard to fund a bunch of labs to basically figure out this new grammar, which is probably closer to alpha school than other things. And the question is, alpha school is a $60,000 version per kid per year. And then what's the $15,000 per kid per year version that charter schools could do? So that's yet an unsolved problem. Yeah, go ahead, go ahead, Mike. I was just gonna, two sort of burning questions I've been dying to ask you, Reed. So one is, I think you've said before you see this acceleration happening, and I'm curious about, I could imagine two scenarios that lead you to the same place. One is, you've seen how difficult and brittle it is in your desire over the years, our shared desire with Andy to improve schools for kids and for kids. And it's just sort of like at some level, there's a certain amount of at least partial exasperation. And like, it's gotta be this. We got like, this AI stuff has to work. And so you're hoping for it and indexing, that might be one plausible story. The other plausible story is you're just seeing AI win in so many use cases. You're like, I think K-12 will be like the other use cases and AI will win and ultimately triumph in helping kids make really large learning gains. I'm just wondering how you come to, is it more with exasperation or is it more like, no, I'm just sitting here and watching all of the wins that AI is putting on the board and it just seems to be inevitable that it will happen here in this space too? Well, we're both Larry Cuban fans. And for any audience member that hasn't read the 30 year ago book, Tinkering Towards Utopia, it's a marvelous account to say the introduction of kindergarten and all the revolutionary zeal around kindergarten, which you'll be able to relate to as the change agent. And it puts it in context of, and gives you respect frankly for the bureaucracy, which is, it's a set of systems designed for, well, it has evolved, wasn't designed, but towards satisfying a set of societal needs while parents work, while kids have opportunity. And I have a profound respect for the system. People wanna say it's totally failing and all of that. And clearly America is this rapidly growing, we're double the economy compared to Europe over the last 20 years. And so it's not like our education system is holding us back or it's a total train wreck or failure. There's a lot of places it could work better and we see that but it's super easy to oversimplify of the quality. And again, if you look at Mississippi in the last five years, by focusing on phonics and being very steady and less political, now the NAEP reading scores for Mississippi, black children, white children, everyone are way, way up above. And so they've got like a mass solution to one problem, which is reading. Unfortunately, that solution doesn't then extend itself to how do you teach seventh grade math and how do you teach ninth grade history? But at least in one pretty technical aspect of learning reading, they've really cracked it. And of course a lot of states are now following them, which is great. So to answer your question, I come at it all from the opportunity side. Like I look at self-driving cars and I drive a Tesla Y and when I say I drive, I punch in my address and I'm just so fascinated because Tesla promised this a decade ago and it took way longer, it was way harder, but it's really, really impressive. Not yet perfect, but really, really impressive. And so I think of things like that, which is if it's succeeding in these other sectors, again, over time and probably overhyped initially, but over time succeeding, that will be an opportunity. And self-driving cars will open up so much for, again, people who were not safe driving, older people, younger people who are not allowed to drive, bringing down the highway death rates, all kinds of super positive things. I think in teaching, it potentially allows this new grammar to emerge, this individualized grammar, but we've got to prove it out. So I would imagine there's a decade of, say charter schools, if it's a powerful tool. So I'm guessing that it is, using it to demonstrate vastly superior results. And then that will infiltrate the thinking around school districts. It will also be possible out, there's sort of outside of the US, like in a developing world, school systems and teaching is much harder and class sizes are larger and it may have an even bigger proportional impact there. And it does here. Again, we'll see over time. I think for the audience, one of the interesting things that we're funding blooms the 1984, if you can believe it, study that showed how tutoring was so much better to standard deviations of improvement. And when you go back and read the paper, it's kind of a joke. I mean, it's like six weeks of study of astronomy and it's like, wow, this is not really trustworthy at all. And so a colleague of mine, Ben Rosen is running a study, sort of called Modern Bloom. And we're taking over time, it'll be 50 or a hundred relatively random public school kids, median kids out and they get full-time one-on-one with a human to see how much can they learn kind of map scores, things like that in one year. And so that should be really set a high watermark. And then back to your thing, it's when does the AI match that? And when does it match it across really a representative sample of kids where motivation is the dominant factor? And a friend of mine, Luis runs Duolingo and he's like, it's all about motivation and engagement. Tactical things about learning theory barely make a difference. Things like the streak and how we do make a big difference on the language learning for adults because it's motivating. So it's not just kids. I think generally learning is work and yeah, there's a small group that are intrinsically motivated around that. And then the breakthrough for society comes when it's engaging to do a lot more. Why are the politics though gonna be any different here? Like I think the self-driving car is a really interesting example. The technology is pretty good now. The politics have not caught up. And so it gets banned because it's an easy thing to ban. It's fearful. I've seen people who just don't even wanna get into Waymo. And so people are very leery of this idea that you have self-driving cars that people have. And so it's fairly easy for that still to get bottled up politically. And so it's why Waymo is even struggling to get into a lot of cities. Education politics are just tortuous. And so why is this gonna be the irresistible force that gets us over? Part of the grammar of schooling is a political adherence to the way we do things and powerful interest groups. So why is this gonna be the thing that's gonna succeed where most other things have bounced off? Well, I think we need enough surface area, essentially test cities in the Waymo example, which I think charter schools will form that. Charter schools are so focused on getting more results for kids that if it works, they'll be willing to pioneer it. And then your question is, well, how do we get school districts to do it? That will be a big challenge. Maybe it's some states move ahead of other states because they really lean in on this like the Mississippi type example. So we'll have to see, but it has to be a big positive effect to make rewiring the schools worthwhile. I don't think we're ready to engage at that basis now. Now it's the stage of just doing some lab schools. And again, I give credit to Alpha School, which frankly barely uses AI, because it doesn't yet work. But what they do is they have a tutor for every five students. And then they use pretty simple software, IXL, DreamBox, Khan, this kind of thing. And it's human provided motivation. And they hire the guide primarily on their motivational ability. Okay, back to Mike's point, which is the hard problem here is motivation, not so much learning science. Yeah, I wanna add, I think there's three components of that motivation where I think like AI could make progress, but there's sort of three pieces from what I've seen. And this is stuff that my colleague Sean Gary and I, we used to run research for this company called QMath based in India, which has 3000 human tutors. So we watched a lot of video of humans trying to motivate the kid. And you can see the mom kind of over here off the screen, like kind of nudging her kid. So it was like, Indian middle-class families trying to get their kids ahead. And there were sort of three ways to win, I think as a tutor, all of which maybe futuristic AI could work on. One is, I think people often misconceive of tutoring as explaining. And sometimes that's true. And Reid, you were sort of saying, sometimes you're just presenting a new idea. Hey, today we're gonna learn about photosynthesis. You don't know what that is. I'm gonna explain it. We're gonna learn it. I'm gonna tell you. But a lot of times you're trying to get the kid to blurt out the half-formed gray mist idea in their head so that you can help clarify. It's like they're sort of stuck and to get them unstuck. AI is a very good explainer and I'll stipulate it's gonna get infinitely better. But already to me, it's so much better at explaining than most experts I can talk to. Very good at that kind of synthesize. The behavior we see in the randomized trial so far is the kid simply will not reveal their misunderstanding to the AI, which again is what separates a good human tutor and a mediocre human tutor. They also don't tell the mediocre human tutor what's going on. So the mediocre human tutor just kind of blah, blah, blah. Just explains, here's how you find the slope of a line, kid. The kid kind of tunes it out. I think the second way to win that AI already does some of, which is it can notice things about you. You mentioned Fitbit, but this is sort of visually you can see it's like during any moment when a kid starts to lose steam and you can pick that up, that body language, a lot of times human tutors miss that because they're spending so much intellectual power just trying to figure out the academic part of it that they miss some of the cues. And that's an area where I think AI could pick up on, hey, I noticed you're kind of tuning out. Let's take a break. Let's reset. Let's go back to this. Let's do something fun. I could imagine AI getting good at this. I think the third way to win though that I feel like will be hard is some kids describe, I will try hard for Ms. Smith, the tutor, because I just respect her so much. She has presence. I will not try hard for Mr. Goldstein, the tutor. Not that I don't like it, it's just sort of eh. I just, he's very meh. And one of the things that seems to generate effort you mentioned before is not just the learned behavior of respecting the adult, but a genuine sort of earned respect or like of. And I've seen people make the case that AI will be delivered in a lot of different formats and there'll be like the cute robot version and so forth. I'm curious about that, but I wonder if that will come anywhere approaching what each of the three of us had experienced at different points in our lives where we really try hard for a particular teacher or coach and that unlocks motivation where they just have outlier respect from us and we really want to succeed so they appreciate us and see us. And so I wonder how AI will progress in all of those areas. There was just a study out on your second point that was like showed some positive effects just for like helping kids, just with basically executive things like, hey, slow down, you're going too fast, do this again. And so it wasn't so much instructing as sort of nudging for sort of behaviors you want to see around learning and had a, we'll throw that in the chat, had an effect. It wasn't compelling overall, but did have an effect on that, helping kids slow down, stay on task, which is good. A teacher can't do that across a class of 28 or 30. That seems like a benefit. Yeah, there's a new Yale paper to that effect too, which is sort of, some of this is you're noticing and negotiating with the kid at all times when you're trying to get them to learn. Well, and one of the companies that I'm working with is Elo, hello without the H. And they started as a reading company and doing well in that consumer. And then they've evolved their AI infrastructure to do a generalized learning engine. And it's mostly conversational. So it's mostly talking back and forth with the six-year-old, the seven-year-old, the eight-year-old and that's what's providing the engagement. So we've also got to get out of the, the AI is not just reading on a screen because that's very flat and alienating. And to the degree that it's a back and forth conversation and you probe it, you know, think of teaching or learning as laying down different neural trackways and you're trying to reinforce certain things. And one of the best ways is to get kids to explain to you, the tutor, you know, what the thing is. It's engaging for them, it's fun for them. So, you know, they're in beta on this now. So you'll be able to play with it kind of by the end of the year. And, you know, it's just the amount, the way the kids light up because the computer is really good at listening to the kid. And again, consumer market has to be very safety controlled and there's a lot of conversations you can't have if the kid or it's really like an adult who tries to trick the system into, you know, pretending it's a kid. But it's got good safety on that aspect. And then it's really good at teaching the primary school curriculum. So there's a lot of innovation that's happening in different parts. If the front camera was always on and watching the kid's face, then you get, as you said, Mike, the reading of what's, you know, when are the kids excited? When are they not? Being able to tune that. And, you know, we'll just progress on that year by year. Again, I'm sure there'll be some people who claim, you know, that it does solve every problem and every kid will thrive. And we've kind of been through those hype cycles, but just because it gets overhyped doesn't mean it won't work. And the clearest example of that self-driving car, which got vastly overhyped, but finally does work. What's the timetable here? In other words, let's say we want to make a friendly bet that Andy adjudicates for a beer in X years. And to win- Well, this is Cal-She. I mean, you know, it's for the production market. Oh, no, yeah. It'll be a cocktail at equal measure. Reed's been doing some product placement. So I'm gonna, I gotta do a little too. The cocktail at equal measure in Boston. What's the timing, Reed? Like, when would you think would be a reasonable point at which we could sort of say, either we'll see kind of large gains that you're describing you think are gonna happen from the doubling of the AI itself, or we would sort of dial it back and say, we've sort of seen some kind of ceiling of friction that we haven't expected. Okay, so I'm trying to think how to frame the bet in a verifiable, you know, it's one thing to say the AI can do it. It's another thing to say the AI is doing it. So- I can think of two bets, by the way. I'll offer them to you. One would be a NAIP bet, but that sort of describes scale, and that might be sort of adding in a political layer that's not. I think a more modest one would be one of the charter type ones that you're describing, a pilot that has a good control group. And you see the kids that have access to the AI making these really large academic canes that weren't true of the control group that doesn't have the same access. Okay, so let's say there's at least one public school, probably a charter, you know, with at least 500 kids where there's no mass teaching going on. There's just small group and tutoring and, you know, so like a pure play in that. I would say three years from now, there'll be schools laddered up that have 500 kids that are in that modality. Yeah. I think the results will be, you know, 0.3 standard deviations higher than, you know, in kind of map progress. So not the full potential yet, but, you know, but enough to like, wow, that's a big result, scale it. And then the key thing potentially of scaling it is as less and less dependent on the quality of the adults and more and more dependent on the quality of the software. Got it. So, and another bet would be, and I'd be interested to hear your all's response, what percentage of kids just nationally in the country of school-aged kids would be getting like what Kim Smith would call like H3 learning, horizon three kind of learning in a decade which what would the penetration of that look like? What's the definition of horizon three? You know, so individualized, very highly customized individual one-on-one kinds of learning and student experiences. So essentially the kind of high variability. Given the infrastructure dynamics less than 1% in five or 10 years. Okay. So, you know, to rewire schools in that way, to find roles for teachers, if this becomes a teacher replacement strategy, it's probably doomed at least for a couple of decades. It has to be a new role for teacher that's much more fun where the teachers are getting to stimulate individual kids and see the joy of that. And so it has to be really teacher led for it to be widely adopted. And in fairness, that's a part of what Kim's talking about. It's not a pure teacher replacement play, but it's definitely, it's a different modality of delivery play. Mike, how would you answer that? I guess I would channel Larry Cuban. I think it'll go a lot slower. I mean, the same reasons that Reid was saying upfront, but I do think that there's something to, one of the interesting things to me has been what's the parent demand side for micro schools for some version of this experience right now. And I'd see numbers like in the hype cycle where the New York Times was saying, well, yeah, like a million kids are in micro schools, but like I know the guys running all the different micro schools chains and they're just sort of like, yeah, we got like 3000 kids. We've got like, you know, Alpha's got like their thousand or whatever. Like it's not a lot of people running to these schools. And I think part of it when you talk to kids and when you talk to kids who've been inside micro schools and then returned, let's say to a traditional high school, there's something about the vibe of the shared journey of like, all right, if we got to learn photosynthesis, it's more tolerable when we're in a group and there's also other social dynamics that have nothing to do with core academics happening. And that's why I think that the teacher with the room of 20-ish kids might persist for longer than we think, because it has an appeal separate from the efficiency of learning, but also because that grind of being roughly alone for that many hours a day, it's just hard to persist. I think that's what- Yeah, I think if we architect the new models as rough, the kids are roughly alone, it's doomed. It has to feel like, no, you've got a companion all the time and you've got both other kids. And I do think you're right that, think of it as we're trying to help kids get a big fact base and reasoning proclivity, but we're also trying to help them with social emotional. And yet the SEL part is a very background force that somewhat happens in some schools kind of by accident. But imagine if a school sat down and said, okay, the software is gonna take care of photosynthesis and we're gonna really work on managing emotions and SEL and sort of the Valor type work as a core focus. I mean, we might be able to draw a lot of teachers into a very positive vision in that way of their role. And that stuff that will be much, much harder for AI to teach than photosynthesis. So it may be that we find a hybrid, but for sure school or kids want to be around other kids. Okay, and so whatever these new models are, have to be highly kind of kid to kid fun oriented rather than something like, you know, do the crossword on your own, you know, in a dark cube. So high student agency, but within that some of the structures of school is basically what I'm hearing. Well, let's see, it's definitely not homeschool. I think parents want the custodial aspect of dropping the kids off in a building. So that may be obvious to the three of us, and micro schools are included in that, you know. So, but it's dropped the kids off for six to eight hours. And then the kids get some time outside the family. So it's positive for society in terms of the broad socialization. And the kids get time with a lot of other kids, which is, you know, important. And so finding out how to do those dynamics where the tutoring is individualized, or at least small group, for example, you might have two kids like, you know, find the problems with photosynthesis, when, how does it work on the moon? You know, the set of provocative type questions. Okay, so a couple of lightning round questions. We're getting some really good questions in the chat. One kind of goes on this issue. It's sort of adjacent to this issue that you all were discussing. Gregory McGinnity is one who we all know is wondering, like, oh, sorry, it was Zach Gregory. It was Dan Meyer started with this question. What are the indicators that you all are looking, what are the leading indicators that you're looking at to see if AI is actually gonna transform the classroom? What's it gonna transform first? What are you watching for? One thing I'm watching for, oh, go ahead, Reid. No, you go. I was just gonna say one obvious thing would be if customers start to substitute paid tutoring for, you know, they start using AI instead. So for example, if AI becomes pretty good at this, then you would think parents that are spending on the tutor, particularly very price sensitive parents would substitute towards Claude and other tutors. So that would be one leading indicator that we're making progress. The AIs are making progress. Sure, LO would be a good, if LO is a commercial success because parents are spending $50, $100 million a year in total that's either net new funding or it's reading coaches and primary school coaches substitute. So is there a growth in the consumer market? Is a good one. I would say, you know, the other ones that there's a lot of experimentation and back to, I think of alpha school as like the Tesla roadster. I know I'm filled with Tesla examples here but the Tesla roadster was $100,000 electric car was short range, but very fast. And it basically said the electric car is not a golf cart. Okay, it was a repositioning of the category. So alpha school, because it's marketed as AI school sort of says, okay, this individualized AI driven where then you only need two or three hours of traditional photosynthesis. And the rest of the day we get to do cool stuff is a very powerful message. And it's redefining the top end because in many ways, that's a better school to go to than a traditional private school with its kind of stodgy rhythms and formality. So I think alpha type models will sweep the high end of the private school market and then start moving down to the middle of the private school market. So that's basically alpha growth would be another habitable hypothesis. All right, so Jack Shaw wants to know and he's with Comprendo, one of the founders there. Jack wants to know if the educator role shifts from delivering instruction, so less about you're going to do less about the photosynthesis to use the example we've been using and more about relationships, social experience and all of that. What do the educators actually do all day and how do you even select and train for that role? How does that job change? Well, I think it changes somewhat, but again, teachers care about their kids' social emotional state quite a bit already. Whether the training and teacher education evolves, that may happen, but they'll just be the study, I think, of the models if we make it important. We probably do need some assessments of SEL. So it goes, you know, when you test it, it becomes more validated or sort of, I'm sure there are ways of testing, at least in high school kids. So it's developing that whole field, which is being more thoughtful. Today, our goal for schools is to create economic actors. So they have economic opportunity. That's kind of the number one real goal that parents have. Now notionally society, we have the goal of creating good citizens, but like no one really cares that much and no one makes the link that strongly. But I think that's an important one, especially as, you know, our society continues to change. We need, it would be advantageous for us to have more and more citizen preparation and trying to think that through. And then there's this emotional preparation, which is how do you understand emotional self-regulation? And again, probably Valor has done some of the best work on it, but we need like 20 different approaches like that to try to work out what are some of the modalities that when you focus on a SEL, you get really good results. So I would add two things. One is, I think, you know, you could imagine that more in the elementary. In the high school, a lot of the things that drew a teacher to wanna teach at the high school level is often love of subject. So if they're kind of moved out of teaching science or teaching history, that job becomes less appealing to them. It's just something to keep an eye on. I think that the fascinating thing is even separate from schools, we have this huge issue where so many teenagers have substituted away from doing real life things, whether it's sports or arts or, you know, reading novels or hanging out with their friends, part-time jobs, volunteering, all this stuff to read. And if schools become more efficient, I think that's one of the core pieces of Alpha and others is like, you can get a reasonable amount of stuff done in a two or three hour school day that frees up a lot. Well, then we really have to figure out how we're gonna reverse this behavior change because we're sort of saying already with their 3 p.m. and on hours, kids aren't doing, many kids, there's a lot of teens doing plenty, but there's a lot that aren't. And now you're adding more discretionary hours to the day. It could be used beautifully, but if we're looking at the trend lines, it's actually people moving away from that behavior. So I think an important part if we succeed and I'm hopeful that, you know, I'm sort of betting the negative case, but I'm hoping to be wrong because it'd be so good for kids. Kids are learning more efficiently, they're making more progress in all these things. We've got to get the real life side better fine-tuned and I'm not sure what the right mechanism is there because the attention, the screens have been doing a good job at winning that battle. Gregor McGinnity, this is the question he wanted, it was a more political question. He asked, if children are designed to please adults and the adults they are closest with to begin with in school see no value of striving to learn, does that end up being a counteracting force here to the power of AI to improve learning and outcomes? I'm not getting the supposition there. The supposition is we have sort of lack of focus on achievement results and so forth. We're coming along saying this is going to deliver better achievement results that the adults who are closest to it are going to be there, they're resistant to that. You can certainly challenge the premise of the question, but you can see where he's going with the question. There are going to be a lot of second and third order effects if you introduce, I mean, individualized like all the 12 year olds learning different topics at the same time, that's such a radical change. There's going to be chaos and there's going to be other things going on. I'm not sure that that particular effect will be the dominant one, but I think we have to keep an open mind about, as we introduce frankly the biggest change in a couple of hundred years of education, that we got to take it step by step and be very clear eyed about which parts are positive and which are negative. But my macro point is all the work we've done, whether you were in the systemic reform side and test the kids a lot or the teacher prep side or the charter side has basically been doing classroom education better. And that gets us some progress, there's no question, but this may get us a lot more if Bloom was right. That is that kids when individually tutored by a human learn at much higher rates. And then the question is, well, we can't quite afford that. So can we get AI as good or better than a human like we are getting in medicine, but not yet quite there, but like we are getting in driving. And so it does follow that parallel. And I'm hearing your other claim is very much, be patient, this is moving fast and we're gonna continue to see potential and that velocity. And so don't sort of judge it on the present technology, like it is getting better. It is such a self-improvement cycle that it's going to surprise us. Correct, yeah. The underlying improvement of AI on the essentially the scaling laws, but let's just think of it as it's probably gonna continue to get twice as good every year for a very long time. And that's an unbelievably rare and powerful phenomenon. Mike, do you agree this is gonna be the most transformative thing we've seen in education a couple hundred years? And you wanna summarize sort of your central claim? Yeah, my central claim, I guess, I grew up near Amish country, but I'm not Amish. I love AI, I use it all the time. And I think when you're motivated to learn something, this thing is incredible. And I think it'll continue to iterate. So really, this is really a threshold question about what happens when the typical 12 year old doesn't really feel like learning how to add fractions. Is this the game changer in that context? And I'm hopeful, but I'm betting sort of closer to the null case, I do think we can see, I very much believe in the idea of islands of excellence. And there's gonna be some charter schools with awesome that figure out how to make this particular thing work. Or private schools. I am also fascinated by and wanna take the bet on the behavior of wealthy parents and whether they start to substitute away from Milton Academy type traditional private school, alpha or similar. Be fascinated. Okay, so that bet is, does alpha grow? There were like eight schools last year, I think. Let's see, the bet in three years, is there at least 100? But would you take the under on that? And I'll take the- Well, no, because the cost of customer acquisition is only they don't mind spending infinite because Joe is underwriting the whole thing. So I feel like that's a weird- No, but they still have to attract a customer. I mean, they might say, you might say there's big marketing behind it. What's the number of enrollment? Let's say there are about a thousand kids now. Okay, yep. Are you saying like 12,000 kids in a few years, something like that? I'll take the over on any variant of this. On any variant of it? Yeah, that one, it's sort of, it's a tougher one to me because it's sort of like about product market fit of a really narrow class of kids who are self-starting kids who would be fine in a traditional private schools, but their parents might want to lean on it. I don't know. I was trying to get it as an indicator of sort of AI as aspirational because they're marketed as the AI school. They are, I agree with you. They're not really the AI school. Not yet, but they're basically, they're writing a lot of software and they're becoming it rapidly. So it's a great case of, you know, sell the vision and then scramble to build the implementation. Yeah, they all have to outrun the curve. There's always fear of new technology and then people embrace it. I think with AI, we're not, we haven't even gotten to the peak of the fear cycle and how that's gonna play out. We're coming up on time. So I would take the under on that one, incidentally. We're coming up on time. So last question, and it's a big one, but just like, you know, you take a look at a lot of technology 101, I think it's hard to argue that like the personalization of news and content and algorithmic delivery of information has actually been particularly good for us from a democratic standpoint, small D democratic. And Reed, you just mentioned like schools do have an important role in helping sort of create transmit values, traditions, customs between generations. So just stepping back, make the case, both of you like that this is actually good for us. And this is something we should embrace rather than this is something that people, I think a lot of people hear this individualization of education and personalization. They see it as like, this hasn't been good. This is probably something we should try to bottle up rather than enthusiastically expand. I can go first with that. Cause I do think this one is straightforward. It's very painful to work with a kid, middle school kid, a high school kid who's really struggling in math and English and so forth academically. They're often very frustrated. They want to be better at those things. It's just been hard for them to unlock their potential. And so if we have this tool come along and meaningfully change achievement for a lot of kids, I feel like that upside sort of speaks for itself. Hard in a scenario where, now I realize there's a lot of, as Reed said, second and third order effects, but if achievement in the country and the world can rise a lot, I think that's a pretty strong case in favor of this being a good thing. Similar cancer therapy today. If you're unlucky to get a cancer, it's very individualized and your oncology is very specific to your genes and tremendously powerful in being adapted that way. So I think when you're trying to achieve a specific goal, you get that. I do think that social networking is polluting the culture because you get these anti-screen bans in schools. And if the school is running good software, that should be like a lot of time. And if the school is letting the kids do Instagram, maybe not, but that is getting mushed together, unfortunately, in these screen bans. And you get that a little bit, probably with personalization. So I think where it works well, for example, in cancer therapies, it's incredibly powerful and you would want it. Yeah, and the issue of quality versus quantity with screen time bans is a huge one. And we're gonna be actually doing some work on that, which is a good way to land. If you wanna learn more about our work on AI, which includes advising companies, large and small, and a lot of policy and field-facing work, check out our website. Thank you for joining us. And thank you to Mike and Reed. I mean, I read, and we put them in the chat. I was reading your takes on these really interesting and important conversations. So really grateful that you're both willing to take a morning out of your summer and sit down and discuss this. So thank you both very much. And thanks for everybody who joined us, either watching now or coming along later to watch. Thank you all. Thank you.