Teaching the Machine to Teach: The Quest for a K-12 AI Tutor

[MUSIC] This session was recorded live at the 2026 ASU GSV Summit in San Diego. [APPLAUSE] >> I'm Jill Barche at the Heckenjee Report. We're an education only news outlet. And I've been spending the last few years covering how well AI is working in classroom, so in laboratory studies. And one of the main things we've been looking at is AI tutors. I'm very excited about this panel because we have four completely different types of companies, big established ones you've heard of, new startups. We have companies that can tutor any subject to any student, any time. We have ones that are laser focusing on one subject only and they're all built differently. So I wanted to begin with Maureen, Heyman's, who's of Google. Give us a little snapshot about how your AI tutor is actually being used in the classroom right now. >> Yeah, I know it's always a funny question for Google because we have so many products and students already come to us for so many of the educational need. And we are really trying to evolve each of those products to become a great tutor. But maybe I'll pick one example because of course it goes from search where students comes with a lot of questions about concept they're trying to understand to YouTube where they consume a lot of great content for learning and the teacher also leverage to notebook LEM, to Gemini and many other tools. But one tool that I think has really seen a lot of great adoption by teacher and students is notebook LEM. And notebook LEM. Yes, and so teachers can create a notebook by uploading a bunch of contents from the curriculum, from the students, so I had the study notes, all the contents so that the tutor is really grounded in the class context. So that it's really kind of aligning to where the students is learning in the classroom. And then thanks to notebook LEM, we can make learning about that content much more engaging because you can get multiple formats. One of the features that I think a lot of people started using notebook LEM for is audio overview. The podcast. Yeah, the podcast. And so you can listen to that podcast while running or going driving. And so I think that's, but you know, when there is the podcast, there is like video overview slides. And so it really enables you to learn in multiple formats, which we know is super helpful for students. You can also have a conversation, like you can actually set it up to use this learning guide mode, which is a much more so credit conversation, which has open-ended questions that break down problems. And so really make sure that you build this deep understanding. And how do you turn on the learning mode? It's a toggle in the conversation setting, and so that's something you can do. It's also some setting that you have in Gemini, and that we are trying to bring in more places, depending on the context you can decide to adopt this small, so credit, yeah, tutor. Thank you. Caleb, you're also, you're with school AI. This is Caleb Hicks. And he also, their company also has a tutoring system that can handle any subject. Can you give us an example of it in classrooms right now? Yeah, one of my favorite use cases is actually supporting multi-lingual learners. One of the stories that we hear over and over again, but I'll give one from January this year, is a student that moved here from Guatemala, moved to the United States from Guatemala, and oftentimes these students, when they're learning English, they get tracked way behind where they actually are. And so on day one, her teacher gave her a school AI sidekick, and that was speaking to her in Spanish. And they very quickly realized that she was in the wrong math class, and the wrong language arts class, and they got her into the right level. Just because she was able to speak in her language and get support in her language, and what one of the things that we do that's really special is that isn't just create the tutor and give it to the student, it is we have agents that are watching the conversation and finding insights and opportunities to support the student and giving those to the teacher. So the teacher didn't have to know Spanish either to go and read and understand how the student was doing, it was getting again kind of highlighted back to her, where this young lady was, and how she as the teacher could go and help. So that feedback loop of AI tutor with the student, very powerful, but also when you can start connecting the dots back to the teacher is where I get the most excited about AI in the classroom. John Woodard at Wolfram is taking a completely different approach. Azer focused on one subject and working hard to iterate it a bit more before it gets distributed widely. You're still, I believe, in pilot beta testing. Give us a snapshot of what it looks like as you're testing it in the classroom. Sure. So we actually have classroom usage, and then outside of classroom usage. So it's trying to have an iterative loop between the teacher, like a use case, in the classroom where the teacher is directing a specific subject, and actually there to actually help the student while they're using it. And then at home, where the student is able to, if the parents are interested, actually use with their parents. So that's been really exciting because what we actually noticed in one of our pilots was that there was a student that really didn't seem to be all that engaged in the classroom where the teacher and the principal kind of had this impression of the student that he wasn't quite interested in the material. But then it turned out when he was actually going home that he went through, I think, almost 75% of the material, when we'd only expected them to go through about 20%. Because he was just so engaged and so interested. So that gave them a completely different idea around engaging, carlo, and interacting with him, and knowing that he was actually really interested in the material. We have a Cartic Mongolum. Am I pronouncing your last name correctly? Yes. This is our startup on the stage right now. And you're famous for some test prep tutoring, but you also have a math product for K-12. Can you give us a snapshot of what that looks like in the classroom right now? That's right. So our math product is new. We are in our second year of free our public usage. And it's called EverTutor. And we do only math and we do it in a way which I think is state standard align, correct and so far we have about a million hours of math instruction done with EverTutor. In particular, something that I've been very excited about is continuous assessments with AI tutoring. So typically the standard assessments that happen at the end of the year or even throughout the school happen in testing windows. And those are lagging indicators of what is actually the case. And it's much more of like you have after the car crash report instead of being able to guide the student through the year. And an AI tutor much like what Killa was saying is able to not only do the job of a supplemental tutor, but also keep a close watch on how the students are performing and then navigate accordingly. And also surface those signals back up to the teachers and the relevant authorities for intervention moving them around either to a higher grade level or to a lower grade level or intervention and so on, which otherwise is just lost with traditional software. Right. Let's do a quick round robin for you to get a sense of scale right now. How many students, how many schools start with you, Kartik? Give us a scale. So we are in 13 states this year we are about to crack into I believe 10 million hours of tutoring and so on. How many students? We have at least 8,000 students with recommended dosage and about 25,000 students total. Morning. I mean Google is a global company and so I would say you here depending on how you come. It's billions of user but again, different level of engagement. Do you have any sense of your tutoring modes, what the usage is on those right now? Right a few millions, yeah. A few million, nothing more specific than that. Okay, Caleb. We are used in about 1 and a half million classrooms. And school AI altogether. What about the tutoring part? The special, the reason why people use us is the tutoring part. We have teacher productivity and so like that, but I'm talking about people that use spaces is what we call them. And that's with sidekick, right? Yes. And so say that again, how many schools? Like 1 and a half million classrooms, we have partnerships in every state in the US. We have the entire province of New Brunswick. So we have users in every country except for North Korea. - Okay, and John? - Yeah, so we have two pilots and maybe about 100 students. Total. - Now I'd like to switch to pedagogy. These AI tutors are built differently. John, how are you thinking about pedagogy, like whether it's inquiry or step-by-step explicit instruction, who comes up with the curriculum? - Yeah, so I mean, that's a really good question, and that's like how we have gone from originally when we were in the Learning and Engineering Virtual Institute or Levy program to now. Like we came in with a lot of assumptions about how we were going to do things in terms of both the pedagogy and curriculum, and it turned out we were kind of all wrong. Because what we were thinking was we were going to dictate to parents and educators and administrators about what they were going to do, which doesn't really go over well when you start interacting with actual people who are going to use your software. So over time we went from trying to create our own curriculum to actually being able to be coherent with what's being instructed in the classroom. So that's why our pilots are more in depth, more critical, and they take longer to kind of set up and do, because that's really our aim. Everyone comes to the style of teaching. It's exactly the same thing. It's trying to match that instruction as opposed to trying to propose our model. Because again, it's the same thing that if you're a school district, a parent, administrator, you've already made those decisions. You've already invested in the infrastructure that you're going to use to actually teach your kids. And it should be you who makes all those decisions not us. We're really kind of-- we're seated into the background and become more about enabling technology to do all those things and promote all those decisions that people have already spent. I mean, that's you guys expertise. How about you, Caleb at school AI? What's the pedagogical approach? I think-- so my experience is in learning design personally. So the way I set it up was-- we're a little over three years old now. And most teachers had not even played with AI themselves yet. And so we set up a bunch of modes, similar to learning mode, like we've got a secratic mode. And we have a direct instruction mode. And so teachers can just pick something that fits their style. But really what we've done is we've said, here's a box of LEGOs teacher. How do you want this tutor to interact with your students? We let them pull in the state standards or upload their own curriculum. But the teachers deciding the curriculum-- Teacher directed for the most part. And if they don't want to be prompt designers or AI specialists, then they can just pick one of the modes. I think the-- It's most popular are people which mode. It would be between the secratic and the direct instruction. I think one of the things we introduced last year was we called them power-ups. But there essentially apps that sit next to the tutor. So imagine where the AI is role-playing as Abraham Lincoln. And the student is interviewing Abraham Lincoln and writing the newspaper article that would be written after the Gettysburg address. And so some of it is just like, that's not really socratic. That's not necessarily direct instruction, but just a new type of experience. Alaborative learning, maybe. Yeah. And so, Maureen, I would imagine with Nopak Alam, those teachers are uploading the curriculum, right? Yes, it's grounded on the curriculum. Now, of course, large language models are built up of authoritative web source. But one thing we really did because we have such a diverse set of product is we work with pedagogy experts both internally and externally to define principles that we thought were critical to be really in few steep, foundational models. So we learn a lot. We define principles like managing cognitive load, deepening, meta-cognition, encouraging active learning. And so we really made sure that through a full product development process, we are evaluating that we are following those principles. And that's really deep in our models. And so active learning is a great example of that. But we really try to make sure that it's not just about passive consumption. It's really about interacting, having a socratic open-ended discussion with the tutor, but also being able to interact through simulation and 3D diagrams, being able to practice with adaptive quizzes. And so that's one of the key principles that we are embedding. Sort of a hybrid of several methods. When you talk about practice, that's not the same thing as socratic, right? I mean, it's a combo. Yeah, I mean, I think it's maybe because we get so many diverse needs from both in the classroom setting, but also teacher continuing to learn at home. And so we are trying to bring all those different principles together. With ever tutor, how do you think about pedagogy? Yeah, so we are, as I said, just math. And within math, there are at least 23 different pedagogies that are-- Everyone argues about them. And they're inconsistent with each other. And everybody thinks theirs is the best. So we try and stay out of that as far as possible. Because what we do is-- and we are kind of unique in our approach-- we mix in hybrid AI instruction with human instruction at the same time. Explain that. What that means is that as the teachers are teaching the usual way, whether they're using slides or whiteboards or notepads, the AI is happening live at the same time in the context of the instruction. So the instruction, for example, happens-- let's say on a whiteboard-- the whiteboard-- real whiteboard. The teachers using expo markers. Still whiteboard with a digital whiteboard, not a real whiteboard, but a digital whiteboard. While they're teaching, it is available to all the students on their devices. They could either just see the whiteboard or they could see their device. But if they want to get a harder problem instead of the one the teacher is working, or ask more questions, they do it naturally on the whiteboard as if they were asking the human. And the AI comes in and helps them right there on the whiteboard. And then they go back to the class. So it's happening on top of live in-person instruction. So that's what I mean by fully hybrid model of human and AI at the same time. So that takes care of pedagogy because teacher is fully free to teach naturally the way they usually do. And then in the off-lines, in the off-trot the class setting, when the teachers are not there, we align to the pedagogy of the curriculum that they're using. And so you're working with completely different pedagogies. Yeah, the pedagogies come from either the teacher or from the curriculum. And the personalities, which is something interesting to happens in the first session off of a tutor. So when the students first come in, they get an ever-tutor class. And in that class, they talk with the AI about how they like to be challenged. Do they like to play sports? How do they get over? Do they want like a cheerful warm tutor? Do they want a little bit coach like tutor? And those things carry forward for the rest of their time. I've seen that a lot in the AI tutors that people can choose the personality that they want. John, you're at Wolfram, which is known for computational thinking and is very well respected in math. And I've fascinated that other tutoring companies have gone big and at scale in the marketplace. And your company that knows math so well is taking much longer. What is it that you are trying to solve for that's not so easy? So partially, when we initially started, we started again with this Lemmy program. And that kind of had its own track. And in general, our goal was just essentially to take our time. Because it's one of those things where it's almost hurry up and wait. And a lot of people are able to come out really quickly and do lots of things. But when you kind of peel them back, they kind of end up being very thin wrappers around the different LLMs that are out there. And so come when we look at it, we look at our experience with Wolfram Alpha. Hey, it turns out that we already have something that's really exciting and solves a lot of problems for people. If you're thinking about math or STEM subjects from a question-answer perspective, and when it comes to our tutor, we're actually trying to go after kind of a larger problem in a way because we're trying to emulate more of human instruction, more of asking questions, more of being able to lead instruction as opposed to having just kind of relying on the material that's kind of in the LLM or using techniques like fine tuning or rags or other stuff. But thinking more from, how do we create something where you have more of the material that's actually used in the classroom and how a teacher actually teaches that material. And that took a lot of time. And we kind of had some missteps on how we went about doing it. As I said earlier, we kind of initially thought about doing it with our own curriculum. That that would be the way that we'd know exactly how a teacher would approach a curriculum. But again, it turns out when you start talking to teachers and administrators and other folks, to realize not only does that not scale, but it's not something that they actually want because they've made a lot of those. decisions. So that's really kind of what's taken like a longer time is we came in with some assumptions and some of them turned out to be really faulty and then we kind of went back and we tried new approaches and we learned a lot from people and learned more about what they want. Again, people have made a lot of their own decisions, they've invested a lot in what they're going to do and it's more of supporting that as opposed to coming in with something different. I'm curious to hear from the other panelists how much are your tutors directed by the questions students come up with and what they begin asking and how much do your tutors steer the students what to do next. Caleb. I'm laughing because I've spent three years working on exactly that which is which is at the beginning when you are more of just like a thin wrapper right. A student will ask a question or a curious student will say the famous ignore all previous instructions right and they'll get the AI to just go off on the rails right. So one of the things that I'm sure we've all had to do is how do you stop that and it isn't just including the instructions to the AI to not go off topic right. You've got to do things along the way the student sends a message every time a student sends a message on school AI we have six different agents that are looking at it and planning a response together and then responding. And so yeah you have to engage very deeply beyond the thin wrapper. Right how about you more in Google just Google steer the practice problem for the student because they you know that that's the right problem for them to try next or is it all led by student inquiry. I mean it depends with you know sometimes it might come because they're stuck on a homework question right and so there we try to help them build an understanding you know providing them with explanation on how to solve them themselves and then and then maybe we might get into like practicing similar problems so they can really build that understanding. In some other case I mean we launched this feature where students can see help me prepare for my test whether this is a standardized test like the ACT or whether this is hey I'm uploading my study guides and I want you to help me you know create some some practice problems and so in this case we are really going to grant it in the contents that they are trying to prepare for you know we know that students wants to be quiz on the same questions that will get in the classroom and so in that case we really try to make sure this is you know following the same it's really grounded on the content that the provider so understand a nice exam curriculum and then we try to learn you know what's this trends what's the mistake and so we also provide them with insights on what they want to practice small and what they're really good at and then you know really make it adaptive so that it progressed with them and they can see that progress and not just within a session but also doing it across sessions so they can remember. It remembers. Yes and you know and and of course you know bringing the right remediation and providing the right tools to help them fill those knowledge gaps. Right Karthag I often hear researchers saying that the problem with having a question driven tutoring from the students is the student often doesn't know what they don't know yet and a good tutor like helps direct them through the topic that they need to learn how do you think about that with ever tutor. So I think you said at the start that we did some test prep stuff before. Say that again. That's preparation products before ever tutor. You're still going very well how familiar students use it. Those were all the previous generation of AI which is question driven. I think that's almost such a stark difference that should be counted as whole another generation. No more questions. Yeah and it should be not about questions but it should be about building proficiency in the skills. Whether it's built through questions whether it's built through direct instruction whether it's built to peer-to-peer learning or collaborative work it should be whether you actually get the necessary skills required for that standard and it turns out there is a large quantity of research that figures out what's the best way to build a skill for different kind of learners. There are some kind of learners which for example they have a long memory that they can keep in work like a working memory while they're talking and for those folks auditory conversations work fine but there are folks which actually don't have that TikTok brain and they would rather do well with the social peer group and for them they should have a method of working with their friends with an AI and these people are not segregated in different classrooms they all sit next to each other. So whether the question approach works or not is probably going to work for 10% of the students but it's just going to be 10-20% of the students. Right so the most important question of course is effectiveness. How are you measuring whether your AI tutor is helping students learn better than they did before and how let's start with Google and Maureen how are you thinking about measuring effectiveness? I mean this is of course the Holy Grail question right everybody is trying to measure efficacy and so we are trying to look at different skill and different time frame so so first of course we are going to run those longer studies like RCT and you know really trying to partner with a lot of you know school and experts outside to really see what's really the learning outcomes and then we evaluate a lot of dimension you know you don't have the RCTs yet. Now we have a few so for example we did this experiment with ED on map tutoring and so in that one we actually were able to to show that thanks to using LearnLM you know and our guided learning experience students were able to apply what they had learned to novel problems so those skills were more durable and interestingly also is that we had teacher in the in the loop right that we're reviewing the answer and actually a bunch of them told us that they learn by looking at the Socratic tutor they learn some new ways to engage the students and to answer the questions so so that's the type of study we are trying to run. Now those sticks longer and we want quicker feedback loop right so we also have pedagogy emails where we run with experts and you know we evaluate whenever we launch a new product we really try to make sure it happens through the full product development process and so there we might ask or read us whether this those responsible and courage curiosity will help with metacognition and so those are expert emails and then we also build auto emails which is a really quick feedback loop which again tries to evaluate those same dimension but do it by stimulating students or do it by emulated students yes and and so it but it's like but of course you try to to make sure they correlate with those longer studies but you try to have those different you know evaluation process that can give you quick feedback so you can quickly iterate you know more expert feedback but then also the four RCT that will really show you the end outcome right and John how are you thinking about as you're in this pilot state measuring effectiveness so right now it's mainly just doing a lot of co-design actually working with teachers and administrators and thinking more from that perspective we hope to do like more sophisticated trials and actually built into the system where we have both kind of styles of testing both formative end and somewhat of assessing and also are able to give both teachers and parents inside into how the students are using things on a daily basis and their performance within the tutor environment so that then teachers can use that material to actually go back and plan for things in the classroom the hope is to go and do longer kind of gold study kind of testing but it's you kind of need to start and have like an overall kind of idea for the product and actually work and interact with those folks who are actually going to be using it on a daily basis actually went to a session yesterday from the Hala foundation with I think it was Isabelle how was from Stanford right and there's really yeah exactly it was really great because it kind of tests a lot of assumptions like we kind of want from the technology side for AI tutors to have this amazing kind of impact and she just kept emphasizing well you know what it's actually all these things are actually about relationships and having whether it's kids or adults actually being able to build relationships with their teachers with their parents with their peers that are going to help them learn things like far beyond any of the any of the AI related tutors at least at this point so it's kind of like we've kind of stumbled into some of those things in a way and thinking about like how we could do that from like a co-design perspective but I think that's like the thing that I've already learned at this conference to like go back and be like oh wow this is like putting the work that we're doing kind of in the proper perspective of how that is going to work in the students life the teacher and parents to hopefully foster that connection and a team that's actually working with the student so how can we work to make that better one thing I think about a lot is how there have been several studies showing how the AI tutor can hold a child's hand too much and even if it's programmed not to give away the answers and be so critic the student can keep saying I don't know I don't know I don't know and you know after a few tries get the answer given to them I mean I've seen it over and over again. And so my question for the panelists is, how do you think about this problem that the students can be spoon-fed by sort of gaming the interaction with the AI tutor? Well, Caleb, have you thought about that? - I think one of the hardest things for AI in classrooms over the last three years has been that the primary use case of AI is being an assistant, right? ChatGBT does things for you, cloud does things for you, and we're all up here trying to make AI not do it for you, right? And finding the balance of when to help and how much to scaffold is the work when you are tying to a human tutor, someone that is sitting next to a student and working them through, human tutors get this wrong too. - Absolutely. - They'll give the student too much. And so, zone of proximal development, we'll talk about some variation of that. It's like you've got to keep the student, I will not say we get this right every time. What I would say is one of, I mentioned those six agents that are kind of navigating and negotiating the response to the students. One of them is that like, let's get the answer in the zone of proximal development. Let's make sure that they are at the right desire to difficulty level. But I would say it is hard because a student sometimes, in the same way that is like, Google's dealing with jail breaking LLM anyway, right? So it's open AI and anthropic and all of us are. So it's an ongoing cat and mouse. - Karthik, how are you thinking about this spoon feeding and hand holding problem? - So, as you said, even the best models today would give away the answer if somebody's very persistent amongst, I don't know. But I think the issue is a little more nuanced actually. I think, for example, there is also to think about why the student is doing so and what's going on in the environment that's promoting that behavior. Which is I think the real problem for most cases. It's not that the AI tutor is giving the answer. It's that the AI tutor hasn't prepared the student and the environment has forced the tutor on the student to the point that it has to come down to, I don't know, I don't know, I don't know, right? And so when we see those things, which inevitably happen, they happen in the actual classrooms, we try and intervene with a better implementation of the system. - A better implementation. - Yeah, which is, maybe these kids are not on boarded properly, maybe the teachers are not trained properly, like the implementation of the system in the classroom is usually at fault when those things happen too much. Also, there is frustration on regards to the student. If they're genuinely, they don't know, then they don't know. (laughing) - But they need to be shown sometimes step by step, right? - They need to be instead of breaking it down and constantly asking problems, sometimes they need a pass. They just need to be like, here's how it's done. We will win this battle another day. Let's go to the next problem. - Right. - Sorry, day and a wrap. - It's a wrap. - Things you said is like, sometimes there is a deeper thing, right, sometimes they don't know. And one of the things we've found very often in our research and working with schools is, there's a classic Maslow's hierarchy thing here, right? We're talking about AI tutors and we're all talking about academic outcomes, but I would say probably for all four of us and anyone else building in this space, the more powerful things are the qualitative things. Students are desperate for someone to hear them. And this is complicated, right? Like I'm up here bringing up a very complicated issue. It's like students will share, how many of us in here have shared something with Chatchee B2U or Claude that we would not first go and ask someone else, right? Me, right? And students, they're afraid of getting made fun of. They're afraid of what an adult is gonna say to them if they ask a complex question. So I'm gonna say something that's like, I know we're gonna get into kind of safety stuff, but we sent out more than 10,000 critical safety and wellness alerts last month. These are 10,000. These are kids sharing serious mental health issues, serious safety, like bullying and things like that. We've been the first people to know that kids are planning on bringing weapons to school. That's not what I designed this platform to be, but because it is an earnest listener and engaged with the student, like those are the types of things that we see in here. And it's too, what Karthik was saying was they're like, how do you get the feedback loop to the right person? - Right, sorry, and I wanted to add something, which is, I mean, really to bring back what Karthik was saying. You know, we also want to understand why this does the student not engaging, right? And I want to bring the book from Rebecca Rindrop about disengaged teen. And so I think it's also about making sure that should the AI, should make sure there's the right productive struggle happening as part of the conversation, but how do you really make the learning engaging? How do you make students more curious? And I think that's really again an opportunity, with the tutor, to bring to make those disengaged teen much more engaged because now they understand why they are learning. Maybe you can connect that physical concept that really is struggling with to some real world problems that they really care about. You know, whether this is basketball, climate change, and so I think that's a really true power of AI, is that a lot of the time the students are stuck or not even trying, because that just disengaged, right? They just don't understand why they need to go to school. And so if you can't really make learning much more engaging and spark that curiosity and making much more tangible and relatable, I think that's how we can solve a lot of that. - That's one of my concerns with the research that I've seen, that where we've seen the AI tutors achieve pretty good learning gains, it's tended to be with very high achieving, highly motivated students. There's a famous AI tutor in the Harvard Physics Department. There's been some ones of high school students in Taiwan. And I just feel like, well, that's not the American context. But also I question the idea that the AI tutors going to suddenly motivate the student. - Yeah, I mean, I need to jump in in a way. I think we're putting a lot onto the potential of AI tutors that may not be there and then actually go back to, I think how is about how is talking about things, because there was a really good model there. Again, it's really about the human relationships and that these AI tools are going to be incredible to make parents potentially even better teachers. Because a lot of times, at least, so I'm going to talk about our product lead, Theodore Gray, who did, I don't know if there are folks out here, did the elements book and app on Apple and has done a lot of educational related stuff. And he actually invented the notebook. So when we're talking about notebook LAM, he actually invented that. So, invented that concept. As he would say, it's like since been lovingly copied by Jupiter notebooks. But let me get back, right? So the critical thing are how can you have that experience where parents can perform better, having an AI tutor just like a book that could go through and tell them everything they need to know about a subject and sit there with the child. So when you talk about that motivation question, maybe that's not for actually an AI tutor platform to solve. It's more of how do we actually enable the teacher to be able to handle a bunch of different students and engage with them all at the same time. Or a parent, even better, because I think there was another person in the same thing that was like, I think it was Springboard. I've forgotten the guy's name, but prepare parents to then be able to be better teachers, to do high impact tutoring, which is what all of these AI's are really trying to be able to do. High impact tutoring. If you make parents be able to sit there with the child, they don't necessarily need to know the subject at the same level as a teacher in the classroom and can do a lot of work. Let's give everyone a final word. I'm curious one thing you would like to improve about your AI tutor in the coming year. I'll start with you, Karthik. I'd like to make our tutor much more emotionally intelligent to the students' struggles. It almost comes off a little tried when it tries to be emotional. And so it's like the uncanny valley and we try not to go there. It also relates to motivation. I mean, maybe going back to what John is doing, I think we really need to make sure the AI is enhancement to the teacher. And really, that human connection is so important in the classroom. And so making sure that the AI is really working in partnership with the teacher so that the teacher can be the architect of transformation. I think it's really what we-- it's top priority, I think, for all of us. On the whiteboard above my garage, when I founded school AI, I wrote, what do I want AI to do in classrooms and to steal from them, but also for myself, three and a half years ago, it was like magnify connection. And it's not because AI is going to do that. It's because AI can be almost a new sensor or place of engagement where you can find out how students are doing and get that to the right person. John, how do you want to be? >> Very quick. Figure out how to lower the price drastically to make it available to as many people as possible. >> That's a good one. I want to thank all of you for coming late on a second day. I'm sure everyone's exhausted from conference going, but thank you for coming here and learning a little bit about what's at the cutting edge of AI tutors. [Music]