AI as Air: When Learning Becomes Ambient

[MUSIC] This session was recorded live at the 2026 ASU GSV Summit in San Diego. [APPLAUSE] >> Hello everybody, welcome. This is the panel on AIS Air. When learning becomes ambient, we're super excited to have this amazing group and panelists here today. We're going to just go ahead and start jumping right into the topic. And it's a one that's been very top of mind, I'm sure all of you have heard ad nauseam about AI and what AI is doing and how it's interacting and learning. But I think what we're hoping to talk about today is kind of the future. When we have what is called ambient AI or basically an ability to collect information and have all sorts of different forms of data points and structured data and structured data and being able to start interpreting more of the world around us. And I'm going to kind of just try to level set at the beginning with what as you guys experts in the field are, believe ambient AI is. And if you see examples in the real world around you today that people can kind of use and understand to get a raft. So I'm going to start at the end here with you, Chris. And go ahead. >> So the question is what is ambient AI? >> I mean is it here with us now? I think that's what a lot of people are thinking right especially when they think about AI. And for me, thinking about it from the simplest way of thinking about it, it's, it's, think about Siri. But think about Siri as a very early ambient AI. And the reason why is while it's there, it doesn't understand the context except maybe location. So I think of ambient AI as something that comes to you instead of you coming to it and having that context of what that is. Now an example of that for me would be a simple example of that for me would be something like fraud detection, something that is happening with your financials. And it comes to you and says, hey guess what, we think that this is probably an issue. Would you like to take a look at it? That's a very simple and basic example. >> Yeah, so at Google we've had a lot of different product surfaces. And so we probably become very biased initially with kind of where we think a lot of these AI systems sort of live. I just took a vacation last week and one area I really hadn't thought much about was maps. I mean there's obviously a lot of AI built into maps. But I think to Chris' point, like having the insight that weather would impact travel and being not just sort of made aware that that would impact travel. But the kind of synthesis or analysis that's done by the AI is becoming more, I guess helpful in many respects. But I was one of the very early kind of home automation adopters too. So I very much bought into the kind of Siri and home assistant type stuff early on. And I think that we're transitioning from, again, more a world where ambient AI is like, it's aware again that it might rain or that the weather might change. And then it's able to even prompt things based on other kind of, maybe the number of people in your house or number of windows and be much more informed in context rich. >> From my perspective, I'm thinking about AI as ambient within education. What does that look like or what is it going to look like? And I think of it as being embedded in all of the processes from being in front of staff, faculty, students, and admin. And fixing all of those gaps where we've never been able to do it before. We have this 400-year-old classroom model that we're still using. Our structure of our education system is very similar to how it's always been. But with AI, we can meet that system where it is and plug in AI to fill in all of these gaps. And I think we'll know that it's ambient when we don't notice that it's AI anymore. You're just doing the thing faster and better than you could have ever done it before. And products and vendors like NECTER are not advertising their tool as AI. They're just advertising as a solution. What are they going to come fix for you on your campus? I think that's what we'll know that we have reached the point of ambient AI in education. Before I make a comment, is that night? Can I ask a show of a hand? How many of you are current classroom teachers? Were we classroom teachers? Oh, terrific. Thank you. This is a very helpful. In my mind, ambient AI means we change the way how we interact with AI in the education system. Nowadays, for many educators, you probably know your workflow maybe is you have a need, you open up a tool, you tap in your prompt, your general response, and then you're the one who moved data from one system to another and trying to complete your workflow. So for ambient AI particular for college AI, we're in the key 12 space. What we're trying to accomplish will be the AI is learning about you, about your students, about the environment, about the content, and the AI will do data analysis. You don't have to be the one who breaks the gaps manually. And you are the one AI also give you recommendation about your action, give you insights. Then you make the ultimate decision of what to do in terms of interacting with your students. So that's where to sum it up. I feel like ambient AI from our product perspective is it should, it should be embedded into your workflow, right? So what you mentioned, like it doesn't need to be like you are where the AI exists the concentrate, right? Embedded in your workflow at the same time. It's no longer a bit collecting data, it's actually giving you intelligence. Awesome. So let's bring this a little bit further deeper into the education space and what this actually means practically. Where do you guys see the opportunities within education? What workflows, what work that needs to be done, are you saying that schools start thinking about from an implementation standpoint, practically, right, as well as areas that you think might make less sense, given kind of existing needs or processes? We want to tackle that first. Go ahead. I can jump in. I serve a continuum with the conversation for us. The workflow that been taking on very quickly by educators by school is from a formative assessment to differentiate material to help teacher iterate instructional iteration. So what we see is now we can connect, understand, school curriculum and for a given lesson, we know teachers learning objective, we can generate exit ticket, a formative assessment. And then educators now can distribute those sometimes even gamified assessment that along with their curriculum to students, once the students turn in their digital handwritten work or follows, then AI will be able to analyze and be able to capture those 10 minutes conversation in the classroom and those very close to students thinking, reasoning, learning process and then turn this into teacher's actionable insights. We can rep command to teachers next day. You may want to modify your warm up activity because today this is your student's strengths area for improvement. Here are the topics you may want to review. Here are group of students. You can give them more advanced learning material can stretch their thinking. That from the curriculum exit tickets assessment to action for the next day for differentiation. That's where we see like a quick late taking off. And that would require a lot of interoperability connections. And one of you guys want to maybe compete at you can start with like, what does that take for a school to even start thinking about tackling this problem. Yeah, that's a great question. I think before you can even think about does it integrate with the LMS or the SIS. The problem that we're seeing a lot of schools have is that there are just no AI governance right now. There was a UNESCO study that came out this year that showed that one out of every three institutions globally has no AI governance. There's no guidelines of how AI should or shouldn't be used on a campus. And we have so many schools that will come to us and say, I'm ready. I want AI infrastructure. How fast can you scale it across the campus? And that's not the hard part. They're thinking that the tough part is going to be, oh, how can I scale it across the campus? We can do that in under two weeks. We've done it many times before. The hard part is is there governance across the campus that everyone is aware of students and faculty and staff of how you should be using AI, where the guardrails should be, what is acceptable and unacceptable uses of AI. And if you look at the Gallup and Luminous study that just came out a couple weeks ago, 53% of students said that their institution either discourages or outright bans the use of AI. And then in that same study, 52% of students said that they're using AI for their coursework on a weekly basis. 21% said they're using it daily for their school. So there's this giant gap between where students are actually at with AI right now and how they're using it and what schools think is happening and there stands on AI. And that, I think, is the biggest problem we need to address first is that gap widens on a daily basis. Not only do students need that AI literacy to be prepared for the workforce, but also if you have no insight into how they're using AI, that's when we start falling in the trap of losing your critical thinking skills, getting used to being overly reliant. on it just feeding you an answer and not double checking that work. We have to get schools to a place where they know how they want AI to be used and find a vendor that they trust that they can work with that has peer-reviewed research that shows that it's actually leading to the outcomes that they're claiming and that's when we'll be able to even the playing field and get the right kind of safe AI into every school. But we're still at the point of we need to get governance across the system first. Chris, you want to expand on that given Salesforce's connectivity and the ecosystem? Well, and I think it's so I thank you for that and I think it's so much more than connectivity and integration because if I think about it again I always try to go back to the simplest view of this. I think about what are the moments that matter and for all of you that rose your hand that said that you were a teacher or someone in education that was serving students for the same group and also for the rest of the people in this room. How many of you are in some type of AI transformation right now at your institution or at your school? And when you think about the moments that matter a lot of times those are the moments that matter to the administrators. When we think about the moments that matter to the students or we think about the moments that matter for all those wonderful teachers who have bigger hearts on the rest of us because they get to touch students every single day what are the moments that matter for them? So just as we think about an integration between two different pieces of software what about the integration between the moments that matter and your moments that matter might be different than the other teachers moments that matter which may be different than the students moments that matter. An example of that would be what about a student that is just about to seek help and then they back down and then they're just about to seek help and then they back down and then a third and a fourth time they never ask for that help. When you think about the way that AI can be used to identify when someone might need to be help and maybe they've taken 70% of the steps to ask for help but then they just can't ask for that piece of help for some reason maybe it's a confidence issue or they're scared or they're worried that someone's going to think something different about them that is a moment that's going to matter for that student for the rest of their lives. So I think about it much more than just integration between products or pieces of software. I think about it as integration between these different moments because all those teachers that rose their hands that is a moment where you can step up and help that student. That's a moment where you can do that so it kind of brings me to the topic of thinking about how do we understand each person as a person and how do we give them agency over how they're understood as a person and for me you know we talked about this maybe a couple weeks ago when we were on the phone and we were thinking about how do we understand someone. Now in the technical world we call that a model. You all think of this and many others do too even those of us on the stage. You all know what a large language model is where you think about Chachi BT or Cloud or something else. That's where the world's information is all into a model and you can ask it questions and you can understand it and do these things and have it right for you and ask questions. Well why can't we have a small model for those people that need those moments that matter so that we understand that person or we understand that teacher or we understand that administrator who's making you go through an AI transformation process. How do we link these moments that matter to these individuals and for us we do that through what's called a small model. I'm really curious how that works at scale because I'm coming at this from the angle of you know I've been in the space for eight years selling and implementing these tools within institutions for anyone that's on a campus you know how long it takes to get a tool that you love across the entire campus if that ever happens. So even if everyone is on board with it and it sounds like a great idea. How does that actually work like give me what that would look like in real life in a classroom or across a campus where every student has a personal model that knows you know how they learn and where their knowledge gaps are and can sort of follow them as their closest advisor and learning partner. I love the idea of it. I'm curious how that works in execution. Well maybe I can answer and then I'd love to hear some others as well but when you think about it and when you think about these consumer products today so I mentioned chat GBT and I talk about and I talked about Claude and you know we have Gemini and we have a whole bunch of other things. They're not just looking at that large language model but they're also adding to the memory of how they understand the person who's using the large language model. So how many of you have used chat GBT or Gemini or Claude or something like that. You all have small models with your large language model. Bring that to the student, understand that with the teacher, bring that to the administrator, bring that to the institution and then doing so in a way where the person who's adding to this model has agency over that information so that it's not being shared with the large language model. It's not being shared with technology vendors like all of us up on the stage and how it's protected but I'd love to hear from my peers as well about small models. So we've actually been exploring the small model kind of concept for a long time. Anyone at this point in time especially those with pixel phones can actually go use something called Gemma which is one of our genuinely smaller models but also a great model for lots of different use cases. There's that domain which is in our world we think of it as kind of distilling or taking from a very large model all the really important parts that that matters so it might be pedagogy, it might be helping students understand math both visually and through text. So this kind of concept is not necessarily new but it's something I think we're on the hyperscaler side figuring out more and more to make not just efficient but more approachable for educators and for systems that are looking to kind of roll them out. But I think what you pair with that to Chris' point is all of these kind of memories and memories the term that the industry is using right now and think of memory is just the conversations that many of us are having with models and what's interesting is that more and more we start identifying by pinning or by doing other things the conversations that really help us understand something or help us navigate a problem and then through that you get these things called artifacts and so what you start seeing is this kind of ecosystem and sometimes you'll hear words like harness but these things kind of attaching to the underlying models whether they're large or small and then they create these kind of experiences whether they're learning experiences or they're just helpful for administrators and those components really truly live with the individuals in the same way the you know your documents or spreadsheets or whatever it might be you know can live in your own kind of files or folder systems. So we're certainly seeing that I think longer term you know we've done a lot of research work just around personalizing the models themselves and what that looks like is taking a subset of those memories or those other artifacts that exist around the model and start embedding them directly into the model and there's a lot of benefits to that including how it sort of initially responds to you versus having to prompt a model over and over again but that's what we're continuing to kind of innovate there and do a lot of research but it's really exciting because it also allows you to start using the models offline and yeah. Yeah I was just trying to connect the dots and go back to your original question about interoperator ability right as well as what you were mentioning about adoption system throughout there are different layers what we can we can share more about the technical layers right how do we build product how do we build models that is one layer and there's layer of our self the second layer is how do we build into the technological ecosystem that's the second layer of interoperator ability the last layer of what is what you were mentioning once we roll the product out in the actual system either a campus or a school district system there are so many other human factors structural system factors come in right how do we make this product how do we make people trust about these and make people feel like this is not too much invasive to their personal privacy at the same time they see the value and so we actually can roll it out right so there are three different layers yeah and I really like that framing I think it was very helpful man as you think about you know what are the memories is the technology being able to actually understand the content memories within context of the interoperability of the connections the systems or kind of the barrel over ecosystem where do you guys see the greatest barriers or challenges amongst those like is it any one of those three layers is yeah where do you see kind of either schools or people who are trying to implement stuff you know kind of running to headwinds for us the first barrier is to be able to link these three big components together for us we need to link three big component one is about the content in terms of what the mastery of a concept or scale like what is the prerequisite knowledge about multiplication what is like those different components for kids to learn about multiplication what is the next step the knowledge relationship concept relationship so this is all about the content the second is we need to learn about learners what is their baseline and do students have i.e.p. there are they multi-lingual learners there are disability status and then more important this is the part we add on is we collect this task-based students learning activities so we can unpack learn about their reasoning process, their verbal expression, and what is their misconception about the multiplication, for example, right? So that we can offer the next step, which is action layer, like we can give teacher recommendation action for that layer, we make sure to give a teacher editing right. So they modify the edit, they accept the reject, so we store those data on our platform, we learn from those as well. To link these three things together, we have to connect with the students information system, we have to connect with the learning management system, where we're actively linking with Google classroom, and to be able to get the content capture the conversation, everything like that. So that is pertain to the second layer, the second in term of link to the ecosystem, we're using one roster to do the linkage, and we're trying to connect with other system API, that has been a lot of work negotiation of different parties to figure out how to link our system together. And the last layer, which is I think going to be later on, the most difficult layer is built human by end, and teachers not offloading their decision making to AI, once they have the options, and once they have those actions, how do they decide, what is the most effective way of interacting with the students, so they're not losing their knowledge about their learners, their students, the implementation pieces where we're doing a lot of experiment of how to work with the school system. I think that was a beautiful way of talking about the challenges, and I agree with every single one of them, and I think for the technical challenges, the technical challenges will be figured out by technical people. I think that there's two huge limitations you talked about, one of them about the human by end, but here in the United States and also globally, we have regulatory issues as well. I think regulatory is a major issue by itself. When we think about the technical pieces, I do think that technical people will figure out the technical items, and will deliver to you a beautiful product. But I think that the by end, which you mentioned, and the regulatory frameworks, which will need to be adjusted and modified and accepted in order to do that. Here's an example. How do we think about FERPA when it comes to being able to think about a learner record with a student and the institution is the conversation with the ambient AI of the future? Is that something that is governed by FERPA? Is that something that's governed by something else here in the United States? I think these are things that will have to be challenged at the regulatory level as well, which will also require human by end, just like you mentioned. The last thing I was going to say is this challenge has been addressed in other industries. When we think about other highly regulated industries just here in the United States, one of those is healthcare. Our doctors, I had a doctor's appointment a couple of weeks ago. He had an ambient AI tool that he put on here, and I'm like, what is that? What is that? He told me, this is just listening to our conversation, is it stored? No, it's not stored. He shared that as we're talking, it can access my patient file and understand, well, Chris is blood pressure has been high in the past. Why? I think it's a really interesting way of thinking about how this could be used in the future, especially when we think about those small models. I didn't mean to interrupt the two of you. I was just going to add, I think that's 100% right. We're going to learn a lot from other industries. We're talking a lot about outcomes, but we're not always describing what the outcome is. It's over-specs the outcome is helping a learner achieve a goal, and it might be graduation, it might be, you know, obtaining new skills, working on soft skills. Because so much of the current set of systems with technical systems, regulatory systems are based on a lot of variables, much of which we've studied, which is all great. But they'll ultimately have to change, I think, to accommodate a lot of the new skills that will sort of come online and become really important. I don't know that we necessarily know what all of those will look like, but I think if you focus on the outcome, I think that it starts to kind of upstream or downstream really change the way we kind of think about interoperability as a whole. I think to add in just a little bit of boots on the ground data of what's happening right now and where current systems need to be integrated for AI to actually work today in the classroom. Nectors currently leading the largest AI deployment in the nation with the California Community College System, 2.1 million students, 116 community colleges across California. And so we have a lot of data from the last two years of doing this with them, of what's actually working at scale and what's not, and what leads to the success that we're looking for. The number one outcome of whether AI tool is successful in the classroom or campus right now from our data is whether it's integrated properly in the LMS or not. The LMS is the one evergreen tool in every K12 and higher ed campus. It is whether they like it or not where students have to go every single day and so do teachers. And so if you are trying to put any new tool no matter how great it is in front of the students and you want them to actually use it, it has to be integrated inside of the LMS. It's not just integrated to where, okay, it pulls in all of the course content, which is a key part of this. The only way that you're going to get students to use your AI versus chat GPT or CLAWD is if it's actually grounded in their course material so that it is more specialized than any generic AI tool they're going to go to. That's important. But what's even more than that, it's little things like eye framing it. Like the actual tool should be able to be used inside of the LMS. Because if you're going to ask a student to add another tool to their daily stack or their workflow, it's just not going to happen. Again, it doesn't matter how great that tool is. And so when we looked at the data from this implementation of the California Community Colleges, we saw that being the number one distinction between is it actually being used in the classroom? Forget success. Is it even being used or not? Yes, if it's integrated in the LMS, no, if it's not. Let me. I want to challenge that a little bit and see if there's any other opinions on stage. I mean, the LMS is probably one of those products with the lowest MPS score that exists. I don't know anybody who doesn't hate their LMS. Like let's say you just put it in the positive terms. Apologies on stage for anybody who's currently working on it. It's just being recorded. Yeah, exactly. But I think there's a question of how does that core fundamental management system exist or change? Does it need to be redefined? Does it need to be rebuilt, given kind of the new capabilities there? And I'm curious, let's start with you, Charles, as someone who's also been building a deep classroom integrated tool, your opinion on the future of the LMS. Well, I mean, first off, I think literally every system that has either a smaller substantial number of processes baked into it are prime for disruption because the processes are both deterministic in many respects. We know that there's a start and end point, but also they provide guard rails. I think we're finding more and more, and you could say, "Agentic frameworks," you could say, "things like the content that the agents are working with is more dynamic and so traditional LMS or content management system isn't the right fit." I think all of those arguments are true, but again, I think go back to the outcomes comment. If the outcome is start anywhere and end up kind of learning a new concept or really refining certain skills, then it doesn't necessarily fit the same box every time. Really what you're talking about is a system of sorts, I guess, that sort of evolves with the learner, but it retains some aspect of the data in terms of interactions. I don't know that it's necessarily an LMS problem. It's a mini-systems problem, but I think on our side what we see a lot of requests for, and this is both in the Gemini app and elsewhere, it's just on the fly, generate something, and a completely different modality based on some text content. I just don't know that an LMS is always the best place for that. Can I build upon what you mentioned about the outcome oriented and then we go backward mapping, right, in terms of what technology we need to offer to schools, to learners, as well as how do we capture learners learning progression, and how do we present the return investment to both learners and school district, whoever purchased the technology, right? So from that standpoint, like a learning management system, it's really good right now because it makes the communication between teacher, learner, parent, transparent. They can all know what is the content, what is assessment, how learner learn in terms of a great book, right? So that's where we've been asked by school district. You have to think back to the learning management system so that we can let parents know how students learn by view the great book. So that is one aspect, but you can think of it in the future. The outcome no longer be the great students learning in a traditional way. The outcome is to capture students thinking process, strategic thinking, problem solving schools, then the interface is going to be different. The task they need to perform on the learning management system will be different as well. And that requires a redesign in my mind. On the other hand, what we think is going to be sustained and cannot be immediately replaced is the student's information system. Because that is authoritative record of student learning. AI is a probability based on the model. Probability made the model mean that's not 100% sure. It cannot be the authoritative record of students learning. So that's where we feel like the learning management system may be able to be distributed soon. But then the student's information system, that's where we still need the traditional software, like what those traditional, like more of record of students learning to be correct. Yeah, and I have to agree completely with what men said. You know, when I have a 20 year old who's in high education and I have a daughter of 16 year old who's in high school and for those of you who have parents who are in one school system or the other, how many of them just by a show of hands, how many of you have to have your children log into more than one system every single day? How many of you have to have your students log into more than five systems a day? How many of you, and we can keep going and play this game for a while? But I think my son, I counted it one time, he has to log into 17 different systems for his campus. And I think two of those like we created sales parts, but at least two of them maybe more actually, but I mean, quite a few for you by the way. But I think it's interesting because when we think about the LMS, like what is the LMS of the future? Now obviously there has to be some type of learning management system. But when you think about that single point of entry, and I'm not talking about something like Octa or I'm not talking about logging in one place and having access to 20 different places, but the interface itself may be the same. When you think about that interface, how does that interface interact with the person and then going back to the small model conversation we were having earlier? Is how does that interface adapt to the person who's using it? So if we were to think 10 or 20 years down the line, it could be a very interesting dynamic about how people access their own information and access learning. Going beyond how do people access their own information, maybe we can also discuss who should own that information and what degree of portability should exist. As we think about the intriguing aspect of ambient AI is the ability to correct just infinitely more information than most who are aware of comfortable with how do we think about the ownership of it and the portability. We're very particular about who can see the interactions are happening inside of Nectar AI. So it's meant to be this AI infrastructure that is based rooted in students having a 24/7 support system both in and out of the classroom. So you can imagine they're asking questions from everything from financial aid to I just lost my scholarship, how am I going to pay for tuition in five days all the way to I have a midterm tomorrow morning and I have not gone to a single lecture this term. I don't even know where to begin studying. So there's the intent for us is to give them a place where they can go and ask all of the questions that they would normally be way too scared to raise their hands and ask or even go into office hours and ask. And so when you're creating a system like this, inherently it has to be private to the student. I think that is one of the most key parts of it is the student has to feel like there's no one looking over their shoulder at the kinds of questions that they are asking. There should always be someone who is looking at the kinds of answers that they're getting and not a whole separate conversation of guardrails around AI. But when it comes to the student's data and especially the memory and personalization that we were talking about over time, AI starts to collect a lot of information about a student's strengths, weaknesses, areas of opportunity. That's the kind of data that we believe the student should be able to see, not necessarily anyone else on campus. There's a lot of data that everyone thinks, oh well if I could see that, it could help me prevent that student from dropping out. Yes, absolutely. There should be early alert systems that are embedded within this that catch students that might be at risk. But I think we have to give the agency to the students first. And that means giving them the tools to be able to get their questions answered, figure out financial aid, paperwork, study and pass their classes on time. If we give them those systems to be able to answer their questions and help themselves, they will. And what that means in turn is that we trust them with their own data. And we're not ever going to surface those conversations in entirety to the faculty staff or admin of that campus. And that's a really strong position that we take. I mean, I guess one thing that we have conversation earlier in the room next door about knowledge graphs. I think we are still early days of understanding what that data looks like. You know, when I think about like, again, navigating with a map, like the surface is the map. I'm used to seeing my little car on a map going down the highway. And I see it that way because it's helpful. The context is incredibly helpful. But when we're talking about capturing, whether it's AI observing or teachers or some combination of that, like, is it traditional SIS and LMS systems like we were saying before? Or is it perhaps a new technology stack or a set of like tools? I think that plays certainly from the technical point of view a really important role in how we think about interoperability and how you can take the information. I know any Strava users in the room. Like I remember early days Strava was in still, I guess, to some extent still pretty kind of siloed in terms of how it captures information. But it's not a whole lot different in terms of human performance data. You want to make sure that I can take the run that I did on a certain app and port it over to my health app. So I definitely think it's important and we should prioritize it across all the different tech providers out there. But I don't know that the format has been fully decided. And I think that's going to play a big role. We've introduced something called A2A, which is the agent agent protocol. And I do think there's a lot of work underway right now, whether you have four agents helping you with your kind of educational outcome or two to kind of continue standardizing both on the security aspects of agent protocol or agent conversations. But also how that plays a role here too. Because I think in addition to porting any of your data anywhere and owning it, making sure that the agents can truly work with it at a high fidelity high quality levels. I think the ATA comments actually great. So we're a member of that network as well, which you actually put together. And when you think about being able to share information between different agents and sharing that information so that handoff to the other agent understands that, that actually brings exact question that you just talked about as who owns that data. So let's talk about it from a student perspective. Does a student on the data or the institution on the data or the technology provider on the data or the school district owns the data or the conglomerate or the state of California owns the data? It's a really interesting approach. And that is a regulatory issue. It's not just a technology issue. For us, Chris, what you just mentioned pretty much sum up in terms of our response of who owns the data, schools and district own the data. Because by contract, they have a student's data privacy agreement we have to sign. And in that data agreement is specify school district own their own users data as well as some derivative metadata they own as well. The part way on is the algorithm is the model. In the last few minutes, it's rare to have a panel full of builders, I feel very fortunate. Based off this conversation, what are you going to take back as you think about your next iteration of product? I see our job being to take all of the incredible technology that Silicon Valley is building and make it accessible for schools. Again, for all of you that are on a campus or have sold a product to a campus or tried to integrate a product there, it really, schools usually end up being the last place the technology goes. It's a really archaic system that we still have. But now with AI, we have an impetus to get AI in front of these students and teachers in the right way as quickly as possible because it's most likely those students in college whose jobs are actually going to be replaced first by AI. So we need to give them that AI literacy as soon as possible. So what I'm taking away from this is there are so many cool technologies that are on the horizon like personal models. That's amazing. I would love for students to have access to that. And then again, how do we let those students take that data afterward so that they can use it in the workforce? Have all that data that, you know, the AI knows everything about them that they can now put it into another system and not have to start from scratch. So I think my job is to go home and figure out how do we package this in a way where schools can actually start using it tomorrow, not five years from now. I think for me, and I like that. I think for me it's the opposite. I think that I'm the conduit to the technology companies of your voice. When we have our conversations and meetings and everything else strained this week, it's my job to represent you in what they're building. So that I can adapt just like what you said as a flip side of it is I can take what's being built and then deliver it to you. But for me, I'm the conduit between you and the technology companies. I represent you. That's what I feel like my job is. For me, as some of you mentioned, where are builders? So building technology doesn't intimidate me. We can build. We can get better and better. The concept, I feel like for me, I'm constantly thinking, experimenting in schools as how do we preserve human intelligence, how do we grow human intelligence. We're at the beginning stage of learning, which is key 12 period. And we need to really capture what measures, what learning tasks of what learning activity that preserve grow human intelligence in the future. Our kids can find jobs, they have a thriving life. So that is the second part, as I mentioned previously already, like how do we work with educators, work with teachers? They still know how to interact with their students. If we already give them options, educators need to be able to say, "I still know what Johnny's struggling was, so I'm not just letting the AI grating and doing all the work, I lose track of my students learning. I still need to be on top of it, I still need to know how do I build Johnny so that he wants to come to school, enjoy school, and in the future live a fruitful life." That, the human part is where I put most of my energy and thinking. I mean, I think from my point of view, and this is both conversation on stage and even prior, just doing more to make sure that I think a lot of the use cases are approachable. Policy obviously is really important not just for us, but for our policy makers. And so, you know, part of working with them and working with folks in that area is making sure that the conversations like this get distilled into really, you know, helpful, actionable kind of ideas and use cases. And hitting the others continue to kind of build on things like the A to A protocol where we can, you know, hopefully build a world where people are able to own the data that's helpful in their learning journeys. And so I know many of us at Google are focused on that. Thank you, everyone, for joining us today. Thank you. [APPLAUSE]