The Omniscient Classroom Assistant | feat. Luyen Chou, CEO of Dewey Learn

Speaker A: If I started the conversation with, okay, here's what we're going to do. We're going to put cameras in every classroom, and we're going to feed it to AI. I probably wouldn't even get through that sentence before being thrown out. If I lead with, let me show you the kind of insights that we can gather about your teachers, your learners, then the first question out of those folks is, how do we do that?

Speaker B: Hi, and welcome to it's not the End of the World Everyday Use Cases for AI. My name is Bobby Maclausich, and I'm the head of AI integration here at, uh, Quite Frankly Productions in New York City. I'm the creative director, and I'm also a former teacher. Now, one of my remits in my teaching profession was to, uh, use technology to enhance education, and I was in charge of, uh, bringing other faculty members on board with new technologies. So I'm particularly excited about my guest today because he is literally at the forefront of that very, what would you say, intersection of tech, technology and education. His credentials are wide ranging, having studied at some of the most prestigious universities in America, namely Harvard and Columbia. And so, uh, without further ado, let me introduce you to Lian Cho Lien. Thank you for joining me on the podcast.

Speaker A: Thanks so much for having me.

Speaker B: All right. Was, um, that introduction, ah, accurate? Did I kind of, uh, give you.

Speaker A: I think you nailed it. I think you nailed it.

Speaker B: All right, well, maybe you can kind of, um, give us a bit more of how you would frame your, uh, background and where you reside now in the world.

Speaker A: Yeah, uh, well, so I began my professional career as a teacher. Uh, so we have that in common. But, uh, going back to the early 90s, I started as a teacher in 1989. And it's been almost an accidental journey because I never thought I would be, uh, a teacher. It was kind of a quick stop on the way to other things, but fell in love with teaching and learning and education and classrooms and students. But I was also really, uh, always have been, ah, kind of a frontier technologist. Got, um, very interested in how we could combine, uh, technology and education. And then over the years have. You know, one of the things I've been very mindful of is that, um, education evolves at a very slow rate in comparison to other industries. And there are a lot of reasons for that. But you see incredible pockets of innovation in classrooms, but they really struggle to scale. And so one of the things I got very interested in was the concept of, uh, product development and products, um, because products are the way that we scale innovation in a free market economy. Um, and I thought very early about how could we apply that kind of concept to educational innovation. And so I've really lived my career, my working life at the intersection of education, technology and product mindset. Um, so that's sort of how I would characterize who I am and what I do.

Speaker B: Okay. The intersection of product, education and technology. Okay. So that's obviously quite conceptual. So bring it into kind of, you know, bricks and mortar.

Speaker A: I am the CEO of a company called Dewey Learn. Uh, D E W E Y L E A R N. And uh, Dewey Learn is a, it's a, it's. So Dewey Learn is named after John Dewey. He wrote a seminal book called Democracy, uh, and education, uh, frankly has never been more relevant actually than today given, uh, larger cultural, uh, and societal trends. And if you go to ed school now, the first book you read is Democracy in Education. And so I named the company after Dewey. But it's also pun Dewey Learn. Um, but there's also a more personal connection, which is that, um, in 1923, John Dewey, uh, starting the first education program at University of Chicago, and his wife got fascinated by China and Chinese history and they actually traveled to China in 1923. I mean, unheard of for an American to travel to China to identify promising young women from rural parts of China to study this new field called education at the University of Chicago. And, uh, one of the lucky young women that they identified was my grandmother and she came to the States to study with John Dewey at the University of Chicago. Um, and so John Dewey is literally the reason I am in the United States and I'm doing what I'm doing today. So that's the origin story behind the name Dewey Learn. Um, uh, and I know you weren't asking about that, but I thought important to connect the dots on the name and the personal story.

Speaker B: It's a beautiful human kind of story and you've almost got quite a spiritual connection to education. I mean it's really the bedrock of your personality and who you are. So am I understanding then that Do We Learn is it's about trying to find new ways to educate people? Is that what you're telling me?

Speaker A: Yeah. So the thing that we set out to do, so my partner Dirk Liebish, uh, who's um, a, uh, brilliant, uh, ML AI cto, has been in the AI space for a long, long time.

Speaker B: I'm going to just say, I'm just going to use those, expand those terms.

Speaker A: So machine learning, machine learning Artificial intelligence. He comes from the business intelligence financial services sector, uh, using early versions of AI and now more advanced versions of AI to do predictive analytics. I've always been interested in how do you take that kind of set of capabilities that we use in the financial services sector and apply it to learning how can we make learning more efficient, uh, how can we drive better outcomes, uh, through the sorts of tools that we use in financial systems to drive, ah, financial outcomes. You know, when we started working together, we started to see the tsunami of the large language model and the transformer, uh, sort of, uh, approach, um, the explosion of power of neural networks. Um, and this is even before the fall of 2022 with the launch of ChatGPT. We could see the tidal wave coming.

Speaker B: When did you guys start working together?

Speaker A: We started working together probably around 2020. Um, and then more earnestly in 2122.

Speaker B: Right.

Speaker A: And you could see that, you could,

Speaker B: you could see that becoming. Because you guys were tapped into the computer science space. Yeah. How did you come to be aware that there was this tsunami on the horizon? Yeah.

Speaker A: So, you know, I had been working in, in AI and some form of AI since I was a high school senior in 1985. Um, and like a lot of people, I kind of gave up on it because we just, it felt like running through mud. We could not achieve what we wanted to achieve. Um, and it just felt like it was a dead end. It was an endless slog and we'd hit a ceiling and we had to find new approaches. And I think, you know, I lost the thread of the argument for a while. People like Dirk didn't, they stuck with it. And I think when Dirk is your business partner. Dirk's my business partner. Dirk Liebisch. Uh, I think if you, what happened is you see this incredible intersection of the advances in pure computing power.

Speaker B: Right.

Speaker A: So the Moore's Law phenomenon that we have this doubling every two years of just raw computer capacity with the birth of the Internet and the digitization of all human knowledge. Now you have a training database that this incredibly highly capable technology can feed on. Um, and then uh, you have new algorithmic techniques. So like the transformer. So most people don't know that GPT. The GPT is generalized pre trained transformers. Right. The T is transformer. And a transformer technology was this idea of um, you know, it started with companies like Google. Google invented this trying to, you know, do a better job at uh, uh, autocomplete. Right. Or how do I predict the next word? And what they realize is to Predict the next word in a sentence actually requires a contextual understanding far beyond the sentence, because the next word that I'm using here with you is informed by the whole conversation that we've had. And so this notion of how can we actually create long term context to inform next word prediction turn into a breakthrough in the way we think about neural networks. And this is geeky stuff. So only people like Dirk or me are looking at this back in 2019, 2020, and we could see, we knew right away because of all of the failures that we had encountered in the past, like, this is the unlock, this is going to change everything. Um, and so it was kind of a drop everything moment. And then even getting past that, um, you know, Even back in 2022, 21, 22, both Dirk and I realized that the short term, the early stage of large language model interaction and our human interaction with AI would be text based. But really like we're here doing a podcast where I'm not texting you my responses, you're not texting me the questions. Because as humans, looking into each other's eyes, seeing you nod as I'm speaking, listening to the tone of my voice is as important, if not more important, than the semantic context of what I'm saying. And so what Dirk and I got very excited about in the educational context was what we call multimodal AI. What happens when an AI is not just responding to a text message? What happens when AI has access to the fully multimodal real time context of an exchange between a learner and a teacher, or between two learners or any two human beings? And the very practical problem we set out to solve was in schools today, in many schools. And as a teacher, you probably are familiar with this. The kind of gold standard of assessing and evaluating and measuring teacher effectiveness is not a standardized test. It's having a skilled master teacher come in and observe your class. And if you go back even to the early 2000s with, uh, the Gates foundation and Harvard having done this study on the measures of effective teaching, what they proved was that the most accurate predictor of teachers success is not the number of degrees they have after their name, the number of years of service, where they came from, what language they speak. It was how other teachers observe their

Speaker B: practice, how other teachers observed their practice,

Speaker A: the score that a master teacher gave to them when they walked into their classroom and they graded their class. It was human feedback on how good you were, Bobby, as a teacher, through a structured lens, right? And it would be things like, do kids raise their hands do they look engaged? Are they focused? Are they paying attention? Do they nod in affirmation because they understood what the teacher was talking about? That if you could measure those things, you knew whether a teacher was delivering learning effectively.

Speaker B: Right. Right. Are, uh, you saying, uh, just to clarify your point, are you saying that that would be a predictor, uh, of whether the teacher would be a good teacher?

Speaker A: Yes. It would be both a measure of high quality teaching and more so even a predictor of student success. If you had a high quality teacher who taught high quality, delivered high quality instruction, that was a better predictor of student success than almost any other measure out there. The problem is you can't scale it. It's just too expensive and there are too few educators who can provide that kind of observational feedback. Um, so the first test, but also.

Speaker B: Sorry to keep interrupting. But also if the master educator who goes into the lesson and observes that lesson is there, it sounds like there's also. You said it's not possible to scale that. So it sounds like there's also a benefit to the person being observed that they will improve their teaching, their pedagogical practice as a result of being observed. Is that what you're sort of insinuating? Okay, but the problem with scale is that there's not an master teachers to go around.

Speaker A: Yeah, exactly, exactly. It's not enough observers and coaches, but there also aren't enough skilled teachers. And what we found is the fastest way to scale the teacher's capability is through reflection on their practice. And you get that, um, through these observations. So if you could unlock that, you can create a force multiplier in classrooms. But how are you going to do that? It's a catch 22. We don't have enough good teachers to become great coaches to then go coach the teachers. Right. What if multimodal AI could do that?

Speaker B: Interesting.

Speaker A: What if you put a camera in a classroom and it would give feedback on what was going that was better than what a master human educator, uh, observer could provide to a teacher.

Speaker B: Okay, well, good question. Good rhetorical question. Well, what if. Well, let me ask you that question then. Well, first, what if. And second, but more importantly, how, how can you do that? Okay, and let me just, by the skeptic's mind, let me, let me voice those concerns. Well, the, the master teacher is a, ah, human being. And they're going in and they're saying, you know, they have the years of experience and they're able to organ, you know, from their years of experience, they're able to impart their wisdom onto the, you know, this naive teacher that is doing their best in the classroom. And through their sensitivities and empathy, they're able to extract the best qualities of this, um, naive teacher. So how can we augment that with AI?

Speaker A: Yeah, so, you know, that was a study question that we set out to answer. We didn't know, uh, what the answer was, what it could do. And I'll get to the how, but let me start with what happened, which is, you know, we dropped a couple cameras into classrooms. We created this thing called UE Learn that extracted the features from the video that we were important to educators. And I was, uh, the trainer, right? I told Dewey learn how I would observe a class. And we were just blown away by the quality. Blown away. And as someone who's probably observed and done evaluations of hundreds of classrooms, uh, it was far better and more consistent and more detail oriented than, than I ever was. Uh, that was a hard thing to admit, but it was true. Uh, and to your point about the how, you know, I mean, this is part of the secret sauce of Dualearn, but it really is, uh, we've created a system that can learn from master educators. What do you look for? How do you measure? What kinds of metrics do you use? What are your look fors in the classroom? And dualearn is able to create its own approach and its own, what we call framework based on those conversations with educators, and then apply it to the video with great, great consistency and accuracy.

Speaker B: Okay. All right, so let me synthesize this and put this into my own words. So much like how the GPT models, the large language models would take corpuses of text and then predict the next word. And in order to do that would need to understand what makes, um, ah, a sentence or a paragraph or a piece of writing sing. So too does Dewey learn. It is able to take its training on what a good lesson looks like, what good teaching looks like, and then through machine learning, reinforcement learning, I imagine, is that what's happening here? It's then able to add two and two together and get the correct answer. So it looks at, uh, the videos placed in the room and observes the lesson. It has seen a gazillion million lessons before, and an expert has said, this is what a good lesson looks like. Tick, tick, tick. And it's able to say, okay, I know what a good lesson looks like, and I'm watching a good lesson. Now, is that kind of essentially what was happening?

Speaker A: You got it? Yep.

Speaker B: Right. Okay, now, so as a former Teacher. One of the most stressful experiences for a teacher is when they're about to be observed. Right. So, you know, I'm like, I know that in England, I don't know if you're familiar with Ofsted, you worked for Pearson, uh, which is obviously, uh, an exam board that, you know, provided examinations for. Certainly in England. I don't know if in America is Pearson in America too. So our listeners will know who Pearson are. Um, uh, yeah, when Ofsted's coming in, that's the, uh, again, for the listeners. Ofsted is the government's examination board. That would come in and examiners would go into your classroom and that week was the most stressful week of a teacher's life because they are making sure all of the books are marked, making sure all of the kids know what levels they're working at. And you're trying to make sure that every one of your lessons is an outstanding lesson, because God forbid it's not an outstanding lesson. And the school's reputation resides on whether or not you, you managed to deliver that outstanding lesson. I went through two Ofsted, uh, processes during my career and sure enough, they were extremely stressful experiences. I do, I will just take the opportunity to say I did get an outstanding rating, for what it's worth. I just want to get that on record. But, uh, it was painful and I could imagine teachers and teachers unions even being, uh, somewhat resistant to a, uh, camera being placed in every classroom. Because Lord knows we can't be outstanding 100% of the time. We're already teaching five lessons a day and we're already overworked and there's already so much admin. And I need to just have one lesson where I let the kids watch, um, the Avengers and list every noun that they see and every adjective so I could mark, um, my 16 year old, uh, you know, coursework. So, um, yeah. How do you respond to that? I know that's maybe a little bit of a kind of.

Speaker A: I mean, I think actually this is the perfect solution to that because this can be implemented in a non high stakes way. You know, if we really wanted to help teachers, we would put that, that observer in their classroom all 200 days of school and provide them the feedback. We can't do it. So what happens? They only go in three times and it's a high stakes activity, just like high stakes testing of students. What you really want to do is you want to be assessing them constantly through their learning and not have three exams that determine the fate of their future. So the use Cases that we use ue Learn, A uh, teacher can have it on all the time and never show anyone the feedback. They can ask the most vulnerable question, did ue learn? Because it's a chatbot about how they could have done something differently or better or, you know, another way to do something. And then from an accountability standpoint, my vision and we haven't quite implemented this with our partners, but instead of having, you know, Luyen come into Bobby's classroom, it happens to be the day that, that, that you know, Bobby was having a family emergency at home, the kids in his classroom didn't eat the night before, and the observer, Lu Yan, uh, was in a bad mood. Why don't I have the teacher say, you know, I've been using ue Learn all semester long. I'm going to pick the three videos and my conversations with ue Learn that I would like to be representative of my practice and sit down with the observer and actually have a conversation with ue Learn with the video and talk about what actually happened. And so actually what we find is that it takes a lot of the stress and the burden off of these high stakes. Very few, you know, m like moments in time where you have to perform and makes it just more part of the warp and the woof of what a teacher does.

Speaker B: Mhm. Okay. So you also get more as like a teaching assistant that isn't necessarily a, like the headmaster of the school isn't observing every class, but in your ideal world, you know, while that might be happening in your ideal world, this is actually the teacher's tool.

Speaker A: Yeah.

Speaker B: That they are using to get real time feedback from this um, expert in teaching, this well trained expert. And I can see how that would be valuable because one of the ways I use ChatGPT or all of m my other large language models, I uh, subscribe to all of them at this point. But one of the ways I use them is for reassurance. Um, and more than ever I noticed that, uh, the therapeutic benefit of using these large language models is to, rather than necessarily get an answer to anything, it's more just to get reassurance that what I'm thinking isn't necessarily stupid and I can just keep moving in this direction, whatever that might be. I won't share all of my insecurities in this life, but there are enough of them that I'm able to talk to a large language model about an insecurity. I've started to realize, oh, what I needed was just a pat on the back and say, you're Doing all right. This is fine. Keep going. When I was a teacher, one of the things biggest sources of anxiety was, am I doing my job well enough? You know, um, am I serving these kids well enough? You know, and there weren't enough hours in the day. It was a real, you know, the weight, the burden of responsibility when you're a teacher, certainly in the early years, is that I don't have enough time to mark all of these kids, work and plan all of these lessons. And I want to do the best lessons possible. I've been told in my training that a good lesson looks like A, B and C. And so I have to deliver my lesson in that way. And I could imagine being able to talk to an AI that says, bobby, you're doing a good job and it's good enough. You can have the kids sit down and read for 20 minutes, half an hour in the lesson while you are preparing something else. And not only that, but all of the other teachers in your school are doing the same thing. So don't beat yourself up if you're not doing a, uh, super interactive lesson every time you see this class.

Speaker A: Yep.

Speaker B: And by the way, extremely helpful.

Speaker A: Yeah. And by the way, you know, you may not have noticed, Bobby, that Lu Yan asks a meaningful question in every class. He's his hand is up for the last three classes. He hasn't raised his hand. You might want to ask him why.

Speaker B: Right.

Speaker A: So being that extra pair of eyes and ears in a classroom, super valuable. You know, we built a version of UE Learn now that is a Zoom Teams Google Meet plugin. Because a lot of the use cases now are online learning. Imagine that you plugged you invited UE Learn the way you could a fathom or an otter. But it's seeing and hearing, not just transcribing to what we're doing right now. And you had a companion in UE Learn that could tell you, bobby, you know, you're doing a great job as a podcaster. Here are examples of things that really drew a question or something deep out of your guest. Uh, and I've watched all your guests. This is the thing you do that most gets to the core of what they're about. This is why your audiences are doing Learn can do all of that. Um, it's watching the expression of you, your interviewee, uh, the voice, tonality, the content. Uh, it becomes your companion in understanding. I like to think we've moved beyond a sort of classroom technology. I really think that Dewey Learn Superpower is helping humans understand. Human understanding is the Way I would

Speaker B: put it is that you feel like that's the real mission at the heart of this. Helping humans understand humans. Interesting. Uh, I've got two kind of, like, routes I want to go down. I guess the first one again is all these sort of logistical questions, which is where my mind goes, because obviously, you know, I've been integrated in that space, in the education space for a long time. Well, it was for a long time. And one of the issues was always compliance. I know this is, like, very logistical. Again, into the weeds. But I'm like, how you said you've got a few partners that you work with, and in England, having a camera in the classroom was very, you know, I couldn't have filmed my kids, for example, let's say I wanted to do. I was a, uh, vlogger. I remember back in the day, you know, obviously vlogging was some sort of piece of my destiny. And here I am now doing a podcast. I remember back when I was a teacher thinking about, like, you know, I was experimenting with new techniques. I was experimenting with behavioral management techniques. And at the time, I was like, oh, I would love to do a blog about this. Maybe I'd like to film something and just show how I'm doing this thing in the classroom. But there would have been no way, right? Like, all of the kids identities were sacrosanct. And even to film a kid at, say, at sports day to share with their parents would, um, have been something that just either you needed a ton of permissions or for more of the parents and so on. It was just a bit of a nightmare. Um, I'm wondering, how does that kind of work with you guys?

Speaker A: Yeah, I mean, I think we're all navigating this as a society. You know, we have actually had fewer issues than I would have imagined. And we're working with major charter school programs, with teacher education programs, I think, you know, and, and we're, you know, there's a lot going on in this space. Right? So, uh, most U.S. school districts and municipalities and states have their own, um, data privacy agreements, and particularly when it deals with students and minors. And a lot of, you know, practitioners in this domain, whether it's a school district or state, they now have data privacy agreements that they require their vendors to sign. But so far we've been able to comply with all of those. I, um, think the, the. What I have found is, um, these sort of data privacy agreements and norms and standards which I actually believe in. I think, I think we need to protect ourselves. I think Government has a role in making sure these, these technologies are being used in a way that doesn't compromise our safety. But what I find is this is, it's not a, it's not a single sided equation, it's a, it's a cost benefit analysis. And so when I go into a district or I talk to an uh, educational partner, I don't start with. If I started the conversation with, okay, here's what we're going to do. We're going to put cameras in every classroom and we're going to feed it to AI. I probably wouldn't even get through that sentence before being thrown out. If I lead with, let me show you the kind of insights that we can gather about your teachers, your learners. Then the first question out of those folks is how do we do that? And it totally changes the conversation. And now they are actually working with us to figure out what is the compliance needed in order to be able to do it. And there are a lot of more tactical things we do. So we mask uh, names, identities of uh, students. We don't take the student information system data in. We don't tie it to a, uh, student record. We can blur faces when necessary. We adhere to the most rigorous standards of um, you know, information security and data storage and data retention.

Speaker B: Is the, is the other, are the recordings like saved? Does the, is the teacher able to watch back or is it more that. It's just like, it's just like the AI has watched it and has a memory of what happened. And then I can talk to it,

Speaker A: we can do it either way. The system works perfectly well if it analyzes a video and then never sees the videos, you know, deleted or, or the customer keeps the video and we have no access to it. You get a better experience if the video is available because you can go from the insight back to the video. Right? If it says at minute 1534, the students seemed disengaged, they were all looking down at their toes. Being able to click and see that part of the video is useful, but it's not required.

Speaker B: Sounds like an incredible piece of technology. So, um, how was it developing it?

Speaker A: Uh, incredible. I've been an entrepreneur for 35 years and I've built multiple companies and I have never experienced the velocity that we experience now because, so I often say this, um, to people like, we have eight human employees in the company now and we have something like 12 AI employees. All the software development, the coding is being done by AI agents now. And the system is improving itself much faster than we could as humans improve it.

Speaker B: Okay, when did you found Dewey Learn?

Speaker A: Uh, so we founded it in the spring of 2024 and uh, we raised our first seed round in almost just about a year ago in February of 2524.

Speaker B: Did you say spring 24? So April 24 through to February 25. How did you. To get that seed, you had to have a prototype. Did you have to have a prototype?

Speaker A: We had a working pilot. We were piloting with customers.

Speaker B: Uh, right. Okay, so how, how did you get that first piece of technology that was ready to be deployed into the classroom?

Speaker A: Up and running, you know, Dirk doing his magic and me acting as product manager and a lot of IT agent. Even back then we were using agentic AI to build these tools, um, really, really fast. And uh, you know you can, you can scale these things incredibly fast now. And you know, we have a very, you know, lean, agile mantra. You know, let's build just good enough to deploy and experiment and get it into the hands of customers.

Speaker B: Okay, I'm going to pivot now into this more like practical, uh, part of the podcast now, which is, yeah, I want to hear about this kind of day to day building and how that all works. So you were a product manager and you were able to start building, were you building it yourself? Did you have uh, some devs that you worked with?

Speaker A: We were building it ourselves. It was mostly Dirk. Then we hired one other developer, but very early on we were agentically coding. We were using kind of the predecessor to OpenAI Codex and predecessor to Claude code.

Speaker B: Um, okay, m. Tell me about that, um, relationship. So Claude code is obviously the sort of darling of the moment and Codex meanwhile, what's the difference between Codex and Claude code? Why were you using them both?

Speaker A: Uh, well, it's just we'll use whatever is the best of breed and these guys are all leapfrogging each other. Right. So it's really more that than anything else. I mean originally, um, uh, you know, chatgpt or GPT and originally we were using even things like cursor, um, uh, just produced really good code, um, fast and um, there were a lot of issues and you really had to bird dog it and manage it and there were a lot of techniques to get it to work. Right. Has very bad memory, you know, all sorts of things like that. And then when Claude code launched, uh, it was way better than anything else and better than GPT. And so we kind of switched over uh, to Claude code and then Claude, really the next step is it's one thing to have one Agent helping you to code, um, and then you want it to be even more capable as an agent. So you don't want to have to go and set up the GitHub, uh, repo and deploy stuff to the web. You want the agent to be able to do all that. So that got more and more capable. And then I think what started to happen was people started to realize like, I don't just want one agentic coder, I want multiple agentic coders. Because I want one really focusing on the front end of our system. I want one focused on the data stack, I want one focused on uh, the AI, uh, analysis. And then actually I don't just want one, I want three for every one of those because I want them to jury their results so that the result I get is actually what three agentic, uh, programmers decided was the best thing. So then you start to see the emergence of these agentic coding frameworks. You need some sort of harness, some sort of air traffic control to manage these multiple agents. And originally Dirk, my partner, wrote our own kind of agent framework to manage that. And then Claude code built uh, their, uh, their SDK API stack and that helped to manage things. And now the new version of codex from uh, OpenAI is exactly that. It's really a management framework for multiple, uh, agentic coding, um, sort of programs. But it manages them like air traffic control and the human job, like our jobs, Dirk job, my job, uh, our senior developer, Bojan. They're almost like code managers. They're like dev managers. They're looking at the code that the agents bring to them and they're validating them and sort of enabling them to check in that code or rejecting the code. Um, and so they're kind of acting as the maintainers rather than as the coders.

Speaker B: Okay, so, all right, that's a very interesting insight into just what it is to build a product from the ground up in this new space. I've also experimented with vibe coding. We're developing apps here at our uh, business to just streamline workflows. And yeah, it's been kind of amazing and eye opening. My apps are very simple. It sounds like what you're creating is something quite complicated. I'm here, I'm. And I'm wanting you to elaborate on that shortly because what I'm imagining, there's a camera in a classroom that is feeding into some sort of AI model that then the teacher needs to, or the user needs to be able to interact with whatever has been put into it. So then there needs to be an interface for the teacher to kind of interact with. I mean that's just at the basic level and I imagine that there's other sort of management and UI requirements. So help me understand this tap, help me understand what you've built. I guess. Yeah, if you don't mind.

Speaker A: So at a high level what we do is we take an input, call it a video of a classroom, um, or any kind of learning. We have multiple models that we. So we want to basically decompose that video into features that are relevant for an educator.

Speaker B: Right.

Speaker A: So if you take a typical video, if you think about how computer vision works. Right. So computer vision processing of a video, it's taking every one of 24 frames per second in whatever 1,080 pixels and it is extracting every feature from each frame. So you can imagine how massively compute and storage intensive that is. But on top of that, a just out of the box computer vision model doesn't know that. Actually in an educational setting saying, ah, uh, that's a teacher, that's a student. The student's raising his hand is more important than oh, there's a plastic fern in back of this person. Right. So the first step is we create a framework, uh, kind of like a uh, fine tuning framework for even though we use an off the shelf vision model, we enhance it to say, here are the sorts of things you need to be paying attention to, here are the things you should not pay attention to. I liken it. It's a little bit like, not to get too geeky, but it's kind of like a, ah, compression algorithm. Like if you think about MPEG or video compression, what it's doing is saying between these two frames, what's the important information to maintain? And so that's what we're doing but through the lens of learning.

Speaker B: So you're building that algorithm yourself, you haven't used a foundational model like an API from.

Speaker A: No, no, we do a foundational model for the pure vision processing. Uh, we don't build our own vision model, but we have grafted onto it this additional intelligence of use your powers to extract these sorts of things. And then we do the same thing with what we call prosody. So speech tonality, you know, here are the things to listen for, uh, here are the things to extract. And then we do the same thing with the semantic processing. Right. So now you have semantics, vision prosody.

Speaker B: Wait, define semantic processing.

Speaker A: So semantic is just the content, the linguistic content. If we had a transcript of our podcast here, that's the semantic content Right. If you have an audio tape, uh, and you're listening to me get excited about something, that's the prosodic layer, what we call prosody. And then the visual is me raising my hands like this, smiling to you, Right?

Speaker B: Sure. And it's the same AI model that you use for the base of each of those.

Speaker A: Three different, for each of those we're using like seven different models. Uh, uh, but we're grafting this intelligence into each of those to extract the features that are meaningful and important for our application. And then that extraction ends up getting stored in what we call a rag. Uh, a rag is effectively think of it as an underlying database for an AI model to, uh, query. Right. And it has these visual prosodic semantic components in it. So now if you put ChatGPT on top of that, and you say, hey, were students engaged? It can go in and say, you know, well, I understand student engagement means students are. The look for is, are students looking forward? Are their eyes on the teacher? Are their hands raised? Are they asking questions? Now it can extract that insight based on this incredibly synthesized database that we've created.

Speaker B: Right. So, all right, let me just describe that pyramid that you've just described. At, ah, the base of the pyramid, there are multiple, um, different models that specialize in the different disciplines, whether it's m, a visual model, an audio model, or I guess you called it, um, you say prosaic semantics.

Speaker A: Yeah, prosody and semantics.

Speaker B: Right. So that's the language. So those you've got several models and maybe more besides in this kind of at the base here, then you've built on top of that, uh, you described it as your own intelligence, but it sounds like you've coded things to help uh, digest what those models have picked out. So look at the student raising his hand in the air. Uh, don't look at the fern in the background.

Speaker A: What do you pay attention to? It's all about attention, right?

Speaker B: And then it reports and then that all gets distilled into the next layer, which is, let's just say ChatGPT, a large language, a traditional large language model that the user can interact with. The large language model that digests all of the information from below in the lower tiers of the pyramid. So my question to you is, did you vibe code your own intelligence? Was it just a case of just prompting, let's just say Claude code for the sake of argument. Like we need to tell our uh, foundational models to look out for the kid putting his hand in the air and not the Fern, can you write that into whichever script you need to write that into?

Speaker A: Yeah, it's a little more complex than that. In that, um, the coding is done by the machines. But we have, you know, Dirk and his team kind of saying, like, uh, hey, we want to take this architectural approach. You know, the research says we should take this approach to extracting these features. We want to architect it this way. And then they're using AI to build the building blocks, the components of that. But it's guided by an architecture that they have defined.

Speaker B: It's the power or it's the power of these tools in people that know how to use them. That's right, is basically what I'm hearing here. I can't walk in and just start building this thing that you've described because I don't have a development background. But if you're a computer scientist or a developer, uh, you're supercharged, which just spreads across all industries. In video production, there's this moment of, oh, my God, you can now generate a video in VO3 or midjourney or whatever your choice is. But as the, uh, kind of the dust is starting to settle, you realize these tools are very powerful in the hands of people that know how to use them. And the professionals are the people that know how to use them, and they can use them very, very effectively. Meanwhile, those that aren't professionals can produce maybe cool stuff, but just not as cool stuff. You got it. And I'm finding that in my own development, uh, adventures, I can build this application at work that's actually very simple and it's great and people can use it. But I'm constant thinking to myself, I wonder what the kind of the people that really know what they're doing are going to be able to do. And it sounds to me that, uh, one example is they're going to be able to build an eye in the sky in a classroom that can act as an omnipotent observer of a lesson and then give extraordinary feedback to the person delivering that lesson. That's basically what I'm hearing here.

Speaker A: Yeah, yeah, you got it. That's exactly right. And I think what it points at is, boy, is there still room for human expertise. Human expertise is a force multiplier, uh, to a set of tools that are highly democratized but can be used very poorly as well as effectively. That human vision, insight, expertise, knowledge is what actually is. It's the opposite of what we thought. It's not that the VIVE coding is a force multiplier for humans. It's the humans are the force multiplier for these agentic capabilities now.

Speaker B: Yeah, that's a good point. You're sort of speaking to the existential dread that everybody, uh, experiences when they talk about this technology taking jobs and all the rest of it. But, yeah, I take your point and I want to expand on that because, yeah, obviously these tools don't exist without the humans in the first place. And everything that they are doing is a reflection of the humans that A, built it and B, want to use it in order to do more things. And these tools have no motivation on their own. And they're not gonna just start taking jobs for the sake of it. They are working for somebody. Uh, and in a utopian vision, all of us can have all of our wildest dreams and imaginations realized, because now I have an idea and I can execute it with, um, my super intelligent friend that has been trained on all of the previous people that came before me and all of their knowledge and information. Uh, again, it's a reflection of, of the humans, not some alien kind of, you know, unique thing that does.

Speaker A: Yeah, you gotta.

Speaker B: Nowhere. Well, Leanne, I feel like we're starting to come towards the end of our time, but this has been a really, really fantastic exploration of this topic. I've really enjoyed the conversation. But before we go, we are going to move to the. The most, uh, what's the word? Exciting part of the podcast. The fun bit is a, uh, fan favorite part of the podcast. We are swimming in fans at this point, at least. Double figures.

Speaker A: How many are AIs?

Speaker B: There might be a few bots in the mix. There's a few Molt Bots.

Speaker A: Yeah, I was going to say any

Speaker B: molt bots in there on Maltberg is huge. It's blowing up.

Speaker A: Yeah, that's a podcast unto itself.

Speaker B: But, uh, the quick tips section of the podcast, uh, is what I'm trying to get at. So, I mean, you've kind of, you know, been delving into superpowers related to agentic coding for sure. But, uh, now this is the bit that I want, you know, viewers just tuned into one bit and they wanted to get some practical kind of takeaways for themselves. This is what this bit's all about. So I'm going to ask you three questions. One for a general tip, just like a general, you know, everybody could use it for this or try using this or this functionality or whatever it might be. And when I say this, I'm largely kind of revolving around large language models. The second tip, I'm going to ask you for four is your own personal superpower or hack, and then third, a surprising or unexpected use case. But you know what, before we get onto those three things, because there's another sort of piece of this puzzle that I always like to look at just before we get to this kind of to finish. I want to finish on an optimistic note. But I also am curious to know whether you have any misgivings about the shape of things and if. Is there anything here that you know, any watch outs, any concerns, anything about. You're talking about a very optimistic vision for how do we learn could help in the classroom. But are there any like kind of pitfalls or minds or tripwires that you're trying to navigate your way around? Yeah, I guess I'm speaking to the realist in you. The perhaps the.

Speaker A: No, I worry about this a lot. I mean I think there's a lot of concerns. I think, you know, for one, I think there's a world in which these agentic AIs are a huge threat to the profession of teaching. You know, I get questions all the time about do we really need teachers? Can we just have agents as teachers? And I'm a firm believer that humans teaching humans is like a, ah, fundamental, uh, value. Um, it's not optional. Um, but that's one view. And other people have different views. And how for those of us who believe in the primacy of human to human interaction, how do we carve space for that? You know? And so part of the urgency I have around ue learn is how can we make human teachers better, faster and in a more human way so they're not outstripped by the agents? And so I worry about that. I think another big area of concern I have is what are the durable skills? What should we be teaching in a world where um, ChatGPT can ace every single academic exam that we could give a human student, including graduate level students, including now 40%, uh, proficiency on humanities last exam, which we, we scale AI built as the last bulwark of defense of humanity versus computers. And in the less than 12 months computers can already beat it at 40%. What are the durable skills that we need to be thinking about? And they're probably not things like reading, writing and arithmetic and factual recall. They're probably things like empathy and creativity and collaboration and storytelling and ethics. Um, so how do we. I think we need to really radically rethink what the value proposition is of our formal educational institutions. And that is both hopeful and worrisome because I think it's hopeful because I think we have incredible opportunity to raise the ceiling of value for education. It's worrisome because educators are sadly, and I will include myself in it, the most resistant to change and evolution. And what do we do about that? How do we accelerate change in education is something I worry about. And I also think I worry tremendously as an educator about what I now call truth decay. Not uh, tooth decay, but truth decay, which is the erosion of factual truth, our ability to know what's real. And this is a trend that started pre AI uh, with moral equivalency in the political sphere. Uh, and now we're in a world where everyone can see a news headline and not actually know if it's true or not, um, and take their own meaning out of it. And how do we navigate that as educators, um, and just more broadly as citizens of civil society? I'm really deeply concerned about all of that.

Speaker B: Yeah, well you started this conversation as saying something about the slow pace at which education space has evolved and that was something that struck you. And it does seem like we're on the kind of verge, uh, of education having to wrestle with that very, very quickly. I'm hearing that in some of the more prestigious universities they are starting to administer their exams orally rather than in any type of written capacity, which uh, seems like a one smart building block in this future. And I too, I happen to be quite optimistic about the teaching profession. I can see how do we learn is an um, adaptable technology that no matter which direction teaching goes in, whether that is teachers using traditional methods of standing at the front of the classroom and trying to engage 30 kids or whether because every human being who's been put out of work is now funneled towards right human to human based jobs of which I see teaching as being one of them. I think we will all want our children to be raised by humans. You can imagine that do we learn would be very useful in helping that new influx of humans into the profession to learn how to administer or convey these new skills which we have to teach. Now what are those skills? Still remains to be seen. I feel fairly confident, confident now that those skills are, you know, human to human life skills, empathy, etc. I even think, I believe I was. There's an npr, um, article, uh, episode, there was an interview with where they were talking about those coming out of economics education. The, the economics courses in universities. I'm butchering this, but the economics courses are now re. Pivoting less towards the kind of data science and more towards modules on People management, because that is going to be the skill of big business in the future. I think all of that feels like what the kind of the landscape is going to look like. And I, for one, am super excited about.

Speaker A: Yeah, me too.

Speaker B: The impact that AI is going to have on education because teachers have for too long been overworked and the class sizes have been getting bigger and bigger, uh, the students more and more disengaged. And I just see this sort of bright shining light of potential to differentiate lessons so that every kid gets their own personalized education. And the teacher can kind of focus on making sure that the lessons are engaging and delivered well, even if that lesson is planned to some degree by an AI. And maybe technology like do we learn? Can help me get better at kind of recognizing when a kid needs intervention or whether, uh, my own skills need to be sharpened. Um, how do you respond to that?

Speaker A: As my own little diatribe, completely 100% agree. I think if I were starting my career as a teacher, I would be over the moon that I was at the dawn of the reinvention of one of the most profound institutions in human history that is quite recalcitrant, but is now being forced to rethink and change. And I could, as a young teacher, have an impact that I never thought possible. I do see the flip side of that. I spend a lot of time talking to schools, universities, educators. And I think you cannot underestimate, um, the fear, uh, and the conservatism in traditional educational institutions. And so I think my biggest hope is with the new generation, uh, folks who are coming into the teaching profession now will have different expectations of what that looks like and what tools they have available to them, and they will push for innovation. I think teaching and learning are going to be more important parts of our labor economy. So if you think about it, what is a foundation model company actually doing? They're teaching a model to deliver value and align with human goals. Every foundation company, every, every frontier, uh, AI company is hiring people with, if you look at the skill sets, it's a teacher they're hiring. They're hiring educators, they're hiring learning specialists. And I think the field of education is going to broaden in scope and actually be a much larger part of the global labor economy over the next decade than it ever was before. But a lot of that learning may not be humans. It might be humans teaching machines to be better at interacting with humans. And so I am very helpful. But I think it's a battle, right? It's going to take Persistent effort.

Speaker B: Yeah, no, I can completely relate to that. And it makes me think about some of the other stories I've seen in the Economist, is that, yes, all of these models are going to be super capable, but we're not taking into account the adoption rate of these institutions. And in my experience, when I was a teacher, I was would. As, uh, I said at the beginning of this podcast, one of my remits was to use technology and explore technology that could enhance education. And then once I found something I thought was effective, roll it out across the faculty and educate other teachers and say, you should try using this new technology. It was almost impossible. It was so hard.

Speaker A: Not because of the technology.

Speaker B: No, no. But you know me, just some young punk straight out of university who's trying to tell this teacher who's done the job for two decades, hey, have you tried using this technology? It's really great. They're like, uh, don't try and tell me how to do my job. Thank you very much, Sonny Jim. You know, Bobby, I don't see that.

Speaker A: To give you a sense of how old I am, um, the school that I worked at, which was my alma mater in New York City, my first job out of college, uh, I created their first ed tech program there. And m. My team and I were the first to wire the building to this thing called the Internet. And you should have seen the reactions of the teachers in the school. They were like, what are you guys doing? Why? How do we kill it? That was the reaction.

Speaker B: Yeah. Wow. I mean, even thinking about it now, you know what I mean? I went to school in 1995, and we had a computer science lab. It's called a computer science lab. There was just a room full of computers. And when I was a teacher, I was a teacher 10 years ago. And still it was rooms full of computers, and still kids were doing everything by hand. On. You know, when you think about it, it's so obvious that eventually that type of way of doing work, I mean, obviously it has its benefits. And I'm not saying that we should do away with pen and paper. You know, clearly everything's online, everything's digitally. There's some sort of mismatch.

Speaker A: I taught history. My first year teaching was teaching history, ninth grade history. So the first year of high school. So it was a course I had taken eight years prior. Right. Or less than that, I guess. Whatever. Yeah. Eight years, right. So I'd gone through high school, college. I was coming back to teach the course I took as a high school freshman, and the textbook was the exact same textbook that I had studied from and it was published in 1968. Students were learning about the world from a single information source that was from 1968. And I didn't remember anything from that course which tells you the value of a history education textbook bound history education. I was rereading the book that I had Learned from in 1985, from 1968. And I was teaching the same things to these students.

Speaker B: You know, I had a pediatrician on the podcast, um, and you sort of talk about Moore's Law. Well, there's a certain Moore's Law that applies to um, health literature. And he was saying that uh, they say it's like every two years or every 18 months, the amount of information to do with the uh, doubles. Yeah, the information coming out of the literature. And so he's like, even during your doctor training, your healthcare training at university, things are going out of date immediately. And he was saying, he was saying that's one of the benefits of ChatGPT is because you don't need to know everything as a doctor, you just need to have the critical thinking skills. But you can go to ChatGPT and get actual up to date information about whatever's happening in the world. But the inability and the. Yeah, it speaks across the board here. It speaks to the need, as you say, to know how to learn, to know how to investigate, to have critical thinking skills and then use these tools to amplify that, uh, or augment that, uh, I think is sort of obvious here. The other point I wanted to bring up or make based on what you're saying is so, because of this institutional lag, I suppose, of adoption, I guess it is going to be with those newly qualified teachers. And back to your point, I think it is a really exciting time for a teacher to be entering the profession because they have the opportunity to completely be the vanguards of this kind of new way of teaching. And uh, that's just so exciting because it's not going to be the old guards. You know, a lot of, you know, speaking from experience, the people that are kind of five years from retirement are like, well, I don't want to learn a new way. So it's going to be the kind of, if you're 20 something going into teaching now, um, I mean, what a great time to start experimenting with how to teach your kids.

Speaker A: Totally, totally.

Speaker B: Right. Well that was a little like a little extra. Do you still have time to do the fun part of the podcast?

Speaker A: Let's do your question. So ask me the first one again.

Speaker B: All right, general tip.

Speaker A: General tip. Uh, pay for your AI.

Speaker B: Okay. I love that. Go on. I mean, so does Sam Altman.

Speaker A: Um, yeah, so whether it's Claude or it's ChatGPT or Google Gemini, pay for it. And there's multiple. There's three real reasons. I always tell people this because everyone's like, why should we pay for this? As another service, I pay for so many things, and I can get it for free. So, number one, for anyone who has the paid versions, the paid versions are light years better than the free versions in every respect. Number two, you have more safety, privacy, security assurances, uh, in the paid version. But the real reason I say that is what we learned from the social media revolution was that if you don't pay for something, you are the product. Um, and the only way there is they're giving this to you for free is because they're extracting all the information from you and selling it to their real customer. Who are the corporates that thrive on that information. We are at a pivotal moment in the history of these foundation models where they are trying to figure out what their business model is. And I don't know if you've seen. It's gone viral, the Claude, the anthropic television commercials that they're going to air

Speaker B: on the super bowl, which explain it to our guests.

Speaker A: Yeah. So OpenAI has now announced that ChatGPT is going to have advertising, advertising, advertising in it. And on one level, you're like, so what? Right, like there's going to be an ad, and now I can use ChatGPT for free. Uh, and the anthropic, uh, which makes Claude, has committed to never having advertising. And they've launched a series of commercials that tell, in a very poignant but snarky way, what your experience looks like, uh, at ChatGPT with advertising. And I think that we look to these models as the arbiters of truth. And as soon as you integrate advertising into it, you have completely violated the boundaries of church and state, in my view. And so I want to encourage these companies to see that the real business model, the real value is in convincing customers like us that it's worth paying for. And if you're the customer, you have leverage. If you're the product, you have no leverage. And so I tell people, just cancel your Disney plus, cancel your Netflix, and pay for your ChatGPT. That's the first thing I would tell people to do.

Speaker B: All right, good tip. Good tip. I like it. All right, um, well, before we move on from that, what do they get? Just, let's just talk in practical terms. What do you get when you pay for one of these models other than leverage?

Speaker A: So you get far more tokens. So you get more interaction before it says, sorry, uh, you're out of tokens. Uh, you don't get to talk to me for another 12 hours. Right. And if you're coding or you're doing any kind of problem solving, the last thing you want is to get interrupted in the middle of that. And so you get more usage, uh, you get more features, the ability to plug it into other, um, applications, uh, access to coding environments, and just a lot more power. The quality of the result, you generally can use more sophisticated models. And, um, you get, get safety and security and privacy assurances that you don't get from the free model.

Speaker B: Right? You can turn, you can turn off the sharing your data with the, uh, the owner, which I guess is part of your leverage that you're describing. I'll also add to that that as more of a kind of a, uh, you know, call to arms, as, you know, amplifying what you've already been saying. But these are going to be. It's going to be through some sort of large language model that we will be interacting with our devices in the future. And it will serve everybody to get better at interacting with an AI and learning how they work. And the more you can integrate it with all of your various different workspaces, the better the, um. Sooner the better, really, to do that. You're at a disadvantage if you're not learning that. That said, um, if anyone has gotten to this part of the podcast and isn't already, uh, subscribing to one or another, then I will be, uh, amazed. But leave a comment in the comments if you are not subscribing.

Speaker A: Yeah. Or if you're subscribing, which one are you subscribing to and why? It would be interesting to know, too.

Speaker B: You should. You need to do my job. All right, uh, so, Lian, um, next one is your own personal superpower or a hack or some sort of interesting, um, quirky way that you can use one of these tools that other people might not have thought of.

Speaker A: Yeah, I am, um, a big believer that context is key, that the quality of the results you get back from your AI increase exponentially with how much it knows about you, you. And so, you know, I don't know how many of your listeners do this, but I'm very intentional about what I put into the system prompt, uh, of my AIs. So I will go into the preferences setting and I will be really thoughtful about how I queue it. You know, the, the banal, like don't use em M dashes to the more, you know, sophisticated. Don't, don't be sycophantic, be a skeptic, challenge everything that I, uh, say don't compliment without a really good reason. You know, those sorts of things. Uh, and I also tell it a lot about myself and I always turn on memory because I want it to learn as m much as it can about who I am. And in the system prompt I say refer to your memory first. Make sure you're using your memory about me before you answer anything.

Speaker B: Okay, I'm going to just uh, make that super clear to our listeners. So also we're talking about paid features here. Memory I'm pretty sure is only available if it's a paid feature. But to our listeners you've talked about two things there. One is that there's a, in settings there is a area where you can, the uh, text box where you can give your AI a, uh, kind of like a base level of knowledge telling it how do I want to be talked to? Here are the key things that you need to know in every interaction with me. And you know, like don't use EM dashes. The other part of it is memory where it stores everything from every conversation. And that is a super powerful way to have a whole bunch of context. So it will say, you know, oh Bobby, I know you've got a dog and um, so you should think about that when you're planning this trip that you've asked me for details about. Um, I have a question for you though because firstly, which is your chat bot of choice? Which, what's your main, which one do you tend to use?

Speaker A: It changes all the time. I actually just uh, downgraded my ChatGPT Pro account to plus and upgraded my Claude to Ultra. So Claude is kind of my, my tool of choice right now. I use both a lot and I like Google Gemini a lot too for certain things.

Speaker B: We're in the same boat. I happen to be, I happen to have just downgrade. ChatGPT has become my third now and I've got Gemini in the middle and Claude is up top. But one thing I always noticed was that when I was doing system prompts, I haven't gone back to it for a long time, but it wasn't very good at kind of.

Speaker A: Mhm.

Speaker B: I don't know if it was read, it wasn't reading them I'm wondering, when you ask it not to do EM dashes, is it affecting?

Speaker A: Yeah, for me it is.

Speaker B: Uh, okay, that's great.

Speaker A: But the other thing just related to the. Just in a little extension of that, and this is particularly useful for those who are doing, like, vibe coding. The thing to remember is memory and context are very precious in an LLM, uh, um, interaction, even with, like, a lot of context window. Right. Which is the context window. So, meaning the amount that your LLM can store in its memory when it interacts with you has grown tremendously. Knowing what's actually. It's the same thing as dualearn. Knowing what's important in that context window is hard without, uh, guidance.

Speaker B: Can you define context window?

Speaker A: Yeah. So the context window is anything that your large language model adds to your prompt based on its memory system prompting documents that you upload. It's its awareness of the larger context, and that is literally think of it as a text document. Right. Context windows have a certain size limit in terms of literally the number of characters in that context, and that is increasing. It used to be very, very small. If you had three, uh, prompts, conversations, uh, with ChatGPT, it would forget what you did in the first one immediately. That context window's gotten bigger and bigger and bigger. But with that getting bigger, the same problem that I mentioned with dewilearn starts to take place. It doesn't know what feature is important in that context window. So one of the tricks I use is, particularly with coding is I actually manage the important stuff in an external document. So I will tell Claude when we're. Every time we get to a major stopping point, major milestone in this coding, I want you to update a change log that explains how we arrived at where we arrived at and what the lessons are that we learned. And then what I tell Claude to do is every time I fire up a new session, check that change log, and that results in radically better outcomes.

Speaker B: Check the change. So you start a new conversation every time you do a coding session, and just make sure there's an updated change log. That's very handy.

Speaker A: It has to read that. It has to read that document. So actually now, now Claude has kind of built that in. So there's a Claude MD file when you code, it's a markdown file, and it's like the system prompt, but you have to go further than that. I say build an additional file. I call it Lessons, or I call it changelog, like Lessons md Distill. Use your intelligence to distill what are the core Lessons you learned from the last 72 interactions documented in this document. Document. And every time I start you up, read that again and that surfaces that to the top of its attentional window.

Speaker B: That's really great. It's really great. That's a really useful term as well. Context window. It's something that I've been. I hadn't fully understood the definition of that, but that I knew that, uh, you know, these long chats can start going a bit loopy the longer they get. And I understood it to be that it's uh, reading everything in the chat before and your prompt, but I hadn't realized it was also documents that you upload as well. And that M that helps me understand why NotebookLM, M& Gemini with these massive context windows are so great at, uh, uploading corpuses of text that you need answers about. But when I upload say 10 interview transcripts to ChatGPT, it can only give me answers around about three of them.

Speaker A: That's right.

Speaker B: And it's the context window.

Speaker A: The context window. And that's the idea of rag. I mean that's very much what we're doing with Julilearn is the way around that is take these podcasts one at a time and extract the things that are really important and ask me if these are the right things. And let's create a smaller document that just consists of the important stuff. And now you've got a memory nugget that is a distilled essence that's really important stuff across all of those pieces of content.

Speaker B: Be aware of context windows and figure out how to hack it it work around it. Okay, great. Um, all right, that's very good. Practical tip. This is some fantastic practical use cases we're getting out of you here. All right, finally, your fun or unexpected use case.

Speaker A: Oh God, so many great.

Speaker B: Well, look, just uh, fire at will.

Speaker A: So, you know, one of them is you mentioned NotebookLM. Uh, so you know, I love NotebookLM, but not for what most people love it for. Not for creating a podcast. I use it to create the infographic that helps me understand what's going on. And so if I have a complex doc research paper or something I'm trying to understand or whatever, I'll feed that into NotebookLM and I'll say give me an infographic or give me a five page deck that explains this, uh, in the crispest terms possible. And it has been a huge benefit to my ability to learn stuff. I don't even have time for the podcast anymore, let alone reading the document. But if you build the infographic for me. A visual learner. Boom. Uh, like, total hack, you know? Total.

Speaker B: That's amazing. From a, uh, Harvard educated person. Harvard.

Speaker A: Quite embarrassing, actually.

Speaker B: And Columbia alumni. You heard it here. Those infographics. I love those infographics. I absolutely. I did a, like an episode on the infographics alone, and I just. Mind completely blown. My wife's a bit of an AI skeptic, but I started making infographics for various things to do with our childcare. And, uh, uh, her words. In the end, I said I could also turn this into a spreadsheet. I think it was like we were doing a trip to England. I turned it into an infographic, like the plan. And I said, I can also do this as a spreadsheet if you want. And she said, no, Bobby, we're infographic people now.

Speaker A: I love it. I love it. That's awesome.

Speaker B: Uh, I said, yes, we are infographic people now. I love it. Love that, though. I mean, a big, complicated computer science paper. I'm going to have to try and chuck one of those into, uh, Notebook LM and see if I can get an infographic that helps me level up my understanding of all of these complex topics.

Speaker A: Blown away, uh, by that. Uh, and the last thing I would just say the other one is if you take Dewey Learn and everything I told you about it, um, I've been taking old videos of my kids, uh, where I had. There's one where I do, um, a math problem with my son when he was like three and a half years old, four years old. He's now 23 years old. He works for my company. Company, But I just found it. Like, think about how many billions of hours of video parents have of their kids up on the web. I've been feeding those into ue Learn one by one, and extracting incredible insights about my kids and about me as a parent and about what, uh, the real meaning of our exchanges are that have, like, like, had a profound, profound impact on my perspective on my relationship with my kids. And I share it with them. And they're like, dad, you never understood this. Dewey Learn got this. You never understood. And I just think it makes me realize we've documented all of our lives digitally and we've never actually extracted the real latent value in that. And with something like ue Learn, we have an opportunity to finally do that. And that has been like, a big mind blower for me.

Speaker B: Okay, so a bit of a. Maybe also another product on your hands. Uh, there. Um, that's nice. Can we end this podcast with a heartwarming, um, profound moment. Are you able to share any of those insights that it gave you?

Speaker A: Yeah, actually. So there's a video of, uh, my son shout out Bailey. And, uh, I was such a bad dad. He would sit down to eat his little chicken nuggets. And I would immediately go into, all right, Bailey, we're going to do a Star wars math problem. There are two Death Stars, and each Death Star has five star cruisers in it, right? And, like, how many Star cruisers? And, you know, God bless this kid. He was always game for these things. Like there's 10 and here's mine. And, you know, I fed it the first time I fed the. This, uh, literally it's a four minute video. I put it into Dewey Learn almost as a lark. And I said, tell me something I don't know about this, that I might not know about this learner. And it said, because in the video he starts to get a little confused and his explanation gets convoluted about something and it says, at these moments in the video, the student, the learner's hands go up and he gets agitated. His voice volume goes up by 12%, the speed of his phrasing increases by 35%. And then he expresses frustration. But rather than indicating that he is frustrated because he doesn't know something, he gets the right answer 100% of the time right after that. And it explained that what he's actually doing is not expressing frustration or a lack of understanding. He is emotionally priming himself to get the energy to then do the mental math to get the answer, which is actually classic behavior, uh, of people with particular cognitive processes. Like, it's a very classic ADHD behavior. And I got frustrated as a parent because I didn't understand why he was getting frustrated. And it was only after seeing Dewey Learn's analysis of it, I realized that I shouldn't have been frustrated, that this was the way that he was getting to the right angle cancer. And when I showed it to my son, he was like, exactly, yes. Why didn't you understand this? And it was a very poignant moment because on the one hand it was like this amazing epiphany. And on the other hand, I'm like, God, if I had had Dewey learn 20 years ago, I would have actually parented him differently.

Speaker B: It could have helped you see your child more clearly.

Speaker A: That's right.

Speaker B: Yeah. Well, that's lovely. That's a really lovely practical. I think that's a great place to wrap up sums everything that was spoken about up in a very kind of heartwarming and clear way with a visual reference. So thank you. Thank you, Leanne. Really, really great having you on the show. And, um, yeah, I think we'll have in the show notes any links to do we learn and anything else that you'd like us to share. But for now, I'd just like to say thank you to you and to our, uh, listeners and viewers. Thank you, too. I'll see you next time on it's not the End of the World. Unless it really is is the end of the world.