DeepMind’s Learnings in Developing an AI Tutor

Hey, this is Diane, and you're about to hear a conversation that Michael and I had with Irina jarenka from Google DeepMind. She's the AI research lead for education there. And I think you're going to love this conversation. It was fascinating for us to talk with someone who is literally working on the large language models from the education perspective, and at Google, no less, one of the most ubiquitous edtech products in the world at this point, and her perspective on, you know, where AI is going, where her work is going, how it's going to be, how she imagines it's going to transform schools or not transform schools, and what's important turns out to be a really interesting dialogue. I think you're going to love it. Hey, Michael.

Hey, Diane. It's good to see you.

It's good to see you, Michael. I'm really excited for the conversation we're going to have today. I find that while almost everyone is talking about AI, almost no one seems to know what they're actually talking about, especially in the circles that I think we sometimes run in. And so I've always found that technology is a bit of a black box to many educators, and I think AI is exacerbating that. But today we get to talk with someone who works on and in the black box, if you will, and understands its intersection with learning. She just understands that just about as well as anyone I know. And so bringing both of them together is Irina Derenka, and she's joining us on the show today. Welcome, Irena.

Thank you.

Irena is the research lead for AI and education at Google DeepMind, and we'll unpack all that in a minute to help people understand what that means There, she's exploring how generative AI can truly enhance teaching and learning, and it's not just by providing answers, but also by helping people learn more effectively and equitably. She recently led a landmark study called Towards Responsible Development of Generative AI for Education, which looks at what it takes to design AI tutors that are actually good teachers. And before DeepMind, Irina earned her doctorate in computational neuroscience at Oxford, studying how the brain processes speech and learning. Her work beautifully bridges neuroscience, machine learning, and education, all in the service of a simple but powerful goal helping every learner reach their potential. We're so excited to be in dialogue here with you, Irina, welcome.

Thank you. I'm really excited to be here.

I thought we would just start with some really basic things to help people understand what you do. And so let me start with asking, is it fair to Say that you're both a learning scientist and a technologist. Is that how you think of yourself and will you explain to us what a research engineer does or is and help us to understand sort of your team and what you do?

Of course. So I actually don't think of myself as a learning scientist. I would say maybe I'm a beginner learning scientist. I'm definitely just starting to learn about this field. But I'm very lucky to be working in a company where we do have learning scientists on the team and we also work very closely with teachers. So we actually just hired a teacher on the team and there is another teacher who is consulting us and we work closely with the academic field as well. So Kemp, Codinger and others are advising us, which we're very privileged to be in a position to have such amazing advisors. My role in education is relatively recent. I only started this project around three years ago.

You know, we hear the term research engineer and you're a research lead. What does that mean? You know, I think a lot of us are accustomed to the terms, you know, software engineer, but in the age of I AI now we hear this term research engineer. So I'm wondering if you can help us understand.

Of course. So I work at Google, DeepMind, right. And DeepMind has always been effectively an academic lab. So When I joined 10 years ago, it was a very small group, it was incredibly academic. So I joined as a research scientist and essentially my job was to do a foundational AI research and publish papers. That's kind of where we are coming from. And now we kind of much more integrated within Google, but we continue on the same mission. So what DeepMind brings to Google is this research expertise. So on my team we have scientists and engineers, but really like the. The line between them is. Is blurred. And what our job is to really think about what are the fundamental scientific problems around language models and in our case, on intersection with education, where we really need to do this foundational scientific work to understand what are the big problems? How do we find tractable solutions and also work out the solutions to these kind of big scientific problems?

So I guess one question I know, Michael, that has been coming up for you and some of the conversations you've been having having is do you engage with or interact with or directly influence the products at Google? So many of us in education are so familiar with so many Google products and what is the intersection of your work and for example, Google Classroom or many of the other products that we in the education field use.

So I find it Very exciting to be in a company as big as Google, where there are so many amazing teams doing incredible work. We really focus on the research because that's the value add that, you know, my team can bring. We do work closely, obviously, with the products because they build on top of the foundational models, in our case, Gemini. We talk, we advise, we help explain kind of what Gemini is capable of, how to best elicit the kind of more pedagogical capabilities out of it. But of course, the teams are amazing and they mostly bring kind of work on the products in their own teams.

Got it. That makes sense. It's been just about three years since most of us in the world were first really introduced to AI via the first release of, or not the first, the one that we're familiar with of ChatGPT. And you shared with me in previous conversations that the arc of your team's work over those three years has been really interesting. You just referenced the three years again. Tell us what you've been working on and what's emerging from that.

Yeah, we actually started the project about six months before the ChatGPT moment. So, yeah, things have definitely changed for us. Something stayed the same. So from the beginning we saw that the biggest impact we could make, or Genai, the newly emerging gen AI we thought could make, was through AI tutoring. There are. I can go into details, like why we think it's the most impactful thing, but maybe we can address that later. But when we started, language models were very different. Yeah. So I remember doing our first demo about six months after we started, and it was very hard to keep the AI on track. So we had to practice a lot, find the right kind of queries to ask. And even then, you know, it was always. We had to be at the kind of tip of our seat just to make sure that it doesn't go off the rails when doing a demo. And of course, now the AI is so much more powerful. You know, we have launched now products like Guided Learning on Gemini app, which millions of users are already engaging with, and it's mostly staying on track. You know, we haven't seen any major problems so far, so it's kind of just the technology itself has changed so, so much and we kind of had to keep up with these things. So when we first started, a lot of our work was trying to deal with kind of the very rough language models and make them do something useful in learning. And you will know how the stakes are so much higher in learning than in other use cases. So we really had to think, how do we control this kind of unruly beast beneath us? And now, of course, you know, a lot of that work essentially had to bin because it was no longer necessary. And we really concentrate on how do we bring the layer of pedagogy and adherence to learning science principles into Gemini to make sure that it really works towards increasing learning outcomes rather than negating them.

Irina, I'd love to jump in there because at first I think it's fascinating that you guys do this much foundational research because we always hear that sort of the domain of the universities. But here's DeepMind and then Google after the acquisition, right. Investing in over a decade of foundational research. There's nothing near term about that. And I'm interested in this work on the tutoring because a lot of the sort of critics, I guess I would say, of the AI tutors are sort of one of two approaches. Either, oh, who's actually going to use that? Why won't they just default to Gemini straight on? Right. And two, even when they do use it, it's maybe overly procedural is the critique. I hear a lot. And so I'm sort of curious, what are you learning about the actual usage in the wild? What are the guardrails that you've thought are important? What's been surpr against those critiques that are everywhere these days?

Yes, thank you for asking this. We are definitely very aware of the negative perceptions and maybe negative use cases of chatbots in learning. So I was just reading an article from the Guardian earlier today where they surveyed 8 to 18 year olds in the UK and what was interesting, I think just over 60% of first responders said that they perceived AI as being a negative addition to their learning journeys. And it's for all the reasons that we're already aware of. AI is just too keen to give away answers. It takes away the cognitive load. Well, not. It takes away the productive struggle. It leads to cognitive offloading of tasks. We know that that's not helpful for learning. So we saw this trend when we started this project because I think at the end of the day, AI is optimized to be an assistant, right? So it's successful when it takes away the burden from you as the user. And we know that in learning the opposite is true. You have to engage, you have to put in the struggle to actually see a difference to your learning outcomes. So what we realize is that if we don't do the foundational work and the research to make sure that AI can deal with these two very different Use cases, the kind of capabilities of AI in learning won't just emerge from all the other work.

It's so interesting. So stay with it then, like as you've been putting it in this very specific use case right around the tutor. I curious, like, why did you choose tutoring given that it is so different from the other LLMs. Right. Sort of that assistant purpose and how are you constraining it to make sure that it's the most useful tutor. Right. That a student could have access to as opposed to maybe its natural instincts based on its foundation.

So we chose the tutoring use case because the way we call it, it's kind of learning or education complete. So what that means is that in order to tutor well, you kind of need to know all the different type subtasks or capabilities that are important for education. So you need to be able to plan a lesson, you need to be able to ask good questions and provide good feedback and check the students work, among many other kind of things about metacognition and active engagement. So if we really manage to figure out the tutoring use case, then the resulting kind of underlying model Gemini can then be used for all these other tasks. A great way to bring one single goal to optimize for what can then result in broader benefits for learning.

Super interesting. Let me ask this question then. How does it integrate with Google Classroom? Because you all have this incredible install base effectively right across schools. I think probably in terms of K 12 schools, you'll correct me if I'm wrong, but I think the largest install base of sort of learning management system instances. So how does this tutor that you've built, how is it integrating with Google Classroom to actually directly serve students? And what are the guardrails? You're putting around that as well.

So in terms of guardrails, I just want to say that we really take this very seriously. There's a wide range of safety and guardrails work happening, happening across Google, DeepMind and Google at all levels. So from the model to the products and Gemini in itself has a lot of safety and kind of trust and safety work going there. What our team actually does is brings an educational and learning specific angle to this to Gemini model. As an example, when we try to optimize the model for tutoring, we kind of realized that a good tutor really engages the learner, they ask a lot of questions. So we brought this bias towards asking more questions to the model, but that resulted in an unintended consequence in the sense that not only does a tutor want to ask Questions. It also wants to encourage questions from the student. And then a student might ask a question that's actually harmful. So it could say something really toxic. You know, ask a question about that and what the chooser would do before we did the kind of the work to mitigate it. It would say, oh, that's an amazing question. I'm really glad you're thinking about this. And now. And then it will kind of bring it back and say, actually maybe there are other things to consider here. But that initial statement was just not helpful. So we had to then go in and kind of bring extra supervised data and kind of take that unintended behavior out of the model to make sure that's actually safe. This additional layer of work is really important. And of course then there's the product layers and other ways to kind of mitigate safety issues.

That's super interesting. I so appreciate the example. And I was going to ask you about, you know, what is this? What are you seeing now? Three years in, you know, you talked about at the beginning you had these sort of rough models and they would kind of, it sounded to me like get distracted and kind of, you know, go off. But, but three years in, it sounds like a, you're learning a lot of things and so you're iterating. So what is it, what is it looking like now and how are you feeling about the learning that's happening when young people are engaging or people are engaging with the products now and then maybe we can talk about where you think it's going as well.

The work that we've done is on the Gemini model side. What we hope that comes out of it is that Gemini is useful for learning products, both for Google, but also for external parties who build, like ad techs who build on top of Google for the internal products. Our team in particular really worked in collaboration with the Gemini app to bring the guided learning experience to users. We really wanted to bring an easy way for anyone out there to get kind of this more pedagogical behavior out of the models without having to engineer a very complex prompt. And so with guided learning, it's really a one click way to get the model to act more like a tutor so to guide you through the information rather than just give it to you as a wall of text. And we worked with learning science experts to make sure that this experience really adheres to the five kind of learning science principles that we have identified as important. Again, our hope is that this actually helps students internalize the information much better. And we are working Kind of very closely to try to measure the efficacy of how well that's actually coming together.

But.

But what I want to give you is a personal anecdote on how I ended up trying guided learning. And I was actually in Stanford a few months ago, and I saw this statue of the Burgess of Calais, which I've never heard of that story before. So I was curious to learn more. So first I kind of just pulled out my phone and used Gemini to ask about this historical event. And it just gave me kind of the standard answer of kind of a longish response. Think of it as like a Wikipedia kind of type answer. So I read through it. It was interesting, but I realized I'm actually personally really bad with history, in fact. So I realized that that information went into kind of into my head and immediately left it, and I didn't remember anything. About 10 minutes later, I was trying to tell the story to somebody else, and I realized I don't remember anything. So I again pulled out Gemini, but this time I switched on guided learning to see what. How different would be the experience. The difference is, like, guided learning doesn't just give you the answer. It kind of engages you in a dialogue. It brings you in over kind of maybe five to ten turns of conversation. It kind of walked me through the same information, but this time I realized actually that I remembered it like a week later. I could still remember the facts. I remember kind of the interesting things it brought in. It kind of brought the connection to the War of the Roses, which the first article didn't bring in, just because of how I selected kind of the options of what where my curiosity led me. To me, it was very visceral how I tried to kind of learn the same thing from the Gem, like a vanilla experience and guided learning. And one of them actually made me remember better without me even trying.

Interesting.

That's really cool. One really quick question on that. Like, what are the five learning science principles that you guys have prioritized to create that sort of experience just so we can enumerate it?

So the five learning science principles that we've identified is to inspire active learning, to manage cognitive load, to adapt to the learner, stimulate curiosity, and deepen metacognition. And we realized that this is not the comprehensive list. There are other important areas of learning science that we are considering to bring to kind of the forefront of what we're optimizing for. But these are the first and the most important five learning science principles that we have been working towards so far.

Irina, one of the Things that I like about talking to you about is that you talk about pedagogy and you said up front, you know, you had this hypothesis about tutors being sort of the way to go. I'm curious about that because we've also talked about there's other kind of ways to learn. And so I'm curious if you guys are exploring other ways and how you think about that and why tutors and, and yeah, anything you can share around that?

Yes, I'm actually curious to hear from you given your experience, what you think would be exciting other ways of learning for us to consider. The reason why we started with AI tutoring is because of, I guess this is where the strength of current gen AI model lies. It's kind of text chat base interface that we're all familiar with. So we thought okay, how can we leverage what's already mature to make a difference in education. But we also realized that as new capabilities in AI are emerging and also maturing, for example, we have these demos of live experience where it's kind of video and audio and you essentially can just talk to AI in the same way as you would talk to your human teacher. We are also thinking about how to kind of how to bring that to users in an interesting learning experience. But yeah, I would be very curious to hear from you what you think would be a good thing.

Well, I mean I think when I think about it and it's hard to really parse how different this might be from a tutor, but I think about this type of learning more in kind of the factual content vocabulary, the, the, the what I would call the content knowledge you need for learning. And then I often, you know, kind of crudely though separate from skill development. So how do you actually communicate effectively or write effectively or analyze problems? And you know, I historically have taken a pro project based or a problem based approach to that. So you start with kind of a big problem that you want to solve or a big question that you have and then you engage in, in a project that gets you an outcome or a product. And so that was pretty long winded. But maybe, maybe the most immediate would just be like maybe a tutor can do this but really helping to teach someone to write effectively or communicate effectively. I think right now at least I, and I think other people are using it to just take what I've written and write it better. But I'm not sure that it's really teaching me yet, giving me that guided practice and that feedback and whatnot. So that might be the more near term version that I'm thinking about.

Yeah. So first, for the skills. Skills acquisition, we really hope that guided learning type experience could actually help with that. So your example of helping you rewrite a. A piece of text with guided learning, it won't just rewrite for you. It will guide you towards how to rewrite it so that you do it. So it will ask you to think about certain things, ask you certain questions. So hopefully a student can learn from just that experience. Another thing I mentioned earlier, that kind of metacognitive abilities are important to us, to kind of make sure that the tutor optimizes those things as well. That's kind of another layer where hopefully students will kind of be able to take a step back and understand, okay, how did I get to this? How did I rewrite this? What was important? How did I think about it so that next time they can actually almost, like, guide themselves and won't need the tutor anymore.

That's so interesting. Last season, Michael and I interviewed the woman who leads the Harvard Writing center, and what you just described was her concern of what was not happening and what would be missed. And so it's interesting, the evolution, I think. I don't know, Michael, if you're.

If you're tracking the same thing. Yeah, I think. I mean, I think it's interesting, right? And it's all a question of. I think. And this may be where you want to go, Diane, like, how do we put this in the hands of teachers and students, right, in productive ways so that they're not just jumping to the shortcut, but actually engaging in the difficult learning that you all are creating these experiences for?

I think that might be a good place for us to go and sort of, you know, bring this conversation, at least for now, to a conclusion. So, Irina, you. You are a new mother. And I know that when I became a mom, it changed how I viewed my work as an educator in ways that I couldn't even have imagined. And so, you know, I'm curious, what, if anything, has changed for you in that. But. But even more so, what do you imagine your child's education will be like? You know, when you think about the next 5, 10, 15, 20 years, will it look the same? Will it look different? What do you want for it? What do you hope for it? You know, how do you think about that?

That is a great question. And I have this at the top of my mind. I think we are in a very unique situation, and kind of we're living through a very interesting period of time where the pace of change is so fast. I think even for us working in this industry, it's kind of head spinning and it's even hard for us to catch up with all the progress. It's very hard to predict where AI will be in five to 10 years, what the role of education will be. We are actively thinking about this. I think what's becoming clear is the importance of human connections and building, kind of making sure that our next generation grows up as complete humans so that they're not just automatons who, you know, provide prompts to AI and just kind of live in this AI driven world where AI really is still a tool that helps human flourishing and helps prove and increase human connections. So I think for my child, I would want him to still go to school and to still have experiences learning how to communicate with his peers, how to talk to his teachers and be inspired by his teachers. I hope that AI can be something that helps him maybe learn faster and learn more and kind of really personalize his learning so that when he's really passionate about something, he can go off and go deeper with AI and maybe, you know, be able to do these projects that are not supported at school, but he can do at home with his peers and AI can serve as kind of this facilitator and help them again achieve more interesting outcomes with their projects. But at the end of the day, I think I want him to have the breadth of experiences and knowledge and just learn how to be a good human.

That's a beautiful place, I think, to wrap.

This season of Class Disrupted is sponsored by Learner Studio, a non profit motivated by one question. What will young people need to be inspired and prepared to flourish in the age of AI as individuals, in careers and for civil thriving? Learner Studio is sponsoring this season on AI in Education because in this critical moment, we need more than just hype. We need authentic conversations asking the right questions from a place of reality, curiosity and learning. You can learn more about Learner Studio's mission and the innovators who inspire them at www.learnerstudio.org. i was going to say we have this section, Irina, where we wrap up, where we share something that we've been reading book wise or watching on TV or movie, podcast, whatever. And so I'll, because we didn't prep you beforehand, we'll let Diane go first with hers and then, uh, we'd love to hear what's on your, what's on your bedside table or in your ear or something like that, if you wouldn't mind, Sharon.

So, so this is a kind of a funny one. I'm. I'm listening to Slash, reading a book called the five types of wealth, a transformative guide to design your dream life by Sahil Bloom. And if you're wondering about the types of wealth according to Sahil, they are time, social, mental, physical, and financial. And I'm actually reading this with a group of other people who are sort of in our. Depending on who you talk to, last half, last quarter of life. And we're exploring this question this year of how do I do my quote, best work in these chapters? So this is one of many things that we're using as a prompt to sort of create a rubric for ourselves, if you will, and self evaluate to. And I'm reading it now in prep for our next get together, so. Fascinating. I'm not sure if I'm like a wholehearted recommendation on it, but, you know, it's kind of. It includes a lot of the ideas that I think exist in a lot of other places. And it's a good reminder.

Fair enough, Fair enough. Well, it'll get marked down either way. We'll track it. Irina, what about you?

So I will be honest. I am struggling to find time to read, given that I have a one year old. But I actually did manage to get through a book recently, and it was Neal Stevenson's novel Diamond Age. I've been recommended it many, many times, given the work I'm doing. So I finally managed to read it. And so just if you haven't heard about it, it's about this world of the future where somebody designs essentially an AI tutor. So it's kind of this book that is given to a young girl and the book essentially teaches her everything kind of throughout her life. And I think what's interesting, like my takeaway from this was that there were three kind of maybe original versions of the book that were given to kind of three girls. And then they kind of made, I guess, like a copy that was given to everyone which wasn't as good. And the difference was that in the three original versions, there was a human who was essentially voicing out the text to the girls. And in the other versions it was like 100% AI. And what was interesting is that the human behind the book, even though they were just voicing what the text that the AI was producing according to the book made a difference. Those three girls, especially the main character, who had this consistent one person who was guiding her throughout her whole life, actually built a connection with that person and grew up to be much more successful, much better individual than anyone else. And it's like this importance of still having a human in the loop.

Very cool. I love it. I've heard a ton about that book, so I need to add it to my list. I think now I'll just say I'm going to shock Diane here because we always make fun of me for not being current on stuff, but I actually not only did I watch seasons one and two of the Diplomat over the summer, season three came out and I'm already done, so I'm ahead. And so I am going to, you know, stay in my Netflix bingeing I guess at the moment. But I'm feeling rather impressed with myself and that I got my Google sweatshirt from back when I lived in Silicon Valley on for this recording. So with that, Irina, I think Diane and I could both talk to you all day and just like learn from this. So really appreciate you joining and scratching the surface with us of all the things going on DeepMind. And for all of you tuning in, we'll see you next time on Class Disrupted.