Transforming learning with AI tutors - Gustav Grimberg (Ezri)
Gustav Grimberg is the co-founder of Ezri, a company dedicated to developing AI tutors that transform online learning.
The full transcript of the podcast can be found below for those who prefer reading rather than listening. The podcast is hosted by Frank Albert Coates.
You can also subscribe to the newsletter summary of the podcast on our Substack!
The Guest
AI tutors will play a very important role in the future of education. What exactly they will look like and how they will work and how they will fit into the learning process are still questions that need to be addressed fully. But, at Ezri, we believe and we work towards a future where AI tutors will be a learning companion that all students will have.
Gustav Grimberg is the co-founder of Ezri, a company dedicated to developing AI tutors that transform online learning. With a background in human-computer interaction, Gustav has long been passionate about creating technologies that are natural and intuitive for human beings to interact with, especially in education. In October 2023, he co-founded Ezri to bring this vision of affordable, personalized education to life.
Recently, Ezri launched its first AI tutors in Germany through a white-label model in collaboration with the K-12 publishers Cornelsen and Duden Learnattack. Their machine learning systems converted tens of thousands of videos, exercises, and articles into interactive AI tutors, offering personalized student support tailored to the content.
The company is now focused on releasing its next generation of AI tutors, which work in multiple modalities to deliver even more effective personal tutoring.
Company profile
Segment: K-12 (to start with)
Business model: B2B
Geo: Global
Year started: 2023
Funding: Bootstrapped
Location: Paris, France
The view of the Garage
With rapid advancements in artificial intelligence and machine learning, the potential for AI tutors to become integral to the learning process is immense. Will AI tutors become indispensable learning companions for all students? There are a few challenges left to tackle (outside of obvious the Large Language Models issues with hallucination and bias):
1. Accessibility: AI tutors have the potential to bridge the gap when there is no teacher or tutor available to support students. However, the current interface and user experience set a huge barrier to general accessibility and usage (as discussed in the episode with Nick from 360Learning). The arrival of multimodality and particularly the ability to dialogue through natural voice, will open up the market as long as quality and relevance is high.
2. Personalized learning: One of the greatest strengths of AI tutors is their ability to offer personalized learning experiences. These tutors can adapt to individual student needs, learning speeds, and styles, thereby making education more effective and engaging. The current integrations of AI tutors depend on having the context, history and overall understanding of the user for personalization which depends on getting as much relevant user data as possible. For publishers, schools and education they might want to share only limited amount of user data with external providers for both competition reasons and for…
3. Data privacy and security: Robust data privacy and security measures must be in place to protect student information and build trust. Especially in K-12 this data is sensitive and any breach will have serious consequences both for learners, schools and external providers. How can startups provide such guarantees and build trust to ultimately provide better personalization?
Extracts
Human tutors vs AI tutors
So how do you compare an AI tutor to a human tutor? There are similarities and there are differences, of course. And I think that going forward there will be ways that these two modes of learning will complement each other.
So in similarities between AI tutors and human tutors, there's the active learning part, there's understanding the gaps in knowledge, there's promoting critical thinking, promoting thinking about your thought processes. Human tutors are human and there's definitely something to explore about how exactly and if it can be replicated in AI tutors.
On the other hand, AI tutors, they are very scalable. They are very affordable. They can be made available at a fraction of the price of a human tutor and they're available 24/7.
So no matter how you prefer to learn, no matter how long time it takes, you don't have to worry about a human sitting there being impatient. You don't have to worry about saying something stupid because they won't judge you, the agents.
Developing personal relationships
An important feature of human tutoring is a mutual understanding and feeling some kind of harmony or some kind of connection with the tutor. And this is something that is also important to explore in AI tutoring systems. If you should make this or try to develop this interpersonal connection with the tutee, with the student.
From what we see, it is definitely a path to go down to try and develop these more close relationships where the tutor has a better understanding of the student, has an ability to not just tutor, but also, go off topic sometimes to develop a more close connection to the student because that in the end can foster and promote better learning.
The interface and user experience
Anything that keeps you up at night? It is thinking about the interface, how learners going forward should interact with AI systems and language models. So what happens from the user either uploads their homework or ask the first question to they at some point end up at having understood the content that they didn't understand before. How to deliver the content or how to deliver the messages, voice, real time, asynchronous. How long the interaction should be, how to plan a good tutoring session automatically is something that keeps me up at night.
And something that I've spent a lot of time thinking about is how to make good dialogue management systems for tutoring, which is very much where we see that large language models right now they lack some quality. They're good at delivering text, but they're not so good at developing, delivering tutoring sessions and bridging this gap to make large language models good at tutoring is a technical challenge, for sure, that we spend a lot of time thinking about right now.
Dynamic vs static tutors
I guess it was interesting to discuss and think about what an AI tutor is exactly. So whether an AI tutor, is only this kind of tutor that we're developing at Ezri, where it's like completely dynamic and interactive or whether it can also be a bit more static, you know, adapting your learning experience, suggesting content, even suggesting questions that can have predefined answers and so on.
If we also allow that category as a tutor, there's a lot of work being done. Both in Europe and in America. For the kind of smaller subset of AI tutors that we are working on that are these very dynamic and personalized tutoring interactions, there haven't been that many AI tutors released now that are used on a big scale in Europe.
And we are seeing interest in this. We're seeing that there are established players that are interested in going into this, both publishers, schools and private tutoring companies. And we definitely expect that this will be the path forward at Ezri.
Voice system and dialogue management
I would love to have a really good voice system that works very well, especially with K-12 students. That delivers a very human like voice interaction. And then just find this dialogue manager that can really just steer the interaction as the best tutor in the whole world. So if I could order one thing and plug into our tutors, it would definitely be the best tutor dialogue management system, which we are working towards at Ezri.
Full transcript
FA: So, we're going to talk to you a lot about AI today and tutors. Could you start with the 30 second pitch of what Ezri is?
GG: Yeah, definitely. So, at Ezri we develop AI tutors and the infrastructure to power them at scale.
That among other things can be integrated into online learning platforms that have content libraries of videos, exercises and articles. We automatically process all this content, turn them into interactive AI tutors that can then be reintegrated into the online learning platforms to deliver personalized learning as the students go through the material of the platform.
FA: Very cool. So let's take a step back and then talk a bit about how you got into this industry of EdTech and why you started up the company.
GG: Yeah, definitely. I co founded Ezri with Leon Mächler, who is a computer scientist and a deep learning expert. And myself, I have a background in human computer interaction, in particular in education.
Throughout my graduate studies, I did research on different things, and I was very lucky to end up in a research lab at INRIA, which is the National French Institute of Computer Science Research here in Paris and other locations in France too. I did research with Justine Cassell and Claire Clavel, and they have been pioneering effective computing, which is research into how to make computers that can understand interpersonal relations and human computer interaction so how to develop good technologies for human computer interaction.
And Justine Cassell, in particular, was interested in how you can develop rapport in educational context with computers. So how you can develop this harmonious interaction between a computer and a learner that can resemble the harmonious interaction that appears when two human beings, they teach each other something.
And this was very interesting research. I got very interested in how to use technology for good and in particular in education. And then Leon and I, we then over the past years explored a bit how to develop learning technologies with language technologies.
And then, you know, a lot of things happened over the past two years in particular language technologies a lot of people got interested in. And in education, there were big questions and big discussions around how this would, you know, impact learning now and in the coming decades. And we saw that as a good time for us to, you know, introduce some of the things that we've been working on to get feedback from teachers, from educational institutions.
And we started sending out these AI tutors that we have been developing. First to teachers and then to publishers and other established education institutions. And we got a lot of good feedback, which eventually led into this partnership that we established in Germany with Cornelsen and Learnattack, that is an online learning platform, where we introduced and launched our first AI tutors earlier this year. And we're now, yeah, you know, a company operating and selling our AI tutors all over Germany.
FA: Fantastic. So lots of things to dive into. But just first sort of the basics. So what size are you at and, you know, where are you at in terms of funding? If you want to give some of the sort of data of where you're at.
GG: So we are at the moment, three people here in Paris. We haven't raised any funding yet. We can work with what we have now and the tutors that we're selling, and we continue to iterate quickly and see how exactly these AI tutors will work. And then we will take, you know these considerations as they come.
But for now we are self funded and bootstrapped.
FA: So, one of the questions I always ask is sort of why will you succeed? What's your secret sauce, if you want to start with that perhaps?
GG: So AI tutors are, you know, a completely new category in online learning. There are some offers out there, but it's still a field that is very early in its days and where there's a lot of stuff to learn. At Ezri, we had some very quick learnings from launching these AI tutors quickly and getting a lot of feedback.
And what really makes our tutors special are two things. It's, you know, deep integration into the online learning platforms that the students already find themselves in. So they don't have to navigate necessarily to completely new sites. They are already in the platform, watching the videos, solving the exercises.
And as they're doing that, they can get personalized support on the side. And we have done a lot of things around user experience and interaction optimization, where we have custom models that do dialogue management, for instance. So we quickly saw that language models, they can do a lot, but there's also a lot they cannot do.
When we develop our AI tutors, we like to approach or get inspiration from human tutoring. And what you often see is that language models that take a very passive tutoring role when you just ask them to, you know, teach you something. Whereas human tutors, they can be a bit more you know, encouraging, help you understand the gaps in your knowledge.
So at Ezri, we try to approach or approximate how human tutors, they work when they are tutoring. And what you see is that human tutors they take a much more active role in the learning experience. And behind this active role is dialogue management to figure out what to say when, how to say it, how long the answers should be, how short they should be. And for this, we have custom dialogue management models that guide the student from initial misunderstanding or initial confusion to the end goal that is learning the content that is in the curriculum.
FA: The big, big question is, are AI tutors actually, are they providing learning? Are they providing, you know, the same amount of teaching you can say? How do you see that in terms of how far we've come with AI tutors today and where they want to go? You have what you have today, but you know, how effective is it as a learning tool?
GG: So if they provide learning, I think the short answer is yes, that we see that from the students that interact with it and the feedback we get that students, they definitely learn things and they learn it in ways that weren't possible before.
So they're sitting alone at home and they're watching these videos and if they hadn't, you know, had a parent next to them before, they couldn't really ask any questions. Now we see that, you know, if there is a misunderstanding or if there is a confusion about, for instance, the content of a maths video, they can now quickly ask a question about the content of the video and get an answer based on the content of the video back.
So the short answer is yes, it provides learning. How it compares to existing learning options and so on is a harder question. So how do you compare an AI tutor to a human tutor? There are similarities and there are differences, of course. And I think that going forward there will be ways that these two modes of learning will complement each other definitely.
So in similarities between AI tutors and human tutors, there's the active learning part, there's understanding the gaps in knowledge, there's promoting critical thinking, promoting thinking about your thought processes. Human tutors are human and you know, there's definitely something to explore about how exactly and if it can be replicated in AI tutors.
On the other hand, AI tutors, they are very scalable. They are very affordable. They can be made available at a fraction of the price of a human tutor and they're available 24/7. So, you know, no matter when you learn or how I prefer to learn, you can access one of, you know, our AI tutors and you can get personalized and immediate help and they're very patient.
So no matter how you prefer to learn, no matter how long time it takes, you don't have to worry about a human sitting there being impatient. You don't have to worry about saying something stupid because you know, they, they won't judge you, the agents.
FA: Yeah, that's a really good point. And then you were talking about personalization. So how personalized are they and how much do they know about the learner so they can adapt and then propose the right level of interaction with the learner?
GG: So they can know how much information you give them. So useful information would be the grade they are in. So if they are fifth, sixth, seventh grade, what topics they're studying, what topics they have studied before.
So, you know, sometimes these learning platforms, they have a tree structure of content. So this content is content for the content you're currently looking at. And after that, so if you understand the progression through the content, you can deliver a tutoring interaction that will build on the content that was before and lead to the content that is coming up. If you can see that the student is from 6th grade or 7th grade, you can, or you know, even 4th grade, you can deliver responses that are very easy to understand, that are very short or longer for the older students and a bit more complicated for the older students.
There's also knowing, you know, where you're from in the world. So whether you are a German student or Danish student and how the curriculum feeds into this. But often, yeah, I said these learning platforms, they have this codified, the curriculum so you can understand what grade the student are in, what they've looked at, what they will look at going forward, and then customize the interaction.
FA: So you have the history of what they've done already, to some extent.
GG: You have sort of the progress of the student through the learning platform, and you have previous interaction too, sometimes depending on anonymization and prioritization of the data. So in education, and especially when you have to do with younger students, there's obviously a lot of data that is complicated to store and where we at Ezri actually do not store all the data.
So this is some data that we stored on the learning platform. And then we provide mostly the AI tutoring interaction where we get anonymized questions from the students and some anonymous metadata, such as the grade that they're in, what topic they start studying, potential progress that they've done, if they have, you know, studied a lot of other things, or if they are a new student on the platform, and then basically that we can send a response back.
FA: Okay. So you, you wouldn't have the data directly if like in the last interaction, the students, you know, didn't complete this or this well. So it depends on the platform you are interacting with.
GG: Yeah, so the platform can either choose or not, to send this data. And if they send this emitted data to us, this is something that we will integrate into the interaction too.
And then deliver this personalized over longer interactions. So either you can have personalization in one interaction, and that is if you asked a question previously, you know, it will obviously know what questions you've asked in that interaction.
And it's customization across interactions is something that is definitely interesting to explore. So how to develop long term relationships with the students and also something that we're exploring at Ezri right now. But we're still in the early stages of these AI tools and there's a lot of stuff to be done around, for instance, personalization across interactions and over time, that we are, you know, looking into and that we're working on making good technical solutions for.
FA: One question around the actual human interaction is, so today in the AI tutors, you mostly converse written, and you need like a keyboard to do that. How do you see the advancement of multi modality and what will come in the next years to the way that students and other learners will interact with these systems?
GG: There are a lot of open and interesting questions around the user experience and how to exactly design the interface. And this is something that we spend a lot of time on right now at Ezri, you know, thinking about how exactly to deliver the best user experience and the best user interaction.
Right now, Ezri is text based, so the students interact with Ezri by writing messages and reading messages from Ezri. I think this is a good start, but it has its limitations, you know, cognitively sometimes, especially for younger students it can be a lot to read long messages.
So in the next iteration of our AI tutors, we are integrating a lot more voice support. So sending voice messages, getting voice messages back. And we're experimenting with what the best approach should be, whether it should be real time voice, so you interact actually as if you would be talking to a human being. You know, real time answers from the AI tutor that is fast and tailored to exactly the questions you're asking, or whether you can deliver asynchronous voice messages. So, you know, it takes a minute or 30 minutes to deliver a really well thought through response to the student's question, similar to having an interaction on WhatsApp, for instance, you know, so you send a message, you get a long message back that you can then listen to.
I think we would definitely move beyond text based interfaces. That was very important for making people understand how powerful these AI systems are. People are also getting used to, you know, even more advanced interactions with computers. And, you know, a lot of people are doing some very interesting stuff around voice technology even. Image technology, how to make avatars that are tutors.
And this is also the direction that we at Ezri will move into. So you have a voice based interactions with our tutors still tailored to the content of the learning platforms. You know when you watch a video, for instance, you can just ask a question as you would have, you know, your older sibling or your parent next to you would ask, ask a question in natural language, just with voice and you will receive a voice response back from Ezri tailored to the video watching or the exercise you're solving.
FA: And then if we look at sort of the, the market today for AI tutors, how do you see that evolving with, the first that came out on the market like Khan Academy having their tutors, which sort of sat some of the standards maybe in the industry first a year ago, how do you see the market evolving and sort of competitors and the overall market space?
GG: There's a lot of interest in AI in EdTech and also in AI tutors. So we mostly work right now in K-12 EdTech and K-12 AI tutors. And there have been, as you mentioned with Khan Academy and Khanmigo, some interesting work.
I guess it was like interesting to discuss and think about what an AI tutor is exactly. So whether an AI tutor, is only this kind of tutor that we're developing at Ezri where it's like completely dynamic and interactive or whether it can also be a bit more static, you know, adapting your learning experience, suggesting content, even suggesting questions that can have predefined answers and so on.
If we also allow that category as a tutor, there's a lot of work being done. Both in Europe and in America. For the kind of smaller subset of AI tutors that we are working on that are these very dynamic and personalized tutoring interactions, there haven't been that many AI tutors released now that are used on a big scale in Europe.
And we are seeing interest in this. We're seeing that there are established players that are interested in going into this, both publishers and schools and private tutoring companies. And we definitely expect that this will be the path forward at Ezri. I mean, obviously, since we're working on this full time, we believe that in some point in the future, all kids can and probably will have an AI tutor that they interact with at home to some extent. And, we are quite happy to see that a lot of other players, both established publishers that are big in the education space in Europe, but also smaller startups, you know, share this belief, and also our early customers that are using our AI tutors daily. And from what we've already seen about how it can help even the simple or the, you know the first version that we have released now, we are very optimistic about AI tutors going forward.
FA: So if we look at, what you've done since you started talking about sort of what has been the challenges? What's been your biggest challenge since you created the company?
GG: I think education is, and no one will say otherwise, is a notoriously hard industry to get into. And rightfully so, it's a very important industry. There are a lot of risks associated with releasing educational technologies and so on. I think we've spent a lot of time in the beginning, thinking about how to actually, you know, market these AI tutors, whether it should be direct to schools or they should be direct to customers or should be directly to the publishers and then establish some partnerships with established educational tech companies and then distribute these AI tutors with them.
We quite quickly learned that the path forward probably wouldn't be us going directly to the schools and sell the AI tutors. That it can be challenging for young startups also to sell directly to the consumers if you don't really have the brand name. At some point in the process, we decided to go with this path forward that we would establish, or you know, partner with established EdTech companies.
But yeah, exactly figuring out what our go to market approach should be was something that was challenging in the beginning. And now we have to sort of steadily and consistently work towards this change in consumer behavior also in tutoring where people have to get accustomed and used to a completely new category of tutoring.
And this will also take a bit of time. Which we are, you know, happy with, with waiting, but coming out of academia and coming out of a university where things say, you know, can go very quickly and you can launch a product in one week, then seeing how some of these product development cycles could work in EdTech. I think, you know, our mindset, was a bit challenged there, but it's also a good learning experience. And we are happy to see that we can affect some, you know, consumer thinking in AI and tutoring.
FA: I guess the most important is that you already have users and you already have, you get that feedback, right?
So you can iterate. Definitely. Also since you started the company, any particular sort of resources that you found very, very helpful and then it can be sort of anything that you've experienced that's really helped you push the company and you're thinking further.
GG: There are, you know, typical start up resources for tech companies that have been nice. So we've been following tech founders on Twitter and LinkedIn and, you know, a lot of advice about getting a product out there quickly, get a lot of feedback and iterate quickly. That's something that we have learned from previous tech companies.
And this is not so EdTech related necessarily. We have been lucky to work with prominent EdTech companies in Europe. And Cornelsen, for instance, a publisher in Germany, is a very esteemed and big publishing house with a long history and they have, from working with them, we have seen how the education space can work and how you, you launch products in education.
And that has been you know, useful for us too, to work with domain experts in education. So one thing is that we develop these AI tutors and that we are technical, we know how to do it on, you know, on the machine learning and AI side. Another thing is also understanding that they fit into some larger educational context where there has to be quality assured also educationally.
And there we've benefited from feedback from learning designers at Cornelsen that have been thoroughly testing our AI tools just before they get released to students and getting feedback from them on what they think a good learning experience should look like.
FA: So a lot of support both from the, what do you say, the clients and the ecosystem.
Okay. If you're looking forward to the next six to 12 months anything that keeps you up at night?
GG: It is thinking about the interface, how learners going forward should interact with AI systems and language models. So what happens from the user either uploads their homework or ask the first question to they at some point end up at having understood the content that they didn't understand before.
How to deliver the content or how to deliver the messages, voice, real time, asynchronous, as we discussed earlier too. How long the interaction should be, how much, you know, how to plan a good tutoring session automatically is something that keeps me up at night and, you know, something that I've spent a lot of time thinking about how to make good dialogue management systems for tutoring, which is very much where we see that large language models right now they lack some quality.
They're good at delivering text, but they're not so good at developing, delivering tutoring sessions and bridging this gap to make large language models good at tutoring it's a technical challenge, for sure, that we spend a lot of time thinking about right now.
FA: Hmm. And that resonates with what OpenAI announced with their new model a few weeks ago. That their new model 4o has sort of a possibility to feel or understand the emotions. So when you speak at a certain pitch, at a certain volume, they can actually do that and sing and, you know, all things related to that. So how, how do you see that sort of possible understanding of emotions in these models?
GG: So understanding interpersonal relations is something that is important for tutoring. And then it's about what you exactly define as emotion and what you define as interpersonal phenomena. But going back to the research I did before co founding this company, an important feature of human tutoring is, you know, a mutual understanding and feeling some kind of harmony or some kind of connection with the tutor. And this is something that is also important to definitely explore in AI tutoring systems. If you should make this or try to develop this interpersonal connection with the tutee, with the student. And if that is the path forward, if it is important to develop these, you know, interpersonal connections, how to exactly do so.
From what we see, it is definitely a path to go down to try and develop these more close relationships where the tutor has a better understanding of the student, has an ability to not just tutor, but also, go off topic sometimes to, you know, develop, a more close connection to the student because that in the end can foster and promote better learning.
FA: And that also makes me think of the developments around smaller models that can, you know, live on your phone or on a device and then do a first sort of instance to interpret the data or the interaction. How do you see that evolving as well?
GG: There is a big privacy consideration in education, and especially in Europe, where using models from big American tech companies is not always an option, especially not for the state markets. And there, smaller, self hosted models, either on European servers or on edge devices such as computers or phones, is a viable path forward, and it's something that we are working on at Ezri too.
To find the right mix between using self automated models to handle private or, you know, potentially sensitive parts of the interaction and combining them with bigger models that are more capable at maths, for instance, in the background, to balance, you know, good human tutoring with an essential consideration for privacy.
FA: And then another question, if you had like a magic wand today and you could, you know, change the product however you want it. What would you do?
GG: So that's a good question. Also because what you actually see now as a product is pretty different from what we have coming up. So, you know, we have, since we launched, I said, you know, done a lot of things. I would love to have a really good voice system that works very well, especially with K-12 students.
That delivers, you know, a very human like voice interaction. And then just find this dialogue manager that can really just steer the interaction as the best tutor in the whole world. So if I could, you know, order one thing and plug into our tutors, it would definitely be the best tutor dialogue management system, which we are working towards at Ezri and we are doing good things towards, but, you know, instead of having it, whenever you reach it, having it tomorrow would be very useful.
FA: Okay, good, good to put on the wishlist. For other startup founders in EdTech, is there any specific advice that you would give them when they're just getting off the ground and creating their company and service?
GG: I'm a young startup founder myself. I think one of the things that I had to learn and we had to learn was that, things did naturally take some time in education and just stay focused. Remember why you chose education. I think it's the best field to work in. I think it's one of the most meaningful fields to work in. And remembering why exactly it should develop education technologies is very useful when sometimes it can take, you know, a bit longer to get user feedback or to launch products.
And then, you know, the standard tech advice for startup founders, you know, just try and get this user feedback very quickly. Launch something that you can iterate on quickly. And then, I mean, don't be afraid of walking away from, you know, bad ideas. Sometimes you work on something that is a bad idea and, you know, realizing that, you should have, you should choose another path is very useful. And, you know, don't put yourself over the product. Just let the product and the users speak, the language and then just work on getting the users in the product closer and really make them work well together.
FA: Good ones, good ones. Anything else that you want to share that you haven't shared with our audience yet?
GG: I think just maybe summarizing a bit what I've said so far, or just reiterating some of my points that I believe that AI tutors will play a very important role in the future of education. What exactly they will look like and how they will work and how they will fit into the learning process are still questions that need to be addressed fully.
But, at Ezri, we believe and we work towards a future where AI tutors will be a learning companion that all students will have. And, yeah, we are very excited to be part of this big shift in education.
