An Interview with Korbit’s CEO, Iulian Serban (Part 2/2)

May 24, 2019 — In this second half of the interview, Iulian talks about the technology behind the company and the new course on machine learning taught by Korbi, the new state-of-the-art AI-tutor.

May 24, 2019 — In this second half of the interview, Iulian talks about the technology behind the company and the new course on machine learning taught by Korbi, the new state-of-the-art AI-tutor.

Part 2 – Interview with Iulian, CEO of Korbit

This is the second part of the interview with the CEO. In case you missed the first part, you can find it here. In this second part of the interview, Korbit’s CEO, Iulian Serban, discusses the more technical aspects of Korbit’s AI tutor, Korbi, and sheds some light on the recently launched course, created in collaboration with Mila, the Quebec Artificial Intelligence Institute: a Free Machine Learning course taught by an AI-tutor with lectures from World Renowned Professors.

You can sign up here → korbit.ai/machinelearning

Mihai, CMO (on the left) and Iulian, CEO/CTO (on the right) at Korbit’s Launch Party


“I think a lot of the people who are new in the field, tend to dive straight into programming instead of trying to understand what’s going on and when it breaks [they] don’t know how to fix it because [they] don’t know how it works.” Iulian Serban, CEO of Korbit

One-on-one Chat with Iulian – Part 2

Mihai: Allright, so let’s move on to the second section, which is essentially the technology that powers Korbi, the AI-tutor. Can you tell us a bit about it, maybe not dive too much into the details because I know how much you love talking about it and we don’t have that much time. Maybe just give us an overview of the technology that powers this amazing system.

Iulian:  From my background at the Mila lab in machine learning, obviously there are lots of machine learning models in there. We use deep learning, we use some very simple form of reinforcement learning and many of these algorithms. It’s really a complicated piece of machinery. Whenever you write something, Korbi will analyze what you’re writing, will build up a student profile of how you learn, what concepts you understood and what concepts you haven’t understood. Then it will go on Wikipedia and will scrape lots of articles and try to find the right content, the right thing for you. So there are really many different models in there and they are really complicated to build. We spent over two years working on this already, as you know. And they all go together and behind the scenes, the system is always trying to optimize for you as a student. How can I help students succeed to learn this content in the best way possible and engage them so that they don’t give up? Those are the two things we’re trying to achieve. The system is completely optimized for student learning and engagement.


“Our vision is basically to transform education. You’re going to see more and more blended learning, flipped classrooms and teachers working on projects rather than reciting content.” Iulian Serban, CEO of Korbit

Mihai: So we touched a bit about the technology behind Korbi. Let’s talk about the scalability aspect. I know there are a lot of machine learning software engineers that are building their own projects using theory. Everyone starts off with theory. Can you tell us about how you go from theory to actually creating a live product, a real-world product?

Iulian: Coming from research, one of the big obstacles was moving from a single isolated model to building a full fledged system. As a researcher, you spend a lot of time working on a very particular problem. You work years on that problem and try to make progress. In my case it was generating text. How do we generate natural language text on the fly? Which is now a big component of Korbi. But moving into this mindset of we’re not building one component, we’re building a full system and this full system isn’t just working in isolation. It’s working in a world with lots of students, teachers, the internet and Wikipedia is changing every single day and the system has to adapt to all these things. Students coming in with random queries, Wikipedia changing, teachers creating courses. And so really getting into a different mindset of systems thinking, I think is one of the challenges for me and maybe also some of the other people on our team. What you realize is that it’s not really the most complicated model that wins. It’s sometimes the union of those things that work together. Two simple models that do their job very well could do better than a complicated model that does it even better, but which might break in other places. It’s about taking a simple approach on a system-wide level. I have a really theoretical background and I think some engineers miss some theory. People tend to dive into programming. “Let’s program a neural network.” Yea, you can do that. You can do it in an hour if you’re very good or do it in a day if you’re just starting. But what have you learned if you just run a bunch of code lines? You have to actually understand what is happening underneath it. I think a lot of the people who are new in the field, they tend to dive straight into programming instead of trying to understand what’s going on and when it breaks you don’t know how to fix it because you don’t know how it works. And if it doesn’t work right away, you don’t know how to improve it. It’s like you were treating it as a black box. I think this is one of the big problems in general in machine learning now. A lot of people are coming in, so I think this is one of the advantages of studying our course. It’s that you get really deep in it. Obviously taking programming tutorials as well and doing our course, you get a deeper level of understanding machine learning. You’ll be able to build new models and solve new problems.

Mihai: Does that mean students can solve actual coding exercises as well on the platform? Is that something we want to explore in the future?

Iulian: We’re exploring building programming activities, it’s not something you can do easily but there are lots of programming resources out there that people can take already which are great. But what is really hard for most people is to learn the theory of why things work the way they do. And that’s really the main advantage of student’s understanding this will know why it actually works.

Mihai: We touched a bit on the technology and the scalability aspect of taking something from research and bring it into the real world. I think it’s a good segway into the next section which is this new course that we’ve just launched on machine learning. Can you tell us about how you worked with the Mila professors and put this course together. What can people expect after taking this course?

Iulian: It’s a course with four modules, an introduction to machine learning. So basically, anybody who has a linear algebra or a stats background, if you just took your first linear algebra course, you can start learning machine learning right away. The course starts with this introduction to machine learning that builds on linear algebra and stats, this is taught by Audrey Durand, soon to be a professor at Laval university. The second lesson is given by Yoshua Bengio himself, introducing neural networks, essentially the building blocks. What is a neural network, what is a unit, what are activation functions. The third lecture is given by Laurent Charlin, and it’s about training neural networks. Now you have a neural network, how do you take some data and make it work. He has a lot of advice, practical advice we can use and he shows a couple of resources online that you can use on your own too. The last lecture is given by myself actually and Aaron Courville, another professor at Mila, which is about convolutional networks and recurrent neural networks which are two very popular types of networks for processing image data and video data, audio data and text. They work really well on a lot of real-world problems so the students who really want to use machine learning for real-world problems should study that too. If you finish this whole course and you do a couple of programming exercises you should be in a good position to start your career.

Mihai: Even before joining Korbit, in the last few years, I kept hearing about artificial intelligence and it was very intimidating because I’m not a math guy and I really saw this AI thing as a black box and something that was way too advanced for me. So this course could potentially be useful for people like me. As a side note, I’m actually enrolled in the course [laughs]. With Korbi’s help and the interactivity, I’m sure it’s going to help build confidence and understand the fundamentals of machine learning.

Iulian: Yeah, Korbi also tries to figure out if there are any prerequisites you are missing. If there are concepts you haven’t learned or forgotten, she tries to target that and give you hints, show you diagrams and explanations. I did want to mention this is really day one for us. We just launched the course. Korbi gets better with every student, but you have to bare in mind that she’s also learning. She’s also a student, right. Over time, she gets better and better. Just think about where Google was on day one. Well it wasn’t what it is today. It’s very good today, but that’s after millions of people have spent years using it and it has learned a lot. Just keep that in mind as you learn with her. You try it today and she says some good things and some bad stuff and if you try it next week, she’s learned a lot and she’ll keep learning all the time.

Mihai: This course, if I understand correctly, is going to keep improving? The more students take it, the more it will improve.

Iulian: And hopefully, the more the course improves, the more students want to enroll in it and tell their friends to take it. It’s like a Korbit cycle [laughs].

Mihai: What advice would you give somebody trying to get into machine learning? Is this course the first resource that somebody should use when getting into this industry? What about if they already have some experience already?

Iulian: I think, definitely, if you’re new, start with our course. It’s going to help you even if you have knowledge gaps and forgotten some concepts. This is a great way to start. Then you can study programming tutorials. If you’re already a veteran, you’ve done some machine learning, but you want to brush up on theory, you can also take our course and skip forward to the hard parts. Or maybe there’s some stuff you thought you knew: “I thought I knew linear regression, but Korbi’s teaching me in a new way and giving me a question about linear regression I can’t solve.” You might discover you can learn new stuff about what you thought you knew already because of the knowledge illusion that we mentioned earlier. You take online courses, you think you’ve understood it, but you haven’t really. Now korbi can fill in those gaps. I think this course is for beginners and veterans, they can both benefit from it.

Mihai: I really love that point on reviewing the material. Because, as you mentioned, even though you might think you know it, there’s always room for improvement. Maybe we can go back to the beginning and talk about our experience with online learning. Where do you see the future of online learning? Where do you see Korbit fit into this new vision for education?

Iulian: As I already mentioned earlier, we see a world where people are studying more and more online on these platforms. Millions of students around the world, especially people who, before couldn’t afford it or have access to this kind of education. We see it also in changing classrooms and teachers. In a sense, I think the teacher’s role should move away from just repeating the textbook and should move towards helping the students one-on-one, doing group projects and helping with team dynamics. There are a lot of skills a computer can’t teach but definitely teaching basic concepts and reciting content is something that an AI-powered tutor can do. Once an AI-powered tutor has taught a million students the same concept, it has a pretty good idea of how to teach the next million and one student. In a way that’s even optimal to that student and most teachers, if they have a classroom of 30 students, just don’t get to spend that time that Korbi does. In a way it can do things very effectively in some cases. Our vision is basically to transform education. You’re going to see more and more blended learning, flipped classrooms and teachers working on projects rather than reciting content. This is really the shift we’re seeing.

Mihai: So Korbi is not going to replace teachers?

Iulian: No. Teachers will never be replaced. They will just change their roles from reciting machines to facilitators and people who will be coordinating and helping students one-on-one with tasks the AI can’t do or the social aspects.

Follow us on social media and get the news first:
FB | TW | LI| IG

More Articles for You

Linear Regression in Machine Learning Explained

What is Linear Regression? Linear regression is a statistical method that allows us to summarize and study the relationships between …

Applications of Machine Learning in Healthcare

Data Abundance According to Wikipedia, the term information explosion refers to the rapid increase in the amount of published information …

We’re Hiring: Machine Learning Engineer

Open Position: Machine Learning Engineer for EdTech Startup We are looking for an extremely motivated full-time machine learning engineer to …

We’re Hiring: Data Scientist/Curriculum Manager

Open Position: Data Scientist and Curriculum Manager for Education Startup We are looking for an extremely motivated full-time data scientist …

We’re Hiring: Back-End Engineer

We are looking for an extremely motivated full-time back-stack engineer to help us build the future of online education