Introduction

Hello everyone, I'm Vinitra, you can also call me Vinny, and today I'm going to be talking about building the future of AI native learning, essentially what learning can look like in the AI era.

So let's see whether...

So who am I, why am I here, what are we doing?

So my name is Vinitra, I come from Lausanne for the last five years, before that the States if you can tell from the accent.

We, me Me and my co -founder Paola started a company about eight months ago towards the future of what learning can look like for the modern workplace.

Before that, we both did our PhDs in a very niche area of AI for education, so essentially very much on how we can personalize education to each individual's background, interests, etc.

And before that, I spent a few years at Microsoft AI, before that at UC Berkeley, where we really worked on this problem of education at scale.

But it was quite obvious when we were founding a company, it would be about exactly this. And it's not just us, what I'm going to present, it's actually a lot of people from a lot of wonderful places, so I wanted to give them their airtime as well.

Who I Am and Why This Topic Matters

Alright, so how did this all start?

A Turning Point: Scaling Data Science at UC Berkeley

Way back in 2015, this was kind of a revolution that was happening at Berkeley. So now we call it AI, before it was deep learning, before it was machine learning, and before Before that, it was data science, and this was the data science era.

When I started there as an undergrad, this course was 100 people. By the time I left, it was 1 ,700 students a semester, and I also ended up teaching it as their first lecture.

And so what we saw was how to grow this kind of program at scale, both in person, but also by the time we left, it became a 60 ,000 -person MOOC. Now this course is used in 60 universities around the world, and it was really about about how to bring data science to everyone.

50 -50 gender ratio, 60 majors represented, fastest growing class in Berkeley's history to date.

The Key Insight: Different Learners Need Different Examples

And the core insight from learning how to scale learning in this way was quite obvious. It's actually that, hey,

there's a lot of these different people from political science, from economics, et cetera, that were getting really excited about data science and then machine learning and then AI.

And it was clear that they all had different needs that they were trying to figure out why this was useful to them.

So diverse students have diverse data science needs and more notably different examples in this class talk to students in different ways.

So for example the folks that were interested in astronomy in the class we had a exercise on estimate the age of the universe from nearby stars or those

that were interested in the judicial system there they had a real court case of was this jury fair, was it selected in a fair way, was it statistically significant that this jury looked so unlike the population that they were serving.

And so many others. I think a favorite in the class was also estimate NBA players' salaries from their heights. There were a lot of people for which these examples really caught on.

And so when you think about teaching, when you have a teacher, they usually think that, hey, to explain this concept, bootstrap sampling, or how does a neural network work, they have a perfect analogy and example in their head. head. But this perfect analogy is not actually the perfect analogy for everyone.

If you teach the same example to everyone, maybe it doesn't really work. But if you try to tailor a different example for every single person in this room, or if you think 100 ,000 person company, simply is not scalable. You can't do this in the form of workshops or something like this.

The Personalization Challenge: Bloom’s Two-Sigma Problem

And how many of you have seen this diagram before? Okay, a few. A few. Do you want to name what it is? I see some folks in the back.

Okay, Bloom's two -sigma problem is what this is. So if you look at the way that a modern classroom works, it's usually one instructor to about 30 students. And this is the graph of the summative assessment scores, essentially how students are performing in one of these courses.

And you can see it's a Gaussian distribution, like people perform well, some people perform badly. You've been in a class, you know how

how it goes but this was the promise of personalized tutoring so when it becomes one teacher to one student you actually get a distribution that everyone moves up on like this is this is the holy grail of what education can look like essentially personalized tutoring and of course this is not scalable even with the Berkeley class of a thousand seven hundred people it's not possible but um

yeah in this case a teacher would be quite sad but maybe the promise of this this, the promise of personalized learning, why everyone is excited about AI for education right now, is that actually the promise of humans working with AI can create this kind of personalized learning system that actually makes sense.

And we can find a way to figure out the right example at the right time in the right way that makes the learning actionable and useful and understandable for everyone.

What Online Learning Looks Like Today (and Why It Fails)

Now, if you look at what current solutions look like in this space, so if you're thinking about probably the thing that's top of mind is how do people learn about AI effectively? How do they learn how to use AI in their jobs, et cetera? How do they learn computer science?

Where do people, and especially even with the context of MindStone, where do, for example, non -technical people be able to kind of close this gap of seeing, okay, you've been hearing about this AI thing, how is it actually useful to you?

MOOCs and the Dropout Reality

This course is probably one of the most famous CS courses around the world. It's Harvard CS50, and this is what it looks like. And because it kind of looks like this,

a MOOC, how many of you have taken a course on edX or Coursera? How many of you have dropped out of a course on edX or Coursera? Yeah, so it's 90%. You're in the majority. I'm also guilty of it.

And we spent most of our PhDs studying learner behaviors on these kind of online platforms to understand what exactly do learners interact with, and essentially we decided that maybe it's not the right way to go because of this dropout rate.

Khan Academy, 7 % retention. A lot of the platforms that you know and love have these similar drop -off problems.

And it's hard to say motivated because because maybe the reason you signed up for this course was actually week two's content or week eight's content. You're not gonna sit through six weeks of stuff just to get there.

And so if I can paint the space a bit more, what online learning looks like today? On one side, one size fits all learning, right? edX, Udemy, Coursera, LinkedIn Learning.

You guys have probably tried these. The problems and the videos stay the same no matter who you are. So you go in, the course is the same, you do the course.

And maybe there's an AI coach on the side that you can ask some questions to. But on the other side, there's a lot of new players, right?

AI “Study Modes” Today: Helpful, but Missing Core Pieces

ChatGPT has a study mode, Gemini has guided learning, it's called LearnOM, Codd has a study mode, and the rest definitely do too, and it's actually, there's about five labs in the world that work on this stuff, and we know all of them, so it's a lot of friends at all of these places, people

are just mingling. But there are some problems that I feel with this way of doing education, for example, ChatGPT study mode.

One, you have to know what you you want to learn before you start learning, which is already kind of hard. There's usually like a curriculum of skills that you're trying to build, but you don't even know what you

should ask for first. The second is that there's no notion of multimodal content right now. You're just kind of learning by like text hitting you, right?

It's not videos and interactive tasks and problems, et cetera. And maybe there's different modalities that work better for you.

That's a solvable problem. The other problem is context, right? Is it in the context of your tools and tasks, is it in the context of your difficulty level?

Right now, if you look at these kind of tools, they're mostly a prompt co -designed with teachers, right? Hey, teach this Socratically, don't reveal the answer.

But the biggest problem is that there's no notion of skill tracking. So usually, you're not trying to just learn one thing at a time, you're trying to learn a set of different skills.

And you want to see like, through this conversation, how am I progressing at this skill? There's no visibility for that for you.

Or if you think about adults, especially in in the workplace, no visibility to your manager, et cetera. This learning doesn't really count. And it's not grounded in the company context, et cetera.

So there are a lot of problems.

What Research Already Knows: Learner Models and Knowledge Tracing

But most notably, the last 15 years of research in our field has been about a concept called a learner model.

Essentially, it's a concept called knowledge tracing, but it's a way to estimate students' aptitudes and levels as they're answering previous questions of a certain skill and use that to decide what the right level of difficulty

of content is or what the right content to show in the first place is to maximize their learning gains.

And none of these guys are doing it like this, and we don't know why, but it's maybe because these research and implementation communities are not actually talking so much to each other.

What AI-Native Learning Should Be

So what do we believe learning should be?

A Knowledge Graph of Multimodal Content

Well, we think that there should be a knowledge base of videos, podcasts, problems, reading, slides, all of this stuff that should be ingested in something like a knowledge base, a graph of sorts.

And not everything in this graph is relevant for everyone to learn. I mean, you can witness that even from the dropout rates, so people aren't learning it anyway.

But if we can find the right part of the graph that makes sense for a specific person and adapt it into the context of their tools and tasks, then we have something that actually makes sense.

And so our company, Skolay, is towards that. Essentially, we're starting with the use case of using AI to teach AI effectively, especially through agentic lesson delivery and the like.

But what does Skolay mean for learners? Well, yeah, content that's actually relevant to you and your job. we're pricing actionability over everything else.

And the second is that curriculum that adapts to you.

We've co -designed the solution over the last six or seven months with Swisscom, as well as Decathlon, Canal Plus, and Coop. And our first larger base is Harvard University, whose agentic AI courses were recently named by Forbes as the number one way to learn agents in 2026.

All of the people around the world who are taking these courses are using Skolay in the evenings and they're from a ton of different companies so we hope that we've in some way solved this generalization problem of what your role in your job and what is relevant for

you to learn for such that every little example every little problem is actually useful and so

Making Content Relevant to Your Job

what do we mean by content relevant to your job well essentially we have this large graph of validated data and AI learning content updated every week from the Harvard courses from Berkeley from EPFL and from a bunch of others.

But we recognize that, hey, no one wants to watch a one -hour lecture video, but maybe the right one minute of the right explanation at the right time from even your company's own AI video might be useful for you to learn this properly.

So of the 60 things you could learn about AI, maybe for you, three of these things are particularly relevant.

And maybe for HR, they should be able to create these custom modules on demand from their own documents, their own needs, assign them to people, et cetera.

An Adaptive Curriculum: Difficulty, Modality, and Context

And what we mean by curriculum that adapts to you, so I was talking about the notion of a learner model, essentially a way to estimate people's skills. We use that to create the right bite -sized chunk of content for you, either text or video or audio or something like this, upon which we generate a questionnaire task, upon which the next chunk is conditioned upon, right? right?

Because you got that question or task wrong, because it's already at this level, because we estimated your skills to be here, this is what the next chunk of content should look like.

So we're dynamically adjusting the difficulty and the modality and the contextualization of the content, examples, and tasks.

And this has to do with the learning science concept called the zone of proximal development. Basically what that means is that you want to be, maybe

when you're learning, if you think about your favorite way of learning, it's probably different from the person sitting next to you. But there's an optimal zone of frustration for everyone.

If it's too easy, you're kind of bored. And if it's too hard, you're like, yeah, man, this is like, not for me.

And that optimal zone is around 80 % correctness. And that's what we're talking about in terms of the zone of proximal development.

And so diving directly in to what the solution could actually look like, I would be super happy to show you. But maybe before that, that if you have any questions.

No questions. Just demo. Wonderful. Cool.

Demo Walkthrough: How Skolay Personalizes Learning

So, I made a new user account here.

But let's say, hey, we want our learning to be as personalized and engaging as possible, blah, blah, blah.

Onboarding: Company, Role, Tools, and Preferences

What company do you work for? We can choose any company. For example, let's choose SBB. Because we're here in Switzerland.

And maybe any role. like it could be the person who is literally working on conducting the trains, it could be the person being the ticket checker, it could be the person being X or Y or Z.

So about SBB or any company, we have some context about it because there is a lot of information available in the world. Behind the scenes we're using Bing company search, but if it's a new company you can edit it, you can rewrite it with AI etc if it's a lesser -known.

And maybe your role, let's say hey I work in marketing for SBB, but it could be anything right like maybe I I actually work on data science for SBB or maybe I actually work on store operations or something like this.

From this, we make a guess at what we think your role is about. And from this guess, we try to identify your tools and tasks. What we think you do on a daily basis such that we can understand what might be useful for you.

So maybe you do use Adobe Creative Suite to make images, et cetera, for being a marketing person. Or maybe you use Figma for example and we make a guess at what we think your goals are

and then from this notion we kind of go to a skill rating hopefully the internet is working seems promising we make a guess at what we think your your experience levels etc are but we allow you to override like for example you could say hey I don't actually like videos I do like explanations I do want it to be energetic not academic I want it to be this way etc where I have these learning goals and it should be about that and maybe I'd like five to ten minute lessons not the 15 to 20 minute okay.

From here we basically make a guess at what we think are the most relevant parts of AI and data science for you because there could be a ton right maybe there's we have about 60 for example different lessons you could say hey we think prompt engineering is actually useful for you but maybe you want to learn agents maybe you want to learn LLM architecture intuition or maybe you don't etc and you can reorder this change your curriculum in any way etc but let's say for example this is what you want to

learn and we have a notion of a learning journey so in this case let's say that you want to learn about agents but you want to learn about agents in the context of being a marketing person at SBB or whatever this specific account is so in this case it's understanding agents but in the context of SBB and and marketing.

There are specific suggestions you can make.

Agentic Lesson Delivery and Learner Control

And you have the option to DJ your lesson. So you could have just started the lesson right away, which is a bunch of agents working together.

There's an agent that orchestrates the lesson, and then an agent that pulls the right questions at the right time, and the right videos at the right time, and the right tasks at the right time.

Or you can change it a little bit. And this has to do with a concept called self -regulated learning. If you do about 30 seconds, one minute of planning, it actually makes you much more engaged with the lesson.

If you say, hey, actually, this is not relevant to me. I don't want this part. I want it to be this way. I have a meeting in 10 minutes, I'm on my morning commute, there should be a podcast.

But let's say, hey, this lesson is fine, and you're ready to go.

So where this differs from something like ChatTPT, there is actually a lesson plan. It is grounded in trustworthy documents.

So if we are working with a company, we have all of their documents, et cetera. If not, we have this knowledge base of Berkeley and Harvard and a whole bunch of other really high -quality curated data science and AI learning materials, from which we have a lesson plan that you can change at any time.

So our ethos is AI should provide suggestions that you can modify. It should be easy to just continue and get the ball rolling, but you should be able to have control still to learn the way you want and how you would like.

A Team of Learning Agents: Tutor, Examples, and Tasks

And what we have here is a bunch of these little agents that kind of interact with each other. So we have Ole, because we're Skole, who is like kind of our primary metacognitive learning tutor, where it not only gives you some parts of the conceptual content, but it also asks you, you know, how you're feeling, how you're doing, what is going on, does this make sense and how you relate to your work, etc.

And then we have a few others, like, for example, examples, analogies, connections to responsible AI policies at your company. Those are given by Atlas.

We have a notion of Quinn, which gives a lot of tasks that are different kinds at the right time. So, for example, you might get a quiz question. In this case, it could be multiple choice. But then there's also fill in the blank.

There's interactive visualizations. There's open -ended questions.

With prompt engineering, for example, you get to actually get a situation that makes sense. But you can see even here, it's in the context of SBB and in the context of what might make sense in this world of SBB as you're learning marketing, etc.

We generate some videos, we pull the right one minute of video sometimes. In this case, we do have a bit of like generated explainer videos.

And based on how you interact, based on what you skip and what you don't skip, actually the lesson itself changes and the next lesson will be even better suited to you.

Interactive Tasks and Generative UI

And then kind of in the back end, one cool thing that I wanted to show you is this notion of interactive visualizations.

So for example, maybe you want like a decision tree or some kind of scenario.

Of course these pop up in the middle of the lesson, but these generative UIs actually change based on who you are and what you're doing to make learning a little bit more interactive than like, hey, content coming at you. You have to obviously answer questions, et cetera. But you also have to interact in a slightly more interesting way.

So, for example, we have a pilot going on with Cope with, for example, cheese stacking. They had a huge food hygiene thing and maybe instead of just hearing okay you have to stack these cheeses this way and these cheeses this way you should actually practice like figuring out which cheeses to stack and which wheeze.

And more than this there are a ton more things we could do with generative UIs, a ton more ways that we can make this extremely interactive and personalized to you such that eventually what you're giving to the tool you also get out. But overall

Where This Goes Next: Learning in the Flow of Work

the one huge concept that I guess I could leave you with is kind of the direction we're going moving forward.

So, in this case, this is what, like, an AI -native learning platform might look when built with AI thinking about it from the ground up. But

what we would like to get into is learning in the flow of work. So, we're building out a browser extension, but what we believe as a team of AI researchers is that a personal assistant is going to win on your computer, right?

You have Siri now, but Siri's not kind of useful. It's going to be better, right?

Or OpenAI will win or Google will win, someone will win, and this will plug into an agent -to -agent marketplace. And in that world,

Closing the Loop From Training to Practice

schooling makes a lot of sense, because we could say, hey, remember that training you did three months ago? This thing that you need help with? This is your perfect chance to use that.

Here's your two -minute refresher, knowing what your misconceptions about it were, knowing your tools and tasks.

And here's your company's internal tool. So hey, do this on Gemini. Don't do it on OpenAI.

Or use Canva. Don't use whatever. And we're able to now close the loop, because we can see what parts of the learning actually made it to practice.

And that actually, in turn, improves our learning, but shows that actually the learning was actually useful. And so that kind of loop is what we're trying to work on.

Analogical Learning Across Tools

And the second thing is analogical learning. So let's say you've learned an AI feature in Excel and now your company is forcing you to use Tableau. Okay, cool.

But we have context of what we think your tools and your tasks are. So there's no point in doing a generic tutorial ever again.

We could say, hey, like you know how to do this thing in Excel, this is how you do it in Tableau, but this is the key difference, and this is where this tool can actually help you more.

This horizontal learning layer across everything, kind of like Clippy, is what we're trying to build over at Skole.

Conclusion

And by the way, it's named Skole because it's an ancient Greek word that means leisurely learning, like learning in a relaxed way with your friends for fun, which

is what we think learning should be, and not the scary feeling you get whenever someone tells you you have to do corporate learning so there you have it thank you so much