This is the future envisioned by Joel Hellermark founder of Sana Labs envisions, a 22-year-old entrepreneur who hails from one of the most innovative tech-savvy countries on earth Sweden, home to Spotify, Klarna and Skype.
It would be an understatement to call Hellermark ambitious; this is a kid who by 13 enrolled at Stanford online to start coding, founded his first company at 16, raised his first round by 18, skipped college to establish Sana Labs and in the process found fans in Mark Zuckerberg and Tim Cook all by 21.
Call it precocious, but we think he has the makings of the next true industry hellraiser. Hellermark wants to disrupt a $6 trillion market, one that has become overburdened with bureaucratic problems, massively outdated and at this point a blunt tool in desperate need of a makeover.
Today, millions of students out there have been left behind in classrooms, and Sana Labs wants to get them up to speed. Sana technology is currently being rolled out in 75% of schools across Sweden and there are plans to globalise the brand further.
We all know something about the potential of AI, but not perhaps of its potential to revolutionise learning, allowing it to become more seamless, agile and cost-effective. And with figures showing that students who received personalised content had 50% better learning outcomes in a year, who wouldn’t want it. But how will this affect teachers, schools and the future of education? We sat down with Joel to find out.
You seem to be a very ambitious kind of person. Who are the heroes that you want to emulate, people like Steve Jobs etc.?
I’ve always been very ambitious. I started programming when I was 13, the ideas I started tinkering with were not ideas that were going to have an incremental impact. I was always drawn to ideas that could potentially fundamentally change the world.
It’s been more about ambition, and I’ve idolised a lot of historical figures. Not necessarily Steve Jobs or Jeff Bezos but people like Leonardo Di Vinci, Michelangelo and Edwin Land whose genius lay at the intersection of combining insights from different fields.
They weren’t just mathematicians, or philosophers, or artists, but their genius lay in combining different insights from different industries. So I idolise polymaths a lot, whose genius lies in taking insights from maths and combining that with the arts, and I think that’s what we try to do with Sana as well.
When you see people like Mark Zuckerberg of Facebook or Brian Chesky of Airbnb, do you think they’ve set the bar too high for people that are young entrepreneurial leadership figures like yourself?
I think that’s allowed us to think on a whole different scale. Our ambition at Sana Labs is that there shouldn’t be a learning product in the world that’s not in some way using our technology. Historically I think before figures such as Brian and Mark you weren’t allowed to think at that scale yet. I think we’ve proven that ambitious young founders with an exceptionally meaningful idea can create some of the world’s most important companies. And I think increasingly you’re seeing that kind of ambition, as a function of that, in young founders.
So why education?
I think education is at the stage where it’s making a sea change to online. It’s at this shift now where we’re taking this one size fits all model and transferring it online, but I think the next step now is, how can you scale personalisation?
That’s what artificial intelligence is good at. Taking these processes that were historically hard to scale and making them available at scale.
It’s not scalable to have a personal teacher for every single student globally, but you can take a lot of the aspects of personalisation that a teacher would have historically been given such as personalisation of the curriculum, recommending which exercise you should see next, and you could provide that autonomously through AI-powered systems. I see that as the next stage of online education.
“I think teachers are going to be one of the last jobs to get automated.” – On automation paranoia
Looking back at the 20th century model of education, it feels outdated compared with what we need it to do today. What do you think is the biggest issue with the education system today.
First of all, I think there’s a lot of good things about the education system today.
We’ve essentially taken what has historically only been accessible to a very few, before the industrial revolution, and we’ve created this factory model where we can put in kids and everyone, to a large degree, gets access to classroom education.
However, there has been this difficulty in providing personalisation at a scale which has allowed in particular the students that have fallen behind to fall increasingly behind over time.
If you only master 60 per cent of year one, and you only master 70 per cent of year two and only 50 per cent of year three, that means you’ve compounded knowledge gaps over time. And it will be increasingly hard for you to keep up. This is largely because of the lack of personalisation.
We haven’t had the ability for each student to be provided with a personalised curriculum and I think that’s what you’re going to see now with this paradigm shift.
Let’s say Sana Labs is the gold standard for AI in education 50 years from now, everyone’s using it, much like Spotify is for streaming or Facebook is for social interaction. I walk into a classroom and sit down, I’m in college or I’m in an adult learning centre in a nighttime course. What does the Sana Labs classroom look like?
Ok well, you will see a lot of the students having already mastered the basic underlying concepts before they come into the classroom. The teacher standing in front of the class knows precisely which student needs help and has actions of how to help each student which they can act on with a single tap.
Teachers will have the tools to understand their students more deeply; the students will be managed in a highly personalised way. The class will be spent mostly on project-based learning.
Our ambition is ubiquity. That’s why we took that approach that rather than building our own products we wanted to power all of the world’s learning products.
So when you walk into a classroom or sign into that online learning product the curriculum becomes entirely personalised for how you learn. Let’s say you’re answering questions incorrectly, we can identify your knowledge gaps and recommend materials to address those knowledge gaps.
If you’re progressing quickly on some concepts, we can make sure that you get increasingly hard questions or assignments.
Where I started looking at how to solve this problem was to deduce it from neuroscience. To study how the brain learns, and look at how we form memories. I spent one summer in Cambridge under one of the professors of neuroscience there and from there looked at how can we create better learning experiences and algorithms to personalise education. However, we still know very little about how we learn.
“I think we’ve proven that young, ambitious founders with an exceptionally meaningful idea can create some of the world’s most important companies.” – On being an entrepreneur
There is a lot of talk and paranoia about discrimination in AI at the moment? You’ve engineered these systems yourself. I’m not trying to point the finger; I’m just trying to work out how you might deal with that issue.
I think the biggest problem with bias in machine learning is when the system does not reflect the users. If you have a machine learning model that doesn’t adequately reflect all of the users that are going to interact with that model. So if you have a system that is predominantly trained on datasets of white middle-aged men, it’s subsequently going to be used by them. But if you look at the datasets we train our models on, it’s datasets of the users of the products. So if it’s a product that’s predominantly used by dyslexics, we’ll try to model how they learn most effectively, and custom tailor their educational experience.
But our understanding of AI and the brain is still limited, right?
There are so many intricacies to how we learn that deducing them from neuroscience is going to take several lifetimes. One testament to this is that no one in the team knows how we learn languages. We’re by no means specialists in this.
We recently took part in a global AI benchmark by Duolingo. They were out to find the world’s top approaches to modelling students learning, and in that benchmark there were professors at Cambridge linguistics, NYU’s cognitive science lab, a Chinese company which has just raised tens of millions which specialises only on language learning modelling, and we ended up winning every single category on all evaluation metrics.
That was because they had taken the approach of rules-based systems and heuristics, they encoded everything they knew about how we learn languages. However, what we did was to apply machine learning to determine from the data.
So an adaptive, evolving kind of learning, as opposed to the heuristic side of machine learning.
Exactly. Yes. Rather than manually encoding rules and dependencies we went for an approach where those were learned directly from raw data. Allowing us to transcend the limits of traditional approaches.
I think exploiting deep neural networks is the future. But my biggest issue is that we’re still at a very infant stage of AI, still learning how the brain works, technology has only started to catch up to our imagination in the last 50-60 years.
I’m worried we’re jumping into an abyss that we don’t know much about? Once Sana Labs starts to accelerate how will it exploit our educational alignment?
I would say we have a very different setup, in that we have learning objectives such as you should master trigonometry. Then we create a personalised path for you, how to effectively learn trigonometry the fastest.
So we have a very controlled setup whereas a lot of the problems with those systems were that they were highly uncontrolled. With Facebook, for example, they had an objective function of maximizing engagement combined with user-generated content which clearly resulted in bad outcomes.
But if you take our objective function, essentially to help people learn and master as quickly as possible, and we have content created by expert educators, that environment doesn’t have a lot of variables that are hard to control.
“We still know very little about how we learn.” – On the mysteries of the brain
You talk about how you want to make learning the great equaliser. You’ve said that you want to help create personalised education ubiquitous for everyone. But we can’t ignore that fact that there are vast swathes of the population across the world that aren’t sponges for learning. There will still be gaps.
I’ll give you a good example, the MOOC courses (A massive open online course). 80 per cent of people drop out. So what about people with dyslexia or other learning difficulties?
If you look at these online courses, to a significant degree they drop out is because the content is too hard or too easy, so if you take students learning languages on MOOC, and you look at their attention, it’s highly correlated by the probability of them completing an assignment.
When students had 70% probability of answering a question correctly, that’s when they were most engaged, when their retention was the highest.
So I think one of the largest problems we see which hasn’t allowed online education to be truly democratised is the lack of personalisation where you can’t account for these very different knowledge levels that the students enter these platforms with.
You can have these systems adapt and recommend to address your knowledge gaps. If you take examples, in Kenya the average ratio for the number of school books to students is 1 to 14. So 14 students per textbook.
But to a large extent, most of the students are getting access to Android phones which is costing them 15 dollars. So what you see there is rather than them sharing a textbook with 14 students, they can get access to online learning services through smartphones.
We’re not replacing teachers; we’re not creating AI teachers. What we’re building are systems to automate a lot of their conceptual understanding. But there is going to be so many other aspects that are important.
But there is a lot of automation paranoia going on at the moment; teachers will be the first to go once AI gets to a certain level. Are you helping to build that paranoia?
I think teachers are going to be one of the last jobs to get automated. The education system is shifting towards a project-based model. So for example, if a knowledge gap emerges and the students are struggling in the classroom, you will be able to notify the teacher and give them suggestions about how to help the students move forward. The education system is shifting towards project-based learning which will mean that teachers become increasingly important.
“You can have these systems adapt and recommend to address your knowledge gaps.” – On AI education
Project managers basically?
Yes, mentors and project managers.
Right. So you took this idea, and you wrote to Mark Zuckerberg and Tim Cook. Is that true?
Yeah, and I think that was a bit of my naivety and ambition as a kid. There’s this way of getting access to phone numbers. When someone registers a website, they always have to register a phone number. You can then look through and get people’s phone numbers. So in my contact list, I had Kanye West’s phone number…
Kanye West registered his phone number?
Yes because they usually register these random domain names which you first have to find and then when you see them you can get access to their phone numbers. So as a kid I used to call them. Usually, I had very little to say, but it’s a fun hack that I used to do.
More recently after founding Sana Labs, we ended up reaching out to Mark and Tim, what’s interesting is there’s this concept of “on the internet, nobody knows you’re a dog.” You can reach out, and if you have a meaningful idea and ambition, then some of the world’s most high-quality people are drawn to that. And that was the case for both of them.
Did you meet them?
We met their teams. We haven’t disclosed anything yet, but we’ve been working a bit with the Chan Zuckerberg foundation. How they can provide personalisation at scale is really one of their key objectives, and also with Apple’s education team.
You want to disrupt a 6 trillion dollar market. What is the biggest obstacle that you face becoming a ubiquitous tool for education worldwide?
We’re working a lot now on scaling to various infrastructures. We’re also working a lot on building trust, publishing case studies, participating in benchmarks, proving that the approaches we use are beneficial to students. So one of the biggest ones that will be of use is making our platform very easy to adopt and integrate independent of which system you use.
So do you think this will be assisting with a new conversation about us and AI in general?
I’m going to be talking with AI, and AI will become my mentor and as you said the teacher almost becomes the outsourced project manager in a way. Have you thought about the ethical design implications of interacting with AI every single day?
Yes of course. But I think we’re very soon going to stop thinking about it as interacting with AI. We don’t think about interacting with electricity every day.
This interview has been condensed and edited for clarity purposes.