In the late ’90s, when Sameer Maskey was a college student, artificial intelligence was an esoteric field. Yet Maskey, who “fell in love with computers” during his freshman year at Bates College, took an immediate interest in AI, particularly the idea of speech-to-speech translation. In 2000, after research stints at Carnegie Mellon and the California Institute of Technology, he built the first Nepali speech synthesizer and released it open source at a conference in Nepal.
Back then, “nobody took any interest,” says Maskey. “Not many people were talking about AI in Nepal.”
Today, Maskey is the founder and CEO of Fusemachines, a company that creates AI tools and provides teams of engineers to businesses in the US that want to build their own. Quiet and driven, he has converted his years of research and teaching — as a PhD from Columbia University, a speech-to-speech-translation research scientist at IBM Watson, and an assistant adjunct professor at Columbia — into the foundation for a 250-person company.
Fusemachines began with the aim of producing customer service products, like automated dialogue systems (the first of which Maskey developed after locking himself in a room and writing 20,000 lines of code). The New York City government was its first client. But as the shortage of AI engineers became a prominent issue in the US, where talent is swiftly snapped up by Silicon Valley giants, Fusemachines’ brand has shifted toward teaching AI in developing countries, where, as Maskey puts it, the company is “democratizing AI.”
This interview has been edited for length and clarity.
Why teach AI and machine learning? Why start in Nepal?
In the past five years, the demand for AI engineers in the US has gone through the roof. They are being sought after by all the big tech companies, but there aren’t enough of them, which has created a huge gap in the market. Newspapers were talking about how getting AI engineers was becoming a lot like sports hiring. Companies in the US that aren’t at the scale of Google and Facebook can’t afford AI talent. Colleges and universities are increasing their capacity, but they can’t keep up: the demand for engineers is in the hundreds of thousands.
After starting Fusemachines, I realized the market size is quite big, and that the technology would be useful for companies not just in the US. We need engineers: where do I find them? When I was in Nepal on one of my yearly trips, I thought, Let’s see what happens when I teach students here.
There was a cost advantage to building a team in Nepal, of course. It gives us more runway with the same amount of money. But from an impact perspective, it was more interesting to train engineers from Nepal. A lot of the AI technology today gets built by large companies and is geared toward the US: if you put a self-driving Tesla on a road in Nepal, it’s not going to work that well. High-quality AI education in other countries means graduates will build tech adapted to their local issues and contexts.
Is teaching AI abroad any different than teaching the subject in the U.S.?
At the beginning, I didn’t really have many expectations. But I started to see sparks of quality talent. I taught the same course as the one I taught at Columbia University. I soon realized my students in Nepal were equally good.
Going into it, I knew there would be some issues with rote learning, but not at the level I saw. Students have read about topics — from a syllabus perspective they have done all the math — but they don’t understand it well enough conceptually to do the coding. A lot of students weren’t ready to take the first course, so our only option was to come up with our own set of foundation courses and teach those beforehand. Our students in Rwanda and the Dominican Republic are similar. And other than the students from the top schools, it’s the same in India too.
How does the AI training industry compare in the U.S.?
The bootcamps are expensive — $24,000 or so for a typical course — but bootcamps have a lot of infrastructure costs, so margins are narrower than for a typical tech company. In traditional coding bootcamps, the students are not computer science students. They’re not making much money yet. In AI bootcamps, students are already software engineers; they already have good jobs. They don’t want to stop making money and enroll in a bootcamp. This might be another reason there aren’t as many AI bootcamps in the U.S.
Do you think, then, that the talent shortage in the U.S. will continue?
There are a couple of things that might reduce it. There are the bootcamps, and also, AI tools are being commoditized — you can build basic versions of a system. But if you want to make any improvement, then it’s a black box. Probably there will be some point in the future where the shortage will start to taper, but it hasn’t happened yet.
Could online courses teaching AI — such as those offered by fast.ai or deeplearning.ai — help satisfy the demand?
Those are mostly taken by computer science students, but the problem with purely online, self-paced courses is that the completion rate is less than 10%. We have online classes now too, but they’re live classes, which makes a huge difference. Most students need more hand-holding, and there’s a lack of faculty. The foundation of AI, unlike programming in general, is all math. You can learn Python programming in six months, but you can’t learn all the foundations of math in the same amount of time. It just takes more time and more effort, from students and from teachers. There’s no shortcut.
We want to contribute to that by democratizing quality AI education, which will speed up the adoption and production of AI engineers. We are thinking of scaling up beyond our current reach by rolling out a program called Teach AI for the World. Like Teach for America, talented AI engineers would travel to different countries and teach for one or two years. We want to see thousands of teachers teaching thousands of AI programs around the world.
You mention “democratizing AI” often. What does it really mean?
There are three core things. The first is bringing AI technology to underserved communities, such as Nepal. Our most successful product has been a platform for AI-assisted learning called Fuse Classrooms, which was adapted from the platform we created to teach fellows. It can inform teachers how students in their classes are doing, what material they’re struggling with, and what they should study to improve. It’s being sold in some markets for as cheap as a dollar, and it hasn’t been long, but altogether, we already have students logging half a million minutes a day.
The second is bringing high-quality AI-education programs as well, so that people in these communities can learn how to build AI themselves. People who live in underserved markets understand how and why AI can make a big impact.
The third thing is enabling the skilled engineers we train to either find jobs or build their own AI technologies. When we accept fellows, our big focus is on training engineers further so they can impact their community. They will eventually become managers and create more demand for AI solutions and engineers.
How do you ensure your students aren’t just the elites in the developing countries where you work?
In Nepal, we’re already starting to partner with schools outside of Kathmandu, and we’re planning to pilot larger live online classes. In the long run, we want to make AI education — really good AI education — available to everyone, getting to the bottom of the pyramid. But we’re not there right now, because we don’t have enough resources. For example, there are students without access to a mobile device, and we can’t afford to buy them these devices. Right now, we mostly train people who studied computer science, most of whom are from Kathmandu. They’re elite in the sense that they can afford a phone and a laptop.
We’re creating offline platforms, so that if students don’t have internet access but the school does, they can come to the school and download what they need for two or three weeks at a time and continue studying through the app. We also want to go downstream. In Nepal, there are 8 million students in school but only 400,000 in undergraduate programs. More than 7 million students just disappear, so we want to start teaching students in high school or earlier.
But you’re also a for-profit company. Where is the trade-off between that and the goal of truly democratizing AI?
The way we look at it is that we have two bottom lines. We have a for-profit company, so we need to make money and grow. We have investors who seek returns on their investment. But our other bottom line is democratizing AI such that it makes lives better for a lot of people. I think there are ways to achieve both; you have to find a balance. For example, our Fusemachines fellowships were open to anyone who wanted to learn, and half of the students from several cohorts were already employees at other companies. That didn’t make sense from a business perspective, but it did from a democratization perspective. We will probably seek grant funding to democratize beyond what we are able to do as a company.
Where do you think tomorrow’s AI talent will come from?
There will be a huge uptick in AI engineers coming from developing countries, the way India became a big center for software development in the late 1990s and early 2000s. The trend will probably be similar, if not even more accelerated with AI, because now there is better access to resources. Especially post-Covid, there will be a lot more development in distance learning, so engineers in developing countries will get almost the same-quality education as students in the U.S.
Did your bet on AI talent pay off?
Right now, Fusemachines has around 120 machine-learning engineers, and a big chunk of them came from our fellowships — we hire more than half of the graduates. We provide AI talent solutions, so many of our engineers are dedicated to clients. Our clients in the U.S. include Enhatch and Barkbox. The companies we’ve worked with have gone on to raise more than half a billion dollars in funding after Fusemachines engineers helped them build their core products.
Besides direct revenue, there’s the intellectual property we have been able to create as a company. We’re in the process of filing six U.S. patents, and a lot of the engineers behind them were former AI fellows who went on to create cutting-edge technology. Engineers in Nepal wouldn’t even dream of their names being listed on a U.S. patent. It’s the kind of moment I have been waiting for.