A Q&A with Deepmind Fellow William Falcon
Recently, CDS PhD student and Deepmind Fellow, William Falcon recently raised an $18.6 Series A for his latest entrepreneurial endeavor, Grid AI. CDS social media team member, Colton Laferriere, spoke to William about his background, his entrepreneurship and how to get your ideas off the ground.
The following was edited for clarity
CL: Can you tell me a bit about your background and what your journey to CDS was like?
WF: I started in the US military. I was going through Navy SEAL training as an officer and came out of hell week with injuries. I ended up going to a SEAL Team for a few years, basically to heal up and I ended up doing supporting operations on the intelligence side. Unluckily, this is around 2010 or 2011 when they downsized the military. And the first places they cut their budget from were people like me who hadn’t completed training.
They gave me the option to do another job or leave the Navy. I decided to leave because I had joined to be a SEAL and wasn’t super interested in being a pilot or something like that. Ultimately, I ended up getting an offer from Merrill Lynch and later, I got accepted to Columbia to do undergrad, so I went back to school. While I was there I was focused on math, computer science and all of that. And then during that time, machine learning started to become a more popular thing and during undergrad I got recruited by Goldman Sachs. I spent about a year working on really cool internal products and eventually found myself trying to use machine learning and deep learning on the trading floor. When that didn’t work out, mainly because the banks were not interested in using deep learning at the time, I left.
Which led to starting a company called NextGenVest with my co founder, Kelly Peeler. She’d been running it for a few years before that, but hadn’t really taken off. So we partnered to revamp the whole thing and came up with a different approach using ML/NLP and we ended up selling that company about two years later. It’s a company that helps low income students figure out how to pay for college.
After that, I applied to PhD programs and from the amazing options I had, I decided to come to NYU. Kyunghyun Cho was my top choice to work with.
CL: Why was that?
WF: I think Professor Cho is probably the most promising researcher today. Today, his contributions are already huge, but I think that we will look back 10 years from now, and trace major developments in the field to Kyunghyun’s work.
Anyway, I took a month off after my company was sold and ended up joining CDS. I was excited to finally have some time off from startup stuff to instead focus on reading, learning and research. It was crazy.
CL: My understanding is that you came up with Pytorch Lightning while you were at CDS.
WF: When I started CDS. I wanted to make sure that I could move through research quickly. I had spent years dealing with engineering issues of deep learning and I wanted to focus on research instead. I wanted to try new crazy ideas quickly. I spent probably the first few months just coming up with some internal tooling for myself to do this, which was a first , raw version of Pytorch Lightning. Then I open-sourced it the next Spring and started at Facebook AI a few months after that.
CL: Did GridAI come about in the same way of you building something for your own use or was it a different process?
WF: It was more intentional. I’ve had a broader vision about how the process of AI should be for years, probably four or five years, when I started doing deep learning at a neuroscience lab at Columbia. I already knew the trouble that, for instance, if you’re not an engineer, but if you’re just a biologist or neuroscientist you don’t know anything about scaling up machine learning. And it’s a problem for science because why should a neuroscientist take four months to train their code, simply because they don’t know how to use GPUs.
WF: That’s slowing the process down. They’re not going to be able to publish their work as fast. I was thinking, “How do I make their stuff faster as well?” A lot of my vision so far has been around that. It’s like, how do you take people who could use this stuff and remove those barriers. And it’s not just about advancing research at companies. Obviously Grid AI is more corporate focused, but it is really meant to help scientists as well to advance stuff very quickly.
We just had a lot of downloads at the end of last year and I started chatting with my friend who is running machine learning at Glossier and I said, “Hey, man, this thing’s taken off and you know a lot of stuff about production. I’ve done it, but not at your scale. So I asked, why don’t we partner to figure out what other problems we can solve with this? And then we kind of came up with the concept of Grid AI at that point.
CL: How does your academic work influence your entrepreneurial work? You’re kind of doing them in parallel to one another and at times they intersect. Do you think being in an academic environment changed the way you approach business?
WF: I think having the startup experience before coming to academia has allowed me to see where the parallels are. If I just started in academia, I wouldn’t have made the links. I think that the main thing to notice is research is the same whether or not it’s for an academic purpose. You have a problem you want to solve and you’re going to take an iterative, scientific approach to solve that problem. That’s the research process.
In academia, the outcome of that is a paper or some sort of publication but in the startup world the outcome of that is a successful business. It’s the same kind of process where we have a thesis and we’re going to be wrong a million times until we’re right. It’s very iterative. I think that the PhD program has helped me become more rigorous in my thinking. Before, I didn’t have such a structured approach and now I do. Because of that, I can think more mechanically and more scientifically about the stuff that I’m doing but move at the speed of startups.
CL: Do you have any advice for other CDS students who have an idea and are trying to find ways to get it off the ground?
WF: As engineers and scientists we tend to overthink a lot. We tend to try to de-risk and we’re very risk averse in general. But for startups it’s just easier to come up with an idea and maybe de-risk it a bit, think about it a bit, and then just go implement it, make it work, and then get it out there and see if people care about it. Of those people who do care about it, figure out why they care about it and see if there are other people that you can expand to. You want to find a small audience who really cares about what you’re doing. Even if it’s 100 people, they have to really care about what you’re doing because they’re going to give you a lot of feedback and then from there you expand.
Once you’ve done that, then think whether there is a real business opportunity. If there is, you might need to raise money. People often do it the other way and fail. They come up with some made up idea, for a problem that doesn’t really exist, they haven’t validated the product need, they put a deck together and try to raise money. Either you spend six months building something that’s useful, or you are going to spend six months trying to raise money and maybe not get money out of it. So spend six months building something useful and spend a week fundraising instead of six months fundraising and then, thinking about the problem for a week.
