“We are in a crucial battle over whether the US keeps its military advantage,” says Sean Gourley. “The consequences are immense. It’s an arms race with artificial intelligence.”
For his 2006 PhD in physics at the University of Oxford in the UK, the New Zealand native applied artificial intelligence to the dynamics of insurgencies. The work was “on the fringes of the physics mainstream,” Gourley says, and journals where he submitted his work told him it was political science. In the time since, “modeling social systems and complexity has become a full member of the physics space,” he says.
Gourley went on to become an entrepreneur. He is currently the CEO of the software company Primer, the second business he founded. It offers artificial intelligence–driven products and services to defense, intelligence, and commercial clients and has offices in London, San Francisco, and Washington, DC.
PT: What drew you to physics?
GOURLEY: When I started university, I thought I’d probably be a lawyer. Ultimately, the thing that got me with physics was being able to uncover a deeper structure in the world. There is beauty in that. And there were the people. Whatever industry you go into, you are making a decision about the kind of people you are going to surround yourself with. And physicists tended to be the most interesting people.
PT: Describe your educational path.
GOURLEY: My master’s research at the University of Canterbury [in New Zealand] was on electronic properties of nanoclusters and self-assembly of systems. A by-product was that deposited bismuth clusters were good sensors for various chemicals. The research contributed to New Zealand’s first nanotechnology company.
A Rhodes Scholarship then took me to Oxford. There I met Neil Johnson, who became my supervisor. I remember sitting down with him and hearing about modeling everything from financial markets to epidemiology. It felt like a whole new branch of problems that I could relate to. It got me excited.
The other professors at Oxford had amazing pedigrees, but Neil still had to make his name. I knew he would have to bet on me, too, and that hunger is often strongest for people who are not yet big names. It took us five years to get our first Nature paper, which is sort of a prerequisite for star power at universities.
PT: How did you get into modeling insurgencies?
GOURLEY: We looked at financial markets. We looked at the growth of fungal networks. But the Iraq war was unfolding, and it was front and center in conversations. At the same time—this was 2003, 2004—we were seeing the emergence of blogs and social media. The confluence of having the mathematical tools to model complex systems and access to data you’d never had access to before got us started on this journey to model insurgent dynamics.
A team of us graduate students were dancing around the edges asking, “Is there something we can do here?” And our answer was, “Yes, absolutely. We can dive in and see what’s there.” I think these journeys are always guided by a sense, a feeling, that there is something interesting here. You don’t necessarily know more than that. But that’s generally how, in my experience, these things start.
PT: What have been your main findings about the patterns of war?
GOURLEY: We’ve learned how insurgencies work, why they are so effective, and what some of their weak points are.
If you want to disrupt an insurgency, the core thing is that you can’t let the small groups evolve to become large groups. You can’t attack the small groups, because there are too many of them; you wouldn’t know where to start. And you don’t want the large groups fragmenting back to small groups. So you want to ossify the large groups—to keep them effectively structured and force them to be quasi-political. They may hold land, manage finances, and act like a local government. But don’t attack them, because they are the ones you will ultimately negotiate with. You want to attack midsize groups to stop them from becoming large groups.
PT: Why do you need to learn this through simulations and artificial intelligence?
GOURLEY: You don’t see the same things from the ground that you see in the data. And the patterns emerge only when the data are aggregated.
PT: Have you shared this advice with the US military?
GOURLEY: In 2005 and 2006 I was able to brief officers at West Point who were about to head to Iraq. My team also had a chance to brief Central Command at the Pentagon, the Iraqi ambassador, and the deputy prime minister of Iraq. Getting science to merge into policy, and then into action, takes time. But I think our work has had an impact.
PT: How did you get access to West Point, the Pentagon, et cetera?
GOURLEY: Through the Rhodes Scholarship community. It’s a weird community, a group of people who have done interesting things and found themselves in interesting places. Connections are informal. Part of it is being in the right place at the right time. But you also have to have something to say. I think ultimately, if you see the world differently and you’ve taken a different approach, you will have answers to questions that are not going to sound the same as everyone else’s. And that affords you the ear of people who are trying to make important decisions.
PT: Are you applying the same approaches to things other than insurgencies?
GOURLEY: The dynamics of insurgency are probably most applicable to the world of cyberattacks and online extremism and misinformation. You see a lot of the same patterns emerging. In the early 2000s, it was about insurgencies. Now it’s more about nation versus nation—Russia’s invasion of Ukraine, and the tensions unfolding in the South China Sea.
PT: How did you become an entrepreneur?
GOURLEY: I remember walking out of a briefing at the Pentagon in the early 2000s and seeing people, probably government contractors, and thinking, “I bet they are selling something these people can use, not just a theory.” Something clicked, and I thought, “If I want to have an impact or make a difference, I have to build something.” The idea may get you 90% of the way there, but ultimately it’s about tools. That was a light bulb moment for me.
Someone told me that it was the right time in my life to take risks. He said, “It’s never going to be easier. You’ve come from being a grad student and living on next to nothing. You could get a corporate job and learn a bunch of stuff, but your risk tolerance will go way down.”
I turned down a very nice, high-paying corporate job and came out to Silicon Valley with a suitcase in my hand and what was left of my graduate stipend—no more than a few thousand dollars. This was 2008, during the financial crisis. I slept on people’s couches for a month while I figured out what to do.
PT: What did you do?
GOURLEY: Based on my experience looking at insurgencies, I got really into visualizing and manipulating high-dimensional data. My first company, Quid, used network and graph theory with visual interactions to let people explore landscapes of data. The company was acquired and is still functioning as part of a larger company.
The new company, Primer, which I started in 2015, reflects the advancement of artificial intelligence. We structure unstructured information. We sell software that you can train to identify different pieces of structure, whether it’s weapons or points of interest or calls to action. We have connectors that let you plug into different data streams, whether it’s audio from radio, PDF documents, or emails. We’ve got those connectors, we’ve got the models, and then we’ve got the applications that run on top of those. It’s quite a complex set of components. The software maintains and generates the self-updating knowledge base for our users.
That has huge implications for strategic analysis in defense and intelligence. It has huge implications for operators who take information fusion from data sources and try to understand what is unfolding in Ukraine. And it has huge implications for detecting whether there’s been a foreign disinformation campaign.
PT: What data do you use to train artificial intelligence?
GOURLEY: First, there are what are called foundational models for language. These are models that are trained with internet-scale data and then predict the next word in a sentence. What actually happens is that the foundational model encodes a whole set of weights and connections in a neural network. That foundational model is then tuned by humans, who label things.
So if you are training a model to detect weapons, the human labels things—that’s a weapon, that’s not a weapon. From thousands of examples the system learns what a weapon is. For audio, the model learns from the way we talk about what a weapon is. You can do the same thing with images.
Now you deploy that model against Russian radio communications in Ukraine. Instead of having thousands of analysts listening to soldiers talk about coffee and baseball—and they do talk a lot about coffee—you can zoom straight into the conversations about weapons, or points of interest, or calls to action. We no longer have to listen to everything.
Another example: Let’s say you know a terrorist organization is emerging and you want to prepare a briefing on it. Going through all the documents to find people, affiliations, actions that have been attributed to the organization, planned attacks, and so on is a huge analyst problem. It’s a day’s work, or a week’s work. With us, a client can effectively hit a button and have it generate something like a Wikipedia page referencing all the different bits and places where it got information. It becomes a self-generating, self-updating page. The analysts go in and do the final edit on top. But it no longer takes a week.
PT: Who are your clients?
GOURLEY: About 70% is defense and intelligence. The other 30% is commercial. For example, Walmart has teams of people who scour through news articles and social media to try to get insights into people’s beverage choices. How does Walmart make decisions about whether to stock, say, zero-alcohol beverages? The structure has been very manual, but they can use our software to automate the generation of consumer insights. Users can hit a button and a briefing will be generated for them.
PT: What are the challenges for you?
GOURLEY: The architectures for artificial intelligence that exist today didn’t exist when we started Primer. And in three years’ time, the dominant neural architectures that we work with today will have changed. The landscape we operate in moves faster than any scientific field I’ve ever seen. In addition, the geopolitical space changes. The world we are moving into is changing rapidly.
PT: You testified before the US Chamber of Commerce AI Commission in July 2022 that the biggest impact artificial intelligence will have is in warfare. Can you elaborate?
GOURLEY: We are still at the very earliest stages of artificial intelligence. The performance we are able to get and create with unencrypted radio content has been mind-blowing. Repeat that for images, for satellite data, for systems to avoid being shot down with UAVs [unmanned aerial vehicles]. You have swarms of robots and intelligence capabilities, but you’ve also got disinformation. You’ve got self-driving cars, so you don’t need people to drive convoys of tanks, and strategic insights and analysis to predict moves in, say, the South China Sea. If my artificial intelligence is better than yours, I’m going to knock your drones out, and now I’ve got air dominance. The fact that the best artificial intelligence wins the battles, and will win the war, hasn’t been fully internalized by our defense and intelligence communities.
PT: How does being a physicist factor into your current work?
GOURLEY: Physics says that even when the world we see may seem unintelligible, there are structures that govern it, and with the right instrumentation and the right theories, you can find the patterns. Ultimately, what physics has done is allow us to manipulate the world to do our bidding.
As a CEO in a fast-moving technology space, I think one of the hardest things to do is make the right bets. I keep up with the latest research papers; so much is seeing what the science means for the commercial world. My physics background helps me accurately weigh the science and to make bets with conviction.
