In his office at Argonne National Laboratory in southwest suburban Lemont, just a few floors above some of the world’s most powerful supercomputers, Ian Foster is talking about poetry. “When the first new generative AI techniques came out, they were doing things like generating useful texts,” says Foster, the director of Argonne’s Data Science and Learning Division. “I used to ask it to write sonnets, which it did amazingly well.”
But the New Zealand native, who joined Argonne in 1989, also quickly realized that AI had the potential to supercharge scientific research. Now Foster is playing a key role in making that happen. Argonne is at the center of the Genesis Mission, the Department of Energy’s new effort to double the rate of American scientific discovery within a decade. The White House has declared the initiative — involving public labs, universities, and private tech companies — “comparable in urgency and ambition to the Manhattan Project.”
Foster is helping design computing systems that will allow researchers to harness AI throughout the entire scientific process, including the design and execution of tests. “We’re even experimenting with humanoid robots to control apparatus that aren’t set up to be automated,” he adds. With the help of AI, a new world is possible: abundant zero-carbon energy, ultraprecise extreme weather predictions, answers to the biggest questions about the universe.
How does AI accelerate scientific discovery in practice?
Say someone’s trying to design a new battery that has high storage capacity and doesn’t depend on some rare material. Scientists might ask the AI, “Well, tell me all that’s known about this class of materials.” That’s something the AI can do very quickly. Then we might say, “Come up with some potential alternative approaches.” And the AI can come up with thousands of them. The human scientists apply their judgment, ask other questions, come up with a set of candidates. We can have an array of these self-driving labs test hundreds or thousands of materials at the same time.
The projects at Argonne funded by the Genesis Mission touch on nuclear energy, the power grid, weather prediction, quantum, and more. What excites you most?
More Top Picks Air Mattresse
Over the last several decades, humanity as a whole has gotten increasingly good at predicting weather. But now we’re leveraging AI to accelerate that by orders of magnitude so that we can get very detailed predictions of what a particular storm event might do to a particular part of the country, to its energy supply, to its water runoff, etc. That’s going to be transformational in terms of how we deal with these mad storms that we’ve been getting. There’s also a strong interest in fusion energy, which is a wonderful technology that has been developing fairly slowly. But now people believe it can be accelerated by using AI to design fusion reactors that can generate very large amounts of energy with low pollution and low cost.
Where does Argonne sit within the Genesis ecosystem?
Argonne has some unique capabilities. It has, of course, the supercomputers. It has the scientific instruments like the Advanced Photon Source. But the central role we play is in creating the underlying brain for the Genesis Mission. We lead something called the Transformational AI Models Consortium, which is the multilab group that is building out the models that will be applied within the mission.
Are we getting to the point where AI agents will be able to run experiments on their own?
They’re certainly going to be running experiments on their own. It’s a question of where human judgment is engaged. The goal of autonomous discovery is to have more of the tasks that we view as mundane being handled by the AI systems. But I see little evidence that human judgment is going to disappear.
How do you ensure the work these systems are doing is trustworthy and foolproof?
Nothing’s ever foolproof. It’s a very old problem in science. We’re trying to understand very complex and subtle things. And so science has got this very rigorous process of skepticism and verification. In a sense, things don’t change just because we’ve got AI in place.
“Increasingly, we want these [AI agents] to learn by experience. We want them to run experiments, observe the results, then choose what to do next.”
How do the AI agents you’re building differ from AI I could use at home?
Typically, an AI agent has one of these language models — the sort of brain you can ask questions of — plus a harness around it that will ask the model, “What should I do to tackle this problem? If I want to buy a fare to Acapulco, what do I do?” And it will say, “You need to go to this website,” and it will automate going to the website and buying the travel fare. Here, we’re doing similar things, but for scientific purposes. We’ll be controlling scientific apparatus or looking up databases. So there’s a whole set of additional concerns that arise: preparing the data, putting the guardrails in place so that it performs safe experiments, experiments that are economically justified, and so forth.
So is the difference just about the data AI agents are trained on?
That’s part of it. Increasingly, we want these things to learn by experience. We want them to run experiments, observe the results, then choose what to do next.
When you read the more alarming predictions about the existential risk AI could pose for humanity, what’s your honest reaction?
I don’t take them very seriously. There are real dangers — for example, misinformation. And we should be concerned about job losses in certain areas. But this notion that they’re going to become conscious and eliminate humans is just not-very-well-informed fearmongering.
Hearing you say that does make me feel a bit better.
More Top Picks Skullcandy Jib
It’s good that we’re thinking about the worst case. But these AI models are particularly hard to evaluate just because they seem so human. If they didn’t generate natural language, then I don’t think we’d feel threatened in the same way.
Has an AI system ever done something that genuinely jolted you?
Its ability to write software is quite astonishing. Complex code that we might not have thought we could produce, just because of the amount of time involved, now it can turn out in a few hours.
What is an AI-driven scientific discovery at Argonne that shows the technology’s potential?
We have some people designing new enzymes to drive various forms of biological reactions [for developing medicines]. They’ve built AI systems trained on huge amounts of information about biology, and they’ve come up with new enzymes that no one had thought of looking at. That’s genuinely new knowledge that I think wouldn’t have occurred without AI.
How big a deal is Genesis for Chicago?
We’re bringing all these institutions together — Argonne, Fermilab, University of Chicago, Northwestern, University of Illinois — to work on these AI developments. Students can come here, they can work with cutting-edge stuff, they can join companies. They don’t have to go to Silicon Valley.
What is it like to be at the forefront?
This is the most exciting time I’ve had at Argonne, because we’re in the middle of a transformational moment. Every day we see new discoveries. We can do things we couldn’t do the day before.