What if we could predict the timing of solar flares, missile attacks and volcano eruptions? Maybe A.I. can help. In November our editor chatted with Paul Werbos, the mathematician whose work on neural networks is the basis of the new artificial intelligence.
METTA SPENCER: Tell me about this new initiative of yours.
PAUL **WERBOS:**We have a name — ProjectAmaterasu. Agroup of us get together on Friday mornings to discuss how to implement a concept I articulated. The goal is twofold: to demonstrate a radically new breakthrough in internet and quantum technology, which I’m responsible for. I call it Quantum Artificial General. Intelligence — QAGI. Ron Turner, who works with the Air Force, suggested we could use a testbed for a problem that he was interested in — the prediction of solar flares.
Solar flares are important to the security of electric powered arrays and also to spacecraft and people in space. I said, “Okay, I’ve got a technology to try out. You’ve got a testbed to try it on. Let’s bring them together.”
SPENCER: Why study solar flares?
WERBOS: I used to run Electric Power Research at the National Science Foundation.It was the Number One program in the United States for advanced research on how to manage and protect electric power grids. Under earlier directors, it was responsible for the whole power grid system in the United States. They did major things and I improved it. One of the things we worried about is whether an electric power grid could be taken down by a solar flare.
SPENCER: I’ve heard that one could knock out the power grid for the hemisphere.
WERBOS: Canada has experienced some very bad solar storm events in the past, knocking out electric power. There was a famous event— the Carrington Event. When I was when I was running this program, I was asked to look into this. A Congressman, Trent Franks from Texas, was put in charge of the Committee in the House that overviews all the worst threats to North America—s very survival. He set up a bipartisan meeting for about 30 people from around the country. In it, he said: ÔI get all the classified information about anything scary that we’re not supposed to know about. In my first six months I could hardly sleep because there was always something new. But it—s now my opinion that the Number One threat to the survival of people in the United States is another Carrington event. Let me tell you why I’m worried about it and how we can prevent it. Let me talk about what we can do to make this country secure.
That meeting was an introduction to me. I got information from many sources in my job. An article inScience— a pretty credible magazine — estimated a Carrington-level event should happen about ten percent of the time when we have a sunspot cycle, which is about every twelve years. It—s a random distribution, but expect something like 25% probability that we will have another Carrington event in the next twenty years. As they added, the world economy has changed. We are more dependent on electricity than in1859,the time of the Carrington. That created a giant mess, but we survived. Now, the most conservative estimate is $2 trillion worth of damage overnight. In these meetings with Trent Franks, people came in to explain to us how another Carrington Event is a threat to life itself in North America. The guy from the Air Force says, “If we can predict these kinds of flares, we can take advanced action and protect the power grid. But you need to know: Is it coming?” That was our first goal — to predict when it’s coming. And by the way, if you’re an astronaut in space when it comes, you’re fried. And if you have assets in space, they’re dissolved too. It’s only one of the many important things you can do with the new technology. But we think a new type of AI could predict solar flares much better than any of the existing technologies.SPENCER: What variables do you take into account?
WERBOS: That’s actually a very practical question. We had our technical meetings every week. We got into the databases. And this guy from Project Air Force gave us a wonderful talk. What does anybody in the world know about predicting these flares? How good is it? What are the stages? By the end of his talk, I concluded that the best data is from space activities — satellites and telescopes where they are looking at the sun. They can see the sunspots. What’s the probability that a sunspot will suddenly grow up to be a flare? When will there be any ejection? They can see the sunspot bubbling. They have mathematical models, but it’s hard to predict when we’ll get an eject.When I look at that data, I say: My god, this is our kind of data!
There are two major kinds of data. There is data kind of like a TV picture — an array of pixels. For each square in the image, you can see what the current variables are, and they have a list of standard variables for space. It’s like a video record. And Task Number One is: How do you predict a video stream? There are new ways to sense the nuclear emissions and I’m excited by that technology — but it’s not what you can do this week. Machine learning can learn to recognize time series in general. We have developed universal learning systems that can learn to predict any time series and there are extensions of those that I published that will extend it to things like grids of arrays dynamically changing with time. Basically, the new quantum technology lets us use the new video prediction technology with more power. You need more power because it’s very complex. When you try to understand the dynamics of a complicated nonlinear video stream like an image of the sun, and you try to predict a model of that stuff, you have to train a network to minimize a measure of error. This is a very difficult mathematical problem. And with a new quantum technology, we can solve those problems a million times faster than you could with any classical computer. The usual quantum computers can’t do it either. You need a new generation of quantum computing. Find the error minimizer and learn how to predict the time series. That’s the technology we want to prove. And, by the way, that technology has many other uses. Pictures of the sun are not the only things we might want to predict. If you look at the sky and you want to know whether there—s a missile coming to hit you, this technology can be used for that. This month, a lot of people are wondering how we get a warning of missiles. We can look in the sky and do pattern recognition and prediction much better than we could with old technology. So, I’m trying to get that video camera learning technology deployed.SPENCER: That’s the electromagnetic pulse — EMP, electromagnetic pulse, right? For 10 years. I’ve been meaning to get an article about that for the magazine.
WERBOS: Solar flares are only one of the major uses for this. There’s a national security aspect. There are nuclear weapons and terrorist weapons.
SPENCER: Could you use it to predict volcano eruptions? There—s one in Iceland now that may blow up, and they would like to know when.
WERBOS: Maybe. I don—t know enough about volcanology to say, but our near goal is to predict time series with a graph of spatial dimensions, based on learning. Anything Deep Mind can predict (which is a lot!), this can predict better — but how much better? That depends on the system and the data.
SPENCER: You make it sound like this is just a little thing that you can create in a couple of hours that could be up and running next week. But I don’t think so. Tell me about reality.
WERBOS: The reality is that people in Israel are worried about missiles this week. A friend of mine from Israel talked about their fear of being hit by missiles. The goal is to save lives. We’re not talking about attacking people.
I’d prefer to go straight to the United Nations, but the United Nations is slow. We have had discussions between Project Amaterasu and Indian investors and semiconductor people. If we got the right people to work for us are — and some of them are doing it pro bono — it would only take a month or two to demonstrate on a quantum simulator that this hardware has the benefits that we claim. And once you have that, you can turn it over to quantum simulators that exist. And my guess is, if we really try and use what we can do, we’re like three or four months away from being able to detect these missiles.SPENCER: What are they expecting to happen in Israel?
WERBOS: Israel and the US have made a big deal. The US will send them the Iron Dome. On TV I could see the Prime Minister of India, Narendra Modi, this guy Netanyahu, and I could see Blinken, the Secretary of State, shaking hands. Yeah, we’re going to send you the Iron Dome, without money going through Congress. As soon as we get Congress together, we’re planning to move as fast as we can get to get money to pay for Iron Dome, to go to Israel to expand their ability to resist all of the missiles.
Hamas is flooding Israel with missiles now. Some of my friends tell me that the missiles are coming from Iran. The existing Iron Dome system isn’t good enough to meet Israel’s needs. It can save some lives. And it’s not just Israel. We need a better defense system to protect all the people of the world and, right now, Iron Dome isn’t the best. I know what the technology is. I was in touch with a ballistic missile defense people when they were putting it together. They didn’t put together the second generation. And what we have now is many generations better than that. So, if we can deploy the new technology quickly it might save millions of lives. A lot of the solution is in Canada, which has the option to become the world leader in this right now. The Indians are angling for it, but Canada could actually take the lead in this technology, which has commercial applications as well as defense applications. I know a lot of the people who don’t want to be quoted, but let’s just say the world is in a very dangerous position right now. It involves missiles from Hamas, and missiles from everybody who gets them from Iran. I think Iran has an axis with six players and every one of those six has access to Iranian missiles. Right now, the Houthis are also proudly sending missiles now from Yemen to Israel — flying right now, killing people right now. Hezbollah has promised to come in and they’re pretty serious. And then there are three others. If they have more missile defense, they’ll be able to wait a little longer. And if they don’t have missile defense, there will be missiles in all directions. Anything that reduces the damage done by missiles will be better for all of humanity.I went to the UN first. You were cc—d on some of my emails. The Secretary General said: “Let’s build a new UN agency for AI.” And some of us said, “Why don’t we put a missile warning system into the new UN agency? Why don—t you make the technology available, open, and transparent to the entire world? If everybody has warning, everybody will be safer. Let’s do it that way.” But ironically, Russia and China have UN representatives who are political bureaucrats. They say, “No, the UN can’t have this technology. We don’t want people to know if we attack them.” I don’t think they understood that the technology is a monopoly owned by the United States and a few of its allies, so Russia will get less warning than it would as an open global system. If the technology is open and global, Russia and China would be better off but they wouldn’t let it happen. I’m hoping the people in Moscow and Beijing will start realizing, Let’s look at the world as a whole. Let’s talk to the terrorists. But first we’ve got to develop the technology. And the best way to develop the technology is to go fast. US India, Canada, New Zealand. Australia. We all have great capabilities.