High Performance Computing and Storage Complex II (HRSK-II) during the official opening of the new data center of the Lehmann Center (LZR) of the Dresden University of Technology in Dresden, eastern Germany, Wednesday, May 13, 2015. AP / JENS MEYER
What the CIA’s Tech Director Wants from AI
Dawn Meyerriecks says staying ahead of Russia and China isn’t as hard as getting U.S. leaders to listen to their own artificial intelligence analysis.
Should the U.S. fear growing Russian progress in artificial intelligence? Last week, Vladimir Putin told students , “Whoever becomes the leader in this sphere will become the ruler of the world.” That caught the interest of noted AI phobe / profiteer Elon Musk who tweeted, simply and ominously: “It begins…”
But the CIA's head of technology development has a different take. Dawn Meyerriecks is less worried about rival nation states might use AI to outflank the United States than about getting U.S. leaders to believe what AI is telling them. “If I want to increase [ certainty in a particular AI-aided assessment] what goes into it? What do I need in order to make a really good assessment on the back-end because that tells me what sort of collection I need to raise confidence to go address national leadership?"
The CIA currently has 137 pilot projects directly related to artificial intelligence, Meyerriecks, the CIA’s deputy director for science and technology, told the Intelligence and National Security Summit in downtown DC. These “experiments” include everything from automatically tagging objects in video (so analysts can pay attention to what’s important) to better predicting future events based on big data and correlational evidence.
“Can we back into correlations with cause and effect that will allow us to be more predictive with what’s about to go down, like the North Koreans are about to launch this or about to do this. We have that in pockets,” she said.
When asked whether she was worried that U.S. prowess in artificial intelligence was falling behind that of China and Russia, Meyerriecks said, “I bet on the innovation and on the systems engineering of the United States every time. One of the things we talk about with In-Q-Tel [the CIA’s venture capital arm] is, as long as we’re going faster than everybody behind us, I don’t want to think about how we find them off [from catching up to the U.S. in AI]. I just want to go faster than they can keep up. If there’s a bear in the woods, you just have to be faster than the slowest person."
Russia’s interest in AI is both well-established and overhyped . Consider, on Monday, Kremlin-backed news site Sputnik ran a story about a new Russian supercomputer for deep learning, part of a “conversational” AI project called iPavlov .
“The computing power is fundamentally important for deep learning,” Mikhail Burtsev of the Moscow Institute of Physics and Technology told Sputnik. “The more powerful hardware we have, the more complex neural network architecture we will be able to work with. The complexity of the model often allows us to make a revolutionary breakthrough in solving practical problems," he said.
The hardware powering the effort was the 170-teraflop DGX-1 server from NVIDIA, the California-based maker of graphics processing unit chips. Jeff Herbst, NVIDIA’s vice president of business development, seated next to Meyerriecks, seemed amused by the news. “We sell those to everyone,” he said.
Even if U.S. industry, if not the government, retains a technological lead in artificial intelligence development, there are real contrasts in the way the U.S. and other rival nations use it. Russia has shown an increased willingness to mix artificial intelligence and guns and lethal ground robots.
The U.S. government’s approach to AI is much more conservative. For Meyerriecks, the biggest challenge in applying cutting-edge AI products and techniques to intel collection and reporting is convincing leaders in government – including the President – to accept intel that comes, at least in part, from a robot, she said.
“We produce a presidential daily brief. We have to have really, really good evidence for why we reach the conclusions that we do. One of the things that’s a challenge for the current AI community, one of the things I’m positive will get addressed, is…you can’t go to leadership and make a recommendation based on a process that no one understands.”