When it comes to artificial intelligence, the United States has a tradition of betting on crazy ideas.
This week, the Defense Advanced Research Projects Agency (DARPA) showcased projects that are part of a new five-year, $2 billion plan to foster the next round of out-there concepts that will bring about new advances in AI. These include efforts to give machines common sense; to have them learn faster, using less data; and to create chips that reconfigure themselves to unlock new AI capabilities.
Speaking at the event, Michael Kratsios, deputy assistant to the president for technology policy at the White House, said the agency’s efforts are a key part of the government’s plan to stay ahead in AI. “This administration supports DARPA’s commitment, and shares its intense interest in developing and applying artificial intelligence,” Kratsios said.
President Trump signed an executive order last month to launch the US government’s AI strategy, called the American AI Initiative. Kratsios, who is also the government’s deputy chief technology officer, has been the driving force behind White House strategy on AI. The American AI Initiative calls for more funding and will make data and computing resources available to AI researchers. “DARPA has a long history of making early investments in fundamental research that has had amazing benefits,” Kratsios said. “[It] is building on this success in artificial-intelligence research.”
Since DARPA’s inception in 1957, it’s had something of a mixed track record, with many projects failing to deliver big breakthroughs. But the agency has had some notable successes. In the ’60s, it developed a networking technology that eventually evolved into the internet. More recently, it funded a personal-assistant project that led to Siri, the AI helper acquired by Apple in 2011.
But many of the algorithms now considered AI were developed many years ago, and they are fundamentally limited. “We are harvesting the intellectual fruit that was planted decades ago,” says John Everett, deputy director of DARPA’s Information Innovation Office. “That’s why we’re looking at far forward challenges—challenges that might not come to fruition for a decade.”
Through its AI Next program, DARPA has launched nine major research projects meant to tackle those limitations. They include a major effort to teach AI programs common sense, a weakness that often causes today’s systems to fail. Giving AI a broader understanding of the world—something that humans take for granted—could eventually make personal assistants more helpful and easier to chat with, and it could help robots navigate unfamiliar environments.
Another DARPA project will seek to develop AI programs that learn using less data. Training data is the lifeblood of machine learning, and algorithms that can ingest more of it can leap ahead of the competition. An innovation in this area could knock out a key advantage of tech companies operating in China, for example, which thrive on their access to an abundance of data. Other projects being funded focus on designing more efficient AI chips; exploring ways to explain the decision-making of opaque machine-learning tools; and making AI programs more secure.
To some degree, though, the AI Next initiative shows how tricky it is to gauge progress and prowess in AI. Much has been made of China’s efforts, and its government has declared an ambitious plan to “dominate” the technology. Other countries have also announced AI plans, and are pouring billions into them. But the US still spends more than any other nation on technology research and development.
Total investment matters, of course—but it’s only one part of the equation. The US has long been focused on funding emerging research through academia and agencies like DARPA. And that, in turn, has shaped the technological landscape in ways that weren’t always evident at first.
Take self-driving cars, for example. A decade ago, DARPA organized a series of driverless-vehicle contests in desert and urban settings. The competitions triggered a wave of excitement about the potential for automated driving, and a huge wave of investment followed. Many researchers who took part went on to start Google’s driverless-car effort. It’s still unclear how automated driving will change transportation, but some cars, such as those sold by Tesla, already offer limited forms of automation.
“Without DARPA coming in, [the self-driving-car boom] probably wouldn’t have happened at that scale at that time,” says Peter Stone, a professor at the University of Texas who took part in the car contest. He believes it’s vital for the US government to identify an unsolved AI problem and tackle it. “It may not happen, but if it works it will have huge implications,” he says.
This new data poisoning tool lets artists fight back against generative AI
The tool, called Nightshade, messes up training data in ways that could cause serious damage to image-generating AI models.
Rogue superintelligence and merging with machines: Inside the mind of OpenAI’s chief scientist
An exclusive conversation with Ilya Sutskever on his fears for the future of AI and why they’ve made him change the focus of his life’s work.
Unpacking the hype around OpenAI’s rumored new Q* model
If OpenAI's new model can solve grade-school math, it could pave the way for more powerful systems.
Generative AI deployment: Strategies for smooth scaling
Our global poll examines key decision points for putting AI to use in the enterprise.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.