Thanks to advanced new machine learning techniques, artificial intelligences are better at performing human tasks than ever. AIs can tell you what’s in your photos, beat you at chess, design typefaces, dream up entirely new cities, and even tweet like Donald Trump–often better than the average person.
But AIs aren’t as smart as people. Not yet. That’s because AIs lack general intelligence. They can’t apply what they’ve learned from one problem to another–which is why even the best AIs are idiot savants: really smart in one arena, and dumb as sticks in all others.
So how can AIs reach this elusive general intelligence? OpenAI–an artificial intelligence research nonprofit backed by Microsoft, Elon Musk, and Peter Thiel–thinks it involves AIs playing video games. Thousands of video games. And that’s just the beginning.
OpenAI’s new software platform, Universe, allows AI agents–any artificial intelligence that is programmed to act upon its environment and work toward goals–to test itself on free Flash games, many provided by the Flash game portal Newgrounds, as well as more popular games like Grand Theft Auto V, Portal, World of Goo, RimWorld, Mirror’s Edge, and more. Basically, it works by allowing AIs to connect to a virtual machine installation of a game and control it like a person, scanning the pixels to determine when it is winning and when it is losing. Better yet, it does it without any special programming on the developer’s part for an agent to play a game: All an AI needs to be able to do is use its “eyes,” just like a person. An agent can even keep track of its high score, using that as a reward to gauge its own performance. It’s like an obstacle course for training AI.
Since the Universe platform essentially just gives AI agents a framework to control virtual machines–which can run any operating system or software–the future of Universe isn’t limited to games. It can technically run any software, including productivity apps and websites. So programmers won’t necessarily just train their agents on Grand Theft Auto V; they’ll train them on Excel, Amazon, and Photoshop. Android support also works, and iOS is coming soon. Anything that runs on a computer is something that Universe can potentially train an AI on, and anyone–or at least, anyone with an AI–can access it for free.
“It’s not just games,” says OpenAI’s Catherine Olsson, an alum of MIT’s Brain & Cognitive Science group. “Our goal is that anything a human can do on a computer, an AI agent should also be able to do.”
So is the point to create a single super AI, one robot to rule them all? Not really, says Jack Clark, communications director at OpenAI. Rather, it’s to give multiple companies and institutions working on machine learning the ability to test whether or not their agents have general intelligence. “Look at this way,” says Clark. “If I asked you: How would you characterize your intelligence? No matter your answer, it would be totally subjective.” In a way, Universe removes the subjective guesswork, because if someone develops an AI that can play video games as well as it can book a flight, “it must have general intelligence.” And if it has general intelligence, an AI is on the path to being as good as a person.
That opens the door to all sorts of possibilities. Digital assistants who are as good as people at scheduling your tasks, even if you’ve never completed a particular task before. Software that can test websites, or provide UX insights for designers and developers. Self-driving cars that can navigate traffic patterns they have never encountered. Even AIs that can adapt alongside humans, instead of just falling apart the second you ask them for something they don’t know how to do. In other words, it’s a step toward flexible computer intelligence.
Universe won’t teach AIs to do these things by itself. Computer researchers will still need to figure out how to program them to be as good at learning as humans. But it will provide the obstacle course. And if everything goes well? It could help establish the first truly “human” AIs, although it’s worth noting that there are many–including one of OpenAI’s backers, Elon Musk–who worry that AIs that are as smart as humans might prove uncontrollable, and therefore, dangerous. HELLO PROFESSOR FALKEN. SHALL WE PLAY A GAME?