OpenAI released a new software platform on Monday to help A.I. learn how to interact with games, apps, and websites like we do. It’s called Universe, and it’s supposed to create more well-rounded A.I. to replace the more limited systems of today.
Microsoft, Nvidia, and others have granted Universe access to their games. This will allow A.I. to play the games, learn how they work, and apply whatever knowledge they’ve gleaned to other titles. The hope is that this will make it so A.I. doesn’t have to be re-trained every time it has to do something new.
“Our goal is to develop a single A.I. agent that can flexibly apply its past experience on Universe environments to quickly master unfamiliar, difficult environments,” OpenAI’s developers said in Universe’s announcement, “Which would be a major step towards general intelligence.”
OpenAI explained that it’s not enough for A.I. to beat the best human Go player in the world at his own game, which Google’s DeepMind did in March, if it can’t also master games like Magic: The Gathering or even grok how a simple board game works.
Those limitations exist because A.I. is currently trained to do one thing very well. Researchers teach one A.I. how to recognize objects, for example, while teaching others how to read lips. The tasks are related but one A.I. is unlikely to be able to do both.
OpenAI will create more generalized A.I. by teaching it how to analyze pixels, interpret what the computer is trying to convey, and use a virtual keyboard and mouse to interact with it. The A.I. is essentially getting good at playing video games, not mastering a single title to the exclusion of all the others.
Universe will also teach A.I. to apply those same skills to using websites. Right now the platform is focused on simple tasks like clicking buttons, for example, but OpenAI said it wants to teach them how to do more complicated things like Google for new information or manage somebody’s email and calendar.
This project grew out of OpenAI Gym, which tasked developers with building A.I. that could set new high scores for Atari 2600 games. Gym was meant to help coders focus their attention on diverse problems and use a common set of tools when programming A.I. systems.
Universe and Gym are part of OpenAI’s efforts to better understand A.I. systems. The group identified four key areas to research: finding covert systems; building simulations where A.I. can interact with each other; using A.I. for cyber defense; and teaching A.I. to program.
OpenAI also wants to help develop friendly A.I. that won’t have a bad outcome, as Tesla CEO and [OpenAI financial supporter](https://goo.gl/BkUFBj Elon Musk put it, which is where Universe and Gym come in. Those programs teach A.I. how to play games, not act like SkyNet.
Universe is open to developers now. OpenAI is asking more game companies, website operators, and others to give their permission for the platform to use their properties. The group plans to release a transfer learning benchmark “in upcoming weeks” to make sure Universe works.
