Taking A.I. to School with OpenAI


Artificial intelligence (A.I.), like human beings, needs a little instruction in order to reach its full potential. To that end, OpenAI, a non-profit “artificial research company” tasked with shepherding A.I. so that it’s friendly to humanity, has built a “school” of sorts: Universe.

Universe supports a variety of apps and games, including “slither.io” and “Grand Theft Auto V.” After the A.I.’s developer selects a particular program as a training environment, Universe launches it within a VNC (Virtual Network Computing) environment. The A.I. agent can use a virtual keyboard and mouse to control the action onscreen, and hopefully learn something in the process:


“Our goal is to develop a single AI agent that can flexibly apply its past experience on Universe environments to quickly master unfamiliar, difficult environments, which would be a major step towards general intelligence,” read OpenAI’s blog posting. If history is any indication, the ability to do anything from play a game to edit a spreadsheet could drastically accelerate the learning curve of A.I. agents; experiments with unleashing such agents on datasets have resulted in A.I. error rates dropping quickly.

Developers interested in helping out with the effort can give OpenAI permission to use their games or apps as part of Universe; they can also train A.I. agents using Universe, which will help improve the platform.

The ultimate goal here is improving A.I.’s generalized intelligence. At least for the moment, according to the OpenAI blog, “systems with general problem solving ability—something akin to human common sense, allowing an agent to rapidly solve a new hard task—remain out of reach.” While some A.I. agents have gotten very good at specific tasks, such as playing “Go” or chess, the skills they learn aren’t often transferrable to the broader world.

OpenAI’s researchers theorize that, with a higher degree of general intelligence, A.I. platforms will apply abstract problem-solving techniques to whatever issues they encounter in the wild, making them capable of tackling an ever-growing variety of tasks. Something like Universe, with its large portfolio of games and browser-based apps, seems like an ideal platform for doing so. If you’re a tech pro interested in artificial intelligence, that reason alone makes it well worth checking out.

Want to give running Universe a shot? You’ll need some knowledge of Docker to run an environment; if you want to write your own A.I. agent, you’ll also need to know your way around frameworks such as TensorFlow and Theano. OpenAI’s blog features steps for getting started. OpenAI also offers its Gym, a toolkit for building reinforcement learning (RL) algorithms, which govern agents’ decision-making and motor-control capabilities.