Kick back and watch our AI learn to play Atari games in real-time

We may earn a commission from links on this page.

Building artificial intelligence is easier than ever before, thanks to new open-source tools. We’ve decided to tinker a little ourselves by using OpenAI’s Universe to build an AI bot that learns to play Atari games.

We’ve been streaming the AI’s progress on TVs around the Quartz office for the past few weeks and thought the world might like a peek under the hood of an AI as it gets smarter. Ms. Pac-Man is the most fun to watch, because everybody knows how the game should be played, and watching the bot get eaten by ghosts over and over can be reassuring: Not everything can be perfectly automated.

How it works

The bot is running on a regular computer. Universe, free software recently released by OpenAI, acts as a bridge between the bot and any other software on the computer. You can use one of their algorithms (like we’re doing), or build your own AI and connect it to the platform. The AI can then interact with games, internet browsers, or really anything a human would do on a PC.

Universe is meant specifically for reinforcement learning, which means the AI learns through trial-and-error over millions of iterations. In a game like Ms. Pac-Man, the AI slowly—very slowly, dying many times—learns how to navigate through corridors, the correlation between dots and points, and eventually to avoid ghosts. After it dies, the AI internalizes what worked and tries again.

For the nerds, we’re running Universe on the quad-core i5 processor of a Dell XPS with Linux Ubuntu. You don’t need a high-end GPU or rack space in a server to get started, and we set it up in just a day.

This project came out of the Quartz Bot Studio, which is experimenting with uses of bots and AI for journalism. (It’s funded in part by a grant from the Knight Foundation.) If you’re looking for more Quartz bots, check out our tech reporter Marvin Prime on Twitter