Two new AI training grounds — one from Google, the other from the non-profit OpenAI — hope to help jumpstart general AI that can quickly gain competency across a variety of tasks, from driving a car to juggling your calendar for you.

The first is Universe, from OpenAI. With just a few lines of code, you can set a given AI agent to work on almost any application.

Universe does this by essentially creating a VNC environment that is piped to the agent, meaning that the only input the computer gets is what any user would see through the screen, while it uses the same input (key presses, mouse movement) a regular computer would use.

A lot of the integrations are focused on gameplay: OpenAI stated that it already has permission to integrate games like Portal and Mirror's Edge into the platform.

But the tasks also cover more routine office work, such as booking flights, responding to emails, and more. A number of these challenges are lumped into a category called the "Mini World of Bits" :

Mini World of Bits. We first set out to create a new benchmark that captures the salient challenges of browser interactions in a simple setting. We call this benchmark Mini World of Bits. We think of it as an analogue to MNIST, and believe that mastering these environments provides valuable signal towards models and training techniques that will perform well on full websites and more complex tasks. Our initial Mini World of Bits benchmark consists of 80 environments that range from simple (e.g. click a specific button) to difficult (e.g. reply to a contact in a simulated email client).


Google opened up its own training environment today as well. Dubbed DeepMind Lab (originally called Labyrinth), the Lab is a 3D environment where the AI is given a variety of tasks such as finding apples in a maze.


  • DeepMind Lab will be open sourced this week, according to Google , and the researchers behind it encouraged others to use it as a training ground for their own virtual agents.