LET A THOUSAND virtual worlds rain down from the clouds. Or rather, the cloud. That’s the call from Google as it gets behind a tiny British startup called Improbable.
Founded by two Cambridge graduates and backed by $20 million in funding from the venture capitalists at Andreessen Horowitz, Improbable offers a new way of building virtual worlds, including not just immersive games à la Second Life or World of Warcraft, but also vast digital simulations of real cities, economies, and biological systems. The idea is that these virtual worlds can run in a holistic way across a practically infinite network of computers, so that they can expand to unprecedented sizes and reach new levels of complexity.
So far, the startup has shared its technology with just a handful of coders and companies. But today, Improbable joined forces with Google to offer its creation, called SpatialOS, to anyone who wants it.
You can think of SpatialOS as a cloud computing service for building virtual worlds, whether they run on desktop computers or VR rigs like the Oculus Rift. This service runs atop the Google Cloud Platform, the tech giant’s growing cloud computing empire, and the two companies just opened a SpatialOS alpha program that lets coders prototype and test their own virtual worlds. When the beta launches in the first quarter of next year, a separate program will provide coders with free time on Google’s cloud as they hone these virtual worlds for release onto the internet at large.
On one level, this partnership allows Google to promote its cloud services as it challenges rivals like Amazon Web Services and Microsoft Azure. In providing of the cloud infrastructure that underpins Pokemon Go, Google has seen the thirst for virtual and augmented reality firsthand, and now, with Improbable, it hopes to push even further into this burgeoning market. But this partnership also points to something bigger down the road: the future of AI.
As developers build more complex virtual worlds, this provides AI researchers with better ways of training the next generation of artificial intelligence. Games have long offered a proving ground for AI, but SpatialOS can help expand these proving ground, providing a way not only for AI agents to learn the successor to Second Life, but to navigate real city streets or even trace the path of contagious disease.
If AI agents set loose in virtual simulations of the real world sounds like Gibsonian science fiction, consider Universe, an AI training ground just recently unveiled by OpenAI, the lab bootstrapped by Tesla CEO Elon Musk and Y Combinator president Sam Altman. Universe is a software platform where researchers can train AI agents to use any application, from games to web browsers to protein folding simulations—anything humans can do on a computer. In theory, you could train agents to navigate any of the beefed-up virtual worlds built with Improbable.
That opens AI research to a few frontier. Game designers Dean Hall (creator of Day Z) and Henrique Olifiers (CEO of Bossa Studios, maker of World Adrift) say Improbable allows massively multiplayer games to achieve unprecedented complexity and scale. And in an effort to understand the impact of autonomous cars, a UK startup called Immense Simulations is using the service to model entire cities. “We can cover really large geographical areas,” says CEO Robin North, “but still keep a high level of detail.”
In the end, such simulations could also provide training grounds for those autonomous cars. Craig Quiter, an engineer at Otto, the robo-vehicle company owned by Uber, is training AI agents on Grand Theft Auto as a stepping stone to more advanced self-driving cars. Swap Grand Theft Auto for a simulation of the city of Manchester, and you get even closer to that goal.
Improbable CEO Herman Narula stresses that today his service is mainly a way of building games. But he too sees it as a path to better AI, hinting that his company is already working with others toward this goal. If a thousand virtual worlds take shape, so too can a thousand AIs.