Microsoft trained an AI model on a game no one played

World models — AI algorithms capable of generating simulated environments — represent one forefront of machine learning. Today, Microsoft published new research in the journal Nature detailing Muse, a model capable of generating game visuals and controller inputs. Unexpectedly, it was born out of a training set Microsoft built from Bleeding Edge.

If, like me, you had completely erased that game from your memory (or never knew it existed in the first place), Bleeding Edge is a 4 vs. 4 brawler developed by Ninja Theory, the studio better known for its work on the Hellblade series. Ninja Theory stopped updating Bleeding Edge less than a year after release, but Microsoft included a clause in the game’s EULA that gave it permission to record games people played online. So if you were one of the few people who played Bleeding Edge, congratulations, I guess: you helped the company make something out of a commercial flop.

So what's Muse good for anyway? Say a game designer at Blizzard wants to test an idea for a new hero in Overwatch 2. Rather than recruiting a team of programmers and artists to create code and assets that the studio may eventually scrap, they could instead use Muse to do the prototyping. Iteration is often the most time-consuming (and expensive) part of making a video game, so it’s easy to see why Microsoft would be interested in using AI to augment the process; it offers a way for the company to control runaway development costs. That’s because, according to Microsoft, Muse excels at a capability of world models the company calls persistency.

"Persistency refers to a model’s ability to incorporate (or 'persist') user modifications into generated gameplay sequences, such as a character that is copy-pasted into a game visual," says Katya Hofmann, senior principal research manager at Microsoft Research. Put another way, Muse can quickly adapt to new gameplay elements as they’re introduced in real-time. In one of the examples Microsoft shared, you can see the "player" character immediately react as two power-ups are introduced next to them. The model seemingly knows that the pickups are valuable and something players would go out of their way to obtain. So the simulation reflects that, in the process creating a convincing facsimile of a real Bleeding Edge match. 

According to Fatima Kardar, corporate vice president of gaming AI at Microsoft, the company is already using Muse to create a "real-time playable AI model trained on other first-party games," and exploring how the technology might help it bring old games stuck on aging hardware to new audiences. 

Microsoft says Muse is a "first-of-its-kind" generative AI model, but that’s not quite right. World models aren’t new; in fact, Muse isn’t even the first one trained on a Microsoft game. In October, the company Decart debuted Oasis, which is capable of generating Minecraft levels. What Muse does show is how quickly these models are evolving. 

That said, there's a long way for this technology to go, and Muse has some clear limitations. For one, the model generates visuals at a resolution of 300 x 180 pixels and about 10 frames per second. For now, the company is releasing Muse's weights and sample data, and a way for researchers to see what the system is capable of.

This article originally appeared on Engadget at https://www.engadget.com/ai/microsoft-trained-an-ai-model-on-a-game-no-one-played-160038242.html?src=rss https://www.engadget.com/ai/microsoft-trained-an-ai-model-on-a-game-no-one-played-160038242.html?src=rss