Microsoft Xbox Unveils an Impractical AI Model for ‘Gameplay Ideation’

Yesterday marked an intriguing move by Microsoft Xbox as they introduced Muse, a new tool described as a “generative AI model for gameplay ideation.” They also released an article on Nature.com, published a blog post, and even put together a YouTube video to shed light on this innovation. If you’re scratching your head at the term “gameplay ideation,” Microsoft clarifies that it means creating “game visuals, controller actions, or perhaps both.” However, let’s be real—its practical applications fall short of transforming the game development process as we know it.

Despite some limitations, the data behind Muse is rather fascinating. The team used H100 GPUs, and it took around a million training updates to stretch a single second of real gameplay into nine extra seconds of simulated gameplay, closely mimicking the game engine’s behavior. What’s more, the training data came mainly from existing multiplayer gaming sessions.

Microsoft didn’t just run this on an ordinary computer. Instead, they harnessed a formidable cluster of 100 Nvidia H100 GPUs. This setup is far from cheap and demands substantial power, yet it only manages a low-resolution output of 300×180 pixels for those extra nine seconds of gameplay.

One of the more compelling demonstrations of Muse was its ability to replicate existing props and enemies, making copies that behave just like the originals. It makes you wonder about the point of all this power and cost when standard development tools can easily introduce new game elements.

Part of what’s intriguing about Muse is how well it managed to maintain the original game’s object permanence and behavior. But even with these achievements, it feels like an extravagant way to approach game development compared to the reliable and efficient methods already in place.

Looking ahead, Muse may evolve to do much more, but it currently joins a growing list of projects attempting to simulate gameplay through AI alone. While there’s something to be said for its engine accuracy and ability to maintain object permanence, the approach seems far from practical. I find myself puzzling over why anyone would opt for this method, even after diving deep into the material.