Microsoft Xbox recently introduced Muse, their latest generative AI model crafted for what they call “gameplay ideation.” Accompanied by an article on Nature.com, a blog post, and a YouTube video, this announcement has certainly created some buzz. But if you’re scratching your head over “gameplay ideation,” you’re not alone. Microsoft explains it as generating game visuals, controller actions, or a combination of both, but don’t expect it to revolutionize or bypass the traditional game development process anytime soon.
However, the technology behind it offers a glimpse into the future. Picture this: the model was trained on a massive scale using H100 GPUs. Imagine needing one million updates to stretch a mere second of actual gameplay into nine additional seconds of reactive, engine-accurate gameplay. The data fueling this process mainly came from existing multiplayer sessions.
What’s truly staggering is the setup required to make Muse tick. Rather than using a single PC, Microsoft unleashed a hundred Nvidia H100 GPUs for this training—a venture costly in terms of both finances and power, yet it only churns out nine seconds of gameplay at a modest resolution of 300×180 pixels.
One of the most striking demonstrations of Muse was its ability to clone existing props and enemies, complete with functional mimicry. It’s a fascinating capability but begs the question: why not just use conventional development tools to insert these elements?
Ultimately, while Muse can admirably uphold object permanence and mirror the original game mechanics, these innovations seem extravagant compared to the traditional, more efficient game development methods.
Although Muse might evolve to achieve more impressive feats in the future, right now, it joins a long lineup of projects aiming to craft gameplay purely using AI. While there’s some appreciation for the engine’s accuracy and preserved object permanence, the method remains an inefficient and perplexing approach to game development, testing, or even play. After delving into the details, it’s hard to fathom why anyone would prefer this AI-driven path.
