Microsoft Xbox Unveils an Overly Complex AI Model for ‘Gameplay Ideation’

Microsoft Xbox has taken a step into the future with the introduction of Muse, their new generative AI model crafted for what they call “gameplay ideation.” Accompanying this launch is an elaborate article on Nature.com and a blog post featuring a YouTube video for a more comprehensive dive. If “gameplay ideation” leaves you scratching your head, Microsoft describes it as the ability to generate game visuals and controller actions. However, before you get too excited, let’s clarify—Muse isn’t about to replace traditional game development processes anytime soon.

Even with its limitations, the data surrounding Muse offers some intriguing insights. The model was trained on a large scale using H100 GPUs, and it took around a million updates just to extend one second of real gameplay into a nine-second stretch of simulated gameplay that mirrored the game engine’s accuracy. Most of the training material came from existing multiplayer sessions, which is noteworthy.

Rather than relying on a single PC setup, Microsoft went all-out to train Muse across a cluster of 100 Nvidia H100 GPUs. This substantial investment in resources and energy only yielded nine additional seconds of extrapolated gameplay at a modest resolution of 300×180 pixels.

A significant achievement showcased was Muse’s ability to replicate existing props and enemies within a game environment, mimicking their original functions. Yet, the cost and effort involved beg the question: why go through all this when conventional development tools can simply spawn new enemies and props?

Muse’s capacity to maintain object permanence and accurately duplicate original game behaviors might be noteworthy. However, when you consider the well-established, efficient processes currently employed in game development, Muse’s applications seem extravagant and not quite justified, at least not yet.

While it’s true future iterations of Muse might hold more promise, for now, it rests among a myriad of other ambitious projects attempting AI-driven gameplay simulation. Sure, there’s a certain level of engine precision and object permanence present, but it seems an inefficient solution for game development or testing. Despite hours of examining related materials, I struggle to see its practical use in the gaming sphere today.

If you’re curious to stay updated with cutting-edge tech and detailed reviews, subscribing to Tom’s Hardware newsletter ensures you won’t miss a beat.