Just yesterday, Microsoft Xbox stepped into the spotlight with the unveiling of Muse, a “generative AI model designed for gameplay ideation.” They also released an article on Nature.com, accompanied by a blog post and a YouTube video. If you’re scratching your head over what “gameplay ideation” actually entails, Microsoft describes it as crafting “game visuals, controller actions, or both.” However, don’t expect Muse to magically replace the entire game development process just yet. Its practical uses remain pretty limited for now.
But despite these constraints, there are some intriguing insights to be had. The model was trained on H100 GPUs, and it took about a million training updates for Muse to stretch a single second of real gameplay into nine seconds of simulated, engine-precise gameplay. Most of the training input came from real multiplayer gameplay sessions, which is a fascinating choice.
Rather than running this on just one machine, Microsoft had to utilize a formidable setup of 100 Nvidia H100 GPUs. This approach is substantially more costly and energy-intensive, delivering an output resolution of 300×180 pixels, all to extend the gameplay by a mere nine seconds.
In one of the more noteworthy demonstrations with Muse, they showcased how it could clone existing props and enemies within a game’s environment, even mimicking their functionality. With so much hardware and energy consumption, the question arises: why not just use traditional development tools to introduce new assets?
Yes, it’s impressive that Muse could eventually keep track of object permanence while copying original game behaviors. However, in the grand scheme of things, this approach seems inefficient compared to the proven methods used in game development today.
Muse might have the potential for exciting future applications, but for now, it joins a long list of projects striving to recreate gameplay solely through AI. Although there’s a level of engine accuracy and object permanence at play here, this method leaves much to be desired when it comes to developing, testing, or playing games. After spending hours delving into all the available material, I’m still left wondering why anyone would choose to use this approach over traditional methods.
