Yesterday, Microsoft Xbox introduced Muse, a cutting-edge AI model aimed at sparking new ideas for gameplay. This launch was accompanied by an explanatory article on Nature.com, a detailed blog post, and a YouTube video. But what exactly does “gameplay ideation” mean? Microsoft describes it as generating game visuals, controller actions, or both. However, the model’s practical applications seem fairly limited and don’t really replace the traditional game development workflow.
Despite this, some aspects of the technology are intriguing. Muse was trained using H100 GPUs, needing roughly 1 million training updates just to stretch one second of actual gameplay into nine seconds of simulated gameplay that accurately mimics the engine. The training data for Muse primarily came from existing multiplayer game sessions.
Instead of utilizing a single PC for this immense task, Microsoft relied on a cluster of 100 Nvidia H100 GPUs. This setup is significantly more costly and energy-intensive, yielding an output of just 300×180 pixels for nine additional seconds of gameplay enhancement.
One of Muse’s fascinating capabilities is its ability to duplicate in-game props and enemies, maintaining their original functionality. Given the hefty costs and energy requirements, it’s curious why such complex AI training is utilized instead of traditional game development tools to achieve similar results.
While Muse’s ability to preserve object permanence and replicate game behavior is noteworthy, its potential applications appear limited compared to conventional video game development methods.
It’s worth considering that while future versions of Muse might offer more groundbreaking possibilities, it’s currently among many projects attempting to simulate gameplay purely through AI. Despite achieving some engine accuracy and object permanence, developing, testing, or playing a video game in this manner seems inefficient. It’s baffling why anyone would opt for this method after thoroughly examining the available information.