Muse: Why Microsoft's AI study doesn't have much to do with game preservation
According to Xbox boss Phil Spencer, a Microsoft study shows how old games can be brought to new systems using AI. But it's about something very different.
Game material generated by Muse in original resolution
(Image: Microsoft)
Porting old games to new systems is work and costs money. So why shouldn't an AI simply learn the games and generate them from scratch on any system? This is how Xbox boss Phil Spencer envisions the future of preserving older video games. However, a recently published Microsoft study that prompted Spencer to make this prediction for the future is actually about something completely different.
Microsoft's research team in Cambridge has just published a paper in the renowned Nature magazine, in which it describes a new AI model called Muse. Muse can generate or “predict” videos paired with controller inputs based on learned game material and associated inputs – in the demo case of the “Bleeding Edge” study by Microsoft studio Ninja Theory –.
This is somewhat reminiscent of Google's research work, GameNGen. In it, Google describes how a playable version of “Doom” is generated entirely by AI. Microsoft's Muse is also intended to show how theoretical players would behave if the developers were to tweak the game design.
(Image:Â Microsoft)
Tool for trying out ideas
The core element of Microsoft's research work is the ability to insert objects into AI-generated image sections and observe how the model deals with them. In the Nature study, the researchers describe how a battery cell or a jump pad, for example, can be incorporated into the AI-generated game section. Developers can then observe how this affects the generated gameplay section.
In essence, Muse is a tool that could help developers come up with ideas. New ideas could be tried out using the AI, at least in a rudimentary way, without having to write the objects into the code at great expense. In principle, Muse could also be helpful when writing music or filming videos, the Microsoft researchers write.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externes YouTube-Video (Google Ireland Limited) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (Google Ireland Limited) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
The study does not suggest areas of application for end users. Nevertheless, Microsoft is also trying to make Muse appealing to its customers. As part of Copilot Labs, it will be possible to play “interactive adventures” generated by Muse in the future. Players will be able to play short AI-generated game sections in real time with a controller and simultaneously drag various objects into the game environment. So they do exactly what developers would do with Muse – just for fun.
Videos by heise
A far cry from playability
The AI-generated sections don't have much to do with a real video game: they run at just 10 FPS at a resolution of 300 × 180 pixels. Statements by Xbox boss Phil Spencer, who places Muse in the context of game preservation in a video, are causing irritation in the community: “You can imagine a world in which a model learns games based on gameplay data and videos and makes them portable for every platform,” says Spencer. This could be practical for older titles that no longer run on modern systems, for example.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externes YouTube-Video (Google Ireland Limited) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (Google Ireland Limited) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
This is not only another application that Microsoft's actual research work does not envisage – but also somewhat uncharitable towards old games, according to some gamers. What's more, this is simply science fiction: previous models for AI-supported imitation of video games are far from being properly playable. This applies to Microsoft's Muse, but also to Google's GameNGen or the Oasis replica of Minecraft, which you can try out for yourself via the link. All of these models spit out a few pixels at an unusable frame rate. They also have major problems building consistent game worlds in the long term.
These AI models are therefore likely to take a while before they can imitate video games in such a way that the difference to the original is indistinguishable. Who knows: maybe it will work when “The Elder Scrolls 6” is an old game.
(dahe)
