VDe Vector

Google DeepMind unveils model for interactive worlds

Genie 4 creates realistic worlds with visuals, sound and interaction, according to the company, but is initially available only on a limited basis.

Google DeepMind has announced Genie 4, an experimental model that can generate interactive worlds and, according to DeepMind, adapt them during use. The system is initially available to trusted testers and has not yet been released broadly to consumers.

According to DeepMind, Genie 4 is what is known as a world model: a system that not only creates images but also attempts to model how an environment responds to movements and instructions. DeepMind says users can describe changes to a simulation in natural language, such as different weather, a new camera setting or an additional object.

As well as visuals, the model generates ambient sound, music-like elements and interactive dialogue. DeepMind says the simulations move more smoothly and retain greater visual coherence when users navigate through the environment.

For now, the technology is being rolled out as a research preview for trusted testers. DeepMind says it is collecting feedback from testers with a view to potentially making Genie 4 available to consumers and developers. A broad market launch has therefore not yet been announced.

DeepMind cites applications in game development, cinematic previsualisation, education, robotics and simulations. Developers could explore ideas quickly before working them out in a traditional game engine. These are potential applications, not proven productivity effects.

The company presents Genie 4 as an important step towards so-called world models and artificial general intelligence. That characterisation comes from DeepMind itself. The announcement is not independent peer-reviewed research and therefore does not in itself provide independent confirmation of the model’s performance.

DeepMind says generated video and audio will receive an invisible SynthID watermark. The company also reports that the model was tested for safety aspects before the limited rollout. The public announcement gives no complete overview of the test results or of performance in misuse scenarios.

The development may be relevant to the media sector and games industry because generative models are affecting increasingly more parts of production processes. At the same time, questions remain about copyright, the recognisability of synthetic images, control over violence or misinformation, and the cost of running such models.

For now, Genie 4 is primarily a company demonstration with limited access. Whether the technology is reliable enough for professional applications will have to emerge from independent tests and experiences outside the group selected by DeepMind.

Respond to this article

← Back to the edition