Google DeepMind is developing Genie 4, an updated world model designed to generate interactive virtual environments in real time. The system improves on last year’s Genie 3 in several areas, including physics simulation, environmental consistency and user control. Google says the technology could eventually support robot training and rapid creative prototyping, although access to Genie 4 may initially be limited to organizations willing to pay for it.
Genie 4 generates interactive worlds in real time
World models have become a major focus for AI research labs since Google DeepMind unveiled Genie 3 last year, an announcement that sent gaming stocks sharply lower. While Genie’s output can resemble a playable video game, DeepMind presents the technology as a component of broader efforts toward artificial general intelligence (AGI).
According to Google DeepMind, a world model can create a virtual environment without a pre-rendered setting or separately programmed interactive elements. Users describe what they want, and the model generates the experience frame by frame in real time as they explore it.
DeepMind says the technology could provide virtually unlimited training environments for physical robots and create new possibilities for rapid prototyping in creative fields. Those applications depend on the model’s ability to simulate requested environments accurately. Genie 4 introduces several improvements intended to make those simulations more useful, supported in part by updates to Google’s Gemini model.
Gemini helps maintain the simulation
Genie 4 uses Gemini 3.8 Flash as a guiding agent. The system breaks goals into smaller steps and evaluates each generated frame, helping keep the simulation on course as users move between environments without losing the original context.
There are still limits to that consistency. DeepMind says Genie 4’s generated world remains coherent for only a few minutes before the model loses track of how events began.
Users can change the world while it runs
Like Genie 3, the new model supports changing parameters during a simulation. Users can alter the weather, add objects, move the camera and make other adjustments, while Gemini helps guide the model through each update. DeepMind compares this process with the way inference models generate additional internal prompts to steer the selection of appropriate output tokens.
Source: arstechnica.com