Google DeepMind unveils Genie 4 world model with audio and real-time steering

What's Happening

• Google DeepMind released Genie 4 in a research preview, introducing real-time audio generation and the ability to simulate human motion and voices.
• Users can alter simulations mid-stream using natural language prompts to change camera angles or alter the environment without restarting.
• The system leverages Google's Gemini 3.8 Flash model to follow the action, plan next steps, and adjust the environment when conditions change.
• Google plans to offer the model to Google AI Ultra subscribers soon and opened a waitlist for early access.
• Developers can use Genie 4 to generate test environments on demand, giving AI agents a place to practice multi-step tasks before real-world deployment.

Why It Matters

World models are advancing from video generators into interactive environments. Google positions Genie 4 as a foundational tool for robotics training, game prototyping, and film pre-visualization.

The Context

• Google released the predecessor, Genie 3, over a year ago, which caused consternation in the gaming industry due to its ability to generate playable video games.
• The broader world model sector is experiencing consolidation, with AMD acquiring World Labs and Rivian purchasing Third Dimension for self-driving simulations.
• Competitors like Runway launched their own playable models that respond dynamically to user inputs.

What They're Saying

• Google research director Tejas Kulkarni described a move toward hybrid world models where Gemini writes code and text and Genie turns it into a running scene.
• Google wrote in a blog post that understanding and simulating the physical world represents a crucial step toward building artificial general intelligence.

What to Watch

• Monitor how filmmakers and game developers adopt the tool for rapid prototyping, as early testers use it to maintain creative flow during pre-visualization.
• Watch for improvements in physics and object permanence, as Google acknowledges the model struggles to maintain visual consistency and accurate physics during longer sessions.
• Monitor how developers evaluate AI agents in these simulations, as unpredictable physics complicates testing for tasks that require remembering locations or returning to earlier states.

The Bottom Line

Google's Genie 4 combines real-time interactivity, audio generation, and Gemini-powered reasoning into a single world-building engine for developers and AI agents.