Google's Genie 4 Creates Interactive Worlds with Speech

Imagine describing a world and watching it come to life around you, frame by frame, with realistic physics and people who can actually talk back to you.

That's exactly what Google DeepMind's new Genie 4 can do. The upgraded AI world model creates interactive virtual environments on the fly, without any pre-programmed elements or rendered graphics.

While it might look like a video game, DeepMind sees Genie 4 as something much bigger. The company believes world models are a crucial stepping stone toward artificial general intelligence, offering infinite training environments for physical robots and rapid prototyping tools for creative professionals.

Genie 4 brings major upgrades over last year's Genie 3, which sent shockwaves through the gaming industry. The new version uses Google's Gemini 3.8 Flash as a guiding agent, breaking down goals and evaluating each frame to keep simulations consistent for several minutes.

The physics improvements stand out immediately. Genie 4 handles aerodynamics and collisions far more realistically than previous versions, making the simulations both more useful and visually appealing.

Human characters got a major upgrade too. People in these AI-generated worlds now move more naturally and respond believably to their changing environments, eliminating much of the uncanny valley effect that plagued earlier versions.

Perhaps the most exciting addition is native audio. Genie 4 can generate ambient background sounds and respond to audio prompts, but the real breakthrough is interactive human speech that adapts as you modify the simulation.

You can still change parameters mid-stream, adjusting weather, adding elements, or moving the camera angle. Gemini helps steer these updates smoothly, similar to how reasoning models generate internal prompts to find the right outputs.

The Ripple Effect

The implications extend far beyond entertainment. Robot developers could use Genie 4 to test machines in countless simulated scenarios before risking expensive hardware. Architects might walk clients through buildings that exist only as descriptions. Filmmakers could explore scene concepts without building sets.

Access won't come cheap or quickly. Google plans to roll out Genie 4 through its Gemini API and AI Studio for developers, with pricing still under wraps.

Consumers can join the waitlist through Project Genie, but access will start with Google's $100-per-month AI Ultra plans before potentially expanding to cheaper tiers. The company is working with a small group of trusted testers now, with wider release coming in phases.

DeepMind emphasizes that world models remain an active research area, and Genie 4 simulations still lose consistency after a couple of minutes. But each version brings us closer to AI that doesn't just understand our world but can recreate and reimagine it on command.

The future might just be whatever we can dream up and describe.