Google DeepMind announce Genie 4 on October 27 as experimental model for interactive simulated worlds. The company say e fit generate visual scenes together with their sound, including background noise, music and dialogue wey dey respond as things change inside the scene.
Users fit navigate with keyboard and direct events with normal-language instructions. The demonstrations show camera movement, weather change and addition of people or objects. Google's account say Gemini models dey help maintain the scene context while Genie dey produce the changing picture and sound experience.
One research arrangement use Gemini 3.8 Flash to observe incoming frames, divide goal into steps and check progress. DeepMind also describe experiments wey connect Genie with Gemini 4 Argon and robotics research. Na simulated demonstrations; dem no prove say physical robot fit reliably do every task wey video show.
Trusted testers dey explore film previews, interactive game ideas, historical environments and engineering scenarios. Such uses fit help teams examine idea before dem build am. But simulation wey look convincing no automatically become verified reconstruction of historical event or engineering result wey experts don validate.
DeepMind clearly say visual details and the way objects remain fit drift, especially for longer sessions or repeated complex edits. E also list physical interaction, correct audio, readable text, spatial accuracy and long-term consistency as research wey never finish. Those limits remain part of the preview announcement.
Access first go to trusted testers. Google say Project Genie for AI Ultra subscribers and early access for Gemini API and AI Studio partners go come later; waitlist don open. The company also say generated sound and video go carry SynthID watermark to help identify content wey AI make.