DeepMind uses Gemini to guide Genie 4 simulations frame by frame

Google DeepMind's Genie 4 uses Gemini 3.8 Flash to break down user goals and evaluate each generated frame, helping its interactive simulations retain their original context, Ars Technica reported. The model generates environments in real time as users explore them, without requiring pre-rendered scenes or programmed interactive elements. Users can change weather, add elements or move the camera during a simulation, with Gemini steering subsequent output. DeepMind says world models could provide training environments for physical robots and support creative prototyping. Their usefulness still depends on simulation accuracy: DeepMind says Genie 4 worlds remain consistent for only a couple of minutes before losing their initial context.