Google Deepmind's Genie 4 puts a modern spin on agent designs that predate LLMs
decoder ·
Google DeepMind has introduced Genie 4, an experimental interactive world model that simulates environments frame by frame in real time. Unlike previous versions, this iteration directly generates matching background audio and character dialogue alongside visual elements. Gemini models operate in the background to plan multistep tasks and ensure the simulation remains coherent as users interact via text prompts or keyboard inputs. The system allows for dynamic changes, such as altering weather conditions, while characters react naturally to their surroundings and user actions. Despite these advancements, the model currently struggles with long sessions, complex physics, and object persistence. Access is initially limited to select testers, with plans to expand to Google AI Ultra subscribers and API partners. All generated content will include invisible SynthID watermarks to identify AI origin. This technology represents a step toward hybrid world models that combine language reasoning with pixel-level simulation for future robotics training.
decoder