Last year, Google DeepMind's reveal of Genie 3 caused gaming stocks to tumble, sparking industry concern over AI-generated environments. Since then, world models have become a major focus for AI labs globally. Google is now rolling out Genie 4, an updated model that improves on its predecessor in almost every way.
Access to the model may require paying Google, but it promises more realistic physics and human simulations. DeepMind positions world models as a key element in the march toward artificial general intelligence (AGI). Unlike traditional engines, Genie 4 generates experiences frame-by-frame in real-time based on user prompts, creating infinite training environments for robots or new prototyping avenues for creatives.
Key improvements stem from Google's updated Gemini models. Genie 4 leverages Gemini 3.8 Flash as a guiding agent to break down goals and evaluate each output frame, helping keep simulations on track as users move through varying environments. However, consistency currently lasts only a couple of minutes before the model 'forgets' the original context.
Like Genie 3, the new version supports changing parameters midstream, allowing users to alter weather, add elements, or move the camera, with Gemini steering the simulation through updates similar to reasoning models refining internal prompts to zero in on correct output tokens.
Source: Ars Technica · Summarized by HeadlinesBriefing