Recently, Google announced the formation of a new team dedicated to developing AI models that simulate the physical world. This team is part of Google's AI research division, DeepMind, and will be led by Tim Brooks, one of the co-leaders of OpenAI's video generator Sora. Brooks joined DeepMind last October and shared this news on platform X.
Brooks stated in his announcement that DeepMind plans to build large generative models to simulate the real world, and he is currently recruiting for the new team. According to the job listing link provided in his announcement, the new team will work closely with Google's Gemini, Veo, and Genie teams to address "new critical challenges" and push the computational scale of these models even further.
Gemini is Google's flagship AI model series used for tasks such as image analysis and text generation; Veo is Google's own video generation model; and Genie represents Google's efforts in the realm of world models, capable of simulating games and 3D environments in real time. The latest Genie model previewed by Google in December can already generate a wide variety of playable 3D worlds.
The job listing mentions that given the importance of scaling AI training on video and multimodal data for achieving artificial general intelligence (AGI), the new team will focus on developing "real-time interactive generation" tools based on these models and researching how to integrate them with existing multimodal models like Gemini. AGI refers to AI capable of performing any task a human can do.
World models hold potential across various fields, including visual reasoning and simulation, planning for physical agents, and real-time interactive entertainment. Currently, multiple startups and major tech companies, including World Labs founded by renowned AI researcher Fei-Fei Li, Israeli startup Decart, and Odyssey, are pursuing world model technologies. These companies believe that world models could be used to create interactive media such as video games and films, as well as run realistic simulations like training environments for robots.
However, this technology has sparked mixed reactions among creative industry professionals. A recent survey found that studios like Activision Blizzard are using AI to cut costs, increase production efficiency, and compensate for staff turnover. Meanwhile, a 2024 study commissioned by the Hollywood animators' and cartoonists' union predicts that over 100,000 jobs in film, TV, and animation industries in the US will be impacted by AI technology by 2026.
Some emerging startups in the world model space, like Odyssey, have pledged to collaborate with creative professionals rather than replace them. Whether Google will follow suit remains to be seen.
Moreover, copyright issues remain unresolved. Some world models seem to be developed using gameplay footage, which could lead to legal action against companies for using unauthorized videos. Although Google, as the owner of YouTube, claims it has the right to use YouTube videos for model training according to its terms of service, the company has not disclosed which specific videos were used for training.
