Google DeepMind is forming a new artificial intelligence (AI) research group that will develop AI models that can simulate physical environments to train robots and create realistic game universes. The initiative will reportedly be led by Tim Brooks, former co-leader of the Sora project at OpenAI, who joined DeepMind back in October.
Image source: Google DeepMind / Unsplash
«World modeling is a relatively new area of AI that can find applications in various fields. The direction can be used to create real-time interactive media environments for video games and movies, as well as to develop realistic training scenarios for robots and other AI systems.
DeepMind is currently actively seeking research engineers and scientists to work in its laboratory by posting vacancies on the Greenhouse website. The team’s main tasks will include training models at large scale, curating training data, and exploring ways to integrate models with multimodal language models. “We believe that scaling pre-training on video and multimodal data is a critical step towards artificial general intelligence,” the job description states.
Despite its ambitious plans, DeepMind has several competitors that already have an advantage in developing “world modeling” technology. These include Nvidia’s Cosmos platform for the development of physical AI and the World Labs startup created by Fei-Fei Li, who is called the “godmother of AI.” The new DeepMind team will work alongside Google’s existing projects, including flagship AI models Gemini, video generator Veo, and Genie, a previously developed world model for simulating 3D gaming environments in real time.
It is worth saying that Google is striving to achieve AGI before its competitors and the race to be first in harnessing the power of superintelligence is gaining momentum. Thus, OpenAI CEO Sam Altman recently said that the company is close to achieving AGI, and that autonomous AI agents could begin to be actively included in workflows in the coming year.