Tim Brooks, a former OpenAI employee, has moved to Google DeepMind, where he is forming a new team focused on world simulation. The goal is to develop an AI model capable of representing the entire physical world and acting accordingly. This endeavor is seen as a potential path towards achieving Artificial General Intelligence (AGI).
Brooks announced the formation of this team on platform X and shared job openings, indicating that the team is not yet complete. He mentioned that the plans are ambitious and require large generative models. The team collaborates closely with those working on Google’s Gemini models, the video generator Veo, and Genie, a foundation model that can create playable 3D worlds from a single image.
The job postings emphasize the need to solve “novel problems” and scale models as much as the available computing power allows. One of the positions they are hiring for is a “Research Engineer, World Modeling” at their main location in Mountain View.
Google continues to pursue the scaling hypothesis, which suggests that generative AI models can be continually improved through scaling until they become a representation of the world. More parameters and larger datasets have historically led to better models. However, critics argue that scaling has reached its limits, as models are not significantly improving despite more data. Additionally, the world’s data is finite, and the environmental impact of ever-larger models is increasing. Some AI experts believe that achieving AGI requires new and different architectures.
One job listing explicitly states, “We believe that scaling to video and multimodal data is a crucial step on the path to artificial general intelligence.” Google anticipates that world models will be used in various areas, such as embodying AI agents and facilitating real-time interactive conversations, including in video games.
In summary, Tim Brooks is leading an initiative at Google DeepMind to create a comprehensive world simulation model. This model aims to represent the entire physical world and potentially lead to the development of AGI. The project relies on scaling generative AI models, despite some skepticism about the limits of this approach. Google believes that expanding models to include video and multimodal data is essential for progress towards AGI, with potential applications in AI agent embodiment and interactive experiences.