Skip to content

Title: Google Mülls Over Creating AI World Simulation

In contemplating the notion, is the world truly worth replicating?

Title: Google Mülls Over Creating AI World Simulation

Google's DeepMind research lab is beefing up its team to construct generative models that can mimic the real world. The brainchild of Tim Brooks, a former OpenAI colleague who famously built Sora, the video generator, this project is a key element of Google's quest for artificial general intelligence. According to job postings related to the new team, Brooks and his crew are eager to develop colossal generative models capable of simulating the physical universe.

Brooks and his team have audacious ambitions to craft mammoth generative models that mirror the real world. Based on the role descriptions, their endeavor to build world models will span various domains, such as visual reasoning, simulation, planning for embodied agents, and real-time interactive entertainment.

Essentially, a world model endeavors to replicate how the universe operates. Whilst generative models like Sora can recreate objects they've encountered in training data, they lack a genuine comprehension of why these things happen. As a result, they can successfully produce a baseball-throwing video, but they don't understand the underlying physics. World models aim to endow AI with sufficient information to get a grasp of how actions unfold and the possible outcomes.

During a speech at Hudson Forum, Meta's chief AI scientist Yann LeCun defined world models as follows: "A world model is your personal understanding of how the world operates. It allows you to envision a series of actions you might take, and your world model will predict the implications of this sequence of actions on the world."

World models are challenging to build for several reasons, including the colossal computational power necessary to run a model and the lack of sufficient, accurate training data. However, it appears that DeepMind's team is bent on extending the realm of world models. Their vision includes creating real-time interactive generation tools derived from the models and potentially integrating their world model into Google's Gemini, a large language model.

It's plausible that DeepMind will target the video game industry to put their world models into action. The new team's job descriptions reference collaboration with Google's Genie and Veo teams. Genie represents Google's Sora-like video generator, while Genie is an existing world model capable of simulating 3D environments in real-time. The video game industry has shown a growing interest in adopting AI tools, which could displace thousands of workers. A survey by CVL Economics revealed that over 86% of gaming firms have already adopted generative AI, and nearly 15% of all gaming jobs could be disrupted by 2026.

The emphasis on creating colossal generative models by Brooks and his team is deeply rooted in their ambition to incorporate artificial intelligence into the realm of technology, specifically focusing on the future of artificial-intelligence-driven world models. These models, when successfully built, will not only mimic the real world but also understand the underlying physics, surpassing the limitations of current generative models like Sora.

Read also:

    Comments

    Latest