World simulation: Google founds new AI team
An AI that can simulate the entire physical world – Google's new "ambitious task" to achieve AGI.
(Image: Outflow_Designs / Shutterstock.com)
Tim Brooks left OpenAI last fall and moved to Google DeepMind. There, the AI expert is building a new team that will work on a world simulation. The idea behind this is nothing less than an AI model that can map the entire physical world – and act accordingly. It is a possible path to Artificial General Intelligence (AGI).
Brooks has announced the formation of the team at X and at the same time published job advertisements, so the team is not yet complete. Brooks also writes there that these are "ambitious plans" that require huge generative models. The team is, of course, working closely with those responsible for Google's Gemini models and the Veo video generator, as well as Genie, the Foundation model that can create playable 3D worlds from a single image. The job advertisements state that "novel problems" need to be solved and models need to be scaled as far as the greatest possible computing power allows. For example, the company is looking for a "Reasearch Engineer, World Modeling" at its headquarters in Mountain View.
Videos by heise
Google has therefore not abandoned the so-called scaling hypothesis. This assumes that generative AI models can be continuously improved through scaling until they become a reflection of the world. More parameters and larger amounts of data have so far resulted in better and better models. However, critics say that scaling is already reaching its limits. Models are no longer really getting better despite more data. What's more, the world's data is finite and the environmental impact of ever larger models is also increasing. Some AI experts say that new, different architectures are needed to achieve AGI.
However, one of the job postings specifically states: "We believe that scaling to video and multimodal data is a critical step on the path to Artificial General Intelligence." Google assumes that world models will be used in numerous areas, such as for the embodiment of AI agents and interactive conversations in real time, for example in computer games.
(emw)
