DeepMind taps former Sora lead to build world models

The mastermind behind OpenAI's Sora is ready to build something even bigger at Google.

Tim Brooks, who co-led Sora's development before jumping ship to Google DeepMind in October, just announced that he is assembling a brand-new team dedicated to building massive generative models that simulate the physical world.

According to job postings shared by Brooks, the new modeling unit will join forces with Google's flagship Gemini, its video engine Veo, and the Genie interactive world project to push models to the highest levels of compute.

A straight shot to AGI

DeepMind isn't being shy about what it is building.

One job description plainly states that scaling training on video and multimodal data is on the critical path to artificial general intelligence. In plain English, the idea is that an AI cannot truly understand the world just by reading words; it needs to understand physics, space, and real-time cause and effect.

The new team wants to build real-time interactive generation tools directly on top of these models, while figuring out how to plug them into existing multimodal models like Gemini.

Target use cases span visual reasoning, physical planning for embodied robots, and real-time interactive entertainment.

Google isn't alone in this pursuit. Heavyweight startups like Fei-Fei Li's World Labs, Decart, and Odyssey are already racing to build world simulators for automated game design, virtual movie production, and robot training grounds.

Yet the creative backlash is brewing fast. A recent Wired report highlighted that studios like Activision Blizzard are leaning on AI to cut corners amid heavy layoffs, while a study from the Animation Guild warned that AI could disrupt more than 100,000 film and animation roles by 2026.

Then comes the copyright headache. While rivals risk lawsuits for scraping gameplay footage without permission, Google leans on its ownership of YouTube, claiming terms of service grant it training rights, though it still won't say which videos are being fed into the machines.

Why it matters

We think this is Google finally turning scattered experiments into an enterprise-grade war machine.

Until now, DeepMind had disparate pieces: Gemini handled textual and visual reasoning, Veo handled polished video clips, and Genie generated simple playable 3D games. Bringing Brooks in to tie these threads together shows that Google sees world simulation not just as a parlor trick for Hollywood, but as the foundational operating system for physical robotics and next-gen gaming.

For developers and creators, true world models could eventually collapse production timelines from months to minutes. Instead of coding complex 3D environments and physics colliders by hand, you could generate interactive virtual sandbox tests on the fly. That said, nobody needs to rush to overhaul their pipeline today; the compute costs for real-time physics generation remain monstrous, meaning broad access is still down the road.

The real friction will be YouTube data leverage. Google holds the ultimate multimodal vault, but using independent creators' footage to train an engine that might ultimately replace junior artists will inevitably trigger fresh legal and cultural showdowns.