Biking on a narrow mountain road in India, driving through volcanic terrain and wingsuit flying over the Alps—these are just a few of the experiences that can now be virtually simulated by Google’s newest A.I. model, Genie 3. The so-called “world model” generates vast, interactive 3D environments for both humans and A.I. systems to explore, a development Google DeepMind describes as a “key stepping stone” toward advanced forms of A.I.
“This line of work (and world models in general) is very close to my heart,” said Demis Hassabis, CEO of Google DeepMind, in a post on X. “Back in the 90s when I was designing [simulation] games we could only dream of one day having tech like this,” added Hassabis, who began his career as a video game programmer.
Genie 3’s uses extend far beyond gaming. With its ability to create unlimited environments, DeepMind says the model can train A.I. agents to navigate real-world scenarios. “We expect this technology to play a critical role as we push towards AGI, and agents play a greater role in the world,” said the company.
Across Silicon Valley, companies are racing toward artificial general intelligence (AGI)—A.I. systems with human-level capabilities—by releasing ever more powerful models. Earlier this week, OpenAI unveiled GPT-5, its fastest and most advanced model to date, which the company claims demonstrates “Ph.D.-level” performance in areas like writing and coding.
World models vs. large language models
Unlike large language models (LLMs) like GPT-5, Genie 3 doesn’t generate text or code. Instead, it uses prompts to create virtual worlds that can train physical A.I. agents, such as robots and autonomous systems, for deployment in the real world.
Its capabilities include simulating an industrial bakery, where an agent must learn how to approach an industrial mixer or move to cooling racks. Through text prompts, users can instantly alter aspects of the environment, such as weather conditions, and add “what if” scenarios, such as a herd of deer crossing a ski slope, to test how agents handle the unexpected.
Genie 1 and Genie 2, released last year, could generate new environments for training agents. Genie 3 is the first version to allow real-time interaction and offers improved realism and consistency.
Google isn’t alone in advancing A.I.-powered simulations. Nvidia earlier this year launched its own world-model platform to train self-driving cars and robots. And World Labs, the startup founded by A.I. pioneer Fei-Fei Li, has raised about $230 million to fund technology that turns 2D images into interactive 3D worlds.
Despite Genie 3’s leap forward, DeepMind acknowledges current limitations. Its geographic accuracy isn’t flawless, and it struggles with enabling multiple agents to interact in the same environment. Real-time responsiveness may also raise safety concerns. For now, Google will release Genie 3 only to a small group of academics and creators to study potential risks before a wider rollout.
<
