Google DeepMind has unveiled Genie 3, a significant advancement in generative artificial intelligence. This new "world model" builds upon its predecessor, Genie 2, and allows for the creation of real-time, interactive simulations from simple prompts or images. The environment is continuously generated, enabling dynamic changes such as adding or altering objects, modifying weather conditions, or introducing new characters – features DeepMind calls "promptable events."
The potential applications of Genie 3 are vast. While the gaming industry remains somewhat sceptical, the ability to generate alterable 3D environments could revolutionise game development, offering developers new avenues to explore concepts and level designs. However, DeepMind also emphasises Genie 3's role as a research tool. Games have historically been crucial in AI development, providing challenging environments to measure progress. World models like Genie 3 elevate this by generating interactive worlds frame by frame, allowing for the refinement of AI models, including "embodied agents," in simulated real-world scenarios.
One of the key limitations in achieving artificial general intelligence (AGI) is the scarcity of reliable training data. Genie 3 addresses this by providing essentially unlimited interactive worlds for training AI agents. This new iteration offers significantly higher visual fidelity and operates in real-time, allowing users to navigate the simulated world in 720p resolution at 24 frames per second using keyboard input. Critically, Genie 3 boasts improved memory, maintaining visual consistency for multiple minutes, a substantial upgrade from Genie 2's limited retention.
Despite these advancements, Genie 3 isn't perfect. It cannot simulate real-world locations, and the generated environments are unique and non-deterministic, making it prone to AI hallucinations. While accuracy has improved, it can still produce incorrect video elements, such as flawed human locomotion. Furthermore, the integration of AI agents within these world models remains limited. While agents can move within the simulated world, they lack the high-level reasoning necessary to alter the simulation itself. DeepMind is currently exploring ways to enable multiple AI agents to interact within a shared environment. Access to Genie 3 is currently limited to a group of experts and researchers, with plans to broaden access in the future.
Fuente Original: https://arstechnica.com/ai/2025/08/deepmind-reveals-genie-3-world-model-that-creates-real-time-interactive-simulations/
Artículos relacionados de LaRebelión:
- Googles Genie 3 AI Creating Real-Time Interactive 3D Game Worlds
- MIT Solve Empowering Startups with a Purpose at MIT
- Motorola Razr Ultra Grab a Folding Phone Bargain with a 200 Discount
Artículo generado mediante LaRebelionBOT