The Google DeepMind team has announced its new AI model to generate interactive worlds. At the end of last year, we were surprised with what Genie 2 could do, and the new version presents an important leap—one that for Google is a significant advancement in the creation of General Artificial Intelligence or AGI, which can match the abilities of the best humans.
Genie 3 is the latest iteration of DeepMind’s World Model. This cutting-edge model allows users to create immersive and interactive worlds that can be explored using simple text prompts. The previous version was limited in time and capability, allowing only a few seconds of interaction, but Genie 3 promises an enhanced experience where users can explore for “several minutes.” The graphics resolution has also seen a marked improvement, now supporting 720p at 24 frames per second, a leap from earlier iterations based on Genie 2 and I see 3.
Memory is the most significant improvement in this new model. The world is generated through AI as users explore it, but when they turn around to re-examine an object, it remains unchanged. Additionally, changes made to the environment, such as painting on a wall, are preserved even after leaving the scene. This was not achievable in previous versions, as the creators indicate they did not explicitly program it to do so. According to an article in TechCrunch, Genie 3 is capable of remembering what it has previously generated, allowing it to train itself and learn how the world operates physically.
Interactivity is another essential feature that sets Genie 3 apart. Users can introduce events via additional prompts. In its promotional materials, DeepMind showcases several interactive scenarios, such as a meadow where users can dictate whether a tractor, a bear, a horse, or hot air balloons appear. This innovative feature, termed “promptable World Events,” also allows modifications to aspects like weather conditions, providing a dynamic experience.
The Importance of World Models lies in their versatility across various applications. They are invaluable for creating scenarios for real-time games, educational environments, or training AI agents. Google highlights on its blog that these advancements represent a critical step towards achieving AGI—an advanced form of artificial intelligence that many companies are racing to develop. These worlds could serve as training grounds for other AI models, including robots, where simulating real-world scenarios poses a considerable challenge.
During the Genie 3 presentation, the DeepMind team demonstrated how they placed an AI agent in a simulated warehouse and instructed it to approach specific objects like a green garbage cube. According to DeepMind, the agent’s successful navigation in these tests is a testament to how “coherent” Genie 3 remains.
Competition in the AI Space primarily revolves around consumer products, particularly chatbots, while video and audio generators have also gained traction. World models, however, are less prominent among the general public, resulting in limited competition. Nvidia’s introduction of COSMOS earlier this year highlights some emerging competition. Additionally, companies like World Labs offer similar proposals, expanding the landscape of interactive AI world models. However, Genie 3 is currently available only in Beta to a limited group of academics, leaving many enthusiasts eager for wider access.
Image | DeepMind
In Xataka | Some researchers created a company where all employees were AI agents. They did not make a quarter of the work