One core axis of Google DeepMind's research is 'SIMA', a general-purpose AI agent that perceives its environment and acts autonomously like a human. Unlike single-purpose AIs like 'AlphaGo' or 'AlphaStar' that show superhuman performance in specific games, SIMA is being developed for 'generality' to adapt to various environments. As such, SIMA does not recognize game code or text, but learns by watching 'pixels'—the visual information on the screen—and operating via keyboard and mouse, just like a human.

It also possesses the ability to understand human 'natural language commands'. If you tell it, "Get on the spaceship and fly," it recognizes the situation and executes the command even in a 3D graphic environment it has never seen before. To expand its testing environment, they partnered with the developers of the famous MMORPG 'EVE Online', testing cognition, memory, and multi-player interactions in a persistent world that requires economic systems and long-term planning, thereby advancing its cognition and response capabilities.

The other axis, 'Genie', is a 'world model' that provides environmental physics simulations. First unveiled in February 2024, Genie has evolved rapidly, with its next version, titled Genie 3, appearing just 18 months after version 1. The key to Genie 3, as mentioned by Director Mufarej, is its ability to generate fantasy and realistic 3D environments that allow for real-time interaction at 24 frames per second from a single line of text or image prompt. It demonstrates unparalleled performance by simulating complex physical laws like light reflection, gravity, and the flow of water and wind, while showing long-term memory capabilities for objects that briefly disappear from the screen or for user actions.

The core of Google DeepMind's future vision begins with these two axes interlocking. It is about creating an 'AGI research flywheel' where the agent and the world model perform actions on each other to learn mutually. This mechanism operates by having the general-purpose AI agent SIMA autonomously explore and learn from the highly novel environments created in real-time by Genie. The vast data from SIMA's trial-and-error process is then used to train and improve Genie. This feedback loop accelerates AI learning within a controlled virtual environment, completely eliminating real-world risks.

Regarding the future of AI research for games, Director Mufarej noted, "We don't assume that some or all of the repetitive AI functions will necessarily lead to a fun experience. However, we must treat them as questions, hypotheses, or ideas, testing them and repeating the process until we find the fun." He concluded by emphasizing that to advance the community and provide new gameplay experiences, we must move beyond mere words and take action, noting that the human role in inspiring the entire gaming industry to explore new experiences and share them with the world is paramount.