Google DeepMind unveiled SIMA 2, an advanced AI agent employing the Gemini model to learn and adapt in 3D video games, marking a significant leap from its predecessor. SIMA 2 can set high-level objectives, interact with users, and self-teach new skills, having honed its abilities in popular games like Goat Simulator 3. This innovation is integral to DeepMind’s quest for general AI, with aspirations for real-world applications. By transitioning from a command-based tool to a collaborative partner, SIMA 2 enhances user interaction. Training consisted of human demonstration videos alongside Gemini-generated labels, fostering a versatile agent capable of adapting to diverse game environments. Notably, SIMA 2 features self-improvement through trial-and-error, initiated by human guidance, and can practice in AI-generated worlds. This self-directed learning is crucial for scaling AI capabilities. While the focus remains on gaming, the skills acquired aim to facilitate future AI endeavors in real-world settings, underscoring the need for cautious optimism in practical applications.
Source link
Share
Read more