Google DeepMind’s New AI Agent Learns, Adapts and Plays Games Like a Human

Google DeepMind has launched SIMA 2, a new AI agent designed for 3D worlds, which represents a significant advancement towards Artificial General Intelligence (AGI). Unlike its predecessor, SIMA 1, this version enables the AI to reason, explain its plans, and learn through self-directed play. SIMA 2 utilizes Google's Gemini model to interpret high-level goals, engage in conversations, and adapt to new games automatically. It has shown a 65% task completion rate, significantly higher than the 31% achieved by SIMA 1, and can successfully operate in newly generated 3D environments. Through iterative learning from tasks it creates, SIMA 2 enhances its capabilities continually. While the AI shows promise, challenges remain, such as managing very long tasks and interpreting complex visuals. Nonetheless, DeepMind anticipates that advancements made with SIMA 2 could translate into future robotics applications, paving the way for more integrated AI embodiments in daily actions.

Source 🔗