Google DeepMind has unveiled SIMA 2, a new AI agent that integrates Gemini, Google’s language model, to excel in virtual environments. SIMA 2 represents a significant advancement over its predecessor, with enhanced reasoning abilities and self-improvement features.
Trained on diverse video game data, the previous SIMA 1 agent demonstrated proficiency in basic tasks but struggled with complex challenges. In contrast, SIMA 2 showcases remarkable progress, boasting improved success rates and the capacity to navigate novel scenarios autonomously.
Powered by the Gemini 2.5 flash-lite model, SIMA 2 embodies the pursuit of artificial general intelligence (AGI). DeepMind defines AGI as a system capable of diverse intellectual tasks, learning new skills, and applying knowledge across domains.
Emphasizing the importance of embodied agents, DeepMind researchers highlight the significance of agents interacting with physical or virtual environments. This approach mirrors human-like interactions, enabling AI to observe inputs and take actions akin to robots or humans.
With SIMA 2, DeepMind aims to push the boundaries of AI capabilities, paving the way for advancements in general-purpose robots and AGI systems. The integration of Gemini propels AI research towards broader applications and fosters innovation in autonomous decision-making.
Source: TechCrunch