The Google DeepMind team introduced the new version of its AI agent SIMA (Scalable Instructable Multiworld Agent), initially launched in March 2024. The new SIMA 2 now utilizes the Gemini artificial intelligence for the first time.
According to the company, the agent can now go beyond merely following instructions, “understanding the overall user goal, performing complex reasoning, and skillfully executing goal-oriented actions in games.” Moreover, it can handle even those games it hasn’t encountered before. The agent can describe steps taken to achieve goals, respond to user questions, and evaluate its behavior and environment.
SIMA 2 understands tasks composed of multiple steps and multimodal prompts, such as sketches, multilingual text, and even emojis. Furthermore, the agent can transfer learned concepts, for example, applying knowledge about “resource harvesting” from one game to “crop harvesting” in another.
With recent insights, SIMA 2 has shown significant improvements in multitasking and cross-contextual learning capabilities, making it a powerful tool for developers aiming to create sophisticated gaming environments. As of now, SIMA 2 is available to a select group of researchers and game developers as a limited introductory version. Industry experts have hailed it as a breakthrough in AI gaming, citing its potential to revolutionize user interaction with more intuitive and dynamic gameplay experiences.
