Google DeepMind has introduced SIMA 2, the latest iteration of its general-purpose artificial intelligence research. Building on last year’s SIMA model, Google DeepMind SIMA 2 combines the ability to follow instructions with reasoning and goal-directed actions. This new agent stands out with its ability to perform complex tasks in virtual environments.

Key Capabilities of Google DeepMind SIMA 2

SIMA 2 represents a significant advancement in artificial intelligence. The new model integrates with Google’s large language model, Gemini. This integration allows it to interpret users’ high-level goals. It also has the ability to explain its actions by reasoning about its actions. The agent can describe the steps it takes in detail and evaluate its environment.

Shiftdelete.net

The AI ​​can successfully execute complex instructions even in games it has never seen before during training. Environments such as the Viking survival game ASKA and MineDojo serve as examples of this. SIMA 2 can understand long, multi-step tasks. It can also respond to multimodal prompts such as sketches, different languages, and even emojis. For example, it can use the “mining” knowledge learned in one game to “harvest” in another.

SIMA 2 has the ability to evolve independently after initial training based on human demonstrations. It generates experience data by playing games on its own. This cyclical process allows the AI ​​to learn increasingly complex tasks. However, the agent is currently limited in tasks requiring long-term planning, precise low-level actions, and strong visual understanding. SIMA 2 is currently available as a limited research preview for academics and game developers.

So, what are your thoughts on the future of AIs like Gemini-powered SIMA 2 in gaming and virtual worlds? Share your thoughts with us in the comments!