Google’s new artificial intelligence, SIMA 2, considerably boosts the potential power over its predecessor, SIMA 1. As recently featured in the hit game No Man’s Sky developed by agent creator Joe Marino, the new agent valiantly tackles cluttered virtual spaces. This innovative development is an exciting benchmark for what AI technology can achieve.
SIMA 2 impressively explained its environment during the demonstration, even while traversing a rugged planetary landscape. It was only from there that it figured out what to do next by finding and approaching a distress beacon. It was the beginning of a third wave of AI advancement. SIMA 2 demonstrated its capability in traversing dynamically produced photorealistic environments with Gemini, DeepMind’s state-of-the-art world model.
The new SIMA 2, powered by Gemini 2.5 flash-lite model, SIMA 2 This smart system unlocks advanced reasoning and decision-making capabilities that are far superior to SIMA 1. SIMA 2 features even more powerful capabilities. It can autonomously detect and recognize other objects in its environment, such as benches, trees, and most impressively of all, even butterflies. This self-improving, experimental agent not only surpasses the previous performance metrics but again doubles them.
The intelligent improvements SIMA 2 introduces go much further than just their gameplay function. According to Marino, “SIMA 2 is a step change and improvement in capabilities over SIMA 1.” This increase is reflected in its performance on complicated multi-step tasks, where it achieves human-level proficiency (71%) for complex tasks! SIMA 2 nevertheless took the strong groundwork laid by SIMA 1 as their starting point. This robust proof‐of‐concept model gave the new agent the room and flexibility to grow into an effective improvement engine.
DeepMind considers SIMA 2 a critical step forward. It intends to lay the groundwork for general-purpose robots that can address open-ended, real-world tasks. The real-world applications are far-reaching, as SIMA 2 is able to execute commands from intuitive emojis. As Marino points out, “You instruct it 🪓🌲, and it’ll go chop down a tree,” demonstrating its capacity to interpret user commands contextually.
Frederic Besse elaborated on the system’s capabilities, highlighting two essential components necessary for performing tasks in real-world scenarios: understanding user intent and responding appropriately. He stated, “If we think of what a system needs to do to perform tasks in the real world, like a robot, I think there are two components of it.”
While we’re proud of our development, it’s just the beginning, and SIMA 2 represents a major step forward for artificial intelligence. Importantly, it can reason internally using the Gemini model. This development has given it the ability to engage with its environment, navigate, and even adapt to newly generated environments entirely on its own. As you’ll see, this ability is beyond cool. It amps up their promise and potential to develop more sophisticated AI agents, while requiring minimal data input from humans.
