🤯 AI Learns Like You: No Man’s Sky 🔥

Tech & Science

DeepMind’s latest advancement, SIMA 2, represents a significant leap forward, building upon the success of its predecessor, SIMA. Initially revealed in March, SIMA demonstrated promise by learning to play 3D games much like a human, trained on hundreds of hours of gameplay data – a process that ultimately saw SIMA 1 struggling with complex tasks, achieving only a 31% success rate. SIMA 2, however, is designed to truly understand and interact with its environment, going far beyond simply following instructions. At its core lies the Gemini 2.5 flash-lite model, which combines Gemini’s powerful language and reasoning abilities with the embodied skills developed through its training. This allows the agent to not only understand what’s being asked but to respond in a meaningful way; for example, when tasked with finding a red house, it can reason through the problem, recognizing that ripe tomatoes are red and therefore the house should be red. DeepMind researchers believe this approach – working with “embodied agents” – is vital for achieving genuine intelligence. Unlike a non-embodied agent that might simply manage your calendar, SIMA 2 interacts with the world through a virtual or physical “body,” observing and acting. The team has even enabled the agent to learn entirely on its own, using Gemini to generate new challenges and a reward system to gauge its performance – a process of trial and error mirroring human learning, guided by AI intelligence. This focus on self-improvement, combined with the agent’s ability to understand emojis and navigate entirely new, photorealistic worlds, marks a significant advance in the quest for truly intelligent machines. While a firm release date for a preview of SIMA 2 hasn’t been announced, the team’s objective is to demonstrate DeepMind’s progress and explore potential partnerships and applications.