Google DeepMind’s New AI Agent Learns, Adapts and Plays Games Like a Human

Google DeepMind has launched SIMA 2, an advanced AI agent designed for 3D environments, marking a significant step towards Artificial General Intelligence (AGI). Built on the Gemini model, SIMA 2 can reason about objectives, explain its strategies, and perform tasks in unfamiliar games. It learns through self-driven play and adapts to newly generated environments by the Genie 3 project. Compared to its predecessor, SIMA 1, which executed basic commands, SIMA 2 can engage in dialogue and think independently, achieving a 65% completion rate on complex tasks versus 31% for SIMA 1. It effectively processes detailed instructions and transfers skills learned in one context to new tasks in different games, showcasing enhanced cognitive capabilities. Despite these advancements, it still faces challenges with lengthy, multi-step tasks and visual interpretation in 3D settings, but DeepMind views SIMA 2 as a promising platform for future robotics applications and AGI development.

Source 🔗