Google DeepMind unveils SIMA 2, a generalist agent that plays without code access

DeepMind marks fifteen years since its founding with a recap that traces its research trajectory from Atari to EVE Online, emphasizing that games have remained the laboratory’s primary testbed. Co-founder Demis Hassabis arrived from the video-game industry, and many team members share that background, making collaborations with studios a natural fit rather than a marketing requirement. This year the lab announced a major research partnership with Fenris Creations and the EVE universe, alongside ongoing work with Hello Games, Coffee Stain Studios and Foulball Hangover.

The first milestone came in 2015 with a Nature paper on Deep Q-Network (DQN), a neural network that learned to play 49 Atari 2600 games directly from raw pixels, without any game-specific engineering. DQN launched the modern era of deep reinforcement learning. Subsequent breakthroughs included AlphaGo’s 2016 victory over world champion Lee Sae Dol, highlighted by the famous move 37; AlphaGo Zero, which surpassed all earlier versions through self-play alone; AlphaZero, which generalized the approach to chess, shogi and Go; MuZero, which learned to play without prior knowledge of the rules; and AlphaStar, which reached Grandmaster level in StarCraft II in 2019, handling real-time complexity and partial information.

Those successes demonstrated that AI can maximize a score when the objective is clearly defined, but the real world offers no explicit rulebook or scoring system. The next question was whether an agent could comprehend a game world and act like a human using only visual input, without an API or source code. DeepMind’s answer is SIMA—Scalable Instructable Multiworld Agent. Unlike earlier models that optimize for high scores, SIMA is designed to function as an interactive companion: it “sees” the screen, understands natural-language instructions, and operates through a standard keyboard and mouse.

SIMA 2, the second generation, runs on the front-end models of Gemini and supports real-time conversation and reasoning. According to DeepMind, the agent achieves human-level performance in complex three-dimensional environments and titles such as No Man’s Sky, Valheim and Hydroneer. For developers, the implication is that AI can engage with existing games without any code modifications, opening avenues for entirely new gameplay experiences, from intelligent companions to mechanics that were previously infeasible.

The research ethos that drove breakthroughs in gaming has spilled over into other domains. AlphaFold applied the same underlying principles to solve the fifty-year-old protein-structure prediction problem, a contribution that earned the team the 2024 Nobel Prize in Chemistry. DeepMind presents this continuum as evidence that games are more than convenient benchmarks; they serve as engines that generate general capabilities which translate into applied scientific advances.