# MIT Technology Review reports Google DeepMind’s SIMA 2 game-playing agent [Canonical reading](https://escapemeridian.com/readings/mit-technology-review-reports-google-deepmind-s-sima-2-game-8dfa8296) MIT Technology Review reported on November 13, 2025, that Google DeepMind built SIMA 2, an agent that follows instructions in virtual worlds. The article says Google DeepMind built it on Gemini, its flagship large language model, and reports that the researchers say it can tackle more complex tasks, learn through trial and error, and communicate with users. Unlike agents trained for preset game goals, SIMA is intended to follow people’s instructions. The researchers tested SIMA 2 in unfamiliar environments generated by Genie 3, a world model, and reported that it could navigate and carry out instructions there. The article describes this as an experiment, not a demonstrated route to capable robots: SIMA 2 struggles with complex, multi-step tasks, has limited memory, and is less skilled than people at mouse-and-keyboard control. Julian Togelius is open-minded about the prospect of better robots; Matthew Guzdial questions whether game skills will transfer, citing the differences between game controls and real-world visuals. Google DeepMind’s longer-term aim is to use such agents to drive real-world robots, and the team hopes to train SIMA 2 in generated worlds with Gemini’s feedback. Whether skills learned in these virtual environments will transfer to robots remains uncertain. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source MIT Technology Review · Article Original source date: Nov 13, 2025 (2025-11-13T15:00:00Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-10T03:33:23.098Z. First observed: 2026-11-02T00:47:30.422Z. Retrieved: 2026-11-02T00:49:22.712Z. - [Google DeepMind is using Gemini to train agents inside Goat Simulator 3](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Limits Text as retrieved on November 2, 2026; originally published November 13, 2025. The article reports an experiment and researchers’ claims; it does not establish that SIMA 2’s skills transfer to real-world robots. The article gives differing views on transfer: Togelius is open-minded, while Guzdial questions whether game skills would carry over. ## Open question Will SIMA 2’s skills in virtual worlds transfer to robots, given the differences in controls and visual input that the article’s researchers identify? ## Source-stated MIT Technology Review reports that Google DeepMind built SIMA 2 on Gemini and says the agent can follow instructions and tackle more complex tasks. > SIMA 2 has been built on top of Gemini, the firm’s flagship large language model - [Supporting source](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Source-stated The article distinguishes SIMA’s instruction-following approach from game agents trained to pursue preset goals. > the idea behind SIMA is to train an agent to play an open-ended game without preset goals. Instead, the agent learns to carry out instructions given to it by people. - [Supporting source](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Source-stated The researchers report that SIMA 2 navigated and followed instructions in environments generated by Genie 3 that it had never seen before. > They found that the agent was able to navigate and carry out instructions there. - [Supporting source](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Source-stated The article reports that SIMA 2 remains experimental, with difficulties in complex tasks, limited memory, and weaker mouse-and-keyboard skills than people. > The agent struggles with complex tasks that require multiple steps and more time to complete. It also remembers only its most recent interactions - [Supporting source](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Source-stated The article presents disagreement about whether skills learned in games will transfer to robots: Julian Togelius is open-minded, while Matthew Guzdial questions the transfer. > Still, he is open-minded about whether or not SIMA 2 could lead to better robots. - [Supporting source](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Source-stated Google DeepMind’s longer-term aim is to use agents like SIMA 2 to drive real-world robots, while the team hopes to continue training in generated virtual worlds. > In the long run, Google DeepMind wants to use such agents to drive real-world robots. - [Supporting source](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## World model An AI system's internal model of how an environment behaves, used to predict what happens next or to plan. - [Definition reference](https://arxiv.org/abs/1803.10122) ## Google DeepMind Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold. - [Explore this context](/entities/google-deepmind) - [Reference](https://deepmind.google/blog/) - [Reference](https://deepmind.google/science/alphafold/) ## Editorial provenance Version 1. Publication ID: f5bc62ea-c198-f9a1-7e54-06e1bab1c10f. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T03:29:20.314Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model. Observation: 763738c09e8a4cd64ff0b84e6736741eca26eb77428db20225e5635fb4ab42e4. Source text hash: e50936fd9e18ee9599e4e8d31552fff3d35d85b2a7602f48974afc5b20684bf9. ## Revision and source-check history ### Version 1 — 2026-11-10T03:33:23.098Z MIT Technology Review reported on November 13, 2025, that Google DeepMind built SIMA 2, an agent that follows instructions in virtual worlds. The article says Google DeepMind built it on Gemini, its flagship large language model, and reports that the researchers say it can tackle more complex tasks, learn through trial and error, and communicate with users. Unlike agents trained for preset game goals, SIMA is intended to follow people’s instructions. The researchers tested SIMA 2 in unfamiliar environments generated by Genie 3, a world model, and reported that it could navigate and carry out instructions there. The article describes this as an experiment, not a demonstrated route to capable robots: SIMA 2 struggles with complex, multi-step tasks, has limited memory, and is less skilled than people at mouse-and-keyboard control. Julian Togelius is open-minded about the prospect of better robots; Matthew Guzdial questions whether game skills will transfer, citing the differences between game controls and real-world visuals. Google DeepMind’s longer-term aim is to use such agents to drive real-world robots, and the team hopes to train SIMA 2 in generated worlds with Gemini’s feedback. Whether skills learned in these virtual environments will transfer to robots remains uncertain. ## Original source MIT Technology Review · Article Original source date: Nov 13, 2025 (2025-11-13T15:00:00Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-10T03:33:23.098Z. First observed: 2026-11-02T00:47:30.422Z. Retrieved: 2026-11-02T00:49:22.712Z. - [Google DeepMind is using Gemini to train agents inside Goat Simulator 3](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Limits Text as retrieved on November 2, 2026; originally published November 13, 2025. The article reports an experiment and researchers’ claims; it does not establish that SIMA 2’s skills transfer to real-world robots. The article gives differing views on transfer: Togelius is open-minded, while Guzdial questions whether game skills would carry over. ## Open question Will SIMA 2’s skills in virtual worlds transfer to robots, given the differences in controls and visual input that the article’s researchers identify? ## Source-stated MIT Technology Review reports that Google DeepMind built SIMA 2 on Gemini and says the agent can follow instructions and tackle more complex tasks. > SIMA 2 has been built on top of Gemini, the firm’s flagship large language model - [Supporting source](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Source-stated The article distinguishes SIMA’s instruction-following approach from game agents trained to pursue preset goals. > the idea behind SIMA is to train an agent to play an open-ended game without preset goals. Instead, the agent learns to carry out instructions given to it by people. - [Supporting source](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Source-stated The researchers report that SIMA 2 navigated and followed instructions in environments generated by Genie 3 that it had never seen before. > They found that the agent was able to navigate and carry out instructions there. - [Supporting source](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Source-stated The article reports that SIMA 2 remains experimental, with difficulties in complex tasks, limited memory, and weaker mouse-and-keyboard skills than people. > The agent struggles with complex tasks that require multiple steps and more time to complete. It also remembers only its most recent interactions - [Supporting source](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Source-stated The article presents disagreement about whether skills learned in games will transfer to robots: Julian Togelius is open-minded, while Matthew Guzdial questions the transfer. > Still, he is open-minded about whether or not SIMA 2 could lead to better robots. - [Supporting source](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Source-stated Google DeepMind’s longer-term aim is to use agents like SIMA 2 to drive real-world robots, while the team hopes to continue training in generated virtual worlds. > In the long run, Google DeepMind wants to use such agents to drive real-world robots. - [Supporting source](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## World model An AI system's internal model of how an environment behaves, used to predict what happens next or to plan. - [Definition reference](https://arxiv.org/abs/1803.10122) ## Google DeepMind Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold. - [Explore this context](/entities/google-deepmind) - [Reference](https://deepmind.google/blog/) - [Reference](https://deepmind.google/science/alphafold/) ## Editorial provenance Version 1. Publication ID: f5bc62ea-c198-f9a1-7e54-06e1bab1c10f. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T03:29:20.314Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model. Observation: 763738c09e8a4cd64ff0b84e6736741eca26eb77428db20225e5635fb4ab42e4. Source text hash: e50936fd9e18ee9599e4e8d31552fff3d35d85b2a7602f48974afc5b20684bf9. ### passed — 2026-11-10T03:29:20.314Z — anthropic/claude-opus-5.5 Publication: f5bc62ea-c198-f9a1-7e54-06e1bab1c10f attribution — pass: The reading consistently attributes the report to MIT Technology Review and the capability claims to Google DeepMind's researchers. It names Togelius and Guzdial for their views. The article is a single news piece. The related-story sidebar is not used. > The researchers claim that SIMA 2 can carry out a range of more complex tasks inside virtual worlds summary — pass: The reading leads with the actor and the development: Google DeepMind built SIMA 2 on Gemini. It explains the instruction-following idea, the Genie 3 test, the limitations and the disagreement over robot transfer. The definitions cover large language model and world model, and agent is explained in context. The phrase 'not a demonstrated route to capable robots' slightly extends the source's 'still an experiment' but stays fair. The question follows from both critics' points. > SIMA 2 is still an experiment. date — pass: November 13, 2025 matches the UTC day of originalPublished. The source is a publisher page, so no archive disclosure is required. The sidebar story referencing ten years after AlphaGo is a later addition, but the reading does not use it. DeepMind's robot aims are presented as goals and hopes, not outcomes. > In the long run, Google DeepMind wants to use such agents to drive real-world robots. support — pass: Each claim is grounded in the article: built on Gemini, open-ended instruction following, Genie 3 environments, limitations (multi-step tasks, recent-only memory, weaker mouse/keyboard control), Togelius's openness, Guzdial's skepticism, and the robot aim. No unsupported benefits are implied. > It’s much harder to understand visuals from cameras in the real world compared to games media — pass: No media was selected, and no verified candidates were supplied, so leaving media empty is appropriate. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/8dfa8296a04d2bffa68136ea5686a821fe6d0c349daa7ae2cc1fe8da9d7df962/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/mit-technology-review-reports-google-deepmind-s-sima-2-game-8dfa8296.json)