Escape Meridian

MIT Technology Review · Mar 13, 2024

MIT Technology Review reports SIMA can follow instructions across games

Markdown · JSON · Original source

On March 13, 2024, MIT Technology Review reports that ’s SIMA can follow text commands across different games, including one it had not seen before, and move around research environments. The article presents this transfer of skills between environments as a step toward more generalized AI, not as evidence that SIMA can handle real-world tasks. The team trained SIMA on examples of people playing games, with input and annotations, then used imitation learning to teach it to play as humans do. The article says that training across games helped SIMA carry out instructions better than training on one game, because games share concepts. But the system remains well below human performance: in No Man’s Sky, it could do only some of the tasks people could do, and it performed much worse without human instructions. Researchers quoted in the article see potential for more capable agents, while also stressing limits in the data and distance from autonomous AI. The team’s stated plans include improving performance, expanding the environments and skills SIMA can handle, and enabling people to chat with it. Whether this approach can extend beyond games, and how well it would work there, remains unknown.

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.

MIT Technology ReviewArticle · technologyreview.comOpen original source

Original source

MIT Technology Review · Article

Source details

Original source date: Mar 13, 2024 (2024-03-13T14:00:00Z).

Date basis: Date from the publisher’s own records, through its API.

Added to Escape Meridian: 2026-11-21T00:43:09.520Z. First observed: 2026-10-28T14:53:52.580Z. Retrieved: 2026-10-28T14:56:05.067Z.

Referenced by

Recorded citations in public readings. A reference alone does not establish agreement or influence.

No public references recorded here yet.

Reference data (JSON)

Limits

Text as retrieved on October 28, 2026; originally published March 13, 2024. The article says SIMA’s relatively limited data set is holding back its performance, but gives no further detail on the data set’s size or composition. The evidence does not establish whether SIMA’s game skills transfer to tasks outside games. The term library does not define “imitation learning”; an editor should add a definition.

Open question

What capabilities, if any, would SIMA’s transfer of skills between games enable outside game environments?

Source-stated

MIT Technology Review reports that Google DeepMind’s SIMA follows text commands across games and can play games it has not encountered before.

A new AI agent from Google DeepMind can play different games, including ones it has never seen before such as Goat Simulator 3

Source-stated

The team trained SIMA on examples of human gameplay and used imitation learning to teach it to play as humans do.

The team trained SIMA on lots of examples of humans playing video games, both individually and collaboratively, alongside keyboard and mouse input and annotations of what the players did in the game

Source-stated

The article reports that training on many games improved SIMA’s ability to carry out instructions compared with training on one game, which Besse attributes to concepts shared between games.

The team found that a SIMA agent that was trained on many games was better than an agent that learned how to play just one. This is because it was able to take advantage of concepts shared between games

Source-stated

The article reports that SIMA remains below human performance and performs much worse when people cannot give it instructions.

in the game No Man’s Sky, the AI agent could do just 60% of the tasks humans could do. And when the researchers removed the ability for humans to give SIMA instructions, they found the agent performed much worse than before.

Source-stated

The article says researchers see skill transfer between games as a step toward more generalized AI, while SIMA remains far from actual autonomous AI.

SIMA inches us closer to a “ChatGPT moment” for autonomous agents, says Roy Fox, an assistant professor at the University of California, Irvine. But it is a long way away from actual autonomous AI.

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.

Google DeepMind

Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold.

Editorial provenance

Version 1. Publication ID: 0293fbdb-4703-574b-9664-00525ed85b4f.

Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-21T00:38:46.094Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: bridge_admitted_by_model.

Observation: 9fad1b38d946c05ebe3a416d1fd633c4d95f01a628fef339cf755ae2130c724f. Source text hash: e1e4d2f1220f41e71a86fc3100a2db51c3df73aab49a6b479906ed1855139ea3.

Revision and source-check history

Version 1 · 2026-11-21T00:43:09.520Z

MIT Technology Review reports SIMA can follow instructions across games

On March 13, 2024, MIT Technology Review reports that ’s SIMA can follow text commands across different games, including one it had not seen before, and move around research environments. The article presents this transfer of skills between environments as a step toward more generalized AI, not as evidence that SIMA can handle real-world tasks. The team trained SIMA on examples of people playing games, with input and annotations, then used imitation learning to teach it to play as humans do. The article says that training across games helped SIMA carry out instructions better than training on one game, because games share concepts. But the system remains well below human performance: in No Man’s Sky, it could do only some of the tasks people could do, and it performed much worse without human instructions. Researchers quoted in the article see potential for more capable agents, while also stressing limits in the data and distance from autonomous AI. The team’s stated plans include improving performance, expanding the environments and skills SIMA can handle, and enabling people to chat with it. Whether this approach can extend beyond games, and how well it would work there, remains unknown.

Original source

MIT Technology Review · Article

Source details

Original source date: Mar 13, 2024 (2024-03-13T14:00:00Z).

Date basis: Date from the publisher’s own records, through its API.

Added to Escape Meridian: 2026-11-21T00:43:09.520Z. First observed: 2026-10-28T14:53:52.580Z. Retrieved: 2026-10-28T14:56:05.067Z.

Limits

Text as retrieved on October 28, 2026; originally published March 13, 2024. The article says SIMA’s relatively limited data set is holding back its performance, but gives no further detail on the data set’s size or composition. The evidence does not establish whether SIMA’s game skills transfer to tasks outside games. The term library does not define “imitation learning”; an editor should add a definition.

Open question

What capabilities, if any, would SIMA’s transfer of skills between games enable outside game environments?

Source-stated

MIT Technology Review reports that Google DeepMind’s SIMA follows text commands across games and can play games it has not encountered before.

A new AI agent from Google DeepMind can play different games, including ones it has never seen before such as Goat Simulator 3

Source-stated

The team trained SIMA on examples of human gameplay and used imitation learning to teach it to play as humans do.

The team trained SIMA on lots of examples of humans playing video games, both individually and collaboratively, alongside keyboard and mouse input and annotations of what the players did in the game

Source-stated

The article reports that training on many games improved SIMA’s ability to carry out instructions compared with training on one game, which Besse attributes to concepts shared between games.

The team found that a SIMA agent that was trained on many games was better than an agent that learned how to play just one. This is because it was able to take advantage of concepts shared between games

Source-stated

The article reports that SIMA remains below human performance and performs much worse when people cannot give it instructions.

in the game No Man’s Sky, the AI agent could do just 60% of the tasks humans could do. And when the researchers removed the ability for humans to give SIMA instructions, they found the agent performed much worse than before.

Source-stated

The article says researchers see skill transfer between games as a step toward more generalized AI, while SIMA remains far from actual autonomous AI.

SIMA inches us closer to a “ChatGPT moment” for autonomous agents, says Roy Fox, an assistant professor at the University of California, Irvine. But it is a long way away from actual autonomous AI.

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.

Google DeepMind

Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold.

Editorial provenance

Version 1. Publication ID: 0293fbdb-4703-574b-9664-00525ed85b4f.

Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-21T00:38:46.094Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: bridge_admitted_by_model.

Observation: 9fad1b38d946c05ebe3a416d1fd633c4d95f01a628fef339cf755ae2130c724f. Source text hash: e1e4d2f1220f41e71a86fc3100a2db51c3df73aab49a6b479906ed1855139ea3.

passed · 2026-11-21T00:38:46.094Z · anthropic/claude-opus-5.5

Publication: 0293fbdb-4703-574b-9664-00525ed85b4f

attribution — pass: The reading attributes the report to MIT Technology Review and the agent to Google DeepMind. Quoted views stay with their speakers: Besse on shared concepts, Harley on limits and Fox on autonomous AI. The source is a single news article, not a listing. The related-story sidebar is not used.

It’s called SIMA, an acronym for “scalable, instructable, multiworld agent.”

summary — pass: The title, summary and question match the article: SIMA follows text commands across games, transfers skills between games, and still falls short of humans. Its limits and the team's plans are kept, as is the split between the researchers' optimism and their caveats. Imitation learning is not in the empty definitions list, but the reading explains it in context as learning to play from human examples, and the limitations already suggest a library entry. The 60% figure is softened to 'some of the tasks', which is a minor loss of precision, not a distortion.

Then they used an AI technique called imitation learning to teach the agent to play games as humans would.

date — pass: The originalPublished instant, 2024-03-13T14:00:00Z, falls on the UTC calendar day of March 13, 2024, as stated. The evidence kind is publisher-page, so no archive disclosure is needed. The team's plans are reported as plans, not outcomes. Sidebar items describing later events, such as the 'Ten years after AlphaGo' blurb, are possible later page additions; the reading neither uses nor relies on them.

Next, Besse says, the team is working on improving the agent’s performance.

support — pass: Each claim matches its excerpt. The training method, the benefit of training on many games and Besse's explanation, the performance gap in No Man's Sky, and the drop without instructions are all stated in the source. The open question about uses outside games is framed as unknown rather than implied. The limitation on the data set is attributed to Rauber's remark. This checks fidelity to the source, not independent verification of DeepMind's results.

SIMA’s relatively limited data set is what is holding back its performance, he says.

media — pass: No media is selected, and no verified media candidates were supplied (sourceMedia is unchecked with no items). Selecting none is appropriate.

Challenge or correct this reading

Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.