Artificial Intelligence
Google DeepMind expands game research with studios
Google DeepMind said it was working with game studios on new gameplay prototypes while recasting its earlier game-playing systems as the basis for newer agents that can use ordinary controls.
- Google DeepMind said it was partnering with game developers to prototype new gameplay experiences.
- The company said its game work dated back to its foundation in 2010 and included a major partnership with Fenris Creations and the EVE Universe, plus work with Hello Games, Coffee Stain Studios and Foulball Hangover.
- It said the Deep Q-Network learned 49 Atari 2600 games from raw pixels, and that a 2015 Nature paper on the system helped drive modern deep reinforcement learning.
- Google DeepMind said AlphaGo beat Lee Sae Dol in 2016, and that AlphaFold later helped with protein structure prediction.
- The company said SIMA and SIMA 2 used keyboard-and-mouse controls and natural-language instructions, with SIMA 2 powered by Gemini and used in games including No Man's Sky, Valheim and Hydroneer.
Google DeepMind said it was working with game developers to build new gameplay prototypes and to explore uses for AI in games. It said games had been central to its research since 2010.
The company described a sequence of earlier systems, starting with the Deep Q-Network, which it said learned 49 Atari 2600 games from pixels alone. It then pointed to AlphaGo, AlphaGo Zero, AlphaZero, MuZero and AlphaStar as later milestones.
Google DeepMind said those systems included AlphaGo’s win over Lee Sae Dol in 2016 and AlphaStar’s Grandmaster-level performance in StarCraft II. It also said AlphaFold contributed to protein structure prediction and that the 2024 Nobel Prize in Chemistry recognised that work.
The company then turned to SIMA and SIMA 2, which it said were built to read the screen, follow natural-language instructions and act through ordinary keyboard and mouse controls. It said SIMA 2, powered by Gemini, had shown human-like play in No Man's Sky, Valheim, Hydroneer and other research environments.
Google DeepMind said such a general agent could support existing games without code changes, including AI companions, adaptive non-player characters and more robust testing during development. It also said it was building a larger portfolio of games for AI research with partner studios.
Named in this story
People
- Demis Hassabis
- was named as one of Google DeepMind’s founders and a former game developer
- Lee Sae Dol
- was the world champion Go player defeated by AlphaGo
Companies
- Google DeepMind
- said it was partnering with game developers to prototype new gameplay experiences
- Fenris Creations
- was part of a research partnership Google DeepMind said it had unveiled earlier in the year
- Hello Games
- was one of the studios Google DeepMind said it had worked with
- Coffee Stain Studios
- was one of the studios Google DeepMind said it had worked with
- Foulball Hangover
- was one of the studios Google DeepMind said it had worked with
Products and systems
- EVE Universe
- was part of the partnership Google DeepMind said it had unveiled earlier in the year
- Deep Q-Network
- was the neural network that learned to play 49 Atari 2600 games
- Atari 2600
- was the game platform used in DeepMind’s early training work
- AlphaGo
- was the system that defeated Lee Sae Dol at Go in 2016
- AlphaFold
- was the system the article said helped solve protein structure prediction
- SIMA
- was described as a scalable instructable multiworld agent
- SIMA 2
- was described as an interactive companion powered by Gemini
- Gemini
- was the model used to power SIMA 2
- No Man's Sky
- was one of the games named as part of SIMA 2’s use cases
- Valheim
- was one of the games named as part of SIMA 2’s use cases
- Hydroneer
- was one of the games named as part of SIMA 2’s use cases
- StarCraft II
- was the game in which AlphaStar reached Grandmaster level
- AlphaStar
- was the system Google DeepMind said reached Grandmaster level in StarCraft II
How the source tells it
A promotional, self-congratulatory register presented Google DeepMind’s game research history as proof of future gains.
- boosterism repeated breakthrough language and a chain of superlatives were used to elevate the company’s work
- triumph past successes were lined up as a success story with each milestone presented as another advance
- speculation and future promise several claims about new gameplay and testing were framed as what these systems could do next