Artificial intelligence is used for everything: writing emails, doing homework, automating processes, finding new medicines, improving vaccines, and, for some time now, playing video games.
And while video games are an unlikely setting for the next big leap in artificial intelligence, Google DeepMind has today revealed an AI program capable of learning to complete tasks in a series of games, including Goat Simulator 3.
For those who don’t know, Goat Simulator 3 is a surreal video game in which players embody a domesticated goat in a series of improbable adventures, sometimes with jetpacks and other times ramming horns into neighbors.
The AI is called SIMA and plays video games better than you
The most impressive thing is that, when the program encounters a game for the first time, it can reliably perform tasks by adapting what it learned from playing other games.
The program is called SIMA (Scalable Instructable Multiworld Agent) and is based on recent advances in AI, which have seen how large language models produce remarkably capable chatbots, such as ChatGPT.
“SIMA is more than the sum of its parts,” says Frederic Besse, an engineer researcher at Google DeepMind who has participated in the project. “It is capable of leveraging shared concepts in the game, learning better skills, and learning to be better at executing instructions.”
While Google, OpenAI, and others compete to gain an advantage in the recent surge of generative AI, expanding the type of data that algorithms can learn from offers a path to more powerful capabilities.
The latest video game project from DeepMind aims to show that AI systems like OpenAI’s ChatGPT and Google’s Gemini could soon do more than just chat and generate images or videos, by taking control of computers and executing complex commands.
It is a dream pursued by both AI enthusiasts and big companies, including Google DeepMind. So, are we one step closer to Skynet and the end of the world?
But, what is DeepMind?
In 2013, before DeepMind was acquired by Google, the London-based startup showed how a technique called reinforcement learning, which involves training an algorithm with positive and negative feedback on its performance, could help computers play classic Atari video games.
In 2016, as part of Google, DeepMind developed AlphaGo, a program that used the same approach to defeat a world champion of Go, an ancient board game that requires subtle and instinctive skill.
For the SIMA project, the Google DeepMind team collaborated with several game studios to collect keyboard and mouse data from humans who were playing 10 different games with 3D environments, including No Man’s Sky, Teardown, Hydroneer, and Satisfactory.
Later on, DeepMind added descriptive tags to that data to associate the clicks and keystrokes with the actions performed by the users, for example, whether they were a goat looking for their propulsive backpack or a human character digging for gold.
