goat simulator 3 is a surreal video game in which players take their domesticated ungulates on a series of incredible adventures, sometimes involving jetpacks.
It may seem an unlikely place for the next big leap in artificial intelligence, but Google DeepMind today revealed an AI program that can learn how to complete tasks in many games, including: I made it. goat simulator 3.
Most impressively, when a program encounters a game for the first time, it can adapt what it has learned from playing other games to reliably perform its tasks. The program is called SIMA (Scalable Instructable Multiworld Agent) and builds on recent AI advances where large-scale language models generate highly capable Chabots like ChatGPT.
“SIMA is better than the sum of its parts,” says Frederic Besse, a research engineer at Google DeepMind who worked on the project. “You can leverage the concepts shared within the game to learn better skills and learn how to better carry out instructions.”
Expanding the types of data that algorithms can learn from provides a path to more powerful capabilities as Google, OpenAI, and others race to gain an edge on the recent generative AI boom.
DeepMind’s latest video game project will soon allow AI systems like OpenAI’s ChatGPT and Google’s Gemini to do more than chat and generate images and videos by controlling computers and executing complex commands. It suggests that it will be possible. It’s a dream pursued by both independent AI enthusiasts and large companies, including Google DeepMind, whose CEO Demis Hassabis recently told WIRED that they’re “investing heavily in that direction.” told.
new way to play
SIMA shows that DeepMind is putting a new spin on Game Playing Agents, an AI technology the company has developed in the past.
In 2013, before DeepMind was acquired by Google, the London-based startup announced that a technique called reinforcement learning (training algorithms using positive and negative feedback about performance) would We showed you how it can help you play Atari video games. In 2016, DeepMind, as part of Google, developed AlphaGo, a program that used the same approach to defeat the world champion of Go, an ancient board game that requires delicate, instinctive skills.
For the SIMA project, the Google DeepMind team collaborated with several game studios to collect keyboard and mouse data from humans playing 10 different games in 3D environments, including: no man’s sky, tear down, Hydroniaand satisfaction. DeepMind then adds descriptive labels to that data, linking clicks and taps to the actions you take, such as whether you’re a goat looking for a jetpack or a man digging for gold. I did.
