In the strange video game Goat Simulator 3, players lead trained ungulates on a number of unbelievable journeys that occasionally include jetpacks.
While Goat Simulator 3 might seem like an unexpected setting for the next big advancement in AI, Google DeepMind recently unveiled a computer that can learn how to fulfill tasks in a number of games.
The most amazing thing is that the computer can consistently complete tasks by using the lessons it has learnt from playing prior games when it comes across a new game. Scalable Instructable Multiworld Agent, or SIMA, is the name of the program, which expands on recent developments in AI that have resulted to the production of amazingly effective chabots such as ChatGPT using massive language models.
“SIMA is more than the sum of its parts,” claims Google DeepMind research engineer Frederic Besse, who worked on the project. “It can leverage the common ideas in the game, develop stronger abilities, and become more adept at following directions.”
The SIMA program from Google DeepMind takes a swing at Goat Simulator 3.
Extending the range of data types that algorithms can process provides a path to more powerful capabilities as Google, OpenAI, and other players compete to acquire a competitive advantage in profiting on the current generative AI boom.
DeepMind’s most recent video game experiment suggests that in the near future, AI systems such as Google’s Gemini and OpenAI’s ChatGPT may be able to do more than simply make photos and videos and talk; they may also be able to take over computers and issue advanced orders. Both large corporations like Google DeepMind, whose CEO, Demis Hassabis, recently told WIRED that the company is “investing heavily in that direction,” and single AI enthusiasts are pursuing this goal.
Linxi “Jim” Fan, a senior research scientist at Nvidia who works on AI gameplay and was involved with an early effort to train AI to play by controlling a keyboard and mouse with a 2017 OpenAI project called World of Bits, says, “The paper is an interesting advance for embodied agents across multiple simulations.” Fan claims that this initiative and a 2022 endeavor named VPT-which involves using agents as learning tools in Minecraft-are both reminded of the Google DeepMind work.
According to him, “SIMA goes one step further and demonstrates stronger generalization to new games.” Although there are still relatively few environments, SIMA seems to be headed in the correct direction.