DeepMind showcases a new artificial intelligence called Scalable Instructable Multiworld Agent, or SIMA, that can play any three-dimensional game just like a human. The goal of creating SIMA is not just to excel at gaming but to demonstrate that AI can receive instructions and carry them out in novel scenarios. SIMA has been trained on a wide variety of games, collaborating with up to 8 game studios and testing environments from Unity in 4 different setups. The initial datasets for SIMA involved players giving commands in one set and viewers suggesting actions based on gameplay videos in another.
Currently, SIMA can interpret 600 commands such as turn left or right, climb stairs, open a box, run towards water, or shoot enemies, typically completing these tasks within ten seconds. The objective of developing SIMA is to build a generalist artificial intelligence that can follow commands even in new and unfamiliar environments. It is anticipated that in the future, AI will be able to handle increasingly complex instructions continuously.
Source: DeepMind
TLDR: DeepMind unveils SIMA, an AI capable of playing any 3D game similar to humans, demonstrating the ability to receive and execute instructions in diverse environments. Development aims to create a generalist AI that can adapt to more complex commands over time.
Leave a Comment