TLDR:
- Google DeepMind’s new AI agent, SIMA, can play various games it has never seen before by watching human players.
- SIMA represents a step towards more generalized AI that can transfer skills across multiple environments.
An AI agent developed by Google DeepMind, named SIMA, has the ability to play different games, including ones it has never encountered before like Goat Simulator 3. The agent was trained on examples of humans playing video games and uses imitation learning to follow text commands and play games as humans would. SIMA can execute approximately 600 instructions and has shown better performance when trained on multiple games compared to just one. While it still has limitations in performing tasks at human level, the researchers are working on improving its performance to learn new skills, chat with people, and adapt quickly to unseen environments. This development represents a significant milestone in AI research, paving the way towards more powerful and autonomous AI systems in the future.