What’s new in the AI world for this week?
Researchers at DeepMind have developed a new artificial intelligence system called Gato that can play 57 different Atari video games at a superhuman level, without any game-specific training.
Gato achieved an average score of over 74% of the human professional gamer score across all 57 games, using only raw pixel inputs and game score as feedback.
This demonstrates Gato’s ability to quickly learn a wide variety of tasks. The system could have broad implications for developing more general AI.
Google Introduces AI System for Generating Code from Natural Language
Google has released a new AI system called PaLM that can generate computer code from natural language descriptions. For example, it can produce code for simple tasks like sorting a list or formatting text given plain English instructions. PaLM is trained on a huge dataset of code examples and corresponding text. While it currently has limitations, being able to translate language into working code could make programming far more accessible.