DeepMind repurposes game-playing AI for efficient coding and infrastructure optimization

Google’s DeepMind has trained game-playing artificial intelligence (AI) to improve code and infrastructure. AlphaGo, the first AI to beat a world Go champion, was trained using human gameplay, while its successors AlphaGo Zero, AlphaZero and MuZero achieved mastery by playing against themselves or without prior knowledge of the game’s rules. The AI has now been put to work analysing data on the usage of Google’s Borg system to produce new ways of predicting load management and reducing underused hardware by 19%. MuZero similarly refined YouTube stream compression, reducing the bitrate of videos by up to 4%. AlphaDev, a relative of AlphaZero, improved sorting algorithms and reduced load times by 30%.