DeepMind, famous for its artificial intelligence that beats humans at chess and go, has demonstrated the new ability of its technology to create algorithms that sort data orders of magnitude faster than the best known human solutions.
To do this, the DeepMind researchers used AlphaDev, an AI-based AlphaZero system that learns to play various games using reinforcement and tree search. They tasked the AI with sorting numbers by size and gave it the ability to select instructions in assembly language, code that is generated from C++ before being translated into machine code.
The results astounded the scientists: AI was able to create algorithms that sorted data up to three times faster than the best known human-made algorithms. These algorithms are already included in two standard C++ libraries and are used by billions of programmers around the world.
“This is an exciting result,” said Emma Branskill, a computer scientist at Stanford University. She added that this approach could be applied to other types of mathematical computing and open up new ways of speeding up computing in some areas.