Playground

by Richard Powers

Deep Blue had beaten Kasparov by brute force. Human programmers specified the best openings, defined strong moves, described how to control the board and gain material, and spelled out how to win the endgame. Deep Blue just took those declarative instructions and applied massive computation to look farther down the branching tree of moves and countermoves than any human could. But the new machines did something wildly different. They were learning on their own, with reiterated reinforcement under limited supervision. They combed through continents of data on their own, finding patterns, generalizing, and drawing conclusions that even their trainers couldn’t see. They were learning how to play simply by playing. And these deep players learned the most extraordinary things. They started to drive cars. Without being told a single thing about cats except whether a given picture showed one, they learned how to recognize any cat from any angle under any conditions. They figured out how to translate text from one language to another with uncanny fluency, without being taught a single rule of grammar or usage. They learned these things the way a child would, by weighing the evidence and adjusting the strengths of the connections in their networks of neurons until their brains began to generalize solutions.

Deep Blue had beaten Kasparov by brute force. Human programmers specified the best openings, defined strong moves, described how to control the board and gain material, and spelled out how to win the endgame. Deep Blue just took those declarative instructions and applied massive computation to look farther down the branching tree of moves and countermoves than any human could. But the new machines did something wildly different. They were learning on their own, with reiterated reinforcement under limited supervision. They combed through continents of data on their own, finding patterns, generalizing, and drawing conclusions that even their trainers couldn’t see. They were learning how to play simply by playing. And these deep players learned the most extraordinary things. They started to drive cars. Without being told a single thing about cats except whether a given picture showed one, they learned how to recognize any cat from any angle under any conditions. They figured out how to translate text from one language to another with uncanny fluency, without being taught a single rule of grammar or usage. They learned these things the way a child would, by weighing the evidence and adjusting the strengths of the connections in their networks of neurons until their brains began to generalize solutions.