Not long ago I wrote about the end of decent paid jobs and the need for basic income. A startling recent advance in machine learning has only heightened my concerns. Last month, Google’s subsidiary, DeepMind, published a paper on AlphaZero, an artificial intelligence (AI) the company designed to play games. The AI started with only game rules. Here’s what happened next:
“At first it made random moves. Then it started learning through self-play. Over the course of nine hours, the chess version of the program played forty-four million games against itself on a massive cluster of specialized Google hardware. After two hours, it began performing better than human players; after four, it was beating the best chess engine in the world.”
—James Somers, New Yorker, How the Artificial-Intelligence Program AlphaZero Mastered Its Games
From “knowing” nothing about the game, in four hours the program became the strongest chess player in the world. AlphaZero also taught itself in a few hours to become the world’s best Go and shogi player.
As a schoolboy I played competitive chess for a few years. Although I haven’t played chess seriously since then, I still have a feeling for the game.
I’d describe AlphaZero’s play as completely solid, interspersed with incredible flashes of brilliance. Great human chess players have an uncanny ability to view a position and quickly select a few plausible moves for deeper study out of the many possible legal moves. The best grandmasters occasionally discover a brilliant and unexpected move in a game. AlphaZero found several during this game.
Having seen this game, I’d describe AlphaZero as the most creative, brilliant, and strongest chess player the world has ever seen.
From a novice to best in the world in four hours, is a level of performance that no human can match.
Now think about what would happen if this kind of performance could be achieved in human work environments such as:
- medical scan diagnosis;
- legal document creation;
- engineering design; and
- stock market trading.
These are only harder problems than playing a game because:
- the problem space is larger; and
- the data needed for learning can’t be self-generated by the AI itself and must be supplied by humans.
But these are not insuperable obstacles. If overcome, many high paid jobs for medical practitioners, lawyers, accountants, and financial analysts would disappear.
Are we moving towards a world where the only available work is in low paid “human service” areas where people are still cheaper than machines? Perhaps.
Until the arrival of robots capable of doing just about everything humans do. What work for humans remains then?
Image attribution: Wired