AlphaZero, machine learning, and the future of work

AlphaZero, machine learning, and the future of work. A photograph of a man waking up in bed with a humanoid robot next to him. Image attribution: Wired

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 about the future of work. 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 to become the world’s best Go and shogi player in a few hours.

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 was shocked watching AlphaZero’s tenth game with Stockfish, the strongest open-source chess engine in the world.

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.

So what is the future of work? 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

2 thoughts on “AlphaZero, machine learning, and the future of work

  1. I think your take is insightful, Adrian. There are many types of AI and the potential is limitless. Do you think the example you saw replaced human creativity? The way I look at it, machines can and will automate mundane, routine tasks that don’t need special “humanness”, freeing us up to spend more time on creative solutions. As far as the future of work goes, there are new jobs being created now that contribute to designing, training and managing AI.

    1. If you think that the very best chess players are creative, then AlphaZero is creative. Ten years ago, no human chess player would describe playing chess as a “mundane, routine task”. What has been happening for quite a while now is that machines continue to master tasks that were formerly thought to need the “special humanness” you mention. When this occurs, we humans redefine the formerly special task as something “routine”.

      For example, realtime natural spoken language processing and translation by machines were considered pie-in-the-sky possibilities until recently. Now we take them for granted. If someone posts in Dutch on FaceBook, I can click a button and understand what they’re saying. And the translations are continually getting better…

      The world of “only humans can do this” is getting steadily smaller, and there’s nothing that ensures that the jobs that continue to be lost will be replaced by as many new jobs that pay as well as the old ones.

Leave a Reply

Your email address will not be published. Required fields are marked *