Google's Go win was a massive step forward for AI, but machines have a ways to go before they're smarter than humans
Considered one of the most challenging games in existence, Go has held out as one of the last traditional board games at which humans are better than machines. So Deep Mind's achievement was no doubt impressive, and Facebook has been making strides in this area too.
But AI still has a ways to go before it exceeds human intelligence in everyday life.Gary Marcus, a neuroscientist at NYU who blogs about AI for the New Yorker, put the achievement into context in a post on Medium.
"With so much at stake as people try to handicap the future of AI, and what it means for the future of employment and possibly even the human race, it's important to understand what was and was not yet accomplished," Marcus writes.
What AI has and hasn't achieved
For starters, Fan Hui, the Go champion bested by Google's AI program, is not a world champion of the game - he's ranked 633rd globally, according to Marcus. DeepMind does have plans for its AI to go head-to-head with the world champion, Lee Sedol, but not until March.
Still, even if DeepMind's AI beats Sedol in March (which many expect it will), that only means the program is extremely good at Go - it says nothing about more general applications of its intelligence.
As Marcus writes, "The real question is whether the technology developed there can be taken out of the game world and into the real world."
So what would that look like? Ultimately, the DeepMind researchers hope to apply their AI to areas like medicine, where it could one day make medical diagnoses, or to climate, where it could help create statistical models of weather patterns. Those are areas that won't be nearly as easy as beating a board game.
IBM's Watson, the AI program that stunned the world by beating the human champions at Jeopardy! in 2011, has been repurposed for medical research. But it's struggling to meet the revenue goals the company set just a few years ago, as The Wall Street Journal reported.
Unlike games like Go or Jeopardy!, the real world doesn't have strict rules. There are no exact answers. Real life requires making decisions based on limited information, something humans are very good at.
Real life requires making decisions based on limited information, something humans are very good at.
DeepMind and many other AI programs today rely on techniques that require them to be trained on massive amounts of data. By contrast, humans can often learn new concepts based on just one or two examples.
Still, there's been some progress. In December, researchers at MIT announced they had created an AI that could learn to draw alphabet letters after seeing just a single example.
But the days when a Terminator-like AI that can rule the world are probably a safe ways off, according to Martin Mueller, a computer scientist at the University of Alberta, in Canada, who's an expert in board game AI research.
So while DeepMind's AI is impressive, Mueller told Business Insider, "it's not Skynet - it's a Go program."