Tech

Why the buzz around DeepMind is dissipating as it transitions from games to science

Key Points
  • DeepMind shot to fame in 2016 when it built a computer program called AlphaGo that learned how to play the board game Go and became better than any human. 
  • The London AI lab, which is owned by Alphabet, is now going through a quieter period, with far less media attention.
  • DeepMind is shifting its focus from building "AI agents" that can play games to building AI agents that can have real world impact, particularly in areas of science like biology. 
Google Deepmind head Demis Hassabis speaks during a press conference ahead of the Google DeepMind Challenge Match in Seoul on March 8, 2016.
Jung Yeon-Je | AFP |Getty Images | Getty Images

In 2016, DeepMind, an Alphabet-owned AI unit headquartered in London, was riding a wave of publicity thanks to AlphaGo, its computer program that took on the best player in the world at the ancient Asian board game Go and won.

Photos of DeepMind's leader, Demis Hassabis, were splashed across the front pages of newspapers and websites, and Netflix even went on to make a documentary about the five-game Go match between AlphaGo and world champion Lee SeDol. Fast-forward four years, and things have gone surprisingly quiet about DeepMind.

"DeepMind has done some of the most exciting things in AI in recent years. It would be virtually impossible for any company to sustain that level of excitement indefinitely," said William Tunstall-Pedoe, a British entrepreneur who sold his AI start-up Evi to Amazon for a reported $26 million. "I expect them to do further very exciting things."

AI pioneer Stuart Russell, a professor at the University of California, Berkeley, agreed it was inevitable that excitement around DeepMind would tail off after AlphaGo.

"Go was a recognized milestone in AI, something that some commentators said would take another 100 years," he said. "In Asia in particular, top-level Go is considered the pinnacle of human intellectual powers. It's hard to see what else DeepMind could do in the near term to match that."

From Go to science

DeepMind's army of 1,000 plus people, which includes hundreds of highly-paid PhD graduates, continues to pump out academic paper after academic paper, but only a smattering of the work gets picked up by the mainstream media. The research lab has churned out over 1,000 papers and 13 of them have been published by Nature or Science, which are widely seen as the world's most prestigious academic journals. Nick Bostrom, the author of Superintelligence and the director of the University of Oxford's Future of Humanity Institute described DeepMind's team as world-class, large, and diverse. 

"Their protein folding work was super impressive," said Neil Lawrence, a professor of machine learning at the University of Cambridge, whose role is funded by DeepMind. He's referring to a competition-winning DeepMind algorithm that can predict the structure of a protein based on its genetic makeup. Understanding the structure of proteins is important as it could make it easier to understand diseases and create new drugs in the future. 

The World's top human Go player, 19-year-old Ke Jie (L) competes against AI program AlphaGo, which was developed by DeepMind, the artificial intelligence arm of Google's parent Alphabet. Machine won the three-game match against man in 2017. The AI didn't lose a single game.
VCG | Visual China Group | Getty Images

DeepMind is keen to move away from developing relatively "narrow" so-called "AI agents," that can do one thing well, such as master a game. Instead, the company is trying to develop more general AI systems that can do multiple things well, and have real world impact.

It's particularly keen to use its AI to leverage breakthroughs in other areas of science including healthcare, physics and climate change.

But the company's scientific work seems to be of less interest to the media. In 2016, DeepMind was mentioned in 1,842 articles, according to media tracker LexisNexis. By 2019, that number had fallen to 1,363. 

One ex-DeepMinder said the buzz around the company is now more in line with what it should be. "The whole AlphaGo period was nuts," they said. "I think they've probably got another few milestones ahead, but progress should be more low key. It's a marathon not a sprint, so to speak."

DeepMind denied that excitement surrounding the company has tailed off since AlphaGo, pointing to the fact that it has had more papers in Nature and Science in recent years. 

"We have created a unique environment where ambitious AI research can flourish. Our unusually interdisciplinary approach has been core to our progress, with 13 major papers in Nature and Science including 3 so far this year," a DeepMind spokesperson said. "Our scientists and engineers have built agents that can learn to cooperate, devise new strategies to play world-class chess and Go, diagnose eye disease, generate realistic speech now used in Google products around the world, and much more."

"More recently, we've been excited to see early signs of how we could use our progress in fundamental AI research to understand the world around us in a much deeper way. Our protein folding work is our first significant milestone applying artificial intelligence to a core question in science, and this is just the start of the exciting advances we hope to see more of over the next decade, creating systems that could provide extraordinary benefits to society."

DeepMind hype

The company, which competes with Facebook AI Research and OpenAI, did a good job of building up hype around what it was doing in the early days.

Hassabis and Mustafa Suleyman, the intellectual co-founders who have been friends since school, gave inspiring speeches where they would explain how they were on a mission to "solve intelligence" and use that to solve everything else.

There was also plenty of talk of developing "artificial general intelligence" or AGI, which has been referred to as the holy grail in AI ­and is widely viewed as the point when machine intelligence passes human intelligence.

But the speeches have become less frequent (partly because Suleyman left Deepmind and works for Google now), and AGI doesn't get mentioned anywhere near as much as it used to.

Larry Page, left, and Sergey Brin, co-founders of Google Inc.
JB Reed | Bloomberg | Getty Images

Google co-founders Larry Page and Sergey Brin were huge proponents of DeepMind and its lofty ambitions, but they left the company last year and its less obvious how Google CEO Sundar Pichai feels about DeepMind and AGI.

It's also unclear how much free reign Pichai will give the company, which cost Alphabet $571 million in 2018. Just one year earlier, the company had losses of $368 million.

"As far as I know, DeepMind is still working on the AGI problem and believes it is making progress," Russell said. "I suspect the parent company (Google/Alphabet) got tired of the media turning every story about Google and AI into the Terminator scenario, complete with scary pictures."

The DeepMind anti-Christ

One academic who is particularly skeptical about DeepMind's achievements is AI entrepreneur Gary Marcus, who sold a machine-learning start-up to Uber in 2016 for an undisclosed sum.

"I think they realize the gulf between what they're doing and what they aspire to do," he said. "In their early years they thought that the techniques they were using would carry us all the way to AGI. And some of us saw immediately that that wasn't going to work. It took them longer to realize but I think they've realized it now."

Marcus said he's heard that DeepMind employees refer to him as the "anti-Christ" because he has questioned how far the "deep learning" AI technique that DeepMind has focused on can go.

"There are major figures now that recognize that the current techniques are not enough," he said. "It's very different from two years ago. It's a radical shift."

He added that while DeepMind's work on games and biology had been impressive, it's had relatively little impact.

"They haven't used their stuff much in the real world," he said. "The work that they're doing requires an enormous amount of data and an enormous amount of compute, and a very stable world. The techniques that they're using are very, very data greedy and real-world problems often don't supply that level of data."