It’s a dilemma for any visionary scientist. If you have a grand, noble goal to make the world a better place with your research, should you take the money of a tech giant like Google to fund your ambitions and see them through?
That’s the bet that DeepMind founders Demis Hassabis and Mustafa Suleyman made in 2014, when they agreed to sell their highly-respected artificial intelligence startup to Google for a reported £400 million ($520 million). Their quest of “solving intelligence to make the world better” however, has been fraught with reported tensions and now faces greater meddling by its parent company in Mountain View, Calif.
Google wants to capitalize on owning one of the world’s highest concentrations of AI talent, particularly in the field of reinforcement learning. But DeepMind’s contribution to Google’s bottom line since 2014 hasn’t been earth-shattering: there’s been a smarter method for cooling data centers, and an open-source breakthrough on making artificial voices sound more real. Meanwhile, DeepMind has been losing hundreds of millions of dollars annually.
On Tuesday, Google moved to take greater control of the London-based AI division. Google will now directly manage DeepMind Health, a division of DeepMind that the company describes as its “first effort” to achieve its social mission. Google’s recently-appointed executive in charge of its disparate health initiatives, David Feinberg, will be in charge.
Publicly, DeepMind appeared to downplay the shift, announcing that “the team behind Streams,” a mobile app that supports doctors and nurses in a number of UK hospitals, would be joining Google. A DeepMind spokesperson suggested this was a reasonable development since the company’s expertise has been in AI research, while Google’s has been in “scaling,” or reaching hundreds of millions of people. But they also confirmed that this meant Google was absorbing DeepMind Health in its entirety.
Despite posting two positive tweets about the shift, it’s hard to imagine that DeepMind co-founder Mustafa Suleyman, who has been overseeing the DeepMind Health initiative for several years now, will be pleased about shifting oversight of the project over to Feinberg in the United States.
A deeper connection to the Google mothership also makes things awkward for some of DeepMind’s partners, in particular the U.K.’s National Health Service (NHS). In 2017 Britain’s privacy regulator ruled that the NHS’s data sharing with DeepMind was illegal because patients had not been properly informed about how their medical records would be used, and yesterday’s news has raised new concerns about what sort of access Google might have to such records. DeepMind says on its website that “no NHS patient data will ever be connected to Google accounts or services.”
Why did it come to this? One competitor to DeepMind says the company had been too fixated on its long-term goal of solving “general intelligence,” which distracted it from working on projects to solve short-term, real-world problems that could have potentially turned into products.
“When it comes to DeepMind commercializing research, there is a big gap,” said Haitham Bou-Ammar, an executive at Cambridge-based AI startup Prowler.io. “What they have built is another university lab, which is good, but at the same time we need to make money in the end.”
Bou-Ammar, who’s startup licenses a decision-making platform for logistics and financial companies and is expecting sales of more than $5 million in 2018, suggested DeepMind needed to shift its focus on building “a universal black box that solves everything” to a “pipeline approach.”
Bou-Ammar’s advice sounds a little like what you’d hear in any decent time-management seminar: rather than fixate on a single goal, break it down into smaller, manageable tasks. But he adds that fundamentally DeepMind hasn’t focused enough on solving real-world problems. “Everyone is saying they need to focus more on commercial applications,” he said of others in the AI community, “and they don’t seem to be doing that.”
DeepMind has also been too focused on one particular approach to artificial intelligence, known as deep-learning neural networks, he added, and it might have helped for the company to hire more researchers with a different viewpoint.
The “deep neural network” approach to AI is inspired by the functioning of the brain and neuroscience, but there are other machine learning approaches. Among them are the multi-agent systems approach, which assigns decision making to individual agents in a simulated environment, evolutionary algorithms or so-called capsule networks being put forward by celebrated AI researcher Geoffrey Hinton.
DeepMind isn’t alone in grappling to commercialize its AI research. Facebook has also struggled to turn early bets on voice-technology, which uses elements of artificial intelligence such as natural language processing, into a viable competitor to Amazon’s Alexa or Apple’s Siri.
Former staffers from Facebook recently told Forbes that the Menlo Park, Calif.-based company lacked a coherent direction for its engineers and research team. Its AI-research division, known as FAIR and run by well-known academic Yann LeCun had turned into a research “enclave” and a “parallel world of research,” according to one former Facebook executive.
Google may have decided that it didn’t want its own parallel world of researchers. But that will likely be at odds with what DeepMind’s founders hoped for when they made that 2014 bet on a wealthy acquirer.
“Google acquired DeepMind in 2014, because they were excited about the potential for our technology,” the company says on its website. “As part of this acquisition, we agreed that DeepMind would continue to operate independently…” It seems that agreement was ultimately not sustainable.