EDMONTON⁠ — Richard Sutton, the research pioneer who was a leading figure in Google’s Alberta AI lab, is parting ways with the company and is considering launching his own new Alberta-based venture.

In his first media interview since DeepMind, Google’s London-based AI research arm, announced it was shutting down its Edmonton artificial intelligence hub, Sutton told The Logic he would no longer be employed by Google, though he said he would maintain “relationships” with the company. A global leader in the field of fundamental AI practices known as reinforcement learning, Sutton said he plans to run a small, nimble independent research body that could potentially take the form of a non-profit, open-source research organization or startup company. 

Sutton had previously said on Twitter that he and the two other key researchers who worked for DeepMind in Edmonton—Michael Bowling and Patrick Pilarski—would remain in Alberta following the announcement, but did not say whether they would continue to work with the company. 

While DeepMind offered to continue to employ the researchers at its Edmonton lab if they relocated to another DeepMind office, Sutton said he thinks many of the 27 who worked there will remain in Alberta and will find new employment in the local AI industry. 

Google Canada spokesperson Luiza Staniec declined to confirm whether the researchers would stay in Edmonton. 

That Sutton is considering launching his own operation, which has not yet been reported, will likely be viewed as a major vote of confidence within Alberta’s AI ecosystem, which was rocked by the news last week that DeepMind was shutting down its local operations. The move was part of a push by Alphabet, Google’s parent company, to trim costs by cutting 12,000 staff from its global payroll. 

Sutton said he hopes his new venture would give him the “flexibility to maintain focus and pace” in his area of study, and relieve him of some of the constraints he said he faced working for Alphabet.

“I do think that DeepMind’s decision to close the lab is really worse for them than it is for us,” he said. “We are freed up, we can do other things, and focus. We were there for five and a half years, and we learned a lot from that, we gained a lot from that. And now we can make new things.” 

Others in the local AI community echoed Sutton’s optimism Tuesday as they gathered at the Edmonton headquarters of the Alberta Machine Intelligence Institute (Amii) to announce $30 million in funding to help Amii and the University of Alberta hire new research chairs. 

“The DeepMind office was a small part of the overall growing ecosystem that’s here,” Amii CEO Cam Linke told reporters. “So the amount of fundraisings that you’ve seen over the last year and a half, the number of people being hired and companies growing, has all been really, really huge.” 

“We’re still one of the most in-demand places to come and do AI research.”

According to the release, the $30 million in funding—from the federal and provincial governments, the Canadian Institute for Advanced Research (CIFAR) and several unnamed public donors—will hire 20 new faculty members to focus on how AI can be applied in areas like food security, biotech and energy. 

DeepMind’s decision to close the Alberta lab comes even as Google faces the most profound challenge to its business in years following the recent launch of ChatGPT, an AI-driven chat bot that answers user prompts in blocks of written text, including in the form of essays or poems. The technology, built by San Francisco-based OpenAI, was met with widespread interest online, and reportedly led Google to kick off a sweeping reassessment of its lucrative search business, given that ChatGPT’s uncanny insights and clean interface could threaten the company’s long-time dominance in the area. 

Sutton said Google’s interest in large language models (or, the deep-learning algorithms that underpin text-based AI technologies like ChatGPT and search) had been growing even before OpenAI released its latest version of the chat bot. 

However, such language models don’t comfortably intersect with Sutton’s area of research, which involves trying to obtain a foundational understanding of the nature of intelligence and AI through areas like reinforcement learning. Reinforcement learning, on which Sutton literally wrote the textbook, is a method of training AI programs through trial and error, and has been used in numerous AI applications including AlphaGo, the first AI program to beat a human player at the Chinese board game Go. 

“Large language models are the flavour of the month, or maybe the flavour of the decade. They’re popular, ” Sutton said. “And I’m not so much interested in trying to pursue that. I don’t see that as the most fruitful way to make fundamental progress.” 

Sutton plans to build his own team around The Alberta Plan for AI Research, a document he, Pilarski and Bowling published last summer that sets a framework for studying “continual” reinforcement learning, a practice in which machines are taught to learn all the time rather than in specific increments. 

“At DeepMind we weren’t entirely our own masters, we couldn’t just decide what to do exactly,” Sutton said. “That’s why there’s a sense that [we’ve been] released to focus more explicitly where we want to be focused, which is on the Alberta plan.” 

Overall, Sutton said there’s a sense that the area of fundamental reinforcement learning has become more diluted in the broader AI space as more capital flows into specific AI applications, and as increased digitization creates higher demand for smarter, more predictive, AI-driven platforms. 

As for large language models, he said an intensifying focus on applying such technologies has caused a broader failure to fully grasp “how they work and what their limitations are.” 

“Even those who promote them will often say, ‘Well, we don’t really understand how they work, but they’re working great.’ So that’s just evidence that it’s not primarily a scientific interest.”

Meanwhile, a more structural study of the nature of intelligence and what growing AI capabilities mean for humanity, of the sort Sutton is pursuing, is as important as ever, he said. 

“It’s an awesome thing we’re trying to do, what the whole world is trying to do, what AI research is trying to do, [and that] is: understand intelligence, understand the mind, understand ourselves. We really should stop and pause to appreciate the magnitude of that goal, that ambition.” 

While Sutton, a self-professed libertarian, said he will remain in his position as University of Alberta professor, he acknowledged his plans to take his research in a more independent direction are perhaps better suited to his natural inclinations than his previous role at a multinational corporation. 

“I’ve always done my own thing, and I‘ve always been somewhat out of favour—it’s more normal to be in the contrarian role. In science, that’s not uncommon. And it’s often those who’ve been in a contrarian role [that] are proven right. That’s an old story.”