CALGARY⁠ — When DeepMind, Google’s London-based AI research arm, announced in 2017 it would open its first international research office in Edmonton, the decision was seen as a ringing endorsement of the local artificial intelligence community. 

DeepMind’s Edmonton office was centered around three prominent University of Alberta AI researchers—Michael Bowling, Patrick Pilarski and particularly Richard Sutton, a luminary in an area of study known as reinforcement learning—whose past academic work formed the foundation for DeepMind applications like AlphaGo, the first AI program to beat a human player at the ancient Chinese board game Go. 

The researchers joined DeepMind as part of the move, splitting their time between the company and their positions with the university and its offshoot, the Alberta Machine Intelligence Institute (Amii), as part of a larger bid to shore up the region’s research capabilities. Amii, along with Montreal’s Mila and the Vector Institute in Toronto, is one of three core centres of excellence within the federal government’s pan-Canadian AI strategy, a $444-million effort to make the country more competitive amid a fast-moving global race to develop the technologies. 

Last week, however, Google parent company Alphabet said it would close DeepMind’s Edmonton office as part of wider staff cuts, but offered researchers there the option to relocate to another DeepMind office. The decision came as a shock to many in Alberta’s tech community, who worry it could be a setback for the province’s efforts to market itself as a destination for world-class AI research. 

The closure of DeepMind’s Edmonton operation comes at a critical—some say existential—moment for Alphabet, whose continued dominance over the search business has come into question with the debut in recent weeks of ChatGPT, an AI chatbot built by San Francisco-based OpenAI. While some experts have dismissed it as old technology in new wrapping, the bot has gained instantaneous popularity for its ability to generate unique answers to complicated questions and display them in paragraphs of easily digestible text.

By Alphabet’s own assessment, ChatGPT’s underlying technology is a game changer for the search giant’s core business—a leap forward from the company’s current search service, which requires users to sift through pages of links to get the results they need. The bot’s popularity reportedly set off a “code red” at the company, which still makes most of its revenue from ads run in tandem with the links that populate its search results. In response, Google co-founders Larry Page and Sergey Brin reviewed the firm’s AI product strategy, which includes 20 new products to be launched this year, sources told The New York Times.

Those products include generative AI tools that would, for example, allow people to generate videos or virtually try on clothing or other products at e-commerce shops, according to an internal company presentation cited by the Times. 

Alphabet’s urgent pivot to new AI-based applications that can bolster the company’s core Google Search business may in part explain why it felt DeepMind’s Edmonton lab—which focused primarily on foundational research in areas like reinforcement learning—wasn’t vital to its larger strategy. ChatGPT and some of the text-based products currently under development at Google are underpinned by deep learning algorithms called large language models, which was not a specific area of focus at the Edmonton office.

“I’m not sure the [DeepMind] team here was doing a lot of work on large language models,” Amii CEO Cam Linke said in an interview with The Logic. 

“I want to be careful about speculating from the outside looking in, but it’s entirely possible that Google and DeepMind and everyone together felt like they needed to have a focus on large language models overall, and that might have to be at the expense of other areas.”

Asked why the company chose to shut its Edmonton office while retaining its Montreal and Toronto offices, Google Canada spokesperson Lauren Skelly said DeepMind Alberta was the company’s only office housed within a non-Google-owned building, and was therefore “more resource-intensive than DeepMind’s other sites to operate.” 

Sutton, Pilarski and Bowling did not respond to The Logic’s requests for comment. 

The three scientists and their collaborators have long been focused on developing the technological underpinnings of reinforcement learning.

Sutton—whom the U of A recruited from AT&T in the winter of 2003; he ran the institute’s reinforcement learning and AI lab until shortly after DeepMind’s arrival—literally wrote the textbook on AI reinforcement training, which uses a system of repetition and rewards to instruct artificially intelligent machines. The method seeks to construct machines in the image of the human mind, repeatedly telling the AI what is good or bad in much the same way parents teach children. 

“We don’t have learning systems; we have learned systems,” he said in a November 2021 interview with The Logic. Current examples of AI “learn in the lab,” he noted. Once installed, “they interact with the public and they do their jobs. They know their environment.” But they don’t learn from it the way a reinforcement learning system would. 

Reinforcement learning can be used to train large language models. But the U of A’s researchers appear to have mostly explored less search-relevant uses like logistics management, water-treatment optimization and prosthetics.

Bowling was part of a U of A research team that built DeepStack, an AI program that uses deep learning to play Texas hold ’em poker at a higher level than most professionals. Pilarski has been working to develop robotic prosthetic limbs that not only move but use AI to predict the precise nature of those movements. 

While Alphabet is trying to put AI to work to fend off the immediate threat to its business, DeepMind Alberta’s principals have been focused on the longer term. In the 2021 interview with The Logic, Sutton described his work at DeepMind as “absolutely basic, fundamental research” that requires a longer time horizon, and said his employer understood the approach. “That’s the way DeepMind thinks about it,” he said, “and the way Alphabet thinks about it.”

Sutton has not yet confirmed whether he will continue to work with DeepMind. 

“It is sad to lose the DeepMind office in Edmonton to the tech layoffs and looming recession,” Sutton said on Twitter shortly after the closure went public. “But AI is not going away, and I am more focused than ever on the Alberta Plan for AI research.” 

He said the three founders—himself, Pilarski and Bowling—would “stay in Alberta,” but didn’t clarify whether they would continue to work with the company. 

He also linked to the Alberta Plan for AI Research, a 12-step blueprint for renovating the foundations of reinforcement learning over the next five to 10 years which Bowling, Pilarski and Sutton published in August 2022. 

The plan is “not concerned with immediate applications of what we currently know how to do, but rather with filling in the gaps in our current understanding,” they wrote.

Regardless of what choices the lab’s researchers make about their professional futures, Linke said Sutton in particular is unlikely to change his area of focus.

“Rich’s obsession in life is understanding intelligence and continuing to grow our understanding of artificial intelligence, and he will continue to do that no matter what happens, to be frank.” 

Skelly did not confirm to The Logic how many people worked at the DeepMind Alberta, which occupied half of the second floor of the Amii building in downtown Edmonton. 

Though he was unsure how many people DeepMind employed, Linke said it was likely fewer than 100 based on the size of the office space—a drop in the pond for Alphabet, which plans to cut around 12,000 staff from its roughly 186,000-person global workforce, but a big number for Edmonton’s research and tech community.

While the DeepMind closure is a “brand hit for the region,” Linke said, the local AI industry and research community has grown to the point that it won’t shut down because of a single closure. 

When Amii first launched in 2017, the organization had 11 research fellows, Linke noted. Now it has 31. Companies including Attabotics, a Calgary-based warehouse robotics firm, and DrugBank Online, an Edmonton-based pharmaceutical database, have put AI at the core of the technologies they offer. 

Reinforcement learning will remain a crucial foundation in AI research, Linke said, and he would be “disappointed” if DeepMind’s closure signified a reduced focus in the area. 

“We’re just starting to see really the beginnings of the impact that reinforcement learning is having.”