EDMONTON — Richard Sutton, the Canadian AI pioneer who until recently led Google’s research efforts in Edmonton, is partnering with a new venture launched by a renowned video-game developer that seeks to address the biggest questions about artificial intelligence.

Sutton will join Keen Technologies, a five-person company founded last year by John Carmack. The former chief technology officer of Oculus VR, a virtual-reality headset maker which Meta acquired in 2014, Carmack is perhaps still best known for co-founding gaming studio Id Software, where he was the lead programmer on the likes of Doom and Quake. At Keen, Sutton will lead research efforts that seek to better understand the core principles underlying AI, including what intelligence is and how it can be learned.

At an announcement Monday evening at the Alberta Machine Intelligence Institute headquarters in Edmonton, Sutton and Carmack—both of whom recently stepped down from prestigious positions at major tech companies—made clear they want Keen to operate outside the narrow confines of current AI ideas, pushing back against the “groupthink” they believe has taken hold of the rapidly growing sector.

Artificial intelligence research has followed familiar tracks of thinking, Sutton said, leading to widespread “copycatting” of past successes. He cited the recent infatuation with OpenAI’s ChatGPT, which has incited a flood of investment into AI startups and has led a host of companies, including Google, to accelerate development of their own large language model chatbots. 

Sutton, a University of Alberta professor, is considered the grandfather of the theories underpinning reinforcement learning, a class of AI research that involves teaching a machine to learn through repetitive feedback. He and other U of A researchers were responsible for the success of AlphaGo, developed by Google subsidiary DeepMind, which achieved a milestone in 2016 when it beat one of the world’s best Go players. 

Alongside Carmack at Keen, Sutton hopes to bring a fresh perspective to research in the field—one that doesn’t rely on past successes. 

“I do think of myself a little bit as a rebel,” Sutton said in an interview. “It’s a bit contrary, and it’s not what everyone else is doing, but we both think this is obviously the right thing to do.”

The Logic was first to report earlier this year that Sutton had decided to part ways with Google after the company closed its DeepMind office in Edmonton. At the time, Sutton also expressed concern over how the AI community’s sudden infatuation with certain machine-learning technologies—including large language models of the kind used in ChatGPT—were drawing crucial attention away from larger existential or theoretical questions. Google’s growing focus on large language models appears partly responsible for Sutton’s decision. 

Upon his departure from DeepMind, Sutton announced plans to establish an independent organization focused on fundamental AI research. The new non-profit, to be called Open Mind Research, is expected to launch this fall and will focus on the Alberta Plan, a 12-step guide for renovating the fundamentals of AI research Sutton published with his fellow U of A professors Michael Bowling and Patrick Pilarski.

Carmack was among the co-founders of Id Software, a Texas-based video-game developer whose classic titles laid the foundation for the modern first-person shooter genre. After joining Oculus VR as chief technology officer, he served as one of Meta’s chief virtual-reality consultants. Following a somewhat messy departure from Meta—he later said the tech giant was hampered by bureaucracy and crippling inefficiency—Carmack plunged into AI with the launch of Keen last year.

In August, Carmack announced Keen had raised US$20 million in seed funding from investors including Shopify co-founder and CEO Tobi Lütke, California venture capital firm Sequoia Capital, Stripe co-founder and CEO Patrick Collison, former GitHub CEO Nat Friedman and Cue founder Daniel Gross. 

Keen intends to develop fundamental algorithms that could feed into an artificially intelligent program by 2030, Sutton said. 

He is one of many experts who have predicted that 2030 could mark a tipping point for AI, when hardware and computing capabilities will be developed enough to allow for true general artificial intelligence. (Whereas current AI technology is only capable of accomplishing very specific tasks, general intelligence would include the human-like ability to solve a much wider array of problems.)  

In an interview Monday, Carmack said Keen loosely aims to develop a prototype by around that time that will have general intelligence equivalent to that of a toddler or young child. To do that, the company first aims to better understand the limits on current AI capabilities and then use that knowledge to develop a blueprint for learning; researchers and companies don’t yet fully understand what questions they need to solve for, Carmack said, let alone apply technology to solve those problems.  

“We don’t even have line of sight,” he said. “Like, we don’t know we need to solve problem A, B and C. We [only] know that there’s a number of problems here, and we have a reason to believe that these problems are tractable—that it’s not something that requires some inhuman level of insight.” 

Sutton, for his part, will continue to contemplate what it means to be intelligent, and what machines need to reach that threshold. 

“We should be thinking about it as addressing fundamental questions, like what it means to be a person and what’s our place in the universe—and what we should aspire to do as a society and civilization.”

Correction: This story has been updated to correct the name of the Alberta Machine Intelligence Institute.