Treatise · The History of AI in Canada
Two of the three founders of deep learning live in Canada — Geoffrey Hinton (1947–) in Toronto and Yoshua Bengio (1964–) in Montreal. With under forty million people and a GDP one-tenth America's, this country leveraged a small program called CIFAR to swing the most consequential paradigm shift in AI history. It was one of the few embers kept alive through connectionism's deepest winter; in 2024 it received the most dramatic Nobel endorsement in AI's young public life.
I. Embers Through Winter: Two Outsiders Travel North
The turning point came in 1987.
Geoffrey Hinton was thirty-nine, on the faculty at Carnegie Mellon. In the mid-1980s, the Reagan administration's defense funding was increasingly tied to military AI; he refused to let his research be used for killing. His wife, meanwhile, disliked the American medical system. In 1987, the chair of the University of Toronto's computer science department called; Hinton accepted a professorship there and moved his family north to a city unusually welcoming to foreign scholars. The move had no obvious meaning at the time; twenty years later it had altered the course of AI. America lost a researcher who would later receive both the Turing Award and the Nobel Prize in Physics; Canada acquired the future "capital of deep learning."
Six years later, in 1993, the thirty-five-year-old Yoshua Bengio completed a postdoc at Bell Labs and joined the Université de Montréal. Born in Paris and raised in francophone Quebec, having earned his bachelor's and PhD at McGill — this was, for him, a homecoming. Montreal would become the city he cultivated for the next thirty years; he founded the LISA lab there, later renamed MILA.
Through the winter, Hinton was in Toronto and Bengio was in Montreal — a thousand Canadian kilometers apart. Almost no one noticed at the time that this Toronto–Montreal axis would become the source of the next AI tidal wave.
II. CIFAR's Foresight: Betting When AI Was Coldest
The Canadian government's decisive contribution did not come through high-profile policy declarations but through a small body called CIFAR — the Canadian Institute for Advanced Research.
Founded in 1982 by entrepreneur Fraser Mustard, CIFAR was created to fund "high-risk, high-reward" interdisciplinary basic research. In 2004, it launched Neural Computation and Adaptive Perception (NCAP), led by Hinton. The program offered roughly one million Canadian dollars per year — trivial by American big-lab standards — but in an age when neural networks attracted almost no investment, it gathered the few remaining true believers worldwide: Hinton and Bengio in Canada, Yann LeCun (1960–) in New York, Sue Becker, Terrence Sejnowski, Bruno Olshausen from elsewhere across North America and Europe. A dozen people met twice a year in Banff for closed-door reviews of one another's unpublished work, then returned to their universities to keep going.
In 2017 the program was renamed Learning in Machines & Brains and continues to this day. CIFAR's president Alan Bernstein recalled in 2024: "We did not ask them to deliver in the short term. There were no KPIs. Each year we asked one question: What do you want to do?" Hinton has repeatedly said in public that CIFAR was the key support that made it possible to keep going through twenty of the field's hardest years.
In 2017, building on years of CIFAR experience, the Canadian government allocated 125 million Canadian dollars to launch the Pan-Canadian AI Strategy — the world's first national AI strategy.
III. Three Pillars: Vector, Mila, and Amii
The strategy materialized as three institutes.
Vector Institute in Toronto was founded in March 2017, with Hinton as Chief Scientific Advisor and Tomi Poutanen as CEO. Anchored at the University of Toronto and University Health Network, it focuses on deep learning and medical AI; its board combines federal and Ontario provincial governments with companies including Google, NVIDIA, and the Royal Bank of Canada. Vector has since trained Aidan Gomez (founder of Cohere), Sara Hooker (head of Cohere For AI), and others.
Mila (the Montreal Institute for Learning Algorithms) was founded by Bengio in 1993 and formally restructured under the name Mila in 2017. It merged the AI groups of Université de Montréal and McGill, becoming one of the world's largest deep-learning academic institutes, with close to a thousand researchers and PhD students at any time. Mila is the birthplace of generative adversarial networks (GANs — proposed by Ian Goodfellow (1985–) during his Mila PhD), early attention-mechanism research, neural machine translation, and the Theano framework.
Amii (the Alberta Machine Intelligence Institute, in Edmonton) has the longest history; its predecessor traces back to the RLAI lab founded in 2002 at the University of Alberta by Rich Sutton and Michael Bowling. Reinforcement learning is Amii's specialty: the Cepheus and DeepStack poker AIs (DeepStack, in 2017, was the first to defeat professional players at heads-up no-limit Texas Hold'em); contributions from Alberta-trained researchers to early AlphaGo; and Sutton and Andrew Barto's Reinforcement Learning: An Introduction, the field's textbook. Sutton and Barto received the 2024 Turing Award.
The three differ in emphasis: Vector leads in applied AI and medical AI, MILA in deep-learning theory and generative models, Amii in reinforcement learning. They share federal strategic funding while preserving academic autonomy — a distinctive "national support plus regional self-governance" structure characteristic of Canadian AI.
IV. Canadian AI Companies: Cohere, Element AI, and DeepMind's Canadian Bloodline
Canada has produced world-class researchers but has long struggled to keep its companies — this is the other half of the story.
The most representative homegrown company is Cohere, founded in 2019 in Toronto by former Google Brain researchers Aidan Gomez (one of the eight authors of Attention Is All You Need), Nick Frosst, and Ivan Zhang. Cohere targets enterprise AI with a focus on RAG, embeddings, and multilingual foundation models, raising more than one billion dollars cumulatively in 2024–2025 at a valuation around 5.5 billion dollars. It chose not to build a consumer chatbot, selling its model as enterprise APIs — an approach distinct from the Silicon Valley mainstream.
Element AI is the other half of the story. Founded in Montreal in 2016 by Bengio, Jean-François Gagné, and Nicolas Chapados, it was once hailed as Canada's first AI unicorn. Its business model wavered between consulting and product, and in 2020 it was acquired by ServiceNow for 230 million dollars at a sharply diminished valuation — an emblematic case of Canada's academia-to-industry break.
DeepMind, technically a UK company, drew heavily from Canadian talent and methodology in its early years — Hinton was for a long time a senior advisor at Google; David Silver, Canadian-trained, was a primary architect of AlphaGo; and the AlphaGo paper that defeated Lee Sedol (1983–) lists multiple authors of Canadian background. DeepMind sits in London, but much of its bloodline is Canadian.
Beyond these, Canada is home to the ServiceNow R&D center, Recursion Pharmaceuticals (split between Montreal and Salt Lake City), and Tenstorrent (Toronto, AI chips, led by Jim Keller).
V. Immigration Policy and Talent Concentration
If one asks "why Canada?", immigration policy gives a concrete answer.
Canada has long run an Express Entry system that scores applicants on skills, allowing STEM-credentialed talent to obtain permanent residency in months. In 2017, the Trudeau government launched the Global Talent Stream, promising work-permit processing within two weeks for high-skilled tech workers — sharp contrast with the American H-1B lottery and the friction of converting student visas to work visas. According to data published by Vector Institute, the number of AI workers in the Toronto area grew more than sixfold between 2017 and 2023.
In 2023, the Trudeau government launched the Tech Talent Strategy, explicitly encouraging U.S. H-1B holders to migrate to Canada. The first three days hit the application cap of ten thousand. Toronto, Montreal, and Vancouver became, after San Francisco, the densest AI-engineering hubs in North America.
But talent inflow does not equal corporate retention. A substantial share of Vector and Mila PhDs still head south after graduation to join Google, Meta, or OpenAI — salary gaps, IPO pathways, and the depth of the M&A market remain Canada's structural shortcomings.
VI. Contemporary Influence: From Academia to Two Nobel Prizes
On October 8, 2024, the Royal Swedish Academy of Sciences awarded the Nobel Prize in Physics to John Hopfield (1933–) and Hinton, "for foundational discoveries that have made machine learning with artificial neural networks possible." Hinton answered the call from Stockholm in his Toronto home, in a bathrobe — an image that flashed around the world. The next day, Demis Hassabis (1976–) and John Jumper received the Nobel Prize in Chemistry for AlphaFold. DeepMind sits in London, but Hassabis's early exchanges with the Toronto school were close, and AlphaFold's methodology grows from the Canadian deep-learning tradition.
Two years earlier, in 2018, Hinton, Bengio, and LeCun had jointly received the Turing Award. In 2024, Sutton and Barto received the Turing Award again, this time for reinforcement learning. In a span of seven years, the Canadian school went from two Turing Awards to two Nobel Prizes, lifting the "university plus CIFAR" model to the highest podiums of world academia.
In May 2023 Hinton resigned as Vice President at Google to warn publicly about AI risks. He later explained: I used to tell myself, "If I don't do this, someone else will"; now I believe that, at least, I can speak freely. His departure was not defection but continuity — an old man insisting on remaining a heretic at a new stage.
VII. Canada's AI Strategy: SCALE AI and the Path Forward
Beyond the Pan-Canadian AI Strategy, the Canadian government in 2018 funded SCALE AI — a 230 million Canadian dollar "AI supply-chain super-cluster" headquartered in Montreal — through the Innovation Superclusters Initiative (ISI), aimed at pushing AI into retail, logistics, and manufacturing. By 2024, SCALE AI had funded more than 130 industrial projects.
In 2024, the Canadian government announced a 2.4 billion Canadian dollar investment in the Canadian Sovereign AI Compute Strategy, targeting a domestic GPU cluster and public compute quotas by 2026 so that companies like Cohere need not depend solely on AWS, GCP, or Azure. That same year, the Artificial Intelligence and Data Act (AIDA) advanced in Parliament, defining compliance obligations for high-risk AI systems and echoing the EU AI Act.
Canada's open question is scale — how to convert world-class research, world-class talent, and world-class policy into world-class companies. Cohere is one answer; one example does not make a pattern. If, over the next decade, Canada cannot incubate at least one AI company at the hundred-billion-dollar scale, the title "deep learning capital" may remain only an academic honorific.
Historian's Note
I have observed half a century of Canadian AI and admire its art of leveraging the small against the large. A vast land with sparse population, industry less abundant than America's, capital thinner than Wall Street's — yet at the deepest moment of the neural-network winter, this country alone dared to keep funding Hinton and Bengio through the laughter of the rest. CIFAR's program, at one million per year, came to no more than twenty million Canadian dollars in two decades — and it incubated a paradigm shift that changed humanity. The return on investment is rarely matched in history. Yet Canada has its pain too: DeepMind landed in London, Element AI was sold to America, the majority of Vector and Mila PhDs travel south. Training world-class talent and retaining world-class companies are two different things. Sima Qian wrote, "where there is impasse there is change; where there is change there is connection." Canadian AI's change has moved from academia to policy, from policy to infrastructure; the next step is from infrastructure to industrial ecosystem. Whether that step can be taken depends not on the laboratories of Toronto and Montreal, but on Ottawa's policy, Vancouver's capital, and the gravitational pull of that vast neighbor not far away.
Eyewitness Accounts
Call for contributions
If you have worked in AI research or built companies in Canada, please contribute on GitHub.
References
- Hinton, G. E., Osindero, S., & Teh, Y. W. (2006). A fast learning algorithm for deep belief nets. Neural Computation, 18(7), 1527-1554.
- CIFAR. (2017). Pan-Canadian Artificial Intelligence Strategy.
- Bengio, Y., Courville, A., & Vincent, P. (2013). Representation learning: A review and new perspectives. IEEE TPAMI, 35(8), 1798-1828.
- Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction (2nd ed.). MIT Press.
- Goodfellow, I., et al. (2014). Generative adversarial nets. NeurIPS.
- Government of Canada. (2024). Canadian Sovereign AI Compute Strategy. ISED.
- Royal Swedish Academy of Sciences. (2024). The Nobel Prize in Physics 2024. nobelprize.org.
- Cohere. (2024). Command R+ Technical Overview.