The artificial intelligence boom presents Canada with unique opportunities and risks as we seek to benefit from a technology that could reshape how we live.

In this special series, Canada’s AI Advantage, The Logic examines how Canadian companies, investors, institutions and workers can gain from the country’s early lead in AI, even as Canada’s pioneers in the field become the world’s most powerful voices of caution.

As the last millennium was giving way to this one, Philippe Beaudoin found himself among about a dozen people watching as a man at the front of a room unfurled a vision of a technology’s future. “It was such a beautiful concept,” recalls Beaudoin, whose voice seems designed to express wonder. “More like alchemy than mathematics and science.” 

Yoshua Bengio was the man, his Université de Montréal lecture hall the room, artificial intelligence the technology. And the concept was neural networks, a methodology modelled on the human brain that was not in vogue at the time. “It was kind of rogue AI, almost,” says Beaudoin, then a master’s student.

Two short decades on, it is the mainstream. Startups are raising billions to build AI systems that respond to queries, develop new drugs and drive cars. Giant corporations are pivoting and partnering to use and sell the technology. Policymakers the world over are scrambling to regulate it.

But moving from alchemy to ubiquitous technology takes time, talent and not a little serendipity. In AI, Canada found itself with all three. A community of researchers led by Bengio and others dedicated to neural networks persevered through the skepticism of their computer science peers, and proved them wrong. They trained a new generation of AI modelled on the mind, and a generation of bright minds that would keep advancing those models. A public research program backed the work, making an unpopular bet that paid out spectacularly. All of it took decades.

Their innovations hold enormous potential—so much that they risk producing a familiar dynamic. Successive generations of Canadian technology companies have been pulled inexorably south, relocating or selling to Silicon Valley to meet their founders’ ambitions, or to deliver the returns their investors sought. And despite all of the country’s research heft, its early AI contenders could not escape that gravity. 

But this time may be different. Really. As nations jostle for the economic benefits—and the brand shine—that comes with AI preeminence, Canada’s claim is based on both its scientific history and its industrial present. Researchers in Montreal, Toronto and beyond demonstrated how the technology could work; many went on to the tech titans and challengers now putting it to work. Beaudoin’s Montreal-based Waverly is among the startups and scale-ups across the country looking to develop AI’s next world-changing, money-spinning application. 

“Why are we Canadian?” says Nicholas Frosst of Cohere, the company he co-founded that is now leading that crop of new contenders. “’Cuz we are.” Staying so is no longer radical. In AI’s modern era, the firm has felt little pressure to move its operation making generative tools to Silicon Valley. “It turns out a lot of people want to be in Toronto,” he says.

The same is true for Montreal, and Edmonton, and Vancouver, and plenty of points in between. But the phase after alchemy has become technology is to turn it into an economy, or at least a significant contributor to it. Canadian policymakers, founders and researchers insist the country can reap the rewards of the AI boom it helped instigate, while managing its risks. Now they just have to prove it. 

---

Ruslan Salakhutdinov was working at a bank in Toronto in 2005 when a chance encounter with an old professor and his math changed his trajectory. “I accidentally run into Geoff on the street, and he [says], ‘Come to my office, I’ll show you something really cool,’” recalls Salakhutdinov, laughing. Geoffrey Hinton, a computer science professor at the University of Toronto, had a fix for a long-standing challenge of neural networks. 

Let’s say you have a folder full of animal photos, and you want AI to pick out the ones with cats. (Awww.) Coders of conventional machine-learning systems might set out features that differentiate the felines from the canines and bovines, and train their algorithms on lots of images labelled by species. Deep-learning models—named for the multiple layers of software they contain—don’t need the tags. Researchers set the system on the images themselves, leaving it to figure out the features on which to focus.

The notion of brain-approximating machine learning has been around for decades. But for a long time, neural networks only “sort of worked, classifying things like corners of images,” says Steve Woods, whose computer science doctoral studies at the University of Waterloo in the 1990s focused on conventional, search-based approaches to AI. Deep learning “just didn’t scale.”

Part of the problem was that training a neural net took an age on the computers of the era. The innovation Hinton demonstrated to Salakhutdinov on that day in 2005 was to do it one layer of the network at a time, a more efficient method. “He showed me some of the mathematical formulations underlying these models,” says Salakhutdinov. “And then I got sold.” Instead of rising the banking ranks, Salakhutdinov returned to U of T to get his PhD in deep learning.

Outside, a neural-net winter had set in. Canadian AI luminaries credit Toronto-based CIFAR, a government-backed research organization, with continuing to fund the field here. In 2004, it committed $2.5 million over five years to Hinton’s plan for a new program called Neural Computation and Adaptive Perception (NCAP). It formed “a very vibrant, global network really pushing the boundaries of fundamental AI research,” says Elissa Strome, an executive director at CIFAR today. U de M’s Bengio was a member, as was Yann LeCun, a professor at New York University. (The duo would go on with Hinton to share the Turing Award, computing’s highest honour, in 2019.)

Every July or August, the NCAP program brought together faculty and students working in the field for five days of summer school. In the early years, the class held 15 or so people in a small conference room at U of T’s computer science department. The professors would give tutorials on new methods and models, and the students would present their research. “In the afternoon we had coding sessions,” says Salakhutdinov.

“I met so many people who now have tremendous careers in AI,” says Hugo Larochelle, then a PhD student under Bengio at U de M, who has his own tremendous career in AI research. “Creating this network was very useful as a junior researcher.” Summer school attendees included the likes of Roland Memisevic, Ilya Sutskever, and Raquel Urtasun, each of whom went on to launch a major AI company.

For all the warmth in those rooms, deep learning was not yet mainstream. Larochelle met a frosty reception from journals when he tried to publish his first paper. “The negative review was along the line, ‘The ideas are interesting. But it’s using neural networks, so no thank you,’” he says. The group persisted, motivated by the feeling that they were onto something.

A series of breakthroughs between the mid-2000s and 2010s—Larochelle and Salakhutdinov each co-authored key papers—thawed conditions. Running on faster graphics processing units (GPUs), deep learning began to beat conventional AI systems, starting with speech recognition. The real mainstreaming began in October 2012, when Hinton and students Sutskever and Alex Krizhevsky won that year’s ImageNet photograph-categorization competition with a neural net. Their paper, “ImageNet classification with deep convolutional neural networks,” has been cited more than 144,000 times.

Silicon Valley paid attention. The following March, Google acquired DNNresearch, the trio’s shell startup. The search giant reportedly paid US$44 million, beating out Baidu, DeepMind and Microsoft to hire the three. Hinton scaled back his academic commitments and started commuting to Google headquarters in Mountain View, Calif. 

Plenty of deep-learning talent followed in the next few years. A slew of young researchers went to Google, either to its internal Brain team or to DeepMind, which Google acquired in January 2014. Others left for the likes of Apple and Microsoft. Salakhutdinov himself left U of T’s computer science faculty in February 2016 for a tenured position at Carnegie Mellon University, and joined Apple part-time soon after. Pittsburgh just had more people working on machine learning, language and robotics. Canada missed out, says Salakhutdinov. “These are the drivers of that technology, and they’re all … in the U.S.” 

---

Canada still held AI aspirations, of course. A crop of startups launched in the last decade with dreams of platforms that would take the technology to more companies and consumers. But for many, their plans proved too expensive or came too early, leading to a string of exits to the titans of Silicon Valley. 

Apple’s launch of Siri in October 2011 gave many humans their first knowing encounter with artificial intelligence. The mobile assistant’s early rival had Canadian provenance: developer Maluuba. The startup in Waterloo, Ont., quickly signed deals with handset makers Samsung, LG and BlackBerry, and the assistant eventually leapt to smart TVs.

Maluuba’s founders wanted to evolve the user-machine interface, so that a conversation with an AI assistant felt like one with a human helper. Think of Her, suggests co-founder Kaheer Suleman. The Spike Jonze film, which debuted the year after Maluuba’s app, is about a man’s relationship with an AI operating system. Samantha, the OS, is the same whether she’s coming from his computer or his phone, and capable of empathy and common-sense reasoning. Romance ensues.    

But building Her in the early 2010s was only possible in fiction. Maluuba needed expensive cloud infrastructure to test its technology. “The researchers always used to complain internally, ‘We don’t have GPUs,’” says then-CEO Sam Pasupalak. 

AI also requires a lot of data on which to train. The startup scraped hundreds of hours of annotated audio from TED Talks to get started, according to co-founder Joshua Pantony, who also cited the difficulty of raising capital in Canada at the time. Silicon Valley investors posed different challenges. Pantony recalls one venture capitalist declining to do a deal after extensive due diligence, saying, “We love your technology, but I don’t want to do board meetings in the cold.”

Maluuba eventually raised US$8.4 million from backers including the venture arms of GM and Samsung, according to PitchBook data. By late 2016, it had a machine learning team of about 25, and a Montreal research lab working with Bengio’s AI institute, Mila. Pasupalak had lined up deals with car companies to introduce the technology into their hands-free systems. 

But the following January, Microsoft announced it would acquire Maluuba for an undisclosed sum. Microsoft got an infusion of technical talent, and Maluuba’s researchers got access to massive cloud infrastructure and huge amounts of text data. “It made a lot of sense,” says Suleman.

The cycle of start and sale played out again and again in the following years. As Maluuba was exiting, another ambitious attempt to build an AI platform was entering the scene in Montreal. In August 2016, a group of researchers and tech executives including Beaudoin and Bengio founded Element AI. The goal was to do for machine learning what Oracle had done for databases, or Amazon for cloud. “It was rooted in a commercial desire to build … a very important and fundamental product that would find application everywhere,” Beaudoin says.  

The company’s machine-learning bonafides quickly brought significant press and investor attention. Element raised US$257 million in total, PitchBook data shows. The company’s site boasted of hosting “the largest privately owned Canadian artificial intelligence R&D lab,” and a network of top faculty advisors. Bengio positioned Element as a bid to stop the brain drain of AI talent to Big Tech. 

But the company was slow to release products and generate significant sales, with revenue initially coming from one-off consulting projects. Others in Montreal’s burgeoning tech ecosystem were put off by its large-scale recruitment of AI researchers and graduates. “It basically created inflation [and] talent shortage in the space,” Louis Têtu, CEO of Quebec City-headquartered Coveo, told The Logic in late 2021. “The only purpose [to] bring together so many data scientists is to sell the workforce.”

Element did ultimately sell, to ServiceNow, a Santa Clara, Calif.-headquartered software firm, in November 2020 for US$228 million. “We were a bit too early,” says Beaudoin, noting no one has yet built Element’s vision of a universal platform for AI. But he hopes the company’s fate doesn’t discourage Canadian entrepreneurs and researchers from dreaming big commercial dreams.

“When it unfolded, the lesson thrown around was, ‘We got scammed’ or ‘Too much money was directed toward that thing,’” Beaudoin says. “I’m not sure. I think we should dare greatly.”  

---

“Why do I find neural nets interesting?” wonders Nick Frosst, considering. “Because I’ve always found them interesting. They just do stuff that you can’t do in another way.”

Here too, Hinton and his math had an influence. Neural networks cross the cognitive and computer sciences. Frosst, raised in the up-river Ottawa suburb of Manotick, Ont., had gone to U of T in 2011 to study both, and took Hinton’s neural-network introductory course. 

A couple of lectures in, the professor demonstrated how the system could be used to classify digits—a six, say. “Which is, first of all, very cool itself,” says Frosst. But then Hinton showed how if you ran the neural net in reverse, “it can draw a six. And it can draw a six that nobody’s seen before.’” 

These days, it’s Frosst’s neural-network creations that are interesting people—and businesses and investors. He’s a co-founder of Cohere, the Toronto-headquartered startup that’s fast become one of the country’s most prominent AI champions. Clients integrate the firm’s large language models (LLMs) into their products to generate text, summarize documents and improve search. 

Cohere’s path runs through Canada’s record of AI research prowess. After university, Frosst spent some time at Google in Waterloo, engineering neural networks for ad prediction tools. Then, a researcher gig opened up at his old professor’s Toronto lab. “Geoff called me: ‘I want two machine-learning geniuses who are also good software engineers,’” recalls Woods, then Google’s senior director of engineering for Canada. 

Frosst answered the call. At Google Brain, he met Aidan Gomez, another young Canadian researcher. During an internship at the firm’s San Francisco outpost, Gomez had been on the team that came up with the transformer, a new kind of model that considered words in context rather than one by one. Their June 2017 paper, “Attention is all you need,” has been cited more than 98,000 times. 

The transformer has gone on to transform the commercialization trajectory of AI, enabling the current boom in generative tools—it’s the “T” in ChatGPT, the viral query-fielding bot that San Francisco-based OpenAI launched last November. But the takeover took time. By mid-2019, Gomez was seeing plenty of applications within Google, but “the tech just wasn’t popping up everywhere like I expected.”

So Frosst, Gomez and co-founder Ivan Zhang struck out on their own with Cohere. In machine-learning terms, language is “general purpose,” says Frosst. Developers can construct a single, mega-model based on it, and set it to all kinds of tasks, like summarizing news articles or reading receipts. But an LLM requires huge amounts of data and compute, the sector shorthand for processing power. “They’re so hard to make, it’s really useful for a small number of companies to set out to try to make the best one they can.”

It’s not that small a number anymore. Over 1,100 privately held generative AI companies have raised a combined US$45 billion since the start of 2020. Cohere’s cohort at the top end of the LLM market includes OpenAI, where Sutskever is chief scientist, and its San Francisco neighbours, Adept and Anthropic. 

The old rules of tech that required promising startups to relocate to Silicon Valley no longer apply, Canadian AI founders say. Capital no longer cares about borders, and anyway, there’s plenty here. 

Cohere has raised US$440 million in total to date, according to PitchBook, and was valued at US$3 billion in the last reported deal involving its stock. Toronto-headquartered Radical Ventures wrote the startup its first cheque; Woods, now a partner at Montreal-based Inovia Capital, led Cohere’s US$270-million Series C in June. “We were very interested in what was going to form around them,” he says, citing opportunities to roll out its technology to Inovia’s portfolio firms.  

Cohere’s co-founders are Canadian. “We like it, and we want to stay,” Frosst says, adding that, “it turns out that Toronto is a pretty fucking great place to start an AI company.” He cites the ability to work with researchers at the city’s Vector Institute, and the stream of machine-learning engineers coming out of local universities. (Cohere is as global as any technology company today, of course, with offices in London and San Francisco.)

A few blocks away, scale-up Ada has also benefited from the country’s AI scene. “I don’t think it would have happened anywhere else,” says CEO and co-founder Mike Murchison. Clients use Ada’s platform to automate certain customer service functions via chatbots and AI phone responses. The theories of big Toronto thinkers like Hinton and Marshall McLuhan shaped how Murchison and his firm approach human-machine interactions. He also credits programs like the Creative Destruction Lab—an accelerator at the University of Toronto’s business school—with helping commercialize more AI research. The 24 firms in Ada’s cohort in the program have raised US$1.02 billion between them, according to PitchBook, although the three unicorns in the class account for three-quarters of the sum.

Ada and Cohere are among a pack of scale-ups building AI businesses here with name-brand clients around the world. Têtu’s Coveo, which makes search and recommendation tools, recently launched its own generative offering. “A lot of things that you hear out there have not quite found a way to monetization, and frankly immediate deployment or practical application,” he said on a May earnings call. “We’re the opposite.”

But the machine-learning universe has far more applications than just generative AI, in fields like drug discovery, climate science and beyond. Scientists who once attended NCAP summer school are now taking those ideas to the real world. In June 2021, Urtasun founded Toronto-based Waabi to raise deep learning’s role in autonomous vehicles from that of a bit player. Her goal is to “bring modern AI to self-driving,” she told The Logic the following February.

More researchers are keen to commercialize their work, says Strome, who manages CIFAR’s central role in the federal government’s $568.8-million Pan-Canadian AI Strategy. Since 2017, the program has helped subsidize universities to recruit or retain more than 120 research chairs, who have been teaching those new cohorts of Canadian AI workers. Strome says CIFAR has seen an increase in the number of researchers working with and starting companies of late.

Some 700 private Canadian AI companies have raised a combined US$7.61 billion since the start of 2020 and this year has already exceeded last, according to PitchBook data. Sector specialists in the venture community say the technology’s transformative potential is enormous, but acknowledge some investors will get burned.

With the advent of LLMs, “there are examples where, in a weekend, you could build an application which is staggeringly impressive,” says Woods of Inovia Capital. “But our view has changed a little lately, back to the traditional.” Startups that create valuable solutions to tangible problems and keep adapting will do well, he says.

Maluuba’s co-founders offer a study in contrast. Pantony left early, for a gig at Bloomberg; when the Microsoft money came in, he launched a new machine-learning firm, Boosted.ai. The Toronto-headquartered company sifts through reams of data to summarize important trends for portfolio managers. (It’s got its own financially savvy LLMs). The firm claims 190 clients, including leading banks and hedge funds. Pantony says he sought out “the biggest possible addressable market” with “a very obvious problem [where] technology is a solution.” In finance and AI, he thinks he’s found a good bet. 

Pasupalak and Suleman are working on a new startup of their own. While it’s still in stealth mode, both say the main product won’t be AI. “It’s very commoditized right now,” says Pasupalak, predicting that few firms now launching to sell applications built atop of others’ LLMs will make it big. “I want to build a real company with real users, solving a real market problem.”

Beyond startupland, corporate Canada is increasingly alive to AI. Foteini Agrafioti has seen the before and after. In September 2016, RBC launched Borealis AI, a lab that conducts research and builds products with the technology. “We recognized that the bank had a role to play in creating jobs for all of these great scientists that were graduating our schools [and] leaving the country,” says Agrafioti, who heads the unit. The lab has since produced stacks of papers. Commercial applications, too. 

In October 2020, RBC’s capital-markets arm launched Aiden, an auto-trading platform. It uses reinforcement learning, an approach that DeepMind had used to best one of the world’s best human Go players. “We knew that science was good for gaming applications,” says Agrafioti, noting that a financial market is a game of sorts. The world’s best human experts in reinforcement learning are at the University of Alberta, the academic home of field pioneer Rich Sutton. Borealis brought some on board, and sent staff back and forth from Toronto to Edmonton to learn. 

It worked. “Today, we send a lot of our trading flow through this platform,” says Agrafioti. 

In the world of bricks and mortar, construction giant EllisDon has used AI to improve cost and timeline projections, working off truckloads of data from hundreds of projects. The Mississauga, Ont.-based firm has piloted technology from startups, like a system that uses computer vision to capture time lapses of a building site, then compares it to the contractor’s models. 

EllisDon is also working with the Vector Institute to add AI tools to its project-management application, though the firm declined to disclose what those tools will do. “Clients are not asking specifically for AI,” says Brandon Milner, senior vice-president of digital and data engineering. But where it can be useful, he says, the company now has the technical chops to apply it. “It’s really a proactive approach.”

---

Canadian AI players say there’s work needed to ensure more of the country’s replenished stock of startups scale, and fewer of its field-changing researchers leave.

The federal government has pledged hundreds of millions in funding via its national strategy and financing for individual firms. But executives are pushing Ottawa to consider procurement and policy, too. Departments could use AI to help deliver services like call centres and immigration processing, says Ada’s Murchison. “That demonstrates to the country how this is technology not to be feared but to be augmented and celebrated, and which actually makes the lives of all Canadian citizens better.”

Murchison is also hoping for a shift in Ottawa’s policy approach. Officials are “very risk conscious,” he says, after participating in recent roundtables. While harmful AI uses must be addressed, he’s pushing for more policy to encourage beneficial ones. 

In September, the Council of Canadian Innovators, a scale-up lobby group, called for Ottawa to move faster on regulating AI and give firms more freedom to launch less-risky applications. Prolonged uncertainty could cost Canada, the group warned—the United States is just a quick re-incorporation away. To slow the flow, it’s also lobbying officials for an AI commercialization and IP strategy. 

(Policy researchers have argued that Ottawa’s own internal automations need close scrutiny and that its proposed AI law doesn’t address the technology’s human-rights impacts.)

Other executives see opportunity in new areas of AI. “I believe in technology being useful to solve important societal problems, and I also want Montreal and Canada to have a voice in that,” says Larochelle. In November 2016, he returned to the city to lead a new Google Brain unit there. (Small world: Bengio’s brother Samy recruited him.) 

Of late, Larochelle and some of his team have focused on sustainability problems. One researcher worked on Project Sunroof, a tool to encourage homeowners to install solar panels by estimating how much they’d save. Others are working on a bioacoustics system designed to identify birds from song, keeping an ear on population numbers to aid conservation. “I’m hopeful that [sustainability] becomes a thriving area in AI,” says Larochelle, now a principal scientist at Google DeepMind.

Canada has long been the domain of business-to-business, software-as-a-service firms. But the world’s biggest tech companies make things for people, so the country needs more founders and funders pursuing consumer AI applications, says Beaudoin. In June 2020, he launched Waverly to build a social platform using “human-first” algorithms.

Amid all the talk of AI applications, CIFAR’s Strome says academia can act as a counterweight to industry. “We’re not all going to trust OpenAI when they say their models are safe,” she says. “They need somebody else to verify that.” And she says fundamental research is still, well, fundamental. CIFAR still runs a summer school; this July, more than 150 AI students attended.  Without the public support that sustained Canada’s deep-learning community through the neural net winter, you may never have gotten to try ChatGPT. 

---

Billie Joe Armstrong, Harry Styles and Gerard Way have all taken the stage at the Garage, a north London music venue of some indie notoriety. At the end of September, Nick Frosst joined that list of iconic frontmen. 

The Cohere co-founder is also the singer of Good Kid, a rockish band with a wholesome vibe composed of five coders. “I don’t use our models to generate lyrics,” Frosst says, answering a question he gets asked a lot. “What I want out of art and artistry and lyric writing is self-expression.” Speeding up the creative process isn’t the point. 

But of late, Frosst has begun using Cohere’s models as a sounding board, inputting his lyrics and asking the system to say what they mean. “I use that sometimes to figure out, ‘Is the metaphor or the imagery I’m using too subtle? Or am I banging it over the head?’”

Allow me one of those banging-over-the-head similes. In it, Cohere could be to AI today what Nortel or BlackBerry were, respectively, to networking and smartphones when those companies were in their primes—Canada’s champion in a strategic and lucrative technology sector. Such firms have a history of falling, or leaving the country.

But this time may be different—many champions instead of one, with fewer reasons to exit. Frosst, for one, doesn’t think of Cohere as analogous to Nortel or BlackBerry—perhaps, he says, “because I know their full stories.” His firm has bigger ambitions than either of those ever did. Besides, he notes, “there are other excellent Canadian AI companies.” 

Still, being the first name to mind when people think of a successful Canadian AI company appeals. “That would be very meaningful to me,” Frosst says. “In a few years, I think we can deliver on that promise. But we still have a lot of work ahead of us.”