So, thank you, Minister Beth Falvi. I particularly want to thank him for many years of his personal leadership, and I think if you check his track record on the work he's done around digitizing the Ontario government, you'll be just as impressed as I am. I'd also like to thank the premier, ministers Crawford, Fideli, and Mol Rooney for their partnership. And a special shout out to IP Ontario who we've partnered with to protect over a hundred assets to date and we've got more coming in the future. Good morning everyone. My name is Glenda Crisp. I'm the president and CEO of Vector Institute. Thank you for joining us here at Remarkable 2026 and welcome to our very large virtual audience as well.
So, I'm excited to dive into our very first conversation because we have someone here who represents exactly what this conference is about. I'm thrilled to welcome Nick Frosst, the co-founder of Cohere, someone whose journey embodies Canadian AI leadership. Nick's story starts in our research community. He studied and worked with Jeff Hinton at Google Brain, contributing to the foundational AI work that positions Canada globally. He then went on to co-found Cohere, turning research excellence into adoption. I guess in his part-time time that he's got available, shockingly, he also leads the Juno nominated band Good Kid, and their debut album's coming out in April. So, I'm going to keep an eye out for that and I encourage you to do likewise. But that's a perfect example of how creative thinking can span disciplines. Cohere has grown to a 7 billion dollar valuation by solving real deployment challenges for large enterprises like RBC, McKinsey, and Bell. They've proven that Canadian AI research can translate into practical solutions that compete globally while maintaining focus on what businesses actually need. AI that works in complex and often regulated environments. Cohere's partnership with the Canadian government on AI sovereignty and domestic compute infrastructure is a great example of how breakthrough research can drive economic benefits that stay here. That's why Nick is such a great person to have with us today as we explore some very key questions such as how do we turn world-class research into economic advantage and what does AI sovereignty actually mean in practice? As someone actively trying to balance research advancement with successful deployment, technical innovation with practical implementation, and enterprise adoption with Canadian values, Nick is uniquely positioned to address these challenges. And I think those challenges will determine whether Canada leads or lags in AI. So, please join me in welcoming Nick to the stage. Thank you.
I'm a little worried they might start playing music to play us off. We'll see how that goes. So Nick, in terms of research going to commercial adoption, if we were to take a time machine to 2030 and then take a look back, what would you say Cohere's impact on Canada was beyond headlines, beyond valuation?
Our objective as a business is to change the way work is done, right? I think we're all pretty familiar and have incorporated bits of AI into our daily lives, right? Like lots of Google searches these days, people are looking at the LLM generated result. People are going to consumer chatbots to help them with little bits of tasks here and there. But if you look at the average knowledge worker, it's still hardly impacted the way they do work. That's not too surprising when you look at the rollout of other technologies, right? There are still places where you have to fax things in. Technology getting into enterprise does take time, but this technology is so useful and there's such a real urgency to get it into enterprise that I think by 2030, if we're successful, anybody who arrives to their desk to do work for the day will open up maybe an email client and they'll open up some terminal for using an LLM. And anytime they don't, there's a boring task that they need to do, they'll get the AI to do it for them. So that's a pretty fundamental change to the way we do work. And that's what we're trying to do. We're trying to do that globally, right? I think if we're able to do that, that will affect the way Canada works. But it will also affect Canada's position as an AI leader, right? Like for the past several decades, most of the world's experience of technology has been defined by Silicon Valley, has been defined by America. And I think at this stage, we're all pretty aware that that isn't great. That the vision of life, that the view of the way the world should be coming out of Silicon Valley is not really the one that a lot of the world wants, especially it's not the one that Canada wants. It's not the one that Europe wants. So, if Cohere is able to be successful and we transform the way work is done, I hope in the process we're able to help build up an AI ecosystem. That means Canada doesn't need to be just following along the vision of tech that comes out of the states, but instead we can be making our own.
That's very inspiring. I'm with you on that. I encourage everybody here. Let's make that happen. So, Cohere makes significant contributions to research while also building commercial products. So, how do you actually balance that? How do you balance research with the pressures from product development?
That's also a great question. So, before this, I was at Google Brain and while I was at Google Brain, this was in 2016, 2017 to 2020. During that time, I took pride in my lack of contribution to the products of Google. Every day that I came home having done nothing of value was a success as a researcher. Because I was like, well, you know, if they're paying me to pontificate, how good must I be at pontification, you know? Pretty awesome. And so we did lots of work that was all very interesting but none of it, like my objective was never cool as am I making a better product. I just wanted to work on pure research and that was really fun and I really encourage people to do some of that in their lives. I think thinking and research for research's sake is a great thing to do and it leads to the types of things that we see now, right? Like if Jeff wasn't stubbornly thinking about the way the brain works or might works, he wouldn't have stayed with neural networks and they wouldn't be here right now, right? So that's really important. But at some point I got bored of that and I was like what I want to do is build a product. I want to build things that people use. It turns out that if you're building AI systems, there's no way to build them without doing research, right? Like there are novel questions that need novel answers and you cannot build a good product for people unless you're tackling those questions. So it's important to carve a distinction between research as in the idea of answering unanswered questions with opening up LaTeX and publishing papers in the right format and like reviewing things. All the work that we do at Cohere is research, all of it. We're constantly answering questions. We're constantly trying to push boundaries, build things that haven't been built before, some of that we publish. And we publish the things when we think we've discovered something that is relevant and interesting. And we do that for the purposes of notoriety and goodwill and attracting good talent and fostering a community. But the tension doesn't really exist, okay? Because the work that we're doing necessitates novel research.
That sounds like it's fairly unique to AI.
I think that's pretty unique to AI. Yeah. I think other companies where, you know, maybe more distinct, where what you need to do is kind of known. You just have to do a good job of it. Maybe there's a little more tension there.
So maybe on some technical, and feel free to go as technical as you want for the researchers in the audience. So Cohere has leaned hard into RAG, embeddings, and flexible deployment rather than chasing a single do-everything frontier model. So what contrarian technical bet did you make that you think is still maybe misunderstood in 2026?
That's a good question. I think our investment in embedding models and RAG has been, well, I don't know if it's contrarian but it's certainly not exciting. Like we have the best, we make the best embedding models, we make the best multimodal embedding models, we make super efficient embedding models that can be run at, we tried to get them down to binary embedding but that didn't quite work so there, but you know, very small footprint of both the model and of the embeddings themselves. And I think there's, I keep seeing people say like, 'Oh, RAG is dead.' Like, you don't need to do search. You just put it all in the context. And that works for cool kind of personal demos. Maybe that works for your own life. You're going to do, you know, you can just have an agent go through all the documents you've written on your computer and like that'll be fine. But if you're operating at scale, if you're operating in an enterprise, that just continues to fail. Like when you're looking at, you know, like 10 million documents, you can't have an agent load them up into context and ask is this relevant to the question? Yes. No. Like you need a better search. And so our investment into embeddings and RAG continues to pay off in the enterprise deployment. And that's not obvious to people using this technology on a personal basis, right?
Yeah. No, that makes sense. So you've also talked about not building a digital god. So what's the vision for AI that you want Cohere and perhaps Canada to be known for?
Yeah, it's a good question too. Yeah, when I talk about not building a digital god, I mean to say a lot of what you hear from the labs these days about like the impending AGI, you know, we're two steps away from turning on something that can do everything. And it's kind of interesting. I feel like we're in a little bit of that arc again. Suddenly a bunch of people got exposed to coding models and then every time the world gets exposed to something that is slightly better than they expected it to be, I see an uptick in the rhetoric of okay, therefore we're going to have a digital god in a second. But we kind of forget like two years ago, 2023, there was routinely discussion about the end of the world. There was routine discussion about people saying, 'We're about to build a thing that will create a bioweapon and kill all people.' I like to ask this question when giving talks, and it's particularly interesting when giving talks like at the Vector Institute, but how many of you think LLMs pose an existential threat to humanity? A show of hands. Not seeing a one. I think I saw two. Two. Okay, cool. Sorry to put you on the spot there, but so yeah. So two. So there was a moment in 2023 where I think if I asked that question even at the Vector Institute, there would have been more hands, right? And that's certainly like you can go back and read breathless articles from any journalist talking about how we are, you know, we're minutes away from this. I don't think that's what autoregressive sequence models are. I don't think they pose a threat like that. I think by the nature of how they are built and how they work, they have fundamental limitations. I think those limitations are drastically underestimated by people and you can build truly like economy-changing, world-changing technology with LLMs and that's what we're seeing right now and people are excited about coding models. And you can imagine how that kind of coding models are going to be able to do stuff like they're doing for coding for every work that's done on a computer. But I don't think we're going to build something that you treat as a person. I don't think we're going to build something that you have an expectation of like that you collaborate with in the same way you collaborate with people. And when I talk about not building a digital god, that's what I'm talking about. I'm saying, 'Hey, we have to acknowledge this as a technology. Super useful, going to fundamentally change the way computers work, the way the economy works, but it is not the AI of sci-fi. It's not just a really smart person in a computer. It's not a god inside a computer. It is something else.' And that's important to focus on. So when like what do I want us to be known for as the years go on and that vision of reality of like we're about to be able to solve all problems. You're going to be left behind. There's going to be no jobs anymore. It's just going to be this one computer sitting in one place doing everything. As that becomes more and more ridiculous. I want us to be known for actually understanding how the tech works and making something useful the entire time.
Right? Because the World Economic Forum last year about this time came out and said they estimated 170 million new jobs with 92 million displaced. So a net of 78 million gain in terms of jobs. So I think the question there is or maybe the point for us is we want to make sure that at least our fair share, if not more, of those new jobs end up here in Canada as opposed to elsewhere.
