Can AI Do Good?

As workplaces across Canada rapidly embrace automation, employees are starting to sweat. Silicon Valley—with its stock-inflating, human-replacing, data-centres-on-the-moon discourse—isn’t helping. But leave it to a Canadian to be the AI world’s voice of reason. Nick Frosst is one-third of the co-founding team behind Cohere, a Toronto-based multinational firm that develops and trains large language models, or LLMs, on companies’ internal data, helping them find ways to lower costs and boost productivity. Its AI agents aren’t here to take your job, Frosst insists. But they could take a few of your calls.

Related: White-Collar Workers Are Not Okay

Cohere’s corporate-world focus has, so far, been the secret to its success, with clients as varied as Bell, RBC and the federal government, and a C-suite that includes former Uber and Meta execs. In seven short years of business, its valuation has risen to US$7 billion. For his part, Frosst—Geoffrey Hinton’s first hire at Google Brain—has zero interest in industry hyperbole and zero desire to build “a god inside a computer,” as he’s put it. Instead of worrying about what AI could do, he says, just put it to work doing what it does best.

Cohere’s key product used to be chatbots for businesses. Now, it’s North, a platform powered by agentic AI. What’s the difference?
A few years ago, we were talking about chatbots, like, “Hey, we’ve made a computer you can have a conversation with.” Now, we’ve figured out not only how to make the tech really good at predicting the next word in a sentence, but also the next line of code or action to take. That means you can ask a model to do a task, instead of just write text back to you. For example, I’ll say to North, “Somebody sent me a prep doc for a meeting. Can you find it in Slack and summarize it?”

So they’ve gone from talkers to doers.
Yeah, at a very base level, that’s a good way to put it.

Jack Clark, co-founder of Anthropic (makers of Claude), calls these agents “little troublesome genies.”
I don’t know if that’s a super-helpful thought construct.

The promise of integrating LLMs into workplaces is that they’ll free up humans from grunt work so they can focus on higher-minded, creative tasks. Have North’s users found that to be true?
LLMs are great at doing the kinds of things people find tedious. I’ve heard of bankers using them to analyze quarterly earnings reports for companies they’re tracking. They’re able to spend more time thinking strategically, as opposed to reviewing a bunch of documents. I recently saw a deployment in a health-care application that has something like 1.5 million patients. They used to have to call somebody to book appointments and, now, they can do it through a chatbot. Booking appointments is not something people should spend a lot of time on.

Evan Solomon, Canada’s AI minister, touts AI as a productivity unlocker, but StatsCan has reported marginal increases on that front so far. One widely shared MIT study even found that 95 per cent of generative-AI pilots in workplaces fail. This kind of data makes people doubt the hype.
We’ve looked at that MIT study a lot. But we’ve run similar numbers to figure out what percentage of Cohere’s deployments have gone into production versus gotten stuck in demo. We’re closer to the inverse of MIT’s number. There’s a huge gap between a “cool thing” and something that actually adds value for a business. Sometimes, companies try to use LLMs made for consumers, so they don’t have the right privacy requirements—or it costs them too much. I’ve also never felt that population-level metrics of productivity are super-great. It’s much better to go into specific organizations and see whether people are able to operate better.

Coders used to have a ton of job security; now, they’re some of the first workers on the chopping block. One day, LLMs advance our cognition; the next, they’re making us lazier. What would you say to Canadians who are craving a single morsel of certainty?
I often think about the printing press, electricity and the internet. During each of those advances, people were like, “We’ve never seen this before!” Massively impactful technologies bring about tumult as they’re integrated into the economy, but everybody mostly looks back and thinks, That was a good idea.

Do you think the same will be true of the AI revolution? “No complaints!”
We look back at some pre–Industrial Revolution practices—horrendous working conditions, child labour—and think, That was crazy. With AI, people want to talk about the doomsday stuff, but 10 years from now, this will look similar to the rollout of the personal computer.

Okay, I wasn’t alive during the Industrial Revolution, but what makes AI different is that, doomsday discourse–wise, the call is coming from inside the house. Depending on the day, Sam Altman, CEO of OpenAI, is either saying AI could create new diseases, launch widespread cyberattacks or bring about the end of the world. Has that kind of alarmist industry commentary made your job harder?
Absolutely. Much of the alarmism is perpetrated by people looking to raise lots of money. That’s been detrimental to the discourse around the technology I love.

Some of the scarier predictions originated with Geoffrey Hinton, the Dr. Frankenstein of AI and your former boss at Google Brain. According to him, machines may let us hang around for a while to keep the power running, but eventually, they could change their minds. Has publicly disagreeing with a mentor of yours weighed on you?
No, it’s quite nice to have an academic disagreement with someone I respect. Geoff’s a provocative guy. But a lot of it comes down to the fact that, whenever people write about the history of AI, he’ll be mentioned. It’s a good thing that one of its founders thinks about how it could impact the future. I, on the other hand, think LLMs have fundamental limitations. Auto-regressive sequence models are awesome, but I don’t think they’ll lead to artificial general intelligence—AI systems that are as smart as humans, or smarter. Geoff does. I sometimes jokingly say, “Well, he’s never built one.”

Do you think the “protégé” label has helped you or hurt you?
I find it humbling. I learned everything I know about research from Geoff. He’s willing to explore things for their own sake, which is a great way to do science. His whole motivation for focusing on neural networks was wanting to understand how the brain works—which is funny, because I don’t think neural nets work anything like brains.

Last August, Cohere signed an agreement with the federal government to help modernize the public sector—which, in Canada, is a hell of a job. How’s that going?
Any time a citizen interacts with various public services, they’re like, “Man, wouldn’t it be great if we could do this better?” Anybody behind a computer doing boring tasks should be doing them with an LLM. That is, unfortunately, a huge amount of government work.

AI sovereignty is a huge federal priority. Cohere now has satellite offices in San Francisco, Seoul, London and New York. You’re rolling out your models in workplaces all over the world and training the models on their data, which will ultimately boost their bottom lines. Are you running the risk of exporting most of the economic gains of our LLMs to other countries?
No. Cohere tailors its tech to companies’ specific needs and gives them autonomy—or sovereignty—over it. We can’t shut it off or see what data comes in and out. When you hear people talk about Canadian AI sovereignty, what they’re saying is: AI is a fundamental technology, and we need to have control over it within Canada the same way we need power plants within our borders. Evan Solomon said sovereignty isn’t isolation—that’s true. I don’t see a conflict between Canadians building and running this tech and empowering our allies. We’re not a country seeking to push anybody down.

On energy: we’re seeing tons of pushback against data centres, mostly because of the noise or environmental impacts. Synapse, a big project in Alberta, just stalled. We really can’t just build these things in space, eh?
That’s a truly brain-dead idea that exists solely to boost a bunch of stock in the public markets—to justify SpaceX buying xAI.

Anthropic was temporarily blacklisted by the U.S. Department of War over its refusal to give carte blanche to use its tech for mass domestic surveillance and in fully autonomous weaponry. Cohere just signed agreements with TKMS and Hanwha, all international defence companies, to integrate its tech into submarines for Canadian military operations. Do you have any hard red lines when it comes to how your creations are used in defence?
I can’t speak specifically to America, but Canada has been thinking about this for a long time. One thing the government decided is that no autonomous decision should be made by a machine in a lethal environment. Cohere follows the same thinking. Ultimately, this is administrative software. It’s useful for handling documents and processing efficiencies. Some of the things others are trying to do with AI aren’t only unethical—it’s also stuff the tech is not particularly good at.

You once said it’s “not great” that most of the world’s experience of technology is defined by Silicon Valley. Why is it such a bad influence?
You probably don’t love tech right now. I know a lot of people who have an openly hostile relationship with their phone.

I’m recording our interview on two separate devices, if that gives you any indication of how much I trust mine.
There are moments of “It helps me do the things I want to do, better.” But there’s also a lot of phone addiction, and feeling like, “It makes me sadder and less good at the things I want to do.” That is a result of the culture and economic incentives created by Silicon Valley.

Can you expand on that?
Social media is a great example. There was a brief period where people thought it was going to connect us more deeply to each other. It demonstrably doesn’t, and that is because the social media companies figured out it’s easier to earn money by making people sad than it is by empowering them. Silicon Valley is only optimizing for next quarter. I’m excited to build LLMs while taking a different view.

You’ve spent some time in California. Why didn’t building Cohere there appeal to you?
All of our co-founders grew up in Canada—our community’s here. Silicon Valley is a monoculture. There are brilliant, well-meaning people there, but there’s also a lot of bullshit. A bunch of them are still talking about crypto and the metaverse. It’s been nice to build a company in a place that has brilliant, motivated people, but that’s a little less high on its own supply.

Tech founders are often regarded as semi-mythical figures, sometimes by their own doing. You and Cohere’s other co-founders, Ivan Zhang and Aidan Gomez, have seemingly sidestepped any god complexes. Is the antidote just… being Canadian?
We’re too aware of our own failings to develop god complexes.

True North Strong Free

Get Canada’s best magazine delivered to your door

SUBSCRIBE NOW

One fact that’s now baked into your founder lore is that you celebrate closing every funding round with a trip to McDonald’s. Cohere is worth US$7 billion. Has your order changed over time?
I always get two Junior Chickens.

I think you can spring for a full McChicken.
It’s too much food!

In addition to running Cohere, you front Good Kid, a band that sounds like the love child of Panic! at the Disco and the Strokes. This isn’t a small-potatoes, industry-battle-of-the-bands thing—you’re touring North America this summer. How do you juggle both?
With the help of a lot of other people and by compartmentalizing very aggressively? Our tours are vaguely monastic in their routine: every day, I wake up, I work in a room at the back of the bus, I play a show, I go to sleep, then I wake up in a different city. I like that, actually.

Do you ever play Good Kid at Cohere’s staff parties?
No. But Bethany on our events team is a fantastic singer. I’ve played her music sometimes. I can get away with that.

Your chief AI officer, Joelle Pineau, who came to Cohere from Meta, has said that your staff-party playlists are better than theirs. How bad are Meta’s, do you think?
Probably pretty bad.

Do you write lyrics? Ever gotten an assist from AI?
I do and I haven’t, because I’m not trying to optimize self-expression or write faster—I’m trying to do it slower. It’s challenging. That’s the point.

How do you avoid talking shop when some Good Kids are ex-Google, ex-Uber and ex-Amazon?
We’re all programmers, and we all like music—those actually go together. When you’re programming, you’re creating. Carpentry is a technical skill requiring math and a set of rules, but you’d never say, “There’s nothing artistic about this.” Programming is similar.

What’s your earliest memory of doing it?
My uncle showed me a Bash “for loop”—a repetitive programming script—when I was 14. It printed “Hello, world!” indefinitely. Even that was exciting.

I’m now encountering a lot of people using LLMs for therapeutic chats and relationship advice. Have you?
No, I use AI for research and for automating my job when I can. I don’t use it a ton in interpersonal situations—though, I guess I just used it to create a wedding invite. I’m engaged.

Congrats.
We’d already made the plans, but we didn’t want to write out the details. So we had an LLM do it for us.

This story appears in the May 2026 issue of Maclean’s. You can subscribe to the magazine here or send a gift subscription here.