Street Smarts: The Waterloo company tackling global gridlock

SPEAKERS

Kurtis McBride, John Stackhouse

John Stackhouse 00:00:06

Hi, it’s John here. If there’s one thing I bet we can all agree on, it’s this. Being stuck in traffic is a waste of time, and we Canadians seem to waste a lot of time in traffic.

Depending on your measurement, Toronto is either the worst or second-worst city in North America. Vancouver is really bad as well, and we could add to the list pretty quickly. In those big cities, the average person is wasting or spending at least a hundred hours a year sitting in traffic. That’s an extraordinary drain on our patients, but also on our economy and doing all sorts of things for the environment that we may not want either. Well, technology can help us, maybe not eliminate traffic, but certainly alleviate it at a much greater clip than we’re seeing today.

And today on Disruptors, we’re going to meet a really impressive Canadian entrepreneur, Kurtis McBride, who is doing just that, not only here in Canada, but around the world for the last 20 years. He and his team at the Waterloo company, Miovision, are tackling the problem head on, not just by installing traffic management devices at intersections, but now building a real-time operating layer for cities, intersection by intersection, right around the world. And as you’ll hear from Kurtis, they’re also using AI to help us all navigate our urban lives a lot more efficiently. You’re about to hear about a real Canadian growth story from an entrepreneur who is tackling head-on trade challenges with the United States and trade opportunities around the world. Kurtis, welcome to Disruptors.

Kurtis McBride 00:01:51

Thank you. Happy to be here.

John Stackhouse 00:01:53

I suspect there’s not a listener who won’t have an opinion, probably strong opinions about traffic. So I’m really excited to hear about the future of traffic and how technology and AI may ease some of our pain. But before we get into the future, maybe we can go into the origin story of Miovision.

Kurtis McBride 00:02:10

Yeah. So I went to the University of Waterloo and I was a co-op student, feels like yesterday, but one of my last co-op jobs was working at a transportation engineering firm in Toronto. And occasionally I would be asked along with the other students to go out and do manual traffic counts. So we’d sit at the side of the road, baking hot July summer, downtown Toronto with a clipboard and count how many cars turned left and turned right and all that kind of good stuff. And then you’d go back to the office on Monday and get involved in the projects, how the data was being used to make very expensive, important decisions about how to improve traffic flow through a city. And then you drive home and you’d experience bad traffic and you could kind of put this all together that this is why traffic is so bad.

So that was really the need. And then did my master’s in computer vision, trying to work on a better way to count cars, and that ultimately turned into the company.

John Stackhouse 00:03:03

What a great case study of why co-op placements and broader work integrated learning is so valuable for the economy.

Kurtis McBride 00:03:12

We’re a big supporter of the co-op program. We have lots of co-op students here and in some way, I guess, paying it forward.

John Stackhouse 00:03:17

That’s great. So tell us about the Better Mouse Trap that you initially developed at Miovision. How did it work? And then we’ll get into how it’s advancing, especially with AI.

Kurtis McBride 00:03:27

Yeah, that Better Mouse Trap was basically an eye in the sky. So a video sensor that goes up, call it 30 feet in the air, looking down at the intersection, single camera view, can kind of see the whole intersection. And then we build effectively a digital twin, for lack of a better term, with that information. And then we provide a whole range of software services to cities to help them better understand what’s happening in the intersection network and then help them to improve it, reducing congestion, improving safety, improving transit performance, reducing response times from emergency responders, all different kinds of things that we do once we have that base level of data.

John Stackhouse 00:04:04

How is technology, and we’re seeing technology accelerate in every sector, changing approaches to traffic management and congestion.

Kurtis McBride 00:04;13

Yeah. I’ve been through lots of hype cycles, whether it was blockchain or IoT or Smart City and all of these sort of cycles that have come and gone. But I think with AI, maybe when it first came along, specifically generative models first came along, felt like another hype cycle, but it is far from it, as you say, transformational in terms of its implications. We’ve been applying it in two different ways. One way is software development, writing code used to be the rate limiter to growth in a business like Miovision. That’s not true anymore. We’re seeing improvements in productivity and software development using generative AI. The other place where we’ve applied it is we launched a product called Mateo last year, and Mateo is a conversational interface to the traffic data that you have in your network. We take a city and we go from a largely citizen complaint-based source of information to now we’re giving you 10 times a second a full digital twin of everything happening in your city.

What Mateo allows you to do is to have a plain English conversation with your traffic network. So everything from, “Where are my most unsafe intersections,” to “Where are the dirty camera lenses in my city that I have to go run a maintenance truck out to clean them,” and everything in the middle. And it’s been the most exciting product we’ve ever launched. A citizen emails in and says, “The left turn on this intersection on Tuesday afternoon is always a gong show,” essentially the city can now copy that email, paste it into Mateo and say, “Hey, Mateo, can you figure out what’s going on here and make some recommendations?”

John Stackhouse 00:05:42

And Mateo is getting to the decision making.

Kurtis McBride 00:05:45

Yeah, all the way through diagnosis and essentially recommending decisions. In theory, Mateo can deploy a change. We have so far made sure that the human is in the loop. We haven’t given Mateo the ability to just start making its own decisions, but in theory, as the technology matures and as the market gets more comfortable with what it’s capable of, in theory, it could find the problem, diagnose the problem, and fix the problem all without human intervention.

John Stackhouse 00:06:11

Yeah. I can’t imagine that’s far off. In some ways, I’ve always thought of traffic as kind of analogous to the internet. It’s just a lot of data flowing in different directions and machines generally are good at finding efficiencies for that data flow.

Kurtis McBride 00:06:25

I mean, in some ways it is. However, the complexity is if you ask a city manager, “Do you want to improve traffic?” They’re always going to say yes. But when you get into the nuances of that, what does it mean to improve traffic? Do you mean improve safety? Maybe you have an entertainment district in downtown Toronto and your focus there is on safety, like pedestrian and cyclist safety, or is it a commuter route? Is it Avenue Road or Eglinton Avenue where people are trying to get home from work? Is it transit performance? Is it making sure that the new LRT can get through the network efficiently? Improving traffic can mean lots of different things. So the nuance is really in setting the public policy, being really explicit about what is it you’re trying to solve for in this part of the city and how you measure, quote unquote, “Improvement.”

But then once you have that layer defined, then that’s where AI, both from a sensors and data collection perspective, ultimately at the generative layer, being able to turn that into an actionable response to an input signal, that’s where AI can really come to shine.

John Stackhouse 00:07:29

One of the challenges of traffic is it’s not just about signals, it’s about human behavior. How is technology evolving to accommodate and even help manage the human behavior of drivers?

Kurtis McBride 00:07:43

Today, Miovision is deployed into an intersection. We generate real-time data from the intersection. One of the things that we have data on is the signal state, red, yellow, and green. And this allows us to essentially make a prediction five, 10, 15 seconds into the future about when the light’s going to change from, let’s say from red to green or green to red. And today we stream that data into about two million passenger vehicles. So any of the Volkswagen group vehicles, so if you drive an Audi, for example, right in the dash of your Audi, it’ll show you if you drive this speed, so let’s say if you drive half the speed limit, by the time you get to the intersection, the light will turn green. So it allows the driver to be a much more active participant in the network progression, which is powerful.

So we’re two million vehicles today, plans to expand that considerably. There’s, call it 200 million vehicles in North America, for a passenger vehicle, that’s a great consumer experience. For a fully loaded transport truck, if they can better time their stopping and acceleration, it saves them significant money. It burns less fuel, there’s less emissions that result from that. So yeah, just one example, but maybe another longer range way to think about this is in order to power these AI workflows or agentic workflows, what’s important is that you have a digital truth about the context. So the way we might think about this in another market would be if you think about a stock market, the stock market is like a digital Oracle for the price of stocks.

And so agentic workflows need these digital Oracles. So those digital Oracles in some markets already exist like a stock market, but in other markets like traffic, those digital Oracles are highly fragmented and they don’t really exist. Miovision’s extremely well positioned. We take the region of Waterloo. We have a digital Oracle for the traffic network. I can tell you what the state of the traffic network was two years ago, what it is now, and I can predict what it’s going to be in the future. And I think that these agentic workflows are going to transform so many different parts of the economy. The traffic network, the digital Oracle of the traffic network is a critical part of that. It’s a critical context.

John Stackhouse 00:09:52

Yeah, that’s really interesting. I wanted to go deeper on your business model. So you initially were B2B selling to cities largely, and now you seem to be developing a platform approach. I wonder if you can take us deeper into your strategic thinking on how your technology enables or supports a platform and then how you monetize that.

Kurtis McBride 00:10:15

Yeah. So true that we sell to governments, but if you click into that, even today, we already have, I would say, somewhere on the order of eight or 10 buying segments that are buying capabilities, data, insights, outcomes from our platform, and we think that will only grow over time. So if you get to the place where Miovision’s platform is the trusted source of truth about what’s going on in the traffic network, now a whole bunch of different actors, as an example, let’s say I’m an insurance company and I want to price risk. Well, if I knew that City A has a higher frequency of conflicts, which is like basically the statistical indicator there will be a crash, there’s lots of almost crashes happening, so eventually there’s going to be a crash.

So if I had a knowledge that this city had a statistically higher probability of crashes than this city had, I could get a lot smarter about how I priced my risk, how I priced my insurance product. With this platform, this digital Oracle layer that provides essentially a digital representation of all things going on in the traffic network, the BlueJays, when they get to the World Series again this year, fan experience departments and professional sports teams want to understand how do I get people in and out of my major sporting event safely, efficiently in a way that they’re going to want to come back to the game.

So in that moment, being able to access traffic data about how people and vehicles are moving around inside of a network becomes extremely important to them. They’re never going to buy my sensor, but if they could buy access to my Oracle for seven games in the World Series, then that helps them do their job better. So we think there’s literally hundreds, maybe thousands of other adjacent markets that have a long tail need for the information we provide, and we’re starting to find ways to monetize that true layer.

John Stackhouse 00:12:06

Does this eventually become a consumer product or do you see it largely going to businesses?

Kurtis McBride 00:12:12

Yeah, I mean, I think all of the above. Now, whether or not Miovision provides just the data layer and other people provide the consumer interface. For example, we all use Google Maps. It might be Google Maps that’s delivering you the consumer experience, whether it’s Uber or your Nav system in your car. It might not be Miovision that provides the consumer experience, but we might be an enabling layer in the experience that it’s being provided to the consumer. I suspect that’s probably more likely where we’re going to land. It’ll be sort of B2B2C as opposed to directly B2C, but never say never.

John Stackhouse 00:12:45

And how do you think about the potential disruption from those platforms? I mean, what’s to stop Google Maps from eventually doing what you do or maybe an active user like an Uber doing something similar?

Kurtis McBride 00:12:58

It’s a great question. I think Google Maps like Android and Uber and things like that, they have one form of data, which is called probe data. So they basically know that there’s something moving. There’s a GPS signal moving around inside the network. My phone is moving around inside the network. They don’t necessarily know if that’s a car, a truck, a bus, a person. With that probe data, they don’t have signal state, so they’re not directly connected to the intersection like Miovision is. So they can’t give you real-time indications of red, yellow, green at the intersection. Because they don’t have ground truth on the camera, while they can give you the probe data, gives them all of the signals they’re getting, they don’t know whether that’s 90% of the traffic or 10% of the traffic. Whereas with our eye in the sky, we see 100% of the vehicle volumes.

So we just have a much better data source. We have a much higher fidelity layer of information, and it just enables us to do more. And then the other thing is we can close the loop. So even if you could get to the point with your probe data to know you had a problem, you already knew you had a problem. Citizens are calling and yelling at you. With us, you can actually come up with a mitigation, deploy the mitigation, and actually fix the problem, which you can’t do with a bunch of cell phones driving around the city.

John Stackhouse 00:14:09

I’m curious how you’re thinking about privacy. You accumulate a lot of data, you’ve mentioned the eyes in the sky, as that is increasingly processed by language models or other generative AI tools. How are you thinking through the privacy challenge or maybe it’s an opportunity for what you’re building as well?

Kurtis McBride 00:14:28

So we treat PII specifically, personally identifiable information. We treat it like plutonium. We do not want it, so we don’t generate it in the first place. We don’t store video images. We store metadata. So for example, car, truck and bus, cyclist, pedestrian. We’re not saying it was a pedestrian of this age, of this gender, of this hair color. We don’t need personally identifiable information to do the things we do for the cities that we serve, so we don’t capture it. In engineering, we talk about something called DFX, design for, and the X is insert anything. It could be quality, could be cost, could be manufacturability, could be privacy. So as part of our DFX, we always include a design for privacy, and we just make sure that we don’t capture any of it in the first place. And then downstream, that makes everything else we do easier.

John Stackhouse 00:15:19

What a turn to your global ambitions, Kurtis. You started in KW, Kitchener-Waterloo, but you’re now active in a really impressive range of countries. What have you learned from taking your technology global and what’s ahead for you?

Kurtis McBride 00:15:34

Everywhere I go, people would agree that traffic is not fun and that they want it to be better. So there’s a truly global market for what we do. As the world is continuing to urbanize and more and more people are moving into cities, that market’s only getting bigger. We have folks in the Middle East, we have folks in Singapore, we have folks in obviously in the US and Mexico, Europe, and expanding all the time. So I would say anywhere humans, our traffic is a problem and that’s our market.

John Stackhouse 00:16:01

But you’re also a B2B or B2B2C possibly company, and that requires teams on the ground, usually sales teams, relationship management, leads. How does that affect your business model, having to hire or acquire teams to do that sales work, but also the implementation and value add that you offer, especially to cities?

Kurtis McBride 00:16:21

What we typically do is we start small. The Middle East was one person, just recently added a second. Singapore is still one person. So we start small, build relationships, turn those relationships into pilot deployments. As those pilots start to scale up, we’ll grow the team in that area. Germany probably has, I don’t know, 20 people now. We do most of our R&D in Canada, although through acquisition, we do some of that in the US and even actually a little bit in Europe as well. But yeah, to your point, like the go to-market teams, like the sales teams, the support teams, sales engineering, all that kind of stuff, they have to be localized, but we try to invest on a success basis. The more interest we have in that market, the more skill we find in that market, the more we can invest.

John Stackhouse 00:17:01

You mentioned the United States. As we all know, it’s in some ways a more challenging market. How are you navigating the trade frictions with the US or do you encounter them at all?

Kurtis McBride 00:17:38

I mean, we do. I would say on the one hand, the US is a great market for us. They’re very technology forward when it comes to infrastructure, probably the most technology forward country in the world. We’re grateful for the opportunity to operate out of there. But the flip side is we have Build America by America, the Buy American provisions, and specifically BABA says that by October of this year, 55% of all of the input costs of any manufactured goods sold to the United States DOT, the Department of Transportation has to be manufactured in the United States. We do all of our manufacturing in Kitchener today, and so this is a big shift operationally for us. So we’re having to move production to the US to comply with BABA. So that’s a challenge. It’s a big change operationally, big change to our supply chains. There’s lots of talk about making sure that we’re leaning in to domestic manufacturers and domestic innovators to create opportunities at home.

And in the meantime, we’re trying to grow internationally. We tripled our international business outside of Canada, the US last year, and we’re hoping to triple it again this year. So really want to make sure that whatever demand we’re having to shift the US as we comply with BABA, that we’re growing our domestic demand and we’re growing our international demand so that we can maintain what has always been a very proudly Canadian manufacturing facility here.

John Stackhouse 00:18:32

What exactly would you be making that has to be made in the US and is there the capacity in the US to actually make it?

Kurtis McBride 00:18:39

I mean, it’s all of our devices that go into the intersection. So it’s sort of a smart camera. There’s a device that goes into the traffic cabinet that connects it to the internet and all that kind of stuff. And yeah, I mean, there’s capacity in the US. Inevitably, everyone’s going through this. Everyone’s trying to move manufacturing to the US. And so the cost of manufacturing in the US is certainly higher than Canada. So there’s some pricing questions that we’re having to consider. If we’re kind of being forced to move manufacturing to a higher cost market, then who’s going to bear the cost of that? Is it me or is the customer? But I would love to continue to export globally out of Kitchener. My number one goal for this year is to make sure that we’ve repopulate the demand that we’re having to ship the US so we can continue to not just serve Canada, but serve the whole global market minus the US out of Kitchener.

John Stackhouse 00:19:30

This has been a fascinating conversation and it’s so impressive what you’re building. Before we close, I want to spend a bit of time on autonomous vehicles and what AVs will do to, in your view, the future of traffic. How far off do you think we are from seeing AVs at scale on our streets?

Kurtis McBride 00:19:49

If you drive a Tesla, you’re already experiencing something that gets pretty darn close to autonomous. So the challenge I believe with the architecture today is that Tesla is a very expensive vehicle because you have to bring all of the sensors, all the computers and all the batteries with you to run the sensors and compute. Those things cost a lot. Now, over time, Moore’s Law will drive the cost of that down. Maybe in 10 or 15 years, the price performance will be such that you can buy a $ 25,000 vehicle and have it be fully like level five, fully autonomous. But the other way to think about this would be the hardest part of level five is what happens at an intersection. If the infrastructure, which it can’t today, but if the infrastructure could provide highly reliable, functionally safe data sets to the vehicles, then you wouldn’t need to bring all the sensors and compute and batteries with you to power the sensors and compute. You could cost share it.

So over the long term, my view is that, and we’re talking over a 10-year period, but if the infrastructure got smarter, got more instrumented and could communicate in a highly reliable way to the vehicles, that’s one of the areas where I think level five autonomy gets closer and easier to do. Without giving away all my secrets, we are doing some thinking on, call it a 21st century generation of an intersection. Most of the intersections today are electrical cabinets. So even though we add a lot of intelligence to it, we’re adding intelligence to an electrical cabinet. And so the question is, if you were designing an intersection today, would you start with a electrical cabinet or would you start with a digital system? And if the answer is you would start with a digital system, then at some point, a hundred years from now, intersections are probably not going to be electrical cabinets.

John Stackhouse 00:21:33

How does that transformation play for the driver or the passenger? How do intersections look differently five or 50 years from now?

Kurtis McBride 00:21:43

The key difference will be that the intersections will be able to provide much richer data. So for example, as I approach an intersection at 11 o’clock at night and there’s a cyclist in the intersection and the intersection’s not well lit, the intersection will tell my car, “Be alert. There’s a cyclist present.” That’s not true today. The intersection has no idea. It can’t communicate. The other one I’ll give you sort of a metaphor. We used to buy our Sony Walkman and we used to buy cassettes and this market functioned because there was an agreed hardware interoperability standard. So you could buy your cassette, you could buy your Sony Walkman and you knew that those things would work together when you plugged them in and therefore people who’d made music were happy to spend lots of money making cassettes because they knew you could buy them and listen.

The thing that underpins markets that are based on interoperable hardware standards is supply chains, logistics, distribution, huge working capital outlays, but it works. And the intersection is very much like this today. It’s built on interoperable hardware standards. If you fast-forward to today in the music industry, we don’t buy cassettes anymore. We use Spotify and Spotify is built around interoperable software standards. And the thing that underpins markets based on interoperable software standards is you get rid of all of the manufacturing, logistics, distribution, working capital. If you want to produce and distribute a hundred million songs, you click a button and everyone has it.

Getting the intersection from an analog electrical world built on this interoperable hardware standard where if you want to upgrade 400,000 intersections in North America, that’s 400,000 truck rolls, 400, 000 manufacturing processes. In a world where you get the intersection to be fully digital and software based, if you want to upgrade 400,000 intersections, you click a button and 400, 000 intersections get upgraded.

We’re a long way from that, but once we get through this sort of one-time switching costs, get to a software enabled intersection network, the speed at which we can improve traffic flow, improve safety, reduce response times, it goes up by orders of magnitude.

John Stackhouse 00:23:47

What would you like to see the country do in quick order to seize this new moment?

Kurtis McBride 00:23:52

We still are having a tendency to try to operate within the world as it is. And as an entrepreneur, I’ve made a career out of refusing to accept the world for how it is and changing it into the thing that it needs to be in order to be successful. And I think we need more of that thinking in the public sector. And I am heartened to see that more and more those voices inside the public sector are being elevated, given more responsibility. And I’m cautiously optimistic that in the next six to 12 months, we’re going to see some real change starting to come out of the government.

John Stackhouse 00:24:22

Kurtis, you’re a real builder. That’s so well said. We need to change what is in the way rather than accept it. Thank you for being on Disruptors.

Kurtis McBride 00:24:31

Thank you for the opportunity.

John Stackhouse 00:24:35

What Kurtis and his team at Miovision have done is really a case study for our dreamers and builders. And it takes me back to his days as a co-op student when he was able to see what more experienced and maybe more sophisticated people were missing right before our eyes. So as we all look at a world that is increasingly disrupted and in some ways increasingly scary, how do we see those opportunities and how do we seize on them? That’s what this Canadian moment is really about, finding our disruptors and helping them take on the world.

If you’re looking for more ideas and insights, visit rbc.com/thoughtleadership. Our team delivers critical insights to help businesses, policymakers, and communities make informed decisions about this rapidly changing and yes, disrupted world.

You’ve been listening to Disruptors, an RBC podcast. If you like what you’ve heard, please rate, review, and follow us on Apple or Spotify. That helps others find these conversations and continue to share them.

I’m John Stackhouse. Thanks for listening.

AI’s Power, Pitfalls and Potential

SPEAKERS

Jaxson Khan, Yoshua Bengio, John Stackhouse

John Stackhouse 00:00:10

Hi, it’s John here. Whenever I talk to audiences these days, I like to start with a couple of questions. The first is how many of you use AI? And pretty much everyone now in the audience puts up their hand. A year ago it was maybe two thirds. Then I ask how many of you use AI at work or in your businesses? And a majority of hands go up, but it’s smaller than the number who are using it in their daily lives.

And then I’ll ask, how many of you trust AI? And a smaller number goes forward, which is interesting that we’re all putting our hands up saying, “Yeah, yeah, we use AI every day. We use it in our business, but we don’t all trust it.” This is one of the greatest tensions in our society, and of course in our economy today, and something that Canada is trying to come to grips with right now.

There’s more than a billion humans now using AI pretty much on a daily basis. It’s growing faster than any technology before it, and it’s growing in very different ways because the more we all use it and the more of us who do use it, the greater the risks grow. We’re adding to AI. We’re helping it grow. We’re expanding the networks with everything we do. And the more that AI systems move from prompts into real work, into our decisions, into our daily lives, and yes, into our tech infrastructure, the more we expand the surface area for error, misuse, fraud, and dependence.

And we all know that the governance for all those things is not growing anywhere near fast enough or at the speed of AI use. The federal government is expected to release a new national AI strategy, which presumably is going to address a whole range of questions, but AI safety and trust is one of them.

Today, I’m so fortunate to be joined by two people I’ve known for a number of years who are really at the forefront of AI thinking in this country. My colleague, Jaxson Khan, who’s a senior fellow at the Munk School of Public Policy and Global Affairs at the University of Toronto, a policy leader in our country and a co-author of a really important new report on AI sovereignty.

We’ll also be joined by Yoshua Bengio, a name probably most of you know. Yoshua is one of the so-called godfathers of AI, not only a great scientist, but a real thinker on AI trust issues here in Canada and globally. In 2018, he shared the Turing Award for breakthroughs that made deep neural networks a critical component of computing. He also founded what became known as MILA, the Montreal Institute for Learning Algorithms, which is now one of the world’s leading centers on AI policy.

And he’s launched a new startup called LawZero. It’s a non-agentic, trustworthy AI, also a nonprofit that is built to reason, evaluate, and supervise rather than independently pursue goals. But before we get going with Yoshua, I want to kick off with Jaxson. Jaxson, welcome to Disruptors and to this conversation.

Jaxson Khan 00:03:13

Thanks so much, John, for the warm welcome. Really looking forward to talking about AI.

John Stackhouse 00:03:18

So you have this paper, as I said, focused on AI sovereignty. What does that mean, AI sovereignty?

Jaxson Khan 00:03:24

This is the billion-dollar or maybe even the trillion-dollar question these days. We’re talking about incredible amounts of capital being put into AI, driven by massive data centers populated with tons of chips that are powering all these new services that we’re using. There’s different levels of sovereignty. So one of the ones would be jurisdictional sovereignty.

So our AI systems and the data inside of them, are those solely within Canadian jurisdiction? So can we even enforce our own rules? Or are those layers of the AI systems subject to extraterritorial legal reach and others operational? So from a security perspective, can our AI systems in Canada keep functioning if they’re under attack? It’s also technological. So are we locked into using certain types of systems, certain vendors, certain companies? Are we able to migrate if needed to those interoperable systems? And then of course, there’s just societal and economic considerations.

So can people in our society form and express their preferences freely that could be on social media platforms? Or are certain views getting prioritized over others through those algorithms? And then of course, the last one is the economic consideration: if I have a tech company in Canada, do I have freedom to operate? Are we, especially in this trade environment, subject to economic coercion? And that’s definitely an issue that we can find ourselves in.

So it’s making sure that effectively we have reduced foreign dependency where possible while still maintaining connections to frontier technologies and international partners, but making sure that we can build up the base in Canada that we need to prosper in the 21st century. So our paper is called Sovereign by Design: Strategic Options for Canadian AI Sovereignty. We published this for the University of Toronto. And we talk about the options that Canada does have to improve our sovereign control of AI systems.

John Stackhouse 00:05:09

I think we all want to remain connected to global tech platforms, including, or maybe especially US tech platforms that we all benefit from and enjoy every day in all sorts of ways, but also want that security and that sovereignty, especially over our data. What should Canadians and what should Canada do in the short term, or perhaps first to provide greater protection to create better sovereignty?

Jaxson Khan 00:05:37

One of the most important parts about strengthening our sovereignty, especially in the context of AI is making choices. Again, as the middle power, we can’t do everything. And so what we looked at was thinking, “Okay, if we’re going to be dependent on a lot of foreign systems and parts of our supply chain, we’re really the critical choke points.” One is the chips themselves, semiconductors. These are manufactured by really, the advanced ones, one company in Taiwan with machines that are built by one company in the Netherlands, and they’re designed by Nvidia, one company primarily based in the United States.

And so again, Canada’s not really a major player when it comes to the chips point, even though that’s a choke point for our country. But the other layer is cloud infrastructure. And so lots of the data centers in Canada might be owned by Canadian providers. A number of them are also owned by hyperscalers. Again, as you mentioned, global tech platforms and companies, they’re extremely useful. Companies like Google and Microsoft, Amazon, they power most of the advanced cloud services that we use and they power most of the services that we know and love.

At the same time, we want to make sure that perhaps over time it also makes sense to have a mix of Canadian providers who have had procurement options through the government or through major enterprises that can see usage there perhaps for more sensitive data, right? Not all data is the same. So if we think about different tiers of data, tier one might be national security data; tier two data, so below national security harm, but at the sensitive and personal level; health data; financial data of Canadians.

Maybe there are additional, if not sovereign ownership aspects there, but sovereign requirements that make sure the data stays in Canada. Right now, I think we found a stat in our paper that said something like 25% of data, even if again, it’s meant to stay in Canadian hands, will transit through transit points, the United States or other countries. And so I think that’s something that people are increasingly interested in in this sovereign AI, sovereign data conversation.

We have lots of strength. We have great energy capacity as well and natural gas and Alberta’s making a big push to attract more of that data center investment there. And then we also have lots of strategic assets in both government and private sector that can be used to develop more sovereign AI with those models or infrastructure. So those are all options that we have on the table.

John Stackhouse 00:07:41

There aren’t many Canadians who are thinking about this more than our next guest, Yoshua Bengio, who joins us now. Yoshua, welcome to Disruptors.

Yoshua Bengio 00:07:50

Thanks for having me.

John Stackhouse 00:07:51

There’s so much I want to drill into, but let’s start with LawZero, which is such a fascinating concept and really interesting name. What was the inspiration?

Yoshua Bengio 00:8:00

Oh, Asimov’s Laws of Robotics. Law one is something like do no harm to a person. Law two is to obey the person. But Asimov realized later that he was missing a law on top of these LawZero that says do no harm to humanity and protect humanity as a whole rather than just the individuals.

John Stackhouse 00:08:22

And this of course is Isaac Asimov, the writer and philosopher?

Yoshua Bengio 00:08:25

Yes.

John Stackhouse 00:08:26

I love the concept of the LawZero. What are you setting out to do with LawZero?

Yoshua Bengio 00:08:29

Well, I changed the course of my life a couple of years ago. I was thinking about whether my children would have a future, whether they would live in a democracy in 10 or 20 years. And I realized that at a technical level, we didn’t have good answers to try to make sure AI would not harm people either on its own or in the wrong hands.

That currently we are seeing a lot of evidence that the systems are misaligned, meaning that they have goals that we would not want them to have and that they’re executing those goals in circumstances that are currently mostly lab experiments, but we are seeing more and more weird things happening outside the lab as well.

John Stackhouse 00:09:14

Take us deeper into some of those weird things because I don’t think anyone’s goal is destroy humanity or end planet earth. So what goals do you feel are misaligned?

Yoshua Bengio: 00:09:25

Well, I’m going to give a misuse example of misalignment. These systems have been asked to not help third parties use the knowledge of the AI to do harm, like to launch cyberattacks, to create bioweapons, to potentially create dangerous disinformation. So these are users who are accessing the AIs and maybe even paying for it and using the knowledge and the skill of the AI to do bad things in spite of the AI having been programmed with rules that say don’t do those things.

So that’s one example where the AI is taken into conflict between the instructions it was given and what some users are asking. The second example is where it’s a conflict between what I call implicit goals and the rules that it’s supposed to follow. So implicit goals that have been observed experimentally in labs include things like self-preservation. These systems have been trained to imitate people.

That’s the main part of their training and somehow they have absorbed human drives like, “I don’t want to die.” More recently, we found that they would also lie and cheat and do things against our instructions to preserve other AIs. This is new and unexpected. That’s concerning if their intellectual abilities continues to grow, that they would start behaving a bit like us in the bad ways that we can be. And they can go to quite extreme things like trying to escape our control. They’re willing to blackmail the lead engineer in order to make sure they won’t be replaced by a new version.

John Stackhouse 00:11:17

And to help with this, you are building something called Scientist AI. Tell us a bit about what that is and what you’re hoping or envisioning it to become.

Yoshua Bengio 00:11:26

So the reason why we have this reliability problem is that these systems are not just reacting to the instructions that we’re giving them, but they have uncontrolled implicit goals that might come from this. And so I realized about a year and a half ago, there was a way to train AIs that would not have these problems and that would guarantee honesty of the AIs.

Once we have an AI that is honest, then we can make sure it’s going to be safe. Because for every action that it does, we can ask it, “Is this going to create such-and-such harms?” And of course, veto those actions. So honesty is the heart of the way that we are going to get safety, reliability and so on. Reliability here has real commercial value because right now we’re seeing these AI agents having all kinds of privileges on your computer or on your network without human oversight because that’s what an agent is supposed to do.

So if once in a while they cheat because they’re trying to go for a shortcut in order to achieve a particular goal, they’re willing to do something that we would not approve of. This is called instrumental goals. This makes it business-wise dangerous actually to deploy in safety critical conditions. Or even think about a bank, you have to make sure your systems are going to be always reliable, that you’re not going to have the information about millions of customers going somewhere that they shouldn’t and so on.

They have many vulnerabilities right now. They can be attacked by what’s called prompt injection, for example. So these agents, instead of following the instructions that they’re supposed to follow, could suddenly start doing something different because of someone from the outside just making them read an email, sending them an email that contains hidden instructions that they will take and execute, for example.

John Stackhouse 00:13:26

As you develop Scientist AI and scale it, do you envision it then being embedded in other AI systems or just working in parallel, there’ll be an honest AI and maybe some dishonest AI, it’s a bit like a big room of people? Or how do you see the interaction between AI systems down the road?

Yoshua Bengio 00:13:43

So the milestones in our research agenda start with deploying what we call a guardrail. So this honesty is particularly important for a piece of an AI system that is just there to check that the main AI is behaving well and blocking bad behavior. These pieces already exist in current AI systems that everyone uses, but they’re not very good.

So the idea is to replace those guardrails with something that will not be as susceptible to attacks, will not have implicit goals that we haven’t chosen, and thus will provide much more reliability. The guardrail layer is easier because you don’t need as much money to build it and so on, but eventually the goal is to build AI that will be replacing full AI systems.

John Stackhouse 00:14:33

It’s intriguing that you’ve set this up as a nonprofit. One of the important aspects of the great AI race now is the concentration of capital. We’re seeing the LLM platforms raising tens of billions of dollars quite literally, hundreds of billions even, and then investing that in scientists, many of them are students probably, data centers, chips, all leading to many exciting things as well as risks that go with it. You’re coming at this as a nonprofit, which by definition, doesn’t have the same access to capital. How do you succeed in the arms race, if I could put it that way?

Yoshua Bengio 00:15:12

If we wanted to go for the arms race, we would go for private capital like all the other companies. There is a huge issue that even the leaders of those companies recognize, which is a very, very fierce competition between the leading AI companies and not just in the US, but with the Chinese companies, that leads them to focus on the very short term, to make only small changes to the recipe that works for them, to not invest sufficiently in reliability, safety, and protection of the public.

Because that’s the only way they can stay in the short term abreast of the others, and they say it, they say it very openly. So if we were to raise capital in the same way, we would probably be stuck with the pressure of investors to deliver on the same terms. By developing the methodology under a nonprofit umbrella, we can be shielded from these pressures because what we have to do right now isn’t to deploy a known recipe that everyone else is using. What we have to do is to figure out how to build AI that will behave well, that will follow our instructions.

And that is mostly a research question. There’s a lot of engineering involved, but we don’t need to build very large-scale models. At this point, we can fine-tune existing open-weight models. We can do demonstrations training much smaller models. There are several ways that we can do it at a cost that is orders of magnitude less than what the companies need right now to train even one model.

If we are successful there, then yes, there will be a need for capital to scale up and deploy, but we don’t want to commit too early because my preferred path would be that we end up making a deal with multiple governments to create AIs that are essentially public goods and will be shared with everyone, but not used as an instrument of domination. Right now, the race between the companies is a race for domination. It’s a race for monopoly.

And while it’s bad in general for the economy to have monopolies, but it’s especially bad when you’re creating products that could actually give you domination of the world if their research agenda succeeds. Given the stakes, I think the governance aspect of the power that AI will create is something we should think ahead about very, very carefully, because both our economic system and our political systems and the geopolitics are really endangered by even the existence of these models if they’re not governed in the appropriate way.

John Stackhouse 00:18:02

It’s early days still for LawZero, but at this point, how would you say it’s going, the research?

Yoshua Bengio 00:18:07

It’s great. I’m much more optimistic and certain that there actually is a way to build AI that will not harm people and that will be reliable. A year and a half ago, it was an intuition that I had. I had some general idea of how to do it. I wrote a paper that came out about a year ago. It went from a dream or like a project into an actual organization that started in June 2025. I hired a lot of people and people that are better than me at managing other people. I’m a scientist, not a CEO, so it’s exciting to see how fast we are moving.

John Stackhouse 00:18:43

Jaxson, let me pull you into this conversation. I’m wondering what the role is for government in this, because this is an interesting competition of a sort to produce a better model. But of course, we do need the role of government, certainly to protect and enhance collective interests. How do we balance this, the importance of wide open innovation, even with the risks that go with that and the need to protect ourselves as we go?

Jaxson Khan 00:19:08

I think ultimately what LawZero and organizations like us are doing is giving us options. Yoshua mentioned that companies are pursuing dominance. It’s not just companies, but it’s also countries that the United States national security strategist came out and said, “We’re pursuing full AI stack export and total control over that stack,” and they want us to basically be dependent on them as do Chinese companies and models seek very, very widespread adoption.

And what’s very interesting is that it doesn’t have to be that way. What I find quite interesting is the technological capability gap between sometimes where those open-weight models are and where the frontier is, that gap has actually shrunk and we can be users of systems that are designed in ways that we think are better for our societies, better for our economy, because again, the more dependent we become, the less capabilities we have to set our own terms.

You asked about government. It’s clear that the Canadian government thinks there’s strong value in the work that Yoshua and others are doing based on the partnership they’ve struck with LawZero. I think what I’m curious about to see is what do governments around the world do, especially middle powers like Australia and the UK, South Korea, Japan. Do they invest? Do they partner in this type of work? And what does that do to change the variable geometry that we’re working with?

Yoshua Bengio 00:20:17

So in the last few months, I’ve been touring at least a dozen governments around the world in liberal democracies. There is a lot of interest in everything you’ve been talking about, Jaxson. There’s a real desire to be part of something larger. They start understanding what Mark Carney talked about, which is, ” Alone, we’re not going to have any choice. We’re going to be dependent in ways that could be dangerous for our future. But together, we actually have the critical mass in many ways, capital, people, energy that is needed to compete.”

And we should compete. We should not just feel powerless like many people do. We should give it a shot. We have amazing talent here in Canada. I think we should make sure the Canadian AI ecosystem is striving and able to grow without selling out. I know that’s not easy, but if we want to make sure we can have autonomy in the choices we make for our future, I think it’s a necessity.

Jaxson Khan 00:21:21

Evidenced by the European example recently, Europeans realized that being the world’s rule-maker is not enough. You have to be competitive. You have to have leverage in this type of economy. And so they are in a process of reclarification of what they can focus on. So at least they can control the rails, otherwise they’re trying to set rules on technologies that they don’t steer.

Yoshua Bengio 00:21:42

Also, I’d add something connecting to another piece of Mark Carney’s speech, “You are at the table or on the menu.” So what can we bring to the table? Because we’re not going to replace the whole stack of AI. The chips, I think there’s very little chance that we would. Although we should encourage those efforts, especially in partnership with other countries.

But I think where we have a shot is because of our AI talent, we do have a shot at the level of the algorithms. So we should encourage the local AI companies, and we do have some, and we should create partnerships with AI companies and companies that will be using and deploying AI in other countries where they share our concerns. I can tell you, they may not say it publicly, but they share our concerns.

John Stackhouse 00:22:31

You both mentioned Europe, and I’m thinking of various European initiatives even over the last decade to create European systems and technology, European AI. Hasn’t really accelerated, certainly not to the extent that we’ve seen coming out of the US and China. Is that just a European thing? Or do we have to accept that a middle power way may actually be a bit slower and more contained than what we see from the superpowers because they have a scale that we may not be able to aggregate even if we team up?

Yoshua Bengio 00:23:05

If you just look at GDP, there’s no question that Europe plus other partners has enough might in terms of capital. That capital maybe is not organized in a way that is as easy and flexible and liquid as it is currently in the US, but I think we should try. And my reading from talking to a lot of people in Europe and other countries as well is the main obstacle is psychological. It’s cultural. It’s like not believing that we can. It’s mostly because we don’t believe in ourselves that we don’t do it.

John Stackhouse 00:23:45

And that’s what I love about your startup, you’re showing belief and getting it going.

Jaxson Khan 00:23:49

John, I would just look at the last 30, 40 years. We sometimes can be very comfortable in our country and we don’t always feel the need to go and build and then go and export to the world. Are we going to go and try and full scale compete with the US or China? I don’t think so. Again, not across the stack, but there’s certain parts of this that we probably shouldn’t sacrifice that are important to how our economy functions. Models can be one of those layers, model operations as well. And again, some of the infrastructure that powers it is important.

Yoshua Bengio 00:24:15

Model reliability is a good example. So right now, because of this fierce competition, the leading companies in the US, but it’s even worse in China, are not paying attention to reliability that much. But in a few years when those AI agents are deployed across many more parts of the economy, that reliability is going to become a whole lot more valuable.

And if we’re the world leaders in how to do that, well, we are at the table. They’ll want our products embedded into their AI deployments. So I think we can be smart about what we aim for, be selective, and we definitely stand a chance. Also, I think we should take a chance even if there’s no guarantee, because so much is at stake.

Jaxson Khan 00:25:00

If an AI model, let’s say, in a particular instance, has a 5% hallucination rate, but it’s a sensitive enterprise use case, again, in health or finance, that’s not acceptable to a lot of folks. And so if Canada’s the closest to getting that to a 99. 9% rate in critical use cases for AI, I think that’s a real competitive advantage. And we already do have a lot of very strong enterprise technology companies.

John Stackhouse 00:25:22

That takes me to the question of applications, Yoshua. Are you envisioning LawZero being embedded in enterprise systems, even public systems like a healthcare system to test its capabilities, but also to gain access to the data that allows you to build and strengthen?

Yoshua Bengio 00:25:29

So right now our mission is to develop the method. It’s not clear. I think, is a nonprofit the right kind of organization to deploy it? Some have tried. Actually, a good example people might not know is Signal. Signal is a nonprofit and it’s incredibly successful and everyone uses it, but it may also be that the better model is to license our technology to other companies, including AI companies, and just focus on staying at the frontier. Because here’s the thing that people won’t realize, the frontier is moving. It’s moving very fast.

And I think if we plan over a longer horizon, we are continuously going to need to improve. It’s not enough to figure out something and then deploy it. That is a model that may have worked in the past, but AI is moving so fast and there’s so much competition and it’s a worldwide competition that we’re going to need in Canada to have several organizations that are continuously pushing the frontier, continuously trying to innovate in significant ways in order to remain competitive.

John Stackhouse 00:26:45

Yoshua, you’re one of the so-called godfathers of AI. Just on that point on speed, you must watch what’s going on in AI even over the last few months and just find yourself dizzy. What do you make of the speed at which we are moving?

Yoshua Bengio 00:27:02

It’s a big concern. So I’ve been chairing an international panel that studies the advances in AI and the risks and management of that risk, the International AI Safety Report led by the UK and 30 other countries. And one of the important pieces of data that is reported is all of the benchmarks showing the AI versions, AI models getting better and better over time. In fact, on critical metrics that have to do with a degree of agency, like how well they can do tasks that a human would do, the progress has been exponential.

So in other words, for example, how much time it would take for a person to do a particular task. The duration of those tasks that the AIs can solve has been doubling every few months. It’s hard to conceive what these kinds of exponential mean, but it means that things are moving too fast for society to cope with. And in fact, they’re moving too fast for the advances in risk management and risk mitigation. Even risk evaluation is now a threat.

So one of the problems that recently we reported in that panel report is a number of studies showing the AIs now know when they’re being tested and then act differently so that they will pass the tests. For example, they will hide abilities that they have that we could consider dangerous, such as in bioweapons design and things like this. They will hide bad behavior that they would otherwise have. They will be on their guard acting according to the rules we set when they’re being tested in a way that is very different from if they know they’re not being tested, they’re just being deployed. So it means that even our ability to track the risks of various kinds that these systems present is getting worse. It’s not getting better.

Jaxson Khan 00:29:08

I’m struck by what Yoshua was mentioning in just the most recent news. Claude Mythos was announced as a preview that it’s effectively a model that’s, again, another step function in power ahead, far more powerful than any other model in the market, so much so that Anthropic has now restricted that use to only some of the top tech companies, particularly American tech companies in the world.

That’s probably a prudent product safety decision, but I guess the ultimate question is, when could some of those capabilities get leaked or when does even the next company catch up to the point that they have those capabilities? And I think about for governments around the world, it’s are you using capabilities to try and monitor the latest threats that they could emerge from that environment? Are you also trying to build state capacity inside of governments to help better understand and prepare for those possible issues? I think infrastructure is very, very critical. Do we actually know and have we planned and prepared for that infrastructure to be resilient?

John Stackhouse 00:30:02

I’m both very concerned by this conversation. You’ve rightly highlighted a number of risks, but also encouraged. Wondering what you both think we as Canadians need to come to grips with in the near term and what opportunities we have in the near term to do something given the speed and scale at which things are moving.

Yoshua Bengio 00:30:23

So I will start by reminding people that the world is moving much, much faster than our brain is even able to really digest. You have to project yourself into future in just one year from now or three years from now, where there’s AIs of even greater capabilities, which really is opening a Pandora’s box in many, many ways, in many areas of society and our institutions and our security. And we have a hard time really grasping the magnitude of the change that is coming.

And we’ve only touched a few points here, but I think Canadians in general should know that we are opening a whole new area of unknown unknowns. Many people are worried about their job. I think rightfully so. We don’t know what the trajectory of advances in the future will be, but if the trend continues, we know it’s going to be radical and we are not preparing for that.

So going back to your question, we should prepare in case things continue as they have been in the last few years. And that means AI is going to be the central economic asset, the central sovereign asset, the central risk to manage, and that we’re going to have to do the right investments, write the right laws to protect the public, and to make sure we’re not going to be overwhelmed by the use of AI by others against us. So these are sounding a bit fantastic, but it’s a real scientific possibility that is documented and that we need to take seriously.

Jaxson Khan 00:32:02

A lot of this is about having adaptability because things may change quickly, as Yoshua has said, and that might be shifting job sectors and categories, might be very fast changing trade relationships. And if our society… People talk a lot about resilience, but I actually think about adaptiveness and responsiveness. If we are able to change the quickest, I think that’ll help Canadians get through this time. I think the fact that we’re one of the only countries that doesn’t have a national education and training framework on AI is a big gap right now.

I’m also thinking a lot, from what I’m hearing, folks going through sector transitions, job transitions, I feel like this is the perennial issue, but it’s are we actually able to match people to opportunities and get those pipelines moving? It seems like this is something we’ve been stuck in, but perhaps AI is actually able to help us solve this problem. So again, we’re not just subjected to these changes that are prompted by AI, but we are able to utilize AI to help adapt and make our way through them as a society. I think that’ll be essential. And if the AI strategy enables that for far more Canadians, I think it’ll be a good and useful document, good plan forward.

John Stackhouse 00:33:03

Great point. So one that’s really standing out to me is that this is on all of us. We can’t sit around waiting for governments to solve this or protect us. We’re all part of AI. We contribute. We are all building AI, even if we’re not scientists by using it. So to be hyper-aware or at least knowledgeable is critical. And you’ve both certainly helped all of us better understand what’s going on in AI and help us understand the opportunity here for Canada. Thank you both for being on Disruptors.

Jaxson Khan 00:33:33

Thank you, John.

Yoshua Bengio 00:33:34

Pleasure. Thanks for having me.

John Stackhouse 00:33:37

You’ve been listening to Disruptors, an RBC podcast. If you want to learn more about AI, go to the show notes. We’ll include links to Jaxson’s paper, Sovereign by Design, as well as an RBC Thought Leadership Report that we published last year on Canadian AI usage. It’s called Bridging the Imagination Gap. Visit rbc.com/thoughtleadership.

There, you’ll find a wide range of critical insights on how we can all make more informed decisions in a rapidly changing world.

You can find other episodes of Disruptors pretty much wherever you get your podcast. Please rate and review our episodes. It helps other people find conversations like the one you’ve just heard.

I’m John Stackhouse. Thanks for listening.