Evidence of meeting #9 for Canadian Heritage in the 45th Parliament, 1st session. (The original version is on Parliament’s site, as are the minutes.) The winning word was content.
A recording is available from Parliament.
Conservative
Kerry Diotte Conservative Edmonton Griesbach, AB
Thank you, Madam Chair.
These questions are for Dr. Kearney.
Doctor, as an Edmonton MP, I'm really pleased you chose to move to Edmonton personally and also to headquarter your company in the city. That's a big decision, obviously. Can you expand on why you decided to do that?
Co-Founder, Artificial Agency, As an Individual
As you know, Edmonton and the province have a long-standing history in innovation. I never had dial-up Internet. I look a bit young, but for my age that's actually surprisingly unique. Long before many places in the United States had high-speed Internet, we had a backbone from Fort McMurray to Calgary.
There's this attitude—and I don't think this is just Alberta but the country as a whole—of pioneering new technology. It's hard for many people to understand, because we're too humble. It's not a story we tell often, but when you think about many of the groundbreaking advancements that have happened in the past 20 years in artificial intelligence, many of the people on the teams that developed that technology either received their training at the University of Alberta or were from the University of Alberta when they were doing their work.
I think about the AlphaGo work at DeepMind where they beat Lee Sedol at Go, which was a groundbreaking technical challenge. Many people from the University of Alberta—trainees—were on that team at DeepMind.
It's that pioneering attitude that I want to bring back here and continue.
When my co-founder Brian and I first started in artificial intelligence, I'd been working in the field for 15 years. He's been working much longer than me. It would have been unthinkable that someone would be able to raise $22 million and start a company building AI technology in Canada, let alone in Alberta. We really strongly believe that to build the future of the country, we need to reinvest, we need to choose to stay here and we need to own it. It's a matter of national importance for us to be at the forefront of the technology that we have a fundamental responsibility for creating.
Conservative
Kerry Diotte Conservative Edmonton Griesbach, AB
You mentioned in passing the Alberta Machine Intelligence Institute.
Can you go into some detail on how that was helpful to you?
Co-Founder, Artificial Agency, As an Individual
Yes, absolutely.
Brian was one of the first AMII students, back before AMII was even a centre that was federally funded. That was back when the provincial government decided that AI research was a thing to really spearhead 25 years ago.
I started working when I was a high school student in research contributing to papers here. That's how I first learned that AI was a thing. I actually was a monetary policy wonk. I wanted to become an economist, but then I saw that we can train these models and they can learn through experience with the environment. You could spend your entire life building models in any discipline that you care to, but by using AI, we can develop systems that are able to continuously learn and adapt and refine themselves based on how the world around us changes. That's a fundamentally powerful thing that 15 years later is being brought to bear.
Basically, the institute there has been an incubator. I chose to leave because there was no place in Canada to do AI at an undergraduate level back when I was pursuing my studies, but, absolutely, there were three great graduate groups, and I chose to come back here because this is where the champions of reinforcement learning, the area that I have my specialty in, are predominantly based.
We have a whole university filled with hundreds of students who are training in this discipline, and those students will come onto the market soon.
Conservative
Kerry Diotte Conservative Edmonton Griesbach, AB
Edmonton is the city of champions, and you're one of them.
I have one final quick question.
What is the best thing the federal government could do to help AI, and what is the worst thing it could do to hurt AI?
Co-Founder, Artificial Agency, As an Individual
I would be very skeptical of someone who tells you that they can have a foolproof means of producing transparent models that give you explainability and auditability at no performance cost.
When you're looking into implementing measures, I would strongly consider introducing technical people into the room. When the previous AI legislation, the AIDA, was being considered, I was in a consulting session and someone said that we need to regulate the technology known as reinforcement learning. My Ph.D. adviser here at the University of Alberta is the grandfather of that field. It's a problem formulation, not a technology, so when you let people in the room who don't necessarily have a basis say this in an unchecked way, you lead yourself into a situation where you could be shooting yourself in the foot.
Again, this is a form of technology that is very nuanced; it's very niche. There are very few people with a great depth of expertise. When they say that this problem formulation needs to be regulated, that means they're saying all of the great work that has come out of the University of Alberta should not have happened. AlphaGo shouldn't have happened. All of the deep reinforcement learning work that has led to some of the fantastic advances that we see today should never have happened where it needs to be tightly controlled.
Moreover, regulate harm—that's the most important thing—not use.
Liberal
David Myles Liberal Fredericton—Oromocto, NB
Wow, this is amazing. Things just got very interesting. Thank you very much. I appreciate this conversation so much. Thank you to all of the witnesses who are here.
Regulating harm.... Everything that was just said was kind of what I was waiting to hear. I feel like what hasn't been said is that transparency would get in the way of this technology to be used even as a tool. I find it interesting because we've talked a lot about monetization. We've talked a lot about this idea of an incentive to create. I would argue that the incentive to create is tied to monetization, in recognizing that when one creates something, one would like to maintain some value in what he or she created.
I think the video game industry is a great example of something that has maintained a high monetary value. There's a reason that it's exported. There's a reason that it employs so many people in Canada. It still sells for $65 to $120 per game. We've watched recorded music and many other mediums follow suit in their demonetization over the last 20 years. Video games have not followed that same track to the same degree. Therefore, they have punched above their weight.
I guess what I'm curious to hear is the idea that no harm.... One would think that demonetizing or taking the creative works of generations of artists would be considered somewhat harmful.
Dr. Kearney, it's so great to have your perspective here because it does enlighten us. I think we have been hearing the importance of transparency. The importance of copyright. You are kind of saying that we wouldn't be here if we were worried about transparency and copyright.
Do you believe that artistic work should be allowed to be used and has been justifiably used and stolen up to this point without harm, or that that's the cost of progress?
Co-Founder, Artificial Agency, As an Individual
I think that's a really honest and fair question.
Again, I focus mostly on agentic AI. By that, I mean systems that are able to make decisions and take actions in their environment. That's very separate from generating text or generating video or audio, which I think is the area that you're predominantly—
Bloc
Martin Champoux Bloc Drummond, QC
Madam Chair, I have a point of order.
It's important to turn off the mikes in the room. Interpretation can't be done properly when they're on.
Liberal
Co-Founder, Artificial Agency, As an Individual
Technology is challenging.
Should I repeat myself, or should I continue where I was?
Co-Founder, Artificial Agency, As an Individual
Your question is fair.
I can't speak to all forms of generative AI. This is why I say, do not treat it as a monolith. I can only speak to the areas that I have expertise in and areas that I work in. That's predominantly agentic AI, systems where an agent is able to take actions based on some goal-oriented behaviour in an environment. That could be a character in a world. That's very different from generating text or generating music or generating images—things of that nature. That is not my expertise.
A comparison I would draw though is.... Let's consider learning systems. People say that these systems just copy. I don't think that's a fair characterization of transformer-based models if that's something that they're talking about. Certainly, you can get these models if you try really hard to produce outputs that look very similar to their inputs, and some people try to do those sorts of things.
I would compare it to something. I like photography. I enjoy computational photography, and when I'm writing programs to edit my photographs, I might have an idea in mind, but if you ask me what artistic influences are pushing to try different things, I can't really identify them. I would draw that comparison to some of the models that we're using and some of the behaviours. If you're using an open-source model, they might have been trained on a rich variety of inputs, but when they are choosing a behaviour in an environment, like choosing to give a player a particular quest because it looks like it might be interesting to them, it's hard for me to say what part of the training data actually led to that decision. In the same way, it's hard for me to sort of look back on all of my experiences and say what made me choose the words that I'm saying right now, if that makes sense.
Liberal
David Myles Liberal Fredericton—Oromocto, NB
It does make sense, but I think I want to go back to the generative.... You know more of the technical language about the artistic side of things, too, so I'm wondering if you see transparency, particularly, as being an impossibility when it comes to generative artistic pursuits, because that's what we're talking about a lot. It's not just the agentic element, but the actual generative songs, artworks, photographs and those kinds of things.
Are you saying that the transparency element is actually a technological impossibility?
Co-Founder, Artificial Agency, As an Individual
What transparency methods are you considering at this point in time? That's the question I would ask back.
Liberal
David Myles Liberal Fredericton—Oromocto, NB
I guess it's more on the inputs. Let's say that I want to write a surf song that sounds like the 1950s. Then, it would come back and say, “We used the Beach Boys.” Is it a falsehood that it would be able to give me a recipe for what inputs came into that output?
Co-Founder, Artificial Agency, As an Individual
You would need to come up with very different models and very different training regimes to be able to facilitate that, in comparison to the models that we have right now, at least the models that I'm presently using, and identify.... Again, it's sort of like saying that Rothko is why I want the colours to be this way in the sunset when I'm putting a photograph together. Am I really right? What actually influenced that? It's not so straight from training data to output. It's not a straight copy.
How the interrelations and the interconnections are made by these models is not as transparent as is made to be seen by that example. I'd be very skeptical.
Liberal
David Myles Liberal Fredericton—Oromocto, NB
I won't touch.... I'm just going to let it go and do its thing.
If we then take out all of the creative works, how does the machine fare?
Co-Founder, Artificial Agency, As an Individual
It depends on what model you're training, and to what end.
Many of the models that I've worked on.... Of course, we use game data from the partner studios that we're working with in this present role that I'm in, but in other cases, where I've been working on, say, bionic limbs for people with limb differences or amputations, and developing control methods so that they can use their arms more naturalistically, that data wasn't creative at all, but it also wasn't intended to be used in this particular space.
Of course, I can understand why it is of deep importance to this committee to consider the inputs of these systems in terms of training data or in terms of prompting. In this case, by removing this data, you're removing the conceptual semantics and the ability to learn these sorts of things.
Of course, these models can be trained on a wide variety of things to achieve different outputs. You wouldn't be able to get the same sorts of results in some cases.
I'll use, actually, a really interesting example of a connection—
Liberal
The Chair Liberal Lisa Hepfner
We don't have time for it right now. I'm sorry.
We'll have to move on.
Liberal
The Chair Liberal Lisa Hepfner
Maybe we will have time in the next session.
Mr. Champoux, you have the floor for two and a half minutes.
Bloc
Martin Champoux Bloc Drummond, QC
I told you that we should continue this meeting until 11 p.m. tonight.
I'll come back to you, Mr. Roberge and Mr. Grenier.
Ensuring the discoverability and presence of francophone and Quebec cultural content on these search engines or these artificial intelligence tools is certainly a major concern. If we can't make Quebec creators' content available or help feed artificial intelligence tools, is there a risk that artificial intelligence will standardize culture and undermine the diversity of content, which we want to protect in Canada? Could there be this perverse effect?
Mr. Grenier, you haven't spoken much in the past few minutes, so you can answer my question.