Radical Talks, Masterclass Edition: Joelle Pineau on Turning AI Research into Real World Impact

Well, well, thank you. So glad to be here with you, Sanja, as well as with all of you here in the room, online as well. It's been such an interesting transition. To give some context, I'm someone who, quite honestly, has never had a career plan. I live in the moment, and right now that moment is characterized by what I would call joyful chaos. And so when I left Meta in May, I did not know where I was going to land. I was hoping to at least enjoy a few months off, but quickly found myself just drawn to explore what could be and what I wanted to do next. Lots of different conversations with a ton of different people. I think really what attracted me to Cohere is a few things. One, I spent a number of years in a very large company, and maybe it's good to rewind a little bit before that. You know, when I went to what was at that time Facebook, when I joined, 2017 was a really interesting point for machine learning where we were starting to see a bit of a divergence in terms of the type of research you could do in an academic lab versus the type of research that you could do in an industry lab. The amount of resources, the access to compute capacity were really starting to change. And so I was very curious to see what it was that you could build when you had access to that level of resources, that amazing concentration of talent as well. And so that's what attracted me there. But you know, after a cycle of eight years, I had a sense that I understood what was possible, what could be done, and I was ready for another challenge. So Cohere is that other challenge in the sense that it's a much earlier stage company. The company's about 400 people, a really talented team, also less resources perhaps, though I think there's a lot of potential to build interesting tech. And already it's a company that has shown that it could build good quality models, whether it's the Command line of models, the Aya line of models, and so on. And so it's interesting to be in something that's earlier stage. It's interesting also for me to sort of spread my wings beyond research. I think at a place like Meta, I'm leading research, and I did so for a number of years. It's not clear how to spread your wings beyond research in this case where there's so many sort of experts of their domains. You know, people with product expertise and so on and so forth. You get into something that's earlier stage, and you know, everyone just gets their hands dirty getting done what needs to be done every day. So that was appealing to me. They're really focused on making AI that works in practice, and it's one thing to be in the consumer space, you ship something, it goes to billions of people, but you don't necessarily get that signal of is AI actually working or not, is it actually bringing those productivity gains that we are promising to everyone. You go into a setting where you're selling AI to enterprise and to governments, and it takes longer to get your foot in the door, but then you get real signal of whether there's value there or not. And in some sense, I feel like I'm going to learn a lot that will inform me in terms of, you know, what is the potential of AI. Maybe at some point I'll want to go back deeper into research and solve some of these problems, but right now I'm just really keen to learn about that.

Yeah, I think I first, you know, came to McGill in 2004. I think a little nod to my friend Doina here, who was influential in bringing me here to Montreal to McGill. And at the time, I had no idea what I was doing. You know, you walk into an academic role, and you know, I knew how to do research on something very narrow called POMDPs. I don't know how many people still know what POMDPs are. World expert on POMDPs. But yet I didn't know how to teach. I didn't know how to write grants. I didn't know how to supervise students. I didn't know how to build bigger collaborations and so on. And so it was a huge learning curve. I found an amazingly supportive, nurturing environment at McGill, in Montreal more broadly. And I spent a number of years sort of building up my research agenda, my program, my lab, and so on, which was great. But there too, I sort of got to a point where I felt a little bit, I had a good sense of how to do that job. I think I'm someone who does best when I feel I don't know how to do this job. Like I show up in the morning and I do not know how to do this job, and I'm back in that space right now, and it actually feels good. And so, you know, 2017, a lot of the research is done in industry, access to increasingly more amount of compute power. The teams are more senior. It's a number of more experienced researchers working together, not necessarily grad students, though of course there's interns and so on. There's a sense that the research you're doing has the potential to go into product. So being able to be part of that transfer from research all the way to product and understanding how that works. You know, in the university lab, I'd worked on some project, for example, on AI and healthcare with clinical partners, but we never got to the stage of bringing this all the way to patients. And so that opportunity was what attracted me there. And I will say, at a place like Meta, you know, over time I got more responsibility. When I started, I was responsible for the Montreal lab, a team that is still up here, a little nod to them as well. And then I took on responsibility for all of the AI research teams globally, which was definitely a bigger challenge, but it came gradually over time. Then spent the last 18 months or so in the company leadership team, the executive committee, Mark Zuckerberg's sort of small leadership team. The company was going through a massive pivot towards becoming AI first, and frankly, he needed people in the room who knew what AI was to help bring the company along and inform the strategy. So that was also an amazing opportunity to bring in what I know to the room, but also to learn from all these other people who are running much bigger teams, who had a long history of having amazing product impact. So that was exciting too.

It's always really hard to judge that gap between the research and what's ready. I would say I am someone who's always been curious to go back and forth between the more theoretical side of research and the more applied side of research, you know, throughout my research career. On the university side, we built a robot wheelchair project. We built wheelchairs. We partnered with clinical specialists, with companies, and so on. Some of the students went on to do a startup with this technology. But I do think going back and forth between the two means that you get that feedback loop of, you know, on the more theoretical side, you can take on problems in a maybe a more distilled form and make some progress, and then you get to confront this with the real world and see what actually works, and you feed that back into the research. I do love being able to access sort of the full cycle of innovation in that way. But this question you ask of how do we judge whether something is ready or not? There's always a big part of uncertainty there, you know, how to judge that gap between what works in the lab, what works in theory, what works with a particular data set before you put it out in the wild. It often comes down to data and evaluations, like setting the right evaluation protocol to give you as much of that signal. But even there, I've seen, you know, you just have to do the leap at some point. It takes someone with conviction, someone who sees the potential, and also someone who's just willing to stomach the risk of doing it. And that risk can come in different forms. Sometimes, you know, you make that jump and it just doesn't go where you want, or it takes longer, or it costs a lot more. You never quite know, and I've faced that many times. It really helps when you're sort of working together towards the success of the project. I've seen cases where the research team is sitting in one place, a product team is sitting in somewhere else. You kind of throw the model over the fence. That does not work well. We evolved the process when we were at FAIR that was framed as a graduation. So often we would have some ideas, technology that would graduate from one team to another. It's a positive way to frame it. And also, you know, that graduation, it wouldn't be just knowledge or assets that would move, like people will move with it too. There's some of that knowledge that resides in the people, and so we would usually have a couple people switch over teams, and that ensured that there was deep knowledge of the technology. It also ensured that like the link to the team that initiated the work was still very strong, and so that was usually a better mark of success. But even there, I want to say it's almost like every one of them is a one-off, but every one of them you learn and you get better every time.

In all honesty, I'm three weeks into my job at Cohere, so it would be premature for me to show up giving anyone any lessons on this one. All I'm going to share is observations in terms of what works. I do think often when you're in an earlier stage company, it feels a little bit harder to take some risks in terms of, for example, like do research that is highly risky. But it feels a little bit easier to take other risks in that the product is less mature. You're earlier in the learning phase, and so there's certain things that you're, some risks you're going to have to take just to learn, and without that you're never going to get better at it. Right? So a lot of what I'm doing in these early days is really just understand how the teams are working now. Understand what research work is happening. Understand there's in this case like a research team, a modeling team, and a product team. And so understand how each of them operates and what is the interface API between these teams.

I think it comes down to having clear priorities. And I'm not yet answering your question, but I find sometimes people are unwilling to have priorities. They want to do everything all at once, right? And so it depends a little bit of like what's the runway you're giving yourself until you need to have like a minimum viable product. For some, it's going to be, you know, nine months, 18 months, three years, I don't know. But you got to decide with your investors what's the runway you have until that moment. That's going to dictate a little bit how much you have to tilt towards more the research or just the prototyping or really having a strong sense of product. What you can't afford to compromise on is, I think, having a strong vision and a narrative, because that's what's going to make sure that everyone's working towards the same north star. And so not having a north star at all, I think, is the worst of these scenarios. Having a very clear north star helps align all the efforts towards that. And then I would balance it based on the timing that you have. Sometimes you have to be opportunistic also. I think when you're very early days, you know, everyone or many people have to be a bit of everything. So having an engineer that really can lean into product means that you may not need a product person quite as early. Depends on what's the complementarity of the team. I do think it's useful to have people who have a high level of trust, who can challenge each other, who have the ability to like hash out the hard decisions. Again, going back to this question of prioritization and clarity, have the ability to challenge each other in a really productive way.

Yeah, this is an interesting one. You don't, you know, again, you know, you get trained as a researcher. You understand how to do research, how to apply the scientific methodology, how to write code, how to write papers, and so on. And no one really teaches you about culture, and yet it matters a lot. And I found, especially in my last few years, I was very much just like the curator of the culture more than anything else. By the time you get a mature organization, really strong bench of leaders, and so on. I was, as I mentioned, at FAIR for eight years. I started, and over those eight years, the field of AI changed tremendously. You know, 2017 to 2025, huge shift in the field. Facebook, Meta as a company changed a lot in those eight years, and FAIR as an organization changed a lot. We started, we were about 100 people. By the time I left, there was about 600 people distributed over 10 locations. There's one thing that didn't change, and that is the set of values of the organization, which feels like so archaic to say, but my very first offsite, fall 2017, I, you know, I met with the leadership team, and we like wrote down on a board a set of values, and we carried them through. And as I was leaving, that still sort of was my parting message to the team: those values carried us through. And in many ways, I think you could ask most of the people in the organization, they could tell you what the values of the organizations were without having to look back at the deck or anything like that. They were deeply ingrained in the culture of the organization, and they really characterized how we did the work. We called it our FRESCO: freedom, responsibility, excellence, scale, collaboration, openness. I think I will go to my grave being able to tell you what those values are. I live them. I breathe them. I sort of brought them daily to how we did the work that we did. And people think that values are like something on a slide. And yet you would be shocked how often they guided strategic decisions that we made in the organization. We had to make hard trade-offs about projects, about resources on a daily basis. And we'd go back to the values. And sometimes it's easy, you look at one value, but sometimes they would be in tension with each other. We wanted to open source something. We were committed to responsibility. It wasn't always easy to do. And so also being willing to live in the tension of these values was something we had to do. But yeah, those values carried us all the way through.