Uh, great. Hi everyone and welcome. I see 148 people here and counting. So exciting to do an event together. So for those of you who are part of our OpenSIZ community, welcome back. I think this is a familiar space, kind of. And for those of you who are new, welcome. Really excited to kick off the Connect Conference, a virtual conference for three days highlighting collaboration and why it's important in today's ecosystem of working this technology and sharing some really fun research projects, but also opportunities for collaboration with everyone. So day one of Connect, it's going to be focusing on — I'll have a short talk in the beginning and then we will dive right into some of our really successful community-driven research projects in the last couple of years. And this is a really good point. There's a lot of us here and probably the chat — I would say just use the chat as much as you want for expressing your emotions and ideas, but the Q&A section of the Zoom call would be a really good place to collect in an organized way your questions about — for different talks that we are going to go and have today. And feel free to just add questions because it's going to be lightning round talks and very quick. So we want to capture every question. Also, my suggestion: check out existing questions and upvote if there's something that is what you want to ask, so that there's a nice collection of questions that we have in the end to go through with at the panel, but also with the speakers during the talks. But also the chat is where you can just express — I'm loving seeing all of the from people from Rwanda, Nigeria, Toronto, UAE, just yeah, see everyone expressing where they are from and how's the weather there and maybe why you are here and what you're looking for to learn here.
Uh, cool. So today, like I said, I will kick off the conference and we will have Joelle here as well. Ah, nice, she's already here. Great. And yeah, so excited to have her. And then we will have lightning talks and again, question answering function for those who want to come up with questions. So let's start. If Britannia, if you can stop sharing maybe for a second so that we can all see each other's faces. Great. Welcome, Joelle, and let's officially kick off the Connect Conference. I'm very happy to introduce Joelle. She joined Cohere as the Chief AI Officer recently and I don't think she needs introduction. She has done an extensive amount of work in research before at FAIR, establishing the fundamental AI research at Meta, and she has done a lot of interesting research in RL, robotics, and at Cohere I will leave to her to share what she's doing here, but we are very excited to also have her as one of the leaders to lead Cohere Labs and the research direction for us. So without further ado, take it away.
Thank you, Marzia. Great to be here with all of you on this opening day of the Cohere Labs virtual conference. I do think this is an amazing community and so it's really an honor for me to be here joining you today. I'm based in Montreal, Canada. We have snow on the ground today, and so love to see all of you who are joining us from warmer parts of the world. I've been in Montreal for a number of years. I'm a faculty member at McGill University where I ran the Reasoning and Learning Lab for a number of years. I think Marzia gave a quick overview of my research interests really. I came out of reinforcement learning. I used to do reinforcement learning there where there were about, you know, 50 people in the world who knew what reinforcement learning was. Things have changed a little bit, but it took many years until we got there. Out of reinforcement learning, I started doing a lot of work in robotics, building a nursing assistant robot, wheelchair robots. I started doing work in dialogue systems. If you can imagine what was the nature of dialogue systems back in the early 2000s, given the state of compute, the state of data, and also the technology back then. I started doing work on applications of AI to healthcare because I do believe there's amazing potential to use this technology in a way that transforms people's lives. And these four themes, I would say — reinforcement learning, healthcare, robotics, dialogue systems — are themes that have accompanied me through a couple decades now in this field. What's really interesting for those of you who are at the early stage of your career is to take some problem, a real problem in the world that if you solve makes a difference, and carry that problem through several years as you see the technology evolving, as we went from predictive AI to generative AI now to agentic systems, every time it unlocks new capabilities for the problem that you care to solve. And so I have no doubt that many of you have real, you know, concrete problems to solve in the world. And I do think having a broad community who looks at the world with different eyes from different vantage points, who sees different problems to solve, is one of the ways in which we're going to be able to really reach the full potential of AI.
I joined Cohere less than two months ago, super early days. A few things really drew me to Cohere, I have to say. One of them is the strong community-building orientation that it has, and in particular the commitment to open science. Eight years prior to this, I joined Meta to lead an AI research team because at the time Meta — was Facebook in those days — was the only company where the AI research team was fully focused on open science, on making sure that the results of the research go out, whether it's through papers, code, models, datasets. And through my eight years or so at Meta, it was an immense privilege to contribute to projects — projects such as Llama, Segment Anything, DINO — projects which now have been downloaded by billions of people around the world, perhaps some of you. And I strongly believed throughout all that time that it was in the interest of large companies to make these models open, free to use for the community. It means from a business point of view the model got better faster. It means that we got to test out things in research before we brought them to product, and we got to benefit from the incredible innovation of the whole community. And that's a lesson that had been learned from the early days of PyTorch and that we carried out through these projects. Now things are a little bit different in 2025, and it turns out that the best place to do open science in a way that actually opens up onto a community is at Cohere. It's one of very few places outside of the US and China where you can actually train models at scale and you can release them openly for people around the world to be using. So to me it was really a no-brainer to be joining this amazing team who does great research. I will also say that one of the advantages of being at a company like Cohere is the — you know, the business model is really focused on making AI, agentic AI that is trustworthy, that enterprises can trust. And so that means a lot of the aspects of building AI in a responsible way, in terms of having security and privacy top of mind, having strong safety guarantees, having transparency standards, is integral to the success of the product. So the way that I care about building AI is the way that the company needs to build AI to succeed. Having that alignment, I think, is really very, very important today. In terms of where I'm excited about going next, you'll hear some of that through the presentations today. I've been working for the last couple of months with the team to establish the key research themes that we will be pushing forward. The Cohere team is relatively small in-house, but magnified through all of its community, I think, can do really great things. One important direction of research is to really carry forward the work that the team has done on multilingual and building multilingual AI systems, and broadening that up to building systems that take into account not only all the languages of the world but also the cultural dimensions of AI. AI are not just language production engines. They're becoming multimodal. They're becoming agentic. And so having a dimension of multicultural aspects modeled is going to be increasingly important beyond just the languages. Of course, work on responsible AI continues to be a huge priority for the team. You'll hear about some of the results of past year, but I do think this is an area that is going to take even more importance. Tackling head-on problems of hallucination, misinformation, problems of safety and privacy, I think continue to be important to build the AI that we want to use. One phrase that really sticks in my head at this vantage point — I have four children who are now, you know, teenagers and young adults — and every day I feel I have to build the AI that my children are going to use. This isn't about just building AI that goes into the cloud. This is building the systems that we want our children to be using in decades to come. And from that prism, I think we have an incredible responsibility in terms of what we are building.
Last but not least, in terms of research areas, I think this notion of building AI not as a singular superintelligence, but building AI that has social intelligence, building AI where the agents work together, where the agents work with individuals, with societies, with institutions, is incredibly important. So this notion of collaboration, cooperation — how do we build that into our AI systems — I think is an important theme of research, and I've had lots of good discussion with the Cohere Labs team. I hope it permeates through the next few days in terms of research topics. This is really like what we do in terms of research. But a really important ingredient is how do we do research, and I would say this is probably the most durable part of the agenda for research labs such as this one. The commitment to doing research in collaboration, the realization that we are building on results from decades of work and we are putting the seeds for decades of work going forward, is really important. So I do think this event, bringing this community together — there's this event, but there's several events throughout the year to bring the group together — is really, really quite important. The collaboration between industry, academia, government, civil society — an important aspect of how we do this work — and the commitment to continuing the tradition of open science and open research. Sharing the artifacts of our work globally to people who can improve them, who can critique them, who can observe them, who can build on them, continues to be incredibly important. I know we have many members of this community who are active in terms of fostering open research. I celebrate the work that you are doing, all of you, towards that. I appreciate it. I'm here to sponsor it, to recognize it, and amplify it going forward. Couldn't be more thrilled to be here to be doing this work with you. I hope we have many more events together where we can interact and learn from each other. With that, let me close. Welcome. Glad to have you here. And I'm going to pass it back to Marzia and the organizing team. Thank you.
Thanks, Joelle. That was great and really excited to have you on board. Nice. So, let me start the — right, I think everyone should be seeing my presentation. Yeah, we have a packed program today. So really excited to get started. And I just wanted to very briefly, for the newcomers who are not really familiar who we are, Cohere Labs, to just briefly introduce ourselves. And adding to what Joelle mentioned about this big part of our mission — mission is doing good research in open science community. So really building a new ecosystem of how research is done right now in AI. And we are a very small team. We have the luxury of a really nice community of researchers who are excited to collaborate. And with our small team, we have published for the past three years over 100 papers at different conferences and released two families of models, in particular models which are the multilingual — one of the first, actually, when we started three years ago, the multilingual models. And I'm very excited to go through some of the lessons learned from these collaborations in the community as part of the research that we did in the past three years today and tomorrow. And we're also really committed to helping emerging researchers to get involved in research who are interested to do that. As part of that, we have this annual research scholars program that we are running every year. And these are some of our biggest research bets that we work with — people who are more on the junior side who might not have the opportunity right away to get involved in bigger institutions, for instance if they want to continue doing their PhDs. But they have a lot of potential, and we are very proud to have been working with people from around the world to provide this opportunity to really create the next generation of researchers who end up joining Cohere or going forward to do their PhDs after finishing the program. And the community, the open science initiative that we have, is a big part of the identity of our lab. A lot of you are from this community and have been for many years. And it's a very active community. There are multiple subgroups and events going on at the same time. If you're interested in a particular niche and area of research or topic, you can find like-minded people in this community. Go to the reading groups, form collaborations there. And as part of this community and working closely with people there, we have also been able to publish a lot of impactful research in the last couple of years. But let me step back and look at why we really want to highlight collaboration and working together as the main theme of this conference. There are many complex systems in the world that almost all of them are not really built individually but really as part of a network, and from biology to economics to AI and research. And why it's really important is that because collaboration will expand the space of ideas. And the more — essentially, what you end up thinking about, the benchmarks that you end up working with, the methods that you think about are going to be limited to what you have experienced so far. But the more you collaborate with people, you are actually moving out of this isolation of the familiar space and expanding what else is out there and what you can learn in terms of just, yeah, what you really want to then explore as interesting research questions. And something that I really believe is that when we look at the last decade in particular, there have been a lot of breakthroughs through scaling of compute, data, models, parameter sizes, but there's also the social aspect of scaling, and that is scaling participation and getting more people involved in creation of this technology. And when we talk about participation, one of the first things that also comes to mind when scaling participation is creation of datasets. But let's not limit participation to just that. Really, how we think about scaling participation is who really ends up framing research questions, who looks at the model behavior and interprets how it's doing, and really what success looks like according to different people who are involved in research. So with that in mind, we intentionally are holding this conference as a virtual and global conference. So this is a map of all of the registered people for the conference, truly from around the world. And our hope is that we can connect people by the end of this conference — people who share the same interest in research, have the same curiosity about what is interesting, what is important to work on right now, without really the limitations of geography and who you are, who you know, who you are familiar with, and really expanding your horizon on the type of problems that you can work on. And really with this cross-region and cross-disciplinary collaboration, the hope and the goal in the end is to increase the surface area of the search space that you will be working on. And just to get a little bit less abstract and more practical, what does really collaboration look like? And it can be on a small scale, just working a couple of people together on ideation, experiments, and iterative thinking. It can also be on a much larger scale, a lot of times involved with data collection or making impact in terms of changes in policy that really needs a larger-scale point of view. But really how it feels like, it's a little bit more different and nuanced. Like a lot of times, collaboration is really two people looking at one paper together and talking about what it looks like, maybe like a few people learning a particular method and then re-implementing the outcome of a paper, or looking at a published evaluation and discuss maybe the anomalies there. So it can be really very small individual talks and discussions that can compound to a more tangible and valuable outcome. And a lot of times, collaboration, especially global collaboration, will look like this: a lot of async alignment and talks, and whether it's like weekly Zoom calls or sharing a document that everyone can leave their feedback and their opinion, and just building a research idea together, and just constant communication between everyone on the team. We started the Aya initiative three years ago, and we learned through that process. The first initiative, it took one year, that there's also this aspect of collaboration that you have to have fun in the end with people that you want to work with. So as part of a lot of the work of Aya was the curation of this dataset and the models that cover 101 languages in the end. But there was also a lot of parts that was about just sharing the music that everyone was enjoying around the world, sharing pictures, doing a recipe exchange. And I'm sharing this here because I think there's this part of the social and connection part of collaboration that might be overlooked sometimes, and we do need some bond in the end to have a successful collaboration, and don't really underestimate the power of connecting to people first before jumping on important research questions that you want to work on together. And as part of this collaboration, we actually ended up creating the Aya initiative, which has become much bigger than we even expected from the beginning, and I'm very excited to go through some of the details of this journey in this conference as well. And other than Aya, which was a really large group of people coming together, in the next three days we are also showcasing the large spectrum of research that is possible through collaboration. So Aya is a really big example, but there are also more smaller groups of people working together on the Global MMLU and the Include project, for instance, or even smaller groups of people working on AfriAya or Maya and NeoBubble paper. So there are all essentially different ways of working together at different scales, and a lot of times we see the paper, which is the end result, but we don't hear the story behind it — how this group of people got together, worked through all of the challenges, and created that research outcome. And our hope for this conference is that we learn that. And like I said, really scaling participation. I truly think that this will be the next breakthrough. So the more people are involved in what research questions are important, how to interpret the results, and where to explore to find interesting research questions next, the more we will get closer to really capturing this human intelligence that with this technology we are hoping to capture. And with that, I'm going to wrap up and just share that this conference is in the end an open invitation for everyone to talk to each other, get inspired, find people who share your same interest as you. Start small. It doesn't have to be the largest end goal in the beginning. And yeah, let's work together. Thanks. And with that, I'm going to pass on to the official start of the conference and the lightning talks. Before that, let me just quickly see — are there any particular questions that we want to go through now? Meline or Britannia, do you see any questions for me or should we just jump into the next?
I think we should jump into the talks. We've got lots to cover today. If there's anything we missed, we'll try to get to it on our community afterwards, but Britannia's just —
That sounds good. Great. So, let's start with the lightning talks. And I'm very happy to introduce the first speakers, Freddy and Wayne. So, we go way back. It's really good to see both of you guys here as part of the original team one and team three in Aya, now three years ago, and they're going to be sharing the story behind building this massively multilingual model and dataset. And without further ado, I'm going to pass on to both of you.
