Hello everybody and welcome to AI Inside, the show where we take a look at the AI that is layered throughout the world of technology. I'm one of your hosts, Jason Howell. Jeff Jarvis unfortunately is out today, so you know, I'm sitting here in my studio flying solo, but I've got an amazing interview for you. My guest is Joelle Pineau. For years she ran FAIR, which we've talked about. Jeff's definitely brought up FAIR a lot on the show. Meta's fundamental AI research lab, one of the most respected research operations in the field. Last year she walked away from that to become the first ever chief AI officer at Cohere, which is a Canadian company that's making a very different bet than a lot of the big US labs out there. They're not, you know, chasing AGI. They're developing enterprise AI for efficiency, security, and in many cases openness. She's a lifelong champion of open science. Her career actually started in healthcare AI, intelligent wheelchairs and seizure detection, which we actually discuss and tie into her current role at Cohere. She's also very outspoken about the need for sovereign AI. And well, the timing could not be any crazier. Just a few days ago, you've probably already heard, the US government forced Anthropic to cut off foreign access to its two most popular models. Anthropic pulled them entirely for everyone. And mind you, this is all at the point of this recording. News may have changed cuz this isn't going to go up immediately, but this is a real argument for sovereign AI. Joelle shares her thoughts on that and how it actually impacts companies like Cohere. And Cohere has been on a tear as well, open-sourcing a powerful coding model just last week, a move into drug discovery, deals putting its agents inside banks. Oh, and Joelle picked up an honorary doctorate and gave a commencement speech about facing uncertainty in the industry, which is where we're going to start. So, let's dive right in to it. Here's my interview with Joelle Pineau. Joelle, thank you so much for hopping on AI Inside today. It is such a pleasure to meet you and have you on the show.
Yeah, and this is, um, this is kind of a... well, you've had a big week. I was kind of looking through your LinkedIn, of course, and you just posted, first and foremost, congratulations on receiving the honorary doctorate from the University of Ottawa just last week. That's pretty... That's amazing.
Yeah, that was on Friday. Ottawa's my hometown. Most, you know, lots of people in my family are graduates of the University of Ottawa, and it's really special to be affiliated with them, and in particular with the School of Engineering there. They do great work, and so it was really an honor, and, and all of course always great to see the graduates walking the stage and all the families.
Yeah. And to be in a position to be able to kind of share some of your knowledge and inspirational, you know, your wisdom and and everything. It's a... It's a cool opportunity to do that. Um, you shared kind of your your entire speech in text on your LinkedIn, and I was reading through it, and you know, the the kind of focus of it is optimism in the face of uncertainty. You said that the graduates are stepping into one of the most uncertain and consequential moments in human history. Certainly something we've talked about, me and my co-host Jeff Jarvis, who could not be here today. He would would have loved to talk with you, but he had something come up. But we've definitely talked about that uncertainty and I have brought up a number of times on the show about what it's like to be a graduate in this moment. Like never in my lifetime has it been the case where you go to school, you work super hard to to make all of these achievements and set yourself up for what your future, what you think your future is going to be. But here is this like moment in technology that really kind of puts a big question mark on things, uh, whether rightfully deserved or not. I'm curious to kind of hear your perspective on that.
Mhm. And and let me say, you know, it is a super interesting time for any of us at different stages of our career, right? I meet people who are 20, 25 years into their career and are wondering like, what does this mean for me? And then you talk to new grads and and they have the same questions. And I have the the immense, you know, privilege to also be a parent of four teenager/young adults, one graduating university this year, one graduating from junior college, two from high school, who are navigating them and their friends these same questions as well. Um, so, what some of the things are, you know, are not changing, right? Like the ability to learn hasn't changed, the ability to be adaptive, to move with agility, to lean on curiosity, to have like a great network of, um, colleagues, uh, all of these elements are still incredibly important and will help people navigate all of that. Um, that being said, you know, like this is the first university cohort that is graduating since the launch of ChatGPT. And so, all of their university studies have been in a time where that technology suddenly was available to them and was changing the the marketplace in which they're stepping in. So, definitely a lot for them to to digest and adapt to.
Mhm. Have you, I mean, being that you were there and and with them and talking with them, and I'm sure you know, have shared experience with them, like is is there anxiety for the folks in these positions or is it excitement? Cuz you know, with with a lot of change comes a lot of unknown directions. It's not, it doesn't all have to be negative or or bad. I don't know, what what was the, what was the tone like?
You sort of see the full spectrum. You see a lot of them, I mean, this is, you know, a joyful affair, like a day of celebration. So, many of them are in fact quite happy and excited. They they see the potential of it. They're eager to jump into the next phase of their lives, are proud of what they've accomplished, but eager to move on to what comes next. Um, some of them, you see it's the families around them who are worried about the choices. It's the families who are bringing in some of their their anxiety and they want some answers. They want to make sure the kids will be okay. Um, but most of the graduates that I talked to were quite positive, um, and excited about about the opportunities ahead. That may not be the case everywhere, but certainly the ones I've been in touch with recently. They're they're curious. They're confident and and they're really eager to contribute and to bring their own way of doing things into the world.
Mhm. Yeah, at a time when the tools and how we do things are shifting so drastically, and you know, you see on the on the side of schools, you see the one approach which is, you know, absolutely not, the way we teach and the the way you learn is a certain way, and so these tools are are, you know, not allowed or not welcome. You see a whole other kind of approach which is, well, but if this is the way of the future, we're doing you a disservice by not teaching you like healthy skills or responsible ways to use these tools, because that's really, that really seems to be the direction that this is all headed. And uh, that you kind of need to know those skills.
Yeah, I do think in many places there's a missed opportunity to, you know, educate in terms of how to use the tools appropriately. Um, you know, you wouldn't let, you know, students through like a machine shop without like proper guidance and proper instructions in terms of like the safe usage of of the machines available. Similarly, you know, today the the technology landscape is such that they have access to very powerful tools that can do all sorts of things. Um, and so not to give them that preparation, I think is not is not helping them. Now, we also have to be understanding that for many of the teachers and educators, the change has been incredibly rapid. They haven't had the time and the support to really digest this technology and to decide how to bring it into the classroom. So, that is adding a lot of a lot of pressure onto onto folks who are already carrying a big workload.
Mhm. Yeah, 100%. It's it's not just the the students trying to figure their way through the the tools. I was just... Yeah, talking with, you know, at a much different scale, but talking to like a high school teacher friend of mine and just kind of the the challenges of of of on one hand recognizing the value of these things and being like, "This is amazing technology. It's it's incredible what we can do with it and what what is possible." But on the other hand, just not having the the built-up kind of facility around how how this like transitionary period between before and after, how this happens. And it's almost like I want to I want to skip forward like 5 years or 10 years down the line and see where things where things land. Where do you think things are going to land like in that regard based on where we're at right now and and the direction that things are headed?
I I think what's most interesting for me is actually the fact that none of us know where we're going to land. If you meet someone who tells you they know, I think you should not trust them. Um, there's an immense amount of uncertainty. And and the other thing that is, you know, quite frankly, um, really interesting for someone in my position is there's really a sense that we can shape that trajectory. And we get to shape that future in in many cases by having a vision of where we want to land, you increase the chances of of that outcome. Um, which was also part of my message to the graduates, right? Like, you know, shaping that trajectory, uh, taking actions even in the face of uncertainty, I think is really important. So, I'm a lot more interested in in figuring out where we think we want this to go, and then figuring out what is the level of agency that each of us has to shape that future.
Yeah, and along those lines, you you told, um, grads, kind of similar to what you were saying, um, say yes to the role that they're not quite ready for.
That That's that speaks to me because I've done a lot of, you know, I used to work for a couple of companies, you know, bigger companies doing podcasts and stuff for them. Only a couple of years ago went independent and faced the kind of difficult and, uh, uncomfortable, you know, uh, scenario of saying yes to things that I'm like, "How am I going to be able to do that?" Well, I'm just going to say yes and figure it out. So, there's a lot to be learned there. Um, so that's very valuable and and valid advice. But, um, you know, at at Cohere and and elsewhere, building agents that are changing what things like entry-level roles actually are, you know, um, North MiniCode, that launched just last week, which congratulations for that. That's awesome. We'll definitely talk about that, but I'm I'm curious like how much of what a like a junior engineer does can already be handled by a tool like North Mini Code today, and where does that leave like the risk in the the optimism under uncertainty equation? Like where does that risk go? Is that risk falling on the hands of the the graduate, let's say, or does that fall on the company that's building the tool? Who should shoulder there's more of that risk, I guess, is what I'm asking.
I think of it a lot in terms of like you should always shoulder the the responsibility of of verifying and ensuring the quality of the work that you're doing. So, whether it's the new graduates, whether it's the company, whether it's the more experienced team member, right? Like there's a piece of work that you're going to take on, there's a project that you're signing up for, there's a role that you're taking on. Having clarity on what you're expected to own and deliver, and then putting a very high bar in terms of the quality of that is incredibly important. Um, and so, you know, part what we see a lot internally, and you know, North is our generative AI platform. It has the ability to use models, generative reasoning models, coding models to accomplish work. We build agents, we build automations. We use it a lot internally. Our employees, most of them use it every day. Um, and so, for many of them it means that there's some piece of work that they can do much faster, there's some piece of work that they can, you know, delegate to automation, especially when it's repeated. Um, there's some piece of work with the AI as a close partner in doing it, but at the end of the day, like the human in the loop is accountable for the quality, for verifying the work. Um, and so, that responsibility hasn't hasn't changed. Um, people feel they can completely off-source the work. No, you have a partner, you have a team member, but you need to figure out whether the work has the quality that's needed.
Mhm. Yeah, I've I've definitely on a on a personal level experienced that myself too. Kind of gone through the general arc of like, "Oh my goodness, this is amazing. I can just... Yeah, you do all that stuff and I'm going to be over here and then I'm going to check in and I'm going to take that and I'm going to going to use it." But removing myself so far out of the equation that like there's no knowledge there. There's no way, you know, on-ramp for me as the human to kind of pick back up with the with the detail there. So keeping the human in the loop like you said, super important. Maybe maybe that's a, I don't know, maybe that's a a transition or a transformation that everybody kind of has to go through as they learn how to work with these models cuz it's really tempting to be pulled deep down the, "Okay, I trust you to do everything. You got it. It's cool." Yeah.
And and we see a lot internally when we have teams working with each other, right? We we have developer teams, you know, you get the AI to write down the specification, you get to the AI to write the code, to write the unit tests, to to to to write the, in some cases, the documentation. Like all of that is totally okay. I'm game for all of this. But then you need to verify and certify all pieces of that. You can't take some work that's been sort of by coded and like throw it a fence over the fence to your partner team and push onto them the burden of verifying your work. And so that's usually how we, you know, we set the contracts between the teams in terms of accountability. We give them a lot of freedom to use AI North platform in their work. But how we phrase the responsibility, you know, goes back to the to the principles we had before.
Mhm. Mhm. Now, um, circling back a little bit to the optimism under uncertainty piece, as you can tell that before the first break, like I'm I'm really kind of dialed into, um, your your speech there cuz a lot of it really really spoke to me. Um, And I And I look at kind of where where a lot of this development is happening right now and this this idea of being optimistic even through uncertainty has the has the possible temptation of leading people to really over explore or never quite, you know, they just keep spending and spending, like I see this a lot in the news. Keep spending, trying new things. Um, there's a lot of that happening in the industry right now and I think looking at the state of the industry, everyone seems to be chasing a lot of the same bets. They're making the same, you know, the same assumptions and over exploring that regard. Do you Do you think that the field is over exploring right now? Is everyone kind of going in the same direction or or like what's your thoughts there?
Mhm. Um, I think the field is definitely in a... [laughter] exploration mode. Is it over exploring? I'm not sure. The The piece that you know, that may get forgotten in this, you know, the the notion of optimism over uncertainty and you know, I didn't get into it but there's like a very rigorous mathematical treatment of this idea which folks who studied reinforcement learning are very familiar with this principle. The idea of that optimism is to, you know, take an action with which has a high upper bound, right? Like high upside potential. But once you've tried this out, you need to learn and either you realize like the outcome is really that good or you realize the outcome isn't so good and then you should start trying something different. And so when you look at it in sort of a little bit more detail, the part of how do you incorporate the learnings from all of that exploration is just as important. Just keep on exploring without aggregating that knowledge is completely wasteful. Quite frankly, it's like, you know, you're just playing the lottery and and that's not the way that you're going to make forward progress.
Mhm. Yeah. Well, and you've talked a lot about about like algorithmic monoculture. You've talked about diversity in these in these models and again, you know, this kind of plays into that, right? Like where is the actual diversity happening right now? It seems like a lot of a lot of companies, a lot of people are focusing a lot of their time on the LLM side of things. You've got folks like Yann LeCun focusing on world models and exploring it from a different perspective. So, there's a little bit of of diversity going on. Like what are your thoughts there? Like what what is truly diverse in in the exploration of AI that's happening right now? Like what's what's kind of catching you?
When it comes specifically to to modeling, I think there's sort of two two angles, right? One angle is trying to solve different problems. And one angle is trying to solve the same problems, but in different ways. Um, so, you know, in terms of solving different problems, we see some folks looking at physical AI. We see some folks looking at, for example, AI for science. There's incredibly interesting problems to be solved in terms of, for example, different verticals, AI for the law, AI for healthcare, and so on. What was most interesting for me coming into Cohere was actually to look at building AI solutions for the enterprise cases. Like so much of what's being discussed and what we see is consumer focused. A lot of the benchmarks are very general intelligence benchmarks. But actually, you know, when you're completely focused on building models for enterprises, you start looking at different types of use cases. There's different pain points. So, we've had to build all of our own testing and evaluation infrastructure. We're building our own environments for these kinds of things. We're building sort of synthetic world models, but our world models are about like how companies and enterprises work. Um, and so that has been really interesting to see and to to spend some time on and and one of the reasons to me that was, uh, really important to do is because when you're dealing with enterprise customers, you get a really clear signal of what works, what doesn't work. You have paying customers, they will tell you if the solutions you're you're building don't don't work much much more quickly than than perhaps some of these external benchmarks. So I do think there's that exploration from like a diversity of use cases perspective. And then there's an exploration with respect to different, um, methods, different hypotheses of how to go about this. You know, should we use LLMs in the classic sense? Should we explore different architectures, different paradigms? Yann LeCun certainly has this hypothesis with respect to... Um, others have their own hypothesis. We've seen a difference between dense model and sparse models. Um, there's a lot of work that's going on in terms of newer architectures that might get us past the transformer architecture. I would say right now for a company like Cohere, the exploration is more in terms of the types of use cases, though we have an exploratory research team that also focuses on new architectures, but we think it's really, you know, looking at new problems is going to give us the signal we need in terms of of what's really bringing value.
Love it. All right, we're going to take a super quick break, um, come back and talk about, you know, you're you've talked a little bit about kind of your background, you know, moving into Cohere. So we're going to talk a little bit about where you came from, where you are now and and what Cohere is up to. So that's coming up here in a moment. Hey, I hope you're enjoying my chat with Joelle Pineau, uh, but really quick, I want to thank you. You know, yes, if you're watching and listening, I want to thank you, but I also really want to thank the patrons because without the Patreon, this show would not be a show. Like it literally helps support the health of this show. So, I want to throw out a huge thank you to Peter and to Casey Kamiyama. Thank you, just a couple of our amazing supporters on Patreon who help us do this show each and every week. You know, they get access to all the extra perks that we have, including a daily podcast called AI Inside Daily. Yes, that's right. If you support on Patreon, you get access to 5 days a week of new news-related episodes on top of the regular episodes that you watch every Wednesday and some of these interview episodes. So, it's a lot of episodes. Thank you in advance for supporting. We couldn't do this show without you. And yeah, so patrons, you're awesome. Thank you so much. All right, going to take a quick break, then come back with more with Joelle.
All right, Joelle, you you left one of the, you know, a very prestigious lab at Meta, the fundamental AI research team. We've talked about it on the show many times, FAIR, to join Cohere just last year. And that was, I don't know, that was a that was a big bet on the highest upside that that we're we're tracking right now. What what made this transition worth betting on for you? This was a big transition for you.
Right. Um, I mean, you know, I spent 8 years at Meta building up the the FAIR team, initially the team in Montreal, over time the the global AI research team. It's It's a group that has done fabulous research. And I would say one of the the most important contributions, perhaps in that work, is to maintain that culture of open science and of sharing that work openly. That was really important to me. We have very strong buy-in from the leadership of the company to keep doing that. And I thought it was very important because I think when you release your work out into the open, that's when you get the fastest and the most accurate signal of the quality of the work. And, um, I always really welcome that that feedback even when critical. I feel that's when we That's when we learn the fastest. Um, after having spent a number of years there, I think I was ready to take on new challenges and in particular get closer to the enterprise space and have a real signal. Like there's so much talk about AGI and pursuing frontier models. But at the end of the day, like the ways in which these model breaks are very different from what the benchmark capture. And so I just felt like a need to get to that, um, signal and get closer to the product side. And a company like Cohere and the particular stage of growth I was walking in, it's a lot easier to get that, uh, to get that signal. And the other thing that to me was important is I I do think we're at a we're at a particular point in time where it's important to have more people rather than fewer people who are building frontier models. Um, there's very few of them currently outside of US and China who have the resources, the talent in order to do that. Cohere is one of them. They have an incredibly talented team. So I thought it was a it was a good time to sort of partner with them and and, uh, put some of my experience behind those efforts and make sure that we have a multiple number of models that people can choose from.
Mhm. Mhm. So on the flip side of that, what did you have to... What like Was there anything that you had to give up? Like that that I don't know that that it comes to mind in that transition. Something that, uh, I don't know, was a little challenging to to give up as you transitioned over?
Well, certainly. I mean, there's a number of, you know, wonderful people that I worked with. Uh, it's it's, you know, always a little bit of, uh, sadness of not seeing them every day, not partnering with them every day. Um, and then we had a number of research projects that were underway. You know, some of these research projects are over 2 years, 3 years, 4 years, and I was stepping away where some of them which I really helped incubate and supported in the early days, and I knew they were about to to hit their strides. And and so, it's been fun over the last year to see the those projects come to fruition. Uh, you know, the SAM 3D project is one of them where I've been involved for a number of years, and sort of you you let them you let them run on their own once you leave. But it's a it's a little bit like, you know, having your your your kids grow outside of the of the house. You're happy for them. You celebrate their success even if you're not there every day supporting them.
Yeah, yeah, indeed. Indeed. Um, now, FAIR was, you know, built around this idea of publishing everything, very like you said, open research culture, uh, that sort of stuff. Cohere, you know, definitely a different kind of focus, selling enterprise security, enterprise control, that sort of stuff. Is Was there any sort of... Well, I don't know, when you when we're talking about the open science part, was there any trading away or or bringing that over with you? Like, what what does that look like?
I would say that part was pretty easy. The co-founders already had a strong commitment to open science. They are all three of them, uh, Aidan, Nick, and Ivan, coming out of the University of Toronto ecosystem. Cohere Labs was well established as an exploratory research entity. And, uh, was already releasing models, publishing, all of that. So, I would say the culture was really there. It that was a very easy conversation. You know, I think we all believe that this is important and that we're going to support that and and we've seen that continue through the year. And so there was no no difficulty there. I'm continuing to support that movement to help them out whenever I can, but I didn't have to push very hard to bring that in.
Yeah, amazing. Good. Well, then you landed at the right place. Um, we talked a little bit about, uh, North Mini Code and, uh, maybe we can, you know, while we're talking about kind of like openness, maybe we can come back to this a little bit. Um, a lot of US labs right now are kind of in a moment where they're tightening access around it and, you know, on the flip side you put the weights out there in the open. A lot of what, uh, Cohere sells, uh, to enterprises, um, is, you know, in some ways can be the opposite of that. Security control, keeping things in house. How do you decide what's worth opening up and then what kind of stays closed? What what does that decision process look like for your team?
Mm, that's an interesting one. And and I would say maybe that's one place where I've been able to to contribute in the last months is to help us articulate more precisely that, uh, that criteria. First First thing to know maybe is in...
In many cases, like the model we open source versus the model we distribute to customers, are not exactly the same one. They're, you know, the pre-training may be similar, but in terms of like the fine-tuning, there's a particular version towards enterprise use cases that we ship in the product, and then there's another version that's more suitable for general use that we open source.
In terms of the decision process itself, really it's about evaluating, you know, what's the value of doing that to the community and what's the value to Cohere. And in terms of value to the community, it's like, is this the type of model that is not already available? Does it bring specific capabilities or capabilities up to a particular efficiency point that could accelerate the work of others? In terms of benefits to Cohere, like, are we going to learn something by releasing that model? Are we going to get some really high quality or very quick feedback from the community? Are we going to have people build on this that are going to help us improve faster than we would otherwise? Are we going to get some good visibility into the community?
And I do think open sourcing plays an important role in terms of, you know, getting that recognition from our community of peers, from developers and researchers, that in return attracts talent. And we are incredibly dependent on having a talented team. It attracts investments in some cases, and it allows us to have third parties verify our claims. Like, we think we have a pretty good coding model for that model size. Again, like, we have to remember this for now, but it's a relatively small model. You know, we can say it all we want, but when we allow third parties to validate it, it gives certainly a lot more credibility to those claims.
Mhm. Yeah, 100%. That's definitely the beauty and the value of openness, and open source and all that stuff. Kind of touching back to your speech a little bit, you said that there's a risk in sharing ideas openly, or there can be a risk anyways, but that the upside is actually worth it. So, when Cohere chooses to open a model or a part of a model, what is the risk that you're accepting in that scenario?
