Radical Talks
00:00:55 · Jordan Jacobs
[00:00:55] S1: I am very pleased today to be able to be joined by Aidan, who I believe is late in the evening where you are. So thanks for doing this. Aidan is the co founder and CEO of Cohere. As some of you will know, as an intern at Google Brain under Jeffrey Hinton and Jeff Dean and others, you can tell us a bit about that history. He was involved in and a co author of the paper Attention is All You Need, which kicked off the modern era of AI and formed the architecture of generative AI. That was back in 2017, way back when you were a teenager. Subsequent to that, in 2019, Aidan co founded Cohere with Nick Frosst and Ivan Zhang. We were lucky enough to participate in helping them to launch the company and writing the first checks. If we could talk a little bit about that history. Most recently, Cohere, it was announced, is going to combine with Aleph Alpha or acquire Aleph Alpha, a German LLM company, which will make Cohere a multi continental business, both Canadian and German. But business is actually global. We'll talk a little bit about that and what that means in a world where sovereignty really matters. Cohere has grown incredibly quickly. Revenues in the hundreds of millions of dollars. Team is grown into the many hundreds, and we can talk about that. Aidan also has a Bachelor of Science in computer science from University of Toronto, has a PhD from University of Oxford, which he managed to do on the side while running Cohere. So for those of you who feel like life is overwhelming when you're in school, imagine doing that plus running a giant global company. He's also on the board of directors of Rivian, the electric car company, which is a very cool side job because he doesn't have enough else to do. Aidan spoke at our master class in 2022. So, Aidan, thank you for coming back, and welcome.
00:03:12 · Aidan Gomez, Jordan Jacobs
[00:03:12] S2: Yeah. No. Thank you for having me back. I think last time I told people to fire people much more. I don't know what kind of impression I left, but thank you for inviting me back. [00:03:24] S1: Yeah. That's okay. We've all forgotten that. And hopefully everyone's in sunnier mood. But let's jump into it. Okay. You co authored Attention is All You Need as an intern. Why did you... And you're you're working at Google. You can... You were working at Google in the Bay Area. You're working at Google in Toronto. What made you leave and decide to start Cohere? [00:03:50] S2: Yeah. I I think it [00:03:51] S1: was a
00:03:51 · Aidan Gomez
[00:03:51] S2: really unique period because we had gone from a world where, machines that could write compelling language were totally out of reach. Like, we could barely get them to spell a sentence properly and, certainly none of it made any sense to almost overnight one where there was pretty much entirely plausible human text that appeared in front of us. And I still remember getting this email from Łukasz Kaiser, who was my manager at Google. The subject was like, Aidan, look at this. And the body was a Wikipedia page. By the way, sorry. I'm I'm sick right now. So if I'm clearing my throat or coughing, I I apologize. Yeah. The body of the email was, a Wikipedia page, and the idol of the Wikipedia article was the Transformer. And then it talked about this, like, Japanese punk rock band that had gotten together, done a few albums, broke up. The basses got replaced. We got back together, did a tour. And then at the bottom of this email after this Wikipedia article on this Japanese punk rock band called The Transformer, Lukash had written, I just wrote The Transformer, and the machine wrote the rest. And that was a really jaw dropping moment. Like, now it's so quaint. Right? Like, of course, you know, you could open up one of 20 different apps and prompt that and, like, obviously, it does... And it will do it better. Right? But back then, it was like getting hit in the face with, something that was out of your dreams. Right? Something that was out of, like, a movie. Like, you just... The night before, you couldn't even by the end of my career, if we had machines that could do that, it would have been an extraordinarily productive life in machine learning. You know, what a contribution to humanity. And it just showed up in front of us. And so it was obvious that this was gonna be impactful, like machines had sort of woken up and could talk back to us. And so that opened up a totally new interface to interacting with machinery, computers. I didn't know how to make money off of that. I didn't know how to build a business around that, but I had a pretty good idea of how to do the tech, how to do the modeling. Right? Like, I knew how these models were built. The trend that seemed to be working was scaling them up, and that's what I was doing my PhD on. And so then when I went to go... For my birthday, I came back home to Toronto.
00:06:38 · Aidan Gomez, Jordan Jacobs
I was, like, you know, hanging out with friends. I was mostly there to party. But then I was gonna give a a talk at TD in layer six because that was, like, one of Toronto's primary ML groups, that was in industry. And I wanted to give a talk on my research to see if they were interested in it so that they might adopt some of this, like, efficient scaling, of neural net techniques that I I was developing. Presented that at layer six, met Tommy. Tommy told me, you know you know who you have to speak to, Jordan. And the next morning, Jordan and I met. And Jordan was like, do it. Let's figure it out. And then later that day, I went to hang out with Nick and Ivan, who are now my cofounders. And I said, like, I had this crazy meeting with this this guy, you know, Jordan Jacobs. Like, he he had... He he thinks this might work. He thinks this technology might work. Should we do it? And, you know, it took some convincing. It took a few months of work, I think, both on my part and and Jordan's part. But, eventually, we were like, alright. Let's quit Google and and do this. [00:07:56] S1: And so... Yeah. And and just to put it in reference, like, we're talking about, you know, almost eight years ago, more than seven Yes. Seven and a half years ago. [00:08:10] S2: At this point, it would be, like, seven and a half. Yeah.
00:08:10 · Aidan Gomez, Jordan Jacobs
[00:08:10] S2: At this point, it would be, like, seven and a half. Yeah. [00:08:13] S1: And so... And just so people understand the context, what's... That's, like, 3.5 before ChatGPT. So it is early. Right? It is... There's there's no such thing as anthropic. There is OpenAI, but they've not put a product out that anyone's really using at scale. [00:08:36] S2: They were doing the robotic hand solving a Rubik's cube, and they were doing Dota. Those were, like, their two their two bets. [00:08:43] S1: Exactly. The robotic hand became covariance, which we invested in, the robotics company, separately. Okay, so you guys decide to leave to do it. And it's true, Aidan was like all in early, and there required some convincing for I would say Nick was the most reticent. Nick's like, I have the most incredible job in the world. I worked for Jeff Hinton at Google Brain. And you know, what could be more amazing than this? So I remember some of those conversations. It was it was fun to try and convince him there was even brighter things ahead and bigger things ahead. In those conversations, there's a couple of things that were considerations that were you made decisions that I think were very consequential. One was we're going to focus on enterprise, not on consumer. And and the second was we're going to do this in Toronto. When you had been working in Mountain View, there's a possibility of doing it in London. Talk us through the decision of those two things. Why focus on enterprise and why do it in Toronto?
00:09:56 · Aidan Gomez
[00:09:56] S2: So enterprise, that was more because we didn't know the product to build. We started, like, well, like like Jordan just said, well before there was any product market fit. No one was buying large language models. No one was really buying models from a third party. They would build stuff internally inside their company, and it might be, like, re rankers for social media, feeds, or, like, yeah, like, recommender systems, basically. That was, like, the predominant actual industry deployment of these models, or Google would build models for itself for translate, that type of thing. And so there was no market. There's just like a totally... And I remember some of these, like, when we were starting to build this, like, how do you price it? How do you charge for this thing? And, like, you know, the unit of compute is a token, but no one's gonna understand what a token is. Should we charge based on characters? Should we charge, you know, like, or something? Like, have some completely different pricing model. It... There was no market. And so the initial bet was we don't know what product to build, and so we're gonna build a platform so that when someone figures out the right product to build, we'll be the place that they go. And that pointed us towards companies, right, and serving b two b. It was... Yeah. I I think it was, like, the only... I think if we had have gone very deep into consumer and we had made some, like, very hardcore product bet at that stage of the technology, we wouldn't be around right now. We we wouldn't be having this conversation. I I don't know if that's transferable advice. You know? I think a lot of people build on spaces where actually you should not build the most general thing, and that's not the fastest path to revenue, and that's not the best strategy. But if you're building on top of, like, an extremely new technology that has no users and no market, that's probably kind of what you're forced to do. And if you bet the whole farm on one specific product, you're you're gonna die with high likelihood. Maybe you'll, you know, get lucky. On being, like, very global from the outset, I think that was inspired by advice from you again, Jordan, which we were like, you know, the default advice, go to the Valley. Like, you're building a tech startup, do it in Silicon Valley.