There's been a sort of slowdown in growth for these developed democratic economies. When that starts to taper off, you see the emergence of things like xenophobia, like that immigrant is taking my opportunity. They're taking my slice of the pie. A lot of people talk about like AI destroying the world. I really don't think AI is going to destroy the world. It might be like our best shot at saving the world.
Welcome to the Upstarts podcast, our weekly show where we talk to emerging startup leaders about their Upstart moment. Upstarts are challengers who punch above their weight and take on the status quo to improve the world, all while building a big business, too. I'm your host, Alex Conrad, founder and editor of Upstarts Media. And I'm delighted to be joined today by Aidan Gomez, co-founder and CEO of Cohere. Thanks for joining, Aidan.
This podcast is brought to you by Mercury Banking Redesign from the ground up. Aidan, you have run Cohere now for six and a half years. You guys have raised 1.7 billion dollars. You work with some big governments, some big companies. When you think about sort of what Cohere looks like today, is it early innings? Is it everything you hoped for six years ago?
I think we're in the takeoff phase. So, we've been around for a while, but this past year was really the breakout year for us in terms of enterprises just adopting AI at scale across the entire company and the importance of all the things that Cohere is good at really sort of hitting the front of mind of executives at companies all over the world. So, privacy, sovereignty, and this focus on critical industries. And so last year all of that came together and we sort of started our takeoff and then this year we expect it to get even bigger as enterprises and the broader global economy really truly adopts AI at scale.
So with some of these big companies or G7 countries that are working with Cohere, what is the problem that they want you to solve or where is your technology actually helping them in sort of a simple overview?
Yeah. Well, it's super broad because we're working with a really diverse range of customers from different industries from like manufacturing to public sector to financial services, energy, and it's as simple as like, you know, get the model to every morning at 6:00 a.m. check my unread emails and send me a text with a summary of like the most important ones that I need to know. So that's like a super general, you know, any company would want that to a hyper specific, you know, within a specific ship. I'm going to plug into the telemetry within that ship, do continuous monitoring on different sensors on board and build automations around if sensors cross certain thresholds, what to do, who to notify, what sort of remediations to take.
I know it's a wide range of use cases, but generally speaking, why often do they need Cohere to exist? Like why is that an unsolved problem for them?
Yeah, I think it's the privacy piece. I think the secure deployment of this platform is something that our customers couldn't do without. It's something that is a hard requirement in their settings. They can't send data out into a cloud. They're highly regulated. They're in these high security settings. And so they need something that's deployed on prem or even airgapped on their infrastructure. There's a deployment level differentiation and the other piece is within the product itself. It's super easy to use. It's very customizable. There's tons of like administrative enterprise controls that are unique to the North platform.
So should we think of Cohere as sort of the base layer or you know this infrastructure layer that then the company is either building internal apps on top of or it's running these agents doing sort of all sorts of other work with sort of the info that your horsepower is providing or what would be the best way to sort of visualize this?
Yeah, I think it's infrastructure for these high security enterprises to adopt agents and integrate it into their secure systems. That's what we do at Cohere. Obviously, we're internal customers of our own product and many of the best automations have been created by or like discovered by folks who don't code and they just, you know, as part of their day-to-day work, they're like, 'Oh, I could totally use North for this.' They build an automation and they share it with people. And so that's what we want to be able to enable is like people just discover killer applications of the system and then it goes viral within an organization and spreads out to all these different teams and North really enables that.
We are in a moment where we're being told you know you can vibe code anything and there are you know some really successful companies that are helping people come up with an app out of an idea. Is the difference here that we couldn't really run a Bolt or a Lovable or one of these tools internally in a sort of sensitive or secure environment with all our data. So using North is just sort of much more viable or what would be sort of the distinction there?
Yeah, I think the vibe coding stuff is very helpful for technical folks who want to get 90% of the way there or even 95% of the way there. It can carry you quite far, but there's still a last mile that you need to close. And often times you do need to be technical to close that last mile. And for North, you don't need to be technical to do that work. You have a sort of like builder view and you can just chat to the model and there's no need for you to be a developer yourself to do the work. So I think that's one thing. The other thing is that for a lot of these applications, you don't need to create a product. You don't need to create a whole new piece of software to solve it. All you really need is like it to have access to the right tools with the right authentication and then you need to be able to build a series of steps that you want it to accomplish. And that's a much simpler problem. And there's an easier way to solve that rather than for each one of these that I want to create, I'm going to have to create an entirely new codebase and program and ask an agent coding agent to go build it for me. So I think it's a more simple and elegant solution to the problem.
Very cool. There's an interesting question of are startups reflecting the personalities or sort of backgrounds preferences of their founders or is it vice versa? And I'm curious is this the kind of business that you would have always imagined running or building?
Yeah, I don't know. I don't know. I definitely like appreciate my own privacy as an individual, but I try to be open. I think like for people to trust you, they have to know you. And so I talk about, you know, where I grew up, how I grew up, my family. So I try to be somewhat open, but I'm certainly I'm not the type I'm not a great marketing or sales guy. Like I'm not out there screaming, you know, my own name or Cohere's name constantly. My background is like I'm an academic, right? Like I came into this through research. And I think there's a very different type of personality that thrives in research than in company building. And I've been, you know, doing my best to adjust to being a CEO and having to be spokesperson and a salesperson for the company. But you can't run from your nature, which is like I'm a, you know, shy nerd. So that doesn't go anywhere. Very tall.
It's tall shy nerd. Yeah. Got it. Well, so when you were in academia, you were privileged or you had earned your way to be at sort of the epicenter of some of the earliest work with LLMs, you know, the famous transformer paper. Could you talk a little bit about how you first, you know, got into this field, why it was exciting to you, and sort of how the journey really starts there?
Yeah, I grew up in the middle of the woods in Canada without like any internet. At some point we got like a satellite dish cuz before that it was a dialup but this was like pre-Starlink and so it was like these like geostationary like if there's a cloud cuz it has to be like super high up. If there's a cloud you're screwed like goodbye internet. And I think that scarcity from technology like it was just so hard to use my computer and like I couldn't watch YouTube. It was impossible. I can like watch pixels load one by one on an image. That scarcity made it super appealing to me. It was like just out of reach and like you know my friends were gaming online and like doing all this cool stuff and when I went home you know we had a computer in the house of course but it wasn't the experience everyone else was getting. So that made me deeply appreciate computers and be very excited about them, invested in them. I was always trying to make my computer faster at home or like let me access the things that other people were accessing through my limited connection.
Was it working? Were you hacking a faster computer?
I mean it got me into like webdev and that type of thing. I wanted to understand why this web page wouldn't load faster. And so you know I inspect element blah blah blah but yeah that's what got me into coding like first it was just like going through websites and trying to understand all this text on my screen and then I took a course because I was interested in that then I started a little small business in my tiny little town of Brighton Ontario building websites for the little businesses. Like there was a really nice lady who had a knitting store and you know yarn and patterns and stuff and so I helped her build like an online catalog for her customers and you know she's been an incredible customer of mine but I started that little business and then I just decided this is what I want to do so I went into computer science at UofT that was my local university right like Toronto was nearest big city and I just happened to go to university at like the epicenter of deep learning. Like Jeff Hinton was there, the guy who won the Nobel Prize and the Turing Award for his work creating deep learning which is like the thing that worked. It was like the breakthrough system that let us reach where we are today.
And were you immediately like this is an area that I want to throw myself into or was it more?
I was obsessed. I just remember every waking minute of my day thinking about it. I would walk around with like a stack of papers because and it would initially to read one paper in the field would take me like a month actually because I would like be going through and I'd run into like a word that I don't understand and then I would have to go start the process of you know researching what the hell is that you know like what's Bayesian or what's a Gaussian.
You didn't have LLM to shortcut.
No, no, it's just Wikipedia pages and like research papers and then gradually I just got more and more familiar and comfortable with the material and I went to a startup in Vancouver where I worked on it's called polyphonic transcription which is like listening to music and transcribing music into sheet music.
So you listen to someone playing and you write down the notes that you're hearing.
Super difficult problem at the time but we used neural networks to approach it and got some pretty good results and I had to this was before JAX before TensorFlow before PyTorch before any of these frameworks really existed and I had to write kernels in Metal which was like Apple's GPU language for forward and backward propagation. So I just had to learn everything at like the lowest level imaginable. You're like, you know, basically writing machine code. Not really, but you know, just super low level, stupidly low level. And then I came back to Toronto and I messaged Hinton and I was just like, 'Hey, I have this idea. This activation function that you've made, I think there's a better one.' And he actually responded and then he sent me in the direction of Roger Gross and others inside the UofT ML group and from there started doing work with them eventually ended up at Google down here in Mountain View and was on the transformer paper and the rest is kind of history. Then I went off to Oxford. I met my co-founders when I came back from the valley to Toronto, ran into Nick and Ivan. Nick was working for Jeff at Google was like Jeff's first report at Google. And then I went off to Oxford to do my PhD. And then from Oxford I called up Nick and Ivan and just said, 'Guys, we have to start something.' I saw the first few results of like language model scaling. And it was very much this zero to one, you know, we could barely get these models to string a sentence together to this was like plausible, fluent, ostensibly human written text coming out of a machine. And now we take that for granted. It's so normal. It's like this is kind of not surprising. But back then it was like it was surreal reading a machine. I think people sort of take for granted the fact that like we're talking to some rocks and metal. Like we're literally talking to our pet rock at this stage, which is it's so insane. Yeah, it's normalized. We do it all every day, but good god, like back then it was just mind-blowing. And so, like, what else could you want to work on?
We talk about the Transformer paper, which was called Attention Is All You Need. I don't know what the best comp would be. Maybe the Magna Carta of AI research, very seminal paper that you were part of. For folks who are not deep in AI world, what would be the TLDR of what this paper was?
Yeah. Well, it was done when I was at Google and in particular was done as a collaboration between Google Brain which was like the AI research department under Jeff Dean and Google Translate and the two teams came together and really the objective was to improve translation, improve the state-of-the-art of translation. The architecture itself was designed to better and more scalably model language. And what I mean by that is, you know, language is a sequence. It's a sequence of characters. It's a sequence of words in order. The order matters. And you want to be able to model that or represent that data in the most efficient way possible. So the transformer was that was the problem we were trying to solve. How do we just make this a more efficient model for language? Turns out the transformer became extremely effective beyond language and you can phrase many different data modalities like video etc. as a sequence. The fact that the model was so efficient made it a very good option as we were going through this process of scaling up models and they're called large language models because they're really big. The transformer as an architecture just scaled better than the alternatives because it was so efficient. And so as we went through this process of the field building bigger and bigger models, training on larger and larger computing clusters, it was the best thing available. And so people poured effort into optimizing chips for it and coming up with better libraries for training these models. And then over time the transformer just kind of took over everything. It took over the domain of images, took over the domain of audio, it took over pretty much the whole field, but it certainly was not designed for that. It was just designed for this one narrow problem of like we want to make translation 3% better.
And now people are saying what comes after the transformer.
No, I've been hearing that for a long time now. Yeah. I don't know. Like I've gotten so excited before. I thought SSM state space models which was like this thing that came out three or four years ago. I thought that was going to kill the transformer. I get excited discreetly.
But would that be like killing your own baby if something...
Not at all? No. Like if you're a researcher you really hope humanity comes up with something better than what you've created otherwise progress has ended.
