Is Cohere the Next AI Powerhouse? | First Time Founders with Ed Elson

Thanks for having me.

So, we're a foundational model company and we are uniquely and singularly focused on the enterprise. So, there's about 10 companies in the world that can make foundational models. Foundational models, the large language models that are largely these days synonymous with AI. If somebody's talking about AI, they're probably talking about large language models. There's about 10 companies in the world that can make them. We are unique amongst them in our singular focus on the enterprise. So we make large language models that are good at the stuff that enterprises need them to be good at. We make them easy to deploy and efficient to deploy for our enterprises. We deploy them securely and privately so that we can't see the data that our customers are passing into the model that allows them to access the truly useful data out there. And we make them easy to work with via an agentic platform. So we do kind of the whole thing in order to get AI to work at work.

If you talk about AI companies, there's a lot of companies building stuff with AI. Mostly when people say they're building stuff with AI, they mean they're building stuff with foundational models. These days there's still a huge amount of work being done on more traditional machine learning, smaller systems. And a lot of those people working on that will still very rightfully call themselves an AI company. But if you are talking to someone and they say, 'Hey, I've got a startup and it's an AI startup or something.' Chances are they mean they're building off of a foundational model. They're making getting a large language model to do something useful for their customers. There's a relatively smaller number of companies that actually make the foundational models. So that actually make the large language models that take in a bunch of words and then predict the next words that should come next. Yeah, there's about 10 of those.

So in short, it's really hard and it's enormously resource-intensive. Building large language models is a lot more like building a rocket than it is like building other computer science projects. It requires a huge number of really smart people who have experience doing it, working in tight unison. There's a whole bunch of different things that need to go well in order for it to be successful. There's a whole bunch of experimentation that needs to get done and there's huge amounts of resources that need to get put into it in order to make the thing work, right? So, you have to get a huge amount of compute. So rent all those chips, you know, that you were talking about. You need to get a huge amount of data. You need to have a huge amount of people helping you create that data, getting data annotators. You need to have a whole bunch of really smart engineers working together in order to make it go well and even then it's still challenging. So yeah, so there's really only about 10 companies in the world that are doing it because of that reason. In the same way that there's not that many companies building rockets either, right?

Before co-founding Cohere with Aidan Gomez and Ivan Zhang, I was a researcher at Google Brain. So I worked with Jeff Hinton for a few years there working on explainability and adversarial examples and capsule networks and stuff. Which was really fun. And it was there that I met Aidan. And Aidan was just finishing up a stint in Google Brain in California where he worked on the paper 'Attention Is All You Need' which introduced the architecture that we still use today. So he helped write that paper in 2017 and you know almost 10 years later we're still using the same architecture. So after he worked on that and when I met him in Google Brain Toronto, he was obviously very excited about the architecture and about what it could do and he showed it to me and I thought it was also really, really exciting. So you know we noticed something about the nature of this new model that created an opportunity and indeed a need for companies to make foundational models. What we noticed was that for the first time in machine learning's history, if you wanted to solve a task like a language task, the best model to solve that task was not a model trained on that task alone. It was a model trained on a whole bunch of tasks. So that was really exciting and that made us realize, hey, there's going to be a need if companies are going to actually make this stuff useful and get this to work for them. There's going to be a need for companies to create really big and really good foundational models that other companies can use. So we had that realization in 2020. And we've been delivering on that since then. We've been trying to make language models useful for the enterprise by making them really effective at the things that they care about.

So I studied with him as an undergrad. So I only have an undergrad, I don't have a masters or a PhD or anything. So I had an undergrad from U of T and I did take his course while I was there and I sat in the front and asked lots of annoying questions. And I really only worked with him closely when I was at Google. So I was a researcher at Google Brain and I worked with him. I was working out of Waterloo for a little bit and then I found out that he was working in Toronto and then we started to work together and then I helped start up the Toronto Brain group with him and worked there for a few years and it's during those three years, four years that I learned most of what I know about research and machine learning and neural nets and I learned it from him. So I learned a huge amount from him, but I don't have a PhD or a masters or anything. As for his contribution, it really can't be understated. Neural nets have been an idea for a while. People have been thinking about a neural net architecture and in particular Jeff's been thinking about neural net architectures since the mid-80s. There was a long time where people thought they were not going to work and there was this whole wave of first perceptrons and that which is just a single layer neural net and people thought that was kind of interesting for a little bit and then some work came out to show that they had some fundamental flaws and that really cooled people down on them. People weren't excited about neural nets. And then people started working on multi-layer neural nets or multi-layer perceptrons and that solved some of the critiques but still people were not excited about it and they generally thought it was a bad idea and if they wanted to build AI they were much better doing things like search or symbolic reasoning or things like that. And so very few people worked on it and largely they thought it was dumb except for a few people, Jeff being one of them. So Jeff tirelessly worked on neural nets in the face of general ridicule for decades. For decades until around 2011, 2012 they were finally able to show that neural nets were suddenly the best at image recognition. That was the first thing that they really knocked out of the park. And that was done at U of T with a bunch of other brilliant U of T students. The reason we are where we are with neural nets in general which of course is the precursor to transformers, right? So there's kind of if you think of it broadly there's AI as a concept, there's machine learning as one strategy for doing that, neural nets as one strategy for machine learning, transformers as one type of neural net. It's kind of where we are. So the neural net part in particular, Jeff can claim a huge amount of responsibility for and it's really his tenacity and his dedication to continuing to work on it even when everybody else around him was saying no this is a bad idea it's not going to work that we have to thank for where we are today.

There have been other AI moments. There have been other times when the whole world's really thinking about AI. This is the first time that I would say it's been this dominant narrative of the economy for the past few years. And that's a first. And it's been the dominant narrative of technology for the past few years and it's been the dominant narrative of the economy even more for the past few years. So that's kind of a first. But there have been moments where people have been as really excited about AI and thinking that they're in some kind of AI moment before. You got to separate AI as a property versus any implementation trying to get at that property. So people have been thinking about artificial intelligence, like what happens if a machine has intelligence the way a person has intelligence, for a really long time. There's a myth that I cite pretty often that was written around the 1500s, 1400s I believe, a Yiddish myth about the Golem which talks about some rabbi imbuing intelligence into a clay man and then he asks the Golem to go get fish from the river and then he leaves his house for a little bit and when he comes back the house is filled with fish and the river is empty. And it's a joke, right? It's effectively a comedic story that's told at that moment. And the joke is, oh, intelligence is complicated and there's nuance in language and if we gave an artificial thing language, maybe it wouldn't understand that nuance. That's like a 500-year-old joke. So, people have been thinking about this for a really long time. More recently, after the computer was invented, there was a whole wave of people thinking about that. Alan Turing was thinking about the Turing test, thinking about intelligence. After that there was search. There was the Deep Blue moment when search algorithms beat Kasparov at chess and that had a similar moment. So people have been thinking about this all the time. This is different. This is a different moment and it's different in its scale. And when people write the history of AI, this is certainly going to be a pivotal moment. And I'm convinced that neural nets are certainly going to be a central component of machine learning and AI going forward. They're so good, they're so fantastic, they do all kinds of things that there's no other way we could get them to do yet. And transformers in particular, large language models are very easy to use for the average person. And that is I think really why this feels different. So if you look at the other moments when people were talking about AI like Deep Blue, let's look at that one as an example, right? There you can read tons of articles about people talking about what's happening with the machines or are computers getting as smart as people. They beat the best chess player in the world. Like what's going on? But if you're an average person, you couldn't really interact with that. Like maybe if you're good at chess, you could try the chess bots and people did. And actually, chess in some ways chess is more popular than it's ever been before. And in part that's because you can be at your home playing against something better than a grandmaster. But you could interact with it that way. You couldn't really interact with a search algorithm like an A* search algorithm in anything else. So your experience of it is pretty limited. Same with machine learning. Like when we made image recognition, the best image recognition model, suddenly, yeah, your phone, you could go on Google Photos and you could search up pictures of dogs and see all the pictures of dogs you've seen over the years. That's new. That's cool. But you couldn't, that's still directive. That's still like somebody made the model that does the thing. It's telling you how to use it. Transformers are the first time that any person without any experience in computer science or AI can go up to the model, open up a chat window, ask it to do something and it'll do it or it will not do it and that'll be interesting itself. But you can interact with it without it being prescriptive of how you interact. And that's I think the reason why this is suddenly so much bigger. It's suddenly so much more interesting, so much more widespread, and why it's become the dominant narrative of tech over the past few years. So when people write the history of AI, and I want to be clear that I think the history of AI is not done, I don't think transformers are going to get us to artificial general intelligence. So I think there's going to be more waves of new independent spontaneous inventions. I'm sure that's going to happen. But I'm convinced that the transformer is going to be a central component of that. And when the history of AI is written 100 years from now, a thousand years from now, this moment will be talked about as relevant and interesting and a moment when a lot of stuff happened really quickly as a result of the tenacity of a handful of people.

I think everything we're experiencing today is largely predictable from around 2020, 2019. Now that's not a coincidence. That's when I left Google to start Cohere with Aidan and Ivan. So that's the reason why I think it was largely predictable around that time is because that's when I predicted it. So I'm sure other people predicted it before but I got on board at that time. At the time I think when I remember telling people I'm going to leave to go create this foundational model. I don't think we used the word foundational model. We just said large language model company. We're going to be a large language model company. We're going to make large language models. I remember everybody saying yeah that's probably a good idea. I don't think anybody was thinking like, oh, that makes no sense. The question was not like, the question was like, oh, is Google just going to do it? Are the other big companies just going to do it? But I think at the time it made sense. Now, it really still wasn't popular. And when we had conversations for the first few years of Cohere's history, the conversation was this is a large language model and here's why we think it can help you. The conversation now is okay cool like why your large language model or like how can this actually help me get into production? How can I have it access my private data without giving that away? Like how can I deploy it in a secure and safe way so that I can handle regulated industries? How can I connect it to my specific data in an enterprise? Those are all the questions we answer now. So it's changed a lot and what changed in particular and a thing I did not predict at the time was the success of chat fine-tuning. So you train this big generic language model. When language models were first created what they did was they just completed the ends of sentences because they were trained on the web. So you really can think of it as a web, they were calling it a large language model but at the time it wasn't a large language model. It was a web text model.

Yeah. So, if you wrote the first part of a website, it would write the second part of a website. Not even the HTML, just the text on the website. And you could do a lot of stuff with that, but it was confusing and weird. And then OpenAI and a few other companies at the same time fine-tuned that large language model on chat dialogue. And that suddenly, suddenly people understood it. And at the time actually I remember thinking I was surprised at how efficient that was because when you think about it you're training a model on the entirety of the web. So a huge amount of language and then you fine-tune it on a relatively small amount of chat and yet actually it learns how to chat pretty well. So that is I think responsible for the difference between 2020 and 2022. It was the data efficiency of chat fine-tuning that allowed people to, like for the model to meet them where they're at. They kind of expected, users kind of expected chat to work when you told them it was a large language model and it didn't. It was this like weird text thing. So then making it work in the way they expected it to work seems to have really woken people up to the effect and the utility of these models.

Let me first talk a little bit about the way we train these models. So the first step is to train them on everything on the web, everything on the open web. So you create a data set of all the text that's available for training from the web and that turns out to be a huge amount of text orders of magnitude more text than you will ever read. Like a thousand people a thousand years reading 24 hours a day volumes of text, that's how much text. So first step is train on that. Then you make a data set with people. So you have people create, like talk to the model and if the model gives a good response they say that's great. If it gives a bad response they say that's bad and they write what the model should have said. If you do that process you'll create both ratings like is it a good response or a bad response and you'll also create what's called supervised fine-tuning data, SFT data. So that's like here's the input to the model and here's a gold standard of what a person wanted, like they wrote out the sentence like that's what the model should have said. That's called SFT data. So then you train the model on that SFT data. After that you can do reinforcement learning which was a type of machine learning invented before transformers where you're training a model without access to the right answer. The model kind of tries stuff and then you say hey this was better or this was worse and you update the weights of the model based on that signal. So then you can do reinforcement learning. Now we do a whole bunch of reinforcement learning with synthetic data. So now we use the model itself to generate data and then do reinforcement learning on that synthetic data. So that's a big component of training now. So there's the data you get from the web, the data you make with people, and then the data you make with the model itself. And all of those are super relevant for making the models that people use today. Your question about models being restricted to the web and missing the stuff in the real world. Is that a blocker to AGI? Like, yeah, definitely that's a blocker to AGI. If when you say AGI, you mean humanlike intelligence. Yes, that's a blocker to AGI. We are embodied creatures. We learn our intelligence through interactions with the real world and intervention into the real world. There's lots of interesting psychological work that suggests learning and interaction are super related. So, interaction is super important. Is that a blocker to AGI? Like, yeah, definitely. But I don't, there's a whole bunch of blockers to AGI and that's just one of them. And the technology as it exists today is massively impactful, massively useful, absolutely transformative on the nature of computers and subsequently the nature of work, massively transformative on the economy in general, then I don't think it's AGI and nor do I think the transformer alone will get us to AGI. Nor do I care. I don't really look out in the world and say, 'Oh, geez, I wish my computer was a person,' right? I look out in the world and I say, 'Oh man, there's so much stuff that a computer should be doing and not me. My time should be free to think strategically, to think creatively. There's so much work that a large language model when connected into the things that I am using can do for me and subsequently allow me to do the interesting and the human and that's what I want to make. I want to make a technology that does that as good as possible.'

Yeah. Yeah. I mean, you said, yeah, look, high priest is a good term. A lot of the thought around AGI and discussion around AGI feels religious to me. It's calmed down a little bit, right? Like if we back in 2023, 2024, my views on this were a little heretical. People would disagree. If I said, 'Hey, AGI's probably not around the corner.' People would disagree and say, 'Why do you think that?' I would get a lot of push back. I don't get much push back these days. I'm like, 'Yeah, guys, guys, guys, we know. Transformers. Incredible. Super awesome. Super good. Can definitely be way better than they are. Need to be deployed correctly. Need to have lots of stuff, you know, to get them into production. That's what we focus on, but AGI like no.' And everybody's like, 'Yeah, yeah, yeah, totally. I get it.' And if you use a large language model, which as everybody does these days, you'll feel that pretty soon. You'll be like, 'Yeah, they're amazing at these things.' And then I ask them some other things and they don't understand at all. Completely different and they have a completely different...

It's very different talking to a language model as it is chatting to a person. People kind of know that when they're grounded in an environment, the focus on it is, I think, a narrative device more so than it is a scientific belief.

The big difference is we are not a consumer company and we're only an enterprise company. So we don't have, we can't pay $20 a month to get access to our tech. We're not trying to build a product that people use in their personal lives. We are instead selling only to large and medium enterprise companies and we create language models and search models and an agentic framework for using them that is tailored to the needs of those companies.

That strategic difference comes from a philosophical difference which is a different view on the technology, right? Like I don't think the technology is going to get us to AGI and I don't think the biggest utility of the models is in people's personal lives. I think the biggest utility of these models is in work. I think that like their ability to augment and automate work at a desk behind a computer is what they are the best at. And so we have that different view of the technology that leads us to think differently about where we can add the most value to the world and that leads us to being a singularly enterprise focused company and that is as mentioned unique amongst the foundational model players.