Found: Getting realistic about AI’s potential with Nick Frosst from Cohere

Enterprise AI is booming so it’s no wonder that, as companies figure out how to implement it, the industry of AI infrastructure is emerging. This week, Equity is bringing you an episode of our sister show, Found. Becca and Dom talk to Nick Frosst from Cohere, the AI company building natural language models for enterprise customers. They discuss why Frosst thinks the AI boom isn’t built on a bubble, whether or not AI companies are building toward a “digital god”, and how AI regulation could be a good thing. Equity is TechCrunch’s flagship podcast, produced by Theresa Loconsolo, and posts every Wednesday and Friday. Subscribe to us on Apple Podcasts, Overcast, Spotify and all the casts. You also can follow Equity on X and Threads, at @EquityPod. For the full episode transcript, for those who prefer reading over listening, check out our full archive of episodes over at Simplecast. Credits: Equity is produced by Theresa Loconsolo with editing by Kell. Bryce Durbin is our Illustrator. We'd also like to thank the audience development team and Henry Pickavet, who manages TechCrunch audio products. Learn more about your ad choices. Visit megaphone.fm/adchoices

• 0:02Speaker 00

  Enterprise AI is booming, so it's no wonder that as companies figure out how to implement it, the industry of AI infrastructure is emerging. Today's guest is helping customers gain access to AI without having to build and store their own models.

• 0:17Speaker 00

  You're listening to Found, TechCrunch's podcast that brings you the stories behind the startups. Today, we're talking to Nick Frost from Cohere, the AI company building natural language models for enterprise customers. I'm Becca Scutac, and here with me is my fabulous co-host, Dominic Midori Davis. And of course, before we get into our conversation with Nick, we have our two truths and a lie. And at the end, we'll tell you which one of these isn't true. So listen carefully. It's a lie that, over a decade ago, Nick thought he was late to the AI game. He thinks AI is on its way to becoming a digital god, or that Cohere is valued at over $5 billion.

• 0:47Speaker 01

  Ooh, listeners, digital gods, being late to the AI game, you're definitely going to want to keep listening to figure out which one of these statements is true and which is not. But before we dive into our conversation with Nick, a quick reminder to please rate and review the show. You can do so wherever you are already listening to this episode, and it helps us out a lot, so we greatly appreciate it. But without further ado, here's our conversation with Nick.

• 1:15Speaker 01

  Hey, Nick, how's it going?

• 1:17Speaker 02

  It's great. How are you?

• 1:18Speaker 01

  Oh, can't complain. It's not a million degrees in New York for once this week, so, I mean, hang it in there. But we are not here to talk about the weather. We are here to talk about Cohere. So why don't you maybe start by telling us a little bit about that?

• 1:31Speaker 02

  Sure, yeah. So I'm Nick, a co-founder of Cohere. And Cohere is an enterprise-focused, large language modeling company. So we exist to try to make AI models useful for enterprise. We started in 2020 and have been doing that since.

• 1:47Speaker 01

  Of course, some of the large players in the AI space, like OpenAI, have not been started since this current wave of interest over the last couple of years. You guys, of course, would fall into that camp as well. But I was looking back at some of the stuff you've done prior to Cohere, and it seems like you've been interested in the AI space for quite a while now. Like, what pushed you toward this area?

• 2:05Speaker 02

  It's hard to not be interested in AI, I'll be honest. I've been interested in AI since right when I started undergrad at University of Toronto. That's where I first met Jeff Hinton, who's one of the founders of this technology. It's very difficult to not be interested in AI. I think it's just exciting to push computers and get them to do things you didn't think they could do. And so I've been interested in neural networks in particular for the past decade, I think, or more now.

• 2:31Speaker 01

  And why neural networks?

• 2:33Speaker 02

  So when I first got started into neural networks, it was right around 2013, I think, 2012. And I was first interested in them because there was a whole bunch of things that you just think computers can't really do at all. At the time, understanding like computer vision, that was a thing we just assumed computers would never really be able to do. And then out of Jeff Hinton's lab in Toronto, they took a neural net, gave it a whole bunch of examples of images and labels for those images. And then without much complexity, it was able to learn how to recognize cats and dogs and cars and things in images. That, looking back, seems almost quaint.

• 3:10Speaker 01

  Right.

• 3:10Speaker 02

  But at the time, was wild. I remember being like, this is crazy. How can you possibly do this? And that had come off the heels of decades of people trying to get this to work in other approaches and just failing completely. It was really exciting. I remember in about 2013, like, learning about this and thinking to myself, oh, man, if only I was here a few years earlier. I've kind of missed the boat on this whole neural net thing. I'm kind of just like a little too late to catch the real excitement. It's always funny to think back to that.

• 3:37Speaker 01

  That is funny. I mean, of course, people were doing research in this area prior to what we're seeing now in venture, obviously. But I was gonna say, you were into AI before it was cool, so it's really interesting to hear that you yourself kind of thought you were late.

• 3:49Speaker 02

  Yeah, no, I totally thought I had missed the boat. I remember looking and being like, wow, if only I had started undergrad like a few years earlier. I could have been in on the ground floor.

• 3:57Speaker 01

  That's too funny. Yeah, it's because I know you then went on to work at Google. And I definitely wanted to ask you a little bit more about why it made sense for you to leave Google and start Cohere. What did you realize you wanted to do that was different than what you could do with what you were working on there?

• 4:13Speaker 02

  Absolutely. Cohere was started based on a technical realization that our co-founder and CEO, Aiden Gomez, had. So Aiden was an intern, actually, working at Google Brain and was involved in a seminal paper that kind of kicked off this whole thing. The paper is called Attention Is All You Need, then it introduced the transformer. The transformer is a type of neural net that happens to be particularly good at text. So after working on that paper, he kind of had this realization that we had entered a moment when a general-purpose model was going to be better than a specific model. So I was talking about image recognition earlier. In 2015, if you wanted to make the best image recognition model for recognizing cats, you should train a model only to recognize cats.

• 4:57Speaker 02

  That was like the best one for recognizing cats was going to be trained to do only that. After the transformer and in the space of language, it was clear that if you wanted a model that was amazing at extracting numbers from a PDF or summarizing a document or automating some process or drafting documents, you were going to be best served by a general-purpose language model that was trained on as much language as possible with as much feedback from people, like all the best language data in general. And that general-purpose model would be better than a model that you train specifically to do that. That's a very interesting technological discovery. That was a totally different paradigm for machine learning models than before.

• 5:33Speaker 01

  Why is that? Just before we move on, like, why would that be a better system?

• 5:37Speaker 02

  No, it's interesting. I think it's because language is actually much more diverse than you might imagine. So unlike recognizing cats, you really only need to know what a cat looks like. And everything else is not a cat. It's either cat or not cat, if that's what you're going for. You just need a whole bunch of examples of cats and a whole bunch of examples of not cats. And that kind of gives you everything you need to know for a machine learning model. Language, if you think your task is specific, you know, let's make it really grounded. I'm going to extract revenue numbers from quarterly earnings reports in PDFs.

• 6:11Speaker 02

  Like it's a really tailored thing. Prior to transformers, people would write, you know, like heuristics or they'd write parsers or they'd try to say like, oh, the numbers often have three digits and then a period or something like that. They'd write these little heuristics to try to do that and that would work a little bit. But if you look at the wide space of the way even something as restrictive as quarterly earnings reports are written, you kind of have to know language in general to be able to do it. We specialize as people, we go into particular fields, right? But the first part of our education is just about how to use language in general. We spend a long time learning language, like you do elementary school and learn how to read and write, and then you do high school and learn language in general, and then undergrad and it's general. It's not until very later that you kind of specify on a particular subfield of language. So there's something kind of similar going on with neural nets

• 6:56Speaker 02

  as well. So yeah, so that was a cool scientific discovery and that gave Aiden the idea of saying, hey, you know, this is an opportunity and indeed a need for a company to be focused on making these models solve problems for enterprise. And we can do that by training a general purpose language model and giving companies access to it.

• 7:14Speaker 00

  That's really interesting. That reminds me of a lot of my college linguistics class when they were telling us that our brains are computers and language was like the code and there are like certain ways to like phrase things so that we can understand and process that. So I guess I never really thought about the fact that of course, when you're training AI models with language, like there would be, I guess, like a linguistic aspect to it.

• 7:33Speaker 02

  Oh yeah, there's a huge link. Well, actually, yes and no. So on the one hand, it's very important to have a good understanding of linguistics when creating language models. On the other hand, ultimately, these models are just statistical models of sequences and you can train it on one language as exactly as easily as you can train it on another language and it doesn't require specific understanding of the grammar or the the structure of the language you're training it on.

• 7:56Speaker 00

  Just one last follow up with the leaving Google piece.

• 7:59Speaker 02

  Hmm.

• 8:00Speaker 00

  Why couldn't you guys do this model there?

• 8:03Speaker 01

  Like, why did you guys have to start another company? Especially if your co-founder had this like great realization like wouldn't Google be all over that? Or why did you guys decide to sort of break off on your own?

• 8:11Speaker 02

  Yeah, we both really enjoyed our time at Google. It's really where I like learned how to be a researcher. And my boss at the time is the guy I credit for kind of teaching me pretty much everything I know about research and machine learning in general. But I think there's a real need for a company like ours to be independent, to be not beholden to any of the cloud providers and to be focused on this and this alone. So we're not trying to make a consumer product like some of our competitors. We're not building a thousand different things at once. We're trying to make language models really useful for enterprise. And that singular focus is not something you can have building inside of a multinational massive corporation.

• 8:53Speaker 00

  I imagine that you probably have seen this field change throughout the past years. Cause I imagine like in 2017 or like 2019, everything seems so different. So I guess I have two questions here, which is like, what is happening on the technological front? Cause AI is not necessarily new, but it seems like there's just been this massive explosion in the past few years of advancement and innovation. What is the technology? What is happening that this is happening now?

• 9:21Speaker 02

  That's a great question, and I think historically a very interesting answer. The technology has not changed a lot in the past few years. From a technological point of view, we're still using the transformer architecture, which Aidan and his co-authors wrote in 2017. We're still training models via getting a whole bunch of language data and training models to read a few words and then predict the next word. That's the same. The scale is different, so we're training significantly bigger neural nets and we're training them on significantly more data and we're training them with significantly more compute. So the scale is really different and that's brought a whole bunch of technological challenges with it. But functionally, it's still a pretty similar technology.

• 10:04Speaker 02

  What caused this like massive public perception change that happened around the release of ChatGPT? It's not a real technological change. But data change. So for the years before ChatGPT, the foundational model companies, of which there were and are a few, were training models on text we could get from the open web. So just training the model on just all text in general. And that meant that you could write the first part of a sentence, and it could do a pretty good job of guessing the second part of the sentence. It could write a first paragraph, and it could do a pretty good job of writing the next paragraph. But if you tried to say something like, "Write me a poem", it wouldn't do that because we're training it to predict the end of sentences that it found on the web.

• 10:49Speaker 02

  And people don't really write on the web, "Write me a poem" and then "write a poem".

• 10:53Speaker 01

  Right.

• 10:53Speaker 02

  They don't write like a declarative. They might say something like, "The following is a poem I wrote. I hope you enjoy it", new line, and then a poem. But they don't say, "Write me a poem, period", and then a poem. So the model couldn't take that request from a user. So the swap to chat models was about fine tuning these general purpose text models, so training them a little longer on data created by people that looks like how people want to interact with the chat model. So now, we train the model on a whole bunch of text from the web, and then we have a whole bunch of people make data where they say, like, "This is the way the bot should respond", and we train the model on that. And so it got a lot better and a lot easier to use. And that's what really excited people about ChatGPT, and that's what awoke them to the possibility of this technology.

• 11:40Speaker 01

  More from this conversation right after a quick break.

• 11:46Speaker 01

  Because you guys had been building already for a few years before things really started to explode in this space, what was it like for you guys going from being a space where, I don't know if you guys had trouble fundraising, but I definitely remember reading the reports about AI funding being down in like 2019 and things like that. Like, times were just very, very different. And what is it like for you guys being in, from the beginning, not in 2013, like you wish you had been, but being in before things really started to swell and like how things have changed since then?

• 12:12Speaker 02

  Yeah, for us, things have changed a lot. All the conversations we used to have were, what is a language model and why does my company want one? And we'd be like, well, you can do all kinds of cool things with it. It's super useful. And that would be a really long conversation. Now, the conversations are, how is your large language model good for us in particular? And we can talk about our focus on data security, on privacy. We can talk about our multilingual capabilities, or models really good at more than English. We can talk about our model's accuracy and how we reduce hallucinations. And we can talk about all the things that make our model in particular good. But we spend no time convincing people that a large language model is useful for their business anymore. That's kind of a given. Now, it's like, what can you guys do in particular?

• 12:48Speaker 01

  That's so interesting about the language piece of it, because I did notice that when I was going through your site earlier, that it occurred to me that I had not really heard much from any of the other AI companies regarding doing things outside of English. LinkedIn notified me about a whole slew of training positions yesterday that opened up at this AI company that's looking to train in other languages. But it has occurred to me that you don't come across this very often. Why is that the case? Why was it important for you guys to add those pieces in when you were building?

• 13:16Speaker 02

  Yeah, so we've been thinking about multilingual and how to make the model good at more than just English, like for a while now. And I think that's a technical differentiator between us and our competitors that stems from a philosophical one. So we are not an AGI company. We're a company trying to make language models useful for enterprises. That's what we're trying to do. So that leads us to invest in things that are not hinting at getting closer to AGI, but that are instead just making them useful for practical applications. We recently announced a partnership with Fujitsu, which is a leading tech company in Japan. And, you know, they've been looking around for a good model in Japanese, and working with us, we're able to build one that's great for them. And that's because we've been investing in making Japanese data for a long time.

• 13:59Speaker 02

  Japan is like one of the largest economies in the world with a vibrant developer community that speaks very little English. So it's been underserved for a while. But we've been thinking about how can we make this useful? Because that's what we're thinking about all the time. We're not

• 14:12Speaker 01

  Right.

• 14:12Speaker 02

  chasing long things. We're just like talking to enterprises, figuring out what's slowing them down and making the model better at that.

• 14:18Speaker 00

  Are there some languages that are harder than others to kind of make a model in?

• 14:22Speaker 02

  There are, but that's because of the prevalence of that language on the web. So if I had, you know, the same amount of language in English for any other language, then all languages would be exactly as easy. And that's because we take that data, we train the model to predict the next word from a sequence of words in that language. And no language would be any harder than any other one. But the web is actually massively overrepresented in English. So if you look at the distribution of languages spoken in the world, it's very different than the distribution of languages written on the web. And so that makes getting some languages harder than others.

• 14:57Speaker 01

  Oh, that's interesting. Of course, you guys are working with enterprises. That's who your main customer is. But I know like sort of the broader conversation around AI startups lately has been the fact that so many are trying to sell into enterprises, but they're having trouble or they're finding that what they're building isn't actually something enterprises want to use or vice versa. There's a lot of interesting conversations happening around that currently. We also had Glean on a few months ago and they were talking a little bit similarly about like the changes of what it was like selling to enterprises as that being their focus as well. I know you mentioned you don't have to explain to people why they need it anymore, but what has it been like selling into this group? When it does seem like you still see stuff all the time that companies are struggling to do that?

• 15:37Speaker 02

  I think there's been a change this year and between this year and last year. Last year was everybody being really excited about the tech and just trying to build a demo and just trying to get anything out there to be like, oh, hey, look, cool, we're doing this cool thing because everybody's so excited about it. This year, there's a lot more focus on like actually going to production, actually saying, okay, well, how can we make something that's actually going to speed up our employees or help our customers or something like that? And so the nature of the conversation has changed a lot. These days, we spend a lot more time talking about, you know, total cost of deployment. We spend time talking about privacy and deployment options. Like we'll deploy our model wherever we'll deploy it on a, on a virtual private cloud or on prem or ever. And those are all things that you care about when you're interested in actually going to production. So recently, we've announced a few like customer stories,

• 16:20Speaker 02

  but one that came out recently that I think is a great example of this is a financial company uses our model now to extract. Yeah, actually, well, I said this earlier, to extract numbers from financial PDX. And in doing so, they were able to speed up their total like analysis time by 50% or something. Like, that's a great example. That's like automating something that is tedious and boring and no one really wants to do. In order to do that, we had to make the model really good at those boring things. We had to figure out the right deployment options, had to do all kinds of things that don't really have a lot to do with AGI or with neural nets even. Some of them are just figuring out how to make the right thing for a business. But this year, those are the conversations we're having. And that makes me really like optimistic and excited because we're now focused on tangible, actual value that these things can deliver.

• 17:05Speaker 01

  Yeah, and I wanted to ask you guys a little bit about funding too, because of course you guys just announced last week this big funding round that you guys were able to secure at a very nice valuation, if you were to ask me. So I'm curious, like, what has fundraising been like for this company? Going back to what we talked about earlier, landscape has definitely changed. Like, how was it raising this round compared to the first few rounds you guys went out to raise?

• 17:26Speaker 02

  Yeah, we're really pleased with the outcome of that funding round. We raised a lot and that really sets us up to continue building our product and building something that's useful for enterprises. This is a resource-intensive industry, right? I mentioned that we're at a different scale now in terms of data and compute and talent. So it's a resource-intensive industry and the funding that we raised really allows us to keep executing on our mission and make this stuff useful. So I'm really pleased about it. How is it different? I think that message of we're interested in enterprise solutions and like trying to make this useful really resonates with both customers and investors. Over the past year, the hype has gotten a little deafening. And I think that our focus on the tangible and the practical and the really valuable is something people are excited about being part of.

• 18:11Speaker 01

  Do you think that this is the AI bubble that a lot of people are talking about?

• 18:15Speaker 02

  I don't think it's a bubble, but I don't think we're close to AGI. So when I say I don't think it's a bubble, I mean, like, that's because I see the real value that our models add. Like, I see the value that we're providing for customers already, and I think we can add so much more than we're already doing. Frequently, I'll run into something where I'll see somebody using our model and they will have enabled some completely new feature that wasn't possible before. Or they'll have automated some process that was really bogging them down and slowing everything up. And like, that's tangible value. It's hard for there to be a complete bubble when you have something so useful adding value already. Now, I think we can add so much more value. Like, there's so many companies that are not doing stuff with LLMs and are wasting time and resources as a result. So I think we can add a lot more value than we're currently adding.

• 18:59Speaker 02

  I don't think we're going to get the artificial general intelligence. I don't think we're going to get to super intelligence. I don't think we're going to have digital gods anywhere anytime soon. I think more and more people are kind of coming to that realization, saying this technology is incredible, it's super powerful, super useful, it's not a digital god, and that requires adjusting how you're thinking about the technology.

• 19:18Speaker 00

  Yeah, no, the digital god thing is actually really funny because I feel like that was everyone's biggest fear. They were like, the AIs are gonna, I don't know, walk and attack us or something like that. And then I think more people, at least this year, I've seen a lot, or I'm on Twitter, so maybe a lot of people on Twitter are starting to like calm down a bit and say, okay, wait, I don't think it's gonna be like an AI invasion, but.

• 19:37Speaker 02

  Yeah, we've seen that change. I think this year, people are a lot more interested in the tangible, actual value that this stuff can deliver, rather than the sci-fi future. And that's been really, I really welcome that. That's been great.

• 19:48Speaker 01

  The AI industry in sort of its rise in interest and popularity in the startup and venture world and beyond, things that keep coming up are around things like sustainability, obviously, training these models and being able to utilize them takes up a lot of energy as well as, oh, are people training models on data that they actually are supposed to have access to, or data that they shouldn't be training on? There just seems like there's always these little like negative trends kind of following the industry around. How do you avoid some of these negative trends that are following the industry? And how is that gonna affect the industry as it continues to kind of progress from here?

• 20:21Speaker 02

  I don't think legitimate questions or criticisms of a technology hurt the industry. No, I think those are both things we should be thinking about. So from the efficiency side of things like, yeah, we do think about how to make our models more efficient, to make running them more efficient, so they use less energy when they're in production, how training them, how we can make that more efficient. And that's from both an energy, but also a cost perspective. So we try to think about what's a good model size we can train that is effective and easy to go to production with. And we've been thinking about that for a while. From a data privacy policy perspective, yeah, we think about that as well. So we're quite confident in the data that we train our models on. So right now, if you use one of our models, we will indemnify you based on the data that it's trained on. And we can do that because we're careful about where we get our data and we're confident that we have the ability to train on it. Yeah, I think those are both legitimate things to be thinking about. I'm glad people are.

• 21:07Speaker 01

  Yeah, it just seems like there's so many, not even just those aspects, there's also the potential for like federal policy or federal regulation of some sort on the industry, which of course, if you're building in many different verticals like healthcare or financial services, you of course are already under regulation of what you can and can't do anyway. But it seems like there's so much, I don't want to say like jump scare factor of like founders here in the US of like some of these terms like, oh, like, well, we'll get to the sustainability piece later and like, we're just trying to build without guardrails and then, you know, like fix issues later. And I'm curious, like, watching those kind of messages, what do you think about that? Those kind of approaches in the space that seem to be quite different from like how you were thinking about it?

• 21:49Speaker 02

  Yeah, this is a complicated question. I think what you're responding to is like some general extreme rhetoric on both sides around the utility of this stuff, right? You're seeing some people say, these things are amazing and they're going to solve everything and every problem is going to be solved by bigger language models. And on the other side, you're seeing people say, these things are useless, or that they can't do this, or they're bad for this, they're bad for that. And I think extreme dialogue and rhetoric on one side brings out the extreme rhetoric on the other. Whereas we're pretty sober about how this technology is useful and what value it can deliver. And to be clear, an insane amount of value. Like, I think these things are super useful. I'm really excited about them. I use an LLM and part of my daily workflow, I use them to, yeah, summarize documents, to help me draft things, to answer questions, to make graphs for me

• 22:33Speaker 02

  and solve complicated research questions. Like, I use them a lot. And I'm really excited for a time that I think will be pretty soon, when every knowledge worker uses one LLM as part of their workflow. I think that's really exciting. So I think it can add a lot of value. But I don't think it's going to bring about the death of all humans. And so we're able to kind of, you know, have this realistic approach that spares us from some of the extreme rhetoric on either side.

• 22:56Speaker 00

  Yeah, I know that the sustainability aspect has been something that I've been seeing, or at least environmental activists talk about, just because stats and stories are coming out about just how much energy and how much natural resources a lot of these, I guess, the chips are using. How worried? Should people be about all of the energy and all of the usage that CEOs and founders are using now during this AI revolution? Because it seems like, you know, people have been making a lot of headway, or it feels like people have been making headway, like in terms of the environment and like the UN SDGs. But now we have this thing that can kind of not bring us back, but change everything in terms of hitting those environmental goals. Because it's like, well, now we have to, you know, use these chips, we have to like innovate, we have to do this.

• 23:40Speaker 00

  But also, if it is impacting natural resources, I'm kind of figuring out like, what is the balance there? Should people be worried? Is this like something regulation handles? Or like, how should people feel about this?

• 23:52Speaker 02

  So I think we as the makers of a model should think about the energy we're using. And we should think about where we get that energy. So I'm sitting in Ontario where right now about 50% of our power comes from nuclear power. And that's awesome. I would love it if it was 100% of our power was coming from nuclear power. And I would love if all the data centers that we could build up here in Canada were cooled by the deep water of the lakes and powered by green nuclear power plants that we can build. That would be great. We need to keep thinking about that stuff. But I think some of the discourse around the electrical usage of these models falls out of thinking we'll continue to train infinitely large models. So we have some people, some of our competitors, going around saying that they're going to train 100 trillion parameters and they're going to spend, you know, eight x billions

• 24:36Speaker 02

  and billions and trillions of dollars to build a compute to do that. And we're not going to do that. Nor do I think that's a particularly good idea. I think that's kind of just playing into the hype and saying, yeah, we'll get there more and more and bigger and bigger and bigger. And like, I just want this to be as useful as I know it can be already. So I don't think when this technology lands and when it's as useful and ubiquitous as I think it is, I don't think it will be a massive component of the electrical usage of the population. It will be a component in the same way that computers are, but we'll have things like transportation and agriculture and the production that will occupy a significantly larger percentage of the energy usage of the people.

• 25:14Speaker 01

  And I do want to ask you a little bit about hiring because thinking of when you guys got started, thinking about now, what has hiring been like? I'm just curious if you've noticed more good tech talent going into this space or going into this area that maybe wasn't prior when you guys were like first looking to hire? Does it seem like there's always been a good pipeline of people who are both talented and trained and want to work on large language models? It's hard to get a gauge. It seems like now everyone knows all of it and can work on these. And I'm just like, was that always the case? Like, I wasn't paying attention then, so I don't know.

• 25:46Speaker 02

  Yeah, there's always been good talent in this. Like I said earlier, it's hard to not be interested in AI. And I maintain like this stuff is kind of inherently interesting. Like, it's very fun to work on. It's very cool to talk to a language model and have it, you know, do a task you didn't want to do. It's very exciting. And I think there's been people who have been excited about that from when we started. So there's always been good talent. It's true that there's a lot more interest and there's lots of people who are, you know, excellent in one field and recently got really excited about this and now are excellent in this field. So we've benefited from people catching up to us on the excitement front and we've been able to continue to hire good people. But I think we've been lucky and had great people to work with right from the beginning.

• 26:25Speaker 01

  And thinking about you guys being based in Canada, I wanted to ask you about that too. You're in Toronto, we're in New York. We're not very far from each other. We're not talking about particularly crazy distances here, but it's been a while since we've heard about like a big startup in Canada. We used to hear about them a lot more often pre the slowdown. The market up there is great, but what is it like building up there, especially watching in the US, companies are like really talking about potential policy, regulation, but you're just slightly removed from that. How do you feel about building this company in Canada, watching the Silicon Valley takeover in this space that's been happening for the last couple of years?

• 27:02Speaker 02

  Yeah, so I think Canada's been a great place to launch a global company. So we're co-headquartered in Toronto and San Francisco. We have an office in New York, actually a lovely office in New York and an office in London. And we recently have a small footprint in Paris and we have people all over the world. The sun never sets on Cohere at this stage. But Canada has been a great place to launch that global company. We now like, not only do we employ people all over the world, but we have customers all over the world. Like I mentioned Fujitsu earlier, but we're working with people everywhere. I think Canada is pretty unique in the culture and I think we've benefited from that. I think we have a pragmatic, practical, tangible approach to things and one that is done with, you know, intense academic interest,

• 27:43Speaker 02

  but also kindness and compassion. And that culture is part of us and has been useful for, you know, extending to the rest of the world.

• 27:51Speaker 01

  Yeah, I just wanted to ask, cause sometimes you'll talk to companies who are literally like, if you're not building AI in San Francisco, like, don't even bother. And I'm like, I don't know if that's true. So it's like always good to be like, talk to companies who are definitely building in other places.

• 28:06Speaker 02

  Well, to be clear, we're also building in San Francisco.

• 28:09Speaker 01

  Right.

• 28:10Speaker 02

  So we're also building AI in San Francisco and it's a great place to be. We've had some really amazing talent that's joined up there and we have some customers that are down there that have been great to work with. But Toronto is a good place to start, right? I mentioned my, my first boss at Google, Jess Hinton, is largely credited with inventing this technology. And his presence here and then the funding that the Canadian government has put into education has created like an awesome ecosystem of people working on ML. So yeah, we benefited from that as well.

• 28:36Speaker 01

  Definitely. And sort of thinking about where you guys are now, you've raised this recent round of funding, this obviously sizable round, and you guys are a couple of years into the journey. Landscape's changed a lot. What's next? What does the next couple of years look like for you guys? Cause I know you mentioned earlier, like you guys are very focused on one area. You're not looking to say, oh, in five years will be the AI that can do absolutely anything for everyone. Where are you guys headed and like, what are you going to be focused on for the next couple of years?

• 29:04Speaker 02

  Yeah, we're going to stay focused on making LLMs useful for enterprise. That's our mission. That's what we want to do. Like I just want this technology to be in a central part of everybody's work life. That's what we think this technology can do and that's what we're focused on delivering. That means we're going to keep making our models better at multilingualities, as we talked about, keep making their deployments more data-secure and private, keep making them better at reducing hallucinations through retrieval augmented generation, make them better at tool use, like there's all of these technological things that we're working on, but ultimately that results in solving the needs of enterprise with large language models. And that's what we've been doing up till now, and that's what we're going to keep doing.

• 29:44Speaker 01

  Well, perfect. Thank you so much for coming on the show, Nick. This has been fascinating.

• 29:47Speaker 02

  Yeah, thanks. I really enjoyed the conversation. Some good questions.

• 29:50Speaker 01

  I'll cut it before I start bugging you about the Toronto Maple Leafs.

• 29:53Speaker 02

  Oh, yeah, no comment on that one.

• 30:01Speaker 01

  And that was our conversation with Nick. Dom, before we dive in, what was the lie?

• 30:05Speaker 00

  The lie is that he thinks AI is on its way to becoming a digital god. He actually thinks you all are overreacting, and it's probably not going to head that way.

• 30:13Speaker 01

  No, I thought that was super interesting because you get a sense that people in the AI industry are getting a little bit more realistic about like how good the technology is and where it can be, but also like what it can't be, which obviously we were not seeing in 2022 and 2023. But it's nice to hear an actual founder be like, no, like, I think this technology is amazing, but not in that way.

• 30:36Speaker 00

  Yeah, definitely. I think, yeah, a lot of people are coming to that realization, or at least the people that I've seen on X. And it was really nice to have a practical person say that because I think that there is a lot of fear mongering, like, or I don't know, at least like I think of like the robots coming to like, take all the jobs and like, I don't know, they're going to march on the streets or something. Like, people have these wild theories about what AI is going to do. And in reality, it's probably just going to be not any of that. It's probably just going to be a really regular thing, as it already is and was.

• 31:06Speaker 01

  I know. And that's what I think it's, like you said, nice to talk to a founder who has that perspective, because talking with other AI companies in the space and things like that, it's like very refreshing to talk to an AI founder who's not only like, no, we have like a legitimate business model, and these are our customers, and this is what we do, and we're in one lane, and we don't, we're not trying to do the be all, end all for consumers and businesses alike. And we talked about diving in here. You liked what he said about the sustainability piece of it, too, something that the AI industry has seemed to largely shy away from.

• 31:37Speaker 00

  Yeah, because I've been reading a lot of articles of at least, I guess, climate people saying, oh, my gosh, you guys are using so much natural resources. We're literally trying to hit the UN SDGs. And, I mean, even I was a little shook. I was like, oh, my gosh, these chips are using how much water? But then it makes sense what he said. People are probably not going to be building like trillions of dollars worth of like, whatever to suck up, I don't know, an entire river. Like, it's probably gonna, once again, be on a much smaller scale. And as the AI industry figures itself out, it probably won't use as much energy uses as computers. Which, I mean, I guess computers and, you know, hardware tech has its own natural resource problem. But I guess that also kind of settled

• 32:19Speaker 00

  my fears, because here I was thinking that like, AI was gonna like, completely drain the Atlantic. And he's like, no, he's like, we're probably not gonna be doing that. And I don't know. I mean, there has to be like, a big fear campaign, I feel, because I was really scared for like, a few days after I read an article.

• 32:37Speaker 01

  No, and it's, I feel like he summed it up so well, too, where he was like, with like, other industries, you build the product, and then you work to make it as efficient as possible, and like, more streamlined, which of course, inevitably uses less power and less resources. So it's like, well, yeah, it's like, kind of the whole, it's just smart business. And like, almost always a smart business to do the more sustainable route, because whoever you're getting there is going to benefit you in some other way. And so it's nice to hear that he was thinking about that, too. But I think the other thing that was crazy for me with talking to him, and thinking about just like, how the AI conversation has changed, because like, when I started covering startups in 2018, literally no one talked about AI. And I remember articles in 2019, and research and data on the market

• 33:20Speaker 01

  and stuff being like, AI startups is coming past, people aren't interested in this space anymore. So it's crazy to hear from him and be like, yeah, I saw this thing in 2013, and like, in neural networks, and like, I thought it was late.

• 33:31Speaker 00

  I know, right? Imagine thinking you were late to AI in 2013.

• 33:35Speaker 01

  I know. It's a good thing he like, kept going.

• 33:38Speaker 00

  If you're an AI company in 2013, stay in line.

• 33:42Speaker 01

  I know. I know, it's just like, especially thinking of like, how Coher has done. And then this is not to say anything about other AI companies in a negative light or anything like that. But like, they've recently raised funding, they're now valued at over $5 billion. They have all these real and legit customers, clients, like they've built this very substantial business. It's crazy to think that he almost like didn't do it at all, or thought he was late. And I'm, yeah, I'm glad he, I'm glad he stayed alive.

• 34:09Speaker 00

  I know. Wait, we asked him about fundraising too. I feel like it's no fun asking AI companies anymore about fundraising. Cause they're all gonna say like, yeah, we went in with like a dollar revenue and got like a $30 billion valuation. And so I'm always like, do we ask about the fundraising journey? Do we not?

• 34:25Speaker 01

  Yeah, they have this couple of interesting backers too. We could have probably spent more time on that, cause like, they're backed by like, some really big organizations in Canada. And less so like huge American VC funds.

• 34:36Speaker 00

  Yeah, now I have regret, we should have gone deeper into that.

• 34:39Speaker 01

  But yeah, it's interesting, the whole like them being based in Toronto. And me trying to play off of that by being like, oh, so you don't agree you have to build in San Francisco? And him being like, oh, actually, we have a lot of engineers in San Francisco. I was like, well, well, well.

• 34:52Speaker 00

  I know, cause I was thinking, I was like, Canada, no offense to Canada. But I was like, Canada, question mark? And I mean, I guess Toronto was a big business hub. Yeah, he was like quickly like, we have a hub in San Francisco, New York, and like, you know, all these things. But, I don't know, I mean, Canada's an interesting place, cause I don't think I know much about policy in Canada. I guess I'm more so thinking like, when it comes to regulation, what we should expect regarding, I guess, AI regulation in Canada? Are they gonna be like, super strict, like the EU, or are they gonna kind of be a little bit looser, maybe like the US?

• 35:27Speaker 01

  Yeah, I don't really know, because there's definitely some areas where they've been like, Canadian policy's been way more in line with the EU and areas where they've seemed to be more in line with the US. But, I mean, they always move faster than the US, I'll give them credit for that. But I think it also depends if they have people in the US, in Canada, and in London. I'm curious where their customers are, because if their customers are in like, all these different places, they would have to follow those regulations anyway. If they're like, building something for a company based in a certain area. So, it definitely is interesting to think about them, because regulation has been such a huge piece of the AI conversation as of late, for US startups. So like, talking to a company like them, that probably already has customers in like, multiple different countries, is like, regulation would impact them in a little bit of a different way.

• 36:11Speaker 00

  Yeah.

• 36:11Speaker 01

  Yeah, I mean, we don't have said regulation to talk about how it'll impact them in a certain way, but it is interesting to think about. It's a lot more moving parts, I guess, for these companies than I had realized, or that was like made clear in a lot of the existing just like conversations around AI regulation.

• 36:25Speaker 00

  But you know what was really something is that he said that he doesn't think AI is in a bubble, which like, I guess, like on my Twitter timeline, I see people go back and forth all the time about it. Well, mostly they're talking about in the video, but like they go back and forth. They're like, AI is totally in a bubble. And he's kind of just like, no, you know, there's a lot of like fluff out there, but you know, people are actually building real companies. I just thought, I mean, do you think AI is in a bubble?

• 36:47Speaker 01

  You know, it's interesting because hearing his answer on that of being like, well, it can't be a bubble really because there are companies who are building real value and like there is real value coming out of this. I was like thinking back on it. Maybe it wouldn't be considered a bubble then because I was like, what are the other bubbles in our space? And it was like, the housing market that let the subprime mortgage that bubble. And then like the dot com boom and bust. And I'm like, I guess there weren't really many winners out of any of those scenarios. And like, I'm sure, of course, there's still some dot com companies around. We obviously TechCrunch is owned by one, Yahoo, but like there really aren't very many. So it's like, maybe we're just like thinking of bubble in the wrong way. Of course, it's not like a set definition, but it definitely made

• 37:30Speaker 01

  me think about it a little more where I was like, oh, well, maybe it wouldn't be a bubble. Maybe he's right. Like if people do actually gain a bunch of value out of it, even if some of the stuff does end up not working out and failing spectacularly, like what makes a bubble a bubble? I don't know. Maybe now I'm thinking about it.

• 37:43Speaker 00

  I like how Nick is having us think so philosophically now about AI. Like he's completely changed the way we think about this.

• 37:49Speaker 01

  Right.

• 37:50Speaker 00

  Now we're asking questions, you know, we're like, what if it's not a bubble? What is a bubble?

• 37:55Speaker 01

  What is a bubble?

• 37:56Speaker 00

  Shout out to Nick. I feel like he's just changed my life.

• 37:59Speaker 01

  He's definitely changed my way of thinking about AI. I will be a little kinder to the industry now, I think, because of Nick.

• 38:07Speaker 01

  Found is hosted by myself, TechCrunch senior reporter Becca Scutac, alongside senior reporter Dominic Midori Davis. Found is produced by Maggie Stamets with editing by Kel. Our illustrator is Bryce Durbin. Found's audience development and social media is managed by Morgan Little, Alyssa Stringer, and Natalie Kreisman. TechCrunch's audio products are managed by Henry Pickavet. Thanks for listening and we'll be back next week.