He built Cohere into a $5.5B AI startup; How to Win in AI; & Why LLMs won't lead to AGI. | Nick Frosst, Co-Founder of Cohere

In this episode, I sit down with Nick Frosst, Co-Founder of Cohere, the $5.5B AI startup that’s targeting the enterprise landscape. 

We go through the origin story of Cohere, the challenges of building foundational models, and why he believes large language models (LLMs) won’t lead to artificial general intelligence (AGI). We also explore the fierce competition in AI, what sets Cohere apart, and Nick’s advice for founders building in AI today. 


Why you should listen

• LLMs are powerful but have clear limitations and won't lead to AGI.
•  Why AI startups need to start with real problems vs leveraging AI for its own sake
• Why ChatGPT was as much of a UI/UX revolution than a technological one
• What tech founders need to do to win in AI

Timestamps:

(00:00:00) Intro
(00:03:21) AI Expectations
(00:06:05)  A Unique and New Moment
(00:09:38)  Resource Intensive Industry
(00:12:03)  Zero to One
(00:15:07)  Base Language Model to Chat Model
(00:17:15)  Carving Out a Niche
(00:21:03)  Open Source
(00:24:00)  The Limits of LLMs
(00:26:18)  Agents
(00:29:30)  AGI
(00:34:04)  A Little Bit of Data
(00:39:05)  Speed of Development
(00:40:47) Finding True Product Market Fit
(00:43:37) One Piece of Advice

Chapters

• 3:21 AI Expectations
• 12:03 Zero to One
• 15:07 Base Language Model to Chat Model
• 17:15 Carving Out a Niche
• 21:03 Open Source
• 24:00 The Limits of LLMs
• 26:18 Agents
• 29:30 AGI
• 34:04 A Little Bit of Data
• 39:05 Speed of Development
• 40:47 True Product Market Fit
• 43:37 A Piece of Advice

Transcript

Nick Frosst (00:00):

Yeah,I just think if we're gonna build something that is smart, like a person, it will probably have to be like a person in some ways. Yeah. But I also don't care. Like, I just don't give a shit. It's not obvious to me that we want AGI, I don't wake up in the morning and say, gee, I wish my computer was a person. So I think this technology is wild. I still get very excited about it all the time, but I don't think it's a clear path to AGI. I don't think we're gonna be making digital gods anytime soon. I think we're gonna be making sequence models. Models that take in a series of words and predict the most likely next word, based on a dataset they were trained on. That's what this technology is. That's what it's always been. That makes 'em extremely useful at a bunch of stuff. Emotionally. I would definitely wanna speak to other first time founders and say  “it's really stressful and regardless of what's actually going on, the emotional valence will be the same. So you should know as you're going on that every challenge you're facing, that's the hardest challenge you've faced so far. And it's gonna feel all encompassing. And then later you're gonna look back and you're gonna say how quaint. But the emotional impact will always be the same. So you gotta, you gotta chill out, you gotta figure out how to handle it. 'cause it won't get easier. 

Pablo Srugo (01:13):

Before we start today, could you just do me a really quick, really small favor. Can you just take your phone outta your pocket, open up Spotify, open up Apple Podcast, open up whatever player you're using, and give the show five star. You can think of it as like your good deed of the day, because it's not just for me, it's for all other founders. If you like the show, iff you get value out of this show, when you review the show, other founders are way more likely to find it. So this is like your philanthropic moment of the day. And yes, if you're thinking to yourself, is this really for me? Is he really talking to me? You must be talking to somebody else. No, I'm actually specifically talking to you. So please take a handful of seconds. If you like this show, rate it, give it five star, write something. Whatever you do, you're gonna be helping. Not just me, but so many other founders. Thank you. And I hope you enjoy today's episode. Well, Nick, welcome to the show.

Nick Frosst (02:06):

Thanks for having me.

Pablo Srugo (02:08):

I have to start with this question. You started like 2019, 2020, we're now like four or five years later. Did you ever think in your wildest dreams that you would be a, that what would happen to AI has happened? And second of all that you'd be now like the CTO and co-founder of $5 billion Startup that just raised one of the largest rounds in Canadian history?

Nick Frosst (02:27):

<laugh>  well, no. To some of those and yes to some of 'em. One, one correction. I'm not the CTO anymore. Um,

Pablo Srugo (02:35):

Oh,Okay

Nick Frosst (02:35):

Yeah, I'm a co-founder. Uh yeah, that's one of actually the great things about being in a company that has grown this quickly is I've gotten to do a whole bunch of different roles at the company. So I was the CTO for a while. I am not anymore. And to be clear, now that we have hundreds of employees, it's been great to bring in people who've had more experience managing teams of that size. To answer your questions,

Pablo Srugo (02:58):

 What is your role now, by the way?

Nick Frosst (03:00):

I jump around a lot. So over the years, I just kind of fill the gaps, and do whatever needs to get done until I find somebody who's better at doing the thing that I'm doing, and then we hire them and I go fill some other gap. Nice. Um, yeah, so that's been, that's actually one of the really enjoyable things of being a founder. You get to move around and get a little bit of experience everywhere. To go back to what you asked, did I ever expect AI to get to where it is now? Yes. It's actually kind of gone pretty much the way I expected it to with the one notable exception of data efficiency, of reinforcement learning from human feedback. So that's the one thing that, when we started this, I knew that language models were gonna be as useful as they are today.

I knew they were gonna be as powerful and as general purpose. I I will also say I knew that they were gonna have the limitations that they have. I did not think you could take a generic web, web text trained language model that's just trained on all the text from the open web and refine it via feedback from people and had such a massive effect on the ease of use of language models. So that's the one thing that caught me by surprise. I think that caught,  pretty much everybody but the people in Open AI by surprise, was that you can have a relatively small number of examples from people fine tune the model on that, and then it's a lot easier to work with. So I don't know if your listeners are familiar with LLMs kind of pre and post chat GPT, but before chat GPT , if you wanted to get them to do something useful, you had to spend a bunch of time prompt engineering and figuring out how to kind of set up this weird prompt to get them, 

Pablo Srugo (04:40):

This is with GPT what? was that GPT three or GPT before, before GPT three, even

Nick Frosst (04:44):

That was before GPT-3. Chat GPT was like a particular fine tuned version. It was fine tuned off of examples of chat, you know, and it was the data efficiency of that that really surprised everybody. And that's kind of what caught the world up with the excitement of the LLM community that they had before. Um, so that was surprising. But otherwise, the utility of them, the ubiquity of them is not surprising.

Pablo Srugo (05:05):

Like even the speed to where we got here.

Nick Frosst (05:08):

Yeah, I think, yeah. I don't think it's that surprising <laugh> I think all of that was pretty obvious after the first few scale ups of transformers. Yeah. Except for this data efficiency thing, this thing really, it is surprising. And just to dive into that a little bit, like when you train a model, when you train a text model, you train it on just a huge amount of, of written text, and that ends up being like terabytes of text. And then when you fine tune a model on feedback from people, it's a comparatively really small amount of data. And like, that's a really scientifically interesting thing. 

Pablo Srugo (05:44):

So, we'll, we'll dive more to that shortly, but maybe then take us back to, to that time, like in 20 19, 20 20, just give us like the, the origin story on, on Cohere, how it all started, particularly interested in, you know, where you thought the market might go at that time, especially since things seem to have played out kind of more or less according to plan.

Nick Frosst (06:02):

So back in 2019,  this was two years after,  the paper “Attention Is All You Need” came out, which is a paper that introduced to architecture that all language models are based off of now. and that paper was written by a bunch of people at Google, one of whom was Aiden Gomez, who is our co-founder and CEO, and he kind of got this whole thing together. So he was involved in that paper, and it was shortly after that that he realized, hey, like this is a unique and new moment in machine learning. And the unique and new moment is we've trained, we're training truly like general purpose machine learning models. Before this, if you wanted to use a machine learning model to do something like, I don't know, recognize cats, you were best served by creating an image recognition model and training it only to recognize cats. That was ideal. Like if you had, you had some vision task, the best model you were gonna get at, that was a model that was trained for only that language models were this new thing. Whereas if you, if you wanted a language model to, to, you know, extract numbers from A PDF or summarize a call transcript or doing something like that, this was the first time where you were gonna be best served by a general purpose language model that was trained on as much like all language in general. And that would be the best model for that task. So that's a really cool and new moment. And it's that realization that prompted Aiden to say, Hey, like this is, there's an opportunity here and indeed a need for a company to train a general purpose language model and make it available for enterprises. And that's what we did. So he, you know, got excited about that and then convinced me and, Ivan, our other co-founder to quit our jobs and start the company. And we've been doing that since.

Pablo Srugo (07:43):

And what was happening at that time, like 2019? I mean, OpenAI was working on its own thing. I mean, I'm sure like the big guys, Google, et cetera, were like, well, what did the market look like then when you decided to do this? But specifically for the enterprise,

Nick Frosst (07:55):

The market didn't really exist at that time. there was nobody selling access to large language models.

Pablo Srugo (08:00):

Were there players that you saw would go a particular way? did you think about where others would go or it was just like totally open space.

Nick Frosst (08:08):

We knew open AI would be going after this. Yeah, yeah. But, it was pretty new. It was still pretty new.

Pablo Srugo (08:13):

And so where did you go from there? What was kind of step one after that, that realization that there was an opportunity there?

Nick Frosst (08:19):

Well, step one was to figure out what to do and how to start it and hire a few people and start building. We started, yeah, work with,  the three co-founders and three founding engineers and started working on language models early 2020.

Pablo Srugo (08:34):

What was your- why the enterprise?

Nick Frosst (08:36):

The enterprise because we've, because we're all very pragmatic people and we all really want this technology to be useful. And where I think it is most useful is inside enterprises. So both for their internal use and for building stuff for customers. Um, but I think that's where it adds the most value. Like, I think it's very fun to chat to A LLM just like as a consumer, I think that's enjoyable, but I don't really do it for fun on a daily basis. I do use an LLM on a daily basis to help me with work, and I think that's really where this technology shines. 

Pablo Srugo (09:11):

And so how long does it take to build- I mean, you can't really MVP this stuff, can you -, what, what is building when you're a foundational AI company, like what does building that first product look like? What, how do you set things up?

Nick Frosst (09:23):

It was a long time ago, and it has, it has moved quickly. So I might get some of the details of our history wrong. But at the time we started and we started setting up the infrastructure and going over data sets to create our first version of a language model. And I think we spent a bunch of time in the beginning working on setting up infrastructure to train really large language models in a way that we knew was probably inefficient, but demonstrated at least our ability. So when we started, we did not have the resources we have now. And this is a very resource intensive industry. Like, it costs a lot of money to train a large language model because you have to pay for the computes, um, you have to get data and you have to get talent. All of those things are expensive. Um, so we did not have the resources we have now. And so we were spending time kind of demonstrating that we knew how to do it, and that if we had more resources, we could do it efficiently. So the first few months we were training large language models in a way that we knew wouldn't scale because we didn't have access to compute, but it was good enough to say, Hey, look, we can do this. And if we had more resources, then we could go and do it quickly and we could train these models in a reasonable timeframe. And we used that to go from our seed round into our series A.

Pablo Srugo (10:30):

And were these like demos or like what, what did the, what did the model do at that time?

Nick Frosst (10:34):

The same stuff, just the way we trained it was really slow. So when we first started training, we would, um, I think we, we started training not with a single kind of supercomputer, but by, with small amounts of GPUs spread out throughout a data center. And that is very, very slow. 'cause you have to communicate between all of them. So it's really, really slow to do it, but it demonstrates, hey, look, we, we have the talent and you know, we've collected the data. So that was enough to say, Hey, well, if we had more resources, we could rent, uh, a lot of GPUs and then we could do this quickly.

Pablo Srugo (11:05):

How do you work backwards from having a commercially available product? Like how, how much time did you guys give yourselves to, to get there?

Nick Frosst (11:12):

Well, yeah, this industry's a little different than probably a lot of your listeners in that it's, it's more like a, it's more like the creation of a resource. Like in some ways, like, I mean, maybe this is an, maybe this is an interesting time to have this conversation, but in some ways, it'd be like asking, you know, when, when, what was the, what was your first MVP of a mine? Or it'd be like, oh, what was an MVP of an oil refinery? Some of these -if you think of them more like a utility, it's, it's a little harder to think about because a language model that hasn't been trained is not useful. There's no MVP of that. I mean, you could train a small, you could train a small language model and you could try to communicate to somebody, Hey, this but better. Like, that kind of works, but it's pretty hard to, to imagine how it's gonna be useful until it's done. So in that way it's, it's more like a, like a utility provider. It's like, you know , what's the MVP of a power plant?

Pablo Srugo (12:02):

No, there isn't. I mean these kind of things go from zero to one. I mean, in a sense, Chat GPT thing is a little different, but from the mainstream's user's perspective, it really was a zero to one moment. Like the thing was kinda like not useful, non-existent, and then it was like ex, you know, incredibly useful and yeah, prevalent everywhere.

Nick Frosst (12:19):

So I think that, yeah, I think that's, I've talked about this a lot, but I think that's a really interesting historical moment, in the history of technology, because everything that you can do with a language model right after chat GPT, you could pretty much do with a language model before it just sucked, sucked in that it was like really hard to do. You had to spend a lot of time prompt engineering and it was like required thinking in a really weird way. So instead of just asking the model to do something like, hey, write me a poem about a fish, you had to write as the prompt. The following is a poem about a fish colon new line, and then get the model to generate. So because it wasn't trained to be a chat model, it didn't know how to chat, it did know how to write the completion of documents. And so if you wanted it to do something, you have to set up the first half of the document to like, trick it into doing the thing you wanted to do. But you could do everything. Refining it on Chatdoesn't make it much better at stuff. It just makes it a lot easier to use. That's a really interesting historical moment that what woke people up to the possibility of language models was an ease of use.

Pablo Srugo (13:19):

So , the models before weren't necessarily worse. Like you're saying Chat GPT’s main, let's say, innovation is almost like a UI, like a UX kind of innovation.

Nick Frosst (13:29):

In some ways,Yeah, yeah. it's, so the base model for language models is just trained a complete document. So you give it a sentence, it writes the next sentence, you give it half a paragraph, it writes the second half of the paragraph, you give it a half a page, it writes the second half of the page. But it does that based on the way people write, you know, write documents that are available for training. So it doesn't, like, we don't run around on the internet and write “ write me a poem about a fish” and then write the poem about a fish instead, if somebody's writing something that they'll probably write the, you know, the following is my favorite poem about a fish colon new line and then the poem or something. So you had to think in that framework, you had to think, okay, imagine the task that I want to solve, somebody has already solved it and then it's on a website, what would the top of that website look like? And then that's how you would trick the model into doing something useful. The chat, GPT and all the models that people use now, they're those base models then further trained on examples of people talking to a chatbot. And now because it's trained on that, like examples of chat logs, it gives it a chat prompt and it will respond the way you would expect it to respond with. That's because it's trained on that data. Yeah. So that doesn't really make it better. It just makes it easier to use. 

Pablo Srugo (14:38):

I'm curious, what was that moment within, 'cause like we all experienced, everybody experienced that moment, you know, differently, but like we're all kind of on, on the outside. Like even me as a vc, obviously we're looking at AI or whatever, but all of a sudden this happens and it's just like this crazy zero to one moment for you. You're actually building a foundational model for years before this you would, you know, in the RD world even before that. And then this happens. Like what, what happens internally at co here, in the like last two months of 22 after Chat GPT.

Nick Frosst (15:07):

Oh yeah, yeah. we, yeah, we were, we were all surprised and excited and the surprise and excitement was that we were like, wow, you really do not need very many training examples to take that base language model and turn it into a chat model. And then the chat model's quite easy to use. So we had been making chatbots before chat GPT, like I was running Discord bots with various chat models that were working already. But the way that I did that was with a really long prompt. So I had many examples of a chat conversation and I had set my prompt up so that it looked like a chat log you might find on the web. And then I would take a new message from a user and put it into the prompt and get the model to respond. So it was like, we were doing that for a long time, but we didn't think, hey, we could just take this base model, train it a little bit more on chat data and get it to be a chat model. So when chat GPT came out, we were all like, wow, that's a new finding. It's significantly more data efficient than we thought. That's really cool and obviously this is the way people want to use it. So we went about creating a data set to do that, but with a particular focus on making it useful in enterprise as opposed to being a chat bot for consumers.

Pablo Srugo (16:12):

Maybe talk to me about that. Like, that's one thing I always think about is like just the, the now like you said, there was no market back then. Now there is a market and there's different people playing in different places. Like you have the open source on one side, like Meta’s LLlama, which we can talk about a little bit. 'cause that's, you know, the, their new model came out or whatever on the closed source size. Um, you've got, you know, whatever open eye philanthropic, Cohere who, like, how do you guys view the world? Like do you, do you see people splitting the pie and taking different pieces of it? Or are you all just kind of competing against each other? Like how, how do you guys think about the market and the way it might develop?

Nick Frosst (16:45):

Yeah, the market is definitely competitive these days, but unlike some of the other industries, there aren't the same positive feedback loops. It's not like a social media in which if you're a single person using a social media, you're having no fun at all. And you're, you know, the more people there are, the better it is for everybody. For, for the most part, that's not the case with a large language model. There are some feedback loops, like more people use your model. You can, you know, get better feedback from them figuring out how to make it better. But it's not quite as massive. So , it's still, it's a competitive market. It, but we have kind of carved out a niche in really only focusing in on what businesses and companies are trying to do with large language models and trying to make it good for them on those purposes. So we're not interested in making a public facing bot for consumers to talk to. We're not interested in chasing AGI, we're not interested in a lot of, there's a lot of like weird public benchmarks that exist trying to measure, you know, how close we are to AGI or how close we're to things. Like we don't really care about any of that. We just care about when we talk to a company who's using our model, is it useful for them? Is it doing the thing they need it to, to do? And that's what we obsess over and that's fairly unique within the foundational model companies. Uh, we're unique in our like, singular focus of that.

Pablo Srugo (17:57):

What does that mean specifically? For an enterprise, what do you like? So I understand what you don't do. Like you don't waste time in in consumer waste time, <inaudible>, you don't spend time consumer, you don't spend time in each I, but like what do you do for an enterprise that might be different than what OpenAI offers or, or Llama three with whatever offers them today, whatever.

Nick Frosst (18:13):

Yeah. what I've described now is not AGI, that's enterprise focused. Like that's a philosophical difference. Like how does that bear out in our actual product? Teah, yeah. Um, it means we have spent a lot of time figuring out how to serve a model in a data privacy preserving way. So we'll, we'll serve our model wherever, we'll you can use our model on our platform or on any cloud provider or on-prem or like wherever we will host our model wherever your data is. That's like one thing. It means we spent a lot of time trying to make the model really good at retrieval augmented generation. That's one thing our model is, is very good. So you can give it a bunch of data and it can answer questions based on that data and give you citations so you know, you can trust the answers. That's like something we've trained the model to be really good at. We've made it really good at multilinguals

Pablo Srugo (18:51):

And enterprises care specifically about that. 'cause they have a lot of data that they wanna feed into these models in a, in a secure way?

Nick Frosst (18:57):

Um, yes, uh, even stronger than that, enterprises really care about that because if they're not giving it their own data, then they're just getting random stuff out of a model, which turns out to be not very useful. <laugh>, it's kind of like a, you know, if you, if you wanted to answer questions about your company, you need to give it information about your company

Pablo Srugo (19:13):

Mainly true. You would say for internal use cases or, or equally so for, for external, for,

Nick Frosst (19:18):

For everything. But like a consumer experiences, where the enjoyment is just like, you know, “do it”. Or for some, some work automation for example, you know, people often use our, our model to, you know, read to format text, you know, to do entity extraction or to do summarization or something like that. And that you can kind of give it the document it needs to be working off of and the answers it. But for, yeah, in internal or external retrieval, augmented generation is hugely important. So we spent a lot of time making the model really good at that. We've made it really good at multilingual stuff. So it's beyond English, it's in, you know, our model is in all major business languages and we'll keep pushing that. And we've made it particularly good at using tools. So it can tell you like, you know, you can, same way retrieval, augmented generation, the model figures out what to search, does the search then answers the question based on what it's found. Using a tool with an LLM is similar. You can give it, you know, access to some database or access to some external function like a calculator or something. And the model can figure out when to use that tool and then use it. So we've made it really good at that. We've also invested a bunch in search in general and embeddings. So we have an embedding model as well, which helps with enterprises who have a whole bunch of data trying to find the relevant stuff. So we've made the models, our embedding models, really good at that. So those are like, that's a handful of, you know, implementation, technical, um, distinctions. But they all come from focusing first on the enterprise. I'm thinking, what, what can we do to make this stuff as useful as we know it can be?

Pablo Srugo (20:42):

What do you make of this latest, uh, LLlama 3.1 and just like open source in general. I mean, there's some people, people that basically every single time Meta puts something new out, they're like, open source is gonna kill everything. They're gonna think, like it's free, right? So like free is a powerful word. And at least when it comes to Chat GPT $20 a month, a lot of people are saying, well that stuff's just gonna go, it's gonna go away. It's gonna go to zero. Do you have a sense of how this stuff shakes out? How it all plays out? 

Nick Frosst (21:03):

Yeah,So I think, I think it's important to draw a distinction between, you know, the open source as it exists now and, and what that word often conjures up in people's minds. So Llama Llama three, like the stuff meta is doing great models, great engineers, really cool stuff. Um, but it's not open source the way Linux is open source. It's not open source the way, you know, I don't know, like Wikipedia contributions are open source in that it's not created by a bunch of hobbyists for fun in their spare time. It's actually a very valuable asset that a large enterprise is choosing to give away for free for the purposes of notoriety, for the purposes of, you know, getting people excited about a really cool technology. It's a little different. It's also not similar to Linux in that it's not building upon itself, right? Like the, the models that were released by like a, as soon as a new model is released, the old ones are kind of irrelevant now.

So when I think about Cohere’s value, it's not in, you know, the weights of any particular model. In fact, we've released our model weights for researchers as well. So if researchers wanna use our model, they can download the weights from Hugging Face and I encourage people to do that, play it out like our value is in the process, which creates those models and improves them. The process which deploys them, makes 'em useful to people, takes feedback and improves them, integrates into your system like, you know, works with search and rag and all of these extra things, the actual, like the weights of any particular model are, are not super useful when you get into it. So I, I think it's really exciting to see meta releasing the weights for models because I, I'm a lover of this technology and I love to see people developing and building on it, but when I think about our commercial value, it's a lot more than just a static weights. Um, and yeah, it's the process which creates them and improves them and year over year makes them better. And it's the way in which we deploy them and give companies access to them.

Pablo Srugo (22:50):

What about on the consumer side? Like you think there's still a world where people pay $20 a month for Chat GPT when the stuff's like in WhatsApp and, and everywhere?

Nick Frosst (22:58):

I don't know. Yeah, I don't know. I I, I don't anymore <laugh>, so I don't know. Yeah, I mean I,partly I don't, because I use our own model, uh, and we do, we do have a, we do have a version where you can use it on a chat with it online, um, though that exists mostly for the purposes of, of demonstrating the utility to enterprises. So I use that as a daily chat model and I, and I use some other products which have, uh, chat built into them. So yeah, I, I don't know how sustainable that's gonna be, but again, that's not, that's not really our focus, right? Like we we're, we're not interested in making a chat bot we're interested in giving this technology to enterprises so they can solve real problems.

Pablo Srugo (23:35):

You, you mentioned, you kind of alluded to this earlier, you, you mentioned that you knew how powerful LLM’s could be and how, you know, what their limits were. Like what, what are their limits? How far can they go? And, and you know, there's, I think there's people on that, on the spectrum of just, you know, how much work these LM’s could take over and others that think, you know, maybe it's, it's a little bit more constrained. Like what are your thoughts on that?

Nick Frosst (24:00):

So I think its technology is wild. I still get very excited about it all the time, but I don't think it's a clear path to AGI. I don't think we're gonna be making digital gods anytime soon. I think we're gonna be making sequence models. I think we're models that take in a series of words and predict the most likely next word, based on a dataset they were trained on. Like that's what, that's what this technology is. That's what it's always been. That makes them extremely useful at a bunch of stuff. Um, and I think we'll get to a point pretty soon where you can, you rely on an, as an individual at work on an LLM for a whole bunch of your daily tasks. I increasingly use an LLM to do stuff for me. Like I'll use our model and I'll say, oh, you know, an answer this question and make a graph to gimme the information of this. So I'll or I'll say, you know, when I'm traveling around, I'll be like, look, look up these statistics about this and create a one page document that summarizes the information I need.

Pablo Srugo (24:52):

And you find it's good, it's accurate. 

Nick Frosst (24:53):

Oh, yeah. It's accurate. And we always provide citations. So I used to always check the citations and be like, oh, where, you know, let me check this. I don't check the citations anymore because I checked them for a long time and they were always right. So now if I see something weird, then I'll check the citations. So I'll be like, oh, is that, is that true? But for the most part, I've, I've come to trust the output of the model. So yeah, I think, I think we'll get to someplace where you use it as a daily part of your job. Like you, you'll go to work, open up slack, open up, you know, an email and you'll open up a chat model and you'll like, use a bunch of, whenever you can offload some mundane task you don't want to do, you'll offload it onto that model. And I'm, we're seeing people go to production with stuff like this already and see huge gains in productivity with it. So that's really exciting. I don't think we're gonna get to a point where you will treat or think about an LLM as a person or a coworker. You will always know that it has very clear limitations and it can be useful when you give it the specific instructions of what to do. But it won't be useful for saying things like, plot out my next business strategy. 

Pablo Srugo (25:50):

Sure

Nick Frosst (25:50):

Yeah. It won't be useful for saying, you know, which stocks are gonna go up tomorrow. It won't be useful for saying, you know,like high level conceptual stuff. that's not stuff that sequence models can do.

Pablo Srugo (26:01):

What about this concept of agents? Like that's been a buzzword lately, just agents that'll come in, replace specific jobs. Like how, how prevalent do you- I feel like it's still most, in most cases people talk about agents, like it's more of an idea than reality. Like does that become reality in many cases?

Nick Frosst (26:16):

Yeah, so I, I <laugh>, I originally pushed back on the term agents 'cause I don't like applying implying agency to sequence models. I don't like that. I think they don't, they don't have agency, they’re sequence models. However, I have lost that nomenclature war. Yes. And everybody uses the term agents, so there's nothing I can do about it. So I gotta lead into it. So when I think of agents, what I think of is language models being used as orchestrators to coordinate other aspects of a computer, of computer processes. So you could like retrieval, augmented generation rag, you could think of that as a really simple agent what does the agent do? It takes in a question from a user, it makes a prediction of what needs to be searched, like searched on the web or searched in a database to answer that question. But then searches, finds the relevant information and answers the question. You could think of that like an agent flow, except it's only got one tool and the tool is search. We can do the same thing by adding in a whole bunch of tools. And you can imagine now, you know, you have access to a search, maybe you have access to an email client, maybe you have access to a, a database, maybe you have access to, you know, some other API to, you know, retrieve information or write information and you could end up asking an agent something like, yeah, make me up, make me a plot of our five biggest customers from the last two months and our like three newest leads from the last week and put that in an email and email my boss and write a brief summary of each of their, each of those businesses. Like, you know, I view businesses use cases and put that into a document in Google Drive. And that LLM could do that. Like what, what I've just described is very like that, that's within the realm of possibility. And that's what I'm excited about, this agent stuff. It's like, it's like giving the model access to things beyond an LLM so it can begin to orchestrate tasks like that.

Pablo Srugo (28:10):

Those are like end-to-end workflows. But do you think that that even takes over like a full job? 'cause a job is always, I mean it's a series of workflows I find like most jobs, I mean there's, there's higher level obviously, but like you think about the lower level jobs, it's a series of workflows plus edge cases, you know, like the stuff that you <laugh> you could really workflow, you're like, yeah, did you do this? And you feel like it's more like this stuff becomes this kind of copilot thing. I'm using buzzwords, you know, trendy things, but like more of a co-pilot where a lot of your workflows get automated, they get a lot faster 'cause of these agents. Or do you think agents can really come in and just do, you know, end-to-end roles just kind of obliterate it as a result?

Nick Frosst (28:42):

Yeah, I think they're gonna more augment and to be clear, I think that augmentation is huge and massively transformative. And when we see people, like we've already seen companies and customers using our tech to augment some workflow and get like a 50% increase in productivity as a result. Like, it's super exciting. I think when you look at the work people do, there's actually a lot of edge cases they encounter all the time and there's lots of things that they do that, uh, could not be done just on the world of the, just on text, just on the world of computers that are actually done like in, in the real world based on communicating to people and reading subtle pews and all kinds of things that can't be replaced by an LLM. So I think we're, we're all about to end, right? Already getting more efficient and uh, through augmentation and that's really exciting to me, but I don't foresee this completely replacing large roles. 

Pablo Srugo (29:29):

Yeah,why not? the other take you have's a little different than others in spaces, like your idea on AGI right? Like that we're not, I mean, not that it's impossible, it's just like, it's much further away than people might think and, and LM’s are not necessarily the way to get there. Why not?

Nick Frosst (29:41):

Yeah, this is an interesting question. I mean, so this, I've been saying this for a long time now and, uh, I think right around Chat gpt , like a year, year and a half ago, it was more contentious and people would be like, Nick, why? Why don't you think, oh, we're gonna get to AGI. It's so obvious as the years have gone on, like more and more people agree with, I don't even get a lot of pushback on this anymore. I'm just like, yeah, this is, this is obviously not gonna get us to AGI Why would you think it would?

Pablo Srugo (30:03):

I don't know if it was, it was the fact that obviously the models have gotten better, but like the original chat GPTI think was maybe less constrained, less safe. And so like some of the answers were so far out there that it personified it, like it made you think like it was a real <laugh>,

Nick Frosst (30:15):

But all but all of that was, was a probabilistic model of sequences, right? Like that's a really, like, that's all, that's all the tech has ever been and that's really powerful, but it's not AGI and I think a year ago people were like, yeah, this is gonna be, you know, this, we're gonna get there. And as the models have gotten better, they've gotten way more useful. I think the economic output of these things is getting bigger and bigger and more and more useful. But I think it's pretty clear to people that just making them better isn't solving, you know, isn't making them more human-like it's making them better sequence models just like awesome. But it's not making them more intelligent in the way that we are intelligent. So yeah, I don't, I don't think we're gonna get that. The fundamental reason for that is when people talk about AGI really what they're talking about is, is it smart in the way that a person is smart and like no, because it learns in a completely different way than you learn, right? Like you do not learn language by hearing as much of it as possible and probabilistically predicting, which is the next most likely word, you do a little bit of that. And if I wanna get really specific, there are two parts of your brain called Vernick area and b Broca's area. They're responsible for language. One of them is responsible for the semantics of language. So if you have an injury in that part of your brain, you're able to speak in a language that is well-formed, but devoid of meaning. The other one is responsible for the syntax of language. So if you have an injury in that part of your brain, you're able to communicate but not with correct syntax. So I think that the syntax part that's like kind of what we've created, like we're kind of really good at automating that. But your brain does so much more than that and you don't learn by just seeing as many examples of this as possible. You learn by interacting with the world, you learn by having agency in it, you learn by, you know, making predictions of the world and, and experiments in the world and validating them. And then you use language as a, as a means of communicating that we're only doing that language part right now and that's all language models are gonna do. So that's super powerful. Like I really can't understate how huge I think this is gonna be and how it's already changed a lot of the way we, we, we work with computers and we'll to continue to change, but I don't think it's gonna get us to AGI 

Pablo Srugo (32:21):

What do you think we need to get? is there a path to there or not? Not really. ِAt least not like in the foreseeable future.

Nick Frosst (32:26):

I'm not a dualist, so I think there's no reason why we couldn't at some point build a computer that is intelligent in the way humans are intelligent. I imagine if we do, it'll probably have to have several of the limitations that people have. So I imagine it will probably have to be embodied. I imagine it may need to be, it may need to be mortal in some sense. I don't, I don't know. “Pablo: Why is that?” I don't know. These we're getting into really like

Pablo Srugo (32:50):

This. Yeah, I know, I know. But it <laugh> it just went that way. I'm surprised by that answer. Yeah,

Nick Frosst (32:54):

I don't, I don't, I just think if we're gonna build something that is smart like a person, it will probably have to be like a person in some ways. Yeah. But I also don't care. Like, I just don't give a shit. It's not obvious to me that we want AGI don't wake up in the morning and say, gee, I wish my computer was a person. I wake up in the morning and I say, gee, I wish I could do my work faster. Like, gee, I wish I didn't have to do this boring thing. Gee, I wish I could get my, you know, I could get an LLM to automate a bunch of stuff I don't want to do. That's what I'm thinking about. That's what we care about. We're figuring out how can we make this useful. That's right.

Pablo Srugo (33:30):

I just want the, I just want Chat GPT or whatever LM to write my LinkedIn post for me <laugh> still.

Nick Frosst (33:35):

Yeah, that sounds great. And you can, you can say like, here, you know, here's the thing I wanna make, gimme a, here's like, here's the bullet point, here's a graph I found. Help me write this. Like, that's great. Or you could even say, you know, these are the types of, yeah, these are the types of things I think are gonna be useful and these are the types of things I care about. The AGI stuff I think is fun to think about, but it's not obvious. It's not an obvious need to me.

Pablo Srugo (33:53):

I think you mentioned earlier that your one big surprise was like the impact of, of reinforcement learning from even feedback. What about that was so surprising?

Nick Frosst (34:01):

It's a lot better than we thought it was gonna be <laugh>. Yeah, it required a lot less data.

Pablo Srugo (34:05):

That that was your, your point there that a little bit of data can make the model so much better. Yeah,

Nick Frosst (34:09):

Exactly.

Pablo Srugo (34:09):

And why, why does that work? 

Nick Frosst (34:11):

Why does that work? It's kind of an interesting question. I mean, it works in that we have sequence models and then we train it a little bit more on a particular sequence and then it's better at those sequences. What's surprising is that it generalizes. So if you have, you know, let's say you, you've trained a model and it's seen a lot of examples of summaries, right? Somebody writes an article on the bottom, they say summary colon, and then they've written the summary. So before, if you wanted a model to get us to write a summary, that's what you'd do. You'd give it an article you'd write at the bottom summary colon and let it predict the summary. And it would do a pretty good job of that. So let's say the model is really good at that and then you'd train it on a bunch of examples of people having casual conversations where somebody asks for something and the model does it, but no, but they've never asked for a summary. They've never said, please write a summary of the article above. If you trade it on just those conversations and it's never been asked to do a summary, it will still do a good job when you ask it to do a summary. So it generalizes to request is never asked, never been asked in the training data. That's cool. That's a bit of a surprise. Why does that work? Mm. There's like complicated answers about the, you know, the mechanisms that LLMs, um, learn while being trained on the data. And it so happens that, you know, changing the weights to make it better at predicting conversational dialogue also makes it better at responding to user requests. Like that's, that's interesting. Um, but the answer is we don't have a great understanding and it mostly comes from just we've, we've trained it on a sequence. And so it's better at modeling that sequence. 

Pablo Srugo (35:37):

This might be a little bit outside scope, but like, you know, a lot of this is just from the vantage point of an early stage founder today and thinking about building in AI, whether like it's truly a nice AI startup or it's like AI is one of the features your SaaS, your SaaS company has. Like, do you have any thoughts about, I guess the point I'm getting at is like, it's a tailwind on one hand because of all of the productivity gains that that at leads to, and generally speaking of like that's good for technology companies, it's also like a really tough place to be because you have so much money and so many players, so much like competition, right? Like you've got to worry about where are the foundational, models gonna go, what new features is like, you know, especially on the consumer front, like open AI and chat GPT gonna put out next, it's gonna completely destroy my company. Even like, here's a silly example, like 11 labs really good at voice and then all of a sudden, you know, chat GPT you know, has a voice thing too. And so like now there's a new competitive threat. Like do you have any high level thoughts on how from, as a na, as a, as a founder today building, you should think about the different players in the AI world from the foundational models to like the incumbents and, and where you can position or, or what are the sort of things you need to think about before going out and building something that, you know, six months from now is redundant because somebody else did it better with more money? 

Nick Frosst (36:55):

Yeah, I think people should figure out what's the problem they're solving and then I think they should, after they've thought about that, think about, hey, could an LLM help me solve this problem? And if the answer is yes, then they should use an LLM to help 'em solve it. I think they shouldn't say, I really want to use an LLM, how am I gonna, how am I gonna use it? That's cool. They should instead say, what's the thing I wanna solve and then figure out if an LLM could help them. And the answer these days is often, yes, often whatever problem you wanna solve with a computer, there's probably a way that an LLM could be helping you, but it's not always. Yes. And there's sometimes they're just like, this is actually totally, totally irrelevant to that the LMS are not helpful for my business at all and you shouldn't use 'em, um, for the, that's less true. And I think less and less true as the years go on. I think at some point we'll get to a situation where companies are not AI companies the way companies are not like app companies anymore, right? Like there was a time when everybody's like, oh, we gotta make an app, we're an app company, we're gonna make an app to do this. And now it's just, you're a company and you're selling something like you, there's a good chance you have an app somewhere for some reason. Or even before that, there was a while where people were like, oh, we're an internet company, we're gonna have a website, we're an internet company. That's the thing. And now you're a company, you definitely have a website. Like it's definitely somewhere, I think we're gonna get to that stage with LLMs. People aren't gonna be differentiating themselves based on being an LLM like an AI company, an LLM company. They're just gonna use an LLM when it makes sense.

Pablo Srugo (38:10):

I think the one challenge, like, so obviously, you know, you need to start with problems and then kind of leverage technology to solve those problems. I think that's always true. Like one of the challenges, at least I see today from my vantage point is the incumbents seem to be, and I'm curious to your take on this, 'cause you're selling to a lot of enterprises, like, but they seem to be so much faster. If you go back to the internet example, like Amazon really could start with, oh here's the internet, like what should we do? We should sell books. And by the time that like, you know, the, the, the bookstores were truly selling online or like by the time Walmart really made a move to online, it was years later, like, so they got that big headstart. Doesn't seem to be the case today. Like today it seems like incumbents are a lot faster to push out these like LLM driven AI driven features. And at least for me as somebody working with a lot of early stage startups, like I'm a lot more worried about incumbent threats than I used to be or than I think others were with, you know, past kind of technology cycles. 

Nick Frosst (39:05):

I think that's an interesting point. I don't know if that point is specific to LLMs. I think that point might just be in general like yeah, the speed of development has increased, you know, and incumbents are aware of their, their ability that they could, uh, their uh, the possibility that they get disrupted and so they're moving faster. I think LLMs might be a good example of that because it's the thing that people are, you know, rushing to these days. But it doesn't sound like the point you're making is specific to this technology. I also think the point is a little true, but kind of overstated. Like we, you know, we heard the exact same thing when I was like, oh, we're gonna go make an LLM company. People were like, what the hell are you gonna do? Like, you know, Google's gonna do it first. Google's gonna encompass like I, there's a bit of a myth, especially around early founders, like the myth you're describing has always existed.

Pablo Srugo (39:49):

Always. I mean, why wouldn't Google do it? Why wouldn’t X do it is always a question. And I guess the question is like, is it more prescient for AI and LM’s or not really?

Nick Frosst (39:58):

I think it's true that everybody needs, everybody's moving faster, right? People are developing things faster. I'm like, that's cool. It's cool to see people build stuff quick. So like that's just true. But the myth you're describing as yeah, has always existed is often perpetrated by the incumbents to be like, oh, you shouldn't, you shouldn't, you shouldn't quit. We'll do it faster. You should stay here and do it here. You know? So that's been going on forever, but I've still, there's lots of companies that have come up, done something really quick and made something cool and moved faster than the incumbents and I don't expect that to change.

Pablo Srugo (40:26):

Well listen Nick, it's been, it's been great having you on the, on the show. we'll end it there. I will ask actually, you know, we tend to end with, with two questions and Yeah, they're different 'cause like what, what you're building is so different than most startups, but I'll ask them anyways, like, was there a moment or when was the moment when you felt like you had kind of true product market fit with Cohere?

Nick Frosst (40:46):

So lemme give you a maybe con contentious answer. I don't- the notion of product market fit I don't love because it's own, it's similar to fitness within the notion of fitness, within evolutionary theory in that it's only post hoc. You can look at an animal and you can say, oh, this is fit for these reasons. Like here's why this is doing really good in this niche. But it's very difficult to look at an animal and say, is it going to be fit? Or is it like, you know, is it, it's really only something you can identify after the fact. Yeah. So I think that's, I often think when early, especially early stage startups are thinking about product market fit, they're thinking like, yes, I've convinced myself we have product market fit, this is the thing. And I'm like, you're never really gonna know until you're looking back. So I think it can often mislead people. I think it's better to focus on a problem that you're passionate about and excited about and you think really needs solving, try to solve it. And one day you'll be like, well there you go. We, we got it. But it's only something you can identify going back. 

Pablo Srugo (41:48):

That's fair. But what was that? Did you have that moment? 

Nick Frosst (41:51):

Yeah, yeah, I did. Yeah, we had that moment, we've had that moment. We have bits of that moment a few times. One, when we first released our API and we saw where there was, there was people who signed up to use it and then we talked to them and they were like, yeah, it's doing this thing for me. And I got especially excited when the thing it was doing for them was boring. That's when I would get most excited. Somebody would be like, oh yeah, I'm using it and I am summarizing a bunch of documents and I no longer have to do that. And they'd be like, whoa, great. Or when somebody is saying, you know, I'm taking all of these,  you know, all of these,  like quarterly earnings reports and I'm extracting the numbers from 'em and I'm putting them in a table. And I used to do that every day. And now I don't do that. Like, those are the things that I would get really excited about. And I'd say, great. That's like real value. We're adding real value there. That's a thing somebody didn't want to do. They don't have to do it anymore. So when we first released our API, I got that a little bit. Then when we started doing private deployments and on-prem deployments and cloud deployments, I got that even more when I would see co companies would be willing to say, yeah, like, we'll take a copy of the model and we'll do this thing and that'll be great. I got that even more as we've released more features, you get that a little bit with each feature. Like we've released, you know, in our API internet search tool, use multilingual, all of these things. And every time we do it and then we see, look into what people are using with it, we're like, oh, great. That's, that's there and exciting. So yeah, we've had like bits of it, but it hasn't come in a single moment. It's come in slow increments of improving what we're building, adding features, talking to customers and them saying, Hey, we're doing this thing and it's helping. Like, that's, that's the moment and that's been, yeah, that's happened several times. 

Pablo Srugo (43:22):

Got it and then my last question, like having been on this like insane rocketship over the last five years, what would be your one, like either a big lesson learned or like one big piece of advice you'd have for yourself if you could go back, you know, four or five years ago when, when you were just starting?

Nick Frosst (43:36):

I would say chill out, <laugh> 

Pablo Srugo (43:41):

Tough. Easy to say tough. 

Nick Frosst (43:42):

Yeah,I know, I know. It's not, it's advice that, it's advice that I know if somebody told me in the first year, I would've ignored.

Pablo Srugo (43:49):

You still wake up with your heart like pounding outta your chest or?- “speaker: No, no, no.” You relaxed. Good

Nick Frosst (43:52):

Good. No, no, I, it's good. I've, I've been having a good time, but it has been. like, emotionally, I would definitely wanna speak to other first time founders and say like, it's, it's really stressful. It's gonna be really stressful. And regardless of what's actually going on, the emotional valence will be the same. So like, I, you know, when we were a six person company, the stakes felt just as high as when we are now a several hundred person company, customers all over the world, offices everywhere. You know, and like, they, the emotional impact feels the same. So you should know as you're going along that every challenge you're facing, that's the hardest challenge you've faced so far. And it's gonna feel all encompassing and then later you're gonna look back and you're gonna say how quaint, but the emotional impact will always be the same.

So you gotta, you gotta chill out, you gotta figure out how to handle it, figure out how to let the stress evaporate sometimes. Figure out how to, you know, connect, be in, in the moment and connect with your family and not, not be thinking about it constantly. 'cause it won't get easier. The stressful, like the, the things you're dealing with will just become more complicated. But if you can figure out a, a, a good cadence, then you, you, in the same way that you dealt with the early stuff, you'll be able to deal with the later stuff.

Pablo Srugo (45:09):

Yeah. I, I love hearing that from you because like, I mean, I, I felt it myself as a founder, but I think from the outside looking in, especially at a company that's gone on the trajectory, you guys have raised so much money. Like a lot of times from the outside looking in, you might think like, oh, I'm sure they're partying. I'm sure they're having a great time. And the reality is like, I mean, you probably are, but it's just, it's just as stressful. It's as hard as anybody else. 

Nick Frosst (45:28):

we're having a good time. But I wouldn't describe it as partying. <laugh> we're, yeah, we're building out and grinding and trying to, trying to keep delivering value. So. Awesome.

Pablo Srugo (45:38):

Cool man, well appreciate having you on the show. it was great. I'm sure founders learned a lot listening to you. 

Nick Frosst (45:44):

Great. Thank you for having me.

Pablo Srugo (45:46):

If you listened to this episode and the show and you like it, I have a huge favor to ask for you. Well, it's actually a really small favor, but it has huge impact. But whichever app you're listening to this episode on, take It Out, go to a product market fit show and leave a review, please. It's going to help. It's not just gonna help me to be clear. It's going to help other founders discover this show because the algorithms, whether it's Spotify, whether it's Apple, whether it's any other podcast player, one of the big things they look at is frequency of reviews. It's quantity of reviews. And the reality is, if all of you listening right now, left reviews, we would have thousands of reviews. So please take literally a minute, even if you're just writing like, great podcast, or I love this podcast, whatever it is, just write a few words. Obviously the longer the better, the more detailed the better. But write anything, leave five stars and you'll be helping me. But most importantly, many other founders just like you, discover the show. Thank you.