The Exchanges

Q How does the focus on enterprise over consumer change the way in which you train and build a model?

A So the models themselves, like transformer architecture, which is the original model that, yeah, that was introduced in Hasn't changed very much, right? Like all, the whole industry is still using transformers. We've changed the way we train them, but the model architecture itself, you know, we're approaching 10 years of, of the same model architecture. Um, when we train our model, we're not training it to be like an amazing conversationalist with you. We're not training it to like keep you interested and keep you engaged and occupied. We don't have like engagement metrics or things like that. We're just training it to Augment you in the workplace. We're just training it to help you do your job. Um, and that means the type of data we train it on is very different. So recently we started doing a bunch on like synthetic data. So we, you know, generate a whole bunch of data, um, to create like fake companies and fake emails between people at these fake companies and fake APIs within those fake companies. And then we train the model in that synthetic environment to help out within that fake business.

AI assessment note: “that means the type of data we train it on is very different.”

Q Now, before we dive into Cohere, I have to ask, you were Jeff Hinton's first hire at Google Brain, and so then you're put in a room with Jeff Hinton, and you get to work with him every day. What was the biggest lesson from working with Jeff, a legend of the industry?

A Yeah, I, I learned, yeah, I love working with Jeff. Um, I learned everything I know about research. Um, from those, those, I think we were there for four years, three years. Um, I think I was very surprised at how creatively and playfully he approaches research. Um, when we would discuss like algorithms or, or, or like optimizers or, or, uh, loss functions, we would discuss them often in like through physical analogy. So we'd spend a lot of time talking about, like, imagine there's, like, a ball here, and, like, an elastic band to this thing, and a pulley here, and, like, this is what the, you know, it's on this kind of a surface, and, like, a lot of it was descriptions in the natural, physical world, and that was very, yeah, like, playful, and a lot of it was approached with, like, oh, what would happen if, you know, with curiosity, um, and I didn't, when working with him, I didn't expect that. Um, I expected it to be much more, like, Like, you know, just, here's the equation. Let's, let's figure out what the derivative is, and let's, let's go from there. Whereas instead, a lot of it's based on, like, intuition.

AI assessment note: “I was very surprised at how creatively and playfully he approaches research.”

Q Is that not only a matter of time?

A No, no, that's not a matter of time. That's, that's fundamentally the, the way that sequence models work. Like we are training statistical models of text. They're phenomenal. They're incredible. And they, they are capable of generalizing across unseen tasks, which is why they're so useful. But the tech, yeah. When you're talking about like the 25 year old marketer or something there, some portion of their work is like writing text on a computer is like all the information's out there. They have this document and that document, this tool, that API, they just need to take that and turn it into another form and like, you know, combine it and then put it out there. Like that's the work. That's some portion of it. That's not the majority of it. Most of it is talking to people like understanding the culture. Understanding the zeitgeist, understanding like what's going to hit, what's relevant, using their intuition and their human experience to understand how they can be helpful in what they can do. And that is not in the data set of text from the internet. So yeah, I, I, yeah, I, I, I fundamentally, I, I, I disagree because you know what they do now?

AI assessment note: “No, no, that's not a matter of time.”

Q Uh, okay, final one. What trait do you love and has contributed to a lot of your success, but you're also quite wary of?

A Um, I think I'm, I'm quite, um, I'm quite curious and contrarian. Um, and that is both an asset and a hindrance. There are times when it's very helpful to be like, oh yeah, I'm like super interested in something and I learn about it and I'm like, everybody thinks this and they're totally wrong. That's super helpful. Um, that's why when Aiden was like, hey, do you want to found a company on language models back in 2019? I was like, yeah, absolutely. You know, that was not a view that was widespread. Not at all. Like even, so I'm, yeah, curious and contrarian, but there's other times when like the whole world has been definitely right, and I've been wrong, and I've been like, oh yeah, I was super excited about this.

AI assessment note: “I'm quite curious and contrarian. Um, and that is both an asset and a hindrance.”

Q How does the focus on enterprise over consumer change the way in which you train and build a model?

A So the models themselves, like transformer architecture, which is the original model that, yeah, that was introduced in Hasn't changed very much, right? Like all, the whole industry is still using transformers. We've changed the way we train them, but the model architecture itself, you know, we're approaching 10 years of, of the same model architecture. Um, when we train our model, we're not training it to be like an amazing conversationalist with you. We're not training it to like keep you interested and keep you engaged and occupied. We don't have like engagement metrics or things like that. We're just training it to Augment you in the workplace. We're just training it to help you do your job. Um, and that means the type of data we train it on is very different. So recently we started doing a bunch on like synthetic data. So we, you know, generate a whole bunch of data, um, to create like fake companies and fake emails between people at these fake companies and fake APIs within those fake companies. And then we train the model in that synthetic environment to help out within that fake business.

AI assessment note: “the type of data we train it on is very different.”

Q Do you not think it's helpful for founders to be very aware of competitive landscapes in case they're asked about them by customers, in case customers are going, hey, why aren't you more open? Why aren't you more closed? Are we, is our data secure if you're?

A This, as in, as with Most things like a middle ground is the right place to be, right? You could spend your whole time as a founder only looking at competitors and being like, oh, why are they doing that? Why are they doing this? What's going on with that? Yeah. Um, and that will, I think not be helpful for you. You could also spend your whole time, you know, with your head in the sand only thinking about what's going on in your company. And I think that would not be helpful either. You have to find some middle ground. The discourse around AI is inescapable. You know, you, you would be hard pressed to ignore it. It is every other headline. You know, it is, it's all over the place. I don't think there are many people who work in the industry who suffer from not enough information about what's going on in AI. Right. I think there's a lot of people who suffer from way too much of it and obsessing over the minute details of like, you know, how, Well, so-and-so got .2% better on this thing, or like, you know, is constant, like constant small changes in businesses out there. Um, and I think that can mislead you from staying grounded in like, what are you actually doing? Who are you actually helping? How is this making, you know, things better for your customers?

AI assessment note: “as with Most things like a middle ground is the right place to be”

Q The question that everyone asks is how do you compete against competitors who have billions and billions of dollars? Do you hate that question and how do you respond to it?

A No, I don't hate that question. I think, yeah, I think that's a fine question. Um, you know, we've, we've announced funding rounds, um, and you can see that they are smaller than some of the other funding rounds out there. Um, yeah, I mean, like we're pretty singularly focused in a way that the other companies who build foundational models, uh, are not, right? Like we don't have a consumer app. We're not trying to get anybody to spend 200 dollars a month on something for their personal lives. We're singularly focused on working with enterprises and businesses and making sure that they get to production with AI. I hope my whole, I'm constantly telling people like, not AGI, ROI, ROI, not AGI. Yeah. Um, and you know, it turns out there's a lot of work that still needs to get done there. And it turns out there's a lot of companies that tried to go to production with AI using something.

AI assessment note: “No, I don't hate that question.”

Q Is that not only a matter of time?

A No, no, that's not a matter of time. That's, that's fundamentally the, the way that sequence models work. Like we are training statistical models of text. They're phenomenal. They're incredible. And they, they are capable of generalizing across unseen tasks, which is why they're so useful. But the tech, yeah. When you're talking about like the 25 year old marketer or something there, some portion of their work is like writing text on a computer is like all the information's out there. They have this document and that document, this tool, that API, they just need to take that and turn it into another form and like, you know, combine it and then put it out there. Like that's the work. That's some portion of it. That's not the majority of it. Most of it is talking to people like understanding the culture. Understanding the zeitgeist, understanding like what's going to hit, what's relevant, using their intuition and their human experience to understand how they can be helpful in what they can do. And that is not in the data set of text from the internet. So yeah, I, I, yeah, I, I, I fundamentally, I, I, I disagree because you know what they do now?

AI assessment note: “No, no, that's not a matter of time. That's, that's fundamentally”

Q Does FTEs not just allow for poor technology?

A What I mean by that is like, no, no, I think that there's this, there's this idea Is this idea sometimes like, oh yeah, you can just make the thing and for every business it'll be, it'll work perfectly and require no engagement. And like, that's not the way, that's the way some technology works. That's the way a lot of consumer technology works. That's not the way a lot of enterprise technology works, right? Like there's, you're selling things to people, um, that are, that have to be like matched to the way their business is set up. Right. And so having engineers go along with it and say, cool, here, here's, here's the model. Here's what we can do to make sure that that's perfect for you in your specific use case is helpful.

AI assessment note: “That's not the way a lot of enterprise technology works”

Q Do you not think it's helpful for founders to be very aware of competitive landscapes in case they're asked about them by customers, in case customers are going, hey, why aren't you more open? Why aren't you more closed? Are we, is our data secure if you're?

A This, as in, as with Most things like a middle ground is the right place to be, right? You could spend your whole time as a founder only looking at competitors and being like, oh, why are they doing that? Why are they doing this? What's going on with that? Yeah. Um, and that will, I think not be helpful for you. You could also spend your whole time, you know, with your head in the sand only thinking about what's going on in your company. And I think that would not be helpful either. You have to find some middle ground. The discourse around AI is inescapable. You know, you, you would be hard pressed to ignore it. It is every other headline. You know, it is, it's all over the place. I don't think there are many people who work in the industry who suffer from not enough information about what's going on in AI. Right. I think there's a lot of people who suffer from way too much of it and obsessing over the minute details of like, you know, how, Well, so-and-so got .2% better on this thing, or like, you know, is constant, like constant small changes in businesses out there. Um, and I think that can mislead you from staying grounded in like, what are you actually doing? Who are you actually helping? How is this making, you know, things better for your customers?

AI assessment note: “as with Most things like a middle ground is the right place to be”

Q The question that everyone asks is how do you compete against competitors who have billions and billions of dollars? Do you hate that question and how do you respond to it?

A No, I don't hate that question. I think, yeah, I think that's a fine question. Um, you know, we've, we've announced funding rounds, um, and you can see that they are smaller than some of the other funding rounds out there. Um, yeah, I mean, like we're pretty singularly focused in a way that the other companies who build foundational models, uh, are not, right? Like we don't have a consumer app. We're not trying to get anybody to spend 200 dollars a month on something for their personal lives. We're singularly focused on working with enterprises and businesses and making sure that they get to production with AI. I hope my whole, I'm constantly telling people like, not AGI, ROI, ROI, not AGI. Yeah. Um, and you know, it turns out there's a lot of work that still needs to get done there. And it turns out there's a lot of companies that tried to go to production with AI using something.

AI assessment note: “we're pretty singularly focused in a way that the other companies... are not”

Q Can I ask you, when we think about, like, problems to solve, I think a lot of people also get worried about the open versus closed document. Um, how do you feel about where the future of efficient AI lands in the balance between open versus closed models?

A So at Cohere, we, we make our foundational models, and then we release the weights for non-commercial usage. So we're somewhere in the, in the between the like open and closed, right? We're a for-profit company. Like we, we exist to make money. Um, and so we don't, we release our weights for scientific and research and like, you know, you can download it on your computer and run it. Um, I think that's a good sweet spot for us as a business that allows us to like, you know, build credibility within the community. If people want to check out our weights, like they can go check them out, right? Like there's lots of companies that started out as open. Um, who no longer released the weights of their models or who never did, right? So we have our models out there. You can go look at them. You can use them. You can validate it. Do they work on my problem? Yes or no. Um, but if you're using them for commercial purposes, you got to talk to us and then we figure out a commercial relationship so that we can, you know, exist as, as a business that works for us. Um, I'm surprised to, I'm surprised there aren't more businesses taking that tact. And more foundational models taking that approach.

AI assessment note: “I'm surprised there aren't more businesses taking that tact.”

Q What's the biggest disagreement that you and Aiden have had?

A We had some different disagreements about like API design. Like that was, there was a brief moment like before RLHF where we were talking about like, oh, we should make an endpoint for summarization, an endpoint for entity extraction or something. And I think we disagreed about that. So we disagreed on like some low level stuff, but beyond that, you know, like I've had the privilege of getting to work with, yeah, both, both Aiden and Ivan, the other co-founders, like I, you know, I have a huge amount of respect for, and we definitely disagree and argue. About, like, the, the, the little things about how to run the business. Should we do this? Or, you know, should we make that policy? But there hasn't been any, like, huge.

AI assessment note: “there hasn't been any, like, huge.”

Q Wow. I love that. That's awesome. Go to McDonald's. What's the worst thing that could happen with regulation towards AI?

A Out of an erroneous understanding of the technology, um, And, uh, you know, like thinking that what we're building is, is digital gods, um, which is like large languages, models are not. Uh, but if you think that that's what they're building, um, then you could think, okay, cool. We need to, you know, come up with benchmarks around existential threats. Um, and I think there are times when looking at those at benchmarks, like, you know, fixation on particular benchmarks, which can be gamed and can be trained either to do way better on or way worse on. Um, are not helpful for establishing how the technology can be used and misused. Um, so I think if you were, like, the worst thing a regulation could do would say, hey, we're gonna pick this random benchmark, we think that represents AGI, and we're gonna, you know, shut down any development on it. I think that would be, that would be a misplay.

AI assessment note: “the worst thing a regulation could do would say, hey, we're gonna pick this random benchmark”

Q When you think about, kind of, the three pillars of compute algorithms and data, which one do you think is most Constrained or the biggest bottleneck?

A It's interesting. I mean, the algorithms haven't changed very much. Like that's interesting. Like they've changed a little bit. You know, when we started this industry, you know, originally we were just training base models, which are not called base models at the time. They were just called large language models. Um, but they weren't trained from human feedback. So all they would do is, you know, take in the first part of a sentence and write the second part of a sentence. Um, but if you tried to have a conversation with them, it wouldn't work because that wasn't the data they were trained on. Um, Since then, you know, now we, we train models in a few different steps. There's like a base modeling step. Then there's a reinforcement learning step from human feedback with SFT data. After that, you know, there, there might, there's a variety of other reinforcement learning techniques you can do. Um, but the algorithms I, I think are not the bottleneck in terms of making those models more useful. I do think a lot of it is still getting good quality data and then making good quality synthetic data From your good quality real data.

AI assessment note: “I do think a lot of it is still getting good quality data”

Q What's the biggest disagreement that you and Aiden have had?

A We had some different disagreements about like API design. Like that was, there was a brief moment like before RLHF where we were talking about like, oh, we should make an endpoint for summarization, an endpoint for entity extraction or something. And I think we disagreed about that. So we disagreed on like some low level stuff, but beyond that, you know, like I've had the privilege of getting to work with, yeah, both, both Aiden and Ivan, the other co-founders, like I, you know, I have a huge amount of respect for, and we definitely disagree and argue. About, like, the, the, the little things about how to run the business. Should we do this? Or, you know, should we make that policy? But there hasn't been any, like, huge.

AI assessment note: “We had some different disagreements about like API design.”

Q When you think about, kind of, the three pillars of compute algorithms and data, which one do you think is most Constrained or the biggest bottleneck?

A It's interesting. I mean, the algorithms haven't changed very much. Like that's interesting. Like they've changed a little bit. You know, when we started this industry, you know, originally we were just training base models, which are not called base models at the time. They were just called large language models. Um, but they weren't trained from human feedback. So all they would do is, you know, take in the first part of a sentence and write the second part of a sentence. Um, but if you tried to have a conversation with them, it wouldn't work because that wasn't the data they were trained on. Um, Since then, you know, now we, we train models in a few different steps. There's like a base modeling step. Then there's a reinforcement learning step from human feedback with SFT data. After that, you know, there, there might, there's a variety of other reinforcement learning techniques you can do. Um, but the algorithms I, I think are not the bottleneck in terms of making those models more useful. I do think a lot of it is still getting good quality data and then making good quality synthetic data From your good quality real data.

AI assessment note: “I do think a lot of it is still getting good quality data”

Q yeah. I mean, I have Benioff on the show from Salesforce two days ago, and he says, like, oh, the same human plus agent. Yeah, yeah, yeah. Are you serious? Most 25, twenty-six-year-old marketing managers or SDRs, I'm sorry to say it, they're not brilliant. They do not love the craft. They are not better than a phenomenal agent will be in the next 12 months. They will be replaced.

A Oh no. Yeah. I actually believe that. Yeah. No, I believe what I said. Sorry. Yeah. Yeah. No, I, yeah, I'm not, uh, yeah, yeah. I fundamentally believe that this technology Is, and I think this is clear if you use it, you know, use it for a while, um, that there are things that it's way better at you, better than you at. Um, but there are still lots of things that people are better at. Like, look, LLMs are incredible people. People have been using them, you know, for years now. There has been no independent breakthrough that an LLM has made. Right. There has been no, like nobody has seen, nobody asked in LLM, Hey, solve this problem. No one's solved before and, and get the, the answer. The breakthroughs are still people.

AI assessment note: “there are still lots of things that people are better at”

Q Can I ask you, when we think about, like, problems to solve, I think a lot of people also get worried about the open versus closed document. Um, how do you feel about where the future of efficient AI lands in the balance between open versus closed models?

A So at Cohere, we, we make our foundational models, and then we release the weights for non-commercial usage. So we're somewhere in the, in the between the like open and closed, right? We're a for-profit company. Like we, we exist to make money. Um, and so we don't, we release our weights for scientific and research and like, you know, you can download it on your computer and run it. Um, I think that's a good sweet spot for us as a business that allows us to like, you know, build credibility within the community. If people want to check out our weights, like they can go check them out, right? Like there's lots of companies that started out as open. Um, who no longer released the weights of their models or who never did, right? So we have our models out there. You can go look at them. You can use them. You can validate it. Do they work on my problem? Yes or no. Um, but if you're using them for commercial purposes, you got to talk to us and then we figure out a commercial relationship so that we can, you know, exist as, as a business that works for us. Um, I'm surprised to, I'm surprised there aren't more businesses taking that tact. And more foundational models taking that approach.

AI assessment note: “I'm surprised there aren't more businesses taking that tact”

Q When you look at Google Brain, and you look at DeepMind, a lot think that really, kind of, Google were asleep at the wheel, given them not being at the forefront in what was the consumerization of it with ChatGPT. Do you think that's fair?

A I don't know. I mean, it's certainly interesting. Look like the transformer was invented at Google, right? Like there was the, in 2017, Aiden, uh, amongst with many other brilliant people in Google brain published the transformer as an architecture. Um, it wasn't, it wasn't then commercialized very quickly within Google. It wasn't scaled up very quickly within Google. Um, a lot of that work had to be done elsewhere and years later. So that's interesting. Like what, why that is like what, what systems are in place to make that be the case? I don't know. I will say there's still a ton of brilliant people in deep mind. I think now it's, it's consumed the rest of it doing great work. Um, and they continue to make good products. Uh, it is interesting that all the people who worked on the transformer left To continue to work on the transformer.

AI assessment note: “It wasn't, it wasn't then commercialized very quickly within Google.”

Q Does FTEs not just allow for poor technology?

A What I mean by that is like, no, no, I think that there's this, there's this idea Is this idea sometimes like, oh yeah, you can just make the thing and for every business it'll be, it'll work perfectly and require no engagement. And like, that's not the way, that's the way some technology works. That's the way a lot of consumer technology works. That's not the way a lot of enterprise technology works, right? Like there's, you're selling things to people, um, that are, that have to be like matched to the way their business is set up. Right. And so having engineers go along with it and say, cool, here, here's, here's the model. Here's what we can do to make sure that that's perfect for you in your specific use case is helpful.

AI assessment note: “no, no, I think that there's this, there's this idea”

Q What will the changes be? Like, what will the company look like, do you think, in five to 10 years?

A I think you will arrive to work, and you will sit in front of a computer, um, and you will predominantly use language to interact with that computer, and any time there's something to do that you know can be done, and you know, like, the information is out there, it doesn't require, you know, creativity or insight, You know that it's there. Um, you just need to do it, and doing it's kind of boring. You will get the model to do it for you. And I think that looks like, that mostly looks like sitting down, speaking to the computer, getting it to do the things you don't want to do, and then spending your time talking to other people, thinking about how it can be useful, whether or not what it did was good. And I, I think that shift is, is Is maybe chaotic. And I would like us as a world to be spending time thinking about, you know, how can we make sure that changes as easy as possible? How can we make sure that making language models allows people to do the stuff that they're good at and that they like? Um, how can we make sure like the labor force is resilient? Uh, how can we make sure, you know, that income inequality doesn't go up as a result of that? Like, those are the types of things I would like us to be talking about. And those are the important things. I want to go back to your earlier question, like when talking about You know, the eight risks of AGI, like I think those e…

AI assessment note: “you will sit in front of a computer, um, and you will predominantly use language”

Q Do you just think then that OpenAI and Anthropik will just seed enterprise?

A I think like right now, you know, those, both those companies have a, are pretty cool. They've both made good consumer products. Um, I think where this technology like adds the most value is in work for like personal reasons. Like that's where I see this technology being the most useful. Um, I don't know if they'll, if they'll start working on that. Um, I know that making models that work in that environment is pretty different than making a model that works in a consumer. Environment. In a consumer environment, you can make the biggest model possible. You can have, like, complicated switches to tell you to go to this model or that model because you're, you're just posting it on a huge amount of GPUs. You can be, like, losing a ton of money on every inference call, but, you know, you're getting, you're getting users and something, and so, like, that works. The types of models you have to build to succeed there are different. Um, I know the work you need to do on the interface, like, you know, we've, we've announced North, which is our agentic framework. It's privately deployable. Customizable, you know, for, uh, knowledge work workers within an enterprise is pretty diff. It looks pretty different than, uh, some of the consumer applications, right? Like a big one is like our models and generate images. Nobody in the workforce is really wanting to generate images as part of their…

AI assessment note: “I don't know if they'll be interested in that at some point.”

Q pace of deployment, we are seeing model evolution so fast and so rapidly that you're essentially seeing this kind of decay rate on models being greater than ever because it's like next one, next one, next one, and actually they're still being trained though on H-one hundreds or Nvidia chips from 18 months ago. Is there a misalignment in terms of the progression of models versus the progression of chips?

A You can cycle through new versions of models quicker. I mean, it's still, it's very slow. It's still, I got, when I was training neural nets in 2000 and I don't know, 11, and it would take like, you know, hours to days. I remember being like, this is crazy. I can't believe this takes so long to train this model. Now we spend months, months training models. So like, you know, that's, that's a time scale I, I didn't anticipate when I was working on this. Um, A long time when I was working on neural nets a long time ago. Um, but that's still very different than the timescale of working on chips, right? Like that's still slow. I think when you talk about like we're seeing all these models iterate so quickly, like yes, on the one hand, we're seeing models iterate really quickly and people are releasing new models. On the other hand, they're still the transformer that was invented in 2017. And they're still sequence models and they still take in words and predict the next word. And we've changed how they're trained a bit. Uh, we've added on steps. Like now there's a base modeling step, then a SFT, like supervised fine tuning from human feedback, or like somebody writes a sentence and then writes the second, the response they want. And we train on that. And then there's a reinforcement learning aspect where the model is generating and you're telling it that's good, that's bad or somet…

AI assessment note: “that's still very different than the timescale of working on chips, right?”

Q Wow. I love that. That's awesome. Go to McDonald's. What's the worst thing that could happen with regulation towards AI?

A Out of an erroneous understanding of the technology, um, And, uh, you know, like thinking that what we're building is, is digital gods, um, which is like large languages, models are not. Uh, but if you think that that's what they're building, um, then you could think, okay, cool. We need to, you know, come up with benchmarks around existential threats. Um, and I think there are times when looking at those at benchmarks, like, you know, fixation on particular benchmarks, which can be gamed and can be trained either to do way better on or way worse on. Um, are not helpful for establishing how the technology can be used and misused. Um, so I think if you were, like, the worst thing a regulation could do would say, hey, we're gonna pick this random benchmark, we think that represents AGI, and we're gonna, you know, shut down any development on it. I think that would be, that would be a misplay.

AI assessment note: “the worst thing a regulation could do would say... shut down any development”

Q What will the changes be? Like, what will the company look like, do you think, in five to 10 years?

A I think you will arrive to work, and you will sit in front of a computer, um, and you will predominantly use language to interact with that computer, and any time there's something to do that you know can be done, and you know, like, the information is out there, it doesn't require, you know, creativity or insight, You know that it's there. Um, you just need to do it, and doing it's kind of boring. You will get the model to do it for you. And I think that looks like, that mostly looks like sitting down, speaking to the computer, getting it to do the things you don't want to do, and then spending your time talking to other people, thinking about how it can be useful, whether or not what it did was good. And I, I think that shift is, is Is maybe chaotic. And I would like us as a world to be spending time thinking about, you know, how can we make sure that changes as easy as possible? How can we make sure that making language models allows people to do the stuff that they're good at and that they like? Um, how can we make sure like the labor force is resilient? Uh, how can we make sure, you know, that income inequality doesn't go up as a result of that? Like, those are the types of things I would like us to be talking about. And those are the important things. I want to go back to your earlier question, like when talking about You know, the eight risks of AGI, like I think those e…

AI assessment note: “I think you will arrive to work, and you will sit in front of a computer”

Q Do you just think then that OpenAI and Anthropik will just seed enterprise?

A I think like right now, you know, those, both those companies have a, are pretty cool. They've both made good consumer products. Um, I think where this technology like adds the most value is in work for like personal reasons. Like that's where I see this technology being the most useful. Um, I don't know if they'll, if they'll start working on that. Um, I know that making models that work in that environment is pretty different than making a model that works in a consumer. Environment. In a consumer environment, you can make the biggest model possible. You can have, like, complicated switches to tell you to go to this model or that model because you're, you're just posting it on a huge amount of GPUs. You can be, like, losing a ton of money on every inference call, but, you know, you're getting, you're getting users and something, and so, like, that works. The types of models you have to build to succeed there are different. Um, I know the work you need to do on the interface, like, you know, we've, we've announced North, which is our agentic framework. It's privately deployable. Customizable, you know, for, uh, knowledge work workers within an enterprise is pretty diff. It looks pretty different than, uh, some of the consumer applications, right? Like a big one is like our models and generate images. Nobody in the workforce is really wanting to generate images as part of their…

AI assessment note: “I don't know if they'll, if they'll start working on that.”

Q So going back, why does it matter that you have a generational company?

A After we detoured it to Adam Smith in the original head. Yeah, seriously. Yeah. Um, one thing that comes to mind is like, you know, like, Look upon my works in despair, like nothing, you know, Ozymandias and the idea of like people obsessing over their legacy and building, you know, you know, some statue to their grandeur. And like one day it will also fall, right? Like one day all that will be left are two legs in the desert, you know, like that's true. Um, and that's true regardless of what you build. Like that's true at some point. Um, but when I think about building it, like what excites me about building Cohere, Um, and when I say like a generational company, um, I mean timescale generations. I don't mean like my generations, you know. Um, I mean the idea of building something that is there for a long time. Um, it's rewarding, and I think it's inherently human. You know, like I think we all like to think about what are we building? How long is it going to be there? And whether that's like a work of art or an actual building or a company or a philosophy or an idea, the idea of building something Or participating in the construction of something that is bigger than you is rewarding.

AI assessment note: “participating in the construction of something that is bigger than you is rewarding.”

Q You don't think it's real that Anthropik are paying 1015, twenty million dollars for great AI researchers?

A I have no idea. I know that there are lots of people who are adding that much value, you know, and there's lots of people who are, like, bringing that much value into the industry. It's a super, it's a super impactful industry, right? Um, so I know that there are people that are bringing that much value, um, and I know that there's lots of brilliant people, and I know that This is a really, it's really demanding work. It's really hard work. It requires a lot of experience, a lot of, uh, ingenuity, um, and a lot of dedication. So I think it's a good place for, for people to be spending their time. And I, and I'm, I think it makes sense that many of them are rewarded very well. That being said, like when I see the stories of, you know, Meta hiring people for like a hundred million, like I, I, I read as many stories of those as I read of people leaving the next day. So I don't, I don't know what's going on over there. Um, I know that people like to work in a place that is, like, that gives them, uh, stability, and people like to work in a place that gives them, like, purpose and is value aligned, um, and they, they like to work in places that make them feel good about what they're doing, and part of that is compensation.

AI assessment note: “I have no idea. I know that there are lots of people who are adding”

Q So going back, why does it matter that you have a generational company?

A After we detoured it to Adam Smith in the original head. Yeah, seriously. Yeah. Um, one thing that comes to mind is like, you know, like, Look upon my works in despair, like nothing, you know, Ozymandias and the idea of like people obsessing over their legacy and building, you know, you know, some statue to their grandeur. And like one day it will also fall, right? Like one day all that will be left are two legs in the desert, you know, like that's true. Um, and that's true regardless of what you build. Like that's true at some point. Um, but when I think about building it, like what excites me about building Cohere, Um, and when I say like a generational company, um, I mean timescale generations. I don't mean like my generations, you know. Um, I mean the idea of building something that is there for a long time. Um, it's rewarding, and I think it's inherently human. You know, like I think we all like to think about what are we building? How long is it going to be there? And whether that's like a work of art or an actual building or a company or a philosophy or an idea, the idea of building something Or participating in the construction of something that is bigger than you is rewarding.

AI assessment note: “participating in the construction of something that is bigger than you is rewarding.”