Latest / The Neon Show / The Al Lab Founder Who Sees The Future First | Karan Goel, Founder & CEO of Cartesia

Transcript

The Al Lab Founder Who Sees The Future First | Karan Goel, Founder & CEO of Cartesia · The Neon Show

0:00
SPEAKER_02: Voice is going to be super important as one of the core interfaces.
0:03
SPEAKER_02: So we have to start by investing in that and understanding it.
0:06
SPEAKER_00: The journey has been pretty interesting.
0:08
SPEAKER_00: DPSRK Purang, then you went to IIT Delhi, came out as a researcher, came to Carnegie to continue your research, came to St.
0:14
SPEAKER_00: Stanford to do your PhD, and then started a company.
0:18
SPEAKER_02: It didn't feel as valuable to do the N plus 1S academic paper and actually felt a lot more important to try to figure out how to use the technology work that we've done and obviously pushed it forward in a more practical way.
0:31
SPEAKER_02: And so that kind of was what brought us into starting the company, which is Cartesia.
0:36
SPEAKER_02: What do I need to build to build the first version of a truly high-quality interactive system?
0:41
SPEAKER_02: A how are you communicating with the user?
0:44
SPEAKER_02: And obviously, for communicating synchronously, voice is the main interface.
0:48
SPEAKER_02: So that was something that was from first principles pretty clear that we needed to do.
1:29
SPEAKER_00: Today I have with me Karan Goyal, founder of Kartisha.
1:32
SPEAKER_00: Karan, welcome on the Neon Show.
1:33
SPEAKER_02: Yeah, great to be here.
1:34
SPEAKER_02: Thanks so much for having me.
1:35
SPEAKER_00: Yeah, uh you know, great thing to know that you know you spent your childhood uh and college in Delhi.
1:41
SPEAKER_00: Like uh I did my internship.
1:42
SPEAKER_00: I built my company out of Delhi and my hometown was like two hours away.
1:46
SPEAKER_02: Yeah.
1:46
SPEAKER_02: No, Delhi is an amazing place to go up.
1:48
SPEAKER_00: Yeah.
1:49
SPEAKER_00: I think apart from the the poor air quality, which is for the last two years, I think everything in in Delhi is about pretty nostalgic.
1:57
SPEAKER_02: Yeah, it is.
1:58
SPEAKER_02: I I like going back and uh the food is amazing.
2:00
SPEAKER_00: The food is still still one of the best, you know, you can get as an Indian anywhere in the world.
2:05
SPEAKER_00: Right.
2:06
SPEAKER_00: Uh your journey has been pretty interesting, right?
2:09
SPEAKER_00: Uh DPS RK Puram, which is famous or infamous school.
2:13
SPEAKER_00: Then you went to IIT Delhi, right?
2:14
SPEAKER_00: Uh came out as a researcher, came to Carnegie uh to continue your research, came to Stanford to do your PhD, and then started a a company.
2:25
SPEAKER_00: Uh I believe you know uh like famous PhDs.
2:30
SPEAKER_00: You completed your PhD, but famous PhD dropouts are like Sergei and Larry, and and they were also doing research.
2:35
SPEAKER_00: So I think it's the journey is pretty incredible.
2:39
SPEAKER_02: Yeah, I think uh, you know, I I um I when I was a kid, I used to be pretty into uh you know gaming.
2:46
SPEAKER_00: Sure.
2:46
SPEAKER_02: Um and that's kind of where it started for me actually.
2:50
SPEAKER_02: Like I used to be pretty focused on uh uh you know, I with my dad I used to build computers and stuff like that, and I I really got into um video games.
3:01
SPEAKER_02: And you know, it's it's pretty fun to, you know, the imagination is is part of uh video games, is something that I was always into.
3:09
SPEAKER_02: That was also a time when I uh first bought an Nvidia GPU, which is yeah.
3:16
SPEAKER_02: Uh back in I think uh maybe 07.
3:19
SPEAKER_02: I I might be off by you know a year or two.
3:22
SPEAKER_00: I think you can still sell it for a few hundred thousand dollars.
3:25
SPEAKER_00: Yeah, yeah, yeah.
3:26
SPEAKER_02: I don't know if it runs hair models pretty effectively, but uh they definitely ran desktop games pretty well.
3:34
SPEAKER_02: Uh and used to buy you know these GPUs out of Nehru Place in uh in Delhi.
3:39
SPEAKER_02: And uh so I used to assemble desktops and and I I used to do a lot of gaming.
3:43
SPEAKER_02: And I think that's also how I got into um uh technology more uh and uh you know I used to actually play games like Counter-Strike on uh multiplayer on uh uh you know India servers, but also Singapore servers in the morning.
4:00
SPEAKER_02: I used to wake up early in the morning before school, get on the server because that's when like it's still you know very early in India, but Singapore is online.
4:09
SPEAKER_02: Uh and uh and that's like how I got you know excited about sort of like uh uh you know doing a lot of these things around even like AI and technology, right?
4:20
SPEAKER_02: Like a lot of I think for a lot of people in AI, games are one of the ways that they got into it.
4:26
SPEAKER_00: Yeah.
4:26
SPEAKER_02: And so I think it was similar for me.
4:28
SPEAKER_00: So building the company, uh, what part of your journey it was clear to you that you that you want to build a company which is focused on research?
4:35
SPEAKER_02: Yeah, it's interesting because like I didn't want to do anything except research for a while.
4:43
SPEAKER_02: Um, you know, when I was in my third or fourth year, I think I had a conversation with my advisor about graduating.
4:50
SPEAKER_02: And um and I think I I sort of actually said that I wanted to stay longer to do the PhD, which seems strange, but I think I felt like I was learning a lot.
5:02
SPEAKER_02: Yeah um and I I like research.
5:05
SPEAKER_02: Yeah, you know, I think research is great as an intellectual pursuit.
5:09
SPEAKER_02: Yeah, I think it's fun to do research.
5:11
SPEAKER_02: You know, it's exciting.
5:12
SPEAKER_02: You are discovering new things, you're trying to push the field forward.
5:16
SPEAKER_02: I think it's just enjoyable, right?
5:18
SPEAKER_02: And I think doing research in academia is super fun, actually.
5:21
SPEAKER_02: So I just wanted to do more of that for a while.
5:24
SPEAKER_02: Uh, but eventually, I think what we realized is we'd made enough progress where um it didn't feel as valuable to do the N plus one S academic paper.
5:34
SPEAKER_02: Yeah.
5:35
SPEAKER_02: And it actually felt a lot more important to try to figure out how to use the technology work that we'd done and obviously push it forward in a more practical way.
5:43
SPEAKER_02: I think that was the time where we were like, we had those, you know, sort of like conversations amongst ourselves, and we were like, yeah, like it feels like we need to do this in a different environment now.
5:56
SPEAKER_02: Um, and so that was kind of what prompted it.
5:58
SPEAKER_02: You know, I think obviously like we're pretty ambitious.
6:01
SPEAKER_02: Yeah.
6:01
SPEAKER_02: Uh all of us want to work on things that are world-changing.
6:05
SPEAKER_02: Um, nothing less gets us out of bed, right?
6:08
SPEAKER_02: So I think for us, like it was pretty important to be oriented towards that level of ambition, which is we are going to be able to use this technology for something that's a lot bigger and uh and very impactful for a lot of people.
6:21
SPEAKER_02: And so that kind of what was what brought us into starting the company, which is Cartesia, and uh and then obviously like building it, right?
6:32
SPEAKER_02: And I think obviously I can talk a lot about that, and I'm sure we will, but that's I think that's where the journey stops for Kern, the researcher.
6:40
SPEAKER_02: Yeah, or at least you know, you're a researcher lifelong, but uh as a as a graduate student, let's say.
6:47
SPEAKER_00: Yeah.
6:48
SPEAKER_00: And and what prompted you Lessive, and why, first of all, why the name Cartesia?
6:52
SPEAKER_02: Yeah, so Cartesia is pretty simple.
6:54
SPEAKER_02: Like um, you know, we are um I'm sure you've heard of the Cartesian plane.
6:59
SPEAKER_02: Yeah.
7:00
SPEAKER_02: And Descartes was uh pretty famous mathematician.
7:05
SPEAKER_02: So, you know, we're technical people, we wanted to have a technical name that's an homage to someone.
7:11
SPEAKER_02: Yeah, it's kind of like that, right?
7:12
SPEAKER_02: Like it's uh it's just an homage to someone that we think is great.
7:15
SPEAKER_00: Yeah, yeah, amazing.
7:17
SPEAKER_00: And at what exact point of time did uh all your co-founders come together?
7:22
SPEAKER_00: Because it's hard to get like five folks together to start a camp.
7:26
SPEAKER_02: Yeah, we actually came together pretty much at the beginning.
7:28
SPEAKER_02: And I think the nice thing was we all knew each other.
7:31
SPEAKER_02: We um incubated in Chris's lab, right?
7:34
SPEAKER_02: And uh Chris is one of our co-founders, as well as like um, you know, obviously we're um his four like grad students.
7:42
SPEAKER_02: So so we spun out of the lab.
7:44
SPEAKER_02: I think uh the four of us really wanted to build this company uh together, and we we sort of like knew each other, right?
7:52
SPEAKER_02: Like pretty well.
7:53
SPEAKER_02: Like me and Albert have worked together since 2019 now, so eight years, close to eight years.
7:58
SPEAKER_02: Um, I worked with all of these other folks for many years.
8:01
SPEAKER_02: So we we we knew each other well, and um and in fact, like you know, we started the company and then I got married, you know, a few months later, and they were all at my wedding.
8:13
SPEAKER_02: So we were pretty good friends, right?
8:15
SPEAKER_02: Outside of work.
8:16
SPEAKER_02: So it was pretty natural for us to come together.
8:18
SPEAKER_02: So you got married like in in your early PhD?
8:21
SPEAKER_02: No, I actually got married right after I graduated.
8:23
SPEAKER_02: Okay.
8:23
SPEAKER_02: Before the company got, you know, yeah, got uh right before I see drown actually.
8:29
SPEAKER_02: Okay.
8:30
SPEAKER_02: But uh, but so we we came together as a group naturally pretty early, and I think we were all excited to work together.
8:37
SPEAKER_00: That was the Andween the four of you, how did you split the responsibilities when you're deciding now there's a real company?
8:44
SPEAKER_02: Yeah, so um, you know, I I um think that one of the cool things about the founders is that we're all super technical, but I think every founder is pretty adaptive and a really good generalist.
9:00
SPEAKER_02: Um, they're able to do many things.
9:02
SPEAKER_02: So when we were early, we um were actually able to sort of like work together pretty effectively, I think, to um take on different responsibilities as needed because all of us have this mentality of like do what it takes.
9:20
SPEAKER_02: Yeah, like none of us are like, hey, like I only do X or Y.
9:24
SPEAKER_02: Like we're very, very open to doing more things.
9:28
SPEAKER_02: And so I think that made it really easy for us to mold ourselves to sort of different problems over time.
9:34
SPEAKER_02: And I I think over time actually we've worked on uh each person has worked on many different things over the the course of the company.
9:42
SPEAKER_02: Albert's our chief scientist, so he might be the one person who is like super focused on research, but the rest of us have sort of like actually uh done a fair few things over the over the last two years.
9:54
SPEAKER_02: Okay.
9:54
SPEAKER_02: Yeah.
9:55
SPEAKER_00: Okay.
9:55
SPEAKER_00: How did your pre-seed happen?
9:58
SPEAKER_02: Um, so you know, we incubated in in Chris's lab, and so we we sort of raised some pre-seed capital as a small research team.
10:09
SPEAKER_01: Okay.
10:09
SPEAKER_02: From a couple of funds like Factory and Conviction, um, and um, you know, um Lightspeed as well was part of that.
10:17
SPEAKER_02: Uh it was just like pre seed capital to get started.
10:21
SPEAKER_02: This was when the company was an idea and less a company.
10:26
SPEAKER_00: September 2023?
10:27
SPEAKER_02: This was, yeah, this was around September 2023.
10:30
SPEAKER_02: So we were still in this phase of like we don't quite know what we want to do yet.
10:36
SPEAKER_02: We know we want to build um a model lab and we want to focus on uh how we can bring new technology to build great, you know, sort of smart interactive models, but we hadn't really decided on the direction.
10:53
SPEAKER_02: And then once we um you know headed into 2024 and I got married, and you know, that brought a lot of clarity.
11:00
SPEAKER_02: Uh, I um uh we we went and did our C round.
11:04
SPEAKER_02: And at that point, we had decided, like, hey, we we're super excited about real-time models.
11:09
SPEAKER_02: We're super excited about interactive models.
11:12
SPEAKER_02: We think that AI that's always on, always running, that's able to communicate synchronously with you is going to be one of the future interfaces for how you uh consume intelligence.
11:25
SPEAKER_02: And so moving from like a turn-based world to like a world where there's continuous interaction happening.
11:32
SPEAKER_02: Like her.
11:32
SPEAKER_02: Like perhaps like her is something that we thought was going to be one of the big transitions, right?
11:39
SPEAKER_02: And and so we wanted to focus on that.
11:42
SPEAKER_02: And we thought that that's gonna be really interesting because there's a lot of unsolved problems there.
11:46
SPEAKER_02: Memory is an unsolved problem, continued learning is an unsolved problem.
11:48
SPEAKER_02: How do you build these systems to be efficient?
11:50
SPEAKER_02: How do you make them feel natural and easy to use?
11:53
SPEAKER_02: These are all big problems in this space.
11:55
SPEAKER_02: So uh so we sort of decided to focus on that.
11:58
SPEAKER_02: And within that, you know, we have a pretty ambitious mission, which is how do we build these interactive models?
12:04
SPEAKER_02: Because they're multimodal models, right?
12:06
SPEAKER_02: You can have audio, you can have text, you can have video.
12:09
SPEAKER_02: Uh, so we needed to focus on some set of problems.
12:11
SPEAKER_02: And we decided let's use voice as an example of this type of system and build that out.
12:18
SPEAKER_02: But build a general uh framework and build the technology in a general purpose way.
12:24
SPEAKER_02: That's kind of how we decided to go after it.
12:27
SPEAKER_00: So choose a voice over text video, like you could have built a video model or only a text-based model.
12:34
SPEAKER_02: Yeah, I think the focus on interaction made it pretty natural.
12:38
SPEAKER_02: So once you think of it as like, what do I need to build to build the first version of a truly high quality interactive system?
12:46
SPEAKER_02: I think the things that you need to focus on are A, how are you communicating with the user?
12:51
SPEAKER_01: Yeah.
12:51
SPEAKER_02: And obviously, for communicating synchronously, voice is the main interface.
12:56
SPEAKER_02: So that was something that was from first principles pretty clear that we needed to do.
13:01
SPEAKER_02: Uh, the second thing is how do you actually uh integrate things like talking, listening, reasoning?
13:07
SPEAKER_02: Again, it was clear that the future is gonna be single single models that are multimodal, can process multiple streams of data, can basically do uh reasoning over that and operate at pretty low latencies.
13:20
SPEAKER_02: So these were kind of obvious from the premise of what we wanted to build.
13:26
SPEAKER_02: And so that's why we said, okay, like voice is gonna be super important as one of the core interfaces.
13:32
SPEAKER_02: So we have to start by investing in that and understanding it.
13:36
SPEAKER_02: And bear in mind, we're not speech researchers.
13:38
SPEAKER_02: Yeah, we're deep learning researchers.
13:40
SPEAKER_02: We don't, I had not beyond, I'd done one project in speech in my PhD, and that was for three months.
13:47
SPEAKER_02: So we didn't, we weren't experts in speech, but we knew a lot about sequence modeling.
13:52
SPEAKER_02: We'd spent a lot of time on in the deep learning side.
13:54
SPEAKER_02: So so we we wanted to sort of bring that in, but we had a lot to learn.
13:59
SPEAKER_00: Let's say uh if you have to uh walk us through in 2035, what would Kartisha look at look like?
14:09
SPEAKER_02: 2035?
14:10
SPEAKER_02: Yeah.
14:10
SPEAKER_02: Um, I think we want to have the best models for interaction.
14:15
SPEAKER_02: Yeah.
14:16
SPEAKER_02: Uh, and we believe that to build them, you need to build them on new architectures.
14:21
SPEAKER_02: That's that's it.
14:22
SPEAKER_02: And so I we would like to be the company that helps the world build more interactive experiences across every surface you can imagine in in the world, right?
14:34
SPEAKER_02: Whether it's in your business, whether it's in your you know personal life, whether it's in you know a train station, you know, airport, uh anything.
14:42
SPEAKER_02: So we just want to be a company that's able to provide developers and uh businesses and builders the ability to build these amazing experiences.
14:52
SPEAKER_02: That's what we want to do.
14:53
SPEAKER_02: And obviously, to make experiences amazing, you need powerful models that are smart, low latency, and they can um do like amazing things when you talk to them and interact with them, right?
15:05
SPEAKER_02: Uh that's that's what we want to do.
15:08
SPEAKER_00: And and and being a model company, can you walk us through like uh because now models have become really popular, walk your audience through the infrastructure required to build a model company?
15:19
SPEAKER_02: Yeah, so um, you know, I think a lot of research companies um need like many different moving pieces to um actually build the product, which is the model ultimately.
15:36
SPEAKER_02: So for us, like it it goes from everything from like infrastructure that's required to build massive data sets to pre-training, to post-training, to um, you know, the architectures that we design and develop, uh, to how they're implemented on you know hardware, to how the training infrastructure actually works and how efficient that is, to you know, how you actually then take these models that we train and inference them.
16:05
SPEAKER_01: Yeah.
16:06
SPEAKER_02: Uh, and then how do you optimize that inference?
16:09
SPEAKER_02: To how do you take that model that can be inferenced on one request and make it possible for it to be used on you know tens of millions of requests?
16:19
SPEAKER_02: Uh, to um, you know, how do you actually then take that and turn it into a great API?
16:25
SPEAKER_02: To how do you then make it easy for people to understand how to use and control the model for what they want?
16:31
SPEAKER_02: So these are all of the pieces that you need to really build a great, great model and to put it in the hands of a user.
16:38
SPEAKER_02: Um I think model training and building models is a bit of an art.
16:45
SPEAKER_02: And the reason it's a bit of an art is in a lot of these new areas, like multimodal models, like there's no known recipe, right?
16:55
SPEAKER_02: Like you can't just go go to the internet and say, like, hey, this is how we're gonna build the model.
17:00
SPEAKER_02: Uh, here's, you know, here's the here's the recipe we can follow, here's a paper you can follow.
17:05
SPEAKER_02: There's a lot of like new science and art to this.
17:09
SPEAKER_02: And so a lot of the work is in how do you build the the muscle for experimentation and how do you have an opinion on how you want to build these systems.
17:19
SPEAKER_02: That is the unspoken part of this, that's like really where the people aspect comes in.
17:25
SPEAKER_02: Um, and so that these are the the big pieces I would say that that we focused a lot on.
17:31
SPEAKER_02: And you know, the the the most undervalued part of this is evaluating your models.
17:36
SPEAKER_02: Because like I think, especially in things like audio, yeah.
17:41
SPEAKER_02: There when we started, the benchmarks were pretty non-existent.
17:47
SPEAKER_02: Even today, I would say the benchmarks are not great.
17:49
SPEAKER_02: And so building a holistic kind of understanding of what does it mean to have a great voice conversation?
17:55
SPEAKER_02: You know, how do people actually perceive voice as a modality?
17:59
SPEAKER_02: That's actually a pretty hard question.
18:01
SPEAKER_02: So uh we need to do a lot of work on that too.
18:04
SPEAKER_02: Uh and so we build a muscle in in a lot of these different areas, yeah.
18:10
SPEAKER_00: And if if we have to explain to our audience, like what are the different parts that a that a model is made up of?
18:23
SPEAKER_02: So I think um and so in in one way a model only has one part, which is uh you know the train the weights, yeah, right?
18:32
SPEAKER_02: The train weights.
18:34
SPEAKER_02: Um I think the the the way you get to a great model is you have to be obsessive about every single part of the stack.
18:41
SPEAKER_01: Yeah.
18:42
SPEAKER_02: Um you have to have really great data, you have to have uh the right architecture, you have to pre-train the model the right amount on a lot of compute, you have to post-train the models in the right ways, and ultimately, if you don't get the experience right through the inference engine.
19:00
SPEAKER_02: So, for example, a Cartesia, one of the things I think we do super well is we co-design everything.
19:05
SPEAKER_02: What do you mean by co-design?
19:06
SPEAKER_02: So, what I mean by co-design is we actually go really deep into understanding how does the architecture impact the data you need?
19:14
SPEAKER_02: How does the architecture impact the way the inference engine will be designed?
19:18
SPEAKER_02: And how does it impact the way you might want to do things like post-training, et cetera?
19:22
SPEAKER_01: Okay.
19:22
SPEAKER_02: And then how does the post-training interact with the pre-training?
19:25
SPEAKER_02: How does the inference interact with um, you know, the way the data has actually been uh structured?
19:32
SPEAKER_02: Like there's all these interdependencies.
19:34
SPEAKER_02: We do a lot of work to understand how to co-design these.
19:38
SPEAKER_02: And when you co-design this stuff, you can actually go after the kind of ambition that we have, which is we always want to have Pareto dominance.
19:46
SPEAKER_02: What I mean by Pareto dominance is we don't like this idea that which is often a popular sort of way of thinking about things, which is that either you have a fast model or you have a good model, yeah.
20:01
SPEAKER_02: Or you have a model that can do, you know, is controllable, or you have a model that's accurate.
20:05
SPEAKER_02: Like we don't like to think of it from that lens of trade.
20:09
SPEAKER_02: Like we want to minimize the trade-offs that our users make.
20:12
SPEAKER_02: And so it's a little bit of an obsession for us to go and say, like, how can we actually build models that are able to do all of these things simultaneously?
20:22
SPEAKER_02: That's really hard.
20:24
SPEAKER_02: The only way you can build models like that is you have to go basically co-design everything.
20:28
SPEAKER_02: Yeah, otherwise, you're you're gonna sacrifice something for something else, right?
20:32
SPEAKER_02: So we've managed to build models, and you know, we we have been launching models that are new now.
20:40
SPEAKER_02: And uh these models are super accurate, they're super fast.
20:44
SPEAKER_02: Every time we release a new model, not only do we make it higher quality, we actually make it faster and cheaper to run.
20:49
SPEAKER_02: So that's the kind of progress we want to make.
20:52
SPEAKER_02: And I think uh that can only happen when you think about this whole system and co-design everything.
20:57
SPEAKER_00: And can you tell us since 2023, you raised your pre-seed, 2024, you raised your seed, uh, the progress of the companies since then or the various milestones that the company achieved?
21:08
SPEAKER_02: Yeah, so we've been uh you know, we've been operating pretty much like a model lab.
21:11
SPEAKER_02: So we've been uh to start with pretty focused on our voice models.
21:17
SPEAKER_02: Um voice has, I think, two major problems.
21:23
SPEAKER_02: How do you understand what somebody is saying?
21:26
SPEAKER_02: And then how do you respond to them, right?
21:29
SPEAKER_02: And then the middle is this big part, which is let's think about what the user is saying and help them accomplish their goals, right?
21:35
SPEAKER_02: So that's those are the pieces when it comes to an interactive model.
21:39
SPEAKER_02: Um, we've been really focused on being super, super good at parts one and three.
21:45
SPEAKER_02: So, how do you listen and how do you respond?
21:47
SPEAKER_02: Yeah, and a lot of our progress in the last couple of years has been A, building this like foundation for interactive models.
21:56
SPEAKER_02: So, for example, uh, there's a lot of like hidden Work that goes into this.
22:01
SPEAKER_02: So, as an example, we've built our own inference engine.
22:04
SPEAKER_02: It's pretty important because we think that these interactive real-time models are going to need to run in a pretty different way.
22:12
SPEAKER_02: So, having the ability to have your own engine that is not designed for LLMs, it's designed for these new class of models is very important.
22:21
SPEAKER_02: So, a lot of foundational work, firstly, we've kind of laid the groundwork on that.
22:26
SPEAKER_02: Secondly, obviously releasing models that are super high quality, very fast for text-to-speech, speech to text, uh being able to build uh uh really, really high quality, pre-to-dominant models for that.
22:42
SPEAKER_02: So that's been one of the set of things we've been releasing.
22:44
SPEAKER_02: And then thirdly, it's um actually putting them in the hands of users and scaling our uh usage on that.
22:52
SPEAKER_02: So obviously, one of our goals is that you know, folks find our models useful, they adopt them, and um and then they build around them, right?
23:01
SPEAKER_02: Like we want to be able to support uh developers and businesses that are building applications in real-time voice.
23:08
SPEAKER_02: And these are all kinds of applications, right?
23:10
SPEAKER_02: Like, as you can, as a lot of uh folks might know, like voice agents as a category is a massive opportunity because um you can now basically build an agent to talk to your users and do work on their behalf uh in pretty much every part of a business.
23:29
SPEAKER_02: So that's been a category that's been taking off, and it relies on having access to models like the ones that we build.
23:35
SPEAKER_02: And so we build some of the best models for that, and that's how we've been uh we've been growing.
23:41
SPEAKER_02: Uh last couple of years.
23:43
SPEAKER_02: And so now we're sort of you know entering this transition of like um uh we built these you know specific models and we want to start bringing them together into uh you know end-to-end models that can start to do more work on behalf of users.
23:59
SPEAKER_02: So ultimately our ambition is again, how do we bring together listening, thinking, and uh response into a single model and just do that continuously?
24:10
SPEAKER_02: Uh and that's that's the next phase.
24:12
SPEAKER_00: The biggest challenge in in voice has been like it's uh I think it's possible to do a great one to two minute conversation.
24:20
SPEAKER_00: Uh but how do you do a 30-minute conversation or an hour-long conversation?
24:26
SPEAKER_00: Because models usually, you know, then don't have a memory uh for for an hour long of conversation and then going back, retrieving context, thinking through.
24:38
SPEAKER_02: Yeah, totally.
24:39
SPEAKER_02: I think um um take this conversation, right?
24:44
SPEAKER_02: This is uh, you know.
24:45
SPEAKER_02: You both are thinking simultaneously.
24:47
SPEAKER_02: Yeah, we're we're doing all of the things uh that I'm describing.
24:50
SPEAKER_02: Like we're talking, I'm thinking, you're thinking.
24:52
SPEAKER_02: Yeah.
24:53
SPEAKER_02: Uh and uh and we have all of this context of this conversation so far.
24:58
SPEAKER_02: And um and so that's something that happens at low latency and supernaturally, actually.
25:04
SPEAKER_02: Um not only this conversation, we're also bringing all of our other context in, right?
25:09
SPEAKER_02: So I think that's something that these systems are not capable of today.
25:14
SPEAKER_02: Now, um, I think part of the reason they're not capable of this is today, when you build a voice agent, it is actually not running a 30-minute conversation, truly.
25:26
SPEAKER_02: Yeah, it is actually running a series of 10-second conversations because it is basically splitting this 30 minutes into small turns, and then every turn is treated as one thing to do, and uh it's engineered, right?
25:42
SPEAKER_02: Every turn is engineered with the context of the previous turn, it's managed in the system.
25:47
SPEAKER_02: So the model is truly not doing that amount of work yet.
25:51
SPEAKER_02: I think the challenge is how do you build really intelligent models that are capable of remembering the context in audio and in text, uh, and then actually can continuously operate at low latency.
26:06
SPEAKER_02: That is actually fairly technically challenging.
26:10
SPEAKER_02: So I think to us, it feels like that's an architectural issue as well.
26:16
SPEAKER_02: Like you can actually improve these models a lot.
26:18
SPEAKER_02: If you can figure out how to do multimodal training better, like the representation learning problems, yeah, there's a lot of issues with how these models are built.
26:27
SPEAKER_02: Like they tend to degrade in terms of their intelligence when you try to build models that can do multimodal reasoning.
26:33
SPEAKER_02: Like text-only models tend to be pretty smart, but when you do multimodal, they tend to get a little bit less smart.
26:39
SPEAKER_02: And so a 30-minute conversation, I think it's a great example because like ideally what we want is for it to run like one conversation.
26:47
SPEAKER_01: Yeah.
26:48
SPEAKER_02: And continuously is listening, talking, and learning from that.
26:53
SPEAKER_02: And not only that, when you stop that conversation, you can then resume something and continue it, right?
27:00
SPEAKER_02: And you don't have to make any sacrifices on uh on context management, personalization, understanding, any of that.
27:07
SPEAKER_02: That is, I think, the grand challenge for this this area uh of of AI.
27:13
SPEAKER_00: Yeah.
27:14
SPEAKER_00: So so if if we think of human businesses, right?
27:18
SPEAKER_00: Let's think in terms of relationships.
27:20
SPEAKER_00: Like I am Yeah, you you have an account with JP Morgan Bank and I'm your account manager, uh, for example.
27:28
SPEAKER_00: Now, why do you prefer to always talk to Fiddles?
27:31
SPEAKER_00: Because he knows the context.
27:33
SPEAKER_00: Uh, you know, let's say when did you last transact?
27:36
SPEAKER_00: Was your preferred home loan rate, etc.
27:39
SPEAKER_00: etc.
27:40
SPEAKER_00: Uh right?
27:41
SPEAKER_00: And and probably he gives you the kind of investment advice that is Taylor made for you because he knows uh you know what current wants to do.
27:50
SPEAKER_00: Uh imagine uh that being done by a modern layer where uh that kind of a memory shield has if that's possible scientifically, then then come the entire account management that companies do can be done by agents.
28:06
SPEAKER_00: I'm not talking about calls, I'm I'm talking about real human experiences.
28:12
SPEAKER_00: And probably it'll it'll only happen when when the agent is able to do an account management, which is 10x better than a human, so that shit is on leave and you're not waiting for shit to come back from his holiday.
28:24
SPEAKER_02: Yeah, no, I think um the ceiling on what should be possible for a lot of these experiences, and some of them are experiences that exist, yeah.
28:35
SPEAKER_02: But you know, some of them are net new.
28:37
SPEAKER_02: So, for example, you know, you may have this kind of capability for ultra high net worse individuals, but actually as a as a bank, you want to provide it to every company.
28:47
SPEAKER_02: Everyone, right?
28:48
SPEAKER_02: Because everybody should have a great experience when they uh transact with you.
28:53
SPEAKER_02: I actually think that's the kind of thing that's super exciting, which is that it opens up the opportunity to provide an experience that was otherwise infeasible.
29:01
SPEAKER_02: Increases the time, right?
29:02
SPEAKER_02: Because the cost of that experience was not possible.
29:04
SPEAKER_02: And so actually it it improves you know your enjoyment of the product.
29:10
SPEAKER_02: Yeah.
29:10
SPEAKER_02: And um and it actually allows businesses to open up new ways to monetize as well.
29:16
SPEAKER_02: Right.
29:16
SPEAKER_02: So you know, one of the interesting applications we've seen, for example, is if you're on a call and you're on hold and you call a business, the when you're on the phone with them, you're already thinking about them.
29:31
SPEAKER_01: Yeah.
29:32
SPEAKER_02: Right?
29:32
SPEAKER_02: You're you're in you're in that zone where you're not gonna be thinking about them day to day.
29:38
SPEAKER_02: But when you're on that phone call, you're thinking about them.
29:41
SPEAKER_02: And that's an opportunity for businesses to actually speak to that customer and either like help them understand how they can have a better experience with their products while they're waiting for maybe a human to pick up the phone or an AI agent to pick up the phone, or even upsell them, right?
29:57
SPEAKER_00: Yeah, you don't need to send the feedback form later on.
30:00
SPEAKER_02: Right.
30:00
SPEAKER_02: So, for example, if I call in and I'm a, you know, I call into a telco and I'm a customer, you have all my information.
30:07
SPEAKER_02: You know that, you know, um I'm maybe like running low on my balance.
30:12
SPEAKER_02: Yeah.
30:13
SPEAKER_02: And so you can actually, I might be calling for a different reason, but you can actually be proactive.
30:18
SPEAKER_02: So the idea of being proactive with customers or being proactive with your employees or being proactive in these interactions is pretty interesting.
30:26
SPEAKER_02: And that's where you see the opportunity to push the experience.
30:30
SPEAKER_02: And so I think part of it is the voice part, but part of it is also just like to your point, like, how do you bring all this context in so that you can actually have a great conversation with someone?
30:42
SPEAKER_00: Yeah, not just solve problems, but make it really feel like it's personalized.
30:47
SPEAKER_02: Exactly.
30:48
SPEAKER_00: Uh for them.
30:50
SPEAKER_02: And and you know, the reality is that um when you talk to somebody on the phone, there are voices that you think sound more personable.
31:00
SPEAKER_01: Yeah.
31:01
SPEAKER_02: There are uh interactions that feel more natural to you, even between human to human.
31:06
SPEAKER_02: So capturing a lot of that is pretty important when you build these systems.
31:10
SPEAKER_02: Like, how do you actually push the experience so that when somebody calls in, it's the experience they would like to have.
31:19
SPEAKER_00: And and today, if if you can share like on on the application side, on the model side, what do you have been able to accomplish?
31:26
SPEAKER_00: Like, what can a developer or a business do with Cartesia model?
31:31
SPEAKER_02: Yeah, so um the way that businesses use us today and developers work with Cartesia is you know, A, you can just go to our website, uh uh play.cartesia.ai, you can sign up for an account, and you know, in a few minutes you can start building an agent using our models that you can talk to.
31:56
SPEAKER_01: Yeah.
31:57
SPEAKER_02: Uh this is obviously one of the biggest use cases right now.
32:00
SPEAKER_02: So um, so we provide you the low latency models you need with the naturalness, the accuracy, the contextual awareness to be able to speak in the right way in these conversations.
32:13
SPEAKER_02: And then the nice thing is we have developers that can sign up and then they can scale on the platform all the way to like very large enterprise in government.
32:22
SPEAKER_02: So you sort of have this unified experience where somebody can uh come in and you know, for our business customers, they uh can basically um deploy our models, whether in our cloud or um in their um, you know, uh, environments, et cetera.
32:40
SPEAKER_02: And so we we kind of offer these different ways of using these models.
32:43
SPEAKER_02: And ultimately, what we're really good at is helping them build the best conversational experiences.
32:49
SPEAKER_02: Uh, we've thought about every part of that stack super carefully.
32:54
SPEAKER_02: Uh, you know, everything from like um how do you actually get to scale, right?
32:59
SPEAKER_02: Like, how do you run millions of calls and make sure that every call is good?
33:04
SPEAKER_02: Because a lot of times people think about demos, yeah, and we can build great demos, but actually, what we care a lot about is we think the technology is so important and impactful that we're actually much more focused on the production problems because we know it's gonna be huge and there's gonna be massive scale.
33:22
SPEAKER_02: We wanna figure out how do you drive the best experience at that mega scale.
33:27
SPEAKER_02: So that's one of the things we're super good at, and that's where all of that core design comes in, right?
33:32
SPEAKER_02: There are invisible systems at play that you don't even see, but they're the reason that you get amazing latency, all of that experience.
33:41
SPEAKER_02: So that's one of the things we do super well.
33:43
SPEAKER_02: And then, of course, like there are things we've obsessed about, like, even like we have a team that's just focused on voice research.
33:50
SPEAKER_02: All they do is try to understand what is the impact of a voice on a conversation and the outcomes that you're trying to drive.
33:58
SPEAKER_02: How do metrics shift because you changed how a voice sounds?
34:03
SPEAKER_02: It has enormous impact in use cases, and different use cases need different experiences, right?
34:09
SPEAKER_02: If you're doing a nursing use case versus you're doing a banking use case versus you're actually doing a use case as a personal assistant, they actually all require different things.
34:18
SPEAKER_02: So, how do you nail that experience?
34:20
SPEAKER_02: That is actually a big part of what we provide.
34:23
SPEAKER_02: Uh, and when we work closely with businesses, we actually help them figure out how to design the experience.
34:28
SPEAKER_02: How do you evaluate the systems?
34:30
SPEAKER_02: This is a big problem, right?
34:32
SPEAKER_02: Again, audio voice is a pretty tricky thing to understand and evaluate.
34:36
SPEAKER_02: So we actually provide a lot of systems, guidance, and uh obviously product to be able to help with that.
34:42
SPEAKER_02: So I think as I said, we obsess about the conversation experience and we help developers and businesses really nail that.
34:48
SPEAKER_00: I think the challenge with voice is uh people need a response immediately, like with extremely low latency, exactly because they're used to the human interactions.
35:00
SPEAKER_00: Whereas in text or chart-based experiences, which are fairly new, I don't think they are more than a decade old.
35:05
SPEAKER_00: People are okay with 30 seconds to even a minute latency for things like deep research.
35:11
SPEAKER_02: Yep.
35:11
SPEAKER_02: No, exactly.
35:12
SPEAKER_02: In fact, you know, there's the deep resource side, which is you know, you're using your favorite LLM and and you expect it to take time and do deep thinking and so on.
35:22
SPEAKER_02: And so there, I think what we're interested in is like how do you compress the time it takes to do thinking, right?
35:26
SPEAKER_02: So that's one of our research problems that we're interested in.
35:29
SPEAKER_02: But um, to your point, like when you're using these systems, think about even support and chat.
35:36
SPEAKER_02: When you're chatting to AI, if it responds too quickly, yeah, it actually feels like it doesn't hasn't done any thinking at all.
35:44
SPEAKER_02: So it it's actually not a it's a it's a feature, not a bug that it can take a few seconds to respond to you.
35:51
SPEAKER_02: Um, I think in voice to your point, it's totally different, right?
35:55
SPEAKER_02: And like actually the problem is more that how can you be as smart as possible in the response in the least amount of time.
36:02
SPEAKER_02: So um so the latency is just core to this.
36:06
SPEAKER_02: Yeah, and um, and the real time aspect is like very, very important to get, right?
36:12
SPEAKER_02: So I think that a lot of what we see when people are building these systems is that um they're struggling to understand how to bring latency down while keeping the overall intelligence of the system as high as possible.
36:26
SPEAKER_02: That is the biggest challenge a lot of these folks face.
36:29
SPEAKER_02: That's part one.
36:30
SPEAKER_02: And then part two is how do I make sure I design the experience so I can drive my outcomes, right?
36:36
SPEAKER_02: And again, depending on the different use cases they might be in, the outcome that they care about is different, right?
36:42
SPEAKER_02: Whether it's CSAT, containment, or it could be uh even like time that somebody spends on the conversation if it's a consumer use case.
36:50
SPEAKER_02: So like all of these things are like depend on the use case.
36:54
SPEAKER_02: So having a broad understanding of that is hard, but that's where we help a lot.
37:00
SPEAKER_00: And and today, if if any industry points to being one of the highest consumers of Kartisha, what would that industry be?
37:08
SPEAKER_02: So we work with a lot of folks across.
37:12
SPEAKER_02: Um I would say healthcare, fin surf, regulated industries is where we see a lot of uh use cases and uh volume.
37:25
SPEAKER_02: I think um, you know, we're uh we're an interesting company in the sense that because we're an enabler, yeah, uh you cannot build voice agents without the voice models.
37:38
SPEAKER_02: Yeah.
37:39
SPEAKER_02: So so we kind of plug into everybody in terms of the stack.
37:43
SPEAKER_02: Um and so, but with these are the areas where we see a lot of the use cases, I think, that are really driving uh a lot of volume and traffic.
37:53
SPEAKER_02: Um, I think that every week we see news use cases pop up.
37:59
SPEAKER_02: So I think that's what keeps it interesting because there's a huge amount of experimentation happening.
38:03
SPEAKER_02: I think the cool thing is there's the the experimentation and then there's the stuff that's going to production.
38:08
SPEAKER_02: And both are like pretty crazy right now.
38:12
SPEAKER_02: Uh, but the experimentation is quite interesting because, as I said, like people are just thinking, like, oh, I have this problem.
38:19
SPEAKER_02: Can I can I use voice to actually make this more fictionless from an experience perspective?
38:24
SPEAKER_02: So, for example, uh, you could uh think about things like internal tools, right?
38:30
SPEAKER_02: Like, how can I give my employees access to more information?
38:33
SPEAKER_02: Yeah, again, a really nice place for you to use voice as an interaction uh interface so that they can actually access answers as as easily as possible.
38:43
SPEAKER_02: So, like dropping the friction with which people can access information and get things done is kind of something that voice is pretty good at.
38:50
SPEAKER_00: And uh Today the fear in many application companies are that the model companies are are you know eating their lunch and as close to to building their end applications as possible, like open-air anthropic, boost announced service as a software in in the last 10 days.
39:12
SPEAKER_00: So where where are you headed?
39:14
SPEAKER_00: Like because uh if if you start competing with with the applications, then the application lose trust in the models.
39:21
SPEAKER_02: Right, right, right.
39:22
SPEAKER_02: Um I think you know, you know, we have a position that there is a lot to do in the models.
39:30
SPEAKER_01: Yeah.
39:30
SPEAKER_02: Because of our research DNA, we think that the actual progress in models that needs to be made over the next 10 years, eight years is actually enormous.
39:40
SPEAKER_02: So I think our focus is primarily on how to be an enabler.
39:45
SPEAKER_02: I think what we see, especially if you take companies like uh, you know, the large Frontier Labs, is often what happens is when they've progressed the model to a point where it's got amazing capabilities, they're often the first to see that new capability because it's you know it's inside the company.
40:05
SPEAKER_02: And then they understand and realize how it can actually change fundamentally an application and they build it then.
40:12
SPEAKER_02: So, for example, I think Cloud Code is a great example of this, where like I think they realize that actually maybe you don't need an IDE anymore.
40:20
SPEAKER_02: That's a pretty provocative perspective.
40:22
SPEAKER_02: Yeah.
40:23
SPEAKER_02: And so if you're inside that business, you're like, hey, we could build this demo that shows what the future could look like.
40:30
SPEAKER_02: A lot of time what happens is that the developers love that future and then it turns into a great product, right?
40:36
SPEAKER_02: And then that's what then becomes disruptive.
40:38
SPEAKER_02: I don't think it comes from a place of trying to compete.
40:41
SPEAKER_02: I think a lot of times the stories around how these products got built is that people inside were amazed by what it was able to do.
40:47
SPEAKER_02: They built a prototype, and the prototype just happened to be something that everybody thought was really cool, yeah, and then it takes off.
40:54
SPEAKER_02: So I think all model companies, I think at some level want to be good to developers and seed like, you know, as many applications in the world as possible.
41:03
SPEAKER_02: That's certainly our like motivation.
41:06
SPEAKER_02: I think the thing that was hard is the information asymmetry that exists, which is that model developers see where the future is before the application developers do.
41:14
SPEAKER_02: So that asymmetry is impossible to sort of like actually remove.
41:19
SPEAKER_02: Uh, but I think um for us, I think the goal is again, how do you push the core technology forward and ultimately arm developers with the right APIs, the right experiences, the right tools to build these great uh applications.
41:32
SPEAKER_00: So in your case, how did you measure that uh you had product market fit and when at what point in time you crossed PMF?
41:41
SPEAKER_02: Yeah, I think um you know you you start to have uh customers and then you realize that uh you um start to get like repeatability across customers, right?
41:58
SPEAKER_02: I think that's one of the key things.
42:00
SPEAKER_02: Um people come to you and they um already know about your product, they know what you can do for them, and and it becomes a lot easier to um for to to get customers in the first place, right?
42:12
SPEAKER_02: So I think there's that you know sort of like natural cadence to this.
42:16
SPEAKER_02: And I think uh especially for a business that's like model oriented, you know, there's a certain you know simplicity to it, right?
42:25
SPEAKER_02: That I think is quite nice.
42:27
SPEAKER_02: Um so so that's one part of it.
42:29
SPEAKER_02: I think um we also started to see obviously like um our you know, consumption, volume, scale on top of our platform, like how much do people actually rely on us for their production workloads, right?
42:45
SPEAKER_02: Because ultimately, like for us, what matters is the consumption that is going through our systems at scale uh for customers globally.
42:55
SPEAKER_02: Um, and so we're seeing customers that are starting to go towards pretty insane volumes now on our on top of us, like anywhere to like um quarter one billion minutes, right?
43:09
SPEAKER_02: Getting to that scale a year uh for one customer.
43:13
SPEAKER_02: And um that's pretty insane, right?
43:17
SPEAKER_02: Um, the amount of um conversations that are happening through Cartesia and the diversity of use cases has just gone up.
43:26
SPEAKER_02: I think we are doing use cases in pretty much every industry.
43:29
SPEAKER_01: Yeah.
43:30
SPEAKER_02: And now we're also you know supporting 42 languages, so it's global.
43:35
SPEAKER_02: Um, you know, obviously US, um, India, uh, Southeast Asia.
43:40
SPEAKER_02: Yeah.
43:41
SPEAKER_02: Uh we have tons of customers that are you know building on us, and um, and they're doing like really interesting use cases at scale.
43:49
SPEAKER_02: So that's one part of how we think we got to product market fit.
43:55
SPEAKER_02: But I think product market fit is also one of those things where it's like, I think.
44:00
SPEAKER_02: The levels to it, right?
44:02
SPEAKER_02: And I think for us, it's like how do you unlock the next level of scale and uh help even more like 10x the number of developers?
44:10
SPEAKER_02: The way our platform teams actually think about it is they're always thinking about how to unlock the next 10x on scale, volume, et cetera.
44:17
SPEAKER_02: Uh, but we have customers that are also like they have consumer bases of more than 100 million people.
44:24
SPEAKER_02: So that's massive scale that I think um that's what gets us really excited.
44:29
SPEAKER_02: Like, how do you how do you build the technology for so many people and deliver it reliably?
44:34
SPEAKER_02: I think that's something that I think gets uh less attention, but I think it's what the big challenge is actually when you work with customers.
44:42
SPEAKER_02: Because customers, you know, once they've realized the value you can offer to the product, I think what they're really excited to work together closely on is ultimately like how do we build a robust application?
44:55
SPEAKER_02: Because this is about their business, right?
44:58
SPEAKER_02: You don't want something that's gonna go down, they don't want something that's gonna be unreliable, and that's where we excel, I think.
45:04
SPEAKER_02: Um, so it's been pretty fun to do that.
45:07
SPEAKER_00: And can you share, like, in any metrics, what would be the approximate scale of Cartesia?
45:13
SPEAKER_02: Yeah, I mean, I think um, you know, we have um uh a lot of customers doing, you know, tens to hundreds of millions of minutes uh a year of conversation, and you know, our customer base has grown to hundreds of uh customers now, enterprise customers.
45:28
SPEAKER_02: So it's uh it's been a good good journey for us uh over the last uh last two years.
45:34
SPEAKER_02: And what does your GTM look like?
45:37
SPEAKER_02: In terms of how we organize ourselves.
45:39
SPEAKER_02: Yes.
45:39
SPEAKER_02: Yeah, so it's it's been we're actually super lean as a as a company.
45:44
SPEAKER_02: Um on the go-to-market side, we've actually um uh got only a small number of folks that are actually um on the sales side and uh a pretty lean team overall.
45:57
SPEAKER_02: Uh our marketing team is super small.
46:01
SPEAKER_02: Yeah.
46:01
SPEAKER_02: Uh we've in fact, we're not really a company that has done a lot of marketing actually.
46:06
SPEAKER_02: And part of the reason is the philosophy we've had is like a great product should not require that much marketing.
46:13
SPEAKER_02: Yeah, so you can spend a lot of time and money on marketing to hide the fact that your product is not great.
46:19
SPEAKER_02: It's better to focus on making the product great and then to focus on marketing after.
46:24
SPEAKER_02: That's sort of our philosophy.
46:26
SPEAKER_02: So we've been really trying to follow that religiously.
46:29
SPEAKER_02: Um, but um, you know, our approach generally is we care a lot about developers and we want to make the experience frictionless for them to adopt the platform.
46:37
SPEAKER_02: So they should be able to come in and get started in five, 10 minutes.
46:42
SPEAKER_02: Um, the way we reach these developers today is honestly just word of mouth.
46:46
SPEAKER_02: We go talk to folks, people talk about us, they like the experience, they go tell their friends, and that's actually been a pretty uh important way that we've grown.
46:55
SPEAKER_02: And we want the model quality to speak for itself, literally.
46:59
SPEAKER_02: So I think that's something that's uh been a big part of it.
47:02
SPEAKER_02: And now, as we've seen like larger enterprises adopt us, I think we've kind of built that um more sales muscle uh that allows us to kind of go uh work with these more complex organizations where they need us uh to help them understand the way in which they should be thinking about the technology.
47:21
SPEAKER_02: What can you do with voice even?
47:23
SPEAKER_02: What are some of the use cases you can unlock?
47:25
SPEAKER_02: And then partnering together with them to uh deploy, you know, whether it's some engineering resources on the forward-deployed side as well as our whole platform to unlock that.
47:36
SPEAKER_02: But uh we we we sort of had this sort of mixed PLG enterprise motion.
47:41
SPEAKER_02: And I think what we've seen is a lot of actually enterprise buyers, even today, are pretty sophisticated.
47:46
SPEAKER_02: Like it's not like they know how to go use your self-served platform and figure out whether things are working the way they're supposed to, especially now where you have coding agents, you have all this stuff, it's really easy for somebody to build a simple demo and test the product out.
48:04
SPEAKER_02: So that's something that we try to enable pretty uh easily for folks, um, and that's been a good driver of like growth for us.
48:12
SPEAKER_00: Uh if you have to uh summarize uh a few things that helped Kartisha reach at the at the scalar today, what would those things be or those uh critical moments be or decisions be?
48:24
SPEAKER_02: Yeah, I mean I think um you know, I think when you're in the process of doing things, you know, building the company, it can feel like a lot of things matter.
48:39
SPEAKER_02: I think in hindsight, very few things matter.
48:43
SPEAKER_02: So that tells you something about there being an imbalance between like the way you tend to view things as important or not versus what is actually important.
48:54
SPEAKER_02: Um, I think for us, like, I think it boils down to having great product and great models.
49:00
SPEAKER_02: Because why do we exist as a company?
49:02
SPEAKER_02: We exist to build amazing models for all of these developers and builders, right?
49:08
SPEAKER_02: I think if you just focus in on that, you realize what is the most important thing for the growth of the company, it is to have great models.
49:16
SPEAKER_02: A lot of the other stuff around that, um, you know, we we we did a lot of things, but I think at the end of the day, I feel like a lot of first time founders feel this way, which is uh if I was to do it all over again, I could do it twice as fast, right?
49:31
SPEAKER_02: Because you kind of learn what are the things that you maybe didn't need to do.
49:36
SPEAKER_02: Um, so I would say that's been that is the biggest thing that we focused on that has actually mattered a lot for us in terms of our growth.
49:44
SPEAKER_02: Because people want it, especially in models, people want to use great models that work for them.
49:51
SPEAKER_02: And uh, you know, I think we've also done this work of actually building a company that has the ability to do great research and actually is building a commercially valuable product.
50:04
SPEAKER_02: Yeah, this is something that a lot of people think is mutually exclusive, right?
50:09
SPEAKER_02: People think either you can be a research lab or you're a product company.
50:13
SPEAKER_01: Yeah.
50:14
SPEAKER_02: We I think what we did was we said, no, no, no, that's not, we don't think those two things are not possible to do together.
50:21
SPEAKER_02: Yeah.
50:21
SPEAKER_02: And we want to prove that it's possible through doing Cartesia and building the company.
50:26
SPEAKER_02: And I think that's the other piece that I think we've done well, which is trying to square these two things.
50:32
SPEAKER_02: And you need a lot of strong opinions to do that.
50:35
SPEAKER_02: Because both of these things have can take on a life of their own.
50:39
SPEAKER_02: The discipline is to co-build it.
50:41
SPEAKER_02: So I think that's the other thing that has contributed a lot to our growth.
50:44
SPEAKER_02: So those are the two big things I would say.
50:46
SPEAKER_02: One is focus on the core product and the models, and the second is a focus on how the organization runs when it's trying to do research and product at once and having an obsession around that.
50:57
SPEAKER_00: When you are four founders, how do you bring alignment between all of you to focus on a single or maybe two directions that match at the same time?
51:05
SPEAKER_02: I think we're uh as founders, we A are like super honest.
51:12
SPEAKER_02: Um I think we try to communicate in a way that's super transparent all the time, actually.
51:20
SPEAKER_02: I don't know what it's like because we never had other people in the room when we're doing these things, but I think we debate pretty vigorously.
51:27
SPEAKER_02: But also we have like a pretty high trust amongst the founders, like super high.
51:32
SPEAKER_02: So we're able to actually have, I think, pretty fast alignment on most things.
51:38
SPEAKER_02: And where we debate, we debate vigorously.
51:40
SPEAKER_02: I think it's just about being uh actually pretty close to all of my co-founders and actually like having pretty high trust there in the relationship where you can like say things to them and they can say things to you, and like you don't take it personally, and everybody's there to win.
51:57
SPEAKER_02: And um, that's pretty important, I think.
52:00
SPEAKER_02: And a lot of companies, I think where they struggle is that the founders don't, you know, sort of like have full alignment.
52:08
SPEAKER_02: I think we're lucky uh that we have uh a founding team that has like uh the humility to align, right?
52:17
SPEAKER_02: They can set aside their egos uh to to get on the same page to ultimately move the company forward.
52:26
SPEAKER_00: And in your opinion, if you have to think about how many companies can become big globally in the YCI space, what what would the number be or how would you try to project it?
52:39
SPEAKER_00: Yeah, it's an interesting um interesting question.
52:42
SPEAKER_02: Um I think one I mean there's many ways to try to understand this.
52:50
SPEAKER_02: I think um I think one way I think about it is like what is the actual amount of time the average consumer will spend with these models that are interactive per day.
53:04
unknown: Yeah.
53:05
SPEAKER_02: Right now the number is super small.
53:07
SPEAKER_00: Maybe one hour, two hours per day?
53:09
SPEAKER_02: Today it's nothing, basically, right?
53:11
SPEAKER_02: But you you want to get to a point where maybe it's one, maybe two, maybe for you know some users actually eight hours a day.
53:17
SPEAKER_02: We we actually don't know how people will work in the future because like you can imagine that the imagine a workplace of the future where actually the way you work is you constantly have this collaborator and you actually redesign the workplace to help people be able to work alongside the AI collaborator.
53:38
SPEAKER_02: So you can't have open office spaces anymore because that's gonna create a lot of disruption.
53:43
SPEAKER_02: And you can imagine that that might involve doing a lot more interaction, right?
53:50
SPEAKER_02: Um, so so if you think of it from that lens, like you you sort of get to pretty insane numbers in terms of how much time is actually spent by people uh when interacting with these applications and models.
54:05
SPEAKER_02: And um and I think from that you can back out.
54:08
SPEAKER_02: I hope somebody does this analysis in your comments.
54:11
SPEAKER_02: You can try to back out how many uh companies could exist.
54:16
SPEAKER_02: I think there's two variables that matter for those companies.
54:19
SPEAKER_02: One is how many minutes of interaction are you driving?
54:22
SPEAKER_02: The other is what is the value of every minute?
54:25
SPEAKER_02: Because in different applications, every minute has different value, right?
54:29
SPEAKER_02: So uh, you know, um for banks in a collections call, a minute might have pretty high value actually in terms of how much uh you know value it drives for that business, versus um how much value does it drive to do character chat, let's say, uh for a minute, right?
54:48
SPEAKER_02: Like so there'll be different ways to monetize and ways to understand that.
54:52
SPEAKER_02: And I think hopefully, I mean, I don't think I've ever seen this type of analysis done before.
54:57
SPEAKER_02: So, you know, again, uh hopefully somebody does this.
55:00
SPEAKER_02: But I think you can back out like across different use cases, industries, uh consumer, you know, enterprise, what are the um the amount of minutes you can expect to have in a specific area and how much value does that drive?
55:17
SPEAKER_02: But I think you can back it out hopefully by making some assumptions about how many minutes.
55:21
SPEAKER_02: I've done some analysis on this.
55:22
SPEAKER_02: I unfortunately don't remember off the top of my head and I don't want to do mathematics uh live.
55:27
SPEAKER_02: Uh, but uh the uh even if you if you take a number that's fairly small per person in the world, yeah, it you actually get incredibly large uh large numbers there in terms of like what's possible.
55:42
SPEAKER_02: So I think that's the fascinating thing, and it's almost a challenge for these companies to make the interaction sticky, right?
55:48
SPEAKER_02: The whole point is can you add value so that somebody wants to do this every day?
55:53
SPEAKER_00: And my last question is you know, before we conclude, what helped you reach each round from the best of tier one investors?
56:00
SPEAKER_00: In your first round, you had you know cognition in and second round, you had index in the third round, you had kinda perkins.
56:08
SPEAKER_02: Yeah, so um, you know, we um um I think I think there's a few things.
56:14
SPEAKER_02: Like one is of course, like having a clear mission and purpose for the company.
56:20
SPEAKER_02: And what's the the mission been?
56:22
SPEAKER_02: For us, the mission was always about building interactive AI and closing the gap between AI and humans in terms of how uh smart and how interactive these models are.
56:33
SPEAKER_02: Yeah, so we were always super focused on real-time interaction.
56:37
SPEAKER_02: And that means also, for example, we never got distracted in doing other things with voice because there's a lot of other voice use cases that are not real-time interaction, but we're like hyper-focus on this.
56:48
SPEAKER_02: Uh, the other piece is having uh the right team.
56:52
SPEAKER_02: Yeah, I think, especially for what we do, which is pretty hardcore research, we need amazing folks in the team.
57:00
SPEAKER_02: And I think that's one of the ways that you can see progress in a company like ours, is our ability to attract and retain the best talent.
57:10
SPEAKER_02: Even as a small company where you know folks have all of these amazing places that they could work, they still choose to come to Cartesia and work here.
57:18
SPEAKER_02: So that's one of the things that I think we were able to demonstrate even from an early on stage.
57:25
SPEAKER_02: And I think progress, right?
57:27
SPEAKER_02: Ultimately, like we've been shipping, we've been building great models, we've been uh, you know, putting them in the hands of users.
57:33
SPEAKER_02: So I think all of these things matter a lot, like clarity of vision, uh, the quality of talent that you can attract and and you know work with, and then uh your uh ability to ship something that matters.
57:46
SPEAKER_02: And so I think these are the three areas that we've really been able to, I think, do pretty well at.
57:51
SPEAKER_02: Um, but I think you know, it's just you can't like the past past performance is not an indicator of future.
57:58
SPEAKER_02: So I think we are constantly just trying to be better at a lot of these things, and and you know, we've been again fortunate to work with the folks at Index and Kleiner, um, as well as Nvidia and and you know, a lot of other folks uh that are pretty big supporters of the company.
58:12
SPEAKER_02: And so you want to keep pushing that forward.
58:15
SPEAKER_00: Well, thank you so much, Kiran.
58:16
SPEAKER_00: It's been an amazing conversation.
58:18
SPEAKER_02: I enjoyed it a lot.
58:18
SPEAKER_02: Thanks so much.
58:19
SPEAKER_02: No, this was great to great to chat.
58:21
SPEAKER_02: Thanks so much for all the questions, and uh yeah, I I I want to see the analysis of uh of the the TAM of the voice space in in the comments.
58:29
SPEAKER_00: Yeah, yeah, definitely.
58:31
SPEAKER_00: I think uh a listeners are student when put out their own analysis on top of it.
58:36
SPEAKER_01: All right.
58:37
SPEAKER_01: Thanks so much.