Rohan Joshi on building Stellon Labs, pivoting from consumer to developer market, and AI voice inflection points
Summary
Rohan Joshi, founder of Stellon Labs, discusses his journey from Meta to launching an AI research lab focused on CPU-efficient open-source voice models. He explains how initial pricing experiments revealed that non-technical users were not the target audience for privacy-focused tools, leading to a pivot toward developers. Joshi highlights the virality of open-source distribution and argues that AI voice models have reached a threshold of naturalness suitable for enterprise applications like customer support and collections. He also outlines future inflection points for voice AI, including bidirectional interfaces and real-time video streaming.
Key takeaways
- 1 Closed-source products for developers often create larger markets for open-source alternatives because developers can customize code and host on-premises.
Unidentified speaker (speaker_1) · 16:39
- 2 Yeah. Yeah. So I like that's another observation that all the open source AI, the center of gravity for that is in China.
Unidentified speaker (speaker_1) · 17:18
Episode sections
- 0:00 – 1:14
Background and Family Influence
Rohan Joshi discusses his upbringing in Mumbai with entrepreneurial parents and his realization that jobs are merely checkpoints toward starting a company. He reflects on the mindset of self-deception as a tool for confidence.
- 1:14 – 13:45
Company Overview and Early Career
Joshi introduces Stellon Labs (KittenML) as an AI research lab building CPU-efficient open-source voice models. He details his education at CMU and his three-year tenure at Meta working on photorealistic avatars for the metaverse.
- 13:45 – 19:35
The Pivot from 11Labs Alternative to Developers
Joshi explains their initial YC application idea: a local, privacy-focused alternative to 11Labs. The pivot occurred when pricing the product at $99/month caused all users to churn, revealing that non-technical users valued price over privacy, whereas developers were the viable market.
- 19:35 – 21:47
Validation Through Open Source and Virality
The speakers discuss using open-source distribution as rapid customer research. Joshi notes that their model went viral on Reddit, leading to such high GitHub stars and downloads that they initially suspected their analytics were broken.
- 21:47 – 23:21
AI in Customer Support and Collections
Joshi argues that AI voice models have surpassed human naturalness thresholds, making them superior for customer support. He agrees with the host that AI agents are effective in financial collections due to their lack of emotion.
- 23:21 – 25:30
Voice AI Inflection Points
Speaker_1 outlines three inflection points for voice AI: current enterprise automation replacing customer support and collections, upcoming bidirectional interfaces for full computer interaction, and a future decade-long shift to AI-generated real-time video streaming.
- 25:30 – 27:10
Resourcefulness and Open Source Leverage
Speaker_1 argues that individuals should 'fool themselves' into believing they can do anything, leveraging resourcefulness and compounding skills. Speaker_0 adds that AI and open source tools provide a leg up, potentially removing the need for years of traditional grinding.
- 27:10 – 28:43
Stellon Labs' Technical Approach
Speaker_1 describes Stellon Labs' focus on ultra-low precision text-to-speech models that run efficiently on CPUs rather than requiring expensive GPUs, aiming to democratize voice AI for developers with limited hardware access.
- 28:43 – 30:52
Geographic Strategy for Tech Founders
Speaker_1 advises that founders building frontier technology should locate in the Bay Area to access the ecosystem and network effects, arguing that companies like Google, Meta, and OpenAI have done more to advance global lives than many local Indian tech companies.
Assets discussed
| Asset | Kind | Stance | Context |
|---|---|---|---|
| DeepSeek | private company | neutral | And we've already seen this with in LLMs. Like if you see the most used AI LLM today is actually DeepSeek and other Chinese models. |
| AlphabetGOOGL | equity | bullish | And one thought experiment you can do is that technologies that have been invented here, like Google, Meta, |
| Meta PlatformsMETA | equity | bullish | And one thought experiment you can do is that technologies that have been invented here, like Google, Meta, |
“Stance” is how the speakers talked about the asset, not a recommendation.
Claims worth verifying
Stellon Labs (KittenML) builds open-source voice AI models that are efficient enough to run on a CPU.
Unidentified speaker (speaker_1) ·1:29
How to verify: Check Stellon Labs website or GitHub repository for model specifications and hardware requirements. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
Rohan Joshi grew up in Mumbai; his father manufactures pharmaceutical machinery (fluid bed dryers) and his mother runs a salon chain and cosmetics brand.
Unidentified speaker (speaker_1) ·2:21
How to verify: Cross-reference with LinkedIn profiles or public biographies of Rohan Joshi. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
Joshi worked at Meta for three years (starting in 2021) building foundation models for 3D avatars.
Unidentified speaker (speaker_1) ·7:54
How to verify: Check Meta's press releases or job history archives for 2026-2026 regarding avatar research teams. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
The company applied to Y Combinator in 2025.
Unidentified speaker (speaker_1) ·10:34
How to verify: Check Y Combinator's public batch lists for 2026. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
The initial product was a local alternative to 11Labs with a simple UI for non-technical users.
Unidentified speaker (speaker_0) ·14:09
How to verify: Check archived YC demo day videos or early press releases from 2026. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
Pricing the initial product at $99/month caused all users to churn.
Unidentified speaker (UNKNOWN) ·15:29
How to verify: This is a specific internal metric; verify via company blog posts or interviews discussing the pivot. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
The open-source model went viral on Reddit, leading to rapid growth in GitHub stars and downloads.
Unidentified speaker (UNKNOWN) ·20:17
How to verify: Check Reddit archives for posts about KittenML/Stellon Labs and GitHub repository statistics. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
DeepSeek is the most used AI LLM today.
Unidentified speaker (speaker_0) ·17:17
How to verify: Check usage statistics from platforms like Hugging Face or industry reports on LLM adoption. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
The Transformer paper was published in 2017.
Unidentified speaker (UNKNOWN) ·5:47
How to verify: Verify publication date of 'Attention Is All You Need' on arXiv. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
Mark Zuckerberg stated in a YC interview that he had millions of users for Facebook but still believed another company would build a social media platform that would take over the world.
Unidentified speaker (UNKNOWN) ·12:35
How to verify: Search for Mark Zuckerberg's Y Combinator interview transcripts. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
AI can expand the number of scenarios dealt with compared to hard-coded chatbots.
Unidentified speaker (speaker_1) ·23:05
How to verify: Review historical documentation of rule-based chatbot limitations versus modern LLM capabilities in customer service logs or technical whitepapers from 2026-2026. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
Voice AI is currently in its first inflection point focused on enterprise automation for customer support and outbound calls.
Unidentified speaker (UNKNOWN) ·23:25
How to verify: Check enterprise AI vendor reports (e.g., Gartner, IDC) or company earnings calls from 2026-2026 detailing adoption of voice AI in BPO sectors. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
Whisperflow is an example of a unidirectional voice interface for electronics.
Unidentified speaker (UNKNOWN) ·23:44
How to verify: Search for 'Whisperflow' product documentation or GitHub repositories to confirm its functionality as a unidirectional voice interface. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
Stellon Labs is building ultra-low precision text-to-speech models that run on CPUs.
Unidentified speaker (UNKNOWN) ·27:28
How to verify: Visit the Stellon Labs website or GitHub repository to verify model specifications, hardware requirements, and open-source availability. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
Running traditional voice models typically requires tens of thousands of dollars a month in GPU costs.
Unidentified speaker (UNKNOWN) ·27:28
How to verify: Compare cloud GPU pricing (e.g., AWS, Azure, Lambda Labs) for high-end GPUs (e.g., H100/A100) against the claimed cost structure. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
Google, Meta, and OpenAI have done more to advance the life of an Indian in India than other Indian tech companies.
Unidentified speaker (speaker_1) ·29:15
How to verify: This is a subjective claim; verify by analyzing economic impact studies or digital adoption metrics in India attributed to these specific US tech firms versus Indian firms. For recent events, verify 2026-08-05 through 2026-11-05. Use another period only when the speaker explicitly names it.
Do your own research
- Verify Stellon Labs' GitHub repository activity and download statistics to assess current traction and community engagement.
- Review the technical specifications of Stellon Labs' CPU-efficient models compared to competitors like 11Labs or open-source alternatives such as Coqui or VITS.
- Analyze the pricing strategies of major voice AI providers to understand market segmentation between enterprise and developer tiers.
- Research the regulatory landscape for AI-generated voice cloning, particularly regarding deepfake detection and consent requirements in customer support contexts.
- Evaluate the competitive landscape of open-source voice models, specifically tracking developments from Chinese labs like DeepSeek or Baidu.
- Check the hiring trends and talent availability in the Bay Area versus other tech hubs to assess the validity of geographic strategy claims for AI startups.
Full transcriptexpand
- 0:00
- Unidentified speaker (speaker_0): But so far listening to you, I feel oh shit, even I can't do this because this guy is so technical and knew all of this for so long.
- 0:06
- Unidentified speaker (UNKNOWN): Like
- 0:06
- Unidentified speaker (speaker_1): the easiest person to fool is yourself. So you should fool yourself into believing that you can do anything. My parents would always tell me that you should do a job for experience. But if you really like decode what that means, like experience for what? What they're tr what it means is that it's like an intermediate checkpoint to like another eventual goal. And that goal was never mentioned, but I never asked because I
- 0:27
- Unidentified speaker (speaker_1): knew what it was, which was to start a company.
- 0:29
- Unidentified speaker (speaker_0): AI plus open source, given that it is commoditized, not commoditized, but available everywhere, so on and so forth, should be taken and looked upon as an asset where potentially the four or five years of grind you had to go through, you don't have to go
- 0:44
- Unidentified speaker (UNKNOWN): through.
- 0:45
- Unidentified speaker (speaker_0): Welcome to Unstarted. What is unique about this show is this is about you. It is for founders by founders.
- 0:54
- Unidentified speaker (speaker_0): Hi, I'm super excited to have a new series of Unstarted in the land where it all started in the Silicon Valley with Rohan of Stellarn Labs, whose product is Kitten ML coming to you from San Francisco Bay Area and focusing on our AI companies. So with
- 1:14
- Unidentified speaker (speaker_0): that. Rohan, welcome to Unstarted.
- 1:17
- Unidentified speaker (speaker_1): Thanks for having me.
- 1:18
- Unidentified speaker (speaker_0): Could you tell us a little bit about very quick background? We'll get into more details, but a little bit about yourself. But first, let's talk about the company, what does it do, and then we'll go into your background. Yeah,
- 1:28
- Unidentified speaker (UNKNOWN): so the
- 1:29
- Unidentified speaker (speaker_1): company is uh Stalin Labs, but we're better known as KittenML among developers. Uh we are an AI research lab that builds open source voice AI that's so efficient that it can run on a CPU.
- 1:39
- Unidentified speaker (speaker_0): Sounds very fancy to me, so we will we will get into details. Uh but it didn't start that fancy, uh Rohan. So one of the goals of this series is to let people who are listening to it know how
- 1:53
- Unidentified speaker (speaker_0): they can also start up, they can also start up an AI, they can do all these fancy things. So let's go all the way back. Did you grow up in a rocket scientist family with a with a computer science background? Take us through your background and then how we ended up here.
- 2:08
- Unidentified speaker (speaker_1): Yeah, absolutely. So I grew up in Mumbai. Pretty much spent all my life in Mumbai until including schooling, college, and came to the US after for my master's. What's
- 2:20
- Unidentified speaker (speaker_0): the family background?
- 2:21
- Unidentified speaker (speaker_1): Yeah, so uh my parents have uh so my fa father has a uh business busin manufacturing. manufacturing business where we manufacture pharmaceutical machinery, so things like fluid bed dryers and other machines that are used to build pharmaceutical tablets. Uh and my mom has a uh a salon chain in in in Mumbai
- 2:41
- Unidentified speaker (speaker_1): for uh for women uh and she started a brand for a cosmetics brand for uh selling to other salons.
- 2:49
- Unidentified speaker (speaker_0): Very entrepreneurial family. So
- 2:51
- Unidentified speaker (speaker_0): I guess one of the questions I normally ask people is kind of obvious answer from you, which is always knew you'd be an entrepreneur?
- 3:00
- Unidentified speaker (UNKNOWN): Yeah,
- 3:00
- Unidentified speaker (speaker_1): pretty
- 3:01
- Unidentified speaker (UNKNOWN): much.
- 3:02
- Unidentified speaker (speaker_1): Yeah, so I
- 3:03
- Unidentified speaker (UNKNOWN): there
- 3:03
- Unidentified speaker (speaker_1): are actually like subtle cues about like when you're growing up in language which is implied. Like one example is that my parents would always tell me that you should do a job for experience. But
- 3:14
- Unidentified speaker (UNKNOWN): if
- 3:14
- Unidentified speaker (speaker_1): you really like decode what that means, like experience for what?
- 3:17
- Unidentified speaker (UNKNOWN): Well,
- 3:17
- Unidentified speaker (speaker_1): they're tr what it means is that it's like an intermediate checkpoint to like another eventual goal. And that goal it was never mentioned, but I never asked because I knew what it was, which was to start a company. So they
- 3:28
- Unidentified speaker (UNKNOWN): would
- 3:28
- Unidentified speaker (speaker_1): always say that you know you should do a job for two or three years to get experience.
- 3:33
- Unidentified speaker (UNKNOWN): And I
- 3:33
- Unidentified speaker (speaker_1): knew what that experience was for.
- 3:34
- Unidentified speaker (speaker_0): And when you were interviewing for the jobs, you told them that? No, I'm
- 3:37
- Unidentified speaker (UNKNOWN): not sure. Yeah.
- 3:38
- Unidentified speaker (speaker_1): Yeah. Yeah, I told
- 3:39
- Unidentified speaker (UNKNOWN): them that
- 3:39
- Unidentified speaker (speaker_1): I've
- 3:40
- Unidentified speaker (speaker_0): But you know, coming from Mumbai,
- 3:42
- Unidentified speaker (speaker_0): the we have so many business families in India, right? And I think this is this is something that you kind of learn on the knee or at dinner tables growing up sometimes. So many of them can insp uh get aspired or inspired to do uh these kind of uh businesses. So you did computer
- 4:02
- Unidentified speaker (speaker_0): science. you came to the US, you came to CMU, which at that time people may not know, this is 2018. 19.
- 4:10
- Unidentified speaker (speaker_0): It was always considered the AI. I did AI by the way, in my IIT,
- 4:16
- Unidentified speaker (UNKNOWN): so
- 4:16
- Unidentified speaker (speaker_0): I know CMU was. So you knew when when did this this is way before Chat GPT?
- 4:22
- Unidentified speaker (speaker_1): Yes.
- 4:23
- Unidentified speaker (speaker_0): This is way before, you know how when AI became so cool. So how did that interest happen?
- 4:30
- Unidentified speaker (UNKNOWN): Yeah,
- 4:30
- Unidentified speaker (speaker_1): so this was at a time when if you said you worked on AI, it means you are not a serious practitioner.
- 4:35
- Unidentified speaker (speaker_0): So
- 4:35
- Unidentified speaker (speaker_1): you had to either pick up vertical, machine learning meant like you're doing theory, computer vision meant you're doing uh AI for images and videos.
- 4:43
- Unidentified speaker (UNKNOWN): So
- 4:43
- Unidentified speaker (speaker_1): I think it happened when I was in college, deep learning just started taking off. So it was the time when neural networks, as we call them, uh just started working at a scale which was like unprecedented.
- 4:54
- Unidentified speaker (UNKNOWN): And
- 4:54
- Unidentified speaker (speaker_1): like many things started coming together. We started having data at a very high scale. GPUs, people figured out that you can use these graphic computing
- 5:04
- Unidentified speaker (speaker_1): processors, which are GPUs. And you actually programmed them to run AI models, which was not what they were meant
- 5:10
- Unidentified speaker (UNKNOWN): for.
- 5:11
- Unidentified speaker (speaker_1): And also you started having advancements in on the algorithmic side, which is like you can scale these neural networks and get them to do things they they couldn't do before.
- 5:19
- Unidentified speaker (UNKNOWN): So
- 5:19
- Unidentified speaker (speaker_1): all these three things started coming together when I was in undergrad. And it was a time when the field hadn't progressed so much, where any motivated individual from any part of the world could just self-teach their way to the top. And basically you could get to the state-of-the-art state of the art in uh the field that you were interested in, with AI for images, let's say. Uh within the span of
- 5:39
- Unidentified speaker (speaker_1): six months you could go from just knowing how to code till code like being able to build the state of the art uh image recognition systems.
- 5:47
- Unidentified speaker (UNKNOWN): So
- 5:47
- Unidentified speaker (speaker_1): I think
- 5:49
- Unidentified speaker (speaker_0): so the Transformer paper came out in 2017. Yeah. When you were doing all of this in 1819, were you familiar and was there a BF was it a BFD even at that time that this is going to change the world?
- 6:01
- Unidentified speaker (speaker_1): Yeah, so before the Transformers paper. like there were uh
- 6:05
- Unidentified speaker (speaker_1): like people had were researching different techniques independently for different modalities, as we call it. So text had their own research fields and their own advancements. Uh images and computer vision, which is like using AI on images and videos, had their own independent stream of research, and similar for speech and other
- 6:25
- Unidentified speaker (speaker_1): modalities.
- 6:27
- Unidentified speaker (UNKNOWN): We
- 6:27
- Unidentified speaker (speaker_1): still had attention mechanisms, but in these silos back then.
- 6:31
- Unidentified speaker (UNKNOWN): It
- 6:32
- Unidentified speaker (speaker_1): was actually after the Transformer paper, which first showed that you could scale these models. And the real breakthrough of the Transformers paper was that you can train them in parallel, which was not possible in the architectures that came before
- 6:45
- Unidentified speaker (UNKNOWN): it. And
- 6:46
- Unidentified speaker (speaker_1): then a few years after the Transformers paper, people started, there were like independent breakthroughs which showed that you can use it for images, you can use it for robotics. So just and that's when we saw like a convergence of the fields. So AI looked very different when I got started.
- 6:59
- Unidentified speaker (speaker_0): Yeah.
- 7:00
- Unidentified speaker (speaker_1): And now it's pretty much like if you uh just studied text, you're basically 80% of the way to imagine.
- 7:06
- Unidentified speaker (UNKNOWN): Yeah, you know, you know, I was thinking
- 7:07
- Unidentified speaker (speaker_0): of that because we said we are trying to bring entrepreneurship down to a level where people can look at you and say, I can do this. But so far, listening to you, I feel oh shit, even I can't do this because this guy is so technical and knew all of this for so long. I know your view is different. You have talked about how the world has changed. We'll come to that. Let's fast forward then into
- 7:27
- Unidentified speaker (speaker_0): post CMU. What was the work experience and how did we get to this idea?
- 7:31
- Unidentified speaker (speaker_1): Yeah, so uh I was working on medical AI research at CMU,
- 7:36
- Unidentified speaker (UNKNOWN): uh,
- 7:36
- Unidentified speaker (speaker_1): which was using AI on computer vision projects. problems for like medical images.
- 7:41
- Unidentified speaker (UNKNOWN): Uh
- 7:41
- Unidentified speaker (speaker_1): so that got me uh like it prepared me for my next role at Meta as a researcher in computer vision, where we worked on building foundation models for 3D avatars. Uh so super crack team. I mean I think probably
- 7:53
- Unidentified speaker (UNKNOWN): Which
- 7:53
- Unidentified speaker (speaker_0): years was this?
- 7:54
- Unidentified speaker (speaker_1): This was 21.
- 7:56
- Unidentified speaker (UNKNOWN): And
- 7:56
- Unidentified speaker (speaker_0): so this was even before the whole metaverse. Meta hadn't even changed its name yet, right?
- 8:01
- Unidentified speaker (speaker_1): I think they changed, yeah. So like the name. of the lab changed a few times, like
- 8:04
- Unidentified speaker (UNKNOWN): yeah
- 8:04
- Unidentified speaker (speaker_0): where
- 8:05
- Unidentified speaker (speaker_1): even we stopped tracking after a while.
- 8:06
- Unidentified speaker (speaker_0): Yeah.
- 8:07
- Unidentified speaker (speaker_1): But yeah, I think it was right about the time when Meta changed its name. It was during COVID, so m metaverse maybe happened.
- 8:13
- Unidentified speaker (UNKNOWN): And
- 8:13
- Unidentified speaker (speaker_0): Metaverse was correct.
- 8:14
- Unidentified speaker (UNKNOWN): Yeah,
- 8:14
- Unidentified speaker (speaker_0): yeah. So
- 8:15
- Unidentified speaker (speaker_1): so the idea then was that uh what would like if metaverse takes off, how can communication change? Like what would be the future of FaceTime? So can we have the same conversation where you're in India, I'm in SF? We put on our headsets and it's indistinguishable from a real conversation.
- 8:30
- Unidentified speaker (speaker_0): Yeah.
- 8:30
- Unidentified speaker (speaker_1): It's almost like
- 8:32
- Unidentified speaker (UNKNOWN): uh
- 8:32
- Unidentified speaker (speaker_1): holographic. I think
- 8:33
- Unidentified speaker (UNKNOWN): it's called
- 8:34
- Unidentified speaker (speaker_1): photorealistic avatars.
- 8:35
- Unidentified speaker (speaker_0): Yeah.
- 8:35
- Unidentified speaker (speaker_1): But for people that are not familiar with it, you can imagine as like a holograms that are like indistinguishable from real conversations.
- 8:42
- Unidentified speaker (UNKNOWN): So how many
- 8:42
- Unidentified speaker (speaker_0): years were you there?
- 8:43
- Unidentified speaker (UNKNOWN): I
- 8:43
- Unidentified speaker (speaker_1): was there for three years.
- 8:44
- Unidentified speaker (speaker_0): Okay. And then after that?
- 8:46
- Unidentified speaker (speaker_1): After that, uh started Stalin Labs. Uh we actually my co co-founder founder was also at Meta with me. and he was also at CMU. Uh so super crack guy. Uh and we were basically pursuing different projects independently with the hope that uh we could turn just running experiments at scale.
- 9:03
- Unidentified speaker (speaker_0): And had you discussed becoming co-founders while at CMU? That you
- 9:07
- Unidentified speaker (UNKNOWN): no, not as so we
- 9:08
- Unidentified speaker (speaker_1): went to CMU together.
- 9:09
- Unidentified speaker (speaker_0): Yeah.
- 9:09
- Unidentified speaker (UNKNOWN): But
- 9:09
- Unidentified speaker (speaker_1): we really became friends at Meta.
- 9:11
- Unidentified speaker (speaker_0): Okay. Uh
- 9:11
- Unidentified speaker (UNKNOWN): so
- 9:12
- Unidentified speaker (speaker_0): I think one lesson I want to call out because it's a pattern that keeps coming, um, is generally people have met that and your co founder is your technical co founder.
- 9:21
- Unidentified speaker (UNKNOWN): Yes.
- 9:21
- Unidentified speaker (speaker_0): I mean it's almost like you followed the perfect script that we keep talking about. You knew him knew him for a while, you worked with him, um, and you were very clear that when you started you would have a co founder.
- 9:33
- Unidentified speaker (speaker_1): Um actually no. So both of us had started building products by ourselves.
- 9:38
- Unidentified speaker (speaker_0): Yeah.
- 9:39
- Unidentified speaker (speaker_1): And we actually realized that you know we're both friends. We both have similar ambitions. Yeah. We're both extremely technical. So if we teamed up, it would be like a super team. Uh so
- 9:49
- Unidentified speaker (UNKNOWN): mainly for
- 9:49
- Unidentified speaker (speaker_0): skill skill reasons. Any other reason, like why not just go ahead and do it yourselves?
- 9:55
- Unidentified speaker (speaker_1): Uh I think that we are yeah, we are an extremely strong team together. Uh and also we weren't waiting for permission, we weren't waiting until we had a co-founder to start. building.
- 10:03
- Unidentified speaker (UNKNOWN): We
- 10:04
- Unidentified speaker (speaker_1): both saw that, that both of us are extremely high agency. So I started building products and running experiments. He was building his products and experiments. So we were kind of following each other's journey and uh like far before we became co-founders.
- 10:16
- Unidentified speaker (speaker_0): So over the course of many episodes, we have heard many reasons, one of which is this complementary Complimentary skill sets and force multiplier. Sometimes it's the emotional support as well.
- 10:27
- Unidentified speaker (speaker_0): You know, because the journey can be up and down, and I'm sure that applies as well. So, when did you guys get started?
- 10:34
- Unidentified speaker (speaker_1): 25 is back when the company started and we applied to YC together and things like that.
- 10:39
- Unidentified speaker (speaker_0): Okay, so with that, I think that's a good time to take the first question. So this is from Ruthwik. Twick. NetNet understands starting up requires immense correct courage and sprinkle of delusion.
- 10:52
- Unidentified speaker (speaker_0): Basically, how did you think about your competence and get comfortable about what you guys are building as being the opportunity? You also mentioned to me offline earlier about thinking about construction industry. So talk through how you guys were you with hindsight
- 11:12
- Unidentified speaker (speaker_0): when you look at it, what part was delusion? what part was conviction? How did you make sure that your conviction was grounded in a lot of research and data, or was there something else?
- 11:23
- Unidentified speaker (UNKNOWN): Yeah,
- 11:23
- Unidentified speaker (speaker_1): I think like there are different schools of thoughts here. Like there's one school of thought which kind of talks about like you should talk to a lot of people, make sure that there's a problem before you try to solve it.
- 11:34
- Unidentified speaker (UNKNOWN): But
- 11:34
- Unidentified speaker (speaker_1): I think a lot of that advice was like before building a prototype and building a product was as cheap as it is today. So today the cost of experimenting as especially for software, I should preface that. Maybe for like other uh
- 11:48
- Unidentified speaker (speaker_0): Yeah, you're
- 11:48
- Unidentified speaker (UNKNOWN): just
- 11:48
- Unidentified speaker (speaker_0): giving the
- 11:49
- Unidentified speaker (UNKNOWN): exact
- 11:49
- Unidentified speaker (speaker_0): opposite advice that I give.
- 11:51
- Unidentified speaker (UNKNOWN): Pretty much.
- 11:52
- Unidentified speaker (speaker_0): But that's good. I do think what I consider tech out businesses,
- 11:56
- Unidentified speaker (UNKNOWN): yeah. We are
- 11:56
- Unidentified speaker (speaker_0): selling sitting in Silicon Valley. I do think the path may be different. So go ahead.
- 12:00
- Unidentified speaker (speaker_1): Yeah. Also in in tech, I would say that. some of the biggest companies we know today started before the founders thought it would be a business. So these were products and maybe like uh cutting
- 12:13
- Unidentified speaker (UNKNOWN): edge tech.
- 12:13
- Unidentified speaker (speaker_0): No, even
- 12:14
- Unidentified speaker (UNKNOWN): tools that were just in
- 12:15
- Unidentified speaker (speaker_0): OpenAI. I
- 12:15
- Unidentified speaker (UNKNOWN): mean
- 12:16
- Unidentified speaker (speaker_0): they thought, you know, but they they started out to be. But
- 12:19
- Unidentified speaker (UNKNOWN): like
- 12:19
- Unidentified speaker (speaker_1): m uh there's an interview by Mark Zuckerberg, I think, in one of the early startup schools at YC, where he said that he had already built Facebook, it had millions of users, and he still thought he was still thinking and telling his friend that one day another big company will build a social media that'll
- 12:35
- Unidentified speaker (UNKNOWN): take over the world.
- 12:35
- Unidentified speaker (speaker_0): Yes. And Facebook, by the way, was
- 12:37
- Unidentified speaker (UNKNOWN): not the first
- 12:37
- Unidentified speaker (speaker_0): one. There was
- 12:38
- Unidentified speaker (UNKNOWN): one before
- 12:39
- Unidentified speaker (speaker_0): that called MySpace. Right.
- 12:40
- Unidentified speaker (UNKNOWN): No, but
- 12:40
- Unidentified speaker (speaker_1): I mean from Mark Zuckerberg's
- 12:41
- Unidentified speaker (UNKNOWN): perspective,
- 12:42
- Unidentified speaker (speaker_1): he was not building with the intention that okay, I'm I'm gonna build something which will become a trillion dollar company.
- 12:48
- Unidentified speaker (speaker_0): Yeah.
- 12:49
- Unidentified speaker (speaker_1): He was building because he just he just couldn't live without building it.
- 12:53
- Unidentified speaker (speaker_0): Yeah. Uh and
- 12:53
- Unidentified speaker (speaker_1): I think a lot of
- 12:54
- Unidentified speaker (speaker_0): Very good. I'm gonna pause you there. Yeah. He was building it because he couldn't live without building it. That's one very good reason to do it.
- 13:00
- Unidentified speaker (UNKNOWN): Yeah,
- 13:00
- Unidentified speaker (speaker_1): so I think a lot of times like you should almost all the builders I know, among friends who are founders too,
- 13:07
- Unidentified speaker (UNKNOWN): like
- 13:07
- Unidentified speaker (speaker_1): at any given time they're just trying building whatever they can build, and they are just building things that itch their own curiosity. So
- 13:14
- Unidentified speaker (UNKNOWN): I would
- 13:14
- Unidentified speaker (speaker_1): recommend like what whatever
- 13:16
- Unidentified speaker (UNKNOWN): like
- 13:16
- Unidentified speaker (speaker_1): whatever phase of life or whatever situation you're in, there's something you can build which pushes some pro makes some problem easier to solve. Uh even if it's just for yourself, you should build it. If it's in software, you can open source it. It's a very easy way to get others to get give feedback on you. Uh, and you might just find that there's more demand than you
- 13:35
- Unidentified speaker (speaker_1): anticipated.
- 13:36
- Unidentified speaker (speaker_0): Okay, super. So let's keep going. So you didn't how did you come up with the idea even with that point of view?
- 13:45
- Unidentified speaker (speaker_1): Yeah, so uh so just just to recap, like what we're building now, we're building open source voice models uh that are so efficient that they can run on a CPU.
- 13:53
- Unidentified speaker (speaker_0): Sorry, just so people know. Now before and this was 2025.
- 13:58
- Unidentified speaker (UNKNOWN): Yes.
- 13:58
- Unidentified speaker (speaker_0): You started with open source because at
- 14:00
- Unidentified speaker (UNKNOWN): that time open source
- 14:01
- Unidentified speaker (speaker_0): wasn't hot.
- 14:02
- Unidentified speaker (speaker_1): Yeah. Actually we didn't. So we applied to YC with uh another idea, which was a local alternative to 11 Labs.
- 14:09
- Unidentified speaker (speaker_0): Yeah.
- 14:09
- Unidentified speaker (speaker_1): But we were not building our own models back then. So we
- 14:12
- Unidentified speaker (UNKNOWN): please
- 14:12
- Unidentified speaker (speaker_0): explain what 11 Labs does, and
- 14:14
- Unidentified speaker (UNKNOWN): please
- 14:14
- Unidentified speaker (speaker_0): explain what local alternative means.
- 14:21
- Unidentified speaker (speaker_1): Models, Things like uh dubbing a sound into a different language, converting text to speech, and other AI models that process sound and human speech specifically.
- 14:32
- Unidentified speaker (UNKNOWN): This
- 14:33
- Unidentified speaker (speaker_1): was in 2025, this was when they just like had everyone knew that 11 Labs is this big incumbent. And open source models started popping up, which did the same thing as 11 Labs, but you can run them locally, you can download it, you can run them on your own servers. And we thought that if we can build a very nice UI to
- 14:52
- Unidentified speaker (speaker_1): make open source models accessible to non technical developers, that would be a game changer because at that time only developers knew how to use open source models. So that was the core insight that open source models, game changing technology, but only restricted to developers. How can you make it accessible to non developers? So we built a local alternative to that. You had a very simple
- 15:12
- Unidentified speaker (speaker_1): UI where you can run the world's best open. source models without ever needing to write code, without ever needing to set up your own environment, and all the processing happens on your laptop. We I think over in a week we got like 600 downloads or so, which
- 15:25
- Unidentified speaker (UNKNOWN): is
- 15:26
- Unidentified speaker (speaker_1): enough to get feedback, and we applied to YC with that.
- 15:29
- Unidentified speaker (UNKNOWN): But
- 15:29
- Unidentified speaker (speaker_1): what we our hypothesis was that people cared about open source models because of privacy reasons. They wanted want it to run on their own hardware. But when we try to price it at the same price at as 11 labs, so that was the experiment. Like if you price it at the same price as 11 labs, do your users churn or do they still uh convert? And if they convert it means that it's they're
- 15:49
- Unidentified speaker (speaker_1): paying for the privacy and other reasons. But all our users churned. The second we started pricing it at like a $99 a month plan. So that's when we realized that the we are were building a very useful technology but for the wrong audience. We were building for non-technical people who were like maybe run digital marketing ads and things like that.
- 16:09
- Unidentified speaker (speaker_1): If we could actually build for developers, developers, we can get bottoms-up adoption there. These developers would work at enterprises, and then that enterprises would come and purchase enterprise licenses for our models. And while we were building this local alternative to 11 Labs, we realized that the open source model is actually not very good.
- 16:29
- Unidentified speaker (speaker_0): But before we go there, so with your hypothesis that if you were to build for developers, XYZ would happen, why couldn't 11 labs do the same thing?
- 16:39
- Unidentified speaker (speaker_1): But one thing to observe for developers is that every closed source product for developers opens up an even bigger market for open source alternatives. And this is because developers they have the skill to make changes, customize products for their own use case. They want to host on their own hardware.
- 16:58
- Unidentified speaker (speaker_1): Some industries have regulations where data can't be crossing through a network. And there are many use cases where actually the open source alternative is an extremely popular and commercially successful product. And we've already seen this with in LLMs. Like if you see the most used AI LLM today is actually DeepSeek and other Chinese models.
- 17:17
- Unidentified speaker (speaker_0): I'm on DeepSeek.
- 17:18
- Unidentified speaker (speaker_1): Yeah. Yeah. So I like that's another observation that all the open source AI, the center of gravity for that is in China.
- 17:24
- Unidentified speaker (speaker_0): Yeah.
- 17:25
- Unidentified speaker (speaker_1): So like if we can pull the center of gravity towards us, that would be a big metric of success.
- 17:32
- Unidentified speaker (speaker_0): So let's take the next question.
- 17:34
- Unidentified speaker (speaker_2): How do I know what our bit is actually working?
- 17:36
- Unidentified speaker (speaker_0): So the question is: how do you how do we know that it's actually working? You've taken some bold calls early, pivoted early, which I think in this world you peep I think there are many lessons for people who are watching you to take away, which is you started with a point of view, you don't necessarily didn't necessarily do
- 17:55
- Unidentified speaker (speaker_0): customer research, which makes sense for tech out businesses. But then you pivoted very early. You had an experiment, you were intellectually honest. to recognize that if it is $99 is making everyone churn something is not working.
- 18:08
- Unidentified speaker (UNKNOWN): Yeah.
- 18:08
- Unidentified speaker (speaker_0): Then open source has happened. So to this question, how do you know it's working? And I know there's there are plans further ahead also. But how do you one? I think you partly answered it, charge for it.
- 18:22
- Unidentified speaker (speaker_1): Yeah.
- 18:23
- Unidentified speaker (speaker_0): And and you realized realize and you pivoted. But any other advice, how do you, as you look for forward, how are you doing
- 18:29
- Unidentified speaker (UNKNOWN): the right? But
- 18:29
- Unidentified speaker (speaker_1): just a correction, I don't mean don't do customer research. Yeah. I mean there are times.
- 18:34
- Unidentified speaker (UNKNOWN): Yes. Please
- 18:34
- Unidentified speaker (speaker_0): do customer research.
- 18:35
- Unidentified speaker (UNKNOWN): There are
- 18:36
- Unidentified speaker (speaker_1): times when building a product is faster than doing customer research. Like instead of
- 18:40
- Unidentified speaker (UNKNOWN): the response. And these are those
- 18:41
- Unidentified speaker (speaker_0): times.
- 18:41
- Unidentified speaker (speaker_1): Yeah. So like if if you're building a software product, let's say you want to build a product which can analyze uh, I don't know, construction legal documents in construction.
- 18:50
- Unidentified speaker (UNKNOWN): Yeah.
- 18:50
- Unidentified speaker (speaker_0): Now
- 18:50
- Unidentified speaker (speaker_1): that is faster to build than it is to reach out to someone in construction, wait for them to respond. And uh But
- 18:56
- Unidentified speaker (speaker_0): this is a very important distinction you're making, Rohan, which I want to bring out for people. Yeah. For example, what you are I think you are saying, and correct me if I'm wrong, that if you're building for developers, build it and put it out there on GitHub or somewhere when you you'll have the real customer research at scale.
- 19:11
- Unidentified speaker (UNKNOWN): Yeah.
- 19:11
- Unidentified speaker (speaker_0): We are just saying it's a different way of getting customer research, right?
- 19:14
- Unidentified speaker (speaker_1): Exactly. And you should do whatever is the fastest way for to make progress. Even if you let's say you want to build something in construction, let's say you should reach out to as many people as you can, try to uh talk to them. But by the time they respond, you should already have a prototype or a product that they can use and give you feedback for. So and like feedback is what like you should do
- 19:34
- Unidentified speaker (speaker_1): whatever gets you the feedback the fastest.
- 19:35
- Unidentified speaker (speaker_0): Yeah. Uh
- 19:36
- Unidentified speaker (UNKNOWN): and
- 19:36
- Unidentified speaker (speaker_0): at scale.
- 19:37
- Unidentified speaker (UNKNOWN): And
- 19:37
- Unidentified speaker (speaker_1): at scale. I mean I think it's hard to get feedback at scale, but uh I think PG has this essay where if you can get hundred people to love it, like that is the fastest way f to build something that a million people can love.
- 19:49
- Unidentified speaker (speaker_0): Yeah. Yeah.
- 19:49
- Unidentified speaker (UNKNOWN): So
- 19:49
- Unidentified speaker (speaker_1): you could you don't need to get feed you don't need to get a million people to
- 19:53
- Unidentified speaker (speaker_0): developers but in particular tend to be the biggest evangelists, right? So
- 19:57
- Unidentified speaker (speaker_1): developers are extremely easy to engage.
- 19:59
- Unidentified speaker (UNKNOWN): Like
- 19:59
- Unidentified speaker (speaker_1): You can put things on Hacker News, you can put things on Reddit, but they are extremely hard to please and impress.
- 20:05
- Unidentified speaker (speaker_0): Yeah.
- 20:05
- Unidentified speaker (speaker_1): Because a lot of like all developers think
- 20:07
- Unidentified speaker (UNKNOWN): they can just rebuild it and
- 20:09
- Unidentified speaker (speaker_1): do a better job.
- 20:10
- Unidentified speaker (speaker_0): Yeah.
- 20:10
- Unidentified speaker (speaker_1): Uh but on the occasion, like what happened with us is that we put out a model on Reddit. It went viral there among developers.
- 20:17
- Unidentified speaker (UNKNOWN): To
- 20:17
- Unidentified speaker (speaker_1): the extent where the traction we were seeing, like we were questioning it ourselves. Like we thought something had gone wrong.
- 20:22
- Unidentified speaker (speaker_0): Yeah.
- 20:23
- Unidentified speaker (speaker_1): Uh we were one of the fastest growing open source projects uh in YC.
- 20:27
- Unidentified speaker (speaker_1): And even like the when we were looking at when we were every time we were refreshing, we would just see like a number pop up in how many GitHub stars or downloads that we were seeing, and we couldn't believe it. We thought something was wrong.
- 20:38
- Unidentified speaker (UNKNOWN): And I
- 20:38
- Unidentified speaker (speaker_1): think maybe that is one way to know things are working, is that
- 20:42
- Unidentified speaker (UNKNOWN): you
- 20:42
- Unidentified speaker (speaker_1): start questioning your analytics.
- 20:44
- Unidentified speaker (speaker_0): Yeah, yeah. No, so Mark Andreessen Andresen has this view on product market fit. Uh And I don't know if you've heard of that podcast very. where he says when you have product market fit, you're not questioning whether you have product market fit. People are
- 20:56
- Unidentified speaker (UNKNOWN): breaking
- 20:57
- Unidentified speaker (speaker_0): it down. You know,
- 20:57
- Unidentified speaker (UNKNOWN): they're
- 20:58
- Unidentified speaker (speaker_0): breaking the doors down. You have that feeling that, like you're saying, is something wrong. Why why are
- 21:03
- Unidentified speaker (UNKNOWN): the the
- 21:03
- Unidentified speaker (speaker_0): results so good? So super. I'm gonna take the last question and then uh continue with with the story.
- 21:11
- Unidentified speaker (UNKNOWN): As
- 21:11
- Unidentified speaker (speaker_2): a founder building an in AI, how do you think AI changes? industries?
- 21:16
- Unidentified speaker (speaker_0): So AI is obviously changing a lot of industries. You should talk. We still didn't get to the construction industry piece of it. So we should talk about what your views on some of those industries. Also, the fact that everyone I many companies I speak to are doing voice, and a lot of people would be saying, I keep hearing
- 21:36
- Unidentified speaker (speaker_0): voice, voice, voice.
- 21:38
- Unidentified speaker (UNKNOWN): You are
- 21:38
- Unidentified speaker (speaker_0): also obviously in the business of doing voice. How is that working? Then we'll come to open source. And specifically, how open source people should think about them differently.
- 21:47
- Unidentified speaker (UNKNOWN): So
- 21:48
- Unidentified speaker (speaker_1): I think let's let me talk about my industry, which is uh building voice AI for developers. Uh so before AI before AI became capable of producing human-like human like speech,
- 22:00
- Unidentified speaker (UNKNOWN): uh our
- 22:01
- Unidentified speaker (speaker_1): industry who had like customer support where we had like call centers filled with humans.
- 22:07
- Unidentified speaker (UNKNOWN): Uh
- 22:07
- Unidentified speaker (speaker_1): you had to train them, you had to make sure that they are following the script, they had to be trained.
- 22:12
- Unidentified speaker (UNKNOWN): I
- 22:12
- Unidentified speaker (speaker_1): think the biggest change in the last few years has been that AI has crossed the threshold of natural naturalness, where a lot of customer support can ac can be done better with AI than it can with a human representative. So you still need cases where humans are intervening, things that AI cannot
- 22:33
- Unidentified speaker (speaker_1): cannot make decisions for, things like issuing a refund. Like could be one of the cases where you still need a human to make a decision. But customer support as a whole has improved massively.
- 22:44
- Unidentified speaker (UNKNOWN): By the way, I
- 22:45
- Unidentified speaker (speaker_0): I'm you m I'm sure
- 22:46
- Unidentified speaker (UNKNOWN): you
- 22:46
- Unidentified speaker (speaker_0): have heard this, but we have companies that uh power financial services industries. where AI agents are better at doing collections because they're just not emotional.
- 22:55
- Unidentified speaker (speaker_1): Yeah.
- 22:56
- Unidentified speaker (speaker_0): You can shout and scream and they'll just keep repeating it in the same way and maybe little bit more warm and it's amazing. The
- 23:04
- Unidentified speaker (UNKNOWN): the
- 23:04
- Unidentified speaker (speaker_0): results are amazing.
- 23:05
- Unidentified speaker (speaker_1): So I think like that is something which uh
- 23:08
- Unidentified speaker (UNKNOWN): was
- 23:08
- Unidentified speaker (speaker_1): not possible before AI. Like we still had like hard hard-coded coded chatbots and we had some companies that did that. But this is like with AI, you can just expand the number of scenarios that you can deal
- 23:18
- Unidentified speaker (UNKNOWN): with. So
- 23:19
- Unidentified speaker (speaker_1): I think that's number one.
- 23:20
- Unidentified speaker (UNKNOWN): And
- 23:21
- Unidentified speaker (speaker_1): you can kind of like draft the story of voice AI in like three different inflection points.
- 23:25
- Unidentified speaker (UNKNOWN): We
- 23:25
- Unidentified speaker (speaker_1): are in the middle of the first one, which is where AI can start replacing a lot of enterprise customer support, enterprise collection, outbound calls, where it is not as natural as a human speaker. but it can kind of automate a lot of tasks. So, like the use case of automation in enterprises.
- 23:44
- Unidentified speaker (UNKNOWN): I
- 23:44
- Unidentified speaker (speaker_1): think the second inflection point, which we will be entering very soon, is that of interfaces. So we already have like Whisperflow is one example of where like a unidirectional interface for voice where all electronics can just be interacted with using voice.
- 23:58
- Unidentified speaker (UNKNOWN): We
- 23:58
- Unidentified speaker (speaker_1): are still waiting for the bi-directional uh version of that, but that's that will be like one of the next inflection points for voice AI.
- 24:06
- Unidentified speaker (speaker_0): Which chat GPT launched, right? With the
- 24:09
- Unidentified speaker (speaker_1): Yes, but that's for interacting with Chat GPT, not with
- 24:11
- Unidentified speaker (speaker_0): a full duplex.
- 24:12
- Unidentified speaker (UNKNOWN): Yeah, yeah. Like
- 24:13
- Unidentified speaker (speaker_1): for example, like uh for the next billion users on the internet, if uh AI is not like you can use they they should be able to use voice
- 24:21
- Unidentified speaker (UNKNOWN): to
- 24:21
- Unidentified speaker (speaker_1): fully interact with their
- 24:22
- Unidentified speaker (UNKNOWN): computer,
- 24:23
- Unidentified speaker (speaker_1): get all education services, government services, anything that they would have. ever need should be accessible by just
- 24:28
- Unidentified speaker (UNKNOWN): voice. And I
- 24:29
- Unidentified speaker (speaker_1): think that's what we'll see in the next few years. And the final inflection point is like a merging of the modalities where video, text, image, all like basically we're just we just have real-time video streaming on our phones.
- 24:41
- Unidentified speaker (UNKNOWN): Like
- 24:42
- Unidentified speaker (speaker_1): right now your iPhone screen looks very similar to mine with like icons and stuff.
- 24:45
- Unidentified speaker (speaker_0): Yeah. I
- 24:46
- Unidentified speaker (UNKNOWN): think
- 24:46
- Unidentified speaker (speaker_1): like maybe in the next decade like all of us will have fully different uh
- 24:50
- Unidentified speaker (UNKNOWN): each
- 24:50
- Unidentified speaker (speaker_1): pixel is being generated by an AI model running somewhere.
- 24:53
- Unidentified speaker (speaker_0): Yeah, yeah. No, super. So
- 24:57
- Unidentified speaker (speaker_0): moving forward from that, so you have some exciting products coming out. Can we talk about it?
- 25:02
- Unidentified speaker (speaker_1): Yeah.
- 25:02
- Unidentified speaker (speaker_0): Yeah.
- 25:03
- Unidentified speaker (UNKNOWN): So
- 25:03
- Unidentified speaker (speaker_0): talk about what you have a point of view on how computing is moving, uh, moving to the edge, what you guys are doing. And then for people watching you, they will say
- 25:15
- Unidentified speaker (speaker_0): Rohan is just born to do this. He was doing this all the time. This is very technically complex. No chance for us. But you have a point of view that open source has changed a bunch of that. So maybe cover both.
- 25:29
- Unidentified speaker (UNKNOWN): Yeah,
- 25:30
- Unidentified speaker (speaker_1): I think let me let me cover the second part first. I
- 25:33
- Unidentified speaker (UNKNOWN): think you
- 25:33
- Unidentified speaker (speaker_1): shouldn't really like the easiest person to fool is yourself. So you should fool yourself into believing. that you can do anything.
- 25:41
- Unidentified speaker (speaker_0): You should.
- 25:42
- Unidentified speaker (speaker_1): You should, yes.
- 25:43
- Unidentified speaker (speaker_0): Okay.
- 25:44
- Unidentified speaker (speaker_1): I think
- 25:44
- Unidentified speaker (UNKNOWN): that at
- 25:46
- Unidentified speaker (speaker_1): any given point, like you should just try to do whatever is possible for you. I think like a term used for this is like resourcefulness. Like whatever resources are available to you, you can kind of use them and
- 25:57
- Unidentified speaker (speaker_1): harness them to your advantage. And
- 26:01
- Unidentified speaker (UNKNOWN): as
- 26:01
- Unidentified speaker (speaker_1): many other things in life, that compounds by itself. Like it takes time to be really good at something. But if you're just focusing on being better than you were yesterday, I think that you can eventually get to a point where others think this is impossible to do or that you were born to do this. Like when you look at someone and feel like they were born to do this, of you attribute to
- 26:21
- Unidentified speaker (speaker_1): talent what has been compounding over many years. So I think like that's how I would kind of that's how I view the second question that you've
- 26:29
- Unidentified speaker (UNKNOWN): got to do. No,
- 26:29
- Unidentified speaker (speaker_0): but that's very good and I do want, I think
- 26:33
- Unidentified speaker (speaker_0): what people may not appreciate fully is how much AI gives you, if harnessed properly, gives you a leg up.
- 26:44
- Unidentified speaker (speaker_1): Yes.
- 26:45
- Unidentified speaker (speaker_0): Uh And now AI plus open source, given that it is commoditized, not commoditized but available everywhere, uh so on and so forth, should be taken and looked upon as an asset where potentially the four or five years of grind you had to go through, you don't have to go through.
- 27:01
- Unidentified speaker (speaker_1): Yes.
- 27:02
- Unidentified speaker (speaker_0): As opposed to saying if I can't if I don't do that grind, I cannot do this. So I think that's that's important for people to take away.
- 27:09
- Unidentified speaker (speaker_1): Yeah.
- 27:10
- Unidentified speaker (speaker_0): And and tell us more about the exciting things happen happening in your world. Yeah.
- 27:14
- Unidentified speaker (speaker_1): So I think that what we're trying to do is we're trying to make voice AI democratized to the point where it's a no-brainer to use it everywhere. And we're doing that by building the world's most efficient voice models that
- 27:28
- Unidentified speaker (UNKNOWN): like
- 27:28
- Unidentified speaker (speaker_1): most people don't have access to a GPU in the world. Most developers are still trying to cram open source models on their local uh CPUs and they don't work as well. And we're going to be building for all of them. So you can take our model for something that would traditionally require tens of thousands of dollars a month to run. You can now take a $20 a month
- 27:48
- Unidentified speaker (speaker_1): DPS server or just on your local laptop and run it there. So we're building a new kind of models called ultra-low precision text-to-speech models, which we have invented and we will be uh open sourcing it. So everyone can run it on their own hardware.
- 28:04
- Unidentified speaker (speaker_0): Superb. One more question because someone asked me this yesterday, Rovan. So let's say there's a guy sitting in Mumbai, in India, somewhere, watching you. Let's say they are they have that uh self-confidence you spoke about. Believe believe you can do anything, they want to do
- 28:21
- Unidentified speaker (UNKNOWN): it.
- 28:22
- Unidentified speaker (speaker_0): Is it important for them to come to the US? Should they be in the valley? And somebody asked. me yesterday specifically about Y Combinator. I'm not going to put you on the spot because they're very good friends of ours, also. But in general, are these kind of programs helpful? Not just Y Combinator, but overall being in the Valley, these programs, how does it change
- 28:42
- Unidentified speaker (speaker_1): what
- 28:42
- Unidentified speaker (UNKNOWN): you're doing?
- 28:43
- Unidentified speaker (speaker_1): So the first question on I think whether you should come to the US or
- 28:46
- Unidentified speaker (UNKNOWN): not. I
- 28:47
- Unidentified speaker (speaker_1): think if you're building at the frontier of tech, let's say you're you're in India, you want to build something which benefits India,
- 28:56
- Unidentified speaker (UNKNOWN): and
- 28:56
- Unidentified speaker (speaker_1): you want to do that by building frontier technology.
- 28:59
- Unidentified speaker (UNKNOWN): And
- 29:00
- Unidentified speaker (speaker_1): people may not like this answer, but I still think that you should be at the epicenter of technology. You should come to the Bay Area. And one thought experiment you can do is that technologies that have been invented here, like Google, Meta,
- 29:15
- Unidentified speaker (speaker_1): OpenAI, have probably done more to advance even the life of an Indian back in India
- 29:21
- Unidentified speaker (UNKNOWN): than
- 29:22
- Unidentified speaker (speaker_1): any other Indian tech company out
- 29:24
- Unidentified speaker (UNKNOWN): there. And
- 29:25
- Unidentified speaker (speaker_1): there's just like an ecosystem of and network effects that happen here, which might make you more like if your goal is to build technology to further the life of an Indian, you might be able to do it, be more successful at it running out of India out of. out of the Bay Area. Because I think the companies in the Bay Area
- 29:42
- Unidentified speaker (UNKNOWN): are
- 29:43
- Unidentified speaker (speaker_1): building technology to advance humanity.
- 29:45
- Unidentified speaker (speaker_0): Yeah.
- 29:45
- Unidentified speaker (UNKNOWN): Whereas
- 29:46
- Unidentified speaker (speaker_1): I think many times in India the ambition becomes I'll do X for India and Y for India.
- 29:50
- Unidentified speaker (speaker_0): Yeah.
- 29:51
- Unidentified speaker (speaker_1): And they usually miss out on
- 29:53
- Unidentified speaker (UNKNOWN): like
- 29:53
- Unidentified speaker (speaker_1): what can I do beyond the current state of affairs.
- 29:57
- Unidentified speaker (speaker_0): Yeah. So the way I would summarize it, I do think many founders are now thinking beyond uh in India as well. But to your point, I'll connect it back to your earlier experience and what your parents told you, which is go do the job for an experience. Come to the US for the experience. Learn for two, three years, you can always go back. And you know the country is
- 30:17
- Unidentified speaker (speaker_0): also trying to get talent back. But super. Super time uh Rohan. Great, great chat. Any parting advice for people watching you and thinking about entrepreneurship and maybe entrepreneurship in AI?
- 30:29
- Unidentified speaker (UNKNOWN): Yeah,
- 30:30
- Unidentified speaker (speaker_1): say fool yourself into believing you can do anything.
- 30:33
- Unidentified speaker (speaker_0): On that note. Super, thank
- 30:35
- Unidentified speaker (UNKNOWN): you.
- 30:36
- Unidentified speaker (speaker_3): You're watching Z47 Moments, the podcast by India's oldest venture capital firm Z47. If you like this episode, consider hitting the subscribe button and the bell icon so you never miss an update from Z47Moments. Z47 moments.