Fireside Chat with Tanay Kothari at Wispr Flow

When we started Whisper, the singular goal that I had with the company was I wanted to build Jarvis. There's a number of companies coming out that are building AI agents for specific workflows, that are building AI agent platforms that can do anything on your behalf. The most important thing that I think most of these companies fail to understand is people don't want you to do 1,000 things. They want, at the very least, for you to say you can do one thing and then do that one thing really well.

Welcome everybody. My name is Sean. I'm one of the co-founders and managing partners at AX Ventures. We are an early-stage venture firm focused on pre-seed, seed, and A. Chris Manning is one of our general partners and co-founders too. And today we're joined by Tanay, who is the CEO and founder, co-founder of Whisperflow.

Thanks for having me here.

Yeah, so great to have you. Welcome to the office. Amazing. To be on this journey and learn from you and excited for the conversation today. Let's jump into it. When I saw that you guys were doing voice dictation, I immediately thought, that's been tried, it's awkward, it doesn't work. There's so many doing that. Why is Whisperflow different?

That is actually a really good question. I had the exact same thought when we stumbled into this problem. Because for people who don't know, WhisperFlow is a product, it's a voice interface that replaces your keyboard. You press a button, you speak, and text shows up. And like every other voice dictation product, it's a space that's been around for the last 20 years. But the thing that we realized was twofold. Number one, when you sit beside somebody and you see what they're doing day in, day out, about 5 to 6 hours of their time is spent typing on their phone, on their keyboard, whatnot. And when I saw a lot of AI companies being built out there that would book a flight for you, that would book an Uber for you, it's like, that is not the biggest problem to solve. That is not the biggest thing that people are doing today that you're starting to replace. What if we could replace that? And you see all the other voice dictation products, and what they do is they take everything that you're saying word for word and they put it out. That is not how it should work because you speak differently than you write. And more than that, it makes mistakes. If you look at any voice product today, you'll see this metric called word error rate, which is the percentage of words that they get wrong. Say it's 90%, 95%, which sounds fantastic, but it's actually terrible because what that means is in every sentence that has 10, 20 words, you're making a mistake. What people want is for it to just reliably get the right thing every single time. And so when we were building Whisper Flow, the key thing we were focusing on is this new metric that we call the zero error rate. What percentage of your messages are perfect? Apple, OpenAI, Deepgram, Assembly, all of these other companies, when you use it for these standard daily tasks, their zero edit rate is about 10%. 10% of your messages are perfect. With Whisperflow, that number is 85%. And when you have something that is that accurate, that has a latency of half a second, and it just works every single time, that is something when we started to see behavior change happen for the first time. And what we see with our user base is 70% of our users, after a few months of using Whisper Flow, use it more than their keyboard. And this is the first time in the history of computing that anything has replaced the keyboard that has been around for the last 200 years.

Okay, so you've in some sense started a bit into the answer to why now, but let me dive a bit more into that. Um, yeah, so I guess I'm the old guy here, So, you know, for me, I mean, there are antecedents even further back, but I kind of think of speech recognition as sort of having a 30-year history at this point because, you know, Windows 95, you could run Dragon Dictate on it or something like that. But I think everyone realizes that speech was terrible in those days. But the world became different, I think, you know, approx recently, 2014, '15. So let's say a decade ago, right? That was when modern neural network speech recognition appeared. And then suddenly there was, you know, what I would have said, speech recognition that worked. You could say stuff to your phone and it got it right. But, you know, that didn't actually lead to speech taking off for most people, you know, they have contexts in which it's useful. You're in your car or something like that. But, you know, most people cling to their keyboards, even if it's the phone keyboard. So why now in terms of why is this the moment where you can fundamentally get a different adoption curve?

That's a fantastic question. And there are two key reasons that we can see today And a third reason that I really strongly believe in is going to show up in the next couple of years. So the first reason is you can start interacting with your computers with natural language. That became a norm. ChatGPT, all of these other tools you can interact with, and that makes the mental model of people today feel more accommodating to like, hey, I can actually speak to this thing because it just feels right. Number 2 is what makes speech recognition good is the context. Even if you are hearing somebody without knowing what they're doing, what they're talking about, there's actually a lot of different ways that— a lot of different things that they could be saying that sound roughly the same even to us people. And so what really powered Whisper is our ability to grab a lot of context and process that really fast, which would not have happened with the GPUs and processing technology of even 5 years ago. Now we can dump all of that context into a large, large, large language model and then actually have it work. The last point that I wanted to get to is people today want to spend less and less time on their phones. If you asked anybody that, that is a big thing that they care about. And we are in the next couple of years going to see new immersive computing devices show up. Could be glasses, different wearables on your wrist, rings, otherwise. And the biggest thing that changes with those devices is they are not going to be display-first anymore. They're going to be voice-first. Display is going to be a secondary interface. And so the only thing that would make those devices be feasible is having a voice interface that you trust. And that is the bigger tailwind that we are driving towards is having Whisper Flow reach hundreds of millions of people around the world, build the habit of voice. So when they start adopting these new tools, Whisper is the default operating system that runs on that.

Okay, so I half believe that. I'm going to push deeper. Yes, let's do it. You know, the thing that you haven't mentioned, um, is that when people my age grew up, um, we spoke, um, where we are now in this world where we've got a whole generation of teenagers that are growing up texting, right? You know, there's been this enormous swing back to text and doing things on phone keyboards. I mean, do you really feel like that can be reversed now?

I actually have a caveat to that, which is that is the generation that is 40 and above that, or like 45 and above, that grew up without keyboards and they're used to speaking. They are also one of the biggest users of Siri today. Then you have the current younger population, which are like right now in middle school and high school and younger, they actually grew up with Alexas in their house. They are used to also speaking. And so when you look at voice adoption, those are the two groups that adopt voice the most. The group of people that has the least adoption of voice are people like me. So you're talking people in the age group of like 22 to like 35. Because when we were growing up, The first thing we had was a computer in our house. And our first memory of a phone is like when I was, when I was what, 9 years old, the first iPhone came out and I got my hands on it for the first time. So I grew up with keyboards more than anybody else. And so that is the age group, which is going to require the most amount of change. Well, the other ones are actually voice-first from birth. Okay.

So this is very helpful, Tanay Kothari. Another thing I've been thinking a lot about is your product strategy and then also the roadmap. Do you see WhisperFlow as the voice entry across many different applications and getting more and more mature there, or do you see yourself going deeper into the workflow and after you have that input, then you're going to the next steps of whatever the user wants to do for various vertical applications. How do you think about that product strategy? And then maybe most importantly, what's next on the roadmap? What are you excited about launching?

Fantastic question. So when we started Whisper, the singular goal that I had with the company was I wanted to build Jarvis. I want to build a personal assistant that understands you, that's with you 24/7, and is something that gets you when you trust day in, day out. And WhisperFlow to me is a way to get that. And so when I'm thinking about here's where we are today, here's where I want to get us, the most important thing when I think about when drafting a product roadmap or planning out for the company is what are the steps to get there? And so we think of Whisper as having 3 core pillars. Number 1, you speak and it writes for you. Number 2 is you speak and it does things for you. And number 3, that most people forget, is it is able to help you proactively because it's learned so much about you. And so when I think about a product roadmap, it is always to build things in these buckets, set them up, and really think about what are the very specific job to be done that we're doing in each of these applications. And have WhisperFlow be the best product there and scaling up from that point.

So, so let's talk about the next thing you're building or the next thing you're excited about launching. Is this a feature that is going to take voice as the input, transcribe, and then actually start doing things for you? And if so, how do you think about like the ChatGPTs, the Perplexities, the Claws of the world and, and what they're trying to do as well?

So The answer to that is yes, that is what it's gonna start doing. Now when I look at, there's a number of companies coming out that are building AI agents for specific workflows, that are building AI agent platforms that can do anything on your behalf. The most important thing that I think most of these companies fail to understand is people don't want you to do 1,000 things. They want at the very least, for you to say you can do one thing and then do that one thing really well. And, uh, like, I love OpenAI, but I've been playing around with ChatGPT's agents. It's amazing. It fails to be up to the mark and has the same issues as Siri, but it promises that it can do 1,000 things. It can maybe do 50 things and can maybe do 2 of those things really well. And that's where the problem comes up. So when I think about what, how we're thinking of building actions into WhisperFlow is same philosophy as before. Look at what are the things that people are doing extremely frequently on their computers and phones. And then we go take one of those workflows and integrate that into the product. So the voice dictation is the one thing that you do 100 times a day. Then we're gonna build the 10 things that you do 10 times a day. And then we'll open up a platform for you to be able to do the 100 things that you do once a day. And that's where it scales up to. So the next couple of things that we're gonna be building is you being able to say, hey, can you tell Sean I'm running 5 minutes late? And because you already use it in your iMessage and your Slack and your email, it's gonna be able to figure out who you're talking to. It knows your writing style, it knows your previous conversation, so it's gonna write that message in your tone and be able to do that for you. 'Cause that is something we do so often and doing that well is gonna be a game changer.

Is it gonna start responding to you to confirm that, you know, Tanay, I'm gonna send a message to Sean Johnson in your contacts that you're gonna be 5 minutes late. Are you gonna start having—

It'll confirm with you the part of that on the specific product design detail that you're going into. These are the design principles, which is one, we want to take away as much grunt work from people as possible. Two, there are certain emotions I want to avoid. Those emotions are anxiety, uncertainty, and lack of trust. So if any of those emotions come up in any product design of ours, we take a look at that, improve that to make sure that that doesn't come up. So anything from a confirmation screen to anything else that gives you peace of mind is what the eventual interface is going to look like.

We've already started going into the future, but I kind of feel like you haven't really covered the present of what makes WhisperFlow so special. So maybe we should actually kind of make sure we've covered that, right? So I mean, I think the thing that we really need to convey is, well, you know, I can just pull my phone out of my pocket and, you know, I'm in an app and I can press the little microphone that Apple gives me and I can do voice recognition with that. Um, but Whisper Flow is more than that. It's much more like having a human assistant in the old days. So I think it'd be really great for you kind of to give a more concrete idea of how there's this extra power that you're getting with Whisper Flow.

For sure.

And I think the, the best way I like people to experience it is just go to the website, w-i-s-p-r- whisperflow.ai, try out the product, give it 1 minute, and you're gonna start to feel why when Chris and Starn started using the product, we wanted to set up this talk. And so the magic that Whisperflow has really comes from the fact that it doesn't transcribe everything you're saying word for word. It actually understands what you're trying to communicate and then puts that down. So for example, if I say, hey Sean, let's meet at 5 o'clock. Actually, you know what, let's do 6 PM. Whisper will write, "Hey, Stronghold, let's meet at 6 PM." It knows my changes in my mind. It knows that that is not exactly what I wanted to say. And it does that hundreds of times a day in all the small things that you're doing. And at the heart of it, it lets you speak and ramble and it puts everything down in the most structured and concise way that makes you feel understood. And that is something that literally no other product on the market does at this point, because there is no product that is built for humans interacting with computers with your voice.

Right, right. So I think that gives a sense of what's really this distinctive different thing that is no longer a voice recognition app that is taking what the person has said. And as you said at the beginning about try to drive, the word error rate to zero of exactly what the person said, that it's really, I guess I call it sort of a translation engine, right? That the person says stuff and you're translating it into what they'd like to have appearing in written form for the email after they've changed their mind twice and decided on a different restaurant or whatever it is like that. And I think that's really the key. Distinction. But well, you know, at least for me, that's where I found it useful. I could blurb some stuff and it'll get turned into a coherent email that I can press send on. But something I totally hadn't realized was that people are using this to write code. The developers are using it. And that didn't sound the same to me. So tell us about developers and how and why they're using Whisper Flow.

Of course, this is actually one of the use cases that really surprised me when I first saw that happen is when Cursor, Windsurf, all of these tools were blowing up in Jan, Feb, we just saw a massive uptick in our user base and we were looking at it, it was all the people who were primarily using Whisper in Cursor and all of these tools. And turns out what I learned from diving deeper into it is two things. One, these systems do a fantastic job when you give them a lot of context and much larger prompts. Two, developers are lazy. I would wanna tell Cursor fix bug and it would not fix bug. But instead, if I'm using Whisper, I can give it the whole context. Hey, I tried this, this happened, that happened, this didn't work out. And then it's just one shots. Everything. So for developers who are adopting new AI tools, WhisperFlow is the one thing that lets them be even faster because they don't have to do that back and forth with AI anymore. It gets them at the first time. And so we started to see all the developers at every top AI company, all the new people who were starting to use developer tools just start to adopt WhisperFlow as one of their main daily drivers. And so most recently, we added a couple of things to make it even better. So for example, if you're an engineer and you're saying like, oh, deploy this on localhost:3000, it knows it's lowercase colon 3000, and it just writes what you expect it to. You can also put in any and all long commands for Docker Compose or whatever into what we have as snippets that just expand out. So you don't have to type those again and configure them in a dozen different places. And finally, if you're using it in Cursor or Windsurf, you can actually just tag files by speaking, which was one of the biggest requests we got from all the developers using it. And now building stuff is a purely hands-free experience 'cause you just speak and do this by pressing the function key on your keyboard and Whisper just does everything for you, which is incredible to see being how the average engineer today is building software.

This is where I'm fit. I'm really feeling like a coding dinosaur these days. So I think I have to start learning about some of these new developer tools and methodologies so I can be as efficient as the developers at the companies that you're mentioning. But let me connect that back to a little bit, I'll geek out a moment and connect it back to why now. I mean, so, so what part of what's different here is if we go under the hood for the moment that, well, when I was saying, oh, there was great neural net speech recognition starting about 2014, '15, that was just sort of, yeah, going from speech to words transduction, whereas for doing all the stuff you're talking about, I guess we have to say LLMs somewhere, right? That you're using much more natural language and context understanding to sort of be rewriting this speech. Do you want to say a little bit more about that?

Yeah, of course. I think at the heart of it, what we realized is there are incredible new tools that we can start using. And so the other thing we also realized is none of the models out there today are built for the kind of job we want them to do. So our ML team actually has a number of incredible PhDs. We do a lot of the model building and post-training in-house, and that is what the core Whisper experience runs on top of. And so my co-founder Sahaj is one of the people who invented diffusion models. And so he leads that entire effort there. And we have essentially new model updates come in every single week that we deploy constantly, adopting a number of different, honestly, like, what do you want to call them? There's a number of different problems that we see happen with existing models, whether they're with OpenAI, open source, you know, flawed, whatever. Everything around, you have hallucinations. Whoever has tried the OpenAI Whisper API, you know half the time it just says, "Thank you for watching," 'cause it was trained on YouTube data. You have very high latencies. Actually on ChatGPT, when you ask a question, it blinks the cursor for 3 seconds, then it starts producing the answer and it actually takes it a solid like 30 to 60 seconds to produce the entire blob of text you're gonna read. And for the kind of product we're building, you cannot expect hallucinations. We're replacing the keyboard, right? It has to be as reliable as the keyboard. You wanna have half a second latency. You wanna speak for 1 minute, 10 minutes, 20 minutes, and you still wanna have half a second of latency. And so that meant that everything from what is happening on the operating system layer to the networking stack to the models that we're running, is all built and managed and custom built for the specific experience to feel truly magical. And that's not even me going and talking about the fact that we support 100 different languages. And so there's a different combination of models that we need to use for all of the ones that we want to work with.

And so one of the special things that you need to handle is person speaking for a long time. You have to start streaming stuff. Rather than waiting right till the end before you do things. So to do that, that's sort of a whole custom architecture that you've built that is not what you get for a regular language model API provider.

No.

And so that's one of the many things that we had to build from scratch. And then if you use something like— and this is now getting more technical— like if you're using WebSockets, most GPUs don't support WebSockets. And so you have to figure something out. And then how are you managing the cache? How are you managing the queuing that is happening inside the model itself? So you're optimally using your GPUs. So those all end up being just extremely hard problems that we have to keep solving day in, day out.

Yeah.

I just wanna zoom out for a minute. There's so much happening with human-machine interfaces these days. You see this big thread around AI agents are going to be using the web more than humans pretty soon. You're pioneering the movement from fat thumbing keyboards to voice. Where do you think these interfaces are going to evolve in the next 3 to 5 years? Are we all going to be using voice and mostly the internet becomes, the web becomes just infrastructure to get our tasks done? Is the internet going away for for humans or like how do you think about the world we're gonna live in in 5 years?

So my answer to this is more nuanced 'cause the way I think about it, there's not a one single way everything is gonna evolve. I look at, okay, what are the different jobs to be done? Why do people go on the internet or on their phones in the first place? You are either getting entertained and consuming content. That is not gonna be voice-purged ever. You're gonna keep scrolling TikTok for the rest of your lives. Like that, it was just stuck in that hole right now and for better or for worse. Like viewers. And so that is one of the things that— okay, that's that. And so what that tells me is humans like getting entertained, so that is going to be a factor of displays in their lives forever. We are not going to ever be in a fully HURR-like situation where it's only an audio interface in and out. Now when we're talking about other things like communicating with others, for the longest time this was in-person as conversation, then we switched to phones and had phone conversations, then texting evolved as a thing. And question now is, I think we might be going back to just having more voice-based conversations because picking up your phone and writing a text to somebody feels honestly a little primitive. And so when you start to see people just voice dictate those messages and be able to just have that conversation back and forth, there's so much more emotion, so much more personality there, and people prefer that. So that's one thing that I think is gonna evolve. And then what I do is I look at each of these segments one by one and think about how that's gonna evolve, how shopping is evolving, how discovery and search of information is evolving. And then at Whisper, we try to design the best experience for what is it that the person really wants to do and what do we as humans just tend towards as behaviors and build for that.

Yeah, very interesting. Um, switching topics again, I'm curious about the founder journey and we've full transparency. We're early investors, uh, but I'd love to hear about like your, your narrative around how this journey has been since day zero. And maybe specifically, uh, people don't know that they started as a hardware company. Um, and, and here you are as, uh, the next-gen Jarvis. Tell us a little bit about this journey.

Yeah. So we actually go way back. Chris was actually one of the professors I really respected and looked up to in college. And so right after I graduated, when we were starting Whisper, he's one of the people I reached out to and I was like, Chris, would love to have you along as a journey. And he said, hey, we are starting this fund called AIX. If you can give me a couple of weeks, you can invest through the fund. And so I think AIX and Whisper pretty much started around the same time, plus minus a month. Yeah. And you all have been one of our first investors, so. Super excited to be doing this with you now. And so when we started, the mission was the same. We wanted to build Jarvis. We wanted to build this kind of personal assistant that feels like an extension of your mind. And the thing that came to our mind— and this is 2021. This is before ChatGPT has been released— was in the next few years, people are going to be talking to their computers. And when you do that, you want it to be private. You don't want to disturb others. And so we built this wearable device this little Bluetooth earpiece that could read your thoughts and convert that to text. And so no microphones on this thing. You silently speak, it converts that, and that essentially felt like a Jarvis in your mind. And so we worked on that for 3 years. We had a team of 40 incredible PhDs in neuroscience, machine learning, signal processing, electrical, mechanical. And we put the system together. Took us 3 years. Took us from August 2021 to August of 2024. And then when this was finally ready, we were like, okay, let's hook it up with ChatGPT. Let's hook it up with Siri, Alexa, and actually get it to do things. And turns out they all sucked for the same problem that they were just made a lot of mistakes and could not reliably do what you wanted them to do. So we started this little project inside the company called Flow which would go from your rambly thoughts and put something that is structured and ready to send. And what was the operating system then, now is Wiffle Flow, the company, our primary product. And what we saw with people was it just started to take off. And we realized that there was this massive need in the world for people to just step away from their keyboards because it was much more effortful and one of those things where It was cars compared to horses that we used before. And so at that point, about a year ago, we pivoted completely to, to this current software product. It was an insanely hard pivot because we went from 40 people to 5 people overnight. There's no more neuroscience. There was no more hardware. There's no more massive data collection operation going on. And for the first couple of months, we didn't know if this is going to work out. And the same question that we started the discussion out with is like, voice dictation doesn't sound sexy. Voice dictation doesn't sound like it is the frontier. But it turns out, if you build it right, it is. If you build it right, it is the tool that a lot of people describe as being the thing that changed their life. It is a tool that helps a lot of people, people with quadriplegia, with dyslexia, with a number of other conditions, be able to type again when they when keyboards was the biggest hindrance to their existence and their work, and is actually, I think, one of the most meaningful things that me and the rest of the team have worked on. And so at this point, there's nothing else I'd rather do.

Okay. Well, I certainly want to hear more about the founder journey and your thoughts for other founders, but I mean, somehow telling that part of the story is from making me think about, in some sense, this sort of this contradiction or change in your thinking in terms of product, right? Because the original idea was that people are going to want to, you know, write their emails or have thoughts that they want to preserve in context where it isn't possible to use speech. And therefore you could do this wonderful thing with your hardware product of subvocalization where people can sort of just whisper, move their mouth to what they were saying, and your wonderful hardware device would pick it up and turn it into text. So that was assuming that you'd be in environments where you couldn't speak, but you wanted to be producing text. Whereas as things have evolved, well, actually producing a speech product. I mean, so what's the resolution to that contradiction? Is it that it just turns out that everyone wears headphones or iPods or AirPods, so it doesn't matter? Or are there still a lot of contexts where speech isn't suitable? I mean, how did your product vision sort of change on that issue?

So what led us to make that pivot with a very clear mind was it boiled down to something really simple, right? We were here in a world where everybody's using keyboards and everybody's typing manually, and you wanted to get to a world where everybody's thinking and stuff happens. We were trying to make this jump directly, but that would require a lot of behavior change, and you need to first go from keyboard to not keyboard. You need to go from, uh, thinking your thoughts, uh, which— and just having it appear, which was again a brand new thing. And so you were asking a lot from users. And the key thing we realized with consumer products or any product in general is you can change one human behavior in time. You cannot and should not do more than that. And so in a world where everybody's tapping on the keyboard, a silent speech device is gonna flop 'cause the first habit change you wanna do from the keyboard is, hey, do everything that you were doing already but just use voice. Cool. Now that people have done that, now it's like, hey, you know this thing you love using every day? What if you could use it even when you're around others? And that is going to have much, much more adoption for a hardware device than we could before. But people, the response we got is like, but I don't use voice for anything. What am I going to do with this hardware device? And that was a very fair question that would have flopped. And so that was the key thing that we realized is actually this journey shouldn't be a big leap, but you need to have steps to get there. And you need to be very clear about what those steps are. And so to me, it wasn't a contradiction for For a couple of months, it felt like I had killed my baby. But at the end of the day, the mission has been the same. The problem we're solving has been the same. And we are now very directly solving a real problem that people face day to day, which is why also you start to see it in the level of product market fit, the user love. The company is one of the fastest growing AI companies right now is I think we really hit home. On something that nobody knew was a problem.

Yeah, that makes sense. Well, that's one very practical piece of advice for other founders and potential founders right there. But you've had an amazing journey. So I'm sure there must be more stories to tell. So you're starting off, you and Sahad, you were dormmates, I think, and that gave you, there's maybe some lessons there. But, you know, then later off, I mean, you mentioned the neuroscientists and the hardware people. I mean, in the early days of Whisper, I mean, I was really impressed by the way you were just sort of getting out there and hiring experts and, you know, really pushing this hardware vision. Now, obviously, that then pivoted to a new vision. But do you wanna share a bit more about, you know, some of what you've learned from being a founder as to what are the most important things and how you achieve them?

The biggest thing that I think doesn't get said enough is when you say you're a founder, you're looking for resources, right? You go to YC, you go to all of these places and it's like, you gotta find the problem, you gotta hit product market fit, you gotta find the market, all of that. But the fact of the matter is, as you build a company, your product's gonna change, Your market's gonna change. The one thing that doesn't change is the people. And when you are a founder, you're as much a leader as you are a manager. You are responsible for the lives of every single person who honestly works at your company. If you give them a work environment that is stressful, guess what? That is gonna, when they go back home, that is gonna come out on their kids and their family. And like, that's not what you want. If you give them a place where they feel a lot of self-efficacy, if they feel like they're doing a fantastic job, just imagine how their personal life would look like. And that is all driven by you personally. And so that responsibility, once you start to realize, makes you make a very different set of decisions where your, your teammates aren't just names in your HR system, but you take them a lot more seriously and you start thinking about the people problems. Because the main reason most startups die and slow down is because of the people problems that they have. And so that led us to have very specific philosophies in terms of the kind of people we hire, what our people philosophy is, and setting that up within the company.

Do you want to say a bit more about Yeah, that. Yes. I mean, let's be frank, quite a few tech founders aren't naturally people people. So there's probably some lessons to bring out there.

Yeah. And honestly, I wasn't either. I had the emotional empathy of a rock, but I did find some incredible mentors along the way who helped me learn those mental models. And turns out everybody's either head-brained, heart-brained, gut-brained. And so you need somebody to explain things to you in a way that makes sense. For me, it was just like very positive ROI to not be an asshole. And so some very specific things people-wise is in our team, we give people a lot of ownership. If you are an engineer, you and one other person are likely, it's a two-person little pod that is working on sometimes a feature, a product, or sometimes even an entire product line. Now on the person side, it is great 'cause you give people a lot more ownership, they feel more bought into the company, a lot of great things happen. From an organizational perspective, the impact of that is you have very little communication overhead. Communication overhead is my biggest pet peeve. It kills companies when you start to have 20 people in meetings and like no clear decision maker. At Whisper right now, it's a 20-person company. We're working on roughly about 25 different projects. And the way all of this works with all of these people is in every project meeting, there's just two people. They make the decisions, they do their thing, and everything runs in a very decentralized fashion. And now my desire to have lower communication overhead within the company that scales as n squared as your team grows also means that I hire for people who are extremely accountable, very responsible, and can do things by themselves a lot. Which tends to us hiring a lot of ex-founders as our team members. And so when you follow this down and you look at the Whisper team today, you're like, okay, like the goal of Whisper, like the way Whisper became this good is we hired a bunch of ex-founders. Like, no, this is the result of a lot of thinking that goes 3 steps back. 'Cause if you just had a bunch of ex-founders and extremely ambitious people and you don't have the right structure for them, it's gonna flop.

Yeah. I mean, that is interesting. I didn't actually realize that. I mean, that seems almost a little bit counterintuitive to me, 'cause I, I think a lot of founders are people who really wanna run their own show. Um, so do you have, uh, problems in sort of making it seem like a real team that works together from that, or?

not, has not happened yet. And that is again a second thing that has to come with it is the set of people that work together keep changing. And so you end up working with other people on the team all the time. And then we have a lot of just company culture that is very warm, very friendly. Like one thing is we have lunch together every day as a team, 12:30, which is like one thing I missed today, which I was a little bummed about, but I was here. Yeah, 12:30 every day our entire team is together having lunch, hanging out. That's the half an hour of the bonding that we have. And overall it's just a place that feels very, very warm and friendly. And so that helps us do a lot of this pod-like structure and make things work that way. And for me personally, I inherently am a person who likes to have a lot of control. And everything, like a lot of founders do. And so there was some dissonance in my head in the early days of like, okay, if somebody's fully managing a project, then I am not able to get into it. But the mentality changed when I started viewing it as, it's a lot more leverage. If I want to do 20 things and I can tell 10 people to do those 10 things, or like if I can only do 10 things at a time, then I can do 20 now 'cause there's 10 people who can do these other 10 things on my behalf. And so I and Whisper as a company can achieve a lot more. And so that is the other way that is pretty much non-refutable to set up. And I think is something that for founders who are also control freaks, whatever, it works. It works and creates a healthy work environment.

Very nice. So we're gonna open it up now, see if there are any questions from the audience, also from anyone online who's joining us. Let's see if any questions are rolling in. I think Lucy had the question.

I have a question. How are you expanding the capabilities of Whisper?

Sorry, expanding the capabilities?

Of Whisper technology to other languages.

Of Whisper technology to other languages?

Yeah, so really good question, and, uh, the answer is it's not as easy as we had initially thought. There's a lot of nuances that come with every language. So what we are doing is basically we're getting, uh, one person from each language and country that we want to improve Whisper in, one by one, uh, to work very closely with the ML team to improve the model in that specific language. We have a list, like priority list of languages where people are already using Flow, and the moment it starts to become really good in that one language, we launch Whisper in that country. And so the biggest thing came from the realization that it is extremely hard and there are so many other nuances of languages that you don't understand if you're not speaking or writing it.

Do you see a future where we'll have more walled gardens? So if my car has its own voice assistant, assistant, my phone has one, my Alexa has one, or will it do one that kind of travels through our lives where, you know, I get in my car and it has all the contacts from the morning and, you know, I get to my office and—

So the initial approach that you said of every single thing having its own voice assistant is what we're starting to see right now. And the fact is every company wants to be the one assistant that's with you all the time. And so I suspect what we're going to see over the next few years is a good amount of consolidation that happens. Because at the end of the day, an assistant is intelligent not because it's smart, but because it knows things about you. And my hunch is in a few years, we're going to be in a place where there are maybe one or two dominant voice players that power a lot of things. In your life and it would be like a Windows, Apple kind of situation that happens because you can expect a few companies to succeed because the market for this is 8 billion people.

And will that be primarily hardware providers that build or will it be software?

So the other thing with this is what I've noticed while both building a hardware product and a software product and kind of diving deep into companies that are doing this, is companies naturally have a place where they excel. Microsoft makes great software. They have historically never made really good hardware. They've just started to after kind of 40 years of building it. Facebook, apart from this Oculus team and the partnership with Ray-Ban, all their other hardware products have just not worked well because the culture that lets them make extremely addictive and successful software products, which is move fast, break things doesn't work in hardware. In hardware, you need to lock your designs 18 months before you want to ship it. And from a lot of friends who work at Facebook, that is not how that team generally likes to, uh, function. And so I, I, I do think you will see some companies that are extremely good at building that software agentic and AI layer of it, and you'll have them partner with, with hardware companies. Um, and so the The Meta Ray-Ban kind of example I think is one of the peak ones that I look at, and we're gonna see a lot more things that look like that. I'm gonna kind of put you on the spot for a hypothetical here. So we're talking about voice interfaces for all sorts of use cases. The ones you're mostly describing are kind of software-oriented actions. On your behalf on the web or email or something along those lines. Where do you think robotics goes and how do I execute that and just that part?

The question is, how do voice interfaces transfer to robotics through the controlling robotics? Is that you talk to your robot?

So when I look at that specific problem and what as a person you want to get done, 90% of it is very similar to software. You want a question asked or a job done. The big thing that comes with robotics is how do you reduce the anxiety and uncertainty, right? On the software layer, it's easy. You show a popup and be like, hey, did you want to message this Sean or this Sean? And easy. In a robotics perspective, that question becomes harder And I don't know what the answer to that is, but that is the biggest thing that is gonna be different is how do you prevent that anxiety and uncertainty feeling that comes up by giving voice commands? 'Cause the thing I realized when I was looking at a lot of the datasets that these robotics models are trained on, especially when it comes to voice is the, so these are datasets where you have a prompt and then you have a series of action that the robot takes. And the funny thing is you look at these prompts and it is not something a human being would say. It has things like, hey, can you take this blue shirt from here and then put it on top of this white shirt or whatever? It's extremely detailed, very precise. And a human's gonna say like, hey, can you organize my shirts? 'Cause you don't wanna give a lot of context. That's not how you speak. And I think that is where a lot of the anxiety and uncertainty comes in. That is right now, what I feel isn't modeled in to that. Similarly for computer use and everything else, I've started to see this problem everywhere 'cause we start to have this dataset of how we see people using Flow. And I'm just like, this is so different from how these datasets are built from what people actually want.

All right, well with that, I wanna thank Tanay for joining us. Thank you, Chris. Thanks everybody.

This was fantastic. Thank you, everyone.