Podcast Episode Summary
The podcast episode features a discussion with three AI company leaders: Gaurav Misra (Co-Founder and CEO of Captions), Carlos Reyna (Chief Revenue Officer of Eleven Labs), and Laura Berkhauser (VP of Product at Descript). The main topic is how these companies have managed to stand out in the competitive AI market, particularly in terms of product design, marketing, and customer retention.
Key Points:
Takeaways:
Participants:
Additional Insights:
Less than two years since the breakthrough of text-based AI, we now see incredible developments in multimodal AI models and their impact on millions of users.
As part of New York Tech Week, we brought together a live audience and three leaders from standout companies delivering AI-driven products to millions. Gaurav Misra, Cofounder and CEO of Captions, Carles Reina, Chief Revenue Officer of ElevenLabs, and Laura Burkhauser, VP of Product at Descript discuss the challenges and opportunities of designing AI-driven products, solving real customer problems, and effective marketing.
From the critical need for preventing AI misuse to ensuring international accessibility, they cover essential insights for the future of AI technology.
Resources:
Find Laura on Twitter: https://x.com/burkenstocks
Find Carles on Twitter :https://twitter.com/carles_reina
Find Gaurav of Twitter: https://twitter.com/gmharhar
Stay Updated:
Let us know what you think: https://ratethispodcast.com/a16z
Find a16z on Twitter: https://twitter.com/a16z
Find a16z on LinkedIn: https://www.linkedin.com/company/a16z
Subscribe on your favorite podcast app: https://a16z.simplecast.com/
Follow our host: https://twitter.com/stephsmithio
Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.
Stay Updated:
Find a16z on X
Find a16z on LinkedIn
Listen to the a16z Podcast on Spotify
Listen to the a16z Podcast on Apple Podcasts
Follow our host: https://twitter.com/eriktorenberg
Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.
Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.
Host
We've existed for about three years and we've passed everybody in revenue in like literally a year and a half. Usage is important, but that does not define the long-term success of an actual customer. I think that daily active use is a pretty terrible metric to uncover customer value. There have been companies built in the past on just great design. There's no reason that they can't be built on the AI side. You end up creating all of these multiple layers that essentially end up building your core defensibility in the market. Retention problems are just activation problems in disguise.
The episode begins with the host setting the scene.
Between June 3rd and June 9th, A16Z ran its second annual New York Tech Week. This week had thousands of people attend a record-breaking 700 plus events, including one event run by our podcast team. This A16Z live recording is exactly what you're about to hear. But first, let's take a quick trip to memory lane.
When ChatGPT was launched in November 2022, it quickly became the fastest growing consumer application in history. But text-based AI was just the beginning. In the next 500 days, a flurry of AI models launched that spanned new modalities, from images to video to audio to 3D, that all yielded an entire ecosystem of applications that have upended, quite frankly, the way we work, learn, create, and even play.
Now here in mid-2024, competition is fierce, but I don't think I have to convince you of that. So for this live recording, we brought in key leaders at three AI companies to discuss how they've managed to stand out amongst the noise, because they have products that reach millions of users.
In this conversation, you'll hear from Gaurav Misra, co-founder and CEO of Captions; Carlos Reyna, chief revenue officer of Eleven Labs; and Laura Berkhauser, VP of product at Descript. Together, we explore what ladders up to AI products that people actually use, including what features really matter, when AI is necessary or distracting, whether you need to own your models, designing for retention and international expansion, and of course, where we all go from here.
I hope you enjoyed this recording as much as I did. As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16Z.com slash disclosures.
The live panel discussion begins.
We're actually less than two years since that, but a lot of people are familiar with text to text. But all three of the products here go into several other modalities, right? We've got audio, we've got video imagery. So I think that's really exciting. But maybe we could actually just start with the why now, and specifically maybe the unlock that we've seen with unstructured data, right? Before we used databases and everything needed to be really structured in order for us to make sense of it. Today, that's not quite the case. So Gaurav, maybe we start with you. And what do you see really today as the why now?
Gaurav Misra
Yeah, I mean, I think it's a really exciting time generally, just because there have been a couple of key breakthroughs in terms of technology with transformers and diffusion models and so on. But I think the key here is we're able to use a lot more data to train these models now than ever before. There are a bunch of things happening, both on the hardware side, the software side, and the data side to enable that to happen. That's why we're seeing some amazing results. If you look at a lot of what the key players in this industry are doing, they're just training these models with more and more data every iteration. That's able to produce reliably better and better results, which is pretty amazing to see. And it's not in sight so far.
Host
Carlos, maybe we'll go to you before we talk about Descript in a second.
Carles Reina
I think it's correct. The key message for our experimentation at 11 Labs has been: if you put garbage in, garbage out. If the quality of the data that you put in is not that great, then what you end up producing is half-baked with lots of mistakes. We can see that with Whisper. How many of you have tried Whisper and it comes out with "subscribe, subscribe, subscribe" and things like that? All the time. It's true. We've seen it all the time.
So for us, there's been a layer. Initially, we trained it with a lot of data. Then over time, we ended up curating the data to make sure it is very high quality. Otherwise, you're not able to achieve the results that you are expecting or that your consumers or your businesses would need. But that's a fundamental change that has happened in the market: the amount of data being used with transformers and LLMs to generate human content, whether that's speech or text or anything else. We're seeing all types of stuff, 3D models.
Host
The reason I wanted to wait to talk to you, Laura, is because I don't know how many of you have used Descript. Any guesses on when Descript started? We talked about ChatGPT, November 2022. Descript has been around since 2017. The reason I wanted to frame that is because obviously the last couple of years have been very exciting, but machine learning and AI have been going since the 50s. There have been unlocks, but I want to get your pulse, Laura, on the importance of putting AI at the forefront. A lot of AI is embedded in the applications that probably people in the room are building as well, but Descript long used machine learning before really saying, "Hey, you're using machine learning, AI, et cetera." So what are your thoughts?
Laura Berkhauser
That's right. Descript is software that lets you edit video just like a text document. So if you can edit a Google document, congratulations, you're also a video editor. You can just download Descript and now you can edit video. It turns out that the technology that undergirds that is in fact AI. But we haven't traditionally come forward and said, "We're an AI video editor." A, there wasn't a huge reward in the hype cycle for saying that, so we didn't have marketers saying it. But also what we found is that customers didn't care. They don't care what the technology is that is creating this value for them. What they care about is that there is value here. This is helpful for me.
That was long our way of designing software, and it probably would have continued that way forever, except that when I think about the thing that is making us change our minds, in addition to some of these cool models that are coming out, it is that the way that humans and computers are interacting is totally different. You can talk to your computer now. You can use human language to communicate more subtle intentionalities that you have for how you want to edit your video or create your video. So as this technology has gotten better, we've thought, "Do we actually want to design AI in the product differently? And if so, how?"
With our latest release, we're bringing all of the AI features that we've long had in the product into the same space and adding a ton of new ones. We had a big discussion with our design team about how to do this. One of the big discussions we had is: is AI a magic wand or is it an entity? One of the big decisions you have to make there is that traditional creators are much more used to interacting with Pro Tools software or creative software in a point-and-click way. So they want a magic wand. But you have this whole new wave of people that are now generating and editing video and audio, and they're used to using more of this entity interaction. They want an entity.
Then you start talking about an entity, and you get into internal discussions like, "I don't know if it's an entity. That might be a bad idea because what about our robot overlords, our inevitable robot overlords?" That's one side of the debate. Then hilariously, you have the other side of the debate: "I don't want an entity because it turns out this technology is really stupid sometimes. If you make it an entity and say, 'Welcome, this is your co-editor,' and it turns out your co-editor is a total moron that makes horrible suggestions sometimes because it's hallucinating." So we're like, "Okay, how do we deal with that?"
What we decided to do with this newest release is we're calling it Underlord. It's a nod to the potentially apocalyptic future of AI while also admitting that right now this thing is a very eager, somewhat competent intern that does a really great job at the first pass of the worst parts of your workflow. So that's some of the story about how we've thought about designing with AI over the years.
Host
I'd love to get both of your takes. How do you think about that same question: what part of AI do I put at the forefront? Or do I just use this really powerful technology and give my users what they want, but not really sell this AI thing too much?
Gaurav Misra
At the end of the day, you have to solve customer problems. That's what we're trying to do. I think the biggest mistake that can be made is to say, "Here's a technology. You can have technology, do whatever you want." People can't just take that and be like, "Okay, I know what to do with this." You have to mold it into a product that solves a problem at the end of the day. Nothing's changed there. It's exactly the same as before. If you're not doing that, you're going to see retention problems. You're going to see people coming in, trying out the thing, not knowing exactly what to do with it, not working perfectly for their use case, and then they'll leave. Tourism is what we're calling it.
But on the marketing side, stepping away from product for a second, there is something to be said about having AI in your message. Here's why: if I just say, "I have a better product, it's so much better, you won't believe it," I'll be saying the same thing that people have been saying for literally 100 years about every product. "Trust me, it's better. Trust me, come on and try it out." This is every single product that exists. But putting that AI term in there, just from the marketing side, this is just tactical, actually lets people understand, "Oh wait, this is going to be a step change." Of course, if you don't meet that expectation when they land on the product, you're going to have a problem. But if you're able to meet that expectation, putting that in does inform people that this is not going to be the better product. It's going to be a step change compared to everything else we've seen.
I do feel like a lot of people are just throwing the AI term in on the marketing side now just to get the eyeballs there. Maybe that message will get lost a little bit. But so far, the innovation has been so strong that the message has remained strong. If it continues this way, the marketing side can continue as well. But at some point, it might get muddled. We'll see.
Host
Maybe I can add on a modifier for you. I think not only do you have to market the product, but if you use this bucket term of AI, that means many different things. Do you own your models? Do you build your own models? Are you an API wrapper? I'd love to hear from you, Carlos, at 11 Labs in particular, in building your own models as well. How does that play into the marketing packaging, thinking about what you share and what you don't?
Carles Reina
We need to be open. We are an AI company. Sorry, guys. And we say it all the time. We say we do AI voices, we do AI sound effects, we're going to be doing AI music in many ways. For us, it's all about the audio sphere. It's that layer infrastructure that allows you to create high-quality engaging content, whether that is with voice or with audio overall.
The way we thought about it: there wasn't really a good quality text-to-speech available before we invented our own site. So we were fundraising initially, it was difficult because the market was not there. "How are you going to be getting customers?" and so on. It was really tough in the early days. But we thought, if you're able to deliver quality, voices that sound engaging, the applications on top of it, then you end up having a market that is just fully on top. So how do you do that? AI voice is simple and plain. And that worked really well.
We started with the LLM pure API play with a very simple UI. That was end of January last year when we launched the product. We thought there would be some content creators that might want to use the UI, but we expected the API side to be quite big because people might want to build their own application on top of it. It worked really well. Since then, we also realized that you cannot expect all of the business to have the capabilities to build their own applications. So what if we go full end-to-end and build our own applications for areas where we really care about? That's how we ended up creating projects like Audio Native or the dubbing product and a bunch of other pieces.
It's been very interesting for us. We always say that it's AI driven because at the end of the day, we're a foundational model that happens to also build applications on top of it. But the beauty of it is that anyone can build anything they fancy on top of the API. Today we power quite a lot of different companies. More than 41% of Fortune 500 companies use 11 Labs. We power a lot of startups, and we're very proud to help all of these companies succeed as well. It's been very interesting having both sides, both motions: the pure API play and the application layer on top of it. It's challenging as well because then you end up having two different profiles on the product side, on the engineering side. You always need to balance it.
Host
Absolutely. Maybe we can actually jump straight to that question of competition. I feel like if there's one question that comes up on this podcast the most, everyone's excited about AI and they're like, "Okay, well, where does differentiation come up? Where do moats arise?" I'd love to probe all three of you on that. I know we're early, but where do you think you can stand out? Do you really need to be building at the model layer? You talked about the infrastructure layer. Or can you really just build a really great UI and capture the app layer? What do you think about that?
Gaurav Misra
I'll start here by saying, again, not much has changed. There have been companies built in the past on just great design. So I think there's no reason that they can't be built on the AI side. But at this point of the journey, there's so much to innovate on, so much to build on. It does help to have models that are foundational and built in-house because it does give you that extra differentiation and that extra step. It is a competitive field. The deeper you can go and the more you can build from the ground up, connecting these different layers together, you can deliver super fast speeds on your models. You can deliver the highest quality that anyone's seen. And you can deliver a great user experience that solves a real problem. Then you have an advantage there.
For consumer companies, which we are, we're used by literally millions and millions of people around the world. People make over 100,000 videos a day published through our platform. For a consumer company, it does matter a lot to have that differentiation at this stage. In the longest term, if you think about what differentiates a consumer company, it's probably just brand. That's what you're building over a period of time. The only way a brand dies is with a generation. It also takes a generation to build a brand too. So I think that's the ultimate goal of where you want to get to. But in the meantime, there are many moats that last different lengths of time, whether that's a data moat or a model moat or a UI/UX moat, whatever it might be.
Laura Berkhauser
At Descript, I would say that we are a horizontal editor and we're a very powerful human editor, which is something that a lot of newer companies that just started in the age of AI, in the second chapter of AI, can't say. It takes a long time to build a really powerful, horizontal, human-driven editor. You can do really complex editing jobs with Descript if you already are an expert who's great at this work. And you can do it really quickly with low barriers to entry if you're new to it.
For that reason, I think the application layer is especially important to us. I almost see us as a mirror to what 11 Labs was saying. In general, we have a "may the best model win" sort of mentality when it comes to all of the different models that we use in our application layer. That's because we're trying to do everything, not just AI voices, but things like eye contact, things like avatars, things like AI speech, transcription, editing video with text. If there's a cool thing happening in AI, when video generation comes out, when Sora comes out, that will be in Descript. We're going to have it.
Generally, we have an attitude that is "may the best model win." We want to give our customers the absolute best experience. If we don't see interesting enough work happening in a space that we want to be in, we'll build that model. I think there are real places for Descript to differentiate because we own so much of the editing workflow and have really great editing workflow data. That may be a place where our models become differentiated. But in general, if you're trying to provide a ton of different services to customers across a ton of different workflows, it can really make sense to not try to build every single one of those in-house, but instead to be very thoughtful about where it makes sense to own versus buy or borrow.
Carles Reina
There's an element here on differentiation in these days. Because the market has evolved a lot from purely foundational picks and shovels, and now there's a transition towards the app side. How I think about defensibility is your users, your consumers, or your businesses. That's what will drive defensibility over the long term. If you think about Instagram or Meta or Facebook in the early days, what was their defensibility? There was literally nothing out there, but they were able to fast grow, outpace everyone in terms of growth, deliver value. The UI was not even that great, but you felt that it was part of the community and it was the experience that you were getting. The defensibility was coming from the actual users versus the product itself.
The transition that we're seeing today from the foundational models toward the app side is very interesting because you're able to engage different types of generations or different types of users. If you retain them and give them the best experience possible, they will stay there for the coming years, whether that is because they're building their own applications on top of it or because they want to use your app overall.
The way we also think about this at 11 Labs is like layers. Having the foundational layer, which is the research that we provide. We do LLMs and we provide the best text-to-speech and AI voices in the market. Fantastic. What else do you have on top of it? The data that we've acquired, that we've licensed from partners, the end-to-end products that we're building, the partnerships that we have, the customers that we have. You end up creating all of these multiple layers that essentially build your core defensibility in the market. That will hopefully sustain us for the coming years. As the market changes, if one of the layers ends up getting replaced, absolutely fine, because then you have all of the other ones that will back you up in the long term.
Host
Something you spoke to here is this new generation. I think we're all trying to figure out what can now be done with AI. You talked about UX even, or designing a new UI. Voice is now in the mix in ways that it wasn't before. But then you also have this question of, do I want to completely reinvent the wheel, show someone a very powerful UI that they're maybe just not familiar with and that you don't retain them? So, Gaurav, I'd love to probe you on retention. Even from the perspective of desktop versus mobile, you do have a mobile app. How do you think about designing for that? Because we've seen over and over the last two years, there's this extreme willingness to try, but then I think someone internally coined this "AI tourist phenomena." People try and then a lot of them do leave. So how do you think about that?
Gaurav Misra
It's something we think about a lot. At the end of the day, you can go by metrics and you can really worry about, "Oh, there's a retention number, it should be at that number." You can get caught up in that a little too much when the reality is those micro optimizations are not going to solve whatever retention problem or any other metric problem that you might have. At the end of the day, it's about the user experience. It's about solving a real problem.
Generally, if you want a complete hit end-to-end, you need to have a breakthrough technology that's applied to solve a very specific problem that a user actually has. Then you need to have an engine that can deliver that solution to people who have that problem as quickly as possible across the world. If you have all those pieces, then you won't have a retention problem or an acquisition problem or any other problem.
The cool thing about this time right now is the technologies are being developed and there's actually a crazy number of technologies out there. I think it's a very unique time from that perspective. For product people, the main problem is, "How do we actually solve problems? Actually solve real problems that people have. And not just sell the technology as technology, like 'Hey, we have technology.' Just that. But actually convert it into a real value delivery for users for a specific use case, even a niche case, whatever it might be."
For marketers, the problem is, "How do we actually educate people that there's a new way to solve these problems?" People may not think the first thing, "Oh, you know what? I'm going to Google AI for this." That might not be the first thing that people think about. They might be searching for just whatever they were normally doing, which may be something that takes a long time. Or they might not be aware that there are new solutions available to these problems. I think if you focus on that at that level, all the other numbers sort of follow on their own. That's what we've seen, both across our desktop app and our mobile apps as well. We're in the consumer space, so retention is definitely a very hard game to crack compared to B2B businesses. But we've been able to do it really well, and I think it's because of that high-level focus across technology, product, and marketing.
Host
Laura, you used to work at Twitter. What are you learning in terms of products that reach so many people? We're talking daily active users. What have you learned from that space that you can apply to AI when you are trying to fix this retention problem?
Laura Berkhauser
I will say that I am so glad to be out of the game of trying to optimize for MDAU, for monetized daily active users. I think that daily active use is a pretty terrible metric to uncover customer value. One of the things that I just love most about working at Descript is being able to identify alternative metrics to think about: have I done right by the customer? Two that I really like to think about, which are a bit in tension with each other and act as guardrails, are time to expression and editing richness.
If Descript is doing its job really well, the amount of time it takes you from starting a project to getting it into a shareable state, whether you're a marketer trying to repurpose a webinar into clips, or a creator trying to make your latest YouTube review, or someone in learning development trying to create a training, I want the amount of time it takes you to create that to go down and down. So you're able to just create more and more content. Is anyone here a creator in any way? Have a YouTube channel or a marketer? Do you know about the gaping maw that can never be fully fed or sated for content that I find so many of our customers are just staring into with despair? Getting their time to expression down is really important.
But one of the ways you do that is by creating worse and worse content that is just a roll with an iPhone and you slap some captions on it. That's great for some use cases, but for others, it's a missed opportunity. You could have done so much more to create really high quality video content. So if Descript is also winning on increasing the editing richness—the number of jobs you're able to do with us and the number of things you're able to do to transform your media and make it really high quality—the interaction of those two metrics is such a great way to drive towards customer value.
I will say that what Gaurav said around good product fundamentals with retention totally resonates with me. My attitude toward the tourists is you've got to triage the tourists. Some component of them just don't have a use case for your software. They want to create a voice clone. They want to see it. They're like, "Ooh, that looks cool." But they don't have anything to do with that voice clone. Great. Let them do that. That's awesome. Maybe one day you'll think about Descript or 11 Labs and come back.
But then who are these tourists who actually have a legitimate use case and they just don't know it yet? They could be using video to communicate within their company. They could be using text-based video editing to create all of their marketing clips. And they don't know that yet. How can I create software that activates really well, that displays all of our use cases and lets them have a good first time? I find that often retention problems are just activation problems in disguise in a trench coat. So what I really try to focus on to improve retention is just the activation experience.
Host
Just having come from a social media background as well at Snap, such a good point about DAU and how that can be such a trap. Social media companies obviously optimized for DAU for a reason because money is coming from a different source. So it's good to be out of that game. And really interestingly, with the generative AI space, it seems like it's having the opposite effect on what it's trying to achieve. Social media on one end is using AI as well, but really to consume time from people as much as possible. Consume as much of your time and it's succeeding. On the other end, generative AI is actually giving back time to people, so they can actually do more. Pretty cool.
We talked about this on a recent episode, how some tools, I'm sure people would resonate with this: if you had one excellent session, it could have saved you four hours of work in five minutes. That's actually more valuable than spending 20 minutes every day in an app. And you don't see that in the same metrics, right? So I love that you brought up different metrics that you're paying attention to, Laura. Carlos, is there anything that jumps to mind there for you in terms of how you might rethink a business model in terms of what metrics you're paying attention to, the way that you're monetizing a product that might be different because the willingness to pay we've also seen is there, even if it is just, I'm using this once a month, once every two months?
Carles Reina
I think it's a really good point. Some consumers actually feel that if they need to do something twice, the product is not working well.
It's that element that we've come from one side to the other side. So probably someone in the middle is what fits well. I was actually in a meeting with a customer, and we presented to a C-level last week. The question they came back with was, "Okay, so how much time am I going to save?" I said, "You're going to save anywhere between 50 to 60 times the time. It's going to be slashed by 50 to 60." They said, "No, that's not possible." I said, "Let's do the math right now." And we did the math, and it was very interesting.
I think there is an emphasis on that side, but sometimes we try to overemphasize the effects of the efficiency you're getting with generative AI, when in fact, generative AI is not perfect. I think that's one of the main reasons why the AI tourists are there and they're very big. Everyone comes with such a big expectation that it's going to solve all of my problems and cook dinner for me tonight as well. Unfortunately, it's not going to cook dinner for you. It's just not going to solve all of your problems. But it's going to help you quite a lot, either because you can do a lot of modern monetization with your customers, you can reach new markets, or you can actually do it much quicker.
I think framing it on what is actually valuable for you as a business or as an individual is much more important. Initially, our metrics were purely about usage. Over the past month, we've ended up switching. Usage is important, but that does not define the long-term success of an actual customer for us. Our innovation side is about what is the use case that you have, how do we measure that over the long term, and how do we understand and try to ensure the use case based on the way you're using the product, so that we can offer you the best tools and the best tips.
For us, those are the key metrics today versus usage. Usage is still super important, but I don't really mind if someone uses the product today and then doesn't do it for a week or two weeks, because I know that if we've nailed it, they're going to come back two weeks later. That's how we are thinking about it. You don't have those social notifications that are like, "A friend of a friend maybe posted something, please come to our app."
Host
All right, well, we're going to open up to questions very soon. So if you have any questions, start thinking about them. But I want to do rapid fire, one or two more. So, the importance of optimizing an application for a specific role or someone's use case: who are you? What are you trying to do? Each of you actually comes from different backgrounds. Gaurav, you've done design and development, you've been an engineer. Laura, you've been immersed in product. Carlos, operations. Those are roles where there's a gosh, I don't know how many other people who fit that subset.
I'd just love to hear your perspective, independent of your company. How do you think of AI in the next five years? What does an AI-powered engineer look like in your case, Gaurav, or an AI-powered operations person? What do you need? What's missing? Are there products out there that actually fit that use case and are doing it well?
Gaurav Misra
Yeah, I mean, thinking about it from an engineering perspective or even from a design perspective, I think maybe the closest on the engineering side would be like a tech lead manager. Someone who's actually setting up the overall architecture of whatever is being built. But a lot of the work's been done by AI, and they're coming in, they're making edits. They're like, "Maybe we need to change this," reviewing stuff. Same on design: giving high-level instructions like, "Let's have this. Let's maybe use this style over here. Let's change these components." And getting that output back and reviewing it, leaving comments the same way that a manager might. And being able to produce, hopefully, a lot more value and output. That means that companies can be going to a much larger revenue scale with way fewer people, which is going to be interesting.
Laura Berkhauser
Yeah, I think a lot about this. What is the AI product manager? The paradigm that I use is more like, how do I want to interact with AI to do my job better? One of the use cases I'm excited about is a rubber duck who talked back. You guys hear about rubber ducking, where you keep a rubber duck on your desk and you talk through difficult problems with that rubber duck. I think I'm never going to cede control of the creativity and the genius to the entity. Clearly, have you met me? I'm in charge of that. But I think it can be fun to toss the ball around with someone. I'm excited to see how AI continues to develop to be a fun thing to toss the ball around with, and then can take all of the stuff that you're just spewing out, all of the kind of word garbage, and turn it into something crisp and readable and easy to understand. So that's a use case that I'm excited about.
Carles Reina
I think from an operations side, it's even more complex, because there's so many things that you need to do. How do you automate or how do you get someone to help you on that front? Ideally, you end up having a product that helps you do twice as much in the same amount of time. Not from an efficiency perspective, but much more how I can potentially generate more revenue for the business. I think that's where, potentially and hopefully, the market is going to be going. On the sales side, it's much easier because you end up having AI SDRs these days, we'll end up having AI CSMs and all those pieces that can already be there in many cases. But purely on the operations side, there's a lot more complexity. ChatGPT is your friend for sure, or Tropic if you use it, or any of those tools that will help you generate quite a lot of different things on a day-to-day basis. Is that giving you a 2x? Not yet. I still haven't found the right product that would help anyone optimize and become 2x themselves. Maybe someone will build it in the room.
Host
I guess, final question, just anyone feel free to jump in. All three of your products have a lot of customers. People are using it. Seems like maybe you've solved the retention problem. What challenges are you facing? Whether it's regulation, or not having the right models, or hoping that the open source models catch up. Just curious if anything jumps out where you'd like to call out a challenge that you'd like to be solved in the next few years.
Gaurav Misra
Yeah, I'd say for us, it's hiring, actually. It's very traditional, right? But hiring the right people to solve the particular problems that we're having in our company. Problems grow really quickly if the company's going really quickly, and you have to kind of keep an eye on all the different things that are happening, where new needs might come up. Especially with a company like ours where we've existed for about three years, and there are video companies that have been around for a long time. We've passed everybody in revenue in literally a year and a half. With growth at that scale, you just have to constantly be thinking about what are the new problems that are coming up and who can we hire to solve those problems. That's a very traditional answer, and maybe there's some AI recruiters out there, but we have a great team, so I don't think we need them. At least not yet.
Host
Maybe I can help with that.
Laura Berkhauser
I think it's just that we're in the middle of a paradigm shift. We haven't gotten to the end of it. We're in the middle now. What I can tell you is that the way that we're going to edit video and audio in a year or in two years is going to look completely different than how we're doing it right now. But we don't know how yet. On one hand, that's why I'm here. That's why I'm doing this job, because this is a place where the next generation of product managers and designers, we're going to reinvent the way that humans and computers interact with each other. Someone's going to figure it out, and God, I hope it's me or that I'm part of it in some small way. But that's also just a very fragile moment. It's both a challenge and an opportunity. I think it's the challenge of our industry right now.
Carles Reina
I think for us, there's two sides of it. One is definitely hiring. I can relate a lot on that. It's difficult. We've gone from zero to tens and tens of millions in months, not even years, in months. It's really difficult to find people that have experienced that previously, because the market has evolved very quickly in such a time frame. So there's a lot of commitment that we expect from people at the company, and we need to be able to actually keep growing at this pace. On the research side, it's extremely difficult to find the right researchers on the engine side, on the operation side, on sales, even support across the board. That's one side of the equation.
The other side of the equation is preventing misuse. Realistically, that is something that we have an entire team dedicated to, day and night. But every time that we put together something, there are all the different things that people make up to try to game it. It is similar to fraud, where you're always just steps behind. It's really difficult to catch and keep fighting it. Those two elements are the biggest challenges that we're constantly facing as a company. We're winning, but still, it's just a matter of making sure that you're constantly innovating and having resources for something that is important. Otherwise, regulators come, or consumers complain, and people complain.
Host
You need unprecedented people for an unprecedented pace. For questions, Laura, who's our wonderful producer at the a16z podcast, is going to go around. So if anyone does have a question, just raise your hand and she'll come. Thank you.
Audience Member
I'm curious how we are thinking about internationalization or serving users of various levels of digital literacy.
Gaurav Misra
We had an international audience from the beginning, including every country and every region you could possibly imagine. So I think it's been a high priority from the beginning. The interesting thing is that a lot of the development that AI is bringing is not just things that are usable in an English thing or a US thing. It actually brings change in workflows across almost every country and every culture you can imagine. And it actually works. We've gone and launched new markets where we've had zero users and overnight had an explosion of users in that market. But then we learned something about that particular market where, oh, they don't like this particular thing. If you think about, for example, the Middle East, text is written in the opposite way. That changes a lot about the UI and changes a lot about the user experience. We've done a lot of work to make that good and make that as usable and as amazing of an experience as it is in any other language. So those are the types of efforts we've made high priority from the beginning.
Host
Would you say that other countries or regions are actually more readily adopting the products? Because I'm thinking, maybe they can't hire the software engineer or maybe they can't pay for the traditional video editor of those thousands of dollars. So they're actually more readily adopting these technologies because they're bringing the cost down.
Gaurav Misra
Absolutely. I think around the world, people are super open to trying something new to see if they can change their workflow. As long as you can provide something that, once you try it, you can't go back to what you were doing before. That's the difference. If you can provide that experience in any language, any culture, any country, people will use the product.
Carles Reina
For us, internationalization has been since day one. We have a full international team. Everyone is fully remote. There's a very strong correlation, funny enough, between the actual employee profile and the fact that we have multiple countries. Everyone can be based wherever they want and travel. And the actual user time that we've gathered. In the initial days, a lot of our growth came from North America and European markets. But actually, these days, when you look at the entire pie, it's super spread out across the world.
I can relate to that purely on the fact that people want the best tools that will help them on a day-to-day basis. You don't really need to spend thousands of dollars or hundreds of dollars these days to actually produce a video, to produce a podcast, or produce something. You can do it much cheaper using tools. That's the beauty of it. By default, anyone that truly wants to have a cost-efficient solution will end up using any of the tools: Script, Captions, or Labs, or anything else that you have out there. By default, you end up having a strategy that is about the international market. We're dealing with content, trying to engage your audiences where they are, and trying to personalize it to them anyway. Otherwise, you end up having a problem of being very skewed towards one market. Traditionally, it's been, "Oh, you go one market, you conquer it, and then you expand to another one." These days, it's just not. It's evolved quite a lot in our product.
Host
Yep. There's time for maybe one, maybe two more. I see one at the back.
Audience Member
I'm just wondering what barriers or stopgaps you might be putting in place for people who may be using your products for nefarious purposes, thinking about trust and safety.
Carles Reina
From ElevenLabs, we invest millions every single year on actually preventing misuse. We were the first company to implement a fingerprinting system for any content that gets generated. Since we launched, the fingerprinting has been in place. We then opened up the API and the UI to make sure that anyone could check whether something was generated by us or not. Since then, we've also engaged in monitoring the content that our users generate. So if someone is generating things that they shouldn't, then we block them. We've gone as far as to build the no-go voices, which is a model that will prevent anyone that tries to clone a celebrity voice. We're constantly adding all of these layers to try to make sure that we stay ahead of the curve. But as I was saying earlier, it's an uphill battle overall. There are always ways in which you can game it. At the same time, you have open source tools. We can try to do our side of the equation, but anything that is open source, to some extent, you don't really have that much control over those tools. But I think it's important as a company, we will keep investing millions every single year, and we can increase it as the market grows as well.
Host
I have to just quickly ask because it's very timely, and I'm sure people in the audience are wondering, with some of the recent news around AI voices—let's just leave it at that—and celebrities. Are you finding there to be a bunch of false positives? Because I feel like that's maybe something that people wonder. You hear a celebrity's voice, but how unique can a voice be? If you're trying to filter out certain people's voices, are you finding that actually our voices maybe aren't that unique?
Carles Reina
That's a really good question. Voices are not as unique as everyone thinks, but however, they're quite unique. You end up having false positives for sure. But if it's a false positive, if it tells you, "Oh, you don't have permission for this voice," automatically it tells you, "Oh, but you can still pass the voice structure," and it will show you the voice capture. So if you pass it because it is your voice, then you're able to actually use your own voice. I have a twin brother, for the ones that don't know. We do sound exactly the same. Even my parents sometimes make mistakes. Truly, I could be talking, but you could be thinking it's my twin brother. We have exactly the same voice. That is a challenge we have as a company and a society we have. But if you end up building layers as a product from a product perspective to help filter those false positives, I think people understand that you're trying to go from everything being free-for-all where you can misuse it as much as you wanted, to let's put some controls. Even if there's some false positives, people understand it down the line.
Laura Berkhauser
There's something about the product side of this too, which I do think is super important: to build the safety features from the product from the ground up. That's the difference between offering a technology versus offering a product. If you just say, "Hey, come to our website, make deep fakes," that's offering a technology. Some people might be out there doing that, I don't know. But if you build that into a product, like for example, we have the language translation feature, which can translate whatever you're speaking to a different language and changes your lip movements as well. Yes, that's using the same technology, but in a very opinionated way that you can't change what was said, but you can change what language it was said in. That limits the scope of abuse immediately quite a bit. And then all the traditional methods can be used on top of that as well.
Gaurav Misra
I agree. With Descript, you can create a voice clone of yourself and sort of intermingle. We have this thing called Overdub where if I say the wrong word, I can go back and, with the text, say the word that I actually meant to say. Then it will, with my voice clone, kind of create that. Obviously, there are a lot of misuses there. So whenever we launch a product, we launch it with protections in place, do a bunch of testing, hire outside people to try to crack it, and try to make sure that we do our very best to make sure that it's ungameable. But like you said, if people are extremely determined to crack through security, they will always find new ways to do it. This was the case when I was in social media, too, where you do all kinds of things to try to protect your platform. Bad actors, they get up every morning and grind just as hard as you do. You're just sort of in the eternal struggle. I think every single tech product should be thinking about how people are going to misuse us and making sure that they're responsibly providing a bunch of resources to stay in the fight.
Host
So as VP of Revenue at ElevenLabs, how do you view the role of open source? Because as a developer myself, I would rather use, for example, Falcon 70B, which is $1 in and $1 out per million tokens, as opposed to GPT-4, which is $30 and $50 out. Do you think that open source is a threat to your business, especially as companies like Meta are kind of taking a scorched earth approach to releasing models?
Carles Reina
I think it's complementary, actually. You always end up having businesses or people that can go and use open source, and they have the means and the tools and the knowledge to make that work. And then you end up having quite a lot of different people that don't really have those means or knowledge. It just ends up becoming different sides of the business or different sides of the market, however you want to segment it.
When I think about voices, we've been talking to each other as humans for the past 50,000 years. There wasn't really a good technology that was able to replicate how we talk as humans. The fact that as a platform, or even open source, you're able to actually replicate people's voices with their permission, make it sound natural, engaging, and then power a new type of communication and platform and experience—that market is massive. By default, you need to have both sides to be able to actually counterbalance each other and push each other.
But open source also comes at a cost, which is that the number of features you will have is more limited. You will end up having less voices. So what's your preference? You don't have the UI. So what's your preference as a business or as an individual? Is it purely building on top of it? Then maybe open source is a good way. Today, the quality is not there yet, but I'm sure that within the next three years, the quality is going to be matching anything that is private. So it's going to be more about the actual ecosystem that you build around it to make sure that people start using it in a much easier way and then embed it. But I actually think it's complementary. Without one, we cannot have the other one, purely because the market needs both sides.
Host
Just to follow up, would you say that it's important for picks and shovels companies, closed source, to build an application layer on top to stay competitive?
Carles Reina
100%. I don't think anyone that has actually built a full LLM, if they're not able to build applications on top of it to make life easier for consumers and businesses, you will end up struggling down the line. Whether that is in six months' time or in 18 months' time, you will struggle. Because at the end of the day, I want to launch my own application, or my product, or use the product like this immediately. If I need to spend the next month coding and building the UIs and everything, I might give up and go somewhere else, even if it's more expensive, especially if I don't even know whether I have product-market fit.
Product-market fit, we always think about startups, but big corporates might not even have product-market fit. If you want to iterate quickly and then go to market as quickly as possible, you might want to have a stack that is truly readily available for you. But once you're ready and you've tested it and the technology feels good enough with other LLMs or open source, then you might end up looking to switch. We've seen that with OpenAI, the big migration from developers that started using OpenAI, ChatGPT APIs, and GPT 3.5, and then now they migrate to Anthropic, Mistral, or Llama. That's been happening for the past six months.
It will continue happening. You start to validate that everything goes well, and then you figure out whether there are alternatives, whether it's renegotiating pricing or open source.
Host
If you liked this episode, if you made it this far, help us grow the show. Share with a friend, or if you're feeling really ambitious, you can leave us a review at ratethispodcast.com slash a16z. You know, candidly, producing a podcast can sometimes feel like you're just talking into a void. So if you did like this episode, if you liked any of our episodes, please let us know. I'll see you next time.