Live at Tech Week: Delivering AI Products to Millions

0:00:01 (upbeat music)
0:00:03 – We’ve existed for about three years
0:00:04 and we’ve passed everybody in revenue
0:00:06 in like literally a year and a half.
0:00:09 Usage is important, but that does not define
0:00:12 the long-term success of an actual customer.
0:00:14 – I think that daily active use
0:00:19 is a pretty terrible metric to uncover customer value.
0:00:21 – There have been companies built in the past
0:00:22 on just great design.
0:00:25 There’s no reason that they can’t be built on the AI side.
0:00:27 – In upgrading all of these multiple layers,
0:00:29 they’ll essentially end up building your core
0:00:31 defensibility in the market.
0:00:34 – Retention problems are just activation problems
0:00:35 in disguise.
0:00:37 – Between June 3rd and June 9th,
0:00:41 A16Z ran its second annual New York Tech Week.
0:00:43 Now this week had thousands of people attend
0:00:46 a record-breaking 700 plus events,
0:00:49 including one event run by our podcast team.
0:00:51 Now this A16Z live recording
0:00:53 is exactly what you’re about to hear,
0:00:57 but first let’s take a quick trip to memory lane.
0:00:59 When ChatGPT was launched in November, 2022,
0:01:03 it quickly became the fastest growing consumer application
0:01:07 in history, but TechSpace AI was just the beginning.
0:01:08 In the next 500 days,
0:01:12 a flurry of AI models launched that spanned new modalities,
0:01:15 from images to video to audio to 3D,
0:01:18 that all yielded an entire ecosystem of applications
0:01:20 that have upended, quite frankly,
0:01:23 the way we work, learn, create, and even play.
0:01:27 Now here in mid-2024, competition is fierce,
0:01:29 but I don’t think I have to convince you of that.
0:01:30 So for this live recording,
0:01:33 we brought in key leaders at three AI companies
0:01:35 to discuss how they’ve managed to stand out
0:01:36 amongst the noise,
0:01:40 because they have products that reach millions of users.
0:01:41 So in this conversation,
0:01:43 you’ll hear from Gora Misra,
0:01:45 co-founder and CEO of Captions,
0:01:46 Karla Sarena,
0:01:48 chief revenue officer of 11 Labs,
0:01:52 and Laura Birkhauser, VP of product at Descript.
0:01:55 Together, we explore what ladders up to AI products
0:01:57 that people actually use,
0:01:59 including what features really matter
0:02:03 when AI is necessary or distracting,
0:02:04 whether you need to own your models,
0:02:07 designing for retention in international expansion,
0:02:10 and of course, where we all go from here.
0:02:13 I hope you enjoy this recording as much as I did.
0:02:17 As a reminder, the content here
0:02:19 is for informational purposes only,
0:02:21 should not be taken as legal, business, tax,
0:02:22 or investment advice,
0:02:25 or be used to evaluate any investment or security,
0:02:26 and is not directed at any investors
0:02:29 or potential investors in any A16Z fund.
0:02:31 Please note that A16Z and its affiliates
0:02:33 may also maintain investments
0:02:35 in the companies discussed in this podcast.
0:02:36 For more details,
0:02:37 including a link to our investments,
0:02:41 please see a16z.com/disclosures.
0:02:48 And so we’re actually less than two years since that,
0:02:51 but a lot of people are familiar with text-to-text,
0:02:53 but all three of the products here
0:02:55 go into several other modalities, right?
0:02:57 We’ve got audio, we’ve got video, imagery.
0:02:59 So I think that’s really exciting,
0:03:02 but maybe we could actually just start with the why now,
0:03:05 and specifically maybe the unlock that we’ve seen
0:03:07 with unstructured data, right,
0:03:09 before we use databases and everything needed
0:03:10 to be really structured
0:03:12 in order for us to make sense of it.
0:03:13 Today, that’s not quite the case.
0:03:14 So Gaurav, maybe we start with you,
0:03:17 and what do you see really today as the why now?
0:03:19 – Yeah, I mean, I think it’s a really exciting time,
0:03:21 generally, just because obviously there’s been
0:03:22 a couple of key breakthroughs,
0:03:25 just in terms of technology with transformers
0:03:28 and diffusion models and so on and so forth.
0:03:31 But I think the key here is we’re able to use a lot more data
0:03:33 to train these models now than ever before, right?
0:03:35 And there’s a bunch of things happening,
0:03:38 both on the hardware side, the software side, right?
0:03:40 And the data side to enable that to happen.
0:03:42 And that’s why we’re seeing amazing results, right?
0:03:44 If you look at a lot of what the key players
0:03:46 in this industry are doing,
0:03:47 they’re just training these models
0:03:50 with more and more and more data every iteration, right?
0:03:52 And that’s able to produce reliably better
0:03:53 and better and better results,
0:03:55 which is pretty amazing to see.
0:03:57 And it’s not in sight so far.
0:03:58 – Carlos, maybe we’ll go to you
0:04:00 before we talk about description in a second.
0:04:00 – I think it’s correct.
0:04:02 The key message for us is experimentation
0:04:04 for 11 lamps has been like,
0:04:05 if you put garbage in, garbage out, right?
0:04:07 If the quality of the data that you put in
0:04:09 is not that great,
0:04:10 then essentially what you end up producing
0:04:13 is half-baked with lots of mistakes and things like that, right?
0:04:15 And we can see that with Whisper,
0:04:16 how many of you have tried Whisper
0:04:17 and it comes out that like,
0:04:19 subscribe, subscribe, subscribe,
0:04:20 and things like that, right?
0:04:21 All the time.
0:04:24 That’s true, we’ve seen it all the time.
0:04:25 But so I think like for us,
0:04:27 like there’s been a layer and initially we trained it
0:04:29 with a lot of data and then over time,
0:04:30 we ended up curating the data
0:04:33 to make sure that like it is very high quality.
0:04:34 Otherwise you’re not able to achieve the results
0:04:36 that you are expecting or that your consumers
0:04:38 or your businesses would need, right?
0:04:39 But that’s a fundamental change
0:04:41 that has happened in the market.
0:04:43 Amounts of data being used with transformers
0:04:46 and alarms to generate this like human content generated,
0:04:49 like whether that’s speech or text or anything else, right?
0:04:51 – Yeah, 3D models, we’re seeing all types of stuff.
0:04:54 So the reason I wanted to wait to talk to you, Laura,
0:04:57 is because I don’t know how many of you have used Descript,
0:05:00 but any guesses on when Descript started?
0:05:03 We talked about Chat GPT, November 2022.
0:05:06 So Descript has been around since 2017.
0:05:07 The reason I wanted to frame that
0:05:09 is because obviously the last couple of years,
0:05:12 very exciting, but machine learning, AI,
0:05:13 in the ’50s is when this really got going.
0:05:15 And obviously there have been unlocks,
0:05:17 but I want to get your pulse, Laura,
0:05:20 on the importance of putting AI at the forefront.
0:05:22 A lot of AI is embedded in the applications
0:05:25 that probably people in the room are building as well,
0:05:27 but Descript long-used machine learning
0:05:28 before really saying, “Hey,
0:05:30 you’re using machine learning, AI,” et cetera.
0:05:32 So what are your thoughts?
0:05:33 – That’s right.
0:05:36 So Descript is software that lets you edit video
0:05:37 just like a text document.
0:05:41 So if you can edit a Google document, congratulations.
0:05:43 You’re also a video editor.
0:05:44 If you can just download Descript,
0:05:46 and now you can edit video.
0:05:47 And it turns out that the technology
0:05:50 that sort of undergirds that is in fact AI,
0:05:52 but we haven’t traditionally come forward
0:05:55 and said, “We’re an AI video editor.”
0:05:57 A, there wasn’t like this huge reward
0:05:58 in the hype cycle for saying that.
0:06:00 So we didn’t have marketers saying it,
0:06:03 but also what we found is that customers didn’t care, right?
0:06:05 They don’t care what is the technology
0:06:07 that is creating this value for me.
0:06:09 What they care about is there is value here.
0:06:10 This is helpful for me.
0:06:14 And so that was long hour way of designing software,
0:06:16 and it probably would have continued that way forever,
0:06:19 except that actually when I think about the thing
0:06:20 that is making us change our minds,
0:06:23 in addition to some of these cool models that are coming out,
0:06:26 it is that the way that humans and computers
0:06:27 are interacting is totally different.
0:06:28 So you can talk to your computer now.
0:06:31 You can use human language to communicate
0:06:33 more subtle intentionalities that you have
0:06:36 for how you wanna edit your video or create your video.
0:06:40 So as this technology has gotten better, we thought,
0:06:42 well, gosh, do we actually wanna design AI
0:06:43 and the product differently?
0:06:45 And if so, how?
0:06:46 And so with our latest release,
0:06:48 we’re actually bringing all of the AI features
0:06:51 that we’ve long had in the product into the same space
0:06:53 and adding a ton of new ones.
0:06:55 And we had a big discussion with our design team
0:06:56 about how do we do this?
0:06:57 And one of the big discussions we had is,
0:07:01 is AI a magic bond or is it an entity?
0:07:04 And one of the big decisions you have to make there
0:07:06 is that traditional creators are much more used
0:07:08 to interacting with Pro Tools software
0:07:11 or creative software in a point and click way.
0:07:12 And so they want a magic wand.
0:07:15 But you have this whole new wave of people
0:07:18 that are now generating and editing video and audio
0:07:20 and they’re used to using kind of more
0:07:22 of this entity interaction.
0:07:24 They want an entity.
0:07:26 Then you start talking about an entity, right?
0:07:28 And you get into internal discussions like,
0:07:31 I don’t know if it’s an entity that might be a bad idea
0:07:33 because what about our robot overlords
0:07:35 are inevitable robot overlords, right?
0:07:37 That’s kind of like one side of the debate.
0:07:39 And then hilariously, you have the other side
0:07:41 of the debate that I don’t want an entity
0:07:43 because actually it turns out this technology
0:07:44 is really stupid sometimes.
0:07:47 And if you make it an entity, you said like,
0:07:49 hey, welcome, this is like your co-editor.
0:07:51 And it turns out your co-editor is like a total moron
0:07:53 that makes horrible suggestions sometimes
0:07:54 because it’s hallucinating.
0:07:57 And so we’re like, okay, how do we deal with that?
0:07:59 So what we decided to do with this newest release
0:08:02 is we’re actually, we’re calling it underlord.
0:08:06 And it’s a nod to the potentially apocalyptic future of AI.
0:08:07 Well, also admitting that right now,
0:08:10 this thing is kind of like a very eager,
0:08:12 like somewhat competent intern
0:08:15 that does a really great job at the first pass
0:08:17 of the worst parts of your workflow.
0:08:18 So that’s some of the story
0:08:20 about how we’ve thought about designing with AI over the years.
0:08:22 – I’d love to get both of your posts.
0:08:24 Like, how do you think about that same question?
0:08:27 What part of AI do I put at the forefront?
0:08:29 Or do I just use this really powerful technology
0:08:31 and kind of give my users what they want
0:08:34 but not really sell this AI thing too much?
0:08:36 – So I’d say, at the end of the day,
0:08:37 you have to solve customer problems.
0:08:39 That’s what we’re trying to do, right?
0:08:41 I think the biggest mistake that can be made is to say,
0:08:43 hey, here’s the technology.
0:08:45 You can have technology, do whatever you want.
0:08:47 People can’t just take that and be like, okay,
0:08:49 I know what to do with this, right?
0:08:50 I think you have to mold it into a product
0:08:52 that solves a problem at the end of the day.
0:08:53 So I think that’s like traditional.
0:08:54 Nothing’s changed there, right?
0:08:56 It’s exactly the same as before.
0:08:57 And if you’re not doing that,
0:08:59 then essentially you’re gonna see retention problems.
0:09:01 Where you’re gonna see people coming in,
0:09:02 trying out the thing,
0:09:03 not knowing exactly what to do with it,
0:09:06 not working perfectly for their use case.
0:09:07 And then they’ll leave, right?
0:09:09 Kind of tourism is what we’re calling it, right?
0:09:11 But I think at the same time on the marketing side,
0:09:13 like stepping away from product for a second,
0:09:17 there is something to be said about sort of having AI
0:09:19 in your message on the marketing side.
0:09:20 Here’s why.
0:09:23 If I just say, I have a better product,
0:09:25 it’s so much better, you won’t believe it.
0:09:26 I’ll be saying the same thing
0:09:30 that people have been saying for literally 100 years
0:09:31 about every product, right?
0:09:33 Like, yeah, trust me, it’s better, right?
0:09:34 Trust me, come on and try it out.
0:09:37 This is every single product that exists, right?
0:09:40 But putting in that AI term in there,
0:09:42 just from the marketing side, this is just tactical,
0:09:44 actually lets people understand,
0:09:47 oh wait, this is gonna be a step change, right?
0:09:48 Of course, if you don’t meet that expectation
0:09:51 when they land in the product, you’re gonna have a problem.
0:09:53 But if you’re able to meet that expectation,
0:09:55 putting that in kind of does inform people about like,
0:09:58 okay, this is not gonna be sort of like the better product,
0:10:00 it’s gonna be a step change
0:10:01 compared to everything else we’ve seen.
0:10:03 So that’s the general guide.
0:10:04 I do feel like a lot of people
0:10:06 are just throwing in the AI term in the marketing side now
0:10:08 just to kind of get the eyeballs there.
0:10:11 And maybe that message will kind of get lost a little bit.
0:10:14 But so far, the innovation has just been so strong
0:10:16 that the message has kind of remained strong.
0:10:17 And if it continues this way,
0:10:19 the marketing side can continue as well.
0:10:23 But at some point, it might get muddled, we’ll see.
0:10:25 – Maybe just I can add on a modifier for you
0:10:28 because I think not only do you have to market the product,
0:10:30 but if you use this bucket term of AI, right?
0:10:31 That means many different things.
0:10:32 Do you own your models, build your own models?
0:10:34 Are you an API wrapper?
0:10:35 And so I would love to hear from you, Carlos,
0:10:37 at 11 Labs in particular,
0:10:40 like in building your own models as well,
0:10:42 like how does that play into it?
0:10:43 Is it a whole marketing packaging
0:10:45 thinking about what you share and what you don’t?
0:10:46 – Yeah, we need to be open.
0:10:48 Like we are an AI company, sorry guys.
0:10:49 And we say it all the time, like we say,
0:10:52 like we do AI voices, we do AI sound effects,
0:10:54 we’re gonna be doing AI music in many ways.
0:10:57 So for us, like it’s all about the audio sphere, right?
0:10:59 So it’s like that layer infrastructure
0:11:01 that allows you to create high quality engaging content,
0:11:05 whether that is like with voice, with like audio overall.
0:11:06 And the way we thought is, well, actually,
0:11:10 there wasn’t really a good quality text-to-speech available
0:11:12 before we invented our own site.
0:11:14 So we were fundraising initially.
0:11:16 It was difficult because the market is not there,
0:11:18 like how are you gonna be getting customers and so on.
0:11:21 So it was like, it was really tough in the early days.
0:11:23 But we thought, look, if you’re able to deliver quality
0:11:25 that voices that sound engaging,
0:11:27 the applications on top of it,
0:11:30 then you end up having market that is just fully on top, right?
0:11:33 So how do you do that? AI voices, simple and plain, right?
0:11:34 And that worked really well.
0:11:37 So we started with like the LLM pure like API play
0:11:40 with a very simple UI that was end of January last year
0:11:42 when we launched the product.
0:11:43 And we thought, well, actually,
0:11:45 there’s gonna be like some pieces of like some content creators
0:11:47 that might want to use the UI,
0:11:50 but we expect on the API side, it’s gonna be quite big purely
0:11:53 because like people might want to build their own applications
0:11:53 on top of it.
0:11:55 And it worked really well.
0:11:57 And since then, what we also realized, like,
0:11:59 well, you cannot expect all of the business
0:12:01 to have the capabilities, build their own application.
0:12:04 So what if we end up going full end to end
0:12:05 and we build our own applications
0:12:08 for areas where we really care about?
0:12:10 And that’s how we end up creating like projects
0:12:13 or audio native or like the dubbing product
0:12:14 and a bunch of other pieces, right?
0:12:16 So it’s been very interesting for us.
0:12:19 And of course, we always say that it’s AI driven
0:12:21 because at the end of the day, we’re a foundational model
0:12:24 that happens to also build applications on top of it.
0:12:26 But I think like the beauty of it
0:12:30 is that anyone can build anything they fancy on top of the API.
0:12:32 And today we power quite a lot of different companies,
0:12:35 more than 41% of Fortune 500 companies use 11 Labs.
0:12:37 We power a lot of startups
0:12:39 and we are very proud to help all of these companies
0:12:40 like succeed as well, right?
0:12:42 So it’s been very interesting,
0:12:44 like having both sides, both motions,
0:12:46 like the pure API play and the application layer
0:12:48 on top of it, it’s challenging as well.
0:12:50 Because then you and I’m having two different profiles
0:12:52 in terms of like on the product side,
0:12:53 on the engineering side and everything, right?
0:12:55 So you always need to balance it.
0:12:57 – Absolutely, maybe we can actually jump straight
0:12:58 to that question of competition.
0:13:00 I feel like if there’s one question
0:13:02 that comes up on this podcast the most,
0:13:03 everyone’s excited about AI and they’re like,
0:13:06 okay, well, where does differentiation come up?
0:13:08 Where do moats arise?
0:13:10 I’d love to prove all three of you on that.
0:13:11 I know we’re early,
0:13:12 but where do you think you can stand out?
0:13:15 Do you really need to be building at the model layer?
0:13:17 You talked about the infrastructure layer,
0:13:19 or can you really just build a really great UI
0:13:20 and capture the app layer?
0:13:22 What do you think about that?
0:13:24 – Maybe I’ll start here by saying,
0:13:26 again, not much has changed in terms of like,
0:13:28 there have been companies built in the past
0:13:29 on just great design.
0:13:31 So I think there’s no reason
0:13:33 that they can’t be built on the AI side.
0:13:36 But at this point of the journey,
0:13:39 there’s so much to innovate on and so much to build on.
0:13:41 It does help to have models
0:13:43 that are foundational and built in-house
0:13:46 because it does give you that extra differentiation
0:13:46 and that extra step.
0:13:50 It is a competitive field and the deeper you can go
0:13:54 and the more you can build from the ground up really,
0:13:56 connecting these different layers together, right?
0:13:59 You can deliver super fast fees on your models.
0:14:02 You can deliver the highest quality that anyone’s seen, right?
0:14:04 And you can deliver a great user experience
0:14:06 that solves a real problem.
0:14:07 Then you have an advantage there.
0:14:10 So I would say though for consumer companies,
0:14:11 which we’re a consumer company, right?
0:14:14 Like we’re used by literally millions and millions
0:14:15 of people around the world
0:14:17 and people make over a hundred thousand videos a day
0:14:19 published through our platform.
0:14:21 For a consumer company,
0:14:24 it does matter a lot to have that differentiation
0:14:25 at this stage.
0:14:26 I think in the longest term,
0:14:29 if you think about what differentiates a consumer company
0:14:32 in the longest of terms, it’s probably just brand, right?
0:14:35 And that’s kind of what you’re building over a period of time.
0:14:38 And the only way a brand dies is like with a generation.
0:14:40 It also takes a generation to build a brand too, right?
0:14:43 So I think that’s kind of the ultimate goal
0:14:44 of where you want to get to.
0:14:45 But I think in the meantime,
0:14:48 there’s many modes that last like different lengths of time,
0:14:51 whether that’s the data mode or a model or like,
0:14:54 whether it’s a UI, UX mode, whatever it might be.
0:14:57 – So at Descript, I would say that we are a horizontal editor
0:14:59 and we’re a very powerful human editor,
0:15:02 which is something that I think a lot of kind of newer
0:15:04 just started in the age of AI,
0:15:06 in the second chapter of AI companies can’t say
0:15:07 because it takes a long time
0:15:09 to build a really powerful,
0:15:11 horizontal human driven editor.
0:15:14 So you can do like really complex editing jobs with Descript.
0:15:17 If you already are like an expert who’s great at this work
0:15:21 and you can do it really quickly with low barriers to entry.
0:15:22 If you’re new to it.
0:15:24 So that reason, I think the application layer
0:15:25 is especially important to us.
0:15:27 And I almost see it as a mirror
0:15:29 to kind of what 11 Labs was saying,
0:15:31 where I think like in general,
0:15:35 we have a may the best model win sort of mentality
0:15:37 when it comes to all of the different models
0:15:39 that we use in our application layer.
0:15:41 And that’s because we’re trying to do everything,
0:15:44 not just AI voices, but things like eye contact,
0:15:48 things like avatars, things like AI speech, transcription,
0:15:50 editing video with text.
0:15:53 If there’s like a cool thing happening in AI
0:15:55 when video generation, when Thora comes out,
0:15:57 that will be in Descript, we’re gonna have it.
0:16:00 And so I think like generally we have an attitude
0:16:02 that is may the best model win,
0:16:04 we wanna give our customers the absolute best experience.
0:16:08 If we don’t see interesting enough work happening
0:16:11 in a space that we wanna be in, we’ll build that model.
0:16:14 And I think there are real places for Descript to differentiate
0:16:16 because we own so much of the editing workflow
0:16:19 and have really great editing workflow data
0:16:20 that like that may be a place
0:16:23 where our models become differentiated.
0:16:25 But in general, if you’re trying to provide
0:16:27 a ton of different services to customers
0:16:29 across a ton of different workflows,
0:16:32 it can really make sense to not try to build
0:16:34 every single one of those in-house,
0:16:36 but instead to be like very thoughtful
0:16:39 about where it makes sense to own versus buy or borrow.
0:16:42 – I think like there’s an element here on,
0:16:45 if you think about purely about differentiation in these days,
0:16:47 like ’cause the market has bought a lot
0:16:49 from purely foundational picks and shovels.
0:16:52 And now the transition towards the app side,
0:16:53 what you end up thinking about
0:16:56 or how I think about defensibility is fear about your users,
0:16:58 your consumers or your businesses.
0:17:00 That’s essentially what will drive defensibility
0:17:01 over the long term.
0:17:03 And if you think about Instagram or Meta
0:17:05 or like a Facebook in the early days,
0:17:06 what was their defensibility?
0:17:08 There was literally nothing out there,
0:17:10 but they were able to fast grow,
0:17:12 outpace everyone in terms of growth, deliver value.
0:17:15 And then the UI was not even that great, right?
0:17:17 But it was actually like you were feeling
0:17:18 there was part of the community
0:17:20 and it was like the experience that you were getting, right?
0:17:22 So defensibility was coming from the actual users
0:17:24 versus the product itself.
0:17:26 And I think like the transition that we’ve seen today
0:17:30 from the foundational models sort of like app side,
0:17:31 it’s actually very interesting
0:17:33 because then you’re able to engage different type
0:17:35 of generations or different type of users
0:17:37 that like if you retain them
0:17:39 and you give them the best experience possible,
0:17:41 they will stay there for the coming year, right?
0:17:43 Whether that is because they’re building their own applications
0:17:45 on top of that because they’re essentially like,
0:17:47 “Well, I want to use your app overall.”
0:17:49 And the way we also think about this at 11 apps
0:17:51 is like layers, right?
0:17:53 So having the foundational layer,
0:17:55 which is like the research that we provide, right?
0:17:58 We do LMS and essentially we provide the best text to speech
0:18:01 and AI voices in the market, fantastic.
0:18:02 What else do you have on top of it?
0:18:05 The data that we’ve acquired that we’ve licensed from partners,
0:18:07 the products end-to-end products that we’re building,
0:18:10 the partnerships that we have, the customers that we have.
0:18:12 So you end up creating all of these multiple layers
0:18:13 that essentially end up building
0:18:16 your core defensibility in the market
0:18:19 that hopefully will sustain us for the coming years, right?
0:18:20 As the market changes,
0:18:22 if one of the layers like ends up getting replaced,
0:18:24 absolutely fine because then essentially you have
0:18:25 all of the other ones that will back you
0:18:27 over the long term, right?
0:18:28 – Yeah, and something you spoke to here
0:18:30 is just like this new generation.
0:18:32 And I think we’re all kind of trying to figure out
0:18:34 what can now be done with AI
0:18:37 when you talked about UX even or designing a new UI.
0:18:39 Voice is now in the mix in ways that it wasn’t before,
0:18:41 but then you also have this question of,
0:18:43 “Do I want to completely reinvent the wheel?
0:18:45 “Show someone a very powerful UI
0:18:47 “that they’re maybe just not familiar with
0:18:48 “and that you don’t retain them.”
0:18:50 So Gaurav, I’d love to probe you on retention.
0:18:52 I mean, even just from the perspective of desktop
0:18:54 versus mobile, you do have a mobile app.
0:18:56 How do you think about designing for that?
0:18:59 Because we’ve seen over over the last, let’s say two years,
0:19:01 there’s this extreme willingness to try,
0:19:03 but then I think someone internally
0:19:05 and coin this like AI tourist phenomena, right?
0:19:08 It’s people try and then a lot of them do leave.
0:19:09 So how do you think about that?
0:19:11 – Yeah, I mean, it’s something we think about a lot
0:19:12 because at the end of the day,
0:19:14 I think you can kind of go by metrics
0:19:15 and you can really worry about like,
0:19:18 “Oh, there’s retention number, it should be at that number.”
0:19:20 And you can kind of get caught up in that a little too much
0:19:23 when the reality is like those micro optimizations
0:19:25 are not going to solve whatever retention problem
0:19:27 or any other metric problem that you might have, right?
0:19:29 At the end of the day, it’s about the user experiences.
0:19:31 It’s about solving a real problem.
0:19:35 I think generally, if you want a complete hit end to end,
0:19:37 you need to have a breakthrough technology
0:19:40 that’s applied to solve a very specific problem
0:19:42 that a user actually has, right?
0:19:43 And then you need to have an engine
0:19:45 that can deliver that solution
0:19:47 to people who have that problem
0:19:49 as quickly as possible across the world, right?
0:19:51 If you have all those pieces,
0:19:53 then you won’t have a retention problem
0:19:54 or an acquisition problem
0:19:56 or any other problem basically, right?
0:19:58 Now, the cool thing about this time right now
0:19:59 is the technologies are being developed
0:20:02 and there’s actually a crazy number of technologies out there,
0:20:02 right?
0:20:04 I think it’s a very unique time from that perspective, right?
0:20:07 And for product people, the main problem is,
0:20:09 “Hey, like how do we actually solve problems, right?”
0:20:11 Actually solve real problems that people have, right?
0:20:13 And not just sell the technology as technology, right?
0:20:16 Like, “Hey, we have technology, just that, right?”
0:20:19 But actually convert it into a real value delivery
0:20:20 for users for a specific use case,
0:20:21 even an issue’s case, right?
0:20:23 Whatever it might be, right?
0:20:24 And then I think for marketers,
0:20:26 the problem is how do we actually educate people
0:20:28 that there’s a new way to solve these problems, right?
0:20:30 Like people may not think the first thing,
0:20:31 “Oh, you know what?
0:20:33 “I’m gonna Google AI for this, right?”
0:20:35 That might not be the first thing that people think about, right?
0:20:36 They might be searching for just
0:20:37 whatever they were normally doing, right?
0:20:39 Which may be something that takes a long time.
0:20:41 And, or they might be like not aware
0:20:43 that there’s new solutions available
0:20:44 for these problems, right?
0:20:46 So I think that’s sort of the end-to-end.
0:20:49 I think if you focus on that at that level,
0:20:51 like all the other numbers sort of follow on their own,
0:20:53 and that’s kind of what we’ve seen,
0:20:56 both across our desktop app and our mobile apps as well.
0:20:57 And we’re in the consumer space,
0:21:01 so retention is definitely a very hard game to crack
0:21:03 compared to, say, B2B businesses.
0:21:05 But we’ve been able to do it really well.
0:21:08 And like, I think it’s because of that high-level focus
0:21:11 across technology, product, and marketing.
0:21:13 – Yeah, maybe Laura, you used to work at Twitter.
0:21:15 What are you learning in terms of products
0:21:17 that reach so many people?
0:21:19 We’re talking daily active users.
0:21:20 What have you learned from that space
0:21:22 that you can apply to AI
0:21:25 when you are trying to fix this retention problem?
0:21:28 – I will say that I am so glad to be out of the game
0:21:31 of trying to optimize for MDAU
0:21:33 for monetized daily active users.
0:21:34 I think that daily active use
0:21:38 is like a pretty terrible metric
0:21:40 to uncover customer value, right?
0:21:42 And so one of the things that I just love most
0:21:43 about working at Descript
0:21:46 is being able to identify alternative metrics
0:21:49 to think about how they’re done right by the customer.
0:21:51 Two that I really like to think about
0:21:53 that are a bit in tension with each other.
0:21:54 They act as guardrails,
0:21:57 it’s time to expression and editing richness.
0:22:00 So I think if Descript is doing its job really well,
0:22:03 the amount of time it takes you from starting a project
0:22:05 to getting it into a shareable state,
0:22:06 whether you’re a marketer
0:22:09 who is like trying to repurpose a webinar into clips
0:22:12 or someone who is more of a creator,
0:22:14 trying to make your latest YouTube review
0:22:16 or you’re someone in learning development,
0:22:18 trying to create a training.
0:22:20 I want the amount of time it takes you to create that
0:22:22 to go down and down.
0:22:25 And so you’re able to just create more and more of the content.
0:22:26 Is anyone here a creator in any way
0:22:29 have a YouTube channel or a marketer?
0:22:31 Do you know about just like the gaping maw
0:22:35 that can never be fully fed or stated for content
0:22:37 that I find so many of our customers
0:22:39 are just staring into with despair?
0:22:42 And so getting kind of their time to expression down
0:22:43 is really important.
0:22:45 But one of the ways you do that is just like
0:22:47 by creating worse and worse content
0:22:49 that it’s just a role with an iPhone
0:22:50 and you slap some captions on it,
0:22:53 which is great for some use cases,
0:22:56 but for others just like a missed opportunity,
0:22:57 like you could have done so much more
0:23:00 to create really high quality video content.
0:23:02 And so if Descript is also winning on increasing
0:23:04 the editing richness,
0:23:06 the number of jobs that you’re able to do with us
0:23:08 and the number of things you’re able to do
0:23:10 to transform your media and make it really high quality,
0:23:12 the interaction of those two metrics
0:23:15 is such a great way to drive towards customer value.
0:23:17 I will say that like what Gorav said
0:23:20 around just like good product fundamentals
0:23:23 with retention totally resonates with me.
0:23:24 My attitudes for the tourists
0:23:27 is you’ve got to triage the tourists.
0:23:29 Some component of them just don’t have a use case
0:23:30 for your software.
0:23:32 They want to create a voice clone.
0:23:33 They want to see it.
0:23:34 They’re like, oh, that looks cool,
0:23:36 but they don’t have anything to do with that voice clone.
0:23:38 And it’s like, great, let’s let them do that.
0:23:39 That’s awesome.
0:23:41 Maybe one day you’ll think about Descript
0:23:43 or 11 Labs and come back.
0:23:45 But then who are these tourists
0:23:48 who actually have a legitimate use case
0:23:49 and they just don’t know it yet.
0:23:50 They could be using video
0:23:52 to communicate within their company.
0:23:54 They could be using text-based video editing
0:23:56 to create all of their marketing clips
0:23:57 and they don’t know that yet.
0:24:01 And how can I create software that activates really well,
0:24:03 that displays all of our use cases
0:24:05 and lets them have a good first time?
0:24:08 And I find that like often retention problems
0:24:11 or just activation problems in disguise in a trench coat.
0:24:13 And so what I really try to focus on
0:24:18 to improve retention is just like the activation experience.
0:24:20 – Just having come from a social media background
0:24:23 as well at Snap, such a good point about just DAU
0:24:25 and like how that can be such a trap.
0:24:28 – I think social media companies obviously optimized DAU
0:24:31 for a reason because money’s coming from a different source.
0:24:33 And so actually it’s good to be out of that game.
0:24:36 And really interestingly with the generative AI space,
0:24:38 it seems like it’s kind of having the opposite effect
0:24:40 on what it’s trying to achieve.
0:24:42 Like social media on one end is using AI as well,
0:24:46 but really to consume time from people as much as possible.
0:24:49 Consume as much of your time and it’s succeeding.
0:24:51 And on the other hand, generative AI is actually kind of
0:24:54 giving back time to people so they can actually do more.
0:24:55 So pretty cool.
0:24:57 – Yeah, we talked about this on a recent episode,
0:25:00 how some tools, I’m sure people would resonate with this.
0:25:03 If you had one excellent session,
0:25:06 it could have saved you four hours of work in five minutes.
0:25:07 That’s actually more valuable
0:25:09 than spending 20 minutes every day in an app.
0:25:11 And you don’t see that in the same metrics, right?
0:25:13 So I love that you brought up different metrics
0:25:15 that you’re paying attention to, Laura.
0:25:18 Charles, is there anything that jumps to mind there for you
0:25:21 in terms of how you might rethink a business model
0:25:23 in terms of what metrics you’re paying attention to,
0:25:25 or the way that you’re monetizing a product
0:25:28 that might be different because the willingness to pay
0:25:30 we’ve also seen is there, even if it is just,
0:25:33 I’m using this once a month, once every two months even.
0:25:36 – Yeah, and I think it’s a really good point, right?
0:25:39 Some consumers actually feel that if they need to do something
0:25:42 twice, the product is not working well, right?
0:25:44 It’s that element that we’ve gone from one side
0:25:45 to the other side.
0:25:47 So probably like someone in the middle
0:25:48 is what it fits well.
0:25:50 I was actually like in a meeting with a customer
0:25:52 and we presented a C level last week.
0:25:54 And the question they came back with was like,
0:25:56 okay, so how much time am I gonna save?
0:25:58 And I was like, well, you’re gonna save anywhere
0:25:59 between 50 to 60 times the time.
0:26:01 Like it’s gonna be like 50 to 60,
0:26:03 like it’s slashed by 50 to 60.
0:26:04 And they were like, no, that’s not possible.
0:26:06 And I was like, let’s do the math right now.
0:26:07 And we did the math.
0:26:08 And it was very interesting.
0:26:11 So I think there is an emphasis on that side.
0:26:14 But I think like sometimes we try to overemphasize
0:26:16 the effects of like the efficiency
0:26:18 that you’re getting with Genitive AI
0:26:20 when in fact, Genitive AI is not perfect, right?
0:26:22 I think like that’s one of the main reasons
0:26:25 why the AI tourists are there and they’re very big.
0:26:28 It’s because everyone comes with like such a big expectation
0:26:30 that he’s gonna be solving all of my problems
0:26:32 and it’s gonna be cooking dinner for me tonight as well.
0:26:34 And unfortunately it’s not gonna cook dinner for you.
0:26:36 It’s just never gonna solve all of your problems.
0:26:38 But it’s gonna help you quite a lot
0:26:40 either because you can do a lot of more modernization
0:26:41 with your customers,
0:26:43 with like you can reach new markets
0:26:45 or you can actually do it much quicker, right?
0:26:47 But I think like framing it on actually
0:26:50 what is valuable for you as a business
0:26:52 or as an individual is much more important.
0:26:54 So like initially our metrics were like beautiful,
0:26:55 like usage, right?
0:26:56 And over the past month,
0:26:59 we’ve ended up like switching to like usage is important,
0:27:02 but that does not define like the long-term success
0:27:04 of an actual customer for us, right?
0:27:06 It’s one of our like, yeah, activation side
0:27:09 is about actually what’s the use case that you have
0:27:11 and how do we measure that of the long-term
0:27:13 and how do we understand, try to insert the use case
0:27:16 based on the way you’re using the product, right?
0:27:17 So that we can offer you the best tools
0:27:19 and the best tips and all that stuff.
0:27:21 For us that that’s essentially those are the key metrics
0:27:22 to the best like usage.
0:27:24 Usage is still super important,
0:27:27 but I don’t really mind if someone uses the product today
0:27:30 and then doesn’t do it for like a week or two weeks
0:27:32 because I know that like if we’ve nailed it,
0:27:34 they’re gonna come back two weeks later, right?
0:27:35 I think that’s how we are thinking about it.
0:27:37 – You don’t have those social notifications
0:27:38 that are like a friend of a friend
0:27:42 maybe posted something, please come to our app.
0:27:43 All right, well, so we’re gonna open up
0:27:44 to questions very soon.
0:27:47 So if you have any questions start thinking about them,
0:27:48 but I wanna do rapid fire one or two more.
0:27:52 So the importance of optimizing an application
0:27:54 for a specific role or someone’s use case,
0:27:56 who are you, what are you trying to do?
0:27:59 So each of you actually comes from different backgrounds,
0:28:00 right?
0:28:01 So Gora, you’ve done design and development,
0:28:02 you’ve been an engineer,
0:28:04 Laura, you’ve been immersed in product,
0:28:05 carless operations.
0:28:07 And so those are roles where there’s a gosh,
0:28:11 I don’t know how many other people who fit that subset.
0:28:13 So I’d just love to hear your perspective,
0:28:15 independent of your company,
0:28:17 how do you think of AI as let’s say the next five years?
0:28:20 What does an AI-powered engineer look like
0:28:21 in your case, Gora,
0:28:23 or like an AI-powered operations person?
0:28:24 What do you need, what’s missing?
0:28:25 Are there products out there
0:28:28 that actually fit that use case and are doing it well?
0:28:30 – Yeah, I mean, thinking about it
0:28:32 from an engineering perspective
0:28:34 or even from a design perspective,
0:28:37 I think maybe the closest on the engineering side
0:28:39 would be like a tech lead manager,
0:28:42 someone who’s actually setting up the overall architecture
0:28:44 of whatever’s being built, right?
0:28:45 But a lot of the work’s been done by AI
0:28:47 and they’re coming in, they’re making edits.
0:28:48 They’re like, maybe we need to change this,
0:28:49 reviewing stuff, right?
0:28:50 Same on design, right?
0:28:52 Like kind of giving high level instructions
0:28:54 and like, let’s have this,
0:28:56 let’s maybe use this style over here,
0:28:57 let’s change these components, right?
0:28:59 And getting that output back
0:29:01 and kind of reviewing it, leaving comments
0:29:03 the same way that a manager might, right?
0:29:06 And being able to produce hopefully a lot more value
0:29:07 and output.
0:29:10 So that means that companies can be going
0:29:12 to a much larger revenue scales with way fewer people,
0:29:14 which is gonna be interesting.
0:29:15 – Yeah, I think a lot about this.
0:29:17 What is the AI product manager?
0:29:19 The paradigm that I use is more like,
0:29:22 how do I wanna interact with AI to do my job better?
0:29:24 One of the use cases I’m excited about
0:29:27 is a rubber duck who talked back.
0:29:28 You guys hear about like rubber ducking
0:29:29 where you keep a rubber duck on your desk
0:29:32 and you talk through difficult problems with that rubber duck.
0:29:35 And I think like, I’m never going to cede control
0:29:39 of the creativity and the genius to like the entity.
0:29:41 Like clearly, have you met me?
0:29:42 I’m in charge of that.
0:29:45 But I think like it can be fun to toss the ball around
0:29:47 with someone and I think I’m excited to see
0:29:51 how AI continues to develop to be like a fun thing
0:29:55 to toss the ball around and then can take all of the stuff
0:29:58 that you’re just like spewing out all of the kind of word
0:30:02 garbage and turn it into something crisp and readable
0:30:04 and easy to understand.
0:30:07 So that’s a use case that I’m excited about.
0:30:08 – I think it’s from an operation side,
0:30:10 it’s like even more complex, right?
0:30:12 Because like there’s so many like things that you need to do.
0:30:13 Like how do you automate
0:30:16 or how do you get someone to help you on that front, right?
0:30:18 So ideally you end up having a product
0:30:22 that helps you to twice as much in the same amount of time.
0:30:23 Not because I’m thinking about it
0:30:24 from an efficiency perspective,
0:30:27 but much more of how I can potentially generate
0:30:28 more revenue for the business, right?
0:30:30 I think that’s where potentially
0:30:32 and hopefully like the market is going to be going.
0:30:34 I like on the sales side, it’s much easier
0:30:37 because you end up having AISDRs these days.
0:30:39 We’ll end up having AISDSMs in all of those pieces.
0:30:42 Like that can be already there in many cases, right?
0:30:44 But purely on the operation side,
0:30:45 there’s a lot more complex.
0:30:46 Chagivity is your friend for sure, right?
0:30:49 Or on topic if you use it, or like any of those tools
0:30:51 that will help you generate quite a lot of different things
0:30:52 on a day-to-day basis.
0:30:54 Is that giving you a 2X?
0:30:55 Not yet, right?
0:30:58 So I’m not sure, like I still haven’t found the right product
0:31:00 that like would help anyone optimize
0:31:02 and become like 2X themselves.
0:31:03 – Maybe someone will build it in the room.
0:31:04 I guess final question.
0:31:06 Does anyone feel free to jump in?
0:31:08 All three of your products have a lot of customers.
0:31:09 People are using it.
0:31:12 Seems like maybe for the retention problem,
0:31:13 what challenges are you facing?
0:31:17 Whether it’s like regulation or not having the right models
0:31:19 or hoping that the open source models catch up
0:31:21 or just curious if anything jumps out
0:31:22 where just calling out a challenge
0:31:24 that you’d like to be solved in the next few years.
0:31:26 – Yeah, I’d say for us, it’s hiring actually.
0:31:28 It’s very traditional, right?
0:31:29 But I think hiring the right people
0:31:30 to solve the particular problems
0:31:31 that we’re having in our company.
0:31:33 And problems go really quickly,
0:31:35 or the company’s going really quickly, right?
0:31:36 And you have to kind of keep an eye on
0:31:38 all the different things that are happening,
0:31:39 where new needs might come up,
0:31:41 especially with a company like ours,
0:31:43 where we’ve existed for about three years
0:31:44 and there’s video companies that have been around
0:31:46 for a long time.
0:31:47 We’ve passed everybody in revenue
0:31:49 in like literally a year and a half.
0:31:51 And with growth at that scale,
0:31:53 you just have to constantly be thinking about
0:31:55 what are the new problems that are coming up
0:31:57 and who can we hire solve those problems, right?
0:32:00 So I think that’s like a very traditional answer.
0:32:01 And maybe there’s some AI recruiters out there,
0:32:02 but we have a great team.
0:32:04 So I don’t think we need them, at least not yet.
0:32:05 – Maybe AI can help with that.
0:32:09 I think it’s just that we’re in the middle
0:32:10 of a paradigm shift, right?
0:32:12 Like we haven’t gotten to the end of it.
0:32:13 We’re in the middle now.
0:32:14 And what I can tell you is that the way
0:32:18 that we’re going to edit video and audio in a year
0:32:22 or in two years is going to look completely different
0:32:24 than how we’re doing it right now.
0:32:27 But we don’t know how yet.
0:32:29 And on one hand, like that’s why I’m here.
0:32:32 That’s like why I’m doing this job,
0:32:34 because this is a place where the next generation
0:32:37 of like product managers and designers
0:32:38 we’re going to reinvent the way
0:32:41 that humans and computers interact with each other.
0:32:42 Someone’s going to figure it out.
0:32:44 And God, I hope it’s like me
0:32:46 or that I’m part of it in some small way.
0:32:50 But that’s also just like a very fragile moment, right?
0:32:52 Like it’s both a challenge and an opportunity.
0:32:54 And I think it’s like the challenge
0:32:56 of our industry right now.
0:32:59 – I think for us, it’s like there’s two sides of it.
0:33:02 What is definitely hiding, I can relate a lot on that.
0:33:03 It’s difficult.
0:33:06 We’ve gone from like zero to like tens and tens of millions
0:33:08 in months, not even years, in months.
0:33:10 And it’s really difficult to find people
0:33:12 that have experienced that previously,
0:33:13 also because like the market has evolved
0:33:15 very quickly in such a timeframe.
0:33:16 So that’s one side.
0:33:18 So there’s a lot of commitment that like we expect
0:33:19 from people at the company
0:33:23 and we need to be able to actually keep growing at this stage.
0:33:24 And on the research side,
0:33:26 it’s extremely difficult to find the right researchers
0:33:28 on the engineering side, on the operation side,
0:33:31 on sales, even support like across the board, right?
0:33:33 But that’s one side of the equation.
0:33:36 The other side of the equation is preventing misuse, right?
0:33:37 And I think realistically,
0:33:38 that is something that we have
0:33:39 an entire team dedicated to that
0:33:41 a day and night in the fourth, seven.
0:33:43 But every time that we put together something
0:33:45 that is windy or the different things
0:33:47 that like people make up to try to game it.
0:33:49 And it is similar to fraud,
0:33:51 where like you’re always like two steps behind
0:33:53 and it’s really difficult to cut and like keep fighting it.
0:33:55 So I think like about those two elements
0:33:56 are like the biggest challenges
0:33:58 that we constantly facing as a company.
0:34:00 Like we’re winning, but still it’s just a matter
0:34:02 of making sure that you’re constantly innovating
0:34:05 and having resources for something that it is important.
0:34:07 Otherwise like regulators come
0:34:08 or like consumers don’t blame
0:34:11 and things like that and people complain, right?
0:34:12 – Yeah, you need unprecedented people
0:34:14 for an unprecedented pace.
0:34:16 Quick question is Laura,
0:34:18 who’s our wonderful producer at the A16Z podcast
0:34:19 is gonna go around.
0:34:22 So if anyone does have a question, just raise your hand
0:34:24 and she’ll come find you.
0:34:28 – I’m curious how are we thinking about internationalization
0:34:32 or serving users of like various levels of digital literacy?
0:34:34 – We’ve had an international audience from the beginning,
0:34:36 including every country and every region
0:34:37 you could possibly imagine.
0:34:40 So I think it’s been a high priority from the beginning,
0:34:41 right?
0:34:44 Because the interesting thing is a lot of the development
0:34:48 that AI is bringing is not just things that are usable
0:34:50 in like, oh, it’s just an English thing
0:34:53 or oh, it’s just like a US thing or something.
0:34:56 It actually brings change in workflows
0:34:57 across almost every country
0:34:59 and every culture you can imagine.
0:35:00 And it actually works, right?
0:35:03 Like I think we’ve gone and launched new markets
0:35:04 where we’ve had zero users
0:35:08 and overnight had an explosion of users in that market.
0:35:10 But then we learned something about that particular market
0:35:13 where, oh, they don’t like this particular thing
0:35:15 or if you think about, for example, the Middle East, right?
0:35:17 Text is written in the opposite way.
0:35:19 And so that changes a lot about the UI
0:35:21 and changes a lot about the user experience, right?
0:35:23 And we’ve done a lot of work to make that good
0:35:26 and make that as usable and as amazing of an experience
0:35:28 as it is in any other language.
0:35:30 So those are the types of efforts
0:35:32 we’ve made high priority from the beginning.
0:35:34 – Would you say that other countries or regions
0:35:37 are actually more readily adopting the products
0:35:38 because I’m just thinking through,
0:35:40 well, actually maybe they can’t hire the software engineer
0:35:42 or maybe they can’t pay for the traditional video editor
0:35:44 or those thousands of dollars.
0:35:48 So they’re actually more readily adopting these technologies
0:35:49 ’cause they’re bringing the cost down.
0:35:50 – Absolutely.
0:35:52 I mean, I think around the world people are super open
0:35:53 to trying something new
0:35:55 to see if they can change their workflow, right?
0:35:56 I think as long as you can provide something
0:35:58 that is once you try it,
0:36:00 you can’t go back to what you were doing before.
0:36:01 That’s it.
0:36:02 That’s the difference, right?
0:36:05 If you can provide that experience in any language,
0:36:06 any culture, any country,
0:36:08 people will use the product.
0:36:10 – I mean, I think for us internationalization
0:36:11 has been like since day one there.
0:36:13 We have a fully international team,
0:36:14 everyone is fully remote.
0:36:16 So that actually, there’s a very strong correlation,
0:36:19 funny enough between the actual employee profile
0:36:21 and the fact that we are multiple countries,
0:36:22 everyone can be based whatever they wanted
0:36:24 and traveling and all of that stuff.
0:36:26 And the actual user type that we’ve got it, right?
0:36:28 So yes, in the initial days,
0:36:31 like a lot of our growth came from North America
0:36:32 and European markets.
0:36:34 But actually these days, when you look at the entire pie,
0:36:37 it’s like super spread out across the world.
0:36:39 I can relate to that purely on the fact
0:36:41 that people want the best tools
0:36:43 that will help them on a day-to-day basis, right?
0:36:45 And you don’t really need to spend these days,
0:36:48 like thousands of dollars or like hundreds of dollars
0:36:51 to actually produce a video or to produce a podcast
0:36:52 or produce something, right?
0:36:55 You could do it much cheaper using tools.
0:36:56 And that’s beauty of it.
0:36:59 So by default, like anyone that truly wants to have
0:37:01 a cost efficient solution will end up like using
0:37:04 any of the tools, script, captions or labs
0:37:06 or anything else that you have out there.
0:37:08 So by default, you end up having the strategy
0:37:09 that is about international markets,
0:37:11 with doing well content,
0:37:14 like trying to engage your audiences like that’s where they are
0:37:17 and trying to personalize it to them anyway.
0:37:20 Otherwise, I think like you end up like having a problem
0:37:22 of being very skewed towards a market,
0:37:23 traditionally it’s been always that,
0:37:25 oh, you go one market, you conquer it
0:37:27 and then you expand to another one.
0:37:28 And this day it’s just not,
0:37:31 it’s just that it worked quite well enough.
0:37:34 – Yep, it’s time for maybe one, maybe two more.
0:37:35 I see one at the back.
0:37:38 – I’m just wondering what barriers or stop gaps
0:37:40 you might be putting in place for people
0:37:43 who may be using your products for nefarious purposes
0:37:46 and thinking about trust and safety.
0:37:48 – I think like from 11, we invest like millions
0:37:51 every single year on actually like preventing misuse, right?
0:37:53 And we will start somebody to implement
0:37:54 like a fingerprinting system
0:37:56 for any content that gets generated.
0:37:58 So since we launched the fingerprinting has been in place,
0:38:01 we then opened up the API and the UI,
0:38:02 make sure that anyone can check
0:38:05 whether something was generated by us or not.
0:38:07 And since then we’ve also essentially engaged
0:38:10 on monitoring the content that our users generate.
0:38:12 So that essentially if someone is generating things
0:38:14 that they shouldn’t, then essentially we block them.
0:38:17 We’ve gone as far as to build the Nogo Voices,
0:38:20 which is a model that will prevent anyone
0:38:23 that tries to clone a celebrity voice for instance, right?
0:38:25 We’re constantly adding all of these layers
0:38:27 to try to make sure that we stay ahead of the curve.
0:38:28 But as I was saying earlier,
0:38:31 like it’s an uphill battle overall, right?
0:38:33 There is always ways in which you can game it.
0:38:35 But at the same time, like you have open source tools, right?
0:38:38 So we can try to do our side of the equation,
0:38:40 like anything that is open source
0:38:43 and to some extent you don’t really have that much
0:38:45 like control over those tools, right?
0:38:47 But I think it’s important as a company,
0:38:49 we will keep investing like millions every single year
0:38:52 and we can increase it as the market grows as well.
0:38:55 – I have to just quickly ask because it’s very timely
0:38:57 and I’m sure people in the audience are wondering
0:38:59 with some of the recent news around AI voices,
0:39:01 let’s just leave it at that and celebrities.
0:39:05 Are you finding there to be a bunch of false positives?
0:39:07 ‘Cause I feel like that’s maybe something
0:39:09 that people wonder, you hear a celebrity’s voice,
0:39:12 but how unique can a voice be?
0:39:15 And so if you’re trying to filter out certain people’s voices,
0:39:19 are you finding that actually like our voices maybe aren’t
0:39:20 that unique?
0:39:22 – That’s a really good question, right?
0:39:24 The voices are not as unique as everyone thinks,
0:39:25 but however they quite unique.
0:39:28 So you end up having like false positives for sure,
0:39:30 but we end up thinking like, if it’s a false positive,
0:39:31 if it tells you like, oh,
0:39:32 you don’t have permission for this voice,
0:39:34 automatically it tells you like, oh,
0:39:36 but you can still pass the voice structure
0:39:38 and it would show you the voice structure.
0:39:40 So if you pass it because it is your voice,
0:39:43 then you’re able to actually like use your own voice, right?
0:39:46 I have a twin brother for the ones that don’t know.
0:39:48 We do sound exactly the same.
0:39:49 And even my parents actually,
0:39:51 they sometimes they made mistakes, right?
0:39:52 So truly like, I could be talking,
0:39:54 but you could be thinking that it’s my twin brother.
0:39:56 We have exactly the same voice.
0:39:58 And that is a challenge that as a company we have
0:40:00 and a society we have, right?
0:40:01 But I think like the end of building layers
0:40:04 as a product from a product perspectives
0:40:06 to help filter those false positives.
0:40:07 I think like people understand that like,
0:40:10 you’re trying to go from like everything is free for all
0:40:11 and then you can misuse as much as you wanted.
0:40:13 There was like, let’s put some controls
0:40:15 and even if there’s some false positives,
0:40:17 people understand it online.
0:40:19 – Something about the product side of this too,
0:40:21 which I do think is super important
0:40:23 to sort of like build the safety features
0:40:25 from the product, from the ground up,
0:40:26 like in the product from the ground up.
0:40:27 And that’s kind of the difference
0:40:28 between offering a technology versus offering a product.
0:40:31 If you just say, hey, come to our website,
0:40:31 make deep fakes, right?
0:40:33 That’s offering a technology.
0:40:35 And some people might be out there doing that, right?
0:40:36 I don’t know, right?
0:40:38 But I think if you build that into a product,
0:40:41 like for example, we have the language translation feature,
0:40:43 right, which can translate whatever you’re speaking
0:40:45 to a different language, change your lip movements as well.
0:40:48 And yes, that’s using the same technology,
0:40:49 but in a very opinionative way
0:40:51 that you can’t change what was said,
0:40:53 but you can change what language it was set in, right?
0:40:55 And so that limits the scope of abuse
0:40:57 immediately, quite a bit, right?
0:40:59 And then all the traditional methods
0:41:01 can be used on top of that as well.
0:41:02 – Bri, I mean, with these group,
0:41:04 you can create a voice clone of yourself
0:41:06 and sort of like intermingle.
0:41:08 We have this thing called Overdub,
0:41:09 where if I say the wrong word,
0:41:11 I can go back in with the text,
0:41:13 say the word that I actually meant to say,
0:41:14 and then it will with my voice clone,
0:41:15 kind of create that.
0:41:18 But obviously there are a lot of misuses there.
0:41:20 And so whenever we launch a product,
0:41:22 we launch it with protections in place
0:41:26 and do a bunch of testing and hire outside people
0:41:28 to try to crack it and try to make sure
0:41:30 that we do our very best to make sure that it’s ungamable.
0:41:35 But like you said, if people are extremely determined
0:41:36 to crack through security,
0:41:38 like they will always find new ways to do it.
0:41:40 And this was the case when I was in social media too,
0:41:42 where like you do all kinds of things
0:41:44 to try to protect your platform.
0:41:47 And bad actors, they get up every morning
0:41:49 and grind just as hard as you do.
0:41:52 And so you’re just sort of in the eternal struggle.
0:41:54 And I think like every single tech product
0:41:55 should be thinking about like,
0:41:57 how are people going to misuse us
0:42:00 and making sure that they’re responsibly providing
0:42:02 a bunch of resources to stay in the fight.
0:42:06 – So as VP of Revenue at 11,
0:42:09 how do you view the role of open source?
0:42:11 Because as a developer myself,
0:42:13 I would rather use, for example, Falcon 70B,
0:42:15 which is a dollar in dollar out per million tokens,
0:42:18 as opposed to GPT-4, which is 30 and 50 out.
0:42:21 So do you think that open source is a threat to your business,
0:42:23 especially as companies like Meta
0:42:24 are kind of taking a scorched earth approach
0:42:26 to releasing models?
0:42:29 – I mean, I think it’s complementary actually.
0:42:31 You always end up having like businesses
0:42:34 or like people that like can go and use open source
0:42:35 and they have the means and the tools
0:42:37 and the knowledge to make that work.
0:42:39 And then you’re having quite a lot of different people
0:42:42 that like don’t really have those means or knowledge, right?
0:42:45 So it just ends up becoming like different sides
0:42:47 of the business or different sides of the market, right?
0:42:49 However you want to segment it.
0:42:50 When I think about voices,
0:42:52 we’ve been talking to each other as humans
0:42:54 for the past 50,000 years, right?
0:42:56 And there wasn’t really a good technology
0:42:58 that was able to replicate how we talk as humans.
0:43:02 So the fact that like as a platform or like even open source,
0:43:05 you’re able to actually replicate people’s voices
0:43:08 with their permission, make it sound natural, engaging,
0:43:10 and then power a new type of communication
0:43:12 and like platform and experience.
0:43:13 The market is massive.
0:43:16 So by default, you need to have both sides
0:43:18 to be able to actually like counterbalance each other
0:43:20 and push each other.
0:43:23 But it comes also the open source at a cost,
0:43:25 which is like the number of features that you will have
0:43:27 is like more limited, right?
0:43:30 So you will end up also having like less voices.
0:43:31 So what’s your preference?
0:43:32 Like you don’t have the UI.
0:43:35 So what’s your preference as a business or as an individual?
0:43:37 Is it purely building on top of it?
0:43:39 Then maybe open source is a good way.
0:43:41 Like today, the quality is not there yet.
0:43:43 But I’m sure that within the next three years,
0:43:45 the quality is going to be like matching anything
0:43:46 that is like private, right?
0:43:48 So it’s going to be more about like the actual system
0:43:52 that you build around it to make sure that like people start
0:43:55 like using it in a much easier way and then embed it anyway.
0:43:57 But I actually think it’s like complimentary.
0:43:59 Like without one, we can not have the other one purely
0:44:02 because the market like needs both sides.
0:44:05 – So just a follow-up, would you say that’s important for,
0:44:08 I guess, picks and shovels, companies,
0:44:10 closed source to build an application layer on top
0:44:12 to stay competitive?
0:44:17 – I don’t think anyone has actually built a pool like LLM.
0:44:20 If they’re not able to build applications on top of it,
0:44:23 to make life easier for consumers and businesses,
0:44:25 you will end up struggling down the line.
0:44:26 Whether that is in six months time
0:44:28 and that is in 18 months time, you will struggle.
0:44:29 Because at the end of the day,
0:44:31 like I want to launch my own application
0:44:34 like my product to use the product like this immediately,
0:44:35 right?
0:44:37 And if I need to spend the next like like coding
0:44:39 and building the UIs and everything
0:44:41 might give up and go somewhere else.
0:44:43 Even if it’s more expensive,
0:44:44 especially if I don’t even know
0:44:46 where they have product market fit.
0:44:48 And product market fit, like we always think about like
0:44:50 actual startups, but like big corporates
0:44:52 might not have even product market fit.
0:44:54 So if you want to iterate quickly
0:44:57 and then go to market as quickly as possible,
0:44:59 then you might want to have a stack
0:45:01 that is like truly readily available for you.
0:45:03 But once you’re ready and you’ve tested it
0:45:06 and the technology fits good enough with other LMS
0:45:07 or like open source,
0:45:09 then you might end up looking to switch.
0:45:11 And we’ve seen that with OpenAI,
0:45:13 like the big migration that like from developers
0:45:15 like that started using OpenAI,
0:45:18 such as GPT, APIs and GPT 3.5.
0:45:20 And then now they’re migrating towards like Anthropic
0:45:22 and like Mithral or Lama.
0:45:24 That’s been happening for the past six months.
0:45:25 It will continue happening, right?
0:45:28 So you start to validate that everything goes well
0:45:30 and then you figure out whether there is alternatives
0:45:32 or that is like really negotiating pricing
0:45:33 or like open source.
0:45:35 (upbeat music)
0:45:39 If you liked this episode, if you made it this far,
0:45:40 help us grow the show,
0:45:43 share with a friend or if you’re feeling really ambitious,
0:45:48 you can leave us a review at ratethisfodcast.com/asixz.
0:45:51 You know, candidly producing a podcast
0:45:54 can sometimes feel like you’re just talking into a void.
0:45:55 And so if you did like this episode,
0:45:58 if you liked any of our episodes, please let us know.
0:46:01 I’ll see you next time.
0:46:03 (upbeat music)
0:46:06 (upbeat music)
0:46:08 (upbeat music)