Dan Kokotov: Speech Recognition with AI and Humans | Lex Fridman Podcast #151

The following is a conversation with Dan Kakotov, VP of Engineering at Rev AI, which is by many metrics the best speech to text AI engine in the world. Rev in general is a company that does captioning and transcription of audio by humans and by AI. I've been using their services for a couple of years now and am planning to use Rev to add both captions and transcripts to some of the previous and future episodes of this podcast to make it easier for people to read through the conversation or reference various parts of the episode. Since that's something that quite a few people requested, I'll probably do a separate video on that with links on the podcast website so people can provide suggestions and improvements there. Quick mention of our sponsors Athletic greens all in one nutrition drink Blinkist app that summarizes books, Business Wars Podcast and Cash app. So the choice is health, wisdom or money. Choose wisely my friends, and if you wish, click the sponsor links below to get a discount and to support this podcast. As a side note, let me say that I reached out to Dan and the Rev team for a conversation because I've been using and genuinely loving their service and really curious about how it works. I previously talked to the head of Adobe Research for the same reason. For me, there's a bunch of products. Usually it's software that comes along and just makes my life way easier. Examples are Adobe Premiere for video editing, Izotope Rx for cleaning up audio, Auto Hotkey on Windows for automating keyboard mouse tasks, Emacs as an IDE for everything, including the universe itself. I can keep on going, but you get the idea. I just like talking to people who create things I'm a big fan of. That said, after doing this conversation, the folks at Rev AI offered to sponsor this podcast in the coming months. This conversation is not sponsored by the guest. It probably goes without saying, but I should say it anyway, that you cannot buy your way onto this podcast. I don't know why you would want to. I wanted to bring this up to make a specific point that no sponsor will ever influence what I do on this podcast or to the best of my ability, influence what I think. I wasn't really thinking about this, for example, when I interviewed Jack Dorsey, who is the CEO of Square that happens to be sponsoring this podcast. But I should really make it explicit. I will never take money for bringing a guest on. Every guest on this podcast is someone I genuinely am curious to talk to or just genuinely love something they've created. As I sometimes get criticized for, I'M just a fan of people and that's.

Who I talk to.

As I also talk about way too much. Money is really never a consideration in general. No amount of money can buy my integrity. That's true for this podcast and that's true for anything else I do. If you enjoy this thing, subscribe on YouTube, review on Apple, Podcast, follow on Spotify, support on Patreon, connect with me on Twitter lexfriedman. And now here's my conversation with Dan Kakotov.

You mentioned science fiction on the phone. So let's go with the ridiculous first. What's the greatest sci fi novel of all time in your view? And maybe what ideas do you find philosophically fascinating about it?

The greatest sci fi novel of all time is Dune. And the second greatest is the Children of Dune. And the third greatest is the God Emperor of Dune. I'm a huge fan of the whole series. I mean, it's just an incredible world that he created. And I don't know if you've read the book or not.

No, I have not. It's one of my biggest regrets, especially because the new movie is coming out. Everyone's super excited about it. It's ridiculous to say, and sorry to interrupt, is that I used to play the video game, used to be Dune. I guess you would call that real time strategy, right?

Right. I think I remember that game.

Yeah, it's kind of awesome. 90s or something. I think I played it actually when I was in Russia.

I definitely remember it. I was not in Russia anymore. I think at the time that I used to live in Russia, I think video games were about like the suspicion of Pong. I think Pong was pretty much like the greatest game I ever got to play in Russia, which was still a privilege, right? In that age.

So you didn't get color, you didn't get like.

So I left Russia in 1991. Right.

19:1. Okay.

So I wanted to feel like a kiss because my mom was a programmer. So I would go to her work, right. I would take the metro, I'd go to her work and play like on I guess the equivalent of like a 286pc, you know, nice with floppy disks. Yes.

So, okay, but back to Dune, what do you get?

Back to Dune. And by the way, the new movie I'm pretty interested in. But are you skeptical? I'm a little skeptical. I'm a little skeptical. I saw the trailer. So there's a David lynch movie, Dune, as you may know. And I'm a huge David lynch fan, by the way. So the movie is somewhat controversial, but it's a little confusing. But it captures kind of the mood of the book better than I would say most any adaptation. And Dune is so much about kind of mood and the world. But back to the philosophical point. So in the fourth book, God Emperor of Dune, there's a sort of setting where Leto, one of the characters, he's become this weird sort of God emperor. He's turned into a gigantic worm. I mean, you kind of have to read the book to understand what that means.

So the worms are involved.

Worms are involved. You probably saw the worms in the trailer, right?

And in the video game.

So he kind of, like, merges with the swarm and becomes this tyrant of the world, and he, like, oppresses the people for a long time, right? But he has a purpose, and the purpose is to kind of break through kind of a stagnation period in civilization, right? But people have gotten too comfortable, right? And so he kind of oppresses them so that they explode and go on to colonize new worlds and kind of renew the forward momentum of humanity, right? So to me, that's kind of, like, fascinating, right? You need a little bit of pressure and suffering to kind of, like, make progress, not get too comfortable. Maybe that's a bit of a cruel philosophy to take away, but that seems.

To be the case, unfortunately. Obviously, I'm a huge fan of suffering. So one of the reasons we're talking today is that a bunch of people requested that I do transcripts for this podcast and dude captioning. I used to make all kinds of YouTube videos, and I would go on upwork, I think, and I would hire folks to do transcription, and it was always a pain in the ass, if I'm being honest. And then I don't know how I discovered Rev, but when I did, it was this feeling of like, holy shit, somebody figured out how to do it just really easily. I. I'm. I'm such a fan of just when people take a problem and they just make it easy, right? You know, like, just. There's so many. It's like there's so many things in life that you might not even be aware of that are painful, then Rev, you just, like, give the audio, give the video. You can actually give a YouTube link, and then it comes back, like a day later or two days later, whatever the hell it is, with the captions, you know, all in a standardized format. That was. I don't know. It was. It was. It was. It was truly a joy. So I thought I had, you know, Just for the hell of it. Talk to you. That one other product, it just made my soul feel good. One other product I've used like that is for people who might be familiar is called Izotope rx. It's for audio editing. And like that's another one where it was like you just drop it. I dropped into the audio and it just cleans everything up really nicely. All the stupid like the mouth sounds and sometimes there's background like sounds due to the malfunctioning the equipment. It can clean that stuff up. It can. It has like general voice denoising. It has like automation capabilities where you can do batch processing and you can put, put a bunch of effects. I mean it just, I don't know, everything else sucked. For like voice based cleanup that I've ever used, I've used Audition, Adobe Audition. I've used all kinds of other things with plugins. You have to kind of figure it all out. You have to do it manually here. It just worked. So that's another one in this whole pipeline that just brought joy to my heart. Anyway, all that to say is Rev. Put a smile to my face. So can you maybe take a step back and say what is Rev and how does it work?

And Rev or Rev.com RevRev.com same thing. I guess we do have Rev AI now as well, which we can talk about later.

Like do you have the actual domain or is it just the actual domain?

But we also use it kind of as a sub brand. So we use Rev AI to denote our ASR services. Right. And Rev.com is kind of our more human and to the end user services.

Brilliant. So it's like WordPress.com and WordPress.org they actually have separate brands that I don't know if you're familiar with what those are. Yeah, they provide almost like a separate branch of a little bit.

I think with that it's like revolution. Now WordPress.org is kind of their open source. Right. And WordPress.com is sort of their hosted commercial offering.

Yes.

And with us the differentiating is a little bit different, but maybe similar idea.

Yeah. Okay, so what is Rev?

Before I launch into what is Rev, I was going to say you're talking about like Rev was music to your ears. Your spiel was music to my ears. To us, the founders of Rev. Because Rev was kind of founded to improve on the model of Upwork that was kind of the original or part of their original impetus. Like our CEO Jason was an early employee of upwork, so he's very familiar with their.

Upwork the company.

Upwork the company. And so he was very familiar with that model and he wanted to make the whole experience better because he knew, like, when you go. At that time, upwork was primarily programmers. So the main thing they offered is if you want to hire someone to help you code a little site, you could go on Upwork. You could browse through a list of freelancers, pick a program or have a contract with them and have them do some work. But it was kind of a difficult experience because for you, you would kind of have to browse through all these people and you have to decide, okay, well, is this guy good or is somebody else better? And naturally you're going to upwork because you're not an expert, right? If you're an expert, you probably wouldn't be getting a programmer from upwork. So how can you really tell? So there's kind of like a lot of potential regret, right? What if I choose a bad person? They're going to be late on the work. It's going to be a painful experience. And for the freelancer, it was also painful because half the time they spent not on actually doing the work, but kind of figuring out how can I make my profile most attractive to the buyer? Right. They're not an expert on that either. Actually. Started with translation, and then we added audio transcription of it later. And we'll just make it a website. You go, give us your files, we'll give you back the results as soon as possible. Originally, maybe it was 48 hours, then we made it shorter and shorter and shorter.

Yeah, there's a rush processing too.

There's a rush processing now. And we'll hide all the details from you.

Right.

And like, that's kind of exactly what you're experiencing, Right. You don't. You don't need to worry about the details of how the sausage is made.

That's really cool. So you picked like a vertical. By vertical you mean basically a service.

A service category.

Why translation? Is rev thinking of potentially going into other verticals in the future? Or is this like the focus now is translation transcription, like language?

The focus now is language or speech services. Generally speech to text language services. You can kind of group them however you want. But originally the categorization was work from home. So work that was done by people on a computer. We weren't trying to get into TaskRabbit type of things and something that could be relatively standard, not a lot of options. So we could kind of present the simplified interface. So programming wasn't a good fit because each programming project is kind of unique. We're looking for something that transcription is. You have five hours of audio. It's five hours of audio translation is somewhat similar in that you can have a five page document and then you just price it by that and then you pick the language you want. And that's mostly all it is to it. So those were a few criteria. We started with translation because we saw the need and we picked kind of a specialty of translation where we would translate things like birth certificates, immigration documents. Immigration, things like that. And so they were fairly even more well defined and easy to kind of tell if we did a good job.

So you can literally charge per type of document. So what is it now? Is it pro word or something like that? Like, how do you measure the effort involved in a particular thing?

So now, like. So for audio transcription, right, it's per audio unit?

Well, that, that, yes, for a translation.

We don't really actually focus on that anymore. But, you know, back when it was still a main business of revit was per page. Right. Or per word, depending on the kind of.

Because you can also do translation now on the audio. Right?

Like subtitles. So it would be both transcription and translation. That's right.

I wanted to test the system to see how good it is to see, like, how. How well is Russian supported.

Think so, yeah.

It'd be interesting to try it out. I mean, one of the.

But now it's only in like the one direction, right? So you start with English and then you can have subtitles in Russian. Not really. Not really the other way.

Got it. Because it's. I'm. I'm deeply curious about this. When Covid opens up a little bit, when the economy, when the world opens up a little bit.

You want to build your brand in Russia.

No, I don't.

First of all, I'm allergic to the word brand. I'm definitely not building any brands in Russia. But I'm going to Paris to talk to the translators of Dostoevsky and Tolstoy. There's this famous couple that does translation. And, you know, I'm more and more thinking of how is it possible to have a conversation with a Russian speaker? Because I have just some number of famous Russian speakers that I'm interested in talking to. And my Russian is not strong enough to be witty and funny. I'm already an idiot in English. I'm an extra level of like awkward idiot in Russian. But I can understand it. Right. And I also, like, wonder how can I create a compelling English, Russian Experience for an English speaker. Like if I. There's a guy named Goli Perlman, who's a mathematician, who obviously doesn't speak any English. So I would probably incorporate like a Russian translator into the picture. And then it would be like a. Not to use a weird term, but like a three. Like a three, three person thing where it's like a dance of. Where, like, I understand it one way, they don't understand the other way. But I'll be asking questions in English. I don't know. I don't know the right way.

Complicated.

It's complicated, but I feel like it's worth the effort for certain kinds of people, one of whom I'm confident. Vladimir Putin, I'm for sure talking to. I really want to make it happen because I think I could do a good job with. But the. The right, you know, understanding the fundamentals of translation is something I'm really interested in. So that's why I'm starting with the actual translators of, like, Russian literature, because they understand the nuance and the beauty of the language and how it goes back and forth. But I also want to see, like, in speech, how can we do it in real time? So that's like a little bit of a baby project that I hope to push forward.

But anyway, it's a challenging thing. So just to share. My dad actually does translation. Not professional. He writes poetry that was kind of always his. Not a hobby, but he had a job, like a day job. But his passion was always writing poetry. And then we got to America and he started also translating. First he was translating English poetry to Russia. Now he also goes the other way. You kind of gain some small fame in that world anyways, because recently this poet, like Lewis Clark, I don't know if you. No. So American poet, she was awarded the Nobel Prize for literature. And so my dad had translated one of her books of poetry into Russian. And he was like one of the few. So they asked him and gave an interview to Radius Voboda, if you know what that is. And he kind of talked about some of the intricacies of translating poetry. So that's like an extra level of difficulty, right? Because translating poetry is even more challenging than translating just, you know, interviews.

Do you remember any experiences and challenges to having to do the translation that stick out to, like, something he's talked about?

I mean, a lot of it, I think, is word choice. Right. The way Russian is structured is first of all quite different than the way English is structured. Right. Just there is inflections in Russian and gender as and they don't exist in English. One of the reasons actually why machine translation is quite difficult for English to Russian and Russian to English because there's such different languages. But then English has like a huge number of words. Many more than Russian actually. I think so. It's often difficult to find the right word to convey the same emotional meaning.

Yeah. Russian language, they play with words much more. So you're mentioning that Rev was kind of born out of trying to take a vertical and upwork and then standardize it.

So we're just trying to make the. The freelancer marketplace idea better. Right. Better for both customers and better for the freelancers themselves.

Is there something else to the story of Rev finding Rev? Like what did it take to bring it actually to life? Was there any pain points?

Plenty of pain points. I mean that's often the case. It's with scaling it up. Right. And in this case the scaling is kind of scaling the marketplace, so to speak. Right. Rev is essentially a two sided marketplace. Right. Because there's the customers and then there's the revvers. If there's not enough revvers. Revvers are what we call our freelancers. So if there's not enough revvers, then customers have a bad experience. Right. It takes longer to get your work done, things like that. If there's too many, then the revvers have a bad experience because they might log on to see what work is available and there's not very much work. Right. So kind of keeping that balance is a quite challenging problem. And that's like a problem we've been working on for many years. We're still refining our methods. Right.

If you can kind of talk to this gig economy idea. I did a bunch of different psychology experiments on Mechanical Turk, for example. I've asked to do different kinds of very tricky computer vision annotation on Mechanical Turk. And it's connecting. Connecting people in a more systematized way, I would say, you know, between task and. And what would you call that worker is what mechanical terror calls it. What do you think about this world of gig economies of there being a service that connects customers to workers in a way that's like massively distributed? Like potentially scaling to. It could. It could be scaled to like tens of thousands of people. Right. Is there something interesting about that world that you can speak to?

Yeah, well, we don't think of it as kind of gig economy. Like to some degree, I don't like the word gig that much. Right. Because to some degree it diminishes the work being done. Right. It sounds kind of like almost amateurish. Well, maybe in like music industry, like gig is the standard term, but in work it kind of sounds like it's frivolous to us. It's improving the nature of working from home on your own time and on your own terms. Right. And kind of taking away geographical limitations and time limitations. Right. So many of our freelancers are maybe work from home moms. Right. And they don't want the traditional 9 to 5 job, but they want to make some income. And rev kind of allows them to do that and decide exactly how much to work and when to work. Or by the same token, maybe someone wants to live the mountaintop life, cabin in the woods, but they still want to make some money. And generally that wouldn't be compatible. Before this new world, you kind of had to choose. But like with rev, like you feel like you don't have to choose.

Can you speak to like, what's the demographics, like distribution? Like where do revvers live? Is it from all over the world? Like, what is it? Do you have a sense of.

All over the world? Most of them are in the US that's the majority, you know, because most of our work is audio transcription and so you have to speak pretty good English. The majority of them are from the US So we have people in some other of the English speaking countries. And as far as like us, it's really all over the place. For some years now we've been doing these little meetings where the management team will go to someplace and we'll try to meet revvers. And pretty much wherever we go, it's pretty easy to find a large number of revvers. The most recent one we did is in Utah, but anywhere really.

Are they from all walks of life? Are these young folks, older folks?

Yeah, all walks of life really. Like I said, one category is the work from home mom students who want to make some extra income. There are some people who maybe have some social anxiety. So they don't want to be in the office. This is one way for them to make a living. So it's really pretty wide variety. But on the flip side, for example, one revver we were talking to was a person who had a fairly high powered career before and was kind of like taking a break and just wanted. She was almost doing this just to explore and learn about, you know, the gig economy, quote unquote. Right. So it really is a pretty wide variety of folks.

Yeah, it's kind of interesting through the captioning process for me to learn about the revvers because like Some are clearly, like, weirdly knowledgeable about technical concepts. Like, you can tell by how good they are at like, capitalizing stuff, like technical terms like a machine learning and deep learning.

Right.

Like, I've used rev to annotate to caption the deep learning lectures or machine learning lectures I did at mit. And it's funny, like, a large number of them were like. I don't know if they looked it up or were already knowledgeable, but they do a really good job at like.

No, they invest time into these things. They will, like, do research. They will Google things, you know, to kind of make sure they get it right. But to some of them, it's like, it's actually part of the enjoyment of the work. Like, they'll tell us, you know, I love doing this because I get paid to watch a documentary on something.

Right.

And I learn something while I'm transcribing. Right. Pretty cool.

Yeah. So what's that captioning transcription process look like for the revver? Can you maybe speak to that to give people a sense, like, how much is automated, how much is manual, what's the actual interface look like, all that kind of stuff?

Yeah. So we've invested a pretty good amount of time to give reverse the best tools possible. So typical day forever. They might log into their workspace. They'll see a list of audios that need to be transcribed, and we try to give them tools to pick specifically the ones they want to do. So maybe some people like to do longer audios or shorter audios. People have their preferences. Some people like to do audios in a particular subject or from a particular country. So we try to give people the tools to control things like that. Then when they pick what they want to do, we'll launch a specialized editor that we've built to make transcription as efficient as possible. They'll start with a speech rec draft. So we have our machine learning model for automated speech recognition. They'll start with that, and then our tools are optimized to help them correct that.

So it's basically a process of correction.

Yeah, it depends on, I would say the audio. If audio itself is pretty good, probably our podcast right now would be quite good. So the SR would do a fairly good job. But if you imagine someone recorded a lecture in the back of an auditorium where the speaker is really far away, and there's maybe a lot of crosstalk and things like that, then maybe the SR wouldn't do a good job. So the person might say, you know what, I'm just going to do it from scratch. Do it from scratch.

Yeah.

So it kind of really depends.

What would you say is the speed that you can possibly get? Like, what's the fastest can you get. Is it possible to get real time or no. As you're like listening, can you write as fast as real time would be pretty difficult.

It's actually a pretty. It's not an easy job. You know, we actually encourage everyone at the company to try to be a transcriber for a day, transcriptionist for a day. And it's way harder than you might think it is. It is, right. Because people talk fast and people have accents and all this kind of stuff. So real time is pretty difficult.

Is it possible, like, there's somebody. We're probably going to use rev to caption this.

They're.

They're listening to this right now. What's, what's, what do you think is the fastest you could possibly get on this right now?

I think on a good audio, maybe 2 to 3x, I would say real.

Time, meaning it takes two to three times longer than the actual audio of the, of the podcast. This is, this is so meta. I could just imagine the reverse working.

On this right now. You're way wrong. You're way wrong.

This takes way longer. But yeah, it definitely doubted me.

Could do real time. Yeah.

Okay, so you mentioned ASR. Can you speak to what is ASR? Automatic speech recognition. How much? Like, what is the gap between perfect human performance and perfect or pretty damn good ASR?

Yeah. So ASR, Automatic speech recognition. It's a class of machine learning problem, right. To take speech like we were talking and transform it into a sequence of words, essentially.

Audio of people talking.

Audio. Audio to words. And there's a variety of different approaches and techniques which we could talk about later if you want. So we think we have pretty much the world's best ASR for this kind of speech, right. So there's different kinds of domains for ASR. One domain might be voice assistants, right. So Siri very different than what we're doing. Right. Because Siri, there's fairly limited vocabulary. You might ask Siri to play a song or order a pizza or whatever. And it's very good at doing that. Very different from when we start talking in a very unstructured way. And Siri will also generally adapt to your voice and stuff like this. So for this kind of audio, we think we have the best and our accuracy right now, I think it's maybe 14% word error rate on our test suite that we generally use to measure. So Word error rate is one way to measure accuracy for ASR.

Right. So what's 14%?

So 14% means across this test suite of a variety of different audios, it would get in some way 14% of the words wrong. 14% of the words wrong. So the way you kind of calculate it is you might add up insertions, deletions and substitutions. Right. So insertion says like extra words. Deletions are words that we said but weren't in the transcript. Substitutions is you said Apple. But the ASR thought it was able something like this human accuracy. Most people think realistically it's like 3%, 2%. Word error rate would be like the max achievable. So there's still quite a gap. Right.

Would you say that? So YouTube, when I upload videos, often generates automatic captions. Are you sort of, from a company perspective, from a tech perspective, are you trying to beat YouTube? Google, it's a hell of a Google. I mean, I don't know how seriously they take this task, but I imagine it's quite serious. And they, you know, Google is probably up there in terms of their teams on, on ASR, just nlp, natural language processing, different technologies. So do you think you can beat.

Google on this kind of stuff? Yeah, we think so.

Google just woke up on my phone. This is hilarious.

Okay, now Google is listening, sending it back to headquarters. Who are these rough people?

But that's the goal.

Yeah, I mean we measure ourselves against like Google, Amazon, Microsoft, you know, some smaller competitors. And we use like our, our internal tests. We try to compose it of a pretty representative set of audios. Maybe it's some podcasts, some videos, some interviews. Some interviews, some lectures, things like that. Right. And we beat them in our own.

Testing and actually rev offers automated. Like you can actually just do the automated captioning. So like I, I guess it's like way cheaper, whatever it is, whatever the rates are.

Yeah, yeah.

So it's a. By the way, it used to be a dollar per minute for captioning and transcription. I think it's like a $15 or something like that.

$25.

25. $25 now. Yeah, it's pretty cool. That was the other thing that was surprising to me. It was actually like the cheapest thing you could. One of the. I mean, I, I don't remember it being cheaper. You could on upward get cheaper. But it was clear to me that this, that's gonna be really shitty.

Yeah.

So like you're also competing on price. I think there were services that you can get like similar to Rev kind of feel to it, but it wasn't as automated. Like the drag and drop, the entirety of the interface. It's like the thing we're talking about. I'm such a huge fan of like Frictionless, like Amazon's Single Buy button. Whatever.

Yeah, that one click, the one click.

That'S genius right there. Like, that is so important for services, that simplicity. And I mean, Rev is almost there. I mean, there's like some. I'm trying to think. So I, I think I've. I stopped using this pipeline, but Rev offers it and that I like it, but it was causing me some issues on my side, which is you can connect it to like Dropbox and it generates the files in Dropbox so like closes the loop to where I don't have to go to Rev at all and I can download it. Sorry, I don't have to go to Revitall and to download the files. It could just like automatically copy them.

Right. You're putting a Dropbox and you know, a day later or maybe a few hours later.

Yeah, depending on the giggle.

Right. Just shows up.

Yeah, Yeah. I was trying to do it programmatically too. Is there an API interface you can. I was trying to through like through Python to download stuff automatically, but then I realized this is the programmer in me. Like, dude, you don't need to automate everything like in life, like flawlessly because I wasn't doing enough captions to justify to myself the time investment into automating everything perfectly.

Yeah. I would say if you're doing so many interviews that your biggest roadblock is clicking on the Rev download button. Now you're talking about Elon Musk levels of business, but for sure we have a variety of ways to make it easy. There's the integration you mentioned, I think is through a company called Zapier, which kind of can connect Dropbox to Rev and vice versa. We have an API if you want to really customize it, if you want to create the Lex Friedman, you know, CMS or, or whatever for this whole thing.

Okay, cool. So can you speak to the, the ASR a little bit more? Like, what is it? What does it take? Like, approach wise, machine learning wise? How hard is this problem? How do you get to the 3% error rate? Like, what's your vision of all of this?

Yeah, well, the 3% rate error rate is definitely. That's, that's the grand vision. We'll see what it takes to get there. But we believe in ASR, the biggest thing is the data.

Right.

That's true of a lot of machine learning problems today. The more data you have and the higher quality the data, the better label the data. That's how you get good results. And we at Rev have kind of like the best data we have.

Like you're literally model is annotating the data.

Our business model is being paid to annotate.

Being paid to annotate the data.

So it's kind of like a pretty magical flywheel. And so we've kind of like ridden this flywheel to this point and we think we're still kind of in the early stages of figuring out all the parts of the flywheel to use, you know, because we have the final transcripts and we have the audios and we train on that. But we in principle also have all the edits that the revvers make. Right.

Oh, that's interesting. How can you use that as data?

That's something for us to figure out in the future. But we feel like we're only in the early stages. Right.

But the data is there. That'd be interesting. Like almost like a recurrent neural net for fixing. For fixing transcripts. I always remember we did segmentation annotation for. For driving data. So segmenting the scene like visual data and you could. You can get all. So it was drawing people drawing polygons around different objects and so on. And it feels like it always felt like there was a lot of information in the clicking, the sequence of clicking that people do, the kind of fixing of the podgones that they do. Now there's a few papers written about how to draw polygons, like with recurrent neural nets to try to learn from the human clicking. But it was just like experimental, you know, it was one of those like CVPR type papers that people do. Like a really tiny data set. It didn't feel like people really tried to do it. Seriously. Yeah. I wonder. I wonder if there's information in the fixing that's. That provides deeper set of signal than just like the raw data.

The intuition is for sure. There must be. Right. There must be in all kinds of signals and how long you took to make that edit and stuff like that. It's going to be up to us. That's why the next couple of years is super exciting for us. Right.

So that's what the focus is. Now you mentioned Rev AI. That's where you want to.

Yeah. So Rev AI is kind of our way of bringing this ASR to the rest of the world. Right. So when we started we were human, only then we kind of created this TEMI service, I think you might have used it, which was kind of ASR for the consumer, Right? So if you don't want to pay $1.25, but you want to pay now, it's 25 cents a minute, I think. And you gather the transcript, the machine generated transcript, and you get an editor and you can kind of fix it up yourself. Right? Then we start using TSR for our own human transcriptionists. Then the kind of rev AI is the final step of the journey, which is we have this amazing engine. What can people build with it? Right? What kind of new applications could be enabled? If you have SpeechTrack, that's that accurate.

Do you have ideas for this, or is it just providing it as a service and seeing what people come up with?

It's providing it as a service and seeing what people come up with and kind of learning from what people do with it. And we have ideas of our own as well, of course, but it's a little bit like, you know, when AWS provided the building blocks, right. And they saw what people built with it and they try to make it easier to build those things. Right. And we kind of hope to do the same thing.

Although AWS kind of does a shitty job of, like, I'm continually surprised, like, Mechanical Turk, for example, how shitty the interface is. We're talking about, like, rev making me feel good. Like, when I first discovered Mechanical Turk, the initial idea of it was, like, it made me feel like rev does. But then the interface is like, come on.

Yeah, it's horrible.

Why? Why is it so painful? Does nobody at Amazon want to, like, seriously invest in it felt like you could make so much money if you took this effort seriously. And it feels like they have a committee of, like, two people just sitting back, like, like a meeting. They meet once a month. Like, what are we gonna do with Mechanical Turk? If it's like, two websites make me feel like this, that and craiglist.org, whatever the hell it is. Yeah, it feels like it's designed in the 90s.

Well, Craigslist basically hasn't been updated pretty much since the chiropractor.

Do you seriously think there's a team? Like, how big is the team working on a Mechanical Turk? I don't know.

There's some team, right?

I feel like there isn't. I'm skeptical.

Yeah. Well, if not possible, they benefit from, you know, the other teams, like, moving things forward. Right. In a small way. Possible, but. No, I know. You mean we do. We. We use Mechanical Turk for a couple of things as well. And yeah, it's painful. Ui, it's painful.

But yeah, it works.

I think most people, the thing is most people don't really use the ui, right? Like we, for example, we use it through the API, right?

So yeah, but even the API documentation and so on, like it's super outdated. Like, yeah, it's, it. I don't, I don't even know what to. I mean, the same criticism, as long as we're ranting. My same criticism goes to the APIs of most of these companies, like Google, for example, the API for the different services. It's just the documentation is so shitty. Like, it's not so shitty. I should, I should actually be, I should exhibit some gratitude. Okay, let's practice some grass gratitude. The, the, you know, the documentation is pretty good. Like most of the things that the API makes available is pretty good. It's just that in the sense that it's accurate, sometimes outdated. But like the degree of explanations with examples is only covering I would say like 50 of what's possible. And it just feels a little bit like there's a lot of natural questions that people would want to ask that doesn't, doesn't get covered. And it feels like it's almost there. Like it's such a magical thing. Like the Maps API, YouTube API.

There's a bunch of stuff I gotta imagine. It's like, you know, there's probably some team at Google, right, Responsible for writing this documentation. It's probably not the engineers, right? And probably this team is not where you want to be.

Well, it's a weird thing. I sometimes think about this for somebody who wants to also build a company, I think about this a lot. You know, YouTube, the, you know, the service is one of the most magical. Like I'm so grateful that YouTube exists and yet they seem to be quite clueless on so many things like that everybody is screaming them at like it feels like whatever the mechanism that you use to listen to your quote unquote customers, which is like the creators is not very, very good. Like there's literally people that are like screaming, why? Like their new YouTube studio, for example. There's like features that that were like begged for for a really long time. Like being able to upload multiple videos at the same time that was missing for a really, really long time now. Like there's probably things that I don't know, which is maybe for that kind of huge infrastructure. It's actually very difficult to build some of these features. But the fact that that wasn' it and it felt like you're not being heard. Like, I remember this experience for me, and it's not a pleasant experience. And it feels like the company doesn't give a damn about you. And that's something to think about. I'm not sure what that is. That might have to do with just, like, small groups working on these small features, on these specific features. And there's no overarching, like, dictator type of human that says, like, why the hell are we neglecting, like, Steve Jobs type of character? That's like, there's people that we need to speak to, the people that want to love our product, and they don't. Let's fix it.

Maybe at some point you just get so fixated on the numbers, right? And it's like, well, the numbers are pretty great, right? Like, people are watching, you know, doesn't seem to be a problem, right? And you're not like, the person that, like, built this thing, right? So you really care about it. You know, you're just there. You came in as a product manager, right? You got hired sometime later. Your mandate is, like, increase this number, like, you know, 10%, right? And you just.

That's brilliantly put. Like, if you. This is the. Okay, if there's a lesson in this is don't reduce your company into a metric of, like, how much. Like you said, how much. How much people watching the videos and so on. And. And. And, like, convince yourself that everything's working just because.