Synthesia Co-founder and COO, Steffen Tjerrild: the Future of Video, Trust, and Deepfakes

0:00Steffen Tjerrild:I think we'll start to have to trust BBC even more than we do today. Now they will be able to kind of create hundreds of BBCs down the line, right? Because the production, the journalistic things, it's going to be easier and easier and easier. And the barriers to entry is going to be lower and lower and lower. Hello, and welcome to Giant Ideas with me, Cameron McLean, and me, Tommy Stadlin. We're co-founders of Giant Ventures, which builds and backs purpose-driven companies. At Giant, we're lucky to meet extraordinary people with Giant Ideas that are changing the world. This podcast brings you behind-the-scenes access to those ideas and the inspiring stories of the people behind them.

0:35We explore how one giant idea can kickstart a billion-dollar company, shape culture, and transform life as we know it. Today on the podcast, we have Synthesia co-founder and COO Stefan Gerald. Synthesia is an AI company that allows you to create professional videos with just a few lines of text. It's now used by over 1 million customers, including 80 % of Fortune 100 companies. Synthesia was recently valued at over$2 billion. And at the heart of this exponential growth is Stefan. As the company's COO, Stefan has seen the company through its inception to the latest funding round. Today, we're going to talk about the magic of creating video from a few lines of text and what that means for human creativity.

1:17Welcome, Stefan. Thanks for having me. Thanks so much for being here. Stefan, you and I met back in 2018 when you were pitching for your seed round for Synthesia. I thought you were awesome, thought your co-founder, Victor, was awesome. But probably like many other people, made the bad mistake of passing. It was at my previous firm, Hummingbird Ventures. I thought you two were awesome. I could see that you were going to do something special. But candidly, I didn't see the giant idea at that moment in time. I think you were pitching more around dubbing for media in Hollywood at that time. But fast forward to now, you have built a company that's valued at$2 billion.

1:49It's worth, it's doing over$100 million of revenue, I believe. and you've built a remarkable magical product that essentially turns text into video like magic and you're working with over 80 % of the fortune 100 company so it's been quite a journey so tell us you know how you started and i guess give the audience an essence of what this giant idea is

2:08Steffen Tjerrild:yeah thanks a lot for having me good to reconnect yeah of course been a few years but uh but yeah back then it was obviously less clear kind of how big and kind of know how this kind of category would unfold so i think there's definitely to not to only your fault but i think the pitch was not as clear a clear as kind of what the outcome has kind of gone back then but uh but back then it was kind of originally stemmed from like a research project so we kind of partnered up with the two professors who kind of for the first time showcased how you can change and update video after the fact so when it comes off camera and the first kind of you know use cases of the things that kind of came to mind there was around kind of movie dubbing enriching and kind of you know taking content that has been shot with a camera and actually kind of you know translating that with the lip movements but back then it was very much like a very unscalable solution and we had to kind of film somebody to then kind of you know create a new piece of content and it was kind of you know inherently unscalable and not really kind of you know something that was almost kind of easy backable back then yes um but actually full circle so we actually launched that product yesterday where you can just one click up an existing piece of content with lip sync, with all these things that historically kind of required a full kind of five PhDs, kind of, you know, changing every frame to kind of create just a few seconds of new video.

3:25Steffen Tjerrild:So it's kind of becoming full circle since then. And it's pretty awesome to kind of, you know, revisit that product and kind of, you know, getting that actual to life with this new kind of offering. So let's talk about the idea itself, the current idea, the big idea, the giant idea. Text to video. Is it as simple as that? That is the huge idea. Yeah, I think if you kind of boil it down to kind of this notion of these kind of, you know, the technicalities of kind of typing in text and getting video out, right, I think that's kind of, you know, pretty magical. But I think what is also kind of allowing companies and individual creators, individual storytellers is also kind of really democratizing kind of, you know, movie creation and storytelling per se, right?

4:08Steffen Tjerrild:I think if we parallel it out to other creative disciplines, such as music creation, we're sitting in a music studio now, right? I think 10, 15 years ago, you needed to go to a Hollywood recording studio to create a top charting hit. Today, you can synthesize any tune in the universe on your laptop. And we think that the same kind of creative disruption or creative democratization is going to happen to video, especially with these magical, amazing video models that's coming out. I think we're sitting right next to DeepMind who kind of came out with the amazing ViewFree. And it's getting closer to this vision.

4:41Steffen Tjerrild:When we started the company, right, we had this vision that you can create a Hollywood style movie on your laptop. And that is actually starting to really kind of come to fruition now. We said back then 10 years. So it's actually kind of, you know, pretty close now, which is pretty kind of, you know, amazing to see that everybody in the world with a good story to actually kind of express that with the techniques that has kind of been a kind of monopoly in Hollywood. with hundreds of millions of dollars for actors and movie studios and so forth. I think it's going to, you know, change what humans are not that good at in historic kind of terms, right, with repeatable processes.

5:15Steffen Tjerrild:I find this very rudimentary and very kind of repeatable to something that is kind of more of creative nature where humans are just kind of, you know, much, much better at some of those tasks, right? So I think there's no doubt that I think by allowing more people to kind of get into the field and allowing them to have the techniques to kind of do these type of things, I think will be a positive force for sure. But I think it will have some kind of replacement of jobs in the long run for sure. But I think that's very general for a lot of AI today. Okay. Tell us a little bit about where creativity will thrive then in the world of Synthesia, where it is about storytelling.

5:50So kind of, you know, why would Cameron's storytelling prompt for Synthesia be better than mine? What is it about how he's going to write that that's going to be more impressive than mine? Is it great writers will kind of thrive and create amazing videos? Will it be more the concept? Will it be quite technical things about how exactly you prompt the level of detail? Who wins in that world?

6:11Steffen Tjerrild:I think it's people who are great storytellers in general have a lot of creativity, have a lot of good things around how they kind of tell stories, right? But I think it's going to be those people who's working with these systems. There's also going to be bouncing off ideas with these systems a lot, right? So I think it's important that you kind of continue to kind of work with this system and really kind of, you know, are trying to feed back off it because that's kind of where you can kind of really put these technologies to the edge of where it's kind of, you know, good enough or where is it not good enough for today.

6:43Steffen Tjerrild:But I think it will be very different from where you are and what is the audience. And I think also like storytelling will fragment a lot. I think we'll kind of go from kind of very broadcast where we all kind of consume the same media in the cinemas to much more individualized feeds as we do on our phones with TikTok. where we have all unique feats, right? We are all consuming very unique information, which is based on our interest, our kind of way of want to kind of, you know, consume information, right? And I think this will also just allow, kind of, you know, storytellers to kind of, you know, have much more individualized, the much more kind of, you know, fragmented kind of stories to distribute.

7:18How close are we to creating video that is indistinguishable from video that was created by humans or are you already there?

7:26Steffen Tjerrild:I think we're getting there. Getting there? I think both with kind of the avatar technology we have, I think we're really kind of starting to get there. We're launching Express 2, which is our latest model, which is like full hand gestures that is in sync with what you're saying. So some of the earlier versions of the technology was like waist up where it doesn't have hands. Now they're kind of full body. They can kind of gestulate in the language and kind of know the body language that is kind of tied to what you're saying. right and then also with these magical kind of foundational models which is you've seen all over the internet right i think it's also just getting increasingly good enough for a lot of these things they're still kind of you know hard to manage and hard to kind of you know uh kind of tame and they're also kind of pretty short form but i think give this a few years right i think it'll get to to to a level where you think it's a few years you think a few years till we hit the uncanny valley when it's indistinguishable it depends on kind of what type of content if it's like corporate content what we do i think we are there okay but if it's like things you watch in the cinema it will still be a few years and when these when this happens and these lines blur how do you think the world will change and you guys are at the real forefront how will we deal with deep fakes i think for we have like had the system since we started the company since we met met you back in back in the early days right this framework around the free c's so consent collaboration and control which means that you can't just upload anybody to our system and make a video of them saying something they have ever said.

8:54Steffen Tjerrild:So we have pretty strong consent systems. So every avatar, every person on the platform has given explicit consent to giving their likeness and digitizing their likeness. We also have really strong content moderation which means that you can't just type anything out in the platform. So we are pretty strict on what you can actually produce. And if you think about content moderation on the internet, that's typically only moderated on the point of distribution. So that's kind of when you actually upload it to Facebook or TikTok or whatever. That's kind of the point of moderation where we actually moderated on the creation, right?

9:29Steffen Tjerrild:That's kind of in the Word document. It's kind of being moderated. You can't even type out hate speech and this type of thing. That's kind of how strict we are with trying to kind of enforce this on the level of creation. And then we also have like, you know, collaboration, right? I think one thing is our system not being abused and misused for some of these things. But it's obviously, as these technologies democratize, it becomes easier and easier to produce these things with open doors and so forth. It's obviously important that we think about this more on an internet-wide scale. So how do we make sure that we believe what we see and have the trust in the internet and the information that we consume?

10:04Do you have a second giant idea for solving that problem?

10:06Steffen Tjerrild:I think it will be less about, has this been manipulated or changed by AI? It will be more, is this the original meaning with the piece of content? So it's the media provenance. If you think about all photos on the internet, right? I think 99 % of them have been through PhotoFilter or Photoshop. And even when you take a picture on your iPhone today, it actually does changing the light systems and the manipulation. Is that AI too? It's less about has it been manipulated or not. It's more, is this the original meaning? And if that provenance changes, right? So if I take a news clip from BBC and changing the anchors' motions on that, then the provenance breaks.

10:47Steffen Tjerrild:And when I upload that to YouTube, it'll use a Shasham-style understanding. Has this clip been uploaded before and has the original meaning been altered? I think that's how we think about how this will be solved at scale. There's this great BBC drama about deepfakes and the real problems in that start when you get live deepfakes. So this person goes on BBC News and does a live interview, and the real politician is talking, giving one message, and then is going out broadcast live, a slightly different, well, very different message, right? How close are we to live deepfakes, or it doesn't have to be deepfakes, but I guess live Synthesia?

11:23Steffen Tjerrild:Yeah, I think we're working on these capabilities now. I think that's also kind of a big picture where we think this is moving from this kind of one directional broadcast into much more kind of conversational pieces, right? You're also kind of talking to AI today with your Grog or with your OpenAI or whatever you're using, right? You're kind of speaking and using the voice mode. And the initial next step is also kind of having a face to some of those capabilities, right? So I think we are getting closer to kind of that real-time world. And I think kind of, you know, the security and the provenance obviously will kind of have to kind of, you know, move with it.

11:58Steffen Tjerrild:And I'm pretty secure, feel pretty good about how we kind of think in and tackling these type of things today. Do you feel good about it? I mean, obviously this is massively positive. You talked about the creativity you're going to unleash for the world with your technology. And so clearly there is so much good there. But it is worrying, right? The world of deep fakes and what effect that can have on democracy. What part of you sort of lies awake at night thinking, am I going to be responsible for something kind of horrific? Or in the world, do you worry about that at all? You've got a young child.

12:30Yeah, I actually don't worry about it.

12:32Steffen Tjerrild:I think we've kind of gone a long way. And I think some of the things we've seen of how, I think we all talk about technology will disrupt a lot of things and AI will have a huge kind of impact on our society. And we all agree. I think it will actually strengthen our institutions. I think we'll start to have to trust BBC even more than we do today because now they will be able to kind of create hundreds of BBCs down the line, right? Because the production, the journalistic things, it's going to be easier and easier and easier. And the barriers to entry is going to be lower and lower and lower.

13:02Steffen Tjerrild:and that's kind of why we'll kind of gravitate to some of these institutions that we're really kind of trusting today and i think being able to kind of you know thinking and changing our worldview from default trusting to default being skeptical unless it's kind of coming from critical sources i think that's kind of where in the world we are moving to right now i think this kind of provenance thing is going to be something that has to be rolled out on an internet-wide scale to to kind of bring that kind of security and that kind of trust back what's the craziest possible future we could imagine with the digital avatars?

13:35Steffen Tjerrild:I think there's obviously kind of, I know the things that we're working on now is kind of this more conversational piece, right? Where you can think about a lot of your business operations are actually being automated with kind of conversations. There's a lot of like structured calls and businesses that is very repeatable, that has very low value, can kind of be onboarding calls and kind of be cybersecurity It can be security training. It can be customer education. Some of these kind of mechanics that is requiring a lot of very rudimentary kind of human power, but it's valuable enough to kind of put a human in front of other humans to make sure that they can understand certain concepts.

14:12Steffen Tjerrild:I think that's kind of the world where we're moving to, where it's less about creating MP4 files, fees cheaper and faster, but it's actually delivering full understanding, right? where we're actually certifying that you've understood the information by being quizzed, by making sure that you understand the materials. And the thing that's going from these passive communication flows that are happening every day, where people are zoning out, they're watching things on 2x speed, they're clicking through security awareness training without understanding anything. And I think that's kind of where the world, where you're actually certifying people have understood it before you can complete stuff.

14:49So it's sort of like the LLMs, the intelligence of open AI matched with, you know, digital avatars that feel very real and kind of that interaction back and forth. I want to go back to that idea of this personalized feed. So it's one thing for a producer to serve someone else, you know, a consumer, a video which they made with AI. But this idea potentially of us all as consumers being able to create our own entertainment, essentially, right? We're sort of scrolling mindlessly through Netflix and not finding something for 20 minutes that you want to watch. you can just go i want to watch a drama about you know two cops in london who are running around king's cross trying to solve a murder like how close are we to consumers self-serving and self-creating the content that they want to watch which is netflix level quality yeah i think it'll still

15:32Steffen Tjerrild:be a few years before we kind of get to that level but will we get there yeah we'll 100 % get there yeah i think to get that with a cost and kind of inference cost i think it'll be maybe five years where you can kind of really produce Netflix-style content, let it just continues, where you can make 10 more seasons of whatever kind of, pick your favorite show, right? But I think it's interesting, right, where we kind of, you know, going from this one-size-fits-all, which is kind of the movie theater, right, to a much more personalized, contextualized, localized version, right, where every kind of piece of content can kind of be very unique and it can also be unique to your situation, right?

16:09Steffen Tjerrild:So if you are driving your car, you probably want to hear the information as a podcast. If you're sitting home at nine, you probably want to like Netflix style movie. If you're at work, it could be like, you know, a condensed kind of document with all the kind of, you know, the things, right? So then you'll have this ability to kind of, you know, transform the information to be consumed in different formats. And you can even get your slide deck as a Netflix video, or you can have your whatever brand pitch book or pitch ideas, kind of different formats, right? and i think as you also get closer to how the audience wants to kind of perceive and understand that you can also kind of you know tailor that on the on the recipient and even though that the creator might have some idea of how this should be created there can be that recipient can also kind of you know peripheralize it and contextualize that information in the way they want i don't think great new shows will be built that way but personally i kind of love that idea i love the more seasons of the show you always love i think your gut it is over exactly i think i think if already been made, then they could extend the season.

17:08Steffen Tjerrild:Or change the ending from don't kill her, kill him or whatever. I mean, this is sort of a slight tragedy of what's going on with AI right now is that it needs the internet, well, the LLMs need the internet to feed their models, but the incentives for creating that content are going away, right? So it's sort of like a victim of some success. It's interesting how that element will play out because I think you need original content to feed the AI beast, right? I think with synthetic data and some of these things, that is really starting to kind of become good enough. I think we have so much data already.

17:46Steffen Tjerrild:I think we've seen that with the data, right, we can create incredible things. But of course, original ideas is going to be the ultimate test for AGI, right? How can you make neophysics breakthroughs? how can you actually kind of, you know, really kind of, you know, take the world understanding to a next level with kind of, you know, new capabilities. Just for listening to synthetic data, tell us about that. So I think as you're kind of, you know, running out of kind of humanly written content or movies produced on YouTube or whatever, there's kind of, you know, going to be a need to kind of, you know, train your AI models, AI systems on synthetic data that is not created by humans.

18:24Steffen Tjerrild:So synthetically generated data, right, that will either kind of, you know, change or kind of, you know, re-mask some of the biases your model have to kind of, you know, fill the gaps. But it can also kind of, you know, take your models and take your AI systems to the next level by feeding its systems data that is not really have enough in the original world and producing that with synthetic data that then trains AI systems. And that's happening now. You're already there. Already doing that. It's starting there. It's starting. Yeah, this is already kind of a really big field. Cool. Well, shifting gears a little bit to the business itself, many of our listeners are entrepreneurs.

18:57I think they'd love to hear a bit more about your plans for the business. One thing we were discussing before this chat is that, you know, when the marginal cost of creating video is dropping to almost zero, even in this format, how do you think you will create and capture value moving forward?

19:13Steffen Tjerrild:Yeah, I think especially for us, right, we're very focused on the enterprise, so working with the biggest companies in the world, right, and I think what they're buying from Synthesia is less about a video clip generator. It's actually a full workflow, right? Yeah. Was it originally a video clip generator? It was, yeah. We were kind of delivering clips for a long time. And the early versions of the platform, you created an avatar, you built kind of, you know, the full video edits, right? Where there's more to a video than just somebody talking to you. There's animations, transitions, music, all this stuff that actually makes a video, a compelling video, right?

19:48Steffen Tjerrild:So we kind of built, we built kind of the full kind of product experience around this paradigm that generated media is different from camera recorded content. And that's kind of, you know, the product experience that we've kind of been tailoring for today and also leveraging all the facts that every pixel is software rendered as opposed to be sensor captured, right? And that has kind of, you know, been kind of the product experience and the product paradigm that we're building for now. But I think what we've seen is our kind of mode as we kind of, you know, going and maturing as a business from a research lab, Can we actually make the technology work to, oh, wow, can we actually build a unique product experience that is native to the fact that this is kind of AI?

20:30Steffen Tjerrild:And then next thing is, oh, okay, wow, we have such a great distribution. We have a license to talk with all these great companies. They have a lot of problems they want to solve with visual information. How can we take and build a custom workflow that is actually delivering really high-end business outcomes for our customers, right? So I think there will be, as I know, video creation drops in costs. I still think it actually is on the rise because you are constantly also evolving the capabilities. We're still not at Netflix style, like customer onboarding in any kind of company. But I think we probably want to, right?

21:03Steffen Tjerrild:If we can get like a LeBron James style kind of onboarding video if I'm a new employee at Nike, right? With a really high production kind of quality of how to actually become a Nike employee, right? We're still kind of a little bit away from that. So I still think in the short to medium term, the cost will rise, but then it will kind of eventually kind of go to zero, towards zero, right, at least. But I think at that point, right, we'll kind of be much more embedded in workflows and there's more value than just kind of, you know, the frames, but actually kind of inserting ourselves in more business impact where today, right, we are delivering MP4 files that is an input to an outcome you want to produce.

21:43Steffen Tjerrild:If I go to any of the CFOs in the Fortune 500 and say, hey, you're creating fast and cheaper videos with Synthesia, they're going to say, so what? I want to kind of, you know, have lower customer support costs. I want to have higher kind of, you know, productivity in my sales force with better enabled salespeople. Tying into business outcomes. All these things, right? And I think that's kind of where we're moving to now, where instead of just delivering MP4 files, actually delivering business outcomes. So you're 100 million in revenue today, or thereabouts,$2 billion valuation. How do you become a$100 billion company?

22:17Steffen Tjerrild:I think it's kind of moving into, we think there's actually a pretty big opportunity to build a really big standalone. Because the paradigm for generated media is so different than camera recorded content. And if you think about the internet today, right, everything is kind of built for camera recorded content. You shoot something with a camera, you put it into the Adobe products, and you put it on YouTube. Those three kind of processes are completely air-gapped and completely different, right? And you can't really kind of go to a YouTube video and say, hey, make me 200 versions of this and A-B test my ad.

22:50Steffen Tjerrild:That's not possible today because you can't talk to the video editing and you can't talk to the camera, right? But what Synthesia has built is kind of this fully connected thing where you can also create the content, right? But you also have the editing and the collaboration kind of capabilities. And then you have the distribution of it. And that means that when you unify those three things, you can also start to build reinforcement or human self-improving systems. So if I can see that people are dropping off after five seconds of my video, I can render a new version that is becoming better at solving those problems.