Jacob Goldstein
Jacob Goldstein spent more than a decade as co-host of the Planet Money podcast. He's also the author of the book Money: The True Story of a Made-Up Thing, which the New…
Join Pushkin+
Gain access to ad-free versions of 20+ podcasts from the Pushkin library along with exclusive bonus episodes and other member benefits.
We’ve all heard that nothing ever gets deleted from the internet. Luke Arrigoni says that just isn’t true. His company deletes thousands of deepfakes every day.
Luke is the founder and CEO of Loti, and his problem is this: How do you turn the Internet into a place where people have more control over their name, image and likeness?
Tens of thousands of famous people pay Loti to scour the internet and take down deepfakes and unapproved endorsements. But in Luke’s telling, this is just the beginning. The company will soon allow celebrities to license their likeness for use in AI videos, and it also plans to start offering its services to ordinary people.
In this episode, Luke explains:
- The harm that online deepfakes and impersonations can cause
- How Loti automates tens of thousands of takedowns per day
- Why the law may soon make it easier to fight deepfakes
- Why celebrities are beginning to license their likenesses to AI platforms
- AI’s role in the future of human creativity
Connect with us:
- Follow Jacob Goldstein on LinkedIn, X and Instagram
- Email us at problem@pushkin.fm
- Follow Pushkin on Instagram, LinkedIn or X
- Listen to Jacob’s other show, Business History
- To listen to the show ad free, sign up for Pushkin+
Pushkin. I'm Jacob Goldstein, and this is What's Your Problem? My guest today is Luke Aragoni. He is the co-founder and CEO of a company called Lodi. And Lodi is an acronym, and it stands for Likeness on the Internet. Tens of thousands of famous people pay Lodi to scour the internet and take down deepfakes and unapproved endorsements. If you've heard of a person, if a person is a little bit famous, there is a very good chance they are one of Lodi's clients. But in Luke's telling, this is just the beginning. The company will soon start offering a service for celebrities to license the use of their image. In other words, ordinary people will be able to pay Lodi to create stuff with AI using the faces and voices of real celebrities. Lodi also plans to start offering its services to ordinary people. The plan is to include a free tier for people who are not famous, but who are worried about what's happening with their likeness on the internet. At a macro level, the biggest level, Luke's problem is this. How do you turn the internet into a place where people have more control over their name, image, and likeness? In our conversation, Luke and I talked about the technology behind Lodi. We talked about federal legislation that may soon make it easier for people to fight deep fakes. And we also talk about AI and the future of creativity. To start, I asked Luke to give me the state of play for what's happening right now with unauthorized use of people's likeness on the internet.
01:52
Speaker 2
So you have kind of two different categories of people. If you're really famous and your likeness is going to get used to sell random things. It could be to sell a scam. It's like an impersonation, a fake Instagram of a celebrity, and they're selling you anything from VIP access to the next tour to a romance scam that they're falling in love with you, right? And all of them have some kind of economic extraction from it. Or it's kind of more straightforward. It's like, hey, I've come up with a brand new shampoo, and this famous celebrity uses the shampoo. Here's a deepfake video talking about how they love it. So that's probably the most common way in which it's used. And then you have a kind of more incendiary thing where everyone else, like me, for example, what we have to be concerned about is someone deepfaking us for any kind of authorization reason. And then there's an even darker way in which it's used that we saw at some high schools this year, which is a lot of folks are getting deepfaked to make images of them that were never taken, but have a very harmful nature to them and their reputation.
02:55
Speaker 1
Do you mean revenge porn? You're talking around revenge porn there.
02:59
Speaker 2
Yes. We like to call it non-consensual intimate images, which is a lot better of a term than revenge porn. But colloquially, we all call it revenge porn, and so we even have to kind of mix the terms together. But just to kind of re-educate a lot of people, NCII is a much better term for it.
03:14
Speaker 1
Okay, fair enough. So this is a potential problem for everybody. What's like... What's the scope of it right now? Like, if I just pick some random A-list celebrity, will there, like, 100% be deepfakes of them scamming on the internet right now?
03:33
Speaker 2
Yeah, we track about 30,000 or so celebrity accounts, people, personas.
03:40
Speaker 1
And those are your clients, or that's just you kind of understanding the landscape?
03:44
Speaker 2
Two-thirds, a little bit more, a little over 20,000 are clients.
03:48
Speaker 1
So that's way below the A-list. That's way more famous people than anybody could name. Oh, yeah.
03:53
Speaker 2
No, it is. It's pretty much every famous person that you can think of, even if you don't remember their name. And so, yes, for those folks, you're going to get nearly 100% coverage on impersonations and other scams.
04:04
Speaker 1
When you say nearly 100% coverage, you mean for all of those people, there are going to be deep fakes.
04:09
Speaker 2
Of them on the Internet? Yes, and impersonations. So it's like a real picture of them on an Instagram account, but not their real account.
04:16
Speaker 1
Right, so that's not a deepfake in the AI sense. You can just take the picture of the person and be like, look, this person loves our soap, and that's a scam and a lie.
04:25
Speaker 2
Absolutely.
04:26
Speaker 1
And just to be clear, like, that is illegal because people have a right to their name, image, and likeness? Like, that's the fundamental business and kind of intellectual property thing that we're talking about?
04:36
Speaker 2
Yeah, it's called a right of publicity. And ironically, it became a bigger conversation when the camera came around. Uh-huh. And so, you know, 100 plus years ago when people started taking pictures of anyone, someone was like, well, I'll put this on the front of my cereal box. This is just like an average person. They're like, no, that person has, all right, you can't just make, it really is the commercial component. You can't make money off their likeness without their approval or their endorsement.
05:02
Speaker 1
Okay, so that's the lay of the land. Your company sort of walks into this universe as it is growing a few years ago. Like, Basically, what's your company do?
05:14
Speaker 2
What we do is we download the internet every day, and then we run face recognition over it looking for our clients. And when we find them in a spot that they don't belong, whether it be a scam, impersonation, NCII, or any other type of thing, we figure out how we get this thing removed. So the removal part is another cool piece of tech because it's fully automated. But we can reach out to adult sites and tell them they are creating liability for their merchant accounts or, you know, they're domain provider, or a number of other ways. And then on social media, we kind of have this great access and great partnership across everyone but X. And they just give us a kind of API access to issue those takedowns.
05:54
Speaker 1
And so does that mean that it's fully automated, like your AI, essentially your facial recognition software, spots these unauthorized uses of people's likeness and then pings the software or the AI of whatever, YouTube or TikTok, and says, hey, take this down, and then the machine just takes it down? Like there's no human needs to.
06:18
Speaker 2
There's no human needed for this.
06:20
Speaker 1
And presumably that's because the volume is too high. Like what order of magnitude, how many, you know, how many things has your company removed, whatever, today or this week?
06:32
Speaker 2
We do tens of thousands of things a day.
06:35
Speaker 1
Tens of thousands a day.
06:37
Speaker 2
Yeah, it's pretty insane.
06:39
Speaker 1
That is insane. So, like, at every moment, at every second, someone is posting a scam likeness. I guess presumably at this point it's machines. Like, you don't need a person. Like, I could just build a bot to create a million scam likenesses.
06:55
Speaker 2
Correct. We called it an industrial game of whack-a-mole.
06:58
Speaker 1
Yeah. Like, what's the frontier? What are you struggling with right now?
07:03
Speaker 2
Hmm. I think a lot of people don't understand that you can have delete for the internet. And so we think about how we position the company into a greater context around like, okay, there's a thing of me online. It's going to be online forever. And we kind of have to interject ourselves and say, well, this won't be online forever. I know everyone was told that for 30 years, but now there's a really efficient way for us to manage that.
07:28
Speaker 1
What's the business model?
07:29
Speaker 2
So for celebrities, we charge a monthly fee to go scraping and removing this content. And for consumers, it'll be more freemium. They can find everything of themselves online, and then they get a certain number of takedowns if they upgrade to Pro. Or we actually teach them how they can go do the takedowns themselves, so we're not trying to gatekeep their privacy.
07:49
Speaker 1
When's it going to be available for consumers?
07:51
Speaker 2
Probably early October.
07:53
Speaker 1
Technically, how is it going?
07:56
Speaker 2
Oh, it's both a technical nightmare and a marvel. It's insane. because the problem set of downloading like 600 million assets a day, you have database problems and disk problems.
08:09
Speaker 1
Why do you need to download it?
08:12
Speaker 2
Because that's how you run the face recognition on it. Okay.
08:15
Speaker 1
Interesting. Yeah.
08:16
Speaker 2
And so you download it, you run your face recognition on it, and then that gives you the capability of saying, this person's coming to me. Can you find everything in my online? We look into our database. We say, oh, we can now identify these rows as you. The other things we toss out.
08:30
Speaker 1
Yeah. And are the Bad people doing things to make it harder for you?
08:37
Speaker 2
No, they're making it easier. They're trying to make their AI-generated faces more like the person, which actually is an alignment to us. It's like the more their deepfakes look realistic, the easier it is for us to catch them.
08:50
Speaker 1
Oh, that's interesting. And so presumably, I mean, the next step is figuring out what are the legitimate uses, right? Correct. In an automated way. How do you do that part?
08:59
Speaker 2
So that's more of the context understanding of what the image is. We actually do that discovery as well. We have this thing that we call a semantic understanding of every image and video where it comes through and we say, what's going on in this? Is someone running in this park? Are they walking in the park? Are they doing drugs? Are they doing something they shouldn't be doing in this picture or image? And then we have this massive collection of, we know that it's this celebrity doing this thing. And was there like an expectation of privacy here? Or More importantly, is this authorized? And they can give us their rules around what should or shouldn't be authorized.
09:34
Speaker 1
And presumably the giving the rules of what should and shouldn't be authorized also needs to be automated, right? Like presumably the clients need to say, these are the things I am endorsing. It's sort of a white list. Is that the system?
09:46
Speaker 2
Yes. Absolutely.
09:51
Speaker 1
Are there any technical problems you're trying to solve or is it technically a sort of solved issue for you?
09:55
Speaker 2
Every couple of months we we trip on something that is kind of insane. We're like, that's oh, I didn't realize we'd even have to solve that problem.
10:03
Speaker 1
What's an example? What's an example?
10:05
Speaker 2
Well, when we went to process video, we found out the reason why no one had done what we do before us, like no one did this in 2005 was because it's expensive to process video. So like to take a video in and actually like look at all the frames. even without great machine learning or AI at that time, it was just really hard to do that. And so we were broke and we couldn't afford to just spool up more servers like some really large AI companies are able to do. And so necessity was the mother of invention. And we invented a method of processing videos that were like a thousand times faster than anyone else. And we just did this thinking like, well, this is what you have to do, right? Like this is just whatever the next step is. And it came to find out that we, uh, We had to build something that no one else had done that others had tried to do and failed. Now it's like, it's one of the most guarded, protected algorithms at our company because before it was just kind of like, it was like floating everywhere.
10:56
Speaker 1
Did you patent it or is it a trade secret?
10:58
Speaker 2
Oh, interesting. It's a trade secret.
11:00
Speaker 1
So you can do a thing and you're not telling anybody how to do it.
11:03
Speaker 2
Correct. And people know that we do it. They still can't figure it out.
11:06
Speaker 1
Oh, that's interesting.
11:07
Speaker 2
We can prove it. And that's the only part that matters.
11:10
Speaker 1
And what it allows you to do is to essentially... have AI look at a video much more efficiently and cheaply than anybody else can or than anybody else could before? Interesting.
11:21
Speaker 2
That's right.
11:24
Speaker 1
And presumably you chose not to patent it because when you patent something, you have to tell the world how you did it and then people can copy it.
11:30
Speaker 2
And then they can tweak one part of the mathematics and essentially have something like that. And it's hard to patent like a pure math solution.
11:38
Speaker 1
Yeah. Has anybody figured it out? Can you tell?
11:41
Speaker 2
No, it's been two or three years and No competitor we've had is able to— they all stumble on video. Like, the volume that we could demonstrate is much higher.
11:52
Speaker 1
So we've been talking about what you have called elsewhere defense, right? Protecting yourself, protecting ourselves from online fakes and scams, people using our likeness without our permission. I've also heard you talk about offense, which is— you know, choosing under certain circumstances to license our image and likeness for use on the internet. I like, I've heard you talk about that as this big, you know, underrated asset class that is going to blossom. Tell me about that.
12:28
Speaker 2
All right. So if you are a big gen AI company right now, you either have the choice of trying to do a bunch of one-off deals with celebrities, which is kind of painful. It's time-consuming. And there's a lot of red lines. And then most of the time they get stuck in the fact that like, well, who's watching you to make sure that bad stuff isn't being made? And the tech companies end up saying, well, I'll watch myself.
12:49
Speaker 1
There is an example where OpenAI for a while had this video generation app called Sora, right? They made this deal with Disney where people using Sora could use Disney characters, could generate videos with Disney characters. This is the kind of thing you're talking about, right?
13:06
Speaker 2
Yes. And this is actually a great example because that particular deal never really got off the ground because it took them months to figure out, well, how do the guardrails work for that?
13:14
Speaker 1
I thought it never got off the ground because OpenAI decided to shut down Sora and focus on other things.
13:19
Speaker 2
Well, they shut it down three months after. Right. So within the three-month period, they still didn't get a single character out the door.
13:26
Speaker 1
Uh-huh.
13:26
Speaker 2
And so a lot of that had to do with the guardrail conversations around, like, well, who protects what?
13:31
Speaker 1
Why is this a deal everybody wants to make? Like, what's the context here?
13:35
Speaker 2
Oh, yeah. Because people want to generate the real world, right? So I think that we are going to use this as a way to collaboratively dream. Like, let's build this story together. Yeah. Or let's build a product together or a company or something. And you'll be using Gen AI tools to dream those things with others. It'll be important in that context to use what the real world looks like. Otherwise, the fake version won't quite be the dream that you had in mind.
14:02
Speaker 1
So you'll... you'll want to make a little movie with a friend using generative AI, and you maybe will want to have a famous person in it, or you'll want it to be at whatever, Starbucks, or at some particular store that exists in the world. And if you're actually going to sell that or post it online and have, you know, have ads run in the middle of it, you're not allowed to do it. You're not allowed to have it be at an actual Starbucks unless you get clearance from Starbucks.
14:31
Speaker 2
Correct. Yes.
14:33
Speaker 1
So how, like, right now are your celebrity clients selling their name, image, and likeness, like, broadly? And are you working on that with them?
14:43
Speaker 2
Yeah. So you have kind of a bifurcation, which is half are like, no, nothing with AI. And the other half are like, it's coming and it's here and this is something we have to get ahead of. So let's make rules and set pricing that makes sense for me. And so the interchange system, the clearinghouse system works for both. One just blocks everything and the other one says, yes, you can make it with these rules.
15:05
Speaker 1
So are there examples that are public of people who have allowed their likeness to be used in generative AI?
15:11
Speaker 2
So like right now we're basically onboarding everyone. We have two different gen AI providers that are going to be using it. So the announcements for who you can generate where will have to come at a different point.
15:23
Speaker 1
Give me a sense of what that's going to look like. Will there be like lots of movie stars I have heard of who are saying it's okay to use my likeness? Is Gen AI subject to certain restrictions?
15:35
Speaker 2
I think you'll see a certain tier, for lack of a better term, we'll say like B celebrities, a large group of them. And you'll have kind of a sprinkling of a few A stars. And then you'll be able to just go to the Gen AI platforms and just type in this person's name instead of it getting immediately rejected. It says, okay, you can do that. But instead of it being $ 0. 30, it's $ 0. 35 online.
15:56
Speaker 1
Uh-huh. So basically you're saying that this platform charges you extra to use a real famous person in the video that you're generating and the real famous person gets some of that money and you get a little tiny bit of that money, presumably.
16:11
Speaker 2
Um, well, I mean, we, we actually don't charge our clients. We charge the gen AI platforms.
16:15
Speaker 1
Uh-huh.
16:16
Speaker 2
So if it's 5 cents, that, that goes to the client. And then we have a different mechanism of how we charge the gen AI companies.
16:23
Speaker 1
And what kind of restrictions to your clients will there be on the use of their likeness in generative videos?
16:29
Speaker 2
There's kind of like a belief system. So you have things like religion. And then more tangibly, it's things around sex, drugs, and alcohol. And then they're all natural language. So you can turn on drugs and say cannabis is just fine. Like I'm okay with this. And there's financial products. Like some people, they're consistently made to make kind of like Bitcoin scams. And so you can have that turned off entirely or say, yes, but it's only for this one app that I'm being sponsored for. And then the last part is literally the sponsorships, right? So you have, uh, you know, Nike is sponsoring me. So that always goes on a, on a sneaker or, you know, I did a deal with hot pockets a few years ago. So you can't ever make some kind of claim of me saying that hot pockets burn my mouth for the next year and a half.
17:16
Speaker 1
Yes. What if I want to make a video of a famous person saying that they love my shampoo? Uh, that I'm selling. Can I do that?
17:25
Speaker 2
No, we do understand the difference between whether it's commercialized or not. And this is where the defense part comes in, which is, let's say that you have them say, happy birthday, and you go start a car wash called the happy birthday car wash. There's no point when you generate it that I would know that that was what the intent was. But when it actually gets used out on Instagram, I will see, oh, this is used in a commercialized way. I'll look up your license and say, oh, that's That wasn't how it was purchased. This was a trickery thing. So then we go ahead and delete that. So they end up having to buy the correct version, but maybe they get away with it for an hour or two.
18:02
Speaker 1
We'll be back in just a minute. Are there any instances where deepfakes are legal? Or the First Amendment might be relevant?
18:22
Speaker 2
So this is tricky. I would say that no fakes is coming down the pipeline, which is a federal act. And that is going to require authorization from the person that you're deep faking. I think this is a really solid place for us to operate from. So between now and when that passes, and it looks like it's going to pass, there's really strong support on both sides of the aisles from tech companies and entertainment. It's kind of a massive collaborative multi-year effort, and it looks like it'll probably pass early next year. It's going to force tech companies to have to ask for authorizations from my neighbor or a celebrity down the street. Everyone is going to need some kind of authorization system for that. And between now and then, you can make a case that it's illegal for right of publicity reasons. If you generate a deepfake endorsing someone, that's like a Lanham Act violation because it's false endorsement. So there's the ways in which it's not about the tech.
19:16
Speaker 1
Well, the endorsement is the easy case. I mean, the one I'm asking you, like, say you do a politician, a satirical video of a politician.
19:23
Speaker 2
Right.
19:24
Speaker 1
Like, yes, we know that satire tends to have more First Amendment protection. We know that discussion of political figures tends to have more First Amendment protection.
19:34
Speaker 2
Yes. Right.
19:34
Speaker 1
Like. I don't know. I can understand why you could make a case that a deepfake is not like a political cartoon, but you could make a case that it is.
19:44
Speaker 2
I think I would probably more make the case that it is. I think there will be protections there. I know that when we talk to both sides of the aisle on who we protect, making sure that we get First Amendment right is a big concern for everyone. And us included.
19:59
Speaker 1
I mean, it's a hard one. It's a hard one because.
20:03
Speaker 1
As a matter of sort of public interest, you want people to understand what is real and what is not real.
20:08
Speaker 2
Yes. Right?
20:09
Speaker 1
Like, I like satire. I'm a reporter. I like the First Amendment. But, like, people being confused about whether the president is really saying something or it's a deepfake of the president, that seems quite bad.
20:20
Speaker 2
I do think that it's going to be. So when you did a political cartoon before, it was in the political cartoon section. of the newspaper or The Economist or whatever.
20:28
Speaker 1
And it's a drawing. Like, no one is going to be confused. Yeah.
20:31
Speaker 2
Well, I think that there'll just be kind of a time and place in the internet for that, which is there'll either be a certain part of the site or it will be labeled clearly on the video. Right? There'll be a restriction.
20:41
Speaker 1
Yeah, like a label. A label is good.
20:43
Speaker 2
If you labeled it, it would be just fine.
20:45
Speaker 1
Is that functional? Like, will that work?
20:48
Speaker 2
Well, this is This is where, you know, and not to harp on us or to focus on what we do, but anyone really that wants to build a business like ours, what you would say is what the time it goes to generate it, you'd say, oh, I see that you're actually making a satirical thing. That doesn't cost you anything because that's your first amendment, right? So we're not going to charge you. You get a bypass on it. But then when we see it out in the wild, if you don't have labeling on it, or if you're not using it in a satirical way, you're using it to like, say, look at what the president said, right? You're trying to like make a stance on it and trying to, you know, push it off as being real. then we'll do a takedown on it. So there's a way that you kind of need both systems to work well.
21:24
Speaker 1
That makes sense. You mentioned the federal legislation. I've heard you talk elsewhere about state laws as well. Is the federal legislation going to sort of preempt or render irrelevant the state laws? Are there state laws that are important that are doing other things that are worth talking about? I mean, I seem to recall California being kind of at the vanguard of name, image, and likeness stuff in the 20th century because of Hollywood, basically, right?
21:49
Speaker 2
Yeah. And then, oh, Tennessee is right now with the Elvis Act.
21:52
Speaker 1
Huh.
21:53
Speaker 2
They've made the No Fakes Act essentially for the state.
21:56
Speaker 1
I see.
21:56
Speaker 2
So if I have a client in Nashville, I have a lot of power for them to go and get things taken down. If someone made just a benign someone walking down the street in AI, if they live in Tennessee, it gets taken down. If they don't, I need to make a pretty good case about why this harms that person and not use AI rules to do it.
22:16
Speaker 1
Oh, interesting. So just to be clear- Outside of Tennessee, which is to say in most of the country at this point, if I make a deep fake of whatever, George Clooney, and he's just like, whatever, reciting Shakespeare or something. And, well, is that okay?
22:36
Speaker 2
Yeah, I mean, it would be allowed. Like, we wouldn't be able to necessarily take that down.
22:40
Speaker 1
If I sell ads against it, if they're just like programmatic ads that show up in the middle of it, now can you take it down?
22:46
Speaker 2
That's right. It's a violation of right of publicity.
22:48
Speaker 1
Because I'm making money off it.
22:50
Speaker 2
Right, off of their likeness.
22:51
Speaker 1
But if this bill becomes law, then it won't be okay anymore.
22:55
Speaker 2
Right. Those people will be able to pick what gets made of them. With a 1A carve-out, right?
23:00
Speaker 1
The First Amendment carve-out, potentially.
23:01
Speaker 2
Yeah.
23:02
Speaker 1
It's interesting that it's okay to do it now. I think that is not intuitive. Yeah.
23:07
Speaker 2
Well, I don't want to encourage people to do it. Um, you literally, like these people are upset that you were stealing part of their private nature, right? Because the people are also doing it for people that aren't celebrities. Um, you know, doing it, there's deep fakes made of, uh, teachers saying crazy things and then them losing their jobs. Like there's still harm, right? That comes from it. So I don't want to really tell people go on the boom rush now and go make it while it's legal. Um, you still are creating harm, even if it may not be a law.
23:34
Speaker 1
Is there a universe where the big platforms just do this themselves? Like, why when I sign up for whatever, Instagram, can I not just say, no deep fakes of me? Why do we need an extra company?
23:46
Speaker 2
So every platform, most platforms, and we're good partners with them, so I don't want to be disparaging about this, but they've tried over the last few years to do this, and there's issues with doing it. One is, it's hard for them to set up similar rules across all systems. Two, it's hard for anyone to believe that they will properly gatekeep themselves.
24:04
Speaker 1
Because the platforms want more content. They're not incentivized to take things down. Right.
24:09
Speaker 2
And so you're like, hey, you as the gatekeeper, can you generate things? And then if you're a celebrity, it's even more complicated because you're saying a biometric release is needed for that. And most humans do not want to have Google give a full biometric release because they know everything about you, right? You also don't need to know what you sound like and look like. And these large tech companies have been sued for face recognition usage in the past because of how they have this weird interstitial data set of like everything you've ever gone to online and what you look like in real life. Places like Illinois and Texas have sued them. So platforms aren't in a great place to say, I'll be the person that steps up and does this because the technology is largely handicapped for them because of this past litigation. And then the other part is people don't want to trust them with the biometrics. Those releases are pretty insane to think about. For us, like this is our only job, right? It's easy to give Lodi those biometric releases and Because all we're doing is protecting your biometrics.
25:01
Speaker 1
I've heard you talk about user agreements as weirdly important in this broader conversation. Why? What's going on there?
25:11
Speaker 2
There's a land grab for AI rights right now, for likeness rights. So if you're a tech company, it doesn't matter. It's not even the big ones, even small startups. They're trying to figure out how to get as many people as possible to do releases on their data, their likeness, their biometrics, because that data is really valuable.
25:27
Speaker 1
And like if I'm just a normal guy with an iPhone, like when I do unlock my phone with my face, am I participating in this without knowing it?
25:36
Speaker 2
No, but actually Apple, if you use an iPhone, they are aware of the same kind of problem sets that we are, which is they make sure that that's all local to the phone.
25:43
Speaker 1
Oh, interesting. So like Apple doesn't know what my face looks like. It's just my phone that knows what my face looks like. And that's good.
25:50
Speaker 2
There's a world that's coming in which all that's going to get legislated.
25:53
Speaker 1
I see.
25:54
Speaker 2
I think that's the next thing after no fakes. Now we're all going to decide what constitutes the data that goes into making a fake, right? I think it's coming and it's probably going to hurt a lot of smaller companies, but the bigger ones now are starting to see that is what's on the horizon and they might as well just design their program accordingly.
26:11
Speaker 1
Yeah, I mean, often regulation... ends up being a moat for big companies, right? It is essentially a barrier to entry. It makes it more expensive to do things. And so ultimately, it's good for the big incumbents.
26:23
Speaker 2
That's absolutely true. You have that plus regulatory capture. And it's definitely something that we'll see in the AI industry soon.
26:31
Speaker 1
Interesting. Let's talk more about the future. We've talked about legislation. We've talked about technology. We've talked about rights. But like, When you think, uh, you know, about whatever amount of time it makes sense to think about five years out, I don't know, 10 years, who knows what's going to go on in 10 years, right? It's so crazy. It's so fast. Um, first of all, what do you worry about?
27:01
Speaker 2
Man. Um, I worry about us being stuck in 2024 art and culture for a long time because of how machine learning is trained that, uh, You know, you basically have the corpus of human creativity that built models a couple of years ago that we all started using. And then everything that comes out of those models will be like what the past thousands of years of creative look like. And what we won't have is something that kind of jumps out. A lot of machine learning is based on training on the top of the bell curve of items, like what looks like the most, like 50th to 75th percentile.
27:36
Speaker 1
Sort of the average, the average. Yeah.
27:39
Speaker 2
So you have the average of 2024 that's coming out of models now. And you have to have something that looks truly unique. And AI doesn't want to make that. It doesn't know how to. And that's where the crap lives on one side. And that's where like the really amazing stuff lives on the other.
27:53
Speaker 1
Of the bell curve. You're talking about sort of the tails of the distribution, things that are really bad and things that are really good.
27:58
Speaker 2
Correct. Right. So you have like this mathematical preventative thing that says, hey, I don't know how to do math that allows the really cool stuff in without letting the bad stuff in. So I'm just going to go ahead and say all this big stuff in the middle is what I can produce. So the fear I have is that we've kind of walked down this path not really understanding that this is a feature of the mathematics. It's not like someone at OpenAI is making the decision. It's that they are using math that is forcing this decision. And then it's in 10 or 20 years and we're basically generating the same kind of content because nothing can really like pop out without getting hammered down by either algorithms or or by the generative models themselves.
28:36
Speaker 1
Well, if that's true, there's a weird, happy, a silver lining to that story, which is that's an opportunity for talented humans, right? If it's true that AI can only generate kind of mid-creative work... That's good for people who are weirdos and make things that are maybe really bad or maybe really great or a few people love even though most people don't. If that's the case, that's actually a happier outcome than I'm worried about.
29:05
Speaker 2
And I have a 15-year-old, a 13-year-old, and a 9-year-old, and that's what I tell them, is that the world is going to get really average quickly, which is going to be good for you because you're all very weird. You kind of have to figure out what your voice is. And in a world that wants to collapse you down, that wants to hammer the nail that sticks out, just know that your competitive advantage is that you won't.
29:26
Speaker 1
I mean, I feel like there's a universe where you can play with the AI and get it to make stuff out in the tails, right? You can use AI to make weirder stuff by kind of iterating, right? And to what extent you need a person doing that, you know, like to what extent it's just one shot where it's like make something really weird and different in this way versus to what extent it's like, well, I write something and then I ask the AI about it and it gives me some ideas and I write some more. is to me kind of the open question and a meaningful open question because the less you need of me, maybe the worse at some level.
30:05
Speaker 2
So this is where I get optimistic about AI is because I do think that humans can push AI to make something extraordinary, but it's the difference between being the painter or the paintbrush, right? Like if AI is the paintbrush, you're good, right? But as long as you, as soon as you say, well, actually AI write a script for me or paint the painting for me, Um, and it becomes more of the painter. You kind of are screwed. You're in that spot of like, it's going to make something that's average, but you're right. If you're a human that's involved in the process and you kind of push it beyond its boundaries, you are then taking it out of like the architect role and more into the, like the engineer, like just do this thing. I know it looks weird or sounds weird, but do it anyway. Um, that's like taste, right? That that's human creativity getting imbued into the AI product are very bullish on that. I think that that's where AI will actually help our species. I just don't like the idea that some people will say, well, they'll be writing movies soon. Like, hopefully not. Or they'll be very mid. Or they'll be writing music soon. Hopefully a human that wants to make a movie or wants to make a song says, like, there's something that is so out of this world I have to use AI for that clip of it. That's cool. I'm all for that.
31:11
Speaker 1
Yeah, I mean, I hope so. You used the word taste in there, which is a word people have been using a lot, right? People have taste. I don't know. That's like an interesting big... conversation about the future of sort of human creativity and AI. If we think more narrowly about the business that you are in, you know, people's own likeness in this world of exploding AI, what do you worry about in the, you know, next few years?
31:41
Speaker 2
I would say the problems I want to solve for are making sure that there's very low friction so people can control their likeness, whether you are a celebrity or you're like me and you're just some average guy walking around I think that that's a difficult problem to solve. Like, how do you do that at scale? Which friend of mine is allowed to generate me as well as a weird problem?
32:00
Speaker 1
I'm going to just say none.
32:02
Speaker 2
Yes. You might say none or, or what you might actually find out is, um, you have a parent that is ill and they're like, actually, um, I do want to make sure that even after I'm gone, I'm reading books to your kids. Right. And then you're like, okay, that's an exception.
32:18
Speaker 1
Right.
32:19
Speaker 2
And then you'll end up finding these kinds of outliers where you're like, can my partner make something? They had a dream of me and they want to like show me the exact dream they had of me. And it's going to be cathartic for them. Yeah, sure. I don't mind if my partner, I trust them with everything. Might as well trust them with that. I think there'll be slivers of things. You're going to have like high school kids that'll be able to give their likeness because their friends always love to create crazy stuff.
32:38
Speaker 1
Right.
32:39
Speaker 2
Um, And you need good guardrails around that too. And so I think there'll be pockets of things that people that even traditionally would be staunch, be like, I'm not going to do anything with AI, will say, well, actually, I think that that's cool. And then it'll just grow from there.
32:56
Speaker 1
You do seem optimistic. So give me your happy vision of the world five or 10 years from now.
33:01
Speaker 2
The happy vision is one in which we don't ever use AI as the painter. It's humans that have figured out how to push the boundary of creativity with this new kind of paintbrush. That's the first part. And then the second part is people don't see the internet as this big wild west of anything can get uploaded of you whenever, wherever, and can destroy your life overnight. And so I think both of these things, like the privacy component and the control and autonomy and agency, hopefully will come to the internet. And then the second part, of course, being creativity, getting solved for, but solved in my mind is more paintbrush and less painter.
33:41
Speaker 1
We'll be back in a minute with the lightning round. Okay, let's do a lightning round. What's your favorite deepfake?
34:01
Speaker 2
The bunny rabbits on the trampoline. Totally harmless. It got a million people laughing and enjoying and seeing that AI could make some funny, cool stuff and tricked a lot of people into thinking they were real as well. So I actually did like that.
34:13
Speaker 1
I'll have to check it out. I haven't seen it. What celebrity likeness would you license to advertise for Lodi, to be a spokesman for Lodi?
34:23
Speaker 2
Cannot say. Yeah. It's coming out actually in the next few weeks, so cannot say.
34:29
Speaker 1
Okay.
34:29
Speaker 2
Yeah.
34:32
Speaker 1
I'm very interested to see. Is there one or will there be a bunch?
34:37
Speaker 2
There's one that's on the forefront of all of this. Okay. Someone that we deeply respect in this space.
34:44
Speaker 1
I'm very interested. How do you want your kids to use AI?
34:51
Speaker 2
I told my 15-year-old that he could probably cheat on everything with AI. And that there's probably no way that me or a teacher could figure that out. Because I know that kids are really crafty. So that he has to make the active choice that if he wants to be successful, he has to do hard work. And that's a completely different skill set. So it really is coming down to just, especially if you have a smart kid, just tell them, look, I know you can do it. But here, let me make the case for why you should work harder instead. Is if it's in your head, you could then leverage it to do things that other people cannot. So it's a little competitive, but It's essentially to tell my kids, don't shy away from it. Like my 15-year-old's on the $ 200 a month cloud plan. Like he uses AI a ton, right? He's like building apps and stuff. It's crazy. I'm really proud of him. But at the same time, he knows that you can't use it to fill out all your homework without you robbing yourself of something.
35:42
Speaker 1
Yes, right. The point of homework is actually not the whatever, the math you're learning. It's learning to just sit there and do it, right? That's actually the work.
35:53
Speaker 2
Yeah, those are the... So a lot of our discussions around AI are actually the reasons why you wouldn't want to use it in a home that deeply uses it. Yeah. So it's, but I think that that's like the, the brake and gas pedal kind of thing. Like you need both to run the car.
36:06
Speaker 1
Yeah.
36:06
Speaker 2
And so a lot of the conversation is around the more complicated use of a brake system. What?
36:12
Speaker 1
I was trying to think, frankly, of a good question about like AI and the future of movie stars. Like, I don't know the first movie with an AI genre, but I couldn't quite get the right question. Like, make a prediction about AI and the future of movie stars.
36:28
Speaker 2
I think that there's going to be, there's going to be a demand for like authentic performances. That's what they'll be called. Something to that effect, some euphemism for it's really them in front of film. And I think they're going to make movies that portray that a little bit better. Like it'll be like human expressions that AI is not able to get or more complicated scripts that land in a way that's not like uncanny valley. This is a good example. There's the, The Silver Springs, Fleetwood Mac, where the lead singer and the guitarist basically are in this long-term situationship, and they're essentially singing about themselves on stage, and everyone can kind of feel that tension. It's a human thing. And if you were to swap them out for robots, you probably would not feel the same way as a human. I think there's something about knowing that that person suffered in the way that you've suffered that makes you identify with their art that is going to be much bigger, and we talk about Taste, I know, is a cheesy word, but still is something real here. I think taste for that will go up because it'll be more rare. And then I think there's going to be a bunch of superhero action movies that are made in AI and no one's going to care.
37:33
Speaker 1
I mean, superhero action movies are already largely made in AI.
37:37
Speaker 2
Correct. That's it. Yes. Yes.
37:40
Speaker 1
Like, it is interestingly not as binary as as we say, right?
37:46
Speaker 2
If your art is making really generic romance songs for radio, you probably are in for some competition with AI. But if you have a unique sound, unique bass, whatever it is that you bring to the table that is truly not like something else we hear everywhere else, the demand for that's going to go through the roof.
38:05
Speaker 1
That's encouraging. I appreciate your time.
38:08
Speaker 2
Of course.
38:16
Speaker 1
Luke Garagoni is co-founder and CEO of Lodi. Our show is produced by Gabriel Hunter Chang and Trina Menino. Our editor is Lydia Jean Cott, and our engineer is Sarah Bruguier. I'm Jacob Goldstein, and we'll be back next week with another episode of What's Your Problem?
The Host
