David Friedberg: Okay, everybody. There's a little bit of backstory to this episode. I actually taped this a couple of months ago. And we taped this episode because we had read a New York Times article called The Secret of Company That Might End Privacy As We Know It. I know it's a dramatic link baiting headline. And the story was about a company called Clearview AI. They do facial recognition. And they sell that software to government agencies and police departments. And they do that, obviously, with good intent to try to help catch criminals. Well, the morning

SPEAKER_02: I interviewed him happened to be, in terms of timing, the day after George Floyd was tragically murdered. And I'm using the term murdered because it felt like murder. It is murder. And it's

David Friedberg: unacceptable. I think we all understand that. So once the protests started in America, and we were watching these anti-racism protesters, we decided we might hold the interview because it didn't feel like the right time. And maybe things would settle down and people could think about the software as something that would theoretically help police departments, as opposed to maybe helping them do something like identify peaceful protesters. And so we've really been thinking about when is the right time? Well, just this past Sunday, last night, and this was August 24th, 2020, a 29-year-old black man named Jacob Blake was shot in the back and tasered by police officers in Wisconsin. He was unarmed. And he remains in critical condition. And we're praying that for him and for his family

SPEAKER_06: that he pulls through. And we saw it. And I tweeted about it. And listen, this is coming from somebody

SPEAKER_02: who's got a family in law enforcement. And I have great respect for the police. But this was not the way this should have gone down. And it's heartbreaking. And I'm now at the point where I talked to my team

David Friedberg: about it. It tragically feels like there's never going to be a right time to release this episode. And we really have a lot of work to do on race in this country. And we're dedicated to that just like you are. And we felt we needed to put this episode out for you to listen to it because it brings up a lot of issues. And in fairness to Juan, who's the founder, he appeared on the podcast. And in good faith, he came on and he answered very difficult questions for me. And I think I actually believe he has good intentions. That's a belief I have. You might disagree. But I think it's a very important topic. Facial recognition is a very important topic. The world is changing. And we know every time one of these technologies comes out, there's pros and there's cons to it, whether it's GPS, which can track you or can get you to your location on time, or it's facial recognition, which could catch a criminal or be used to track a peaceful protester in a bad way and compromise people's privacy. These are important discussions. We're going to have them here. And I hope this, this episode is taken in the spirit in which I and the team at This Week in Startups that works very hard on this podcast intended, which is to discuss important issues. Black lives matter. We all know that. And I hope you enjoy the podcast.

SPEAKER_10: This Week in Startups is brought to you by LinkedIn Jobs. A business is only as strong as its people. And every hire matters. Get $50 off your first job post at linkedin.com slash twist. And Coors Light. When you want to reset this summer, reach for the beer that's made to chill. You can have Coors Light delivered by going to get.coorslight.com and finding local delivery options near you.

SPEAKER_13: Hey, everybody. Welcome to another episode of This Week in Startups. We're taping in late May 2020,

SPEAKER_15: somewhere south of Market Street in San Francisco, in a socially distanced empty office where my investment company launch once was a beehive of activity. And now I come to this office for the last two or three months, and it's empty. And there's a layer of dust that we clean off and nobody has been to work here in months. It's a very weird feeling. But...

SPEAKER_16: I know. It is a weird time, isn't it, Jason?

SPEAKER_15: It's a very weird time. And our guest today is Juan Tantat, who you just heard. He's in New York City. He's socially distancing, I hope.

SPEAKER_18: That's correct. Very distanced.

SPEAKER_15: Juan Tantat is the CEO and co-founder of Clearview AI, which is involved in facial recognition in order to help law enforcement catch criminals. The company has had a little bit of, let's call it, controversy or notability around the technology and the fact that it exists. My opening preamble, this technology has existed for a long time. Facial recognition and anybody who's involved in a crime or a victim of a crime, I should say, would very much like to have the perpetrator caught. And if they were caught on a CCTV, a closed circuit television system, which we have all over the place now in cities. London has, I think, more CCTVs than people. You would want that person caught and brought to justice. However, we all know that any system that can be abused will be abused. And you don't want people having access to a tool like this who might be a stalker or might want to harass somebody. And certainly the same technology used by a group of people in an authoritarian country, let's say China, could be used to round up hundreds of thousands and millions of people in a certain demographic to be sent to re-education camps.

SPEAKER_21: A re-education camp in China might be viewed by the majority of the Western world as more akin to a prison

SPEAKER_22: or a concentration camp where people are tortured and where people are starved and where people are abused because of their views of the world or something as simple as religion. So with that backdrop,

SPEAKER_21: Juan reached out to us actually and said, hey, I'd love to talk. I guess you're a fan of the podcast and you wanted to come on and talk about your technology and what you're working on.

SPEAKER_15: I'm curious, you have some notable backers, Peter Thiel and my friend Naval. How much of the controversy around your company do you think is because of Peter Thiel's involvement? Because he's such a lightning rod and we've got such a polarized right and left kind of world right now that I think when Peter invests in something, it kind of can bring a little bit of lightning

SPEAKER_26: to the founder. How much of it is related to Peter Thiel being a bit radioactive these days?

SPEAKER_28: Thanks, Jason, for having me on your show and a very long preamble about the issues we are facing

SPEAKER_29: and a lot of the controversy. So Peter is a great investor. He's a super smart person. We all know that. But I think fundamentally there is something controversial about facial recognition, also privacy, the privacy debate, and its use in policing. So some of the misconceptions that are out there is

SPEAKER_30: this is going to be a tool that's used to without any oversight regulation or in a real-time way. So

SPEAKER_29: in China, they have a lot of real-time surveillance, but the way Clearview AI is used, it's after the fact investigative tool. So that's the misconception that a lot of people have. So when there's probable cause for a crime, so your car window was smashed and there's a person in surveillance footage and you don't know who it is, this is a tool to help get a lead. You still have to obey all the protocols that are in place. And when we think about our technology and how powerful it is, we always think about how best to apply it. So it has, you know, the best upside we can think of in terms of

SPEAKER_30: solving crime, but also minimizing the downside and abuse. So, you know, we're kind of living in the

SPEAKER_29: future. We have over 2,400 police agencies in the United States using it. And we've had so many crimes from child sexual abuse being solved, a lot of murder cases, financial fraud rings that are massive. And it's really our belief that the upside completely outweighs the downside. We haven't had

SPEAKER_32: any instances of abuses or people wrongfully arrested from the technology, which is the number

SPEAKER_34: one fear that people have. What is the, what does it cost a police department to have this? How do you charge them? Do you charge them a yearly fee? Do you charge them per police officer? Do you charge them by

SPEAKER_13: the density of the city? Do you charge them by the search? How does it work?

SPEAKER_30: Yeah. So right now it's a SaaS business, depending on how many seats or licenses a police agency wants. And it varies if they're federal, local, or state. But it's pretty inexpensive, especially compared to what's come previously. So previously, yeah, ballpark, you know, $2,000 a year per officer, all the way down to be more inexpensive. Got it. And you might only need to have one or two officers

SPEAKER_43: in a, in a police department to have access to this. You don't want all police officers to have

SPEAKER_46: access to this. You want it limited to like detectives or a very specific group of people

SPEAKER_29: that are vetting the searches, correct? Correct. So right now it's really used with detectives for after the fact, after the fact crimes. So they might be in a crime center. They might be at the financial fraud division of a police department. They might be, you know, investigators who

SPEAKER_43: use all different kinds of tools to build cases. So when a detective in, is it public which regions have it? Or is it part of like their spend, like that they report that they use this? I would have

SPEAKER_29: Yeah. There's a Freedom of Information Act. So people have to comply with that when it comes to

SPEAKER_13: the use of their tools. Got it. So what's an example of a jurisdiction using this and having

SPEAKER_29: great success? It's used to buy, you know, New York City? Yeah. So we've had people in the New York region use it to great success. Correct. So when somebody uses it in the New

SPEAKER_15: York region, let's just say somebody's using it in Albany or something. I'm just picking a place. Some town, New York City, New York. In some town, New York, a detective wants to use it. Now when they take a picture, they get a picture from a drop cam of somebody or an S cam or something, security camera, of somebody who broke into a house. They put it into the database.

SPEAKER_43: And you record that they did that search. Do they need to have a, in this case, do they need to have a warrant like somebody would need for getting somebody's phone records? Or there's no need

SPEAKER_58: for a warrant because this is all public data, correct? Yeah, correct. So we had one of the

SPEAKER_29: best legal minds, Paul Clement. He was Solicitor General under Bush. And he wrote a legal reading of the Fourth Amendment and how Clearview is used. So all the data inside Clearview is publicly available on the internet. So it's like a Google search for faces. You put in a face, you get a lead. It's like

SPEAKER_30: Googling someone's name, right? You have to make sure you get the right person. And then you build your case. Then you go to the judge and say, I have all the evidence that this person was here. He lives in the New York region. This is a car that was stolen. And you now have his name. Can we go forward? So it's before, you know, anything's open. Which is what a detective would do. They

SPEAKER_43: would take the screen grab of that person breaking into the house and they would go to the local bar, they would go to the local cafe, they would go to the local grocery store, take out the picture and say, hey, you're the bartender. You're the checkout person at this coffee shop. Have you ever seen this person? And the person says, yeah, that guy comes by every Tuesday. Exactly. So it's,

SPEAKER_66: it's more, much more efficient than going around and it's much more accurate. So the tool is so much

SPEAKER_29: more accurate than the human eye now. So you reduce all these people that might be pulled aside. We've had one really interesting story from Alabama where they were looking for African-American woman, maybe in their thirties, who beat up a grandma. And they were able to look at the surveillance footage, run the photo. And they actually found mugshots of African-American male. Turns out they were dressed up as a woman and they stopped them from going through and pulling over instead of, yeah. So it's

SPEAKER_30: the opposite of what people expect. It's so accurate that it would stop them from, you know, interrogating five or six African-American women of that age. So I think that the accuracy is one of the selling

SPEAKER_29: points. Um, because it's not just about locking up bad people, but also making sure that the

SPEAKER_15: innocent, uh, are not wrongfully detained or arrested. And how would you prevent a police officer, a detective rather in this, any town, New York from taking a picture of their, I don't know, ex girlfriend, ex their, their, their ex wife, whatever it is, or just somebody they wanted to

SPEAKER_43: harass a friend of theirs, ex wife or something, and putting it into the system and saying, show me every picture of this person. And then finding out, Oh, this, uh, woman who I used to date was, is now in a photo on somebody else's Instagram that you scraped or Facebook you scraped or some other public photo, and then using that to harass them. How would you prevent that from happening? Just like we see in every detective novel or every crime procedural television show, you know,

SPEAKER_21: somebody's like, Hey, I know a detective. I know a private investigator. I can get them to run that license plate, right? Like the running of the license plate for a friend or for some mafia guy who leans on a cop to run a license plate. How do you prevent that from happening with your software?

SPEAKER_58: Yeah, it's a great question. And our goal is to get the best out of the technology and minimize the

SPEAKER_29: abuse completely. So for each police department that uses it, they are nominated a, um, someone who

SPEAKER_30: audits the logs of every search, and they can opt to say every search must have a reason or a case number with it. And these, that kind of transparency where the police officers, they know they're trained. These are the searches. This is what you can use it for. There's not evidence in court. And by the way, please get us an administrator. So they, so they can, so they're aware that there's

SPEAKER_29: an audit log. Absolutely. And I think that's the key thing here is to, uh, you know, build systems that are secure and can be audited. And that just knowing that exists can make everyone feel at ease

SPEAKER_75: in terms of, yeah, you would be less likely to run the license plate if you knew that it was tagged to your login. Now, of course, the issue would be is if there's a shared login, and a bunch of different people could use it, maybe somebody could sneak it in and say, I didn't do it. Yeah, yeah, but there's also

SPEAKER_29: ways to get around shared logins. All these logins are from the same IP or device, we can detect that eventually, uh, I think that because of the power of our facial recognition, there'll be, we, you know, how you have two factor authentication, which we do have now for our service, there'll be three factor, you know, your email, your text, but also your face. So I think that, so the person doing

SPEAKER_43: the search would then have to turn on the webcam, have their picture taken to say, I am a detective,

SPEAKER_83: I'm detective Calacanis, and I'm doing the search and my picture is taken by the computer that does

SPEAKER_29: the search that we haven't, we don't have that yet. But, uh, pretty great. I think that's going to be the future of a lot of authentication, because you can have account takeovers with SMS and email, but it's hard to take over a face. So, uh, even the SMS is pretty hard to do. I mean, let's face

SPEAKER_75: it, like who's giving it is hard away. Yeah, but you know, people do get, people do get targeted,

SPEAKER_29: they get, you know, the call to Verizon. And actually, we see this a lot, we have a lot of financial fraud detectives. And one of the cases we helped solve with was 35 million in fraud that was recovered, uh, between this agency and a big bank. And it was 19 people and they were getting the photos of the people stealing identities from the ATM. So I would go to the criminal would go to a bank and say, hi, I'm Jason Calacanis. Um, this is my social, this is my phone number. Uh, and they'd say, all right, show me your ID. And they would make a fake ID, but that have your face on it. Right. And so they would steal the ID, withdraw all the money from the bank or the ATM. And these fraud rings are massive. And we were just shocked that 35 million fraud, that's just one case that we know of. And it's all after the fact. So when we kind of think of, uh, the impact of the

SPEAKER_28: technology and the, uh, ramifications of all this, imagine if that's a deployed at scale everywhere.

SPEAKER_43: Do you have a central log file? In other words, I hear that you have like an ambassador on the local police force or a, a budsman as it were, an auditor, but do you keep a log? Like in other words, if that any town Albany did these searches, can they, is that log file something that you maintain as well? So if they try to alter their log file, there's some conspiracy locally, you still have the backup to it, or do they maintain their life? Right now it's a SAS service on our service.

Chamath Palihapitiya: It's very hard to deploy on-prem because we have billions and billions of photos, but we do not

SPEAKER_28: look at the logs. It's up to the agency to enforce their- So you don't keep a log of their,

SPEAKER_75: what their usage is? Yeah, we don't look at it, but it's- Well, no, but do you keep it is what I'm

SPEAKER_43: saying. Do you keep, do they, cause then that would be a two layer of it that the local police department knows, and the police officers using that system know that there's a local level. But then if there was, like we've seen many times in the United States, there's a conspiracy at a local level where DAs and police are in cahoots, then they would know, hey, wait a second, Clearview has a log of everything. Do you have a log of everything in case there's local abuse like

SPEAKER_29: that or not? Yeah, it's all on our servers at this point in time. Right, so you would know. But we don't, it's not our job to really police the police. Right, but a judge, if a judge came and

SPEAKER_43: said to Clearview, hey, listen, we've got a dirty cop and a dirty prosecutor in this region, which has happened before, where they railroaded people. Yeah, we need to see the Clearview logs and see what they did. You would be able to produce that. We would comply with any legal

SPEAKER_102: orders. We wouldn't be complying. All right, when we get back from this quick break, I want to know

SPEAKER_15: what does facial recognition still not do right? What are the false positives? And what are the

SPEAKER_23: things that have not yet been solved, if any, with facial recognition when we get back on this

David Friedberg: being started? All right, let's get down to brass tacks. I've got $50 to you from LinkedIn jobs. The market is coming back. I know it's really weird to say, but people are actually hiring because people, entrepreneurs are resilient and they figure things out. They figure out how to save their businesses. And I'm going to give you $50 towards hiring that next great hire. And we got an amazing testimonial from one of our listeners, Jay, who was the founder of something called 10 Golden Rules. It's a boutique digital marketing agency. And he used our $50 credit for the first job post, which I'm going to give you the secret URL for in a moment. Well, he used that and he put up an account manager position and he quickly received, and we hear these stories all the time

SPEAKER_109: about LinkedIn jobs, over 150 qualified applications, not 150 noisy applications, which makes your life miserable, qualified. And after identifying his two top targets, Jay noticed that he shared mutual connections on LinkedIn with both of them. So boom, he can go vet those hires. And that's the number one thing that I see founders don't do. They don't vet. They don't do reference checks. Those are built into LinkedIn. We know that he hired his favorite candidate over Zoom and it's worked out great. Congratulations to Jay. Be like Jay. There are over 690 million people waiting for you on LinkedIn right now to do that work, to join your team. I want to give you $50 right now.

David Friedberg: LinkedIn.com slash twist. That's right. Go to LinkedIn.com slash twist to get $50 off. Terms and conditions apply because they're giving you $50 because you listen to this podcast, the greatest podcast in startup history. LinkedIn.com slash twist. Thank you, LinkedIn,

SPEAKER_02: for helping me hire so many great people. I cannot thank you enough. It is the best.

SPEAKER_15: Okay, let's get back to this amazing episode. All right, Juan Tontat is with us. He is the CEO and co-founder of Clearview AI. You may have read about them in the press. And the New York Times piece in January of this year, 2020, the secret of company that might end privacy as we know it, kind of

SPEAKER_113: painted you out to be a bit of a Bond villain. What was the response to that piece like? And

SPEAKER_43: do you feel the New York Times piece was sensational, fair, unfair, and in what ways?

SPEAKER_102: Thanks for the question. It's a very interesting process going from a company that had no media

SPEAKER_29: exposure to being on the front page of the New York Times. And it's an honor to be at the center of the debate now and to talk about privacy. So at the end of the day, we're very honored to be at the forefront of this debate. And I think that regarding the New York Times, they were actually extremely fair. When we're going into it, I wasn't very sure of what to do and how to respond. But we ended up responding to Cashmere Hill and her questions and engaging with it. And we thought we made the story a lot better off, we could show her a lot of examples of a lot of success stories from Indiana State Police, FBI, Homeland Security, etc. So it's totally worth engaging with the media. And since then, there's been a lot of controversy. But fundamentally, this is something that's such a great tool for society.

SPEAKER_15: And, you know, you kind of glossed over the question I asked you at the top, which was,

SPEAKER_103: how much of this blowback do you think is because Peter Thiel's name is associated with it, obviously him being a controversial character? Do you get a little blowback from that? Were you a Thiel fellow?

SPEAKER_28: No, I wasn't a Thiel fellow. How did you meet Peter Thiel? I met him in Silicon Valley. But I think that it's a small part of the controversy around it. I

SPEAKER_29: think it is a lot more around the privacy and the law enforcement part of it. So think about it as it was a shock to some people. And I understand that as well. But overall, like it's been

SPEAKER_103: But he was a seed investor, he put 200k into the company. It's not like he's on the board of the company, correct?

SPEAKER_121: No, he's not.

SPEAKER_43: Right. And then you raised another, you've raised about 7 million from investors. Any of those like VC firms, like a proper VC firm or no?

SPEAKER_28: Yeah, we have some institutional investors for the Series A. Who led it?

SPEAKER_29: Yeah, it's a firm here in New York City. And we've also had...

SPEAKER_125: Which one? They don't want to have their name out there as an investor or something? I think it's a pretty savvy investment.

SPEAKER_29: Oh, thank you. We're very thankful to all our investors. They've been very supportive

SPEAKER_30: behind the scenes. And they really believe in the mission of the company, which is to reduce crime and fraud all across the United States.

SPEAKER_129: Wait, did you mention who the... Is it Keurig? Keur and Naga partners?

SPEAKER_48: Keur and Naga were part of the Seed and the Series A. Got it. They're the lead. Yeah. So what was interesting about all the feedback is...

SPEAKER_131: I've never even heard of that firm.

SPEAKER_29: We have two different storylines here. So there's a lot of the things that are in the public around the privacy, which people forget this is all publicly available information. This is anything you can find in a Google search, right? There's nothing controversial about Google.

SPEAKER_15: No, but let's answer the question about what facial recognition hasn't figured out yet.

SPEAKER_43: There's been a lot of... Race has been inserted into the facial recognition discussion because certain instances of facial recognition software were better at identifying certain ethnicities than others. I'm curious, being a neophyte in this, is there an actual issue where Irish people look more similar to each other than, say, Italians? And that's actually an issue for facial recognition? Or do you, in your estimation, as an expert on this running a company on it, was the issue that the people who made the first ones were white males in Silicon Valley who didn't take into account

SPEAKER_134: maybe the characteristics of other ethnicities?

SPEAKER_29: Yeah, what I would say now, and I'd love to put you through a demo, so you can see the software,

SPEAKER_30: is that we've created a technology that is way more accurate than anything before. It's better than the human eye. You can search out of billions and billions of photos, and it picks out the right person from different angles, with beards, with glasses, and we made sure to train it on every ethnicity. So that's some of the problems that...

SPEAKER_21: To my question, what was that initial problem set, do you think, when you look at it? Was it that

SPEAKER_43: the people building Microsoft or Google's or whatever companies it was just had a blind spot to an ethnicity? Or is it, in fact, that Irish people all look the same?

SPEAKER_29: No, everyone's unique. That's the funny thing. There's 7 billion or 8 billion people in the world. Everyone's face is unique.

SPEAKER_43: Are some races more unique than others, in other words, I guess is the question.

SPEAKER_29: I don't think so, unless you have an identical twin. Everyone has a unique face. And what happened, I think, is the earlier companies, in the super early days, they would not even use neural networks. They would actually just measure the distance between the eyes in a manual way. And then you had the neural network phase. And what we did is we made sure that the training set had people of all different races in it. And that was part of why ours is even more

SPEAKER_30: accurate. Because a lot of the training sets that people get are just celebrity training sets.

SPEAKER_48: So they're not fully representative of the whole population. So I'll just show you how accurate.

SPEAKER_15: Wasn't there something with Face ID for Apple where I think Asian people, Asian folks were able to

SPEAKER_142: unlock each other's phones? Wasn't that like the, yeah, the weakling? I can't remember the story.

SPEAKER_48: That was a headline.

SPEAKER_97: That was a headline. Yeah. Was it an accurate headline? Or was it just like, I don't know. But I think I've used Apple ID. And

SPEAKER_48: it's just been pretty accurate.

SPEAKER_43: Well, the question is, if your friend who was also Vietnamese used it, would they have a?

SPEAKER_48: Yeah, well, my sister doesn't unlock my phone. It's pretty good. So I can just give you a demo. So I think what really happened is we basically solved the issue of accuracy. Because that's been

SPEAKER_30: something that's been an issue before in the past, but we've got kind of like broke the sound barrier. So I'll show you a demo right now, Jason. Great.

SPEAKER_29: And just share it by the screen. And people can see how it works. So here is how the web version of Clearview works. And you can just upload a photo. So I took a screenshot of you from before.

SPEAKER_30: There you go. There we go. All right. So for people who are listening, he's doing a search

SPEAKER_15: on the Clearview.ai website. He uploaded a picture of me in a leather jacket looking like some Irish loan shark from Hell's Kitchen. And she found the photos of the, somebody made business cards. And check this out. There's a picture of me for Phil Hellman's home game when I played at Phil Hellman's home game and they put me in a caricature. So there's, there's a bunch of photos of me.

SPEAKER_151: 565. And we'll find some. I have a public figure. So it's. Yeah, there's no false positives.

SPEAKER_21: There should be. Yeah, it's kind of hard.

SPEAKER_104: And then you can click on the other person if you don't remember who they are. Right.

SPEAKER_21: So this would be something I would love to have is the ability to zip through and see these photos. But I can do this. Can I do this on Google reverse image search? Or is that not doing facial recognition? That's just doing the specific image?

SPEAKER_36: Yeah, they're doing exact image search. Exact image reverse search.

SPEAKER_29: If your face is in a different angle or your face is.

SPEAKER_15: Is there a public, is there a public facial recognition service that I can use? Does Amazon provide one?

SPEAKER_29: All Amazon provides is an API called Amazon recognition. So you'd still have to, you'd have to try it yourself. I think it works pretty well, but it doesn't have a data set.

SPEAKER_163: Well, I mean, it's your direct competitor. I mean, does it, does it work?

SPEAKER_29: Well, they're not really. Yeah, they're not really a direct competitor. Like I said, they just sell an API. They're not selling a product to law enforcement. So. Got it. So they, they developer, you'd have to hire a developer.

SPEAKER_165: But if you send them a photo, they will then look for photos on the web that they've scraped?

SPEAKER_28: No, they would only look for photos in a database you would provide.

SPEAKER_83: Ah, so that, that's the key difference is you have the database.

SPEAKER_29: So basically everyone in the space previously has had matching software. They sell as an API. So it's very hard if you're a user, like in law enforcement to just get started.

SPEAKER_02: So let me ask you a legal question here. Yeah. And I'm not a lawyer, but I'll ask a, I'll, I'll ask it anyway. And we'll, we'll kind of

SPEAKER_113: work backwards. So I publicly put that photo, uh, that you're showing of my incubator class

SPEAKER_43: by the golden gate bridge on my blog. You have it in your database. You never asked me permission to have it on my database. It's a, it's a copyrighted photo on my website. So if I ask you to remove it, will you remove it from your database?

SPEAKER_82: So we have, uh, we comply with the...

SPEAKER_43: And then why didn't you ask for my rights to it?

SPEAKER_28: Well, there's two parts to that. So there's a fair use. I'm not, uh, um, speaking completely as a lawyer here, so I'm not going to too much, but this is the thing is fair use. And this is in also, it's also in Google search engine.

SPEAKER_83: Okay. But just because Google has, it doesn't mean you get to have it. So do you believe that

SPEAKER_46: fair use gives you the ability to take my photo and put it in your database that you took, you

SPEAKER_29: scraped from my site and then same with Google and all the other search engines. If you were to say, uh, remove that, uh, then there'd be no Google. There'd be no Bing. No, I know, but, but do I have

SPEAKER_102: the legal right to ask you to remove it or not? So yeah, we do comply with all privacy laws that

SPEAKER_13: what if I asked you to remove all the photos of me that weren't even mine? Am I allowed to do that?

SPEAKER_28: Am I allowed to opt out of your service? Yeah. So we do comply with CCPA, GDPR and other,

SPEAKER_29: you know, laws. What is that? CCPA, California Consumer Privacy Act.

SPEAKER_15: So anybody in California can email you and say, here's my name and my photo. I want to have all photos, all 500 of those photos removed. Yes. We want to be compliant with every law that's out there. So, so a savvy criminal knowing that this existed could send you a letter and say, remove all my photos or just somebody as a privacy person. Do people do that actually in reality? Do

SPEAKER_177: people? Yeah. There's some people who do some people who do. Yeah. Yeah. Oh, you hear that sound.

SPEAKER_108: You know what time it is. That's the sound. It's that crisp course light. I can a crisp course light opening because you've been on five, six, seven, eight zoom calls a day, just like me, and you're losing your mind and you need to relax. And you need to relax with the crisp, cold course light. You close that laptop, you close your Dell laptop and you just sit back. Maybe you put on a Netflix, a Disney, watch the Mandalorian for the second time. Maybe you hang out and chill with your friends, socially distance, and you crack open that Rocky mountain, cold course light. Yes. You know, born in the Rocky mountains of Colorado in 1978 course light is refreshing, crisp, and only 102 calories. Very important for me because you know, I've been hitting that Peli Peloton and I'm trying to lose weight. So I got to go with the course light. It's so crisp. It's so delicious. I tell you, I work so hard. You guys know how hard I work on this podcast and in life. And then that five o'clock whistle, six o'clock whistle blows, just close the laptop and I crack open a cold one. It's brewed at the ice cold Coors Brewing Company in Golden, Colorado, where they were made to chill. So close that laptop and chill out with a crispy Coors light. There's no doubt. Summer's totally different. We know that. It seems like everything's been canceled. You know what hasn't been canceled? That crisp, cold Coors light. So go ahead. It's okay. You do a social distance hike, you come back and you crack open that crisp Coors light. And you can even get Coors light delivered right now. Just go to get dot Coors light.com. So that's simple. G E T get dot Coors light.com. And you'll find a local delivery option like I did. And that 12 pack comes. Crack it open. You ready? You want to hear it? There it is. Coors light mountain cold refreshment made to chill. Of course, as always,

SPEAKER_109: I want you to celebrate responsibly. Okay, let's get back to this amazing episode.

SPEAKER_185: All right, everybody. If you're posting your pictures on the internet publicly,

SPEAKER_15: you can be 100% certain that they are in a database in the Communist Republic of China's

SPEAKER_21: secret police database. You can be sure Putin has them. And in all likelihood, Clearview AI has a copy of them and the FBI as well. Everybody is scraping publicly available data, but Clearview AI is doing it as a SAS service for local police departments so that they can solve crimes. They are not doing it as a public service. You cannot subscribe to this as an individual. Can a detective, I'm sorry, a private investigator, subscribe and give you money?

SPEAKER_58: We considered it in the beginning, but it's something we're not doing.

SPEAKER_29: Why? Because we want to make sure that it's for law enforcement. There's a lot more procedures there and regulatory oversight and people don't want to abuse it as much. If a law enforcement official does, their whole career is on the line. And so we made the decision just to stick to law

SPEAKER_189: enforcement. But private investigators do want to use your software? I'm sorry.

SPEAKER_29: I think a lot of people want to use the software, but it doesn't mean they get to. Because everything we do, we have to think about what the implications are and how it's used and what's the upside and

SPEAKER_43: the downside. All right. In all my conversations with Google, they said, listen, Matt Cutts would

SPEAKER_21: say, if you don't like the way we're indexing and using your data, you have the option to opt out of being in the index. So your defense of what you're doing has been thus far fair use. And hey, Google's using it, Google's doing it, Google's doing it. And that is true. But it's also true that

SPEAKER_43: Google will put a robots, they will respect a robots.txt saying do not scrape our site. So if a website that was a social network or Flickr said, hey, we're publicly information,

SPEAKER_194: but we don't want to be indexed in search engines, they could opt out of being in Google, but they could not opt out of being in your service, correct?

SPEAKER_29: Sure, we do comply with robots.txt for our open web crawler for the millions of other sites we do scrape. And the other thing is with Google, you can't you don't do that with Instagram and Facebook.

SPEAKER_104: Sure, we we have but we're only accessing publicly available information.

SPEAKER_21: Right. But if they said we want to opt out of it, this is where I think your argument is challenged. You know, you say Google is allowed to do it. But Google respects the robot.txt. What you're

SPEAKER_83: telling me is you respect the robots.txt, except in cases where it's publicly available information, which means you're not respecting robots.txt.

SPEAKER_29: Yeah, but the courts have ruled in LinkedIn versus HiQ very clearly. Okay, but to be clear, you're not taking the Google approach. Yeah, but the other approach is this. Can I go to Google and say, please remove these links of myself that I don't like? Today, on this day, you cannot say, remove these articles that I don't like on other people's sites. So if you're another site that is a publisher, and they publish a photo of you, you don't like-

SPEAKER_02: Well, right to be forgotten in the EU is giving people that right, is it not?

SPEAKER_29: Yeah, that's why we're compliant with GDPR. Now, if you use the right to be forgotten with Google, and apply an opt out request, they don't always satisfy that. You can't just get your stuff removed. So there is a little bit of judgment that comes into place. I think your point that a criminal could go in there and remove their information, there is a trade off. So Google doesn't just take you out of an index or take your links out if you don't like them, right? Say there's a link of you that you don't like, an article that someone published. How do you get a take-

SPEAKER_43: There's a lot of those, yeah. No, I mean, you're not going to be able to do that. There's freedom of speech, exactly. But if it's your server and the service you're providing vis-a-vis Flickr,

SPEAKER_113: or if I just had a small website, let's say I had a website for people who were, it was a social network for a specific group of people, a specific ethnic group of people. And they said, hey,

SPEAKER_21: you can't search our TXT, you can't index us, Clearview. You'd be like, okay, I respect that.

SPEAKER_83: But if it was Instagram, you'd say, we don't respect that. That's just to be clear.

SPEAKER_113: You don't respect the fact that Instagram said, don't scrape our stuff.

SPEAKER_175: Yeah, we have custom crawlers for certain sites. You believe you're within fair use to do that.

SPEAKER_29: Yeah. And also within all the case law that's happened, LinkedIn versus IQ. And finally, if you take the use case of this technology and the upside to what it's providing,

SPEAKER_30: we're taking public information, we're solving crimes with it, we're saving children. There's a lot of really good upside that people are not talking about or thinking about. So when you take all that, into context, it's a very positive thing overall for society.

SPEAKER_15: Got it. And are there things that facial recognition has not gotten perfect yet?

SPEAKER_78: Yeah. And what's left? Yeah. Yeah. I think what we've gotten to now is we at Clearview have developed such an accurate technology

SPEAKER_30: that can pick people that searches over 3 billion faces in the data set and can pick you out perfectly, no false positives to the point where it can be used now for so many good things.

SPEAKER_21: But your facial recognition technology is software that you've written or are you using some open source software? What's the state of it?

SPEAKER_29: Yeah. So we developed our own algorithm to do the matching. We also made our own database to search as vectors. So each face we turn into 512 points and then we made a database.

SPEAKER_185: What's an example of those vectors for people who don't understand that concept?

SPEAKER_29: Yeah. So we want to take the image of a face maybe and we resize it to 110 by 110 pixels. And then we convert it into all these different points, the interesting points around the face.

SPEAKER_30: And then we search that throughout other 3 billion plus vectors. And we can do that pretty quickly. So that's the other technology we developed. We have developed our own crawlers and search engine, which is really hard to do. But when we talk about those vectors,

David Friedberg: what's an example that a human could understand? Is it the distance between eyes is the one people typically give? Is it the length of a person's nose? Is it the forehead? I mean, what vectors are you

SPEAKER_15: trying to, is the neural network trying to understand? Yeah. So in the early days of facial recognition,

SPEAKER_29: people would measure things manually. Things are the difference between the eyes, etc. What we do is

SPEAKER_30: we have a ton of examples of one person, say like yourself, Jason, and then a hundred or a thousand examples of George Clooney and a thousand examples of Brad Pitt. Can a computer pick me and Brad Pitt

SPEAKER_22: yet? Because this happens all the time. I'm walking down the street and people ask for something.

SPEAKER_29: I know. They can't. Yeah, they can't. Yeah. So, and it just learns the things that are different between these training examples. And then when you have a new face that hasn't seen before, it kind of puts it in a different category. Got it. So when the neural network learns these 500

SPEAKER_113: photos of Jason, these 500 photos of Clooney, these 500 photos of Brad, then when 500 photos of one show up for testing, it's like, we know it's not those three people. Exactly. And it puts them in a

SPEAKER_29: different coordinate space, so to say. That's what it's 512 dimensions. It's a stupid question,

SPEAKER_15: but can't you run your vector and algorithms against an Instagram photo,

SPEAKER_225: but not scrape it as such, and not put it into your database, but just keep the vector data?

SPEAKER_48: It's possible. And then point to it. Yeah, I mean, but that's what Google does as a search engine.

SPEAKER_29: It downloads the whole internet, and then it makes an index of all common keywords that you do that point to the original web page. So yeah, you do have to store it on a technical level in order

SPEAKER_15: to search it. But if you were to process it, and then never store the photo, but store the, let's call it the metadata, the 500 vectors, you would have your own conception of what this image

SPEAKER_43: is that you took from my site process, but never actually permanently store it. In other words,

SPEAKER_21: it was in short term memory was in a short term cache, you don't actually have a copy of the photo,

SPEAKER_113: you have a pointer to the photo, if it gets removed, then you can't be helped by it. But you have this composite of it. Would that not be a way to route around this issue of the scraping issue?

SPEAKER_29: Yeah, we believe we're doing everything 100% compliance with the law. And you know, Google has a cache of each web page it crawls as well. So you can opt out of the cache too. Yeah, that's another. Yeah, we have the same thing we do. If your page is deleted,

SPEAKER_15: we have a place where you can opt out of the cache. I noticed on Reddit now, the deleting of the web is such a acute issue deleting of objects on the web, that there are these mirroring sites. And what

SPEAKER_113: these mirroring sites do, and I don't know who runs them, but I'm guessing they're run out of

SPEAKER_15: jurisdictions where it's very hard to pursue folks, i.e. Russia or North Korea, etc. They will look at somebody post a photo like this, you know, one of these Karen videos or something,

SPEAKER_113: and they know it's going to get taken down, where there's a potential of that. And it just makes a copy of it mirrors it and then they automatically put like three mirrors to it. It's pretty interesting thing that's occurring in the world. Could you not mirror the could you not scrape the mirrors that are

SPEAKER_26: exist in the world as opposed to scraping the primary sites? Or do you do that?

SPEAKER_29: Yeah, we have a open web crawler that just goes from site to site, you know, millions of different domains that's continuously finding websites with photos on it and indexing it.

SPEAKER_15: So yeah, these bots that do it are pretty interesting as a concept of just anything that can, and we also have the, what's the web archive called here in San Francisco? Archive.org. Archive.org, The Wayback Machine. I found so many

SPEAKER_43: old episodes of Calacanis cast and my magazine in there. Somebody had PDF'd one of my early magazines. I was like, I didn't give them the rights to do this. And I was like, yeah, they just mirror stuff

SPEAKER_113: that's going to disappear on the web. So what is the acute issue that's not been solved for facial recognition? That we still haven't, I still haven't gotten kind of an answer on that of like, what's left? I mean, you say it could be improved, but is there something specific that's hard? Like

SPEAKER_189: are masks hard? Are sunglasses hard? Are wigs hard? What's hard?

SPEAKER_29: Yeah, sunglasses can be a little bit of a problem and masks, but it works pretty well if you cover your mouth or beard. If you have a beard, it matches you to non-beard photos, different angles.

SPEAKER_30: So there's always room for improvement. But my overall point is it's reached the tipping point in terms of being extremely accurate, useful. What do CIA agents and other spies and people

SPEAKER_15: who want to avoid facial recognition do to avoid it? Well, you'd have to ask them, but I think they stay out of focus. Well, I mean, you have to go then reverse it. So when somebody is on the lam and they're on the run, when police are looking for them and they

SPEAKER_113: decide they're going to wear a wig, they're going to put on a prosthetic nose, they're going to put on sunglasses, they're going to wear a fake mustache. Does that stuff actually work or not? That's sort of what I'm getting at. Because you must get this request from people where they're like, this person's on

SPEAKER_29: the lam, but we know they're using a disguise. Yeah, we've had crazy success stories. One was someone from a federal agency. The first search they ran was someone from the most wanted list, and they could find them on the lam since 93, and they could find a lead to them from photos. So it's been very accurate so far. So it's a neural network. So it's trained on a lot of examples. So it can learn to ignore things like the beard and glasses, things that are-

SPEAKER_243: Have you had instances where people are repeatedly mistaken for a criminal? We had this with the

SPEAKER_113: databases after 9-11, the famous Do Not Fly list. Somebody's name would be Muhammad and some other common last name, and they would be a professor at a university. But they happen to have the same generic John Doe type name, except of an ethnicity of people from Saudi Arabia, let's say, that were

SPEAKER_15: involved in the 9-11 attacks, the murders. And they would get mistaken over and over and over again. They'd be like, listen, I know my name is Muhammad, but I'm not that Muhammad. Do you have that issue with facial recognition or not, where people repeatedly are false positive?

SPEAKER_30: No, we don't. What's great about it is there's a lot of common names out there, a lot of Jason's, a lot of Mohammed's. But with a face, it actually just does purely matching on the face. So if there's no similar match in the database, it turns up zero results. So a lot of the times they're running a photo and it actually gets zero results, would rather not give a false positive. So we've been, we think it's a great tool. In these kinds of things, if your name's Muhammad and someone's on the treasury list of, you know, or a terrorist list, and they get detained at the airport, it's not a good experience at all. And we don't want that to happen. I think facial wreck is a great tool to help, like again, catch the bad guys, but not get false positive.

SPEAKER_103: Who do you come up against as a competitor? Is this something, isn't this what Palantir,

SPEAKER_113: isn't what you're doing a subset of what Palantir does? Like they just

SPEAKER_103: provide people with intelligence on how to find people on the web and public data.

SPEAKER_29: Is that who you come up against? No, we, we're in a very unique spot because we're such a new product. And we have this, these legacy competitors in facial recognition that, that sell products that aren't that accurate. But Palantir doesn't provide this product? No, they don't have anything like it. Yeah, they have, they have more data related products and they do a lot of custom things for different agencies, but it's more name based or entity

SPEAKER_15: based when we're doing things purely on the face. And so what about real time facial recognition?

SPEAKER_21: Could a police officer's camera on their dashboard be watching people cross the street in real time with your software and just tell them who that person is in order to find a suspect who committed

SPEAKER_43: a homicide the night before. And you're like, I think this person's still in town. So we're going to park 10 squad cars with the software. And in real time, watch everybody go by to see if we have a match.

SPEAKER_151: Yeah, we don't have any real time surveillance. This is all off to the fact, you know, the person

SPEAKER_32: has done something wrong. You're looking at surveillance footage, then you run the photo.

SPEAKER_59: Got it. So you don't do real time. Isn't that the Holy Grail? Is there, is there a real time outside of like, let's say the Chinese Communist Party's real time tracking of their citizens?

SPEAKER_29: I think some people are offering it and tests. Yeah. In the West, I think UK has a real time uh, surveillance thing, but it's not something we do. So does it work? Is that, is that harder to do than what you do the real time? Um, depends, you know, you have to have a good depends on the algorithm and it depends on the placement of the cameras, but it's something that,

SPEAKER_15: uh, I think does work. Yeah. So is that what happens when we all go through customs? There's

SPEAKER_113: cameras everywhere when we're sitting there, um, online and they're doing video of us and we're going through, um, even if you're going through quickly, it's taking a picture of you putting it into the database and attaching it to your passport. I think there's some systems already like for

SPEAKER_29: entry exit that just match your face to the passport to make sure it's the same, right? That's all they

SPEAKER_113: do. Do you think they're keeping like a record of like every time I've gone through it? So they have

SPEAKER_43: 20 pictures of me. So they have like, now they have their data set to match to the one photo on my

SPEAKER_29: passport. Uh, I'm not sure how those systems work actually, but I mean, the facial rec's been around for a long time, right? I think about 20 years and now it's just got to the point where it's super accurate. Um, and we have a big breakthrough here. Do you believe in police should have the ability to,

SPEAKER_15: let's say go to a concert or, uh, you know, a train station if they're looking for a perpetrator and just take everybody's photo and then just run every photo against the database. Should police be

SPEAKER_29: about not to you that yes or no? Um, it's a good question. I think there, you have to balance the privacy trade-offs between, uh, the, you know, catching the criminals. So what would you do?

SPEAKER_260: What do you believe your personal belief? Yeah. My personal belief. I think that police should

SPEAKER_29: have the right tools, but it should be balanced with any kind of like auditing and stuff like that.

SPEAKER_21: So there's no, so in this case, you believe police should be allowed to take a picture of every person coming into a subway station. If they were looking for a murderer,

SPEAKER_113: check it against the database, even if there's no probable cause for any of those people, they're just doing a dragnet. You believe that that's in the best interest of society.

SPEAKER_48: Well, I, right now, personally, personally, I think that the right now police, what they do is

SPEAKER_30: they look through footage, say, for example, the Boston bomber, right? Correct.

SPEAKER_29: A classic case. They, they had his photo, they put it, uh, out there and there was so many people misidentified, um, from that photo. People would send in tips. They couldn't find him for seven days,

SPEAKER_30: 14 days. And here in New York city, there was the pressure cooker case, same kind of thing. Some guy left pressure cookers at the subway. And you know, with the use of our tool, many different agencies ran his photo and found him in less than an hour. Oh, they use your tool for that one. Yeah. We had, uh, multiple agencies run the photo and matched to a previous arrest record of him. Uh, yeah. And so that's the difference between the technology.

David Friedberg: Okay. So that's when you have a picture of that. But I guess the, the nuance here that I'm trying

SPEAKER_21: to get at is if, if, and, and it's hard to use the Boston bombing case or nine 11 because the magnitude of the suffering and the pain and the terrorizing nature of it would lean everybody to say, or any reasonable person to say, we have to catch those people before they harm more people. And it's

SPEAKER_113: understandable. It's sort of like the, somebody has got a nuclear bomb. It's about to go off. Do you believe in torture or not? If the person knows the location of the nuclear bomb,

SPEAKER_43: putting aside that edge case in a general situation where a homicide occurred, should the people who go through that train station all have their privacy, uh, compromised in order to

SPEAKER_122: catch that criminal sort of what I'm getting at. And you believe, yeah, no, I don't think so. I think there should be probable cause before you do anything, right?

SPEAKER_29: Got it. So then probable cause means what in this instance? So if for example, someone is running out of the subway after committing the act and the police officer is going through the footage and they're like, oh, I think that's the guy. We think he, he matches a description, but we don't know who he is. All we have is the face. Let's run the face. It wouldn't make sense to run everyone's face in real time. That's not really the society we want to live in. Got it. So the real-time nature of it makes it bad because why? It's more of a dragnet. Um, like you said before, but also we want to build technology that fits into our society and how we want to live. So in China, there are no, there is no rule of, there's no probable cause. There's no like legal system. There's no checks and balances here. There's a presumption of innocence. So I think that in order to make this technology work for Western societies, the way we've approached it as after the fact crime solving and only searching public information, it really fits in because you can't just take technology and just, you know, apply it somewhere else to make it work. You have to really fit in with the system. So,

SPEAKER_99: so we're thinking through this, this homicide case, the detective could say,

SPEAKER_15: give me the tape today. They could say, give me the tapes of, you know, we think that this murderer takes the R train at, you know, uh, you know, uh, the, uh, 18, the 69th street station in

SPEAKER_21: Brooklyn. Let's take the video of the last five days and we'll give it to an intern or a rookie and say, just look for this description, a white male with blonde hair, who's, uh, six foot two. And they would just find each one of those clip them, then run it through your software to look for them as opposed to just taking the video, having the video auto clip, every best shot of a face and running everybody through it and saying, here's everybody who goes through that

David Friedberg: station. And then we're going to talk to those people. Cause we know this person has to get to

SPEAKER_29: work at this time. Sure. I mean, if you're a detective, you're trying to be efficient, right? Right. You're not trying to like look at everyone and question. You don't have time to question everybody. What this can really do is narrow down the possible set. It really doesn't expand the possible set of suspects. It narrows it down. If you're a detective, you're smart, you realize, okay, this is the time he's got, you know, back and forth to work in a lot of cases, using something like Clearview AI narrows down the possibility set. It doesn't really expand it

SPEAKER_165: because they do that. Can they, can you upload a video clip and say, just pull all the faces from

SPEAKER_29: this video clip? Is that the feature of the software images right now? It's just images. Yeah. You upload an image. So it's not, it's not something where you're trying to do a drag net.

SPEAKER_28: Their job is to actually find the person. So they rely on a lot of,

SPEAKER_99: sometimes they might want to find the witnesses too. So I wonder if find, if you said, Hey,

David Friedberg: listen, this homicide occurred at this time, 8 0 5 PM. I want everybody in and out of that station in the hour before and the hour after. And they took the video, they clipped it, and they just got all of the possible witnesses. And then they went down the list. In society, we would be okay with that or not okay with it? Would you be okay with that or not okay with that? There was a murder on your

SPEAKER_21: subway station by your house. They, you happened to be there an hour before and your spouse was there an hour later. And both of you got pulled into questioning because you're possible witnesses. Would that be good or bad in your mind for society?

SPEAKER_29: I mean, if it leads to solving a crime, then it's obviously a good thing. And I think there's a lot of cases where people could be wrongfully arrested, right? You got the wrong guy. And now it's maybe some, theoretically, a defense counsel could get footage and say, Hey, I can identify that witness

SPEAKER_30: and bring him to the stand. So it's about finding the truth. And it's about protecting the innocent, but also catching the bad guys. So in these kinds of cases, detectives are, they're trying to narrow

SPEAKER_29: the possibility space down. They're not really trying to expand it.

David Friedberg: Here's the thing though, if they have a fixed amount of time, you keep bringing that up. And I think that's like a very astute observation. So what technology will allow them to do is instead

SPEAKER_21: of having the witness spend the afternoon looking through mug shots, which leads to a lot of false

SPEAKER_15: identifications, because they're like, here's a, here's a book of people who've committed crimes, pick one, that man, that picked the one that most matches the person who robbed you. It's like,

SPEAKER_113: is that real? That kind of leading the witness, right? To give them a stack. And that's why people

SPEAKER_21: become repeat offenders in some cases, is they're just, they're already in the system. But this makes a detective bionic, if they could use software to clip every shot,

SPEAKER_43: then run every shot, they can do more with less time. So it actually expands the range of what

SPEAKER_21: they're able to do, while compromising on the margins, people's privacy. I think that's probably the

SPEAKER_29: real world reality of this. The reality is that police departments, there's too much work to do. There's so much crime, they don't get to all the cases they want to get to. For example, the financial fraud case that I mentioned before, 35 million recovered, 19 people,

SPEAKER_30: and they got it from ATM photos. They would have not been able to close any of these cases as stacks and stacks of cases, unsolved things that are sitting there, and they're able to go through them more efficiently. So I think overall, it's very, very positive.

SPEAKER_15: In terms of tricking it, what have you learned in terms of people trying to trick facial recognition?

SPEAKER_29: I mean, people have tried face paint. For us, it still works most of the time. People have tried all kinds of stuff. But yeah, it's very, very accurate.

SPEAKER_15: I know this is really James Bond-ish. Does people, given your data set, if I said, show me everybody who's had a nose job and what date you think they had it on, you'd be able to actually do that. You have all these photos of me. You could figure out when I had

SPEAKER_292: my nose done. Probably not at this stage. But yeah, it's just a search engine for faces. We're

SPEAKER_99: not doing anything super sophisticated like that. And so I wonder if this whole James Bond villain of

David Friedberg: them, people getting plastic surgery actually does defeat the facial recognition. Has that happened

SPEAKER_22: where people committed a crime and then went and undergone facial? I wonder if there's a case of this. There must be in the world where some criminal, who aren't always the brightest people, but went and

SPEAKER_113: got a massive facial reconstruction surgery, in fact, to be obscure in the future. I know there's a

SPEAKER_29: Columbo episode with this. It's possible. And I think that some very sophisticated criminals would

SPEAKER_292: do that. But for the most part, it's a lot of pain to go through facial reconstruction. So

SPEAKER_15: not everyone's going to do that. And for your company, we had San Francisco ban the use of facial recognition, I believe. Maybe Boston did too. What's the state of local municipalities saying

David Friedberg: we're going to preemptively ban the use of facial recognition technology?

SPEAKER_29: Yeah, there's a few local municipalities, San Francisco, Somerville, Massachusetts, maybe one other, Oakland that have banned the use of facial recognition for law enforcement. But we see that most of America, most communities, they want to stay safe and they're okay with the use of this.

SPEAKER_15: So in Oakland and San Francisco, and I think San Francisco is number one in crime, by the way, so congratulations. So we made this official ban, I think last year in May, and they preemptively did this. What was the impetus for this, do you know? Yeah, a lot of impetus is around the accuracy of

SPEAKER_30: the technology and potential misuse. And I think we addressed both of those completely. But they're basing it off really old facial recognition algorithms that aren't accurate. And I think a lot of fear and hysteria around the use of this technology, without thinking about the upsides.

SPEAKER_29: I mean, San Francisco has a big crime problem, a lot of other cities do. And if anything can bring

SPEAKER_22: it down, I think it's a good thing. Yeah, it's kind of crazy that the Oakland Police Department, the San Francisco Police Department, where we have massive numbers of break-ins,

SPEAKER_15: stabbings, murders, violence, would not want to keep, would not want to be able to recognize who committed those crimes. It's almost, I mean, I understand people are sensitive to

SPEAKER_21: the issue, but we take mugshots and pictures of people taking crimes. Police do this all the time. They keep a database. Do those police departments, when they are taking those pictures of criminals and trying to build indexes and taking pictures of gang tattoos, are they using some technology like yours to keep an index of these folks and then track them?

SPEAKER_32: Most police departments don't. And that's where we can come in and help them search their existing

SPEAKER_30: datasets more accurately. But a lot of the things that police are doing, taking mugshots of people, most wanted, wanted photos. We have pretty funny stories, too, about people just taking a photo of someone who's wanted and actually being able to solve the case right away. So you can take a photo of a photo on Clearview. So yeah, I think that every city is different with their approaches to crime. But for the vast majority of America wants to stay safe. And we can really help with that. And there's an interesting Pew study, actually, about the acceptability of facial recognition in the police and law enforcement world, which is about 60-something percent who approve of it. Technology, maybe 30 percent. And advertising is like seven or something like that. So the general public is really fine with the use of facial recognition to solve crime because it's such a good trade-off. When you look at the privacy security trade-off, it's one of the best that's out there.

SPEAKER_15: Yeah. I was thinking about this in relation to license plate readers.

SPEAKER_22: There is open source software out there that does license plate reading. I know this because I was looking into, or there was a thread on a Nextdoor where people were talking

David Friedberg: about in the peninsula which jurisdictions were using this software and these services. Many

SPEAKER_21: jurisdictions right now are tracking all of the license plates that come in and out of their

SPEAKER_113: neighborhoods and looking for, when a unique license plate happens, they can be alerted. This is a license plate that's never been in the neighborhood before. A license plate that's never been in the neighborhood before could be a rental car or somebody drove up to see grandma. It could also be somebody who wants to rob houses. It could also be an Uber driver. Who the heck knows? But what are

SPEAKER_15: your thoughts on license plates are public? That's the definition of public. You're driving publicly on a public road. Should people be able to keep databases of license plates and then you would be able to

SPEAKER_113: correlate license plates with the driver at some point? Is that on the roadmap for you? What are your

SPEAKER_29: general thoughts on license plates? We don't do anything around license plate reading, LPR as they call it. It's already an industry that's matured. When you look at the evolution of LPR, it's been around for quite a while. A company, Vigilant Solutions, is a leader for that. I think 10 years have been around and they went through a lot of the stuff in the beginning where people were worried about the privacy aspects, but eventually it was adopted. There were some moratoriums for LPR and then everyone came

SPEAKER_30: around to it and said, this is a very good tool. When you look at the history of LPR, it's a more mature market. Technology has been around for longer. Overall, people are okay with license plate tracking because once you understand the trade-offs here and the fact that they're really locking up some really terrible bad guys, that people are okay with it. Yeah, I was thinking about it. I was like,

David Friedberg: I was driving in my Model 3 and the accuracy now of those cameras is insane. I was just thinking

SPEAKER_21: either ways, which is on your mobile phone, on your dashboard, if you had an Uber-like driver, like a mount, like the Uber drivers have or Lyft drivers have on their dashboard, which is what I use. It's really great for keeping your eyes on the road, although it looks dorky in a car.

SPEAKER_113: Those could very easily read every license plate. You're watching the Tesla Autopilot,

SPEAKER_21: know the difference between a truck and a car, know the difference between a bicycle and a motorcycle, and it could pick up a cone, like a traffic cone. It could very easily keep a database of every license plate you ever saw. We could have license plates recorded everywhere. And I was just thinking,

SPEAKER_15: wow, if technology can do this, there must be hackers out there right now that are tracking every

SPEAKER_21: license plate on the highway in every different direction. And this technology, how difficult would it be for somebody to set up an instance of a scraper that just took every photo on Twitter, every photo on Instagram? How difficult is it to make just a giant database of that right now? Isn't it an easy task?

SPEAKER_175: It's still a lot of work to code all this stuff.

SPEAKER_113: Well, coding the algorithm. I'm talking about just scraping every Instagram photo, like it's publicly available on the web. Could they stop you?

SPEAKER_78: I mean, yeah, maybe it's possible, but we're not doing any LPR stuff.

SPEAKER_113: Yeah, yeah. But I'm just talking about just in general, like the concept of a group of three

SPEAKER_21: hackers on a weekend project doing a decided they want to scrape Instagram, and then put that data

SPEAKER_113: set onto Amazon service. They could do that relatively easy. There are public scrapers out there as a public database tool for Amazon could be done pretty easily, right?

SPEAKER_29: Yeah, I think to get started and prototype these things is not super hard. But to build like a

SPEAKER_30: really large scale database like we have, and super high accuracy and billions of photos, we have our own infrastructure. We're not using AWS for all the storage. So that gets the cost down. So to do it well, to do it at scale is a totally different task.

SPEAKER_314: Do you scrape tick tock yet? And video?

SPEAKER_30: Yeah, we don't do any video, but we do a whole variety of different sites and customers come to us with like suggestions.

SPEAKER_113: Ah, so they say, hey, get me all the tick tocks. But you know, you could take a screenshot of a video playing on there and then just use that's enough. You don't need a video.

SPEAKER_29: Yeah, you don't need full videos. And maybe one day we'll do that. But we also are at the very

SPEAKER_30: early stages of how much information is out there. There is 30 trillion pages on the internet about what in 10 has a photo of face on it. So I think with Clearview AI, which is still at the very beginning of collecting as much information as publicly available. Just thinking about the number of stock

SPEAKER_15: images on the web as well. Like there's somebody who has the most images on the web? LeBron James,

SPEAKER_113: Jesus? Is there a stock photo model? Who has the most images on the, who has the larger, who do you have the largest data set on LeBron James?

SPEAKER_317: Um, I'd actually have to find out. Yeah.

SPEAKER_43: Uh, so, uh, companies profitable now companies close to profitability. I mean,

SPEAKER_194: you've raised very modest amount, not a gigantic amount.

SPEAKER_29: Yeah, we think, uh, we have a great business opportunity ahead of us. So after the New York Times story, we've had just so much interest from,

SPEAKER_243: What did that do double, triple your user base?

SPEAKER_28: Yeah, I think it, I think it doubled, tripled it, something like that. And also,

SPEAKER_22: So leaning into the, in a, in a, as a founder, leaning into the controversy and having a defined position is the best PR tactic for you guys?

SPEAKER_29: I don't think it was a PR tactic. I think we wanted, we built what we built and we have our

SPEAKER_30: own beliefs and other people have different beliefs and that's what makes life interesting. Uh, but we're not out to court controversy. We want to be something that's like as good as possible. And, um, you know, live up to the highest standards.

SPEAKER_113: Yeah. I was just thinking in terms of like what a PR person would advise in crisis communication, which is a thing that occurs, right? And, and wouldn't necessarily qualify the New York Times wanting to do a story about it as crisis. Um, but it could be a crisis if the story went poorly.

SPEAKER_15: So do you engage a PR firm to do all this for you and try to think through how to get the public to understand what you're doing and communicate it? And then what's their advice?

SPEAKER_29: Uh, I have a wonderful person who helps with, uh, PR and she, it's been a great mentor to me in terms of how to prepare for interviews, how to, uh, answer questions and all that stuff. But fundamentally it's something new.

SPEAKER_26: What was her advice to you on how to deal with this? Cause it's like such a hot topic.

SPEAKER_29: Yeah. Yeah. To engage. I think it's a very important to engage with people and explain your side of the story. I think there's a lot of people who come in who might be very against it,

SPEAKER_30: but you know, after talking to me or hearing outside of the story can understand that it does have a lot of value. So.

SPEAKER_113: See, I think engagement is the right thing. Uh, some people take the, just don't even respond. I, um, I think we've invited Pallinger on the podcast many times. I just don't think they would want me to ask them the questions the way I asked them to you. And you were very honest in

SPEAKER_21: your answers to me, even when I forced you to like, Hey, answer the question one more time. You know, what is it like to have Peter as an investor, right? Like you, I can tell you've, yeah, you've practiced answering some questions and, but I was trying to steer you to get to the

SPEAKER_75: more real and you got to the more real and I appreciate it.

SPEAKER_102: Yeah. I don't, I think there's, there's a, there's a value to engaging. And I think, uh,

SPEAKER_29: that's one thing that, um, that has served us. Well, I mean, not all the media is positive and it's never always going to be a hundred percent positive. That's like a myth. So, but engaging

SPEAKER_30: with people with being able to convince a lot of people that this thing is a great tool. It's saving lives for saving kids. And, uh, fundamentally it has a place and a reason to exist. I think that it's important to engage with people who are different from you. Uh, a lot of people in the media

SPEAKER_21: have their views, uh, regulate the media is, uh, there's a new term like, um, anyway, they have an, uh, agenda would be a negative way to say if they have a point of view, they have a position, they have something they're championing. So, you know, do you find that the media is coming with,

SPEAKER_328: Hey, they've kind of already written the story and now you're just trying to make sure that your view is included in their version of reality. Yeah. A lot of the times they've already had their

SPEAKER_151: mind made up, they've already written most of the story, but you can engage with them. And

SPEAKER_29: if you're thinking it on a longterm basis, they're always going to be around and we're always going to be around. And so it's just the beginning of a conversation. So I think that a lot of, you have to really ask yourself with journalists is like, can, would they, you know, change your mind if you showed them your side of the story? And most of them want to be honest and do that.

SPEAKER_30: So I think there's, there's less trust in media now than ever before. But I think part of that is people aren't always willing to engage. So do people write stories about you without

SPEAKER_21: ever contacting you in the press? They just basically write a story and then you have to go

SPEAKER_113: in retrospect and, uh, after they've written this, then try to convince them and

SPEAKER_29: sometimes, but, uh, if it's a follow-on story from someone else, but most of the original stories, they do reach out and we do have a chance to comment. So that's the thing I find weird now is

SPEAKER_99: it used to be in my day when I was coming as a journalist, I'm old now, but in the nineties,

SPEAKER_22: they said, you can't run the story unless you get a comment from the first person.

SPEAKER_21: You have to let them know you're doing a story. You have to let them know what facts are in that story. Um, you have to let them respond to the quote. So if there was somebody who had a negative

SPEAKER_113: take on it, you want to give them the chance to respond. And now I don't even see that happening. People just write the story unilaterally, pick 10 facts. And I know this because I'll, I'll be mentioned in a story and I just get a Google alert. I'm like, I wonder why the journalists didn't even just, my DMS are open. My email is my first name at my last name. It's literally all over the web. Like it's not hard to get me and they don't even bother to like, say,

SPEAKER_21: Hey, we're, we're quoting you in this story from your podcast or from this wondering what you think of X, you know, they just don't even take the time to do this weird. Yeah. I think for the most part,

SPEAKER_48: they've done it with us, at least for the big stories where they're doing original research,

SPEAKER_30: other people who do follow on stories might not, but, um, it's part of the process and we had to

SPEAKER_29: learn how to deal with it. And it's, I think it was an important thing to do.

SPEAKER_02: How do you know your, uh, employees are not abusing the system?

SPEAKER_104: So yeah, we make sure that we only use it for testing purposes. So, um,

SPEAKER_02: How do you know someone like doesn't get compromised and then, uh, abuse the system?

SPEAKER_29: Yeah. So we routinely check the audits of our people, making sure they're using it for demo purposes. So an employee might need to use it when they're doing a zoom demonstration or Google Hangouts

SPEAKER_30: demonstration to a potential client. So we give them, these are the images you try and is the images you use. Um, and so it's very easy to check. So you audit all your employees use of the

SPEAKER_02: system. It would be hard for them to, yeah, because this is something I don't, you know,

SPEAKER_15: people really don't talk about, but we had, um, spies from the kingdom of Saudi Arabia working at

SPEAKER_78: Twitter. You must've seen the story, correct? Yeah, it's fantastic story. Uh, fascinating stuff.

SPEAKER_189: How do you know you don't have anybody in your sort of hen house who is a international spy? And do you think about that as a founder and how would you know?

SPEAKER_29: Yeah, I actually do think about weird stuff like that. Yeah, I would think so, right? Like organized crime, spies, uh, things like that, where it could be a target. And now that we're out there, they were more of a target. So we've made security a big priority in, in everything. So employees know that they have to follow the highest security

SPEAKER_30: practices. We have invested a lot into more auditing to factor off more things like that to make sure the systems are secure. So something we think about all the time. Now that there's all the scrutiny, we don't want to kind of compromise people going in there and, uh, and using the system.

SPEAKER_113: I was explaining this to somebody and they were like, it's impossible. I said, listen, Alexa, I'm sure Alexa, the, um, you know, Google, uh, I'm sorry, Amazon, uh, audio assistant. I said,

SPEAKER_21: I'm sure that's been compromised to what extent, you know, I'm not certain, but, and they're like,

SPEAKER_113: you're just a conspiracy there. I was like, well, here, let me run something by you. Uh, have you ever seen the show, the Americans where they get compromised on somebody and then they say, listen,

SPEAKER_21: uh, we have this video of you in this compromised position that we're going to publish. And all we want you to do is just go in and change this piece of code or just, you know, download this

SPEAKER_113: person's data and send it to us. And that's the last thing we're going to ask you to do. But what the person doesn't realize is that dumb drive that they asked him to put it on also has a worm

SPEAKER_21: that downloads everybody's information or puts in some backdoor. It sounds crazy. That's exactly what is happening today in Silicon Valley with Chinese spies and spies from the kingdom of Saudi Arabia.

SPEAKER_15: And I'm sure we're doing it to other countries. That's the game. And then you as the founder is

SPEAKER_29: responsible for if this happens in your company. Yeah, it's true. Uh, we have, uh, quite a small team. So that's easier when you're smaller, you really know everybody in your company, but as it

SPEAKER_30: grows, it's something that we have to be very wary of in terms of, you know, who's administering

SPEAKER_15: everything. So CIA and FBI, like they must be on top of what you're doing. What's your relationship with the three letter agencies? And do they have these kinds of concerns that your information could

SPEAKER_151: be compromised, et cetera? Yeah. Some of them are our clients. FBI child victims unit have had a lot

SPEAKER_29: of success identifying, um, pedophiles with the software and victims of pedophiles. So yeah,

SPEAKER_348: we have good relationships with people in law enforcement. And that was the, that's the real,

SPEAKER_29: that's the real thing I get out of this, uh, is the stories we have every day, the psychic reward.

SPEAKER_30: Like when we were talking to some of these people, we had no idea that they were actually solving really horrendous crimes. One case was, uh, in a child pornography video, there was, there was an adult male in the background for a few frames and they couldn't find it for six months, seven months. And they put him through us and they found him on someone else's social media in the background, in a mirror working at the gym. So they were able to go to the gym and say, Hey, have you seen this guy? And the, the agent has, yeah, there he is. He's on the elliptical. Yeah, no, but the agent had

SPEAKER_29: to convince them. Like, uh, we don't give out information. They said, this is for like a really terrible, um, child abuse case. Eventually you could identify him. He's now doing 35 years in jail and they saved a seven year old girl, seven year old. We have thousands of stories like this. So my

SPEAKER_220: question to other people is like, uh, how many of these bad criminals go away?

David Friedberg: I, for one, I'm glad you're out there. I'm glad you reached out to be on the pod

SPEAKER_21: and I commend you for taking the hard questions, even though I asked them two or three times and you know, I'm not going to let you go on them. I think that it's important that people who run companies like yours allow themselves to be scrutinized and have an open discussion about what they're doing. And I asked you very granular questions here and I give you an A plus on your, uh, performance today because I was challenged you with very obscure stuff like the robots.txt or should there be every single person, you know, for one hour before the crime, should they all be there? I mean, you had very thoughtful responses to that. I think it makes me feel better

David Friedberg: about you being the person running the company because you obviously care, right?

SPEAKER_29: Yeah. Thanks. I appreciate it, Jason. And I appreciate that you're asking these harder questions. I think that, um, it's been an honor. All right. Listen, great job on the company. I'm glad

SPEAKER_15: you're out there doing this and catching criminals. It's super important. Um, in a mirror, just like in Blade Runner, by the way, the in Blade Runner, that's how they found, uh, one of the replicants. It was somebody in a mirror at a, you remember when he was using the, it was, it was the original

SPEAKER_21: scene. Decker's in his apartment and he says, you know, trim, go left, go right, zoom in, zoom out.

SPEAKER_113: And he's zooming in and out with that like system that goes, clack, clack, clack, clack, clack. And you're like, it's pretty funny that the system's got like a mechanical clack, clack, clack, clack, clack, to zoom in on a photo. Like it's using like a steam engine or something. But he, that's how he,

SPEAKER_02: that's literally what happens in Blade Runner. Yeah. All right. Listen, great job. Uh, really appreciate you coming in the pod and, uh, we'll see you all next time on This Week in Startups.