Inside the AI Deepfake War

Transcript generated automatically from the episode audio and lightly cleaned up (filler words removed, misheard names corrected where clear). Speaker labels were added afterward and may contain errors.

Rob May: Hi everyone, I'm Rob May. I'm the co-founder and CEO at Neurometric.

Anna Kirk: Hi, I'm Anna Kirk. I'm sales lead at Thread AI. And this is the AI and NYC show where we talk about all things happening in New York City, which is the applied AI capital of the world.

Rob May: So, our guest today is Ben Colman from Reality Defender. Ben, welcome to the podcast.

Ben Colman: Thank you.

Rob May: And before we go into your background and everything else, I want to start with a question. You guys solve the deep fake problem. So, for our listeners, what is a deep fake? And then, how do I make one of Anna saying something funny that she would never say in real [laughter] life?

Ben Colman: Oh, yeah. It's a good example versus a lot of more dangerous examples.

Rob May: Yeah, we're gonna stay away from those.

Ben Colman: So, Reality Defender, we build models to detect AI generated audio, video, images, and texts, focusing on, detecting human-based impersonations. Everything from, AI generated calls to banks to take over an account to commit wire fraud to identity fraud, kind of a image you upload, it's you, but it's actually not you. To things a lot more dangerous and scary like AI generated non-consentual pornography or AI generated child sexual abuse material. All of it's AI generated and the real challenge right now is to your question how do you do it? There are over half a million tools available online, most of them free that allow anybody, whether it's my 80-year-old parents, my 8-year-old son, to create anything limited only by your imaginations. Which means there's a lot of great things, you know, AI for productivity, AI for health and science, but also in a very dangerous edge case, AI for very dangerous things. And right now, the absence of well, simple regulations, it is the wild west.

Anna Kirk: Ben, you are solving one of what I would consider to be the most relevant and important problems today. What tell us more about the story behind Reality Defender and kind of the moment you realized that this was a big enough problem to start a company around it.

Ben Colman: So we started the company about 5 years ago. We had some support from some government grants the AI foundation and Microsoft and we came out saying that the 2020 election was going to be the deep fake election. And we worked with Microsoft to launch what was then called the Microsoft video authenticator. A lot of press and then there were no deep fakes. There were deep fakes on one hand because the buzzword wasn't there yet. So we said we were detecting AI avatars and virtual humans but there were no AI avatars and virtual humans. And so we're very early. So to your question when we realized it, we still hadn't realized it yet. Now we were certain of it but it still hadn't happened yet. Yeah. And so what then happened over a few years suddenly a company called OpenAI launched at ChatGPT and then the world caught up to us in a very very big way. And where we are today is that whatever you see whatever you hear almost all of it has touched AI in some way. And the question is how do you tell without being a computer science expert or PhD if something is indeed real or AI generated or manipulated?

Rob May: I mean, I remember this. I don't know if you if you remember this story, but you guys actually pitched. I was at a venture firm at the time. And you guys pitched and my partners were sort of like this doesn't that this is probably 2021 maybe

Ben Colman: or maybe we had 190 people say the same thing. What is this thing? Like what? This isn't a problem. Then go back. You should work at a bank. It's more stable, which [laughter] is not not true. We had a lot of folks say what what is this? Where is the problem? We also had folks saying, "Why don't you pivot into NFT authentication?"

Rob May: Remember that? That was a thing.

Ben Colman: Oh, yeah. Yeah. I don't know if it's back.

Rob May: I heard something about NFTs this morning. I don't think it's back. And Bitcoin's down like 25% just in the last couple of weeks.

Ben Colman: There's no NFTs. But as a as a cyber nerd, my view has always been that fraudsters and hackers are the best product managers. So, if there's any tool or widget or technique or social engineering playbook giving you a little bit extra edge, obviously you'll do it. You're not competing for, you know, podcast points. You're competing just to commit a little more fraud.

Anna Kirk: So in that same vein, this is such a moving target like new models, new techniques are getting shipped every single day. How do you stay proactive in that type of world that's only speeding up?

Ben Colman: And so the question is like is this a cat-and- mouse game? Is it a game of whack-a-ole? It absolutely is. It's also true of all of cyber security, al all physical security as well. For us, we stay ahead in a few different ways. One of them is that we have our own red team that tends to break our models better faster than anything out in the market. Now, new models are coming out. I wouldn't say they're trying anything particularly novel since the most recent diffusion models came out. And so for us, we develop different kind of internal offensive capabilities in the lab that try and break our models so we can constantly retune and retrain. Other side of it is we partner with a lot of the leading generative companies. So some of them as we can name are 11 labs reeer play AI they want to be on the kind of the right side of history and so they give us early access to models and our data teams work with them just to ensure if there's anything unique or interesting that it is represented in our data set. We haven't seen anything that novel yet but it is good to show that we're kind of partnering both on the generation side and the detection side.

Rob May: So what do you see this do do you see this going long term a as people build more models? Do you think number one they're eventually going to get good enough that your problem becomes too hard and you need a totally different approach like maybe being built into a tool like 11 labs somehow or or or something or there needs to be regulation or you know do you think this can do you think that the technical solutions will work for at least the foreseeable future.

Ben Colman: So right now all the JAI platforms that are creating audio or video or images they're trying to replicate the way the human eye might observe things. They're not trying to replicate how I don't know photonics how physics actually work.

Rob May: Yeah.

Ben Colman: So you know we're confident that for at least the next 5 10 years we're fine. Now beyond that we have an idea of how different models can become kind of a net new foundational model to solve for that because while fakeness is getting better and better realness is not and so by attacking both sides of the equation we believe we can provide a kind of a superior solution. It's not one model it's an ensemble of models which we're constantly retuning retraining removing readding. We push updates probably once a week right now. The goal is to do it a lot more often.

Rob May: Yeah. Can you talk about some of the use cases that like obviously detection is an obvious use case but like from a enterprise perspective right so I come like why do I buy your product and what specific thing am I trying to solve and like who's the kind of like I mean I'm sure all the VCs ask you like who's the buyer

Ben Colman: you're yeah you're like yeah used to be

Rob May: and you're like how who who who does the you know who's who's the title of the person you sell to you know sort of stuff like that where do you live in an organization

Ben Colman: so the first step is that we are API first so that any develop veloper can go on our site, grab our API, a range of SDKs, two lines of code, and they can build real time detection into whatever solution they have.

Rob May: Oh, that's awesome.

Ben Colman: So, that's one key item. We have about 700 independent companies, developers coming every month, building it, testing it. We give free access to anybody. Three months free access to be specific. As far as our enterprises, the main use cases started off within call centers. So, real time audio scanning for call centers, for IVRs, for for desk phones. And then it moved into real-time video. So we integrate right now into about 85% of leading collaboration and call center tools obviously Zoom and Teams. And so whether you want to build us directly or just go into the app store on Zoom or Teams or Genesis you could grab our app, click on it if you're an admin and permit all your team members to use it. And what's key here is that we are what's called providence free and contextf free meaning we don't touch any personal data. We wouldn't know whether you're Rob or Ben. We wouldn't know whether you're AI or not.

Anna Kirk: I imagine part of this next wave is also going to be an educational problem. Like how do you educate people that there's even a possibility that they're looking at a deep fake? I'm curious for for the people listening who maybe aren't in this world. What are some of the common like gotchas that you should look for? What are the tricks? How can you stay ahead of this as an individual?

Ben Colman: So, if we were speaking a year ago, I'd give you all kinds of tips on symmetry and anti- aliasing and hourglassing of the face kind of on a face swap. But the real scary truth right now is that there's nothing that you can do. Even the PhDs from top schools who work for our company, majority of our 50 people are PhD researchers, engineers, they can't tell the difference.

And so, I'll use the same analogy as antivirus software. You know, you get an email, Gmail or Outlook flags it as a potential being a virus or ransomware or a Trojan horse or something else. Are you reading the code? Probably not. You're trusting the antivirus to actually check for it. It's really the same thing in this space. Even the experts can't tell the difference. How can average people stand a chance? Well, we need platforms like Reality Defender to be baked into every single email client, communications tool, browser session, and beyond. And the good news is we are doing this right now. And we're hoping that regulations will catch up. We right now only really serve large enterprises and corporates and developers. Consumers right now don't really have a tool here. We're hoping that'll change. And so working with you know, telecoms and banks to extend our solution, but with most bleeding edge cyber security, it starts kind of at the high side with government with enterprise and then hopefully becomes a bit more consumerized.

Anna Kirk: Yeah, when you think about, you know, the prospect of this technology eventually being regulated and kind of the morality and the legality around that, do you take a particular stance in whether or not people should be able to have the liberty of creating deep fakes?

Ben Colman: So, we're we're very much pro AI and pro AI acceleration. We're also pro knowing whether or not you're seeing AI or not. So, we're not saying whether it's good or bad. We're not saying whether it's truthful or untruthful. We are saying that you should have the right to say I do want to see AI or I don't want to see AI. The same way that on social media or streaming, you can flag different things. You could say I don't want to see violence. I don't want to see nudity. Football games, I'm a big football fan, but they're a little violent. And so when my 5-year-old using his tablet with all of its parental controls, he's not going to see any of those.

And when you upload a video onto, you know, streaming or social media, it's checking for, you know, violence, nudity, cam, it's checking for, I don't know, do you upload the latest Taylor Swift song? They're not doing that because it's from the kindness of their heart. It's because it's simple regulations. And so this next stage should be is there AI or not? And letting consumers make their own determination. The same way when you get, you know, a potential spam call, your phone is saying, "Hey, Rob, this might be spam." in the next few months you'll get a ping that says hey this call may be AI maybe a robocall from the airlines last few weeks have been exciting with delays or [laughter] it might be a fraudulent call or a voice agent either trying to waste your time or trying to commit fraud.

So the more that we can kind of evangelize this the more that folks understand the problem. It's not just education. It's you have to scare people a little bit. You have to show them how scary it is with their own voice, with their own face. And that's what we do really every day is we demonstrate how scary it is to our our clients.

Rob May: Yeah. I've been getting a bunch of calls almost like one a day. It's always a different number so it's hard to block and it's a small business credit line call and it's like so annoying and you've been selected.

Ben Colman: Yeah.

Rob May: I never answer it and it's always AI related but it's a big annoyance like every day I have to delete a voicemail, you know, which takes 30 seconds out of my day.

Ben Colman: I'm sure you're on a list somewhere and so you're just getting spammed left and right. I turn off my voicemail and if it's a number I don't recognize I just don't pick it up, which makes it very inefficient.

Rob May: But yeah, that's pretty interesting. So, you know, this podcast is about AI companies in New York. You guys are based here, which is really interesting because you're we think of this as the applied AI capital of the world compared to San Francisco where it feels like everybody's working on AGI and deep tech, but this is a pretty hardcore deep tech company as well. What's your experience been like hiring and building in New York? Like first of all, why did you make the decision here? And then have you had trouble hiring here or has it been awesome? Like what what do you think? What do you recommend to people?

Ben Colman: So we started the company kind of at the middle of co we could have started it anywhere. And the reason why we chose New York and not us I'm from California originally. I had every excuse to be in California was we believed at the time that this is where all the clients are. And so if it's an option of being where the clients are or other things, I'd always pick the client side. What we realized was that given all the clients are, you know, big banks with their own, you know, quant teams of researchers, there's a very captive audience of potential hiring talent here and they tend to have real experience. They don't jump from company to company every, you know, 6 to 12 months. I think San Francisco is phenomenal. I think it's also an incredibly challenging place to hire given the propensity people to jump from company to company to company. But also, we're building something much more applied. It's not theoretical. We found it very important to have very intimate design partners with our clients. And so our our first client was JP Morgan. You couldn't do that from San Francisco. You had to be here, you know, a Uber ride or a subway ride away from your your main stakeholders.

Rob May: Yeah. No, that makes sense. And it's part of what I like about New York is because it's got this culture where it's dense and you run like I run into people that I've worked with or pitched like all the time, right? It's like just in a coffee shop, you walk in to get coffee and there's people and they're having a meeting or at a dinner or just on the street and it's really it's really productive and it helps you stay top of mind with people particularly if you're out a lot and about and it helps you learn about problems that can be solved. You're not you're not a bubble with only people who are saying what if what if what if.

Ben Colman: Yeah, you know, I lived in SF. I was with friends who were all doing what I was doing. And often times it was a bit of an echo chamber. I can't tell you how many Uber for X ideas that I that I heard.

Rob May: Yeah. And tell us a little bit about what what did you do before reality defender? Like what was your background and like how how did you stumble on the need for this problem?

Ben Colman: So this is really all I've ever done and I've worked in kind of data science and cyber security. You know I've worked at Goldman Sachs focusing on cyber security commercialization and strategic investment. I've done some work for some government defensive groups. I'll I'll leave it at that. And I worked at Google in grad school. Kind of in a similar role thinking about data science on top of different kind of data. And so for me it's a continuation of what I've always done. It just happens there's a new venue for it which is you know this wide world of AI fraudsters will follow the technologies. You know maybe it's AI today, maybe it's quantum tomorrow.

Anna Kirk: How far away do you think we are from the quantum wave actually hitting as a material risk?

Ben Colman: I mean every day it says a different a week away or a year away or 10 years away. I was chatting with Worly from Strange Works. I think you might know him in the quantum space. I don't know that their view people smarter than me their view is we're probably a year or two away.

Anna Kirk: Wow.

Ben Colman: My excitement is there's a whole new generation of quantum secure cryptography which will solve for a lot of the new issues that quantum will bring to security.

Anna Kirk: Talking about quantum today does kind of remind me of talking about AI in 2018, you know, like there was buzz around the horizon, but it's just not quite there yet.

Rob May: Yeah. Yeah. Do we want to get in the into the news for this week?

Anna Kirk: Yes. So, let's do it.

Rob May: What article do you want to talk about first?

Anna Kirk: I mean the big one on people's mind is the release of Gemini 3. Have you guys played around with it at all?

Rob May: I have. I'm a Gemini user over most other I we use Cloud at the office, but I use Geminis for a lot of stuff because the number one thing I use these tools for is to create presentations because I hate that and I don't care that they're kind of like a little bit janky. Like they're close enough and they're they're fast, but you can give it a long prompt and they can't create PowerPoints because it's a proprietary format. So none other tools can really do that. There's probably some tool out there that's licensed and can, but like none of the Claude can't do it, you know, and ChatGPT can't do it. So, but Google will save it directly to Google Sheets after it creates it, whereas Claude will give you some like HTML presentation thing that's annoying. You got to move it. Yeah. Yeah. So, yeah. So, I so I use Gemini a lot and I've played around with it, but I also I don't have like a I don't have like one like a lot of the people have they they go and they run a bunch of prompts and see like I just like what ask it what it can do that's new, you know?

Ben Colman: I mean, I have small kids, eight and five, so they have some pretty bizarre ideas to try. [laughter] Sometimes I have to be a little more focused in, you know, being my own moderation on on what comes out. But, yeah, it's a lot of fun, but also really scary of what you do.

Anna Kirk: Yeah, I'm I personally was a ChatGPT user before switching to my current role and, I've played around with Gemini a little bit. The one thing that I don't like about it is they don't have the functionality to like toggle on and off deep research within a chat. Sometimes it's slow, sometimes it's fast.

Rob May: Is that that kind of challenge?

Anna Kirk: Yeah. Yeah. And like that's really nice with ChatGPT is that you can kind of do it natively in one chat window. But anyway, I mean the reception is generally pretty positive. It's exceeded almost every single performance metric out there. And the biggest piece of news that really resonated with me this week was around the fact that they did so using only their own TPUs. So they didn't use any of Nvidia's GPUs in order to do that.

Rob May: Yeah. CNBC this week had a couple days ago had a like Google be the Nvidia killer because now their TPUs are advanced enough. They're starting to get a lot of other usage from other companies. They're starting to I don't know if they'll sell them like to AWS or Microsoft Azure or anything like that. They'll probably stay on Google Cloud for the near term, but maybe I'm wrong, but it's I don't think anybody's going to kill Nvidia. Like there's so much demand for AI compute. Like we're a long way from even there needs to be a lot of increase in supply to get there.

Anna Kirk: Yeah, I don't think anyone will kill Nvidia, but like it is a good signal that like there are other players out there that can now be self-sustaining. I think Google is probably the biggest one to watch in my opinion.

Ben Colman: It's probably great for Nvidia. Like real competition really advances innovation for us. Our clients will deploy us on anything. So, we've had to be a little bit flexible. We are default biased toward Nvidia, though.

Rob May: Yeah. Well, of course. Yeah. You know, they they they've been the market leader for a long time. What which speaking of like one of our articles this week was Nvidia came out and like with their earnings two days ago, a day ago, whatever when we're recording this, seemed pretty good.

Anna Kirk: Yeah. I mean, they beat expectations again and part of my thing was like how do you beat expectations when a company has consistently beat expectations? Like at some point you need to raise your expectations, [clears throat] right?

Rob May: So I've heard it's like expectations of the expectations. It kind of gets really I don't know.

Anna Kirk: Very meta.

Rob May: But I you know I think it's interesting because it ties back into this idea that we've talked about several times on here about the AI bubble. Yeah. And there's a lot of pieces to that. You know, I wrote a piece two weeks ago on the idea that it's hard to talk about an AI bubble because AI is a very broad thing. And you know, do you mean do you mean data centers? Do you mean chips? Do you mean you know, power and utility companies? Do you mean gas? Yeah. There's parts of those are bubbly and parts of those are less bubbly. But I'm interested related to the bubble conversation. What do you see in terms of like do you think the demand for AI services you see it outstrips supply or they're evenly matched or it's where you would expect for a startup or it's too much or like what what's your view on

Ben Colman: I mean honestly on the cyber security side we need more compute like there's just not enough compute like we've calculated out if we scan every single phone call every single video conferencing there's just not enough compute in the data centers even if you think beyond you know cloud to edge to I don't know on device as well so for us you know I'm sure there's bubbles in some places. I think that here we still have a lot of room to go.

Rob May: Yeah. Will Reality Defender run at the edge at some point? You like will you guys be able to run on a microcontroller in a you know or NPU in some kind of small device?

Ben Colman: Right now we have quantized our models. So we've compressed them to run in lower compute environments and the future is compute wherever the compute is.

Anna Kirk: Interesting. When I look at like the flow of cash within certain players within this ecosystem, it does start to get a little bit bubbly when you know I've mentioned this multiple times but to beat a dead horse here like Nvidia has invested a hundred billion into OpenAI so that OpenAI can go back and give them more money right

Ben Colman: it's not always money someone's investing compute or I don't know

Anna Kirk: exactly yeah sometimes it's a little bit more of a favorable business arrangement but I do think you know where if the long lead time for these GPUs loses around what like 12 months then we're this is such a lagging indicator that is it even indicative of where we are right now like yes it beat expectations but it to your point beat expectations from the demand that we had a year ago so what does that what does that say about the state of the market

Rob May: yeah I think I mean I think AI demand is as re I if people are worried about the bubble because they think the demand's not there or the ROI is not there. I think they're wrong. I think you'll see that if they're worried that the valuations are getting out of whack comp like that could happen in any market at any time, right?

Ben Colman: Just where people are the interesting part is the financing of it. You know, if you're something has a year lifetime, you're financing it on a kind of four or five year depreciation schedule. Someone's having some challenges. It's great for the companies that are financing it might be a challenge for the lenders.

Rob May: Yeah. So I want to talk about we don't we don't normally do research papers here but there was a really interesting one that came out that's cyber security related which is the title is adversarial poetry as a universal single turn jailbreak mechanism in large language models. So basically if you make the model craft a poem you can put it in a state where it'll tell you something it's not supposed to tell you because of the poem.

Anna Kirk: Yeah. Because of any poem or a specific poem.

Rob May: Well, it says in here 62% for handcrafted poems and 43% for metaprompt conversation. I don't know what that means, but yeah, but anyway, it's but it it gets to your point about where this is going, which is people are getting more creative. You have hackers now who are like, "Oh, the model won't disclose this information. What if we ask it to write a poem and the poem is supposed to you know it's like

Ben Colman: I mean we have a chat on our website and we found out that somebody was communicating it using meme images and we're like well one is happening and then two wow they're actually communicating the bot is responding to the meme which is bizarre. So anyhow clearly a lot of things going under the hood there.

Rob May: Yeah. Yeah. What's the most interesting deep fake or related thing that you've seen happen? Like is there one story that you always tell when you speak publicly where you're like, "Man, these people did the strangest thing to try to hack somebody or That's [clears throat] interesting."

Ben Colman: I mean, for us it's the ones that are the most apparent. Like we try and stay away from like the really scary ones which we don't really want to talk about publicly but the ones that move the stock market those are the ones that I am fascinated with because you could release them in such a way that they just go viral quickly and there was one that looked like an explosion of the Pentagon that led to a hundred billion dollar flash crash of the market and the fact that no one stopped it. Yeah, we could have if they platforms required to scan for it. Just shows how quickly things can affect beyond that, you know, defig audio being injected into Ukrainian walkie-talkie traffic, things like that. I mean, wherever there is a real-time communication, there's nothing stopping average people with average technology to try to see how they could move the market or affect geopolitics.

Rob May: As the world gets more automated, the viral impact across all kinds of groups is going to could happen and scale and toggle faster, right? From a deep fake or whatever. So, yeah.

Ben Colman: Or a voice agent.

Rob May: A voice agent.

Ben Colman: We're now working with voice agent companies that want to make sure they're not talking to another voice agent. If you never heard ever heard I talk in AI, it quickly devolves. But also that's a lot of expensive waste when you should have a data data connection.

Anna Kirk: Yeah. How are you beginning to quantify something like this? Like how how do you even begin to know how prevalent of a problem it is if like no human can detect it?

Ben Colman: I mean a few years ago we couldn't prove that it was a problem. We knew it was theoretically possible but only when we started scanning a lot of data on behalf of our clients and we can't see their data. So then they have to let us know what they're finding with our solution. But I think the most recent data is like one out of every 100 calls has some level of AI. Maybe a robocall, maybe an agent voice, maybe a deep fake. Video's still a little bit behind the scenes because the compute requirement is still a little bit too high. But, you know, to your point about evolution of GPUs, this time next year, real time video deep fakes will be just as prevalent as audio. So, turning on the news is going to have a whole different experience.

Rob May: Like, is this even real? What do you do you have any suggestions for I mean you mentioned something a little bit earlier in the in the podcast about just maybe like that implied like maybe we need just to introduce more friction in the process of trust and some of this stuff like what's the advice that you give you know co-workers, friends, kids about like like should you literally just are we going to have to change the way we view this stuff as a society and just be like yeah you know

Ben Colman: there's no silver bullet. Reality defender defect detection is not the beall end all. It's a bit of a perimeter strategy. You need multiple perimeters to protect yourself. So, even with my my kids, we just got our eight-year-old a watch. It's a very locked down watch. It's not an Apple Watch. You can't make calls to only I don't know, my wife and I and our and our parents. And so what we've discussed with him is that if you ever hear me, you just ask me some question. AI can fake my voice, my face, but it can't access what I know.

Rob May: Yeah. So ask it about dinner last night.

Ben Colman: You know, you could say have a code word. That gets kind of hard in a heat of the moment. But just to have a conversation. Yeah. It might sound like me, but it won't know what I know.

Rob May: Yeah. Unless I publish it online or in a podcast, then it's or you wear one of those pins people are developing that records everything you do all day. But like, yeah, that's kind of scares me. There's a I think I do think you're going to see people want to keep some humanness in as the world becomes more this way. And I think you're going to see I you've seen it in a lot of things, right? You've seen it with the return of people coming back to paper books. Yep. You've seen it with the return. We went through a phase where people didn't wear watches, right? Because you were like, "Oh, my time's on my cell phone." And we just wore watches to, you know, to tell time and now they've come back as a status symbol or fashion.

Ben Colman: I tried to go digital books and my kids said, "No, can we get some of those books like you used to have?"

Rob May: Yeah. Wow. [laughter] I'm really old. And printed those those snap cameras that automatically print the polaroids have come back like

Anna Kirk: Yeah. My on the flip side, we are making AI generated custom books with our kids in the book, which is super cool. They never had that when I was little. It's fun that they have a bunch of books where they're somehow a character in it.

Rob May: Yeah, that's pretty awesome.

Ben Colman: It does. It does raise a lot of questions for me about like how do you be a parent in this day and age? It's a technology balance and I don't think we figured it out ourselves, but every day there's more technology.

Rob May: Yeah. So I think one final thing that we want to talk about today is Trump made an executive order on AI to sort of take a lot of the power away from the states and sort of have it centralized and his you know quote here that Anna put in here for for us was you can't go through 50 states. You have to get one approval. 50 is a disaster because you'll have one woke state and you'll have to do all woke. [laughter]

Anna Kirk: you'll be a direct, you know, I know like I didn't put it in there because it was eloquent. [laughter]

Rob May: Yeah. To me to me wokeness is the wrong reason to do this or whatever. It's also funny because this is a very sort of anti- conservative rightle leaning strategy. Like typically, you know, right-leaning people are federalists or libertarians and believe that states should have the rights and you should run a lot of experiments and that's good for the economy. What's your so what's your take? Have you guys you know looked at these executive orders and do you have any opinions on

Ben Colman: Yeah. Mean I can't speak to the politics but I can say that any regulations on anything digital they they just don't work on a state-by-st state level. You know we have rules around you know different levels of non-consensual AI pornography at state level they just don't work because the internet doesn't stop at the edges of the state. And so we end up spending a lot of time with state officials but even more time with you know folks in DC. And what's really fascinating is that what we're doing is truly bipartisan. You know, last year, oh yeah, I gave testimony in the Senate to Senate Judiciary Committee. We had Senator Hawley, Senator Blumenthal, who tend to be opposite a lot of things. And this one, they were arm-in- arm. They allowed us to do a live deep fake of them. In this area, it's a universal feeling that like we need something now. It's just taking a little bit longer than we would have we would have hoped.

Rob May: Yeah. Any thoughts on that, Anna?

Anna Kirk: On the I mean I think usually when you have regulatory inconsistency across jurisdictions it does make it more difficult for people not only to develop technologies around but also interpret right I'm so I'm very pro a federal mandate actually Trump and I are on the same page on this one

Rob May: okay for once [laughter] that was not a deep fake she actually said that cool and actually one more article we have today and this will be interesting to see if this applies to reality defender. Yann LeCun is leaving Meta and it's because he does not believe in the future of LLMs and he wants to create what are called world models which are a little more grounded in physics and how the world works and everything else. This going to be a problem for is this going to help deep fakes or not? Because you mentioned that we're already at a point where or you mentioned something earlier in the podcast that we they they don't understand the physics of maybe how people move or how things are put together. And so you think world models are going to make it harder.

Ben Colman: I think it's going to actually create more opportunities for the defenders than for the attackers because for us, the more compute we have, the more things we can do to identify different kinds of features we can extract from what we're seeing or what we're hearing. But he's brilliant. He's going to need a lot of compute. I know that.

Rob May: Yeah, it'll be interesting.

Anna Kirk: The only one. Yeah. Do you want to wrap it up?

Rob May: So Ben, typically we wrap up these podcasts with a New York specific question. So it'll be like your new your favorite New York X. And this week is place you are most likely to stop before you go home when it's past midnight after a fun evening.

Ben Colman: Oh my goodness.

Rob May: So if you need a minute, I have Well, you know, this is a this would have been an interesting one. I wish Ryan was here this time because what what we lose when Ryan's not here is Ryan gives very sophisticated answers to these.

Anna Kirk: I feel like is he past midnight a lot?

Rob May: No, no, no. Just just to all our New York, he's just like a deep thinker and he has like a reverence for New York that I think I don't necessarily have and I don't know if Anna has it, but like like if you ask like what do you love about New York, right? I always give a stupid answer like there's pizza places every block or whatever. And Ryan's always like well the history of the things that have happened here and how they tied a human and we're always like I'm always like wow much more deeply thoughtful. I always wish I'd given Ryan's answer and [laughter] he's not here for this one so it' be interesting to see what he would say.

Anna Kirk: We didn't even We have not made fun of Ryan at all this entire podcast.

Rob May: Yeah, I know. Well, we got we got started a little bit late

Ben Colman: cuz we could make a deep fake if you guys help.

Rob May: Yeah, let's do that. Put him right here. Hey, Ryan. [laughter]

Anna Kirk: Yeah, Ghost of Ryan.

Rob May: I'll give my answer. So, I So, so it depends. If if I'm going out in West Village or Soho and I'm walking back cuz I live in Chelsea, it's normally I'm I'm normally pretty and it's late at night. Like I'm normally fight. I normally just walk back, right? For some reason if I'm coming from, you know, upper east side, upper west side, Midtown or whatever, and I'm walking back to Chelsea, man, that Two Brothers Pizza at 25th and I guess 6th.

Ben Colman: I know that one.

Rob May: Yeah. It's like it's always open like And then there's a crowd like after. Yeah. And it's like they got both sides open cuz they're on the corner. So, and it's I just, you know, and then I can get a slice of pizza. I walk a few more blocks. It's like perfect timing to eat the slice. It's just it's my kryptonite, man. It's like I

Anna Kirk: How many pieces do you got?

Rob May: Usually just one. So, I am I am usually okay. And plus, I'm walk like, you know, if I'm walking from, you know, 45th Street, you know, it's I probably burn off a slice of pizza.

Ben Colman: So, my my first apartment was on 25th and 6th, so I am deeply deeply familiar with that with that pizza place. The bad part though is when you don't have a block to walk it off. Like, when you're just shuffling back and forth between your apartment and the pizza place, it's bad news. [laughter]

Anna Kirk: I Okay, so I'm a proud Upper Westside resident. Tiki Chick on 85th in Amsterdam. Shout out to them. They have the best hot honey chicken sandwich you will ever have, and you can get it for like $5 any time of night.

Rob May: That doesn't sound like something that's going to help you sleep well if you eat that after midnight.

Anna Kirk: Well, pizza won't either.

Rob May: Oh, wow. [laughter]

Ben Colman: I mean, I'm a big fan of Mexican food. I'm from California, and it's sometimes hard to find it in New York, especially after midnight. There is a bodega restaurant called Benonuto and Tribeca that makes a wrap that's kind of burrito-ish and you know they have every salad ingredients. I end up I don't know making kind like a Frankenstein version of what I want with different kinds of tomatoes and onions and everything else and I can taste it right now. Now the challenge is they always want to cut it in half. It's always a race saying please don't cut it. They'll heat it up. Yeah. And then I don't want to cut because then it's a mess.

Rob May: Yeah. You got to say it right at the right time.

Ben Colman: Yeah. Multiple times.

Anna Kirk: Yeah. It's dangerous that I know about that. I'm definitely gonna be using that one. Oh, no. Oh, just that place. Oh, yeah. Yeah.

Rob May: Awesome. Ben, if people want to learn more about Reality Defender, what's the best way?

Ben Colman: If they're a developer, go to realitydefender.com. If they email us or comment on our chatbot, we'll give them additional credits to scan. Otherwise, feel free to reach out to me. LinkedIn, email. I'm benrealitydefender.com.

Rob May: Ben, thanks so much for having or for joining us today. Yeah. I know you're you're a busy person, so we really appreciate it.

Ben Colman: Well, that chicken sandwich pizza's still good, but that sounds pretty wild. [laughter]

Rob May: It it was worth it to do the podcast for that tip.

Anna Kirk: Yeah, we should have started with that. Yeah. For those of us for those of you listening, send us your swag. We will wear it on the show. I know I'm not representing Rob's representing Neurometric today.

Rob May: Well, this is my own company.

Anna Kirk: A little bit. Yeah, [laughter] me sad.

Rob May: And also, I will say, our Christmas show, our holiday show, at the end of the year is going to feature a handful of startups and companies. Reach out to us if you want to be featured on that show, and we'll start taking applications in the next couple of weeks. And then subscribe, like us, all the other stuff that you're supposed to do, listen to our other shows, and go check out go check out Reality Defender. Thanks for thanks for watching.

Ben Colman: Thank you guys.