1 00:00:00,080 --> 00:00:01,760 Puma, welcome to the NRC podcast. 2 00:00:02,080 --> 00:00:04,200 Thanks for having me. Awesome, you want to introduce 3 00:00:04,200 --> 00:00:06,720 yourself briefly. Yeah, so I'm Co founder CEO of 4 00:00:06,720 --> 00:00:10,080 succinct succinct applied cryptography company. 5 00:00:10,880 --> 00:00:15,480 We probably best known for making the fastest 0 knowledge 6 00:00:15,480 --> 00:00:19,600 virtual machine ZKVM for short in the world known as SP1. 7 00:00:20,240 --> 00:00:24,240 For those of you who aren't aware, ZK is this really 8 00:00:24,240 --> 00:00:27,720 powerful cryptography technique where it lets you prove to 9 00:00:27,720 --> 00:00:30,320 someone else something is true without revealing all the 10 00:00:30,320 --> 00:00:33,240 details. Kind of the canonical real world 11 00:00:33,240 --> 00:00:37,560 example is that if I can prove to you that you know I'm over 21 12 00:00:38,000 --> 00:00:40,680 without revealing my birthday or my home address or anything like 13 00:00:40,680 --> 00:00:42,400 that. So instead of showing a driver's 14 00:00:42,400 --> 00:00:45,320 license at a bar, you can just show them a proof of you know 15 00:00:45,440 --> 00:00:51,600 that you're of age. We built this ZKVM which how can 16 00:00:51,600 --> 00:00:53,320 I describe it? I would say it's somewhat like a 17 00:00:53,320 --> 00:00:55,000 foundation model for cryptography. 18 00:00:55,000 --> 00:00:58,520 So if you want to prove really complex statements in ZK, such 19 00:00:58,520 --> 00:01:01,320 as a roll up state transition function or like very complex 20 00:01:01,320 --> 00:01:04,560 predicates, the thing we built makes it super easy. 21 00:01:04,560 --> 00:01:07,360 You just write normal code, you stick it in to ask P1 and 22 00:01:07,360 --> 00:01:10,720 outcomes the proof. And then, yeah, we made it super 23 00:01:10,720 --> 00:01:12,480 fast and really easy to use, which is awesome. 24 00:01:13,600 --> 00:01:15,680 And I would say right now, succinct. 25 00:01:15,880 --> 00:01:18,440 Although historically our ZKVM has been used mostly for 26 00:01:18,440 --> 00:01:21,120 approving block chains and proving, proving roll up state 27 00:01:21,120 --> 00:01:22,840 transition functions and things like this. 28 00:01:22,840 --> 00:01:26,080 And like Ethereum and other chains, right now we're really 29 00:01:26,080 --> 00:01:29,640 excited about the potential of cryptography to solve a lot of 30 00:01:29,640 --> 00:01:33,000 the problems that AI poses. I think Bologi has been an 31 00:01:33,000 --> 00:01:35,640 extensive tweeter about this topic for many years. 32 00:01:35,640 --> 00:01:37,360 You're very ahead of your time, honestly. 33 00:01:38,640 --> 00:01:42,280 And so, yeah, I think honestly, that's probably one of the most 34 00:01:42,280 --> 00:01:44,240 important things cryptography can do right now. 35 00:01:45,240 --> 00:01:47,320 And there's like a finally a clear catalyst. 36 00:01:47,320 --> 00:01:50,440 Every model released where the image stuff or video stuff gets 37 00:01:50,440 --> 00:01:52,760 better and better. It's like we need cryptography 38 00:01:52,760 --> 00:01:54,840 as a defense. So I think it's time for 39 00:01:54,840 --> 00:01:56,960 cryptography to be on like a societal stage right now. 40 00:01:56,960 --> 00:01:58,680 It's like a solution to all day I stuff. 41 00:01:59,840 --> 00:02:01,920 So I'm very excited about that. Awesome. 42 00:02:01,920 --> 00:02:07,920 So yeah, I actually actually many years ago, it partly 43 00:02:07,920 --> 00:02:12,240 actually because of AI, but also with social media when people 44 00:02:12,240 --> 00:02:15,640 are talking about misinformation, disinformation 45 00:02:15,640 --> 00:02:19,640 and so on and so forth. Years ago, I remember I tweeted 46 00:02:19,640 --> 00:02:22,840 something and I was like, oh, so you want to ban lies on the 47 00:02:22,840 --> 00:02:25,680 Internet? OK, give me a function that says 48 00:02:25,680 --> 00:02:29,720 whether the Reman hypothesis is true, right. 49 00:02:30,000 --> 00:02:33,880 And so, you know, that's a reduction ad absurdum where we 50 00:02:34,120 --> 00:02:36,720 we don't know whether it's true and it could be true and it's 51 00:02:36,720 --> 00:02:39,120 plausible that it's true. But there's many things in math 52 00:02:39,120 --> 00:02:42,920 which have really arcane counter examples that, you know, you get 53 00:02:42,920 --> 00:02:45,840 up to N equals whatever and it it's actually not true. 54 00:02:47,320 --> 00:02:53,240 And so, but as I thought about that, I was like, well, how 55 00:02:53,240 --> 00:02:58,760 would you code Trugal TRUGLE? You know, if you were actually, 56 00:02:59,080 --> 00:03:01,400 you know, going to do it, right, how would you do it? 57 00:03:01,760 --> 00:03:05,680 And The thing is, LLMS get you some of the way towards that, 58 00:03:05,960 --> 00:03:07,120 right? Because they will take a 59 00:03:07,120 --> 00:03:09,680 statement and they'll do at least a probabilistic search at 60 00:03:09,680 --> 00:03:11,920 the literature and pull things up and so on and so forth, 61 00:03:13,040 --> 00:03:14,320 right? And The thing is though, of 62 00:03:14,320 --> 00:03:16,600 course, then those assertions themselves need to be 63 00:03:16,600 --> 00:03:20,560 underpinned, the citations. And then that's how you get to 64 00:03:20,560 --> 00:03:25,960 on chain everything. And so my view is like with LMS, 65 00:03:26,440 --> 00:03:28,800 actually you can, you can kind of show a version of this today. 66 00:03:29,080 --> 00:03:36,080 If you ask any LLM to summarize some major crypto hack, it will 67 00:03:36,080 --> 00:03:39,600 show you probably some link that shows some on chain block 68 00:03:39,600 --> 00:03:41,200 explorer record among other things. 69 00:03:41,920 --> 00:03:45,680 And so that's currently only used to document financial 70 00:03:45,680 --> 00:03:48,480 things like the on chain transaction, you know, during, 71 00:03:48,600 --> 00:03:51,000 let's say FTX had a hack or whatever during that period, 72 00:03:52,360 --> 00:03:54,640 right? But as more and more things get 73 00:03:54,640 --> 00:03:58,720 logged on chain, then more and more references from L Lems will 74 00:03:58,720 --> 00:04:01,960 point to on chain events and we get what I call the Ledger of 75 00:04:01,960 --> 00:04:03,240 record. And I think succinct could be 76 00:04:03,240 --> 00:04:04,960 maybe a big part of that. So you had some slides. 77 00:04:04,960 --> 00:04:06,280 So maybe you you want to go through your slides? 78 00:04:06,840 --> 00:04:08,240 Oh yeah. What's your background by the 79 00:04:08,240 --> 00:04:12,320 You're born in the US, You what's? 80 00:04:12,320 --> 00:04:18,519 Your What's your spiel? The USI went to school at MIT, 81 00:04:18,519 --> 00:04:20,320 was a double major in math and CS. 82 00:04:20,320 --> 00:04:22,840 I've always really loved math. So that's kind of how I got into 83 00:04:22,840 --> 00:04:26,720 ZK. And yeah, I mean, ZK and 84 00:04:26,720 --> 00:04:29,680 cryptography have a lot of fascinating math, and that was 85 00:04:29,680 --> 00:04:33,840 really big the draw for me. And actually before Zika, I was 86 00:04:33,840 --> 00:04:36,800 doing some AI stuff. So I was like at Google Brain 87 00:04:36,800 --> 00:04:39,400 doing research into like early LLM. 88 00:04:39,400 --> 00:04:42,920 So this is pre GPTI was doing some stuff with Bert back then. 89 00:04:43,320 --> 00:04:45,840 So I'm familiar with that world. And now it's like exciting to 90 00:04:45,840 --> 00:04:47,480 see the synthesis. Yes. 91 00:04:47,480 --> 00:04:51,440 I mean, you know, The thing is, I, you know, I was actually also 92 00:04:51,440 --> 00:04:56,600 in machine learning prior to the deep learning era, but from the 93 00:04:56,600 --> 00:05:00,040 standpoint of genomics and diagnostics and, and whatnot. 94 00:05:00,560 --> 00:05:02,720 And, you know, just to digress on that for a second before we 95 00:05:02,720 --> 00:05:06,640 get into the ZK stuff like, you know, all the stuff with hidden 96 00:05:06,640 --> 00:05:08,960 markup models and conditional random fields. 97 00:05:08,960 --> 00:05:16,720 And, you know, it was surprising to me that Transformers worked 98 00:05:16,720 --> 00:05:20,640 as well as they did to get long range context in there. 99 00:05:20,840 --> 00:05:24,480 It's even more surprising to be that diffusion models work and 100 00:05:24,560 --> 00:05:28,840 and yet they do that, you know, you wouldn't necessarily intuit 101 00:05:28,840 --> 00:05:32,320 from the equations that they would work as well as they do in 102 00:05:32,320 --> 00:05:33,880 practice. I don't know if you have any 103 00:05:33,880 --> 00:05:35,800 thoughts on that. Maybe talk about that and then 104 00:05:35,800 --> 00:05:37,960 go to the next. I mean, when you're working on 105 00:05:37,960 --> 00:05:41,720 Bert, did you think obviously there was there were people who 106 00:05:41,720 --> 00:05:45,080 had the graphs of scaling and here's how it's going to go 107 00:05:45,360 --> 00:05:47,120 right. So there was some intuition that 108 00:05:47,120 --> 00:05:51,800 it could maybe get there, but but it but it worked a lot. 109 00:05:51,920 --> 00:05:55,680 I mean, the jump between GPT 2 and GPT 3 and then ChatGPT in 110 00:05:55,680 --> 00:06:02,000 terms of usability was very non linear I think from, you know, 111 00:06:02,040 --> 00:06:02,800 right, go ahead. Maybe. 112 00:06:03,400 --> 00:06:05,960 Were you surprised by that? Oh, I mean, absolutely. 113 00:06:06,560 --> 00:06:09,320 I think even the close people closest to the metal on this 114 00:06:09,320 --> 00:06:13,560 stuff seemed surprised at like how well it's going. 115 00:06:13,560 --> 00:06:16,280 And even today, like the level of math problems they're solving 116 00:06:16,280 --> 00:06:20,400 and stuff like that. Yeah, I, I would say I, I was 117 00:06:20,480 --> 00:06:25,640 very surprised that this process, which, you know, with 118 00:06:25,640 --> 00:06:27,640 deep learning, there aren't really that many proofs. 119 00:06:27,640 --> 00:06:30,840 Like in cryptography, everything we do is proven like it's 120 00:06:30,840 --> 00:06:34,120 deterministic. You have very concrete bounds 121 00:06:34,120 --> 00:06:38,000 and proofs for everything. Yeah. 122 00:06:38,000 --> 00:06:41,160 And alarms is just, it's like, why does deep learning work kind 123 00:06:41,160 --> 00:06:44,120 of vibes based? But right now the vibes are 124 00:06:44,120 --> 00:06:45,240 really good. It works really well. 125 00:06:45,680 --> 00:06:48,000 Right. And I think it's funny because, 126 00:06:48,240 --> 00:06:49,800 well, actually go through your talk and let's talk. 127 00:06:49,920 --> 00:06:52,120 Go prove what's real. OK, cool. 128 00:06:53,920 --> 00:06:59,320 So, yeah, I mean, I think you one of your favorite quotes is 129 00:06:59,520 --> 00:07:01,960 AI makes everything fake. Crypto makes a real gun. 130 00:07:03,840 --> 00:07:07,480 And yeah, I think like the problem is very clear now. 131 00:07:07,720 --> 00:07:10,520 AI makes everything fake and we're seeing that every single 132 00:07:10,520 --> 00:07:12,240 day. So it's like whether it's this 133 00:07:12,240 --> 00:07:15,440 DoorDash driver, you know, kind of faking the delivery. 134 00:07:15,520 --> 00:07:18,320 Maybe that's a little bit of a trivial example that went quite 135 00:07:18,320 --> 00:07:22,080 viral on X to something like, oh, is Jeffrey Epstein still 136 00:07:22,080 --> 00:07:24,120 alive? Or it's like the White House 137 00:07:24,120 --> 00:07:27,720 posting digitally altered pictures or it's politicians 138 00:07:27,720 --> 00:07:31,360 getting deepfaked or, you know, celebrities getting deepfaked 139 00:07:31,360 --> 00:07:34,320 doing all sorts of things. Like the problem is I think 140 00:07:34,320 --> 00:07:38,360 today is like extremely clear and it's only getting to be 141 00:07:38,360 --> 00:07:40,480 worse and worse as these models get better and better. 142 00:07:41,320 --> 00:07:44,320 So like the seed dance release recently of like the really good 143 00:07:44,320 --> 00:07:49,000 video models that like, you know, can really impersonate any 144 00:07:49,000 --> 00:07:52,200 celebrity or any person. Like it's very clear that AI 145 00:07:52,200 --> 00:07:54,360 makes everything fake is kind of like a huge problem for the 146 00:07:54,360 --> 00:07:57,040 Internet. So yeah, that's like the problem 147 00:07:57,040 --> 00:08:01,400 statement. And then I think like one really 148 00:08:01,400 --> 00:08:03,160 interesting thing. So people have identified this 149 00:08:03,160 --> 00:08:05,520 problem statement. It's not, you know, that hard to 150 00:08:05,520 --> 00:08:07,280 understand why it's so problematic. 151 00:08:08,200 --> 00:08:11,280 I would say the state-of-the-art right now for trying to detect 152 00:08:11,320 --> 00:08:15,920 AI is use AI. So people have like, train these 153 00:08:15,960 --> 00:08:19,920 AI detectors to train models to say, hey, is something real or 154 00:08:19,920 --> 00:08:21,960 is like something from the model. 155 00:08:22,400 --> 00:08:26,160 And recently it's a thing. We did this, you know, 156 00:08:26,160 --> 00:08:29,880 benchmarking study to, you know, evaluate those claims and say, 157 00:08:29,880 --> 00:08:31,920 hey, does AI detection actually work? 158 00:08:32,440 --> 00:08:37,840 And we published, you know, this data set of realistic AI images 159 00:08:37,840 --> 00:08:41,360 and tried to benchmark all the leading commercial detectors. 160 00:08:42,039 --> 00:08:45,480 And you can actually go to the website AI detection dot sync 161 00:08:45,480 --> 00:08:48,680 dot XYZ. But the resounding answer was 162 00:08:48,680 --> 00:08:52,200 like, yeah, the AI detection stuff is not robust and it just 163 00:08:52,200 --> 00:08:54,960 does not work. I'm going to slightly argue with 164 00:08:54,960 --> 00:08:59,320 you on this maybe, which is to say on text as opposed to 165 00:08:59,320 --> 00:09:05,200 images, right? So much so this reminds me a 166 00:09:05,200 --> 00:09:09,320 little bit of I'm not, I'm not really arguing with the results 167 00:09:09,320 --> 00:09:11,640 of your paper, but on the macro thing, right? 168 00:09:11,800 --> 00:09:19,320 And it's so with Snapchat, you know, it has a deterrent to 169 00:09:19,320 --> 00:09:22,720 someone taking a screenshot of, you know, like like a 170 00:09:22,760 --> 00:09:26,080 disappearing message. Now, of course you or I or 171 00:09:26,080 --> 00:09:27,880 someone who's a computer scientist will say, well, 172 00:09:27,880 --> 00:09:30,760 there's still the analog hole. You can just hold up another 173 00:09:30,760 --> 00:09:34,560 phone and record it. And you know, if you want to, 174 00:09:34,560 --> 00:09:36,240 you can just take a second phone and record it. 175 00:09:36,240 --> 00:09:38,960 And that doesn't have the you know you can defeat it with a 176 00:09:38,960 --> 00:09:41,520 sufficiently motivated attacker relatively easily, right? 177 00:09:42,000 --> 00:09:46,120 However, most people aren't that motivated, and so the simple and 178 00:09:46,120 --> 00:09:50,240 dumb screenshot detection thing sets the norm and makes it 179 00:09:50,240 --> 00:09:52,320 relatively hard to do screenshots, right? 180 00:09:52,640 --> 00:09:56,280 Similar to how you know, you could people could work around 181 00:09:56,280 --> 00:09:59,520 the Twitter 140 character limit by pasting in screenshots of 182 00:09:59,520 --> 00:10:01,160 140. You know, more than 140 183 00:10:01,160 --> 00:10:03,560 characters, but they didn't for a long time, right? 184 00:10:04,120 --> 00:10:07,880 And my view is that there's a lot of AI text on X, for 185 00:10:07,880 --> 00:10:10,800 example, that at least I can trivially detect. 186 00:10:11,120 --> 00:10:14,560 It's not this, it's that and the M dashes and and so on and so 187 00:10:14,560 --> 00:10:16,560 forth, right? And there's certain people who 188 00:10:16,560 --> 00:10:19,960 just are clearly AI posters because of just the style. 189 00:10:19,960 --> 00:10:22,000 It's like this overdramatic kind of style. 190 00:10:22,000 --> 00:10:25,080 It's, it jumps out to you immediately when you see it 191 00:10:25,080 --> 00:10:27,920 because you see it a lot. It's like seeing the same person 192 00:10:27,920 --> 00:10:32,040 writing over and over, you know, and pangram.com or something 193 00:10:32,040 --> 00:10:37,200 like that feels pretty good at detecting ChatGPT type slop, 194 00:10:37,200 --> 00:10:39,640 which you see a lot of and clawed and ChatGPT, for whatever 195 00:10:39,640 --> 00:10:42,560 reason, a very similar text voice, I think right on this 196 00:10:42,560 --> 00:10:44,880 kind of thing. Yeah, I mean, I guess they're 197 00:10:44,880 --> 00:10:46,600 trained on the same data to a certain. 198 00:10:47,280 --> 00:10:54,160 Yes, whereas images, you know, maybe I, I guess it depends on 199 00:10:54,160 --> 00:10:57,520 the class of image. I mean, obviously with hands and 200 00:10:57,520 --> 00:10:59,280 things like that. Gymnasts, they're finally 201 00:10:59,280 --> 00:11:02,640 starting to get good with gymnastics with, with sea dance 202 00:11:02,640 --> 00:11:05,560 because those are unusual poses, but they do like physics 203 00:11:05,560 --> 00:11:07,320 simulations, I guess to to train them. 204 00:11:07,320 --> 00:11:08,640 I don't know, maybe you have a thought on that. 205 00:11:08,640 --> 00:11:11,800 You understand my point, right? Like AI detection may not work 206 00:11:11,800 --> 00:11:15,480 100% of the time, but for text, I think it currently works well 207 00:11:15,480 --> 00:11:21,000 enough to get a lot of the ChatGPT type slop at a fairly 208 00:11:21,000 --> 00:11:22,240 high. Like you can certainly see it 209 00:11:22,240 --> 00:11:25,080 visually. You know, like a human can see 210 00:11:25,080 --> 00:11:28,200 it, then it then if it, if it's unsubtle enough for us to see. 211 00:11:28,280 --> 00:11:32,120 Let me pause there. Yeah, I I do agree that the tech 212 00:11:32,120 --> 00:11:36,320 stuff, at least right now, there are these watermarks almost like 213 00:11:36,320 --> 00:11:39,080 the M dash or the patterns you were saying like, oh, it's X, 214 00:11:39,520 --> 00:11:45,000 this is X, not Y. But I, I still think that 215 00:11:46,440 --> 00:11:49,360 similar to images actually, like the study we did basically was 216 00:11:49,360 --> 00:11:51,680 you take an image that an AI generates. 217 00:11:51,680 --> 00:11:53,120 And by the way, these things are pretty good. 218 00:11:53,120 --> 00:11:56,800 Like I actually, I've gotten personally fooled a bunch of 219 00:11:56,840 --> 00:12:00,960 times. So empirically it seems to be 220 00:12:00,960 --> 00:12:04,520 really good. And then we did the study where 221 00:12:04,720 --> 00:12:07,120 you basically perturb the image a bit. 222 00:12:07,120 --> 00:12:11,920 So you blur it or you crop it or you add some like indiscernible 223 00:12:11,920 --> 00:12:15,320 Gaussian noise to the image and then the AI detectors all 224 00:12:15,320 --> 00:12:17,880 completely break. And I think even in text, that's 225 00:12:17,880 --> 00:12:20,840 kind of true, right? And I mean, who's to say using 226 00:12:20,880 --> 00:12:23,840 AI to help you write some of your tweets, maybe that's not 227 00:12:23,840 --> 00:12:25,400 even a bad thing necessarily, right? 228 00:12:25,400 --> 00:12:28,040 Like ultimately, like content is content and maybe you're saying 229 00:12:28,040 --> 00:12:29,720 something interesting with the AI help. 230 00:12:29,960 --> 00:12:32,320 But like a lot of people do these tricks where they're like 231 00:12:32,520 --> 00:12:35,600 get the output from GPT and then they tell GP remove all the M 232 00:12:35,600 --> 00:12:39,760 dashes and then then it's not detectable. 233 00:12:40,000 --> 00:12:41,760 No, it's true. I I guess The thing is. 234 00:12:41,760 --> 00:12:44,640 So here's my view on that. It's my emerging view. 235 00:12:44,640 --> 00:12:46,640 So at least here's our current standard on this. 236 00:12:47,000 --> 00:12:56,920 We so at NS our our rule is no public undisclosed AI, right. 237 00:12:57,080 --> 00:12:59,440 So why do I say that? Well, first is people can just 238 00:12:59,440 --> 00:13:04,240 go full AI and full AI means like because AI is a shortcut. 239 00:13:05,080 --> 00:13:10,240 Yeah, and as a shortcut, I think it's a good term because people 240 00:13:10,240 --> 00:13:12,640 can take too many shortcuts and they fake it and they don't know 241 00:13:12,640 --> 00:13:15,040 what they're doing and so on. The more expert you are, the 242 00:13:15,040 --> 00:13:17,200 more legitimate it is to take a shortcut because you know how to 243 00:13:17,200 --> 00:13:20,960 do it the normal way, right? And it's like writing down a 244 00:13:20,960 --> 00:13:23,440 theorem without doing the full proof every time it's right 245 00:13:23,440 --> 00:13:25,840 using a function call route that there's a reason that people use 246 00:13:25,840 --> 00:13:27,600 shortcuts. OK, but they can overuse them 247 00:13:27,600 --> 00:13:30,680 fine. So the alternative is no AI, 248 00:13:30,680 --> 00:13:32,600 which is a lot of people actually are going to go to. 249 00:13:32,920 --> 00:13:36,480 And there's like an anti AI fine, but no public undisclosed 250 00:13:36,480 --> 00:13:39,600 AII think in 4 words. It captures so you can use 251 00:13:39,600 --> 00:13:42,560 private AI and that's undisclosed because you're going 252 00:13:42,560 --> 00:13:44,280 and editing your own stuff, right? 253 00:13:45,320 --> 00:13:47,360 I mean, you're, you're or like you're editing code. 254 00:13:47,360 --> 00:13:48,960 Who cares? You're using it for yourself, 255 00:13:48,960 --> 00:13:52,120 right? Public disclosed AI when whether 256 00:13:52,120 --> 00:13:55,480 it's a watermark at the bottom right or it's like an animation, 257 00:13:55,480 --> 00:13:59,200 a comic, a movie, something like that, no one can get mad because 258 00:13:59,200 --> 00:14:01,600 you're not trying to pull one over on on somebody, right? 259 00:14:02,240 --> 00:14:06,360 It's public undisclosed AI that gets people mad. 260 00:14:06,720 --> 00:14:12,360 And at least if I analyze my own reaction on that, I don't when 261 00:14:12,360 --> 00:14:16,120 someone is sending me something that's obviously AII think they 262 00:14:16,120 --> 00:14:20,080 are either stupid or lazy. Why? 263 00:14:20,360 --> 00:14:26,760 They're stupid because they can't see the obvious AI tells 264 00:14:26,880 --> 00:14:30,680 like they send an AI slop slide deck or they have AAI web page 265 00:14:30,960 --> 00:14:34,560 that has a lot of like it's one thing if they say, hey, this is 266 00:14:34,560 --> 00:14:36,880 a prototype, check it out. OK, fine. 267 00:14:37,080 --> 00:14:41,360 Right, But that's disclosed AI if it's undisclosed and it's 268 00:14:41,360 --> 00:14:45,000 just got like a wall of AI cause AI tends to, you know, in AI 269 00:14:45,000 --> 00:14:48,240 images, they're more full of people than normal images by 270 00:14:48,240 --> 00:14:49,520 default. You know, if you've noticed, 271 00:14:49,520 --> 00:14:50,960 unless you actually pull that back, right? 272 00:14:50,960 --> 00:14:54,520 Like their outdoor scenes have too many people often, right? 273 00:14:56,000 --> 00:15:01,560 And that's like 1 tell, right? And similarly, AI pages and AI 274 00:15:01,560 --> 00:15:05,240 slide decks are not succinct. Yeah. 275 00:15:06,440 --> 00:15:09,720 They're just really right. And so either they're dumb and 276 00:15:09,720 --> 00:15:14,000 they can't tell what's good, or they're lazy and they're hitting 277 00:15:14,000 --> 00:15:16,920 a few keys and then sending me a bunch of slop and I have to go 278 00:15:16,920 --> 00:15:19,360 through it. And fundamentally they're taxing 279 00:15:19,760 --> 00:15:21,040 the other side. It's like someone leaving a 280 00:15:21,040 --> 00:15:22,640 voice memo for. You. 281 00:15:23,120 --> 00:15:25,760 You know, like I have to verify everything because they didn't 282 00:15:25,760 --> 00:15:28,360 verify everything. And so when they whenever I get 283 00:15:28,360 --> 00:15:33,920 an AI message from somebody, I downweight them because of that. 284 00:15:35,320 --> 00:15:38,160 And I and I down with them as a poster and so on and so forth 285 00:15:38,160 --> 00:15:43,840 because they, they just, they're taking shortcuts in a way that 286 00:15:45,360 --> 00:15:46,960 makes me question their judgement. 287 00:15:48,640 --> 00:15:50,840 Yeah, if he gets good enough that go ahead, say what you're 288 00:15:50,840 --> 00:15:52,960 going to. Say, Well, I think one reason 289 00:15:52,960 --> 00:15:58,040 the fake images and fake text is a little different is I mean, 290 00:15:58,040 --> 00:16:00,480 historically you could just write whatever words you want. 291 00:16:00,840 --> 00:16:03,520 Even pre AI you could just write a bunch of things that were not 292 00:16:03,520 --> 00:16:06,400 true, like you could lie. So I think humans are very used 293 00:16:06,400 --> 00:16:09,840 to critically evaluating the text they see because like 294 00:16:10,360 --> 00:16:12,360 people can always just write whatever. 295 00:16:13,440 --> 00:16:16,440 I think we're much less used to being able to critically 296 00:16:16,440 --> 00:16:20,800 evaluate images we see. Historically, it was pretty 297 00:16:20,800 --> 00:16:22,840 hard. I mean, OK, you had things like 298 00:16:22,840 --> 00:16:26,600 Photoshop and this and that, but like, you know, it'd be pretty 299 00:16:26,600 --> 00:16:31,480 hard to really fake something elaborate or like fake, for 300 00:16:31,480 --> 00:16:34,320 example, the president of the United States doing like a one 301 00:16:34,320 --> 00:16:36,400 minute long video saying whatever. 302 00:16:37,480 --> 00:16:39,000 You just could not have done that in the past. 303 00:16:39,000 --> 00:16:41,760 And now with the AI tools, it's very easy to do that. 304 00:16:41,920 --> 00:16:44,960 So, so it's really interesting you say that and, and I want to 305 00:16:45,240 --> 00:16:48,720 continue your presentation. My I agree and I'll give a 306 00:16:48,720 --> 00:16:53,280 partial counter argument, which is I actually think most of the 307 00:16:53,280 --> 00:16:57,320 images and videos people have seen are television or movies 308 00:16:58,400 --> 00:17:03,200 until recently. And those were actually all 309 00:17:03,200 --> 00:17:08,359 fictional and synthetic. And so they kind of live within 310 00:17:08,359 --> 00:17:12,359 a world where some significant fraction of their inbound 311 00:17:12,359 --> 00:17:16,200 training data is fictional, as seen by the extent to which 312 00:17:16,200 --> 00:17:20,640 people reference, I don't know, Star Wars or The Handmaid's Tale 313 00:17:20,640 --> 00:17:23,640 or something like that, like the, you know, Harry Potter. 314 00:17:23,880 --> 00:17:27,319 That's actually more real for many people than actual history. 315 00:17:28,800 --> 00:17:32,680 But that's like disclosed. It is disclosed, but I don't 316 00:17:32,680 --> 00:17:34,840 think they can actually go ahead say, say, say, say. 317 00:17:35,480 --> 00:17:36,440 Yeah, I was just going to say it. 318 00:17:36,440 --> 00:17:39,320 That is disclosed in that you know it's not, you know, it's 319 00:17:39,320 --> 00:17:41,960 made-up. I, I think you and I know it's 320 00:17:41,960 --> 00:17:46,880 made-up, but I, but I think, OK, here's my argument on this and 321 00:17:46,880 --> 00:17:50,280 let's continue. But the I call it Jurassic 322 00:17:50,280 --> 00:17:54,640 Ballpark, like, you know, Jurassic Park has the scene 323 00:17:54,640 --> 00:17:57,480 where, see, I'm actually referencing A fictional movie 324 00:17:57,480 --> 00:17:58,760 scene to explain fictional movie scenes. 325 00:17:58,760 --> 00:18:01,920 Very meta. OK, so Jurassic Park has the 326 00:18:01,920 --> 00:18:07,000 scene where the dinosaurs have amphibian DNA spliced in because 327 00:18:07,480 --> 00:18:09,880 the scientists didn't know what to make of that part. 328 00:18:09,880 --> 00:18:13,640 So they spliced in amphibian DNA and that leads to, you know, the 329 00:18:13,640 --> 00:18:18,480 the dinosaurs reproducing. The point being that when we are 330 00:18:18,480 --> 00:18:21,360 dealing with a situation that we don't have personal experience 331 00:18:21,360 --> 00:18:26,000 of, like we don't have personal data on, you implicitly rely on 332 00:18:26,000 --> 00:18:30,920 some movie you've seen about that area to tell you how it's 333 00:18:30,920 --> 00:18:33,600 like. For example, unless you've 334 00:18:33,600 --> 00:18:38,840 actually been if, unless you've worked at CIA or you know, 335 00:18:38,840 --> 00:18:42,160 people at Palantir, you don't really understand what the 336 00:18:42,280 --> 00:18:45,640 actual CIA is as opposed to the movie version. 337 00:18:45,640 --> 00:18:49,160 You think the movie version is in the ballpark and even if it's 338 00:18:49,160 --> 00:18:51,000 like more dramatized or whatever, right? 339 00:18:51,840 --> 00:18:54,360 And, and it's often just totally not. 340 00:18:56,120 --> 00:18:58,360 And so that's why I mean, like, you're right that we kind of 341 00:18:58,360 --> 00:19:01,640 know it's fictional, but we don't know what reality is. 342 00:19:02,040 --> 00:19:05,240 And so often we think that the fictional is just a jazzed up 343 00:19:05,240 --> 00:19:07,920 version of the real as opposed to like totally, totally, 344 00:19:07,920 --> 00:19:12,760 totally off. So anyway, so the reason, the 345 00:19:12,760 --> 00:19:14,560 reason I say that is I think there's a huge opportunity. 346 00:19:14,560 --> 00:19:17,400 One of the things I want to fund at some point is people taking 347 00:19:17,400 --> 00:19:20,640 actual history and then using AI to dramatize it. 348 00:19:22,920 --> 00:19:25,360 So now it's actually more fictional. 349 00:19:25,360 --> 00:19:28,120 It's fictional but factual, fictional depiction of real 350 00:19:28,120 --> 00:19:29,880 events, you know? Anyway, keep going. 351 00:19:30,120 --> 00:19:31,400 I mean to digress. Keep going. 352 00:19:31,520 --> 00:19:34,440 There's all these examples and like, it's actually pretty 353 00:19:34,440 --> 00:19:37,000 fascinating, so. You're the receipts. 354 00:19:37,520 --> 00:19:39,120 Yeah. I mean, even, I mean, this is 355 00:19:39,120 --> 00:19:42,160 like kind of maybe a mundane example, but we did all these, 356 00:19:42,480 --> 00:19:45,320 we had a bunch of different categories of like real cases 357 00:19:45,320 --> 00:19:47,800 where AI deepfakes could be somewhat harmful. 358 00:19:48,000 --> 00:19:50,160 And 1 is just, you know, receipts and like 359 00:19:50,600 --> 00:19:53,160 reimbursements. And we had AI generate a bunch 360 00:19:53,160 --> 00:19:56,640 of images that were like taking a real receipt and modifying the 361 00:19:56,640 --> 00:20:00,040 numbers to be much greater than they actually were, like by an 362 00:20:00,040 --> 00:20:02,320 order of magnitude. And then we put them through 363 00:20:02,320 --> 00:20:06,760 these AI image detectors and. It it's turns out like, you 364 00:20:06,760 --> 00:20:09,760 know, they're OK They're like, oh, this is a 36% chance it's 365 00:20:09,880 --> 00:20:11,880 AI. This is a 44% chance it's AI. 366 00:20:11,880 --> 00:20:13,720 Maybe what's the original somethings? 367 00:20:15,520 --> 00:20:18,720 These are actually real photos. No, I mean, but did what did the 368 00:20:18,720 --> 00:20:24,440 original come up as 0%? Oh, I don't have those numbers 369 00:20:24,440 --> 00:20:27,440 here, but I think it was like pretty accurate so. 370 00:20:27,600 --> 00:20:30,800 Because the reason, yeah, the reason I ask is I'd love to see 371 00:20:30,800 --> 00:20:32,640 that data if you can pull it at some point. 372 00:20:32,640 --> 00:20:38,280 Because even if the detector was saying 36% if it could, if it 373 00:20:38,280 --> 00:20:41,240 had variance, you could rescale the axis, you know what I mean? 374 00:20:41,640 --> 00:20:45,680 Like if the if the real photos were left shifted relative to 375 00:20:45,680 --> 00:20:48,360 the fake photos, you could recast it as. 376 00:20:48,800 --> 00:20:50,600 You know, like a binary classifier problem. 377 00:20:51,200 --> 00:20:54,240 Yeah, like basically the distribution of real, like the 378 00:20:54,240 --> 00:20:57,480 distribution of fake. But then the problem is if you 379 00:20:57,480 --> 00:21:01,560 just do simple perturbations to the AI generated stuff like you 380 00:21:01,560 --> 00:21:05,200 add a simple blur or noising. And I mean if you are looking at 381 00:21:05,200 --> 00:21:07,600 the video of this and not the podcast, you can see these 382 00:21:07,600 --> 00:21:10,480 basically look pretty identical to the human eye. 383 00:21:10,760 --> 00:21:14,600 The AI detector says 4% chance this is AI. 384 00:21:14,600 --> 00:21:17,720 So it's just not robust. Wow, Interesting. 385 00:21:17,880 --> 00:21:20,400 OK. And that's like, that's true 386 00:21:20,400 --> 00:21:23,640 across a variety of examples. And then that's even true across 387 00:21:23,640 --> 00:21:26,000 a variety of problems. So we did like, some other 388 00:21:26,000 --> 00:21:28,400 examples, OK, This one's maybe a little more higher stakes. 389 00:21:28,640 --> 00:21:31,080 You take a picture of a car that's not damaged. 390 00:21:31,400 --> 00:21:34,480 You add AI to like, add dents in scratches. 391 00:21:34,480 --> 00:21:36,000 Maybe you're doing insurance fraud. 392 00:21:37,080 --> 00:21:40,600 Again, similar story. The AI says, hey, OK, like the 393 00:21:40,600 --> 00:21:44,120 original version when you just do naive, like, hey, grok, tell 394 00:21:44,120 --> 00:21:47,680 me like add dents, the AI detector will say, hey, it's 395 00:21:47,680 --> 00:21:50,240 like 44% chance or something like that. 396 00:21:50,480 --> 00:21:53,440 But when you add some trivial blurring and noising, the AI 397 00:21:53,440 --> 00:22:00,200 detector goes down to like 2%. So OK, there's, there's other, 398 00:22:00,240 --> 00:22:02,840 you know, then there then we took pictures of like real 399 00:22:02,840 --> 00:22:05,360 editorial photos. So you can imagine like war 400 00:22:05,360 --> 00:22:09,040 zones or like, you know, other journalism or famous political 401 00:22:09,040 --> 00:22:12,280 leaders and like kind of similar story across all these different 402 00:22:12,280 --> 00:22:15,640 categories of images. And so our conclusion from this 403 00:22:15,640 --> 00:22:21,600 study was that AI detection is a dead end, like fundamentally. 404 00:22:21,800 --> 00:22:24,120 And I think if you think about how these models are trained, it 405 00:22:24,120 --> 00:22:26,680 kind of makes sense. Like when you're training these 406 00:22:26,680 --> 00:22:31,080 models, you're optimizing some sort of lost function from like 407 00:22:31,080 --> 00:22:34,520 the generation to like the manifold of real data. 408 00:22:35,200 --> 00:22:37,600 And you're literally optimizing so that the things you spit 409 00:22:37,600 --> 00:22:41,360 outlook statistically very similar to the real data. 410 00:22:41,960 --> 00:22:46,320 And so it's not that difficult to imagine that it's going to be 411 00:22:46,320 --> 00:22:50,080 very hard to detect what's real and what's fake because the 412 00:22:50,080 --> 00:22:52,040 models are being trained to minimize that. 413 00:22:53,160 --> 00:22:56,080 And there's actually like a bunch of work without going into 414 00:22:56,080 --> 00:22:58,200 too much detail. And also, I, I mean, obviously 415 00:22:58,200 --> 00:23:00,960 I'm no longer an AI researcher, so I'm not like super in the 416 00:23:00,960 --> 00:23:04,160 weeds here, but there's a bunch of work done at MIT and by a 417 00:23:04,160 --> 00:23:07,520 bunch of other people on adversarial examples where 418 00:23:07,520 --> 00:23:10,760 basically they had these detectors back then it was these 419 00:23:10,760 --> 00:23:13,680 image net classifiers. And then they added, they did a 420 00:23:13,680 --> 00:23:17,480 similar study, they added like some simple noise and stuff like 421 00:23:17,480 --> 00:23:19,080 that. And then they found that the 422 00:23:19,080 --> 00:23:23,560 image classifiers more robust to these adversarial perturbations. 423 00:23:23,640 --> 00:23:27,560 So you could always kind of find some perturbation of an image. 424 00:23:27,560 --> 00:23:30,080 Like you would take an image of a panda, you would add some 425 00:23:30,080 --> 00:23:31,880 simple blurring. It would look the same to a 426 00:23:31,880 --> 00:23:35,720 human, but then the classifier would flip from panda to like 427 00:23:35,720 --> 00:23:36,880 dog. Right. 428 00:23:37,240 --> 00:23:41,480 And this is this is to do with basically just the fact that you 429 00:23:41,480 --> 00:23:47,720 would never actually see a point of that kind in the like the 430 00:23:47,720 --> 00:23:51,640 manifold of where pandas live. You could perturb it out to the 431 00:23:51,640 --> 00:23:55,760 manifold where dogs live because there was no training data along 432 00:23:55,760 --> 00:23:57,960 that vector. Typically it's like very thin on 433 00:23:57,960 --> 00:24:01,520 that axis. Yeah, again, I, I don't like, I 434 00:24:01,520 --> 00:24:05,280 wasn't in this research line. So my naive like way I think 435 00:24:05,280 --> 00:24:08,600 about it is like just these are such high dimensional decision 436 00:24:08,600 --> 00:24:11,080 boundaries. Like we're going to mess up at 437 00:24:11,080 --> 00:24:13,600 some point and like there's going to be some point in the 438 00:24:13,600 --> 00:24:16,760 decision boundary where you think it's a dog, but like too 439 00:24:16,760 --> 00:24:19,680 human, it looks like a panda and like it's just inevitable 440 00:24:19,680 --> 00:24:23,160 because like the you're just operating over such like a high 441 00:24:23,160 --> 00:24:25,360 dimensional space. That's kind of how I think about 442 00:24:25,360 --> 00:24:27,360 it. There's a. 443 00:24:27,480 --> 00:24:31,400 There's actually a like the pedal width versus length thing 444 00:24:31,640 --> 00:24:36,280 the I like there's this Irish data set in in R I'll bring this 445 00:24:36,280 --> 00:24:36,640 up. Here. 446 00:24:36,640 --> 00:24:38,480 Oh yeah, yeah, it's the famous Yeah, I I've heard. 447 00:24:38,520 --> 00:24:40,000 Of the famous 1, you know what I'm talking about, right? 448 00:24:40,000 --> 00:24:45,760 And so it's like like something like this is probably a 3D 449 00:24:45,760 --> 00:24:47,640 actually, you know what a better one is, like Swiss roll or 450 00:24:47,640 --> 00:24:52,600 something like that, right? In 3D, basically. 451 00:24:52,640 --> 00:24:57,840 Let's see if I can pull this up. So something like this, So Swiss 452 00:24:57,840 --> 00:25:03,200 roll, right, is sort of something where you have like 453 00:25:03,600 --> 00:25:06,720 the yellow category and the green category, aquamarine, 454 00:25:06,960 --> 00:25:09,640 light yellow, blue, right? And in three space, they're 455 00:25:09,640 --> 00:25:14,200 clearly distinct. But if, and let's say this was, 456 00:25:14,480 --> 00:25:19,240 you know, the panda and this is the dog or something like that. 457 00:25:19,640 --> 00:25:23,280 If you put a vector and you perturbed it in such a way that 458 00:25:23,280 --> 00:25:26,520 you had a point that was, I don't know, 60% of the way 459 00:25:26,520 --> 00:25:30,040 towards this blue part and there's no normal points that 460 00:25:30,040 --> 00:25:34,600 existed here in image space. That's my intuition for how the 461 00:25:34,600 --> 00:25:36,080 perturbation works. I should look that up. 462 00:25:36,080 --> 00:25:38,600 But that's that's certainly how it works with low dimensional 463 00:25:38,600 --> 00:25:41,120 things and probably something like that works with higher 464 00:25:41,120 --> 00:25:45,320 dimensional. And you know, similarly to the 465 00:25:46,040 --> 00:25:49,040 this, the pedal went with one over here. 466 00:25:50,200 --> 00:25:52,320 Anyway, I want to I want to get into succinct because this is 467 00:25:52,320 --> 00:25:53,480 the probabilistic. Let's get into your 468 00:25:53,480 --> 00:25:56,520 deterministic go, go, go. So this this also, if you had 469 00:25:56,520 --> 00:25:59,520 something over here that'd be outside of the training set, you 470 00:25:59,520 --> 00:26:04,880 could misclassify it as, you know, as a circle when it was 471 00:26:04,880 --> 00:26:06,120 actually a triangle or vice versa. 472 00:26:06,520 --> 00:26:09,680 OK, go, go, go. Yeah, yeah, we're we. 473 00:26:09,840 --> 00:26:12,920 So we fully established that, yeah, the AI detecting AI. 474 00:26:12,920 --> 00:26:14,320 So it was like not going to work. 475 00:26:14,960 --> 00:26:17,080 That seems bad. So it's the same. 476 00:26:17,240 --> 00:26:19,600 Well, OK, AI makes everything fake. 477 00:26:19,600 --> 00:26:22,280 That's what you said. What's the what's the solution? 478 00:26:22,480 --> 00:26:27,080 Crypto makes it real again. So we're big believers of that. 479 00:26:27,080 --> 00:26:28,520 It's the same. Doesn't apply to cryptography 480 00:26:28,520 --> 00:26:30,560 company. So now like, let's dive into 481 00:26:30,560 --> 00:26:34,840 what that actually means. So today, like, how does content 482 00:26:34,840 --> 00:26:38,200 actually get posted online? I mean, basically first it gets 483 00:26:38,200 --> 00:26:41,640 captured whether it's on like a smartphone or a camera or a 484 00:26:41,640 --> 00:26:43,920 microphone for audio or some other sensor. 485 00:26:44,400 --> 00:26:47,480 Then it goes through some editing, whether it's like 486 00:26:47,480 --> 00:26:50,880 Photoshop or these AI editing tools, and then it gets 487 00:26:50,880 --> 00:26:52,920 published. So it's like across social 488 00:26:52,920 --> 00:26:57,280 media, news services, news wires, traditional media, 489 00:26:57,280 --> 00:26:59,440 YouTube, and then it gets consumed. 490 00:26:59,440 --> 00:27:03,320 So you look at the content and you say like you, you just look 491 00:27:03,320 --> 00:27:04,920 at the content. So that's kind of like the 492 00:27:04,920 --> 00:27:08,120 current life cycle. And yeah, throughout all of 493 00:27:08,120 --> 00:27:09,720 this, there's like no verification. 494 00:27:09,720 --> 00:27:13,320 So it would be impossible for you to tell if something's real 495 00:27:13,320 --> 00:27:18,120 or something's fake. Now, how does crypto help with 496 00:27:18,120 --> 00:27:19,840 this? So this is like what we're 497 00:27:19,840 --> 00:27:23,920 building at Succinct, but we think there's this notion of 498 00:27:23,960 --> 00:27:26,960 basically what we call the provable technology stack. 499 00:27:27,320 --> 00:27:30,960 So at every point in this like capture, edit, publish, consume, 500 00:27:30,960 --> 00:27:34,360 life cycle, you insert in cryptography and provable 501 00:27:34,360 --> 00:27:39,360 technology to prove it's real. So to start when you capture. 502 00:27:39,360 --> 00:27:44,000 Something yeah, this is exact. This is you must have taken some 503 00:27:44,000 --> 00:27:46,920 of my content and and maybe yeah, yeah, OK. 504 00:27:47,280 --> 00:27:50,280 A lot of it is very inspired by, like, a lot of your work, yeah. 505 00:27:50,360 --> 00:27:52,560 OK, well, this is great. So basically there's a crypto 506 00:27:52,560 --> 00:27:55,640 camera and then chain of custody Ledger of record public 507 00:27:55,640 --> 00:27:57,080 verification. Exactly this. 508 00:27:57,080 --> 00:28:00,840 Exactly the stuff that I've wanted out there for whether 509 00:28:00,840 --> 00:28:02,480 it's scientific experiments or something. 510 00:28:02,720 --> 00:28:03,640 OK, keep going. I'm listening. 511 00:28:03,640 --> 00:28:05,840 I, I, I know this, but say, say what you're going to say. 512 00:28:06,040 --> 00:28:08,520 Yeah, yeah. I mean, and yeah, like all 513 00:28:08,520 --> 00:28:11,520 credit words do, I think you identified that this is the 514 00:28:11,520 --> 00:28:14,520 solution maybe like 5 years ahead of its time, five years 515 00:28:14,520 --> 00:28:17,240 ahead of the problem. But and you're, you're always 516 00:28:17,240 --> 00:28:19,480 very ahead of your time. So a lot of this stuff is like 517 00:28:19,480 --> 00:28:22,000 very inspired by your work. And I, I think there's a lot of 518 00:28:22,000 --> 00:28:24,440 other like, I think Mark Andreessen has talked about this 519 00:28:24,440 --> 00:28:25,800 actually. And I'll get to this later. 520 00:28:25,800 --> 00:28:28,680 Like the head of Instagram is now talking about this. 521 00:28:29,480 --> 00:28:32,320 But yeah, OK, just to get into what is approvable tech stock. 522 00:28:32,560 --> 00:28:36,080 So I capture things are captured on hardware devices. 523 00:28:36,760 --> 00:28:40,640 Hardware devices can have private keys that are binded to 524 00:28:40,640 --> 00:28:42,720 the device. So you have a cryptographic chip 525 00:28:42,760 --> 00:28:44,320 with the key. That's kind of how you can think 526 00:28:44,320 --> 00:28:47,080 of it. And basically like as the raw 527 00:28:47,080 --> 00:28:50,480 sensor data is coming into the camera, the cryptographic chip 528 00:28:50,560 --> 00:28:54,520 signs like the content of the raw sensor data and it binds 529 00:28:54,520 --> 00:28:58,520 like the content being captured to the specific device time and 530 00:28:58,520 --> 00:29:01,080 location. So that's cryptographic capture. 531 00:29:02,200 --> 00:29:06,320 Then as the content gets edited, you have this like chain of 532 00:29:06,320 --> 00:29:10,480 custody and chain of edits. So there's a cryptographically 533 00:29:10,480 --> 00:29:13,400 signed manifest for every transformation you do, whether 534 00:29:13,400 --> 00:29:16,480 it's like cropping or color correction or grading or things 535 00:29:16,480 --> 00:29:18,240 like that. And you basically keep this 536 00:29:18,240 --> 00:29:21,280 append only record of what's going on to the image. 537 00:29:22,000 --> 00:29:26,800 And then finally, you publish the piece of content and the 538 00:29:26,800 --> 00:29:31,200 manifest of the original signature when it got captured 539 00:29:31,320 --> 00:29:35,280 to the chain of edits and you publish that to a unbiased 540 00:29:35,280 --> 00:29:37,760 permanent Ledger, which is like this Ledger record. 541 00:29:38,800 --> 00:29:41,640 And then when the content actually gets consumed, so it's 542 00:29:41,640 --> 00:29:45,440 like in some front end, whether it's YouTube or Instagram or X, 543 00:29:46,840 --> 00:29:51,640 the front end integrates with the Ledger and it basically 544 00:29:51,840 --> 00:29:54,280 verifies all the signatures, verifies they're real and 545 00:29:54,280 --> 00:29:56,120 displays that information to the user. 546 00:29:56,320 --> 00:29:58,920 And you know, if the user wants more information, they can just 547 00:29:58,920 --> 00:30:01,360 click and like verify all the signatures for themselves. 548 00:30:01,600 --> 00:30:04,720 So today on most content platforms, we have the blue 549 00:30:04,720 --> 00:30:07,000 check mark for like your verified identity. 550 00:30:07,120 --> 00:30:09,920 You can imagine in the future, maybe all content comes with a 551 00:30:09,920 --> 00:30:13,560 pink check mark that says, hey, this content is like actually 552 00:30:13,560 --> 00:30:15,480 real. And like, here's the device and 553 00:30:15,480 --> 00:30:17,480 here's like the series of transformations that happened to 554 00:30:17,480 --> 00:30:20,080 it. Very cool. 555 00:30:20,440 --> 00:30:23,520 So OK, keep going. So, yeah, this is the provable 556 00:30:23,520 --> 00:30:25,720 tech stack and this is like all the stuff we're building at 557 00:30:25,720 --> 00:30:29,520 succinct. And yeah, I, I think to your 558 00:30:29,520 --> 00:30:32,680 point, you talked about this for a really long time, which is 559 00:30:32,680 --> 00:30:36,520 like very cool. And I think finally, like other 560 00:30:36,520 --> 00:30:39,400 people are starting to catch on 'cause like the problem is 561 00:30:39,400 --> 00:30:43,040 finally very evident. So there's this quote from Adam 562 00:30:43,040 --> 00:30:45,600 Maseri, who runs. Draftily sign a capture? 563 00:30:45,640 --> 00:30:48,080 Yep. Yeah, he posted at the at the 564 00:30:48,080 --> 00:30:50,760 start of 2026. He posted like, hey, here's 565 00:30:50,760 --> 00:30:52,880 Instagrams, like kind of what we're thinking about, what I'm 566 00:30:52,880 --> 00:30:57,840 thinking about right now. And he says that basically we're 567 00:30:57,840 --> 00:31:00,960 going to move from assuming what we see is real by default to 568 00:31:00,960 --> 00:31:04,360 starting with skepticism. So he's kind of identifying this 569 00:31:04,360 --> 00:31:06,760 like AI mix everything fake problem. 570 00:31:07,240 --> 00:31:10,400 And then he said, OK, platforms like Instagram will do good work 571 00:31:10,400 --> 00:31:13,080 identifying AI content, but they'll get worse at it over 572 00:31:13,080 --> 00:31:15,000 time. As AI gets better, it will be 573 00:31:15,000 --> 00:31:18,320 more practical to fingerprint real media than fake media. 574 00:31:18,680 --> 00:31:22,120 And then this is kind of like the thesis of Prove What's real 575 00:31:22,240 --> 00:31:25,920 and all this cryptography stuff. Camera manufacturers will 576 00:31:25,920 --> 00:31:28,640 cryptographically sign images that capture, creating a chain 577 00:31:28,640 --> 00:31:33,320 of custody. So, yeah, I mean, even like 578 00:31:33,320 --> 00:31:37,200 people like Adam who are running Instagramers saying that crypto, 579 00:31:37,240 --> 00:31:40,960 crypto is going, cryptography is going to be the solution to this 580 00:31:40,960 --> 00:31:44,600 like AI, the problems that AI creates for like content 581 00:31:44,600 --> 00:31:45,320 platform. That's right. 582 00:31:45,320 --> 00:31:48,800 Now, I think actually crypto, social and AI are all 583 00:31:48,800 --> 00:31:52,600 interlinked here because another piece of this which is actually 584 00:31:52,600 --> 00:31:54,960 implicit in like the first part of what he's saying, starting 585 00:31:54,960 --> 00:31:57,120 with skepticism, pay attention to who is sharing something and 586 00:31:57,120 --> 00:32:01,040 why. I think actually AI and crypto 587 00:32:01,040 --> 00:32:07,080 together are going to result in, you know, like you know, I think 588 00:32:07,080 --> 00:32:09,240 the future is China versus the Internet. 589 00:32:09,320 --> 00:32:11,720 Did we talk about that? I have you heard me say it. 590 00:32:11,720 --> 00:32:13,560 Talk right about that. I've heard you say a little bit 591 00:32:13,560 --> 00:32:16,800 about it. So I think the future is a 592 00:32:16,800 --> 00:32:20,000 billion person Chinese super state or 1000 million person 593 00:32:20,000 --> 00:32:21,280 network states. Why? 594 00:32:21,640 --> 00:32:25,400 Because everybody thinks about AI improving productivity. 595 00:32:25,440 --> 00:32:29,760 But that was only true within a tribe where you can trust, you 596 00:32:29,760 --> 00:32:34,360 know, you can share information and whether we call it indexing 597 00:32:34,360 --> 00:32:36,960 or surveillance, right? Because 1 is good and one is 598 00:32:36,960 --> 00:32:38,920 consensual and one is bad and one is not, right? 599 00:32:39,480 --> 00:32:43,080 So it is indexing everything and it's learning everything and it 600 00:32:43,080 --> 00:32:47,080 doesn't really miss like a single remark somewhere AI can 601 00:32:47,080 --> 00:32:50,240 pull out a remark from like 3 years ago and surface it and 602 00:32:50,240 --> 00:32:54,040 synthesize in a way that no human you know, or, or you'd 603 00:32:54,040 --> 00:32:56,800 have to have a very attentive smart human human. 604 00:32:56,840 --> 00:32:59,120 It was human limited that level of surveillance from before, 605 00:32:59,120 --> 00:33:02,800 right? So or that level of synthesis, 606 00:33:03,200 --> 00:33:07,160 you know, the to look over every commit and find security holes 607 00:33:07,160 --> 00:33:08,960 from years ago. It's amazing, right? 608 00:33:09,440 --> 00:33:12,720 But that operates within the tribe, outside the tribe. 609 00:33:13,360 --> 00:33:16,560 It's spam, it's scams, it's slop, right? 610 00:33:17,120 --> 00:33:22,080 And so basically the cost of production goes way down, but 611 00:33:22,080 --> 00:33:24,520 the cost of verification goes way up. 612 00:33:25,520 --> 00:33:28,280 And so this part about paying attention to who is sharing 613 00:33:28,280 --> 00:33:31,440 something and why, I think another big piece of this is Web 614 00:33:31,440 --> 00:33:37,200 3 of trust. So you take web of trust like I 615 00:33:37,200 --> 00:33:40,400 trust you because I know you and I've known you in person. 616 00:33:40,720 --> 00:33:44,400 And when you cryptographically sign something on a camera, 617 00:33:44,680 --> 00:33:47,760 there is the human part of that as well as the machine part. 618 00:33:48,120 --> 00:33:52,360 Like ultimately, if I wasn't actually there with you in the 619 00:33:52,360 --> 00:33:57,320 room, I have to trust at some point some human assertion that 620 00:33:57,320 --> 00:34:00,960 this data, because I can see it on chain, that it was stamped at 621 00:34:00,960 --> 00:34:04,280 this time. And there's various proofs that 622 00:34:04,280 --> 00:34:06,440 one can put on there, like proof of location, proof of this, 623 00:34:06,440 --> 00:34:10,320 proof of that. But ultimately at like you as a 624 00:34:10,320 --> 00:34:13,719 human have to tell me that you didn't manipulate it before you 625 00:34:13,719 --> 00:34:16,800 cryptographically signed it. Like you, you, because you could 626 00:34:16,800 --> 00:34:19,760 do something upstream, like the analog hole upstream, you know, 627 00:34:19,760 --> 00:34:21,199 the equivalent of putting something in front of the 628 00:34:21,199 --> 00:34:25,199 camera, right? And we can make it hard to do 629 00:34:25,199 --> 00:34:26,639 that. We can make it impossible to do 630 00:34:26,639 --> 00:34:31,199 that and unless like every single camera has one of these. 631 00:34:31,199 --> 00:34:34,480 And I think maybe it'll get there eventually, but there'll 632 00:34:34,480 --> 00:34:36,920 also be a demand for those things that don't have these 633 00:34:36,960 --> 00:34:38,719 kind of like burner phones, you know what I mean? 634 00:34:38,800 --> 00:34:41,719 Right. And so and, and there's so many 635 00:34:41,719 --> 00:34:43,480 phones out there. There's billions of phones that 636 00:34:43,480 --> 00:34:47,000 do not have crypto chips in them that you, you know, like, just 637 00:34:47,000 --> 00:34:49,880 like you can get an old laptop, you could get a fake able phone, 638 00:34:49,880 --> 00:34:52,040 you know, right. And people will also revolt 639 00:34:52,040 --> 00:34:54,960 against too much tracking or what have you. 640 00:34:54,960 --> 00:34:56,400 You know, they wanted to be free, whatever. 641 00:34:56,840 --> 00:35:00,920 Anyway, I think that's another piece of this is the full supply 642 00:35:00,920 --> 00:35:03,720 chain of custody includes the person who's sending it to you. 643 00:35:04,080 --> 00:35:07,440 And so who is sharing something and why if they're within your 644 00:35:07,440 --> 00:35:12,200 crypto tribe, crypto thinks tribally natively and AI is 645 00:35:12,200 --> 00:35:15,800 going to make people think tribally necessarily. 646 00:35:17,880 --> 00:35:20,920 And so everything reduces to digital tribes where digital 647 00:35:20,920 --> 00:35:22,840 borders and physical borders become the same. 648 00:35:23,120 --> 00:35:26,680 And China's the biggest digital tribe of all because they can 649 00:35:26,680 --> 00:35:30,120 centrally moderate all of their chat apps and so on and so 650 00:35:30,120 --> 00:35:32,440 forth. Like they just whatever AI 651 00:35:32,440 --> 00:35:35,080 detection stuff they roll out and WeChat, they can force human 652 00:35:35,080 --> 00:35:37,680 verification and so on. And say I have just a central 653 00:35:37,680 --> 00:35:39,960 choke point where basically a billion people get on boarded 654 00:35:39,960 --> 00:35:43,880 into whatever AI detection prevention, fake detection 655 00:35:43,880 --> 00:35:46,400 things they want. But the rest of the world 656 00:35:46,400 --> 00:35:49,520 doesn't have the same level of I mean, Google and others can roll 657 00:35:49,520 --> 00:35:52,360 out certain levels of things, but they've almost opted for a 658 00:35:52,360 --> 00:35:55,240 more anarchic standard because of the whole freedom of speech 659 00:35:55,240 --> 00:35:59,080 fight, right, which I get, but there's a there's a under 660 00:35:59,080 --> 00:36:00,960 correction and an over correction on anything. 661 00:36:00,960 --> 00:36:04,320 And what you want is consensual moderation, I think anyway. 662 00:36:04,440 --> 00:36:05,600 So it's a compliment to what you're saying. 663 00:36:05,600 --> 00:36:08,480 Keep going. Yeah, I, I think what you're, I 664 00:36:08,520 --> 00:36:11,640 think in the future, like it's not, I don't imagine a feature 665 00:36:11,640 --> 00:36:14,760 where every photo posted on Instagram is required that it's 666 00:36:14,760 --> 00:36:18,680 real 'cause like, I mean some AI pictures are really cool or like 667 00:36:18,680 --> 00:36:21,440 really interesting. I think it's more like to your 668 00:36:21,440 --> 00:36:25,080 point of consensual moderation, it's like if you want to prove 669 00:36:25,080 --> 00:36:27,600 something's real, and I think a lot of people deeply care about 670 00:36:27,600 --> 00:36:29,960 that. Well, then now you finally using 671 00:36:29,960 --> 00:36:32,080 cryptography actually of the tools to do that. 672 00:36:33,600 --> 00:36:37,560 And then, yeah, if you want to and if you want to follow 673 00:36:37,560 --> 00:36:40,640 content creators that have those capabilities or only post real 674 00:36:40,640 --> 00:36:43,200 stuff, you can do that. And then, you know, social media 675 00:36:43,200 --> 00:36:46,160 is one thing, but obviously for things like journalism, I think 676 00:36:46,200 --> 00:36:49,000 Nikita Beer, who's the head of product at X tweeted about this. 677 00:36:49,160 --> 00:36:51,920 There's these accounts posting totally fake pictures from the 678 00:36:51,920 --> 00:36:54,920 Iranian war and it it's like pretty bad. 679 00:36:54,920 --> 00:36:56,360 And like people are getting misinformed. 680 00:36:56,360 --> 00:36:58,280 And so obviously that's like not OK. 681 00:36:58,960 --> 00:37:01,880 And I think this sort of technology, I, I'm hopeful will 682 00:37:01,880 --> 00:37:05,920 help with much higher stakes situations or, you know, 683 00:37:06,120 --> 00:37:08,560 political ads or like what the president is saying or things 684 00:37:08,560 --> 00:37:10,720 like that, I think will be really important to like prove 685 00:37:10,720 --> 00:37:12,400 what's real there. Great. 686 00:37:12,920 --> 00:37:14,160 OK, cool. All right, keep going. 687 00:37:14,800 --> 00:37:16,440 Cool. I mean, yeah, I think the rest 688 00:37:16,440 --> 00:37:20,120 of the slides are just like a little more detail about what's 689 00:37:20,120 --> 00:37:23,640 going on. So already today you said that, 690 00:37:23,640 --> 00:37:26,440 you know how many cameras actually have this cryptographic 691 00:37:26,440 --> 00:37:28,040 chip. Well, fun fact, every single 692 00:37:28,040 --> 00:37:30,920 iPhone does have a secure enclave that has this 693 00:37:30,920 --> 00:37:34,800 capability. And so, you know, interacting 694 00:37:34,800 --> 00:37:37,720 with these enclaves across all the device types is really hard. 695 00:37:37,720 --> 00:37:40,760 And so we've built this SDK to kind of provide a unified 696 00:37:40,760 --> 00:37:42,680 experience for people. Who is it free? 697 00:37:42,800 --> 00:37:44,080 How far? How's how's it cost? 698 00:37:44,080 --> 00:37:45,760 How much does it cost? Yeah, yeah. 699 00:37:45,920 --> 00:37:48,200 The SDK, well, it's not published yet, but we're going 700 00:37:48,200 --> 00:37:50,880 to publish it. OK, I want to try I I will 701 00:37:51,040 --> 00:37:53,120 Commission some apps on this once you publish this. 702 00:37:53,520 --> 00:37:55,360 Oh, OK. Yeah, yeah, that'll be cool. 703 00:37:56,800 --> 00:37:59,880 That is actually the foundation of a new kind of media. 704 00:38:01,160 --> 00:38:04,520 Yes, Yeah, yeah, I, yes, yeah, I think there's a lot of potential 705 00:38:04,520 --> 00:38:08,320 there. I think incentivizing 706 00:38:08,320 --> 00:38:12,680 decentralized media collection in an AI, first crypto, first 707 00:38:12,680 --> 00:38:16,800 social, first mobile, first Internet, first way, this is 708 00:38:16,800 --> 00:38:21,320 like a missing piece of that where we have all of these quote 709 00:38:21,320 --> 00:38:25,440 reporters from around the world and on any topic that we care 710 00:38:25,440 --> 00:38:28,920 about. We can incentivize first party 711 00:38:28,920 --> 00:38:33,560 reporting where we pay in crypto and we verify in crypto, where 712 00:38:33,560 --> 00:38:36,080 we pay in cryptocurrency and verify with cryptography. 713 00:38:36,840 --> 00:38:38,680 We essentially have like a decentralized news outlet. 714 00:38:39,280 --> 00:38:42,000 So this is something that I want to get going and maybe we can 715 00:38:42,000 --> 00:38:43,960 collaborate on this. We can talk about this right 716 00:38:43,960 --> 00:38:46,000 after this. Yeah, that would be very cool. 717 00:38:48,160 --> 00:38:52,880 And yeah, but citizen journalism like you kind of, well, 718 00:38:52,880 --> 00:38:55,280 especially now with the AI generation stuff, you actually 719 00:38:55,280 --> 00:38:57,960 do need a way to verify that it's actually real. 720 00:38:57,960 --> 00:38:59,800 And so I I totally agree with this. 721 00:39:00,000 --> 00:39:04,560 And I think we would focus it on the news of the network state 722 00:39:04,920 --> 00:39:07,600 and startup societies and cryptocurrency and technology 723 00:39:07,760 --> 00:39:13,000 biotech areas that I think are not well covered, but should be 724 00:39:13,600 --> 00:39:17,080 because they are for tech decision makers. 725 00:39:19,640 --> 00:39:22,920 So because The thing is, news is a huge topic, right? 726 00:39:23,240 --> 00:39:26,040 And rather than the news of the state, we focus on the news of 727 00:39:26,040 --> 00:39:29,840 the network. And those types of things that 728 00:39:30,320 --> 00:39:33,880 are like with a relatively small amount of money, you could get 729 00:39:33,880 --> 00:39:36,040 much more coverage of them because they're more important 730 00:39:36,040 --> 00:39:39,440 for technical decision makers. And that's kind of the niche 731 00:39:39,440 --> 00:39:43,560 that all of these tech outlets basically abdicated. 732 00:39:44,240 --> 00:39:48,760 And actually in part the reason they abdicated is because a full 733 00:39:48,760 --> 00:39:54,480 time journalist is like a professional journalist is often 734 00:39:54,480 --> 00:39:57,800 somebody who doesn't actually know technology because if they 735 00:39:57,800 --> 00:39:59,520 did, they wouldn't be a full time journalist. 736 00:39:59,520 --> 00:40:02,520 They'd be actually like a player on the field, right building. 737 00:40:03,560 --> 00:40:06,440 So moreover, by being a, quote, full time journalist, they're 738 00:40:06,440 --> 00:40:09,040 loyal to the journalist tribe as opposed to technologist tribe. 739 00:40:09,360 --> 00:40:11,560 And technologist tribe is taking away revenue from journalist 740 00:40:11,560 --> 00:40:12,840 tribe. So a lot of their coverage is 741 00:40:12,840 --> 00:40:15,520 very hostile. So the way we solve both of 742 00:40:15,520 --> 00:40:18,440 those problems in my view is rather than one full time 743 00:40:18,440 --> 00:40:23,600 journalist making, I don't know, 50K, whatever it is, we have 50 744 00:40:23,920 --> 00:40:28,160 part time journalists who earn $1000 bounties for writing up 745 00:40:28,160 --> 00:40:30,920 what they know. And because they have domain 746 00:40:30,920 --> 00:40:33,400 knowledge, if they write up one article a year, we're good, 747 00:40:34,560 --> 00:40:36,640 right? So that's like NS News. 748 00:40:37,080 --> 00:40:40,640 And so maybe we can integrate. Yeah, yeah, Yeah, we OK. 749 00:40:40,640 --> 00:40:41,840 Yeah, we should talk about that. Yeah. 750 00:40:41,840 --> 00:40:43,240 You can build that with our stuff now. 751 00:40:43,400 --> 00:40:45,400 It's like pretty. The whole point of that I see is 752 00:40:45,400 --> 00:40:47,000 it makes it easy. OK, great. 753 00:40:47,240 --> 00:40:48,440 Go keep going and let's talk more. 754 00:40:49,120 --> 00:40:50,160 Cool. Yeah. 755 00:40:50,160 --> 00:40:53,600 Then there's the provable edit history part where after you get 756 00:40:53,600 --> 00:40:56,960 the provable capture, you do all the stuff you want to do with 757 00:40:56,960 --> 00:40:58,800 it. And there's actually these 758 00:40:58,800 --> 00:41:02,080 existing standards for it called C2 PA, which kind of tracks, 759 00:41:02,120 --> 00:41:04,400 which is a metadata standard that kind of tracks. 760 00:41:04,640 --> 00:41:07,560 OK, who, what series of edits did you do? 761 00:41:07,560 --> 00:41:10,160 What production did you do? And then it appends it to a 762 00:41:10,160 --> 00:41:13,120 manifest. And then finally, after you've 763 00:41:13,120 --> 00:41:17,640 kind of compiled the proof of capture, the proof of edits, it 764 00:41:17,640 --> 00:41:21,600 gets published to this thing which you came up with this 765 00:41:21,600 --> 00:41:26,200 name, the Ledger of record, which is this like open, 766 00:41:26,200 --> 00:41:31,760 unbiased, you know, place where all this content gets published. 767 00:41:31,760 --> 00:41:35,160 And then that's where all the content that gets displayed in 768 00:41:35,160 --> 00:41:37,520 front ends. So for example, Instagram or X 769 00:41:37,520 --> 00:41:40,760 or whatever, it can read from this Ledger, which is basically 770 00:41:40,760 --> 00:41:44,920 just a database of like what is actually real or not. 771 00:41:46,800 --> 00:41:51,160 So that's kind of our vision for provable technologies and like 772 00:41:51,160 --> 00:41:55,160 the whole stack. And yeah, we kind of imagined 773 00:41:55,160 --> 00:41:59,760 that this stuff will show up one day in every single app, every 774 00:41:59,760 --> 00:42:03,200 single real picture on the Internet will have a pink check 775 00:42:03,200 --> 00:42:07,280 mark that says it's real with all the signatures and all the 776 00:42:07,280 --> 00:42:10,040 cryptographic proof. And you can, anyone viewing a 777 00:42:10,040 --> 00:42:12,600 picture can just look at that and know what's actually real. 778 00:42:12,600 --> 00:42:16,520 So that's our kind of vision for how cryptography is going to 779 00:42:16,760 --> 00:42:18,720 solve a lot of the problems posed by AI. 780 00:42:19,160 --> 00:42:26,320 Amazing. OK, so which people should go to 781 00:42:26,320 --> 00:42:29,080 succinct dot XYZ? Yeah, people can go to Succinct 782 00:42:29,080 --> 00:42:35,240 dot XYZ or follow us on Twitter at Succinct Labs and we will be 783 00:42:35,240 --> 00:42:38,040 posting, we're building this whole stack and we're going to 784 00:42:38,040 --> 00:42:41,600 be releasing like a lot of products and related 785 00:42:41,720 --> 00:42:44,680 technologies, you know, in the in the coming months. 786 00:42:44,880 --> 00:42:48,000 OK, awesome. OK, so. 787 00:42:48,240 --> 00:42:50,840 It'd be interesting to hear kind of your vision for how you think 788 00:42:50,840 --> 00:42:55,600 this is going to, like be put into the world or like, yeah, 789 00:42:56,120 --> 00:42:57,840 you've been talking about these ideas for so long. 790 00:42:57,840 --> 00:43:00,720 I'm just curious to know like more about why I got what got 791 00:43:00,720 --> 00:43:02,880 you excited about it and like how you think this is going to 792 00:43:02,880 --> 00:43:04,720 like proliferate? Sure. 793 00:43:04,800 --> 00:43:08,040 So. Well, what got me excited about 794 00:43:08,040 --> 00:43:14,560 this, you know, in the 90s, like, you know, when I was, I 795 00:43:14,560 --> 00:43:16,920 was a kid then. So I'm about maybe 10-15 years 796 00:43:16,920 --> 00:43:18,000 older than you, something like that. 797 00:43:18,480 --> 00:43:21,760 Or I would never presume to know your age or whatever. 798 00:43:21,760 --> 00:43:23,400 Just saying like probably, probably in that ballpark, 799 00:43:23,400 --> 00:43:27,880 right? In the 90s, nobody cared about 800 00:43:27,880 --> 00:43:31,800 politics. It was something where it was 801 00:43:31,800 --> 00:43:34,240 literally being interested in politics was like being 802 00:43:34,240 --> 00:43:38,920 interested in the train tables or the bus schedule or something 803 00:43:38,920 --> 00:43:42,480 like that, you know, and it was genuinely something, why would 804 00:43:42,480 --> 00:43:44,200 you care about this legislative this and that? 805 00:43:44,200 --> 00:43:47,280 No one cares. And you cared about music, 806 00:43:47,280 --> 00:43:49,800 movies, sports, video games, whatever, right? 807 00:43:49,800 --> 00:43:51,480 It was just a vacation from history. 808 00:43:52,640 --> 00:43:57,960 And so, like, for much of my life, I was essentially just an 809 00:43:57,960 --> 00:44:04,200 apolitical academic and all I care about was math, computer 810 00:44:04,200 --> 00:44:07,200 science, fine formatics, all that kind of stuff. 811 00:44:08,400 --> 00:44:14,320 And then after, you know, essentially the full political 812 00:44:14,320 --> 00:44:17,640 breakdown of arguably you can, you can argue when it started 813 00:44:17,640 --> 00:44:22,040 2000 and one 2008, 2015, 2020. Everyone's got a different 814 00:44:22,040 --> 00:44:27,280 moment, right? The, you know, there's a saying 815 00:44:27,560 --> 00:44:30,800 which is amazing Tweet. If the news is fake, imagine 816 00:44:30,800 --> 00:44:33,600 history. OK. 817 00:44:34,200 --> 00:44:37,160 And you actually start, you know, realizing a lot of the 818 00:44:37,160 --> 00:44:41,960 movies in the 90s, we're almost like the collective unconscious 819 00:44:42,400 --> 00:44:47,640 was putting out movies like The Matrix, Eternal Sunshine of the 820 00:44:47,640 --> 00:44:55,320 Spotless Mind, The Game, Dark City, Fight Club, 12 Monkeys, 821 00:44:55,760 --> 00:44:59,000 all of which were essentially about your, you know, memento, 822 00:44:59,120 --> 00:45:01,080 right? Your memory playing tricks on 823 00:45:01,080 --> 00:45:03,840 you. And in some sense, the world was 824 00:45:03,840 --> 00:45:06,680 not what it seemed, right? The Truman Show, right? 825 00:45:07,000 --> 00:45:09,520 The Truman Show, The Matrix, all of these are like you're living 826 00:45:09,520 --> 00:45:13,120 in a constructed world, right? Memento, your memory of the past 827 00:45:13,120 --> 00:45:18,040 isn't the same, right? And it was as if, like almost 828 00:45:18,040 --> 00:45:21,200 the collective world was waking up to realize that the 829 00:45:21,200 --> 00:45:24,680 centralized century of the 20th century was an illusion in some 830 00:45:24,680 --> 00:45:26,920 ways and that there's more to the past. 831 00:45:26,920 --> 00:45:29,760 And they've, they'd sort of been, you know, hypnotized, 832 00:45:29,760 --> 00:45:34,520 zombified or what have you. And so putting those together, 833 00:45:35,800 --> 00:45:38,520 you know, I started asking questions like, how do we 834 00:45:38,520 --> 00:45:39,840 actually know what's really true? 835 00:45:40,080 --> 00:45:42,960 Like, let me give an example, maybe a seemingly trivial 836 00:45:42,960 --> 00:45:44,720 example, but this is in the network state book. 837 00:45:45,080 --> 00:45:50,280 Let's even take F equals MA. How would how would you actually 838 00:45:50,280 --> 00:45:52,760 know that's true? If you track it all the way 839 00:45:52,760 --> 00:45:56,680 back? Ultimately, there are scatter 840 00:45:56,680 --> 00:46:01,520 plots, you know, when people rolling balls down incline 841 00:46:01,520 --> 00:46:06,760 planes, right, where they are taking X and YS and correlating 842 00:46:06,760 --> 00:46:11,720 them and then effectively doing a line fit that is then 843 00:46:11,720 --> 00:46:14,040 generalized into this deterministic physical law, 844 00:46:14,040 --> 00:46:16,040 right? Underpinning everything that we 845 00:46:16,040 --> 00:46:20,160 think is true is ultimately a set of observations that you 846 00:46:20,160 --> 00:46:22,640 could track all the way back to Newton, like, you know, the 847 00:46:22,640 --> 00:46:25,880 famous, you know, apocryphal apple falling down, right? 848 00:46:26,360 --> 00:46:28,360 Like what you think you know is true. 849 00:46:28,760 --> 00:46:32,440 If you can track back all the citations all the way back to 850 00:46:32,440 --> 00:46:34,280 root. That's the reason that we think 851 00:46:34,280 --> 00:46:36,240 it's true. Why is that actually sometimes 852 00:46:36,240 --> 00:46:40,360 important to do? Well, I'm, I'm forgetting this 853 00:46:40,360 --> 00:46:43,080 is a whole complicated story and I, I think it's something like 854 00:46:43,800 --> 00:46:46,960 there's the story about vitamin CI believe in medicine. 855 00:46:46,960 --> 00:46:49,480 I'm probably getting this wrong and I'll look it up, but it's, 856 00:46:49,560 --> 00:46:53,760 it's like vitamin C supplementation, but it's what 857 00:46:54,320 --> 00:46:56,680 was it spinach? There was like a there's a whole 858 00:46:56,680 --> 00:47:01,520 medical story. Hold on, let me find this. 859 00:47:01,920 --> 00:47:04,520 The iron myth, right? Spinach is a good source of 860 00:47:04,520 --> 00:47:08,640 iron, right? And this is one of those things 861 00:47:08,640 --> 00:47:15,120 where somebody tried to track it back and it was, it was either 862 00:47:15,120 --> 00:47:17,880 this or something else where when you try to track the 863 00:47:17,880 --> 00:47:24,920 citations all the way back, it was a complicated mixture of 864 00:47:25,440 --> 00:47:28,480 multiple mistake and citations on top of each other. 865 00:47:28,880 --> 00:47:45,200 I think this is it here. Let's look at this a sudden 866 00:47:45,440 --> 00:47:46,480 thing. I think this is it. 867 00:47:48,560 --> 00:47:53,600 Basically the complex and convoluted myths is one call for 868 00:47:53,600 --> 00:47:57,280 want of a less complex name, the iron decimal point error myth. 869 00:47:57,800 --> 00:48:10,440 And essentially it is a decimal error knowledge gap. 870 00:48:11,320 --> 00:48:13,200 It's like it's like a myth piled on top of a myth. 871 00:48:13,200 --> 00:48:15,160 It's like something complicated enough that I have to go and 872 00:48:15,160 --> 00:48:17,320 remember it right. But you can look at this, this 873 00:48:17,320 --> 00:48:22,800 document, the point being that that is a concrete example of 874 00:48:22,800 --> 00:48:24,880 something where someone literally dug through every 875 00:48:24,880 --> 00:48:28,480 citation going all the way back and they found that the thing 876 00:48:28,480 --> 00:48:31,920 that people thought was solid was actually based on nothing, 877 00:48:32,080 --> 00:48:34,960 right? All kinds of social sciences 878 00:48:34,960 --> 00:48:37,640 failed the reproducibility crisis in this way, right? 879 00:48:38,160 --> 00:48:43,560 So all kinds of political science, history, social science 880 00:48:44,160 --> 00:48:49,640 is something we're now with LLMS, we can back solve and go 881 00:48:49,640 --> 00:48:52,920 all the way back, right? Because it's, it's much better 882 00:48:52,920 --> 00:48:55,080 search right. So you can track it all the way 883 00:48:55,080 --> 00:48:58,640 back to all the original citations behind a claim, right. 884 00:48:58,640 --> 00:49:00,720 You can push it pretty hard to do that deep research, whatever 885 00:49:00,720 --> 00:49:03,440 you want to call it. Like, you know, the team of 886 00:49:03,440 --> 00:49:06,240 agents saying that Grok has can pull like 1000 sources or 887 00:49:06,240 --> 00:49:08,040 something like that much faster than a human can. 888 00:49:08,040 --> 00:49:10,920 And so now we can really remember that true goal thing 889 00:49:10,920 --> 00:49:14,080 that I was talking about. We can really start 890 00:49:14,080 --> 00:49:17,040 interrogating. It's almost like the, you know, 891 00:49:17,040 --> 00:49:20,520 the Bertrand Russell program in math of really trying to put 892 00:49:20,520 --> 00:49:23,600 math on an axiomatic basis, right? 893 00:49:23,600 --> 00:49:26,200 And really trying to have as few axioms as possible when he 894 00:49:26,200 --> 00:49:27,800 builds the whole thing from set theory. 895 00:49:28,080 --> 00:49:31,680 And and I think it's like on page 347 he says and thus we 896 00:49:31,680 --> 00:49:34,000 proved that 1 + 1 = 2. You. 897 00:49:34,600 --> 00:49:35,560 Know the thing I'm talking about, right? 898 00:49:36,080 --> 00:49:37,640 It's like, it's like a famous thing in math, right? 899 00:49:37,880 --> 00:49:44,200 So you're probably aware of it. So, so like that I, I wanted to, 900 00:49:44,440 --> 00:49:49,040 I realized how ignorant I was about what had actually happened 901 00:49:49,040 --> 00:49:51,720 in the past, about what scientific facts were actually 902 00:49:51,720 --> 00:49:54,440 true, about how scoped my knowledge was. 903 00:49:54,440 --> 00:49:56,960 And I started to ask what I know there's a longer answer than you 904 00:49:56,960 --> 00:49:59,040 wanted, but this is what led me to this, right? 905 00:49:59,520 --> 00:50:03,960 It's like I was like, you know, as a research scientist and 906 00:50:03,960 --> 00:50:05,200 you're a research scientist also. 907 00:50:05,800 --> 00:50:09,120 We're in the unusual position of being pre headline people. 908 00:50:10,240 --> 00:50:13,760 What I mean by that is like this was more true on the Twitter of 909 00:50:13,760 --> 00:50:17,040 like five years ago, but there's a fair number of, let's call 910 00:50:17,040 --> 00:50:22,640 them normie NPC type people who genuinely cannot believe 911 00:50:22,640 --> 00:50:25,400 something is true until it's appeared in the headline. 912 00:50:27,240 --> 00:50:32,440 That is to say, until NYT or the State Department or something 913 00:50:32,440 --> 00:50:36,440 like that, their implicit epistemology was is a reputable 914 00:50:36,440 --> 00:50:40,880 source saying it. If so, then true, If not, then 915 00:50:40,880 --> 00:50:44,200 false. Now this was always bizarre to 916 00:50:44,200 --> 00:50:47,480 me because as a research scientist you're used to figure 917 00:50:47,480 --> 00:50:49,840 out if something is true on your own using logic and reason. 918 00:50:50,200 --> 00:50:51,840 And eventually I was able to figure out the difference 919 00:50:51,840 --> 00:50:54,040 between pre headline people and post headline people. 920 00:50:54,600 --> 00:50:58,560 A pre headline person, you have some scientific finding and you 921 00:50:58,560 --> 00:51:01,440 are going to publish it and you are actually the upstream source 922 00:51:01,440 --> 00:51:03,760 of that finding. Like the press release will be 923 00:51:03,760 --> 00:51:08,360 based on your paper, right? Or conversely, you have some VC 924 00:51:08,360 --> 00:51:12,080 investment round and you know something is true before the 925 00:51:12,080 --> 00:51:16,120 world knows it's true. So you're actually upstream. 926 00:51:16,120 --> 00:51:19,400 It's like a miner, you're mining truth before it's being sold at 927 00:51:19,400 --> 00:51:22,440 the market. However, you realize that 928 00:51:22,440 --> 00:51:25,320 actually the guy who's a post headline person has some wisdom 929 00:51:25,320 --> 00:51:28,840 all his own because he implicitly, I'm not saying 930 00:51:28,840 --> 00:51:31,360 they're doing this explicitly, they kind of know that you can 931 00:51:31,360 --> 00:51:33,440 only be a pre headline person in so many areas. 932 00:51:34,520 --> 00:51:39,520 Like you, you can't be an expert on Turkish and Japanese and I 933 00:51:39,520 --> 00:51:42,520 don't know, Brazilian iron ore and so on and so forth. 934 00:51:42,800 --> 00:51:46,360 Much of what you're sensing is going to be essentially on some 935 00:51:46,360 --> 00:51:49,280 web 3 of trust, which is based on some information supply 936 00:51:49,280 --> 00:51:52,880 chain, right? Anyway, it was through thinking 937 00:51:52,880 --> 00:51:55,480 through things like this and how we build a higher standard of 938 00:51:55,480 --> 00:51:57,800 truth that got me to where we were. 939 00:51:58,000 --> 00:51:59,800 Let me pause here. Interesting. 940 00:52:00,960 --> 00:52:03,760 Yeah, I guess you've been thinking about this for a really 941 00:52:03,760 --> 00:52:06,520 long time. And then I think we've been 942 00:52:06,520 --> 00:52:10,720 thinking about this for, honestly maybe the past three to 943 00:52:10,720 --> 00:52:15,240 six months as we saw the AI problem get worse and worse and 944 00:52:15,240 --> 00:52:16,560 there's this interesting asymmetry. 945 00:52:16,560 --> 00:52:18,400 I mean, our team is based on our software. 946 00:52:18,600 --> 00:52:22,120 There's so many smart people working on accelerating all this 947 00:52:22,160 --> 00:52:24,720 AI stuff, which is really good. There's obviously incredible 948 00:52:24,720 --> 00:52:29,200 positive externalities, but I think if there's a techno, if 949 00:52:29,200 --> 00:52:32,560 there's a technology that's genuinely so powerful, obviously 950 00:52:32,560 --> 00:52:34,320 it's going to have negative externalities. 951 00:52:34,800 --> 00:52:38,760 And I think there's very few people focused on combating 952 00:52:38,760 --> 00:52:43,280 these negative externalities. And I think in this domain with 953 00:52:43,280 --> 00:52:48,440 the pictures and images and fake audio and you know, that poses 954 00:52:48,440 --> 00:52:51,000 real problems. And I do think this is an area 955 00:52:51,000 --> 00:52:53,680 where the combination of cryptographic hardware, 956 00:52:53,680 --> 00:52:58,400 cryptographic software, chain of trust, custody Ledger record, 957 00:52:58,640 --> 00:53:02,880 provable technology broadly as a category can actually like help 958 00:53:03,120 --> 00:53:06,080 some solve those negative externalities. 959 00:53:08,080 --> 00:53:10,280 So amazing. Yeah, that's how I got to it. 960 00:53:10,280 --> 00:53:13,280 But you got to it much earlier than all of us, which is like 961 00:53:13,480 --> 00:53:15,960 kind of your specialty, which is which is very cool. 962 00:53:16,320 --> 00:53:19,600 Well, thank you and but I but I appreciate you also grinding 963 00:53:19,600 --> 00:53:22,280 through all the details to actually build the SDK and so 964 00:53:22,280 --> 00:53:24,160 on, because obviously that's non trivial. 965 00:53:25,320 --> 00:53:30,000 So let's talk more about that. And Uma, thank you for coming 966 00:53:30,000 --> 00:53:31,080 on. Never see a podcast? 967 00:53:31,720 --> 00:53:32,600 Thank you for having me.