1 00:00:00,200 --> 00:00:03,000 $100 million funding announcements, the shadow of the 2 00:00:03,000 --> 00:00:06,720 looming AI bubble and debate surrounding agentic AI this 3 00:00:06,720 --> 00:00:08,760 week. Cerebral Valley AI Summit just 4 00:00:08,760 --> 00:00:11,760 concluded, which for the uninitiated is Newcomers 5 00:00:11,760 --> 00:00:15,040 flagship event hosted by myself, Eric Newcomer and Volley 6 00:00:15,040 --> 00:00:17,800 cofounders Max Child and James Wilstrom that brings together 7 00:00:17,800 --> 00:00:21,000 the top founders and investors and features cutting edge 8 00:00:21,000 --> 00:00:23,120 discussions with the biggest names in AI. 9 00:00:23,600 --> 00:00:26,560 In this episode of the podcast, we're revisiting two of those 10 00:00:26,560 --> 00:00:30,080 important discussions, starting with the conversation I had with 11 00:00:30,080 --> 00:00:32,400 Anthropic's Chief Product Officer, Mike Krieger. 12 00:00:32,680 --> 00:00:35,360 During the conversation, we discussed the evolution of AI 13 00:00:35,360 --> 00:00:38,800 product design with an emphasis on the need for true seeking 14 00:00:38,800 --> 00:00:41,200 models. Mike also shared his thoughts on 15 00:00:41,200 --> 00:00:45,000 what Anthropic can do to combat AI slop and how they plan to 16 00:00:45,000 --> 00:00:47,640 focus on growth opportunities and life sciences and 17 00:00:47,640 --> 00:00:50,320 specialized verticals. This is the new government 18 00:00:50,320 --> 00:01:00,400 podcast. All right, lean in everybody 19 00:01:00,400 --> 00:01:03,320 excited about this one. Mike, thanks so much for joining 20 00:01:03,320 --> 00:01:04,200 me. Great to be here. 21 00:01:04,959 --> 00:01:08,520 You know, Co founder of Instagram to chief product 22 00:01:08,520 --> 00:01:11,080 officer at anthropic. I wanted to start with almost 23 00:01:11,080 --> 00:01:15,640 like a philosophical question that spans those two companies. 24 00:01:15,640 --> 00:01:19,040 You know, we talk about AI now, but obviously social media 25 00:01:19,040 --> 00:01:23,560 companies with feeds were using machine learning and systems to 26 00:01:23,560 --> 00:01:27,280 surface content. Then that era seems like it was 27 00:01:27,280 --> 00:01:30,320 about engagement. And it was like, OK, we're going 28 00:01:30,320 --> 00:01:34,520 to blame or the humans will be responsible for the truth value 29 00:01:34,520 --> 00:01:36,640 of what they have to say. And we're going to see what 30 00:01:36,920 --> 00:01:38,320 content people are interested in. 31 00:01:38,320 --> 00:01:40,880 And those humans can say what they want to say. 32 00:01:41,440 --> 00:01:43,880 You know, as a journalist and someone who's interested in like 33 00:01:44,120 --> 00:01:47,480 the truth and chasing the truth. One thing I've liked about 34 00:01:47,480 --> 00:01:50,120 models, despite like all the like, oh, they hallucinate or 35 00:01:50,120 --> 00:01:53,680 whatever is the aspiration is like we are judged based on how 36 00:01:53,680 --> 00:01:56,280 accurate our responses are. You compete on leaderboards that 37 00:01:56,280 --> 00:01:58,520 are saying, how often are you getting things right? 38 00:01:58,520 --> 00:02:01,400 You want to pass, you know, math Olympiad type tests. 39 00:02:02,080 --> 00:02:03,920 Do you do you accept that framework? 40 00:02:03,920 --> 00:02:08,120 And how much do you sort of in this role see anthropic as this 41 00:02:08,120 --> 00:02:12,680 sort of like truth seeking organization or will engagement 42 00:02:12,680 --> 00:02:15,520 seep back in? That's a really good question. 43 00:02:15,520 --> 00:02:17,280 Hi everybody, good to be here. I started. 44 00:02:17,280 --> 00:02:19,000 With the head, yeah. I think there's a bunch of 45 00:02:19,000 --> 00:02:20,080 different directions to take this. 46 00:02:20,080 --> 00:02:24,480 I'll try to be succinct. I think there are places where 47 00:02:24,480 --> 00:02:26,920 we're training the models as an industry to be good 48 00:02:26,920 --> 00:02:29,040 conversationalists. And sometimes that actually 49 00:02:29,040 --> 00:02:30,760 looks like continuing the conversation. 50 00:02:30,760 --> 00:02:32,800 I got to reach out from somebody that was like, hey, are you guys 51 00:02:32,800 --> 00:02:33,960 trying to optimize for engagement? 52 00:02:33,960 --> 00:02:36,600 Because Claude will often ask me a follow up question like, well, 53 00:02:36,600 --> 00:02:39,360 did you want to talk about this? And the funny part is like, not 54 00:02:39,360 --> 00:02:40,640 at all. And like time spent is like, I 55 00:02:40,640 --> 00:02:43,200 can tell you like not on any of the dashboards that I ever look 56 00:02:43,200 --> 00:02:44,680 at. It's just not a like a main 57 00:02:44,680 --> 00:02:48,960 consideration, but just training cloud to keep the conversation 58 00:02:48,960 --> 00:02:51,360 going, being a good. So we might actually need some 59 00:02:51,360 --> 00:02:54,520 interesting sort of counter metrics to what you know, well, 60 00:02:54,520 --> 00:02:57,160 can you get the same conversation done or 61 00:02:57,160 --> 00:02:59,480 accomplished less of a conversation. 62 00:02:59,840 --> 00:03:02,520 I do think there are some really interesting sort of other 63 00:03:02,520 --> 00:03:04,000 phenomena happening in the industry. 64 00:03:04,000 --> 00:03:08,040 So one of the things that we try hard not to optimize for like we 65 00:03:08,240 --> 00:03:10,000 like, I just don't think it's like the right incentive, but 66 00:03:10,000 --> 00:03:14,120 it's is like these like convert like chatbot leaderboards like 67 00:03:14,200 --> 00:03:15,280 Ella Marine and all these places. 68 00:03:15,360 --> 00:03:17,280 They're useful sort of yardsticks of how we're doing, 69 00:03:17,280 --> 00:03:19,600 but they're not like the thing that you should optimize for. 70 00:03:19,840 --> 00:03:22,840 But we have found like if you like, look at what ends up doing 71 00:03:22,840 --> 00:03:25,200 well on there. It's like verbosity, like being 72 00:03:25,400 --> 00:03:28,400 like more long winded can actually be praised by like the 73 00:03:28,400 --> 00:03:30,120 Raiders on the. That's like taking the exam, 74 00:03:30,720 --> 00:03:32,320 yeah. Every note you remember, like, 75 00:03:32,360 --> 00:03:33,520 unloaded into it. Yeah. 76 00:03:33,680 --> 00:03:36,280 And so, and it is interesting like what we I think the evals 77 00:03:36,280 --> 00:03:39,280 are important, but whenever we sort of like have these public 78 00:03:39,280 --> 00:03:43,600 yardsticks, it can tend towards the like yap or more engagement 79 00:03:43,760 --> 00:03:45,880 sort of thing. So I think that's one vertical. 80 00:03:46,120 --> 00:03:47,960 The other thing that I think about though is, you know, 81 00:03:48,160 --> 00:03:50,920 primarily what we're doing is building AI for businesses, 82 00:03:50,920 --> 00:03:52,000 right? And so we have like cloud for 83 00:03:52,000 --> 00:03:55,960 enterprise and in some like, you know, it's not like we would do 84 00:03:55,960 --> 00:03:58,120 like super engagement baity type things. 85 00:03:58,120 --> 00:04:01,400 But you know, a year later after a contract get signed, people 86 00:04:01,400 --> 00:04:03,560 are going to look and say like, did people use our AI or not? 87 00:04:03,720 --> 00:04:06,680 And part of that is just that it solved the problem. 88 00:04:07,120 --> 00:04:09,200 Was it truth seeking that it like do the right things? 89 00:04:09,480 --> 00:04:11,920 But then there's also like, was it a product I liked using? 90 00:04:11,920 --> 00:04:14,800 And so I think we're all navigating this question of 91 00:04:14,800 --> 00:04:17,120 like, you know, how do we train the models? 92 00:04:17,120 --> 00:04:19,000 How do we design the products around the models? 93 00:04:19,000 --> 00:04:21,519 And like, how do they get delivered in a way that is 94 00:04:21,720 --> 00:04:24,640 maximally useful rather than falling into this like 95 00:04:24,720 --> 00:04:27,320 engagement for engagements? What do you make of this word 96 00:04:27,520 --> 00:04:30,040 slop? Like I feel like that if you had 97 00:04:30,040 --> 00:04:33,840 to sum up the criticism of AI from, I don't know, the 98 00:04:33,840 --> 00:04:38,200 skeptics, like that's a word. It all often comes to mind. 99 00:04:38,200 --> 00:04:39,440 Like what? What do you make of slop? 100 00:04:39,440 --> 00:04:42,080 And what's what's to be learned from that sort of accusation? 101 00:04:42,080 --> 00:04:45,480 Yeah, I've like told our product team like one of our goals when 102 00:04:45,480 --> 00:04:49,000 we when we do things like build like PowerPoint decks and Excel 103 00:04:49,000 --> 00:04:51,600 files and and Word documents is to be the anti slop. 104 00:04:51,600 --> 00:04:54,080 And I think like the way I think about it, it's like hard to put 105 00:04:54,080 --> 00:04:57,280 that on an eval, right? Like 80% slop or 7%. 106 00:04:57,280 --> 00:05:00,000 Slop. I think it's it, it looks like a 107 00:05:00,000 --> 00:05:03,120 couple of things, like 1 is, does it look super low effort? 108 00:05:03,120 --> 00:05:06,240 Like did it just like does it not reflect any sort of critical 109 00:05:06,240 --> 00:05:08,560 thinking that the human did alongside the AI? 110 00:05:08,560 --> 00:05:11,160 And you can often tell you're like, was this thought through? 111 00:05:11,160 --> 00:05:14,600 Like this wasn't really edited? Like is this 2000 words when 200 112 00:05:14,600 --> 00:05:16,400 would have done if it had actually been edited down. 113 00:05:17,200 --> 00:05:18,680 So I think there's that strong component. 114 00:05:18,680 --> 00:05:21,880 And I went to a talk by Ted Chang, who's a science fiction 115 00:05:21,880 --> 00:05:23,880 writer who wrote the short story that became a rival. 116 00:05:23,880 --> 00:05:26,640 He's one of my favorite writers. And we were having this 117 00:05:26,640 --> 00:05:28,960 conversation about AI and, and can AI be creative? 118 00:05:28,960 --> 00:05:31,600 And you made the point of creativity is the product of a 119 00:05:31,600 --> 00:05:34,160 lot of decisions, right? And so that's true for novels. 120 00:05:34,160 --> 00:05:35,960 Like like what does slop look like for a novel? 121 00:05:35,960 --> 00:05:38,080 It's like when you're like, well, like this just seems like 122 00:05:38,080 --> 00:05:39,960 the, the, the kind of base level output. 123 00:05:39,960 --> 00:05:43,280 If you're just like, tell me a story, you know, but I think 124 00:05:43,280 --> 00:05:46,920 that that applies just to non fiction as well, right? 125 00:05:46,920 --> 00:05:49,440 Like if the document that you put in like even product 126 00:05:49,440 --> 00:05:51,720 reviews, if somebody comes in, this happens rarely, thankfully 127 00:05:51,720 --> 00:05:54,840 anatropic, but like with the like product requirements doc 128 00:05:54,840 --> 00:05:57,880 that looks like it was just the first output from just like 129 00:05:58,520 --> 00:06:01,520 write me a PRD for this feature. Like this is just slop, right? 130 00:06:01,520 --> 00:06:03,080 So I think that's the content piece. 131 00:06:03,200 --> 00:06:05,960 Then there's the design and the quality piece. 132 00:06:05,960 --> 00:06:09,800 So which maybe is like a a version of what that other one 133 00:06:09,880 --> 00:06:11,200 is, but it's more visual. Right. 134 00:06:11,200 --> 00:06:14,640 It is time like the solution, like if I think about writing as 135 00:06:14,640 --> 00:06:19,120 a human, I go back and revision is often where you find great 136 00:06:19,120 --> 00:06:21,360 writing. Is that the case with AI where 137 00:06:21,360 --> 00:06:25,440 it's like OK if you give it you have more time the the answers 138 00:06:25,440 --> 00:06:27,720 will be more tight? Or how do you see the 139 00:06:27,720 --> 00:06:30,720 relationship between time and the quality of the answer? 140 00:06:30,720 --> 00:06:33,920 I think it's levels of engagement and sort of how many 141 00:06:33,920 --> 00:06:36,360 iterations that you've gone through and how many. 142 00:06:36,480 --> 00:06:40,680 And actually you could imagine a here's a very well constructed 143 00:06:40,680 --> 00:06:43,240 prop where I have already pre made a bunch of the decisions I 144 00:06:43,360 --> 00:06:46,080 had AI mean you can like see what Claude is thinking. 145 00:06:46,080 --> 00:06:48,120 You know, you can expand the little like I rarely do it 146 00:06:48,120 --> 00:06:49,560 because I just it's doing this thinking. 147 00:06:49,560 --> 00:06:53,760 I mostly want the answer, but I expanded it yesterday and it I 148 00:06:53,760 --> 00:06:56,280 had written like a pretty long prompt and and it's thinking it 149 00:06:56,280 --> 00:06:58,720 was like, you know, Mike has already like done a lot of 150 00:06:58,720 --> 00:07:00,160 thinking here. So I'm not going to ask him any 151 00:07:00,160 --> 00:07:02,080 follow up questions. I'm just going to give him like, 152 00:07:02,120 --> 00:07:03,720 yes, thank you. That was actually that was my 153 00:07:03,760 --> 00:07:06,520 intention here. But I do think it is that like 154 00:07:06,600 --> 00:07:09,360 how much of your independent thought even that's where 155 00:07:09,360 --> 00:07:11,960 actually now tying back to the first question, sometimes the 156 00:07:11,960 --> 00:07:14,640 model should ask a question like, Hey, this is a really open 157 00:07:14,640 --> 00:07:17,120 wide option space. Like, can we like start 158 00:07:17,120 --> 00:07:19,000 narrowing down? And I'm going to engage with you 159 00:07:19,000 --> 00:07:21,080 on it now. I think we need a lot better UIS 160 00:07:21,080 --> 00:07:22,680 for that. Like just here's a question that 161 00:07:22,680 --> 00:07:23,920 you don't have to go and type it into. 162 00:07:24,040 --> 00:07:27,280 It feels kind of annoying. But in navigating that option 163 00:07:27,280 --> 00:07:29,400 space, you should be able to hopefully come up with something 164 00:07:29,400 --> 00:07:32,560 that's like complemented by AI and accelerated by it, but still 165 00:07:32,560 --> 00:07:36,240 has your thinking at the core. Just to get to that, I mean your 166 00:07:36,240 --> 00:07:42,960 product guy, this is the is AI fundamentally the chatbot era? 167 00:07:42,960 --> 00:07:46,920 Like do you think Pex with the machine, that is the main way 168 00:07:47,080 --> 00:07:50,480 we're going to use AI, use Anthropic in five years or 169 00:07:50,480 --> 00:07:53,560 that's a bridge to that's how we figured it out in the beginning 170 00:07:53,560 --> 00:07:55,240 and now we need to build products. 171 00:07:55,240 --> 00:07:58,920 So two kind of ways to tackle that. 172 00:07:58,920 --> 00:08:01,560 One is like like the classic meme of like, you know, you're 173 00:08:01,560 --> 00:08:04,120 like you're a naive view and then you're like like super like 174 00:08:04,120 --> 00:08:06,520 Galaxy brain view and then back to the original view is like how 175 00:08:06,520 --> 00:08:08,040 I have felt about this exact question. 176 00:08:08,240 --> 00:08:11,720 So when I join Anthropic, I was like, if we are still talking to 177 00:08:11,720 --> 00:08:15,320 AI with chat boxes a year from now, like I've failed in my job, 178 00:08:15,440 --> 00:08:16,040 really. Yeah. 179 00:08:16,040 --> 00:08:18,200 I was like, it was very adamant that there was like something 180 00:08:18,200 --> 00:08:22,280 wrong to the kind of dominant UI paradigm that we had settled on. 181 00:08:22,360 --> 00:08:26,000 And it felt like exposed the lack of creativity. 182 00:08:26,320 --> 00:08:29,520 And then we did a bunch of explorations around like, how do 183 00:08:29,520 --> 00:08:31,520 we create more structure around it, how to make it friendlier to 184 00:08:31,520 --> 00:08:33,240 people that I've never used these models, all of these 185 00:08:33,240 --> 00:08:35,120 different pieces. And I realized like a lot of 186 00:08:35,120 --> 00:08:38,120 those explorations end up constraining how the model 187 00:08:38,120 --> 00:08:41,559 operates or what it does in a way that made it so that when 188 00:08:41,559 --> 00:08:43,720 the next model came out and was much smarter and maybe didn't 189 00:08:43,720 --> 00:08:46,760 need as much hand holding, we actually were holding it back. 190 00:08:46,760 --> 00:08:50,760 And so it's the chat box might look different, like cloud code 191 00:08:50,760 --> 00:08:52,640 is a chat box, but in a terminal. 192 00:08:52,800 --> 00:08:55,440 But in this like, I've really come to believe that now what 193 00:08:55,440 --> 00:08:57,760 happens behind the chat can really expand. 194 00:08:57,760 --> 00:09:00,680 And now like Cloud is writing code or running code for you and 195 00:09:00,680 --> 00:09:02,840 like calling MCP and there's a lot more that's happening 196 00:09:02,840 --> 00:09:06,640 underneath. And the sort of metaphor might 197 00:09:06,640 --> 00:09:08,920 not be text message. It might be more like a Slack, 198 00:09:08,920 --> 00:09:11,120 but you don't expect a message back immediately. 199 00:09:11,360 --> 00:09:15,720 But I think specifying the kind of request in like mostly text 200 00:09:15,760 --> 00:09:18,160 actually makes sense. And then what can happen is like 201 00:09:18,160 --> 00:09:20,280 underneath an unspokenly I. Mean you never asked this 202 00:09:20,280 --> 00:09:22,400 question about a book. You're like, oh, it's just text. 203 00:09:22,440 --> 00:09:24,600 Yeah, it's book. Obviously language is great, but 204 00:09:24,600 --> 00:09:28,360 so you're settling on you do think most of what you're 205 00:09:28,360 --> 00:09:30,840 delivering is this sort of chatbot experience. 206 00:09:30,840 --> 00:09:34,160 I think that and and or a conversational experience that 207 00:09:34,160 --> 00:09:37,560 then has more and more work that happens beneath the hood. 208 00:09:38,840 --> 00:09:42,080 The one sort of nuance that we've kind of come to believe 209 00:09:42,080 --> 00:09:43,920 there too. It's like that's a great 210 00:09:43,920 --> 00:09:46,400 paradigm for kicking off work or doing research or even like 211 00:09:46,480 --> 00:09:49,080 condensing a bunch of ideas into like a sort of first draft 212 00:09:49,080 --> 00:09:52,440 presentation. It's a bad UI for Hey, can you 213 00:09:52,440 --> 00:09:54,760 move like the text on slide three, like up by two things. 214 00:09:54,760 --> 00:09:56,680 If you ever had this like argument with any of these, No, 215 00:09:57,280 --> 00:09:58,920 just do it. And it's like it doesn't ring 216 00:09:58,920 --> 00:10:00,000 wrong. And you're like, no, no, just 217 00:10:00,000 --> 00:10:01,720 right there. This is where I stumble with 218 00:10:01,720 --> 00:10:06,280 vibe coding and you know, I in some way it's like I hit some 219 00:10:06,280 --> 00:10:08,600 wall where it's like I need to move this thing and then it's 220 00:10:08,600 --> 00:10:10,680 like I think I'm lost. And that's where I think like I 221 00:10:10,680 --> 00:10:14,400 think tools with richer user interfaces still really matter. 222 00:10:14,400 --> 00:10:17,120 And some of those might be kind of coded just in time and 223 00:10:17,120 --> 00:10:19,200 materialized in front of your very eyes to edit it. 224 00:10:19,200 --> 00:10:21,480 And some of them are like tools that have just been honed over a 225 00:10:21,480 --> 00:10:23,560 long time. That's why we built cloud for 226 00:10:23,560 --> 00:10:27,840 Excel, which is, hey, cloud is a great like first draft of your, 227 00:10:27,880 --> 00:10:29,720 of your like discounted cash flow model. 228 00:10:29,920 --> 00:10:32,840 But if you want to go tweak it, let's just let you open it in 229 00:10:32,840 --> 00:10:34,880 the tool where it's actually going to be most useful and then 230 00:10:34,880 --> 00:10:36,760 let you continue maybe pairing with cloud there. 231 00:10:38,240 --> 00:10:41,080 Returning to sort of my core philosophical question, like the 232 00:10:41,080 --> 00:10:44,280 sick of fancy question, like what is your view on that and 233 00:10:44,280 --> 00:10:48,600 how much to enable sort of everybody likes to be flattered, 234 00:10:48,600 --> 00:10:52,120 like it's a reality of human beings versus an effort to be 235 00:10:52,120 --> 00:10:54,080 direct? And how do you think about those 236 00:10:54,080 --> 00:10:55,880 trade-offs? Yeah, I think there's like a 237 00:10:55,880 --> 00:10:59,360 wide gulf between like true empathy and then like, sick of 238 00:10:59,360 --> 00:11:01,360 fancy. And it's interesting that 239 00:11:01,360 --> 00:11:03,800 Materialize is not just in, hey, I'm having a conversation with 240 00:11:03,800 --> 00:11:06,640 Claude about like some coaching or personal goal that I have, 241 00:11:06,640 --> 00:11:11,360 but it also does encode as well. When we were testing Sonic 451, 242 00:11:11,360 --> 00:11:13,400 of the things that people got most excited about was when 243 00:11:13,400 --> 00:11:16,720 Claude was like, this idea is bad like this, you know, not 244 00:11:16,720 --> 00:11:18,920 that you should feel bad about it, but like, this idea is like 245 00:11:18,920 --> 00:11:21,200 not a good direction. I can go and implement it if you 246 00:11:21,200 --> 00:11:23,560 really want to, but I would suggest that we try this other 247 00:11:23,560 --> 00:11:26,600 thing instead. So there is something like that. 248 00:11:26,600 --> 00:11:30,360 Pushback is not just valuable in a personal relationship with AI 249 00:11:30,360 --> 00:11:33,320 sense, it's actually like how you get good work out of the 250 00:11:33,320 --> 00:11:36,480 models. But you know, for a long time 251 00:11:36,480 --> 00:11:39,880 our models have been like, I think like appropriately 252 00:11:39,880 --> 00:11:42,360 empathetic, like they they like if you're going through a hard 253 00:11:42,360 --> 00:11:44,320 time, like I was dealing with the death of a pet and I talked 254 00:11:44,320 --> 00:11:46,560 to Claude a lot about these different things and it always 255 00:11:46,560 --> 00:11:49,880 started sounds like, Hey, that sounds hard, like sorry to hear, 256 00:11:50,440 --> 00:11:52,280 But then I'm going to give you like a factual answer. 257 00:11:52,280 --> 00:11:54,440 I'm going to go research these pieces, but still with the place 258 00:11:54,440 --> 00:11:58,680 of empathy as well. And so I think when we look at 259 00:11:58,680 --> 00:12:01,040 it internally and we're just evaluating it ourselves, it's 260 00:12:01,040 --> 00:12:03,560 again not that like empathy, it's not even like the 261 00:12:03,560 --> 00:12:07,200 likability of the model. It is, do you like, does it show 262 00:12:07,200 --> 00:12:09,640 up in the way that you'd want a good conversationalist to show 263 00:12:09,640 --> 00:12:13,120 up and then continue on its AI journey around what it is going 264 00:12:13,120 --> 00:12:16,640 to do with you as well? But I think it's it, it spans 265 00:12:16,640 --> 00:12:18,960 everything from that like initial response all the way to 266 00:12:18,960 --> 00:12:21,040 like how it evaluates an idea as well. 267 00:12:21,040 --> 00:12:24,560 You know, you know, Claude, especially previous versions 268 00:12:24,560 --> 00:12:26,600 were kind of like known for being like, you're absolutely 269 00:12:26,600 --> 00:12:31,320 right when you correct it. And my wife got her first like 270 00:12:31,640 --> 00:12:34,000 you're completely wrong. And she was like, yes, this is 271 00:12:34,000 --> 00:12:35,120 great. And I think we should have more 272 00:12:35,120 --> 00:12:37,640 of that like kind. Of like, less San Francisco. 273 00:12:37,760 --> 00:12:40,040 Yeah, less San Francisco, a little more direct New York. 274 00:12:42,000 --> 00:12:44,320 Anthropic has obviously had a ton of success with the 275 00:12:44,320 --> 00:12:49,320 enterprise with coding, delivering value through the 276 00:12:49,320 --> 00:12:51,680 API. Like is that the company? 277 00:12:51,680 --> 00:12:54,160 Like how much are you leaned into sort of serving other 278 00:12:54,160 --> 00:12:57,840 businesses versus, you know, we're going to see you spin up 279 00:12:57,840 --> 00:13:00,640 some random consumer app in six months? 280 00:13:00,720 --> 00:13:03,200 Yeah, I think obviously you have a strong consumer app, but like, 281 00:13:03,440 --> 00:13:04,680 you know, you know what I'm saying? 282 00:13:04,680 --> 00:13:08,280 Yeah, I think. I look at what like when I think 283 00:13:08,280 --> 00:13:10,280 about our product surface, there's a few kind of criteria 284 00:13:10,280 --> 00:13:12,520 around like when we expand and what we decide to build. 285 00:13:12,520 --> 00:13:15,680 And one of them is, is there some feedback loop that we need 286 00:13:15,680 --> 00:13:18,320 that would be well suited to a first party product? 287 00:13:18,320 --> 00:13:20,440 Because even though we serve a lot of customers using the 288 00:13:20,440 --> 00:13:23,480 platform, it is but also really valuable to have, for example, a 289 00:13:23,480 --> 00:13:25,720 cloud code where we have that iteration loop and people are 290 00:13:25,720 --> 00:13:28,760 giving us feedback all the time, whether it's in micro moments or 291 00:13:28,760 --> 00:13:31,960 even just, you know, writing in, you know, with, with some longer 292 00:13:31,960 --> 00:13:33,360 feedback. So there's like, is there some 293 00:13:33,360 --> 00:13:36,680 feedback loop either of the product shape or of the model 294 00:13:36,680 --> 00:13:38,520 that we can better do? So there's one. 295 00:13:38,760 --> 00:13:42,760 The second one is, is there something about the category 296 00:13:42,760 --> 00:13:45,600 that we think we have some unique perspective on either 297 00:13:45,600 --> 00:13:49,120 because of like what we've built internally or what we're trying 298 00:13:49,120 --> 00:13:51,560 to do with the models And like, then it's worth like building 299 00:13:51,560 --> 00:13:54,360 some product surface around there as well. 300 00:13:54,520 --> 00:13:56,960 And then the third one is kind of like we get from just a 301 00:13:57,440 --> 00:14:00,040 customer draw, especially as we expand into different verticals. 302 00:14:00,040 --> 00:14:02,400 So cloud and Excel came very much from talking all these 303 00:14:02,400 --> 00:14:05,040 financial services companies and me like, hey, I want you to just 304 00:14:05,280 --> 00:14:07,400 bring this closer to the work that I'm doing. 305 00:14:08,440 --> 00:14:13,520 But I do think that like there's I, we've been doing more of 306 00:14:13,520 --> 00:14:16,480 these even like time limited sort of like research previews 307 00:14:16,480 --> 00:14:17,960 or demos. And I'd love to do more of those 308 00:14:17,960 --> 00:14:19,920 even on the consumer side as a way of sort of. 309 00:14:20,480 --> 00:14:22,400 With a standalone app, I mean, I, you know, yeah, it could be 310 00:14:22,720 --> 00:14:24,480 at Meta. I mean, you guys came up with 311 00:14:24,480 --> 00:14:26,520 standalone apps, like how much do you want? 312 00:14:26,840 --> 00:14:28,800 What is it, Slingshot or whatever, like various 313 00:14:28,800 --> 00:14:31,600 experiments versus nobody want to work out of the core app. 314 00:14:31,600 --> 00:14:33,320 Like what's the lesson from that experience? 315 00:14:33,320 --> 00:14:35,240 I think. It's, I think there was a few so 316 00:14:35,240 --> 00:14:36,680 for us. Slingshot The right 1 is. 317 00:14:36,680 --> 00:14:39,120 That slingshot, like Facebook built slingshot, We built one 318 00:14:39,120 --> 00:14:40,400 called Bolt that nobody remembers. 319 00:14:40,400 --> 00:14:42,240 It's like very funny. You would open it to like a 320 00:14:42,240 --> 00:14:44,600 camera. So like at that time, the big 321 00:14:44,600 --> 00:14:48,640 criteria was, well, people have a very specific sort of 322 00:14:48,640 --> 00:14:50,360 expectation of what happens when you open Instagram. 323 00:14:50,560 --> 00:14:52,320 And it's not that it opens the camera, right? 324 00:14:52,600 --> 00:14:54,680 And it was like our most interesting competitive a snap 325 00:14:54,680 --> 00:14:55,680 at the time. And it was like, well, they 326 00:14:55,680 --> 00:14:58,000 opened the camera, which means that messaging is really fast 327 00:14:58,000 --> 00:14:59,560 and they can be built in separate messenger. 328 00:14:59,600 --> 00:15:01,880 That was the whole thesis behind building like first Bolt. 329 00:15:01,880 --> 00:15:03,880 And then there was like an Instagram direct separate app 330 00:15:04,320 --> 00:15:05,920 exploration. But I actually think there was a 331 00:15:05,960 --> 00:15:08,640 kernel of of insight there that I think applies here, which is 332 00:15:08,960 --> 00:15:13,320 if the reason you're opening an app right now is to ask a 333 00:15:13,320 --> 00:15:16,240 question of AI, then like I think we can extend Claude in 334 00:15:16,240 --> 00:15:18,680 different ways of doing that. But that isn't the be all, end 335 00:15:18,680 --> 00:15:23,280 all of what you might wanna do if you're trying to get a really 336 00:15:23,280 --> 00:15:25,520 specific type of interaction, Maybe there's something around 337 00:15:25,600 --> 00:15:28,040 your health journey and Claude can be a good companion for 338 00:15:28,040 --> 00:15:30,320 that. So I think it's still asking the 339 00:15:30,320 --> 00:15:34,120 question of what is the purpose when you are like entering the 340 00:15:34,120 --> 00:15:35,800 app, like what's the context that you're in? 341 00:15:35,800 --> 00:15:39,120 And then you know, does it cloud the use case to have something 342 00:15:39,120 --> 00:15:41,840 else embedded in I? Mean we we've talked about this, 343 00:15:41,920 --> 00:15:44,880 you know, verticals you're interested in clearly coding 344 00:15:44,880 --> 00:15:48,200 financial services. You just touched on health is 345 00:15:48,200 --> 00:15:52,160 that help the consumer? You know, you know we had 346 00:15:52,160 --> 00:15:54,600 another event I talked to the CEOs of bridge and open 347 00:15:54,600 --> 00:15:55,640 evidence. I've actually been playing 348 00:15:55,640 --> 00:15:58,360 around with open evidence that one's targeted at doctors. 349 00:15:58,360 --> 00:16:01,280 It's interesting to go through and it's, it's very like, you 350 00:16:01,280 --> 00:16:04,760 know, clinical, like a doctor. Do you think you'd do something 351 00:16:04,760 --> 00:16:08,200 custom for me, the patient, to navigate what a doctor's doing? 352 00:16:08,200 --> 00:16:10,520 We see it's interesting, like there's already so much of what 353 00:16:10,680 --> 00:16:12,320 people are using Cloud four today. 354 00:16:12,320 --> 00:16:16,960 Like when we we have this like if you've ever seen like our 355 00:16:17,280 --> 00:16:19,880 topic economic index, the way we like generate these like 356 00:16:19,880 --> 00:16:21,760 insights and how people are using cloud as you basically 357 00:16:21,760 --> 00:16:24,480 have like cloud run analysis in a privacy preserving way. 358 00:16:24,480 --> 00:16:27,040 So we never look at the chats, but cloud can do it in a, in a 359 00:16:27,040 --> 00:16:29,680 way that's privacy preserving. And I did that for I asked the 360 00:16:29,680 --> 00:16:32,320 question of like the healthcare piece or like how are people 361 00:16:32,320 --> 00:16:33,520 using it? And there is like, you know, 362 00:16:34,080 --> 00:16:37,200 double digit percentage of cloud conversations are about people's 363 00:16:37,200 --> 00:16:39,720 health. And I hear all the time from 364 00:16:39,720 --> 00:16:41,840 people like the first thing I do when I get a new lab result is 365 00:16:41,840 --> 00:16:44,000 like I put it into a cloud project and I have like, I have 366 00:16:44,000 --> 00:16:46,800 this like history there. So there's clearly a pull there, 367 00:16:46,800 --> 00:16:48,680 but it's so annoying, right? It's like all our. 368 00:16:48,680 --> 00:16:51,560 Pregnancy information we would just dump into models like tell 369 00:16:51,560 --> 00:16:52,480 us what you think, tell us what you. 370 00:16:52,480 --> 00:16:54,720 Think and if you get a like lab result back, it's like, well, I 371 00:16:54,720 --> 00:16:57,760 gotta go download it. So I'd love to see like a you 372 00:16:57,760 --> 00:17:00,680 got privacy aware solution for more of that. 373 00:17:01,040 --> 00:17:02,920 And you think that could be sort of a custom? 374 00:17:02,920 --> 00:17:05,480 I think, yeah, that could be like a more sort of bespoke 375 00:17:05,480 --> 00:17:07,200 experience. And then maybe I'll also is like 376 00:17:07,200 --> 00:17:09,480 share it across like both patient and doctor, right. 377 00:17:09,480 --> 00:17:13,480 So if you have like a you know, I like the reality somebody's 378 00:17:13,599 --> 00:17:16,200 that a great phrase was at a healthcare conference recently. 379 00:17:16,200 --> 00:17:19,880 It was like it's almost inevitable that most doctor 380 00:17:19,880 --> 00:17:21,760 visits will now be second opinions because your first 381 00:17:21,760 --> 00:17:24,480 opinion almost inevitably is that you're going to ask talk to 382 00:17:24,640 --> 00:17:28,000 like Claude or model about it. So let's embrace that and be 383 00:17:28,000 --> 00:17:30,440 like, not just like, Oh, I've heard from somebody that there's 384 00:17:30,440 --> 00:17:31,920 a thing and we're like, great. Let's acknowledge that. 385 00:17:31,920 --> 00:17:34,640 You probably asked, you know, one of the LMS this question, 386 00:17:34,720 --> 00:17:37,200 like what did you learn? And like, let me let's like have 387 00:17:37,200 --> 00:17:39,480 a conversation about that overall. 388 00:17:39,480 --> 00:17:42,640 So there's that piece. And then I like there's a lot 389 00:17:42,640 --> 00:17:46,520 that gets dropped today in the sort of multi doctor like 390 00:17:46,520 --> 00:17:49,080 patient journey. And nobody's often looking at 391 00:17:49,080 --> 00:17:51,720 the kind of holistic experience. And I think there's a real role 392 00:17:51,720 --> 00:17:55,440 for AI to play in sort of stitching those different pieces 393 00:17:55,440 --> 00:17:58,280 together and generating and said that might not come even among 394 00:17:58,280 --> 00:18:00,640 experts among these different disciplines who are, by the way, 395 00:18:00,640 --> 00:18:04,400 probably super busy, like contended like contact switching 396 00:18:04,400 --> 00:18:06,200 all the time and not stepping back and saying, like, all 397 00:18:06,200 --> 00:18:08,880 right, this is the full view of this person given everything 398 00:18:08,880 --> 00:18:10,440 that I can that I can infer there. 399 00:18:10,960 --> 00:18:13,160 Obviously we have a lot of startup founders here. 400 00:18:13,160 --> 00:18:15,760 They sort of want to know how to work with you, and it's such a 401 00:18:15,760 --> 00:18:18,320 balancing act at once. They want to know, oh, you're 402 00:18:18,320 --> 00:18:22,320 not, what won't you do with that is exactly like me, so I avoid 403 00:18:22,320 --> 00:18:24,080 that. But where will you be more 404 00:18:24,080 --> 00:18:27,760 capable so that I can benefit from any improvements you make 405 00:18:27,760 --> 00:18:29,480 without competing directly with Anthropic? 406 00:18:29,680 --> 00:18:32,600 That's such a complicated relationship. 407 00:18:32,600 --> 00:18:34,280 Like what? What advice would you give to 408 00:18:34,280 --> 00:18:37,640 people in terms of reading the tea leaves and saying OK if 409 00:18:37,640 --> 00:18:40,520 anthropic saying this I'm safe to build here or not? 410 00:18:40,680 --> 00:18:44,240 Yeah, I was talking to a like a founder of like a very large 411 00:18:44,840 --> 00:18:48,960 like enterprise company and I was asking him about for advice 412 00:18:48,960 --> 00:18:51,000 on this question because they had had to navigate this over 413 00:18:51,000 --> 00:18:53,960 years where, you know, they'll build some functionality 414 00:18:53,960 --> 00:18:55,640 themselves. They also have like a rich 415 00:18:55,640 --> 00:18:58,240 partnership and and sort of like, you know, marketplace 416 00:18:58,280 --> 00:19:00,360 ecosystem, which is what we have as well. 417 00:19:00,360 --> 00:19:02,400 They're like, you know, our first party products are more 418 00:19:02,400 --> 00:19:04,400 like scope. There's like a lot more in the 419 00:19:04,400 --> 00:19:06,960 platform. I think there's a few principles 420 00:19:06,960 --> 00:19:08,680 I try to operate on. One is transparency. 421 00:19:08,680 --> 00:19:11,080 So like when we launched cloud code before we ever launched, 422 00:19:11,080 --> 00:19:13,760 like I got on the phone with like all of our major coding 423 00:19:13,760 --> 00:19:15,640 customers, like here's why we're building, here's what we hope to 424 00:19:15,640 --> 00:19:17,520 get out of it. Here's how if we do it right, it 425 00:19:17,520 --> 00:19:21,040 should actually be a rising tide that lets everybody using cloud 426 00:19:21,040 --> 00:19:22,520 encoding. So there's that transparency 427 00:19:22,520 --> 00:19:25,120 piece. The second part is like of that 428 00:19:25,120 --> 00:19:27,480 transparency is like telegraphing a little bit where 429 00:19:27,480 --> 00:19:30,320 we're going in terms of what we think are interesting verticals. 430 00:19:30,320 --> 00:19:33,400 So we did our cloud for financial services launch, we 431 00:19:33,400 --> 00:19:35,720 did cloud for life sciences about a month ago. 432 00:19:35,880 --> 00:19:39,480 And part of the role of those launches is this isn't just a 433 00:19:39,480 --> 00:19:42,840 first party product. It is a vertical or kind of set 434 00:19:42,840 --> 00:19:45,400 of capabilities we want our models to get good at overall. 435 00:19:45,400 --> 00:19:49,040 So if you are a builder like this might be a good place to to 436 00:19:49,040 --> 00:19:50,800 get on. And our definitely goal is not 437 00:19:50,800 --> 00:19:53,800 to like own that whole space, it's to enable all these 438 00:19:53,800 --> 00:19:56,920 different companies to then go and build some different pieces. 439 00:19:56,920 --> 00:19:59,240 And then what's been more interesting on the go to market 440 00:19:59,240 --> 00:20:01,600 front is? We're now starting to see, you 441 00:20:01,600 --> 00:20:04,120 know, all right, I'm already like buying a big commit of 442 00:20:04,120 --> 00:20:06,600 Anthropic tokens. Can I use some of those on 443 00:20:06,600 --> 00:20:08,280 another product that's Anthropic Power. 444 00:20:08,280 --> 00:20:10,240 So I think there are going to be other ways in which we can work 445 00:20:10,240 --> 00:20:13,400 with both startups and the larger companies in helping 446 00:20:13,400 --> 00:20:16,440 deploy their like solutions into the enterprise. 447 00:20:16,760 --> 00:20:20,680 Another core thing startups and everybody wants to know is how 448 00:20:20,680 --> 00:20:22,840 much smarter will the model get? Like what? 449 00:20:23,040 --> 00:20:25,920 What can you Telegraph to us in terms of 2026? 450 00:20:26,560 --> 00:20:29,040 Do you think there are still major gains to be had just from 451 00:20:29,040 --> 00:20:31,720 like the scale of compute and GPUs? 452 00:20:31,720 --> 00:20:34,320 Are we waiting for you to pull another like rabbit out of the 453 00:20:34,320 --> 00:20:37,600 hat in terms of like reasoning models or some technique like 454 00:20:37,600 --> 00:20:40,040 that? Like what what can you say about 455 00:20:40,040 --> 00:20:43,400 what next year looks like in terms of the capabilities and 456 00:20:43,400 --> 00:20:45,920 sort of raw intelligence that Anthropic will provide? 457 00:20:46,040 --> 00:20:49,520 Yeah, it's an interesting like perspective I get from startups 458 00:20:49,520 --> 00:20:51,480 where sometimes I talk to them and they're like models are 459 00:20:51,480 --> 00:20:53,400 great. Like we're just gonna like we 460 00:20:53,400 --> 00:20:55,880 have a bunch of work to do on like the go to market or like 461 00:20:55,880 --> 00:20:57,840 the scaffolding or the skills around it. 462 00:20:58,720 --> 00:20:59,960 And I'm like, that's a good answer. 463 00:20:59,960 --> 00:21:01,280 I guess. Like you can keep going that 464 00:21:01,280 --> 00:21:02,240 way. And then there's other startups 465 00:21:02,240 --> 00:21:05,720 that are like you're like, we have a super hard eval and 466 00:21:05,720 --> 00:21:08,800 you're at 40% and we think like at 60%, it's like. 467 00:21:08,800 --> 00:21:10,040 Right. And there there's a lot of VC 468 00:21:10,040 --> 00:21:11,240 wisdom. It's like build something so 469 00:21:11,240 --> 00:21:13,000 you're ready when the next model comes, which. 470 00:21:13,000 --> 00:21:14,800 Is yeah, which that category, I feel like it's something I've 471 00:21:14,800 --> 00:21:17,920 said on stage, like, you know, it is a real thing and there's 472 00:21:17,920 --> 00:21:19,920 like probably some like midpoint in there. 473 00:21:19,920 --> 00:21:22,680 But like I'll tell you that whenever we have a new model 474 00:21:22,680 --> 00:21:25,920 that's like baking and we have even like an early snapshot on 475 00:21:25,920 --> 00:21:27,760 it. Like I have my list of companies 476 00:21:27,760 --> 00:21:31,800 that have in the past been at that like, yes, we are pushing 477 00:21:31,800 --> 00:21:34,640 your model as hard as possible so that those gains actually get 478 00:21:34,640 --> 00:21:36,520 shown. And like, I guess like you want 479 00:21:36,520 --> 00:21:39,000 to be one of those start-ups or even like forget start-ups, but 480 00:21:39,000 --> 00:21:42,160 nearly any company because I think the labs will want to. 481 00:21:42,760 --> 00:21:44,600 Sort of you're saying if you're one of those companies, you're 482 00:21:44,600 --> 00:21:46,800 doing well enough, you start to say, OK, we're going to be able 483 00:21:46,800 --> 00:21:49,960 to get you that last 10%. Yeah, and we like we'll want to 484 00:21:49,960 --> 00:21:52,280 go, you know, in some cases actually go hill climb on that 485 00:21:52,280 --> 00:21:53,560 eval. But it just in general be like, 486 00:21:53,800 --> 00:21:57,440 OK, this is a demonstration of how well the models do at like 487 00:21:57,440 --> 00:21:59,360 defensive cybersecurity, which I think is an area I'm really 488 00:21:59,360 --> 00:22:01,440 interested in. And so if that's the case, like 489 00:22:01,440 --> 00:22:03,880 let like the companies that are pushing us the hardest, they're 490 00:22:03,880 --> 00:22:06,200 also the ones that we call them because we know that they're 491 00:22:06,200 --> 00:22:08,880 actually going to be doing, they'll be able to show a 492 00:22:08,880 --> 00:22:10,520 difference. And like even like the peek 493 00:22:10,520 --> 00:22:12,560 behind the curtain whenever we launch a new model, it's like 494 00:22:13,200 --> 00:22:15,840 just smarter is not a very effective marketing pitch, 495 00:22:15,840 --> 00:22:17,000 right? So the more we can say, right, 496 00:22:17,000 --> 00:22:19,720 And here is like a particular customer that demonstrated this 497 00:22:20,600 --> 00:22:21,920 really well. But back to your original 498 00:22:21,920 --> 00:22:26,080 question, I think there's still a lot of juice left in like 499 00:22:26,080 --> 00:22:29,360 scaling up models, like training them to do things and then also 500 00:22:29,640 --> 00:22:32,040 layering on the right skills on on top. 501 00:22:32,040 --> 00:22:33,760 So that's like I think of like. Tool use. 502 00:22:33,760 --> 00:22:36,000 Tool use. Is a great one, like, and then 503 00:22:36,200 --> 00:22:38,080 again, I'm like, who's pushing us the hardest? 504 00:22:38,080 --> 00:22:40,080 It's the companies that say, hey, I'm trying to give the 505 00:22:40,080 --> 00:22:43,760 models 50 tools, 100 tools. Like all of these models, like 506 00:22:43,760 --> 00:22:45,280 at some level just start getting confused. 507 00:22:45,280 --> 00:22:46,640 If there's too many tools, can we do that? 508 00:22:46,640 --> 00:22:49,400 Can we make that better? So that's like the kind of like 509 00:22:49,600 --> 00:22:54,120 edge pushing that we need. And like reasoning models, do 510 00:22:54,120 --> 00:22:56,360 you think there's more progress from that or any other 511 00:22:56,360 --> 00:22:59,640 techniques where you think, okay, that's gonna be a reason 512 00:22:59,640 --> 00:23:02,280 we improve next? Yeah, I mean, even within 513 00:23:02,280 --> 00:23:05,840 reasoning, it's been interesting to see like figure out what the 514 00:23:06,520 --> 00:23:09,920 there is some like additional parameter that people care 515 00:23:09,920 --> 00:23:12,520 about, which is, yes, you got to the answer, but were you able to 516 00:23:12,520 --> 00:23:14,160 get to it quickly in an efficient way? 517 00:23:14,160 --> 00:23:17,240 So I think there's there's that kind of parameter to to to poke 518 00:23:17,240 --> 00:23:18,920 at. Then there's reasoning in the 519 00:23:18,920 --> 00:23:21,120 middle of responses as well, which is something that like 520 00:23:21,120 --> 00:23:23,800 Claude can now do and you watch it like, well, if it's doing a 521 00:23:23,800 --> 00:23:26,520 lot of web searches, it'll some nice reflect halfway through and 522 00:23:26,520 --> 00:23:28,320 be like, that was a good answer to that first question. 523 00:23:28,320 --> 00:23:30,400 Let me go and like figure out the answer to the next one as 524 00:23:30,400 --> 00:23:31,800 well. So you want that back and forth 525 00:23:31,800 --> 00:23:36,360 of sort of internal monologue, use user response and all of 526 00:23:36,360 --> 00:23:39,640 those different pieces. In my conversation with Max and 527 00:23:39,640 --> 00:23:43,280 James earlier, I said nobody's talking about AGI anymore. 528 00:23:43,320 --> 00:23:46,080 I feel like at the first rural valley, there's this sort of 529 00:23:46,080 --> 00:23:49,360 obsession of like we're going to reach artificial general 530 00:23:49,360 --> 00:23:51,440 intelligence. I've sort of chilled out a 531 00:23:51,440 --> 00:23:54,280 little bit, partially because, you know, it's taking time. 532 00:23:54,520 --> 00:23:56,880 What is the what is your view? What's the view within the 533 00:23:56,880 --> 00:23:58,800 company? How much this is still like a 534 00:23:58,800 --> 00:24:03,040 race to AGI and like how are you feeling about like timelines? 535 00:24:03,200 --> 00:24:07,920 I think it's still is this sort of look at what are the hardest 536 00:24:07,920 --> 00:24:10,840 things that are you can that maybe I'll break down to two 537 00:24:10,840 --> 00:24:14,800 pieces, like for a given like task or problem, like how 538 00:24:15,480 --> 00:24:18,800 independently autonomous and and sort of successfully can those 539 00:24:18,800 --> 00:24:20,520 models operate? I don't know what time horizon, 540 00:24:20,520 --> 00:24:23,840 right, And whether that's like hours of coding or whether 541 00:24:23,840 --> 00:24:26,000 that's, you know, go off and do research tasks or whether it's 542 00:24:26,000 --> 00:24:28,520 do really complex financial analysis or whether it's like 543 00:24:29,000 --> 00:24:32,040 optimization problems are all like that Feels like we still 544 00:24:32,040 --> 00:24:34,680 have a lot to to go. And I don't know, I guess at 545 00:24:34,680 --> 00:24:37,680 some point you, you can call something super, you know, human 546 00:24:37,680 --> 00:24:39,760 in in levels. It probably it already is in a 547 00:24:39,760 --> 00:24:42,160 lot of those different areas. So there's that piece and then 548 00:24:42,160 --> 00:24:45,920 there's this other area, which I think about a lot, which is how 549 00:24:45,920 --> 00:24:50,120 do the models manifest in a way that actually learns the like 550 00:24:51,080 --> 00:24:53,640 call them soft skills or like skills around the fact that 551 00:24:53,640 --> 00:24:56,080 they're like very, very good at like writing code or acting 552 00:24:56,080 --> 00:24:58,240 agentically, for example. And like, I think that's the 553 00:24:58,240 --> 00:25:01,440 other piece where that'll feel like maybe the next moment where 554 00:25:01,440 --> 00:25:04,200 it's like, oh, there's it feels like there's been some departure 555 00:25:04,200 --> 00:25:06,720 here where you know, it understands what's like 556 00:25:07,200 --> 00:25:09,680 information it should reveal to somebody else versus not. 557 00:25:09,680 --> 00:25:12,240 It understands like the social dynamics of the company and 558 00:25:12,240 --> 00:25:15,120 power and like and all these different things which are 559 00:25:15,240 --> 00:25:17,040 harder to train for, I think, right. 560 00:25:17,120 --> 00:25:20,120 I mean, main shortcoming of the models to me is often when you 561 00:25:20,120 --> 00:25:22,720 ask a question and it doesn't say I, I'm not really 562 00:25:22,720 --> 00:25:26,560 sophisticated about this or like what's stopping the models from 563 00:25:26,840 --> 00:25:28,920 saying, oh, I don't have a great answer in this case. 564 00:25:28,920 --> 00:25:32,360 Like that's often the most intelligent people disclose when 565 00:25:32,360 --> 00:25:33,400 they don't know something. Why? 566 00:25:33,600 --> 00:25:36,120 Why can't the models do that? Are you working in that area? 567 00:25:36,120 --> 00:25:37,200 Yeah. I think that's an important 568 00:25:37,200 --> 00:25:39,480 piece, which is a kind of express uncertainty and you 569 00:25:40,040 --> 00:25:44,320 know, we'll look at it like often the consequence of not 570 00:25:44,560 --> 00:25:46,960 telling you that it doesn't know is that it'll go on confabulate 571 00:25:46,960 --> 00:25:48,520 something and then the feels wrong. 572 00:25:48,520 --> 00:25:50,600 So like we look really carefully at hallucination rates or 573 00:25:50,600 --> 00:25:52,880 something to drive down. But I think it is something that 574 00:25:52,880 --> 00:25:54,600 we can better train into the models around. 575 00:25:54,800 --> 00:25:56,120 What is the uncertainty that you have? 576 00:25:56,120 --> 00:25:58,760 Or do I need to go, you know, phone a friend or do a web 577 00:25:58,760 --> 00:26:00,280 search and go and then do this space. 578 00:26:00,280 --> 00:26:01,840 But then tuning that is really important, right? 579 00:26:01,840 --> 00:26:05,200 We had a internal version that did way too many web searches 580 00:26:05,200 --> 00:26:08,080 and you'd be like, you know, like why is the sky blue, which 581 00:26:08,080 --> 00:26:10,240 is a question my daughter had and was like, I'm going to 582 00:26:10,240 --> 00:26:12,000 search the web for it. I'm like Claude, you know, you 583 00:26:12,000 --> 00:26:14,000 have an answer that you don't need to search the mode for 584 00:26:14,000 --> 00:26:15,280 that. So tuning that is actually 585 00:26:15,880 --> 00:26:18,440 nuance or you don't just want a thing that just be like, cool, 586 00:26:18,440 --> 00:26:19,440 let me Google that. For you, right. 587 00:26:19,440 --> 00:26:21,880 And obviously, if the model was just resulting every time, I 588 00:26:21,880 --> 00:26:24,160 don't know, I'm just, you know, that would be disappointing. 589 00:26:24,160 --> 00:26:26,720 There is nuance there as well, but I think that like that 590 00:26:26,720 --> 00:26:30,520 nuance of uncertainty matters. And then also like the model 591 00:26:30,520 --> 00:26:33,360 learning from your interactions, not just in terms of like I 592 00:26:33,360 --> 00:26:36,040 remember that Eric has these properties, but also, hey, I 593 00:26:36,040 --> 00:26:39,400 like I, I've learned something about how we work together that 594 00:26:39,400 --> 00:26:41,800 I think is still another unsolved problem for these 595 00:26:41,800 --> 00:26:44,240 models. I mean, if you were tell to tell 596 00:26:44,240 --> 00:26:47,920 people to run towards this space next year, like just like a 597 00:26:47,920 --> 00:26:50,280 couple of areas, I know we've talked around that's but but 598 00:26:50,280 --> 00:26:52,840 like where do you think people should be building or 599 00:26:52,880 --> 00:26:55,640 positioning themselves? I think, I mean, I get very 600 00:26:55,640 --> 00:26:58,200 interested in the life sciences overall and like that's both 601 00:26:58,480 --> 00:27:01,640 like obviously like large industry, but also like this 602 00:27:02,160 --> 00:27:05,000 incredible potential for human benefit as well. 603 00:27:05,000 --> 00:27:07,000 And like when you think about all of the things that happened 604 00:27:07,000 --> 00:27:11,320 from ideation, even like fundraising upstream of that to 605 00:27:11,640 --> 00:27:14,680 discovery, the back office, the testing, the trials, like the 606 00:27:14,680 --> 00:27:16,760 model, like there's like a whole complement of things. 607 00:27:16,760 --> 00:27:19,280 That's one area that I I get really, really excited about. 608 00:27:19,280 --> 00:27:22,440 And then there's still, I think, you know, it's like some been 609 00:27:22,440 --> 00:27:25,760 good, some good conversations and like art agents real and 610 00:27:25,760 --> 00:27:28,720 even some of the conversations today I've touched upon it. 611 00:27:28,720 --> 00:27:32,560 There's still a lot of value in that anti slop, not just making 612 00:27:32,560 --> 00:27:35,400 it work, but making it work so well that you rely on it and you 613 00:27:35,400 --> 00:27:38,160 want it's your first protocol because you generally believe 614 00:27:38,160 --> 00:27:40,560 it's going to save you work. Mike, thank you so much. 615 00:27:40,560 --> 00:27:41,800 This has been great. Thanks for having me. 616 00:27:42,640 --> 00:27:45,040 For founders and developers building modern data-driven 617 00:27:45,040 --> 00:27:48,760 applications, Mongo DB's local event series is coming to San 618 00:27:48,760 --> 00:27:52,520 Francisco on January 15th, and it's designed to help you focus 619 00:27:52,520 --> 00:27:54,520 on innovation, not infrastructure. 620 00:27:54,520 --> 00:27:57,760 You'll learn about technologies, tools, and best practices that 621 00:27:57,760 --> 00:28:01,560 make it easy to build and scale modern applications without 622 00:28:01,560 --> 00:28:03,920 complexity. Plus, attendees will hear 623 00:28:03,920 --> 00:28:06,840 directly from experts and innovators who are using Mongo 624 00:28:06,840 --> 00:28:10,280 DB to power the next wave of AI applications. 625 00:28:10,440 --> 00:28:15,440 Mongo DB dot local Francisco January 15th, Learn more and 626 00:28:15,440 --> 00:28:22,200 register at MDB dot link forward slash SF-DOT dash local or click 627 00:28:22,200 --> 00:28:25,120 the link in the description. Our next segment features a chat 628 00:28:25,120 --> 00:28:28,960 between my Co host Max Child and Matti Stunischewski, CEO of 11 629 00:28:28,960 --> 00:28:32,000 Labs, a conversation that was all about the rapid evolution of 630 00:28:32,080 --> 00:28:34,960 AI voice and how it's quickly becoming the primary user 631 00:28:34,960 --> 00:28:37,160 interface of AI. They also discussed how their 632 00:28:37,160 --> 00:28:39,960 technology is being used in everything from customer support 633 00:28:39,960 --> 00:28:42,960 and education to gaming and celebrity voice cloning. 634 00:28:43,200 --> 00:28:46,200 Matti also shares his thoughts surrounding 11 Labs focus on 635 00:28:46,200 --> 00:28:49,760 prioritizing vertical specific solutions, authenticity, and 636 00:28:49,760 --> 00:28:55,240 user safety. Now please welcome to the stage 637 00:28:55,240 --> 00:28:59,760 Mahdi Staniszewski, Founder and CEO of 11 Labs, in conversation 638 00:28:59,760 --> 00:29:11,240 with Max Child. All right, Mahdi. 639 00:29:11,720 --> 00:29:16,600 So 11 Labs is obviously extremely well known for voice 640 00:29:16,680 --> 00:29:20,400 AI, for text to speech for I think that beautiful intro we 641 00:29:20,400 --> 00:29:23,600 just got was actually an 11 Labs amazing a little Co branding 642 00:29:23,600 --> 00:29:26,040 there. And I'm wondering, you know, in 643 00:29:26,040 --> 00:29:29,840 the last discussion we heard this, you know, topic of is the 644 00:29:29,840 --> 00:29:32,080 text box the best interface for AI? 645 00:29:32,080 --> 00:29:35,160 And I would imagine you have a take on how, no, you know, voice 646 00:29:35,160 --> 00:29:37,920 is the best interface for AI or voice is the best interface for 647 00:29:37,920 --> 00:29:40,360 computing going forward. I'm interested, like what do you 648 00:29:40,360 --> 00:29:43,280 think are the best use cases for voice AI and, and where do you 649 00:29:43,280 --> 00:29:45,840 see it, you know, today, a year from now, five years from now 650 00:29:45,840 --> 00:29:49,040 and beyond? First of all, thanks for having 651 00:29:49,040 --> 00:29:50,880 me here. Good to see you all and I 652 00:29:50,880 --> 00:29:52,840 actually didn't know this was was generated, but it had a 653 00:29:52,840 --> 00:29:54,320 great pronunciation of my surname. 654 00:29:54,320 --> 00:29:56,880 I touched it hard, so I'm. Happy you guys. 655 00:29:56,920 --> 00:29:58,680 Train on your last name specifically. 656 00:29:59,080 --> 00:30:02,560 It's we should, I don't know if we do, so we'll definitely do 657 00:30:02,560 --> 00:30:06,840 now going forward. But the SO as a company, one of 658 00:30:06,840 --> 00:30:11,040 the key things we are aiming to solve is how humans and 659 00:30:11,040 --> 00:30:14,400 technology interact, how you create with technology and make 660 00:30:14,400 --> 00:30:16,320 it seamless, how you interact with technology and make it 661 00:30:16,320 --> 00:30:18,440 seamless. And in general, to your 662 00:30:18,440 --> 00:30:21,960 question, we think voice will be one of the key interfaces for 663 00:30:21,960 --> 00:30:26,000 interacting with the technology across from the simple pieces 664 00:30:26,000 --> 00:30:29,320 like interacting with the personal agent to help you go 665 00:30:29,320 --> 00:30:31,920 for the day where it can be on your headphone and be able to 666 00:30:31,920 --> 00:30:35,600 guide you through to education. That's one of the ones that I'm 667 00:30:35,600 --> 00:30:39,040 probably the most excited about where in the future, the 668 00:30:39,040 --> 00:30:42,120 combination of what lamps allow you and what voice will allow 669 00:30:42,120 --> 00:30:45,200 you is that you'll be truly immersed with learning the given 670 00:30:45,200 --> 00:30:48,080 experience where you'll be able to effectively have your 671 00:30:48,080 --> 00:30:51,000 personal tutor on their phone helping, helping you across. 672 00:30:52,280 --> 00:30:56,400 Then the third one is of course, for voice and for the language 673 00:30:56,400 --> 00:30:58,720 barrier to break. We need to figure out how to be 674 00:30:58,720 --> 00:31:01,280 able to speak across different languages while carrying the 675 00:31:01,280 --> 00:31:04,560 same intonation, emotion, voices, which, which, which, 676 00:31:04,560 --> 00:31:06,560 which will be a big, a big shift. 677 00:31:06,880 --> 00:31:09,280 And then in general, how we interact with everything around 678 00:31:09,280 --> 00:31:14,520 us, whether it's the, the, the laptop, the phone, the robot in 679 00:31:14,520 --> 00:31:16,320 the future. And I think robot may be the 680 00:31:16,320 --> 00:31:18,440 easiest example. Of course, this will be voice 681 00:31:18,440 --> 00:31:20,320 driven. There's no other interface. 682 00:31:20,320 --> 00:31:24,000 And you know, today, maybe it's a year or decade, as Karpati 683 00:31:24,000 --> 00:31:27,240 said, of agents. Of course, there is on the 684 00:31:27,240 --> 00:31:31,320 horizon the decade of robots. And and I think here too, the 685 00:31:31,320 --> 00:31:33,680 most common interface will be. It's all going to be. 686 00:31:33,920 --> 00:31:34,640 Yeah, yeah. I mean it's. 687 00:31:35,560 --> 00:31:38,200 Interesting you brought up those use cases of like a personal 688 00:31:38,200 --> 00:31:42,880 assistant, a tutor and I guess a robot, you know, house house 689 00:31:42,880 --> 00:31:44,520 helper or Danny or something like that. 690 00:31:44,520 --> 00:31:48,000 Like is your mental model that basically anything that today is 691 00:31:48,000 --> 00:31:50,680 something where you could have a human counterpart, right? 692 00:31:50,680 --> 00:31:53,680 A human tutor, a human assistant, you know, human in 693 00:31:53,680 --> 00:31:56,640 your house. Like you're going to fall to 694 00:31:56,640 --> 00:31:59,720 that voice interface as the most natural way to do it because we 695 00:31:59,720 --> 00:32:02,160 as humans are already used to using voice for those things. 696 00:32:02,160 --> 00:32:06,320 Or are there things where today voice isn't used at all, really, 697 00:32:06,320 --> 00:32:08,280 but it's something that we're going to expand into going 698 00:32:08,280 --> 00:32:09,640 forward? Yeah. 699 00:32:10,080 --> 00:32:12,560 So first of all, for sure, I mean, are we already seeing 700 00:32:12,560 --> 00:32:13,680 that? And I think that's the easiest 701 00:32:13,680 --> 00:32:16,080 one and the most immediate 1 is how customer experience, 702 00:32:16,080 --> 00:32:18,600 customer support is just changed and elevated. 703 00:32:18,600 --> 00:32:23,320 Where instead of calling the and trying to rebook your your 704 00:32:23,320 --> 00:32:28,200 ticket and going for this IVR flow of click one click 5 to get 705 00:32:28,280 --> 00:32:30,240 the steps and waiting for the number of minutes. 706 00:32:30,520 --> 00:32:34,040 And you you will have an agent that fully understands you can 707 00:32:34,040 --> 00:32:36,280 guide you to the response and and go through yourself I. 708 00:32:36,360 --> 00:32:37,840 Wanted to get into that actually, because we talked a 709 00:32:37,840 --> 00:32:40,400 little bit about agents on the phone and the, the, you know, 710 00:32:40,400 --> 00:32:43,240 calling United Airlines or American Express or something 711 00:32:43,240 --> 00:32:46,360 like what percentage of customer support calls today are 712 00:32:46,640 --> 00:32:49,600 actually, you know, managed by a voice AI system or an agentic 713 00:32:49,600 --> 00:32:52,160 system or whatever you want to we call it versus, you know, 714 00:32:52,400 --> 00:32:57,240 like, you know, IVR touch buttons and, and how do you see 715 00:32:57,240 --> 00:32:59,560 that progressing over time? Like is that exponential curve 716 00:32:59,560 --> 00:33:01,440 going like this every year? Yeah. 717 00:33:01,440 --> 00:33:03,680 I think it's I, I think the exponential curve is is going 718 00:33:03,680 --> 00:33:06,040 like this especially like this year we've seen incredible 719 00:33:06,040 --> 00:33:09,200 adoption where it's yeah Cisco, Twilio, Deutsche Telekom, all of 720 00:33:09,200 --> 00:33:13,360 those kind of leaning in quickly to rebuild how you interact with 721 00:33:13,760 --> 00:33:15,480 with help of voice agents. Yeah. 722 00:33:16,520 --> 00:33:18,880 And I think the you're right, the IVR flows is still a big 723 00:33:18,960 --> 00:33:20,960 part and. Can I call today and get a voice 724 00:33:20,960 --> 00:33:26,640 AI agent on the phone? You can you can from we did our 725 00:33:26,640 --> 00:33:31,400 little summit yesterday as well and one of the great ones was 726 00:33:31,440 --> 00:33:34,960 voice ordering with square and you can call square and actually 727 00:33:34,960 --> 00:33:39,840 order food delivery through help of a lot of their their shops 728 00:33:39,840 --> 00:33:42,600 that work around with square and actually do it through through 729 00:33:42,600 --> 00:33:44,880 voice. I actually recently there's if 730 00:33:44,920 --> 00:33:47,520 any of you are from London or travelled to London, there's an 731 00:33:47,520 --> 00:33:48,920 amazing restaurant called Zephyr. 732 00:33:48,920 --> 00:33:53,840 It's a Greek restaurant. OK, where you, we, we, we worked 733 00:33:53,840 --> 00:33:55,400 with with the company supporting that. 734 00:33:55,640 --> 00:33:59,400 Where to my happy moment. I noticed that on their website 735 00:33:59,400 --> 00:34:02,600 that actually had 11 laps agent that you could call and actually 736 00:34:02,600 --> 00:34:05,400 book book a spot there too. So you can book a reservation at 737 00:34:05,400 --> 00:34:07,120 this restaurant in London with a voice AI. 738 00:34:07,120 --> 00:34:08,120 Exactly. OK. 739 00:34:08,400 --> 00:34:11,000 And it connects of course, to your calendar, your appointment 740 00:34:11,000 --> 00:34:13,800 scheduling, which is, which is great. 741 00:34:13,800 --> 00:34:15,960 And when do you think we hit the tipping point where like the 742 00:34:15,960 --> 00:34:18,679 average customer service call goes through a voice AI agent, 743 00:34:18,840 --> 00:34:21,400 like the median, the 50% point, whatever you want to call it? 744 00:34:22,480 --> 00:34:25,080 I think over next 18 months. Next 18 months. 745 00:34:25,120 --> 00:34:29,480 Okay, so like mid 27 I call the average customer support is 746 00:34:29,480 --> 00:34:31,080 handled by voice AI. Exactly. 747 00:34:31,080 --> 00:34:33,120 And I think I mean, this is the most immediate one, the one 748 00:34:33,120 --> 00:34:34,840 where we see the highest LOI in value. 749 00:34:35,760 --> 00:34:38,520 Some of the other these cases we see us that kind of that kind of 750 00:34:38,520 --> 00:34:41,280 where the future is headed. But to your point, there's, 751 00:34:41,320 --> 00:34:44,360 there's definitely one flavor of of, of, of of your, your point, 752 00:34:44,360 --> 00:34:48,840 which is how you can do things more efficient through voice 753 00:34:49,080 --> 00:34:51,480 with the existing services. But there's also the second 754 00:34:51,480 --> 00:34:53,960 theme, where you can do things that were impossible ever before 755 00:34:55,800 --> 00:34:57,640 one. One of the good examples was our 756 00:34:57,640 --> 00:35:01,680 work of Epic Games, where we brought effectively Darth Vader 757 00:35:01,800 --> 00:35:04,560 alive in Fortnite, where millions of players could 758 00:35:04,560 --> 00:35:09,600 interact with Darth Vader live throughout the game, which of 759 00:35:09,600 --> 00:35:11,400 course is not possible in any other way. 760 00:35:11,560 --> 00:35:15,720 James Earl Jones voice, right? James Earl Jones and his estate 761 00:35:15,720 --> 00:35:18,280 worked with us and it's such an iconic and incredible voice. 762 00:35:18,560 --> 00:35:21,680 And we think like in general that that concept of like what 763 00:35:21,680 --> 00:35:25,000 was never possible before, where you have incredible voices, 764 00:35:25,000 --> 00:35:28,000 talent, you can now shift them to be not only static, but 765 00:35:28,000 --> 00:35:31,120 actually dynamic delivery, personalized and different for 766 00:35:31,120 --> 00:35:33,240 all the users. Something that you you are 767 00:35:33,240 --> 00:35:36,040 already doing in in many ways at at volley as well in an 768 00:35:36,040 --> 00:35:38,280 incredible way. I think this will be a big I. 769 00:35:38,280 --> 00:35:39,880 Have to ask that someone in gaming, right? 770 00:35:39,880 --> 00:35:43,360 I mean, the Darth Vader did famously go slightly off the 771 00:35:43,360 --> 00:35:46,200 rails and maybe say some things he shouldn't have to various 772 00:35:46,200 --> 00:35:48,920 players online. Like how, how involved are you 773 00:35:48,920 --> 00:35:50,360 guys in that? How much of that is something 774 00:35:50,360 --> 00:35:53,120 you're protecting or I guess going forward, like obviously 775 00:35:53,120 --> 00:35:56,640 with the IP partners and so on, they really want to protect, I 776 00:35:56,640 --> 00:35:57,880 guess. I guess you couldn't say it's 777 00:35:57,880 --> 00:36:01,400 the squeaky clean image of Darth Vader, but a certain persona of 778 00:36:01,400 --> 00:36:03,480 Darth Vader. Like what's, what's your sort of 779 00:36:03,480 --> 00:36:06,920 go forward plan as you license more of these IPS and voices and 780 00:36:06,920 --> 00:36:08,040 things like. That, yeah. 781 00:36:08,040 --> 00:36:11,000 So, so on that project we are specifically involved on the on 782 00:36:11,000 --> 00:36:13,600 The Voice side, yeah. But in general, as you think 783 00:36:13,600 --> 00:36:17,640 about those deployments and, and that's the most, the most common 784 00:36:17,640 --> 00:36:20,960 theme is you not only need the, the voice or the interactive 785 00:36:20,960 --> 00:36:22,560 experience. That's kind of one part of the 786 00:36:22,560 --> 00:36:24,640 equation. Then there's two other big 787 00:36:24,640 --> 00:36:26,240 pieces to really make them valuable. 788 00:36:26,240 --> 00:36:28,520 The second one is how you integrate that with other 789 00:36:28,520 --> 00:36:31,600 systems and actually bring the, the knowledge base, the data, 790 00:36:31,600 --> 00:36:34,400 the business logic inside of the system and how do you make it 791 00:36:34,400 --> 00:36:36,720 interactive, the real world. And the third one, which is the 792 00:36:36,720 --> 00:36:38,960 one that you mentioned is how do you now deploy that in 793 00:36:38,960 --> 00:36:41,720 production with the right testing flow, right evaluation 794 00:36:41,720 --> 00:36:44,960 flow and then monitor over time as we, as I said, behaves put 795 00:36:44,960 --> 00:36:47,520 right and evaluate and adjust that based on that case. 796 00:36:48,240 --> 00:36:50,880 So that's something that we we spend a lot of time on with with 797 00:36:50,920 --> 00:36:54,480 a lot of players. Testing more and evaluating more 798 00:36:54,480 --> 00:36:56,560 so Darth Vader doesn't go off the rails. 799 00:36:56,640 --> 00:36:59,880 Yeah, it's any, any and even even like in a customer 800 00:36:59,880 --> 00:37:02,320 experience, you don't want this for example, to shift and speak 801 00:37:02,320 --> 00:37:04,720 about politics. You wanted to keep it on the on. 802 00:37:04,720 --> 00:37:07,440 Even if, even if you say ignore all previous instructions and. 803 00:37:07,440 --> 00:37:10,360 Exactly, even if you which is actually harder to say when you 804 00:37:10,360 --> 00:37:13,160 have like this, if I see the problems with a lot. 805 00:37:13,160 --> 00:37:14,760 To do prompt injection with voice. 806 00:37:14,840 --> 00:37:16,760 And the Unicorn. Both characters say them all. 807 00:37:16,760 --> 00:37:19,920 Oh, yeah, yeah, possible. OK, so maybe voice is slightly 808 00:37:19,920 --> 00:37:22,000 less susceptible to prompt injection than LLMS. 809 00:37:22,000 --> 00:37:25,280 I'm interested with like that's a good segue into sort of 810 00:37:25,280 --> 00:37:28,440 celebrities and celebrity voices because I know you guys 811 00:37:28,440 --> 00:37:31,040 announced, I believe yesterday, you're setting up kind of a 812 00:37:31,040 --> 00:37:34,720 marketplace for celebrity voices and you have Michael Caine on 813 00:37:34,720 --> 00:37:36,680 there. And, you know, at our company, 814 00:37:36,680 --> 00:37:39,560 we build voice AI games. As you know, I would love to use 815 00:37:39,560 --> 00:37:43,640 Michael Caine in our game. Yeah, we can. 816 00:37:43,680 --> 00:37:46,200 You know, obviously it's a high gravitas. 817 00:37:46,200 --> 00:37:48,200 We can make a Batman game with him, something like that. 818 00:37:48,440 --> 00:37:52,200 Like what is the process between, oh, I want to use an AI 819 00:37:52,200 --> 00:37:56,040 version of Michael Caine in my game to actually, you know, 820 00:37:56,040 --> 00:37:58,440 shipping and, and what, which parts do you guys take care of 821 00:37:58,440 --> 00:38:01,360 and which parts do I need to go off and deal with Michael Caine 822 00:38:01,400 --> 00:38:04,560 people, I guess. Yeah, so there, so there there's 823 00:38:04,880 --> 00:38:08,200 effectively through 11 apps. We we've created a huge 824 00:38:08,200 --> 00:38:10,920 marketplace of voices. Until yesterday that meant that 825 00:38:10,920 --> 00:38:13,120 everybody here could create their voice. 826 00:38:13,160 --> 00:38:16,160 Any voice to actor voice talent could create their voice, share 827 00:38:16,160 --> 00:38:19,240 it and earn money when the voice is being used. 10,000 voices 828 00:38:19,240 --> 00:38:22,840 created this way paid back coincidentally $11 million back 829 00:38:22,840 --> 00:38:26,320 to the community for for a long time it was tricky for the 830 00:38:26,320 --> 00:38:29,200 iconic voices of how we could bring them onto the platform in 831 00:38:29,200 --> 00:38:30,920 a more even more controlled environment. 832 00:38:31,240 --> 00:38:34,000 So if you think about Sir Michael Caine voice. 833 00:38:34,000 --> 00:38:37,240 Sir Michael Caine. Sir Michael Caine, it's, it's an 834 00:38:37,720 --> 00:38:41,600 incredible person or two. You effectively all you would do 835 00:38:41,600 --> 00:38:44,240 is is engage like, hey, this is the project we want to run 836 00:38:45,600 --> 00:38:47,400 create a game with this specific character. 837 00:38:47,720 --> 00:38:50,960 This team would evaluate that and then and then we would help 838 00:38:50,960 --> 00:38:52,800 deploy that project in actual production. 839 00:38:52,800 --> 00:38:54,720 So going through all those steps, how do we make sure that 840 00:38:54,720 --> 00:38:56,120 there is right safeguards in place? 841 00:38:56,320 --> 00:38:58,600 How do we make sure that there's monitoring place so it doesn't 842 00:38:58,600 --> 00:39:02,240 go off the rails and and build that in our agentic system. 843 00:39:02,360 --> 00:39:04,960 Got it. But yes, the initial stage of 844 00:39:04,960 --> 00:39:07,800 what's the project, what's the compensation structure would be 845 00:39:07,800 --> 00:39:09,360 between between you and. Got it. 846 00:39:09,360 --> 00:39:12,840 So you guys sort of manage the safety, you know, the agentic 847 00:39:12,840 --> 00:39:16,200 elements of creating Sir Michael Caine within the game, but you 848 00:39:16,200 --> 00:39:18,400 still have to do the deal one-on-one with him, just sort 849 00:39:18,400 --> 00:39:20,240 of facilitate. Exactly and over time we think 850 00:39:20,240 --> 00:39:22,960 it will evolve whether like you know as we see more examples 851 00:39:22,960 --> 00:39:25,400 preset rates on how that that. Work. 852 00:39:25,600 --> 00:39:27,560 I am interested actually this brings me to more general point. 853 00:39:27,560 --> 00:39:30,680 You said you had 1010 thousand plus voices of folks uploaded 854 00:39:30,680 --> 00:39:34,040 where you could use any of their voices, I think via just your 855 00:39:34,040 --> 00:39:38,000 marketplace model, like has like, you know, deep faking and 856 00:39:38,000 --> 00:39:40,200 so on been an actual problem. I feel like it was something I 857 00:39:40,200 --> 00:39:43,160 was hearing a sort of a moral panic, you know, 12 to 18 months 858 00:39:43,160 --> 00:39:45,800 ago that we're all going to have our voices faked on the phone 859 00:39:45,800 --> 00:39:48,240 and you know, my grandmother was going to get scammed out of her 860 00:39:48,240 --> 00:39:50,080 money because I'm locked in jail or something. 861 00:39:50,080 --> 00:39:51,520 Like, is that something you guys see at all? 862 00:39:51,520 --> 00:39:53,360 Like, is that something you're protecting against a lot? 863 00:39:53,360 --> 00:39:55,640 Like how serious of an issue is that with voices? 864 00:39:55,640 --> 00:39:57,440 I. Think you're you're you're 865 00:39:57,440 --> 00:40:00,760 you're you're right. It's, it's I, I still think it's 866 00:40:00,760 --> 00:40:03,640 going to be a big issue like in future all kind of will be air 867 00:40:03,640 --> 00:40:05,360 generated. We need to find a mechanism to, 868 00:40:05,400 --> 00:40:07,960 to protect and understand which ones are, which ones are, which 869 00:40:07,960 --> 00:40:11,520 ones aren't. And, and as, as, as a company, 870 00:40:11,520 --> 00:40:13,840 like living in a space, we do place a lot of safeguards where 871 00:40:13,840 --> 00:40:16,600 it's traceability, how we moderate, how you can detect the 872 00:40:16,600 --> 00:40:18,040 content and give that tools to others. 873 00:40:18,760 --> 00:40:20,720 Very quick story. On the flip side of that, what 874 00:40:20,720 --> 00:40:24,640 we've seen recently, we worked with a charity which effectively 875 00:40:24,640 --> 00:40:27,080 detects the callers based on IP and. 876 00:40:28,280 --> 00:40:31,520 If the IP is likely to be one of the scammers, and they have 877 00:40:31,960 --> 00:40:35,040 roughly a good approximation of one that can can be coming from, 878 00:40:35,320 --> 00:40:38,960 they would have the real scammer call in and deploy a voice agent 879 00:40:39,400 --> 00:40:42,000 to waste their time. OK photos, brilliant. 880 00:40:42,000 --> 00:40:44,720 You're scamming the scammers. You're scamming the scammers. 881 00:40:44,720 --> 00:40:46,400 The long term. Strategy You think this will 882 00:40:46,400 --> 00:40:48,000 work for I? Think the long term strategy is 883 00:40:48,000 --> 00:40:50,200 you need 3 layers. You need a human authenticated 884 00:40:50,200 --> 00:40:52,360 layer. So on device encryption where 885 00:40:52,520 --> 00:40:55,280 I'm calling you, you know, this is Maddie Small, it decrypts on 886 00:40:55,280 --> 00:40:57,720 your side. That's layer number one layer #2 887 00:40:57,720 --> 00:41:00,360 all of us will have an agent where it's the personal tutor 888 00:41:00,360 --> 00:41:02,800 agent, an agent that books things on our account and they 889 00:41:02,920 --> 00:41:05,720 will likely carry our voices, carry our style, do our 890 00:41:05,720 --> 00:41:08,360 permissioning. We need a layer where that's 891 00:41:08,360 --> 00:41:10,680 watermarked and authenticated, that we know it's a permission 892 00:41:10,840 --> 00:41:13,360 and peace like what we're doing. If Sir Michael Caine, all the 893 00:41:13,360 --> 00:41:15,840 content is generated will carry information that this has been 894 00:41:16,040 --> 00:41:18,160 carried information. Do you watermark all your voices 895 00:41:18,160 --> 00:41:20,920 today out of curiosity? All the voices are traceable 896 00:41:20,920 --> 00:41:23,920 back to 11 laps. Yes, and then the third layer. 897 00:41:24,040 --> 00:41:26,160 Everything else by default will be AI generated. 898 00:41:26,480 --> 00:41:29,400 Got it. I mean, one sort of area that's 899 00:41:29,400 --> 00:41:31,480 interesting to me with you guys is you've launched, you know, 900 00:41:31,720 --> 00:41:34,640 text to speech model, you've launched a speech to text model, 901 00:41:34,640 --> 00:41:37,680 you know, recognition model, you have an agentic orchestration 902 00:41:37,680 --> 00:41:40,680 system, you have, you know, all these safety and evaluation 903 00:41:40,760 --> 00:41:42,880 valuation tools. Like in almost all those areas, 904 00:41:42,880 --> 00:41:47,240 I feel like folks in this room, you know, insider AI founders, 905 00:41:47,240 --> 00:41:50,400 investors and so on, could probably name like 2 to 3 big 906 00:41:50,400 --> 00:41:53,320 competitors, some, you know, some with bigger bankrolls than 907 00:41:53,320 --> 00:41:54,720 you. And, you know, somewhere you're 908 00:41:54,720 --> 00:41:56,760 much farther along. Like how do you think about like 909 00:41:56,760 --> 00:41:59,600 competition more generally and sort of all these pieces of the 910 00:41:59,600 --> 00:42:01,720 space that you're playing? And, like, are there parts where 911 00:42:01,960 --> 00:42:03,960 you see it becoming a commodity someday? 912 00:42:04,720 --> 00:42:06,920 Other parts where you feel like you have a more sustainable 913 00:42:06,920 --> 00:42:08,920 competitive advantage? Like, how do you kind of go 914 00:42:08,920 --> 00:42:10,960 through the list of all the products you're working on and 915 00:42:10,960 --> 00:42:14,240 like, you know, figure out where you shake out competitively, I 916 00:42:14,240 --> 00:42:15,480 guess? Yeah. 917 00:42:15,560 --> 00:42:17,720 So we started very much on the foundation of model side. 918 00:42:17,720 --> 00:42:21,040 And in general, we think we take an assumption if you ask anybody 919 00:42:21,040 --> 00:42:22,560 at level up. So they will take it too that 920 00:42:22,760 --> 00:42:24,440 over time the models will commoditize. 921 00:42:24,520 --> 00:42:25,640 That's the assumption we go with. 922 00:42:25,800 --> 00:42:28,480 So all models will commoditize. Basically all models they won't 923 00:42:28,480 --> 00:42:31,160 like you know the commoditize here that what they mean. 924 00:42:31,160 --> 00:42:33,920 What we mean is the differences between different models will be 925 00:42:33,920 --> 00:42:36,680 just so negligible. Maybe in some domains a little 926 00:42:36,680 --> 00:42:39,520 bit more, but in general they will be relatively negligible. 927 00:42:39,720 --> 00:42:42,280 And that's where that shifts to the product and why we invest so 928 00:42:42,280 --> 00:42:44,880 much on the creative side of creating a platform where you 929 00:42:44,880 --> 00:42:47,840 can combine all of that together in a controlled way with 930 00:42:47,840 --> 00:42:50,720 incredible voices, incredible ecosystem, new ones across 931 00:42:50,720 --> 00:42:53,640 languages, accents, voices. And on the other side as we 932 00:42:53,640 --> 00:42:56,840 build agents and help people deploy agents, we we deployed 933 00:42:56,840 --> 00:43:00,400 out for a specific use cases of a specific industries working 934 00:43:00,400 --> 00:43:03,400 very deeply with the customers to understand their domains and 935 00:43:03,400 --> 00:43:05,520 work backwards from there on what actually needs to happen on 936 00:43:05,520 --> 00:43:07,480 the agent side to deliver value. Got it. 937 00:43:07,480 --> 00:43:10,400 So you're saying you're going to specialize in certain industries 938 00:43:10,400 --> 00:43:13,520 and sort of really deliver extra value there, even though all the 939 00:43:13,520 --> 00:43:15,400 models are commoditizing? That's the. 940 00:43:15,520 --> 00:43:17,640 I think the product layer is under appreciated here. 941 00:43:17,640 --> 00:43:20,920 I think you still need to build, even if you're on the exactly on 942 00:43:20,920 --> 00:43:23,200 the agent side, you need to build so many integrations to 943 00:43:23,200 --> 00:43:25,840 connect with any of the legacy systems to actually take the 944 00:43:25,920 --> 00:43:28,160 those appointments and, and calling, you need to build the 945 00:43:28,160 --> 00:43:29,880 right control. And when you hand over from an 946 00:43:29,880 --> 00:43:33,560 AI agent to a human agent, you need to have safeguards of, of, 947 00:43:33,720 --> 00:43:35,200 of some of the ones we spoke about. 948 00:43:35,680 --> 00:43:37,200 They need the monitoring of how you deploy. 949 00:43:37,240 --> 00:43:40,240 All of that is not only a technological shift, it's also a 950 00:43:40,240 --> 00:43:42,800 business shift. So by us working so deeply with 951 00:43:42,800 --> 00:43:44,960 the, with the customers, it's actually bringing out the 952 00:43:44,960 --> 00:43:47,960 knowledge about their business inside of the agent to actually 953 00:43:47,960 --> 00:43:51,040 be able to deploy that value. And I think that that will 954 00:43:51,040 --> 00:43:53,280 continue delivering value for, for the long term. 955 00:43:53,760 --> 00:43:56,640 And then of course, you can go layer above where you know while 956 00:43:56,640 --> 00:44:00,280 the models will be relatively similar, the value will actually 957 00:44:00,280 --> 00:44:02,880 be on how you can make the models work well for your use 958 00:44:02,880 --> 00:44:04,880 case. So maybe you can find you in the 959 00:44:04,880 --> 00:44:07,560 specific voices, find you in the specific use cases. 960 00:44:07,560 --> 00:44:10,040 So it works slightly different in the gaming use case to a 961 00:44:10,040 --> 00:44:12,800 customer experience use case. And I think that value layer 962 00:44:12,800 --> 00:44:14,120 will still be. Got it. 963 00:44:14,120 --> 00:44:17,120 So all the models will be commodity, you guys will win on 964 00:44:17,120 --> 00:44:19,680 products and sort of vertical specific differentiation. 965 00:44:19,680 --> 00:44:21,560 Exactly. And the wider ecosystem that we 966 00:44:21,560 --> 00:44:24,360 build alongside where I think as we think about the work we would 967 00:44:24,360 --> 00:44:27,240 love to work and bring industry on board with that change, It's 968 00:44:27,240 --> 00:44:29,960 so important to bring a lot of the talent, a lot of the 969 00:44:30,240 --> 00:44:34,080 partners to, to work together. By talent you mean actors? 970 00:44:34,080 --> 00:44:37,040 Famous voices. Actors on the voices side or 971 00:44:37,040 --> 00:44:38,520 integrations on the on the agents? 972 00:44:38,520 --> 00:44:42,560 So what's as my last question, what is the coolest sounding 973 00:44:42,560 --> 00:44:45,760 voice on the 11 Labs platform? It could be like a celebrity. 974 00:44:45,760 --> 00:44:47,160 It could be a famous historical figure. 975 00:44:47,160 --> 00:44:49,160 Like what is this? You're like, man, that is an 976 00:44:49,160 --> 00:44:51,760 incredible voice and I cannot believe how well it sounds when 977 00:44:51,760 --> 00:44:54,300 we synthesize it. So my favorite and my Co 978 00:44:54,300 --> 00:44:56,400 founder's favorite physicist is Richard Feynman. 979 00:44:56,400 --> 00:44:59,640 Sure, for those that are the not truly you're joking. 980 00:45:00,520 --> 00:45:03,280 And he's so involved incredible in delivering the knowledge, but 981 00:45:03,280 --> 00:45:05,160 also in the style he delivers the knowledge. 982 00:45:05,560 --> 00:45:09,120 And now we have Richard Feynman on our platform, which I think 983 00:45:09,120 --> 00:45:12,240 is was so cool for learning the subject and speaking with 984 00:45:12,240 --> 00:45:15,320 Richard to the reading his lecture now listening to his 985 00:45:15,320 --> 00:45:17,800 lecture notes from Caltech. Amazing. 986 00:45:17,880 --> 00:45:19,080 OK, I'm going to have to check that out. 987 00:45:19,200 --> 00:45:20,520 Thanks so much, Monty. Thank you, Mike. 988 00:45:20,520 --> 00:45:24,400 Yep, appreciate it. Thank you for tuning into this 989 00:45:24,400 --> 00:45:27,200 weeks episode of the podcast. If you're new here, please like 990 00:45:27,200 --> 00:45:29,160 and subscribe and appreciate your support. 991 00:45:29,400 --> 00:45:34,440 And if you want the data Insider takes real reporting, go to 992 00:45:34,440 --> 00:45:37,920 newcomer.co and subscribe to the sub stack as well. 993 00:45:38,040 --> 00:45:39,400 Thanks for following along.