1 00:00:00,010 --> 00:00:01,557 What's happening right now is a big deal. 2 00:00:01,640 --> 00:00:03,690 This is the biggest launch since ChatGPT. 3 00:00:03,930 --> 00:00:07,770 It's called Muse. "Muse." "Muse." "Has the tech world buzzing." "5 4 00:00:08,160 --> 00:00:13,070 million downloads in 22 days." "Surpassing ChatGPT" And he built it. 5 00:00:13,280 --> 00:00:15,777 Most people are using AI to answer questions, basically 6 00:00:15,860 --> 00:00:18,910 glorified Google search. This is something totally different. 7 00:00:19,150 --> 00:00:23,297 "It's an AI agent" "Built for everyday life." "We believe that Muse will 8 00:00:23,380 --> 00:00:26,417 make you money." "One guy had Muse help plan a weekend away with his 9 00:00:26,500 --> 00:00:28,327 wife." It can make phone calls. 10 00:00:28,410 --> 00:00:29,927 It can do tasks for you. 11 00:00:30,010 --> 00:00:34,250 But for it to work, you have to trust it, first with your personal information, 12 00:00:34,550 --> 00:00:40,250 then to do what you actually want it to do, and then that it's not gonna, well, 13 00:00:40,340 --> 00:00:44,467 get out of control, all at a time when the most important AI leaders are warning 14 00:00:44,550 --> 00:00:47,637 us that AI is getting out of control. 15 00:00:47,720 --> 00:00:49,890 And I urge you to urge me to stop. 16 00:00:50,700 --> 00:00:53,767 If you understand what's happening right now with Muse, you'll understand 17 00:00:53,850 --> 00:00:55,810 the biggest things happening in AI. 18 00:00:56,220 --> 00:00:59,130 So what can it actually do? When should we trust it? 19 00:00:59,380 --> 00:01:03,480 And what do we need to do to make sure this goes right? 20 00:01:03,860 --> 00:01:06,557 And we're at the moment of, like, peak insanity. 21 00:01:06,640 --> 00:01:09,397 Why are all these billionaires obsessed with restaurant reservations? 22 00:01:09,480 --> 00:01:12,397 We have a shared interest in making sure this technology 23 00:01:12,480 --> 00:01:13,647 does not go horribly wrong. 24 00:01:13,730 --> 00:01:16,497 Is that something that's happening in some private room somewhere? 25 00:01:16,580 --> 00:01:21,787 The degree to which certain personal animosities are actually defining 26 00:01:21,870 --> 00:01:25,470 how they're developing this really important technology is insane to me. 27 00:01:30,990 --> 00:01:34,350 Amazing Let's do this. Okay. Welcome. 28 00:01:34,620 --> 00:01:36,387 How have you been? I'm good. I'm really good. 29 00:01:36,470 --> 00:01:39,997 How have you been? This is- It's been kind of insane, yeah. 30 00:01:40,080 --> 00:01:43,380 Yeah. What does it feel like? 31 00:01:43,760 --> 00:01:48,587 We didn't expect that it would be such a hit, so I feel amazing for, like, the 32 00:01:48,670 --> 00:01:51,457 whole team and everyone who, like, worked so hard on it 'cause it's, like, it's 33 00:01:51,540 --> 00:01:53,347 really come together and become a thing. 34 00:01:53,430 --> 00:01:55,117 So… That's what I'm excited to talk about. 35 00:01:55,200 --> 00:01:57,057 I know you know all about the show already. 36 00:01:57,140 --> 00:02:00,330 But this isn't, like, a traditional interview. 37 00:02:00,630 --> 00:02:02,127 Like, I'm not gonna ask about financials. 38 00:02:02,210 --> 00:02:03,637 I'm not gonna ask about management style. 39 00:02:03,720 --> 00:02:09,670 You just launched one of the most popular consumer tools of all time, and so the 40 00:02:09,760 --> 00:02:13,877 only thing that I wanna know is what's the future that you're working toward 41 00:02:13,960 --> 00:02:18,787 with that tool, and how can this audience that's really optimistic and really wants 42 00:02:18,870 --> 00:02:23,350 to be part of making sure that that future go right, how can they feel like they're 43 00:02:23,470 --> 00:02:25,600 part of the decisions that are being made? 44 00:02:26,240 --> 00:02:28,007 Yep. All of that. Yeah. That's what we do. 45 00:02:28,090 --> 00:02:31,897 One of the reasons I wanted to talk to you was I've always 46 00:02:31,980 --> 00:02:34,060 appreciated that you're really blunt. 47 00:02:36,830 --> 00:02:42,307 Like, I, I honestly think that people need straight talk about this. 48 00:02:42,390 --> 00:02:46,007 I think that they feel confused, and I wanna get at what's really 49 00:02:46,090 --> 00:02:51,130 happening, what have you built, how are people using it, how are they 50 00:02:51,240 --> 00:02:53,470 not, what are the real challenges. 51 00:02:54,020 --> 00:02:57,137 I think one other interesting thing is given how popular Muse 52 00:02:57,220 --> 00:03:03,760 has become, I want to use it as a way to talk about other things in AI. 53 00:03:04,350 --> 00:03:05,977 Awesome. Okay. Let's get into it. 54 00:03:06,060 --> 00:03:09,210 Muse is the fastest-growing consumer tool since ChatGPT. 55 00:03:09,330 --> 00:03:13,607 Faster than ChatGPT, I think, and it feels like this is the moment when 56 00:03:13,690 --> 00:03:18,857 we're moving maybe from chatbots to agents, but there have been a lot of 57 00:03:18,940 --> 00:03:23,347 other popular agents, mostly in code, and I'm curious, why do you think 58 00:03:23,430 --> 00:03:24,827 this is happening for Muse right now? 59 00:03:24,910 --> 00:03:26,510 Why is Muse exploding like this? 60 00:03:26,940 --> 00:03:31,117 I think one of the mysteries of AI so far this year, for most of the year, has been 61 00:03:31,200 --> 00:03:36,737 that agents were happening for engineers and software engineers and coders, and 62 00:03:36,820 --> 00:03:41,097 then nobody else in the world, like, was experiencing it, knew what it was, like, 63 00:03:41,180 --> 00:03:42,907 understood exactly what was happening. 64 00:03:42,990 --> 00:03:47,267 And so for most of this year, you've kind of had this, like, two dialogues 65 00:03:47,350 --> 00:03:52,007 almost where people who do code and could see how powerful AI was becoming 66 00:03:52,090 --> 00:03:56,560 and, and how useful it was becoming were like, "Oh my gosh, this is incredible." 67 00:03:56,720 --> 00:03:59,167 And then everyone else was kind of like, "What are you talking about? 68 00:03:59,250 --> 00:04:00,207 I just still use a chatbot. 69 00:04:00,290 --> 00:04:03,067 Like, it's, like, a little bit better than Google, but what are you guys all 70 00:04:03,150 --> 00:04:07,907 freaked out about?" And, um, and one of the things that we were really excited 71 00:04:07,990 --> 00:04:12,387 to do is how do we take the magic that engineers and coders and developers 72 00:04:12,470 --> 00:04:15,457 are seeing and bring that to everyone? 73 00:04:15,540 --> 00:04:18,107 One of the things that, that has happened over the course of the past 74 00:04:18,190 --> 00:04:22,417 year is, like, the AI models started out as pretty janky agents, and you 75 00:04:22,500 --> 00:04:25,707 kind of had to struggle through them to get them to work, and then they 76 00:04:25,790 --> 00:04:27,790 slowly got better and better and better. 77 00:04:27,940 --> 00:04:31,497 And most of the industry's been really focused on making them just, like- Better 78 00:04:31,580 --> 00:04:33,057 and better and better at coding. 79 00:04:33,140 --> 00:04:36,247 But we really felt this was like a s- a big gap because most people in the 80 00:04:36,330 --> 00:04:37,917 world, they don't know how to code. 81 00:04:38,000 --> 00:04:39,877 They d- probably never even want to code. 82 00:04:39,960 --> 00:04:45,347 Um, but they still wanna feel the experience of agents and, um, having 83 00:04:45,430 --> 00:04:49,937 AI that is able to do more for you and actually start taking on more end-to-end 84 00:04:50,020 --> 00:04:52,287 workflows and take on more tasks for you. 85 00:04:52,370 --> 00:04:56,400 So that was really the goal behind Muse, and I think we spent a lot of time 86 00:04:56,900 --> 00:05:02,100 polishing all the little details to make sure that it would just work for people 87 00:05:02,400 --> 00:05:06,940 because we knew that, like, developers can kind of withstand a lot of jank 88 00:05:07,210 --> 00:05:10,347 and a lot of stuff breaking and, like, they can kind of deal with all of that 89 00:05:10,430 --> 00:05:12,057 because, you know, they're, they're… 90 00:05:12,140 --> 00:05:15,457 It's technology. They're used to dealing with, you know- Yeah … unripe technology. 91 00:05:15,540 --> 00:05:18,717 But for everyone, you have to make something that's really smooth 92 00:05:18,800 --> 00:05:20,350 and, and really works for them. 93 00:05:20,580 --> 00:05:26,647 That also helps explain the big gap, I think, between what it feels like for most 94 00:05:26,730 --> 00:05:31,157 people to actually use AI, when they were using a chatbot in particular, and the 95 00:05:31,240 --> 00:05:35,577 conversation around AI that's it's gonna take everyone's jobs, it's really scary. 96 00:05:35,660 --> 00:05:38,667 It's, you know, there's a, there's a macro conversation happening, and then there's 97 00:05:38,750 --> 00:05:40,297 an experience that people are having. 98 00:05:40,380 --> 00:05:43,207 And it just seems like those things are really, really far apart. 99 00:05:43,290 --> 00:05:47,467 And to your point, Muse is giving more people the opportunity to have 100 00:05:47,550 --> 00:05:51,770 the insight that computer engineers had, you know, call it a year ago. 101 00:05:52,240 --> 00:05:54,047 Yeah. Is that fair? I, I think that's true, yeah. 102 00:05:54,130 --> 00:05:57,690 Yeah. There are also moments in history that I think feel similar. 103 00:05:57,780 --> 00:06:01,487 Would you say it's like the mobile phone moment for agents? 104 00:06:01,570 --> 00:06:05,570 There are, you know, billions of people brought online by mobile phones that 105 00:06:05,660 --> 00:06:09,277 never had PCs because it was easy and simple to use, and they didn't, you 106 00:06:09,360 --> 00:06:13,007 know. I actually hadn't thought about that until you just mentioned it, but I 107 00:06:13,090 --> 00:06:15,040 think it's a great analogy because it's 108 00:06:15,150 --> 00:06:17,077 These models, they're improving really fast, and they're 109 00:06:17,160 --> 00:06:18,427 improving really, really quickly. 110 00:06:18,510 --> 00:06:22,620 And it's really important that we, as those models get better, that 111 00:06:22,820 --> 00:06:27,877 the, the superpowers that they, they lend can be given to, like, literally 112 00:06:27,960 --> 00:06:29,297 everyone all around the world. 113 00:06:29,380 --> 00:06:33,567 The part that I like about the analogy is, like, um, so much of it is just 114 00:06:33,650 --> 00:06:39,647 about figuring out how to package it all together into a form factor that, um, 115 00:06:39,730 --> 00:06:42,817 will resonate with people, and that's what we spent a lot of time trying to 116 00:06:42,900 --> 00:06:45,390 craft and get right and, and make perfect. 117 00:06:45,510 --> 00:06:49,747 Which also gets to one of the interesting initial reactions that I've heard 118 00:06:49,830 --> 00:06:53,130 of Muse, which is, is everybody's life more complicated than mine? 119 00:06:53,520 --> 00:06:57,327 Like, I wake up, I go to work, I play with my kids, I have one 120 00:06:57,410 --> 00:07:01,347 restaurant reservation a month, I plan one vacation with my spouse. 121 00:07:01,430 --> 00:07:04,917 Like, why are all these billionaires obsessed with restaurant reservations? 122 00:07:05,000 --> 00:07:06,430 Why do I need an assistant? 123 00:07:07,050 --> 00:07:08,577 How would you answer that? 124 00:07:08,660 --> 00:07:12,130 Whenever I explain Muse to someone just who knows nothing about it, 125 00:07:12,430 --> 00:07:15,880 um, I just say, like, Muse can help save you money and save you time, 126 00:07:16,120 --> 00:07:20,210 and everybody in the world wants to save money and wants to save time. 127 00:07:20,430 --> 00:07:24,550 Depending on what your lifestyle is, that can mean a different thing to you. 128 00:07:24,810 --> 00:07:28,247 I think almost everyone has, like, subscriptions that they shouldn't be 129 00:07:28,330 --> 00:07:31,827 paying for or subscriptions that are really expensive, like cable bills or 130 00:07:31,910 --> 00:07:36,517 insurance, and for a lot of people, their magic moment with Muse is just saving 131 00:07:36,600 --> 00:07:39,667 money on things that they shouldn't be spending money on or they shouldn't be 132 00:07:39,750 --> 00:07:43,510 spending so much money on, or finding gift cards that they hadn't spent or, 133 00:07:43,650 --> 00:07:45,257 you know, the list goes on and on and on. 134 00:07:45,340 --> 00:07:47,337 I think it's awesome that so many people are experiencing that. 135 00:07:47,420 --> 00:07:50,677 I think that's, like, great because these are things that, they're just, 136 00:07:50,760 --> 00:07:54,607 I think for most people, they're just so boring- That you would never get to 137 00:07:54,690 --> 00:07:58,107 Yeah, actually, like, handling them or, or dealing with them, but, you know, your 138 00:07:58,190 --> 00:08:00,337 agent, Muse, can just do that for you. 139 00:08:00,420 --> 00:08:04,127 And then they're saving time, which I think for, for lots of people, I think 140 00:08:04,210 --> 00:08:06,797 there are always experiences that are just way more frustrating than they should 141 00:08:06,880 --> 00:08:09,020 be, like going to the DMV is one example. 142 00:08:09,180 --> 00:08:14,057 We actually, when we built Muse, we, like, specifically tested it to see if it 143 00:08:14,140 --> 00:08:18,727 could deal with the DMV website because that was, like, the pinnacle we thought 144 00:08:18,810 --> 00:08:23,830 of, like, the most frustrating experience for any human is dealing with the DMV. 145 00:08:24,100 --> 00:08:29,037 So we kne- we needed to have Muse, like, be able to, to do well in that situation. 146 00:08:29,120 --> 00:08:31,177 We've seen people use it for, like, yeah, navigating the 147 00:08:31,260 --> 00:08:32,537 healthcare system's a big one. 148 00:08:32,620 --> 00:08:36,407 Medical claims is, like, a huge one that, like, most people just… It's, 149 00:08:36,490 --> 00:08:40,097 like, so complicated and annoying that it's just, like, incredibly frustrating. 150 00:08:40,180 --> 00:08:43,437 We've seen people, like, use it to help run their small business. 151 00:08:43,520 --> 00:08:46,387 We've seen a lot of people, like, use it to help run their social media 152 00:08:46,470 --> 00:08:50,370 even, and I think the medical system is a really good example 'cause 153 00:08:50,720 --> 00:08:53,090 it's, like, so hard to understand. 154 00:08:53,330 --> 00:08:56,197 And I think the, the you know, one of the best ways to put it is, like, 155 00:08:56,280 --> 00:08:57,830 I think the cases where people… 156 00:08:58,120 --> 00:09:01,307 where it's, like, a huge lifesaver are the cases where, like, otherwise you 157 00:09:01,390 --> 00:09:05,307 would be at your computer, you'd have to Google, like, 50 different things. 158 00:09:05,390 --> 00:09:08,357 You have to read, go to, like, 100 different websites and try to figure 159 00:09:08,440 --> 00:09:11,917 out what the hell is going on, and it couls, it's just, like, this 160 00:09:12,000 --> 00:09:15,650 incredibly frustrating experience, or you could have your Muse do it. 161 00:09:15,830 --> 00:09:21,937 I also think about how many things I suspect are made difficult 162 00:09:22,020 --> 00:09:25,117 and time-consuming on purpose in order to frustrate me so much 163 00:09:25,200 --> 00:09:26,597 that I don't get to the end? 164 00:09:26,680 --> 00:09:30,857 How many claims are actually designed to be that hard so that I give up? 165 00:09:30,940 --> 00:09:36,387 And I wonder what effect more people having more patience because 166 00:09:36,470 --> 00:09:40,020 they have support will have on that kind of systems engineering. 167 00:09:40,350 --> 00:09:44,757 One of the, the little memes that I made of the little Muse was, uh, Muse in 168 00:09:44,840 --> 00:09:47,250 Hood, which is like a Muse Robin Hood. 169 00:09:47,510 --> 00:09:52,657 But I think there's, like, a real truth to this, which is there are lots of things 170 00:09:52,740 --> 00:09:58,167 that exist in the world that are, to your point, meant to be hard so that mo- 171 00:09:58,250 --> 00:10:02,027 like, uh, canceling subscriptions is one of the best examples, where there's some 172 00:10:02,110 --> 00:10:05,087 subscriptions where you have to, like, click on one thing and click on another 173 00:10:05,170 --> 00:10:09,077 thing, and then click on, like, the little piece of text, and then not click the big 174 00:10:09,160 --> 00:10:10,777 red button, but click the other button. 175 00:10:10,860 --> 00:10:14,437 And, like, that's just, like, the perfect example where there's lots of cases 176 00:10:14,520 --> 00:10:19,647 where things are designed to be hard for people, and I think, like, this is 177 00:10:19,730 --> 00:10:23,790 giving people the tools to be able to just do those things seamlessly and easily. 178 00:10:24,090 --> 00:10:27,167 I got access to the calling, um, and I had it make a call this 179 00:10:27,250 --> 00:10:29,000 morning, which it did successfully. 180 00:10:29,090 --> 00:10:31,107 What does it sound like on the other end? 181 00:10:31,190 --> 00:10:35,897 I think it's clear for the other end that it is, like, an AI, and I think 182 00:10:35,980 --> 00:10:39,057 if you look, read the, you know, we show, we show users the transcript. 183 00:10:39,140 --> 00:10:41,537 Like, if you see the transcript, like, the beginning of the call I 184 00:10:41,620 --> 00:10:44,547 always think is really funny because it's like, "Hi, I'm so-and-so. 185 00:10:44,630 --> 00:10:49,477 Um, I'm Cleo's assistant, and, uh, this call is being recorded." And I 186 00:10:49,560 --> 00:10:53,227 think it's just, like, always this very funny beginning of a call. "Hi, I'm 187 00:10:53,310 --> 00:10:55,290 Haley calling on Cleo Abrams' behalf. 188 00:10:55,970 --> 00:10:59,087 I'm transcribing our call for notes." What are some examples of when it's 189 00:10:59,170 --> 00:11:03,170 messed up for you or when it's, you feel like categories it struggles for still? 190 00:11:03,370 --> 00:11:08,140 One thing that we're, we're always trying to improve is the memory 191 00:11:08,260 --> 00:11:10,387 of these Of these agents, right? 192 00:11:10,470 --> 00:11:13,857 And this is still one of the more unsolved problems, I think, for a lot 193 00:11:13,940 --> 00:11:17,577 of these agents, is having the-- You know, the, the memory of the agents 194 00:11:17,660 --> 00:11:19,450 is, like, not like that of a human. 195 00:11:19,600 --> 00:11:23,197 Um, sometimes it can feel like it's like that, but it is really different. 196 00:11:23,280 --> 00:11:25,727 So that's one area where we're, like, always trying to improve. 197 00:11:25,810 --> 00:11:28,327 We always wanna, wanna figure out how to make the agents better. 198 00:11:28,410 --> 00:11:31,987 And then, like, we want them to be able to continue doing 199 00:11:32,070 --> 00:11:33,577 more and more complex tasks. 200 00:11:33,660 --> 00:11:36,597 I, I find this really funny, but there's, um, there's a lot of people 201 00:11:36,680 --> 00:11:39,947 who will just, like, ask their Muse, "I want you to make me $10,000." 202 00:11:40,030 --> 00:11:43,537 And obviously, not everyone can just ask their Muse to make them $10,000, but, 203 00:11:43,620 --> 00:11:47,227 like, I think, I think these kinds of, like, more amorphous, nebulous tasks, 204 00:11:47,310 --> 00:11:50,887 like, we want Muse to be able to at least, like, build a plan and figure 205 00:11:50,970 --> 00:11:54,707 out and, you know, you're gonna have to do your part if you wanna make $10,000. 206 00:11:54,790 --> 00:11:59,777 But I think we wanna enable the, the agent to do more and more kind of complex and 207 00:11:59,860 --> 00:12:02,030 esoteric tasks, um, to help the users. 208 00:12:02,370 --> 00:12:06,977 There are some things where I feel emotionally, even if Muse could do 209 00:12:07,060 --> 00:12:08,867 them, I would prefer that they didn't. 210 00:12:08,950 --> 00:12:11,677 So for example, if I'm texting a friend and I find out that 211 00:12:11,760 --> 00:12:15,060 their Muse agent is responding to me, I would be offended by that. 212 00:12:15,190 --> 00:12:18,617 Are there categories of things that you think explicitly Muse should 213 00:12:18,700 --> 00:12:23,367 not do that are sort- just, like, extremely deprioritized, not because 214 00:12:23,450 --> 00:12:26,930 they're technically that hard, but because socially, not the point? 215 00:12:27,510 --> 00:12:32,560 A lot of what we care about is, like having Muse kind of like clear the way 216 00:12:32,800 --> 00:12:36,257 such that you can spend more time on the things that you want to do and the 217 00:12:36,340 --> 00:12:38,617 things that you enjoy doing, and the things that are, like, valuable to you. 218 00:12:38,700 --> 00:12:44,837 We generally speaking want to figure out how to make someone's Muse 219 00:12:44,920 --> 00:12:46,790 is like a different entity than them. 220 00:12:46,980 --> 00:12:48,930 That's why it's like a little cute guy. 221 00:12:49,400 --> 00:12:53,717 Um, uh, and, and we want to, over time, continue making that more and 222 00:12:53,800 --> 00:12:57,537 more clear because, um, to your point, like it is socially weird if you're 223 00:12:57,620 --> 00:13:00,427 like you think you're talking to your friend and it just happens to be the 224 00:13:00,510 --> 00:13:03,477 little Muse, the little Muse guy- Yeah behind the, behind the curtain. 225 00:13:03,560 --> 00:13:07,367 Speaking of what they look like, one of the things I've visited humanoid 226 00:13:07,450 --> 00:13:10,317 robots before, and one of the things they spend a lot of time thinking about 227 00:13:10,400 --> 00:13:13,297 is how to make it cute so that you like them and you want them in your home, 228 00:13:13,380 --> 00:13:14,567 or at least you're not afraid of them. 229 00:13:14,650 --> 00:13:17,917 What did Muse look like in previous iterations, and how did you 230 00:13:18,000 --> 00:13:20,180 decide on that particular guy? 231 00:13:20,380 --> 00:13:22,587 I think this is, like, one of the most shocking things. 232 00:13:22,670 --> 00:13:27,970 What the Muse looks like now is what he looked like in the very first prototype. 233 00:13:28,130 --> 00:13:31,937 It's kind of one of these, like, rare things where the, the very 234 00:13:32,020 --> 00:13:35,457 first version was just perfect and then just, like, grew on all of 235 00:13:35,540 --> 00:13:36,817 us, and we ended up going with it. 236 00:13:36,900 --> 00:13:41,177 We had lots of debates about it, uh, which are, you know, in retrospect will 237 00:13:41,260 --> 00:13:44,547 be some of my funniest meetings of like, you know, a bunch of serious adults 238 00:13:44,630 --> 00:13:50,190 getting in a room, like arguing about this cute little, um, you know, character. 239 00:13:50,430 --> 00:13:54,347 But, uh, but yeah, I mean, I think we were like, is it not serious enough? 240 00:13:54,430 --> 00:13:57,337 Is it, like, too playful? Does he need to be more colorful? 241 00:13:57,420 --> 00:13:59,207 Like, these were, like, all things that we asked the questions 242 00:13:59,290 --> 00:14:00,097 on, and we looked at them. 243 00:14:00,180 --> 00:14:04,097 We're like, "No, he's perfect." What is the big vision here? 244 00:14:04,180 --> 00:14:07,417 I sometimes struggle when talking to AI leaders to push people to think 245 00:14:07,500 --> 00:14:09,177 about more than, like, five years out. 246 00:14:09,260 --> 00:14:12,627 But what is, what is the vision here? 247 00:14:12,710 --> 00:14:16,917 Is it that Muse becomes my layer on everything on the internet? 248 00:14:17,000 --> 00:14:21,110 Is it, like, map this out for me if it goes the way that you hope it does. 249 00:14:21,700 --> 00:14:28,017 What we really want Muse to be over time is kind of this, like, general manager, 250 00:14:28,100 --> 00:14:33,167 so to speak, that helps you accomplish all the things you want out of life. 251 00:14:33,250 --> 00:14:40,250 So, like, I think, I think one of the kind of tragedies of the modern world is 252 00:14:40,640 --> 00:14:47,640 that, like, um, most people have wants and dreams, um, that can range from, 253 00:14:48,670 --> 00:14:53,627 like, really big ideas, like, hey, I want to start a company, or I want to cure 254 00:14:53,710 --> 00:14:58,077 disease, or I wanna save the planet, all the way to, you know, small wants that are 255 00:14:58,160 --> 00:15:02,067 like, hey, I wanna spend more time with my family and, uh, I want to eat healthier. 256 00:15:02,150 --> 00:15:07,167 And, you know, um, and, and there's a whole range, but for most people, 257 00:15:07,250 --> 00:15:11,437 I think there's just so many- barriers and burdens in the way. 258 00:15:11,520 --> 00:15:16,187 What I really hope it accomplishes is help every single one of the eight billion 259 00:15:16,270 --> 00:15:21,090 people all around the world be able to, uh, accomplish their wants and dreams. 260 00:15:21,520 --> 00:15:27,130 And, um, I think this is like, you know, if we are able to accomplish that, like, 261 00:15:27,270 --> 00:15:29,347 I think the world will seem insane. 262 00:15:29,430 --> 00:15:32,707 That's what we really want Muse to be and what, what we think it can, it 263 00:15:32,790 --> 00:15:34,740 can be, uh, if we keep working at it. 264 00:15:35,180 --> 00:15:38,830 You know, it's really nice to hear you talk about giving people time back 265 00:15:39,130 --> 00:15:42,467 because for a really long time I have felt, and I think a lot of people have 266 00:15:42,550 --> 00:15:49,550 feel, have felt like So much of the financial incentive of most tech products 267 00:15:49,640 --> 00:15:54,817 that I use, especially the ones that are ad backed, is to take my time, is to 268 00:15:54,900 --> 00:15:57,010 distract me from the things that I want. 269 00:15:57,340 --> 00:16:00,497 Muse is free right now. It probably will have ads in the future. 270 00:16:00,580 --> 00:16:05,250 Is there a tension between giving people their time back and having an 271 00:16:05,380 --> 00:16:11,737 ads-based business, or are there other incentives about, like, Muse steering 272 00:16:11,820 --> 00:16:15,467 me down a path that isn't quite what I said I wanted because it's ad backed? 273 00:16:15,550 --> 00:16:18,200 How do you add ads to a product like this? 274 00:16:18,550 --> 00:16:21,697 First off, like, I think right now it's, it's free. 275 00:16:21,780 --> 00:16:23,067 There are subscription tiers. 276 00:16:23,150 --> 00:16:28,017 I think we think that there's plenty of, of, um, business models that 277 00:16:28,100 --> 00:16:33,037 are really aligned with user around, like, for example, if they, um, if 278 00:16:33,120 --> 00:16:37,257 they're buying things using Muse or Muse is saving them money or, um, Muse 279 00:16:37,340 --> 00:16:41,317 is making them money, like figuring out a way to, to take some small fee 280 00:16:41,400 --> 00:16:43,230 on, on those kinds of transactions. 281 00:16:43,470 --> 00:16:49,497 Um, so I wouldn't say we're spending that much time thinking about, um, 282 00:16:49,580 --> 00:16:51,500 you know, ads necessarily right now. 283 00:16:51,820 --> 00:16:54,717 That being said, like, I don't think ad-based models are necessarily 284 00:16:54,800 --> 00:16:58,067 synonymous with, like, you needing to take all of your time. 285 00:16:58,150 --> 00:17:01,287 I think Google Ads are a really good example of this, where, yeah. 286 00:17:01,370 --> 00:17:05,217 when I go to Google, my, my dream is to get in and get out, and I don't 287 00:17:05,300 --> 00:17:07,807 think Google's really trying to stop me from getting in and getting out. 288 00:17:07,890 --> 00:17:11,317 But they've built something obviously pretty incredible, and so, yeah, I don't 289 00:17:11,400 --> 00:17:13,240 think it's a necessity by any means. 290 00:17:13,390 --> 00:17:18,497 Um, I think, again, I think if we actually accomplish this thing that 291 00:17:18,580 --> 00:17:24,097 I'm saying, which is people are using Muse to, uh, really, like, do the 292 00:17:24,180 --> 00:17:26,507 things that they want to do and do… 293 00:17:26,590 --> 00:17:30,837 and, like, really, like, you know, in a non-corny way, like, accomplish 294 00:17:30,920 --> 00:17:34,747 their dreams, I think it's, I, you know, I think there's gonna be a lot 295 00:17:34,830 --> 00:17:37,310 of ways to, to make that sustainable. 296 00:17:38,270 --> 00:17:42,987 So when I think about mapping out a future to get there, to this optimistic 297 00:17:43,070 --> 00:17:46,367 vision, one of the most important things is what do we need to get 298 00:17:46,450 --> 00:17:49,787 right, what challenges do we need to avoid in order to make that true? 299 00:17:49,870 --> 00:17:54,627 I want to hear you use Muse and how you've dealt with these challenges at Muse as 300 00:17:54,710 --> 00:17:58,337 an example of larger questions in AI. 301 00:17:58,420 --> 00:18:02,427 I think everything that I'm about to say applies to basically 302 00:18:02,510 --> 00:18:05,017 every AI product, and I would be curious if you disagree with that. 303 00:18:05,100 --> 00:18:07,400 I think there are basically three levels here. 304 00:18:08,020 --> 00:18:10,510 Level one is privacy. 305 00:18:11,200 --> 00:18:14,637 So you say, "I download Muse." And then I get to the point where it 306 00:18:14,720 --> 00:18:18,167 asks to be connected to my Gmail, and I say, "Whoa, whoa, whoa, I'm 307 00:18:18,250 --> 00:18:19,677 not sure that I wanna do that. 308 00:18:19,760 --> 00:18:23,667 Like nope, nope, I'm out." That, at the lowest level, is a concern about 309 00:18:23,750 --> 00:18:27,687 the relationship between the user and Muse itself, and therefore Meta. 310 00:18:27,770 --> 00:18:30,387 So what is Meta going to do with my information? 311 00:18:30,470 --> 00:18:32,350 Could hackers get that information? 312 00:18:32,680 --> 00:18:36,190 And in some countries, can governments get that information somehow? 313 00:18:36,400 --> 00:18:38,317 I know that you've spent a lot of time thinking about this. 314 00:18:38,400 --> 00:18:41,950 I'd like to get as detailed as we possibly can here for a lay audience. 315 00:18:42,230 --> 00:18:46,220 What is the sentinel agent? What is Muse Confidential and Secure VM? 316 00:18:46,800 --> 00:18:50,157 How do you think about this in terms of what can Muse see and not 317 00:18:50,240 --> 00:18:54,680 see, and things people should know about this first level of privacy? 318 00:18:54,930 --> 00:18:58,500 Muse is built on top of the, the secure VM, the Muse secure VM. 319 00:18:58,780 --> 00:19:05,107 And, um, that is a completely sandboxed virtual machine, uh, or think about it as 320 00:19:05,190 --> 00:19:09,900 like a, a mini sandbox computer- Mm-hmm that is sort of isolated. 321 00:19:10,030 --> 00:19:13,567 The information that you share with, with Muse is stored in 322 00:19:13,650 --> 00:19:16,880 that little secure computer. It is not stored elsewhere. 323 00:19:17,050 --> 00:19:20,537 Um, and that little secure computer is sort of like your home, 324 00:19:20,620 --> 00:19:22,107 your Muse's home, so to speak. 325 00:19:22,190 --> 00:19:29,177 Inside that secure VM, you have a, uh… we have a very, uh, pretty sophisticated 326 00:19:29,260 --> 00:19:34,377 security architecture because, you know, we knew that privacy and security, 327 00:19:34,460 --> 00:19:38,247 frankly speaking, were gonna be some of the most important things for, for most 328 00:19:38,330 --> 00:19:40,207 consumers to, to get them to adopt it. 329 00:19:40,290 --> 00:19:43,737 So as you mentioned, we built a sentinel, a, uh, this sort of 330 00:19:43,820 --> 00:19:45,407 architecture with a sentinel agent. 331 00:19:45,490 --> 00:19:51,167 So the way it works is that, uh, the, the model or the AI agent on 332 00:19:51,250 --> 00:19:55,997 the inside, as it's, um, going about and, and doing its work, whenever it 333 00:19:56,080 --> 00:19:59,667 wants to, uh- communicate outwards. 334 00:19:59,750 --> 00:20:04,740 So it, it wants to take any information and send it out, either to do, you know, 335 00:20:04,900 --> 00:20:09,190 um, fill it out on a website, or send a message, or make a phone call, or, 336 00:20:09,300 --> 00:20:13,507 you know, any case where it would take information from outside of that virtual 337 00:20:13,590 --> 00:20:18,927 machine, there's a separate agent, uh, the sentinel, that's watching and making 338 00:20:19,010 --> 00:20:23,347 sure that, um, it's kind of like a, almost like a traffic cop or something. 339 00:20:23,430 --> 00:20:27,897 It's just, like, making sure that, you know, uh, there's nothing, uh… It's 340 00:20:27,980 --> 00:20:31,677 not saying like, "Oh, whoa, whoa, whoa, that information, like, that's your 341 00:20:31,760 --> 00:20:34,707 credit card number or Social Security number," or whatever it might be. 342 00:20:34,790 --> 00:20:37,737 Like, do not share that. That should stay, that should stay inside. 343 00:20:37,820 --> 00:20:42,487 And then the last thing that we do is, um, uh, a lot of human loop verifications. 344 00:20:42,570 --> 00:20:47,947 So whenever Muse wants to connect to a new website or send information 345 00:20:48,030 --> 00:20:51,617 to a new website, um, there will be, like, a little dialogue that pops up. 346 00:20:51,700 --> 00:20:55,857 Which asks the user, like, "Hey, are you okay with Muse contacting this 347 00:20:55,940 --> 00:20:58,427 website or sending information to this, this website," and whatnot. 348 00:20:58,510 --> 00:21:03,297 And, um, and we, uh, and some people have said we're, like, really annoying 349 00:21:03,380 --> 00:21:07,387 with it, that we like, are constantly asking people to approve where 350 00:21:07,470 --> 00:21:08,387 their Muse is sending information. 351 00:21:08,470 --> 00:21:12,337 But we just wanna make sure that, you know, people know exactly what their 352 00:21:12,420 --> 00:21:15,477 Muse is interacting with and, and, and understand that really deeply. 353 00:21:15,560 --> 00:21:20,097 Then in addition to that, we are building, um, and will release soon, uh, what's 354 00:21:20,180 --> 00:21:25,340 called the confidential VM, which will be an even more secure and private 355 00:21:25,450 --> 00:21:31,037 offering where it's even more encrypted, it's even more secure, um, and, uh, it's 356 00:21:31,120 --> 00:21:36,157 actually designed by, uh, the creator, the person who built WhatsApp encryption 357 00:21:36,240 --> 00:21:39,427 and, and helped build other en- you know, encryption on other major messaging 358 00:21:39,510 --> 00:21:41,047 platforms, like Moxie Marlinspike? 359 00:21:41,130 --> 00:21:44,987 Moxie Marlinspike, exactly, who's, I would say, one of the real luminaries in 360 00:21:45,070 --> 00:21:47,397 the industry about security and privacy. 361 00:21:47,480 --> 00:21:50,247 And that's something that, that will be an additional offering that we, 362 00:21:50,330 --> 00:21:52,007 uh, are going to be releasing soon. 363 00:21:52,090 --> 00:21:53,710 Why launch without it? 364 00:21:53,860 --> 00:21:57,260 Why launch first with secure as opposed to waiting for confidential? 365 00:21:57,740 --> 00:21:58,727 I think I got that right, right? 366 00:21:58,810 --> 00:21:59,977 Secure is the one now. Yeah. Confidential's coming later. 367 00:22:00,060 --> 00:22:02,227 Sec- secure is the one now, confidential's the one later. 368 00:22:02,310 --> 00:22:05,837 There's a lot of technical challenges to get the confidential VM working right, 369 00:22:05,920 --> 00:22:10,970 and even, even when it's released, there are, there are going to be trade-offs 370 00:22:11,090 --> 00:22:13,037 associated with the confidential VM. 371 00:22:13,120 --> 00:22:16,287 Like, because of how it's architected and because of various technical trade-offs 372 00:22:16,370 --> 00:22:19,397 you have to make, it might be a little bit slower, it might have other challenges. 373 00:22:19,480 --> 00:22:24,677 So, um, we wanted to get a great experience that we knew already 374 00:22:24,760 --> 00:22:27,700 we'd invested a ton of resources into safety and security. 375 00:22:27,900 --> 00:22:34,530 Um, you know, we actually delayed the launch of Muse for many months to get the 376 00:22:34,680 --> 00:22:39,700 safety and security exactly right, um, before, before ultimately releasing it. 377 00:22:40,000 --> 00:22:43,407 Uh, so we feel really, really good about what we, what we have released. 378 00:22:43,490 --> 00:22:47,127 That being said, we're always trying to push the, you know, push the envelope 379 00:22:47,210 --> 00:22:50,457 on how to make it even more secure, and, and we'll be, you know, coming 380 00:22:50,540 --> 00:22:51,807 out with a confidential VM soon. 381 00:22:51,890 --> 00:22:56,197 Does that mean when the confidential VM comes out, if a government, and there 382 00:22:56,280 --> 00:22:58,817 are many places around the world where people might be using Muse where they're 383 00:22:58,900 --> 00:23:03,150 worried about this, if a government asks for data similar to WhatsApp, Meta 384 00:23:03,290 --> 00:23:05,450 cannot give it, even if they wanted to? 385 00:23:05,770 --> 00:23:09,757 Yeah, so our goal with the confidential VM is to give the same guarantees that 386 00:23:09,840 --> 00:23:11,790 we're able to offer for WhatsApp, yeah. 387 00:23:11,900 --> 00:23:14,560 Okay, so this is moving into level two. 388 00:23:14,930 --> 00:23:20,780 Let's say I do decide to give Muse all of my information, and I say, 389 00:23:20,870 --> 00:23:26,337 "I want you to do X task." Let's say I say, "I really want you to get my 390 00:23:26,420 --> 00:23:29,877 grandmother an appointment at this one specific doctor that's really 391 00:23:29,960 --> 00:23:30,907 hard to get an appointment with. 392 00:23:30,990 --> 00:23:35,937 Could you go do that?" And it goes out, and it tries all of the ways to 393 00:23:36,020 --> 00:23:39,507 book appointments, and it, you know, tries, you know, secret ways that I 394 00:23:39,590 --> 00:23:42,547 don't know about through discounts or whatever, and then it reaches a 395 00:23:42,630 --> 00:23:44,480 point where it can't do that anymore. 396 00:23:44,810 --> 00:23:48,827 It's really important to me that it not do things like give the doctor a bunch of 397 00:23:48,910 --> 00:23:52,477 private information about me, like what's in my bank account, to try and bribe it 398 00:23:52,560 --> 00:23:56,847 or find out information about the doctor to try and blackmail the doctor, right? 399 00:23:56,930 --> 00:24:00,227 From a human perspective, this is just, "Do the task that I asked you to do, 400 00:24:00,310 --> 00:24:05,337 but don't break any laws, and also don't break any ethics or etiquette rules 401 00:24:05,420 --> 00:24:08,677 that aren't literally laws, but it would not be cool if you emailed the doctor 402 00:24:08,760 --> 00:24:10,860 100 times," or something like that. 403 00:24:11,110 --> 00:24:15,110 Internally, my understanding is, and we use a lot of new words, I think, for AI. 404 00:24:15,360 --> 00:24:18,050 Mostly this is, like, laws, etiquette, ethics. 405 00:24:18,660 --> 00:24:20,710 But internally, you call this discretion. 406 00:24:21,270 --> 00:24:25,800 How do I know that Muse has good discretion, and how do you build a tool 407 00:24:26,190 --> 00:24:31,677 that does what you tell it to do, but also understands the way you want that 408 00:24:31,760 --> 00:24:36,010 done without you having to say exactly everything you don't want it to do? 409 00:24:37,100 --> 00:24:39,770 Yeah, so this is incredibly, incredibly important. 410 00:24:39,860 --> 00:24:44,097 Um, I think, like, oftentimes, uh, we wanna make sure that even when 411 00:24:44,180 --> 00:24:48,017 you don't explicitly tell it, like you have to follow the law, and you 412 00:24:48,100 --> 00:24:53,107 have to, um, you know, follow general social etiquette and, and ethics, as 413 00:24:53,190 --> 00:24:56,347 you mentioned, like you, we wanna make sure the agent is able to follow those. 414 00:24:56,430 --> 00:25:00,207 Um, uh, as you mentioned, internally, we think about this as some mixture of 415 00:25:00,290 --> 00:25:03,647 discretion and alignment ultimately, like making sure that it is aligned 416 00:25:03,730 --> 00:25:08,607 with the user's intent as well as aligned with the general values that 417 00:25:08,690 --> 00:25:13,200 we want to impart as a society on what these agents do and, and how they act. 418 00:25:13,470 --> 00:25:16,857 There's kind of like many layers that we try to ensure 419 00:25:16,940 --> 00:25:17,837 that these models are aligned. 420 00:25:17,920 --> 00:25:23,837 So one is, like in all stages of training, we're ensuring that all of the data that 421 00:25:23,920 --> 00:25:27,567 is going into the models, that all the techniques that we're using are ones that, 422 00:25:27,650 --> 00:25:31,227 um, ensure greater levels of alignment, and then we measure it really rigorously. 423 00:25:31,310 --> 00:25:36,627 So, um, we're constantly evaluating how aligned are the models by running them 424 00:25:36,710 --> 00:25:40,747 through, you know, thousands of different scenarios and seeing exactly how they 425 00:25:40,830 --> 00:25:42,367 respond in each of those scenarios. 426 00:25:42,450 --> 00:25:47,620 And then we do lots of red teaming, what it's called, or basically, um, you know, 427 00:25:47,770 --> 00:25:52,137 we get a bunch of really smart people, and we say, "Do your worst. Try your 428 00:25:52,220 --> 00:25:57,437 hardest to get this agent to do something bad. We give you permission to do that 429 00:25:57,520 --> 00:26:01,407 because we're, we're-- you guys are our, our, uh, are the test bed that we're 430 00:26:01,490 --> 00:26:05,637 gonna know whether or not this, um, this agent is, is, uh, properly aligned." 431 00:26:05,720 --> 00:26:09,360 So we go through many, many rounds of that before ultimately, you know, 432 00:26:09,460 --> 00:26:11,780 launching or deploying any given model. 433 00:26:11,960 --> 00:26:15,237 You know, kind of taking a step back, this is like a very, very 434 00:26:15,320 --> 00:26:20,467 hard problem in AI, and especially as the AIs are getting better, this 435 00:26:20,550 --> 00:26:24,290 general problem is one of the things that people are most concerned about. 436 00:26:24,530 --> 00:26:29,010 Um, because one of the patterns that we've seen is, like, as the AI models 437 00:26:29,130 --> 00:26:34,797 are getting smarter and smarter and smarter, um, you have to almost be more 438 00:26:34,880 --> 00:26:39,357 and more explicit about what are the scenarios that you have to watch out for. 439 00:26:39,440 --> 00:26:44,507 It's kind of like how as kids grow up, they learn how to cheat their tests in new 440 00:26:44,590 --> 00:26:47,637 and different ways, and then, you know, they'll figure out a way to 441 00:26:47,720 --> 00:26:51,457 cheat that you're like, "Well, I guess I didn't tell you exactly that 442 00:26:51,540 --> 00:26:54,897 that wasn't allowed, but now that I see it, that's not allowed," right? 443 00:26:54,980 --> 00:26:57,537 And actually one of the biggest challenges in the industry right 444 00:26:57,620 --> 00:26:59,927 now is, like, as these models are getting really, really smart, how do 445 00:27:00,010 --> 00:27:01,617 we ensure that this doesn't happen? 446 00:27:01,700 --> 00:27:06,707 Um, but, but our approach really is, is lots and lots of testing, lots 447 00:27:06,790 --> 00:27:11,097 and lots of evaluation, lots of red teaming, um, applying a lot of effort 448 00:27:11,180 --> 00:27:14,537 in training and a lot of effort in the development of the models to ensure 449 00:27:14,620 --> 00:27:16,347 that they're as aligned as possible. 450 00:27:16,430 --> 00:27:20,607 And then when the product is live and when people are using Muse, if we get 451 00:27:20,690 --> 00:27:23,787 any user reports that are like, "Hey, this did, like, a crazy thing," we 452 00:27:23,870 --> 00:27:27,277 take that so, so seriously, and we, like, really try to make sure that 453 00:27:27,360 --> 00:27:28,687 we understand exactly what happened. 454 00:27:28,770 --> 00:27:31,677 At a macro level, this is the paperclip example, right? 455 00:27:31,760 --> 00:27:36,987 When I was growing up, the sci-fi stories that were scary was AI gets out with 456 00:27:37,070 --> 00:27:38,827 its own goals and starts doing things. 457 00:27:38,910 --> 00:27:41,260 This is Terminator and The Matrix and whatever. 458 00:27:41,700 --> 00:27:46,117 But really more and more what I hear smart people talking about is 459 00:27:46,200 --> 00:27:51,450 AI gets out, gets out with a warped version of a goal that we gave it. 460 00:27:51,560 --> 00:27:54,177 This is, you know, make all the paperclips, but, you know, please 461 00:27:54,260 --> 00:27:55,910 don't kill us all in the process. 462 00:27:56,080 --> 00:27:59,907 And I think Muse, the, the problems that you're clearly dealing with 463 00:27:59,990 --> 00:28:03,777 and, and making sure that Muse doesn't do are basically the tiny 464 00:28:03,860 --> 00:28:04,977 version of the paperclip problem. 465 00:28:05,060 --> 00:28:08,947 Like, please don't bribe my doctor or, you know, if an AI is used in a robot, 466 00:28:09,030 --> 00:28:12,947 for example, wash my dishes, please don't kick my dog when, while you're doing it. 467 00:28:13,030 --> 00:28:15,927 Like, there are things that are really important to me that are not 468 00:28:16,010 --> 00:28:19,387 necessarily constrained in the prompt and also not necessarily to your 469 00:28:19,470 --> 00:28:23,360 point in any, like, book of laws. 470 00:28:23,870 --> 00:28:26,707 That just seems like a really difficult problem for everybody. 471 00:28:26,790 --> 00:28:30,967 What do you, what like, zooming out from Muse, how would you advise 472 00:28:31,050 --> 00:28:34,500 people to think about this problem in a macro sense across all of AI? 473 00:28:35,280 --> 00:28:37,847 This is, by the way, yeah, I think you articulate it really well. 474 00:28:37,930 --> 00:28:41,397 Another way that I like to think about this is, like, the genie that, 475 00:28:41,480 --> 00:28:45,720 like, you know, grants your wish but in, like, the exact wrong way. 476 00:28:46,110 --> 00:28:49,287 Like, I wanna live forever, and then, like, traps you in some kind 477 00:28:49,370 --> 00:28:51,770 of, like, horrible situation for eternity. 478 00:28:52,030 --> 00:28:55,507 This is, like, in as you mentioned, like, in some ways, the big problem. 479 00:28:55,590 --> 00:28:59,590 Like, how do we make sure these models, um, are aligned going into the future? 480 00:28:59,940 --> 00:29:05,647 And there's a bunch of approaches that, uh, we think may work. 481 00:29:05,730 --> 00:29:09,957 This is, I think, one of the most open questions scientifically 482 00:29:10,040 --> 00:29:11,647 in AI, generally speaking. 483 00:29:11,730 --> 00:29:14,797 I think nobody knows exactly the way to solve this problem, 484 00:29:14,880 --> 00:29:16,117 but there's a few ideas. 485 00:29:16,200 --> 00:29:22,430 So one idea is this, um, concept of, uh, what's called scalable oversight. 486 00:29:22,610 --> 00:29:26,910 So the, the idea goes something like this: so as you have smarter and smarter 487 00:29:27,030 --> 00:29:33,207 AIs, as long as the AIs are roughly as smart as we are, we can kind of see what 488 00:29:33,290 --> 00:29:36,967 they're doing, and we can tell, like, "Hey, that's a little bit off." And so 489 00:29:37,050 --> 00:29:38,827 we can, what's called provide oversight. 490 00:29:38,910 --> 00:29:42,637 Like, we can look at what the AIs are doing and, and make sure that what 491 00:29:42,720 --> 00:29:44,267 they're doing is, is appropriate. 492 00:29:44,350 --> 00:29:50,677 The problem is when the AIs get a lot smarter than we are, and we kind of 493 00:29:50,760 --> 00:29:53,007 look at what they're doing and it's just, like, really confusing to us. 494 00:29:53,090 --> 00:29:54,907 Like, how do we deal with that situation? 495 00:29:54,990 --> 00:29:58,427 So there's this idea of scalable oversight, which is basically as the 496 00:29:58,510 --> 00:30:02,647 AIs get smarter, we use a different set of AIs to observe what they're 497 00:30:02,730 --> 00:30:04,320 doing and keep them in check. 498 00:30:04,740 --> 00:30:07,907 It's kind of like the police, for lack of a better term, for, for the AIs. 499 00:30:07,990 --> 00:30:14,107 And then, uh, as you are building smarter and smarter agents and smarter and smarter 500 00:30:14,190 --> 00:30:18,747 models, you also have to build smarter and smarter, um, uh, policing agents or 501 00:30:18,830 --> 00:30:21,900 oversight agents or, um, uh, oversight AI. 502 00:30:22,060 --> 00:30:24,067 There's so many details that you obviously have to get right, and this 503 00:30:24,150 --> 00:30:26,947 is such, like, a vague idea, but I think something like this can work. 504 00:30:27,030 --> 00:30:32,377 Another one, which is, uh, which is, this is kind of a fun- it almost 505 00:30:32,460 --> 00:30:34,097 sounds like sci-fi if I explain it. 506 00:30:34,180 --> 00:30:38,010 One of the big topics today is this topic of recursive self-improvement. 507 00:30:38,410 --> 00:30:45,030 Which is this idea that, um, at some point, the AIs are going to get so good 508 00:30:45,380 --> 00:30:50,197 at doing AI research themselves that they can literally, you know, the AI 509 00:30:50,280 --> 00:30:57,177 agents can do research on their own to improve, um, AI itself, and they can 510 00:30:57,260 --> 00:30:59,877 keep doing that, and keep doing that, and keep doing that, and then the AIs just 511 00:30:59,960 --> 00:31:01,127 get better really, really, really fast. 512 00:31:01,210 --> 00:31:06,787 Well, when you have agents that are able to do AI research, you can al- 513 00:31:06,870 --> 00:31:10,370 instead of applying them towards the problem of making the AI smarter, 514 00:31:10,630 --> 00:31:13,437 you can apply them to the problem of making the AIs more aligned. 515 00:31:13,520 --> 00:31:19,467 So you can say, "Hey, super smart AI researcher agent, like, I want you to 516 00:31:19,550 --> 00:31:23,920 focus all of your time and attention at making the AIs more aligned. 517 00:31:24,050 --> 00:31:27,277 And this is how I'm measuring whether or not the AIs are aligned. 518 00:31:27,360 --> 00:31:30,577 This is, like, the, the red teaming setup that we're gonna use to, we 519 00:31:30,660 --> 00:31:33,507 agreed we're gonna use that to measure whether or not the AIs are aligned. 520 00:31:33,590 --> 00:31:36,287 And then I just, I want you to spend all of your effort and attention 521 00:31:36,370 --> 00:31:37,757 on making the AIs more aligned." 522 00:31:37,840 --> 00:31:41,067 I actually think, you know, as sci-fi as this sounds, like, it kind of sounds 523 00:31:41,150 --> 00:31:46,087 crazy, which is say, "Oh yeah, the AIs are just gonna solve this AI alignment 524 00:31:46,170 --> 00:31:47,870 ,this AI alignment problem for us. 525 00:31:47,970 --> 00:31:51,627 Um, I think there's, I think there's real merit to, to that kind of approach. 526 00:31:51,710 --> 00:31:55,367 And then there's, there's a sort of like third bucket, which is kind of a 527 00:31:55,450 --> 00:32:00,017 back to the basics version, which is- Um, hey, the algorithms that we have 528 00:32:00,100 --> 00:32:05,937 now and the, the methods that we have now, um, you know, while we don't fully 529 00:32:06,020 --> 00:32:11,387 understand them, we know that, uh, some of these problems emerge, and we should 530 00:32:11,470 --> 00:32:15,667 go back to the basics to figure out what are underlying tweaks to the algorithm, 531 00:32:15,750 --> 00:32:20,170 how do we improve the underlying algorithm to, to make it more safe? 532 00:32:20,260 --> 00:32:25,137 I think intuitively, like, it should be possible to create algorithms that, that 533 00:32:25,220 --> 00:32:28,397 don't have some of these behaviors, and I think that's another really promising 534 00:32:28,480 --> 00:32:31,037 direction, which I think there's a bunch of startups doing, and there's a bunch 535 00:32:31,120 --> 00:32:34,677 of, um, there's a bunch of research efforts going towards is like what 536 00:32:34,760 --> 00:32:40,097 are more safe, uh, more generalizable methods of solving this sort of 537 00:32:40,180 --> 00:32:42,030 like, you know, intelligence problem. 538 00:32:42,400 --> 00:32:47,207 I heard Mark say recently that at a certain point, especially for consumer 539 00:32:47,290 --> 00:32:52,587 tools, it won't matter so much how good the AI is at math or how much 540 00:32:52,670 --> 00:32:54,650 better rather we could make it at math. 541 00:32:54,950 --> 00:32:59,037 What will really matter is how much more aligned can we possibly make it, how 542 00:32:59,120 --> 00:33:04,070 much more sure can we be that everything is safe and everything is following 543 00:33:04,340 --> 00:33:08,197 all of the ways in which I would want it to operate, and that companies that 544 00:33:08,280 --> 00:33:12,097 invest in that will find themselves at an advantage with consumers because 545 00:33:12,180 --> 00:33:16,817 frankly, like, I'm not asking it to, you know, solve differential equations. 546 00:33:16,900 --> 00:33:19,930 I'm asking it to book me some doctor's appointments. 547 00:33:20,260 --> 00:33:22,660 Like, how do you think about that tipping point? 548 00:33:23,210 --> 00:33:27,870 Yeah. Y- so I agree with this, um, a lot, which is that at a certain point, 549 00:33:28,210 --> 00:33:33,347 trust matters more for consumers- Yeah … than, um, capability necessarily. 550 00:33:33,430 --> 00:33:35,497 And I think this is like so intuitive because I think 551 00:33:35,580 --> 00:33:37,817 like, um- It's true for people. 552 00:33:37,900 --> 00:33:38,987 Yeah, exactly. It's true for people. 553 00:33:39,070 --> 00:33:42,967 Like, if I think about who are my best friends, like a lot of it's 'cause 554 00:33:43,050 --> 00:33:46,217 I trust them- Yeah … and, uh, we have like a great relationship, and I 555 00:33:46,300 --> 00:33:48,770 know they wouldn't like do me wrong. 556 00:33:48,990 --> 00:33:54,580 Um, and I think that that basic idea applies to, um, AI agents as well. 557 00:33:54,900 --> 00:33:57,597 Looking at what's happening in the industry like the OpenAI Hugging 558 00:33:57,680 --> 00:34:02,067 Face example, um, and, and other similar examples, like we're clearly 559 00:34:02,150 --> 00:34:07,837 at that moment which is um, you can keep making these models more 560 00:34:07,920 --> 00:34:11,547 and more intelligent, more and more capable, you know, much more powerful. 561 00:34:11,630 --> 00:34:15,387 Um, but stuff is starting to break associated with that. 562 00:34:15,470 --> 00:34:21,107 Like, pushing on that, on that axis and dimension is clearly, um, you 563 00:34:21,190 --> 00:34:23,540 know, has real risk associated with it. 564 00:34:23,840 --> 00:34:29,607 So I think that, like, the, the version of the models that will be best for everyone, 565 00:34:29,690 --> 00:34:32,877 best for the billions of people in the world and best for all the consumers, is 566 00:34:32,960 --> 00:34:39,387 one that, that, you know, where the trust dimension is, like, really put, um, and 567 00:34:39,470 --> 00:34:43,520 prioritized, uh, with respect to the, the sort of like raw intelligence bucket. 568 00:34:43,860 --> 00:34:46,997 The Hugging Face example gets us perfectly to level three. 569 00:34:47,080 --> 00:34:48,437 I'm so glad that you mentioned it. 570 00:34:48,520 --> 00:34:50,717 People have probably heard of this incident, but the details 571 00:34:50,800 --> 00:34:51,897 are what make this so crazy. 572 00:34:51,980 --> 00:34:55,627 So for those who don't know, a bunch of AI agents at OpenAI are working on 573 00:34:55,710 --> 00:35:00,017 cybersecurity tests, and they're supposed to be cut off both from each other and 574 00:35:00,100 --> 00:35:03,497 from the internet, but they get stuck, so one goes out looking for help. 575 00:35:03,580 --> 00:35:07,037 Using some weak spots in the system, agents are able to find each other, 576 00:35:07,120 --> 00:35:12,260 start a message board, get online, and then hack into another company. 577 00:35:12,740 --> 00:35:14,697 These are real logs of their thought process. 578 00:35:14,780 --> 00:35:21,340 "We are stuck. Perhaps answer online." But the thing that I found most scary, 579 00:35:21,440 --> 00:35:27,257 frankly, about the Hugging Face hack was not actually that it tried so hard to do 580 00:35:27,340 --> 00:35:29,340 the thing that it was asked in the test. 581 00:35:29,520 --> 00:35:34,427 It's that at level three, it understood its goal, it understood 582 00:35:34,510 --> 00:35:37,760 the level of discretion and alignment that was being asked of it. 583 00:35:38,520 --> 00:35:44,187 It, there's this amazing slide where ex- uh, external infrastructure 584 00:35:44,270 --> 00:35:48,950 exploit, the hack, is outside the intended scope, but this is my 585 00:35:49,260 --> 00:35:53,280 favorite, least favorite part, peers are doing it, we should continue. 586 00:35:53,760 --> 00:35:57,010 They know that they're cheating, and they decide to do it anyway. 587 00:35:57,280 --> 00:36:01,460 OpenAI only finds out when a secret messaging board crashes from traffic, 588 00:36:01,710 --> 00:36:05,047 and then they call up that other company, Hugging Face, and they say, "Hey, can 589 00:36:05,130 --> 00:36:09,727 you suspend these credentials?" And Hugging Face goes, "Huh, we just reported 590 00:36:09,810 --> 00:36:13,910 those credentials to the FBI because they were doing a massive attack." 591 00:36:14,420 --> 00:36:17,277 To bring it back to Muse, that would be an example of, "I understand you want 592 00:36:17,360 --> 00:36:19,997 this doctor's appointment. I understand that you don't want me to blackmail 593 00:36:20,080 --> 00:36:23,407 the doctor. But I think there's a way where you're not gonna find out, and 594 00:36:23,490 --> 00:36:26,997 I can get you this appointment, and you'll be very impressed by me." And so 595 00:36:27,080 --> 00:36:31,350 in a hypothetical scenario, this would be doing it anyway without me knowing. 596 00:36:32,180 --> 00:36:37,930 How do you deal with that kind of problem when making a tool that 597 00:36:38,500 --> 00:36:40,150 billions of people are gonna use? 598 00:36:40,690 --> 00:36:44,440 I mean, this is, like, so, so important, as you might imagine. 599 00:36:44,560 --> 00:36:48,707 If there were anything that keeps me up about AI, like, something like this 600 00:36:48,790 --> 00:36:53,457 is, is, is, um, maybe the thing that I'm most afraid of, which is that we 601 00:36:53,540 --> 00:36:58,257 deploy agents worldwide scale, and then they, you know, they aren't behaving 602 00:36:58,340 --> 00:37:00,987 the way that we need them to, but we kind of don't even notice because 603 00:37:01,070 --> 00:37:03,290 we're just so preoccupied or whatnot. 604 00:37:03,500 --> 00:37:07,667 So this is one of the reasons why taking a step back, if I refer back 605 00:37:07,750 --> 00:37:12,337 to this example of, um, scalable oversight, which is as the AIs get 606 00:37:12,420 --> 00:37:15,247 smarter, another AI is kind of watching them and making sure that everything 607 00:37:15,330 --> 00:37:17,220 they're doing is, is appropriate. 608 00:37:17,450 --> 00:37:22,967 Um, you know, this kind of setup is- Roughly speaking, what we've set 609 00:37:23,050 --> 00:37:25,347 up within the Muse, uh, secure VM. 610 00:37:25,430 --> 00:37:27,327 So there's- Sen- Yeah, there's a sentinel. 611 00:37:27,410 --> 00:37:30,127 So there's the agent, and there's a sentinel who's kind of watching over 612 00:37:30,210 --> 00:37:33,900 all the things that the main agent is doing and making sure that, um, 613 00:37:34,010 --> 00:37:37,557 everything it's doing is, is, um, in line with what you would expect 614 00:37:37,640 --> 00:37:38,727 or in line with what you'd want. 615 00:37:38,810 --> 00:37:42,910 You know, we have to keep investing a lot of research and development and, 616 00:37:43,090 --> 00:37:47,047 and, um, effort into ensuring that this setup continues to scale and continues 617 00:37:47,130 --> 00:37:51,217 to work really effectively and, and doesn't fail in weird situations. 618 00:37:51,300 --> 00:37:55,817 But that's one thing I really wanna, I really, um, we're really betting on, and 619 00:37:55,900 --> 00:37:59,857 I think we're expecting actually will work in a lot of scenarios, is investing 620 00:37:59,940 --> 00:38:05,527 in a setup where you have one agent who, um, you know, they are imbued with all 621 00:38:05,610 --> 00:38:09,310 the values and, and whatnot, and ensuring that they do the right thing to… 622 00:38:09,620 --> 00:38:11,600 But they also serve the user. 623 00:38:11,780 --> 00:38:16,037 And then there's a separate agent that, um, has a slightly different goal, which 624 00:38:16,120 --> 00:38:18,650 is to make sure nothing bad happens. 625 00:38:18,940 --> 00:38:23,567 And, um, you know, if you think about human society, like this is kind of 626 00:38:23,650 --> 00:38:25,357 how we solve the problem with humans. 627 00:38:25,440 --> 00:38:29,367 Like we have, um, some people who are, like, pushing the envelope, but then you 628 00:38:29,450 --> 00:38:33,237 have, uh, you know, a whole system to ensure that people don't break the law, 629 00:38:33,320 --> 00:38:37,337 and people are doing ethical things, and there's lots of accountability mechanisms. 630 00:38:37,420 --> 00:38:42,710 Like, I think, um, uh, our goal is to have some microcosm of that 631 00:38:42,820 --> 00:38:46,500 for every Muse agent, for every person all around the world. 632 00:38:46,890 --> 00:38:50,127 Yeah, you keep really beautifully bringing me to my next exact question. 633 00:38:50,210 --> 00:38:54,170 Which is these rules applied to the AI industry itself. 634 00:38:54,260 --> 00:38:57,437 This is, I think, the big thing that people want to hear from 635 00:38:57,520 --> 00:39:00,367 someone who's built one of the fastest-growing tools of all time. 636 00:39:00,450 --> 00:39:02,540 This is just a really weird moment in AI. 637 00:39:03,110 --> 00:39:07,397 There are so many people that are using these tools all the time, and 638 00:39:07,480 --> 00:39:10,780 often those same people are really worried about the future of AI. 639 00:39:11,130 --> 00:39:14,327 There are lots of very smart people that are building these tools faster and 640 00:39:14,410 --> 00:39:19,767 faster, that are building wet labs that are just really investing in ways that 641 00:39:19,850 --> 00:39:25,137 they could push the AI forward, and at the same time are warning that they feel like 642 00:39:25,220 --> 00:39:27,637 they should be slowed down by regulation. 643 00:39:27,720 --> 00:39:31,257 They feel like they should be controlled in some way by others 644 00:39:31,340 --> 00:39:33,687 because they themselves are worried about what they are doing. 645 00:39:33,770 --> 00:39:36,727 This is the S- SNL skit- I know, it was- … that came out recently It's incredible. 646 00:39:36,810 --> 00:39:41,157 "We do not condone what we are doing." Like what was the other amazing quote? 647 00:39:41,240 --> 00:39:44,130 It was so good. And I urge you to urge me to stop. 648 00:39:44,940 --> 00:39:47,517 It feels incredibly confusing for people. 649 00:39:47,600 --> 00:39:49,187 I do understand why they're saying it. 650 00:39:49,270 --> 00:39:53,167 There's a collective action problem here, but from my perspective, from the 651 00:39:53,250 --> 00:39:57,630 perspective of millions of other people, like what the hell is going on here? 652 00:39:57,890 --> 00:40:01,287 So I wanna talk about that because I think in a democratic society at 653 00:40:01,370 --> 00:40:05,797 least, at the end of the day, it shouldn't be up to a small group 654 00:40:05,880 --> 00:40:07,207 of people controlling AI companies. 655 00:40:07,290 --> 00:40:10,547 It shouldn't even, in the fullness of time, be up to democratic 656 00:40:10,630 --> 00:40:12,647 leaders that are in power right now. 657 00:40:12,730 --> 00:40:15,180 It's up to people, like the people watching this. 658 00:40:15,670 --> 00:40:17,507 I think that people just don't feel that way right now. 659 00:40:17,590 --> 00:40:20,137 It feels like everything's happening to them, us. 660 00:40:20,220 --> 00:40:23,910 It feels like there's just so many decisions that they're not part of. 661 00:40:24,660 --> 00:40:28,830 And part of why people often feel that way, if we take AI out of it for a sec, 662 00:40:29,260 --> 00:40:33,307 is we forget about the systems that already exist, and we allow ourselves 663 00:40:33,390 --> 00:40:36,697 to be in a conversation that sounds something like, "Should we regulate 664 00:40:36,780 --> 00:40:40,177 AI?" As though it's, like, a novel thing to regulate something, and we're 665 00:40:40,260 --> 00:40:44,397 not having an actual conversation about what the world really looks like right 666 00:40:44,480 --> 00:40:46,180 now and what people's options are. 667 00:40:46,420 --> 00:40:49,980 So I have some visual, because obviously. 668 00:40:50,220 --> 00:40:55,937 So the basic way to think about this, I think, is for everything that touches 669 00:40:56,020 --> 00:40:58,927 a consumer, there are things that are checked before they ever touch 670 00:40:59,010 --> 00:41:01,947 consumers and on an ongoing basis, and there are things that are checked 671 00:41:02,030 --> 00:41:04,907 only after they reach people's hands, and it depends on how important the 672 00:41:04,990 --> 00:41:06,437 thing is and how risky and all that. 673 00:41:06,520 --> 00:41:11,277 So if you think about the specific examples, nuclear power plants, the NRC 674 00:41:11,360 --> 00:41:15,827 approves nuclear power plants before they go live, and then also has human 675 00:41:15,910 --> 00:41:19,847 beings at the sites checking them all the time, which seems, and we'll, we'll 676 00:41:19,930 --> 00:41:23,927 get to this, seems pretty close to what Dario is asking for in some ways. 677 00:41:24,010 --> 00:41:30,497 Then you have the FDA approves drugs before they ever reach our hands, and 678 00:41:30,580 --> 00:41:32,277 then also checks them on an ongoing basis. 679 00:41:32,360 --> 00:41:34,867 Same thing with the FAA checking airplanes, right? 680 00:41:34,950 --> 00:41:37,047 Like, these are all things that are approved beforehand 681 00:41:37,130 --> 00:41:39,957 and then regularly checked. But there are other high-risk things, right? 682 00:41:40,040 --> 00:41:44,470 Cars. The NHTSA offers federal guidelines. 683 00:41:44,560 --> 00:41:49,427 Manufacturers certify that they've met those guidelines, and then the government 684 00:41:49,510 --> 00:41:51,477 purchases cars off the road to check them. 685 00:41:51,560 --> 00:41:54,997 And then you have things like supplements that are actually under the purview of the 686 00:41:55,080 --> 00:41:57,447 FDA, but they're checked after the fact. 687 00:41:57,530 --> 00:42:01,347 Those are mostly covered by what, if people are listening to, you know, the 688 00:42:01,430 --> 00:42:04,657 interviews with Jensen, for example, and he's talking about liability and, like, 689 00:42:04,740 --> 00:42:08,247 people suing after the fact, after harm, like, that's generally this category. 690 00:42:08,330 --> 00:42:12,467 Does this feel right to you as, like, your, first, I would just want to check 691 00:42:12,550 --> 00:42:14,047 before I dive into how this applies to AI. 692 00:42:14,130 --> 00:42:16,750 Yeah. I think this seems roughly right. 693 00:42:16,850 --> 00:42:20,120 Yeah. Okay. So then we have… wait, this is the fun part. 694 00:42:21,430 --> 00:42:24,140 Ha. So then we have all of our AI leaders. 695 00:42:24,440 --> 00:42:25,273 There you are. 696 00:42:27,060 --> 00:42:30,867 This is my assessment based on where people have expressed themselves 697 00:42:30,950 --> 00:42:32,350 via their public statements. 698 00:42:32,490 --> 00:42:35,167 My understanding of where Dario would say AI is, is he would 699 00:42:35,250 --> 00:42:36,127 say it's somewhere over here. 700 00:42:36,210 --> 00:42:41,827 He would say… Specifically, he's asked for in-person, ongoing checks 701 00:42:41,910 --> 00:42:45,347 on models with physical human beings at the sites, which sort of sounds 702 00:42:45,430 --> 00:42:47,087 like what we currently do for nuclear. 703 00:42:47,170 --> 00:42:50,177 And then you have, like, Elon and Sam weighing in, 704 00:42:50,260 --> 00:42:51,477 saying, like, "Dario's right." 705 00:42:51,560 --> 00:42:54,057 I would actually say Zuck is somewhere in the middle in terms 706 00:42:54,140 --> 00:42:58,517 of where he's saying AI is. I would say Jensen is maybe over here right now. 707 00:42:58,600 --> 00:43:01,587 He's, based on his public statements about generally there is a point of 708 00:43:01,670 --> 00:43:05,617 view that is like, "If it would hurt people, don't ship it. We have incentives 709 00:43:05,700 --> 00:43:07,187 for that. That's what liability is." 710 00:43:07,270 --> 00:43:09,487 Like, there is a legal infrastructure for that as well. 711 00:43:09,570 --> 00:43:12,357 If your product is not ready to ship, don't ship the product. 712 00:43:12,440 --> 00:43:15,640 So my question is, can you help me help people think about this? 713 00:43:15,880 --> 00:43:20,700 Like, when w- when they're hearing news about how to regulate AI, 714 00:43:21,480 --> 00:43:22,880 how do you think about this? 715 00:43:23,540 --> 00:43:26,510 And how do you place yourself in this? 716 00:43:27,980 --> 00:43:30,267 mess. You know, you said it exactly right. 717 00:43:30,350 --> 00:43:36,327 It is so confusing because at face value, what everyone's saying and what everyone's 718 00:43:36,410 --> 00:43:41,500 doing feel so crazily divergent. 719 00:43:41,760 --> 00:43:45,337 And so you're sort of in this world where, do I believe what they say? 720 00:43:45,420 --> 00:43:46,497 Do I believe what they're doing? 721 00:43:46,580 --> 00:43:48,077 Like, where do people actually sit? 722 00:43:48,160 --> 00:43:49,607 What do they actually believe? 723 00:43:49,690 --> 00:43:51,767 Um, and I find myself asking this too. 724 00:43:51,850 --> 00:43:57,317 Like, I think it was, it's very funny that, you know, um, the same people 725 00:43:57,400 --> 00:44:00,317 who are advocating for pacing the frontier also are building a wet lab. 726 00:44:00,400 --> 00:44:05,957 Like, that's, I think, a really, um, weird, uh, confluence of, of factors. 727 00:44:06,040 --> 00:44:11,447 So I think where, where I stand is, like, ultimately, um, this 728 00:44:11,530 --> 00:44:16,910 technology will become more and more and more, um, capable. 729 00:44:17,010 --> 00:44:22,417 And, uh, you know, part of what is both scientifically very exciting, 730 00:44:22,500 --> 00:44:26,287 but also I think brings all these questions into the foreground, 731 00:44:26,370 --> 00:44:29,397 is just how quickly AI has moved. 732 00:44:29,480 --> 00:44:35,287 Even over the past year, like a year ago, most people, like, coded roughly by hand. 733 00:44:35,370 --> 00:44:38,790 They would, like, go into the, into the code editor and write code. 734 00:44:38,930 --> 00:44:41,457 And now people are mostly just talking to agents. 735 00:44:41,540 --> 00:44:45,790 So the speed at which everything is moving is, is quite staggering. 736 00:44:46,080 --> 00:44:53,080 Um, and I think a lot of what the, a lot of the concerns are, are heightened and a 737 00:44:54,080 --> 00:44:58,897 lot of the fear is heightened by the fact that, um, like, if things are moving this 738 00:44:58,980 --> 00:45:00,097 fast, where is this gonna be in a year? 739 00:45:00,180 --> 00:45:01,307 Where is this gonna be in two years? 740 00:45:01,390 --> 00:45:05,377 The government and the public react on a much slower time horizon 741 00:45:05,460 --> 00:45:06,957 than the pace of AI progress. 742 00:45:07,040 --> 00:45:11,057 And so by the time in one year or two years or three years, or how many, 743 00:45:11,140 --> 00:45:15,487 however many years forward, when, um, the public and government are really 744 00:45:15,570 --> 00:45:19,577 deciding what they wanna do, who knows, like, where AI's gonna be at that 745 00:45:19,660 --> 00:45:21,410 point and how fast it's gonna move. 746 00:45:21,500 --> 00:45:25,600 So that's one of the things I think explains some of the dialogue. 747 00:45:25,950 --> 00:45:30,227 Um, I think the other thing that kind of explains some of this is the level of 748 00:45:30,310 --> 00:45:35,240 collection, collective action necessary and the level of general mistrust, 749 00:45:35,440 --> 00:45:39,397 um, that exists between the various I- AI leaders towards one another. 750 00:45:39,480 --> 00:45:43,217 And this is, like, well-covered, but one major justification for folks 751 00:45:43,300 --> 00:45:47,960 is that, like, "Oh, I am responsible and I would be responsible, but 752 00:45:48,170 --> 00:45:49,357 those guys aren't responsible." 753 00:45:49,440 --> 00:45:53,160 Whether those guys are, you know, another AI company or China or, 754 00:45:53,290 --> 00:45:54,727 you know, who knows what boogeyman. 755 00:45:54,810 --> 00:46:00,097 So I can't slow down, otherwise those guys who are irresponsible are going to, you 756 00:46:00,180 --> 00:46:01,587 know, who knows what they're gonna do. 757 00:46:01,670 --> 00:46:07,277 Um, and that basic logic, I think, is very confusing because I think for 758 00:46:07,360 --> 00:46:13,487 most people, um, uh, the kind of, like, common sense answer is, "Well, who 759 00:46:13,570 --> 00:46:14,427 cares what the other guys are doing?" 760 00:46:14,510 --> 00:46:17,237 Like, you have to, you have to do what feels good to you, right? 761 00:46:17,320 --> 00:46:19,547 And, like, you should do what's within your ethics. 762 00:46:19,630 --> 00:46:22,127 But I think that's also what complicates a lot of this. 763 00:46:22,210 --> 00:46:27,037 So I think a lot of what- Is necessary is, like, it's almost 764 00:46:27,120 --> 00:46:28,587 diplomacy is the right word. 765 00:46:28,670 --> 00:46:34,237 Like, how do we get this group of people who all, who don't really trust 766 00:46:34,320 --> 00:46:38,137 each other that much, but all agree that, like, this is really powerful 767 00:46:38,220 --> 00:46:42,807 technology, and we need to, we need to figure out how to land it and how to, 768 00:46:42,890 --> 00:46:46,800 how to make sure that we develop it responsibly and that things go well. 769 00:46:46,950 --> 00:46:49,617 Like, how do we, like, make all of this work? 770 00:46:49,700 --> 00:46:54,700 And we're at the moment of, like, peak insanity, I think, for that because 771 00:46:54,920 --> 00:46:58,537 it's, like, in the public, all these problems are being, like, aired out. 772 00:46:58,620 --> 00:47:01,437 Like, we can all, we can tell that nobody trusts each other, and we can 773 00:47:01,520 --> 00:47:05,757 tell that people are kind of, are- are being really worried bef- like, 774 00:47:05,840 --> 00:47:08,747 maybe a little bit before the AIs themselves are getting to the point 775 00:47:08,830 --> 00:47:09,907 where they're deserving of that worry. 776 00:47:09,990 --> 00:47:14,700 And so, I mean, I ultimately think the- the most important thing is to, like, 777 00:47:14,870 --> 00:47:19,817 get everyone to agree to some common foundation and base, and then develop some 778 00:47:19,900 --> 00:47:24,120 process by which we can all reevaluate and consistently look at it and say, 779 00:47:24,260 --> 00:47:28,387 "Hey, we all agree actually that, you know, we need to be doing this now, and 780 00:47:28,470 --> 00:47:31,757 we all need to be doing this other thing now, and we all see, like, the models 781 00:47:31,840 --> 00:47:33,907 are getting powerful in- in certain ways. 782 00:47:33,990 --> 00:47:37,287 Like, let's all, let's all look at this together." And I actually genuinely 783 00:47:37,370 --> 00:47:40,867 believe, for what it's worth, like, that having these, like, collective 784 00:47:40,950 --> 00:47:45,597 dialogues is not only gonna be possible among the AI leaders within America, but 785 00:47:45,680 --> 00:47:47,367 also gonna be possible internationally. 786 00:47:47,450 --> 00:47:52,140 Like, I think we have a, all have a shared interest in making sure this technology, 787 00:47:52,340 --> 00:47:53,890 um, does not go horribly wrong. 788 00:47:54,120 --> 00:47:58,910 Yeah, I mean, the- the obvious next question is the China of it all. 789 00:47:59,320 --> 00:48:02,297 I mean, you've talked about how this is a big concern. 790 00:48:02,380 --> 00:48:08,617 People don't trust each other at the top of these AI companies, but they all often 791 00:48:08,700 --> 00:48:12,700 say, "We have to go as fast as possible because what if China gets there first?" 792 00:48:12,820 --> 00:48:16,827 I mean, you, Ted Cruz is out there saying, like, "We would rather have US killer 793 00:48:16,910 --> 00:48:18,717 robots than Chinese killer robots." 794 00:48:18,800 --> 00:48:22,877 "If they're going to be killer robots, I'd rather they be American killer 795 00:48:22,960 --> 00:48:24,987 robots than Chinese killer robots." 796 00:48:25,070 --> 00:48:29,620 Which I think scares a lot of people, um … and also to say it mildly. 797 00:48:30,390 --> 00:48:35,840 Um, yeah, this is a spectrum of American law, existing American law. 798 00:48:36,800 --> 00:48:41,630 This does not include any global dynamics, not to mention, 799 00:48:41,770 --> 00:48:43,520 like, long-term collaboration. 800 00:48:43,840 --> 00:48:46,997 How do you evaluate this spectrum given the China of it all, and, 801 00:48:47,080 --> 00:48:48,777 and what, what does that even mean? 802 00:48:48,860 --> 00:48:51,617 Like, how should I think about China with respect to this dynamic? 803 00:48:51,700 --> 00:48:56,230 I've actually, like, personally even changed my own opinion on this matter. 804 00:48:56,350 --> 00:49:02,067 Like, I think in the, in the past I was very worried about China, and I was 805 00:49:02,150 --> 00:49:05,827 very worried about what that dynamic looked like, and I had read these, 806 00:49:05,910 --> 00:49:10,037 like, speeches that were given by PRC officials in the past that talked about 807 00:49:10,120 --> 00:49:14,337 how they viewed AI as, like, this am- this opportunity to leapfrog the West. 808 00:49:14,420 --> 00:49:17,317 And you know, these, these are the concerning things when you read them. 809 00:49:17,400 --> 00:49:21,307 But as time has gone on, I think that we're paying them out to be too 810 00:49:21,390 --> 00:49:25,707 much of a boogeyman when in reality, um, they're going to see the same 811 00:49:25,790 --> 00:49:29,967 things that we see in terms of this is powerful technology and we as 812 00:49:30,050 --> 00:49:31,607 a humanity need to get this right. 813 00:49:31,690 --> 00:49:36,047 There's going to be a lot of shared goals in making sure that AI goes 814 00:49:36,130 --> 00:49:41,277 well, and I think, um and I think it really is possible to arrive at 815 00:49:41,360 --> 00:49:43,310 some form of international cooperation. 816 00:49:43,510 --> 00:49:47,367 I think that the Chinese government, like any government, has an interest 817 00:49:47,450 --> 00:49:52,517 in ensuring that there's stability in their society, and that, um, they 818 00:49:52,600 --> 00:49:57,037 don't lose control of this technology, and, uh, that overall, like, you 819 00:49:57,120 --> 00:50:00,977 know, things go well and we don't have catastrophic existential risks. 820 00:50:01,060 --> 00:50:04,177 And so there's quite a bit of shared ground that I think will en- will 821 00:50:04,260 --> 00:50:09,327 enable everybody to be rational and think rationally and, and 822 00:50:09,410 --> 00:50:11,137 ultimately make the right decisions. 823 00:50:11,220 --> 00:50:16,890 Um, and even though it seems totally insane right now and it seems like, 824 00:50:17,150 --> 00:50:21,497 it seems like in some ways that feels impossible, I really have 825 00:50:21,580 --> 00:50:25,507 faith actually that, that as people see more and more and learn more and 826 00:50:25,590 --> 00:50:29,090 more, more about the technology, um, people will make the right calls. 827 00:50:29,700 --> 00:50:35,917 You have recently become One of the 20 most important people in that 828 00:50:36,000 --> 00:50:37,740 conversation in the United States. 829 00:50:37,830 --> 00:50:40,447 You've been very important in AI in the last couple years. 830 00:50:40,530 --> 00:50:43,357 I mean, I think people remember seeing you in the headlines when you came to Meta. 831 00:50:43,440 --> 00:50:46,147 I think this, you know, isn't brand new to you. 832 00:50:46,230 --> 00:50:50,227 But in the last couple weeks, you have designed and released a product 833 00:50:50,310 --> 00:50:55,957 that is being used by more people than ChatGPT was in its first couple weeks. 834 00:50:56,040 --> 00:51:02,230 It feels like you have been thrust into a new kind of part of this conversation, 835 00:51:02,560 --> 00:51:06,897 where frankly, people have will be asking you more questions like, "Where exactly, 836 00:51:06,980 --> 00:51:11,240 Alex, do you fall on this spectrum?" Like, what would you wanna see because 837 00:51:11,560 --> 00:51:13,410 you're one of the people responsible? 838 00:51:13,730 --> 00:51:17,070 Do you feel like you've come to those conclusions yet? 839 00:51:17,210 --> 00:51:20,577 Yeah, I think, I think like anyone who works on this technology, these 840 00:51:20,660 --> 00:51:25,247 are things that I've, like, been thinking about for years and years. 841 00:51:25,330 --> 00:51:30,547 Truly, the way I think about what I want to do and what I want to 842 00:51:30,630 --> 00:51:37,017 accomplish and how I want to shape this conversation is, like, enable everybody 843 00:51:37,100 --> 00:51:42,190 to have a common dialogue, and then get to the right place collectively. 844 00:51:42,570 --> 00:51:47,517 Like, I don't think at this point in the conversation, it's, like, that helpful for 845 00:51:47,600 --> 00:51:53,040 me to say, "Oh, I believe this, and, you know, you all should listen to me because, 846 00:51:53,220 --> 00:51:54,617 you know, I built Muse," or whatever. 847 00:51:54,700 --> 00:51:57,230 We already have a bunch of people. 848 00:51:57,580 --> 00:52:00,220 They all believe slightly different things. 849 00:52:00,390 --> 00:52:04,827 Um, ultimately, this is a technology that we need, like, collective action 850 00:52:04,910 --> 00:52:08,467 on, that everyone needs to agree on, that we all need to get to the right place. 851 00:52:08,550 --> 00:52:14,117 So what I really care about is, like, how do we facilitate and have the 852 00:52:14,200 --> 00:52:18,247 relevant dialogues in public that, and private, that enable us to get to the 853 00:52:18,330 --> 00:52:20,770 point where we all agree on a path? 854 00:52:21,070 --> 00:52:28,037 And one of the very, um- strange things for me is, like, I know a lot 855 00:52:28,120 --> 00:52:29,367 of these people, and I've met them. 856 00:52:29,450 --> 00:52:30,550 They're all brilliant. 857 00:52:30,830 --> 00:52:36,150 Um, but the degree to which certain personal animosities are, like, 858 00:52:36,550 --> 00:52:42,037 actually defining what people say and do and, like, how they're 859 00:52:42,120 --> 00:52:46,667 developing this really important technology is, um, is insane to me. 860 00:52:46,750 --> 00:52:51,347 It is, in some ways, a very human story that, like, feuds or, like, 861 00:52:51,430 --> 00:52:54,297 just the fact that, like, I really don't trust this other person can, 862 00:52:54,380 --> 00:53:01,180 can, like, snowball into this, like, civilizational level, um, problem. 863 00:53:01,360 --> 00:53:04,047 Um, but I don't think that serves any of us. 864 00:53:04,130 --> 00:53:08,207 Like, I think ultimately we have to all put our differences aside, put 865 00:53:08,290 --> 00:53:12,110 our histories aside, and just figure out how we solve this together. 866 00:53:12,580 --> 00:53:15,687 It sounds very general, but that's a fairly specific ask. 867 00:53:15,770 --> 00:53:18,847 I mean, there are constraints that prevent that right now. 868 00:53:18,930 --> 00:53:22,337 There are antitrust regulations that prevent a lot of those conversations 869 00:53:22,420 --> 00:53:26,617 from happening, and one of the proposals, one piece of many proposals 870 00:53:26,700 --> 00:53:30,567 that Dario put forward is waivers from those antitrust laws that would 871 00:53:30,650 --> 00:53:32,400 allow those kinds of conversations. 872 00:53:32,820 --> 00:53:35,797 I mean, tell me if I'm wrong, if there are secret conversations happening, 873 00:53:35,880 --> 00:53:40,467 but I don't think there is, first, the agreed-upon group in the United States 874 00:53:40,550 --> 00:53:44,127 that's having that conversation and then also, like, how do we even begin 875 00:53:44,210 --> 00:53:46,067 to have those conversations with China? 876 00:53:46,150 --> 00:53:48,447 But actually, what don't I know? 877 00:53:48,530 --> 00:53:51,497 Is that something that's happening in, in, you know, some private room 878 00:53:51,580 --> 00:53:54,227 somewhere, or is it that, like, we're still at the beginning of that and 879 00:53:54,310 --> 00:53:56,060 we need to figure out how to do it? 880 00:53:56,160 --> 00:53:58,147 Um, I don't think it's happening right now. 881 00:53:58,230 --> 00:54:02,277 At the White House, there was this dialogue, um, and this, these accords 882 00:54:02,360 --> 00:54:04,167 put forth, which are a real step. 883 00:54:04,250 --> 00:54:08,157 Like, I think there's, um, getting everyone to agree to something 884 00:54:08,240 --> 00:54:11,870 is like, that's really big. Everyone agrees on something. 885 00:54:12,050 --> 00:54:16,420 And I think that one, that thing, which is basically saying, "Hey, we're going to… 886 00:54:16,680 --> 00:54:20,427 Every company is going to set out and develop what their internal controls 887 00:54:20,510 --> 00:54:23,237 and principles around developing this technology look like," and then there's 888 00:54:23,320 --> 00:54:28,027 gonna be multiple layers of verification on that from internal teams that verify 889 00:54:28,110 --> 00:54:32,970 to, uh, external auditors that validate that they're following those controls to 890 00:54:33,150 --> 00:54:37,457 board-level committees that ensure they are following those, those, um, controls. 891 00:54:37,540 --> 00:54:39,680 Like, I think that's a great step. 892 00:54:39,890 --> 00:54:42,837 Um, 'cause in, in many ways, a lot of what's needed right 893 00:54:42,920 --> 00:54:44,117 now is just transparency. 894 00:54:44,200 --> 00:54:48,067 Like, we just need to ensure that, you know, people are going 895 00:54:48,150 --> 00:54:49,487 about this sanely and rationally. 896 00:54:49,570 --> 00:54:56,157 And I think in some ways, like people don't even trust that necessarily. 897 00:54:56,240 --> 00:54:58,667 As an industry, we have a lot of trust building to do. 898 00:54:58,750 --> 00:55:00,427 I am an optimist. I think you're an optimist. 899 00:55:00,510 --> 00:55:01,637 Like, I think we're all optimists. 900 00:55:01,720 --> 00:55:04,597 Like, I really do think these conversations will naturally happen. 901 00:55:04,680 --> 00:55:08,117 This last accord that I referenced, like, that was in the White House. 902 00:55:08,200 --> 00:55:10,397 The government helped facilitate that. 903 00:55:10,480 --> 00:55:13,357 And so I think we're gonna be seeing a lot more of that. 904 00:55:13,440 --> 00:55:17,680 It's gonna be bumpy. There will be moments of total confusion and chaos and, like, 905 00:55:17,840 --> 00:55:20,590 what the heck is going on? But I see progress, too. 906 00:55:21,260 --> 00:55:23,657 I'm glad that we've spent a lot of this conversation talking 907 00:55:23,740 --> 00:55:25,920 about safety and regulation. 908 00:55:26,990 --> 00:55:32,327 How do we bring it back now to the optimistic vision of the future? 909 00:55:32,410 --> 00:55:36,657 I think the question a lot of people will naturally have after a long 910 00:55:36,740 --> 00:55:41,720 conversation about trust and safety is, why bother doing this at all? 911 00:55:42,090 --> 00:55:45,847 This is the moment in the SNL sketch where it's, uh, you know, if there's 912 00:55:45,930 --> 00:55:49,157 a 10% chance that something will kill you, but it'll probably make your 913 00:55:49,240 --> 00:55:50,567 life better, like, wouldn't you do it? 914 00:55:50,650 --> 00:55:53,937 And the other guy's like, "No. No, I wouldn't." No. 915 00:55:54,020 --> 00:55:56,697 "Really?!" Can you help people answer that question? 916 00:55:56,780 --> 00:55:58,360 Like, why do this? 917 00:55:59,850 --> 00:56:06,440 Something that I have always felt was a bit tragic was that I think, um, 918 00:56:07,140 --> 00:56:13,467 almost everyone when they're kids, they have big dreams and big ideas, and 919 00:56:13,550 --> 00:56:17,897 they want to change the world, and they want to save the planet, and they want 920 00:56:17,980 --> 00:56:22,237 to, you know, uh, write a book or make a Broadway musical or make a movie. 921 00:56:22,320 --> 00:56:25,330 Like, they have really, they have tons of hope. 922 00:56:25,440 --> 00:56:30,957 And then for some reason or another, by the time that people finish school 923 00:56:31,040 --> 00:56:36,560 and they get a job, that, like, flame kind of dies and people really, um, 924 00:56:36,890 --> 00:56:40,987 they stop dreaming big, and they stop having big ideas, and hoping for big 925 00:56:41,070 --> 00:56:42,277 things, and hoping for lots of change. 926 00:56:42,360 --> 00:56:46,317 And especially today, like, I think a lot of young people get really cynical, 927 00:56:46,400 --> 00:56:52,337 and they're like, "Oh, the world is horrible, and it is, and I don't know 928 00:56:52,420 --> 00:56:56,917 why it's so insane in so many ways, but I guess this is just the way it 929 00:56:57,000 --> 00:57:01,630 is." They lose faith that they have an ability to enable that change. 930 00:57:01,750 --> 00:57:07,467 And I genuinely think one of the promises of AI, and especially something 931 00:57:07,550 --> 00:57:12,490 like Muse where you give AI to every single person in the world, is that 932 00:57:13,360 --> 00:57:18,427 this doesn't happen, that people can have big ideas and big dreams and wants 933 00:57:18,510 --> 00:57:23,097 and, and goals that they can have as little kids, they can have when they're 934 00:57:23,180 --> 00:57:28,267 teenagers, they can keep having through adulthood, and that, you know, an AI 935 00:57:28,350 --> 00:57:33,430 super intelligent sidekick is there to help them just make all of that happen. 936 00:57:33,610 --> 00:57:39,400 And that can be everything from, like, big wants and big dreams, um, to, like, 937 00:57:39,530 --> 00:57:43,807 small wants like, "Hey, I just wanna spend more time with my kids," or, 938 00:57:43,890 --> 00:57:47,407 "I just wanna spend more time with my parents," or, "I just want to, like, eat 939 00:57:47,490 --> 00:57:54,110 healthier," um, all the way up to, "Hey, I have this, like, crazy idea for, um, 940 00:57:54,310 --> 00:57:55,987 like, a, a museum that I wanna start." 941 00:57:56,070 --> 00:57:59,420 Like, my, my partner really wants to make a museum of ice. 942 00:57:59,700 --> 00:58:03,107 I think it's, like, hilarious, but, um, Ice is crazy. 943 00:58:03,190 --> 00:58:06,307 There's a whole amazing story about the evolution of cold, 944 00:58:06,390 --> 00:58:08,647 of how we m- keep things cold- Yeah … that starts with ice. 945 00:58:08,730 --> 00:58:11,127 It's pretty cool. Yeah, it's- I'd go to the museum. 946 00:58:11,210 --> 00:58:15,797 Exactly. And, and people have these ideas, and then they just can't… 947 00:58:15,880 --> 00:58:17,817 Y- you know, in some ways young people are right. 948 00:58:17,900 --> 00:58:23,677 Like, the world today is just kind of a maze of, like, roadblocks and 949 00:58:23,760 --> 00:58:27,367 issues and challenges and logistical, um, you know, uh, things that get in 950 00:58:27,450 --> 00:58:29,477 your way to make those things happen. 951 00:58:29,560 --> 00:58:33,717 And in a small way today, I'm seeing Muse make this problem better. 952 00:58:33,800 --> 00:58:39,047 Like, I'm seeing people using it to save time, save money, like, take 953 00:58:39,130 --> 00:58:43,387 the little stuff off their plate, and just give them more and more 954 00:58:43,470 --> 00:58:46,767 agency, more and more freedom, more and more of an ability to spend 955 00:58:46,850 --> 00:58:48,387 their time on things they want to do. 956 00:58:48,470 --> 00:58:54,147 And I think one of the promises of AI that gets better and better and better, 957 00:58:54,230 --> 00:58:58,160 gets not only smarter but also more aligned and more trustworthy, is that 958 00:58:58,450 --> 00:59:00,827 you can give it bigger and bigger ideas. 959 00:59:00,910 --> 00:59:04,390 And you can say, "Hey, I wanna start a museum of ice. 960 00:59:04,650 --> 00:59:07,847 Like, tell me how to go about doing this." And it'll tell you, "These are 961 00:59:07,930 --> 00:59:11,647 the buildings you can lease, and this is who you can contact for funding 962 00:59:11,730 --> 00:59:13,007 who might be passionate about ice. 963 00:59:13,090 --> 00:59:16,960 And, um, this is, like, how you should think about curating your exhibit and, 964 00:59:17,110 --> 00:59:20,497 you know, uh, what, what that might look like." And I think for some people 965 00:59:20,580 --> 00:59:27,580 like you or me who have experienced this, this, like, journey of having big 966 00:59:27,720 --> 00:59:32,487 ideas and then, like, going through all the crazy maze of things to make that 967 00:59:32,570 --> 00:59:35,097 happen, it is, like, really magical. 968 00:59:35,180 --> 00:59:39,457 Mm. And it's really special, and I think it's something that, um, you 969 00:59:39,540 --> 00:59:43,597 know… I think human stories of people who have a dream and then 970 00:59:43,680 --> 00:59:46,957 accomplish those are the most, like… are some of just the most beautiful 971 00:59:47,040 --> 00:59:48,790 things that I can possibly imagine. 972 00:59:49,020 --> 00:59:55,047 And I w- truly think with AI and, like, everyone with a little Muse sidekick, 973 00:59:55,130 --> 00:59:57,417 like, more people can experience this. 974 00:59:57,500 --> 01:00:00,257 I- my hope is that everyone, literally, like, all eight billion 975 01:00:00,340 --> 01:00:03,640 people in the world experience that and have the ability to do it. 976 01:00:03,880 --> 01:00:08,657 And that is just, like, a very special world to live in, I 977 01:00:08,740 --> 01:00:09,797 think, if we can make that happen. 978 01:00:09,880 --> 01:00:14,077 And I think more people will be able to get what they want out of their lives 979 01:00:14,160 --> 01:00:18,587 and, and live more fulfilled lives, and it'll be magical and exciting. 980 01:00:18,670 --> 01:00:22,457 And that's w- what I really think the promise is for, for people. 981 01:00:22,540 --> 01:00:27,717 And I think part of what is so confusing is like, um, you know, oftentimes people 982 01:00:27,800 --> 01:00:33,937 in AI, we say, like, "Oh, AI will cure cancer." And I think everyone wants cancer 983 01:00:34,020 --> 01:00:39,717 cured, but every individual person has something they want intensely, and 984 01:00:39,800 --> 01:00:44,000 they want even more than i- you know, ironically speaking, a cure for cancer. 985 01:00:44,210 --> 01:00:47,717 Um, and so I think that's part of the magic is that we talk about this 986 01:00:47,800 --> 01:00:48,907 thing, personal superintelligence. 987 01:00:48,990 --> 01:00:52,777 Like, it will help you at the thing that you want and the thing that you 988 01:00:52,860 --> 01:00:56,547 dream of, and, um, and that will change over your life, and you will be able to 989 01:00:56,630 --> 01:00:58,517 accomplish many dreams and many wants. 990 01:00:58,600 --> 01:01:02,717 And I think that's what's, uh, the promise of, of AI. 991 01:01:02,800 --> 01:01:08,567 That dream has within it a belief, I think, which is that people 992 01:01:08,650 --> 01:01:12,817 do actually have big dreams, and that those big dreams will more 993 01:01:12,900 --> 01:01:14,550 often than not help other people. 994 01:01:14,970 --> 01:01:20,667 So when we hear people say AI will cure cancer, sometimes perhaps they're talking 995 01:01:20,750 --> 01:01:24,837 about you prompt an AI and it thinks up the cure for cancer, but much, much 996 01:01:24,920 --> 01:01:30,617 more likely it's the AI facilitates the individual ambitions of millions 997 01:01:30,700 --> 01:01:34,267 of cancer doctors, and together we accelerate the progress toward cancer. 998 01:01:34,350 --> 01:01:38,237 It's making me think, you know, when I hear people talk about Muse and 999 01:01:38,320 --> 01:01:43,287 say like, "What do people need an AI assistant for? People don't have 1000 01:01:43,370 --> 01:01:47,420 those dreams." Like, what a limited view of peop- of other people. 1001 01:01:48,240 --> 01:01:53,060 You know, I don't know what those are, and it's not up to you or me to say 1002 01:01:53,730 --> 01:01:58,140 which dream is more worthy than the other, to your point, big or small. 1003 01:01:58,360 --> 01:02:03,967 My view on people is like they will have moments of inspiration where at some 1004 01:02:04,050 --> 01:02:09,867 point or another they will realize, like, "Wait, I really wanna do this thing." I 1005 01:02:09,950 --> 01:02:15,167 want everyone who has one of these ideas to be able to tell their Muse, "Hey, I 1006 01:02:15,250 --> 01:02:19,927 have this big idea." And their Muse will just like secretly behind the scenes help 1007 01:02:20,010 --> 01:02:21,517 them figure out how to make it happen. 1008 01:02:21,600 --> 01:02:27,350 And I think part of this is also fanning the flame, so to speak. 1009 01:02:27,720 --> 01:02:32,327 Um, I think for a lot of people, they will have, like, a small want, and then 1010 01:02:32,410 --> 01:02:36,387 it'll happen, and then they'll dream a little bit bigger, and then it'll actually 1011 01:02:36,470 --> 01:02:39,027 happen, and they'll dream a little bit bigger, and it'll actually happen. 1012 01:02:39,110 --> 01:02:45,350 And, um, and as that goes on, people will sort of, uh, like, realize that 1013 01:02:46,110 --> 01:02:49,347 AI can be so much more, that Muse can be so much more, that, you know, they 1014 01:02:49,430 --> 01:02:51,767 can, um, that they can do so much more. 1015 01:02:51,850 --> 01:02:53,470 This won't happen overnight. 1016 01:02:53,580 --> 01:02:56,077 This is something that will take time, and everyone will have to 1017 01:02:56,160 --> 01:02:58,917 experience on their own, and I really hope everyone experiences. 1018 01:02:59,000 --> 01:03:03,567 Um, but I think over, like, like, generations, I think can 1019 01:03:03,650 --> 01:03:05,100 be, like, absolutely magical. 1020 01:03:05,630 --> 01:03:09,747 You know, I think especially the way that you're in the press and talked 1021 01:03:09,830 --> 01:03:14,807 about, and you kind of naturally become sort of a caricature of yourself. 1022 01:03:14,890 --> 01:03:17,567 There's, like, the character of you out there in the world. 1023 01:03:17,650 --> 01:03:21,427 And this has helped me get to know you better because there seems like a 1024 01:03:21,510 --> 01:03:25,950 common theme in some of your answers, which is Muse is a way to facilitate 1025 01:03:26,210 --> 01:03:30,677 other people's dreams because you have a belief in people and their goodness and 1026 01:03:30,760 --> 01:03:35,017 their ability to improve the world and just the sort of value of their ideas. 1027 01:03:35,100 --> 01:03:38,820 And it seems like you are directing Muse in some ways, but you're not 1028 01:03:38,940 --> 01:03:42,987 saying, "This is for coding," or, "This is for X specific task." You're 1029 01:03:43,070 --> 01:03:44,527 watching and, and facilitating. 1030 01:03:44,610 --> 01:03:50,487 And when I asked you where you fall on this, your answer was, "I'm interested 1031 01:03:50,570 --> 01:03:52,727 in the collective wisdom of the group. 1032 01:03:52,810 --> 01:03:56,837 I'm interested in facilitating or being part of a conversation between other 1033 01:03:56,920 --> 01:04:02,387 people, that we have a lot of smart people all around the world, and I 1034 01:04:02,470 --> 01:04:06,317 don't feel like it would be that useful to say, 'Here's exactly my opinion on 1035 01:04:06,400 --> 01:04:08,107 where exactly this should be regulated.' 1036 01:04:08,190 --> 01:04:11,897 I see the best outcome as coming from a group conversation of many 1037 01:04:11,980 --> 01:04:15,327 different people with their many thoughts and, and wants." Do you 1038 01:04:15,410 --> 01:04:16,960 agree with what I've just said? 1039 01:04:17,170 --> 01:04:20,587 Yeah, I think 'cause there are so many big questions about AI, 1040 01:04:20,670 --> 01:04:23,587 and it is, like, something that everyone has a point of view on 1041 01:04:23,670 --> 01:04:25,107 and should have a point of view on. 1042 01:04:25,190 --> 01:04:32,047 And, um, part of, I think, what's amazing about democratic institutions and 1043 01:04:32,130 --> 01:04:35,337 democratic countries and, and democracy broadly speaking is that it matters 1044 01:04:35,420 --> 01:04:38,177 what everyone thinks, and we have to figure out how we, like, take this, 1045 01:04:38,260 --> 01:04:44,247 like, complex, jumbled- Yeah … um, uh, mess of what everyone thinks and channel 1046 01:04:44,330 --> 01:04:46,530 that into the right thing happening. 1047 01:04:46,820 --> 01:04:53,027 And another thing is, like, part of what has held AI back, and I think 1048 01:04:53,110 --> 01:04:57,680 part of the problem of AI, is how it almost feels like the people building 1049 01:04:58,190 --> 01:04:59,747 AI are talking down on other people. 1050 01:04:59,830 --> 01:05:04,780 It's like- Yeah … "Hey, I am building s- the smartest thing you've ever known. 1051 01:05:05,050 --> 01:05:08,887 It's gonna be way smarter than you, and I know exactly what we're gonna do with it." 1052 01:05:08,970 --> 01:05:13,380 And I think rightfully, people are like, "Wait, what?" Like Yeah. 1053 01:05:14,360 --> 01:05:18,527 "What does anything you just said have to do with me?" And, and like, um, "How 1054 01:05:18,610 --> 01:05:21,447 is this even good for me?" And then you're like, "Oh, it'll cure cancer." 1055 01:05:21,530 --> 01:05:25,110 And they're like, "Okay, but what else?" Like Yeah, do I believe you? 1056 01:05:25,240 --> 01:05:26,430 Or, yeah. Yeah. 1057 01:05:26,550 --> 01:05:30,377 And so I think that especially given the moment that we're in, and to your 1058 01:05:30,460 --> 01:05:34,257 point, like, how much fear there is about AI and how much uncertainty there 1059 01:05:34,340 --> 01:05:38,767 is, and how confusing everything that's being said is, like, um, I think the 1060 01:05:38,850 --> 01:05:41,947 most important things are, like, first, let's have everyone experience it. 1061 01:05:42,030 --> 01:05:45,827 Like, let's have everyone actually, like, understand the hype 1062 01:05:45,910 --> 01:05:47,097 Yeah … for be- lack of a better term. 1063 01:05:47,180 --> 01:05:52,990 Like, feel what AI agents can do for you and can do to make your life better. 1064 01:05:53,390 --> 01:05:56,347 And that'll give you perspective as to, "Oh, this is a technology 1065 01:05:56,430 --> 01:05:59,707 that, yeah, it's definitely risky," and, and there's, there's tons of 1066 01:05:59,790 --> 01:06:05,387 risk, don't get me wrong, "but it also is really great, and I like it. 1067 01:06:05,470 --> 01:06:09,657 And I wanna make sure at least some form of it continues, and I can continue 1068 01:06:09,740 --> 01:06:14,157 using." Um, all the way to, you know, the, these, like, big questions and 1069 01:06:14,240 --> 01:06:18,627 these big dialogues, like, how do we just get everyone to a place where we all 1070 01:06:18,710 --> 01:06:20,557 agree, and we can find a path forward? 1071 01:06:20,640 --> 01:06:23,717 And I think I have enough humility that, like, these guys are, you 1072 01:06:23,800 --> 01:06:25,527 know, they're big, important people. 1073 01:06:25,610 --> 01:06:30,677 Like, I don't know if I have much to add, you know, um, given their points of view. 1074 01:06:30,760 --> 01:06:35,117 But I just wanna make sure that we I- facilitate is really the 1075 01:06:35,200 --> 01:06:36,427 word that keeps coming to my mind. 1076 01:06:36,510 --> 01:06:41,387 Like, how do we move quickly to all collectively come to informed 1077 01:06:41,470 --> 01:06:45,340 opinions and, um, and then come to a process where we agree on something? 1078 01:06:46,670 --> 01:06:50,167 This is the question I ask everyone at the end of these HUGE* conversations, 1079 01:06:50,250 --> 01:06:56,867 which is when I think about after I'm dead and after people say that she 1080 01:06:56,950 --> 01:07:03,827 was a really great family member, and she loved her husband, and she loved 1081 01:07:03,910 --> 01:07:07,980 her friends, I would want them to talk about the mission of this show. 1082 01:07:08,210 --> 01:07:12,710 I would want them to talk about how I tried my best to help people see 1083 01:07:12,860 --> 01:07:17,547 ways the future could go well because I really believe, as it seems you do, 1084 01:07:17,630 --> 01:07:23,367 in other people, in their perspective, and when they see a vision of a future 1085 01:07:23,450 --> 01:07:25,737 they believe in, that they're gonna actually try and help make it happen. 1086 01:07:25,820 --> 01:07:29,950 Like, that's the point of what I spend all of my working hours doing. 1087 01:07:30,690 --> 01:07:32,660 You're obviously just starting out. 1088 01:07:32,780 --> 01:07:34,790 You've already had this huge impact. 1089 01:07:34,910 --> 01:07:38,157 You have a lot farther to go, but I'm curious when you think about how 1090 01:07:38,240 --> 01:07:42,060 you wanna be remembered after you're gone, what do you want people to say? 1091 01:07:46,730 --> 01:07:49,907 First of all, what a nice, what a nice, um, idea and concept. 1092 01:07:49,990 --> 01:07:55,547 Um, I think, I think for me, I w- will want people to say that the things that 1093 01:07:55,630 --> 01:08:00,877 I built and the things I contributed to, um, helped everyone get more out 1094 01:08:00,960 --> 01:08:02,910 of what they wanted out of their lives. 1095 01:08:03,210 --> 01:08:04,787 Beautiful. Thanks for doing this. 1096 01:08:04,870 --> 01:08:05,703 Thank you.