1 00:00:00,000 --> 00:00:04,677 So, what I think we've not yet conceptually gotten right and a shared 2 00:00:04,760 --> 00:00:06,837 understanding is what is this future of 3 00:00:06,920 --> 00:00:09,437 work going to look like for a tech CEO? 4 00:00:09,520 --> 00:00:11,717 You have to be deep inside of what's the 5 00:00:11,800 --> 00:00:14,637 tech stack. AI is not a technology. It's 6 00:00:14,720 --> 00:00:16,997 the future of the firm. One of the 7 00:00:17,080 --> 00:00:19,437 dictums I have is don't use frontier 8 00:00:19,520 --> 00:00:24,237 models for non-frontier problems. I think in the AI age that is going to be 9 00:00:24,320 --> 00:00:26,357 everything, right? I think I would be 10 00:00:26,440 --> 00:00:28,197 very surprised Reed if you were sitting 11 00:00:28,280 --> 00:00:31,317 here a year from now if the world is not 12 00:00:31,400 --> 00:00:36,957 completely turned on what is my AI supply chain look like. 13 00:00:37,040 --> 00:00:38,237 >> I couldn't be more delighted to 14 00:00:38,320 --> 00:00:42,077 introduce a special episode of Possible 15 00:00:42,160 --> 00:00:45,277 with Satya Nadella, the chairman and CEO 16 00:00:45,360 --> 00:00:47,197 of Microsoft. Satya and I have known 17 00:00:47,280 --> 00:00:50,037 each other a long time, and part of in 18 00:00:50,120 --> 00:00:52,837 this is kind of AI revolution for 19 00:00:52,920 --> 00:00:55,277 humanity. Uh I thought this would be a 20 00:00:55,360 --> 00:00:58,437 great episode to get out. I mean, we 21 00:00:58,520 --> 00:01:01,157 covered all kinds of important topics. 22 00:01:01,240 --> 00:01:04,557 Satya, as always, is elegant, is 23 00:01:04,640 --> 00:01:07,877 cohesive, is smart, is comprehensive, 24 00:01:07,960 --> 00:01:12,637 and above all, humanist in what is our AI future. This will be an amazing 25 00:01:12,720 --> 00:01:15,357 episode. Actually, one of the things that's great with Satya about filming 26 00:01:15,440 --> 00:01:18,597 this here is it it reminds me of the earliest days when we were talking about 27 00:01:18,680 --> 00:01:21,117 Microsoft and LinkedIn. >> That's right, because we did one of our 28 00:01:21,200 --> 00:01:24,117 very, very great conversations here in the Greylock office, so it's like just 29 00:01:24,200 --> 00:01:26,677 awesome to be back. >> I want to start with something that I 30 00:01:26,760 --> 00:01:29,077 don't know as many people realize and 31 00:01:29,160 --> 00:01:30,917 appreciate about you, which is with how 32 00:01:31,000 --> 00:01:34,317 many books of poetry you have in your 33 00:01:34,400 --> 00:01:38,997 house. Um can you say a little bit about your attraction to poetry, favorite 34 00:01:39,080 --> 00:01:41,397 poets, what what the engagement is there? 35 00:01:41,480 --> 00:01:44,077 >> Yeah, I mean, I I actually got into it 36 00:01:44,160 --> 00:01:46,637 in different times of my life. I 37 00:01:46,720 --> 00:01:48,757 remember that, you know, uh as a middle 38 00:01:48,840 --> 00:01:51,197 schooler we had this standard issue 39 00:01:51,280 --> 00:01:53,717 English poetry book, which I've been 40 00:01:53,800 --> 00:01:57,077 trying to reclaim and get all my life, 41 00:01:57,160 --> 00:02:01,157 but unfortunately it's out of print, but it sort of had, you know, even getting, 42 00:02:01,240 --> 00:02:05,037 you know, introduced to Shelley or Wordsworth and 43 00:02:05,120 --> 00:02:09,597 it had even sort of Indian authors like Sarojini Naidu writing in English and it 44 00:02:09,680 --> 00:02:11,997 is I don't know. I felt maybe it was my 45 00:02:12,080 --> 00:02:16,877 attention span or what have you. I was naturally drawn to poetry as a thing to 46 00:02:16,960 --> 00:02:18,717 enjoy and love and and I've always 47 00:02:18,800 --> 00:02:20,717 compared it even to code which is it's 48 00:02:20,800 --> 00:02:24,277 sort of compression in its best form. Um 49 00:02:24,360 --> 00:02:29,077 and so whenever I'm bored, I get to, you know, go read. I'm not great actually at 50 00:02:29,160 --> 00:02:32,877 understanding deeply. I've never studied it. I'm not like so so and so. That's 51 00:02:32,960 --> 00:02:35,317 why my reputation of knowing about 52 00:02:35,400 --> 00:02:38,957 poetry is far exceeds my knowledge of 53 00:02:39,040 --> 00:02:40,517 poetry. Uh 54 00:02:40,600 --> 00:02:44,077 but I still, you know, continue to 55 00:02:44,160 --> 00:02:47,077 um use poetry as perhaps the best 56 00:02:47,160 --> 00:02:51,917 expression of um uh the human experience, right? I 57 00:02:52,000 --> 00:02:53,717 mean if you if you sort of broadly think 58 00:02:53,800 --> 00:02:57,437 about literature um as sort of what 59 00:02:57,520 --> 00:03:01,437 captures more than even history um 60 00:03:01,520 --> 00:03:04,997 the human experience, I think poetry is the compressed form of it. 61 00:03:05,080 --> 00:03:08,197 >> I completely agree. Is there a particular poet that you go back 62 00:03:08,280 --> 00:03:11,637 >> Yeah, so I the thing that I really 63 00:03:11,720 --> 00:03:16,517 I would say the poetry that probably speaks to me most deeply is the Urdu 64 00:03:16,600 --> 00:03:20,557 poetry. That is sort of because I grew up in Hyderabad in India and Urdu is 65 00:03:20,640 --> 00:03:23,357 sort of the in the air um and some of 66 00:03:23,440 --> 00:03:26,517 the Urdu poets both modern and sort of 67 00:03:26,600 --> 00:03:30,877 um you know, people in the 17th, 18th century were just extraordinary. Uh and 68 00:03:30,960 --> 00:03:34,037 in particular, you know, there's there's this 69 00:03:34,120 --> 00:03:37,157 poet uh whose name it goes by Ghalib as 70 00:03:37,240 --> 00:03:39,757 his name um and he's just uh 71 00:03:39,840 --> 00:03:44,317 extraordinary or even a modern poet like Faiz. Um I was also very having grown up 72 00:03:44,400 --> 00:03:46,237 in you know, in Hyderabad, I was very 73 00:03:46,320 --> 00:03:49,917 influenced by Rumi. Um in fact, um you 74 00:03:50,000 --> 00:03:53,957 You I might the high school I went to was fascinating. In fact, we had um you 75 00:03:54,040 --> 00:03:56,397 know, the number of languages which were all taught was obviously there was 76 00:03:56,480 --> 00:04:01,237 English. Um there was Hindi, which is the national language, and then um we 77 00:04:01,320 --> 00:04:04,997 had Sanskrit. Like basically, you know, uh like Latin 78 00:04:05,080 --> 00:04:08,837 here, I guess. And then uh we had the local language Telugu, 79 00:04:08,920 --> 00:04:10,677 which is my mother tongue. 80 00:04:10,760 --> 00:04:15,157 Um but we also had Urdu and uh Persian. 81 00:04:15,240 --> 00:04:17,557 And so, all of them like in fact, when 82 00:04:17,640 --> 00:04:19,437 we would break for what was called 83 00:04:19,520 --> 00:04:22,277 second language, I I my second language 84 00:04:22,360 --> 00:04:25,357 was Sanskrit. Um but you know, they I 85 00:04:25,440 --> 00:04:29,757 could see I had classmates who would go to Persian, Urdu, Telugu, Hindi, and it 86 00:04:29,840 --> 00:04:32,437 was fascinating. >> Well, among the experiences I'm 87 00:04:32,520 --> 00:04:36,077 experiencing a little bit of language envy cuz I a joke that I like is you 88 00:04:36,160 --> 00:04:38,877 know, what do you call a person who speaks three languages? Trilingual. Two 89 00:04:38,960 --> 00:04:42,237 languages, bilingual. One language, American. And unfortunately, I I 90 00:04:42,320 --> 00:04:46,037 resemble this joke. Um all right. So, let's uh we're just coming out of Build, 91 00:04:46,120 --> 00:04:48,797 which was amazing. Um let's kind of 92 00:04:48,880 --> 00:04:50,877 actually start with kind of going back 93 00:04:50,960 --> 00:04:53,077 in history. So, you know, Microsoft's 94 00:04:53,160 --> 00:04:56,517 first product was a basic interpreter. 95 00:04:56,600 --> 00:05:00,117 Um and 50 years later, here at Build, framed 96 00:05:00,200 --> 00:05:05,237 the company's future on what others can build with AI. So, what is the new basic 97 00:05:05,320 --> 00:05:08,877 interpreter >> Yeah. >> in this AI era? That's a great one. So, 98 00:05:08,960 --> 00:05:11,997 in fact, it's it that that I think is 99 00:05:12,080 --> 00:05:16,837 the core of the DNA of the company, right? Which is what can others build, 100 00:05:16,920 --> 00:05:21,037 and therefore what should we build? It's sort of the two questions that we always 101 00:05:21,120 --> 00:05:23,717 ask ourselves. And in the AI era, 102 00:05:23,800 --> 00:05:28,117 to me, the answer is uh the basic interpreter is the hill-climbing 103 00:05:28,200 --> 00:05:30,557 machine. Right? Because you think of you 104 00:05:30,640 --> 00:05:32,917 Now, bring all of this AI, what is it? 105 00:05:33,000 --> 00:05:35,237 It is basically taking an objective, an 106 00:05:35,320 --> 00:05:38,037 outcome, an eval that you have, 107 00:05:38,120 --> 00:05:40,757 and learning how to achieve that by 108 00:05:40,840 --> 00:05:43,637 learning uh using data, 109 00:05:43,720 --> 00:05:47,037 using some reinforcement reward. And so, 110 00:05:47,120 --> 00:05:50,557 the the entire conference was about not 111 00:05:50,640 --> 00:05:52,717 Hey, here is a new frontier model. 112 00:05:52,800 --> 00:05:56,157 It was about helping every developer, 113 00:05:56,240 --> 00:05:58,797 every company, whether it's a AI native 114 00:05:58,880 --> 00:06:01,477 startup or an enterprise, building their 115 00:06:01,560 --> 00:06:04,597 own hill-climbing machine so that they 116 00:06:04,680 --> 00:06:09,037 can operate at the frontier, right? So, I think that that's it. They The idea In 117 00:06:09,120 --> 00:06:11,437 fact, the you know, being very clear 118 00:06:11,520 --> 00:06:14,237 about the evals and the objectives that 119 00:06:14,320 --> 00:06:18,797 you care deeply about, knowing how to evaluate them, right? That it's in some 120 00:06:18,880 --> 00:06:22,317 sense is the most and keeping That's the new IP. 121 00:06:22,400 --> 00:06:24,237 >> Yes. >> Because everything else is pretty 122 00:06:24,320 --> 00:06:27,677 mechanical. Uh but knowing what is the 123 00:06:27,760 --> 00:06:30,437 set of data that you want to train a 124 00:06:30,520 --> 00:06:35,477 model on and how you reward it is probably where the the next level of IP 125 00:06:35,560 --> 00:06:38,517 gets created. >> Well, and one of the things let's dig 126 00:06:38,600 --> 00:06:41,277 into a couple different areas here cuz you and I've had a number of 127 00:06:41,360 --> 00:06:45,797 conversations, you know, with Microsoft strategy. And one of them is kind of 128 00:06:45,880 --> 00:06:50,317 this question about how enterprises keep the advantage and integrity of 129 00:06:50,400 --> 00:06:55,557 their own data. And Microsoft is the most natural company in the entire world 130 00:06:55,640 --> 00:06:58,717 to do this. So, what Say a little bit about how enterprises should be thinking 131 00:06:58,800 --> 00:07:02,677 about like, "We need our own frontier intel- intelligence, but we also need to 132 00:07:02,760 --> 00:07:04,077 maintain control of our data." 133 00:07:04,160 --> 00:07:07,357 >> Yeah, see that I think is the the question, which is 134 00:07:07,440 --> 00:07:10,637 this economy is going to be shaped going forward 135 00:07:10,720 --> 00:07:13,157 by both human capital 136 00:07:13,240 --> 00:07:17,837 and let's call it this token capital, right? That is true for Microsoft, that 137 00:07:17,920 --> 00:07:22,237 is true for a new startup, that's going to be true for any bank that's been in 138 00:07:22,320 --> 00:07:24,397 existence for 100 years, right? Doesn't 139 00:07:24,480 --> 00:07:26,357 matter. All of us will now need to sort 140 00:07:26,440 --> 00:07:29,797 of in fact the interplay between human 141 00:07:29,880 --> 00:07:34,437 capital and token capital and compounding the returns of it is what 142 00:07:34,520 --> 00:07:38,197 you need to do. So, if you sort of frame it that way, then one of the most 143 00:07:38,280 --> 00:07:42,477 important thing is not just even thinking of data in its aggregate sense, 144 00:07:42,560 --> 00:07:44,717 right? See, what is the tacit knowledge 145 00:07:44,800 --> 00:07:48,077 of an enterprise or a firm? It's the 146 00:07:48,160 --> 00:07:51,597 unique ways that you're able to operate, 147 00:07:51,680 --> 00:07:55,277 pass judgment, have taste. All that's 148 00:07:55,360 --> 00:07:57,437 the tacit knowledge, mostly captured 149 00:07:57,520 --> 00:07:59,957 today in the tacit knowledge that is 150 00:08:00,040 --> 00:08:02,677 there with the human capital, and some 151 00:08:02,760 --> 00:08:04,837 artifacts that are digital. 152 00:08:04,920 --> 00:08:08,397 So, now when it comes to AI, the the 153 00:08:08,480 --> 00:08:10,917 model, in some sense, is able to 154 00:08:11,000 --> 00:08:13,277 extract, if anything, that tacit 155 00:08:13,360 --> 00:08:18,557 knowledge through human trajectories, and then code it in a set of weights in 156 00:08:18,640 --> 00:08:21,797 a model. So, what I claim is that every 157 00:08:21,880 --> 00:08:24,197 enterprise now needs to be more mindful 158 00:08:24,280 --> 00:08:28,157 about that interplay of humans and their 159 00:08:28,240 --> 00:08:30,517 digital estate working together, those 160 00:08:30,600 --> 00:08:33,197 trajectories training essentially the 161 00:08:33,280 --> 00:08:36,237 models that they keep as IP versus leak 162 00:08:36,320 --> 00:08:40,317 it. Because if you leak it, it's a one-way door. You're done in some sense, 163 00:08:40,400 --> 00:08:44,037 right? Which is what is unique, what you may have spent a hundred years, right? 164 00:08:44,120 --> 00:08:47,517 In fact, we don't even know how to articulate it, right? Nobody sort of has 165 00:08:47,600 --> 00:08:49,237 a line item in their balance sheet 166 00:08:49,320 --> 00:08:51,837 called tacit knowledge, but we take it 167 00:08:51,920 --> 00:08:54,597 for granted that because we have human 168 00:08:54,680 --> 00:08:59,357 capital, we have it. Now, I believe that can leak. And in 169 00:08:59,440 --> 00:09:02,637 fact, it is leaking, right? If you look at even the number how the model 170 00:09:02,720 --> 00:09:04,237 companies learn, they're essentially 171 00:09:04,320 --> 00:09:07,237 setting up these gyms with rewards, 172 00:09:07,320 --> 00:09:11,997 which are employing employees who worked at your company previously. I mean, 173 00:09:12,080 --> 00:09:15,437 that's sort of should tell you everything what what should not be 174 00:09:15,520 --> 00:09:19,197 happening. Which is So, that's one of the fundamental reasons why we want this 175 00:09:19,280 --> 00:09:24,157 effectively the regime to change or the paradigm to change, where you welcome 176 00:09:24,240 --> 00:09:28,197 the models to come in. They should hill climb inside a machine 177 00:09:28,280 --> 00:09:30,837 that you control. Your data is your 178 00:09:30,920 --> 00:09:33,477 context you feed the model. You collect, 179 00:09:33,560 --> 00:09:35,877 in fact, these traces or trajectories of 180 00:09:35,960 --> 00:09:38,197 how work gets done between humans and 181 00:09:38,280 --> 00:09:40,837 agents inside the enterprise. 182 00:09:40,920 --> 00:09:45,717 But you have a continuous loop of that and you're not letting that leak. That I 183 00:09:45,800 --> 00:09:47,597 think is the fundamental operation. 184 00:09:47,680 --> 00:09:51,437 >> Yep. And actually the related parallel to the tacit knowledge is, and this is 185 00:09:51,520 --> 00:09:53,597 one of the conversations you and I had 186 00:09:53,680 --> 00:09:55,477 actually at the at the Microsoft board, 187 00:09:55,560 --> 00:09:59,077 which was what uh is the future role of 188 00:09:59,160 --> 00:10:03,117 AI employees? Right? Because part of the question is say, "Hey, we're going to 189 00:10:03,200 --> 00:10:08,517 provision not just, you know, kind of uh tools to amplify human work, great for 190 00:10:08,600 --> 00:10:11,637 you know, kind of AI companies, but we're also going to provision at least 191 00:10:11,720 --> 00:10:15,237 specialist employees." But the challenge with that is is that 192 00:10:15,320 --> 00:10:16,997 one of the things that really matters to 193 00:10:17,080 --> 00:10:20,757 enterprises is your employees embody a 194 00:10:20,840 --> 00:10:24,917 lot of, you know, tacit knowledge, how to, what to do. 195 00:10:25,000 --> 00:10:29,237 Do you want that in other companies and provision other places? What's the your 196 00:10:29,320 --> 00:10:31,837 and Microsoft view of the future of work 197 00:10:31,920 --> 00:10:37,597 when it comes down to how to think about AI as also, maybe even specialist 198 00:10:37,680 --> 00:10:39,437 employees? >> Yeah. I mean, I think that that that's 199 00:10:39,520 --> 00:10:42,637 right. So, the way to think about this, you know, there are two ways I come at 200 00:10:42,720 --> 00:10:45,997 this, right? Which is let's let's take one analogy and I'll come back to it, 201 00:10:46,080 --> 00:10:49,637 right? If somebody in the early '80s had come to us and said, "You know what? 202 00:10:49,720 --> 00:10:53,517 There are going to be 4 billion typists uh who are going to wake up every 203 00:10:53,600 --> 00:10:57,117 morning and start typing." And we would say, "What for? I mean, it makes no 204 00:10:57,200 --> 00:10:59,037 sense, right? We have a typist pool, we 205 00:10:59,120 --> 00:11:01,357 have a slide pool, and we're you know, 206 00:11:01,440 --> 00:11:04,437 fine, uh you know, with that." But 207 00:11:04,520 --> 00:11:07,357 except we invented this complete new 208 00:11:07,440 --> 00:11:11,757 thing called knowledge work, where everybody was typing and creating 209 00:11:11,840 --> 00:11:13,957 artifacts and so on. 210 00:11:14,040 --> 00:11:18,957 So, what I think we have not yet conceptually gotten right, and our good 211 00:11:19,040 --> 00:11:22,397 understanding, and our shared understanding, is what is this future of 212 00:11:22,480 --> 00:11:26,117 work going to look like when you have, let's call it, you know, and let's take 213 00:11:26,200 --> 00:11:28,437 Microsoft, we have 20,000 employees and 214 00:11:28,520 --> 00:11:31,077 we have, let's say, 2 million uh agents 215 00:11:31,160 --> 00:11:35,757 or 20 million agents. All in a loop. 216 00:11:35,840 --> 00:11:39,757 What is happening? Uh what is the tacit knowledge that gets 217 00:11:39,840 --> 00:11:41,557 created? 218 00:11:41,640 --> 00:11:45,117 Um what is, you know, the artifacts uh 219 00:11:45,200 --> 00:11:50,037 that are going between agents, between agents and humans, and so on. All that's 220 00:11:50,120 --> 00:11:53,917 now going to be played out in the next whatever even multiple years. You can 221 00:11:54,000 --> 00:11:56,677 see early forms of this in coding, 222 00:11:56,760 --> 00:12:01,757 right? It's a great place to observe, in fact, the the the this even the sort of 223 00:12:01,840 --> 00:12:03,477 social change, right? I mean, think 224 00:12:03,560 --> 00:12:07,197 about We started in a good old IDE and 225 00:12:07,280 --> 00:12:11,757 said, "Hey, AI is in you know, in there." Uh and it's doing code 226 00:12:11,840 --> 00:12:15,757 completion. That itself was useful, easy to understand. We've always had spelling 227 00:12:15,840 --> 00:12:17,797 correction and IntelliSense was there 228 00:12:17,880 --> 00:12:22,837 for in VS Code, you know, from there 15 years ago and and it just got better. 229 00:12:22,920 --> 00:12:26,997 Then we said, "Oh, you know, instead of going out of band and going to a browser 230 00:12:27,080 --> 00:12:31,757 and, you know, going to Stack Overflow and searching, you can now bring all 231 00:12:31,840 --> 00:12:35,757 that coding knowledge to a chat session." Well, that was also also easy 232 00:12:35,840 --> 00:12:40,237 to understand and you still were in the IDE, you had sort of the chat. 233 00:12:40,320 --> 00:12:44,557 Then we said, "Okay, um you now have reasoning models and some 234 00:12:44,640 --> 00:12:48,397 you know, primary you know, preliminary agent loops and so on. And so you can 235 00:12:48,480 --> 00:12:51,717 assign tasks." So we had agent mode, right? So you not only had chat, but you 236 00:12:51,800 --> 00:12:55,637 can give it small tasks and you could see it complete 237 00:12:55,720 --> 00:12:59,277 um and then you could accept what it did and then you could insert it and what 238 00:12:59,360 --> 00:13:00,437 have you. 239 00:13:00,520 --> 00:13:03,797 Then came the big breakthrough of total 240 00:13:03,880 --> 00:13:08,117 autonomy and agentic loop working for long periods of time where you could 241 00:13:08,200 --> 00:13:11,397 literally fire forget, right? Where you could assign a high-level intent and 242 00:13:11,480 --> 00:13:13,757 it'll go off and do the full PR and you 243 00:13:13,840 --> 00:13:15,837 would accept the PR. 244 00:13:15,920 --> 00:13:18,917 So that transition, 245 00:13:19,000 --> 00:13:24,237 right? Is what I think is going to happen across all work, right? And then 246 00:13:24,320 --> 00:13:28,037 now we're seeing it with even in Copilot, we now have chat and we have 247 00:13:28,120 --> 00:13:31,997 co-work and now in fact at build we announced something called scout and 248 00:13:32,080 --> 00:13:37,237 autopilot. So, what is happening in coding will happen even in knowledge work. 249 00:13:37,320 --> 00:13:38,877 Interestingly enough, one of the things 250 00:13:38,960 --> 00:13:42,597 we launched even at build in GitHub was 251 00:13:42,680 --> 00:13:44,517 a new feature called canvas because what 252 00:13:44,600 --> 00:13:46,477 has happened is 253 00:13:46,560 --> 00:13:49,677 as we've all gotten so good at using 254 00:13:49,760 --> 00:13:53,517 these coding agents. In fact, you know, the biggest challenge we now have is I 255 00:13:53,600 --> 00:13:58,757 have a hundred CLI sessions open I'm trying to operate these hundred you 256 00:13:58,840 --> 00:14:00,117 know, agents 257 00:14:00,200 --> 00:14:03,357 and now the cognitive load on me 258 00:14:03,440 --> 00:14:07,237 managing this is so high, right? I mean first of all, I don't I mean they think 259 00:14:07,320 --> 00:14:11,757 about it, right? It's a linear chat session in a command line and I have a 260 00:14:11,840 --> 00:14:13,997 hundred of them. So, guess what? We now 261 00:14:14,080 --> 00:14:17,117 have a back to an IDE, right? So, I have 262 00:14:17,200 --> 00:14:19,277 a new IDE we give it a new fancy name. 263 00:14:19,360 --> 00:14:22,437 It's called an ADE. It's an agentic development environment. That's what the 264 00:14:22,520 --> 00:14:24,717 new GitHub app is. The GitHub app looks 265 00:14:24,800 --> 00:14:29,677 like an inbox except it's an inbox of agents that are working across all the 266 00:14:29,760 --> 00:14:30,917 repos 267 00:14:31,000 --> 00:14:36,877 in in allowing me to do the micro steering of the macro delegation I gave 268 00:14:36,960 --> 00:14:40,477 them. Uh but it's a complete UI for the agents 269 00:14:40,560 --> 00:14:44,957 to deal with me and for me to deal with agents. And so, I feel so we introduced 270 00:14:45,040 --> 00:14:49,357 this new feature called canvas in GitHub where it can like for example, you know, 271 00:14:49,440 --> 00:14:51,917 sometimes it's easy to have a Kanban 272 00:14:52,000 --> 00:14:54,917 board visualization as a way to run down 273 00:14:55,000 --> 00:14:58,957 your PRs. Very useful both for the agents and me 274 00:14:59,040 --> 00:15:01,357 to interact versus some chat session. 275 00:15:01,440 --> 00:15:04,277 So, I think that that type of innovation 276 00:15:04,360 --> 00:15:08,517 is how work will change. In fact, one of the other fascinating things for me is a 277 00:15:08,600 --> 00:15:10,717 line of even AI research would be the 278 00:15:10,800 --> 00:15:14,477 models will also get much more tuned to 279 00:15:14,560 --> 00:15:18,797 learning how to stay the course 280 00:15:18,880 --> 00:15:22,917 um and understand human preferences in 281 00:15:23,000 --> 00:15:27,357 steerability, right? Because that's what I want. And in fact, I want models that 282 00:15:27,440 --> 00:15:32,397 not only do instruction following, but are also really steerable. 283 00:15:32,480 --> 00:15:35,317 Uh and that when you have that, that's when you have confidence in it. So, I 284 00:15:35,400 --> 00:15:38,037 think the work of the future is about 285 00:15:38,120 --> 00:15:40,437 tacit knowledge that gets created by 286 00:15:40,520 --> 00:15:42,557 this interplay. Just like knowledge work 287 00:15:42,640 --> 00:15:46,677 got created because of digital artifacts 288 00:15:46,760 --> 00:15:48,597 and human capital, 289 00:15:48,680 --> 00:15:51,677 the new one would be this AI capital and 290 00:15:51,760 --> 00:15:56,157 the human capital working together, creating in some sense the digital 291 00:15:56,240 --> 00:15:58,637 artifacts. >> And what do you think are the things 292 00:15:58,720 --> 00:16:00,997 that people miss about how important it 293 00:16:01,080 --> 00:16:04,077 is to build out additional kind of 294 00:16:04,160 --> 00:16:06,797 structure for enabling enterprises. So, 295 00:16:06,880 --> 00:16:08,837 one of them is obviously the enabling 296 00:16:08,920 --> 00:16:11,917 humans canvas, you know, ADEs, etc. But 297 00:16:12,000 --> 00:16:17,357 what are the like, for example, notions of like, what is uh security in an 298 00:16:17,440 --> 00:16:19,077 enterprise? >> a great point. It's a great point. So, 299 00:16:19,160 --> 00:16:22,877 in fact, the to your point, one is the experience layer. 300 00:16:22,960 --> 00:16:26,557 Clearly, that's important. And another one which we referenced earlier was, 301 00:16:26,640 --> 00:16:30,437 hey, we need this hill climbing machine as a concept that you needs to be 302 00:16:30,520 --> 00:16:34,757 instantiated. But the third is the manageability of it and the security of 303 00:16:34,840 --> 00:16:36,517 it. Starting with observability, right? 304 00:16:36,600 --> 00:16:39,597 One of the first things we did even, and we talked about it even yesterday, is 305 00:16:39,680 --> 00:16:43,797 something called Agent 365. I need to know, I need to have an inventory. I 306 00:16:43,880 --> 00:16:47,877 said, oh, there may be 20 million agents at Microsoft. I first need to know, what 307 00:16:47,960 --> 00:16:49,677 are these agents? What are they doing? 308 00:16:49,760 --> 00:16:53,117 What are their reasoning traces? They need to be fully inspectable, fully 309 00:16:53,200 --> 00:16:55,957 auditable. Uh and by the way, when the 310 00:16:56,040 --> 00:16:58,157 agents also have this other attribute, 311 00:16:58,240 --> 00:17:00,597 they execute they are running, they can 312 00:17:00,680 --> 00:17:03,837 generate code and execute. So, you need 313 00:17:03,920 --> 00:17:08,077 the environment in which they are executing code with maybe even file 314 00:17:08,160 --> 00:17:13,277 system access, network access, to be think that is governed by policy. So, 315 00:17:13,360 --> 00:17:16,637 therefore, you now need to give them identities, you need to give them 316 00:17:16,720 --> 00:17:19,237 sandboxes, then you need to set policies 317 00:17:19,320 --> 00:17:23,957 to govern them. And so we built this entire thing called Agent 365, which 318 00:17:24,040 --> 00:17:27,277 really has, you know, we've extended Entra so that you have an identity, we 319 00:17:27,360 --> 00:17:31,557 extended Defender so that you have security, we extended Purview so that 320 00:17:31,640 --> 00:17:35,437 you can even label the data that is getting created automatically so that 321 00:17:35,520 --> 00:17:39,397 you can have data protection. So, I think security, containment, 322 00:17:39,480 --> 00:17:44,517 manageability, observability is the way we will have confidence around these 323 00:17:44,600 --> 00:17:46,357 agents. The one other thing I'd say Reed 324 00:17:46,440 --> 00:17:49,037 also was you know salient in the 325 00:17:49,120 --> 00:17:53,637 developer conference was how important it is when you're building these 326 00:17:53,720 --> 00:17:56,877 long-running agents. All right, like if you think about 327 00:17:56,960 --> 00:18:02,157 right, we have always had in programming languages, 328 00:18:02,240 --> 00:18:06,677 you know, programming models for understanding the 329 00:18:06,760 --> 00:18:10,717 very, you know, the verifiability of a program, right? And it's execu- at 330 00:18:10,800 --> 00:18:13,277 runtime. So, one of the attributes of 331 00:18:13,360 --> 00:18:15,437 these long-running agents that we added 332 00:18:15,520 --> 00:18:18,277 to Foundry was something called asserts. 333 00:18:18,360 --> 00:18:22,237 So, this allows us to assert what are 334 00:18:22,320 --> 00:18:26,757 the boundaries, right? So, instead of talking about guardrails as a thing that 335 00:18:26,840 --> 00:18:28,677 is sort of just a classifier of some 336 00:18:28,760 --> 00:18:31,717 form, you really need to now have the 337 00:18:31,800 --> 00:18:34,517 ability during execution 338 00:18:34,600 --> 00:18:38,677 to really have that execution path not 339 00:18:38,760 --> 00:18:42,517 go off rails. And so, there's a lot of engineering sophistication now that's 340 00:18:42,600 --> 00:18:46,397 emerging as we build out the platform, the 341 00:18:46,480 --> 00:18:51,277 runtime, the security layer, the management layer, the observability layer. 342 00:18:51,360 --> 00:18:55,397 >> So, let's go to a little bit of the role of CEO. And I think 343 00:18:55,480 --> 00:18:59,877 Fortune recently described you as acting like a startup CEO inside of Microsoft's 344 00:18:59,960 --> 00:19:03,237 AI teams. But 345 00:19:03,320 --> 00:19:07,037 the more general question because I think a lot of CEOs seek, you know, kind 346 00:19:07,120 --> 00:19:09,957 of technology strategy advice about 347 00:19:10,040 --> 00:19:11,797 what's going on in the world from you. 348 00:19:11,880 --> 00:19:17,077 What do you think the CEOs should be doing around AI? 349 00:19:17,160 --> 00:19:20,717 What um you know, kind of when they're like AI 350 00:19:20,800 --> 00:19:24,637 brings the kind of a re-founding moment to lots of companies because of the 351 00:19:24,720 --> 00:19:29,437 nature. And so, what's your advice to them for how they should engage? 352 00:19:29,520 --> 00:19:31,517 >> Great question because I've been thinking about this, right? Because in 353 00:19:31,600 --> 00:19:33,917 some sense, you know, it clearly for if 354 00:19:34,000 --> 00:19:38,597 you sort of you're a tech CEO, you have to be deep inside of what's the tech 355 00:19:38,680 --> 00:19:42,957 stack. There's no way you could be a tech CEO and not essentially have a 356 00:19:43,040 --> 00:19:44,837 fundamental 357 00:19:44,920 --> 00:19:48,157 uh worldview on where the future is 358 00:19:48,240 --> 00:19:50,157 going and then to be long before it's 359 00:19:50,240 --> 00:19:55,277 conventional wisdom, you have to pass judgment on uh you know, where the 360 00:19:55,360 --> 00:19:59,117 company's going, right? So, that's sort of the our industry's sort of pretty 361 00:19:59,200 --> 00:20:03,277 binary transition. So, I think there's no hope if you don't have a 362 00:20:03,360 --> 00:20:06,357 uh the CEO leading uh from the front and 363 00:20:06,440 --> 00:20:08,957 taking the shot on goal knowing that 364 00:20:09,040 --> 00:20:12,117 these are fairly uh harsh transitions. 365 00:20:12,200 --> 00:20:14,397 But, the interesting thing that now I'm 366 00:20:14,480 --> 00:20:19,637 up to now if you are not a tech CEO, you needed to be a great CEO, which means 367 00:20:19,720 --> 00:20:23,517 you need to be great at banking or healthcare or whatever it is. 368 00:20:23,600 --> 00:20:28,717 And you could always have great partners and technology advisers and what have 369 00:20:28,800 --> 00:20:33,197 you. But, now I'm changing my prior on that, right? Because I think 370 00:20:33,280 --> 00:20:37,117 because I'm noticing even what's happening and I don't think the the rest 371 00:20:37,200 --> 00:20:42,237 of the the the industry and the CEO community has broadly woken up um to 372 00:20:42,320 --> 00:20:45,557 this and I'm like perplexed at it because they're still you know, pretty 373 00:20:45,640 --> 00:20:49,957 happy doing a press release with a tech company and saying, "Yeah, I got my AI 374 00:20:50,040 --> 00:20:53,357 strategy." And you know, pointing to eight agents they built and some 375 00:20:53,440 --> 00:20:54,717 outcome. 376 00:20:54,800 --> 00:20:57,997 But, this is no longer AI is not like a 377 00:20:58,080 --> 00:21:02,357 sort of a technology. It's the future of the firm. 378 00:21:02,440 --> 00:21:05,517 Right? It would be like saying, "Oh, I don't know what the about the human 379 00:21:05,600 --> 00:21:10,117 capital in my firm." Right? So, I think that this is where my thing would be 380 00:21:10,200 --> 00:21:13,437 that you have to now get a deep understanding 381 00:21:13,520 --> 00:21:15,637 of what's your token capital. Can you 382 00:21:15,720 --> 00:21:17,517 answer that question? Right? Can you 383 00:21:17,600 --> 00:21:21,317 concretely say, "Oh, last night 384 00:21:21,400 --> 00:21:24,157 or based on all the work that happened 385 00:21:24,240 --> 00:21:27,077 across all of our operations, 386 00:21:27,160 --> 00:21:30,277 we were able to translate that into a 387 00:21:30,360 --> 00:21:33,717 set of knowledge that is somehow now 388 00:21:33,800 --> 00:21:38,197 part of my token capital. It can be some context. 389 00:21:38,280 --> 00:21:42,557 It can be some skill. It can be some weights in a model. It 390 00:21:42,640 --> 00:21:44,477 doesn't need to be any one thing, but 391 00:21:44,560 --> 00:21:46,437 you need to be able to specific you 392 00:21:46,520 --> 00:21:49,797 know, you need to be able to identify it 393 00:21:49,880 --> 00:21:53,717 clearly as something that you own, you 394 00:21:53,800 --> 00:21:56,517 control, you created, and you're put in 395 00:21:56,600 --> 00:21:59,917 place a system to have that compound. 396 00:22:00,000 --> 00:22:04,117 So, I think that that is going to be probably the toughest change. So, this 397 00:22:04,200 --> 00:22:08,837 change is like unlike how this is like going to a mobile phone era or a PC era 398 00:22:08,920 --> 00:22:12,917 or a cloud era where all I needed to do is to have an IT department that knew 399 00:22:13,000 --> 00:22:15,597 how to deal with a bunch of vendors and 400 00:22:15,680 --> 00:22:20,517 did some smart things to reduce cost or improve my efficiency. It should start 401 00:22:20,600 --> 00:22:21,717 there. I'm not saying that that's not 402 00:22:21,800 --> 00:22:24,517 the place, but this is about what what 403 00:22:24,600 --> 00:22:27,597 happens structurally in your industry 404 00:22:27,680 --> 00:22:30,957 when AI is sort of basically knows 405 00:22:31,040 --> 00:22:35,437 everything that it needs to know about being in your industry. 406 00:22:35,520 --> 00:22:39,437 And if you start there, then I think you will start understanding that this is 407 00:22:39,520 --> 00:22:43,117 not just a tech. Uh it is really about a fundamental 408 00:22:43,200 --> 00:22:47,397 change uh to the firm. >> And one of the things that I think 409 00:22:47,480 --> 00:22:52,197 thinking of the nature of the firm uh is that you've done a great job in 410 00:22:52,280 --> 00:22:55,997 leading Microsoft. One of the many different genius moments and skills is 411 00:22:56,080 --> 00:22:59,437 uh going through acquisitions and partnerships, right? So, it's you know, 412 00:22:59,520 --> 00:23:02,877 whether it's LinkedIn, obviously something close to our hearts. 413 00:23:02,960 --> 00:23:05,677 You know, GitHub, OpenAI, etc. 414 00:23:05,760 --> 00:23:08,677 And part of what I think Microsoft leads 415 00:23:08,760 --> 00:23:12,797 the entire world on is how do you build trust in these other ecosystems? And in 416 00:23:12,880 --> 00:23:15,437 the world of AI, that's going to be extremely important, especially as the 417 00:23:15,520 --> 00:23:19,837 firm changes. So, what are some of that the kind of the 418 00:23:19,920 --> 00:23:23,437 um the kind of lessons you've learned, the 419 00:23:23,520 --> 00:23:26,277 things that you are advocating for in 420 00:23:26,360 --> 00:23:28,637 order to build and maintain that trust? 421 00:23:28,720 --> 00:23:31,317 >> so it's so key because in some sense 422 00:23:31,400 --> 00:23:33,917 I've always sort of thought about what's 423 00:23:34,000 --> 00:23:37,157 long-term stable, right? Um even from a 424 00:23:37,240 --> 00:23:41,597 Microsoft perspective. What is long-term stable is for us to be a tools and a 425 00:23:41,680 --> 00:23:44,077 platform company where we fundamentally 426 00:23:44,160 --> 00:23:46,597 are defined by the amount of value that 427 00:23:46,680 --> 00:23:51,397 gets created on top of the platform, which should far exceed 428 00:23:51,480 --> 00:23:54,797 anything that is captured in the platform, right? It That's the only way 429 00:23:54,880 --> 00:23:56,917 to have stability. 430 00:23:57,000 --> 00:23:59,517 If you do that, then you have trust. 431 00:23:59,600 --> 00:24:01,957 Right? Because then the customer, the 432 00:24:02,040 --> 00:24:05,797 partner knows that it's not a zero-sum. 433 00:24:05,880 --> 00:24:08,637 Uh especially it's not a game theoretic 434 00:24:08,720 --> 00:24:11,597 z- you know, zero-sum with these games 435 00:24:11,680 --> 00:24:13,877 where, "Oh, I'm going to subsidize this 436 00:24:13,960 --> 00:24:16,757 until 2 years or 3 years only to then 437 00:24:16,840 --> 00:24:21,797 eat your lunch" type of games that the tech people are really good at. 438 00:24:21,880 --> 00:24:24,717 Um and so, I've always felt like, "Hey, 439 00:24:24,800 --> 00:24:29,237 look, that is all not the way. The way to think about it is to be very 440 00:24:29,320 --> 00:24:32,677 principled that you are a platform company, 441 00:24:32,760 --> 00:24:38,237 that you will really live and die by your ability to create success on top of 442 00:24:38,320 --> 00:24:40,677 the platform, and when they're successful, you're going to be 443 00:24:40,760 --> 00:24:43,197 successful." And that is the equation 444 00:24:43,280 --> 00:24:45,397 that builds trust and long-term 445 00:24:45,480 --> 00:24:48,317 stability for both sides, right? Um 446 00:24:48,400 --> 00:24:53,077 I think in the AI age, that is going to be everything, right? I think I would be 447 00:24:53,160 --> 00:24:57,197 very surprised, Reed, if you were sitting here a year from now 448 00:24:57,280 --> 00:24:59,397 if the world is not completely turned 449 00:24:59,480 --> 00:25:03,157 on, what is my AI supply chain look 450 00:25:03,240 --> 00:25:07,797 like, where it is helping me as a firm 451 00:25:07,880 --> 00:25:11,197 compound the returns of AI that I can 452 00:25:11,280 --> 00:25:15,597 uniquely point to as my value in a world where we know that these learning 453 00:25:15,680 --> 00:25:17,477 systems by definition don't have 454 00:25:17,560 --> 00:25:19,837 boundaries. They have to have boundaries 455 00:25:19,920 --> 00:25:23,197 because of mostly structures of markets 456 00:25:23,280 --> 00:25:26,557 and society and other mechanisms. 457 00:25:26,640 --> 00:25:29,477 >> Yep. So, we've talked a lot about chatbots 458 00:25:29,560 --> 00:25:34,117 and code, um which have just seen amazing explosive progress. 459 00:25:34,200 --> 00:25:38,557 Um what are some of the next frontiers for LLMs over the next few years? 460 00:25:38,640 --> 00:25:42,037 >> Well, I mean, talking about that, in fact, you know, one of the things that I 461 00:25:42,120 --> 00:25:44,517 know you have gotten very, very excited 462 00:25:44,600 --> 00:25:46,557 about, which I'm quite frankly, you 463 00:25:46,640 --> 00:25:49,357 know, excited for you, but it also has 464 00:25:49,440 --> 00:25:52,037 caused us to you to examine where you 465 00:25:52,120 --> 00:25:53,917 spend your time, and I'll let you speak 466 00:25:54,000 --> 00:25:56,917 to it, but it's sort of unbelievable, 467 00:25:57,000 --> 00:25:59,837 uh to see what's happening in science, 468 00:25:59,920 --> 00:26:04,637 um and and in fact, the company that I, you know, you started with 469 00:26:04,720 --> 00:26:09,397 uh Sid and Monis, you know, it's just a great example, um 470 00:26:09,480 --> 00:26:11,557 I'd say of 471 00:26:11,640 --> 00:26:14,477 what has happened in, let's call it, 472 00:26:14,560 --> 00:26:16,717 coding and knowledge work. 473 00:26:16,800 --> 00:26:19,157 If it start happening in science, in 474 00:26:19,240 --> 00:26:22,117 fact, I would even claim that I wish in 475 00:26:22,200 --> 00:26:24,397 fact, we did it in the reverse order, uh 476 00:26:24,480 --> 00:26:27,557 right? Because the um social permission 477 00:26:27,640 --> 00:26:30,837 for AI would be so much higher if we had 478 00:26:30,920 --> 00:26:33,357 started with some nice discoveries in 479 00:26:33,440 --> 00:26:37,957 science that were having great societal benefits. Uh then people would have 480 00:26:38,040 --> 00:26:41,957 really thought of this as a thing that is really going to be more helpful. So, 481 00:26:42,040 --> 00:26:46,037 I'm really excited uh for you, Reed, what what you're doing. Obviously, 482 00:26:46,120 --> 00:26:49,037 you've spent a lot of time, but maybe you want to talk about that. 483 00:26:49,120 --> 00:26:52,237 >> Well, yeah, so, I mean, it was one of the things that uh I realized over the 484 00:26:52,320 --> 00:26:57,717 last month was that um we're seeing such progress with Manas in 485 00:26:57,800 --> 00:27:01,917 actually in fact we've got an internal description of move 37 for chemistry cuz 486 00:27:02,000 --> 00:27:06,637 we're seeing chemistry that that might actually in fact take shots at really 487 00:27:06,720 --> 00:27:08,797 interesting cancers and other things. 488 00:27:08,880 --> 00:27:10,677 And you know, it's very early, right? 489 00:27:10,760 --> 00:27:14,437 And so you know, these things take a while with all the IDs, but the the fact 490 00:27:14,520 --> 00:27:17,077 that we are already beginning to see like we've got some of the best 491 00:27:17,160 --> 00:27:18,517 computational chemists in the world. 492 00:27:18,600 --> 00:27:21,637 They looked and said that's very interesting. We've never 493 00:27:21,720 --> 00:27:26,557 seen it before and that might work, right? And so that kind of thing and so 494 00:27:26,640 --> 00:27:29,197 you know, Sid and Uzair and I were talking about this and I said, look, I 495 00:27:29,280 --> 00:27:33,957 think I need to get back to founder mode in terms of how to do this and so I need 496 00:27:34,040 --> 00:27:38,317 to be able to kind of focus on this and then you know, part of the you know, 497 00:27:38,400 --> 00:27:41,797 conversation you and I have had this week is to say, okay, 498 00:27:41,880 --> 00:27:46,677 um it's been 10 years on the Microsoft board. It's been a huge honor and 499 00:27:46,760 --> 00:27:50,957 pleasure with you know, and not just LinkedIn, but obviously Open AI, GitHub, 500 00:27:51,040 --> 00:27:54,557 a whole bunch of things. But you know, at the end of the year I 501 00:27:54,640 --> 00:27:56,877 should really be transitioning right now 502 00:27:56,960 --> 00:27:58,757 to being founder mode. So like you know, 503 00:27:58,840 --> 00:28:02,277 it's like and you know, we'll always be 504 00:28:02,360 --> 00:28:06,277 working together. So it's like but but you know, it's time to kind of dig back 505 00:28:06,360 --> 00:28:08,397 into the company. >> Now, first of all, it's a been such a 506 00:28:08,480 --> 00:28:10,357 privilege to have you on the Microsoft 507 00:28:10,440 --> 00:28:12,597 board, work with you, obviously through 508 00:28:12,680 --> 00:28:15,717 all the partnerships and LinkedIn and 509 00:28:15,800 --> 00:28:19,677 you will definitely be missed on the Microsoft board, but I know we'll be 510 00:28:19,760 --> 00:28:24,037 very connected, but I'm also excited for you. You know, when you say you're going 511 00:28:24,120 --> 00:28:27,957 back into founder mode, you know, that means I'm even more 512 00:28:28,040 --> 00:28:32,197 curious about what you're going to be building because it's a as you I think you 513 00:28:32,280 --> 00:28:35,677 rightfully captured, this is the moment where I think the impact of these 514 00:28:35,760 --> 00:28:39,957 technologies more broadly felt in the society is the need of the hour. 515 00:28:40,040 --> 00:28:43,157 >> Yeah, no, exactly. [clears throat] >> Let's come back to you know, kind of 516 00:28:43,240 --> 00:28:45,397 questions around 517 00:28:45,480 --> 00:28:47,877 the patterns of thinking 518 00:28:47,960 --> 00:28:51,237 that happen with AI because one of the 519 00:28:51,320 --> 00:28:55,437 things that I think like I kind of roughly think about these things is 520 00:28:55,520 --> 00:28:57,597 you've got an alien intelligence, right? 521 00:28:57,680 --> 00:29:00,077 A different version of A that has been 522 00:29:00,160 --> 00:29:03,237 built to very much heavily mimic human 523 00:29:03,320 --> 00:29:07,197 intelligence which creates a whole bunch of utility for us. 524 00:29:07,280 --> 00:29:12,317 But it's patterns of reasoning are not the same as ours, right? And people 525 00:29:12,400 --> 00:29:14,557 frequently encounter that and they go, "Oh, that's scary." You're like, "Well, 526 00:29:14,640 --> 00:29:17,717 no, actually that's that's Look, you have to pay attention to make sure 527 00:29:17,800 --> 00:29:21,197 there's alignment, but it's also wonderful because it enables new 528 00:29:21,280 --> 00:29:24,957 things." And so that's part of the thing like the kinds of chemistry we're doing 529 00:29:25,040 --> 00:29:27,917 like the humans hadn't discovered this chemistry at Manus, 530 00:29:28,000 --> 00:29:31,677 right? What are the other kind of thinkings about this pattern of 531 00:29:31,760 --> 00:29:37,317 reasoning, right? That that that kind of amplify human capability. 532 00:29:37,400 --> 00:29:38,677 >> It's a great one. Like, you know, I've 533 00:29:38,760 --> 00:29:41,837 not thought that deeply about sort of 534 00:29:41,920 --> 00:29:44,477 how it improves the I mean, if you sort 535 00:29:44,560 --> 00:29:46,797 of sort of say there's, you know, 536 00:29:46,880 --> 00:29:51,037 induction and deduction that we do. 537 00:29:51,120 --> 00:29:54,717 Being in this cognitive loop with an AI, 538 00:29:54,800 --> 00:29:55,717 um 539 00:29:55,800 --> 00:29:58,477 that is fascinating, right? Because my 540 00:29:58,560 --> 00:30:02,397 exploration space is changing. Um 541 00:30:02,480 --> 00:30:06,277 right? The you know, like even in coding, 542 00:30:06,360 --> 00:30:09,037 one of the most fascinating 543 00:30:09,120 --> 00:30:13,477 um things that one of my colleagues introduced me to is he he wrote a new 544 00:30:13,560 --> 00:30:15,837 skill which I have in GitHub Copilot. 545 00:30:15,920 --> 00:30:18,277 It's called cognitive coverage, right? 546 00:30:18,360 --> 00:30:23,157 It's fascinating. What he's done is you know, whenever an agent does some 547 00:30:23,240 --> 00:30:26,477 work for me, he says, "Just like how we have test 548 00:30:26,560 --> 00:30:29,237 coverage, we now have this new concept 549 00:30:29,320 --> 00:30:31,957 called cognitive coverage where we we as 550 00:30:32,040 --> 00:30:35,877 humans are going to learn from what it 551 00:30:35,960 --> 00:30:36,917 did." 552 00:30:37,000 --> 00:30:39,677 And so it's just creates a quiz. 553 00:30:39,760 --> 00:30:44,397 So because I literally think of learning 554 00:30:44,480 --> 00:30:47,317 uh from and it's sort of it is me 555 00:30:47,400 --> 00:30:50,357 matching my ability 556 00:30:50,440 --> 00:30:52,277 to essentially form a deductive 557 00:30:52,360 --> 00:30:55,997 understanding in some sense um of what 558 00:30:56,080 --> 00:30:59,957 an agent did. Right? So, to your point, it is um 559 00:31:00,040 --> 00:31:01,917 it's a basically it's mimicking. It's 560 00:31:02,000 --> 00:31:04,397 sort of an imitation game on one side. 561 00:31:04,480 --> 00:31:07,997 But, that imitation game does create so 562 00:31:08,080 --> 00:31:11,077 then my ability to then deductively 563 00:31:11,160 --> 00:31:15,037 understand it I think is probably one of the more 564 00:31:15,120 --> 00:31:19,077 important human skills we have to develop. So, this cognitive coverage is 565 00:31:19,160 --> 00:31:22,717 I'm like fascinated by it, right? Which is oh I think we're going to have 566 00:31:22,800 --> 00:31:25,797 something like that, which is the agents are working. They have to be aligned on 567 00:31:25,880 --> 00:31:28,357 one side. But, the other side of it is 568 00:31:28,440 --> 00:31:31,597 what is new will be us knowing, oh we've 569 00:31:31,680 --> 00:31:33,837 cognitively covered what AI did. 570 00:31:33,920 --> 00:31:38,397 >> Yeah. Well, I think that point intersected with your earlier one about, 571 00:31:38,480 --> 00:31:42,917 you know, kind of token capital now AI capital, is that part of the skill set 572 00:31:43,000 --> 00:31:45,037 for the modern you know, kind of human 573 00:31:45,120 --> 00:31:47,757 knowledge worker is to say, okay, how do 574 00:31:47,840 --> 00:31:52,317 I strategize on the use of AI canvas orchestration? 575 00:31:52,400 --> 00:31:55,397 >> Yeah. >> But, also within kind of capital 576 00:31:55,480 --> 00:31:58,637 allocation bounds. Right? Because part of what we've seen is people can go 577 00:31:58,720 --> 00:32:01,117 crazy on spending lots of tokens in 578 00:32:01,200 --> 00:32:05,717 ineffective ways. It's like, you know, buying, you know, go back to bad 579 00:32:05,800 --> 00:32:08,957 Pentagon days, toilet seats for a million dollars. And so, putting your 580 00:32:09,040 --> 00:32:12,357 thinking together with the cognition together with 581 00:32:12,440 --> 00:32:14,757 token management. And what do you think 582 00:32:14,840 --> 00:32:18,157 is part of the question around this 583 00:32:18,240 --> 00:32:22,117 skill of blending the cognitive coverage 584 00:32:22,200 --> 00:32:25,517 with also token intelligence in terms of 585 00:32:25,600 --> 00:32:27,597 how do you amplify the outcomes? 586 00:32:27,680 --> 00:32:31,637 >> one. In fact, I've been thinking about like one of the things that um 587 00:32:31,720 --> 00:32:35,077 I think one has to study deeply is what are all the things that become more 588 00:32:35,160 --> 00:32:39,637 valuable in the age of let's call it token abundance. 589 00:32:39,720 --> 00:32:42,037 One is what you reference, which is how 590 00:32:42,120 --> 00:32:43,957 do you use tokens becomes very valuable. 591 00:32:44,040 --> 00:32:47,077 Whoever figures out that oh, I can use 592 00:32:47,160 --> 00:32:51,677 tokens more efficiently for an outcome that matters in the world is going to 593 00:32:51,760 --> 00:32:55,237 get ahead, right? By definition. So, then how does one build the intuition 594 00:32:55,320 --> 00:32:57,677 for it, right? Which is oh, like this is 595 00:32:57,760 --> 00:33:00,197 where what is that outcome? How do I 596 00:33:00,280 --> 00:33:02,717 measure it? What's a rubric? In fact, 597 00:33:02,800 --> 00:33:05,997 this is where I think the eval the you 598 00:33:06,080 --> 00:33:08,277 know, if it's it's fascinating. 599 00:33:08,360 --> 00:33:12,837 How much time needs to be spent, especially for an RL regime, I think 600 00:33:12,920 --> 00:33:14,637 it's been the clearest where if you 601 00:33:14,720 --> 00:33:16,917 really want you want to set up the 602 00:33:17,000 --> 00:33:20,877 rubric um and the eval dimensions or the 603 00:33:20,960 --> 00:33:23,237 rubrics scoring dimensions such that 604 00:33:23,320 --> 00:33:26,837 they're really capturing the high taste 605 00:33:26,920 --> 00:33:31,797 that you only can define. Because if you did that, then you could be token 606 00:33:31,880 --> 00:33:33,477 efficient, right? In fact, the other 607 00:33:33,560 --> 00:33:36,637 side of the token efficiency is um in 608 00:33:36,720 --> 00:33:39,037 some of the examples we even showed at 609 00:33:39,120 --> 00:33:40,797 build was 610 00:33:40,880 --> 00:33:44,077 let's say you have a um I don't know um 611 00:33:44,160 --> 00:33:48,797 some you're a retailer or uh you're a package goods company and one of the 612 00:33:48,880 --> 00:33:53,157 things you do is handle sort of transactions around 613 00:33:53,240 --> 00:33:56,917 uh trade promotions. Right now, you get all these claims from 614 00:33:57,000 --> 00:34:01,037 all the retailers and you have to process them and the humans are doing it. 615 00:34:01,120 --> 00:34:05,117 You can easily automate that using an agentic workflow. 616 00:34:05,200 --> 00:34:08,957 And you could say oh, I use a frontier model for it. But then one of the 617 00:34:09,040 --> 00:34:13,757 dictums I have is don't use frontier models for non-frontier problems, right? 618 00:34:13,840 --> 00:34:15,477 So, this is not a frontier problem. 619 00:34:15,560 --> 00:34:16,917 You're not trying to discover some new 620 00:34:17,000 --> 00:34:19,757 material science. This is a repeatable 621 00:34:19,840 --> 00:34:22,517 deterministic workflow, 622 00:34:22,600 --> 00:34:25,437 uh but that can benefit from all the 623 00:34:25,520 --> 00:34:30,037 mistakes humans can make and claims and so on. So, therefore you do need uh you 624 00:34:30,120 --> 00:34:32,437 know, intelligence in it, but you can 625 00:34:32,520 --> 00:34:35,957 take, you know, a model like an MAI 5B 626 00:34:36,040 --> 00:34:39,597 and use the traces to hill climb in your 627 00:34:39,680 --> 00:34:43,437 RLE to perform even outperform even a 628 00:34:43,520 --> 00:34:46,277 frontier prompted. Right, so that to me 629 00:34:46,360 --> 00:34:49,477 is token efficiency. So that type of 630 00:34:49,560 --> 00:34:53,557 human understanding of both the limits 631 00:34:53,640 --> 00:34:58,757 of the system, the characteristics of the system, I think become high premium 632 00:34:58,840 --> 00:35:02,637 at this stage. >> Yeah, I agree. One of the other things that 633 00:35:02,720 --> 00:35:06,637 I think is important for people to grok, as you know, cuz we've been in a lot of 634 00:35:06,720 --> 00:35:10,797 these conversations, there's a lot of concern in other countries about, you 635 00:35:10,880 --> 00:35:15,637 know, kind of sovereign AI. And I think that it may be useful here 636 00:35:15,720 --> 00:35:19,757 given we've just kind of gone in a depth about companies and the sovereignty, 637 00:35:19,840 --> 00:35:22,797 like the companies, their information, their employees. 638 00:35:22,880 --> 00:35:26,317 What are the parallels between what we're doing for companies and how 639 00:35:26,400 --> 00:35:31,637 countries should think about kind of also engaging and trust in the AI future? 640 00:35:31,720 --> 00:35:33,397 >> Yeah, I mean this is interesting because I I 641 00:35:33,480 --> 00:35:36,197 one of the reasons why I have really 642 00:35:36,280 --> 00:35:40,997 pivoted to companies is because companies exist all over the world. 643 00:35:41,080 --> 00:35:44,677 Right, that's the good news here, which is because in some sense given countries 644 00:35:44,760 --> 00:35:47,837 thinking abstractly about sovereignty, 645 00:35:47,920 --> 00:35:50,557 they can even make big mistakes where 646 00:35:50,640 --> 00:35:53,917 they ultimately could erode their 647 00:35:54,000 --> 00:35:56,597 comparative advantage that is naturally 648 00:35:56,680 --> 00:35:59,437 there today embodied in the commercial 649 00:35:59,520 --> 00:36:04,277 activity happening in the country through its company formation and 650 00:36:04,360 --> 00:36:05,917 thriving. 651 00:36:06,000 --> 00:36:09,477 So therefore preserving that is the best 652 00:36:09,560 --> 00:36:13,037 form of sovereignty. Sometimes I think people think, oh, if only I had a 653 00:36:13,120 --> 00:36:17,197 firewall or all the data was resident or what have you. 654 00:36:17,280 --> 00:36:22,117 You know, I'm not saying those are not considerations, but this is not about 655 00:36:22,200 --> 00:36:26,037 even any of that, right, which is this is 656 00:36:26,120 --> 00:36:29,477 about making sure that you have an economy. 657 00:36:29,560 --> 00:36:32,237 Uh that means you need to have companies. That means that those 658 00:36:32,320 --> 00:36:35,357 companies have to thrive in a token economy. 659 00:36:35,440 --> 00:36:39,037 Uh that means they need to be able to build that IP. And so therefore, in 660 00:36:39,120 --> 00:36:41,477 fact, having partnerships even with 661 00:36:41,560 --> 00:36:44,397 companies outside that give you the 662 00:36:44,480 --> 00:36:47,437 ability, I think are more important than 663 00:36:47,520 --> 00:36:51,557 sort of suddenly falling behind some frontier, right? And so, I think that 664 00:36:51,640 --> 00:36:54,517 therefore, I think that this is one of those places where again, even the 665 00:36:54,600 --> 00:36:58,677 policy makers have to think about infrastructure. 666 00:36:58,760 --> 00:37:03,077 Right? After all, you know, tokens, they're electrons on one end and tokens 667 00:37:03,160 --> 00:37:04,917 on the other end. So, you want to be the 668 00:37:05,000 --> 00:37:07,957 cheapest, best, and environmentally good 669 00:37:08,040 --> 00:37:13,237 producers of those electrons to token conversion machines called data centers, 670 00:37:13,320 --> 00:37:16,677 right? So, that one I think is going to be very important for every country to 671 00:37:16,760 --> 00:37:19,837 prioritize. Uh of course, the country should even 672 00:37:19,920 --> 00:37:23,317 prioritize whether that's their own semiconductor production and what have 673 00:37:23,400 --> 00:37:25,477 you, right? That's are all very valid 674 00:37:25,560 --> 00:37:27,997 things to do. But beyond that, I think 675 00:37:28,080 --> 00:37:30,837 the most important thing is to take what 676 00:37:30,920 --> 00:37:33,357 Ricardo was always right on, which is 677 00:37:33,440 --> 00:37:36,157 countries have by definition their own 678 00:37:36,240 --> 00:37:38,597 comparative advantage, and now they need 679 00:37:38,680 --> 00:37:43,957 to amplify that using AI. That means best thing is that whether it's a small 680 00:37:44,040 --> 00:37:47,077 business or whether it's the large multinational or even the public sector 681 00:37:47,160 --> 00:37:51,357 efficiency in the country is getting better 682 00:37:51,440 --> 00:37:54,597 uh and is operating at the frontier. The worst thing for any country to in the 683 00:37:54,680 --> 00:37:56,637 name of sovereignty, 684 00:37:56,720 --> 00:37:59,317 if they're off frontier, then that makes 685 00:37:59,400 --> 00:38:01,877 no sense because you're falling behind. 686 00:38:01,960 --> 00:38:05,997 You have to but at the same time, being dependent on one frontier model also 687 00:38:06,080 --> 00:38:08,637 makes no sense because then you're not sovereign. 688 00:38:08,720 --> 00:38:11,637 So, what's the the way to solve it is to be able to 689 00:38:11,720 --> 00:38:14,597 say, "No, we will use models to hill 690 00:38:14,680 --> 00:38:18,437 climb on our own one firm at a time." 691 00:38:18,520 --> 00:38:20,637 And at an economy at aggregate, that's I 692 00:38:20,720 --> 00:38:22,117 think the equation. 693 00:38:22,200 --> 00:38:25,597 >> The the gesture at Silicon actually uh 694 00:38:25,680 --> 00:38:30,077 I think is a good also bring up both for companies and we are 695 00:38:30,160 --> 00:38:33,917 building Mayan Cobalt, but also of course partnering well with 696 00:38:34,000 --> 00:38:35,437 Nvidia and AMD. 697 00:38:35,520 --> 00:38:38,677 So, first what's the strategic job of 698 00:38:38,760 --> 00:38:43,557 Microsoft's own Silicon? What might companies or countries also learn from 699 00:38:43,640 --> 00:38:45,677 that? >> Yeah, so it's interesting if you think 700 00:38:45,760 --> 00:38:50,317 about it right we want it's a great way to observe it even because I what I mean 701 00:38:50,400 --> 00:38:54,037 if you look at what Nvidia has done or what AMD has done or what Intel has done 702 00:38:54,120 --> 00:38:57,077 they built general purpose technology. 703 00:38:57,160 --> 00:39:00,717 Which you know, like in fact when I look at what we're doing with GPUs today 704 00:39:00,800 --> 00:39:05,517 we're of course using GPUs to do model training model inferencing. But we're 705 00:39:05,600 --> 00:39:08,917 also accelerating other workloads. In fact one of the exciting announcements 706 00:39:09,000 --> 00:39:11,957 at Build was using GPUs to accelerate 707 00:39:12,040 --> 00:39:16,077 Nvidia GPUs to accelerate our fabric data warehouse, right? In fact one of 708 00:39:16,160 --> 00:39:19,597 the things that's happening with the agent take workloads is we need more 709 00:39:19,680 --> 00:39:22,637 more performance on everything. And so 710 00:39:22,720 --> 00:39:27,357 that's a great that's because of the general purpose nature, right? GPUs have 711 00:39:27,440 --> 00:39:31,357 you know, Nvidia has CUDA CUDA can be used as a programming model on top of 712 00:39:31,440 --> 00:39:35,357 those to accelerate a variety of workloads. In fact we're using the older 713 00:39:35,440 --> 00:39:39,277 chips of Nvidia to accelerate and this also works out economically for us, 714 00:39:39,360 --> 00:39:41,517 right? Which is a smart for us to take a 715 00:39:41,600 --> 00:39:46,397 fleet and keep using that fleet over a lifetime where we are not only using it 716 00:39:46,480 --> 00:39:50,797 for some cutting-edge AI, but we're also using it to in fact make an old workload 717 00:39:50,880 --> 00:39:54,757 even better performing, right? That's great for Nvidia great for us. 718 00:39:54,840 --> 00:39:58,317 But that also speaks to what's happened in terms of the new workloads, right? 719 00:39:58,400 --> 00:40:03,117 The new workloads of AI are these data parallel synchronous workloads training 720 00:40:03,200 --> 00:40:05,237 inference as well as these new agent 721 00:40:05,320 --> 00:40:10,357 runtime workloads which have very different call patterns. Very different 722 00:40:10,440 --> 00:40:13,797 and and they're highly constant like there's this they didn't exist 3 years 723 00:40:13,880 --> 00:40:18,477 ago right at any scale and now they exist at scale. So, it'd behooves us to 724 00:40:18,560 --> 00:40:21,437 start thinking whether it's not even before we even get to the 725 00:40:21,520 --> 00:40:25,957 semiconductors. I'm building my data center. My civil engineering is in 726 00:40:26,040 --> 00:40:31,237 force. My cooling system, my mechanical systems. In fact, the DC to AC that they 727 00:40:31,320 --> 00:40:32,357 we're trying to make sure that the 728 00:40:32,440 --> 00:40:36,037 electrons are coming in in kilowatts, uh 729 00:40:36,120 --> 00:40:40,797 right? Hundreds of kilowatts straight to the silicon without any losses, right? 730 00:40:40,880 --> 00:40:42,197 So, we're trying to minimize even the 731 00:40:42,280 --> 00:40:44,957 power distribution in a data center. So, 732 00:40:45,040 --> 00:40:49,997 we can optimize to the nth degree for these workloads because they're at such 733 00:40:50,080 --> 00:40:53,157 scale. And that's what we're doing with Maya, right? Maya, for example, is being 734 00:40:53,240 --> 00:40:57,477 co-designed with our MAI models and the OpenAI models to get the best 735 00:40:57,560 --> 00:40:59,597 performance out of them. We're designing 736 00:40:59,680 --> 00:41:03,037 Cobalt, which is our arm-based uh core 737 00:41:03,120 --> 00:41:05,437 uh for compute, and we're designing it 738 00:41:05,520 --> 00:41:10,357 for, for example, using all the agentic traces uh of GitHub, right? Coding, you 739 00:41:10,440 --> 00:41:12,317 know, the call pattern of a coding agent 740 00:41:12,400 --> 00:41:15,437 is pretty different than human apps or 741 00:41:15,520 --> 00:41:19,397 even asynchronous human apps. And so, therefore, we're optimizing and getting 742 00:41:19,480 --> 00:41:23,637 massive latency gains, performance gains, and what have you. And so, 743 00:41:23,720 --> 00:41:25,437 we're going to be a systems company that 744 00:41:25,520 --> 00:41:27,837 continuously to optimizes for the new 745 00:41:27,920 --> 00:41:30,797 at-scale workloads that we have while 746 00:41:30,880 --> 00:41:35,237 using general-purpose technology from our partners uh to maximize the utility 747 00:41:35,320 --> 00:41:37,957 of those. And that flexibility, by the way, 748 00:41:38,040 --> 00:41:41,157 uh I think is what really is good. And that that's where I think your question 749 00:41:41,240 --> 00:41:44,517 was so good, which is countries should think of that, right? Which is if you 750 00:41:44,600 --> 00:41:47,717 really think of there's one thing that answers, right? If I said, "Oh, 751 00:41:47,800 --> 00:41:50,397 everything is Maya and everything is Cobalt," that's probably not the right 752 00:41:50,480 --> 00:41:54,397 thing for Microsoft. Uh but at the same time, if you said we're open to 753 00:41:54,480 --> 00:41:57,837 innovation from the outside, we will innovate inside, we will in fact 754 00:41:57,920 --> 00:42:01,357 benchmark everything, we'll be principled about it ultimately for 755 00:42:01,440 --> 00:42:05,037 better economics. >> Yes, cuz cuz creating the efficient 756 00:42:05,120 --> 00:42:09,157 capital token factories is the create human prosperity is the goal. 757 00:42:09,240 --> 00:42:12,437 >> Exactly. And enabling many of companies to do it. So, 758 00:42:12,520 --> 00:42:15,797 uh different thread, um one of the things that I know from various 759 00:42:15,880 --> 00:42:18,397 conversations with you that you're very thoughtful on 760 00:42:18,480 --> 00:42:20,917 is the topic of children in the modern 761 00:42:21,000 --> 00:42:25,717 age because obviously we're creating products for tutors, co-workers, etc., 762 00:42:25,800 --> 00:42:28,637 etc. And one of the things we're all 763 00:42:28,720 --> 00:42:30,877 exploring together is what should we be 764 00:42:30,960 --> 00:42:35,597 doing as an industry to navigate kind of 765 00:42:35,680 --> 00:42:39,117 being of helping children elevate the right 766 00:42:39,200 --> 00:42:43,357 way and also keeping them safe. What are some of the principles that you know, 767 00:42:43,440 --> 00:42:47,077 Microsoft thinking about, you're thinking about as ways to kind of be 768 00:42:47,160 --> 00:42:51,757 good as it regards, you know, the next generation of humanity through children. 769 00:42:51,840 --> 00:42:55,317 >> Yeah, it's a great question. I you know, the the there's 770 00:42:55,400 --> 00:42:57,997 obviously important things that we have 771 00:42:58,080 --> 00:43:01,077 to do around child safety, right? So, 772 00:43:01,160 --> 00:43:04,957 when any digital technology comes out, I 773 00:43:05,040 --> 00:43:06,917 think we have to now think as first 774 00:43:07,000 --> 00:43:09,997 class, right? There are AI safety issues 775 00:43:10,080 --> 00:43:15,117 around cyber, bio-weapons, alignment, but it's also child safety. So, I think 776 00:43:15,200 --> 00:43:16,837 that that one thing that we want to make 777 00:43:16,920 --> 00:43:20,037 sure is some of the challenges 778 00:43:20,120 --> 00:43:22,677 uh of uh current sort of set of let's 779 00:43:22,760 --> 00:43:25,117 call it chatbots, right? 780 00:43:25,200 --> 00:43:28,197 And their conversations with children in 781 00:43:28,280 --> 00:43:29,997 particular is something that we need to 782 00:43:30,080 --> 00:43:32,237 be very mindful of and make sure that we 783 00:43:32,320 --> 00:43:35,877 are not uh having children uh not have 784 00:43:35,960 --> 00:43:38,997 that agency uh that they need to have in 785 00:43:39,080 --> 00:43:43,917 order to interact with this on their terms versus um be persuaded, for 786 00:43:44,000 --> 00:43:46,877 example, right? So, those are very important things. 787 00:43:46,960 --> 00:43:50,437 But, the other thing that you're pulling on, which I think is a super important, 788 00:43:50,520 --> 00:43:52,757 is what does it mean to even be a child 789 00:43:52,840 --> 00:43:57,597 in a world like this where there is all this abundance of tokens. 790 00:43:57,680 --> 00:44:00,517 And that I think is the the question, right? How do children what should how 791 00:44:00,600 --> 00:44:03,717 should learning happen? Uh 792 00:44:03,800 --> 00:44:06,197 how should we inspire 793 00:44:06,280 --> 00:44:08,597 uh children? What's their sort of 794 00:44:08,680 --> 00:44:14,077 ability to go have a new pedagogical system even that they can enter. Because 795 00:44:14,160 --> 00:44:19,317 I think even like if you take the traditional ways, right? The anxiety 796 00:44:19,400 --> 00:44:24,517 that let's say I I had growing up on learning. 797 00:44:24,600 --> 00:44:26,317 Interestingly enough, 798 00:44:26,400 --> 00:44:28,317 that's a sort of 799 00:44:28,400 --> 00:44:32,437 an artifact of the scarcity 800 00:44:32,520 --> 00:44:35,957 of opportunities, scarcity of good 801 00:44:36,040 --> 00:44:37,917 learning, 802 00:44:38,000 --> 00:44:40,117 and a variety of things which may or may 803 00:44:40,200 --> 00:44:43,157 not be true going forward. So, one of 804 00:44:43,240 --> 00:44:45,677 the first things I think is creating an 805 00:44:45,760 --> 00:44:49,277 learning environment where students and 806 00:44:49,360 --> 00:44:51,757 children in particular from the earliest 807 00:44:51,840 --> 00:44:55,077 of ages don't develop these phobias for 808 00:44:55,160 --> 00:44:57,557 math or science or what have you. In 809 00:44:57,640 --> 00:45:00,397 fact, because everyone's by definition a 810 00:45:00,480 --> 00:45:03,277 lot more curious. It's just that contact 811 00:45:03,360 --> 00:45:06,157 with the world is what, you know, erodes 812 00:45:06,240 --> 00:45:08,797 that curiosity and confidence that 813 00:45:08,880 --> 00:45:11,277 children have in Italy. How do we 814 00:45:11,360 --> 00:45:14,557 develop that even further by giving them 815 00:45:14,640 --> 00:45:17,477 the ability to explore, knowing that 816 00:45:17,560 --> 00:45:21,637 there's no anxiety because after all the expertise is always going to be there in 817 00:45:21,720 --> 00:45:24,317 abundance. It's really your cognitive 818 00:45:24,400 --> 00:45:28,917 coverage of that expertise that's more at a premium. Right? That's if I had 819 00:45:29,000 --> 00:45:31,277 sort of somebody said that to me 820 00:45:31,360 --> 00:45:33,677 as a five-year-old, I think I may have 821 00:45:33,760 --> 00:45:36,557 approached life very differently. 822 00:45:36,640 --> 00:45:40,877 And so, how do we as a society create the necessary conditions for that? I 823 00:45:40,960 --> 00:45:43,357 think it would be probably very important. And I think it's one of the 824 00:45:43,440 --> 00:45:47,117 things, you know, I know from conversations with you intensely is like 825 00:45:47,200 --> 00:45:51,157 this is one of our responsibilities as a tech industry. Right? This is like it's 826 00:45:51,240 --> 00:45:53,677 a No, no, we can't abrogate it. This is 827 00:45:53,760 --> 00:45:58,517 like given the ubiquity of the AI technology, we have to be responsible 828 00:45:58,600 --> 00:46:00,517 for the next human generation. Exactly. 829 00:46:00,600 --> 00:46:02,477 Both both sort of the unintended 830 00:46:02,560 --> 00:46:05,077 consequences is something we think about 831 00:46:05,160 --> 00:46:08,317 from day one and really building safety 832 00:46:08,400 --> 00:46:11,597 guardrails and what have you. Then the 833 00:46:11,680 --> 00:46:16,957 great advantages of new technology, right, have to be democratized. 834 00:46:17,040 --> 00:46:19,997 And then I think there needs to be structural change, right? I don't think 835 00:46:20,080 --> 00:46:24,517 you can say education remains exactly the same and we value the same 836 00:46:24,600 --> 00:46:27,717 credentials. I don't know. I don't think so. 837 00:46:27,800 --> 00:46:31,197 Something's got to change. >> So one of the things we haven't had a 838 00:46:31,280 --> 00:46:34,877 chance to talk about yet, which is very natural in this context, is Pope Leo's 839 00:46:34,960 --> 00:46:40,077 encyclical. Which I thought was actually in fact a magnificent part of leadership 840 00:46:40,160 --> 00:46:42,997 on the behalf on the behalf of the church. I mean, Pope Francis had 841 00:46:43,080 --> 00:46:45,037 actually gotten me engaged in helping 842 00:46:45,120 --> 00:46:48,517 them talk about AI 10, 11 years ago, 843 00:46:48,600 --> 00:46:54,117 right? So like the church has been amazingly front-footed on this. 844 00:46:54,200 --> 00:46:56,517 And actually part of the thing that didn't surprise me at all in the 845 00:46:56,600 --> 00:46:59,757 encyclical was the humanism of it, right? Cuz people are thought, oh, it's 846 00:46:59,840 --> 00:47:02,997 about religion, it's about how you how you pray and it's like, actually in 847 00:47:03,080 --> 00:47:07,197 fact, the the the at least the parts of the church I've interfaced with have 848 00:47:07,280 --> 00:47:09,957 been humanists, right? And 849 00:47:10,040 --> 00:47:15,157 do you have any reflections on the Pope's encyclical? To see the Pope weigh 850 00:47:15,240 --> 00:47:18,997 in and from what I understand is historical precedent on this, right? I 851 00:47:19,080 --> 00:47:21,117 believe the Pope at the time of the 852 00:47:21,200 --> 00:47:26,077 Industrial Revolution also had weighed in on the condition of labor and what 853 00:47:26,160 --> 00:47:27,797 should it do and what have you. And so 854 00:47:27,880 --> 00:47:31,517 to have the Pope sort of come out in 855 00:47:31,600 --> 00:47:35,077 defense of what I think all of us could 856 00:47:35,160 --> 00:47:39,957 be deeply care about, which is human dignity and human agency 857 00:47:40,040 --> 00:47:41,997 in the age of AI, 858 00:47:42,080 --> 00:47:45,797 I think it's so important and I'm glad 859 00:47:45,880 --> 00:47:48,837 that he has advocated what he thinks is 860 00:47:48,920 --> 00:47:53,397 important. Um the other side of this to 861 00:47:53,480 --> 00:47:57,637 me is what is the society like if I look at 862 00:47:57,720 --> 00:48:00,157 what is at least describing what what 863 00:48:00,240 --> 00:48:02,237 when I think about some of the greatest 864 00:48:02,320 --> 00:48:05,397 ways technological advances were 865 00:48:05,480 --> 00:48:07,477 harnessed 866 00:48:07,560 --> 00:48:12,597 to create great prosperity. The story of the West is pretty 867 00:48:12,680 --> 00:48:16,597 unbelievable, right? There's a beautiful book, in fact, that Joel Mokyr and a 868 00:48:16,680 --> 00:48:20,277 couple of other authors have written called two It's called, I think, Two 869 00:48:20,360 --> 00:48:22,917 Paths to Prosperity, which describes the 870 00:48:23,000 --> 00:48:25,717 thousand-year history of China and the 871 00:48:25,800 --> 00:48:30,597 West, essentially. And you know, I'm paraphrasing uh here, 872 00:48:30,680 --> 00:48:32,237 but fundamentally, some of the 873 00:48:32,320 --> 00:48:35,397 constructs, cultural, societal 874 00:48:35,480 --> 00:48:38,677 constructs in the West on how to use 875 00:48:38,760 --> 00:48:40,477 even the scientific revolution, the 876 00:48:40,560 --> 00:48:43,277 industrial revolution, this and they 877 00:48:43,360 --> 00:48:46,157 that necessitated a real change in how 878 00:48:46,240 --> 00:48:50,877 the society was organized so that it could fundamentally take advantage of 879 00:48:50,960 --> 00:48:54,797 this new uh is sort of really massive. 880 00:48:54,880 --> 00:48:58,597 Uh Right? That in fact, defined the modern world. 881 00:48:58,680 --> 00:49:00,917 And so, one of the things I feel we now 882 00:49:01,000 --> 00:49:04,237 need is a similar coming together from 883 00:49:04,320 --> 00:49:06,517 sort of both the moral philosophy. I 884 00:49:06,600 --> 00:49:08,477 think the Pope basically sort of said, 885 00:49:08,560 --> 00:49:11,317 "This needs to be the moral philosophy 886 00:49:11,400 --> 00:49:14,797 that guides uh us going forward." You 887 00:49:14,880 --> 00:49:17,437 combine that with even what is the 888 00:49:17,520 --> 00:49:21,997 market, what is the what is democracy, uh and then what is the 889 00:49:22,080 --> 00:49:23,597 scientific/technological 890 00:49:23,680 --> 00:49:25,717 revolution. So, if we can somehow get 891 00:49:25,800 --> 00:49:30,037 into a virtuous cycle where the morality 892 00:49:30,120 --> 00:49:31,757 and the scientific sort of 893 00:49:31,840 --> 00:49:35,117 breakthroughs, the political system, 894 00:49:35,200 --> 00:49:39,197 and the markets are all reinforcing each 895 00:49:39,280 --> 00:49:42,917 other, then we will have abundance, and 896 00:49:43,000 --> 00:49:47,157 we will have many stakeholders all benefiting. And if you don't, we are 897 00:49:47,240 --> 00:49:48,717 going to lose social permission, right? 898 00:49:48,800 --> 00:49:50,437 So, I just don't think I I this is what 899 00:49:50,520 --> 00:49:55,917 the narrow understanding of the success of the West as just a technological 900 00:49:56,000 --> 00:49:57,397 breakthrough 901 00:49:57,480 --> 00:49:59,357 I think is sort of not the case right 902 00:49:59,440 --> 00:50:03,197 it's a unbelievable coming together 903 00:50:03,280 --> 00:50:06,677 of a multitude of sort of forces in a 904 00:50:06,760 --> 00:50:11,197 virtual cycle it also had really bad parts to it 905 00:50:11,280 --> 00:50:15,237 right we know that and so you have to avoid that right which is and so that I 906 00:50:15,320 --> 00:50:18,837 think was also some things that the Pope himself wrote in the encyclical which I 907 00:50:18,920 --> 00:50:22,837 thought was great which is even to think about sort of what are the bad parts how 908 00:50:22,920 --> 00:50:28,477 do we not repeat it what is the way to be you know advocating for this positive 909 00:50:28,560 --> 00:50:32,797 cycle I think it's beautifully captured >> and I thought it was great that it was 910 00:50:32,880 --> 00:50:37,277 kind of a focus on how do you keep humans at the center how do you elevate 911 00:50:37,360 --> 00:50:39,597 human dignity how do you 912 00:50:39,680 --> 00:50:42,117 address needs not just of the wealthy 913 00:50:42,200 --> 00:50:47,157 countries but the entire world is I think you know a great you know beacon 914 00:50:47,240 --> 00:50:48,397 in terms of how we started to think about it 915 00:50:48,480 --> 00:50:51,677 >> you know I think deeply about that right for example I've always 916 00:50:51,760 --> 00:50:56,077 you know you know been like someone who cared deeply about the global south in 917 00:50:56,160 --> 00:51:01,037 some sense finally having their moment where they can be real catch-up growth 918 00:51:01,120 --> 00:51:05,157 and so I think that there's a real danger now in the age of AI for even 919 00:51:05,240 --> 00:51:07,917 that convergence growth to slow and in 920 00:51:08,000 --> 00:51:10,317 fact go the other way and so what are 921 00:51:10,400 --> 00:51:13,277 the again the global structures because 922 00:51:13,360 --> 00:51:17,837 by the way someone sitting in the United States in you know in Palo Alto may 923 00:51:17,920 --> 00:51:21,557 think that somehow it doesn't impact them but there's no such thing right you 924 00:51:21,640 --> 00:51:24,077 know we share this planet and our 925 00:51:24,160 --> 00:51:25,877 destinies are a lot more tied than we 926 00:51:25,960 --> 00:51:28,997 think we are just because of you know 927 00:51:29,080 --> 00:51:33,157 being far away from what's happening >> well um last question before we get to 928 00:51:33,240 --> 00:51:35,317 rapid fire 929 00:51:35,400 --> 00:51:37,757 as you and I both know 930 00:51:37,840 --> 00:51:40,597 you know in the US and in Europe there's 931 00:51:40,680 --> 00:51:44,957 a lot of AI backlash now part of how as you know I've I've 932 00:51:45,040 --> 00:51:48,677 been trying to address it is like with books like Super Agency, trying to say, 933 00:51:48,760 --> 00:51:52,037 "No, no, this is an opportunity to gain agency. The transition will be 934 00:51:52,120 --> 00:51:55,997 difficult, but embracing the agency and the transformation 935 00:51:56,080 --> 00:51:58,877 ultimately you have to, but if you do it 936 00:51:58,960 --> 00:52:03,637 you with forethought and and kind of like leaning into it, it can be greatly 937 00:52:03,720 --> 00:52:05,077 helpful. 938 00:52:05,160 --> 00:52:07,757 Uh what do you think we should be trying 939 00:52:07,840 --> 00:52:10,917 to help people in the US and the West 940 00:52:11,000 --> 00:52:14,197 understand about why it is that actually 941 00:52:14,280 --> 00:52:18,797 embracing AI is more important. And it doesn't mean that there aren't genuine 942 00:52:18,880 --> 00:52:22,397 issues in the backlash and obviously work transition all the rest will be 943 00:52:22,480 --> 00:52:26,757 real issues, but how do we help people see that this could be an important part 944 00:52:26,840 --> 00:52:27,677 of their future? 945 00:52:27,760 --> 00:52:30,557 >> Yeah, it's I think really at least now 946 00:52:30,640 --> 00:52:33,837 I've come to the conclusion that um 947 00:52:33,920 --> 00:52:39,600 there needs to be real tangible 948 00:52:41,840 --> 00:52:46,637 um outcomes that speak for themselves. Because I 949 00:52:46,720 --> 00:52:50,477 think what has happened, and this is one of those places where quite frankly our 950 00:52:50,560 --> 00:52:53,277 industry even, you know, the way we have 951 00:52:53,360 --> 00:52:58,397 talked about it, I mean when you go out and say, "Hey, you know, all economic 952 00:52:58,480 --> 00:53:00,237 opportunity will go away for knowledge 953 00:53:00,320 --> 00:53:02,317 workers." or you know, "White collar 954 00:53:02,400 --> 00:53:05,557 jobs are gone." or you know, I 955 00:53:05,640 --> 00:53:10,037 And and you And then you're saying, "I'm excited about building that technology." 956 00:53:10,120 --> 00:53:13,757 right? Why would anyone want you to be 957 00:53:13,840 --> 00:53:17,877 successful? Right? I mean, I don't want you to be successful. I mean, this just 958 00:53:17,960 --> 00:53:23,277 makes no you know, no social sense. So, I feel 959 00:53:23,360 --> 00:53:26,277 now um like when, you know, when you have 960 00:53:26,360 --> 00:53:28,237 someone in a college commencement be 961 00:53:28,320 --> 00:53:30,797 booed because they're saying AI is a you 962 00:53:30,880 --> 00:53:35,597 know, means we've now crossed over to people don't believe us. 963 00:53:35,680 --> 00:53:39,077 And and rightfully so. So, therefore I think what do we need to do now is time 964 00:53:39,160 --> 00:53:43,197 to do the hard work. The hard work is if you're building a data center, let's 965 00:53:43,280 --> 00:53:47,797 make sure that that community believes that this data center is great for them. 966 00:53:47,880 --> 00:53:50,077 It's the for their tax base, for their 967 00:53:50,160 --> 00:53:52,317 community efforts, their real estate 968 00:53:52,400 --> 00:53:57,237 value, their schools, their water use, right? Their 969 00:53:57,320 --> 00:54:00,877 electricity prices. It can't be again like, "Oh, I said something." No, it has 970 00:54:00,960 --> 00:54:04,517 to be real, right? That's kind of what the way we earn social permission. Same 971 00:54:04,600 --> 00:54:07,237 thing I would say with employment. We 972 00:54:07,320 --> 00:54:11,677 can't abstractly even say, "Hey, lump of labor fallacy, always there's going to 973 00:54:11,760 --> 00:54:14,157 be new jobs." What are the new jobs? 974 00:54:14,240 --> 00:54:16,997 What are the wages of the new jobs? 975 00:54:17,080 --> 00:54:19,517 Uh that I can now go apply for, train 976 00:54:19,600 --> 00:54:24,117 myself for, and and how do I really start a new company or what have you, 977 00:54:24,200 --> 00:54:27,797 right? I think unless we really get clear. All the third thing which we 978 00:54:27,880 --> 00:54:32,877 unpacked a lot, every firm needs to participate in the frontier ecosystem. 979 00:54:32,960 --> 00:54:36,757 Oh, it's not like, "Oh, I'm just a feeder of data to some foundation 980 00:54:36,840 --> 00:54:39,197 model." Like that is like talk about 981 00:54:39,280 --> 00:54:41,877 sovereignty and dignity both being lost, 982 00:54:41,960 --> 00:54:44,397 right? At simultaneously, right? By for 983 00:54:44,480 --> 00:54:46,677 countries, communities, and companies. 984 00:54:46,760 --> 00:54:51,677 So, I think now we have to go all the way and say, okay, this is a positive 985 00:54:51,760 --> 00:54:55,997 sum. This is in fact the you know, challenges 986 00:54:56,080 --> 00:54:57,797 of some of the technology. We are going 987 00:54:57,880 --> 00:55:00,837 to really actively work it. We can 988 00:55:00,920 --> 00:55:05,317 articulate the tangible benefits and this is where again, what you will do at 989 00:55:05,400 --> 00:55:08,837 Manas and others are also going to be very, very helpful because the world 990 00:55:08,920 --> 00:55:11,157 needs more proof points that this 991 00:55:11,240 --> 00:55:16,397 technology is ultimately helping human condition and our societies broadly, not 992 00:55:16,480 --> 00:55:20,157 narrowly. >> Yeah. No, AI for humanity. So, 993 00:55:20,240 --> 00:55:24,437 you don't have to answer rapidly, but we ask the same questions of all our guests. 994 00:55:24,520 --> 00:55:28,517 So, the first one is, is there a movie, song, or book that fills you with 995 00:55:28,600 --> 00:55:32,317 optimism for the future? >> It's this 996 00:55:32,400 --> 00:55:34,357 parallel paths to prosperity, all right? 997 00:55:34,440 --> 00:55:38,597 Because I I the reason why I like that is because it's a good call to action. 998 00:55:38,680 --> 00:55:42,917 For the next thousand years, what's the path to prosperity? There was 999 00:55:43,000 --> 00:55:46,397 a blueprint um at least that at the part of the 1000 00:55:46,480 --> 00:55:49,437 world got it right in the last thousand. 1001 00:55:49,520 --> 00:55:53,797 Can we now as a an entire planet get it 1002 00:55:53,880 --> 00:55:56,957 uh right for the next thousand? I think that this is where 1003 00:55:57,040 --> 00:55:59,797 uh some of our very best work has to be done. 1004 00:55:59,880 --> 00:56:04,877 >> Agreed. What's a question that you wish people would ask you more often? 1005 00:56:04,960 --> 00:56:08,517 >> Ah, that's a good one. I would love for people to ask me what am I not excited 1006 00:56:08,600 --> 00:56:11,997 about. Because I'm excited about a lot of things. 1007 00:56:12,080 --> 00:56:16,277 Uh but I'm not excited, for example, about us losing permission on 1008 00:56:16,360 --> 00:56:20,397 AI by you know, saying all the wrong things or doing the wrong things even 1009 00:56:20,480 --> 00:56:25,517 and not having a a complete thought on how to truly have a positive sum 1010 00:56:25,600 --> 00:56:29,757 construct. >> Yep. 100%. So, where do you see progress or 1011 00:56:29,840 --> 00:56:34,597 momentum outside of your industry? And given that Microsoft powers a lot of the 1012 00:56:34,680 --> 00:56:37,797 world's industry, um but you know, 1013 00:56:37,880 --> 00:56:41,997 uh it could be robotics, AI, etc. But where do you see progress that inspires 1014 00:56:42,080 --> 00:56:46,597 you? >> I mean I mean to some degree you know, 1015 00:56:46,680 --> 00:56:48,677 the work in bio I mean, if I think about 1016 00:56:48,760 --> 00:56:51,237 it, right? One of the most complex 1017 00:56:51,320 --> 00:56:54,797 systems um that we have to have a better 1018 00:56:54,880 --> 00:56:57,077 understanding of is human biology. So, 1019 00:56:57,160 --> 00:57:01,717 anything any tool that can help humans 1020 00:57:01,800 --> 00:57:04,917 take care of humans is probably the 1021 00:57:05,000 --> 00:57:08,237 thing that will have uh it's it's 1022 00:57:08,320 --> 00:57:10,357 awe-inspiring. There's this one thing uh 1023 00:57:10,440 --> 00:57:12,637 we I recently came across, right? Which 1024 00:57:12,720 --> 00:57:15,357 is this for immunotherapy, 1025 00:57:15,440 --> 00:57:17,997 I believe there's a test um 1026 00:57:18,080 --> 00:57:20,357 that is uh you know, basically a a 1027 00:57:20,440 --> 00:57:23,397 complex costly test that figures out 1028 00:57:23,480 --> 00:57:25,677 whether a particular immunotherapy will 1029 00:57:25,760 --> 00:57:28,637 work on that tumor or not. Um and so, 1030 00:57:28,720 --> 00:57:31,797 Providence and uh some researchers at U 1031 00:57:31,880 --> 00:57:34,477 Dub and Microsoft Research came together 1032 00:57:34,560 --> 00:57:36,997 and built this thing called Giga time, 1033 00:57:37,080 --> 00:57:39,357 which is a cool model, which basically 1034 00:57:39,440 --> 00:57:44,277 is a simulation of that task and reduces the cost of what could get done only at 1035 00:57:44,360 --> 00:57:48,277 sort of, you know, for took a lot of time and a lot of money, now can be done 1036 00:57:48,360 --> 00:57:51,117 by any tertiary sort of hospital in a, 1037 00:57:51,200 --> 00:57:54,597 you know, in any city. And that type of 1038 00:57:54,680 --> 00:57:57,517 economic sort of availability of sort of 1039 00:57:57,600 --> 00:58:01,957 medicine in and and medical practice, I think is just breakthrough. 1040 00:58:02,040 --> 00:58:05,597 >> So, last question. Can you leave us with a final thought on what you think is 1041 00:58:05,680 --> 00:58:10,077 possible to achieve if everything breaks humanity's way in the next 15 years? And 1042 00:58:10,160 --> 00:58:12,717 what's our first step? >> Mhm. 1043 00:58:12,800 --> 00:58:15,117 I think if 1044 00:58:15,200 --> 00:58:17,237 everything breaks our way, I've always 1045 00:58:17,320 --> 00:58:19,797 gone back like what's the dream 1046 00:58:19,880 --> 00:58:22,877 um of the world 1047 00:58:22,960 --> 00:58:27,037 compounding at 10% GDP growth. 1048 00:58:27,120 --> 00:58:31,637 Right? Um that is what can happen, right? You mean 1049 00:58:31,720 --> 00:58:35,717 one thought experiment is if the industrial revolution had reached all 1050 00:58:35,800 --> 00:58:39,437 corners of the world at the same time. 1051 00:58:39,520 --> 00:58:43,157 And every country could express their comparative 1052 00:58:43,240 --> 00:58:46,117 advantage fully. 1053 00:58:46,200 --> 00:58:51,117 Right? That's the maximalist positive sum construct. 1054 00:58:51,200 --> 00:58:55,277 Can we do it? Right? Because we are kind of captive to this, you know, what we 1055 00:58:55,360 --> 00:58:59,237 will describe as, "Hey, that's not how the world works. History is about 1056 00:58:59,320 --> 00:59:02,277 dominance and dominant powers." And And 1057 00:59:02,360 --> 00:59:06,437 I'm not saying we will defy all that, but the bottom line is you should since 1058 00:59:06,520 --> 00:59:09,197 you asked me to dream, I'm dreaming that 1059 00:59:09,280 --> 00:59:11,437 if humanity can get back past their 1060 00:59:11,520 --> 00:59:14,317 bounded rationality and only sort of 1061 00:59:14,400 --> 00:59:16,837 say, "Oh, history repeats itself and we 1062 00:59:16,920 --> 00:59:21,797 are only never going to get better than, you know, I know fighting these wars and 1063 00:59:21,880 --> 00:59:25,437 and what have you." And say, "No, no, no, let's maybe we can change the 1064 00:59:25,520 --> 00:59:28,797 course." Uh then the first step would be 1065 00:59:28,880 --> 00:59:33,717 to accept that that possibility exists is what I would say is what will get us 1066 00:59:33,800 --> 00:59:37,557 down. Otherwise, we're going to back to re-litigating, oh, we are Let me look at 1067 00:59:37,640 --> 00:59:40,357 history and then just try to match. 1068 00:59:40,440 --> 00:59:43,917 >> Beautiful dream. Sateer, always a pleasure, and I'll see you next week at 1069 00:59:44,000 --> 00:59:47,117 the board meeting. >> Thank you so much, Reid. Thank you so much. 1070 00:59:47,200 --> 00:59:50,157 >> This is [laughter] awesome. Such a fun >> Yes, >> conversation. 1071 00:59:50,240 --> 00:59:53,197 >> Yes, exactly. >> Possible is produced by Pallet Media. 1072 00:59:53,280 --> 00:59:57,877 It's hosted by Ara Finger and me, Reid Hoffman. Our show runner is Shawn Young. 1073 00:59:57,960 --> 00:59:59,837 Possible is produced by Tanaz Deelo, 1074 00:59:59,920 --> 01:00:02,197 Katie Sanders, Spencer Strassmore, Imo 1075 01:00:02,280 --> 01:00:05,237 Zoo, Aman Souri, Lexi Kevin, Danny 1076 01:00:05,320 --> 01:00:09,397 Garrison, Trent Barbosa, and Tafadzwa Nemarundwe. 1077 01:00:09,480 --> 01:00:11,637 >> Special thanks to Surya Yalamanchili, 1078 01:00:11,720 --> 01:00:18,560 Saida Sabiyeva, Ian Ellis, Greg Biato, Parth Patel, and Ben Ralis.