1 00:00:01,200 --> 00:00:05,837 has generated $250 [music] billion with a B in market value for Microsoft. Scott 2 00:00:05,920 --> 00:00:07,597 Nadella, chairman and CEO of Microsoft. 3 00:00:07,680 --> 00:00:08,957 >> Since you've been the CEO, three and a 4 00:00:09,040 --> 00:00:11,037 half years, the [music] stock is up 5 00:00:11,120 --> 00:00:14,717 about uh I guess it's about 120%. I'm 6 00:00:14,800 --> 00:00:18,917 good for my 80 billion. I am going to spend $80 billion building out Azure. 7 00:00:19,000 --> 00:00:22,317 [music] Maybe after the industrial revolution, this is the biggest thing. 8 00:00:22,400 --> 00:00:24,397 That's our goal with our frontier model. 9 00:00:24,480 --> 00:00:26,317 Our model should be the [music] best 10 00:00:26,400 --> 00:00:31,437 model that they can use as a base. We create technology so that others can 11 00:00:31,520 --> 00:00:33,277 create more technology. That's who we 12 00:00:33,360 --> 00:00:35,197 are. [music] We're tool maker. 13 00:00:35,280 --> 00:00:38,957 >> Please welcome Satia Nadella. 14 00:00:39,040 --> 00:00:42,120 >> All right. 15 00:00:43,360 --> 00:00:48,237 >> Hi guy. Good to see you coming out. 16 00:00:48,320 --> 00:00:52,717 >> Good to see you. >> Good morning guys. 17 00:00:52,800 --> 00:00:55,837 >> How are you? >> Good. >> Thanks for joining us. 18 00:00:55,920 --> 00:01:00,177 >> Crazy weekend, but here we are. Do we need to paste the frontier? 19 00:01:00,260 --> 00:01:02,280 [laughter] 20 00:01:03,120 --> 00:01:06,077 >> So, let's start with the common sense 21 00:01:06,160 --> 00:01:10,397 part first, which is we should do what 22 00:01:10,480 --> 00:01:13,757 it takes to build stuff that serves 23 00:01:13,840 --> 00:01:17,437 humanity first and is in human control. 24 00:01:17,520 --> 00:01:20,317 You know, it's kind of crazy that we have to start with that level of common 25 00:01:20,400 --> 00:01:22,957 sense, but I think it's a good place. 26 00:01:23,040 --> 00:01:26,557 Then when I think about pacing whatever 27 00:01:26,640 --> 00:01:29,597 the first thing that at least I believe 28 00:01:29,680 --> 00:01:32,077 is the broad diffusion of this 29 00:01:32,160 --> 00:01:34,717 technology is the most critical thing 30 00:01:34,800 --> 00:01:37,677 because the benefits of this tech 31 00:01:37,760 --> 00:01:41,197 showing up everywhere is really what's 32 00:01:41,280 --> 00:01:45,037 all about right so at the end of the day if you sort of say serving humanity let 33 00:01:45,120 --> 00:01:47,677 it actually reach humanity in ways that 34 00:01:47,760 --> 00:01:52,237 it serves humanity and that means you got to have choice you have to have 35 00:01:52,320 --> 00:01:55,757 compet competition. You have to have all kinds of business models whether they're 36 00:01:55,840 --> 00:01:57,357 open weights, close weights, what have 37 00:01:57,440 --> 00:02:00,957 you. Then the other aspect I think that 38 00:02:01,040 --> 00:02:03,197 is not talked about when we talk about 39 00:02:03,280 --> 00:02:06,797 control is actually the control that for 40 00:02:06,880 --> 00:02:09,917 example customers have, enterprises or 41 00:02:10,000 --> 00:02:12,397 businesses have around this technology 42 00:02:12,480 --> 00:02:14,637 because sometimes this is so opaque, 43 00:02:14,720 --> 00:02:17,437 right? I want my privacy. I want to be 44 00:02:17,520 --> 00:02:19,677 able to embed my knowledge in a set of 45 00:02:19,760 --> 00:02:22,797 weights I control. I want to see all of 46 00:02:22,880 --> 00:02:25,837 the coot uh that's being generated. I 47 00:02:25,920 --> 00:02:30,717 want to use it to do fine-tuning of my own models. My IP shouldn't leak. So, 48 00:02:30,800 --> 00:02:35,197 there's an entire body of things that nobody's talking about as much, which is 49 00:02:35,280 --> 00:02:37,437 my I really want to make sure that this 50 00:02:37,520 --> 00:02:41,677 tech is in my control. Then we get to uh 51 00:02:41,760 --> 00:02:45,837 what is I think a real issue of safety 52 00:02:45,920 --> 00:02:49,037 and we should take it seriously which is 53 00:02:49,120 --> 00:02:51,837 we should take all the time we want uh 54 00:02:51,920 --> 00:02:56,957 to test things. In fact I love this idea of having third party testers. Oh wow. 55 00:02:57,040 --> 00:03:00,957 You know I you know I grew up in a company that's always done testing. Uh 56 00:03:01,040 --> 00:03:05,037 so it's novel that we should say wow they're having embedded third party 57 00:03:05,120 --> 00:03:07,917 testers. Why not? It's a great idea. In 58 00:03:08,000 --> 00:03:09,677 fact, the only thing I would say is we 59 00:03:09,760 --> 00:03:12,317 should avoid like these, you know, cozy 60 00:03:12,400 --> 00:03:16,717 arrangements of who's testing what, who has access to what, and it should be 61 00:03:16,800 --> 00:03:20,077 broad. >> Were you were you surprised though when 62 00:03:20,160 --> 00:03:23,037 both the essay landed and then it seemed 63 00:03:23,120 --> 00:03:25,437 like there was a circling of the wagons 64 00:03:25,520 --> 00:03:27,117 amongst the frontier companies? I 65 00:03:27,200 --> 00:03:30,717 >> I I think that it comes my my suspicion 66 00:03:30,800 --> 00:03:35,597 is it comes genuinely from this place where when you start seeing in fact it's 67 00:03:35,680 --> 00:03:40,317 fascinating, right? We are when you start seeing reward hacking um and 68 00:03:40,400 --> 00:03:42,237 what's happening in these environments 69 00:03:42,320 --> 00:03:45,277 right with these agent swarms there is 70 00:03:45,360 --> 00:03:50,157 the mundane there is some DevOps error where somebody misconfigured 71 00:03:50,240 --> 00:03:52,957 a container [laughter] >> right right or these API keys 72 00:03:53,040 --> 00:03:57,277 >> or an API keys or yeah exactly there's no monitoring uh there's internet access 73 00:03:57,360 --> 00:03:59,437 there's sort of classic I would call it 74 00:03:59,520 --> 00:04:02,157 basic devops and then there is real 75 00:04:02,240 --> 00:04:04,717 novel new stuff right which is what is 76 00:04:04,800 --> 00:04:09,357 this uh reward board hacking uh that you know with these persistent agents and so 77 00:04:09,440 --> 00:04:13,517 on and that's a place where I'll admit that the science is not there it's I 78 00:04:13,600 --> 00:04:17,757 thought Yakob's post which is a good one which he said he called it we're growing 79 00:04:17,840 --> 00:04:19,837 intelligence not building intelligence 80 00:04:19,920 --> 00:04:22,797 so it's an experimental science and so 81 00:04:22,880 --> 00:04:27,757 the more experimental sciences uh then you really need to make sure you're 82 00:04:27,840 --> 00:04:32,077 doing those experiments in controlled environments if anything the place where 83 00:04:32,160 --> 00:04:34,077 I would love is taking even the hugging 84 00:04:34,160 --> 00:04:39,277 face incident in other places more transparency on what would it take in 85 00:04:39,360 --> 00:04:43,277 fact one of the fascinating things right now is the insider risk I mean think 86 00:04:43,360 --> 00:04:46,397 about it right if you're sitting in an enterprise this is all test time compute 87 00:04:46,480 --> 00:04:49,277 by the way right so it's not like oh it's going to only happen when in some 88 00:04:49,360 --> 00:04:52,237 training run it can happen for a very 89 00:04:52,320 --> 00:04:55,517 mundane task uh that I give one of these 90 00:04:55,600 --> 00:05:00,077 frontier models inside an enterprise uh where I say you know I don't I was you 91 00:05:00,160 --> 00:05:01,757 know telling David this suppose I say 92 00:05:01,840 --> 00:05:04,157 hey go optimize my working capital it 93 00:05:04,240 --> 00:05:08,877 may fake my books uh right [laughter] because this is like a new type of 94 00:05:08,960 --> 00:05:12,557 insider risk >> and so what is the way to do that I 95 00:05:12,640 --> 00:05:16,957 would say oh go build a maybe a causal model like a semantic model that 96 00:05:17,040 --> 00:05:18,877 actually checks and verifies so I think 97 00:05:18,960 --> 00:05:21,997 there's a lot of product building um I I 98 00:05:22,080 --> 00:05:26,797 would say making things more robust which is classic engineering that we 99 00:05:26,880 --> 00:05:30,317 should be talking a lot more about transparently versus saying hey this is 100 00:05:30,400 --> 00:05:33,757 so mystical that you know we can't figure this out. Do you do you buy this 101 00:05:33,840 --> 00:05:35,197 argument that it's mystical? 102 00:05:35,280 --> 00:05:38,797 >> I I mean I I buy the argument that we do 103 00:05:38,880 --> 00:05:43,757 not understand the latent space. Uh right other than I thought you know as 104 00:05:43,840 --> 00:05:47,597 you said like do we understand the brain? We don't. We do functional MRIs 105 00:05:47,680 --> 00:05:51,917 and do neuroscience and we're trying to figure this out continuously getting a 106 00:05:52,000 --> 00:05:54,077 little better understanding. So I do 107 00:05:54,160 --> 00:05:56,237 think that in that sense we don't 108 00:05:56,320 --> 00:05:59,037 exactly uh have a complete un that's why 109 00:05:59,120 --> 00:06:01,917 by the way I also I don't believe in new 110 00:06:02,000 --> 00:06:06,397 release right so that's why I think making sure that the coots are in 111 00:06:06,480 --> 00:06:11,357 language that we can all understand in fact they're transparent so that when 112 00:06:11,440 --> 00:06:15,277 when I go back to an enterprise that's using all these models and if you have 113 00:06:15,360 --> 00:06:18,637 the full coot uh then you can >> chain of thought 114 00:06:18,720 --> 00:06:22,157 >> chain of thought and so then you can really go look at it deeply in fact you 115 00:06:22,240 --> 00:06:26,477 can have multiple models uh and you can look at the coot across those I think 116 00:06:26,560 --> 00:06:28,797 these are all things that I think will become very important 117 00:06:28,880 --> 00:06:30,557 >> satia you've worked with you've worked 118 00:06:30,640 --> 00:06:33,437 with technologists for decades 119 00:06:33,520 --> 00:06:35,997 uh and when you see as a leader of one 120 00:06:36,080 --> 00:06:39,597 company Microsoft which has very crisp 121 00:06:39,680 --> 00:06:42,717 communications with the public uh and 122 00:06:42,800 --> 00:06:44,637 you see what's happening with Daario and 123 00:06:44,720 --> 00:06:47,277 his team people coming out saying 10% 124 00:06:47,360 --> 00:06:50,157 chance we all die uh what do you think 125 00:06:50,240 --> 00:06:55,117 is going through those technologists minds. Do you believe they actually 126 00:06:55,200 --> 00:06:58,797 believe that this is going to kill humanity or are they going through some 127 00:06:58,880 --> 00:07:03,117 psychosis or are they seeing something working on those frontier models that is 128 00:07:03,200 --> 00:07:05,997 terrorizing them? You're not a psychologist, but you have worked with 129 00:07:06,080 --> 00:07:09,997 technologists for a long time. Handicap what's going on in these organizations 130 00:07:10,080 --> 00:07:12,797 that's all the making people feel the 131 00:07:12,880 --> 00:07:15,437 need to resign and say we're all going 132 00:07:15,520 --> 00:07:19,997 to die. >> Yeah. you know, it's it's hard for me to 133 00:07:20,080 --> 00:07:21,917 speak to what's happening in any of 134 00:07:22,000 --> 00:07:25,037 these places, but let let's just say uh 135 00:07:25,120 --> 00:07:29,117 how we I grew up even inside of Microsoft, you know, for example, you 136 00:07:29,200 --> 00:07:32,957 know, one of the biggest things you learn as an early sort of engineering 137 00:07:33,040 --> 00:07:35,997 lead is how to deal with a showstopper bug. 138 00:07:36,080 --> 00:07:38,717 >> Yeah. >> Right. I mean, that's kind of like 101, 139 00:07:38,800 --> 00:07:40,397 right? Which is why you're faced, you're 140 00:07:40,480 --> 00:07:43,677 like, you know, you have a bug. Um what 141 00:07:43,760 --> 00:07:47,277 do you do? do you stop uh and fix or you 142 00:07:47,360 --> 00:07:50,397 defer or you go in and say hey this is 143 00:07:50,480 --> 00:07:52,317 such an edge case that's kind of the 144 00:07:52,400 --> 00:07:57,997 judgment so I do think and as the stakes go up you want to like transaction 145 00:07:58,080 --> 00:08:01,517 processing I remember working on databases right you know wow like you 146 00:08:01,600 --> 00:08:04,077 know you got to take very seriously any 147 00:08:04,160 --> 00:08:07,117 bug uh where if the transaction is going 148 00:08:07,200 --> 00:08:09,597 to get lost right data loss is a thing 149 00:08:09,680 --> 00:08:13,037 that you stop the thing for so I feel a 150 00:08:13,120 --> 00:08:15,197 little bit culture culturally in the AI 151 00:08:15,280 --> 00:08:17,517 industry rediscovering maybe because 152 00:08:17,600 --> 00:08:20,077 when you see and it's possible that they 153 00:08:20,160 --> 00:08:22,957 see stuff which are showstoppers before 154 00:08:23,040 --> 00:08:25,757 the rest and if you see a showstopper 155 00:08:25,840 --> 00:08:29,037 stop the show um right [laughter] to fix 156 00:08:29,120 --> 00:08:32,317 the bugs yeah when you saw the the 157 00:08:32,400 --> 00:08:35,437 hugging face run and it was super 158 00:08:35,520 --> 00:08:39,757 performative Dwaresh did his whole post civilizations 159 00:08:39,840 --> 00:08:44,317 what do you think what what's your take on that testing they ran because they 160 00:08:44,400 --> 00:08:48,477 could have run a test where they had 3,000 agents defend a bunch of websites. 161 00:08:48,560 --> 00:08:53,357 Instead, they instructed them to hack websites and you know the hiding of 162 00:08:53,440 --> 00:08:56,317 information all this anthropomorphicizing 163 00:08:56,400 --> 00:08:58,797 whatever of the agents. I mean the way 164 00:08:58,880 --> 00:09:00,637 at least I understand it was it was 165 00:09:00,720 --> 00:09:03,677 actually you know basically trying to uh 166 00:09:03,760 --> 00:09:08,717 do an eval uh for cyber gym and um as I 167 00:09:08,800 --> 00:09:12,157 understand it given that eval it sort of 168 00:09:12,240 --> 00:09:14,317 figured out a way to say let's just say 169 00:09:14,400 --> 00:09:17,117 reward hack uh and that's what led it to 170 00:09:17,200 --> 00:09:19,917 hugging phase in fact it speaks to I 171 00:09:20,000 --> 00:09:24,477 think what's the pre you know clear issue right now which is you can have 172 00:09:24,560 --> 00:09:28,317 these things if they're are longunning persistent agents 173 00:09:28,400 --> 00:09:30,557 become essentially like new insider 174 00:09:30,640 --> 00:09:35,677 risks. Uh and so that I would start from the very basics of saying okay what is 175 00:09:35,760 --> 00:09:38,877 containment look like. So for example like one of the things that I think is 176 00:09:38,960 --> 00:09:41,197 going to be really an issue and a thing 177 00:09:41,280 --> 00:09:44,557 that needs great solutions is true 178 00:09:44,640 --> 00:09:48,477 aggressive monitoring of agent activity. 179 00:09:48,560 --> 00:09:52,317 Uh that's behavioral >> evidence >> evidence and so everything has got to be 180 00:09:52,400 --> 00:09:56,237 auditable. uh and then every object it access, right? If it goes and gets a 181 00:09:56,320 --> 00:09:59,197 secret, oh, it's going to go chain a couple of things, you should be able to 182 00:09:59,280 --> 00:10:01,597 see it when it's starting to chain a 183 00:10:01,680 --> 00:10:03,917 couple of uh vulnerabilities uh to go 184 00:10:04,000 --> 00:10:06,637 hack. And so I think that these are the 185 00:10:06,720 --> 00:10:09,597 ways um that you really have to sort of 186 00:10:09,680 --> 00:10:12,317 deal with these situations versus saying 187 00:10:12,400 --> 00:10:15,597 um in in fact I think the core of my 188 00:10:15,680 --> 00:10:19,677 take is we will have to get the 189 00:10:19,760 --> 00:10:23,517 engineering process around building out 190 00:10:23,600 --> 00:10:27,117 this experimental science to be more robust. 191 00:10:27,200 --> 00:10:29,357 >> Yeah. >> Thanks. 192 00:10:29,440 --> 00:10:31,597 >> So I think I think that's a great point. 193 00:10:31,680 --> 00:10:36,157 I love how you uh differentiated in the HuggingFace uh episode between the 194 00:10:36,240 --> 00:10:39,997 mundane things they got wrong like the misconfigured sandbox and HuggingFace 195 00:10:40,080 --> 00:10:43,837 had credentials just sitting in a public repository and there was no monitoring 196 00:10:43,920 --> 00:10:47,437 and then you have the genuinely novel behavior, the swarms of agents, the 197 00:10:47,520 --> 00:10:51,357 reward hacking. That's the stuff that has everyone freaked out. I agree that, 198 00:10:51,440 --> 00:10:55,837 you know, we have to now figure out how to fix the bugs or, you know, fix the 199 00:10:55,920 --> 00:10:59,197 deeper problem that's coming from that reward hacking. What what do you think 200 00:10:59,280 --> 00:11:02,957 that means for and and and and I think to their credit I think what the 201 00:11:03,040 --> 00:11:05,677 Frontier Labs are saying is we are now 202 00:11:05,760 --> 00:11:08,877 going to slow down the pace of let's say 203 00:11:08,960 --> 00:11:13,037 raw power and shift towards reliability and predictability and you know what 204 00:11:13,120 --> 00:11:17,117 they call alignment which I think is good business practice I guess what do 205 00:11:17,200 --> 00:11:21,037 you think that means for what we see in terms of new products for the next year 206 00:11:21,120 --> 00:11:25,677 or two does it mean we just kind of improve what we already have or do we 207 00:11:25,760 --> 00:11:28,397 see new capabilities what do you think this going to mean A great question, 208 00:11:28,480 --> 00:11:30,877 David. I I do think there's already a 209 00:11:30,960 --> 00:11:33,197 massive model overhang, right? I mean, 210 00:11:33,280 --> 00:11:35,677 um capability overhang in the sense of 211 00:11:35,760 --> 00:11:39,277 the models are very good except the 212 00:11:39,360 --> 00:11:44,477 broad diffusion uh requires a lot of things, right? even requires uh 213 00:11:44,560 --> 00:11:48,477 essentially if you're compressing workflows and changing workflows to 214 00:11:48,560 --> 00:11:53,197 happen differently u the amount of change management that needs to happen 215 00:11:53,280 --> 00:11:55,037 in order to even incorporate these 216 00:11:55,120 --> 00:11:57,917 systems is sort of what's taking time so 217 00:11:58,000 --> 00:12:01,037 to some degree I would say the and also 218 00:12:01,120 --> 00:12:04,957 uh the the ability to create these new form factors right I mean if you think 219 00:12:05,040 --> 00:12:09,757 about coding agents and coding agents became really usable when you discovered 220 00:12:09,840 --> 00:12:13,997 that you could have an agent loop with a file system uh and that was the 221 00:12:14,080 --> 00:12:18,317 breakthrough that just made coding agents work. Um and I think now maybe 222 00:12:18,400 --> 00:12:20,797 with KUA right so which is with Astra 223 00:12:20,880 --> 00:12:23,757 with KUA uh could be a way for us to 224 00:12:23,840 --> 00:12:26,717 even do computer use or we just use long 225 00:12:26,800 --> 00:12:28,957 trajectory tasks that can get completely 226 00:12:29,040 --> 00:12:31,357 automated. So I think these type of 227 00:12:31,440 --> 00:12:34,397 product innovations where the model plus 228 00:12:34,480 --> 00:12:37,997 the harness allow us to do things that 229 00:12:38,080 --> 00:12:40,317 then lead to broad adoption. Right? I 230 00:12:40,400 --> 00:12:42,237 even go back to the chat GPT moment for 231 00:12:42,320 --> 00:12:45,677 me, right? Which was it was that RHF at 232 00:12:45,760 --> 00:12:49,917 the very end that made a chat conversation possible. Mhm. 233 00:12:50,000 --> 00:12:52,157 >> Uh and so I think that yes, so there's 234 00:12:52,240 --> 00:12:55,837 some science, there is some form factor 235 00:12:55,920 --> 00:13:00,717 that then leads to broad diffusion and we now need to find the next level of 236 00:13:00,800 --> 00:13:05,117 these things that are doing real work in the real enterprise. Um and in that 237 00:13:05,200 --> 00:13:09,117 context by the way the other thing is it's going to be a multimodel world 238 00:13:09,200 --> 00:13:12,237 right so at this point just out of resilience right I mean think about 239 00:13:12,320 --> 00:13:16,397 right every enterprise now comes to me and says hey this model does refusals 240 00:13:16,480 --> 00:13:20,637 here this model I want weights here I don't and so the people are going to 241 00:13:20,720 --> 00:13:23,117 want multiple models so one of the other 242 00:13:23,200 --> 00:13:25,117 things that we have to get right is some 243 00:13:25,200 --> 00:13:27,997 standards of interop right like even KV 244 00:13:28,080 --> 00:13:32,477 cache like why the heck can't I use multiple model families and have KV 245 00:13:32,560 --> 00:13:36,557 cache reuse uh right we've had document standards 246 00:13:36,640 --> 00:13:39,917 you and I lived through it right but we've sort of you know you kind of have 247 00:13:40,000 --> 00:13:43,837 things that are interoperable in the real world everywhere else so I think 248 00:13:43,920 --> 00:13:47,997 this industry also has to wake up and say hey in fact if I were talking about 249 00:13:48,080 --> 00:13:50,317 the most important pressing things is 250 00:13:50,400 --> 00:13:53,197 how do I have more standards on uh 251 00:13:53,280 --> 00:13:57,837 interoperability how do I have a harness that is external to a model so that my 252 00:13:57,920 --> 00:14:01,277 memory is not tied to one model I mean this is the first time you're going to 253 00:14:01,360 --> 00:14:04,477 have a technology where your use of it 254 00:14:04,560 --> 00:14:07,117 and the exhaust in the data could not be 255 00:14:07,200 --> 00:14:11,677 yours. Uh I mean that you know like it's like if I g sold you a database and said 256 00:14:11,760 --> 00:14:15,597 hey the data you put into your database is not yours and it's mine. It goes away 257 00:14:15,680 --> 00:14:18,957 if I took away the license. How would you feel about it? So therefore I think 258 00:14:19,040 --> 00:14:21,437 we have some serious issues like that to deal with. 259 00:14:21,520 --> 00:14:22,557 >> I think that's a good segue. 260 00:14:22,640 --> 00:14:27,437 >> Sorry. Let me just ask one question to connect the um economic incentive 261 00:14:27,520 --> 00:14:29,517 argument on what's going on. The 262 00:14:29,600 --> 00:14:33,037 argument is the Frontier Labs are facing 263 00:14:33,120 --> 00:14:37,437 token compression. 50 bucks for OpenAI's 264 00:14:37,520 --> 00:14:41,037 kind of million token output versus I 265 00:14:41,120 --> 00:14:45,117 think someone estimated Deep Seeks new is like can go as low as 15 cents for a 266 00:14:45,200 --> 00:14:46,637 million tokens of output. Let's call it 267 00:14:46,720 --> 00:14:50,797 60 cents. 99% cost reduction. 268 00:14:50,880 --> 00:14:53,437 If that is the the big kind of economic 269 00:14:53,520 --> 00:14:58,237 crux of what the frontier labs are facing, why would most tokens be paying 270 00:14:58,320 --> 00:15:02,317 50 bucks? Most enterprises pay 50 bucks when they could pay 60 cents for most of 271 00:15:02,400 --> 00:15:06,237 their tasks. Doesn't that also beg the question, are they in the wrong business 272 00:15:06,320 --> 00:15:09,917 model? And I I asked this for you as the CEO of Microsoft, what's the right 273 00:15:10,000 --> 00:15:13,917 business model? Do you want to be making the frontier model? Do you want to be 274 00:15:14,000 --> 00:15:17,997 running the compute and charging for rent on your compute? Or do you want to 275 00:15:18,080 --> 00:15:20,877 be in the application layer? I know you talk about this a lot, but I just love 276 00:15:20,960 --> 00:15:25,037 your perspective from where we sit today and how this all kind of >> um 277 00:15:25,120 --> 00:15:28,397 >> Yeah, I think the the fundamental thing that I think we're observing is good 278 00:15:28,480 --> 00:15:32,477 old-fashioned competition, right? I mean, for me, if I look back at it, we 279 00:15:32,560 --> 00:15:36,797 were we had like some real great closed source assets, Windows. What was the 280 00:15:36,880 --> 00:15:38,637 check against it? It was of course the 281 00:15:38,720 --> 00:15:41,197 Mac, but Linux 282 00:15:41,280 --> 00:15:45,677 >> uh we had a great closed source product called SQL Server. What was the check 283 00:15:45,760 --> 00:15:50,237 against it? there was always a substitute called Postgress or MySQL. So 284 00:15:50,320 --> 00:15:53,517 I think that's what's happening a little bit of it is there's real competition 285 00:15:53,600 --> 00:15:55,997 between closed source and the open- 286 00:15:56,080 --> 00:16:01,197 source check is real. Um, and that's good quite frankly uh because without it 287 00:16:01,280 --> 00:16:04,637 I don't think we're going to have a broad frontier ecosystem or broad 288 00:16:04,720 --> 00:16:06,637 diffusion because otherwise we'll just 289 00:16:06,720 --> 00:16:08,557 we'll be back to some uh you know 290 00:16:08,640 --> 00:16:11,757 mainframe uh locket that's just not uh a 291 00:16:11,840 --> 00:16:15,197 thing to your point about if anything 292 00:16:15,280 --> 00:16:17,757 given that we will now hopefully 293 00:16:17,840 --> 00:16:21,277 continue to have a much richer choice in 294 00:16:21,360 --> 00:16:23,917 every layer. Right. So to me hopefully 295 00:16:24,000 --> 00:16:26,077 we can start building these AI because 296 00:16:26,160 --> 00:16:29,597 today the royalty of an AI product all 297 00:16:29,680 --> 00:16:32,637 going to just the model layer doesn't 298 00:16:32,720 --> 00:16:36,637 make sense if you really want to build a product company right it just cannot be 299 00:16:36,720 --> 00:16:39,517 in fact if anything like that's the same thing right which is if you take the 300 00:16:39,600 --> 00:16:44,397 database if there was no open-source check on closed source uh the prices 301 00:16:44,480 --> 00:16:47,677 wouldn't have been at a place where people could have built the app tier 302 00:16:47,760 --> 00:16:51,917 successfully and the with a margin and so I think the apps are going to become 303 00:16:52,000 --> 00:16:53,997 you know much more viable economically 304 00:16:54,080 --> 00:16:56,957 which is great for the ecosystem. uh 305 00:16:57,040 --> 00:17:01,117 there are going to be all these other layers of middleware call it right which 306 00:17:01,200 --> 00:17:03,677 is hey what's my memory system what's my 307 00:17:03,760 --> 00:17:08,157 harness and orchestration layer so there's going to be a very rich tools 308 00:17:08,240 --> 00:17:12,557 ecosystem there the model companies will do fine uh in fact you know the paro 309 00:17:12,640 --> 00:17:15,117 they can manage the token pricing based 310 00:17:15,200 --> 00:17:19,837 on their model family if anything I want them to work on even the KV you know 311 00:17:19,920 --> 00:17:23,997 these these standards >> such that we can use multiple model f in 312 00:17:24,080 --> 00:17:28,477 fact it's better for them in fact I worked on Windows interrupt with Unix 313 00:17:28,560 --> 00:17:31,117 first. >> In fact, it was counterintuitive, right? 314 00:17:31,200 --> 00:17:34,957 We used to think, oh my god, this interrupt means we'll be less used 315 00:17:35,040 --> 00:17:39,597 except we were more used. >> In fact, we became weirdly enough 316 00:17:39,680 --> 00:17:44,077 because there were so many variants of Unix at that time that Windows interrupt 317 00:17:44,160 --> 00:17:46,237 made Unix better and Windows better. And 318 00:17:46,320 --> 00:17:48,637 in fact, we were able to penetrate the 319 00:17:48,720 --> 00:17:51,277 enterprise primarily because we did that 320 00:17:51,360 --> 00:17:53,997 interrupt work. And so that's at least 321 00:17:54,080 --> 00:17:55,997 how I think about it. Satya one of these 322 00:17:56,080 --> 00:17:58,557 we're in this interesting moment where 323 00:17:58,640 --> 00:18:01,037 on the one hand you have these experts 324 00:18:01,120 --> 00:18:04,717 asking for regulation asking for 325 00:18:04,800 --> 00:18:07,517 oversight governance it typically always 326 00:18:07,600 --> 00:18:11,117 leads to some restriction of freedom 327 00:18:11,200 --> 00:18:14,637 and general society 328 00:18:14,720 --> 00:18:18,797 are put in a position where now we have to opine on whether this is right or 329 00:18:18,880 --> 00:18:21,517 wrong but then on the other side most 330 00:18:21,600 --> 00:18:23,997 people's lived experience 331 00:18:24,080 --> 00:18:26,397 is not this magical productivity boost 332 00:18:26,480 --> 00:18:29,437 of AI. At best, it's integrating our 333 00:18:29,520 --> 00:18:31,597 Apple Eyewatch data to tell us why we're 334 00:18:31,680 --> 00:18:33,677 sleeping less. That's like functionally 335 00:18:33,760 --> 00:18:36,557 the bar for most people. Or why is my 336 00:18:36,640 --> 00:18:41,837 kid an into chat GPT? Uh so can 337 00:18:41,920 --> 00:18:45,597 you just help us bridge this? I mean, you see so many enterprise applications. 338 00:18:45,680 --> 00:18:49,757 Where's the magic? Like where is the where are the gains in profits? Where 339 00:18:49,840 --> 00:18:52,077 are the huge upside breakthroughs that 340 00:18:52,160 --> 00:18:57,997 AI is creating that will somehow make all of this tension understandable for 341 00:18:58,080 --> 00:18:59,917 everybody? >> Yeah, it's a great it's a great point. I 342 00:19:00,000 --> 00:19:03,357 mean, I think this is the real question 343 00:19:03,440 --> 00:19:06,157 which is how do we truly see this in the 344 00:19:06,240 --> 00:19:10,637 productivity stats? How do we really see it in the GDP growth? That's broadbased. 345 00:19:10,720 --> 00:19:12,637 It's not just supplier 346 00:19:12,720 --> 00:19:15,357 >> or supply side. Um I mean the the one 347 00:19:15,440 --> 00:19:18,317 example that I I love and I get back to 348 00:19:18,400 --> 00:19:19,997 in fact healthcare is a good one right 349 00:19:20,080 --> 00:19:23,757 if you think about um health care and 350 00:19:23,840 --> 00:19:26,317 even the simple doctor patient 351 00:19:26,400 --> 00:19:27,997 interaction in our case we have this 352 00:19:28,080 --> 00:19:30,957 thing called DAX copilot um that's the 353 00:19:31,040 --> 00:19:33,517 place which is the most tangible example 354 00:19:33,600 --> 00:19:38,077 I can always point to when a doctor can spend more time with the patient caring 355 00:19:38,160 --> 00:19:40,237 for them versus just the entry into an 356 00:19:40,320 --> 00:19:43,037 EMR system that's a good productivity 357 00:19:43,120 --> 00:19:46,797 gain If it can triage uh the inbox for 358 00:19:46,880 --> 00:19:48,637 the doctor so that they can be more 359 00:19:48,720 --> 00:19:52,557 responsive uh that's helpful for uh uh 360 00:19:52,640 --> 00:19:57,197 for the patient and the care system the administrator in fact keying like the 361 00:19:57,280 --> 00:19:58,717 insure like because it's the 362 00:19:58,800 --> 00:20:02,157 triangulation of the pay patient and the 363 00:20:02,240 --> 00:20:05,117 health system. Yeah. Uh that's of all in 364 00:20:05,200 --> 00:20:07,197 fact most of healthcare is sort of all 365 00:20:07,280 --> 00:20:10,397 workflow cost. Uh so taming of that 366 00:20:10,480 --> 00:20:14,397 workflow complexity that's a helpful thing. But do you see that in Microsoft 367 00:20:14,480 --> 00:20:15,677 with the people that you're helping? 368 00:20:15,760 --> 00:20:19,437 >> Yeah, absolutely. We see that and and by the way even in in simple co-pilot 369 00:20:19,520 --> 00:20:24,317 cases, right, which is if you look at the amount most people think about jobs 370 00:20:24,400 --> 00:20:27,357 which I think there is going to be displacement there is but the bottom 371 00:20:27,440 --> 00:20:29,437 line is what are the new jobs that get 372 00:20:29,520 --> 00:20:32,477 created uh is going to be one of the key 373 00:20:32,560 --> 00:20:35,117 aspects of it. But also a lot of 374 00:20:35,200 --> 00:20:40,397 knowledge work unfortunately is drudgery right who you know I get up in the 375 00:20:40,480 --> 00:20:41,997 morning and I think about like man all I 376 00:20:42,080 --> 00:20:44,397 do is email triage right you know 377 00:20:44,480 --> 00:20:47,997 >> uh what if uh even just these workflows 378 00:20:48,080 --> 00:20:52,077 that are taking away time from things that you could be spending time on 379 00:20:52,160 --> 00:20:54,477 >> okay well you're bring you're bringing up this great point if you go all the 380 00:20:54,560 --> 00:20:57,757 way back to like the turn of the century the industrial revolution when we had a 381 00:20:57,840 --> 00:21:01,517 7-day work week you know a lot of people forget why did we introduce the weekends 382 00:21:01,600 --> 00:21:05,197 it was to sort of manage the tension between different uh religious groups 383 00:21:05,280 --> 00:21:06,717 that had to work in the same factory. 384 00:21:06,800 --> 00:21:11,757 And then when you look at long run GDP outside of some exogenous events, it 385 00:21:11,840 --> 00:21:15,117 sort of is, you know, between two and 400 basis points. 386 00:21:15,200 --> 00:21:18,637 >> And so what happens is as productivity boosts come in, 387 00:21:18,720 --> 00:21:23,277 >> human work steps back and you kind of accomplish the same amount of work. 388 00:21:23,360 --> 00:21:27,117 >> Do you think that that happens here? Is that is there a risk that we have a 389 00:21:27,200 --> 00:21:29,277 three-day work week and we're just still 390 00:21:29,360 --> 00:21:34,477 growing at two and a half%. Yeah, that's a great qu or will we find new things 391 00:21:34,560 --> 00:21:38,717 and this is where the excitement at least I have for what the real impact of 392 00:21:38,800 --> 00:21:40,957 AI would be is instead of just thinking 393 00:21:41,040 --> 00:21:44,237 about hey it has helped me augment some 394 00:21:44,320 --> 00:21:46,957 workflow or simplify something that's 395 00:21:47,040 --> 00:21:49,277 happening today is it inventing new 396 00:21:49,360 --> 00:21:52,077 things uh is it speeding up drug 397 00:21:52,160 --> 00:21:56,397 discovery um is it taking the u I don't 398 00:21:56,480 --> 00:21:59,917 know let's again go back to my example of okay the working capital management 399 00:22:00,000 --> 00:22:02,797 of a small business has become so much 400 00:22:02,880 --> 00:22:05,917 more efficient uh that suddenly it's no 401 00:22:06,000 --> 00:22:10,397 longer just oh I have an ERP or a QuickBooks like thing but I truly am 402 00:22:10,480 --> 00:22:12,797 making decisions based on the ability to 403 00:22:12,880 --> 00:22:15,677 introspect my invoices my emails and 404 00:22:15,760 --> 00:22:17,677 what have you and some somehow optimize 405 00:22:17,760 --> 00:22:20,477 my working capital that's productivity 406 00:22:20,560 --> 00:22:24,157 that didn't exist and so I do hope that 407 00:22:24,240 --> 00:22:29,517 we will start seeing GDP growth which we did see in the industrial era um during 408 00:22:29,600 --> 00:22:31,117 the first phase of it. >> Yeah. 409 00:22:31,200 --> 00:22:33,997 >> Right. So, so that I think is what is 410 00:22:34,080 --> 00:22:35,757 needed, right? Which is in order for all 411 00:22:35,840 --> 00:22:38,717 of this to play out quite frankly, we do 412 00:22:38,800 --> 00:22:42,877 need to see at least 7 8% GDP growth 413 00:22:42,960 --> 00:22:45,757 that is real and that's broad-based. 414 00:22:45,840 --> 00:22:48,877 >> What's the what business is Microsoft in 415 00:22:48,960 --> 00:22:51,837 in relation to AI? Obviously, Azure has 416 00:22:51,920 --> 00:22:53,757 been crushing it. you're turning away 417 00:22:53,840 --> 00:22:56,797 customers uh and you're doing $175 418 00:22:56,880 --> 00:22:59,197 billion in capex buildout, but your 419 00:22:59,280 --> 00:23:02,397 capex is far below what Meta is doing, 420 00:23:02,480 --> 00:23:06,797 far below what Google's doing. They're doing secondary raises and raising debt, 421 00:23:06,880 --> 00:23:10,397 350 billion. The Frontier Labs are spending 500 billion. You were so early 422 00:23:10,480 --> 00:23:12,957 to the party with the precient open AI 423 00:23:13,040 --> 00:23:16,317 investment, but then co-pilot didn't 424 00:23:16,400 --> 00:23:19,997 exactly land. I don't think it didn't get great reviews. You don't have a 425 00:23:20,080 --> 00:23:21,677 frontier model. What's the business? 426 00:23:21,760 --> 00:23:24,877 Please come back. [laughter] >> No, but what's the business here? What's 427 00:23:24,960 --> 00:23:29,037 the >> Do you need to have a frontier model? 428 00:23:29,120 --> 00:23:31,757 >> Did Did we tell you there was one journalist on the panel? [laughter] 429 00:23:31,840 --> 00:23:35,277 >> No, no, no. It's I mean I mean it sincerely because I'm just curious. 430 00:23:35,360 --> 00:23:39,517 You're a great strategist. We know that about you. Microsoft missed the mobile 431 00:23:39,600 --> 00:23:42,477 revolution. >> Is Microsoft going to miss the AI 432 00:23:42,560 --> 00:23:44,957 revolution? You don't have a frontier model? Because I always found it 433 00:23:45,040 --> 00:23:48,157 perplexing that you didn't. And what's the strategy there in all seriousness? 434 00:23:48,240 --> 00:23:51,357 Like do you think open source is going to win? you should have that play. 435 00:23:51,440 --> 00:23:55,357 >> Yeah. So, let me walk you uh through the sort of where we are and what we're up 436 00:23:55,440 --> 00:23:59,037 to on each of these. By the way, on the capex side and the buildout side, we 437 00:23:59,120 --> 00:24:01,757 started early. So we if you sort of 438 00:24:01,840 --> 00:24:05,597 cumulatively look um it's a good I'm not 439 00:24:05,680 --> 00:24:09,197 sort of saying you know right right now speaking about a lot of capex is not a 440 00:24:09,280 --> 00:24:12,877 feature it's a bug but that [laughter] said but if you really go actually add 441 00:24:12,960 --> 00:24:17,117 up the math uh given when we started because we started multiple years before 442 00:24:17,200 --> 00:24:20,397 people woke up to even actually needing to build and so that's kind of one 443 00:24:20,480 --> 00:24:24,557 aspect of it. The other aspect of it is we are calibrating our capex in such a 444 00:24:24,640 --> 00:24:28,637 way that we don't we don't want to build for one or two customers right so we 445 00:24:28,720 --> 00:24:32,397 want to build for the long tail right because that's I think most important 446 00:24:32,480 --> 00:24:36,157 and that's I mean that if you're a hyperscaler you're not a supplier to two 447 00:24:36,240 --> 00:24:40,637 model companies that's not a business uh you have to sort of basically build a 448 00:24:40,720 --> 00:24:45,917 system that is great for lots of third parties uh and our own one in that 449 00:24:46,000 --> 00:24:49,517 context we're pretty thrilled with the progress we're making uh with even 450 00:24:49,600 --> 00:24:53,117 copilot if you sort of look at the subscriber numbers we gave which is this 451 00:24:53,200 --> 00:24:56,237 is goes back in fact to Chamat's fundamental point which is these are 452 00:24:56,320 --> 00:25:00,877 real enterprises using it for real workflows u and the fact that we now 453 00:25:00,960 --> 00:25:05,597 have 30 plus million not over forum remember the total knowledge worker base 454 00:25:05,680 --> 00:25:09,037 right where most people talk about 3 billion people 4 billion people on the 455 00:25:09,120 --> 00:25:11,357 internet the entire office 365 or 456 00:25:11,440 --> 00:25:14,237 Microsoft 365 is the the the sort of the 457 00:25:14,320 --> 00:25:18,717 standard when it comes to knowledge work there's 450 million that's including all 458 00:25:18,800 --> 00:25:20,157 students in the world >> oh wow 459 00:25:20,240 --> 00:25:22,797 >> right So when we talk like the market 460 00:25:22,880 --> 00:25:26,077 quote unquote as defined is maybe 300 uh 461 00:25:26,160 --> 00:25:29,517 250 even of real enterprise users and of 462 00:25:29,600 --> 00:25:31,837 that we've got the penetration of close 463 00:25:31,920 --> 00:25:33,597 to 30 million on that and it's growing 464 00:25:33,680 --> 00:25:36,957 and so on. The aspect on the model side 465 00:25:37,040 --> 00:25:41,197 is we're thrilled about obviously our investment in open AAI the access we 466 00:25:41,280 --> 00:25:44,957 have to their IP which we have for a long time we're going to use that but we 467 00:25:45,040 --> 00:25:46,797 are well on our way building our MAI 468 00:25:46,880 --> 00:25:49,277 models right if you look at it we have a 469 00:25:49,360 --> 00:25:54,397 flash cyber model that you know with our harness orchestrating other models 470 00:25:54,480 --> 00:25:55,917 outperforms 471 00:25:56,000 --> 00:25:59,277 um on cyber gym even a mythos uh same 472 00:25:59,360 --> 00:26:03,917 thing we're seeing in coding same thing we're seeing in uh knowledge work right 473 00:26:04,000 --> 00:26:08,557 So our goal is to basically hill climb from the bottom by the way uh not 474 00:26:08,640 --> 00:26:10,797 distilling anything. So from the very 475 00:26:10,880 --> 00:26:13,997 bottom using our RLES our data uh and 476 00:26:14,080 --> 00:26:17,917 then also have a differentiated position with enterprises going back to 477 00:26:18,000 --> 00:26:21,837 addressing some of the things that they want which is hey can I have the weights 478 00:26:21,920 --> 00:26:27,197 can I have the weights that I can then add to my knowledge uh these are the 479 00:26:27,280 --> 00:26:30,477 things that we will do with our foundation. Your best advice I think to 480 00:26:30,560 --> 00:26:35,197 enterprises is AI sovereignty is important. Putting your data into a 481 00:26:35,280 --> 00:26:38,717 frontier model probably not a good idea and then you're going to be that harness 482 00:26:38,800 --> 00:26:40,797 for them to to help them. So my 483 00:26:40,880 --> 00:26:43,677 implement my advice is more like use all 484 00:26:43,760 --> 00:26:46,317 but be independent of all. So for 485 00:26:46,400 --> 00:26:49,757 example my asset test is you should 486 00:26:49,840 --> 00:26:51,757 always eval 487 00:26:51,840 --> 00:26:53,757 that matter to you right. So what's the 488 00:26:53,840 --> 00:26:56,957 outcome you want? you should go run that 489 00:26:57,040 --> 00:27:01,677 outcome through all the models. Then here's the test I would do. I would pull 490 00:27:01,760 --> 00:27:03,837 out a model and see whether I can retain 491 00:27:03,920 --> 00:27:06,717 the eval. If I can't, that means you 492 00:27:06,800 --> 00:27:11,037 really are dependent on something that may or may not be yours. 493 00:27:11,120 --> 00:27:14,237 >> Right. >> Right. That's so so my fundamental 494 00:27:14,320 --> 00:27:18,637 enterprise architecture would say you should have a model system that 495 00:27:18,720 --> 00:27:21,197 fundamentally allows you to be able to 496 00:27:21,280 --> 00:27:23,837 continuously hill climb on your own on 497 00:27:23,920 --> 00:27:25,917 eval 498 00:27:26,000 --> 00:27:29,837 uh while using all models closed open u 499 00:27:29,920 --> 00:27:33,917 if you want you can even fine-tune any of these models but you can even 500 00:27:34,000 --> 00:27:36,637 substitute models >> s just to build on Jason's question you 501 00:27:36,720 --> 00:27:40,957 had this um incredible moment I think we put it here where you said you know 502 00:27:41,040 --> 00:27:42,877 we're good for our 80 billion. But just 503 00:27:42,960 --> 00:27:45,597 to expand the question, um there's 504 00:27:45,680 --> 00:27:48,237 effectively this sort of bank of AI that 505 00:27:48,320 --> 00:27:52,957 has emerged and there's this financing mechanism that just is so important to 506 00:27:53,040 --> 00:27:57,117 the entire ecosystem and now broadly to the entire economy. But you've been very 507 00:27:57,200 --> 00:28:00,877 disciplined. You have an enormous balance sheet. You're also an investment 508 00:28:00,960 --> 00:28:05,037 grade issuer. So you could do what Jensen did, but you've taken a very 509 00:28:05,120 --> 00:28:08,637 different capital allocation approach, much larger bets, very concentrated, and 510 00:28:08,720 --> 00:28:12,317 you've kind of stayed into your own ecosystem. just talk us through your 511 00:28:12,400 --> 00:28:14,237 mindset as a capital allocator at 512 00:28:14,320 --> 00:28:16,317 Microsoft and that balance sheet. Yeah. 513 00:28:16,400 --> 00:28:20,397 So the way I'm sort of looking at our 514 00:28:20,480 --> 00:28:21,997 book of business whether it's the hypers 515 00:28:22,080 --> 00:28:26,077 scale our model or our app tier and the 516 00:28:26,160 --> 00:28:31,037 shape of the demand um and then what's the way to build out for it. And so if 517 00:28:31,120 --> 00:28:34,077 you think about these assets right there are two classes of it. There are the 518 00:28:34,160 --> 00:28:37,677 long lead um long duration assets like 519 00:28:37,760 --> 00:28:40,317 the the land power cold shell let's call 520 00:28:40,400 --> 00:28:43,597 it. Then there is the kit. the kit is 521 00:28:43,680 --> 00:28:46,957 the short-term uh asset uh that you can 522 00:28:47,040 --> 00:28:51,517 much more you know uh be demand driven in other words right I have to forecast 523 00:28:51,600 --> 00:28:54,477 let's say two years three year out demand and then and then also 524 00:28:54,560 --> 00:28:57,757 >> the kit means the racks the chips >> the racks the chips and what have you 525 00:28:57,840 --> 00:28:59,597 and that's 60% of the cost or what have 526 00:28:59,680 --> 00:29:01,597 you right so therefore so what we do is 527 00:29:01,680 --> 00:29:05,837 we go build as much um we lease we even 528 00:29:05,920 --> 00:29:07,837 rent now right now we're even renting 529 00:29:07,920 --> 00:29:09,837 quite a bit because we kind of were 530 00:29:09,920 --> 00:29:13,677 short on supply uh But the overall goal 531 00:29:13,760 --> 00:29:17,117 is to build more lease some and then if 532 00:29:17,200 --> 00:29:19,277 really need to surge we will even rent 533 00:29:19,360 --> 00:29:22,317 that's kind of on the on the on the uh 534 00:29:22,400 --> 00:29:25,917 assets and then the chips themselves we 535 00:29:26,000 --> 00:29:28,477 will try to be first of all make sure 536 00:29:28,560 --> 00:29:31,197 that we're matching demand and as I said 537 00:29:31,280 --> 00:29:35,837 my goal is not to have just two customers three customers uh it's great 538 00:29:35,920 --> 00:29:39,197 to have openi being one of our largest customers it's great that they're 539 00:29:39,280 --> 00:29:41,357 growing uh but we need more uh is the 540 00:29:41,440 --> 00:29:45,197 kit over earning right now and do do we 541 00:29:45,280 --> 00:29:48,157 need is the is the industry pushing for 542 00:29:48,240 --> 00:29:51,517 diversification more silicon more memory 543 00:29:51,600 --> 00:29:52,717 more vendors 544 00:29:52,800 --> 00:29:56,157 >> yeah what's happening is the workloads 545 00:29:56,240 --> 00:29:59,597 that are now at scale uh they obviously 546 00:29:59,680 --> 00:30:03,517 grew up from what GPUs were there but 547 00:30:03,600 --> 00:30:06,877 now the the shape is so well understood 548 00:30:06,960 --> 00:30:11,677 uh that you're able to optimize for a very different world right So you can 549 00:30:11,760 --> 00:30:13,597 sort of start building 550 00:30:13,680 --> 00:30:15,597 >> um and saying well you know there are 551 00:30:15,680 --> 00:30:18,957 these multiple phases in um an inference 552 00:30:19,040 --> 00:30:23,677 or a training phase so why not build silicon that's optimized for these uh 553 00:30:23,760 --> 00:30:27,197 and that's just going to lead to a systems architecture that I think is 554 00:30:27,280 --> 00:30:31,677 going to by definition have a lot more uh diversity uh I mean I know you have 555 00:30:31,760 --> 00:30:35,997 Jensen coming he himself if you look at his own architecture is changing quite 556 00:30:36,080 --> 00:30:39,437 drastically >> quite drastically >> um and so I think that there is going to 557 00:30:39,520 --> 00:30:43,997 be a lot more choice even there in that layer. So ours we have Jensen stuff 558 00:30:44,080 --> 00:30:48,477 which is I think our primary thing. We have our own uh OpenAI is building their 559 00:30:48,560 --> 00:30:50,157 chip so that's also going to be there. 560 00:30:50,240 --> 00:30:55,117 AMD is in there. So we I I my thing is to run whether it's the OpenAI models, 561 00:30:55,200 --> 00:30:58,237 the anthropic models or our own models on a heterogeneous kit. 562 00:30:58,320 --> 00:31:00,557 >> Sax I want to let you get in here before we run out of time. 563 00:31:00,640 --> 00:31:02,477 >> Yeah. So you know we've heard now from 564 00:31:02,560 --> 00:31:05,117 the the various frontier lab leaders Sam 565 00:31:05,200 --> 00:31:07,997 Dario Elon Demis that we need to 566 00:31:08,080 --> 00:31:10,237 prioritize alignment like we're talking 567 00:31:10,320 --> 00:31:13,197 predictability reliability robustness uh 568 00:31:13,280 --> 00:31:16,157 as opposed to maybe just say raw raw 569 00:31:16,240 --> 00:31:18,157 power. Do you think the Chinese labs 570 00:31:18,240 --> 00:31:20,237 will follow suit? 571 00:31:20,320 --> 00:31:23,437 I think that that's the dialogue um that 572 00:31:23,520 --> 00:31:25,277 is I think should be prioritized right 573 00:31:25,360 --> 00:31:28,237 so because at some level my own premise 574 00:31:28,320 --> 00:31:31,357 would be that 575 00:31:31,440 --> 00:31:36,237 that China should also deeply care uh 576 00:31:36,320 --> 00:31:38,957 about the same safety concerns if the 577 00:31:39,040 --> 00:31:43,917 United States uh cares about them right why should it be different for them it's 578 00:31:44,000 --> 00:31:47,197 not like they won't have the same hacking problem 579 00:31:47,280 --> 00:31:50,477 >> uh it's not as if uh they don't want to 580 00:31:50,560 --> 00:31:55,277 make sure that their citizens um are benefiting from AI just like we will 581 00:31:55,360 --> 00:31:57,437 want our citizens to benefit from AI. So 582 00:31:57,520 --> 00:32:00,477 I think that there's a possibility of 583 00:32:00,560 --> 00:32:02,957 international norms around it. If we 584 00:32:03,040 --> 00:32:07,597 really are concrete about what's the risk, why is this risk so idiosyncratic 585 00:32:07,680 --> 00:32:12,637 that the only people who are worried about it is the Americans. Uh it doesn't 586 00:32:12,720 --> 00:32:15,917 make sense, right? It's not like a thing that is sort of said, "Oh, I'm going to 587 00:32:16,000 --> 00:32:19,997 only show up in the United States. I'm going to be something. If it is going to 588 00:32:20,080 --> 00:32:22,877 go wrong, it's going to go wrong everywhere at the same time." So I think 589 00:32:22,960 --> 00:32:27,117 the Chinese should care. I mean they're they are a superpower. 590 00:32:27,200 --> 00:32:30,317 >> Well that's you use the word idiosyncratic and I think that is the 591 00:32:30,400 --> 00:32:32,877 right word is I don't think we know yet 592 00:32:32,960 --> 00:32:37,677 is this um you know conversation we're having in the US over the past week. Is 593 00:32:37,760 --> 00:32:39,757 it idiosyncratic to us because we have 594 00:32:39,840 --> 00:32:45,197 you know the strong I guess you could say doomer type uh school of thought or 595 00:32:45,280 --> 00:32:47,917 is it something that the rest of the world will basically feel as well? 596 00:32:48,000 --> 00:32:50,237 >> It's a great question >> and if they do then presumably they'd 597 00:32:50,320 --> 00:32:54,717 want to act on it as well. Yeah, I I just feel my my take there is that we 598 00:32:54,800 --> 00:32:56,317 are ahead 599 00:32:56,400 --> 00:32:59,997 >> and we are who we are which is we argue 600 00:33:00,080 --> 00:33:03,117 we sort of we compete uh we are more 601 00:33:03,200 --> 00:33:07,117 transparent which is all by the way virtues as far as I'm concerned so 602 00:33:07,200 --> 00:33:11,117 therefore the fact that this debate is happening here the world will be better 603 00:33:11,200 --> 00:33:13,277 off for it right so to some degree us 604 00:33:13,360 --> 00:33:15,917 setting if anything I would love a US 605 00:33:16,000 --> 00:33:19,997 set us to lead in the norms that allow 606 00:33:20,080 --> 00:33:22,717 us to defuse use this technology broadly 607 00:33:22,800 --> 00:33:27,677 and create safety standards uh that work for the world including China. But 608 00:33:27,760 --> 00:33:31,197 >> what do you think we should be doing that we're not doing and what are you 609 00:33:31,280 --> 00:33:35,757 doing at Microsoft to change the narrative the populist 610 00:33:35,840 --> 00:33:39,277 sentiment that we have to shut down super intelligence stop building data 611 00:33:39,360 --> 00:33:40,397 centers 612 00:33:40,480 --> 00:33:44,557 >> etc. So, so to me I think this is I am 613 00:33:44,640 --> 00:33:47,197 squarely focused on one of the to 614 00:33:47,280 --> 00:33:50,317 answering Chamat's question from earlier 615 00:33:50,400 --> 00:33:53,037 which is whom is it benefiting and give 616 00:33:53,120 --> 00:33:55,517 me concrete stories right uh we talked 617 00:33:55,600 --> 00:33:58,077 about the productivity benefits a bit uh 618 00:33:58,160 --> 00:34:00,557 whether it's in healthcare or in general 619 00:34:00,640 --> 00:34:05,517 knowledge work coding but I'll give you another example right I was looking at 620 00:34:05,600 --> 00:34:07,517 data centers because after all we didn't 621 00:34:07,600 --> 00:34:10,557 talk much uh today on that but there's a 622 00:34:10,640 --> 00:34:12,557 challenge on how does one earn 623 00:34:12,640 --> 00:34:15,597 permission uh to open a data center in a 624 00:34:15,680 --> 00:34:18,797 region. In fact, we just have some of 625 00:34:18,880 --> 00:34:21,837 the best longitudinal data now for a 626 00:34:21,920 --> 00:34:26,157 data center we built out in Quinsey, Washington, uh for 20 years, close to, 627 00:34:26,240 --> 00:34:28,477 you know, 2008 is when we started it. 628 00:34:28,560 --> 00:34:31,277 And when I look at that data and what it 629 00:34:31,360 --> 00:34:33,117 has meant for that community, right, 630 00:34:33,200 --> 00:34:35,837 where uh the tax revenues have gone up 631 00:34:35,920 --> 00:34:39,997 12 times, uh the paidin taxes have gone 632 00:34:40,080 --> 00:34:44,637 down by a third. Um the growth is higher 633 00:34:44,720 --> 00:34:47,357 than Seattle in Quinsey. This is a rural 634 00:34:47,440 --> 00:34:50,717 town. Uh they have a new school, a new 635 00:34:50,800 --> 00:34:54,717 hospital, a new town center, a new aquatic center. Wow. 636 00:34:54,800 --> 00:34:57,997 >> Uh we have two and most people say, "Oh, there not that many jobs." In fact, 637 00:34:58,080 --> 00:35:03,037 there have been 1,200 construction jobs in that region all through that 20-year 638 00:35:03,120 --> 00:35:06,317 period, right? Because it's not like you just build it and leave. You 639 00:35:06,400 --> 00:35:08,957 continuously refurbishing, building, expanding. 640 00:35:09,040 --> 00:35:10,477 >> And how big, how big is that data center? 641 00:35:10,560 --> 00:35:14,477 >> Uh I think it's now going to be at least 4 or 500 megawatt 642 00:35:14,560 --> 00:35:16,717 >> and it sort of will keep expanding. 643 00:35:16,800 --> 00:35:21,037 >> Um and so so these are uh so that's a 644 00:35:21,120 --> 00:35:25,437 real like that community. So earning it like just not saying hey these are all 645 00:35:25,520 --> 00:35:28,397 the benefits but seeing it >> but how do you get people to tell that 646 00:35:28,480 --> 00:35:30,957 story because that's what's missing today is those stories aren't being 647 00:35:31,040 --> 00:35:34,797 organically told and if a Microsoft executive gets on stage and says don't 648 00:35:34,880 --> 00:35:36,717 worry it's good for the community. 649 00:35:36,800 --> 00:35:40,157 >> Yeah. No I don't think Yeah. So I think storytelling is one thing. The other one 650 00:35:40,240 --> 00:35:42,957 is I think we just need more people 651 00:35:43,040 --> 00:35:47,597 outside of the tech industry to say yeah because if you go to Quinsey Washington 652 00:35:47,680 --> 00:35:49,757 they will tell you thank god for this 653 00:35:49,840 --> 00:35:51,677 data center. It's part of like you know. 654 00:35:51,760 --> 00:35:54,957 So to me that's like when it's tangible 655 00:35:55,040 --> 00:35:58,717 uh like that uh because that's the only way to earn permission because at some 656 00:35:58,800 --> 00:36:01,117 level the skepticism of any of us in the 657 00:36:01,200 --> 00:36:04,077 tech industry just saying things uh is 658 00:36:04,160 --> 00:36:06,557 so high that I think we have to now do 659 00:36:06,640 --> 00:36:09,597 the hard yards of actually doing things 660 00:36:09,680 --> 00:36:13,677 in the world uh which allow people to say okay I now believe you. 661 00:36:13,760 --> 00:36:16,397 >> It's a new muscle. >> It's a new muscle. It's a new muscle. 662 00:36:16,480 --> 00:36:18,877 >> So I think you're a good spokesperson to 663 00:36:18,960 --> 00:36:20,797 flex that muscle. I hope you do it more. 664 00:36:20,880 --> 00:36:22,397 Thank you for being with us. >> Thank you so much. 665 00:36:22,480 --> 00:36:26,470 >> We appreciate you. [music] 666 00:36:29,440 --> 00:36:33,760 >> Thank you, sir. Appreciate your time.