1 00:00:00,400 --> 00:00:05,437 models, runtime, and silicon, all working together to create a meaningful 2 00:00:05,520 --> 00:00:09,037 experience. We're working with the GitHub team to 3 00:00:09,120 --> 00:00:13,277 bring our vision for hybrid intelligence to life on Windows. 4 00:00:13,360 --> 00:00:15,517 Earlier this year, GitHub launched Hydro 5 00:00:15,600 --> 00:00:17,917 Fusion, which routes each task to the 6 00:00:18,000 --> 00:00:21,437 right model in the cloud. Today, we're 7 00:00:21,520 --> 00:00:25,677 extending intelligence to Windows, letting Hydrofusion tap into models 8 00:00:25,760 --> 00:00:30,190 running locally. >> [applause] 9 00:00:31,920 --> 00:00:34,077 >> Now let's talk about these models. At 10 00:00:34,160 --> 00:00:37,997 Build, we introduced Miode1 and Flash, a 11 00:00:38,080 --> 00:00:40,477 130 billion parameter coding model built 12 00:00:40,560 --> 00:00:43,597 for real world developer workloads. Now 13 00:00:43,680 --> 00:00:48,477 we're bringing that same intelligence to your device. This is where full stack 14 00:00:48,560 --> 00:00:52,317 engineering comes in. We have taken that model to three bits of precision, 15 00:00:52,400 --> 00:00:55,437 reducing its size by nearly 80% while 16 00:00:55,520 --> 00:01:00,157 preserving coding quality. That is an incredible feat of engineering. But 17 00:01:00,240 --> 00:01:01,677 really, shrinking the model to the 18 00:01:01,760 --> 00:01:05,037 device is only half the story. MAI code 19 00:01:05,120 --> 00:01:08,237 can work at impressive 256K context 20 00:01:08,320 --> 00:01:12,617 window right on your device. 21 00:01:12,700 --> 00:01:16,317 [applause] Now, our goal our goal is simple. 22 00:01:16,400 --> 00:01:19,277 Let AI run locally, stay responsive 23 00:01:19,360 --> 00:01:21,037 while leaving the room for everything 24 00:01:21,120 --> 00:01:23,917 else on your PC. And there are many more 25 00:01:24,000 --> 00:01:26,557 powerful models coming to Windows. 26 00:01:26,640 --> 00:01:28,477 Nvidia is working on a new Neatron 27 00:01:28,560 --> 00:01:31,757 model. At over 70 billion parameters, it 28 00:01:31,840 --> 00:01:34,077 manages to quantize to 2bit and uses 29 00:01:34,160 --> 00:01:36,557 just over 20 gigs of memory, making it 30 00:01:36,640 --> 00:01:42,890 ideal to run on RTX Spark. [applause] 31 00:01:43,200 --> 00:01:45,597 And why why stop there? With advanced 32 00:01:45,680 --> 00:01:47,277 quantization methods, it's possible to 33 00:01:47,360 --> 00:01:50,397 bring more of the frontier to your PC. 34 00:01:50,480 --> 00:01:54,797 At 1.6 6 bits deepsee flash v4 a 284 35 00:01:54,880 --> 00:01:56,717 billion parameter model is running 36 00:01:56,800 --> 00:02:00,637 locally on device fitting in just that's 37 00:02:00,720 --> 00:02:03,597 right it's quite amazing fitting in just 38 00:02:03,680 --> 00:02:06,477 60 gigs of memory this model outperforms 39 00:02:06,560 --> 00:02:09,597 GP5 encoding and reasoning benchmarks 40 00:02:09,680 --> 00:02:14,477 a year ago such capabilities were exclusive to Frontier cloud models now 41 00:02:14,560 --> 00:02:16,237 you can have these capabilities running 42 00:02:16,320 --> 00:02:18,397 on your PC 43 00:02:18,480 --> 00:02:21,917 and this is just the understanding 44 00:02:22,000 --> 00:02:25,997 to power this growing ecosystem of highly capable models, you need a 45 00:02:26,080 --> 00:02:28,397 performant runtime. And this is where 46 00:02:28,480 --> 00:02:31,197 Windows ML comes in. Windows ML is a 47 00:02:31,280 --> 00:02:33,437 scalable runtime to deploy models across 48 00:02:33,520 --> 00:02:35,917 GPUs, NPUs, and CPUs. And we're 49 00:02:36,000 --> 00:02:38,317 continuously evolving it. We're excited 50 00:02:38,400 --> 00:02:43,017 to announce we're bringing Llama CPP to Windows ML. 51 00:02:43,100 --> 00:02:45,120 [applause] 52 00:02:46,000 --> 00:02:50,557 This gives developers more choice. Open source models on day one, the ability to 53 00:02:50,640 --> 00:02:54,717 experiment quickly with emerging models as they land. 54 00:02:54,800 --> 00:02:59,597 With intelligent routing, highly capable models, and a performant runtime, a 55 00:02:59,680 --> 00:03:03,837 clear recipe emerges for what hybrid intelligence looks like. That's the 56 00:03:03,920 --> 00:03:05,997 repeatable pattern we'll apply to light 57 00:03:06,080 --> 00:03:11,037 up experiences across coding, security, creation, and work. Kayla is here from 58 00:03:11,120 --> 00:03:12,557 the GitHub team to show you this in 59 00:03:12,640 --> 00:03:18,520 action. Come on over, Kayla. [applause] 60 00:03:20,560 --> 00:03:24,717 >> Thank you, Pavin. I'm excited to show you how I'm going to use the best of 61 00:03:24,800 --> 00:03:26,877 cloud and local models to do my coding 62 00:03:26,960 --> 00:03:29,917 workflows in GitHub Copilot on Windows. 63 00:03:30,000 --> 00:03:34,717 So, to start, let's just paste in a simple prompt. And I have the Windows 64 00:03:34,800 --> 00:03:39,357 terminal codebase cloned onto my machine. So what I'm asking is I want 65 00:03:39,440 --> 00:03:43,757 our agent to detect unused code, imports, packages, and obsolete feature 66 00:03:43,840 --> 00:03:47,357 flags. And then give me an actionable summary. So this is just basic code 67 00:03:47,440 --> 00:03:49,757 cleanup that I'd like our agent to do. 68 00:03:49,840 --> 00:03:53,917 And then here at the bottom, you might notice in the model picker, I have auto 69 00:03:54,000 --> 00:03:57,997 efficiency enabled. So this is going to select my model depending on the 70 00:03:58,080 --> 00:04:02,077 complexity of the task. And now with our new version of auto, it will 71 00:04:02,160 --> 00:04:05,997 automatically pick a local model depending on if that is a more simple 72 00:04:06,080 --> 00:04:11,197 task. So here we could see on the left hand side our GPU has ticked up and 73 00:04:11,280 --> 00:04:15,997 that's because we're leveraging our local model here and I did run this same 74 00:04:16,080 --> 00:04:17,837 prompt before on a different codebase 75 00:04:17,920 --> 00:04:19,997 and we could see that it leveraged MAI 76 00:04:20,080 --> 00:04:24,237 code 1.1 flash local. Now since my code 77 00:04:24,320 --> 00:04:27,197 is cloned to my desktop and my model is 78 00:04:27,280 --> 00:04:31,437 local we could just put our whole machine into airplane mode and this 79 00:04:31,520 --> 00:04:34,637 process will still continue to turn because everything is right here on the 80 00:04:34,720 --> 00:04:38,077 device. So this is just a simple coding 81 00:04:38,160 --> 00:04:41,757 task right to the local model. But what if we asked something a little bit more 82 00:04:41,840 --> 00:04:44,717 complex? So in this session I asked our 83 00:04:44,800 --> 00:04:47,517 model to or I asked our agent to look at 84 00:04:47,600 --> 00:04:51,837 the issues in the Windows terminal repo that has that have the needs repro 85 00:04:51,920 --> 00:04:55,517 labels. So these are issues that our team hasn't been able to reproduce. And 86 00:04:55,600 --> 00:04:59,197 then I wanted to do some investigation and then hand off to some local sub 87 00:04:59,280 --> 00:05:02,717 agents to do further building and testing and see if we can get more 88 00:05:02,800 --> 00:05:04,797 information here. So this prompt 89 00:05:04,880 --> 00:05:07,597 actually ended up leveraging GPT6 Luna. 90 00:05:07,680 --> 00:05:12,077 So this is a cloud session here. And then we've got three local sub sessions 91 00:05:12,160 --> 00:05:17,197 happening on the lefth hand side leveraging MAI code 1.1.1 flash local. 92 00:05:17,280 --> 00:05:21,837 So this is a really great example of hybrid intelligence. If I hop into one 93 00:05:21,920 --> 00:05:26,397 of these local sessions, we can also see all of the tokens that I got for free. 94 00:05:26,480 --> 00:05:30,637 And on this one in particular, we've got 1.6 6 million being sent to the model 95 00:05:30,720 --> 00:05:34,557 and then 10 and a half thousand coming back from the model. And all of this was 96 00:05:34,640 --> 00:05:39,197 free and this was just one of the sub sessions that I ran. So I have a lot of 97 00:05:39,280 --> 00:05:42,877 sessions going. I like to take a look at what I've got running on my machine and 98 00:05:42,960 --> 00:05:45,037 what I've run before. So we have a 99 00:05:45,120 --> 00:05:47,117 canvas that we created that is a really 100 00:05:47,200 --> 00:05:49,917 great visualizer of my sessions. So a 101 00:05:50,000 --> 00:05:54,157 canvas is a custom UI that works inside the copilot app where both you and your 102 00:05:54,240 --> 00:05:58,637 agent can work back and forth and update the UI accordingly. So here in our 103 00:05:58,720 --> 00:06:02,397 visualizer, all of my purple sessions are local sessions that I've run with 104 00:06:02,480 --> 00:06:05,837 our little leaf and all of my blue sessions with the cloud are cloud 105 00:06:05,920 --> 00:06:09,677 sessions that I've run. And you might notice I've got the same session 106 00:06:09,760 --> 00:06:13,517 happening a few times. I'm doing issue triage a lot and I'm doing it locally. 107 00:06:13,600 --> 00:06:17,357 And that's because issue triage is one of the tasks that takes a while and it's 108 00:06:17,440 --> 00:06:20,637 something that we have to do pretty frequently. So I've actually set this up 109 00:06:20,720 --> 00:06:22,957 as an automation. So automations can run 110 00:06:23,040 --> 00:06:25,597 on a cadence or on a trigger. And I've 111 00:06:25,680 --> 00:06:28,237 set my issue triage one to leverage the 112 00:06:28,320 --> 00:06:32,957 local model explicitly. So I know for a fact that it's going to use just the 113 00:06:33,040 --> 00:06:36,317 model on my machine and it's not going to spend any credits without me knowing. 114 00:06:36,400 --> 00:06:39,517 So I can have this running overnight and I'm still getting all of those tokens 115 00:06:39,600 --> 00:06:43,837 for free even while I'm asleep. So these are just a few scenarios that I've 116 00:06:43,920 --> 00:06:48,397 showed today. Imagine how much more you can save doing these workflows always on 117 00:06:48,480 --> 00:06:53,227 even when you step away. Thank you so much. I'm going to hand it back to Pavin. 118 00:06:53,310 --> 00:06:55,330 [applause] 119 00:06:58,960 --> 00:07:01,517 Thank you, Kayla. That was awesome. 120 00:07:01,600 --> 00:07:05,117 Hybrid intelligence Windows powered by Hydra Fusion will be coming to the 121 00:07:05,200 --> 00:07:08,397 GitHub Copilot app, CLI, and VS Code 122 00:07:08,480 --> 00:07:11,277 starting October 15th. 123 00:07:11,360 --> 00:07:15,757 You've seen how Windows is the most secure platform for hybrid intelligence. 124 00:07:15,840 --> 00:07:18,237 Now, to make AI a part of everyday life, 125 00:07:18,320 --> 00:07:20,637 we need to bring it to places people use 126 00:07:20,720 --> 00:07:23,197 the most into the flow of human 127 00:07:23,280 --> 00:07:27,757 ambition. We're moving toward a future where everyone can simply express what 128 00:07:27,840 --> 00:07:31,837 they want and their Windows PC with their permission can take action on 129 00:07:31,920 --> 00:07:34,317 their behalf. For hundreds of millions 130 00:07:34,400 --> 00:07:36,797 of people each day, search on the 131 00:07:36,880 --> 00:07:40,157 taskbar is where intent begins. So, it's 132 00:07:40,240 --> 00:07:44,957 only natural our story starts there. At its core, the new Windows search 133 00:07:45,040 --> 00:07:48,717 experience is quicker, more efficient, so you can find your apps, settings, and 134 00:07:48,800 --> 00:07:50,957 files faster. 135 00:07:51,040 --> 00:07:53,837 Starting this fall, on Windows 11 PCs, 136 00:07:53,920 --> 00:07:56,397 you can take thousands of quick actions 137 00:07:56,480 --> 00:07:59,597 directly from search on the taskbar. 138 00:07:59,680 --> 00:08:02,157 You can switch to dark mode. You can 139 00:08:02,240 --> 00:08:04,477 turn on do not disturb to help you stay 140 00:08:04,560 --> 00:08:07,597 focused. You can turn up your mic volume 141 00:08:07,680 --> 00:08:10,797 and then use the inline slider to get to 142 00:08:10,880 --> 00:08:14,157 the exact level you want. 143 00:08:14,240 --> 00:08:17,757 All of this as fast as you can type. 144 00:08:17,840 --> 00:08:19,997 And actions can do more than just 145 00:08:20,080 --> 00:08:23,117 settings. You can take a screenshot 146 00:08:23,200 --> 00:08:25,677 directly arrange your windows all from 147 00:08:25,760 --> 00:08:30,717 the taskbar. And my personal favorite, you can send a text message to a friend 148 00:08:30,800 --> 00:08:33,437 or colleague just letting them know you're running a little late without 149 00:08:33,520 --> 00:08:34,717 having to pull your phone out of your 150 00:08:34,800 --> 00:08:36,877 pocket. A quick intuitive way to move 151 00:08:36,960 --> 00:08:42,750 from intent to action. [applause] 152 00:08:44,800 --> 00:08:47,197 And of course, right from the taskbar, 153 00:08:47,280 --> 00:08:49,197 you'll be able to connect to the new 154 00:08:49,280 --> 00:08:51,917 Copilot app when your request calls for 155 00:08:52,000 --> 00:08:54,637 it. Ask a quick question and Copilot 156 00:08:54,720 --> 00:08:57,357 will bring the answer right there to you 157 00:08:57,440 --> 00:08:59,917 in line in search. If you want to have a 158 00:09:00,000 --> 00:09:04,477 richer conversation, you can jump directly from search to the full copilot 159 00:09:04,560 --> 00:09:06,557 app to tell you more about the copilot 160 00:09:06,640 --> 00:09:12,230 experience on Windows. Please welcome Jacob. [applause] 161 00:09:14,720 --> 00:09:17,757 Thanks Bobin. Recently we introduced the 162 00:09:17,840 --> 00:09:19,677 new co-pilot with three core 163 00:09:19,760 --> 00:09:24,477 experiences. home code and autopilot. 164 00:09:24,560 --> 00:09:29,277 Today I'm really excited to talk to you about how Windows hybrid intelligence 165 00:09:29,360 --> 00:09:31,437 supercharges everything that we've just 166 00:09:31,520 --> 00:09:33,597 brought to market. First, let's turn it 167 00:09:33,680 --> 00:09:36,957 on. Copilot now has access to three new 168 00:09:37,040 --> 00:09:40,877 things. First, local context. Local 169 00:09:40,960 --> 00:09:42,957 context means that Copilot can now 170 00:09:43,040 --> 00:09:45,597 ground over all of the files on your PC 171 00:09:45,680 --> 00:09:48,237 the exact same way that work IQ helps 172 00:09:48,320 --> 00:09:50,397 businesses ground all over all of their 173 00:09:50,480 --> 00:09:52,957 business context in the cloud. 174 00:09:53,040 --> 00:09:56,237 Second, local actions. Local actions 175 00:09:56,320 --> 00:10:00,397 means that Copilot can actually take action directly on my machine. It can 176 00:10:00,480 --> 00:10:04,877 move files. It can change a setting. It has the same level of the ability to 177 00:10:04,960 --> 00:10:06,637 control and change things that I do as a 178 00:10:06,720 --> 00:10:09,597 user, but always with permission. Third, 179 00:10:09,680 --> 00:10:12,637 local models. Copilot will still use the 180 00:10:12,720 --> 00:10:15,117 cloud for the hardest tasks, but for 181 00:10:15,200 --> 00:10:17,997 times when cost or privacy matter more, 182 00:10:18,080 --> 00:10:22,317 it can delegate down to local models that run directly on your computer. So, 183 00:10:22,400 --> 00:10:25,437 let's take a look at how this transforms the rest of Copilot. Let's first talk 184 00:10:25,520 --> 00:10:27,917 about home. Home is the new starting 185 00:10:28,000 --> 00:10:29,597 point for C-Pilot. We started by 186 00:10:29,680 --> 00:10:33,597 bringing together chat and co-worker. 187 00:10:33,680 --> 00:10:36,237 We also introduced Office in Copilot, 188 00:10:36,320 --> 00:10:38,797 which brings full rich document editing 189 00:10:38,880 --> 00:10:41,277 and collaboration from Office directly 190 00:10:41,360 --> 00:10:44,877 into the Copilot app. And today, Home 191 00:10:44,960 --> 00:10:49,197 gets even better with hybrid intelligence. Home now knows what's on 192 00:10:49,280 --> 00:10:53,997 my PC. And so the exact same local files that I've been working on locally myself 193 00:10:54,080 --> 00:10:57,837 can now be turned into collaboration ready artifacts directly inside of 194 00:10:57,920 --> 00:11:00,077 Copilot. 195 00:11:00,160 --> 00:11:03,197 Next, let's talk about code. 196 00:11:03,280 --> 00:11:05,837 Code is the easiest way for anyone to 197 00:11:05,920 --> 00:11:10,237 build and share an app. No developer experience required. 198 00:11:10,320 --> 00:11:12,717 Hybrid intelligence makes code better in 199 00:11:12,800 --> 00:11:16,557 three key ways. First, you can now build 200 00:11:16,640 --> 00:11:20,637 a native Windows app from a single prompt. 201 00:11:20,720 --> 00:11:25,437 This means unlimited apps and widgets all native on Windows directly from the 202 00:11:25,520 --> 00:11:27,277 C-Pilot app. Building software for 203 00:11:27,360 --> 00:11:30,397 Windows has never been this easy. 204 00:11:30,480 --> 00:11:33,837 Next, we introduced MXC support. 205 00:11:33,920 --> 00:11:38,157 MXC support means that your code, even locally, is running in the sandbox that 206 00:11:38,240 --> 00:11:42,317 we heard about earlier. This means for the first time, not just developers, but 207 00:11:42,400 --> 00:11:44,797 anyone can write and run code on their 208 00:11:44,880 --> 00:11:47,997 local machine with confidence. 209 00:11:48,080 --> 00:11:50,797 And lastly, local models. 210 00:11:50,880 --> 00:11:53,117 Copilot is now able to hand work down to 211 00:11:53,200 --> 00:11:55,597 a local model as it's coding. This means 212 00:11:55,680 --> 00:11:58,317 that projects move faster and cost less 213 00:11:58,400 --> 00:12:01,437 just like we saw in GitHub Copilot. 214 00:12:01,520 --> 00:12:04,477 These three things 215 00:12:04,560 --> 00:12:07,357 into your personal software factory. We 216 00:12:07,440 --> 00:12:10,477 really can't wait to see what you build. 217 00:12:10,560 --> 00:12:14,237 Lastly, I want to talk about autopilot. 218 00:12:14,320 --> 00:12:17,320 Autopilot. 219 00:12:17,680 --> 00:12:21,917 Enterprisegrade longunning autonomous agent. 220 00:12:22,000 --> 00:12:25,757 Let's see how hybrid intelligence makes autopilot even better. We'll flip over 221 00:12:25,840 --> 00:12:29,997 to a quick demo. So, over the last couple weeks, maybe a 222 00:12:30,080 --> 00:12:33,277 little bit late, I was working with Autopilot to pull together a bunch of 223 00:12:33,360 --> 00:12:36,797 stuff to file my taxes. And so, I'll go ahead and just kick off this simple 224 00:12:36,880 --> 00:12:39,997 prompt. And the first thing that Autopilot will do is it'll go into my 225 00:12:40,080 --> 00:12:43,357 inbox and it'll find the email from my accountant and it'll look for the 226 00:12:43,440 --> 00:12:46,717 specific list of files that we have to pull together. 227 00:12:46,800 --> 00:12:50,637 Now, because of hybrid intelligence, Autopilot can now ground against my 228 00:12:50,720 --> 00:12:54,317 entire PC. And the documents I need to file for my taxes are all over the 229 00:12:54,400 --> 00:12:57,197 place. They're in downloads. They're in documents. Half are organized, half 230 00:12:57,280 --> 00:13:01,037 aren't. That doesn't matter. Autopilot's able to use hybrid intelligence to 231 00:13:01,120 --> 00:13:05,437 ground over absolutely everything with local context. With my permission, it 232 00:13:05,520 --> 00:13:08,957 starts to organize. It moves everything into a new folder, pulls things 233 00:13:09,040 --> 00:13:13,117 together, and actually starts to rename files. Now, all of this organization 234 00:13:13,200 --> 00:13:17,117 renaming is actually happening with local models. my tax details are not 235 00:13:17,200 --> 00:13:20,637 going to the cloud. It creates some subfolders here to give a little bit 236 00:13:20,720 --> 00:13:24,717 more structure and then it finishes by zipping everything that it found into a 237 00:13:24,800 --> 00:13:29,277 single zip file that I can easily send to my accountant. Now, autopilot of 238 00:13:29,360 --> 00:13:32,797 course always goes one step further uh and it also actually gohead goes ahead 239 00:13:32,880 --> 00:13:37,197 and composes an email for me to send directly back to my accountant with the 240 00:13:37,280 --> 00:13:38,957 zip file that I just created attached. 241 00:13:39,040 --> 00:13:42,397 But of course, it pauses here and it makes sure that I'm good to go. I review 242 00:13:42,480 --> 00:13:46,237 everything. This looks great. We can go ahead and send the email. And just like 243 00:13:46,320 --> 00:13:50,317 that, all of these files are now sent off and in my accountant's hands and 244 00:13:50,400 --> 00:13:56,500 we're good to go. So that's Windows [applause] 245 00:14:00,160 --> 00:14:03,677 hybrid. C-Pilot now truly understands 246 00:14:03,760 --> 00:14:07,677 your PC. It's able to take action on your computer. It's able to keep 247 00:14:07,760 --> 00:14:11,837 sensitive work on your device. And at every single step, you're in full 248 00:14:11,920 --> 00:14:13,357 control. 249 00:14:13,440 --> 00:14:16,397 Really, it's C-Pilot built natively for 250 00:14:16,480 --> 00:14:20,637 Windows. Copilot features that are powered by hybrid intelligence are 251 00:14:20,720 --> 00:14:24,237 coming to C-Pilot over the next couple months. And with that, I'll hand things 252 00:14:24,320 --> 00:14:29,900 back to Pavin. [applause] 253 00:14:32,560 --> 00:14:34,477 Thank you, Jacob. We're excited to bring 254 00:14:34,560 --> 00:14:37,197 hybrid intelligence to Copilot. 255 00:14:37,280 --> 00:14:41,677 Windows is the largest open enduser computing platform. So along with great 256 00:14:41,760 --> 00:14:45,757 Microsoft experiences, our broader ecosystem is also taking advantage of 257 00:14:45,840 --> 00:14:49,837 Windows platform capabilities. We're seeing popular agents like Hermes, 258 00:14:49,920 --> 00:14:52,397 OpenClaw, and Perplexity building local 259 00:14:52,480 --> 00:14:57,757 AI experiences on Windows. Here's an example from Perplexity. Let's 260 00:14:57,840 --> 00:15:00,717 say you're an aspiring photographer trying to get off the ground, and you 261 00:15:00,800 --> 00:15:04,957 have a folder of photos from a recent photo shoot that you want to turn into a 262 00:15:05,040 --> 00:15:07,357 professional gallery for your client. 263 00:15:07,440 --> 00:15:09,597 Normally, that means you'd be paying for 264 00:15:09,680 --> 00:15:12,317 a service, resizing images, building a 265 00:15:12,400 --> 00:15:17,117 page, bringing it all together. That takes multiple steps and multiple tools. 266 00:15:17,200 --> 00:15:21,277 With Perplexity's portable computer on Windows, you can just describe what you 267 00:15:21,360 --> 00:15:24,077 want. Resize the photos and build a 268 00:15:24,160 --> 00:15:28,877 local gallery. Perplexity takes it from there, finding the photos, checking the 269 00:15:28,960 --> 00:15:33,437 tools on my PC, and building the gallery. You define the outcome, and 270 00:15:33,520 --> 00:15:35,597 Perplexity figures out the next steps. 271 00:15:35,680 --> 00:15:38,157 The model is running locally on the PC. 272 00:15:38,240 --> 00:15:40,157 Processing stays on the device without 273 00:15:40,240 --> 00:15:42,637 using Proplexity computer credits. And 274 00:15:42,720 --> 00:15:46,557 here it is. 12 photos in a browsable 275 00:15:46,640 --> 00:15:49,997 gallery complete with titles and image 276 00:15:50,080 --> 00:15:52,557 details. 277 00:15:52,640 --> 00:15:55,517 We're also excited to announce Muse is 278 00:15:55,600 --> 00:16:01,610 coming to Windows as a native app with MXE integration. [applause] 279 00:16:05,120 --> 00:16:08,477 To achieve our vision of bringing unmetered intelligence to every desk in 280 00:16:08,560 --> 00:16:10,557 every home, you need a broad range of 281 00:16:10,640 --> 00:16:13,277 PCs and silicon from a mini desktop PC 282 00:16:13,360 --> 00:16:16,237 to a commercial laptop to a supercomput. 283 00:16:16,320 --> 00:16:20,237 Let's start with the mini desktop PCs or claw boxes as people call them which 284 00:16:20,320 --> 00:16:22,877 millions of people use today. It's a 285 00:16:22,960 --> 00:16:27,917 natural home for a device and an agent that needs an always on device. An agent 286 00:16:28,000 --> 00:16:29,597 that keeps working even when you step 287 00:16:29,680 --> 00:16:34,957 away. Today, we're excited to announce a new get started experience in Windows. 288 00:16:35,040 --> 00:16:38,957 I'm excited for this because it's something customers are asking for. A 289 00:16:39,040 --> 00:16:43,437 simpler, more secure way to get started running agents. 290 00:16:43,520 --> 00:16:45,597 This a this experience makes setting up 291 00:16:45,680 --> 00:16:48,477 popular agents fast, just a few clicks 292 00:16:48,560 --> 00:16:53,117 right out of the box. No terminal, no complicated configuration. And for those 293 00:16:53,200 --> 00:16:57,757 of you using OpenClaw, we're introducing a native Windows gateway with a simpler 294 00:16:57,840 --> 00:17:01,677 setup. And because all the popular agents are adopting MXC, you can now get 295 00:17:01,760 --> 00:17:03,437 these agents running securely on 296 00:17:03,520 --> 00:17:06,877 Windows. This is just the beginning. As 297 00:17:06,960 --> 00:17:11,437 we continue to learn from our customers, our ambition is to bring secure agentic 298 00:17:11,520 --> 00:17:13,437 experiences to the billion people who 299 00:17:13,520 --> 00:17:17,517 use Windows every day. Next, Copile Plus 300 00:17:17,600 --> 00:17:20,877 PCs, your PC for everyday AI. We 301 00:17:20,960 --> 00:17:23,437 introduced these thin, light, powerful 302 00:17:23,520 --> 00:17:25,997 PCs two years ago with power efficient 303 00:17:26,080 --> 00:17:29,517 NPUs that run everyday tasks locally. 304 00:17:29,600 --> 00:17:33,437 Today, over 40% of the laptops being built for business are Copilot Plus PCs, 305 00:17:33,520 --> 00:17:37,437 tens of millions of PCs a year. And we're seeing over two trillion 306 00:17:37,520 --> 00:17:39,517 inferences per month running locally, 307 00:17:39,600 --> 00:17:44,637 improving customers experiences across search, imaging, and video calls. The 308 00:17:44,720 --> 00:17:46,317 hybrid intelligence experiences you just 309 00:17:46,400 --> 00:17:48,477 saw from Jacob will take advantage of 310 00:17:48,560 --> 00:17:52,317 local AI on Copilot Plus PCs. 311 00:17:52,400 --> 00:17:56,957 And then there's the builder tier. To create a new class of devices for 312 00:17:57,040 --> 00:17:59,197 builders, we needed to understand what 313 00:17:59,280 --> 00:18:01,757 matters to them and push the boundaries 314 00:18:01,840 --> 00:18:04,317 of hardware engineering. I want to 315 00:18:04,400 --> 00:18:07,037 introduce you to Surface Laptop Ultra, 316 00:18:07,120 --> 00:18:13,480 the most powerful Surface we've ever built. Let's take a look. 317 00:18:35,120 --> 00:18:38,637 I [music] want a little bit of something 318 00:18:38,720 --> 00:18:43,907 so good. I set a fire to watch it burn. 319 00:18:43,990 --> 00:18:46,010 [music] 320 00:18:51,850 --> 00:18:53,870 [music] 321 00:18:57,090 --> 00:18:59,110 [music] 322 00:19:07,680 --> 00:19:10,680 Heat. Heat. 323 00:19:18,050 --> 00:19:20,070 [music] 324 00:19:23,290 --> 00:19:25,310 >> [music] 325 00:19:37,050 --> 00:19:40,797 [cheering] >> This [applause] 326 00:19:40,880 --> 00:19:43,917 is Surface Laptop Ultra. 327 00:19:44,000 --> 00:19:46,637 We built it for developers and creators 328 00:19:46,720 --> 00:19:48,797 pushing the limits of performance who 329 00:19:48,880 --> 00:19:52,397 want portability with exceptional craft. 330 00:19:52,480 --> 00:19:54,557 I've been using Surface Laptop Ultra for 331 00:19:54,640 --> 00:19:56,797 the past few months and it's an absolute 332 00:19:56,880 --> 00:19:59,677 gamecher. Developers push their machines 333 00:19:59,760 --> 00:20:01,997 hard. Here's an intensive developer 334 00:20:02,080 --> 00:20:04,717 workflow with GitHub Copilot. But a 335 00:20:04,800 --> 00:20:06,957 demanding task shouldn't mean putting 336 00:20:07,040 --> 00:20:09,517 everything else on hold. While that work 337 00:20:09,600 --> 00:20:11,917 continues, a film is being mastered in 338 00:20:12,000 --> 00:20:14,797 Da Vinci Resolve. This kind of mastering 339 00:20:14,880 --> 00:20:19,357 takes serious performance. The CPU, the GPU, and memory all working together 340 00:20:19,440 --> 00:20:21,277 while the developer workload is still 341 00:20:21,360 --> 00:20:24,557 running. And 342 00:20:24,640 --> 00:20:27,437 Gears of War, which released yesterday, 343 00:20:27,520 --> 00:20:31,987 has been running on this machine the entire time. And it runs awesome. 344 00:20:32,070 --> 00:20:34,090 [applause] 345 00:20:36,320 --> 00:20:38,717 It runs simply awesome at 60 frames a 346 00:20:38,800 --> 00:20:42,637 second. Work, create, and play. So you 347 00:20:42,720 --> 00:20:47,277 don't have to choose just one. Now, it wouldn't be Windows and Nvidia without 348 00:20:47,360 --> 00:20:49,597 talking a little bit more about gaming. 349 00:20:49,680 --> 00:20:52,477 At launch, hundreds of AAA titles that 350 00:20:52,560 --> 00:20:55,117 you love and the indies that you play 351 00:20:55,200 --> 00:20:58,397 will be great on RTX Spark. 352 00:20:58,480 --> 00:21:00,637 And we're super excited to announce that 353 00:21:00,720 --> 00:21:07,490 in 2027, Call of Duty will be coming to RTX Spark. [applause] 354 00:21:11,680 --> 00:21:14,477 Getting here took Surface, Windows, and 355 00:21:14,560 --> 00:21:17,357 the Nvidia team building a new platform 356 00:21:17,440 --> 00:21:19,997 from the ground up. When three teams 357 00:21:20,080 --> 00:21:24,717 come together, you get something none of us could have built alone. And for that, 358 00:21:24,800 --> 00:21:26,957 I am deeply grateful. 359 00:21:27,040 --> 00:21:29,677 We built Surface Laptop Ultra around 360 00:21:29,760 --> 00:21:33,437 Nvidia RTX Spark. It took end-to-end 361 00:21:33,520 --> 00:21:35,917 engineering, design, and full stack 362 00:21:36,000 --> 00:21:39,197 thinking to unlock what local AI can do. 363 00:21:39,280 --> 00:21:44,397 With up to 128 gigs of unified memory and up to a pedlop of AI compute, you 364 00:21:44,480 --> 00:21:47,597 can run models on this laptop that simply don't fit on a traditional 365 00:21:47,680 --> 00:21:51,437 machine. Surface Laptop Ultra is engineered 366 00:21:51,520 --> 00:21:54,557 beautifully inside and out. It has more 367 00:21:54,640 --> 00:21:58,477 than two and a half times the thermal capacity of our existing laptops. So, it 368 00:21:58,560 --> 00:22:01,437 stays faster even under sustained load. 369 00:22:01,520 --> 00:22:04,157 And you don't have to worry about sacrificing performance when you're 370 00:22:04,240 --> 00:22:06,557 unplugged on battery. It can maintain 371 00:22:06,640 --> 00:22:09,837 over 99% of its power while on battery 372 00:22:09,920 --> 00:22:11,917 power. And that's great because it's 373 00:22:12,000 --> 00:22:14,957 portable, under 18 mm thin, and less 374 00:22:15,040 --> 00:22:17,437 than 4 12 lb. 375 00:22:17,520 --> 00:22:20,477 I also love our new Pixel Sense Ultra 376 00:22:20,560 --> 00:22:23,597 Touch display. It has the brightest peak 377 00:22:23,680 --> 00:22:27,197 HDR of any laptop. It looks incredible. 378 00:22:27,280 --> 00:22:29,277 Our Surface Laptop Ultra has the most 379 00:22:29,360 --> 00:22:32,237 ports on a Surface Laptop. Our customers 380 00:22:32,320 --> 00:22:35,277 have told us they love magnetic charging 381 00:22:35,360 --> 00:22:37,837 and they love the versatility of USBC. 382 00:22:37,920 --> 00:22:41,597 So, we invented magnetic connect, the 383 00:22:41,680 --> 00:22:46,267 world's first built-in magnetic USBC charging. 384 00:22:46,350 --> 00:22:51,080 [applause] I know. 385 00:22:51,440 --> 00:22:56,077 Just bring the connector close and it snaps right into place. Easy magnetic 386 00:22:56,160 --> 00:22:59,837 charging without giving up the USBC port. 387 00:22:59,920 --> 00:23:02,637 But that's right. This kind of awesome. 388 00:23:02,720 --> 00:23:04,740 [applause] 389 00:23:05,840 --> 00:23:09,357 But the real measure of all of this power is what happens when it's in your 390 00:23:09,440 --> 00:23:13,357 hands. When you can close the distance between what you can imagine and what 391 00:23:13,440 --> 00:23:16,317 you can create, that is amazing. That 392 00:23:16,400 --> 00:23:20,557 matters because the future of AI is not about having just one model. It's about 393 00:23:20,640 --> 00:23:24,157 having the right model at the right moment. Frontier capability when you 394 00:23:24,240 --> 00:23:28,557 need it, smaller, faster models when you want speed, responsiveness, and the room 395 00:23:28,640 --> 00:23:30,397 to keep the rest of your workflow 396 00:23:30,480 --> 00:23:33,677 running. Surface Laptop Ultra is built 397 00:23:33,760 --> 00:23:37,517 for that future. It brings new ideas to life with 398 00:23:37,600 --> 00:23:39,997 advanced graphics, unified memory, and 399 00:23:40,080 --> 00:23:42,957 local AI working together. 400 00:23:43,040 --> 00:23:45,517 Let me show you how one of our designers 401 00:23:45,600 --> 00:23:48,477 put all of that power to work, combining 402 00:23:48,560 --> 00:23:51,197 local and cloud AI to bring his creative 403 00:23:51,280 --> 00:23:53,437 vision to life. 404 00:23:53,520 --> 00:23:56,077 >> Hi, I'm Danilo and I'm a senior designer 405 00:23:56,160 --> 00:24:01,277 at Microsoft. In my day-to-day work, I'm building products, but outside of work, 406 00:24:01,360 --> 00:24:05,437 I have a true passion for creating concept art. I love creating concept 407 00:24:05,520 --> 00:24:07,917 vehicles in sci-fi worlds and I'm about 408 00:24:08,000 --> 00:24:09,997 to show you a workflow that used to take 409 00:24:10,080 --> 00:24:15,720 me days and now just takes me a few hours. 410 00:24:17,120 --> 00:24:19,037 One of my favorite tools is Blender. 411 00:24:19,120 --> 00:24:22,877 Here I'm building a spaceship from scratch, modeling all the details, 412 00:24:22,960 --> 00:24:27,037 texturing, lighting, and rendering. I could keep going, but if I wanted to 413 00:24:27,120 --> 00:24:30,317 bring more of my story to life and create an entire world around the 414 00:24:30,400 --> 00:24:32,557 spaceship, it would take me days or even 415 00:24:32,640 --> 00:24:34,957 weeks. That's where local AI can be 416 00:24:35,040 --> 00:24:37,357 really helpful. It lets me explore ideas 417 00:24:37,440 --> 00:24:40,237 faster. Here I'm using Confi, an app 418 00:24:40,320 --> 00:24:44,477 that lets me download and run AI models locally. I started by importing some of 419 00:24:44,560 --> 00:24:49,837 my renderers as reference images and use a Gemma 4, an image to text model to 420 00:24:49,920 --> 00:24:53,117 generate [music] a description of the world and a list of assets that match 421 00:24:53,200 --> 00:24:55,997 the style and elements of my images. 422 00:24:56,080 --> 00:24:57,677 Then I imported another image with 423 00:24:57,760 --> 00:24:59,997 close-up details as reference along with 424 00:25:00,080 --> 00:25:03,277 asset descriptions generated by Gemma 4. 425 00:25:03,360 --> 00:25:07,517 Using Flux 2D dev, an image [music] to image generation model, I create an 426 00:25:07,600 --> 00:25:11,677 image of each asset. After running several generations, I end up with 427 00:25:11,760 --> 00:25:16,397 [music] a folder full of assets. Next, I imported my favorite images into Hoon 1 428 00:25:16,480 --> 00:25:20,317 3D 2.1, [music] an image to 3D model to generate the 3D 429 00:25:20,400 --> 00:25:24,777 assets. I then brought all these assets into Blender so I can start building 430 00:25:24,860 --> 00:25:29,437 [music] my scene. Using a highly capable coding model running fully on my device, 431 00:25:29,520 --> 00:25:32,957 I wrote a [music] basic prompt and generated several layout options using 432 00:25:33,040 --> 00:25:36,957 my assets. After choosing the one I liked, I kept 433 00:25:37,040 --> 00:25:40,637 exploring and refining my scene, adding more elements and tweaking the overall 434 00:25:40,720 --> 00:25:44,157 composition. Once I was happy with it, [music] I 435 00:25:44,240 --> 00:25:48,077 rendered the scene, brought the images into Photoshop, and painting over areas 436 00:25:48,160 --> 00:25:50,557 I wanted to refine, adding elements like 437 00:25:50,640 --> 00:25:52,637 smoke and adjusting [music] the overall 438 00:25:52,720 --> 00:25:55,437 color and mood. Then I took the finished 439 00:25:55,520 --> 00:25:57,837 image back into coffee and used [music] 440 00:25:57,920 --> 00:26:00,797 one 2.2 an image to video model to 441 00:26:00,880 --> 00:26:04,317 generate the final animation. 442 00:26:04,400 --> 00:26:07,037 In my final step, I used GPT6 [music] 443 00:26:07,120 --> 00:26:09,357 Astra in the cloud to take things even 444 00:26:09,440 --> 00:26:11,917 further. 445 00:26:12,000 --> 00:26:16,797 I uploaded my final files and asked the agent to build a scene in Unreal Engine 446 00:26:16,880 --> 00:26:20,477 that could actually walk and fly around [music] in. 447 00:26:20,560 --> 00:26:24,477 So this is the result. This would normally take me days, but now only 448 00:26:24,560 --> 00:26:30,200 takes hours. All on my Surface Laptop Ultra. 449 00:26:30,360 --> 00:26:32,380 [applause] 450 00:26:36,240 --> 00:26:38,797 What you just saw is Dino using multiple 451 00:26:38,880 --> 00:26:43,357 local and cloud models in a variety of different apps, bringing his creative 452 00:26:43,440 --> 00:26:47,597 vision to life, shaping the design, refining the details, staying in control 453 00:26:47,680 --> 00:26:49,917 of his entire workflow. What used to 454 00:26:50,000 --> 00:26:52,957 take Danello days now takes hours. 455 00:26:53,040 --> 00:26:57,117 That's what we mean by hybrid intelligence. The new Surface Laptop 456 00:26:57,200 --> 00:27:01,357 Ultra is built for this new era of creative work. Across the models we 457 00:27:01,440 --> 00:27:05,117 tested, Surface Laptop Ultra delivers faster time to first token performance 458 00:27:05,200 --> 00:27:07,437 for language models and shorter image 459 00:27:07,520 --> 00:27:09,677 generation time than M5 Pro powered 460 00:27:09,760 --> 00:27:12,717 laptops. This is where having both the 461 00:27:12,800 --> 00:27:17,837 large memory and the pedlop of AI compute on a single device is a distinct 462 00:27:17,920 --> 00:27:21,757 advantage. For builders using local AI, that means 463 00:27:21,840 --> 00:27:26,157 less waiting and more creating on Windows. 464 00:27:26,240 --> 00:27:28,317 Since build, many of you asked about the 465 00:27:28,400 --> 00:27:32,397 Surface RTX Spark Devox for built for 466 00:27:32,480 --> 00:27:36,237 those of you who want to push local AI development even further. It ships with 467 00:27:36,320 --> 00:27:40,477 Project Zenith, a ready to code Windows experience preconfigured out of the box 468 00:27:40,560 --> 00:27:45,037 so you can build from day one. I'm excited to share that both Surface 469 00:27:45,120 --> 00:27:47,437 Laptop Ultra and the Dev Box are 470 00:27:47,520 --> 00:27:50,837 available today for pre-order. 471 00:27:50,920 --> 00:27:52,940 [applause] 472 00:27:56,320 --> 00:27:59,517 In addition to Surface, we've built a 473 00:27:59,600 --> 00:28:02,957 broad device ecosystem. NVIDIA has been 474 00:28:03,040 --> 00:28:06,557 our close partner on this journey. RTX 475 00:28:06,640 --> 00:28:08,877 Spark brings powerful AI compute to the 476 00:28:08,960 --> 00:28:12,397 laptop. DGX station takes it even 477 00:28:12,480 --> 00:28:15,437 further. Look at the ecosystem bringing 478 00:28:15,520 --> 00:28:19,757 this to life. Lenovo Yoga 92in1, a 479 00:28:19,840 --> 00:28:22,157 lightweight convertible design with pen 480 00:28:22,240 --> 00:28:25,437 support. Asus Pro RP14, an incredibly 481 00:28:25,520 --> 00:28:29,437 thin design with a 4K OLED display. 482 00:28:29,520 --> 00:28:32,877 Dell XPS16, a seamless glass haptic 483 00:28:32,960 --> 00:28:36,157 touchpad with integrated speakers. MSI 484 00:28:36,240 --> 00:28:38,717 Prestige N16, a versatile flip form 485 00:28:38,800 --> 00:28:41,917 factor with pen and UHD plus OLED 486 00:28:42,000 --> 00:28:43,437 display, 487 00:28:43,520 --> 00:28:47,197 HP Omnibook Ultra 16 with a distinctive 488 00:28:47,280 --> 00:28:49,437 tower hinge design and gesture controls 489 00:28:49,520 --> 00:28:52,077 on the haptic touchpad. And of course, 490 00:28:52,160 --> 00:28:55,117 the Surface Laptop Ultra. All of these 491 00:28:55,200 --> 00:28:57,357 RTX Spark laptops are available for 492 00:28:57,440 --> 00:28:59,677 pre-order starting today and will begin 493 00:28:59,760 --> 00:29:04,990 shipping next Friday. [applause] 494 00:29:07,600 --> 00:29:11,437 And finally, we have the supercomputer. 495 00:29:11,520 --> 00:29:14,397 DGX station for Windows brings data 496 00:29:14,480 --> 00:29:18,877 center class performance together with enterprisegrade security and 497 00:29:18,960 --> 00:29:22,797 manageability. This unlocks the power to run frontier 498 00:29:22,880 --> 00:29:25,277 class models locally from Llama V4 499 00:29:25,360 --> 00:29:28,077 Maverick to Kimmy K2.6 to Deepseek V4 500 00:29:28,160 --> 00:29:31,357 Pro at more than a trillion parameters. 501 00:29:31,440 --> 00:29:36,477 Capabilities that once required renting a cluster now in a deskside supercomput. 502 00:29:36,560 --> 00:29:40,957 With RTX Spark, we're bringing last year's Frontier to your laptop. With DGX 503 00:29:41,040 --> 00:29:45,357 Station, we're bringing last quarter's Frontier to your desk side. Take a look 504 00:29:45,440 --> 00:29:48,317 at this. [applause] 505 00:29:48,400 --> 00:29:50,877 Take a look at this. For the first time, 506 00:29:50,960 --> 00:29:55,837 DGX Station running Windows Live 507 00:29:55,920 --> 00:29:58,237 with Dell Precision Pro. Frontier models 508 00:29:58,320 --> 00:30:00,397 and agents work locally, taking months 509 00:30:00,480 --> 00:30:03,677 of work, complex work into moments with 510 00:30:03,760 --> 00:30:05,997 more control over sensitive data and 511 00:30:06,080 --> 00:30:10,237 cost. With the HP ZGX Fury AI station, 512 00:30:10,320 --> 00:30:15,197 it shows what happens can work and scale simultaneously across 32 agents, a 513 00:30:15,280 --> 00:30:19,917 virtual token factory for development teams. Both coming later this year, 514 00:30:20,000 --> 00:30:24,767 running the Windows platform organizations know and trust. 515 00:30:24,850 --> 00:30:26,870 [applause] 516 00:30:29,520 --> 00:30:33,677 This is an amazing moment. On behalf of Microsoft, I want to say thank you to 517 00:30:33,760 --> 00:30:38,797 all of our partners. Next, Satya Jensen and Shrian Krishna 518 00:30:38,880 --> 00:30:40,237 are going to take the stage in a few 519 00:30:40,320 --> 00:30:46,440 moments to share their thoughts on the opportunity ahead. Thank you. 520 00:30:50,640 --> 00:30:54,397 >> All full of power as we go. Chase the 521 00:30:54,480 --> 00:30:58,557 thing we came here for. Push it further than before. We'll be coming around the 522 00:30:58,640 --> 00:31:02,797 thought we [music] can't let go. We'll be coming around the thought we won't 523 00:31:02,880 --> 00:31:05,437 let go. Making brick and shaping 524 00:31:05,520 --> 00:31:09,307 crafting as we go. Then the pieces 525 00:31:09,390 --> 00:31:15,480 [music] comes to life. Let's see how far. 526 00:31:19,680 --> 00:31:25,240 >> All right. [applause] All right. 527 00:31:25,770 --> 00:31:27,790 >> [applause] 528 00:31:28,240 --> 00:31:32,840 >> Wait, wait, wait. Take the phone. 529 00:31:34,560 --> 00:31:38,270 >> No phone calls on stage here. [laughter] 530 00:31:39,360 --> 00:31:41,837 Um, first off, was that cool or what? 531 00:31:41,920 --> 00:31:46,130 What do you guys think? Huh? >> Yeah. [applause] 532 00:31:47,280 --> 00:31:52,557 >> Huge congratulations to Paw, Jacob, the teams at Microsoft and Nvidia. Um, I 533 00:31:52,640 --> 00:31:55,757 will say they were kind enough to send me a Surface Laptop Ultra for the last 534 00:31:55,840 --> 00:31:59,747 few days. It is amazing, but I think I want the workstation for Christmas. Now, 535 00:31:59,830 --> 00:32:02,957 [laughter] now there's a bunch of stuff about the 536 00:32:03,040 --> 00:32:05,917 exciting future to get into. But before 537 00:32:06,000 --> 00:32:08,477 we do that, I was looking both of you up 538 00:32:08,560 --> 00:32:10,717 last night on YouTube, and it's quite 539 00:32:10,800 --> 00:32:13,117 rare to see both of these gentlemen on 540 00:32:13,200 --> 00:32:15,677 stage together, and I was thinking about 541 00:32:15,760 --> 00:32:20,317 how, you know, you both had a history, these two great companies had a history 542 00:32:20,400 --> 00:32:23,757 going back decades. You both individually had histories of working 543 00:32:23,840 --> 00:32:28,077 with each other going back decades. Ties to me. My first job was at Microsoft. My 544 00:32:28,160 --> 00:32:30,317 first paycheck a GeForce FX to play 545 00:32:30,400 --> 00:32:34,477 Halfife 2. So, so maybe to start us off 546 00:32:34,560 --> 00:32:39,757 and Jensen, what is your favorite memory uh or fun memory of working with Satia 547 00:32:39,840 --> 00:32:41,037 and Microsoft? 548 00:32:41,120 --> 00:32:44,397 >> Well, uh Nvidia was founded because of 549 00:32:44,480 --> 00:32:47,677 Windows. It was 1992, late '92, uh, 550 00:32:47,760 --> 00:32:50,557 Windows 3.1. And we were just started to 551 00:32:50,640 --> 00:32:55,197 see the the vision of what a what a personal computer really can be. And at 552 00:32:55,280 --> 00:32:59,277 the time, frankly, we had never I'd never even see a PC before. I largely 553 00:32:59,360 --> 00:33:03,277 worked on workstations. We had, you know, we had uh file servers and 554 00:33:03,360 --> 00:33:07,437 supercomputers that we used to design chips with. and and the founders and I, 555 00:33:07,520 --> 00:33:10,077 you know, talked about this future where 556 00:33:10,160 --> 00:33:12,717 maybe uh a personal computer with 3D 557 00:33:12,800 --> 00:33:15,037 graphics would enable it to do things 558 00:33:15,120 --> 00:33:18,317 like games and and be a workstation and 559 00:33:18,400 --> 00:33:22,877 design things. And so we we imagined about this world. If not for Windows, uh 560 00:33:22,960 --> 00:33:24,797 Nvidia would have been founded. And then 561 00:33:24,880 --> 00:33:26,797 and of course 562 00:33:26,880 --> 00:33:30,077 and of course we we started in 1993 563 00:33:30,160 --> 00:33:35,437 um uh Windows 95 really made reinvented 564 00:33:35,520 --> 00:33:39,757 Windows 95. >> Wow. 565 00:33:39,840 --> 00:33:42,077 Am I on 566 00:33:42,160 --> 00:33:45,677 >> Windows? Windows 95 567 00:33:45,760 --> 00:33:49,277 made possible the GPU to be connected to 568 00:33:49,360 --> 00:33:53,357 uh uh a personal computer. And what 569 00:33:53,440 --> 00:33:56,557 revolutionized the GPU was DirectX. 570 00:33:56,640 --> 00:34:00,317 Direct 3D uh was the API that really uh 571 00:34:00,400 --> 00:34:03,917 enabled us to re uh reveal the 572 00:34:04,000 --> 00:34:07,037 capabilities of GPUs into applications. 573 00:34:07,120 --> 00:34:09,437 But then DX8 came along and we 574 00:34:09,520 --> 00:34:11,997 co-invented with Microsoft the world's 575 00:34:12,080 --> 00:34:14,477 first G programmable GPU which is 576 00:34:14,560 --> 00:34:16,957 programmable shaders which a few steps 577 00:34:17,040 --> 00:34:21,837 later evolved into uh CUDA as we know it 578 00:34:21,920 --> 00:34:23,677 >> and so if you look at the entire journey 579 00:34:23,760 --> 00:34:28,477 of our company Windows was at the core of it and if not for Windows there would 580 00:34:28,560 --> 00:34:30,477 be no GeForce if no GeForce CUDA 581 00:34:30,560 --> 00:34:32,557 wouldn't come to the world uh without 582 00:34:32,640 --> 00:34:36,797 CUDA Alex Kresevki and Ilia Suscover um 583 00:34:36,880 --> 00:34:39,917 Yan Lakun, you know, uh Andrew Ang, none 584 00:34:40,000 --> 00:34:45,277 of them would have uh found a computer uh to do deep learning. 585 00:34:45,360 --> 00:34:48,397 >> And here we are. >> And Jensen is now going to do a little 586 00:34:48,480 --> 00:34:50,697 transfer of market cap back to me. 587 00:34:50,780 --> 00:34:52,800 [laughter] 588 00:34:53,600 --> 00:34:56,237 >> I do that every day. I do that every day. 589 00:34:56,320 --> 00:34:59,917 >> But it's extraordinary that journey uh 590 00:35:00,000 --> 00:35:01,597 that Jensen recounted. I mean, it's kind 591 00:35:01,680 --> 00:35:04,557 of the the it's the history of computing 592 00:35:04,640 --> 00:35:09,117 in some sense. Um, and the thing, one of 593 00:35:09,200 --> 00:35:12,477 the things, uh, that Jensen sort of 594 00:35:12,560 --> 00:35:14,877 sparked for me was 595 00:35:14,960 --> 00:35:19,117 he has been perhaps the most consistent 596 00:35:19,200 --> 00:35:22,877 with the vision of what accelerated 597 00:35:22,960 --> 00:35:25,677 computing can be. And this going back to 598 00:35:25,760 --> 00:35:27,997 the trajectory, you've been the one 599 00:35:28,080 --> 00:35:30,157 person in our industry who has had this 600 00:35:30,240 --> 00:35:34,797 one idea whose sort of trajectory has 601 00:35:34,880 --> 00:35:37,437 been the most expansive quite frankly 602 00:35:37,520 --> 00:35:40,237 more so than even any single PC form 603 00:35:40,320 --> 00:35:43,037 factor or chip architecture. Because I 604 00:35:43,120 --> 00:35:46,717 look back, one of my first memories was 605 00:35:46,800 --> 00:35:49,517 we were talking about essentially saying 606 00:35:49,600 --> 00:35:52,797 in the cloud, how big can be accelerated 607 00:35:52,880 --> 00:35:54,317 computing or high performance computing, 608 00:35:54,400 --> 00:35:57,357 right? The first deal I remember was in 609 00:35:57,440 --> 00:36:01,197 Azure, how do we make HPC a bigger 610 00:36:01,280 --> 00:36:04,237 thing? Um and by the way if 611 00:36:04,320 --> 00:36:07,437 >> Satio was the first user of Infiniban 612 00:36:07,520 --> 00:36:10,797 inside a cloud which brought the first 613 00:36:10,880 --> 00:36:14,637 supercomputer to open AAI which boom >> that's right if we had not done that 614 00:36:14,720 --> 00:36:18,877 like the fascinating you know like life is full of accidents but if we had not 615 00:36:18,960 --> 00:36:21,517 done that for HPC 616 00:36:21,600 --> 00:36:26,237 we wouldn't have had the supercomputer to give to open AI to go train a GPT 617 00:36:26,320 --> 00:36:29,037 model and rest would have been history I mean it was kind of one of those happen 618 00:36:29,120 --> 00:36:31,997 stance that we had done the HPC see 619 00:36:32,080 --> 00:36:37,037 stuff when we were building an Infinibanbased supercomputer and and 620 00:36:37,120 --> 00:36:41,437 comes Sam and says, "Hey, do you have computing?" And we said, "Yeah, maybe." 621 00:36:41,520 --> 00:36:43,917 And and so it's sort of like but the 622 00:36:44,000 --> 00:36:46,957 thing that I really give Jensen all the 623 00:36:47,040 --> 00:36:50,637 credit quite frankly is that consistency 624 00:36:50,720 --> 00:36:54,477 of vision of what this can be, not 625 00:36:54,560 --> 00:36:57,517 today, not tomorrow, but as a secular 626 00:36:57,600 --> 00:37:00,397 trend. And that's kind of what you know 627 00:37:00,480 --> 00:37:02,477 brings us even to this moment. 628 00:37:02,560 --> 00:37:06,557 >> And then it led to a conversation. It led to a conversation I remember quite 629 00:37:06,640 --> 00:37:08,957 well where Satia and I were talking 630 00:37:09,040 --> 00:37:10,877 about so so what does this mean to the 631 00:37:10,960 --> 00:37:15,917 PC? How do we make uh the the perfect PC 632 00:37:16,000 --> 00:37:19,517 in the post era of all of this stuff and 633 00:37:19,600 --> 00:37:23,117 you know of course the personal computer is the ultimate tool. It's my ultimate 634 00:37:23,200 --> 00:37:24,557 tool and for a whole generation of 635 00:37:24,640 --> 00:37:26,957 people it's our ultimate tool. Um, what 636 00:37:27,040 --> 00:37:30,317 happens in the era of agents, uh, when 637 00:37:30,400 --> 00:37:33,117 when when the agent is on your computer, 638 00:37:33,200 --> 00:37:37,757 it is now your personal assistant, but it wants to have access to all of the 639 00:37:37,840 --> 00:37:39,917 amazing tools that we have, but the 640 00:37:40,000 --> 00:37:42,317 amazing tools are spread out in the 641 00:37:42,400 --> 00:37:46,317 NVIDIA ecosystem into uh gaming PCs, 642 00:37:46,400 --> 00:37:49,757 >> yeah, >> into workstations. And that's where most 643 00:37:49,840 --> 00:37:52,077 of the tools were, right? all the all 644 00:37:52,160 --> 00:37:53,997 the all the do tools and all the 645 00:37:54,080 --> 00:37:57,437 seammen's tools and uh autoesk and adobe 646 00:37:57,520 --> 00:38:01,757 and you saw many of them just now and you know we wanted the world's best 647 00:38:01,840 --> 00:38:05,197 blender the world's best unreal the world's best omniverse the world's best 648 00:38:05,280 --> 00:38:09,277 katillaa everything all sitting on the same PC 649 00:38:09,360 --> 00:38:14,637 >> um as the AI >> and so how do we go create that by the way took four years 650 00:38:14,720 --> 00:38:19,597 >> wow >> that was a four year ago conversation and just incredible 651 00:38:19,680 --> 00:38:23,837 >> now that's very a good segway today and Satia. So I think the last few months 652 00:38:23,920 --> 00:38:27,197 you know we've been sort of inundated with these stories of agents and 653 00:38:27,280 --> 00:38:29,517 security you know we all use agents I'm 654 00:38:29,600 --> 00:38:32,397 sure everybody here is using a muse or 655 00:38:32,480 --> 00:38:35,277 uh instinct and you are giving full disk 656 00:38:35,360 --> 00:38:37,037 access and you're having it act as you 657 00:38:37,120 --> 00:38:39,997 and today when you guys showed MXC it 658 00:38:40,080 --> 00:38:44,957 seemed to me like a fundamental new operating system primitive for agents 659 00:38:45,040 --> 00:38:46,637 which is a break from where we have 660 00:38:46,720 --> 00:38:49,677 been. to maybe talk to us about why the 661 00:38:49,760 --> 00:38:52,637 OS is the right place for this and what 662 00:38:52,720 --> 00:38:54,877 this primitive if you will opens up. 663 00:38:54,960 --> 00:38:57,917 >> Yeah. No, first of all remember this all 664 00:38:58,000 --> 00:39:01,517 started with coding agents all running 665 00:39:01,600 --> 00:39:03,597 on the desktop. I mean in fact if you 666 00:39:03,680 --> 00:39:06,717 sort of look at uh where is the agent 667 00:39:06,800 --> 00:39:10,397 that needs this security primitive the 668 00:39:10,480 --> 00:39:12,957 most? It's the local harness and its 669 00:39:13,040 --> 00:39:17,597 access to the file system and to be able to take action. And so that's why in 670 00:39:17,680 --> 00:39:19,677 fact you know working with Jensen and 671 00:39:19,760 --> 00:39:22,957 team we said look how do we w you know 672 00:39:23,040 --> 00:39:25,197 wire this in into the primitives whether 673 00:39:25,280 --> 00:39:28,397 it's in the cloud or in uh uh in the 674 00:39:28,480 --> 00:39:30,637 operating system locally and given the 675 00:39:30,720 --> 00:39:33,117 fact that essentially coding went there 676 00:39:33,200 --> 00:39:35,597 first where the harness is local even if 677 00:39:35,680 --> 00:39:37,757 you were using a cloud model we needed 678 00:39:37,840 --> 00:39:40,637 to make the desktop uh the most secure 679 00:39:40,720 --> 00:39:45,277 place for agents to execute and so that's where MXC comes from it's 680 00:39:45,360 --> 00:39:50,157 fantastic to see open shell now fully integrated and native with MXC the 681 00:39:50,240 --> 00:39:54,397 controls that Pavin talked about right which is containment itself has to be 682 00:39:54,480 --> 00:39:58,637 done at every level right it's not about sort of doing one thing and leaving the 683 00:39:58,720 --> 00:40:02,317 other right it has to be at the microVM it has to be at the WM it has to be at 684 00:40:02,400 --> 00:40:06,637 the session the piece that I'm most excited about is the Windows session 685 00:40:06,720 --> 00:40:10,957 having containment right it's so so critical then okay containment is 686 00:40:11,040 --> 00:40:14,077 wonderful but then I need full observability how do you get 687 00:40:14,160 --> 00:40:18,557 observability you get it if you have an identity for an agent that I can track I 688 00:40:18,640 --> 00:40:20,877 can then see the entire exhaust of the 689 00:40:20,960 --> 00:40:23,757 agent right all of it uh and then I can 690 00:40:23,840 --> 00:40:25,917 govern it that means I can like classic 691 00:40:26,000 --> 00:40:29,437 set policy and have it govern that is 692 00:40:29,520 --> 00:40:31,277 now what's going to give peace of mind 693 00:40:31,360 --> 00:40:34,797 whether it's to IT ops to sec ops uh and 694 00:40:34,880 --> 00:40:38,797 by the way and the other thing that we should also not forget is the ability to 695 00:40:38,880 --> 00:40:42,957 track the token spend right and where it's happening and so that's the 696 00:40:43,040 --> 00:40:47,037 entirety of what we have tried to together. It's not about any one thing. 697 00:40:47,120 --> 00:40:49,117 It's about stitching all these things. 698 00:40:49,200 --> 00:40:52,957 Having seen all these adoption cycles, whether it's the consumer adoption 699 00:40:53,040 --> 00:40:55,757 cycle, the enterprise adoption cycle, I 700 00:40:55,840 --> 00:40:57,917 would say the thing [clears throat] that 701 00:40:58,000 --> 00:41:00,877 blocks broad diffusion is not having 702 00:41:00,960 --> 00:41:05,197 that complete thought on what is required. Right? So if you do one and 703 00:41:05,280 --> 00:41:09,517 leave the other, there's no trust and what have you. So I feel very very good 704 00:41:09,600 --> 00:41:11,677 that we have brought it together in the 705 00:41:11,760 --> 00:41:14,877 cloud in the in the client across it 706 00:41:14,960 --> 00:41:19,437 sack and fin and it's heterogeneous right it's not about any one piece of 707 00:41:19,520 --> 00:41:21,757 Microsoft technology it's about supporting it all 708 00:41:21,840 --> 00:41:23,997 >> what Satia just said what Satia just 709 00:41:24,080 --> 00:41:27,677 said uh is going to be the foundation of 710 00:41:27,760 --> 00:41:32,477 the next generation of it as we know it if you go the words he used what he 711 00:41:32,560 --> 00:41:36,077 described the architecture deeply profound for the next generation of 712 00:41:36,160 --> 00:41:38,237 computers just as Windows and and 713 00:41:38,320 --> 00:41:41,117 DirectX revolutionized how applications 714 00:41:41,200 --> 00:41:44,237 were built. MXC 715 00:41:44,320 --> 00:41:47,997 is going to revolutionize how agents are 716 00:41:48,080 --> 00:41:51,597 built and deployed >> without it 717 00:41:51,680 --> 00:41:56,077 started without it complete non-starter and so this big big deal 718 00:41:56,160 --> 00:41:59,437 >> now this is I think closely tied to the next piece um for both of you which is 719 00:41:59,520 --> 00:42:02,317 hybrid intelligence now this feels a 720 00:42:02,400 --> 00:42:06,797 little different I think everybody here has uses large model in the cloud you 721 00:42:06,880 --> 00:42:10,557 use it it runs a tool maybe using blender and you get back a response 722 00:42:10,640 --> 00:42:12,237 we're all used to that the 723 00:42:12,320 --> 00:42:14,717 in a few years. This feels different 724 00:42:14,800 --> 00:42:19,517 where you have a choice of models, some running locally, some in the cloud, 725 00:42:19,600 --> 00:42:23,197 you're going back and forth. And one sentence from the presenter who spoke 726 00:42:23,280 --> 00:42:25,517 about GitHub stood out to me was you're 727 00:42:25,600 --> 00:42:27,597 getting these tokens for free, right? 728 00:42:27,680 --> 00:42:32,477 And so maybe talk to us about what this opens up because this feels again 729 00:42:32,560 --> 00:42:36,397 something new where you are running part of the workload here in the locally on 730 00:42:36,480 --> 00:42:40,317 your on your laptop, part of it in the cloud and intelligently routing between. 731 00:42:40,400 --> 00:42:41,837 What does this open up? Yeah. For first 732 00:42:41,920 --> 00:42:46,157 of all, you know, I think this is one of 733 00:42:46,240 --> 00:42:48,477 those moments where 734 00:42:48,560 --> 00:42:52,317 um some of the things that you've been 735 00:42:52,400 --> 00:42:55,117 using all click together, right? In 736 00:42:55,200 --> 00:42:56,477 fact, it's kind of I was thinking about 737 00:42:56,560 --> 00:43:01,117 this in the late 80s when I was in grad 738 00:43:01,200 --> 00:43:05,037 school. I kind of had a a a terminal, 739 00:43:05,120 --> 00:43:10,397 right? My VT0000 and a VAX. Um and um 740 00:43:10,480 --> 00:43:13,997 and then I had a PC at home. >> Mhm. 741 00:43:14,080 --> 00:43:16,477 >> Um and I kind of was using it all. In 742 00:43:16,560 --> 00:43:19,277 fact, it's interesting that on my PC I 743 00:43:19,360 --> 00:43:21,837 had um essentially a modem with a 744 00:43:21,920 --> 00:43:24,077 terminal connecting to my wax and then I 745 00:43:24,160 --> 00:43:26,957 had all my local DOSs and uh you know 746 00:43:27,040 --> 00:43:29,757 early Windows and what have you. 747 00:43:29,840 --> 00:43:31,997 And then suddenly took all and by the 748 00:43:32,080 --> 00:43:34,477 way I had Usenet all of that stuff but 749 00:43:34,560 --> 00:43:37,357 except it took all the way maybe until 750 00:43:37,440 --> 00:43:39,757 95 and even Mosaic right the first 751 00:43:39,840 --> 00:43:43,117 browser to bring it all together as one 752 00:43:43,200 --> 00:43:46,317 seamless experience right it's not about 753 00:43:46,400 --> 00:43:49,917 local cloud what have you I just want to 754 00:43:50,000 --> 00:43:53,917 use my computing power wherever it is in 755 00:43:54,000 --> 00:43:56,397 sort of a seamless flow of what I'm 756 00:43:56,480 --> 00:44:01,597 trying to get done so when I was seeing even what Daniel was showing that demo 757 00:44:01,680 --> 00:44:04,477 is mind-blowing. I mean, think about that moment. I think we're going to come 758 00:44:04,560 --> 00:44:07,197 back years from now and say, "Oh my god, 759 00:44:07,280 --> 00:44:10,237 that was when we saw what he did was he 760 00:44:10,320 --> 00:44:13,037 went to Comfy UI. He brought down a 761 00:44:13,120 --> 00:44:14,877 bunch of models, built a bunch of 762 00:44:14,960 --> 00:44:18,237 assets, took it to Blender, uh, used 763 00:44:18,320 --> 00:44:21,437 that to even enhance those assets, then 764 00:44:21,520 --> 00:44:23,757 used again brought it back to Comfy. Oh, 765 00:44:23,840 --> 00:44:26,237 went to Photoshop, used a local model 766 00:44:26,320 --> 00:44:28,637 perhaps to sort of enhance it, took it 767 00:44:28,720 --> 00:44:31,677 back to Comfy UI, then took it all to 768 00:44:31,760 --> 00:44:35,277 the cloud with Astra with Unreal, but he 769 00:44:35,360 --> 00:44:38,797 just used apps on his desktop, models, 770 00:44:38,880 --> 00:44:41,757 local models in the cloud, it all just 771 00:44:41,840 --> 00:44:45,357 worked. That is what this is about. Like 772 00:44:45,440 --> 00:44:47,197 that's kind of where today was about 773 00:44:47,280 --> 00:44:49,997 multiple fantastic devices, all these 774 00:44:50,080 --> 00:44:53,197 things. But the trajectory for me is so 775 00:44:53,280 --> 00:44:57,677 clear. There will never be a moment where we will go and look and say, "Oh, 776 00:44:57,760 --> 00:44:59,517 this runs locally. This runs in the 777 00:44:59,600 --> 00:45:02,637 cloud." You will expect this hybrid 778 00:45:02,720 --> 00:45:07,277 intelligence to be everywhere and pervasive. And I think that that's what 779 00:45:07,360 --> 00:45:10,717 to me is pretty exciting. >> Yeah, that one application what you just 780 00:45:10,800 --> 00:45:13,677 said, Blender was accelerated, Comfy UI 781 00:45:13,760 --> 00:45:16,477 was accelerated, Astra was in the cloud, 782 00:45:16,560 --> 00:45:18,237 the agents were locally. I mean that's 783 00:45:18,320 --> 00:45:21,117 just the the surface area surface area. 784 00:45:21,200 --> 00:45:26,077 That's cool. The surface area of computing was magnificent. 785 00:45:26,160 --> 00:45:28,397 >> Yeah. >> It is comp it it brought together 786 00:45:28,480 --> 00:45:33,277 literally the entire might of lit every single technology company in human 787 00:45:33,360 --> 00:45:35,277 history came together for that one application. 788 00:45:35,360 --> 00:45:37,837 >> That's right. And and by the way going back to sort of the way you talk about 789 00:45:37,920 --> 00:45:39,837 accelerated computer that was I mean his 790 00:45:39,920 --> 00:45:42,317 point about why he was excited what took 791 00:45:42,400 --> 00:45:45,117 him multiple days takes him hours. like 792 00:45:45,200 --> 00:45:50,077 that the fact that he was accelerating in his ambition and that to me is pretty 793 00:45:50,160 --> 00:45:53,917 cool >> which I I think you know this ties closely to something else because these 794 00:45:54,000 --> 00:45:56,797 models have gotten very good at tool use 795 00:45:56,880 --> 00:46:01,757 but the tools often live locally you know the files live locally by the way 796 00:46:01,840 --> 00:46:03,437 you know Jacob and I are still both of 797 00:46:03,520 --> 00:46:05,437 us are filing taxes at the last minute 798 00:46:05,520 --> 00:46:09,197 apparently um and and I think there is a 799 00:46:09,280 --> 00:46:10,717 aspect here about 800 00:46:10,800 --> 00:46:13,757 >> I want my data to be secure and safe in 801 00:46:13,840 --> 00:46:16,557 control. I want the tool use to happen 802 00:46:16,640 --> 00:46:20,957 near where the data is where if you look at what Jacob showed with co-pilot like 803 00:46:21,040 --> 00:46:22,717 that it really opens up something new. 804 00:46:22,800 --> 00:46:26,077 >> Yeah. I mean to me like one of the things right if you look at it in fact 805 00:46:26,160 --> 00:46:28,397 one of the when the coding agents became 806 00:46:28,480 --> 00:46:31,197 the thing I was excited to see everybody 807 00:46:31,280 --> 00:46:33,277 was in love with you know one the 808 00:46:33,360 --> 00:46:35,437 console uh and everybody was love with 809 00:46:35,520 --> 00:46:38,317 file systems again. It just shows how 810 00:46:38,400 --> 00:46:40,957 powerful of an abstraction a file system 811 00:46:41,040 --> 00:46:46,237 is. I think agents are going to be the largest users of file system. They need 812 00:46:46,320 --> 00:46:49,117 to be secure. >> They need to be because they're a 813 00:46:49,200 --> 00:46:52,157 heterogeneous collection space. Uh and 814 00:46:52,240 --> 00:46:54,637 what his tax example Jacob SP was a 815 00:46:54,720 --> 00:46:56,557 great one. Right. Basically what it did 816 00:46:56,640 --> 00:46:59,837 was it organized from as he said from 817 00:46:59,920 --> 00:47:03,277 downloads from even cloud files, email, 818 00:47:03,360 --> 00:47:05,997 everything into a file system container 819 00:47:06,080 --> 00:47:08,557 and then used models locally and in the 820 00:47:08,640 --> 00:47:11,277 cloud to go do stuff with it. And that 821 00:47:11,360 --> 00:47:15,837 is the new computer, right? So it's a I use the computer, my agents use the 822 00:47:15,920 --> 00:47:19,757 computer, the agents need these primitives as well. And so I think like 823 00:47:19,840 --> 00:47:24,077 one drive, when was the last time I thought of one drive as a cloud thing or 824 00:47:24,160 --> 00:47:26,397 a client thing? It's there everywhere. 825 00:47:26,480 --> 00:47:28,957 And so to me, that is what I think we 826 00:47:29,040 --> 00:47:30,957 are beginning to see is, hey, I want my 827 00:47:31,040 --> 00:47:35,517 primitives for my agents to use 24 by7. 828 00:47:35,600 --> 00:47:39,677 The form factor when I'm using it on my PC should feel like I'm using it on my 829 00:47:39,760 --> 00:47:43,357 PC. When I'm roaming, it should roam with me. 830 00:47:43,440 --> 00:47:47,277 >> You know what's amazing, as you know, is is that these agents are going to be 831 00:47:47,360 --> 00:47:52,077 better users of tools than us. And the reason for that is just we know 10 15% 832 00:47:52,160 --> 00:47:55,677 of the features of [clears throat] any application, but the agent knows every 833 00:47:55,760 --> 00:47:58,637 single feature. >> That's right. And so so the the the fact 834 00:47:58,720 --> 00:48:02,077 that the tools these incredible tools are sitting on your PC, they're they're 835 00:48:02,160 --> 00:48:04,237 accelerated in in this incredible 836 00:48:04,320 --> 00:48:09,277 computer Surface Ultra. Uh now the agent has access to all of that capability. 837 00:48:09,360 --> 00:48:11,997 Who knows what we're going to be able to do. I mean just look at what it did. It 838 00:48:12,080 --> 00:48:17,437 was the Blender, the Unreal, the Comfy AI. It was incredible. 839 00:48:17,520 --> 00:48:20,797 Kind of show of hands. How many people here have tried using one of these apps 840 00:48:20,880 --> 00:48:24,237 locally with the model in the last couple of months or a few months? 841 00:48:24,320 --> 00:48:28,237 >> Great. Right. And so I think this just kind of show goes to show where you can 842 00:48:28,320 --> 00:48:31,997 actually bring the tools locally. Now Jensen I want to ask you about the 843 00:48:32,080 --> 00:48:34,077 hardware piece of this right we very 844 00:48:34,160 --> 00:48:36,877 often see you talking about these big 845 00:48:36,960 --> 00:48:40,237 massive you know Vera Rubins Blackwells 846 00:48:40,320 --> 00:48:43,997 NBL72s you know these you know intelligence 847 00:48:44,080 --> 00:48:49,117 token factories um you know that's you know that power data centers here today 848 00:48:49,200 --> 00:48:52,637 we're talking about something different right like we're talking about the spark 849 00:48:52,720 --> 00:48:57,037 the workstation which I really want for Christmas you know which runs in your 850 00:48:57,120 --> 00:49:01,517 office talk to me a little bit about how you see this evolution about tokens 851 00:49:01,600 --> 00:49:05,917 being produced like right next to you as opposed to you know Ver Rub in data in 852 00:49:06,000 --> 00:49:09,917 the data center. >> Well uh the number of engineers that 853 00:49:10,000 --> 00:49:12,717 worked on this project is incredible 854 00:49:12,800 --> 00:49:15,997 over 4,000 engineering years. This is a 855 00:49:16,080 --> 00:49:20,317 massive project. I mean look what we've done. We've reinvented the personal 856 00:49:20,400 --> 00:49:22,237 computer completely for the first time. 857 00:49:22,320 --> 00:49:24,077 We took all the capabilities what used 858 00:49:24,160 --> 00:49:27,037 to be in a gaming PC, in a laptop, uh in 859 00:49:27,120 --> 00:49:29,277 a workstation, in a supercomput in the 860 00:49:29,360 --> 00:49:31,117 cloud. We unified the architecture and 861 00:49:31,200 --> 00:49:34,237 we made it run on one super chip. This 862 00:49:34,320 --> 00:49:36,477 super chip, you know, it's rectangular 863 00:49:36,560 --> 00:49:38,477 because it's two giant chips fused 864 00:49:38,560 --> 00:49:42,317 together. And so this it's a beast. It 865 00:49:42,400 --> 00:49:43,997 has one pedlops. Just to put in 866 00:49:44,080 --> 00:49:47,357 perspective, the computer that started 867 00:49:47,440 --> 00:49:50,797 the AI revolution called DGX1 was one 868 00:49:50,880 --> 00:49:54,557 pedlop. And it was a quarter of a bill it was a 869 00:49:54,640 --> 00:49:57,197 quarter of a billion dollars and um it 870 00:49:57,280 --> 00:50:00,477 weighed 500 lb and now it's in this in 871 00:50:00,560 --> 00:50:04,317 your laptop. Um but what's what's amazing is this is also the only 872 00:50:04,400 --> 00:50:06,957 computer the only computer in the world 873 00:50:07,040 --> 00:50:09,757 that runs DirectX and all of its 874 00:50:09,840 --> 00:50:14,157 applications OpenGL and every single version of it because we need all of the 875 00:50:14,240 --> 00:50:17,437 tools and design tools, right? Creative tools. We need everything to run. We 876 00:50:17,520 --> 00:50:19,037 need Unreal to run incredibly well. 877 00:50:19,120 --> 00:50:21,517 Also, DirectX has to be amazing and it 878 00:50:21,600 --> 00:50:26,237 runs every single CUDA application incredibly. What what uh Satia and I 879 00:50:26,320 --> 00:50:29,597 were talking about four years ago is you know we're going to have a whole new 880 00:50:29,680 --> 00:50:33,037 generation of software developers and these software developers are going to 881 00:50:33,120 --> 00:50:37,037 be AI software developers. They're either going to be creating AI 882 00:50:37,120 --> 00:50:41,597 applications and AI models or they're going to be using AI to to develop their 883 00:50:41,680 --> 00:50:45,597 software. And so the computer has to be revolutionized and has to be really good 884 00:50:45,680 --> 00:50:48,237 at CUDA from the ground up. And so this 885 00:50:48,320 --> 00:50:50,877 is the only computer on the planet that 886 00:50:50,960 --> 00:50:54,237 runs everything natively on the device. 887 00:50:54,320 --> 00:50:58,717 The first time ever. And so it took thousands of us to do this. But it also 888 00:50:58,800 --> 00:51:03,517 took the entire industry. Look behind us. Every single laptop maker, every 889 00:51:03,600 --> 00:51:07,197 single do desktop maker, every single workstation maker joined us. 890 00:51:07,280 --> 00:51:10,157 >> Yeah. Wow. Yeah. This is the first time 891 00:51:10,240 --> 00:51:14,797 I think first time I think a new piece of hardware new computer architecture 892 00:51:14,880 --> 00:51:19,517 have come together and the entire industry top to bottom left to right 893 00:51:19,600 --> 00:51:23,277 everybody's involved every single geography and so this is a complete 894 00:51:23,360 --> 00:51:26,397 reinvention we understand the seriousness of the work that we're doing 895 00:51:26,480 --> 00:51:28,877 doing here we've now test tested some 896 00:51:28,960 --> 00:51:30,797 over 897 00:51:30,880 --> 00:51:34,317 1,200 applications the most intense 898 00:51:34,400 --> 00:51:36,477 applications in the world all fully 899 00:51:36,560 --> 00:51:39,277 tested performance tested compatibility 900 00:51:39,360 --> 00:51:43,917 tested everything has been optimized we're going to keep on going the support 901 00:51:44,000 --> 00:51:47,517 for this system the the amount of software we're going to bring to it 902 00:51:47,600 --> 00:51:52,077 middleware algorithms I mean it will never end this is just a magnificent 903 00:51:52,160 --> 00:51:55,917 thing that when you when you finally look at it you look at this beautiful 904 00:51:56,000 --> 00:52:00,317 laptop pav stands up here with this beautiful laptop you know but the number 905 00:52:00,400 --> 00:52:03,437 of engineers between our two companies and all of the companies around the 906 00:52:03,520 --> 00:52:06,077 world that did this incredible all 907 00:52:06,160 --> 00:52:10,717 because we believe that the personal computer was our tool tool for people. 908 00:52:10,800 --> 00:52:13,917 Now it's going to be tool for people but is also going to be our personal 909 00:52:14,000 --> 00:52:18,157 assistant that uses the tool. Now one last thing for both of you I think if 910 00:52:18,240 --> 00:52:21,597 you've been on X for the last few days there has been this intense conversation 911 00:52:21,680 --> 00:52:25,197 about I would say personal agents we've been talking about productivity and 912 00:52:25,280 --> 00:52:29,997 maybe enterprise agents here whether it is Muse Instinct there's this question 913 00:52:30,080 --> 00:52:34,237 about hey when this agent goes out and makes a purchase is that me or the agent 914 00:52:34,320 --> 00:52:36,397 who is liable for it and if I'm a 915 00:52:36,480 --> 00:52:39,917 service provider how do I react to this 916 00:52:40,000 --> 00:52:43,757 human being no longer showing up on my website and people are trying to build 917 00:52:43,840 --> 00:52:46,877 standards and as an industry I think we're trying to figure this out. So I'm 918 00:52:46,960 --> 00:52:50,237 sort of curious if you kind of had to peer on the crystal ball a bit. Where do 919 00:52:50,320 --> 00:52:54,477 you see this headed? Right? Is it standards? Is the industry coalissing 920 00:52:54,560 --> 00:52:57,117 around some new behaviors? Where do you see this headed? 921 00:52:57,200 --> 00:52:59,597 >> Yeah, I mean first of all I think it starts with a little bit of the 922 00:52:59,680 --> 00:53:01,757 conversation we had uh in the early part 923 00:53:01,840 --> 00:53:05,757 which is trust. Um and trust starts with 924 00:53:05,840 --> 00:53:07,597 things like hey are these agents 925 00:53:07,680 --> 00:53:10,317 contained? Do they have identity? is the 926 00:53:10,400 --> 00:53:13,117 exhaust uh of the agent with the 927 00:53:13,200 --> 00:53:15,517 identity separate from the user right 928 00:53:15,600 --> 00:53:18,157 which is one of Pin's thing was how do 929 00:53:18,240 --> 00:53:20,877 we make sure the agent ID that's even 930 00:53:20,960 --> 00:53:23,757 working on behalf is separate from the 931 00:53:23,840 --> 00:53:26,397 user and you need it even for everyday 932 00:53:26,480 --> 00:53:29,037 uh usage but you also need it for what I 933 00:53:29,120 --> 00:53:30,957 think you're referencing which is how do 934 00:53:31,040 --> 00:53:33,917 I then trust what do I delegate in fact 935 00:53:34,000 --> 00:53:36,477 if you look at the first decision that 936 00:53:36,560 --> 00:53:40,477 we are making is what is my delegation 937 00:53:40,560 --> 00:53:42,957 um you know authority that I'm giving to 938 00:53:43,040 --> 00:53:47,997 the agent uh on my behalf of right and that is a decision that's going to be 939 00:53:48,080 --> 00:53:49,917 based on trust and it'll keep increasing 940 00:53:50,000 --> 00:53:53,117 the more autonomous it is uh you trusted 941 00:53:53,200 --> 00:53:57,357 and gone so then I think on the consumer space this will come down to even you 942 00:53:57,440 --> 00:54:01,757 know I I think of it as insurance will price this >> insurance 943 00:54:01,840 --> 00:54:05,677 >> because at some level we can all talk about it's not like oath is hard to do 944 00:54:05,760 --> 00:54:07,437 right I think the protocols are good we 945 00:54:07,520 --> 00:54:10,317 should do them and t But ultimately it's 946 00:54:10,400 --> 00:54:15,117 going to be pricing the the essentially 947 00:54:15,200 --> 00:54:19,197 the agent working on my behalf. Uh and 948 00:54:19,280 --> 00:54:22,157 how does the marketplace price it? 949 00:54:22,240 --> 00:54:26,637 >> Your question at the at the foundation of it MXC is so foundational. Open Shell 950 00:54:26,720 --> 00:54:29,917 the work that we're doing together. It's about it's about trust. It's about 951 00:54:30,000 --> 00:54:33,757 containment. It's about observability monitoring. If we can't get that right 952 00:54:33,840 --> 00:54:35,757 that's right. Oh yeah. >> If we just can't get the right. So 953 00:54:35,840 --> 00:54:39,677 that's the first part. We have got to secure that part of it so that the 954 00:54:39,760 --> 00:54:43,597 agents that are operating on our behalf, doing things on our behalf, we give it 955 00:54:43,680 --> 00:54:46,317 minimum rights. It has credentials. It 956 00:54:46,400 --> 00:54:48,877 is monitored and trusted. Only then can 957 00:54:48,960 --> 00:54:52,077 we allow these agents to really go off 958 00:54:52,160 --> 00:54:54,717 and do do uh do tasks on our behalf. 959 00:54:54,800 --> 00:54:57,997 >> You know, today's a powerful moment. I remember when open clock came out from 960 00:54:58,080 --> 00:55:00,797 there with the terminal hacking own um 961 00:55:00,880 --> 00:55:05,277 gateway to here where it is an embedded piece of Windows, you know, with tokens 962 00:55:05,360 --> 00:55:07,277 running accelerated. It's quite something. 963 00:55:07,360 --> 00:55:10,557 >> So much engineering went engineering went into that that the ride today. 964 00:55:10,640 --> 00:55:13,917 >> Amazing. And it is a great device. Let me just tell you that honestly. Well, 965 00:55:14,000 --> 00:55:17,437 Jensen Satia, thank you so much. It's a great day. Thank you for having me. And 966 00:55:17,520 --> 00:55:19,357 I'm going to throw it back to Pavin 967 00:55:19,440 --> 00:55:23,197 >> and Satia and Microsoft team. We have 968 00:55:23,280 --> 00:55:26,477 reinvented the computer as we know it. 969 00:55:26,560 --> 00:55:28,237 >> Thank you so much Jensen. Thanks for the 970 00:55:28,320 --> 00:55:32,397 partnership. Really excited today. And I think as you said like one thing I if I 971 00:55:32,480 --> 00:55:37,517 can end with Jensen just in passing talked about the vision for this was how 972 00:55:37,600 --> 00:55:39,357 do we I mean Windows has always been 973 00:55:39,440 --> 00:55:41,677 about how do we not leave any piece of 974 00:55:41,760 --> 00:55:45,117 code any piece of framework behind and 975 00:55:45,200 --> 00:55:48,877 bring it all. The fact that he talked about hey we're going to bring all of 976 00:55:48,960 --> 00:55:51,437 Windows apps all of the the graphics 977 00:55:51,520 --> 00:55:55,357 apps all of CUDA all of OpenGL and now 978 00:55:55,440 --> 00:55:59,837 with the token factory that is what is mind-blowing. Windows has been with me 979 00:55:59,920 --> 00:56:03,597 for 34 years. Looks like it's going to be with me for 34 more. 980 00:56:03,680 --> 00:56:05,347 >> There you go. Thank you so much. 981 00:56:05,430 --> 00:56:07,597 [applause] Thank you. 982 00:56:07,680 --> 00:56:12,690 Thank you. [applause] 983 00:56:20,720 --> 00:56:24,957 >> That's a bit hard to follow, but uh absolutely incredible moment for our 984 00:56:25,040 --> 00:56:29,517 industry. Just to wrap, we're building the OS for hybrid intelligence. It 985 00:56:29,600 --> 00:56:31,357 provides the most secure platform for 986 00:56:31,440 --> 00:56:34,397 agents, unlocks unmeeded intelligence. 987 00:56:34,480 --> 00:56:38,957 We're bringing AI into the ambition of humans and their flow. And of course, 988 00:56:39,040 --> 00:56:40,877 we're delivering it all on PCs that 989 00:56:40,960 --> 00:56:46,750 bring it together. Thank you all for your time today. Grateful. [applause] 990 00:56:47,350 --> 00:56:49,370 [music] 991 00:56:57,830 --> 00:56:59,850 >> [music] 992 00:57:04,400 --> 00:57:07,050 >> Heat up >> [music] 993 00:57:13,550 --> 00:57:15,570 [music] 994 00:57:22,060 --> 00:57:24,080 [music] 995 00:57:27,680 --> 00:57:33,250 >> here. >> [music] 996 00:57:43,360 --> 00:57:49,630 >> Heat. Heat. [music] 997 00:57:52,560 --> 00:57:55,560 Heat. Heat. 998 00:58:07,910 --> 00:58:09,930 [music] 999 00:58:17,090 --> 00:58:19,110 >> [music] 1000 00:58:22,980 --> 00:58:25,000 [music] 1001 00:59:35,120 --> 00:59:38,120 >> Heat. 1002 00:59:54,320 --> 00:59:58,010 Heat. [music] 1003 01:00:01,230 --> 01:00:03,250 >> [music] 1004 01:00:04,800 --> 01:00:07,800 >> Hey, 1005 01:00:10,640 --> 01:00:13,640 hey, hey.