1 00:00:00,160 --> 00:00:03,040 Speaker 1: If you like this episode, just like, share and follow 2 00:00:03,080 --> 00:00:06,879 this podcast. Thank you, Back to the show, Welcome. 3 00:00:06,559 --> 00:00:10,000 Speaker 2: To the Xai All Hands. We've been a very exciting 4 00:00:10,080 --> 00:00:10,880 presentation for you. 5 00:00:11,279 --> 00:00:14,480 Speaker 1: We're going to start off by recapping the incredible progress 6 00:00:14,519 --> 00:00:17,120 that the Xai team has made in just two and 7 00:00:17,120 --> 00:00:20,800 a half years. It's really remarkable in pursuit of our 8 00:00:20,839 --> 00:00:22,920 goal of understanding the universe. So, just going of our 9 00:00:23,320 --> 00:00:27,440 accomplishments since inception, it's important for bear in mind that 10 00:00:28,120 --> 00:00:30,320 XI is only two and a half years old, basically 11 00:00:30,359 --> 00:00:34,280 a toddler, and we've nonetheless achieved an incredible amount in 12 00:00:34,320 --> 00:00:37,000 a very short period of time. So our competitors are 13 00:00:38,000 --> 00:00:42,679 five ten, some cases twenty years old, they have much 14 00:00:42,759 --> 00:00:45,679 larger teams they started off with for more resources, and 15 00:00:45,719 --> 00:00:49,399 yet nonetheless we have achieved number one in many arenas 16 00:00:50,159 --> 00:00:54,159 in just a few years. So we've achieved number one 17 00:00:54,240 --> 00:00:57,560 in voice, in image and video generation. I think we 18 00:00:57,719 --> 00:01:00,719 now at this point are actually generating more images and 19 00:01:00,840 --> 00:01:03,679 video based on the last numbers I saw then all 20 00:01:03,679 --> 00:01:08,680 of our competitors combined, we are winning in terms of forecasting, 21 00:01:08,920 --> 00:01:11,400 which is one of the key metrics of intelligence. So 22 00:01:11,680 --> 00:01:15,040 up the grock four to twenty forecasting model beat all 23 00:01:15,040 --> 00:01:19,359 the other AIS and forecasting, and we've talked many leaderboards. 24 00:01:19,680 --> 00:01:24,879 We've got now a great app with the Imagine, with 25 00:01:25,000 --> 00:01:28,319 the core Rock, we've made radical improvements to the x app, 26 00:01:28,560 --> 00:01:31,760 and we've launched a Grokipedia which is on its way 27 00:01:31,840 --> 00:01:36,560 too far exceeding Wikipedia and ultimately be Audersmatitude are more 28 00:01:36,560 --> 00:01:39,879 comprehensive and more accurate and have more information as well 29 00:01:40,120 --> 00:01:45,239 as video and image data that simply isn't there on Wikipedia. 30 00:01:45,359 --> 00:01:49,159 So it's it's intended ultimately to be Encyclopedia Galactica, a 31 00:01:49,239 --> 00:01:55,480 distallation of all knowledge, of all knowledge. And we're the 32 00:01:55,480 --> 00:01:59,280 first to achieve one hundred thousand, h one hundred GPU 33 00:01:59,359 --> 00:02:03,040 training clusters, and we're now about to achieve the first 34 00:02:03,120 --> 00:02:06,159 one hundred I should say, one million, h one hundred 35 00:02:06,280 --> 00:02:10,759 GPU equivalents in training. So really an incredible amount of 36 00:02:10,840 --> 00:02:13,319 work in a very short period of time. And it's 37 00:02:13,319 --> 00:02:17,479 important to consider for competitiveness of any technology company. 38 00:02:17,680 --> 00:02:20,240 Speaker 2: What matters. It's not the position at any point in time, but. 39 00:02:20,199 --> 00:02:23,639 Speaker 1: What is your velocity and acceleration And if you're moving 40 00:02:23,680 --> 00:02:27,360 faster than anyone else in any given technology arena, you 41 00:02:27,400 --> 00:02:30,039 will be the leader, and Xai is moving faster than 42 00:02:30,080 --> 00:02:30,840 any other company. 43 00:02:30,960 --> 00:02:33,120 Speaker 2: No one's even close. So let's go to our team. 44 00:02:34,039 --> 00:02:37,120 Speaker 1: As we grow as a company, a natural thing that 45 00:02:37,199 --> 00:02:40,719 happens is you reorganize the company as it scales up. 46 00:02:40,840 --> 00:02:42,639 Speaker 2: So when you first have a startup, you. 47 00:02:42,639 --> 00:02:45,560 Speaker 1: Might have just a few dozen people and they'll just 48 00:02:45,639 --> 00:02:48,759 chat amongs themselves. As you grow to several hundred people, 49 00:02:49,000 --> 00:02:51,800 you have to then add more structure, just like an 50 00:02:51,840 --> 00:02:54,240 organism that grows from a single like we all just 51 00:02:54,319 --> 00:02:56,840 grow from a single cell and then to a blob 52 00:02:56,879 --> 00:03:00,400 of sales. Then you get organ differentiation limbs, you grow 53 00:03:00,400 --> 00:03:04,319 a tail, hopefully the ptail disappears, and then you become 54 00:03:04,360 --> 00:03:07,400 a baby. You go through these stages, and so we're 55 00:03:07,719 --> 00:03:09,639 uh organizing because. 56 00:03:09,439 --> 00:03:11,080 Speaker 2: We've we've reached a certain scale. 57 00:03:11,080 --> 00:03:14,199 Speaker 1: We're organizing the company h to be more effective at 58 00:03:14,199 --> 00:03:17,599 this scale. And naturally, when when this happens, that there's 59 00:03:17,639 --> 00:03:20,400 some people who are better suited for the early stages 60 00:03:20,439 --> 00:03:23,360 of the company and and less suited for the later stages. 61 00:03:23,680 --> 00:03:27,639 And so uh and for the people that have departed, 62 00:03:27,680 --> 00:03:29,759 I'd just like to say thank you for a contribution, 63 00:03:30,280 --> 00:03:32,400 thank you for getting us this far. And we wish 64 00:03:32,479 --> 00:03:34,759 you very well in your future endeavors. So now going 65 00:03:34,800 --> 00:03:37,879 on to h the new structure of the company. Uh, 66 00:03:38,080 --> 00:03:42,759 the companies organized in four main application areas. There's there's 67 00:03:42,840 --> 00:03:45,560 Grock Main in voice, which is really the main Grock 68 00:03:45,599 --> 00:03:48,439 model as much as called Grock Main. Then there's a 69 00:03:48,680 --> 00:03:52,039 coding specific model. There's an image and video model which 70 00:03:52,080 --> 00:03:55,080 is Imagine, and then Macrohard which is intended to do 71 00:03:55,400 --> 00:03:58,960 full digital emulation of entire companies. And then we've got 72 00:03:59,560 --> 00:04:02,080 the Impress ructural ayers. So I'd like to invite members 73 00:04:02,120 --> 00:04:03,560 of the teams come up and talk about each of 74 00:04:03,599 --> 00:04:04,479 their areas. 75 00:04:09,080 --> 00:04:10,080 Speaker 2: Hip. Thanks you long. 76 00:04:10,319 --> 00:04:13,240 Speaker 3: So Grock, Maine and Boys are going to be merged 77 00:04:13,280 --> 00:04:16,279 into one team. And you know on Boys when anecdote 78 00:04:16,319 --> 00:04:19,040 is September twenty twenty four opening, I had this product 79 00:04:19,079 --> 00:04:21,360 you could talk to advanced voice mode, and we had nothing, 80 00:04:21,439 --> 00:04:23,519 no model, of course, in no product. 81 00:04:24,040 --> 00:04:25,480 Speaker 2: We started much after. 82 00:04:25,279 --> 00:04:27,560 Speaker 3: That and in a span of few months, six months, 83 00:04:27,639 --> 00:04:30,079 we developed the model in house from scratch without a 84 00:04:30,120 --> 00:04:32,439 bunch of people who knew audio, and had a product 85 00:04:32,600 --> 00:04:35,199 that was surpassing open air. And six months fast forward 86 00:04:35,240 --> 00:04:37,959 six more months and now we have Grock in more 87 00:04:37,959 --> 00:04:40,279 than two billion test las. We have a Rock Voice 88 00:04:40,319 --> 00:04:42,800 Agent API. You can do all kinds of amazing things. 89 00:04:43,040 --> 00:04:44,839 In a span of one year, we went from nothing 90 00:04:45,000 --> 00:04:47,240 to being leaders. That kind of stuff is only possible 91 00:04:47,279 --> 00:04:49,759 in a place like XCI. We have small teams, committed 92 00:04:49,800 --> 00:04:53,000 mission focus, lots of compute, and we really really want 93 00:04:53,079 --> 00:04:55,920 to keep pushing same story on the chat models. You know, 94 00:04:55,920 --> 00:04:59,040 we've always been at the forefront of reasoning, starting from 95 00:04:59,079 --> 00:05:01,600 Rock one point five two Doc three, and we want 96 00:05:01,639 --> 00:05:04,360 to really move to a world where it's no longer 97 00:05:04,360 --> 00:05:06,959 about this question answering. We want to build and everything 98 00:05:07,040 --> 00:05:08,560 up so you should be able to come to it 99 00:05:08,759 --> 00:05:10,920 and really get done whatever you want. You know, ask 100 00:05:10,959 --> 00:05:13,920 a leave question, make a slide deck, or you know, 101 00:05:14,240 --> 00:05:15,800 solid pizzle stuff like that. 102 00:05:15,920 --> 00:05:18,079 Speaker 4: Yeah, So I really think on the product side, we're 103 00:05:18,120 --> 00:05:20,720 really going to see a huge transformation happening in a 104 00:05:20,800 --> 00:05:21,959 very short period of time. 105 00:05:22,800 --> 00:05:23,920 Speaker 2: We're going to see work. 106 00:05:24,199 --> 00:05:27,439 Speaker 4: The magnitude of amount of work that all knowledge workers 107 00:05:27,519 --> 00:05:30,199 are going to be able to produce increase tenfold in 108 00:05:30,240 --> 00:05:33,160 the next short period of a few months. The models 109 00:05:33,160 --> 00:05:36,360 that we are building out are incredibly amazing, and we 110 00:05:36,480 --> 00:05:38,319 have a lot on the way and we're really excited 111 00:05:38,319 --> 00:05:41,360 to share that with you all. And on a product side, 112 00:05:41,399 --> 00:05:43,759 the goal is to just build that portal that allows 113 00:05:43,839 --> 00:05:46,079 you to accomplish all of your work and how do 114 00:05:46,120 --> 00:05:49,199 we amplify everyone to achieve much much more than what 115 00:05:49,240 --> 00:05:51,800 they can accomplish alone. And we're building that out and 116 00:05:51,800 --> 00:05:54,040 it's going to be an incredibly easy. 117 00:05:53,759 --> 00:05:55,959 Speaker 5: To use experience that just works seamlessly. 118 00:05:56,199 --> 00:05:56,360 Speaker 2: Too. 119 00:05:57,000 --> 00:05:59,639 Speaker 6: Being said, we are hiring and we're looking for intelligent 120 00:05:59,680 --> 00:06:00,399 and smart people. 121 00:06:00,439 --> 00:06:02,439 Speaker 2: This is not an easy place to work guys like 122 00:06:02,519 --> 00:06:02,879 this is. 123 00:06:03,240 --> 00:06:07,319 Speaker 6: It's a grind, but we have I guess like intersellar ambitions, 124 00:06:07,360 --> 00:06:09,519 so it's it's not going to be easy, right. So 125 00:06:10,000 --> 00:06:13,240 I will say, having come to x Ai, uh, it 126 00:06:13,319 --> 00:06:15,959 has been an opportunity for lifetime to work among really 127 00:06:16,000 --> 00:06:19,600 smart and really passionate people. The vibes here are amazing 128 00:06:19,879 --> 00:06:22,040 and it's truly an environment where if you're a smart 129 00:06:22,079 --> 00:06:24,279 person you want to get shipped done, you can get 130 00:06:24,240 --> 00:06:28,120 shipped done. There isn't like organizational overhead getting your way 131 00:06:28,240 --> 00:06:30,399 or kind of I don't know, like having to write 132 00:06:30,480 --> 00:06:32,879 docs and all this kind of stuff. You just do stuff, 133 00:06:32,920 --> 00:06:34,959 at least at least for me, I just you know, 134 00:06:35,120 --> 00:06:37,519 just you can do things here and that's amazing, and 135 00:06:37,720 --> 00:06:39,560 I invite more people to come here and just do 136 00:06:39,600 --> 00:06:40,519 it awesome things. 137 00:06:40,720 --> 00:06:44,040 Speaker 1: Yeah, So with with the grockmin the sort of main 138 00:06:44,560 --> 00:06:48,040 foundation model, the intent is that it's it's genuinely useful 139 00:06:48,160 --> 00:06:49,480 in a wide range of areas. 140 00:06:49,600 --> 00:06:53,560 Speaker 2: So if you're doing engineering or law or tit or. 141 00:06:53,519 --> 00:06:57,600 Speaker 1: Medicine, any anything, it is useful to you in your 142 00:06:57,920 --> 00:07:02,439 in your job. That's essential to understand the universe and 143 00:07:03,000 --> 00:07:06,199 make things as useful as possible, like when God gives 144 00:07:06,199 --> 00:07:07,120 you an answer that you can. 145 00:07:07,000 --> 00:07:09,680 Speaker 2: Count it right, all right, thank you, Thanks. 146 00:07:14,360 --> 00:07:15,399 Speaker 7: Everybody. I'm Macro. 147 00:07:15,800 --> 00:07:18,639 Speaker 8: So the world changed a lot recently in terms of coding, 148 00:07:19,639 --> 00:07:22,319 the coding models. I was always complaining people are trying 149 00:07:22,319 --> 00:07:24,000 to convince me to use a coding model, and I 150 00:07:24,040 --> 00:07:26,959 was like trusting it, and I wasn't really convinced. But 151 00:07:27,240 --> 00:07:31,720 as of recently, the models they actually produce good, decent 152 00:07:31,800 --> 00:07:34,240 quality code. I mean, you still need to review and 153 00:07:34,240 --> 00:07:37,759 give feedback, but you can. It's easy to see how 154 00:07:37,759 --> 00:07:40,120 they can accelerate you quite a lot. So it's not 155 00:07:40,160 --> 00:07:43,879 only about coding. It's like they understand your intuition like 156 00:07:44,160 --> 00:07:47,079 much better than before. Like now when you are. When 157 00:07:47,079 --> 00:07:50,000 I describe a problem, I only have to phrase it 158 00:07:50,079 --> 00:07:52,240 like I would to another colleague engineer who has already 159 00:07:52,240 --> 00:07:54,399 seen the code base. That's a huge change before you 160 00:07:54,519 --> 00:07:56,759 kind of need to handhold a toddler to make a change, 161 00:07:56,839 --> 00:07:59,680 and they don't only write your code, but they also 162 00:07:59,680 --> 00:08:00,839 can bugio code. 163 00:08:00,920 --> 00:08:03,160 Speaker 7: So now we have I do like what we. 164 00:08:03,079 --> 00:08:06,959 Speaker 8: Do, like hours of grog code running continuously to make 165 00:08:07,000 --> 00:08:11,879 sure that a more complex changed the training system actually 166 00:08:11,959 --> 00:08:14,120 works in production. So it's easy to see for us 167 00:08:14,360 --> 00:08:17,879 that this is not only about accelerating us ourselves writing 168 00:08:18,000 --> 00:08:20,480 code and making us ten x more productive, but we're 169 00:08:20,519 --> 00:08:24,439 really on this path for recursive self improvement where the 170 00:08:24,519 --> 00:08:27,680 current generation of grock code is training the next generation 171 00:08:27,720 --> 00:08:30,519 of Crock code. And we see that this path and 172 00:08:30,560 --> 00:08:33,639 on exponential takeoff here this path will continue. So we 173 00:08:33,679 --> 00:08:36,039 are doubling down on coding and making coding one of 174 00:08:36,039 --> 00:08:38,200 the highest party efforts in the company. So if you're 175 00:08:38,240 --> 00:08:41,600 out there and you're excited about coding, and you're inner 176 00:08:41,759 --> 00:08:44,600 either very good at training modeling, or you're a really 177 00:08:44,639 --> 00:08:48,440 good low level software engineer interesting in systems design, this 178 00:08:48,679 --> 00:08:51,320 is the place to work. Like we have a million 179 00:08:52,000 --> 00:08:55,240 h one hundred equivalents to train the best coding medal 180 00:08:55,279 --> 00:08:56,000 in the world right. 181 00:08:55,840 --> 00:08:57,679 Speaker 2: Now, so please join us. 182 00:08:58,080 --> 00:09:01,440 Speaker 9: Yeah, I'm going on a compare with Macro on coding, 183 00:09:01,639 --> 00:09:03,799 so it'll become one and more obvious to us, like 184 00:09:03,840 --> 00:09:05,840 you know, over time, like we are on the past 185 00:09:05,960 --> 00:09:08,879 to singularity at least on coding. So we decided, like 186 00:09:08,879 --> 00:09:10,919 you know, have our best engineer in the company in 187 00:09:11,000 --> 00:09:13,679 Macro to lead the coding, and we are building the 188 00:09:13,679 --> 00:09:16,559 best coding model for everyone to empower everyone to build 189 00:09:16,879 --> 00:09:19,919 and for me, like the man like limiting factor is 190 00:09:19,960 --> 00:09:23,159 probably computer and energy where they can run the best 191 00:09:23,200 --> 00:09:26,840 model to support everyone, to empower everyone and respects. 192 00:09:26,960 --> 00:09:29,279 Speaker 7: Now we are one keeam and we will win on 193 00:09:29,320 --> 00:09:30,720 the compute and we are. 194 00:09:30,679 --> 00:09:34,759 Speaker 9: Wing with space compute and also like for every engineer 195 00:09:34,919 --> 00:09:37,600 right so, if you're like writing kernel, if you're writing 196 00:09:37,559 --> 00:09:40,080 a compiler, to think about like whether is still all 197 00:09:40,120 --> 00:09:41,879 worth it, maybe you should join us, you know, for 198 00:09:41,960 --> 00:09:44,759 coding effort to automate yourself alt the like, to speed 199 00:09:44,759 --> 00:09:45,519 it ourselves up. 200 00:09:46,200 --> 00:09:48,200 Speaker 7: Yeah, I think it's like really a million. 201 00:09:48,039 --> 00:09:51,639 Speaker 9: Year basically with the year to be alive, and I 202 00:09:51,639 --> 00:09:53,879 can already feel the AGI, feel the hi at least 203 00:09:53,919 --> 00:09:54,360 for coding. 204 00:09:54,600 --> 00:09:57,720 Speaker 1: Yeah, yeah, I think actually things will move, maybe even 205 00:09:57,840 --> 00:10:00,600 by the end of this year, to where you don't 206 00:10:00,639 --> 00:10:04,159 even bother do during coding. The AI just creates the 207 00:10:04,200 --> 00:10:07,080 binary directly, and the AI can create a much more 208 00:10:07,080 --> 00:10:11,360 efficient binary than can be done by any compiler. So 209 00:10:11,639 --> 00:10:14,679 just say create optimized binary for this particular outcome, and 210 00:10:15,360 --> 00:10:18,960 and you actually bypass even traditional coding. There's there's no 211 00:10:19,440 --> 00:10:22,720 that that's an intermediate step that actually will not be 212 00:10:22,759 --> 00:10:25,320 needed probably if by i'd say the end of this year. 213 00:10:26,120 --> 00:10:27,799 Speaker 2: And we do expect. 214 00:10:29,039 --> 00:10:31,080 Speaker 1: Code to be stead of the art in two to 215 00:10:31,120 --> 00:10:32,840 three months, so it's happening very quickly. 216 00:10:38,240 --> 00:10:41,519 Speaker 9: Also, do Imagine, so you know, I mean, what will 217 00:10:41,559 --> 00:10:43,000 you do right after post aga? 218 00:10:43,120 --> 00:10:43,279 Speaker 8: Right? 219 00:10:43,320 --> 00:10:45,360 Speaker 9: You probably do, like digital life, So that's what we 220 00:10:45,360 --> 00:10:48,080 are doing here as well. We have the Imagine team 221 00:10:48,279 --> 00:10:50,600 like started pretty much from Squash like six months ago. 222 00:10:50,639 --> 00:10:53,080 We have a few people we decided will to do 223 00:10:53,120 --> 00:10:55,679 the image and we'll do the video gen like, yeah, 224 00:10:55,679 --> 00:10:57,440 look at it. What do we achieved today, Like you know, 225 00:10:57,639 --> 00:11:00,840 like two weeks ago. We'll recently Imagine we actually thought 226 00:11:00,919 --> 00:11:03,679 of a leader ball across like many of them, and 227 00:11:03,799 --> 00:11:06,639 people really love our product, love our model. And we 228 00:11:06,759 --> 00:11:09,399 have many more releases actually this month and next months. 229 00:11:09,559 --> 00:11:13,000 So yeah, to me, there's like a very high chance, 230 00:11:13,240 --> 00:11:16,679 like we actually made view the Metaverse before Mada. Yeah, 231 00:11:16,720 --> 00:11:18,960 I'll also pass to try to to talk Abaul, like 232 00:11:18,960 --> 00:11:21,519 you know, the metrics we have, the products. 233 00:11:21,600 --> 00:11:24,279 Speaker 10: Yeah, yeah, Like like Godong said, it's only been six 234 00:11:24,320 --> 00:11:26,320 months since we started working on Imagine, we had we 235 00:11:26,360 --> 00:11:30,200 had no code internally for Diffusion at all six months ago, 236 00:11:30,480 --> 00:11:34,080 and basically now we've launched Imagine on every product surface 237 00:11:34,159 --> 00:11:37,000 that we have, including seamlessly integrating into X so you 238 00:11:37,039 --> 00:11:38,919 can open the except right now. You can long press 239 00:11:38,919 --> 00:11:40,559 on any image, you can edit the image, you can 240 00:11:40,559 --> 00:11:42,639 make a video out of the image. We also ran 241 00:11:42,679 --> 00:11:45,320 a contest recently where we had some really funny submissions 242 00:11:45,320 --> 00:11:46,559 that I'm sure many of you have seen. 243 00:11:47,240 --> 00:11:49,039 Speaker 5: So Imagine it's growing extremely. 244 00:11:48,639 --> 00:11:51,240 Speaker 10: Extremely fast, and it's because of the speed at which 245 00:11:51,279 --> 00:11:54,120 we iterate. Basically, we do multiple product updates every day, 246 00:11:54,159 --> 00:11:56,960 we do model updates every other week, and effectively what 247 00:11:57,000 --> 00:11:59,519 this has led to is now users are generating close 248 00:11:59,559 --> 00:12:02,000 to fifty million videos every day using Imagine. 249 00:12:02,279 --> 00:12:02,759 Speaker 2: And just to. 250 00:12:02,720 --> 00:12:05,080 Speaker 10: Reiterate what Elon said earlier that to the best of 251 00:12:05,080 --> 00:12:08,039 our knowledge that is more than every other provider combined, 252 00:12:08,120 --> 00:12:10,600 which again is an astonishing place to be compared to 253 00:12:10,600 --> 00:12:13,240 where we were six months ago. We are also generating 254 00:12:13,279 --> 00:12:17,120 six billion images in the last thirty days nano You know, 255 00:12:17,159 --> 00:12:20,240 Google recently posted that, you know, one billion images were 256 00:12:20,279 --> 00:12:23,120 generated using nano Vana in thirty days, So you know 257 00:12:23,120 --> 00:12:26,159 we're six times that, right, And really the goal is 258 00:12:26,440 --> 00:12:28,559 it's not like we don't just want to win. We 259 00:12:28,600 --> 00:12:30,679 want to win like like over a long period of 260 00:12:30,720 --> 00:12:33,559 time and have sustained greatness. And so the goal with 261 00:12:33,679 --> 00:12:35,759 Imagine is to take anything that you can you know, 262 00:12:35,919 --> 00:12:39,000 imagine and turn it into reality. And so that's that's 263 00:12:39,039 --> 00:12:40,879 what we're gonna you know that we're going to speed 264 00:12:40,960 --> 00:12:42,120 run that basically is the goal. 265 00:12:42,279 --> 00:12:43,559 Speaker 5: Yeah, Hey, I'm haatin. 266 00:12:44,279 --> 00:12:48,519 Speaker 11: We As we keep scaling our model capabilities puwing visual 267 00:12:48,519 --> 00:12:53,720 worlds that's indistinguishable from reality, we're also puling systems that 268 00:12:53,799 --> 00:12:56,879 onlocks much more possibility than what we have right now, 269 00:12:57,559 --> 00:13:00,000 never be able to generate the videos that's much longer 270 00:13:00,080 --> 00:13:02,440 and what we have right now with stories or with 271 00:13:02,600 --> 00:13:06,200 souls of your imagine, And by the end of the year, 272 00:13:06,240 --> 00:13:09,240 we likely will be having models that allow you to 273 00:13:09,320 --> 00:13:12,679 generate videos of ten minutes or twenty minutes in one 274 00:13:12,679 --> 00:13:16,720 shot without any intervention. You just need to give your 275 00:13:16,759 --> 00:13:19,360 imagination and our model, our agents. 276 00:13:19,000 --> 00:13:19,759 Speaker 2: Will do it for you. 277 00:13:20,000 --> 00:13:25,159 Speaker 11: And moreover, those are the videos we generate, and we're 278 00:13:25,200 --> 00:13:29,639 also going to allow rendering those. We're already the fastest 279 00:13:29,720 --> 00:13:32,120 in generating the videos, and we're going to keep pushing 280 00:13:32,120 --> 00:13:35,120 the extreme where we're going to render those videos in 281 00:13:35,159 --> 00:13:38,240 real time and you will be able to imagine, build 282 00:13:38,320 --> 00:13:41,960 and interact with your own world, and the world will 283 00:13:42,000 --> 00:13:45,000 respond to you in real time, and it is exciting 284 00:13:45,279 --> 00:13:47,559 future that we are going to build with ourself. 285 00:13:47,679 --> 00:13:52,240 Speaker 1: Absolutely, my friction is that most of AI compute is 286 00:13:52,240 --> 00:13:54,799 going to be real time, but your understanding and real 287 00:13:54,840 --> 00:13:57,639 time video generation and we expect. 288 00:13:57,360 --> 00:13:58,120 Speaker 2: To people leaves in that. 289 00:13:58,679 --> 00:14:01,840 Speaker 1: It's really emphsizing these points that you know, six months 290 00:14:01,840 --> 00:14:04,639 ago we didn't even have. We had basically nothing in 291 00:14:05,759 --> 00:14:09,080 very weak in video and image generation and editing, and 292 00:14:09,159 --> 00:14:12,039 we're in six months to number one spot and in 293 00:14:12,039 --> 00:14:15,080 fact generally more more videos and images than everyone else combined. 294 00:14:15,159 --> 00:14:16,879 We're going to do the same thing with coding, and 295 00:14:16,919 --> 00:14:20,080 we're going to do the same thing with Macrohart and 296 00:14:20,080 --> 00:14:23,320 I think people will be pretty impressed with the clock 297 00:14:23,360 --> 00:14:26,279 four point two models coming out. That's a it's a 298 00:14:26,360 --> 00:14:29,559 it's a significant improvement, and that's really just that's that's 299 00:14:29,600 --> 00:14:33,960 the small version of our new model. So we will 300 00:14:33,960 --> 00:14:36,440 have a medium and a large version that are even 301 00:14:36,440 --> 00:14:37,080 more intelligent. 302 00:14:41,799 --> 00:14:45,919 Speaker 12: All right, Hi everyone, I'm Toby and I work on Macrohart, 303 00:14:46,120 --> 00:14:50,639 the most series of all product names. So arguably, giving 304 00:14:50,720 --> 00:14:53,559 computers to humans was a good idea, So we're doing 305 00:14:53,559 --> 00:14:56,200 the same thing thing for AI. It's kind of like Inception. 306 00:14:56,279 --> 00:15:00,360 We're giving computers to computers. So Macrohart is building fully 307 00:15:00,360 --> 00:15:05,000 capable digital review time, very important human emulator, so it's 308 00:15:05,120 --> 00:15:07,600 able to do anything on a computer that a human 309 00:15:07,679 --> 00:15:11,279 is able to do, including using advanced tools and engineering 310 00:15:11,399 --> 00:15:15,000 and medicine. So there should be rocket engines fully designed 311 00:15:15,200 --> 00:15:16,080 by AI. 312 00:15:16,440 --> 00:15:19,120 Speaker 5: And in a sense, it's one of the last few. 313 00:15:18,919 --> 00:15:22,679 Speaker 12: Remaining areas where AI is significantly worse than humans, which 314 00:15:22,840 --> 00:15:25,279 is why I think it's one of the most exciting 315 00:15:25,360 --> 00:15:29,200 areas to actually innovate in and actually change change the field. 316 00:15:29,440 --> 00:15:32,639 Speaker 13: Hi everyone, So yeah, my name is John and yeah, 317 00:15:32,679 --> 00:15:35,120 so we're building these strong reasoning models which are now 318 00:15:35,240 --> 00:15:36,360 going to control our CLI. 319 00:15:36,799 --> 00:15:38,559 Speaker 5: Like we're actively using these every day. 320 00:15:38,879 --> 00:15:42,399 Speaker 13: They are like tremendous, like productivity bursts to the whole team. 321 00:15:42,440 --> 00:15:44,360 I know, the voice team is like killing on that. 322 00:15:44,679 --> 00:15:46,720 And you know, this is the reason why we need 323 00:15:46,720 --> 00:15:48,840 the compute. You know, we need a large scale computer 324 00:15:48,919 --> 00:15:51,320 run these models to boost our own productivity. 325 00:15:51,519 --> 00:15:55,200 Speaker 5: But you know, to ninety five percent of the world 326 00:15:55,919 --> 00:15:57,519 world software has a GUI. 327 00:15:58,320 --> 00:16:01,360 Speaker 13: So that's like, you know, great representation, and you know, 328 00:16:01,440 --> 00:16:04,840 to truly make people's lives easier, we need to develop models. 329 00:16:04,600 --> 00:16:07,600 Speaker 5: That are capable of solving day to day tasks on GUI. 330 00:16:07,919 --> 00:16:11,639 Speaker 13: So Macrohart, you know, we will emulate a company where 331 00:16:11,639 --> 00:16:13,639 the output is digital and so this is. 332 00:16:13,600 --> 00:16:15,080 Speaker 5: The obvious next step for agents. 333 00:16:15,600 --> 00:16:18,200 Speaker 13: Macrohard will and the able true end to end orchestration 334 00:16:18,279 --> 00:16:21,919 across the desktop, and it will lead to immense economic prosperity. 335 00:16:22,720 --> 00:16:25,120 So yeah, we're entering in there where we need to 336 00:16:25,120 --> 00:16:27,679 tackle the hardest of tech problems. But in order to 337 00:16:27,679 --> 00:16:29,519 solve this, we need to hire the best people. So 338 00:16:30,120 --> 00:16:32,080 you know, think of the smartest people that you've worked 339 00:16:32,080 --> 00:16:35,080 with and and put them forward for a position here. 340 00:16:35,320 --> 00:16:37,799 And if you can't think of anybody, like, go through 341 00:16:37,840 --> 00:16:41,159 your phone book, go for your LinkedIn you'll be surprised 342 00:16:41,240 --> 00:16:43,919 like how big your actual network is. And they just 343 00:16:43,960 --> 00:16:46,840 need three properties obviously that we want to optimize for. 344 00:16:47,200 --> 00:16:50,320 Are they clever, can we solve hard problems? And the 345 00:16:50,320 --> 00:16:52,720 second property is are they driven? Do they have the ambition? 346 00:16:52,919 --> 00:16:55,000 Do they want to win? And the third is are 347 00:16:55,000 --> 00:16:56,840 they a nice person? Like do you want to actually 348 00:16:56,840 --> 00:16:58,200 work with them? 349 00:16:58,480 --> 00:16:58,720 Speaker 14: Yeah? 350 00:16:58,840 --> 00:16:59,559 Speaker 5: So thank you. 351 00:16:59,759 --> 00:17:02,759 Speaker 2: Yeah. The Macrohart project is over time. 352 00:17:02,799 --> 00:17:07,079 Speaker 15: Actually, we'll probably be the most important project because what 353 00:17:07,160 --> 00:17:11,200 we're talking about is emulation of entire human companies. So 354 00:17:11,440 --> 00:17:13,759 when you look at the most valuable companies in the world, 355 00:17:13,799 --> 00:17:14,400 they are. 356 00:17:15,000 --> 00:17:19,319 Speaker 1: That their their output is digital, so they don't actually 357 00:17:19,920 --> 00:17:23,240 make hardware. So it should be possible to completely emulate 358 00:17:23,720 --> 00:17:25,000 any company that where. 359 00:17:24,799 --> 00:17:25,799 Speaker 2: The output is digital. 360 00:17:26,240 --> 00:17:29,240 Speaker 1: And this will usher in an age of prosperity lives 361 00:17:29,279 --> 00:17:31,319 which we can barely imagine at this point. 362 00:17:31,359 --> 00:17:32,680 Speaker 2: You need imagine to imagine it. 363 00:17:33,400 --> 00:17:35,200 Speaker 1: So this is a big This is a big deal, 364 00:17:35,240 --> 00:17:38,160 and this is why the words macrohard are painted on 365 00:17:38,240 --> 00:17:40,960 the roof of the training cluster, because that's what it's 366 00:17:40,960 --> 00:17:41,680 going to build. 367 00:17:41,720 --> 00:17:49,799 Speaker 5: It's also pretty funny, meant to be a joke. It's 368 00:17:49,839 --> 00:17:50,279 me again. 369 00:17:50,519 --> 00:17:53,440 Speaker 12: You might remember me from macrohart in computer used from 370 00:17:53,480 --> 00:17:56,319 a long time ago, but I also actually work on 371 00:17:56,440 --> 00:17:59,319 a core product infrastructure and API. In fact, this is 372 00:17:59,319 --> 00:18:02,759 what I've done for most time at XAI. So anytime 373 00:18:03,039 --> 00:18:05,400 you use any of our products like Cork dot Com 374 00:18:05,440 --> 00:18:09,160 API authentication, you go to status at xot Ai. This 375 00:18:09,240 --> 00:18:12,160 is done by the core product infront team, and a 376 00:18:12,279 --> 00:18:14,519 large portion of them actually sit in London and we 377 00:18:14,599 --> 00:18:17,319 work with himI over there, so we keep the lights 378 00:18:17,359 --> 00:18:21,359 on peak our four pm every day. We get paged 379 00:18:21,440 --> 00:18:23,160 at night when stuff goes down. 380 00:18:23,359 --> 00:18:25,599 Speaker 5: Also, thank you to anyone in paleal to getting paged. 381 00:18:26,319 --> 00:18:30,119 Speaker 12: This is really important work reliability, security, core product infrastructure. 382 00:18:30,279 --> 00:18:33,079 So if you actually if you're really interested in solving 383 00:18:33,160 --> 00:18:36,480 difficult distributed problems with like messy data, this is. 384 00:18:36,440 --> 00:18:37,160 Speaker 5: The team to join. 385 00:18:40,039 --> 00:18:42,400 Speaker 7: Hey everyone, manasiego. Yeah. 386 00:18:42,440 --> 00:18:44,440 Speaker 16: So, I think one of the main bottlenecks in this 387 00:18:44,480 --> 00:18:46,519 next year for these models is gonna be very high 388 00:18:46,559 --> 00:18:48,920 quality emails and training data. And one of the ways 389 00:18:48,920 --> 00:18:51,160 we solve that is by taking the world's foremost experts 390 00:18:51,160 --> 00:18:54,640 in these domains, bringing them here and having them a 391 00:18:54,680 --> 00:18:56,839 value with them up. We do this for domains like medicine, 392 00:18:56,960 --> 00:19:00,920 finance law with voice actors. We have video editors who 393 00:19:00,920 --> 00:19:04,680 contribute daily to making GROW better and uh yeah, we're 394 00:19:04,720 --> 00:19:06,559 going to be continuing to work on very high quality 395 00:19:06,599 --> 00:19:07,119 emails over the. 396 00:19:07,079 --> 00:19:07,640 Speaker 7: Next few months. 397 00:19:07,759 --> 00:19:10,680 Speaker 16: We have some same stuff in you know, the frontier 398 00:19:10,680 --> 00:19:14,440 of useful tast and finance and law. You know, we're 399 00:19:14,440 --> 00:19:16,960 trying to build evls that are are are useful in 400 00:19:17,039 --> 00:19:20,680 training data that that represents useful work and not necessarily 401 00:19:20,759 --> 00:19:23,720 proxies of intelligence for a lot of the open source 402 00:19:23,759 --> 00:19:25,640 emails to today. 403 00:19:25,759 --> 00:19:25,960 Speaker 7: Yeah. 404 00:19:26,160 --> 00:19:30,279 Speaker 1: Yeah, I'd like to say like we're we're shifting from 405 00:19:30,480 --> 00:19:34,319 using these sort of common uh Internet emails, which I 406 00:19:34,359 --> 00:19:37,200 think are actually not a real indicator of usefulness, to 407 00:19:38,039 --> 00:19:43,599 having expert judos in each domain to every domain of engineering, medicine, law. 408 00:19:43,559 --> 00:19:46,279 Speaker 2: Whatever the case may be. And the the. 409 00:19:46,039 --> 00:19:49,079 Speaker 1: The actual evail is, does the expert in that arena 410 00:19:49,400 --> 00:19:51,480 or does our group of experts in that arena human 411 00:19:51,519 --> 00:19:56,039 experts agree that GROCK is extremely useful and that the 412 00:19:56,079 --> 00:19:58,640 results are correct. That's the that's actually the only EVL 413 00:19:58,640 --> 00:19:59,240 it really matters. 414 00:19:59,359 --> 00:20:00,000 Speaker 5: Yeah, exactly. 415 00:20:01,279 --> 00:20:04,000 Speaker 16: In you'll you'll see this in GROW four twenty. But 416 00:20:04,039 --> 00:20:05,920 we've made some improvements because of that type of data 417 00:20:05,960 --> 00:20:08,960 in truth seeking and kind of minimizing political bias. Our 418 00:20:08,960 --> 00:20:13,359 responses are are much more cogent. Yeah, that's exciting. And 419 00:20:13,519 --> 00:20:16,279 we are also working on Grockipedia. So the the goal 420 00:20:16,279 --> 00:20:18,640 of Wrockipedia is to create a distillation vol human knowledge. 421 00:20:19,400 --> 00:20:20,519 I kind of like to think of this as like 422 00:20:20,519 --> 00:20:23,200 a modern day version of the Library of Alexandria. And 423 00:20:23,319 --> 00:20:26,559 in the quest to build Encyclopedia Galactica to will one 424 00:20:26,599 --> 00:20:29,200 day be cult we've gone from essentially having nothing to 425 00:20:29,359 --> 00:20:33,920 around six million articles for context wikipedias around seven million 426 00:20:33,920 --> 00:20:36,920 English articles uh and uh yeah, or we're improving on 427 00:20:37,000 --> 00:20:41,599 hallucination uh. And our our goal is essentially for Rock 428 00:20:41,640 --> 00:20:43,720 five to not have to search out of the data center. 429 00:20:44,759 --> 00:20:51,720 Speaker 7: So yep, I guess. 430 00:20:51,720 --> 00:20:54,759 Speaker 8: So in the infrad team, we are building the training 431 00:20:55,039 --> 00:20:57,720 in France and touring team tooting. 432 00:20:57,440 --> 00:20:59,839 Speaker 7: Software for the company. So to give a new example, 433 00:21:00,119 --> 00:21:01,240 we were training Rock three. 434 00:21:01,599 --> 00:21:05,160 Speaker 8: We built the pre training framework for this and these 435 00:21:05,200 --> 00:21:07,519 are some of the coolest system in my opinion, that 436 00:21:07,559 --> 00:21:09,240 you can build as a software engineer. So it's like 437 00:21:09,480 --> 00:21:11,680 we have one hundred kh one hundreds at the time 438 00:21:12,079 --> 00:21:14,640 and they were just delivered and we. 439 00:21:14,599 --> 00:21:16,119 Speaker 5: Didn't quite have the software. 440 00:21:16,240 --> 00:21:18,599 Speaker 8: We thought we'd have the software, but then at thirty 441 00:21:18,680 --> 00:21:21,799 K scale we realized, actually the software is not quite working, 442 00:21:22,119 --> 00:21:25,960 and it took a major almost I was like halfway 443 00:21:26,240 --> 00:21:29,119 rewrite off the software because there's so much going on 444 00:21:29,160 --> 00:21:31,119 in a data center that you can't actually account for. 445 00:21:32,440 --> 00:21:36,559 Switch switches are flapping, little links are flapping, switches are 446 00:21:36,559 --> 00:21:40,240 going down, GPUs are just burning through. You have numerics issues, 447 00:21:40,559 --> 00:21:43,160 and it's a system where you want really one hundred 448 00:21:43,240 --> 00:21:45,480 k h one hundreds to behave in locksteps. So a 449 00:21:45,519 --> 00:21:48,319 training step is like five seconds and you're going five 450 00:21:48,319 --> 00:21:51,599 seconds in lockstep, but during that five seconds everything can happen. 451 00:21:51,759 --> 00:21:54,079 So you need to write a system that makes progress 452 00:21:54,119 --> 00:21:56,680 despite all these things that can happen in the environment. 453 00:21:56,920 --> 00:21:58,640 And we did a successfully in what's one of the 454 00:21:58,640 --> 00:22:01,880 coolest times in our life where the system was actually running, 455 00:22:01,880 --> 00:22:03,640 and it was running at the same time my son 456 00:22:03,720 --> 00:22:08,000 was born, so that was extra excitement. But these problems 457 00:22:08,079 --> 00:22:10,680 like you don't find anywhere else, like nobody has this 458 00:22:10,799 --> 00:22:13,559 kind of compute and also nobody has this kind of 459 00:22:13,559 --> 00:22:16,079 talent density. So at the time, to give you a perspective, 460 00:22:16,359 --> 00:22:18,640 we were like an overall team in pre training we 461 00:22:18,640 --> 00:22:21,160 were probably like fifteen people. Out of that, maybe like 462 00:22:21,480 --> 00:22:24,519 seven people were working on the actual training system, and 463 00:22:24,599 --> 00:22:27,920 we still maintain that talent density in the team. 464 00:22:28,160 --> 00:22:30,559 Speaker 7: So if you're interested in working on these problems and. 465 00:22:30,480 --> 00:22:32,039 Speaker 8: You don't want to be just like part of a 466 00:22:32,039 --> 00:22:35,000 bigger organization where you're one of like a thousand people 467 00:22:35,039 --> 00:22:36,880 working on this, then this is the place. 468 00:22:37,160 --> 00:22:38,880 Speaker 7: Like we are still a very small. 469 00:22:38,599 --> 00:22:40,640 Speaker 8: Team with me is Lea and Minn from the RL 470 00:22:40,680 --> 00:22:41,400 and inference team. 471 00:22:41,759 --> 00:22:42,519 Speaker 2: Hi, I'm Lemming. 472 00:22:42,720 --> 00:22:45,319 Speaker 17: So at our team we were on our reinforced Malanning 473 00:22:45,400 --> 00:22:48,160 training job and the production inference has a larger scale 474 00:22:48,160 --> 00:22:51,160 on the ours and probably as soon in space, and 475 00:22:51,279 --> 00:22:54,240 we are kind of already designed a lot of seeing 476 00:22:55,039 --> 00:22:58,720 to make it more resilient and scalable, So building a 477 00:22:58,799 --> 00:23:02,359 system to scale from hydro kate ships to millions of ships, 478 00:23:02,519 --> 00:23:06,599 and we aug every aspect of the stack like paradism 479 00:23:06,720 --> 00:23:10,440 pre fiell decault and make resilient to error in them 480 00:23:10,519 --> 00:23:13,640 and the unknown hardware failure. So if you are system 481 00:23:13,680 --> 00:23:18,279 hackers obsessed with extreme performance and reliability, so here is 482 00:23:18,359 --> 00:23:21,319 you will find the most interesting problems to work place. 483 00:23:21,559 --> 00:23:25,240 And I think actually like very similar to all kinds 484 00:23:25,240 --> 00:23:26,319 of things like you. 485 00:23:26,920 --> 00:23:29,480 Speaker 2: It's it's very important for you to first see the. 486 00:23:29,440 --> 00:23:32,680 Speaker 17: Problem and there you will developed the solutions that Noah 487 00:23:32,759 --> 00:23:33,680 else can develop. 488 00:23:33,839 --> 00:23:37,319 Speaker 7: Before I'll handle to the tooling team. Hello, I might 489 00:23:37,359 --> 00:23:39,160 slip from the tooling team. 490 00:23:39,359 --> 00:23:42,680 Speaker 18: Every software needs to have a great interface to be 491 00:23:42,680 --> 00:23:45,079 able to make it useful. So as the Tooling team 492 00:23:45,119 --> 00:23:48,880 were responsible for building the platform sprameworks and infrassection, which 493 00:23:48,880 --> 00:23:52,240 is required for humans as well as agents to be 494 00:23:52,279 --> 00:23:55,279 able to use our products. We started by building out 495 00:23:55,279 --> 00:23:57,680 the Human Data Platform. There's a place where we collect 496 00:23:57,720 --> 00:24:00,200 all of our human data and eventually expect, you know, 497 00:24:00,359 --> 00:24:03,359 to build that internal e general platform through which we 498 00:24:03,480 --> 00:24:08,079 basically run deployments, run evaluations or like look at like 499 00:24:08,119 --> 00:24:11,039 what training results exist. So if you really care about 500 00:24:11,240 --> 00:24:15,000 building a good interface or providing a really useful framework 501 00:24:15,200 --> 00:24:17,960 for researchers, for agents as well as our tutests, we 502 00:24:17,960 --> 00:24:19,240 should definitely join our team. 503 00:24:19,359 --> 00:24:21,799 Speaker 7: So how everyone, I'm you know from the Jacks team. 504 00:24:21,920 --> 00:24:25,279 Speaker 19: So now Jacks XAI is a really small team with 505 00:24:25,680 --> 00:24:29,680 a couple of engineers working on Jack's GPU to optimize 506 00:24:29,680 --> 00:24:33,839 our ultra large scale GPU training. So you can imagine 507 00:24:33,839 --> 00:24:36,240 that training at scale can be very complicated. Even you 508 00:24:36,359 --> 00:24:39,200 run hollow world at scale can be complicated. Right, So 509 00:24:39,960 --> 00:24:44,640 then we're actually responsible for supporting the entire companies from 510 00:24:45,079 --> 00:24:48,920 pre training fundation models rls and also multimodel to scale 511 00:24:50,000 --> 00:24:54,880 things to from from first from ten k hundred kent 512 00:24:55,000 --> 00:24:59,519 then probably one million, h one hundred equivalents GPO scale, 513 00:24:59,759 --> 00:25:02,480 and that we to implement a lot of you know, 514 00:25:02,519 --> 00:25:07,559 practical practical optimizations. We have to uh customize the entire 515 00:25:07,640 --> 00:25:10,359 jacks stack from compiler and brunt times, and there will 516 00:25:10,400 --> 00:25:12,400 be a lot of interesting problems. 517 00:25:13,240 --> 00:25:15,799 Speaker 14: And also if you really want to, uh, you know, 518 00:25:16,359 --> 00:25:20,720 obsessed on optimizing uh the entire stack at scale, that 519 00:25:20,759 --> 00:25:23,279 we are probably the best place to go, uh because 520 00:25:23,279 --> 00:25:26,240 you know, we really have very large scale GIP clusters 521 00:25:26,240 --> 00:25:28,240 and we have a lot of interesting problems to work with. 522 00:25:28,559 --> 00:25:30,480 Speaker 7: Hey, I'm branding from the Kunnels team. 523 00:25:30,599 --> 00:25:32,720 Speaker 18: Basically, the Kunnel team sets at the very bottom of 524 00:25:32,720 --> 00:25:35,400 a training and serving stack. Our code runs inside the 525 00:25:35,440 --> 00:25:38,160 million current GPUs that we have and if you look 526 00:25:38,200 --> 00:25:40,920 inside the gp there's hundreds of thousands of threads and 527 00:25:40,960 --> 00:25:42,880 these threads are tend to talk to each other to 528 00:25:43,000 --> 00:25:46,359 multiply matrices. Computer tens scores and some of them even 529 00:25:46,359 --> 00:25:48,319 talk to the million other gps that we have. And 530 00:25:48,400 --> 00:25:50,359 this is the low level system that we have, and 531 00:25:50,400 --> 00:25:53,160 we like optimizing every single microsecond in this and we 532 00:25:53,240 --> 00:25:55,920 care deeply about squeezing every large top of performance from 533 00:25:55,960 --> 00:25:59,799 these gps. So if you like these low level systems, problems, algorithms, 534 00:26:00,039 --> 00:26:06,279 these join us. 535 00:26:06,079 --> 00:26:08,359 Speaker 20: But you know, we'll try to bring in uh Heiner 536 00:26:08,440 --> 00:26:12,799 and Spencer who are actually at our Tubook compew cluster 537 00:26:13,160 --> 00:26:18,079 in Memphis. 538 00:26:18,720 --> 00:26:21,440 Speaker 7: And I'm high from the computer network. 539 00:26:21,599 --> 00:26:24,480 Speaker 5: Plus such a team we are maybe basic alto by 540 00:26:24,480 --> 00:26:26,599 the day. We're here in Memphis in the super computer. 541 00:26:26,759 --> 00:26:29,720 Speaker 21: So the data center here in Memphis swum the last 542 00:26:29,799 --> 00:26:31,480 tribute class on the planet and it. 543 00:26:31,519 --> 00:26:32,440 Speaker 3: Is still growing. 544 00:26:33,279 --> 00:26:35,880 Speaker 22: Our drugs keep all this from do and running. 545 00:26:36,319 --> 00:26:40,440 Speaker 21: Next Rock and the sir Ai alps us used to 546 00:26:40,480 --> 00:26:40,880 work well. 547 00:26:41,119 --> 00:26:42,200 Speaker 2: I love to agree. 548 00:26:41,839 --> 00:26:45,000 Speaker 1: It's actually just put the mic really close to your 549 00:26:45,000 --> 00:26:47,000 mouth because you're the ambient noises high. 550 00:26:47,039 --> 00:26:49,200 Speaker 2: It's getting out that you go back. 551 00:26:49,880 --> 00:26:52,880 Speaker 21: So I was saying, I'll drop the compute up and running, 552 00:26:52,960 --> 00:26:54,200 try the next model of ruck. 553 00:26:54,079 --> 00:26:55,640 Speaker 2: And serve the so use us. 554 00:26:55,720 --> 00:26:58,079 Speaker 21: So, but you used to work well and at in 555 00:26:58,160 --> 00:27:00,839 greens have to come together, mainly soft yeah and hotware. 556 00:27:01,039 --> 00:27:05,240 So there's all these tb TPUs, nicked switches and hundreds 557 00:27:05,240 --> 00:27:07,960 of thousands of operating systems running. Is one take supercomputer 558 00:27:08,279 --> 00:27:11,359 and what we need Sparks to re understand the notes, 559 00:27:11,440 --> 00:27:13,759 re understand that he may and re understand how computers work. 560 00:27:13,759 --> 00:27:16,160 Brond deep level that is, you reach album X and 561 00:27:16,160 --> 00:27:17,519 I'm heading over ten A. 562 00:27:17,640 --> 00:27:17,839 Speaker 1: Right. 563 00:27:18,720 --> 00:27:22,880 Speaker 22: So we have three hundred thousand TV, three hundred platforms 564 00:27:23,000 --> 00:27:26,680 us here today, still growing, still building eight hundred and 565 00:27:26,759 --> 00:27:29,799 forty seven miles of fiber per Data Hall twelve. 566 00:27:29,920 --> 00:27:30,480 Speaker 5: Data Holls. 567 00:27:30,599 --> 00:27:33,039 Speaker 22: You want to be part of the world's largest supercomputer, 568 00:27:33,319 --> 00:27:34,799 come join us, all right. 569 00:27:35,000 --> 00:27:37,400 Speaker 23: So it's quite marvelous what we've been able to do 570 00:27:37,480 --> 00:27:38,759 in less than one year's time. 571 00:27:38,799 --> 00:27:40,599 Speaker 2: Here we have. 572 00:27:41,319 --> 00:27:44,119 Speaker 23: Once we're completely finished, we'll have north of a giggle out. 573 00:27:44,000 --> 00:27:45,599 Speaker 7: Of power online and running. 574 00:27:45,640 --> 00:27:48,759 Speaker 23: We'll have the largest tessil of megapack system in the 575 00:27:48,799 --> 00:27:52,680 world margin than Hawaii or South Australia. And Zach is 576 00:27:52,680 --> 00:27:54,880 really quickly going to talk a little bit about actually 577 00:27:54,920 --> 00:27:55,799 constructing the data is. 578 00:27:56,279 --> 00:27:58,559 Speaker 24: So behind me you can see Data Hall eleven. So 579 00:27:58,559 --> 00:28:01,200 one of the most incredible things about what we're doing 580 00:28:01,200 --> 00:28:03,799 here at Macrohart's how fast we do it right, So, 581 00:28:03,960 --> 00:28:06,279 like they were saying before, over eight or fifty miles 582 00:28:06,319 --> 00:28:08,759 of fiber and every single data hall, over twenty seven 583 00:28:08,839 --> 00:28:13,039 thousand jews and over a two hundred thousand connections, So 584 00:28:13,240 --> 00:28:15,640 all of this that you can see behind me was 585 00:28:15,680 --> 00:28:17,400 put up in less than six weeks. 586 00:28:17,400 --> 00:28:19,039 Speaker 5: We do that over and over and over again. 587 00:28:19,119 --> 00:28:22,799 Speaker 24: We massively parallelize it. It's pretty much the most complex 588 00:28:23,039 --> 00:28:27,880 and consistent type of engineering, design and construction project you 589 00:28:28,000 --> 00:28:28,839 possibly imagine. 590 00:28:29,240 --> 00:28:30,759 Speaker 2: So come join us. Yes, you know. 591 00:28:30,839 --> 00:28:33,799 Speaker 23: The other really awesome thing about this is that everything 592 00:28:33,839 --> 00:28:38,559 is completely vertically integrated within this team, from architecture, mechanical, electrical, structure, 593 00:28:38,599 --> 00:28:41,200 all the disciplines. And we also care a lot about 594 00:28:41,200 --> 00:28:44,240 efficiency while we're designing all of this too, So it's 595 00:28:44,279 --> 00:28:47,160 not just about getting the most compute online the fastest, 596 00:28:47,200 --> 00:28:51,200 but also achieving the highest PUE in the industry of 597 00:28:51,559 --> 00:28:55,319 using as much power smoothing technology as we can. I'm 598 00:28:55,359 --> 00:28:57,920 being really good partners in the community here in Memphis 599 00:28:58,839 --> 00:29:01,839 with the toss of micropark that we have going. You 600 00:29:01,839 --> 00:29:04,799 can check them out. XAI Memphis back to you with. 601 00:29:05,599 --> 00:29:10,440 Speaker 2: All right, thank you, all. 602 00:29:10,400 --> 00:29:12,759 Speaker 1: Right, So that was live from the front lines in Memphis. 603 00:29:13,440 --> 00:29:18,960 So fundamental to any AI company success is the compute advantage. 604 00:29:19,039 --> 00:29:21,519 And what we've demonstrated over and over again is that 605 00:29:21,680 --> 00:29:24,200 Xai can actually deploy more AI. 606 00:29:23,960 --> 00:29:25,599 Speaker 2: Compute faster than anyone else. 607 00:29:26,079 --> 00:29:29,440 Speaker 1: And actually, as Justin Wong of CEO and Vidia has 608 00:29:29,480 --> 00:29:32,599 said many times in interviews, there is no one faster 609 00:29:32,680 --> 00:29:38,880 at getting AI compute online than Xai. So congratulations, guys. 610 00:29:39,839 --> 00:29:40,960 Speaker 2: Yeah, this is what it looks like. 611 00:29:41,599 --> 00:29:45,160 Speaker 1: So that's a really phase one, which is three hundred 612 00:29:45,160 --> 00:29:49,599 and thirty thousand Grace Blackwells with macrohart on the building 613 00:29:49,640 --> 00:29:52,039 that's on an image that it actually is on the 614 00:29:52,079 --> 00:29:55,079 roof of the building. And then macro Hotter will be 615 00:29:55,119 --> 00:29:57,079 the building that you can see which has got the 616 00:29:57,119 --> 00:29:59,400 macro Hotter with the rockets on it, and that'll be 617 00:29:59,519 --> 00:30:02,400 another turned twenty thousand GP three hundreds. 618 00:30:02,880 --> 00:30:06,480 Speaker 2: So all of this will be training the models that 619 00:30:07,400 --> 00:30:08,000 you experience. 620 00:30:08,160 --> 00:30:13,119 Speaker 1: So it's absolutely fundamental obviously to have large scale training 621 00:30:13,160 --> 00:30:15,119 compute in order to get the best models. Yeah, I'm 622 00:30:15,160 --> 00:30:18,160 sort of reminded of the Jose Amian where you see 623 00:30:18,240 --> 00:30:21,720 one guy digging and there's like seven people watching and 624 00:30:22,000 --> 00:30:25,119 One of the big differences between Xai and other companies 625 00:30:25,160 --> 00:30:26,359 is we are actually Jose. 626 00:30:26,480 --> 00:30:29,400 Speaker 25: Hello, all right, I'm Nikita. You might know me as 627 00:30:29,440 --> 00:30:32,559 a part time ship poster, full time customer support for 628 00:30:32,799 --> 00:30:36,960 x So we're now reaching over a billion people across 629 00:30:37,000 --> 00:30:40,920 our family of apps. Every time news breaks, it just 630 00:30:40,960 --> 00:30:44,200 becomes evident that this is the most important communication tool 631 00:30:44,240 --> 00:30:49,680 of our time. It's where the most influential people come convene. 632 00:30:50,279 --> 00:30:53,960 It's where truth is crystallized. Everything is downstream of x 633 00:30:54,400 --> 00:30:56,440 The reason they say this is going to hit Facebook 634 00:30:56,480 --> 00:30:59,960 in a week because it happens here, and I think 635 00:31:00,119 --> 00:31:04,160 we're only beginning to realize its full potential. We had 636 00:31:04,160 --> 00:31:06,920 a remarkable year for the app. We rolled up our 637 00:31:06,960 --> 00:31:11,359 sleeves and got a ton done. January was our biggest 638 00:31:11,400 --> 00:31:15,680 month ever for the app in terms of engagement, and 639 00:31:15,720 --> 00:31:18,680 then February is on track to beat that. Much of 640 00:31:18,720 --> 00:31:22,279 the credit lies with the algorithm team. They've been putting 641 00:31:22,319 --> 00:31:25,599 in crazy hours and it's clearly paying off, but there's 642 00:31:25,640 --> 00:31:28,240 still a huge amount of work to be done. On 643 00:31:28,319 --> 00:31:31,440 the top of funnel side. First time downloads are up 644 00:31:31,480 --> 00:31:35,519 over fifty percent every month, and we're exhibiting right now 645 00:31:35,599 --> 00:31:39,079 like basically the growth rates of an early stage consumer product. 646 00:31:39,920 --> 00:31:42,039 We also made a ton of headway and solving one 647 00:31:42,079 --> 00:31:44,319 of the like twenty year old problems of the app, 648 00:31:44,440 --> 00:31:47,640 which was ramping up new users. New users are now 649 00:31:47,640 --> 00:31:50,559 spending fifty five percent more time per day in the 650 00:31:50,599 --> 00:31:55,400 app than they were six months ago. And on the 651 00:31:55,400 --> 00:31:59,240 core product side, we're hitting our stride to not only 652 00:31:59,319 --> 00:32:02,200 did we rebuild the algorithm, we rebuilt our onboarding flows 653 00:32:02,279 --> 00:32:05,119 and we're seeing double digit increases on all our key metrics. 654 00:32:05,200 --> 00:32:10,400 We rebuilt notifications, our web browser, U x chat, basically 655 00:32:10,480 --> 00:32:12,559 every surface of the app has been rebuilt to be 656 00:32:12,559 --> 00:32:16,039 better than ever, and it's clear that if we're focused, 657 00:32:16,680 --> 00:32:19,119 we can move mountains and evolve this platform. 658 00:32:19,359 --> 00:32:19,799 Speaker 2: Uh. 659 00:32:19,839 --> 00:32:23,200 Speaker 25: Just last month we did a little push on articles 660 00:32:24,279 --> 00:32:29,000 and articles published are up ten x, articles read are 661 00:32:29,079 --> 00:32:34,200 up seventeen x. And on all other fronts, like over 662 00:32:34,240 --> 00:32:36,799 the holidays, we did a big push on subscriptions. 663 00:32:37,000 --> 00:32:37,799 Speaker 2: We just crossed a. 664 00:32:37,759 --> 00:32:38,960 Speaker 25: Billion dollars in ar R. 665 00:32:39,119 --> 00:32:41,359 Speaker 2: There. I think with the X. 666 00:32:41,319 --> 00:32:44,920 Speaker 25: App, you know, the there's very few unknowns, like the 667 00:32:44,960 --> 00:32:48,279 path for us to win and become you know, the 668 00:32:48,359 --> 00:32:51,559 number one app in the world. Uh, We're it's it's 669 00:32:51,599 --> 00:32:54,519 we we know what to do. The balls in our court. Uh, 670 00:32:54,759 --> 00:32:56,720 it's it's for us to win, and it's just a 671 00:32:56,720 --> 00:32:58,400 matter of us executing yep. 672 00:32:58,599 --> 00:33:01,240 Speaker 2: And yeah, So. 673 00:33:04,519 --> 00:33:07,519 Speaker 1: We've evolved the what used to be the old Twitter 674 00:33:07,599 --> 00:33:11,640 DM stack, which was unencrypted basically just text, to a 675 00:33:11,680 --> 00:33:16,160 fully encrypted messaging system that allows you to do audio 676 00:33:16,200 --> 00:33:19,079 and video calls. Has uh you know, all the things 677 00:33:19,079 --> 00:33:22,559 you'd want from any messaging app for the disappearing messages, 678 00:33:22,599 --> 00:33:24,960 screen screenshot blocks, like, there's a whole all the features 679 00:33:24,960 --> 00:33:27,359 that you'd want want in an app. And we and 680 00:33:27,920 --> 00:33:30,000 we will be open sourcing the code for this in 681 00:33:30,039 --> 00:33:32,079 the next few months. As we are open sourcing the 682 00:33:32,119 --> 00:33:35,519 recommendation algorithm code so people can actually see what we're doing. 683 00:33:35,880 --> 00:33:36,039 Speaker 19: Uh. 684 00:33:36,279 --> 00:33:42,480 Speaker 1: Nothing beats nothing beats uh, transparency for believing in a company. 685 00:33:42,640 --> 00:33:47,119 So we're going to be the only recommendation algorithms that 686 00:33:47,279 --> 00:33:49,720 actually open sources so you can see what it, what 687 00:33:49,759 --> 00:33:52,720 it does, and how it's evolving. With with crock Chat, 688 00:33:52,759 --> 00:33:55,119 it will also be open source so you can actually 689 00:33:55,119 --> 00:33:57,160 see if there are any vulnerabilities. There will be no 690 00:33:57,160 --> 00:34:00,200 hooks for advertising or anything else like that in in 691 00:34:00,519 --> 00:34:03,200 grock Chat, which is really intended to be a generalized 692 00:34:03,200 --> 00:34:06,319 communication system, and in the next few months we'll be 693 00:34:06,359 --> 00:34:10,159 releasing a standalone UH x chat app, So if you 694 00:34:10,199 --> 00:34:11,960 just want to do messaging, you can just you can 695 00:34:12,000 --> 00:34:12,239 do that. 696 00:34:12,320 --> 00:34:15,079 Speaker 2: You don't you don't have to go to the core product, and. 697 00:34:15,119 --> 00:34:20,519 Speaker 1: We'll have desktop sharing and multi user so you can 698 00:34:20,559 --> 00:34:22,280 do you can do video calls with lots of people. 699 00:34:22,400 --> 00:34:26,840 It's really intended to be a fully functional communications system 700 00:34:27,159 --> 00:34:31,159 with x chat. For x money, we're UH. We've actually 701 00:34:31,199 --> 00:34:34,920 had x money UH live in closed beta within the company, 702 00:34:35,360 --> 00:34:38,239 and we expect in the next month or two UH 703 00:34:38,280 --> 00:34:41,840 to go to UH a limited external beata, and then 704 00:34:41,880 --> 00:34:45,760 to go worldwide to all x users. And this is 705 00:34:45,760 --> 00:34:47,920 really intended to be the place where all the money 706 00:34:48,039 --> 00:34:52,880 is the central source of all monetary transactions. So it's 707 00:34:52,920 --> 00:34:54,920 it's a it's really going to be a game changer. 708 00:34:55,719 --> 00:34:58,159 And the reason we say one billion user is actually 709 00:34:58,159 --> 00:35:01,760 over a billion users is that while our monthly users 710 00:35:01,800 --> 00:35:05,400 are on average around six hundred million, the number of 711 00:35:05,400 --> 00:35:08,000 people who have the X app and sold this well 712 00:35:08,000 --> 00:35:10,760 over a billion. It's just that most people only occasionally 713 00:35:11,000 --> 00:35:14,199 come to the X app when there's some major world event. 714 00:35:14,440 --> 00:35:17,000 But as we give people more reasons to use the 715 00:35:17,400 --> 00:35:22,119 x app, whether it's for communications, for rock, or for 716 00:35:23,199 --> 00:35:26,320 X money, whatever the case may be. We wanted to 717 00:35:26,360 --> 00:35:28,159 be such that if you wanted to, you could live 718 00:35:28,159 --> 00:35:30,039 your life on the X app. And as we make 719 00:35:30,079 --> 00:35:33,400 it more more useful, we'll obviously give people reasons, compelling 720 00:35:33,440 --> 00:35:37,079 reasons to use the app every day and have my 721 00:35:37,159 --> 00:35:41,760 expectations well over a billion daily active users now. In 722 00:35:41,880 --> 00:35:45,039 order to understand the universe, you must explore the universe. 723 00:35:45,119 --> 00:35:48,679 There's only so much you can learn from just being 724 00:35:48,719 --> 00:35:52,519 on Earth with telescopes and clatters on Earth. Ultimately, you 725 00:35:52,559 --> 00:35:53,960 have to go out there and you have to explore 726 00:35:53,960 --> 00:35:58,119 the universe to understand it. And that's the motivation behind 727 00:35:58,159 --> 00:36:03,039 the combination of space and XAI, is to accelerate humanity's 728 00:36:03,039 --> 00:36:06,079 future in understanding the universe and extending the light of 729 00:36:06,079 --> 00:36:09,159 consciousness to the stars. So, in the grand scheme of things, 730 00:36:09,199 --> 00:36:12,320 when you look at how much energy Earth is actually 731 00:36:12,400 --> 00:36:15,400 using for civilization, we're only right now using called it 732 00:36:15,519 --> 00:36:19,079 roughly one percent of the potential energy of Earth. And 733 00:36:19,400 --> 00:36:21,840 if we wanted to use even a millionth of the 734 00:36:21,880 --> 00:36:25,039 Sun's energy, that would be roughly a million times more 735 00:36:25,159 --> 00:36:28,960 energy than civilization currently uses. The only way to access 736 00:36:29,079 --> 00:36:31,360 that that energy, the energy of the Sun, is to 737 00:36:31,400 --> 00:36:34,840 extend beyond Earth. Earth is really a tiny, tiny dust 738 00:36:34,880 --> 00:36:38,679 mote in a vast darkness. The Sun is ninety nine 739 00:36:38,719 --> 00:36:41,000 point eight percent of all mass in the Solar system. 740 00:36:41,239 --> 00:36:44,679 So you have to expand beyond the tiny dust mote 741 00:36:44,679 --> 00:36:48,880 that is Earth to make any significant dent in using 742 00:36:48,960 --> 00:36:49,760 the Sun's energy. 743 00:36:49,960 --> 00:36:52,000 Speaker 2: Like says, you'd have to expand roughly. 744 00:36:51,800 --> 00:36:54,239 Speaker 1: A million times just to get to one millionth of 745 00:36:54,719 --> 00:36:58,719 our Sun's energy, and then going beyond that exploring, extending 746 00:36:58,760 --> 00:37:01,639 to the galaxy and maybe someday even to other galaxies. 747 00:37:01,760 --> 00:37:02,480 Speaker 2: So the. 748 00:37:04,320 --> 00:37:08,559 Speaker 1: Next step beyond Earth data centers is our Earth orbital 749 00:37:08,599 --> 00:37:13,039 data centers, and we'll be launching with SpaceX orbital data 750 00:37:13,079 --> 00:37:15,719 centers at the one hundred to two hundred gigawat per 751 00:37:15,800 --> 00:37:19,760 year level, not cumulative, I mean per year. And ultimately 752 00:37:20,159 --> 00:37:22,480 we see a path to maybe launching as much as 753 00:37:22,679 --> 00:37:25,480 a terror wat per year of compute from Earth. But 754 00:37:25,760 --> 00:37:28,400 what if you want to go beyond a mere terror 755 00:37:28,440 --> 00:37:30,239 wide per year. In order to do that, you have 756 00:37:30,280 --> 00:37:32,440 to go to the Moon. So by having factories on 757 00:37:32,480 --> 00:37:36,000 the Moon, building AI satellites and having a mass driver, 758 00:37:36,159 --> 00:37:37,320 which is the kind of thing you really need to 759 00:37:37,400 --> 00:37:40,079 learn about in or read about in science fiction. But 760 00:37:40,119 --> 00:37:42,320 we're going to make it real. We're actually going to 761 00:37:42,400 --> 00:37:45,119 have a mass driver on the Moon. And if you 762 00:37:45,239 --> 00:37:48,159 do that, you can go several orders of magnitude greater. 763 00:37:48,360 --> 00:37:51,119 You can go to one thousand gigawats or more per 764 00:37:51,199 --> 00:37:55,360 year and ultimately get to maybe a millionth and then 765 00:37:55,760 --> 00:37:58,000 a thousandth and maybe even. 766 00:37:57,719 --> 00:37:59,280 Speaker 2: A few percent of the Sun's energy. 767 00:37:59,480 --> 00:38:02,199 Speaker 1: Is simple to imagine what an intelligence of that scale 768 00:38:02,440 --> 00:38:04,880 would think about. But it's going to be incredibly exciting 769 00:38:04,920 --> 00:38:07,039 to see it happen. I really want to see the 770 00:38:07,079 --> 00:38:11,119 mass driver on the Moon that is shooting AI satellites 771 00:38:11,159 --> 00:38:13,400 into deep space. It's going to like shoo shoom, just 772 00:38:13,559 --> 00:38:16,760 one after the other. I can't imagine anything more epic 773 00:38:16,960 --> 00:38:19,280 than a mass driver on the Moon and a self 774 00:38:19,280 --> 00:38:21,519 sustaining city on the Moon, and then going beyond the 775 00:38:21,519 --> 00:38:25,400 Moon to Mars, going throughout our Solar system, and ultimately 776 00:38:26,199 --> 00:38:29,119 going being out there among the stars and visiting all 777 00:38:29,199 --> 00:38:32,639 these star systems. Maybe we'll meet aliens, uh, maybe we'll 778 00:38:32,920 --> 00:38:35,840 see some civilizations that lasted for millions of years, and 779 00:38:35,920 --> 00:38:38,760 we'll find the remnants of ancient alien civilizations. But the 780 00:38:38,760 --> 00:38:40,320 only way we're going to do that is if we 781 00:38:40,400 --> 00:38:42,639 go out there and we explore, and this is the 782 00:38:42,679 --> 00:38:43,760 path to making it happen. 783 00:38:43,880 --> 00:38:51,960 Speaker 2: Thank you. Wow, epic end to a great presentation. 784 00:38:52,280 --> 00:38:52,840 Speaker 5: Yes it's done. 785 00:38:52,920 --> 00:38:54,760 Speaker 2: If that's all he wanted to say, catch you, light up. 786 00:38:54,840 --> 00:38:55,760 Speaker 5: I'm going to share a few of my 787 00:38:55,840 --> 00:38:58,880 Speaker 10: K type Thanks for listening, See you in the next episode.