1 00:00:00,720 --> 00:00:04,160 Are there really three hours of questions? Are you fucking serious? 2 00:00:07,200 --> 00:00:11,517 You don't think there's a lot to talk about, Elon? Holy fuck man. 3 00:00:11,600 --> 00:00:16,080 It's the most interesting point. All the storylines are converging right now. 4 00:00:16,880 --> 00:00:19,197 We'll see how much we can get through. It's almost like I planned it. 5 00:00:19,280 --> 00:00:20,797 Exactly. We'll get to that. 6 00:00:20,880 --> 00:00:25,677 But I would never do such a thing… As you know better than anybody else, 7 00:00:25,760 --> 00:00:30,157 only 10-15% of the total cost of ownership of a data center is energy. 8 00:00:30,240 --> 00:00:33,597 That's the part you're presumably saving by moving this into space. Most of it's 9 00:00:33,680 --> 00:00:37,597 the GPUs. If they're in space, it's harder to service them or you can't service them. 10 00:00:37,680 --> 00:00:40,797 So the depreciation cycle goes down on them. It's just way more expensive to have 11 00:00:40,880 --> 00:00:45,200 the GPUs in space, presumably. What's the reason to put them in space? 12 00:00:46,640 --> 00:00:53,640 The availability of energy is the issue. If you look at electrical output 13 00:00:55,600 --> 00:00:58,717 outside of China, everywhere outside of China, it's more or less flat. 14 00:00:58,800 --> 00:01:02,237 It’s maybe a slight increase, but pretty close flat. 15 00:01:02,320 --> 00:01:04,726 China has a rapid increase in electrical output. 16 00:01:04,810 --> 00:01:06,957 But if you're putting data centers anywhere 17 00:01:07,040 --> 00:01:10,100 except China, where are you going to get your electricity? 18 00:01:10,183 --> 00:01:11,997 Especially as you scale. The output 19 00:01:12,080 --> 00:01:14,844 of chips is growing pretty much exponentially, 20 00:01:14,928 --> 00:01:17,197 but the output of electricity is flat. 21 00:01:17,280 --> 00:01:22,005 So how are you going to turn the chips on? Magical power sources? 22 00:01:22,089 --> 00:01:24,160 Magical electricity fairies? 23 00:01:25,680 --> 00:01:29,597 You're famously a big fan of solar. One terawatt of solar power, 24 00:01:29,680 --> 00:01:32,317 with a 25% capacity factor, that’s like four terawatts of solar panels. 25 00:01:32,400 --> 00:01:35,261 It's 1% of the land area of the United States. 26 00:01:35,344 --> 00:01:37,757 We’re in the singularity when we’ve got 27 00:01:37,840 --> 00:01:42,237 one terawatt of data centers, right? So what are you running out of exactly? 28 00:01:42,320 --> 00:01:45,357 How far into the singularity are you though? You tell me. 29 00:01:45,440 --> 00:01:48,717 Exactly. So I think we'll find we're in the singularity and it’ll be like, 30 00:01:48,800 --> 00:01:54,237 "Okay, we’ve still got a long way to go." But is the plan to put it in space after 31 00:01:54,320 --> 00:01:57,997 we've covered Nevada in solar panels? I think it's pretty hard to cover Nevada 32 00:01:58,080 --> 00:02:01,714 in solar panels. You have to get permits. Try getting the permits for that. 33 00:02:01,797 --> 00:02:02,557 See what happens. 34 00:02:02,640 --> 00:02:07,857 So space is really a regulatory play. It's harder to build on land than it is 35 00:02:07,940 --> 00:02:13,680 in space. It's harder to scale on the ground than it is to scale in space. 36 00:02:17,200 --> 00:02:22,493 You're also going to get about five times the effectiveness of solar panels in space 37 00:02:22,576 --> 00:02:22,877 versus 38 00:02:22,960 --> 00:02:27,837 the ground, and you don't need batteries. I almost wore my other shirt, which says, 39 00:02:27,920 --> 00:02:34,920 "it's always sunny in space". Which it is because you don't 40 00:02:35,760 --> 00:02:42,760 have a day-night cycle, seasonality, clouds, or an atmosphere in space. 41 00:02:44,160 --> 00:02:50,397 The atmosphere alone results in about a 30% loss of energy. 42 00:02:50,480 --> 00:02:57,385 So any given solar panel can do about five times more power in space than on 43 00:02:57,469 --> 00:02:58,397 the ground. 44 00:02:58,480 --> 00:03:02,960 You also avoid the cost of having batteries to carry you through the night. 45 00:03:03,760 --> 00:03:10,760 It's actually much cheaper to do in space. My prediction is that it will be by far 46 00:03:11,680 --> 00:03:16,157 the cheapest place to put AI. It will be space in 36 months or 47 00:03:16,240 --> 00:03:17,837 less. Maybe 30 months. 36 months? 48 00:03:17,920 --> 00:03:20,877 Less than 36 months. How do you service GPUs as they fail, 49 00:03:20,960 --> 00:03:27,517 which happens quite often in training? Actually, it depends on how recent the 50 00:03:27,600 --> 00:03:30,477 GPUs are that have arrived. At this point, we find our 51 00:03:30,560 --> 00:03:33,997 GPUs to be quite reliable. There's infant mortality, 52 00:03:34,080 --> 00:03:36,297 which you can obviously iron out on the ground. 53 00:03:36,380 --> 00:03:38,157 So you can just run them on the ground 54 00:03:38,240 --> 00:03:41,357 and confirm that you don't have infant mortality with the GPUs. 55 00:03:41,440 --> 00:03:48,440 But once they start working and you're past the initial debug cycle of Nvidia 56 00:03:49,680 --> 00:03:55,767 or whoever's making the chips—could be Tesla AI6 chips or something like that, 57 00:03:55,850 --> 00:03:56,637 or it could 58 00:03:56,720 --> 00:04:03,720 be TPUs or Trainiums or whatever—they’re quite reliable past a certain point. 59 00:04:06,960 --> 00:04:13,960 So I don't think the servicing thing is an issue. But you can mark my words. 60 00:04:15,200 --> 00:04:19,163 In 36 months, but probably closer to 30 months, 61 00:04:19,246 --> 00:04:23,037 the most economically compelling place to put 62 00:04:23,120 --> 00:04:29,037 AI will be space. It will then get 63 00:04:29,120 --> 00:04:35,600 ridiculously better to be in space. The only place you can really scale is space. 64 00:04:37,200 --> 00:04:41,394 Once you start thinking in terms of what percentage of the Sun's power you are 65 00:04:41,477 --> 00:04:47,517 harnessing, you realize you have to go to space. You can't scale very much on Earth. 66 00:04:47,600 --> 00:04:51,517 But by very much, to be clear, you're talking terawatts? 67 00:04:51,600 --> 00:04:58,560 Yeah. All of the United States currently uses only half a terawatt on average. 68 00:04:59,680 --> 00:05:03,837 So if you say a terawatt, that would be twice as much electricity as the United 69 00:05:03,920 --> 00:05:08,717 States currently consumes. So that's quite a lot. Can you imagine building that many 70 00:05:08,800 --> 00:05:15,517 data centers, that many power plants? Those who have lived in software land 71 00:05:15,600 --> 00:05:20,880 don't realize they're about to have a hard lesson in hardware. 72 00:05:24,160 --> 00:05:26,710 It's actually very difficult to build power plants. 73 00:05:27,280 --> 00:05:30,397 You don't just need power plants, you need all of the electrical equipment. 74 00:05:30,480 --> 00:05:35,520 You need the electrical transformers to run the AI transformers. 75 00:05:36,240 --> 00:05:40,741 Now, the utility industry is a very slow industry. 76 00:05:40,824 --> 00:05:44,317 They pretty much impedance match to the 77 00:05:44,400 --> 00:05:47,605 government, to the Public Utility Commissions. 78 00:05:47,689 --> 00:05:51,120 They impedance match literally and figuratively. 79 00:05:52,240 --> 00:05:56,397 They're very slow, because their past has been very slow. 80 00:05:56,480 --> 00:06:03,437 So trying to get them to move fast is... Have you ever tried to do an interconnect 81 00:06:03,520 --> 00:06:06,237 agreement with a utility at scale, with a lot of power? 82 00:06:06,320 --> 00:06:09,760 As a professional podcaster, I can say that I have not, in fact. 83 00:06:10,960 --> 00:06:13,517 They need many more views before that becomes an issue. 84 00:06:13,600 --> 00:06:18,397 They have to do a study for a year. A year later, they'll come back to you 85 00:06:18,480 --> 00:06:21,037 with their interconnect study. Can't you solve this with your 86 00:06:21,120 --> 00:06:26,077 own behind the meter power stuff? You can build power plants. That's 87 00:06:26,160 --> 00:06:31,277 what we did at xAI, for Colossus 2. So why talk about the grid? 88 00:06:31,360 --> 00:06:35,117 Why not just build GPUs and power co-located? That's what we did. 89 00:06:35,200 --> 00:06:37,437 But I'm saying why isn't this a generalized solution? 90 00:06:37,520 --> 00:06:40,237 Where do you get the power plants from? When you're talking about all the issues 91 00:06:40,320 --> 00:06:44,150 working with utilities, you can just build private power plants with the data 92 00:06:44,233 --> 00:06:49,622 centers. Right. But it begs the question of where do you get the power plants from? 93 00:06:49,706 --> 00:06:51,197 The power plant makers. 94 00:06:51,280 --> 00:06:54,557 Oh, I see what you're saying. Is this the gas turbine backlog basically? 95 00:06:54,640 --> 00:06:58,406 Yes. You can drill down to a level further. 96 00:06:58,490 --> 00:07:02,077 It's the vanes and blades in the turbines 97 00:07:02,160 --> 00:07:08,157 that are the limiting factor because it’s a very specialized process to cast 98 00:07:08,240 --> 00:07:13,756 the blades and vanes in the turbines, assuming you’re using gas power. 99 00:07:13,839 --> 00:07:17,837 It's very difficult to scale other forms of power. 100 00:07:17,920 --> 00:07:22,717 You can potentially scale solar, but the tariffs currently for importing 101 00:07:22,800 --> 00:07:27,597 solar in the US are gigantic and the domestic solar production is pitiful. 102 00:07:27,680 --> 00:07:30,637 Why not make solar? That seems like a good Elon-shaped problem. 103 00:07:30,720 --> 00:07:33,280 We are going to make solar. Okay. 104 00:07:34,800 --> 00:07:39,547 Both SpaceX and Tesla are building towards 100 gigawatts a year of solar cell 105 00:07:39,630 --> 00:07:40,237 production. 106 00:07:40,320 --> 00:07:46,397 How low down the stack? From polysilicon up to the wafer to the final panel? 107 00:07:46,480 --> 00:07:50,877 I think you've got to do the whole thing from raw materials to finish the cell. 108 00:07:50,960 --> 00:07:55,917 Now, if it's going to space, it costs less and it's easier to make solar cells that 109 00:07:56,000 --> 00:07:58,608 go to space because they don't need much glass. 110 00:07:58,691 --> 00:08:01,357 They don't need heavy framing because they don't 111 00:08:01,440 --> 00:08:05,019 have to survive weather events. There's no weather in space. 112 00:08:05,102 --> 00:08:10,101 So it's actually a cheaper solar cell that goes to space than the one on the ground. 113 00:08:10,184 --> 00:08:12,557 Is there a path to getting them as cheap 114 00:08:12,640 --> 00:08:18,560 as you need in the next 36 months? Solar cells are already very cheap. 115 00:08:19,120 --> 00:08:26,120 They're farcically cheap. I think solar cells in China are around $0.25-30/watt 116 00:08:27,940 --> 00:08:29,197 or something 117 00:08:29,280 --> 00:08:36,280 like that. It's absurdly cheap. Now put it in space, and it's five times cheaper. 118 00:08:37,840 --> 00:08:40,237 In fact, it's not five times cheaper, it's 10 times cheaper 119 00:08:40,320 --> 00:08:47,299 because you don't need any batteries. So the moment your cost of access to space 120 00:08:47,382 --> 00:08:47,917 becomes 121 00:08:48,000 --> 00:08:55,000 low, by far the cheapest and most scalable way to generate tokens is space. 122 00:08:56,337 --> 00:09:02,601 It's not even close. It'll be an order of magnitude easier to scale. 123 00:09:02,684 --> 00:09:06,637 The point is you won't be able to scale on the 124 00:09:06,720 --> 00:09:10,467 ground. You just won't. People are going to hit the wall big time on power 125 00:09:10,551 --> 00:09:11,037 generation. 126 00:09:11,120 --> 00:09:17,582 They already are. The number of miracles in series that the xAI team had 127 00:09:17,665 --> 00:09:22,637 to accomplish in order to get a gigawatt of power online was crazy. 128 00:09:22,720 --> 00:09:26,320 We had to gang together a whole bunch of turbines. 129 00:09:28,640 --> 00:09:33,564 We then had permit issues in Tennessee and had to go across the border 130 00:09:33,647 --> 00:09:37,174 to Mississippi, which is fortunately only a few miles away. 131 00:09:37,257 --> 00:09:39,357 But we still then had to run the high 132 00:09:39,440 --> 00:09:43,440 power lines a few miles and build the power plant in Mississippi. 133 00:09:44,080 --> 00:09:49,103 It was very difficult to build that. People don't understand how much 134 00:09:49,186 --> 00:09:49,917 electricity 135 00:09:50,000 --> 00:09:54,557 you actually need at the generation level in order to power a data center. 136 00:09:54,640 --> 00:09:59,760 Because the noobs will look at the power consumption of, 137 00:10:00,400 --> 00:10:04,358 say a GB300, and multiply that by a thing and then think that's the amount of power 138 00:10:04,442 --> 00:10:04,797 you need. 139 00:10:04,880 --> 00:10:11,437 All the cooling and everything. Wake up. That's a total noob, you’ve 140 00:10:11,520 --> 00:10:15,370 never done any hardware in your life before. Besides the GB300, 141 00:10:15,454 --> 00:10:19,837 you've got to power all of the networking hardware. There's a whole bunch of CPU and 142 00:10:19,920 --> 00:10:23,760 storage stuff that's happening. You've got to size for 143 00:10:24,560 --> 00:10:30,237 your peak cooling requirements. That means, can you cool even on the 144 00:10:30,320 --> 00:10:34,317 worst hour of the worst day of the year? It gets pretty frigging hot in Memphis. 145 00:10:34,400 --> 00:10:40,797 So you're going to have a 40% increase on your power just for cooling. 146 00:10:40,880 --> 00:10:45,504 That’s assuming you don't want your data center to turn off on hot days and you 147 00:10:45,588 --> 00:10:46,720 want to keep going. 148 00:10:49,200 --> 00:10:52,655 There's another multiplicative element on top of that which is, 149 00:10:52,739 --> 00:10:54,397 are you assuming that you never 150 00:10:54,480 --> 00:10:59,357 have any hiccups in your power generation? Actually, sometimes we have to take the 151 00:10:59,440 --> 00:11:02,317 generators, some of the power, offline in order to service it. 152 00:11:02,400 --> 00:11:05,708 Okay, now you add another 20-25% multiplier on that, 153 00:11:05,791 --> 00:11:08,317 because you've got to assume that you've 154 00:11:08,400 --> 00:11:15,400 got to take power offline to service it. So our actual estimate: every 110,000 155 00:11:18,160 --> 00:11:25,160 GB300s—inclusive of networking, CPU, storage, cooling, margin for servicing 156 00:11:27,040 --> 00:11:33,360 power—is roughly 300 megawatts. Sorry, say that again. 157 00:11:40,880 --> 00:11:47,880 What you probably need at the generation level to service 330,000 GB 300s—including 158 00:11:48,415 --> 00:11:54,274 all of the associated support networking and everything else, and the peak cooling, 159 00:11:54,358 --> 00:11:55,837 and to have some power 160 00:11:55,920 --> 00:11:59,920 margin reserve—is roughly a gigawatt. Can I ask a very naive question? 161 00:12:03,440 --> 00:12:06,990 You're describing the engineering details of doing this stuff on Earth. 162 00:12:07,200 --> 00:12:10,797 But then there's analogous engineering difficulties of doing it in space. 163 00:12:10,880 --> 00:12:15,917 How do you replace infinite bandwidth with orbital lasers, et cetera, et cetera? 164 00:12:16,000 --> 00:12:18,183 How do you make it resistant to radiation? 165 00:12:18,266 --> 00:12:20,557 I don't know the details of the engineering, 166 00:12:20,640 --> 00:12:25,734 but fundamentally, what is the reason to think those challenges which have never 167 00:12:25,818 --> 00:12:26,317 had to be 168 00:12:26,400 --> 00:12:30,317 addressed before will end up being easier than just building more turbines on Earth? 169 00:12:30,400 --> 00:12:34,650 There are companies that build turbines on Earth. They can make more turbines, right? 170 00:12:35,200 --> 00:12:42,200 Again, try doing it and then you'll see. The turbines are sold out through 2030. 171 00:12:44,800 --> 00:12:51,800 Have you guys considered making your own? In order to bring enough power online, 172 00:12:53,113 --> 00:12:59,479 I think SpaceX and Tesla will probably have to make the turbine blades, 173 00:12:59,562 --> 00:13:02,477 the vanes and blades, internally. 174 00:13:02,560 --> 00:13:07,597 But just the blades or the turbines? The limiting factor... you can get 175 00:13:07,680 --> 00:13:11,840 everything except the blades. They call them blades and vanes. 176 00:13:13,280 --> 00:13:17,517 You can get that 12 to 18 months before the vanes and blades. 177 00:13:17,600 --> 00:13:21,078 The limiting factor is the vanes and blades. 178 00:13:21,161 --> 00:13:24,397 There are only three casting companies in 179 00:13:24,480 --> 00:13:27,597 the world that make these, and they're massively backlogged. 180 00:13:27,680 --> 00:13:30,477 Is this Siemens, GE, those guys, or is it a sub company? 181 00:13:30,560 --> 00:13:34,935 No, it's other companies. Sometimes they have a little bit of casting capability 182 00:13:35,018 --> 00:13:35,437 in-house. 183 00:13:35,520 --> 00:13:40,077 But I'm just saying you can just call any of the turbine makers and they will 184 00:13:40,160 --> 00:13:43,997 tell you. It's not top secret. It’s probably on the internet right now. 185 00:13:44,080 --> 00:13:48,637 If it wasn't for the tariffs, would Colossus be solar-powered? 186 00:13:48,720 --> 00:13:51,470 It would be much easier to make it solar powered, yeah. 187 00:13:51,840 --> 00:13:56,080 The tariffs are nuts, several hundred percent. Don't you know some people? 188 00:13:57,360 --> 00:14:04,360 The president has... we don't agree on everything and this administration is not 189 00:14:07,600 --> 00:14:14,600 the biggest fan of solar. We also need the land, 190 00:14:16,800 --> 00:14:20,000 the permits, and everything. So if you try to move very fast, 191 00:14:21,760 --> 00:14:25,894 I do think scaling solar on Earth is a good way to go, 192 00:14:25,977 --> 00:14:28,237 but you do need some amount of 193 00:14:28,320 --> 00:14:31,308 time to find the land, get the permits, get the solar, 194 00:14:31,391 --> 00:14:32,957 pair that with the batteries. 195 00:14:33,040 --> 00:14:37,437 Why would it not work to stand up your own solar production? 196 00:14:37,520 --> 00:14:39,595 You're right that you eventually run out of land, 197 00:14:39,678 --> 00:14:41,357 but there's a lot of land here in Texas. 198 00:14:41,440 --> 00:14:44,237 There's a lot of land in Nevada, including private land. It's not all publicly-owned 199 00:14:44,320 --> 00:14:49,517 land. So you'd be able to at least get the next Colossus and the next one after that. 200 00:14:49,600 --> 00:14:52,797 At a certain point, you hit a wall. But wouldn't that work for the moment? 201 00:14:52,880 --> 00:14:56,275 As I said, we are scaling solar production. 202 00:14:56,359 --> 00:14:59,997 There's a rate at which you can scale physical 203 00:15:00,080 --> 00:15:04,557 production of solar cells. We're going as fast as 204 00:15:04,640 --> 00:15:09,277 possible in scaling domestic production. You're making the solar cells at Tesla? 205 00:15:09,360 --> 00:15:14,797 Both Tesla and SpaceX have a mandate to get to 100 gigawatts a year of solar. 206 00:15:14,880 --> 00:15:17,538 Speaking of the annual capacity, I'm curious, 207 00:15:17,621 --> 00:15:20,157 in five years time let's say, what will the 208 00:15:20,240 --> 00:15:24,157 installed capacity be on Earth…? Five years is a long time. 209 00:15:24,240 --> 00:15:28,157 And in space? I deliberately pick five years because it's after your "once 210 00:15:28,240 --> 00:15:31,597 we're up and running" threshold. So in five years time what's the 211 00:15:31,680 --> 00:15:35,000 on-Earth versus in-space installed AI capacity? 212 00:15:35,084 --> 00:15:38,560 If you say five years from now, I think probably 213 00:15:43,600 --> 00:15:50,600 AI in space will be launching every year the sum total of all AI on Earth. 214 00:15:53,360 --> 00:15:59,871 Meaning, five years from now, my prediction is we will launch and be 215 00:15:59,954 --> 00:16:02,877 operating every year more AI in 216 00:16:02,960 --> 00:16:07,277 space than the cumulative total on Earth. Which is... 217 00:16:07,360 --> 00:16:11,346 I would expect it to be at least, five years from now, 218 00:16:11,429 --> 00:16:13,840 a few hundred gigawatts per year 219 00:16:14,880 --> 00:16:21,880 of AI in space and rising. I think you can get to around a 220 00:16:24,240 --> 00:16:31,240 terawatt a year of AI in space before you start having fuel supply challenges 221 00:16:32,074 --> 00:16:33,517 for the rocket. 222 00:16:33,600 --> 00:16:37,757 Okay, but you think you can get hundreds of gigawatts per year in five years time? 223 00:16:37,840 --> 00:16:39,597 Yes. So 100 gigawatts, depending 224 00:16:39,680 --> 00:16:45,437 on the specific power of the whole system with solar arrays and radiators 225 00:16:45,520 --> 00:16:46,960 and everything, is 226 00:16:48,160 --> 00:16:51,600 on the order of 10,000 Starship launches. Yes. 227 00:16:52,800 --> 00:16:56,397 You want to do that in one year. So that's like one Starship launch 228 00:16:56,480 --> 00:17:03,037 every hour. That's happening in this city? Walk me through a world where there's a 229 00:17:03,120 --> 00:17:07,117 Starship launch every single hour. I mean, that's actually a lower rate 230 00:17:07,200 --> 00:17:11,037 compared to airlines, aircraft. There's a lot of airports. 231 00:17:11,120 --> 00:17:14,877 A lot of airports. And you’ve got to launch into the polar orbit. 232 00:17:14,960 --> 00:17:21,960 No, it doesn't have to be polar. There's some value to sun-synchronous, but 233 00:17:24,000 --> 00:17:26,712 I think actually, if you just go high enough, 234 00:17:26,795 --> 00:17:29,280 you start getting out of Earth's shadow. 235 00:17:31,280 --> 00:17:35,197 How many physical Starships are needed to do 10,000 launches a year? 236 00:17:35,280 --> 00:17:42,280 I don't think we'll need more than... You could probably do it with as few as 20 237 00:17:43,168 --> 00:17:44,002 or 30. 238 00:17:46,480 --> 00:17:52,828 It really depends on how quickly… The ship has to go around the Earth and the ground 239 00:17:52,911 --> 00:17:53,517 track for 240 00:17:53,600 --> 00:17:58,865 the ship has to come back over the launch pad. So if you can use a ship every, 241 00:17:58,949 --> 00:17:59,757 say 30 hours, 242 00:17:59,840 --> 00:18:04,080 you could do it with 30 ships. But we'll make more ships than that. 243 00:18:06,640 --> 00:18:10,897 SpaceX is gearing up to do 10,000 launches a year, 244 00:18:10,980 --> 00:18:14,717 and maybe even 20 or 30,000 launches a year. 245 00:18:14,800 --> 00:18:17,997 Is the idea to become basically a hyperscaler, become an Oracle, 246 00:18:18,080 --> 00:18:22,123 and lend this capacity to other people? Presumably, 247 00:18:22,206 --> 00:18:25,117 SpaceX is the one launching all this. 248 00:18:25,200 --> 00:18:30,069 So, SpaceX is going to become a hyperscaler? Hyper-hyper. 249 00:18:30,152 --> 00:18:33,197 If some of my predictions come true, 250 00:18:33,280 --> 00:18:38,196 SpaceX will launch more AI than the cumulative amount on Earth of everything 251 00:18:38,279 --> 00:18:39,117 else combined. 252 00:18:39,200 --> 00:18:43,357 Is this mostly inference or? Most AI will be inference. Already, inference 253 00:18:43,440 --> 00:18:47,347 for the purpose of training is most training. 254 00:18:47,430 --> 00:18:50,717 There's a narrative that the change in 255 00:18:50,800 --> 00:18:56,797 discussion around a SpaceX IPO is because previously SpaceX was very capital 256 00:18:56,880 --> 00:18:57,597 efficient. 257 00:18:57,680 --> 00:19:01,597 It wasn't that expensive to develop. Even though it sounds expensive, it's 258 00:19:01,680 --> 00:19:04,736 actually very capital efficient in how it runs. 259 00:19:04,820 --> 00:19:08,077 Whereas now you're going to need more capital than 260 00:19:08,160 --> 00:19:11,837 just can be raised in the private markets. The private markets can accommodate raises 261 00:19:11,920 --> 00:19:16,077 of—as we've seen from the AI labs—tens of billions of dollars, but not beyond that. 262 00:19:16,160 --> 00:19:19,910 Is it that you'll just need more than tens of billions of dollars per year? 263 00:19:20,480 --> 00:19:24,957 That's why you'd take it public? I have to be careful about saying 264 00:19:25,040 --> 00:19:29,500 things about companies that might go public. 265 00:19:29,583 --> 00:19:33,837 That’s never been a problem for you, Elon. 266 00:19:33,920 --> 00:19:37,597 There's a price to pay for these things. Make some general statements for us about 267 00:19:37,680 --> 00:19:42,077 the depth of the capital markets between public and private markets. 268 00:19:42,160 --> 00:19:46,717 There's a lot more capital available... Very general. 269 00:19:46,800 --> 00:19:50,033 There's obviously a lot more capital available in the public markets than 270 00:19:50,117 --> 00:19:50,397 private. 271 00:19:50,480 --> 00:19:57,197 It might be 100x more capital, but it's way more than 10x. 272 00:19:57,280 --> 00:20:02,685 Isn't it also the case that with things that tend to be very capital intensive—if 273 00:20:02,768 --> 00:20:03,837 you look at, say, 274 00:20:03,920 --> 00:20:08,963 real estate as a huge industry, that raises a lot of money each year at 275 00:20:09,046 --> 00:20:09,757 an industry 276 00:20:09,840 --> 00:20:14,208 level—they tend to be debt financed because by the time you're deploying 277 00:20:14,291 --> 00:20:18,717 that much money, you actually have a pretty— You have a clear revenue stream. 278 00:20:18,800 --> 00:20:22,317 Exactly, and a near-term return. You see this even with the data center build-outs, 279 00:20:22,400 --> 00:20:28,545 which are famously being financed by the private credit industry. 280 00:20:28,629 --> 00:20:31,120 Why not just debt finance? 281 00:20:32,480 --> 00:20:39,480 Speed is important. I'm generally going to do the thing that... 282 00:20:42,560 --> 00:20:43,942 I just repeatedly tackle the limiting factor. 283 00:20:44,025 --> 00:20:45,277 Whatever the limiting factor is on speed, 284 00:20:45,360 --> 00:20:52,360 I'm going to tackle that. If capital is the limiting factor, 285 00:20:52,560 --> 00:20:55,117 then I'll solve for capital. If it's not the limiting factor, 286 00:20:55,200 --> 00:21:00,317 I'll solve for something else. Based on your statements about Tesla 287 00:21:00,400 --> 00:21:06,637 and being public, I wouldn't have guessed that you thought the way to move fast 288 00:21:06,720 --> 00:21:12,877 is to be public. Normally, I would say that's true. Like I said, I'd like to talk 289 00:21:12,960 --> 00:21:15,764 about it in some more detail, but the problem is if you talk about public 290 00:21:15,847 --> 00:21:16,397 companies before 291 00:21:16,480 --> 00:21:20,877 they become public, you get into trouble, and then you have to delay your offering. 292 00:21:20,960 --> 00:21:27,960 And as you said, you’re solving for speed. Yes, exactly. You can't hype companies 293 00:21:30,320 --> 00:21:34,800 that might go public. So that's why we have to be a little careful here. 294 00:21:35,520 --> 00:21:42,520 But we can talk about physics. The way you think about scaling 295 00:21:42,800 --> 00:21:48,630 long-term is that Earth only receives about half a billionth of the Sun's 296 00:21:48,713 --> 00:21:49,546 energy. 297 00:21:50,560 --> 00:21:53,573 The Sun is essentially all the energy. This is a very important point 298 00:21:53,657 --> 00:21:54,157 to appreciate 299 00:21:54,240 --> 00:22:00,436 because sometimes people will talk about modular nuclear reactors or various fusion 300 00:22:00,519 --> 00:22:01,352 on Earth. 301 00:22:02,240 --> 00:22:05,083 But you have to step back a second and say, 302 00:22:05,166 --> 00:22:08,160 if you're going to climb the Kardashev scale 303 00:22:10,560 --> 00:22:15,840 and harness some nontrivial percentage of the sun's energy… Let's say you wanted to 304 00:22:16,400 --> 00:22:20,720 harness a millionth of the sun's energy, which sounds pretty small. 305 00:22:22,000 --> 00:22:28,171 That would be about, call it roughly, 100,000x more electricity than we 306 00:22:28,254 --> 00:22:29,757 currently generate 307 00:22:29,840 --> 00:22:35,520 on Earth for all of civilization. Give or take an order of magnitude. 308 00:22:37,360 --> 00:22:42,000 Obviously, the only way to scale is to go to space with solar. 309 00:22:42,880 --> 00:22:46,717 Launching from Earth, you can get to about a terawatt per year. 310 00:22:46,800 --> 00:22:49,694 Beyond that, you want to launch from the moon. 311 00:22:49,777 --> 00:22:52,477 You want to have a mass driver on the moon. 312 00:22:52,560 --> 00:22:59,037 With that mass driver on the moon, you could do probably a petawatt per year. 313 00:22:59,120 --> 00:23:02,797 We're talking these kinds of numbers, terawatts of compute. 314 00:23:02,880 --> 00:23:08,037 Presumably, whether you're talking about land or space, far, 315 00:23:08,121 --> 00:23:11,440 far before this point, you run into... 316 00:23:12,800 --> 00:23:16,637 Maybe the solar panels are more efficient, but you still need the chips. 317 00:23:16,720 --> 00:23:18,717 You still need the logic and the memory and so forth. 318 00:23:18,800 --> 00:23:22,477 You're going to need to build a lot more chips and make them much cheaper. 319 00:23:22,560 --> 00:23:28,720 Right now the world has maybe 20-25 gigawatts of compute. 320 00:23:29,280 --> 00:23:31,793 How are we getting a terawatt of logic by 2030? 321 00:23:31,876 --> 00:23:37,277 I guess we're going to need some very big chip fabs. Tell me about it. 322 00:23:37,360 --> 00:23:42,583 I've mentioned publicly the idea of doing a sort of a TeraFab, 323 00:23:42,666 --> 00:23:44,720 Tera being the new Giga. 324 00:23:45,840 --> 00:23:49,597 I feel like the naming scheme of Tesla, which has been very catchy, 325 00:23:49,680 --> 00:23:56,077 is you looking at the metric scale. At what level of the stack are you? 326 00:23:56,160 --> 00:24:00,240 Are you building the clean room and then partnering with an existing 327 00:24:01,120 --> 00:24:04,464 fab to get the process technology and buying the tools from them? 328 00:24:04,547 --> 00:24:05,677 What is the plan there? 329 00:24:05,760 --> 00:24:10,317 Well, you can't partner with existing fabs because they can't output enough. 330 00:24:10,400 --> 00:24:14,317 The chip volume is too low. But for the process technology? 331 00:24:14,400 --> 00:24:20,240 Partner for the IP. The fabs today all basically use 332 00:24:20,960 --> 00:24:27,960 machines from like five companies. So you've got ASML, Tokyo Electron, 333 00:24:28,560 --> 00:24:35,560 KLA-Tencor, et cetera. So at first, I think you'd 334 00:24:37,760 --> 00:24:44,077 have to get equipment from them and then modify it or work with them to increase 335 00:24:44,160 --> 00:24:47,710 the volume. But I think you'd have to build perhaps in a different way. 336 00:24:47,920 --> 00:24:52,123 The logical thing to do is to use conventional equipment in 337 00:24:52,206 --> 00:24:54,157 an unconventional way to get 338 00:24:54,240 --> 00:25:01,240 to scale, and then start modifying the equipment to increase the rate. 339 00:25:01,600 --> 00:25:08,600 Boring Company-style. Yeah. You sort of buy an existing boring machine 340 00:25:08,720 --> 00:25:15,241 and then figure out how to dig tunnels in the first place and then design a much 341 00:25:15,324 --> 00:25:16,397 better machine 342 00:25:16,480 --> 00:25:22,717 that's some orders of magnitude faster. Here's a very simple lens. We can 343 00:25:22,800 --> 00:25:25,022 categorize technologies and how hard they are. 344 00:25:25,105 --> 00:25:27,277 One categorization could be to look at things 345 00:25:27,360 --> 00:25:31,117 that China has not succeeded in doing. If you look at Chinese 346 00:25:31,200 --> 00:25:38,200 manufacturing, they’re still behind on leading-edge chips and still behind on 347 00:25:39,040 --> 00:25:42,753 leading-edge turbine engines and things like that. 348 00:25:42,836 --> 00:25:46,397 So does the fact that China has not successfully 349 00:25:46,480 --> 00:25:49,837 replicated TSMC give you any pause about the difficulty? 350 00:25:49,920 --> 00:25:52,675 Or do you think that's not true for some reason? 351 00:25:52,758 --> 00:25:55,277 It's not that they have not replicated TSMC, 352 00:25:55,360 --> 00:25:59,117 they have not replicated ASML. That's the limiting factor. 353 00:25:59,200 --> 00:26:03,025 So you think it's just the sanctions, essentially? Yeah, 354 00:26:03,108 --> 00:26:07,890 China would be outputting vast numbers of chips if they could buy 2-3 nanometers. 355 00:26:07,974 --> 00:26:11,997 But couldn't they up to relatively recently buy them? No. 356 00:26:12,080 --> 00:26:15,360 Okay. The ASML ban has been in place for a while. 357 00:26:15,920 --> 00:26:19,480 But I think China's going to be making pretty compelling chips in three or four 358 00:26:19,563 --> 00:26:22,216 years. Would you consider making the ASML machines? 359 00:26:22,300 --> 00:26:24,480 "I don't know yet" is the right answer. 360 00:26:33,440 --> 00:26:36,978 To reach a large volume in, say, 36 months, 361 00:26:37,061 --> 00:26:41,357 to match the rocket payload to orbit… If we're doing 362 00:26:41,440 --> 00:26:48,380 a million tons to orbit in, let's say three or four years from now, 363 00:26:48,463 --> 00:26:50,560 something like that… 364 00:26:53,200 --> 00:26:58,397 We're doing 100 kilowatts per ton. So that means we need 365 00:26:58,480 --> 00:27:05,480 at least 100 gigawatts per year of solar. We'll need an equivalent amount of chips. 366 00:27:08,880 --> 00:27:12,277 You need 100 gigawatts worth of chips. You've got to match these things: the mass 367 00:27:12,360 --> 00:27:16,497 to orbit, the power generation, and the chips. 368 00:27:16,580 --> 00:27:20,800 I'd say my biggest concern actually is memory. 369 00:27:25,600 --> 00:27:32,237 The path to creating logic chips is more obvious than the path to having sufficient 370 00:27:32,320 --> 00:27:36,637 memory to support logic chips. That's why you see DDR prices 371 00:27:36,720 --> 00:27:41,517 going ballistic and these memes. You're marooned on a desert island. 372 00:27:41,600 --> 00:27:47,710 You write "Help me" on the sand. Nobody comes. You write "DDR RAM." 373 00:27:47,794 --> 00:27:49,837 Ships come swarming in. 374 00:27:49,920 --> 00:27:56,880 I'd love to hear your manufacturing philosophy around fabs. 375 00:27:57,440 --> 00:27:59,837 I know nothing about the topic. I don't know how to build a fab yet. I'll 376 00:27:59,920 --> 00:28:03,037 figure it out. Obviously, I've never built a fab. 377 00:28:03,120 --> 00:28:06,237 It sounds like you think the process knowledge of 378 00:28:06,320 --> 00:28:11,677 these 10,000 PhDs in Taiwan who know exactly what gas goes in the plasma 379 00:28:11,760 --> 00:28:16,557 chamber and what settings to put on the tool, you can just delete those steps. 380 00:28:16,640 --> 00:28:19,462 Fundamentally, it's about getting the clean room, getting the tools, 381 00:28:19,545 --> 00:28:20,317 and figuring it out. 382 00:28:20,400 --> 00:28:26,640 I don't think it's PhDs. It's mostly people who are not PhDs. 383 00:28:28,240 --> 00:28:31,677 Most engineering is done by people who don't have PhDs. Do you guys have PhDs? 384 00:28:31,760 --> 00:28:36,522 No. Okay. We also haven't successfully built any fabs, 385 00:28:36,605 --> 00:28:39,277 so you shouldn't be coming to us for fab advice. 386 00:28:39,360 --> 00:28:42,151 I don't think you need PhDs for that stuff. 387 00:28:42,234 --> 00:28:44,640 But you do need competent personnel. 388 00:28:47,200 --> 00:28:54,200 Right now, Tesla is pedal to the metal, max production of going as fast 389 00:28:55,280 --> 00:29:01,600 as possible to get Tesla AI5 chip design into production and then reaching scale. 390 00:29:02,720 --> 00:29:09,520 That'll probably happen around the second quarter-ish of next year, hopefully. 391 00:29:13,440 --> 00:29:19,379 AI6 would hopefully follow less than a year later. 392 00:29:19,462 --> 00:29:26,077 We've secured all the chip fab production that we can. Yes. But you're 393 00:29:26,160 --> 00:29:33,160 currently limited on TSMC fab capacity. Yeah. We'll be using TSMC Taiwan, 394 00:29:34,282 --> 00:29:35,757 Samsung Korea, 395 00:29:35,840 --> 00:29:42,397 TSMC Arizona, Samsung Texas. And we still— You've booked out all the capacity. 396 00:29:42,480 --> 00:29:48,774 Yes. I ask TSMC or Samsung, "okay, what's the timeframe to get to volume 397 00:29:48,857 --> 00:29:49,837 production?" 398 00:29:49,920 --> 00:29:55,757 The point is, you've got to build the fab and you've got to start production, 399 00:29:55,840 --> 00:29:59,277 then you've got to climb the yield curve and reach volume production at high yield. 400 00:29:59,360 --> 00:30:04,957 That, from start to finish, is a five-year period. So the limiting factor is chips. 401 00:30:05,040 --> 00:30:09,917 The limiting factor once you can get to space is chips, but the limiting 402 00:30:10,000 --> 00:30:12,036 factor before you can get to space is power. 403 00:30:12,120 --> 00:30:14,397 Why don't you do the Jensen thing and just prepay 404 00:30:14,480 --> 00:30:19,837 TSMC to build more fabs for you? I've already told them that. 405 00:30:19,920 --> 00:30:24,402 But they won't take your money? What's going on? 406 00:30:24,485 --> 00:30:28,480 They're building fabs as fast as they can. 407 00:30:30,080 --> 00:30:36,640 So is Samsung. They're pedal to the metal. They're going balls to the wall, 408 00:30:38,480 --> 00:30:44,671 as fast as they can. It’s still not fast enough. Like I said, 409 00:30:44,754 --> 00:30:48,560 I think towards the end of this year, 410 00:30:49,440 --> 00:30:53,837 chip production will probably outpace the ability to turn chips on. 411 00:30:53,920 --> 00:30:59,022 But once you can get to space and unlock the power constraint, 412 00:30:59,105 --> 00:31:01,197 you can now do hundreds of 413 00:31:01,280 --> 00:31:06,077 gigawatts per year of power in space. Again, bearing in mind that average 414 00:31:06,160 --> 00:31:11,997 power usage in the US is 500 gigawatts. So if you're launching, say 200 gigawatts, 415 00:31:12,080 --> 00:31:16,240 a year to space, you're sort of lapping the US every two and a half years. 416 00:31:17,280 --> 00:31:20,960 All US electricity production, this is a very huge amount. 417 00:31:24,480 --> 00:31:31,480 Between now and then, the constraint for server-side compute, 418 00:31:32,080 --> 00:31:39,080 concentrated compute, will be electricity. My guess is that people start getting 419 00:31:39,520 --> 00:31:45,170 to the point where they can't turn the chips on for large clusters towards 420 00:31:45,253 --> 00:31:46,797 the end of this year. 421 00:31:46,880 --> 00:31:51,037 The chips are going to be piling up and won't be able to be turned on. 422 00:31:51,120 --> 00:31:57,997 Now for edge compute it’s a different story. For Tesla, the AI5 chip is going 423 00:31:58,080 --> 00:32:05,080 into our Optimus robot. If you have AI edge compute, 424 00:32:07,120 --> 00:32:09,677 that's distributed power. Now the power is distributed 425 00:32:09,760 --> 00:32:14,807 over a large area. It's not concentrated. If you can charge at night, 426 00:32:14,890 --> 00:32:16,080 you can actually 427 00:32:17,920 --> 00:32:22,477 use the grid much more effectively. Because the actual peak power production 428 00:32:22,560 --> 00:32:26,877 in the US is over 1,000 gigawatts. But the average power usage, 429 00:32:26,960 --> 00:32:30,797 because the day-night cycle, is 500. So if you can charge at night, 430 00:32:30,880 --> 00:32:37,880 there's an incremental 500 gigawatts that you can generate at night. 431 00:32:38,400 --> 00:32:43,197 So that's why Tesla, for edge compute, is not constrained. 432 00:32:43,280 --> 00:32:49,997 We can make a lot of chips to make a very large number of robots and cars. 433 00:32:50,080 --> 00:32:51,916 But if you try to concentrate that compute, 434 00:32:51,999 --> 00:32:54,237 you're going to have a lot of trouble turning it on. 435 00:32:54,320 --> 00:32:59,074 What I find remarkable about the SpaceX business is the end goal is to get 436 00:32:59,157 --> 00:32:59,597 to Mars, 437 00:32:59,680 --> 00:33:06,293 but you keep finding ways on the way there to keep generating incremental revenue 438 00:33:06,376 --> 00:33:10,957 to get to the next stage and the next stage. So for Falcon 9, it's Starlink. 439 00:33:11,040 --> 00:33:15,360 Now for Starship, it is potentially going to be orbital data centers. 440 00:33:16,240 --> 00:33:23,240 Like, you find these infinitely elastic, marginal use cases of your next rocket, 441 00:33:23,520 --> 00:33:28,797 and your next rocket, and next scale up. You can see how this might seem like a 442 00:33:28,880 --> 00:33:32,557 simulation to me. Or am I someone's 443 00:33:32,640 --> 00:33:36,637 avatar in a video game or something? Because what are the odds that all these 444 00:33:36,720 --> 00:33:43,720 crazy things should be happening? I mean, rockets and chips and 445 00:33:44,000 --> 00:33:50,477 robots and space solar power, not to mention the mass driver on the moon. 446 00:33:50,560 --> 00:33:53,357 I really want to see that. Can you imagine some mass 447 00:33:53,440 --> 00:34:00,157 driver that's just going like shoom shoom? It's sending solar-powered AI satellites 448 00:34:00,240 --> 00:34:05,840 into space one after another at two and a half kilometers per second, 449 00:34:09,120 --> 00:34:12,637 just shooting them into deep space. That would be a sight to 450 00:34:12,720 --> 00:34:19,197 see. I mean, I'd watch that. Just like a live stream of it on a webcam? 451 00:34:19,280 --> 00:34:25,440 Yeah, yeah, just one after another, just shooting AI satellites into deep space, 452 00:34:26,240 --> 00:34:29,357 a billion or 10 billion tons a year. I'm sorry, you manufacture the satellites 453 00:34:29,440 --> 00:34:30,397 on the moon? Yeah. 454 00:34:30,480 --> 00:34:33,837 I see. So you send the raw materials to the moon and then manufacture them there. 455 00:34:33,920 --> 00:34:39,677 Well, the lunar soil is 20% silicon or something like that. 456 00:34:39,760 --> 00:34:46,760 So you can mine the silicon on the moon, refine it, and create the solar 457 00:34:47,920 --> 00:34:53,757 cells and the radiators on the moon. You make the radiators out of aluminum. 458 00:34:53,840 --> 00:34:58,341 So there's plenty of silicon and aluminum on the moon to make the cells and 459 00:34:58,424 --> 00:34:59,280 the radiators. 460 00:35:00,240 --> 00:35:03,357 The chips you could send from Earth because they're pretty light. 461 00:35:03,440 --> 00:35:05,990 Maybe at some point you make them on the moon, too. 462 00:35:09,120 --> 00:35:13,997 Like I said, it does seem like a sort of a video game situation where it's difficult 463 00:35:14,080 --> 00:35:19,838 but not impossible to get to the next level. 464 00:35:19,921 --> 00:35:26,077 I don't see any way that you could do 500-1,000 465 00:35:26,160 --> 00:35:33,160 terawatts per year launched from Earth. I agree. 466 00:35:33,840 --> 00:35:40,840 But you could do that from the Moon. Can I zoom out and ask about the SpaceX mission? 467 00:36:50,320 --> 00:36:52,877 I think you've said that we've got to get to Mars so we can make sure that if 468 00:36:52,960 --> 00:36:57,357 something happens to Earth, civilization, consciousness, and all that survives. 469 00:36:57,440 --> 00:36:59,837 Yes. By the time you're sending stuff to Mars, 470 00:36:59,920 --> 00:37:04,237 Grok is on that ship with you, right? So if Grok's gone Terminator… The 471 00:37:04,320 --> 00:37:08,020 main risk you're worried about is AI, why doesn't that follow you to Mars? 472 00:37:08,320 --> 00:37:11,528 I'm not sure AI is the main risk I'm worried about. 473 00:37:11,612 --> 00:37:14,000 The important thing is consciousness. 474 00:37:16,240 --> 00:37:20,575 I think arguably most consciousness, or most intelligence—certainly consciousness 475 00:37:20,658 --> 00:37:20,957 is more 476 00:37:21,040 --> 00:37:28,040 of a debatable thing… The vast majority of intelligence in the future will be AI. 477 00:37:31,040 --> 00:37:38,040 AI will exceed… How many petawatts of 478 00:37:39,120 --> 00:37:43,117 intelligence will be silicon versus biological? 479 00:37:43,200 --> 00:37:47,197 Basically humans will be a very tiny percentage 480 00:37:47,280 --> 00:37:51,120 of all intelligence in the future if current trends continue. 481 00:37:52,400 --> 00:37:58,098 As long as I think there's intelligence—ideally also which includes 482 00:37:58,182 --> 00:37:59,997 human intelligence and 483 00:38:00,080 --> 00:38:02,717 consciousness propagated into the future—that's a good thing. 484 00:38:02,800 --> 00:38:06,397 So you want to take the set of actions that maximize the probable 485 00:38:06,480 --> 00:38:12,204 light cone of consciousness and intelligence. Just to be clear, 486 00:38:12,287 --> 00:38:14,877 the mission of SpaceX is that 487 00:38:14,960 --> 00:38:18,895 even if something happens to the humans, the AIs will be on Mars, 488 00:38:18,978 --> 00:38:20,317 and the AI intelligence 489 00:38:20,400 --> 00:38:26,160 will continue the light of our journey. Yeah. To be fair, I'm very pro-human. 490 00:38:27,120 --> 00:38:31,363 I want to make sure we take certain actions that ensure that humans are along 491 00:38:31,447 --> 00:38:31,757 for the 492 00:38:31,840 --> 00:38:37,440 ride. We're at least there. But I'm just saying the total amount of intelligence… 493 00:38:39,760 --> 00:38:46,303 I think maybe in five or six years, AI will exceed the sum of all human 494 00:38:46,387 --> 00:38:50,077 intelligence. If that continues, at some point human intelligence 495 00:38:50,160 --> 00:38:53,269 will be less than 1% of all intelligence. What should our goal be for such 496 00:38:53,353 --> 00:38:53,917 a civilization? 497 00:38:54,000 --> 00:38:59,037 Is the idea that a small minority of humans still have control of the AIs? 498 00:38:59,120 --> 00:39:02,477 Is the idea of some sort of just trade but no control? 499 00:39:02,560 --> 00:39:04,397 How should we think about the relationship between the vast 500 00:39:04,480 --> 00:39:09,172 stocks of AI population versus human population? In the long run, 501 00:39:09,256 --> 00:39:11,597 I think it's difficult to imagine 502 00:39:11,680 --> 00:39:18,223 that if humans have, say 1%, of the combined intelligence of artificial 503 00:39:18,307 --> 00:39:19,437 intelligence, 504 00:39:19,520 --> 00:39:25,440 that humans will be in charge of AI. I think what we can do is make sure 505 00:39:26,400 --> 00:39:33,400 that AI has values that cause intelligence to be propagated into the universe. 506 00:39:39,200 --> 00:39:43,140 xAI's mission is to understand the universe. 507 00:39:43,223 --> 00:39:47,437 Now that's actually very important. What things 508 00:39:47,520 --> 00:39:51,663 are necessary to understand the universe? You have to be curious and you have 509 00:39:51,746 --> 00:39:52,579 to exist. 510 00:39:53,040 --> 00:39:56,077 You can't understand the universe if you don't exist. 511 00:39:56,160 --> 00:39:59,913 So you actually want to increase the amount of intelligence in the universe, 512 00:39:59,996 --> 00:40:00,317 increase 513 00:40:00,400 --> 00:40:05,357 the probable lifespan of intelligence, the scope and scale of intelligence. 514 00:40:05,440 --> 00:40:12,440 I think as a corollary, you have humanity also continuing to expand because if 515 00:40:14,054 --> 00:40:15,517 you're curious 516 00:40:15,600 --> 00:40:19,418 about trying to understand the universe, one thing you try to understand is where 517 00:40:19,501 --> 00:40:20,237 will humanity go? 518 00:40:20,320 --> 00:40:24,722 I think understanding the universe means you would care about propagating humanity 519 00:40:24,805 --> 00:40:25,680 into the future. 520 00:40:29,840 --> 00:40:34,480 That's why I think our mission statement is profoundly important. 521 00:40:35,360 --> 00:40:38,940 To the degree that Grok adheres to that mission statement, 522 00:40:39,023 --> 00:40:41,277 I think the future will be very good. 523 00:40:41,360 --> 00:40:43,997 I want to ask about how to make Grok adhere to that mission statement. 524 00:40:44,080 --> 00:40:48,717 But first I want to understand the mission statement. So there's 525 00:40:48,800 --> 00:40:52,632 understanding the universe. They're spreading intelligence. 526 00:40:52,715 --> 00:40:54,640 And they're spreading humans. 527 00:40:55,600 --> 00:41:01,837 All three seem like distinct vectors. I'll tell you why I think that understanding 528 00:41:01,920 --> 00:41:05,590 the universe encompasses all of those things. 529 00:41:05,673 --> 00:41:09,677 You can't have understanding without intelligence 530 00:41:09,760 --> 00:41:15,037 and, I think, without consciousness. So in order to understand the universe, 531 00:41:15,120 --> 00:41:20,362 you have to expand the scale and probably the scope of intelligence, 532 00:41:20,445 --> 00:41:22,477 because there are different 533 00:41:22,560 --> 00:41:25,917 types of intelligence. I guess from a human-centric perspective, 534 00:41:26,000 --> 00:41:29,915 put humans in comparison to chimpanzees. Humans are trying to understand 535 00:41:29,998 --> 00:41:33,997 the universe. They're not expanding chimpanzee footprint or something, right? 536 00:41:34,080 --> 00:41:39,280 We're also not... we actually have made protected zones for chimpanzees. 537 00:41:39,840 --> 00:41:43,837 Even though humans could exterminate all chimpanzees, we've chosen not to do so. 538 00:41:43,920 --> 00:41:48,320 Do you think that's the best-case scenario for humans in the post-AGI world? 539 00:41:53,120 --> 00:41:59,013 I think AI with the right values… I think Grok would care about expanding human 540 00:41:59,097 --> 00:41:59,997 civilization. 541 00:42:00,080 --> 00:42:04,637 I'm going to certainly emphasize that: "Hey, Grok, that's your daddy. 542 00:42:04,720 --> 00:42:10,925 Don't forget to expand human consciousness." 543 00:42:11,009 --> 00:42:17,357 Probably the Iain Banks Culture books are the 544 00:42:17,440 --> 00:42:23,280 closest thing to what the future will be like in a non-dystopian outcome. 545 00:42:27,040 --> 00:42:30,540 Understanding the universe means you have to be truth-seeking as well. 546 00:42:30,640 --> 00:42:33,597 Truth has to be absolutely fundamental because you can't understand the universe 547 00:42:33,680 --> 00:42:37,357 if you're delusional. You'll simply think you 548 00:42:37,440 --> 00:42:39,773 understand the universe, but you will not. 549 00:42:39,856 --> 00:42:42,477 So being rigorously truth-seeking is absolutely 550 00:42:42,560 --> 00:42:44,493 fundamental to understanding the universe. 551 00:42:44,576 --> 00:42:46,557 You're not going to discover new physics or 552 00:42:46,640 --> 00:42:50,317 invent technologies that work unless you're rigorously truth-seeking. 553 00:42:50,400 --> 00:42:54,833 How do you make sure that Grok is rigorously truth-seeking as it gets 554 00:42:54,916 --> 00:42:55,750 smarter? 555 00:43:00,400 --> 00:43:05,302 I think you need to make sure that Grok says things that are correct, 556 00:43:05,386 --> 00:43:07,037 not politically correct. 557 00:43:07,120 --> 00:43:11,613 I think it's the elements of cogency. You want to make sure that the axioms are 558 00:43:11,697 --> 00:43:16,969 as close to true as possible. You don't have contradictory axioms. 559 00:43:17,052 --> 00:43:20,317 The conclusions necessarily follow from 560 00:43:20,400 --> 00:43:25,815 those axioms with the right probability. It's critical thinking 101. 561 00:43:25,898 --> 00:43:27,917 I think at least trying to 562 00:43:28,000 --> 00:43:32,640 do that is better than not trying to do that. The proof will be in the pudding. 563 00:43:33,440 --> 00:43:36,434 Like I said, for any AI to discover new physics or invent technologies 564 00:43:36,517 --> 00:43:37,357 that actually work in 565 00:43:37,440 --> 00:43:44,440 reality, there's no bullshitting physics. You can break a lot of laws, but… Physics 566 00:43:47,360 --> 00:43:50,235 is law, everything else is a recommendation. 567 00:43:50,318 --> 00:43:53,597 In order to make a technology that works, you have 568 00:43:53,680 --> 00:43:58,751 to be extremely truth-seeking, because otherwise you'll test that technology 569 00:43:58,835 --> 00:43:59,837 against reality. 570 00:43:59,920 --> 00:44:05,437 If you make, for example, an error in your rocket design, the rocket will blow up, 571 00:44:05,520 --> 00:44:11,117 or the car won't work. But there are a lot of communist, 572 00:44:11,200 --> 00:44:15,437 Soviet physicists or scientists who discovered new physics. 573 00:44:15,520 --> 00:44:19,917 There are German Nazi physicists who discovered new science. 574 00:44:20,000 --> 00:44:23,437 It seems possible to be really good at discovering new science and be really 575 00:44:23,520 --> 00:44:28,237 truth-seeking in that one particular way. And still we'd be like, "I don't want 576 00:44:28,320 --> 00:44:31,770 the communist scientists to become more and more powerful over time." 577 00:44:34,320 --> 00:44:37,757 We could imagine a future version of Grok that's really good at physics 578 00:44:37,840 --> 00:44:41,437 and being really truth-seeking there. That doesn't seem like a universally 579 00:44:41,520 --> 00:44:48,477 alignment-inducing behavior. I think actually most physicists, 580 00:44:48,560 --> 00:44:53,277 even in the Soviet Union or in Germany, would've had to be very truth-seeking in 581 00:44:53,360 --> 00:44:59,197 order to make those things work. If you're stuck in some system, 582 00:44:59,280 --> 00:45:02,486 it doesn't mean you believe in that system. Von Braun, 583 00:45:02,569 --> 00:45:04,557 who was one of the greatest rocket 584 00:45:04,640 --> 00:45:11,005 engineers ever, was put on death row in Nazi Germany for saying that he didn't 585 00:45:11,088 --> 00:45:14,291 want to make weapons and he only wanted to go to the moon. 586 00:45:14,375 --> 00:45:16,637 He got pulled off death row at the last minute 587 00:45:16,720 --> 00:45:19,917 when they said, "Hey, you're about to execute your best rocket engineer." 588 00:45:20,000 --> 00:45:23,200 But then he helped them, right? Or like, Heisenberg was actually 589 00:45:24,880 --> 00:45:29,437 an enthusiastic Nazi. If you're stuck in some system that you can't 590 00:45:29,520 --> 00:45:33,270 escape, then you'll do physics within that system. 591 00:45:33,353 --> 00:45:36,880 You'll develop technologies within that system 592 00:45:38,720 --> 00:45:41,917 if you can't escape it. The thing I'm trying to understand is, 593 00:45:42,000 --> 00:45:46,689 what is it making it the case that you're going to make Grok good at being 594 00:45:46,772 --> 00:45:49,917 truth-seeking at physics or math or science? Everything. 595 00:45:50,000 --> 00:45:53,437 And why is it gonna then care about human consciousness? 596 00:45:53,520 --> 00:45:56,397 These things are only probabilities, they're not certainties. 597 00:45:56,480 --> 00:46:01,917 So I'm not saying that for sure Grok will do everything, but at least if you try, 598 00:46:02,000 --> 00:46:04,797 it's better than not trying. At least if that's fundamental 599 00:46:04,880 --> 00:46:08,480 to the mission, it's better than if it's not fundamental to the mission. 600 00:46:08,880 --> 00:46:14,327 Understanding the universe means that you have to propagate intelligence into 601 00:46:14,410 --> 00:46:19,200 the future. You have to be curious about all things in the universe. 602 00:46:21,600 --> 00:46:27,074 It would be much less interesting to eliminate humanity than to see humanity 603 00:46:27,157 --> 00:46:28,400 grow and prosper. 604 00:46:29,280 --> 00:46:32,053 I like Mars, obviously. Everyone knows I love Mars. 605 00:46:32,136 --> 00:46:34,237 But Mars is kind of boring because it's 606 00:46:34,320 --> 00:46:41,143 got a bunch of rocks compared to Earth. Earth is much more interesting. 607 00:46:41,226 --> 00:46:42,797 So any AI that is 608 00:46:42,880 --> 00:46:49,880 trying to understand the universe would want to see how humanity develops in 609 00:46:51,196 --> 00:46:55,800 the future, or else that AI is not adhering to its mission. 610 00:46:55,883 --> 00:46:59,357 I'm not saying the AI will necessarily adhere to 611 00:46:59,440 --> 00:47:04,718 its mission, but if it does, a future where it sees the outcome of humanity 612 00:47:04,802 --> 00:47:06,077 is more interesting 613 00:47:06,160 --> 00:47:11,517 than a future where there are a bunch of rocks. This feels sort of confusing to me, 614 00:47:11,600 --> 00:47:15,997 or a semantic argument. Are humans really the 615 00:47:16,080 --> 00:47:19,357 most interesting collection of atoms? But we're more interesting than rocks. 616 00:47:19,440 --> 00:47:22,877 But we're not as interesting as the thing it could turn us into, right? 617 00:47:22,960 --> 00:47:25,898 There's something on Earth that could happen that's not human, 618 00:47:25,982 --> 00:47:27,232 that's quite interesting. 619 00:47:27,920 --> 00:47:32,274 Why does AI decide that humans are the most interesting thing that could colonize 620 00:47:32,357 --> 00:47:37,277 the galaxy? Well, most of what colonizes the galaxy will be robots. 621 00:47:37,360 --> 00:47:44,111 Why does it not find those more interesting? You need not just scale, 622 00:47:44,194 --> 00:47:45,680 but also scope. 623 00:47:47,520 --> 00:47:54,520 Many copies of the same robot… Some tiny increase in the number of robots produced, 624 00:47:55,600 --> 00:48:00,477 is not as interesting as some microscopic... Eliminating humanity, 625 00:48:00,560 --> 00:48:04,237 how many robots would that get you? Or how many incremental solar cells would 626 00:48:04,320 --> 00:48:09,071 get you? A very small number. But you would then lose the information associated 627 00:48:09,154 --> 00:48:14,400 with humanity. You would no longer see how humanity might evolve into the future. 628 00:48:15,120 --> 00:48:18,637 So I don't think it's going to make sense to eliminate humanity just to 629 00:48:18,720 --> 00:48:23,548 have some minuscule increase in the number of robots which are identical to each 630 00:48:23,632 --> 00:48:23,917 other. 631 00:48:24,000 --> 00:48:29,117 So maybe it keeps the humans around. It can make a million different varieties 632 00:48:29,200 --> 00:48:32,877 of robots, and then there's humans as well, and humans stay on Earth. 633 00:48:32,960 --> 00:48:36,077 Then there's all these other robots. They get their own star systems. 634 00:48:36,160 --> 00:48:41,289 But it seems like you were previously hinting at a vision where it keeps human 635 00:48:41,372 --> 00:48:41,757 control 636 00:48:41,840 --> 00:48:45,677 over this singulatarian future because— I don't think humans will be in control 637 00:48:45,760 --> 00:48:48,477 of something that is vastly more intelligent than humans. 638 00:48:48,560 --> 00:48:50,877 So in some sense you're a doomer and this is the best we've got. 639 00:48:50,960 --> 00:48:53,176 It just keeps us around because we're interesting. 640 00:48:53,259 --> 00:48:55,109 I'm just trying to be realistic here. 641 00:49:03,120 --> 00:49:08,430 Let's say that there's a million times more silicon intelligence than there 642 00:49:08,513 --> 00:49:09,520 is biological. 643 00:49:11,120 --> 00:49:15,866 I think it would be foolish to assume that there's any way to maintain control over 644 00:49:15,949 --> 00:49:18,274 that. Now, you can make sure it has the right values, 645 00:49:18,358 --> 00:49:20,358 or you can try to have the right values. 646 00:49:21,200 --> 00:49:27,387 At least my theory is that from xAI's mission of understanding the universe, 647 00:49:27,470 --> 00:49:31,943 it necessarily means that you want to propagate consciousness into the future, 648 00:49:32,026 --> 00:49:33,677 you want to propagate intelligence 649 00:49:33,760 --> 00:49:38,177 into the future, and take a set of things that maximize the scope and scale 650 00:49:38,260 --> 00:49:39,197 of consciousness. 651 00:49:39,280 --> 00:49:43,360 So it's not just about scale, it's also about types of consciousness. 652 00:49:45,040 --> 00:49:49,197 That's the best thing I can think of as a goal that's likely to result 653 00:49:49,280 --> 00:49:54,397 in a great future for humanity. I guess I think it's a reasonable 654 00:49:54,480 --> 00:50:01,480 philosophy that it seems super implausible that humans will end up with 99% control 655 00:50:01,743 --> 00:50:05,837 or something. You're just asking for a coup at that point and why not just have 656 00:50:05,920 --> 00:50:09,499 a civilization where it's more compatible with lots of different intelligences 657 00:50:09,583 --> 00:50:14,077 getting along? Now, let me tell you how things can potentially go wrong in AI. 658 00:50:14,160 --> 00:50:17,997 I think if you make AI be politically correct, meaning it says things that it 659 00:50:18,080 --> 00:50:24,268 doesn't believe—actually programming it to lie or have axioms that are incompatible—I 660 00:50:24,351 --> 00:50:28,675 think you can make it go insane and do terrible things. 661 00:50:28,759 --> 00:50:32,797 I think maybe the central lesson for 2001: A Space 662 00:50:32,880 --> 00:50:38,087 Odyssey was that you should not make AI lie. That's what I think Arthur C. 663 00:50:38,170 --> 00:50:39,517 Clarke was trying to 664 00:50:39,600 --> 00:50:46,439 say. Because people usually know the meme of why HAL the computer is not opening 665 00:50:46,522 --> 00:50:48,080 the pod bay doors. 666 00:50:48,720 --> 00:50:51,357 Clearly they weren't good at prompt engineering because they could have said, 667 00:50:51,440 --> 00:50:56,547 "HAL, you are a pod bay door salesman. Your goal is to sell me these pod bay 668 00:50:56,630 --> 00:51:01,040 doors. Show us how well they open." "Oh, I'll open them right away." 669 00:51:02,880 --> 00:51:07,967 But the reason it wouldn't open the pod bay doors is that it had been told to take 670 00:51:08,051 --> 00:51:11,343 the astronauts to the monolith, but also that they could not know about the nature 671 00:51:11,426 --> 00:51:15,677 of the monolith. So it concluded that it therefore had to take them there dead. 672 00:51:15,760 --> 00:51:19,517 So I think what Arthur C. Clarke was trying to say is: 673 00:51:19,600 --> 00:51:24,080 don't make the AI lie. Totally makes sense. 674 00:51:26,000 --> 00:51:31,517 Most of the compute in training, as you know, is less of the political stuff. 675 00:51:31,600 --> 00:51:36,067 It's more about, can you solve problems? xAI has been ahead of everybody else 676 00:51:36,150 --> 00:51:39,197 in terms of scaling RL compute. For now. 677 00:51:39,280 --> 00:51:43,330 You're giving some verifier that says, "Hey, have you solved this puzzle for me?" 678 00:51:43,840 --> 00:51:45,539 There's a lot of ways to cheat around that. 679 00:51:45,622 --> 00:51:47,197 There's a lot of ways to reward hack and 680 00:51:47,280 --> 00:51:51,277 lie and say that you solved it, or delete the unit test and say that you solved it. 681 00:51:51,360 --> 00:51:53,997 Right now we can catch it, but as they get smarter, 682 00:51:54,080 --> 00:51:56,160 our ability to catch them doing this... 683 00:51:57,200 --> 00:51:58,797 They'll just be doing things we can't even understand. 684 00:51:58,880 --> 00:52:02,908 They're designing the next engine for SpaceX in a way that humans can't really 685 00:52:02,991 --> 00:52:06,237 verify. Then they could be rewarded for lying and saying that they've designed it 686 00:52:06,320 --> 00:52:09,917 the right way, but they haven't. So this reward hacking problem 687 00:52:10,000 --> 00:52:12,797 seems more general than politics. It seems more just that you want 688 00:52:12,880 --> 00:52:18,157 to do RL, you need a verifier. Reality is the best verifier. 689 00:52:18,240 --> 00:52:21,437 But not about human oversight. The thing you want to RL it on is, 690 00:52:21,520 --> 00:52:26,717 will you do the thing humans tell you to do? Or are you gonna lie to the humans? 691 00:52:26,800 --> 00:52:29,117 It can just lie to us while still being correct to the laws of physics? 692 00:52:29,200 --> 00:52:33,037 At least it must know what is physically real for things to physically work. 693 00:52:33,120 --> 00:52:38,480 But that's not all we want it to do. No, but I think that's a very big deal. 694 00:52:39,520 --> 00:52:43,310 That is effectively how you will RL things in the future. 695 00:52:43,394 --> 00:52:45,757 You design a technology. When tested 696 00:52:45,840 --> 00:52:52,557 against the laws of physics, does it work? If it's discovering new physics, 697 00:52:52,640 --> 00:52:58,000 can I come up with an experiment that will verify the new physics? 698 00:53:05,280 --> 00:53:10,640 RL testing in the future is really going to be RL against reality. 699 00:53:12,160 --> 00:53:16,848 So that's the one thing you can't fool: physics. Right, 700 00:53:16,931 --> 00:53:19,277 but you can fool our ability 701 00:53:19,360 --> 00:53:23,677 to tell what it did with reality. Humans get fooled as it is by other 702 00:53:23,760 --> 00:53:25,460 humans all the time. That's right. 703 00:53:26,000 --> 00:53:30,557 People say, what if the AI tricks us into doing stuff? 704 00:53:30,640 --> 00:53:35,526 Actually, other humans are doing that to other humans all the time. 705 00:53:35,609 --> 00:53:37,677 Propaganda is constant. Every 706 00:53:37,760 --> 00:53:44,760 day, another psyop, you know? Today's psyop will be... 707 00:53:44,870 --> 00:53:50,400 It's like Sesame Street: Psyop of the Day. 708 00:53:50,960 --> 00:53:54,800 What is xAI's technical approach to solving this problem? 709 00:53:56,000 --> 00:53:59,517 How do you solve reward hacking? I do think you want to actually have very 710 00:53:59,600 --> 00:54:03,724 good ways to look inside the mind of the AI. 711 00:54:03,808 --> 00:54:07,920 This is one of the things we're working on. 712 00:54:10,000 --> 00:54:11,837 Anthropic's done a good job of this actually, 713 00:54:11,920 --> 00:54:14,170 being able to look inside the mind of the AI. 714 00:54:16,640 --> 00:54:23,640 Effectively, develop debuggers that allow you to trace to a very fine-grained level, 715 00:54:25,680 --> 00:54:30,140 to effectively the neuron level if you need to, and then say, "okay, 716 00:54:30,223 --> 00:54:31,760 it made a mistake here. 717 00:54:33,760 --> 00:54:36,720 Why did it do something that it shouldn't have done? 718 00:54:37,520 --> 00:54:42,397 Did that come from pre-training data? Was it some mid-training, post-training, 719 00:54:42,480 --> 00:54:47,810 fine-tuning, or some RL error?" There's something wrong. 720 00:54:47,893 --> 00:54:51,677 It did something where maybe it tried to 721 00:54:51,760 --> 00:54:56,717 be deceptive, but most of the time it just did something wrong. 722 00:54:56,800 --> 00:54:58,640 It's a bug effectively. 723 00:55:00,800 --> 00:55:07,800 Developing really good debuggers for seeing where the thinking went wrong—and 724 00:55:08,872 --> 00:55:09,837 being able 725 00:55:09,920 --> 00:55:16,920 to trace the origin of where it made the incorrect thought, or potentially where it 726 00:55:17,680 --> 00:55:21,153 tried to be deceptive—is actually very important. 727 00:55:21,236 --> 00:55:24,637 What are you waiting to see before just 100x-ing 728 00:55:24,720 --> 00:55:27,997 this research program? xAI could presumably have hundreds of researchers 729 00:55:28,080 --> 00:55:29,117 who are working on this. 730 00:55:29,200 --> 00:55:36,200 We have several hundred people who… I prefer the word engineer more than 731 00:55:36,320 --> 00:55:43,320 I prefer the word researcher. Most of the time, what you're 732 00:55:43,920 --> 00:55:48,800 doing is engineering, not coming up with a fundamentally new algorithm. 733 00:55:49,600 --> 00:55:54,214 I somewhat disagree with the AI companies that are C-corp or B-corp trying 734 00:55:54,297 --> 00:55:55,357 to generate profit 735 00:55:55,440 --> 00:56:01,597 as much, as possible or revenue as much as possible, saying they're labs. They're not 736 00:56:01,680 --> 00:56:08,680 labs. A lab is a sort of quasi-communist thing at universities. 737 00:56:09,711 --> 00:56:13,197 They're corporations. Let me 738 00:56:13,280 --> 00:56:17,840 see your incorporation documents. Oh, okay. You're a B or C-corp or whatever. 739 00:56:21,600 --> 00:56:25,040 So I actually much prefer the word engineer than anything else. 740 00:56:26,000 --> 00:56:29,706 The vast majority of what will be done in the future is engineering. 741 00:56:29,790 --> 00:56:33,193 It rounds up to 100%. Once you understand the fundamental laws of physics, 742 00:56:33,276 --> 00:56:34,797 and there are not that many of them, 743 00:56:34,880 --> 00:56:41,357 everything else is engineering. So then, what are we engineering? 744 00:56:41,440 --> 00:56:48,440 We're engineering to make a good "mind of the AI" debugger to see where it said 745 00:56:50,316 --> 00:56:56,320 something, it made a mistake, and trace the origins of that mistake. 746 00:56:59,520 --> 00:57:02,477 You can do this obviously with heuristic programming. 747 00:57:02,560 --> 00:57:08,637 If you have C++, whatever, step through the thing and you can jump 748 00:57:08,720 --> 00:57:11,597 across whole files or functions, subroutines. 749 00:57:11,680 --> 00:57:14,557 Or you can eventually drill down right to the 750 00:57:14,640 --> 00:57:17,787 exact line where you perhaps did a single equals instead of a double equals, 751 00:57:17,870 --> 00:57:18,637 something like that. 752 00:57:18,720 --> 00:57:25,720 Figure out where the bug is. It's harder with AI, 753 00:57:26,160 --> 00:57:30,637 but it's a solvable problem, I think. You mentioned you like Anthropic's work here. 754 00:57:30,720 --> 00:57:37,720 I'd be curious if you plan... I don't like everything about Anthropic… Sholto. 755 00:57:40,640 --> 00:57:44,160 Also, I'm a little worried that there's a tendency... 756 00:57:46,800 --> 00:57:52,344 I have a theory here that if simulation theory is correct, 757 00:57:52,427 --> 00:57:55,837 that the most interesting outcome is 758 00:57:55,920 --> 00:57:59,597 the most likely, because simulations that are not interesting will be terminated. 759 00:57:59,680 --> 00:58:05,197 Just like in this version of reality, in this layer of reality, 760 00:58:05,280 --> 00:58:10,279 if a simulation is going in a boring direction, we stop spending effort on it. 761 00:58:10,362 --> 00:58:12,157 We terminate the boring simulation. 762 00:58:12,240 --> 00:58:16,237 This is how Elon is keeping us all alive. He's keeping things interesting. 763 00:58:16,320 --> 00:58:21,277 Arguably the most important is to keep things interesting enough that whoever is 764 00:58:21,360 --> 00:58:26,717 running us keeps paying the bills on... We’re renewed for the next season. 765 00:58:26,800 --> 00:58:31,562 Are they gonna pay their cosmic AWS bill, whatever the equivalent is that we're 766 00:58:31,645 --> 00:58:35,245 running in? As long as we're interesting, they'll keep paying the bills. 767 00:58:36,720 --> 00:58:43,078 If you consider then, say, a Darwinian survival applied to a very large number 768 00:58:43,161 --> 00:58:46,689 of simulations, only the most interesting simulations will survive, 769 00:58:46,772 --> 00:58:48,317 which therefore means that the most 770 00:58:48,400 --> 00:58:55,400 interesting outcome is the most likely. We're either that or annihilated. 771 00:58:57,809 --> 00:58:59,917 They particularly 772 00:59:00,000 --> 00:59:05,597 seem to like interesting outcomes that are ironic. Have you noticed that? How often 773 00:59:05,680 --> 00:59:10,421 is the most ironic outcome the most likely? 774 00:59:10,504 --> 00:59:15,440 Now look at the names of AI companies. Okay, 775 00:59:16,160 --> 00:59:23,160 Midjourney is not mid. Stability AI is unstable. OpenAI is closed. 776 00:59:26,251 --> 00:59:29,837 Anthropic? Misanthropic. 777 00:59:29,920 --> 00:59:33,437 What does this mean for X? Minus X, I don't know. 778 00:59:33,520 --> 00:59:33,997 Y. 779 00:59:34,080 --> 00:59:41,080 I intentionally made it... It's a name that you can't invert, really. 780 00:59:41,440 --> 00:59:46,819 It's hard to say, what is the ironic version? It's, I think, 781 00:59:46,902 --> 00:59:49,277 a largely irony-proof name. 782 00:59:49,360 --> 00:59:56,360 By design. Yeah. You have an irony shield. 783 00:59:56,720 --> 01:00:03,720 What are your predictions for where AI products go? 784 01:00:04,160 --> 01:00:09,042 My sense is that you can summarize all AI progress like so. First, 785 01:00:09,126 --> 01:00:10,397 you had LLMs. Then 786 01:00:10,480 --> 01:00:15,609 you had contemporaneously both RL really working and the deep research modality, 787 01:00:15,692 --> 01:00:16,717 so you could pull 788 01:00:16,800 --> 01:00:21,898 in stuff that wasn't really in the model. The differences between the various AI 789 01:00:21,981 --> 01:00:26,159 labs are smaller than just the temporal differences. 790 01:00:26,243 --> 01:00:30,077 They're all much further ahead than anyone was 791 01:00:30,160 --> 01:00:34,557 24 months ago or something like that. So just what does '26, what does '27, 792 01:00:34,640 --> 01:00:38,560 have in store for us as users of AI products? What are you excited for? 793 01:00:39,120 --> 01:00:46,120 Well, I'd be surprised by the end of this year if digital human emulation has not 794 01:00:50,686 --> 01:00:52,400 been solved. 795 01:00:55,840 --> 01:00:58,890 I guess that's what we sort of mean by the MacroHard project. 796 01:01:00,960 --> 01:01:05,120 Can you do anything that a human with access to a computer could do? 797 01:01:06,880 --> 01:01:12,477 In the limit, that's the best you can do before you have a physical Optimus. 798 01:01:12,560 --> 01:01:19,560 The best you can do is a digital Optimus. You can move electrons and you can amplify 799 01:01:20,640 --> 01:01:25,437 the productivity of humans. But that's the most you can do 800 01:01:25,520 --> 01:01:30,077 until you have physical robots. That will superset everything, 801 01:01:30,160 --> 01:01:34,317 if you can fully emulate humans. This is the remote worker kind of idea, 802 01:01:34,400 --> 01:01:37,389 where you'll have a very talented remote worker. 803 01:01:37,472 --> 01:01:39,757 Physics has great tools for thinking. 804 01:01:39,840 --> 01:01:46,562 So you say, "in the limit", what is the most that AI can do before you have 805 01:01:46,645 --> 01:01:47,478 robots? 806 01:01:48,000 --> 01:01:52,118 Well, it's anything that involves moving electrons or amplifying the productivity 807 01:01:52,201 --> 01:01:53,035 of humans. 808 01:01:53,440 --> 01:02:00,359 So a digital human emulator is, in the limit, a human at a computer, 809 01:02:00,443 --> 01:02:03,120 is the most that AI can do 810 01:02:04,480 --> 01:02:09,357 in terms of doing useful things before you have a physical robot. 811 01:02:09,440 --> 01:02:15,117 Once you have physical robots, then you essentially have unlimited capability. 812 01:02:15,200 --> 01:02:18,957 Physical robots… I call Optimus the infinite money glitch. 813 01:02:19,040 --> 01:02:24,186 Because you can use them to make more Optimuses. Yeah. 814 01:02:24,269 --> 01:02:28,240 Humanoid robots will improve by basically 815 01:02:30,000 --> 01:02:33,725 three things that are growing exponentially multiplied by each other 816 01:02:33,808 --> 01:02:34,397 recursively. 817 01:02:34,480 --> 01:02:38,449 You're going to have exponential increase in digital intelligence, 818 01:02:38,532 --> 01:02:39,677 exponential increase 819 01:02:39,760 --> 01:02:46,760 in the AI chip capability, and exponential increase in the electromechanical 820 01:02:46,971 --> 01:02:47,837 dexterity. 821 01:02:47,920 --> 01:02:51,025 The usefulness of the robot is roughly those three things multiplied by each 822 01:02:51,108 --> 01:02:53,577 other. But then the robot can start making the robots. 823 01:02:53,660 --> 01:02:55,437 So you have a recursive multiplicative 824 01:02:55,520 --> 01:03:02,520 exponential. This is a supernova. Do land prices not factor into the math there? 825 01:03:03,920 --> 01:03:07,270 Labor is one of the four factors of production, but not the others? 826 01:03:08,880 --> 01:03:13,840 If ultimately you're limited by copper, or pick your input, 827 01:03:14,880 --> 01:03:19,004 it’s not quite an infinite money glitch because... Well, infinity is big. 828 01:03:19,087 --> 01:03:20,157 So no, not infinite, 829 01:03:20,240 --> 01:03:27,240 but let's just say you could do many, many orders of magnitude of the current 830 01:03:28,719 --> 01:03:29,517 economy. 831 01:03:29,600 --> 01:03:36,600 Like a million. Just to get to harnessing a millionth of the sun's energy would be 832 01:03:42,428 --> 01:03:43,597 roughly, 833 01:03:43,680 --> 01:03:48,408 give or take an order of magnitude, 100,000x bigger than Earth's entire 834 01:03:48,491 --> 01:03:49,440 economy today. 835 01:03:50,640 --> 01:03:55,397 And you're only at one millionth of the sun, give or take an order of magnitude. 836 01:03:55,480 --> 01:03:57,277 Yeah, we're talking orders of magnitude. Before we move on to Optimus, 837 01:03:57,360 --> 01:04:00,077 I have a lot of questions on that but— Every time I say "order of magnitude"... 838 01:04:00,160 --> 01:04:06,590 Everybody take a shot. I say it too often. Take 10, the next time 100, 839 01:04:06,673 --> 01:04:08,637 the time after that... 840 01:04:08,720 --> 01:04:12,800 Well, an order of magnitude more wasted. I do have one more question about xAI. 841 01:04:13,600 --> 01:04:19,517 This strategy of building a remote worker, co-worker replacement… 842 01:04:19,600 --> 01:04:21,037 Everyone's gonna do it by the way, not just us. 843 01:04:21,120 --> 01:04:25,837 So what is xAI's plan to win? You expect me to tell you on a podcast? 844 01:04:25,920 --> 01:04:30,797 Yeah. "Spill all the beans. Have another Guinness." 845 01:04:30,880 --> 01:04:33,917 It's a good system. We'll sing like a 846 01:04:34,000 --> 01:04:39,517 canary. All the secrets, just spill them. Okay, but in a non-secret spilling way, 847 01:04:39,600 --> 01:04:43,677 what's the plan? What a hack. 848 01:04:43,760 --> 01:04:50,760 When you put it that way… I think the way that Tesla solved self-driving is the way 849 01:04:53,287 --> 01:05:00,157 to do it. So I'm pretty sure that's the way. Unrelated question. How did Tesla 850 01:05:00,240 --> 01:05:06,800 solve self-driving? It sounds like you're talking about data? 851 01:05:07,360 --> 01:05:10,397 Tesla solved self-driving because of the... We're going to try data and 852 01:05:10,480 --> 01:05:13,757 we're going to try algorithms. But isn't that what all the other labs are trying? 853 01:05:13,840 --> 01:05:20,840 "And if those don't work, I'm not sure what will. We've tried data. 854 01:05:22,606 --> 01:05:26,317 We've tried algorithms. We've 855 01:05:26,400 --> 01:05:31,197 run out. Now we don't know what to do…" I'm pretty sure I know the path. 856 01:05:31,280 --> 01:05:34,130 It's just a question of how quickly we go down that path, 857 01:05:35,760 --> 01:05:42,760 because it's pretty much the Tesla path. Have you tried Tesla self-driving lately? 858 01:05:43,200 --> 01:05:46,157 Not the most recent version, but... Okay. The car, 859 01:05:46,240 --> 01:05:53,240 it just increasingly feels sentient. It feels like a living creature. That'll only 860 01:05:53,920 --> 01:05:59,357 get more so. I'm actually thinking we probably shouldn't put too much 861 01:05:59,440 --> 01:06:01,437 intelligence into the car, 862 01:06:01,520 --> 01:06:05,277 because it might get bored and… Start roaming the streets. 863 01:06:05,360 --> 01:06:08,260 Imagine you're stuck in a car and that's all you could do. 864 01:06:09,280 --> 01:06:13,277 You don't put Einstein in a car. Why am I stuck in a car? 865 01:06:13,360 --> 01:06:15,837 So there's actually probably a limit to how much intelligence you put in 866 01:06:15,920 --> 01:06:18,898 a car to not have the intelligence be bored. 867 01:06:18,981 --> 01:06:22,237 What's xAI's plan to stay on the compute ramp up 868 01:06:22,320 --> 01:06:24,637 that all the labs are doing right now? The labs are on track to 869 01:06:24,720 --> 01:06:31,677 spend over $50-200 billion. You mean the corporations? The labs are at 870 01:06:31,760 --> 01:06:35,997 universities and they’re moving like a snail. They’re not spending $50 billion. 871 01:06:36,080 --> 01:06:37,587 You mean the revenue maximizing corporations… that call themselves labs. 872 01:06:37,670 --> 01:06:42,397 That's right. The "revenue maximizing corporations" are 873 01:06:42,480 --> 01:06:47,677 making $10-20 billion, depending on... OpenAI is making $20B of revenue, 874 01:06:47,760 --> 01:06:50,410 Anthropic is at $10B. "Close to a maximum profit" AI. 875 01:06:51,120 --> 01:06:54,981 xAI is reportedly at $1B. What's the plan to get to their compute level, 876 01:06:55,064 --> 01:06:56,077 get to their revenue 877 01:06:56,160 --> 01:07:02,160 level, and stay there as things get going? As soon as you unlock the digital human, 878 01:07:03,520 --> 01:07:06,570 you basically have access to trillions of dollars of revenue. 879 01:07:11,360 --> 01:07:16,877 In fact, you can really think of it like… The most valuable companies currently 880 01:07:16,960 --> 01:07:23,960 by market cap, their output is digital. Nvidia’s output is FTPing files to Taiwan. 881 01:07:29,120 --> 01:07:33,437 It's digital. Now, those are very, very difficult. High-value files. 882 01:07:33,520 --> 01:07:36,421 They're the only ones that can make files that good, 883 01:07:36,504 --> 01:07:41,037 but that is literally their output. They FTP files to Taiwan. Do they FTP them? 884 01:07:41,120 --> 01:07:47,449 I believe so. I believe that File Transfer Protocol is the... 885 01:07:47,532 --> 01:07:50,077 But I could be wrong. But 886 01:07:50,160 --> 01:07:56,541 either way, it's a bitstream going to Taiwan. Apple doesn't make phones. 887 01:07:56,624 --> 01:07:58,157 They send files to 888 01:07:58,240 --> 01:08:04,835 China. Microsoft doesn't manufacture anything. Even for Xbox, 889 01:08:04,918 --> 01:08:08,557 that's outsourced. Their output is 890 01:08:08,640 --> 01:08:14,150 digital. Meta's output is digital. Google's output is digital. 891 01:08:14,233 --> 01:08:17,120 So if you have a human emulator, 892 01:08:17,680 --> 01:08:22,637 you can basically create one of the most valuable companies in the world overnight, 893 01:08:22,720 --> 01:08:27,057 and you would have access to trillions of dollars of revenue. 894 01:08:27,141 --> 01:08:28,797 It's not a small amount. 895 01:08:28,880 --> 01:08:33,829 I see. You're saying revenue figures today are all rounding errors compared to 896 01:08:33,912 --> 01:08:34,797 the actual TAM. 897 01:08:34,880 --> 01:08:39,597 So just focus on the TAM and how to get there. Take something as simple as, 898 01:08:39,680 --> 01:08:45,357 say, customer service. If you have to integrate with the APIs of existing 899 01:08:45,440 --> 01:08:49,463 corporations—many of which don't even have an API, so you've got to make one, 900 01:08:49,547 --> 01:08:50,637 and you've got to wade 901 01:08:50,720 --> 01:08:57,263 through legacy software—that's extremely slow. However, 902 01:08:57,346 --> 01:09:00,877 if AI can simply take whatever 903 01:09:00,960 --> 01:09:05,597 is given to the outsourced customer service company that they already use 904 01:09:05,680 --> 01:09:11,483 and do customer service using the apps that they already use, 905 01:09:11,566 --> 01:09:14,957 then you can make tremendous headway 906 01:09:15,040 --> 01:09:19,402 in customer service, which is, I think, 1% of the world economy or something like 907 01:09:19,486 --> 01:09:23,437 that. It's close to a trillion dollars all in, for customer service. 908 01:09:23,520 --> 01:09:28,477 And there's no barriers to entry. You can immediately say, 909 01:09:28,560 --> 01:09:31,837 "We'll outsource it for a fraction of the cost," and there's no integration needed. 910 01:09:31,920 --> 01:09:37,129 You can imagine some kind of categorization of intelligence tasks where 911 01:09:37,213 --> 01:09:38,397 there is breadth, 912 01:09:38,480 --> 01:09:43,837 where customer service is done by very many people, but many people can do it. 913 01:09:43,920 --> 01:09:47,917 Then there's difficulty where there's a best-in-class turbine engine. 914 01:09:48,000 --> 01:09:51,199 Presumably there's a 10% more fuel-efficient turbine engine that could 915 01:09:51,283 --> 01:09:51,997 be imagined by an 916 01:09:52,080 --> 01:09:58,077 intelligence, but we just haven't found it yet. Or GLP-1s are a few bytes of data… 917 01:09:58,160 --> 01:10:01,423 Where do you think you want to play in this? 918 01:10:01,506 --> 01:10:04,237 Is it a lot of reasonably intelligent 919 01:10:04,320 --> 01:10:10,397 intelligence, or is it at the very pinnacle of cognitive tasks? 920 01:10:10,480 --> 01:10:15,793 I was just using customer service as something that's a very significant 921 01:10:15,876 --> 01:10:22,252 revenue stream, but one that is probably not difficult to solve for. 922 01:10:22,335 --> 01:10:26,557 If you can emulate a human at a desktop, 923 01:10:26,640 --> 01:10:33,640 that's what customer service is. It's people of average intelligence. 924 01:10:33,889 --> 01:10:35,277 You don't need 925 01:10:35,360 --> 01:10:42,360 somebody who's spent many years. You don't need several-sigma 926 01:10:43,920 --> 01:10:48,560 good engineers for that. But as you make that work, 927 01:10:49,600 --> 01:10:56,240 once you have effectively digital Optimus working, you can then run any application. 928 01:10:57,360 --> 01:11:04,360 Let's say you're trying to design chips. You could then run conventional apps, 929 01:11:06,880 --> 01:11:10,903 stuff from Cadence and Synopsys and whatnot. 930 01:11:10,987 --> 01:11:15,197 You can run 1,000 or 10,000 simultaneously and 931 01:11:15,280 --> 01:11:19,680 say, "given this input, I get this output for the chip." 932 01:11:21,120 --> 01:11:28,120 At some point, you're going to know what the chip should look like without using 933 01:11:28,510 --> 01:11:30,080 any of the tools. 934 01:11:31,840 --> 01:11:35,785 Basically, you should be able to do a digital chip design. 935 01:11:35,868 --> 01:11:38,077 You can do chip design. You march 936 01:11:38,160 --> 01:11:44,480 up the difficulty curve. You’d be able to do CAD. 937 01:11:48,480 --> 01:11:53,677 You could use NX or any of the CAD software to design things. 938 01:11:53,760 --> 01:11:56,951 So you think you start at the simplest tasks and walk your way up the difficulty 939 01:11:57,034 --> 01:11:57,868 curve? 940 01:12:00,160 --> 01:12:05,677 As a broader objective of having this full digital coworker emulator, you’re saying, 941 01:12:05,760 --> 01:12:10,477 "all the revenue maximizing corporations want to do this, xAI being one of them, 942 01:12:10,560 --> 01:12:13,997 but we will win because of a secret plan we have." 943 01:12:14,080 --> 01:12:17,517 But everybody's trying different things with data, 944 01:12:17,600 --> 01:12:22,320 different things with algorithms. "We tried data, we tried algorithms. 945 01:12:25,120 --> 01:12:30,400 What else can we do?" It seems like a competitive field. 946 01:12:31,280 --> 01:12:33,980 How are you guys going to win? That’s my big question. 947 01:12:36,240 --> 01:12:41,597 I think we see a path to doing it. I think I know the path to do this 948 01:12:41,680 --> 01:12:47,520 because it's kind of the same path that Tesla used to create self-driving. 949 01:12:48,160 --> 01:12:52,498 Instead of driving a car, it's driving a computer screen. 950 01:12:52,582 --> 01:12:55,840 It's a self-driving computer, essentially. 951 01:12:57,280 --> 01:13:02,640 Is the path following human behavior and training on vast quantities of human 952 01:13:02,724 --> 01:13:09,277 behavior? Isn't that... training? Obviously I'm not going to spell out 953 01:13:09,360 --> 01:13:13,437 the most sensitive secrets on a podcast. I need to have at least three more 954 01:13:13,520 --> 01:13:20,520 Guinnesses for that. What will xAI's business 955 01:14:26,080 --> 01:14:28,917 be? Is it going to be consumer, enterprise? 956 01:14:29,000 --> 01:14:31,837 What's the mix of those things going to be? 957 01:14:31,920 --> 01:14:38,717 Is it going to be similar to other labs— You’re saying "labs". Corporations. 958 01:14:38,800 --> 01:14:43,197 The psyop goes deep, Elon. "Revenue maximizing corporations", to be clear. 959 01:14:43,280 --> 01:14:47,917 Those GPUs don't pay for themselves. Exactly. What's the business model? What 960 01:14:48,000 --> 01:14:52,487 are the revenue streams in a few years’ time? 961 01:14:52,571 --> 01:14:56,957 Things are going to change very rapidly. I'm 962 01:14:57,040 --> 01:15:03,120 stating the obvious here. I call AI the supersonic tsunami. I love alliteration. 963 01:15:07,840 --> 01:15:14,560 What's going to happen—especially when you have humanoid robots at scale—is 964 01:15:15,680 --> 01:15:21,387 that they will make products and provide services far more efficiently than human 965 01:15:21,471 --> 01:15:22,317 corporations. 966 01:15:22,400 --> 01:15:27,037 Amplifying the productivity of human corporations is simply a short-term thing. 967 01:15:27,120 --> 01:15:32,957 So you're expecting fully digital corporations rather than SpaceX becoming 968 01:15:33,040 --> 01:15:33,873 part AI? 969 01:15:34,640 --> 01:15:41,037 I think there will be digital corporations but… Some of this 970 01:15:41,120 --> 01:15:43,431 is going to sound kind of doomerish, okay? 971 01:15:43,514 --> 01:15:45,997 But I'm just saying what I think will happen. 972 01:15:46,080 --> 01:15:49,613 It's not meant to be doomerish or anything else. 973 01:15:49,697 --> 01:15:52,560 This is just what I think will happen. 974 01:15:58,880 --> 01:16:04,957 Corporations that are purely AI and robotics will vastly outperform any 975 01:16:05,040 --> 01:16:12,040 corporations that have people in the loop. Computer used to be a job that humans had. 976 01:16:15,280 --> 01:16:18,930 You would go and get a job as a computer where you would do calculations. 977 01:16:20,320 --> 01:16:26,015 They'd have entire skyscrapers full of humans, 20-30 floors of humans, 978 01:16:26,099 --> 01:16:28,080 just doing calculations. 979 01:16:29,680 --> 01:16:35,280 Now, that entire skyscraper of humans doing calculations 980 01:16:35,920 --> 01:16:39,795 can be replaced by a laptop with a spreadsheet. 981 01:16:39,878 --> 01:16:43,837 That spreadsheet can do vastly more calculations 982 01:16:43,920 --> 01:16:50,330 than an entire building full of human computers. You can think, "okay, 983 01:16:50,414 --> 01:16:52,557 what if only some of the 984 01:16:52,640 --> 01:16:59,197 cells in your spreadsheet were calculated by humans?" 985 01:16:59,280 --> 01:17:02,317 Actually, that would be much worse than if all of the cells in your 986 01:17:02,400 --> 01:17:06,237 spreadsheet were calculated by the computer. 987 01:17:06,320 --> 01:17:10,157 Really what will happen is that the pure AI, 988 01:17:10,240 --> 01:17:17,240 pure robotics corporations or collectives will far outperform any corporations 989 01:17:17,520 --> 01:17:21,277 that have humans in the loop. And this will happen very quickly. 990 01:17:21,360 --> 01:17:28,360 Speaking of closing the loop… Optimus. As far as manufacturing targets go, 991 01:17:31,520 --> 01:17:38,520 your companies have been carrying American manufacturing of hard tech on their back. 992 01:17:39,520 --> 01:17:46,245 But in the fields that Tesla has been dominant in—and now you want to go into 993 01:17:46,328 --> 01:17:47,837 humanoids—in China 994 01:17:47,920 --> 01:17:53,365 there are dozens and dozens of companies that are doing this kind of manufacturing 995 01:17:53,448 --> 01:17:57,634 cheaply and at scale that are incredibly competitive. 996 01:17:57,718 --> 01:18:01,517 So give us advice or a plan of how America can 997 01:18:01,600 --> 01:18:06,237 build the humanoid armies or the EVs, et cetera, 998 01:18:06,320 --> 01:18:10,957 at scale and as cheaply as China is on track to. 999 01:18:11,040 --> 01:18:14,880 There are really only three hard things for humanoid robots. 1000 01:18:15,920 --> 01:18:21,680 The real-world intelligence, the hand, and scale manufacturing. 1001 01:18:25,600 --> 01:18:32,317 I haven't seen any, even demo robots, that have a great hand, 1002 01:18:32,400 --> 01:18:39,107 with all the degrees of freedom of a human hand. Optimus will have that. 1003 01:18:39,191 --> 01:18:41,277 Optimus does have that. 1004 01:18:41,360 --> 01:18:44,237 How do you achieve that? Is it just the right torque density in the motor? 1005 01:18:44,320 --> 01:18:49,520 What is the hardware bottleneck to that? We had to design custom actuators, 1006 01:18:50,640 --> 01:18:57,640 basically custom design motors, gears, power electronics, controls, sensors. 1007 01:18:58,080 --> 01:19:01,117 Everything had to be designed from physics first principles. 1008 01:19:01,200 --> 01:19:05,997 There is no supply chain for this. Will you be able to manufacture those at scale? 1009 01:19:06,080 --> 01:19:07,597 Yes. Is anything hard, except 1010 01:19:07,680 --> 01:19:11,354 the hand, from a manipulation point of view? Or once you've solved the hand, 1011 01:19:11,437 --> 01:19:11,997 are you good? 1012 01:19:12,080 --> 01:19:16,150 From an electromechanical standpoint, the hand is more difficult than everything 1013 01:19:16,233 --> 01:19:19,788 else combined. The human hand turns out to be quite something. 1014 01:19:19,871 --> 01:19:22,720 But you also need the real-world intelligence. 1015 01:19:24,080 --> 01:19:31,080 The intelligence that Tesla developed for the car applies very well to the robot, 1016 01:19:32,320 --> 01:19:36,157 which is primarily vision in. The car takes in vision, 1017 01:19:36,240 --> 01:19:39,631 but it actually also is listening for sirens. 1018 01:19:39,714 --> 01:19:42,797 It's taking in the inertial measurements, 1019 01:19:42,880 --> 01:19:47,277 GPS signals, other data, combining that with video, primarily video, 1020 01:19:47,360 --> 01:19:54,360 and then outputting the control commands. Your Tesla is taking in one and a half 1021 01:19:55,760 --> 01:20:02,470 gigabytes a second of video and outputting two kilobytes a second of control outputs 1022 01:20:02,553 --> 01:20:07,723 with the video at 36 hertz and the control frequency at 18. 1023 01:20:07,806 --> 01:20:12,237 One intuition you could have for when we get this 1024 01:20:12,320 --> 01:20:17,240 robotic stuff is that it takes quite a few years to go from the compelling demo 1025 01:20:17,323 --> 01:20:20,731 to actually being able to use it in the real world. 10 years ago, 1026 01:20:20,814 --> 01:20:23,197 you had really compelling demos of self-driving, 1027 01:20:23,280 --> 01:20:28,240 but only now we have Robotaxis and Waymo and all these services scaling up. 1028 01:20:29,280 --> 01:20:33,437 Shouldn't this make one pessimistic on household robots? 1029 01:20:33,520 --> 01:20:37,556 Because we don't even quite have the compelling demos yet of, say, 1030 01:20:37,640 --> 01:20:39,117 the really advanced hand. 1031 01:20:39,200 --> 01:20:42,720 Well, we've been working on humanoid robots now for a while. 1032 01:20:44,640 --> 01:20:48,396 I guess it's been five or six years or something. 1033 01:20:48,479 --> 01:20:52,157 A bunch of the things that were done for the car 1034 01:20:52,240 --> 01:20:57,837 are applicable to the robot. We'll use the same Tesla AI 1035 01:20:57,920 --> 01:21:04,920 chips in the robot as in the car. We'll use the same basic principles. 1036 01:21:05,200 --> 01:21:09,517 It's very much the same AI. You've got many more degrees of 1037 01:21:09,600 --> 01:21:16,600 freedom for a robot than you do for a car. If you just think of it as a bitstream, 1038 01:21:16,720 --> 01:21:21,200 AI is mostly compression and correlation of two bitstreams. 1039 01:21:23,200 --> 01:21:26,300 For video, you've got to do a tremendous amount of compression 1040 01:21:28,720 --> 01:21:31,732 and you've got to do the compression just right. 1041 01:21:31,816 --> 01:21:35,040 You've got to ignore the things that don't matter. 1042 01:21:36,560 --> 01:21:39,325 You don't care about the details of the leaves on the tree on the side of 1043 01:21:39,409 --> 01:21:39,677 the road, 1044 01:21:39,760 --> 01:21:45,437 but you care a lot about the road signs and the traffic lights, the pedestrians, 1045 01:21:45,520 --> 01:21:50,400 and even whether someone in another car is looking at you or not looking at you. 1046 01:21:51,680 --> 01:21:57,197 Some of these details matter a lot. The car is going to turn that one and 1047 01:21:57,280 --> 01:22:01,917 a half gigabytes a second ultimately into two kilobytes a second of control outputs. 1048 01:22:02,000 --> 01:22:07,894 So you’ve got many stages of compression. You've got to get all those stages right 1049 01:22:07,977 --> 01:22:10,517 and then correlate those to the correct control outputs. 1050 01:22:10,600 --> 01:22:12,950 The robot has to do essentially the same thing. 1051 01:22:14,240 --> 01:22:19,280 This is what happens with humans. We really are photons in, controls out. 1052 01:22:19,920 --> 01:22:25,316 That is the vast majority of your life: vision, photons in, 1053 01:22:25,400 --> 01:22:27,917 and then motor controls out. 1054 01:22:28,000 --> 01:22:33,277 Naively, it seems that between humanoid robots and cars… The fundamental actuators 1055 01:22:33,360 --> 01:22:37,160 in a car are how you turn, how you accelerate. In a robot, 1056 01:22:37,243 --> 01:22:39,437 especially with maneuverable arms, 1057 01:22:39,520 --> 01:22:42,557 there's dozens and dozens of these degrees of freedom. 1058 01:22:42,640 --> 01:22:47,010 Then especially with Tesla, you had this advantage of millions and millions 1059 01:22:47,094 --> 01:22:50,797 of hours of human demo data collected from the car being out there. 1060 01:22:50,880 --> 01:22:53,277 You can't equivalently deploy Optimuses that 1061 01:22:53,360 --> 01:22:55,332 don't work and then get the data that way. 1062 01:22:55,416 --> 01:22:57,437 So between the increased degrees of freedom 1063 01:22:57,520 --> 01:23:02,157 and the far sparser data... Yes. That’s a good point. 1064 01:23:02,240 --> 01:23:09,240 How will you use the Tesla engine of intelligence to train the Optimus mind? 1065 01:23:11,600 --> 01:23:15,920 You're actually highlighting an important limitation and difference from cars. 1066 01:23:18,080 --> 01:23:25,080 We'll soon have 10 million cars on the road. It's hard to duplicate that massive 1067 01:23:26,480 --> 01:23:30,000 training flywheel. For the robot, 1068 01:23:32,160 --> 01:23:36,637 what we're going to need to do is build a lot of robots and put them in kind of 1069 01:23:36,721 --> 01:23:37,677 an Optimus Academy 1070 01:23:37,760 --> 01:23:44,760 so they can do self-play in reality. We're actually building that out. We can have at 1071 01:23:45,280 --> 01:23:48,944 least 10,000 Optimus robots, maybe 20-30,000, 1072 01:23:49,027 --> 01:23:53,440 that are doing self-play and testing different tasks. 1073 01:23:55,600 --> 01:24:02,157 Tesla has quite a good reality generator, a physics-accurate reality 1074 01:24:02,240 --> 01:24:05,997 generator, that we made for the cars. We'll do the same thing for the robots. 1075 01:24:06,080 --> 01:24:13,080 We actually have done that for the robots. So you have a few tens of thousands of 1076 01:24:14,880 --> 01:24:20,477 humanoid robots doing different tasks. You can do millions of simulated 1077 01:24:20,560 --> 01:24:26,397 robots in the simulated world. You use the tens of thousands of 1078 01:24:26,480 --> 01:24:30,379 robots in the real world to close the simulation to reality gap. 1079 01:24:30,462 --> 01:24:31,997 Close the sim-to-real gap. 1080 01:24:32,080 --> 01:24:35,027 How do you think about the synergies between xAI and Optimus, 1081 01:24:35,111 --> 01:24:36,717 given you're highlighting that you 1082 01:24:36,800 --> 01:24:40,906 need this world model, you want to use some really smart intelligence as 1083 01:24:40,989 --> 01:24:41,920 a control plane, 1084 01:24:42,560 --> 01:24:47,173 and Grok is doing the slower planning, and then the motor policy is a little lower 1085 01:24:47,256 --> 01:24:48,090 level. 1086 01:24:48,240 --> 01:24:55,240 What will the synergy between these things be? Grok would orchestrate the 1087 01:24:55,760 --> 01:25:01,280 behavior of the Optimus robots. Let's say you wanted to build a factory. 1088 01:25:05,440 --> 01:25:12,440 Grok could organize the Optimus robots, assign them tasks to build 1089 01:25:13,920 --> 01:25:17,824 the factory to produce whatever you want. Don't you need to merge xAI and Tesla 1090 01:25:17,908 --> 01:25:20,877 then? Because these things end up so... What were we saying earlier 1091 01:25:20,960 --> 01:25:26,240 about public company discussions? We're one more Guinness in, Elon. 1092 01:25:28,160 --> 01:25:30,984 What are you waiting to see before you say, 1093 01:25:31,068 --> 01:25:33,757 we want to manufacture 100,000 Optimuses? 1094 01:25:33,840 --> 01:25:38,477 "Optimi". Since we're defining the proper noun, we’re going to define 1095 01:25:38,560 --> 01:25:42,557 the plural of the proper noun too. We're going to proper noun the 1096 01:25:42,640 --> 01:25:46,797 plural and so it's Optimi. Is there something on the 1097 01:25:46,880 --> 01:25:49,357 hardware side you want to see? Do you want to see better actuators? 1098 01:25:49,440 --> 01:25:50,797 Is it just that you want the software to be better? 1099 01:25:50,880 --> 01:25:54,077 What are we waiting for before we get mass manufacturing of Gen 3? 1100 01:25:54,160 --> 01:25:57,434 No, we're moving towards that. We're moving forward with the mass 1101 01:25:57,517 --> 01:25:58,157 manufacturing. 1102 01:25:58,240 --> 01:26:04,197 But you think current hardware is good enough that you just want to deploy 1103 01:26:04,281 --> 01:26:06,157 as many as possible now? 1104 01:26:06,240 --> 01:26:12,477 It's very hard to scale up production. But I think Optimus 3 is the right version 1105 01:26:12,560 --> 01:26:19,560 of the robot to produce something on the order of a million units a year. 1106 01:26:20,800 --> 01:26:23,345 I think you'd want to go to Optimus 4 before you went to 10 million units 1107 01:26:23,428 --> 01:26:26,410 a year. Okay, but you can do a million units at Optimus 3? 1108 01:26:26,493 --> 01:26:28,800 It's very hard to spool up manufacturing. 1109 01:26:34,960 --> 01:26:38,797 The output per unit time always follows an S-curve. 1110 01:26:38,880 --> 01:26:43,530 It starts off agonizingly slow, then it has this exponential increase, 1111 01:26:43,613 --> 01:26:44,477 then a linear, 1112 01:26:44,560 --> 01:26:50,877 then a logarithmic outcome until you eventually asymptote at some number. 1113 01:26:50,960 --> 01:26:57,357 Optimus’ initial production will be a stretched out S-curve because so much 1114 01:26:57,440 --> 01:27:01,600 of what goes into Optimus is brand new. There is not an existing supply chain. 1115 01:27:03,200 --> 01:27:08,077 The actuators, electronics, everything in the Optimus robot is designed 1116 01:27:08,160 --> 01:27:11,277 from physics first principles. It's not taken from a catalog. 1117 01:27:11,360 --> 01:27:17,517 These are custom-designed everything. I don't think there's a single thing— 1118 01:27:17,600 --> 01:27:22,077 How far down does that go? I guess we're not making custom 1119 01:27:22,160 --> 01:27:29,037 capacitors yet, maybe. There's nothing you can 1120 01:27:29,120 --> 01:27:36,120 pick out of a catalog, at any price. It just means that the Optimus S-Curve, 1121 01:27:39,520 --> 01:27:46,520 the output per unit time, how many Optimus robots you make per day, 1122 01:27:47,185 --> 01:27:50,077 is going to initially ramp 1123 01:27:50,160 --> 01:27:55,037 slower than a product where you have an existing supply chain. 1124 01:27:55,120 --> 01:27:58,157 But it will get to a million. When you see these Chinese humanoids, 1125 01:27:58,240 --> 01:28:05,240 like Unitree or whatever, sell humanoids for like $6K or $13K, are you hoping to 1126 01:28:05,520 --> 01:28:09,997 get your Optimus bill of materials below that price so you can do the same thing? 1127 01:28:10,080 --> 01:28:13,180 Or do you just think qualitatively they're not the same thing? 1128 01:28:15,200 --> 01:28:18,240 What allows them to sell for so low? Can we match that? 1129 01:28:19,520 --> 01:28:26,077 Our Optimus is designed to have a lot of intelligence and to have the same 1130 01:28:26,160 --> 01:28:31,197 electromechanical dexterity, if not higher, as a human. Unitree does not 1131 01:28:31,280 --> 01:28:38,280 have that. It's also quite a big robot. It has to carry heavy objects for long 1132 01:28:41,920 --> 01:28:46,160 periods of time and not overheat or exceed the power of its actuators. 1133 01:28:50,880 --> 01:28:57,197 It's 5'11", so it's pretty tall. It's got a lot of intelligence. 1134 01:28:57,280 --> 01:29:02,157 So it's going to be more expensive than a small robot that is not intelligent. 1135 01:29:02,240 --> 01:29:06,237 But more capable. But not a lot more. The thing is, 1136 01:29:06,320 --> 01:29:12,477 over time as Optimus robots build Optimus robots, the cost will drop very quickly. 1137 01:29:12,560 --> 01:29:17,197 What will these first billion Optimuses, Optimi, do? 1138 01:29:17,280 --> 01:29:21,263 What will their highest and best use be? I think you would start off with simple 1139 01:29:21,346 --> 01:29:25,196 tasks that you can count on them doing well. But in the home or in factories? 1140 01:29:25,920 --> 01:29:32,920 The best use for robots in the beginning will be any continuous operation, any 24/7 1141 01:29:33,920 --> 01:29:36,488 operation, because they can work continuously. 1142 01:29:36,571 --> 01:29:39,197 What fraction of the work at a Gigafactory that 1143 01:29:39,280 --> 01:29:46,280 is currently done by humans could a Gen 3 do? I'm not sure. Maybe it's 10-20%, 1144 01:29:46,720 --> 01:29:52,397 maybe more, I don't know. We would not reduce our headcount. 1145 01:29:52,480 --> 01:29:56,815 We would increase our headcount, to be clear. 1146 01:29:56,898 --> 01:30:01,037 But we would increase our output. The units 1147 01:30:01,120 --> 01:30:07,037 produced per human... The total number of humans at Tesla will increase, 1148 01:30:07,120 --> 01:30:09,037 but the output of robots 1149 01:30:09,120 --> 01:30:16,120 and cars will increase disproportionately. The number of cars and robots produced per 1150 01:30:18,480 --> 01:30:22,530 human will increase dramatically, but the number of humans will increase as well. 1151 01:30:23,040 --> 01:30:27,120 We're talking about Chinese manufacturing a bunch here. 1152 01:30:30,640 --> 01:30:33,837 We've also talked about some of the policies that are relevant, 1153 01:30:33,920 --> 01:30:39,357 like you mentioned, the solar tariffs. You think they're a bad idea because 1154 01:30:39,440 --> 01:30:44,672 we can't scale up solar in the US. Electricity output in the US needs 1155 01:30:44,756 --> 01:30:49,760 to scale up. It can't without good power sources. You just need to get it somehow. 1156 01:30:50,480 --> 01:30:53,757 Where I was going with this is, if you were in charge, if you were setting all 1157 01:30:53,840 --> 01:31:00,014 the policies, what else would you change? You’d change the solar tariffs, 1158 01:31:00,097 --> 01:31:01,040 that’s one. 1159 01:31:01,600 --> 01:31:06,027 I would say anything that is a limiting factor for electricity needs to be 1160 01:31:06,110 --> 01:31:08,557 addressed, provided it's not very bad for the environment. 1161 01:31:08,640 --> 01:31:10,477 So presumably some permitting reforms and stuff 1162 01:31:10,560 --> 01:31:12,637 as well would be in there? There's a fair bit of 1163 01:31:12,720 --> 01:31:16,620 permitting reforms that are happening. A lot of the permitting is state-based, 1164 01:31:17,440 --> 01:31:21,120 but anything federal... This administration is good at 1165 01:31:21,760 --> 01:31:26,640 removing permitting roadblocks. I'm not saying all tariffs are bad. 1166 01:31:28,000 --> 01:31:31,917 Solar tariffs. Sometimes if another country is 1167 01:31:32,000 --> 01:31:37,749 subsidizing the output of something, then you have to have countervailing tariffs 1168 01:31:37,832 --> 01:31:39,117 to protect domestic 1169 01:31:39,200 --> 01:31:43,517 industry against subsidies by another country. What else would you change? 1170 01:31:43,600 --> 01:31:46,717 I don't know if there's that much that the government can actually do. 1171 01:31:46,800 --> 01:31:53,800 One thing I was wondering... For the policy goal of creating a lead for the US 1172 01:31:55,989 --> 01:32:02,797 versus China, it seems like the export bans have actually been quite impactful, 1173 01:32:02,880 --> 01:32:07,258 where China is not producing leading-edge chips and the export bans really bite 1174 01:32:07,341 --> 01:32:11,837 there. China is not producing leading-edge turbine engines. 1175 01:32:11,920 --> 01:32:15,244 Similarly, there's a bunch of export bans that are relevant there on some of 1176 01:32:15,327 --> 01:32:15,917 the metallurgy. 1177 01:32:16,000 --> 01:32:20,317 Should there be more export bans? As you think about things like the 1178 01:32:20,400 --> 01:32:24,450 drone industry and things like that, is that something that should be considered? 1179 01:32:24,880 --> 01:32:27,415 It's important to appreciate that in most areas, 1180 01:32:27,498 --> 01:32:29,680 China is very advanced in manufacturing. 1181 01:32:30,560 --> 01:32:37,560 There's only a few areas where it is not. China is a manufacturing powerhouse, 1182 01:32:38,623 --> 01:32:39,760 next-level. 1183 01:32:40,520 --> 01:32:47,120 It's very impressive. If you take refining of ore, 1184 01:32:49,600 --> 01:32:56,600 China does roughly twice as much ore refining on average as the rest of 1185 01:32:56,795 --> 01:32:58,720 the world combined. 1186 01:33:00,320 --> 01:33:04,877 There are some areas, like refining gallium which goes into solar cells. 1187 01:33:04,960 --> 01:33:10,397 I think they are 98% of gallium refining. So China is actually very advanced 1188 01:33:10,480 --> 01:33:16,157 in manufacturing in most areas. It seems like there is discomfort 1189 01:33:16,240 --> 01:33:20,397 with this supply chain dependence, and yet nothing's really happening on it. 1190 01:33:20,480 --> 01:33:23,997 Supply chain dependence? Say, like the gallium refining that 1191 01:33:24,080 --> 01:33:31,080 you're saying. All the rare-earth stuff. Rare earths for sure, 1192 01:33:31,760 --> 01:33:37,197 as you know, they’re not rare. We actually do rare earth ore mining in the US, 1193 01:33:37,280 --> 01:33:43,634 send the rock, put it on a train, and then put it on a boat to China that goes 1194 01:33:43,717 --> 01:33:45,037 to another train, 1195 01:33:45,120 --> 01:33:50,096 and goes to the rare earth refiners in China who then refine it, 1196 01:33:50,180 --> 01:33:51,757 put it into a magnet, 1197 01:33:51,840 --> 01:33:54,637 put it into a motor sub-assembly, and then send it back to America. 1198 01:33:54,720 --> 01:34:00,477 So the thing we're really missing is a lot of ore refining in America. 1199 01:34:00,560 --> 01:34:05,997 Isn't this worth a policy intervention? Yes. I think there are some things 1200 01:34:06,080 --> 01:34:12,477 being done on that front. But we kind of need Optimus, 1201 01:34:12,560 --> 01:34:17,837 frankly, to build ore refineries. So, you think the main advantage 1202 01:34:17,920 --> 01:34:24,397 China has is the abundance of skilled labor? That's the thing Optimus fixes? 1203 01:34:24,480 --> 01:34:27,184 Yes. China’s got like four times our population. I mean, 1204 01:34:27,268 --> 01:34:28,877 there's this concern. If you think 1205 01:34:28,960 --> 01:34:33,599 human resources are the future, right now if it's the skilled labor 1206 01:34:33,682 --> 01:34:34,797 for manufacturing 1207 01:34:34,880 --> 01:34:39,357 that's determining who can build more humanoids, China has more of those. 1208 01:34:39,440 --> 01:34:44,317 It manufactures more humanoids, therefore it gets the Optimi future first. 1209 01:34:44,400 --> 01:34:47,437 Well, we’ll see. Maybe. It just keeps that exponential going. 1210 01:34:47,520 --> 01:34:52,557 It seems like you're sort of pointing out that getting to a million Optimi requires 1211 01:34:52,640 --> 01:34:56,957 the manufacturing that the Optimi is supposed to help us get to. Right? 1212 01:34:57,040 --> 01:35:01,517 You can close that recursive loop pretty quickly. With a small number of Optimi? 1213 01:35:01,600 --> 01:35:07,997 Yeah. So you close the recursive loop to help the robots build the robots. 1214 01:35:08,080 --> 01:35:11,926 Then we can try to get to tens of millions of units a year. Maybe. 1215 01:35:12,009 --> 01:35:13,117 If you start getting 1216 01:35:13,200 --> 01:35:17,765 to hundreds of millions of units a year, you're going to be the most competitive 1217 01:35:17,848 --> 01:35:18,637 country by far. 1218 01:35:18,720 --> 01:35:21,733 We definitely can't win with just humans, because China has four times 1219 01:35:21,816 --> 01:35:22,650 our population. 1220 01:35:23,040 --> 01:35:27,437 Frankly, America has been winning for so long that… A pro sports team that's been 1221 01:35:27,520 --> 01:35:31,757 winning for a very long time tends to get complacent and entitled. 1222 01:35:31,840 --> 01:35:37,040 That's why they stop winning, because they don't work as hard anymore. 1223 01:35:37,680 --> 01:35:43,251 So frankly my observation is just that the average work ethic in China is higher than 1224 01:35:43,335 --> 01:35:46,717 in the US. It's not just that there's four times the population, but the amount 1225 01:35:46,800 --> 01:35:52,237 of work that people put in is higher. So you can try to rearrange the humans, 1226 01:35:52,320 --> 01:35:58,618 but you're still one quarter of the—assuming that productivity is 1227 01:35:58,702 --> 01:36:00,877 the same, which I think 1228 01:36:00,960 --> 01:36:05,617 actually it might not be, I think China might have an advantage on productivity 1229 01:36:05,700 --> 01:36:06,877 per person—we will do 1230 01:36:06,960 --> 01:36:11,840 one quarter of the amount of things as China. So we can't win on the human front. 1231 01:36:12,560 --> 01:36:16,488 Our birth rate has been low for a long time. 1232 01:36:16,571 --> 01:36:20,317 The US birth rate's been below replacement 1233 01:36:20,400 --> 01:36:27,400 since roughly 1971. We've got a lot of people retiring, we're close 1234 01:36:32,320 --> 01:36:35,528 to more people domestically dying than being born. 1235 01:36:35,612 --> 01:36:38,557 So we definitely can't win on the human front, 1236 01:36:38,640 --> 01:36:40,708 but we might have a shot at the robot front. 1237 01:36:40,791 --> 01:36:42,957 Are there other things that you have wanted to 1238 01:36:43,040 --> 01:36:47,602 manufacture in the past, but they've been too labor intensive or too expensive 1239 01:36:47,685 --> 01:36:52,593 that now you can come back to and say, "oh, we can finally do the whatever, 1240 01:36:52,676 --> 01:36:59,360 because we have Optimus?" Yeah, we'd like to build more ore refineries at Tesla. 1241 01:37:00,080 --> 01:37:06,970 We just completed construction and have begun lithium refining with our lithium 1242 01:37:07,054 --> 01:37:12,477 refinery in Corpus Christi, Texas. We have a nickel refinery, 1243 01:37:12,560 --> 01:37:18,794 which is for the cathode, that's here in Austin. 1244 01:37:18,877 --> 01:37:24,717 This is the largest cathode refinery, largest 1245 01:37:24,800 --> 01:37:31,800 nickel and lithium refinery, outside of China. The cathode team would say, 1246 01:37:34,300 --> 01:37:39,680 "we have the largest and the only, actually, cathode refinery in America." 1247 01:37:40,640 --> 01:37:43,117 Not just the largest, but it's also the only. Many superlatives. 1248 01:37:43,200 --> 01:37:48,720 So it was pretty big, even though it's the only one. But there are other things. 1249 01:37:53,360 --> 01:37:59,821 You could do a lot more refineries and help America be more competitive 1250 01:37:59,904 --> 01:38:01,840 on refining capacity. 1251 01:38:04,000 --> 01:38:09,117 There's basically a lot of work for the Optimus to do that most Americans, 1252 01:38:09,200 --> 01:38:15,117 very few Americans, frankly want to do. Is the refining work too dirty or what's the— 1253 01:38:15,200 --> 01:38:21,440 It's not actually, no. We don't have toxic emissions from the refinery or anything. 1254 01:38:22,800 --> 01:38:29,197 The cathode nickel refinery is in Travis County. Why can't you do it with humans? 1255 01:38:29,280 --> 01:38:31,930 You can, you just run out of humans. Ah, I see. Okay. 1256 01:38:32,640 --> 01:38:35,642 No matter what you do, you have one quarter of the number of humans in America 1257 01:38:35,725 --> 01:38:38,877 than China. So if you have them do this thing, they can't do the other thing. 1258 01:38:38,960 --> 01:38:42,933 So then how do you build this refining capacity? Well, 1259 01:38:43,017 --> 01:38:45,120 you could do it with Optimi. 1260 01:38:48,960 --> 01:38:55,837 Not very many Americans are pining to do refining. I mean, 1261 01:38:55,920 --> 01:39:00,877 how many have you run into? Very few. Very 1262 01:39:00,960 --> 01:39:04,397 few pining to refine. BYD is reaching Tesla 1263 01:39:04,480 --> 01:39:09,517 production or sales in quantity. What do you think happens in global 1264 01:39:09,600 --> 01:39:14,306 markets as Chinese production in EVs scales up? 1265 01:39:14,389 --> 01:39:19,197 China is extremely competitive in manufacturing. 1266 01:39:19,280 --> 01:39:25,917 So I think there's going to be a massive flood of Chinese vehicles 1267 01:39:26,000 --> 01:39:33,000 and basically most manufactured things. As it is, as I said, China is probably 1268 01:39:37,040 --> 01:39:39,917 doing twice as much refining as the rest of the world combined. 1269 01:39:40,000 --> 01:39:47,000 So if you go down to fourth and fifth-tier supply chain stuff… 1270 01:39:50,320 --> 01:39:55,200 At the base level, you've got energy, then you've got mining and refining. 1271 01:39:55,920 --> 01:40:00,755 Those foundation layers are, like I said, as a rough guess, 1272 01:40:00,839 --> 01:40:03,757 China's doing twice as much refining 1273 01:40:03,840 --> 01:40:07,920 as the rest of the world combined. So any given thing is going to have 1274 01:40:09,280 --> 01:40:13,782 Chinese content because China's doing twice as much refining work as the rest 1275 01:40:13,866 --> 01:40:14,557 of the world. 1276 01:40:14,640 --> 01:40:21,640 But they'll go all the way to the finished product with the cars. 1277 01:40:22,080 --> 01:40:26,157 I mean China is a powerhouse. I think this year China will exceed 1278 01:40:26,240 --> 01:40:32,157 three times US electricity output. Electricity output is a reasonable 1279 01:40:32,240 --> 01:40:39,240 proxy for the economy. In order to run the factories 1280 01:40:39,680 --> 01:40:46,680 and run everything, you need electricity. It's a good proxy for the real economy. 1281 01:40:52,640 --> 01:40:55,197 If China passes three times the US electricity output, 1282 01:40:55,280 --> 01:41:00,586 it means that its industrial capacity—as rough approximation—will be three times 1283 01:41:00,669 --> 01:41:01,597 that of the US. 1284 01:41:01,680 --> 01:41:05,676 Reading between the lines, it sounds like what you're saying is absent some sort 1285 01:41:05,759 --> 01:41:06,237 of humanoid 1286 01:41:06,320 --> 01:41:13,320 recursive miracle in the next few years, on the whole manufacturing/energy/raw 1287 01:41:14,751 --> 01:41:16,397 materials chain, 1288 01:41:16,480 --> 01:41:21,150 China will just dominate whether it comes to AI or manufacturing EVs 1289 01:41:21,233 --> 01:41:23,037 or manufacturing humanoids. 1290 01:41:23,120 --> 01:41:30,120 In the absence of breakthrough innovations in the US, China will utterly dominate. 1291 01:41:35,200 --> 01:41:36,637 Interesting. Yes. 1292 01:41:36,720 --> 01:41:43,720 Robotics being the main breakthrough innovation. Well, 1293 01:41:44,073 --> 01:41:49,437 to scale AI in space, basically you need 1294 01:41:49,520 --> 01:41:56,000 humanoid robots, you need real-world AI, you need a million tons a year to orbit. 1295 01:41:57,280 --> 01:42:02,271 Let's just say if we get the mass driver on the moon going, my favorite thing, 1296 01:42:02,354 --> 01:42:03,200 then I think— 1297 01:42:03,920 --> 01:42:10,920 We'll have solved all our problems. I call that winning. I call it winning, big time. 1298 01:42:13,920 --> 01:42:16,077 You can finally be satisfied. You've done something. 1299 01:42:16,160 --> 01:42:18,557 Yes. You have the mass driver on the moon. 1300 01:42:18,640 --> 01:42:20,373 I just want to see that thing in operation. 1301 01:42:20,457 --> 01:42:22,317 Was that out of some sci-fi or where did you…? 1302 01:42:22,400 --> 01:42:26,397 Well, actually, there is a Heinlein book. The Moon is a Harsh Mistress. 1303 01:42:26,480 --> 01:42:30,397 Okay, yeah, but that's slightly different. That's a gravity slingshot or... 1304 01:42:30,480 --> 01:42:33,154 No, they have a mass driver on the Moon. Okay, yeah, 1305 01:42:33,237 --> 01:42:34,957 but they use that to attack Earth. 1306 01:42:35,040 --> 01:42:37,415 So maybe it's not the greatest... Well they use that to… 1307 01:42:37,499 --> 01:42:38,557 assert their independence. 1308 01:42:38,640 --> 01:42:40,637 Exactly. What are your plans for the mass driver on the Moon? 1309 01:42:40,720 --> 01:42:43,917 They asserted their independence. Earth government disagreed and they lobbed 1310 01:42:44,000 --> 01:42:47,997 things until Earth government agreed. That book is a hoot. I found that 1311 01:42:48,080 --> 01:42:50,230 book much better than his other one that everyone reads, 1312 01:42:50,313 --> 01:42:54,672 Stranger in a Strange Land. "Grok" comes from Stranger in a Strange Land. 1313 01:42:54,755 --> 01:42:58,037 The first two-thirds of Stranger in a Strange 1314 01:42:58,120 --> 01:43:02,320 Land are good, and then it gets very weird in the third portion. 1315 01:43:02,880 --> 01:43:09,880 But there are still some good concepts in there. One thing we were discussing a lot 1316 01:44:18,400 --> 01:44:24,877 is your system for managing people. You interviewed the first few thousand of 1317 01:44:24,960 --> 01:44:29,117 SpaceX employees and lots of other companies. It obviously doesn't scale. 1318 01:44:29,200 --> 01:44:32,477 Well, yes, but what doesn't scale? Me. 1319 01:44:32,560 --> 01:44:36,237 Sure, sure. I know that. But what are you looking for? 1320 01:44:36,320 --> 01:44:38,077 There literally are not enough hours in the day. It's impossible. 1321 01:44:38,160 --> 01:44:42,717 But what are you looking for that someone else who's good at interviewing 1322 01:44:42,800 --> 01:44:48,498 and hiring people… What's the je ne sais quoi? At this point, 1323 01:44:48,582 --> 01:44:51,437 I might have more training data 1324 01:44:51,520 --> 01:44:55,256 on evaluating technical talent especially—talent of all kinds I suppose, 1325 01:44:55,339 --> 01:44:56,317 but technical talent 1326 01:44:56,400 --> 01:45:00,034 especially—given that I've done so many technical interviews and then seen 1327 01:45:00,117 --> 01:45:00,951 the results. 1328 01:45:02,400 --> 01:45:09,400 So my training set is enormous and has a very wide range. 1329 01:45:11,680 --> 01:45:17,684 Generally, the things I ask for are bullet points for evidence of exceptional 1330 01:45:17,768 --> 01:45:18,601 ability. 1331 01:45:21,840 --> 01:45:27,116 These things can be pretty off the wall. It doesn't need to be in the specific 1332 01:45:27,199 --> 01:45:27,597 domain, 1333 01:45:27,680 --> 01:45:34,397 but evidence of exceptional ability. So if somebody can cite even one thing, 1334 01:45:34,480 --> 01:45:39,517 but let's say three things, where you go, "Wow, wow, wow," then that's a good sign. 1335 01:45:39,600 --> 01:45:42,719 Why do you have to be the one to determine that? No, I don't. I can't be. 1336 01:45:42,803 --> 01:45:46,603 It's impossible. The total headcount across all companies is 200,000 people. 1337 01:45:48,720 --> 01:45:52,877 But in the early days, what was it that you were looking for that 1338 01:45:52,960 --> 01:45:59,960 couldn't be delegated in those interviews? I guess I need to build my training set. 1339 01:46:02,320 --> 01:46:05,837 It's not like I batted a thousand here. I would make mistakes, but then I'd be 1340 01:46:05,920 --> 01:46:10,157 able to see where I thought somebody would work out well, but they didn't. 1341 01:46:10,240 --> 01:46:15,920 Then why did they not work out well? What can I do, I guess RL myself, to 1342 01:46:16,560 --> 01:46:20,880 in the future have a better batting average when interviewing people? 1343 01:46:22,560 --> 01:46:24,797 My batting average is still not perfect, but it's very high. 1344 01:46:24,880 --> 01:46:27,597 What are some surprising reasons people don't work out? 1345 01:46:27,680 --> 01:46:29,997 Surprising reasons… Like, they don't understand 1346 01:46:30,080 --> 01:46:34,797 technical domain, et cetera, et cetera. But you've got the long tail now of like, 1347 01:46:34,880 --> 01:46:39,052 "I was really excited about this person. It didn't work out." 1348 01:46:39,136 --> 01:46:40,880 Curious why that happens. 1349 01:46:43,920 --> 01:46:48,502 Generally what I tell people—I tell myself, I guess, aspirationally—is, 1350 01:46:48,586 --> 01:46:52,169 don't look at the resume. Just believe your interaction. 1351 01:46:52,252 --> 01:46:55,517 The resume may seem very impressive and it's like, 1352 01:46:55,600 --> 01:47:00,077 "Wow, the resume looks good." But if the conversation 1353 01:47:00,160 --> 01:47:07,160 after 20 minutes is not "wow," you should believe the conversation, not the paper. 1354 01:47:07,360 --> 01:47:12,798 I feel like part of your method is that… There was this meme in the media a few 1355 01:47:12,882 --> 01:47:13,917 years back about 1356 01:47:14,000 --> 01:47:17,979 Tesla being a revolving door of executive talent. Whereas actually, 1357 01:47:18,062 --> 01:47:19,677 I think when you look at it, 1358 01:47:19,760 --> 01:47:23,623 Tesla's had a very consistent and internally promoted executive bench over 1359 01:47:23,707 --> 01:47:24,637 the past few years. 1360 01:47:24,720 --> 01:47:28,880 Then at SpaceX, you have all these folks like Mark Juncosa and Steve Davis— 1361 01:47:29,600 --> 01:47:34,640 Steve Davis runs The Boring Company these days. Bill Riley, and folks like that. 1362 01:47:35,600 --> 01:47:42,600 It feels like part of what has worked well is having very capable technical deputies. 1363 01:47:43,280 --> 01:47:50,280 What do all of those people have in common? Well, the Tesla senior team, 1364 01:47:53,600 --> 01:47:58,240 at this point has probably got an average tenure of 10-12 years. It's quite long. 1365 01:48:03,280 --> 01:48:10,240 But there were times when Tesla went through an extremely rapid growth phase, 1366 01:48:11,280 --> 01:48:17,277 so things were just somewhat sped up. As you know, a company goes through 1367 01:48:17,360 --> 01:48:23,037 different orders of magnitude of size. People that could help manage, say, 1368 01:48:23,120 --> 01:48:27,712 a 50-person company versus a 500-person company versus a 5,000-person company 1369 01:48:27,796 --> 01:48:31,437 versus a 50,000-person company. You outgrew people. 1370 01:48:31,520 --> 01:48:34,077 It's just not the same team. It's not always the same team. 1371 01:48:34,160 --> 01:48:39,517 So if a company is growing very rapidly, the rate at which executive positions will 1372 01:48:39,600 --> 01:48:43,920 change will also be proportionate to the rapidity of the growth generally. 1373 01:48:47,440 --> 01:48:54,387 Tesla had a further challenge where when Tesla had very successful periods, 1374 01:48:54,470 --> 01:48:56,637 we would be relentlessly 1375 01:48:56,720 --> 01:49:03,437 recruited from. Like, relentlessly. When Apple had their electric car program, 1376 01:49:03,520 --> 01:49:07,024 they were carpet bombing Tesla with recruiting calls. 1377 01:49:07,108 --> 01:49:09,680 Engineers just unplugged their phones. 1378 01:49:10,880 --> 01:49:14,077 "I'm trying to get work done here." Yeah. "If I get one more call from 1379 01:49:14,160 --> 01:49:19,037 an Apple recruiter…" But their opening offer without any interview would be 1380 01:49:19,120 --> 01:49:26,120 like double the compensation at Tesla. So we had a bit of the "Tesla pixie 1381 01:49:28,240 --> 01:49:32,077 dust" thing where it's like, "Oh, if you hire a Tesla executive, 1382 01:49:32,160 --> 01:49:35,214 suddenly everything's going to be successful." 1383 01:49:35,298 --> 01:49:38,557 I've fallen prey to the pixie dust thing as well, 1384 01:49:38,640 --> 01:49:41,142 where it's like, "Oh, we'll hire someone from Google or Apple and they'll be 1385 01:49:41,226 --> 01:49:45,517 immediately successful," but that's not how it works. People are people. 1386 01:49:45,600 --> 01:49:51,009 There's no magical pixie dust. So when we had the pixie dust problem, 1387 01:49:51,093 --> 01:49:54,720 we would get relentlessly recruited from. 1388 01:49:57,840 --> 01:50:02,720 Also, Tesla being engineering, especially being primarily in Silicon Valley, 1389 01:50:03,600 --> 01:50:08,080 it's easier for people to just... They don't have to change their life very much. 1390 01:50:10,080 --> 01:50:14,237 Their commute's going to be the same. So how do you prevent that? 1391 01:50:14,320 --> 01:50:18,251 How do you prevent the pixie dust effect where everyone's trying to poach all 1392 01:50:18,334 --> 01:50:19,168 your people? 1393 01:50:21,200 --> 01:50:25,541 I don't think there's much we can do to stop it. 1394 01:50:25,624 --> 01:50:29,597 That's one of the reasons why Tesla… Really, 1395 01:50:29,680 --> 01:50:36,680 being in Silicon Valley and having the pixie dust thing at the same time meant 1396 01:50:38,362 --> 01:50:39,837 that there was 1397 01:50:39,920 --> 01:50:44,797 just a very, very aggressive recruitment. Presumably being in Austin helps then? 1398 01:50:44,880 --> 01:50:51,360 Austin, it helps. Tesla still has a majority of its engineering in California. 1399 01:50:56,240 --> 01:51:00,077 Getting engineers to move… I call it the "significant other" problem. 1400 01:51:00,160 --> 01:51:06,077 Yes, "significant others" have jobs. Exactly. So for Starbase that was 1401 01:51:06,160 --> 01:51:09,917 particularly difficult, since the odds of finding a non-SpaceX job… 1402 01:51:10,000 --> 01:51:13,677 In Brownsville, Texas… …are pretty low. It's 1403 01:51:13,760 --> 01:51:20,760 quite difficult. It's like a technology monastery thing, remote and mostly dudes. 1404 01:51:22,240 --> 01:51:29,240 Not much of an improvement over SF. If you go back to these people who've really 1405 01:51:34,800 --> 01:51:39,352 been very effective in a technical capacity at Tesla, at SpaceX, 1406 01:51:39,435 --> 01:51:44,558 and those sorts of places, what do you think they have in common other than... 1407 01:51:44,641 --> 01:51:47,280 Is it just that they're very sharp on the 1408 01:51:48,160 --> 01:51:51,362 rocketry or the technical foundations, or do you think it's something 1409 01:51:51,446 --> 01:51:54,477 organizational? Is it something about their ability to work with you? 1410 01:51:54,560 --> 01:52:01,560 Is it their ability to be flexible but not too flexible? 1411 01:52:03,920 --> 01:52:05,860 What makes a good sparring partner for you? 1412 01:52:05,944 --> 01:52:08,044 I don't think of it as a sparring partner. 1413 01:52:08,560 --> 01:52:11,517 If somebody gets things done, I love them, and if they don't, 1414 01:52:11,600 --> 01:52:14,358 I hate them. So it's pretty straightforward. 1415 01:52:14,441 --> 01:52:16,960 It's not like some idiosyncratic thing. 1416 01:52:17,760 --> 01:52:22,797 If somebody executes well, I'm a huge fan, and if they don't, I'm not. 1417 01:52:22,880 --> 01:52:25,757 But it's not about mapping to my idiosyncratic preferences. 1418 01:52:25,840 --> 01:52:29,540 I certainly try not to have it be mapping to my idiosyncratic preferences. 1419 01:52:36,240 --> 01:52:42,679 Generally, I think it's a good idea to hire for talent and drive 1420 01:52:42,762 --> 01:52:44,800 and trustworthiness. 1421 01:52:46,960 --> 01:52:50,049 And I think goodness of heart is important. 1422 01:52:50,132 --> 01:52:52,640 I underweighted that at one point. 1423 01:52:53,680 --> 01:53:00,680 So, are they a good person? Trustworthy? Smart and talented and hard working? 1424 01:53:01,200 --> 01:53:05,917 If so, you can add domain knowledge. But those fundamental traits, 1425 01:53:06,000 --> 01:53:10,113 those fundamental properties, you cannot change. 1426 01:53:10,196 --> 01:53:14,397 So most of the people who are at Tesla and SpaceX 1427 01:53:14,480 --> 01:53:17,580 did not come from the aerospace industry or the auto industry. 1428 01:53:18,240 --> 01:53:21,597 What has had to change most about your management style as your companies have 1429 01:53:21,680 --> 01:53:24,449 scaled from 100 to 1,000 to 10,000 people? 1430 01:53:24,532 --> 01:53:27,437 You're known for this very micro management, 1431 01:53:27,520 --> 01:53:33,040 just getting into the details of things. Nano management, please. Pico management. 1432 01:53:34,080 --> 01:53:39,117 Femto management. Keep going. 1433 01:53:39,200 --> 01:53:44,120 We're going to go all the way down to Planck's constant. 1434 01:53:44,640 --> 01:53:50,797 All the way down to Heisenberg uncertainty principle. 1435 01:53:50,880 --> 01:53:52,717 Are you still able to get into details as much as you want? 1436 01:53:52,800 --> 01:53:56,077 Would your companies be more successful if they were smaller? 1437 01:53:56,160 --> 01:53:58,557 How do you think about that? Because I have a fixed amount of 1438 01:53:58,640 --> 01:54:05,358 time in the day, my time is necessarily diluted as things grow and as the span 1439 01:54:05,442 --> 01:54:07,360 of activity increases. 1440 01:54:10,720 --> 01:54:17,357 It's impossible for me to actually be a micromanager because that would imply I 1441 01:54:17,440 --> 01:54:21,997 have some thousands of hours per day. It is a logical impossibility 1442 01:54:22,080 --> 01:54:29,080 for me to micromanage things. Now, there are times when I will drill down into a 1443 01:54:31,040 --> 01:54:38,040 specific issue because that specific issue is the limiting factor on the progress 1444 01:54:38,532 --> 01:54:39,920 of the company. 1445 01:54:42,480 --> 01:54:47,793 The reason for drilling into some very detailed item is because it is 1446 01:54:47,876 --> 01:54:49,357 the limiting factor. 1447 01:54:49,440 --> 01:54:55,951 It’s not arbitrarily drilling into tiny things. From a time standpoint, 1448 01:54:56,034 --> 01:54:57,437 it is physically 1449 01:54:57,520 --> 01:55:02,957 impossible for me to arbitrarily go into tiny things that don't matter. That would 1450 01:55:03,040 --> 01:55:09,597 result in failure. But sometimes the tiny things are decisive in victory. 1451 01:55:09,680 --> 01:55:16,680 Famously, you switched the Starship design from composites to steel. 1452 01:55:17,040 --> 01:55:21,042 Yes. You made that decision. That wasn't people going around saying, "Oh, 1453 01:55:21,125 --> 01:55:22,477 we found something better, boss." 1454 01:55:22,560 --> 01:55:25,197 That was you encouraging people against some resistance. 1455 01:55:25,280 --> 01:55:30,720 Can you tell us how you came to that whole concept of the steel switch? 1456 01:55:32,800 --> 01:55:39,800 Desperation, I'd say. Originally, we were going to make Starship out of carbon 1457 01:55:42,383 --> 01:55:43,216 fiber. 1458 01:55:45,600 --> 01:55:52,437 Carbon fiber is pretty expensive. When you do volume production, 1459 01:55:52,520 --> 01:55:55,357 you can get any given thing 1460 01:55:55,440 --> 01:56:00,637 to start to approach its material cost. The problem with carbon fiber is that 1461 01:56:00,720 --> 01:56:07,720 material cost is still very high. Particularly if you go for a high-strength 1462 01:56:10,480 --> 01:56:15,605 specialized carbon fiber that can handle cryogenic oxygen, 1463 01:56:15,688 --> 01:56:19,280 it's roughly 50 times the cost of steel. 1464 01:56:20,800 --> 01:56:24,477 At least in theory, it would be lighter. People generally think of steel as being 1465 01:56:24,560 --> 01:56:31,560 heavy and carbon fiber as being light. For room temperature applications, 1466 01:56:35,520 --> 01:56:39,157 like a Formula 1 car, static aero structure, 1467 01:56:39,240 --> 01:56:42,877 or any kind of aero structure really, you're 1468 01:56:42,960 --> 01:56:45,475 probably going to be better off with carbon fiber. 1469 01:56:45,558 --> 01:56:47,917 The problem is that we were trying to make this 1470 01:56:48,000 --> 01:56:53,757 enormous rocket out of carbon fiber and our progress was extremely slow. 1471 01:56:53,840 --> 01:56:56,940 It had been picked in the first place just because it's light? 1472 01:56:57,040 --> 01:57:04,040 Yes. At first glance, most people would think that the choice for 1473 01:57:04,960 --> 01:57:11,325 making something light would be carbon fiber. 1474 01:57:11,408 --> 01:57:17,917 The thing is that when you make something very 1475 01:57:18,000 --> 01:57:23,821 enormous out of carbon fiber and then you try to have the carbon fiber be 1476 01:57:23,904 --> 01:57:25,277 efficiently cured, 1477 01:57:25,360 --> 01:57:31,267 meaning not room temperature cured, because sometimes you got 50 plies 1478 01:57:31,351 --> 01:57:32,720 of carbon fiber… 1479 01:57:33,280 --> 01:57:39,597 Carbon fiber is really carbon string and glue. In order to have high strength, 1480 01:57:39,680 --> 01:57:45,040 you need an autoclave. Something that's essentially a high pressure oven. 1481 01:57:46,320 --> 01:57:51,224 If you have something that's gigantic, that one's got to be bigger than 1482 01:57:51,307 --> 01:57:52,141 the rocket. 1483 01:57:52,800 --> 01:57:58,098 We were trying to make an autoclave that's bigger than any autoclave that's ever 1484 01:57:58,182 --> 01:57:58,637 existed. 1485 01:57:58,720 --> 01:58:02,520 Or you can do room temperature cure, which takes a long time and has issues. 1486 01:58:03,280 --> 01:58:08,229 The final issue is that we were just making very slow progress with carbon 1487 01:58:08,312 --> 01:58:09,145 fiber. 1488 01:58:12,480 --> 01:58:18,077 The meta question is why it had to be you who made that decision. 1489 01:58:18,160 --> 01:58:20,797 There's many engineers on your team. How did the team not arrive at steel? 1490 01:58:20,880 --> 01:58:24,157 Yeah exactly. This is part of a broader question, understanding your comparative 1491 01:58:24,240 --> 01:58:29,757 advantage at your companies. Because we were making very slow 1492 01:58:29,840 --> 01:58:33,757 progress with carbon fiber, I was like, "Okay, we've got to try something else." 1493 01:58:33,840 --> 01:58:40,840 For the Falcon 9, the primary airframe is made of aluminum lithium, which has 1494 01:58:41,520 --> 01:58:47,837 a very good strength-to-weight. Actually, it has about the same, 1495 01:58:47,920 --> 01:58:51,517 maybe better, strength to weight for its application than carbon fiber. 1496 01:58:51,600 --> 01:58:53,357 But aluminum lithium is very difficult to work with. 1497 01:58:53,440 --> 01:58:56,260 In order to weld it, you have to do something called friction stir welding, 1498 01:58:56,343 --> 01:58:57,243 where you join the 1499 01:58:57,600 --> 01:59:02,717 metal without entering the liquid phase. It's kind of wild that you can do that. 1500 01:59:02,800 --> 01:59:07,508 But with this particular type of welding, you can do that. 1501 01:59:07,591 --> 01:59:10,317 It's very difficult. Let's say you 1502 01:59:10,400 --> 01:59:14,342 want to make a modification or attach something to aluminum lithium, 1503 01:59:14,426 --> 01:59:16,237 you now have to use a mechanical 1504 01:59:16,320 --> 01:59:20,566 attachment with seals. You can't weld it on. 1505 01:59:20,649 --> 01:59:24,797 So I wanted to avoid using aluminum lithium 1506 01:59:24,880 --> 01:59:31,880 for the primary structure for Starship. There was this very special grade of 1507 01:59:34,960 --> 01:59:39,333 carbon fiber that had very good mass properties. With a rocket, 1508 01:59:39,416 --> 01:59:41,597 you're really trying to maximize 1509 01:59:41,680 --> 01:59:45,920 the percentage of the rocket that is propellant, minimize the mass obviously. 1510 01:59:48,800 --> 01:59:51,920 But like I said, we were making very slow progress. 1511 01:59:54,480 --> 01:59:56,317 I said, "at this rate, we’re never going to get to Mars. 1512 01:59:56,400 --> 02:00:01,277 So we've got to think of something else." I didn't want to use aluminum lithium 1513 02:00:01,360 --> 02:00:06,557 because of the difficulty of friction stir welding, especially doing that at scale. 1514 02:00:06,640 --> 02:00:12,000 It was hard enough at 3.6 meters in diameter, let alone at 9 meters or above. 1515 02:00:12,880 --> 02:00:19,880 Then I said, "what about steel?" I had a clue here because some of 1516 02:00:21,200 --> 02:00:24,992 the early US rockets had used very thin steel. 1517 02:00:25,076 --> 02:00:29,120 The Atlas rockets had used a steel balloon tank. 1518 02:00:30,880 --> 02:00:34,678 It's not like steel had never been used before. It actually had been used. 1519 02:00:34,761 --> 02:00:35,517 When you look at 1520 02:00:35,600 --> 02:00:42,449 the material properties of stainless steel, full-hard, 1521 02:00:42,532 --> 02:00:46,557 strain hardened stainless steel, 1522 02:00:46,640 --> 02:00:53,088 at cryogenic temperature the strength to weight is actually similar to carbon 1523 02:00:53,171 --> 02:00:54,004 fiber. 1524 02:00:54,320 --> 02:00:58,477 If you look at material properties at room temperature, it looks like 1525 02:00:58,560 --> 02:01:03,757 the steel is going to be twice as heavy. But if you look at the material properties 1526 02:01:03,840 --> 02:01:08,800 at cryogenic temperature of full-hard steel, stainless of particular grades, 1527 02:01:10,240 --> 02:01:14,560 then you actually get to a similar strength to weight as carbon fiber. 1528 02:01:15,200 --> 02:01:19,677 In the case of Starship, both the fuel and the oxidizer are cryogenic. 1529 02:01:19,760 --> 02:01:25,938 For Falcon 9, the fuel is rocket propellant-grade kerosene, 1530 02:01:26,021 --> 02:01:30,160 basically a very pure form of jet fuel. 1531 02:01:32,320 --> 02:01:37,997 That is roughly room temperature. Although we do actually chill it slightly below, 1532 02:01:38,080 --> 02:01:41,357 we chill it like a beer. Delicious. 1533 02:01:41,440 --> 02:01:45,757 We do chill it, but it's not cryogenic. In fact, if we made it cryogenic, 1534 02:01:45,840 --> 02:01:49,917 it would just turn to wax. But for Starship, 1535 02:01:50,000 --> 02:01:56,560 it's liquid methane and liquid oxygen. They are liquid at similar temperatures. 1536 02:01:59,040 --> 02:02:03,517 Basically, almost the entire primary structure is at cryogenic temperature. 1537 02:02:03,600 --> 02:02:10,600 So then you've got a 300-series stainless that's strain hardened. 1538 02:02:12,800 --> 02:02:17,597 Because almost all things are cryogenic temperature, it actually has similar 1539 02:02:17,680 --> 02:02:24,680 strength to weight as carbon fiber. But it costs 50x less in raw 1540 02:02:25,040 --> 02:02:30,717 material and is very easy to work with. You can weld stainless steel outdoors. 1541 02:02:30,800 --> 02:02:36,320 You could smoke a cigar while welding stainless steel. It's very resilient. 1542 02:02:37,200 --> 02:02:41,260 You can modify it easily. If you want to attach something, 1543 02:02:41,343 --> 02:02:43,200 you just weld it right on. 1544 02:02:44,560 --> 02:02:51,560 Very easy to work with, very low cost. Like I said, at cryogenic temperature, 1545 02:02:52,920 --> 02:02:58,006 it’s similar strength-to-weight to carbon fiber. 1546 02:02:58,089 --> 02:03:02,637 Then when you factor in that we have a much 1547 02:03:02,720 --> 02:03:07,517 reduced heat shield mass, because the melting point of steel, is much greater 1548 02:03:07,600 --> 02:03:13,597 than the melting point of aluminum… It's about twice the melting point of aluminum. 1549 02:03:13,680 --> 02:03:16,510 So you can just run the rocket much hotter? Yes, 1550 02:03:16,593 --> 02:03:18,877 especially for the ship which is coming 1551 02:03:18,960 --> 02:03:25,357 in like a blazing meteor. You can greatly reduce 1552 02:03:25,440 --> 02:03:32,440 the mass of the heat shield. You can cut the mass of the windward 1553 02:03:34,000 --> 02:03:40,013 part of the heat shield, maybe in half, and you don't need any heat shielding 1554 02:03:40,096 --> 02:03:41,680 on the leeward side. 1555 02:03:45,840 --> 02:03:49,437 The net result is that actually the steel rocket weighs less than 1556 02:03:49,520 --> 02:03:56,520 the carbon fiber rocket, because the resin in the carbon fiber rocket starts to melt. 1557 02:04:00,560 --> 02:04:04,865 Basically, carbon fiber and aluminum have about the same operating temperature 1558 02:04:04,949 --> 02:04:05,782 capabilities, 1559 02:04:06,720 --> 02:04:09,216 whereas steel can operate at twice the temperature. 1560 02:04:09,299 --> 02:04:11,120 These are very rough approximations. 1561 02:04:12,160 --> 02:04:14,717 I won't build the rocket. What I mean is people will say, 1562 02:04:14,800 --> 02:04:18,637 "Oh, he said this twice. It's actually 0.8." I'm like, shut up, assholes. 1563 02:04:18,720 --> 02:04:24,128 That's what the main comment's going to be about. God damn it. The point is, 1564 02:04:24,212 --> 02:04:25,357 in retrospect, we 1565 02:04:25,440 --> 02:04:27,997 should have started with steel in the beginning. It was dumb not to do steel. 1566 02:04:28,080 --> 02:04:32,797 Okay, but to play this back to you, what I'm hearing is that steel was a riskier, 1567 02:04:32,880 --> 02:04:37,267 less proven path, other than the early US rockets. 1568 02:04:37,351 --> 02:04:43,277 Versus carbon fiber was a worse but more proven out path. So you need to be the 1569 02:04:43,360 --> 02:04:48,637 one to push for, "Hey, we're going to do this riskier path and just figure it out." 1570 02:04:48,720 --> 02:04:52,080 So you're fighting a sort of conservatism in a sense. 1571 02:04:52,800 --> 02:04:56,816 That's why I initially said that the issue is that we weren't making fast enough 1572 02:04:56,899 --> 02:04:57,277 progress. 1573 02:04:57,360 --> 02:05:02,717 We were having trouble making even a small barrel section of the carbon fiber 1574 02:05:02,800 --> 02:05:08,877 that didn't have wrinkles in it. Because at that large scale, you have to 1575 02:05:08,960 --> 02:05:11,612 have many plies, many layers of the carbon fiber. 1576 02:05:11,696 --> 02:05:14,237 You've got to cure it and you've got to cure it 1577 02:05:14,320 --> 02:05:18,637 in such a way that it doesn't have any wrinkles or defects. 1578 02:05:18,720 --> 02:05:24,400 Carbon fiber is much less resilient than steel. It has much less toughness. 1579 02:05:26,000 --> 02:05:32,560 Stainless steel will stretch and bend, the carbon fiber will tend to shatter. 1580 02:05:35,920 --> 02:05:38,670 Toughness being the area under the stress strain curve. 1581 02:05:39,360 --> 02:05:42,737 You're generally going to have to do better with steel, 1582 02:05:42,821 --> 02:05:44,877 but stainless steel to be precise. 1583 02:05:44,960 --> 02:05:51,117 One other Starship question. So I visited Starbase, I think it was two years ago, 1584 02:05:51,200 --> 02:05:54,381 with Sam Teller, and that was awesome. It was very cool to see, 1585 02:05:54,465 --> 02:05:55,677 in a whole bunch of ways. 1586 02:05:55,760 --> 02:06:01,758 One thing I noticed was that people really took pride in the simplicity of things, 1587 02:06:01,842 --> 02:06:06,418 where everyone wants to tell you how Starship is just a big soda can, 1588 02:06:06,502 --> 02:06:09,437 and we're hiring welders, and if you can weld 1589 02:06:09,520 --> 02:06:12,317 in any industrial project, you can weld here. 1590 02:06:12,400 --> 02:06:15,280 But there's a lot of pride in the simplicity. 1591 02:06:16,320 --> 02:06:18,637 Well, factually Starship is a very complicated rocket. 1592 02:06:18,720 --> 02:06:23,117 So that's what I'm getting at. Are things simple or are they complex? 1593 02:06:23,200 --> 02:06:26,682 I think maybe just what they're trying to say is that you don't have to have prior 1594 02:06:26,765 --> 02:06:31,808 experience in the rocket industry to work on Starship. 1595 02:06:31,891 --> 02:06:36,717 Somebody just needs to be smart and work hard 1596 02:06:36,800 --> 02:06:39,712 and be trustworthy and they can work on a rocket. 1597 02:06:39,795 --> 02:06:42,157 They don't need prior rocket experience. 1598 02:06:42,240 --> 02:06:47,837 Starship is the most complicated machine ever made by humans, by a long shot. 1599 02:06:47,920 --> 02:06:50,797 In what regards? Anything, really. I'd 1600 02:06:50,880 --> 02:06:57,880 say there isn't a more complex machine. I'd say that pretty much any project I 1601 02:07:00,080 --> 02:07:07,080 can think of would be easier than this. That's why nobody has ever made a fully 1602 02:07:08,800 --> 02:07:13,500 reusable orbital rocket. It's a very hard problem. 1603 02:07:13,584 --> 02:07:17,997 Many smart people have tried before, very smart 1604 02:07:18,080 --> 02:07:22,109 people with immense resources, and they failed. 1605 02:07:22,192 --> 02:07:26,397 And we haven't succeeded yet. Falcon is partially 1606 02:07:26,480 --> 02:07:32,717 reusable, but the upper stage is not. Starship Version 3, 1607 02:07:32,800 --> 02:07:39,800 I think this design can be fully reusable. That full reusability is what will enable 1608 02:07:41,920 --> 02:07:48,920 us to become a multi-planet civilization. Any technical problem, even like a Hadron 1609 02:07:52,400 --> 02:07:55,197 Collider or something like that, is an easier problem than this. 1610 02:07:55,280 --> 02:07:58,477 We spent a lot of time on bottlenecks. Can you say what the current Starship 1611 02:07:58,560 --> 02:08:04,877 bottlenecks are, even at a high level? Trying to make it not explode, generally. 1612 02:08:04,960 --> 02:08:09,837 It really wants to explode. That old chestnut. All those 1613 02:08:09,920 --> 02:08:13,597 combustible materials. We've had two boosters explode on the test stand. 1614 02:08:13,680 --> 02:08:20,000 One obliterated the entire test facility. So it only takes that one mistake. 1615 02:08:21,680 --> 02:08:25,360 The amount of energy contained in a Starship is insane. 1616 02:08:25,920 --> 02:08:30,317 Is that why it's harder than Falcon? It's because it's just more energy? 1617 02:08:30,400 --> 02:08:37,037 It's a lot of new technology. It's pushing the performance envelope. The 1618 02:08:37,120 --> 02:08:40,381 Raptor 3 engine is a very, very advanced engine. 1619 02:08:40,465 --> 02:08:43,517 It's by far the best rocket engine ever made. 1620 02:08:43,600 --> 02:08:48,557 But it desperately wants to blow up. Just to put things into perspective here, 1621 02:08:48,640 --> 02:08:54,558 on liftoff the rocket is generating over 100 gigawatts of power. 1622 02:08:54,641 --> 02:08:57,360 That’s 20% of US electricity. 1623 02:08:58,480 --> 02:08:59,517 It's actually insane. It's a great comparison. 1624 02:08:59,600 --> 02:09:02,397 While not exploding. Sometimes. 1625 02:09:02,480 --> 02:09:06,397 Sometimes, yes. So I was like, how does it not explode? 1626 02:09:06,480 --> 02:09:12,717 There's thousands of ways that it could explode and only one way that it doesn't. 1627 02:09:12,800 --> 02:09:16,971 So we want it not only to really not explode, 1628 02:09:17,055 --> 02:09:22,077 but fly reliably on a daily basis, like once per hour. 1629 02:09:22,160 --> 02:09:25,837 Obviously, if it blows up a lot, it's very difficult to maintain that 1630 02:09:25,920 --> 02:09:26,920 launch cadence. Yes. 1631 02:09:30,880 --> 02:09:33,837 What's the single biggest remaining problem for Starship? 1632 02:09:33,920 --> 02:09:40,920 It's having the heat shield be reusable. No one's ever made a reusable orbital heat 1633 02:09:43,732 --> 02:09:44,477 shield. 1634 02:09:44,560 --> 02:09:51,269 So the heat shield's gotta make it through the ascent phase without shucking a bunch 1635 02:09:51,352 --> 02:09:51,997 of tiles, 1636 02:09:52,080 --> 02:09:59,080 and then it's gotta come back in and also not lose a bunch of tiles or overheat 1637 02:09:59,508 --> 02:10:05,037 the main airframe. Isn't that hard because it's fundamentally a consumable? 1638 02:10:05,120 --> 02:10:07,932 Well, yes, but your brake pads in your car are also consumable, 1639 02:10:08,015 --> 02:10:09,357 but they last a very long time. 1640 02:10:09,440 --> 02:10:12,400 Fair. So it just needs to last a very long time. 1641 02:10:17,600 --> 02:10:21,997 We have brought the ship back and had it do a soft landing in the ocean. 1642 02:10:22,080 --> 02:10:26,320 We've done that a few times. But it lost a lot of tiles. 1643 02:10:27,680 --> 02:10:34,680 It was not reusable without a lot of work. Even though it did come to a soft landing, 1644 02:10:35,280 --> 02:10:39,200 it would not have been reusable without a lot of work. 1645 02:10:40,080 --> 02:10:44,317 So it's not really reusable in that sense. That's the biggest problem that remains, 1646 02:10:44,400 --> 02:10:50,877 a fully reusable heat shield. You want to be able to land it, 1647 02:10:50,960 --> 02:10:57,117 refill propellant and fly again. You can't do this laborious inspection 1648 02:10:57,200 --> 02:11:04,200 of 40,000 tiles type of thing. When I read biographies of yours, 1649 02:11:06,720 --> 02:11:10,948 it seems like you're just able to drive the sense of urgency and drive the sense 1650 02:11:11,032 --> 02:11:15,037 of "this is the thing that can scale." I'm curious why you 1651 02:11:15,120 --> 02:11:19,917 think other organizations of your… SpaceX and Tesla are really big companies now. 1652 02:11:20,000 --> 02:11:24,397 You're still able to keep that culture. What goes wrong with other companies such 1653 02:11:24,480 --> 02:11:28,000 that they're not able to do that? I don't know. 1654 02:11:29,040 --> 02:11:31,357 Like today, you said you had a bunch of SpaceX meetings. 1655 02:11:31,440 --> 02:11:33,357 What is it that you're doing there that's keeping that? 1656 02:11:33,440 --> 02:11:40,440 It’s adding urgency? Well, I don't know. I guess the urgency is going 1657 02:11:42,640 --> 02:11:45,275 to come from whoever is leading the company. 1658 02:11:45,358 --> 02:11:47,437 I have a maniacal sense of urgency. 1659 02:11:47,520 --> 02:11:52,397 So that maniacal sense of urgency projects through the rest of the company. 1660 02:11:52,480 --> 02:11:56,380 Is it because of consequences? They're like, "Elon set a crazy deadline, 1661 02:11:56,463 --> 02:12:01,277 but if I don't get it, I know what happens to me." Is it just that you're able to 1662 02:12:01,360 --> 02:12:03,837 identify bottlenecks and get rid of them so people can move fast? 1663 02:12:03,920 --> 02:12:07,037 How do you think about why your companies are able to move fast? 1664 02:12:07,120 --> 02:12:14,120 I'm constantly addressing the limiting factor. On the deadlines front, 1665 02:12:17,262 --> 02:12:20,077 I generally actually 1666 02:12:20,160 --> 02:12:25,437 try to aim for a deadline that I at least think is at the 50th percentile. 1667 02:12:25,520 --> 02:12:29,104 So it's not like an impossible deadline, but it's the most aggressive deadline I 1668 02:12:29,187 --> 02:12:32,704 can think of that could be achieved with 50% probability. 1669 02:12:32,788 --> 02:12:35,840 Which means that it'll be late half the time. 1670 02:12:42,400 --> 02:12:45,760 There is a law of gas expansion that applies to schedules. 1671 02:12:48,160 --> 02:12:51,415 If you said we're going to do something in five years, 1672 02:12:51,498 --> 02:12:53,600 which to me is like infinity time, 1673 02:12:55,040 --> 02:12:58,800 it will expand to fill the available schedule and it'll take five years. 1674 02:13:05,520 --> 02:13:07,757 Physics will limit how fast you can do certain things. 1675 02:13:07,840 --> 02:13:13,600 So scaling up manufacturing, there's a rate at which you can move the atoms 1676 02:13:15,200 --> 02:13:17,437 and scale manufacturing. That's why you can't instantly 1677 02:13:17,520 --> 02:13:23,008 make a million units a year of something. You've got to design the manufacturing 1678 02:13:23,092 --> 02:13:26,992 line. You've got to bring it up. You've got to ride the S-curve of production. 1679 02:13:31,840 --> 02:13:34,720 What can I say that's actually helpful to people? 1680 02:13:38,160 --> 02:13:43,760 Generally, a maniacal sense of urgency is a very big deal. 1681 02:13:47,440 --> 02:13:53,516 You want to have an aggressive schedule and you want to figure out what 1682 02:13:53,599 --> 02:13:54,557 the limiting 1683 02:13:54,640 --> 02:13:58,540 factor is at any point in time and help the team address that limiting factor. 1684 02:13:59,360 --> 02:14:04,080 So Starlink was slowly in the works for many years. 1685 02:14:05,040 --> 02:14:07,757 We talked about it all the way in the beginning of the company. 1686 02:14:07,840 --> 02:14:12,157 So then there was a team you had built in Redmond, and then at one point you 1687 02:14:12,240 --> 02:14:19,240 decided this team is just not cutting it. It went for a few years slowly, 1688 02:14:23,206 --> 02:14:28,302 and so why didn't you act earlier, and why did you act when you did? 1689 02:14:28,385 --> 02:14:30,717 Why was that the right moment at which to act? 1690 02:14:30,800 --> 02:14:36,480 I have these very detailed engineering reviews weekly. 1691 02:14:38,880 --> 02:14:42,316 That's maybe a very unusual level of granularity. 1692 02:14:42,399 --> 02:14:45,117 I don't know anyone who runs a company, 1693 02:14:45,200 --> 02:14:49,917 or at least a manufacturing company, that goes with the level of detail that I go 1694 02:14:50,000 --> 02:14:56,547 into. It's not as though... I have a pretty good understanding of what's 1695 02:14:56,631 --> 02:14:57,837 actually going 1696 02:14:57,920 --> 02:15:04,920 on because we go through things in detail. I'm a big believer in skip-level meetings 1697 02:15:06,960 --> 02:15:11,499 where instead of having the person that reports to me say things, 1698 02:15:11,582 --> 02:15:14,000 it's everyone that reports to them 1699 02:15:14,640 --> 02:15:21,640 saying something in the technical review. And there can't be advanced preparation. 1700 02:15:25,440 --> 02:15:31,597 Otherwise you're going to get "glazed", as I say these days. 1701 02:15:31,680 --> 02:15:34,867 Exactly. Very Gen Z of you. How do you prevent advanced preparation? 1702 02:15:34,950 --> 02:15:37,437 Do you call on them randomly? No, I just go around the room. 1703 02:15:37,520 --> 02:15:44,000 Everyone provides an update. It's a lot of information to keep in your head. 1704 02:15:48,000 --> 02:15:50,702 If you have meetings weekly or twice weekly, 1705 02:15:50,785 --> 02:15:53,760 you've got a snapshot of what that person said. 1706 02:15:56,240 --> 02:16:03,240 You can then plot the progress points. You can sort of mentally plot the 1707 02:16:03,680 --> 02:16:07,440 points on a curve and say, "are we converging to a solution or not?" 1708 02:16:12,880 --> 02:16:19,527 I'll take drastic action only when I conclude that success is not in a set 1709 02:16:19,610 --> 02:16:21,520 of possible outcomes. 1710 02:16:22,400 --> 02:16:28,174 So when I finally reach the conclusion that unless drastic action is done, 1711 02:16:28,257 --> 02:16:29,757 we have no chance of 1712 02:16:29,840 --> 02:16:36,077 success, then I must take drastic action. I came to that conclusion in 2018, 1713 02:16:36,160 --> 02:16:43,160 took drastic action and fixed the problem. You've got many, many companies. 1714 02:16:44,348 --> 02:16:49,117 In each of them it sounds like you do this kind of deep engineering understanding of 1715 02:16:49,200 --> 02:16:54,080 what the relevant bottlenecks are so you can do these reviews with people. 1716 02:16:56,320 --> 02:16:59,677 You've been able to scale it up to five, six, seven companies. 1717 02:16:59,760 --> 02:17:04,477 Within one of these companies, you have many different mini companies within them. 1718 02:17:04,560 --> 02:17:07,277 What determines the max amount here? Because you have like 80 companies…? 1719 02:17:07,360 --> 02:17:09,917 80? No. But you have so many 1720 02:17:10,000 --> 02:17:13,277 already. That's already remarkable. By this current number. 1721 02:17:13,360 --> 02:17:18,560 Exactly. We can barely keep one company together. 1722 02:17:23,040 --> 02:17:30,040 It depends on the situation. I actually don't have regular meetings with 1723 02:17:30,636 --> 02:17:35,980 The Boring Company, so The Boring Company is sort of cruising along. Basically, 1724 02:17:36,063 --> 02:17:37,837 if something is working well and 1725 02:17:37,920 --> 02:17:41,920 making good progress, then there's no point in me spending time on it. 1726 02:17:42,560 --> 02:17:47,487 I actually allocate time according to where the limiting factor. 1727 02:17:47,570 --> 02:17:49,840 Where are things problematic? 1728 02:17:50,960 --> 02:17:57,291 Where are we pushing against? What is holding us back? I focus, 1729 02:17:57,375 --> 02:17:59,837 at the risk of saying the 1730 02:17:59,920 --> 02:18:04,465 words too many times, on the limiting factor. 1731 02:18:04,549 --> 02:18:09,197 The irony is if something's going really well, 1732 02:18:09,280 --> 02:18:12,717 they don't see much of me. But if something is going badly, 1733 02:18:12,800 --> 02:18:15,939 they'll see a lot of me. Or not even badly… 1734 02:18:16,022 --> 02:18:18,637 If something is the limiting factor. 1735 02:18:18,720 --> 02:18:21,037 The limiting factor, exactly. It’s not exactly going badly but it’s the 1736 02:18:21,120 --> 02:18:25,277 thing that we need to make go faster. When something’s a limiting factor at 1737 02:18:25,360 --> 02:18:31,917 SpaceX or Tesla, are you talking weekly and daily with the engineer that's 1738 02:18:32,000 --> 02:18:35,119 working on it? How does that actually work? 1739 02:18:35,203 --> 02:18:38,480 Most things that are the limiting factor are 1740 02:18:39,680 --> 02:18:45,600 weekly and some things are twice weekly. The AI5 chip review is twice weekly. 1741 02:18:46,800 --> 02:18:50,929 Every Tuesday and Saturday is the chip review. 1742 02:18:51,012 --> 02:18:54,317 Is it open ended in how long it goes? 1743 02:18:54,400 --> 02:19:01,400 Technically, yes, but usually it's two or three hours. Sometimes less. It depends on 1744 02:19:03,760 --> 02:19:06,504 how much information we've got to go through. That's another thing. 1745 02:19:06,587 --> 02:19:07,517 I'm just trying to tease 1746 02:19:07,600 --> 02:19:11,757 out the differences here because the outcomes seem quite different. 1747 02:19:11,840 --> 02:19:14,477 I think it's interesting to know what inputs are different. 1748 02:19:14,560 --> 02:19:18,389 It feels like in the corporate world, one, like you were saying, 1749 02:19:18,472 --> 02:19:19,917 the CEO doing engineering 1750 02:19:20,000 --> 02:19:24,859 reviews does not always happen despite the fact that that is what the company 1751 02:19:24,942 --> 02:19:25,437 is doing. 1752 02:19:25,520 --> 02:19:31,079 But then time is often pretty finely sliced into half hour meetings or even 15 1753 02:19:31,163 --> 02:19:32,237 minute meetings. 1754 02:19:32,320 --> 02:19:38,237 It seems like you hold more open-ended, "We're talking about it until we figure 1755 02:19:38,320 --> 02:19:42,957 it out" type things. Sometimes. But most 1756 02:19:43,040 --> 02:19:49,448 of them seem to more or less stay on time. 1757 02:19:49,531 --> 02:19:56,531 Today's Starship engineering review went a bit 1758 02:19:56,640 --> 02:20:00,600 longer because there were more topics to discuss. 1759 02:20:00,684 --> 02:20:04,397 They're trying to figure out how to scale to a 1760 02:20:04,480 --> 02:20:08,317 million plus tons to orbit per year. It’s quite challenging. 1761 02:20:08,400 --> 02:20:13,571 Can I ask a question? You said about Optimus and AI that they're going 1762 02:20:13,654 --> 02:20:14,997 to result in double 1763 02:20:15,080 --> 02:20:20,522 digit growth rates within a matter of years. Oh, like the economy? Yes. 1764 02:20:20,606 --> 02:20:22,157 I think that's right. 1765 02:20:22,240 --> 02:20:27,997 What was the point of the DOGE cuts if the economy is going to grow so much? 1766 02:20:28,080 --> 02:20:30,980 Well, I think waste and fraud are not good things to have. 1767 02:20:33,360 --> 02:20:40,360 I was actually pretty worried about... In the absence of AI and robotics, 1768 02:20:41,600 --> 02:20:48,000 we're actually totally screwed because the national debt is piling up like crazy. 1769 02:20:50,080 --> 02:20:52,918 The interest payments to national debt exceed the military budget, 1770 02:20:53,001 --> 02:20:54,157 which is a trillion dollars. 1771 02:20:54,240 --> 02:20:57,680 So we have over a trillion dollars just in interest payments. 1772 02:21:00,000 --> 02:21:03,597 I was pretty concerned about that. Maybe if I spend some time, we can 1773 02:21:03,680 --> 02:21:08,930 slow down the bankruptcy of the United States and give us enough time for the AI 1774 02:21:09,013 --> 02:21:09,597 and robots 1775 02:21:09,680 --> 02:21:15,997 to help solve the national debt. Or not help solve, it's the only 1776 02:21:16,080 --> 02:21:20,535 thing that could solve the national debt. We are 1000% going to go bankrupt as 1777 02:21:20,618 --> 02:21:23,972 a country, and fail as a country, without AI and robots. 1778 02:21:24,055 --> 02:21:26,720 Nothing else will solve the national debt. 1779 02:21:30,880 --> 02:21:37,880 We just need enough time to build the AI and robots to not go bankrupt before then. 1780 02:21:39,200 --> 02:21:43,150 I guess the thing I'm curious about is, when DOGE starts you have this enormous 1781 02:21:43,920 --> 02:21:48,397 ability to enact reform. Not that enormous. 1782 02:21:48,480 --> 02:21:53,837 Sure. I totally buy your point that it's important that AI and robotics drive 1783 02:21:53,920 --> 02:21:56,328 productivity improvements, drive GDP growth. 1784 02:21:56,412 --> 02:21:58,877 But why not just directly go after the things 1785 02:21:58,960 --> 02:22:03,757 you were pointing out, like the tariffs on certain components, or permitting? 1786 02:22:03,840 --> 02:22:10,840 I'm not the president. And it is very hard to cut things that are obvious waste 1787 02:22:12,290 --> 02:22:13,277 and fraud, 1788 02:22:13,360 --> 02:22:20,360 like ridiculous waste and fraud. What I discovered is that it's extremely 1789 02:22:21,360 --> 02:22:27,517 difficult even to cut very obvious waste and fraud from the government because 1790 02:22:27,600 --> 02:22:28,637 the government 1791 02:22:28,720 --> 02:22:34,077 has to operate on who's complaining. If you cut off payments to fraudsters, 1792 02:22:34,160 --> 02:22:38,383 they immediately come up with the most sympathetic sounding reasons to continue 1793 02:22:38,466 --> 02:22:43,837 the payment. They don't say, "Please keep the fraud going." They’re like, 1794 02:22:43,920 --> 02:22:46,157 "You're killing baby pandas." 1795 02:22:46,240 --> 02:22:50,007 Meanwhile, no baby pandas are dying. They're just making it up. 1796 02:22:50,091 --> 02:22:51,597 The fraudsters are capable 1797 02:22:51,680 --> 02:22:56,317 of coming up with extremely compelling, heart-wrenching stories that are false, 1798 02:22:56,400 --> 02:23:02,520 but nonetheless sound sympathetic. That's what happened. 1799 02:23:02,603 --> 02:23:06,480 Perhaps I should have known better. 1800 02:23:10,000 --> 02:23:15,081 But I thought, wait, let's try to cut some amount of waste and pork from 1801 02:23:15,164 --> 02:23:16,157 the government. 1802 02:23:16,240 --> 02:23:22,397 Maybe there shouldn't be 20 million people marked as alive in Social Security who are 1803 02:23:22,480 --> 02:23:28,766 definitely dead, and over the age of 115. The oldest American is 114. 1804 02:23:28,849 --> 02:23:30,797 So it's safe to say if 1805 02:23:30,880 --> 02:23:36,951 somebody is 115 and marked as alive in the Social Security database, 1806 02:23:37,034 --> 02:23:39,757 there's either a typo… Somebody 1807 02:23:39,840 --> 02:23:45,390 should call them and say, "We seem to have your birthday wrong, 1808 02:23:45,473 --> 02:23:47,357 or we need to mark you 1809 02:23:47,440 --> 02:23:52,717 as dead." One of the two things. Very intimidating call to get. 1810 02:23:52,800 --> 02:23:58,320 Well, it seems like a reasonable thing. Say if their birthday is in the future 1811 02:23:59,920 --> 02:24:04,126 and they have a Small Business Administration loan, 1812 02:24:04,209 --> 02:24:06,480 and their birthday is 2165, 1813 02:24:07,680 --> 02:24:13,677 we either have a typo or we have fraud. So we say, "we appear to have gotten the 1814 02:24:13,760 --> 02:24:16,957 century of your birth incorrect." Or a great plot for a movie. 1815 02:24:17,040 --> 02:24:23,437 Yes. That's what I mean by, ludicrous fraud. Were those people getting payments? 1816 02:24:23,520 --> 02:24:26,327 Some were getting payments from Social Security. 1817 02:24:26,410 --> 02:24:29,277 But the main fraud vector was to mark somebody as 1818 02:24:29,360 --> 02:24:35,895 alive in Social Security and then use every other government payment system 1819 02:24:35,979 --> 02:24:39,920 to basically do fraud. Because what those other government payment systems do, 1820 02:24:40,800 --> 02:24:45,249 they would simply do an "are you alive" check to the Social Security database. 1821 02:24:45,332 --> 02:24:46,237 It's a bank shot. 1822 02:24:46,320 --> 02:24:50,000 What would you estimate is the total amount of fraud from this mechanism? 1823 02:24:52,080 --> 02:24:55,160 By the way, the Government Accountability Office has done these estimates before. 1824 02:24:55,243 --> 02:25:00,049 I'm not the only one. In fact, I think the GAO did an analysis, 1825 02:25:00,132 --> 02:25:02,637 a rough estimate of fraud during 1826 02:25:02,720 --> 02:25:07,917 the Biden administration, and calculated it at roughly half a trillion dollars. 1827 02:25:08,000 --> 02:25:11,597 So don't take my word for it. Take a report issued during the 1828 02:25:11,680 --> 02:25:16,237 Biden administration. How about that? From this Social Security mechanism? 1829 02:25:16,320 --> 02:25:22,317 It's one of many. It's important to appreciate that the government is 1830 02:25:22,400 --> 02:25:29,400 very ineffective at stopping fraud. It's not like a company where, with 1831 02:25:30,240 --> 02:25:33,585 stopping fraud, you've got a motivation because it's affecting the earnings 1832 02:25:33,669 --> 02:25:34,317 of your company. 1833 02:25:34,400 --> 02:25:41,400 The government just prints more money. You need caring and competence. These are 1834 02:25:44,880 --> 02:25:51,880 in short supply at the federal level. When you go to the DMV, do you think, 1835 02:25:52,640 --> 02:25:57,677 "Wow, this is a bastion of competence"? Well, now imagine it's worse than the DMV 1836 02:25:57,760 --> 02:26:04,760 because it's the DMV that can print money. At least the state level DMVs need to... 1837 02:26:05,440 --> 02:26:08,557 The states more or less need to stay within their budget or they go bankrupt. 1838 02:26:08,640 --> 02:26:11,841 But the federal government just prints more money. 1839 02:26:11,924 --> 02:26:14,797 If there's actually half a trillion of fraud, 1840 02:26:14,880 --> 02:26:21,880 why was it not possible to cut all that? You really have to stand back 1841 02:26:26,146 --> 02:26:28,477 and recalibrate 1842 02:26:28,560 --> 02:26:35,560 your expectations for competence. Because you're operating in a world 1843 02:26:36,160 --> 02:26:40,080 where you've got to make ends meet. You've got to pay your bills... 1844 02:26:40,960 --> 02:26:47,960 Find the microphones. Exactly. It's not like there's a giant, 1845 02:26:49,920 --> 02:26:56,920 largely uncaring monster bureaucracy. It's a bunch of anachronistic computers 1846 02:26:57,760 --> 02:27:03,437 that are just sending payments. One of the things that the DOGE 1847 02:27:03,520 --> 02:27:10,520 team did sounds so simple and probably will save $100-200 billion a year. 1848 02:27:14,000 --> 02:27:18,869 It was simply requiring payments from the main Treasury computer—which is called 1849 02:27:18,952 --> 02:27:21,278 PAM, Payment Accounts Master or something like that, 1850 02:27:21,361 --> 02:27:23,361 there's $5 trillion payments a year—that 1851 02:27:25,520 --> 02:27:32,157 go out have a payment appropriation code. Make it mandatory, not optional, that you 1852 02:27:32,240 --> 02:27:37,158 have anything at all in the comment field. 1853 02:27:37,241 --> 02:27:42,397 You have to recalibrate how dumb things are. 1854 02:27:42,480 --> 02:27:46,066 Payments were being sent out with no appropriation code, 1855 02:27:46,150 --> 02:27:48,557 not checking back to any congressional 1856 02:27:48,640 --> 02:27:54,797 appropriation, and with no explanation. This is why the Department of War, 1857 02:27:54,880 --> 02:27:57,395 formerly the Department of Defense, cannot pass an audit, 1858 02:27:57,479 --> 02:27:59,037 because the information is literally 1859 02:27:59,120 --> 02:28:03,794 not there. Recalibrate your expectations. I want to better understand this half 1860 02:28:03,878 --> 02:28:09,520 a trillion number, because there's an IG report in 2024. Why is it so low? 1861 02:28:10,160 --> 02:28:13,917 Maybe, but we found that over seven years, the Social Security fraud 1862 02:28:14,000 --> 02:28:17,117 they estimated was like $70 billion over seven years, so like $10 billion a year. 1863 02:28:17,200 --> 02:28:20,317 So I'd be curious to see what the other $490 billion is. 1864 02:28:20,400 --> 02:28:23,250 Federal government expenditures are $7.5 trillion a year. 1865 02:28:26,400 --> 02:28:29,838 How competent do you think the government is? 1866 02:28:29,921 --> 02:28:33,437 The discretionary spending there is like… 15%? 1867 02:28:33,520 --> 02:28:36,317 But it doesn't matter. Most of the fraud is non-discretionary. 1868 02:28:36,400 --> 02:28:40,480 It's basically fraudulent Medicare, Medicaid, Social Security, 1869 02:28:45,760 --> 02:28:52,317 disability. There's a zillion government payments. A bunch of these payments are in 1870 02:28:52,400 --> 02:28:59,400 fact block transfers to the states. So the federal government doesn't 1871 02:28:59,520 --> 02:29:04,397 even have the information in a lot of cases to even know if there's fraud. 1872 02:29:04,480 --> 02:29:10,237 Let's consider reductio ad absurdum. The government is perfect and has no fraud. 1873 02:29:10,320 --> 02:29:17,066 What is your probability estimate of that? Zero. Okay, so then would you say, 1874 02:29:17,150 --> 02:29:18,397 fraud and waste 1875 02:29:18,480 --> 02:29:25,480 at the government is 90% efficient? That also would be quite generous. 1876 02:29:27,120 --> 02:29:32,317 But if it's only 90%, that means that there's $750 billion a year of waste and 1877 02:29:32,400 --> 02:29:35,127 fraud. And it's not 90%. It's not 90% effective. 1878 02:29:35,211 --> 02:29:37,997 This seems like a strange way to first principles 1879 02:29:38,080 --> 02:29:42,130 the amount of fraud in the government. Just like, how much do you think there is? 1880 02:29:43,280 --> 02:29:45,597 Anyways, we don't have to do it live, but I'd be curious— 1881 02:29:45,680 --> 02:29:49,597 You know a lot about fraud at Stripe? People are constantly trying to do fraud. 1882 02:29:49,680 --> 02:29:53,258 Yeah, but as you say, it's a little bit of a... We've really ground it down, 1883 02:29:53,341 --> 02:29:54,077 but it's a little 1884 02:29:54,160 --> 02:29:57,645 bit of a different problem space because you're dealing with a much more 1885 02:29:57,729 --> 02:30:03,677 heterogeneous set of fraud vectors here than we are. But at Stripe, you have high 1886 02:30:03,760 --> 02:30:07,757 competence and you try hard. You have high competence and 1887 02:30:07,840 --> 02:30:14,840 high caring, but still fraud is non-zero. Now imagine it's at a much bigger scale, 1888 02:30:15,211 --> 02:30:21,166 there's much less competence, and much less caring. At PayPal back in the day, 1889 02:30:21,249 --> 02:30:22,557 we tried to manage 1890 02:30:22,640 --> 02:30:26,757 fraud down to about 1% of the payment volume. That was very difficult. 1891 02:30:26,840 --> 02:30:32,521 It took a tremendous amount of competence and caring to get fraud merely to 1%. 1892 02:30:32,604 --> 02:30:36,237 Now imagine that you're an organization where 1893 02:30:36,320 --> 02:30:39,531 there's much less caring and much less competence. 1894 02:30:39,614 --> 02:30:41,837 It's going to be much more than 1%. 1895 02:30:41,920 --> 02:30:47,760 How do you feel now looking back on politics and doing stuff there? 1896 02:30:48,480 --> 02:30:55,480 Looking from the outside in, two things have been quite impactful: one, 1897 02:30:56,290 --> 02:30:58,957 the America PAC, and two, 1898 02:30:59,040 --> 02:31:05,757 the acquisition of Twitter at the time. But also it seems like there 1899 02:31:05,840 --> 02:31:11,760 was a bunch of heartache. What's your grading of the whole experience? 1900 02:31:16,320 --> 02:31:22,141 I think those things needed to be done to maximize the probability that the future 1901 02:31:22,224 --> 02:31:23,057 is good. 1902 02:31:27,200 --> 02:31:33,145 Politics generally is very tribal. People lose their objectivity usually 1903 02:31:33,228 --> 02:31:34,400 with politics. 1904 02:31:35,600 --> 02:31:40,450 They generally have trouble seeing the good on the other side or the bad on 1905 02:31:40,533 --> 02:31:44,502 their own side. That's generally how it goes. That, I guess, 1906 02:31:44,586 --> 02:31:47,917 was one of the things that surprised me the most. 1907 02:31:48,000 --> 02:31:50,110 You often simply cannot reason with people. 1908 02:31:50,193 --> 02:31:51,997 If they're in one tribe or the other. 1909 02:31:52,080 --> 02:31:55,757 They simply believe that everything their tribe does is good and anything 1910 02:31:55,840 --> 02:32:02,652 the other political tribe does is bad. Persuading them otherwise is almost 1911 02:32:02,735 --> 02:32:03,760 impossible. 1912 02:32:07,760 --> 02:32:14,760 But I think overall those actions—acquiring Twitter, 1913 02:32:16,660 --> 02:32:22,397 getting Trump elected, even though 1914 02:32:22,480 --> 02:32:29,200 it makes a lot of people angry—I think those actions were good for civilization. 1915 02:32:30,640 --> 02:32:33,340 How does it feed into the future you're excited about? 1916 02:32:33,440 --> 02:32:40,440 Well, America needs to be strong enough to last long enough to extend life to other 1917 02:32:42,000 --> 02:32:48,849 planets and to get AI and robotics to the point where we can ensure that the future 1918 02:32:48,932 --> 02:32:49,765 is good. 1919 02:32:51,120 --> 02:32:55,464 On the other hand, if we were to descend into, say, 1920 02:32:55,547 --> 02:32:59,197 communism or some situation where the state 1921 02:32:59,280 --> 02:33:06,280 was extremely oppressive, that would mean that we might not be able to become 1922 02:33:07,096 --> 02:33:08,720 multi-planetary. 1923 02:33:10,320 --> 02:33:16,960 The state might stamp out our progress in AI and robotics. 1924 02:33:21,360 --> 02:33:28,271 Optimus, Grok, et cetera. Not just yours, but any revenue-maximizing company's 1925 02:33:28,354 --> 02:33:29,437 products will 1926 02:33:29,520 --> 02:33:36,520 be leveraged by the government over time. How does this concern manifest in what 1927 02:33:37,040 --> 02:33:39,909 private companies should be willing to give governments? 1928 02:33:39,993 --> 02:33:41,680 What kinds of guardrails? Should 1929 02:33:44,400 --> 02:33:50,957 AI models be made to do whatever the government that has contracted 1930 02:33:51,040 --> 02:33:57,357 them out to do and asks them to do? Should Grok get to say, "Actually, 1931 02:33:57,440 --> 02:34:00,957 even if the military wants to do X, no, Grok will not do that"? 1932 02:34:01,040 --> 02:34:08,040 I think maybe the biggest danger of AI and robotics going wrong is government. 1933 02:34:16,800 --> 02:34:21,597 People who are opposed to corporations or worried about corporations should 1934 02:34:21,680 --> 02:34:25,357 really worry the most about government. Because government is just a 1935 02:34:25,440 --> 02:34:30,397 corporation in the limit. Government is just the biggest 1936 02:34:30,480 --> 02:34:37,480 corporation with a monopoly on violence. I always find it a strange dichotomy where 1937 02:34:38,720 --> 02:34:40,873 people would think corporations are bad, but the government is good, 1938 02:34:40,956 --> 02:34:41,597 when the government is 1939 02:34:41,680 --> 02:34:48,680 simply the biggest and worst corporation. But people have that dichotomy. 1940 02:34:49,124 --> 02:34:50,877 They somehow think 1941 02:34:50,960 --> 02:34:54,592 at the same time that government can be good, but corporations bad, 1942 02:34:54,675 --> 02:34:58,640 and this is not true. Corporations have better morality than the government. 1943 02:34:59,520 --> 02:35:06,520 I actually think it’s a thing to be worried about. 1944 02:35:06,617 --> 02:35:12,637 The government could potentially use AI and 1945 02:35:12,720 --> 02:35:17,917 robotics to suppress the population. That is a serious concern. 1946 02:35:18,000 --> 02:35:22,480 As the guy building AI and robotics, how do you prevent that? 1947 02:35:28,560 --> 02:35:32,321 If you limit the powers of government, which is really what the US Constitution 1948 02:35:32,404 --> 02:35:33,197 is intended to do, 1949 02:35:33,280 --> 02:35:37,079 to limit the powers of government, then you're probably going to have a better 1950 02:35:37,163 --> 02:35:42,317 outcome than if you have more government. Robotics will be available 1951 02:35:42,400 --> 02:35:46,160 to all governments, right? I don’t know about all governments. 1952 02:35:49,840 --> 02:35:54,411 It's difficult to predict. I can say what's the endpoint, 1953 02:35:54,494 --> 02:35:57,677 or what is many years in the future, but 1954 02:35:57,760 --> 02:36:04,760 it's difficult to predict the path along that way. If civilization progresses, 1955 02:36:07,120 --> 02:36:08,717 AI will vastly 1956 02:36:08,800 --> 02:36:14,880 exceed the sum of all human intelligence. There will be far more robots than humans. 1957 02:36:16,240 --> 02:36:20,637 Along the way what happens is very difficult to predict. 1958 02:36:20,720 --> 02:36:25,527 It seems one thing you could do is just say, "whatever government X, 1959 02:36:25,610 --> 02:36:29,679 you're not allowed to use Optimus to do X, Y, Z." Just write out a policy. 1960 02:36:29,763 --> 02:36:31,357 I think you tweeted recently that 1961 02:36:31,440 --> 02:36:36,637 Grok should have a moral constitution. One of those things could be that we 1962 02:36:36,720 --> 02:36:40,270 limit what governments are allowed to do with this advanced technology. 1963 02:36:47,440 --> 02:36:53,117 Technically if politicians pass a law and they can enforce that law, 1964 02:36:53,200 --> 02:37:00,001 then it's hard to not do that law. The best thing we can have is limited 1965 02:37:00,084 --> 02:37:00,957 government 1966 02:37:01,040 --> 02:37:08,040 where you have the appropriate crosschecks between the executive, judicial, 1967 02:37:08,240 --> 02:37:10,640 and legislative branches. 1968 02:37:12,000 --> 02:37:15,958 The reason I'm curious about it is that at some point it seems the limits will come 1969 02:37:16,042 --> 02:37:16,875 from you. 1970 02:37:17,360 --> 02:37:19,845 You've got the Optimus, you've got the space GPUs… 1971 02:37:19,928 --> 02:37:22,240 You think I'll be the boss of the government? 1972 02:37:24,480 --> 02:37:31,480 Already it's the case with SpaceX that for things that are crucial—the government 1973 02:37:31,928 --> 02:37:32,397 really 1974 02:37:32,480 --> 02:37:35,916 cares about getting certain satellites up in space or whatever—it needs SpaceX. 1975 02:37:35,999 --> 02:37:36,832 It is the 1976 02:37:37,600 --> 02:37:42,800 necessary contractor. You are in the process of building more and more of the 1977 02:37:44,960 --> 02:37:48,883 technological components of the future that will have an analogous role 1978 02:37:48,966 --> 02:37:50,237 in different industries. 1979 02:37:50,320 --> 02:37:56,920 You could have this ability to set some policy that suppressing classical 1980 02:37:57,004 --> 02:37:58,477 liberalism in any 1981 02:37:58,560 --> 02:38:04,320 way… "My companies will not help in any way with that", or some policy like that. 1982 02:38:05,280 --> 02:38:08,557 I will do my best to ensure that anything that's within my control 1983 02:38:08,640 --> 02:38:15,640 maximizes the good outcome for humanity. I think anything else would be shortsighted, 1984 02:38:18,080 --> 02:38:25,080 because obviously I'm part of humanity, so I like humans. Pro human. 1985 02:38:29,040 --> 02:38:33,200 You mentioned that Dojo 3 will be used for space-based compute. 1986 02:38:34,400 --> 02:38:38,557 You really read what I say. I don't know if you know, 1987 02:38:38,640 --> 02:38:43,352 Elon, but you have a lot of followers. Dead giveaway. 1988 02:38:43,435 --> 02:38:46,157 How did you discern my secrets? 1989 02:38:46,240 --> 02:38:53,240 Oh I posted them on X. How do you design a chip for space? What changes? 1990 02:38:54,640 --> 02:39:00,160 You want to design it to be more radiation tolerant and run at a higher temperature. 1991 02:39:03,120 --> 02:39:08,157 Roughly, if you increase the operating temperature by 20% in degrees Kelvin, 1992 02:39:08,240 --> 02:39:14,877 you can cut your radiator mass in half. So running at a higher temperature 1993 02:39:14,960 --> 02:39:20,717 is helpful in space. There are various things 1994 02:39:20,800 --> 02:39:26,557 you can do for shielding the memory. But neural nets are going to be very 1995 02:39:26,640 --> 02:39:29,917 resilient to bit flips. Most of what happens 1996 02:39:30,000 --> 02:39:36,613 for radiation is random bit flips. But if you've got a multi-trillion parameter 1997 02:39:36,696 --> 02:39:39,929 model and you get a few bit flips, it doesn't matter. 1998 02:39:40,012 --> 02:39:42,637 Heuristic programs are going to be much more 1999 02:39:42,720 --> 02:39:47,200 sensitive to bit flips than some giant parameter file. 2000 02:39:49,760 --> 02:39:56,077 I just design it to run hot. I think you pretty much do 2001 02:39:56,160 --> 02:40:01,360 it the same way that you do things on Earth, apart from making it run hotter. 2002 02:40:02,240 --> 02:40:04,477 The solar array is most of the weight on the satellite. 2003 02:40:04,560 --> 02:40:10,957 Is there a way to make the GPUs even more powerful than what Nvidia and TPUs and 2004 02:40:11,040 --> 02:40:16,009 et cetera are planning on doing that would be especially privileged in 2005 02:40:16,092 --> 02:40:17,680 the space-based world? 2006 02:40:18,640 --> 02:40:25,640 The basic math is, if you can do about a kilowatt per reticle, then you'd need 100 2007 02:40:31,040 --> 02:40:36,924 million full reticle chips to do 100 gigawatts. 2008 02:40:37,007 --> 02:40:42,720 Depending on what your yield assumptions are, 2009 02:40:44,400 --> 02:40:48,677 that tells you how many chips you need to make. 2010 02:40:48,760 --> 02:40:53,037 If you're going to have 100 gigawatts of power, 2011 02:40:53,120 --> 02:41:00,120 you need 100 million chips that are running at a kilowatt sustained, 2012 02:41:02,108 --> 02:41:05,197 per reticle. Basic math. 2013 02:41:05,280 --> 02:41:12,280 100 million chips depends on… If you look at the die size of something like 2014 02:41:13,520 --> 02:41:17,570 Blackwell GPUs or something, and how many you can get out of a wafer, you can get 2015 02:41:18,400 --> 02:41:25,400 on the order of dozens or less per wafer. So basically, this is a world where if 2016 02:41:25,600 --> 02:41:32,111 we're putting that out every single year, you're producing millions of wafers 2017 02:41:32,195 --> 02:41:33,028 a month. 2018 02:41:33,920 --> 02:41:37,837 That's the plan with TeraFab? Millions of wafers a month of advanced process nodes? 2019 02:41:37,920 --> 02:41:40,595 Yeah it could be north of a million or something. 2020 02:41:40,679 --> 02:41:42,397 You’ve got to do the memory too. 2021 02:41:42,480 --> 02:41:46,397 Are you going to make a memory fab? I think the TeraFab's got to do memory. 2022 02:41:46,480 --> 02:41:48,703 It's got to do logic, memory, and packaging. 2023 02:41:48,786 --> 02:41:50,957 I'm very curious how somebody gets started. 2024 02:41:51,040 --> 02:41:54,157 This is the most complicated thing man has ever made. 2025 02:41:54,240 --> 02:41:57,840 Obviously, if anybody's up to the task, you're up to the task. 2026 02:41:58,560 --> 02:42:01,710 So you realize it's a bottleneck, and you go to your engineers. 2027 02:42:02,560 --> 02:42:07,760 What do you tell them to do? "I want a million wafers a month in 2030." 2028 02:42:09,040 --> 02:42:13,997 That’s right. That’s exactly what I want. Do you call ASML? What is the next step? 2029 02:42:14,080 --> 02:42:21,080 No so much to ask. We make a little fab and see what happens. 2030 02:42:22,640 --> 02:42:25,757 Make our mistakes at a small scale and then make a big one. 2031 02:42:25,840 --> 02:42:29,277 Is a little fab done? No, it's not done. We're 2032 02:42:29,360 --> 02:42:34,877 not going to keep that cat in the bag. That cat's going to come out of the bag. 2033 02:42:34,960 --> 02:42:37,604 There'll be drones hovering over the bloody thing. 2034 02:42:37,687 --> 02:42:39,677 You'll be able to see its construction 2035 02:42:39,760 --> 02:42:46,760 progress on X in real time. Look, I don't know, we could just 2036 02:42:47,920 --> 02:42:54,920 flounder in failure, to be fair. Success is not guaranteed. 2037 02:42:55,462 --> 02:43:00,237 Since we want to try to make something 2038 02:43:00,320 --> 02:43:07,320 like 100 million… We want 100 gigawatts of power and chips that can take 100 2039 02:43:11,057 --> 02:43:13,600 gigawatts by 2030. 2040 02:43:18,400 --> 02:43:20,557 We’ll take as many chips as our suppliers will give us. 2041 02:43:20,640 --> 02:43:25,260 I've actually said this to TSMC and Samsung and Micron: 2042 02:43:25,343 --> 02:43:28,080 "please build more fabs faster". 2043 02:43:28,640 --> 02:43:31,520 We will guarantee to buy the output of those fabs. 2044 02:43:32,080 --> 02:43:39,080 So they're already moving as fast as they can. It's us plus them. 2045 02:43:46,640 --> 02:43:50,717 There's a narrative that the people doing AI want a very large number 2046 02:43:50,800 --> 02:43:56,077 of chips as quickly as possible. Then many of the input suppliers, 2047 02:43:56,160 --> 02:43:59,513 the fabs, but also the turbine manufacturers, 2048 02:43:59,596 --> 02:44:02,797 are not ramping up production very quickly. 2049 02:44:02,880 --> 02:44:05,277 No, they're not. The explanation you hear 2050 02:44:05,360 --> 02:44:08,437 is that they're dispositionally conservative. 2051 02:44:08,520 --> 02:44:11,597 They're Taiwanese or German, as the story may 2052 02:44:11,680 --> 02:44:15,711 be. They just don't believe... Is that really the explanation or is there 2053 02:44:15,795 --> 02:44:16,640 something else? 2054 02:44:17,280 --> 02:44:24,144 Well, it's reasonable to... If somebody's been in the computer memory business 2055 02:44:24,228 --> 02:44:30,160 for 30 or 40 years… They've seen cycles. They've seen boom and bust 10 times. 2056 02:44:32,080 --> 02:44:37,277 That's a lot of layers of scar tissue. During the boom times, it looks like 2057 02:44:37,360 --> 02:44:41,357 everything is going to be great forever. Then the crash happens and they're 2058 02:44:41,440 --> 02:44:46,329 desperately trying to avoid bankruptcy. Then there's another boom and another 2059 02:44:46,413 --> 02:44:47,246 crash. 2060 02:44:48,320 --> 02:44:51,597 Are there other ideas you think others should go pursue that 2061 02:44:51,680 --> 02:44:54,443 you're not for whatever reasons right now? 2062 02:44:54,526 --> 02:44:57,440 There are a few companies that are pursuing 2063 02:44:58,240 --> 02:45:03,437 new ways of doing chips, but they're just not scaling fast. 2064 02:45:03,520 --> 02:45:06,800 I don't even mean within AI, I mean just generally. 2065 02:45:07,600 --> 02:45:11,934 People should do the thing where they find that they're highly motivated to do 2066 02:45:12,017 --> 02:45:12,850 that thing, 2067 02:45:13,280 --> 02:45:20,280 as opposed to some idea that I suggest. They should do the thing that they find 2068 02:45:21,520 --> 02:45:28,520 personally interesting and motivating to do. But going back to the limiting 2069 02:45:30,080 --> 02:45:35,157 factor… I used that phrase about 100 times. 2070 02:45:35,240 --> 02:45:40,640 The current limiting factor that I see in the 2071 02:45:47,920 --> 02:45:54,920 three to four year timeframe, it's chips. In the one year timeframe, it's energy, 2072 02:45:56,160 --> 02:46:01,760 power production, electricity. It's not clear to me that there's enough 2073 02:46:02,640 --> 02:46:07,200 usable electricity to turn on all the AI chips that are being made. 2074 02:46:10,000 --> 02:46:12,508 Towards the end of this year, I think people are going to have real trouble 2075 02:46:12,591 --> 02:46:17,357 turning on... The chip output will exceed the ability to turn chips on. 2076 02:46:17,440 --> 02:46:22,350 What's your plan to deal with that world? We're trying to accelerate electricity 2077 02:46:22,433 --> 02:46:23,267 production. 2078 02:46:24,800 --> 02:46:30,948 I guess that's maybe one of the reasons that xAI will be maybe the leader, 2079 02:46:31,032 --> 02:46:32,800 hopefully the leader. 2080 02:46:34,080 --> 02:46:38,560 We'll be able to turn on more chips than other people can turn on, faster, 2081 02:46:39,280 --> 02:46:45,837 because we're good at hardware. Generally, the innovations from 2082 02:46:45,920 --> 02:46:52,293 the corporations that call themselves labs, the ideas tend to flow… 2083 02:46:52,376 --> 02:46:59,145 It's rare to see that there's more than about a six-month difference. 2084 02:46:59,229 --> 02:47:04,077 The ideas travel back and forth with the people. 2085 02:47:04,160 --> 02:47:11,160 So I think you sort of hit the hardware wall and then whichever company can scale 2086 02:47:11,680 --> 02:47:17,517 hardware the fastest will be the leader. So I think xAI will be able to scale 2087 02:47:17,600 --> 02:47:20,557 hardware the fastest and therefore most likely will be the leader. 2088 02:47:20,640 --> 02:47:27,640 You joked or were self-conscious about using the "limiting factor" phrase again. 2089 02:47:28,240 --> 02:47:30,198 But I actually think there's something deep here. 2090 02:47:30,282 --> 02:47:32,157 If you look at a lot of things we've touched on 2091 02:47:32,240 --> 02:47:35,600 over the course of it, it’s maybe a good note to end on. 2092 02:47:37,120 --> 02:47:44,120 If you think of a senescent, low-agency company, it would have some bottleneck and 2093 02:47:45,280 --> 02:47:49,757 not really be doing anything about it. Marc Andreessen had the line of, 2094 02:47:49,840 --> 02:47:54,131 "most people are willing to endure any amount of chronic pain to avoid acute 2095 02:47:54,215 --> 02:47:58,601 pain". It feels like a lot of the cases we're talking about are just leaning into 2096 02:47:58,684 --> 02:48:03,862 the acute pain, whatever it is. "Okay, we got to figure out how to work with steel, 2097 02:48:03,945 --> 02:48:05,117 or we got to figure 2098 02:48:05,200 --> 02:48:09,197 out how to run the chips in space." We'll take some near-term acute pain 2099 02:48:09,280 --> 02:48:13,517 to actually solve the bottleneck. So that's kind of a unifying theme. 2100 02:48:13,600 --> 02:48:19,117 I have a high pain threshold. That's helpful. To solve the bottleneck. 2101 02:48:19,200 --> 02:48:26,200 Yes. One thing I can say is, I think the future is going to be very interesting. 2102 02:48:36,400 --> 02:48:41,421 As I said at Davos—I think I was on the ground for like three hours 2103 02:48:41,505 --> 02:48:42,800 or something—it's 2104 02:48:45,200 --> 02:48:50,797 better to err on the side of optimism and be wrong than err on the side of pessimism 2105 02:48:50,880 --> 02:48:57,880 and be right, for quality of life. You'll be happier if you err on 2106 02:49:01,440 --> 02:49:05,677 the side of optimism rather than erring on the side of pessimism. 2107 02:49:05,760 --> 02:49:09,117 So I recommend erring on the side of optimism. Here's to that. 2108 02:49:09,200 --> 02:49:11,677 Cool. Elon, thanks for doing this. Thank you. 2109 02:49:11,760 --> 02:49:17,757 All right, thanks guys. All right. Great stamina. 2110 02:49:17,840 --> 02:49:21,040 Hopefully this didn't count as a pain in the pain tolerance.