Elon Musk – “In 36 months, the cheapest place to put AI will be space” (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) So, are there really three hours of (00:00:02) questions or or has are you [ __ ] (00:00:04) serious? (00:00:04) >> Yeah. [laughter] (00:00:07) You don't even talk about Elon, man. (00:00:11) >> I mean, it's the most interesting point. (00:00:13) All the story lines are kind of (00:00:14) converging. Yeah. (00:00:15) >> Right now, so we'll see how much (00:00:18) >> almost like I planned it. (00:00:19) >> Exactly. Well, we'll get (00:00:21) >> I would never do such a thing. (00:00:22) [laughter] (00:00:24) So, as you know better than anybody (00:00:25) else, uh the total cost of ownership of (00:00:27) a data center, only 10 to 15% is energy. (00:00:30) And that's the part you're presumably (00:00:31) saving by moving this into space. Most (00:00:33) of it's the GPUs. If they're in space, (00:00:35) it's harder to service them or you can't (00:00:37) service them. And so, the depreciation (00:00:39) cycle goes down on them. So, like it's (00:00:40) just way more expensive to have the GPUs (00:00:41) in space presumably. What's the reason (00:00:44) to put them in space? (00:00:45) >> Um well, the availability of energy is (00:00:49) the issue. Um so uh I mean if you look (00:00:54) at at electrical output um outside of (00:00:56) China everywhere outside of China it's (00:00:58) more or less flat. It's very you know (00:01:00) maybe a slight increase but for pretty (00:01:02) close flat. China has a rapid increase (00:01:04) in in electrical output. But if you're (00:01:06) putting data centers anywhere except (00:01:07) China where you going to get your (00:01:09) electricity um especially as you scale (00:01:12) uh the output of chips is growing um (00:01:13) pretty much exponentially but the output (00:01:16) of electricity is flat. So how are you (00:01:18) going to turn them chips on? um you know (00:01:21) >> magical power sources, magical (00:01:23) electricity fairies. (00:01:24) >> You mean you're famously [laughter] (00:01:26) you're famously a big fan of solar one (00:01:28) terowatt of solar power. So with a 25% (00:01:30) compat factor like four terowatts of (00:01:32) solar panels it's like 1% of the land (00:01:34) area of the United States and that's (00:01:36) like far in this you were in the (00:01:37) singularity when we've got one terowatt (00:01:38) of data centers right um so what are we (00:01:42) running out of (00:01:42) >> how far into the singularity are you (00:01:44) though [laughter] (00:01:44) >> you tell me (00:01:45) >> yeah exactly so so I think I think we (00:01:47) we'll find we're in the singularity and (00:01:49) like okay we still got a long way to go (00:01:50) [laughter] (00:01:51) >> but is this like a is the plan to like (00:01:53) put it in the space after we've covered (00:01:55) Nevada and solar panels (00:01:57) >> I think it's pretty hard to cover Nevada (00:01:58) in solar panels you get permits from (00:02:00) like the permits for try getting the (00:02:02) permits for that. (00:02:03) >> So space is really a reg it's really a (00:02:05) regulatory play. It's like harder to (00:02:07) harder to build on land than it is in (00:02:08) space. (00:02:09) >> It's it's harder to scale um on the (00:02:12) ground than it is to scale in space. Um (00:02:14) but but also the the you're going to get (00:02:18) about five times the um effectiveness of (00:02:22) solar panels in space versus the ground. (00:02:24) And you don't need batteries. Um I (00:02:27) almost wore my other shirt which says (00:02:28) it's always sunny in space which it is. (00:02:32) [laughter] (00:02:34) So um because you don't have a dayight (00:02:36) cycle or uh seasonality uh clouds uh or (00:02:41) or an atmosphere in space uh because the (00:02:44) atmosphere alone um uh results in about (00:02:48) a 30% uh loss of energy. Um so uh so (00:02:52) you're going for any given uh solar (00:02:54) panels can do about five times more uh (00:02:57) power in space than on the ground and (00:02:58) you avoid the cost of having batteries (00:03:00) to carry you through the night. (00:03:03) Um so it's it's actually much cheaper to (00:03:05) do it in space and I I my prediction is (00:03:08) that (00:03:10) um it will be by far the cheapest place (00:03:12) to put uh AI will be space in 36 months (00:03:16) or less maybe 30 months. (00:03:17) >> 36 months (00:03:18) >> less than 36 months. Um, how do you (00:03:20) service GPUs as they fail, which happens (00:03:21) quite often in training? (00:03:24) >> Actually, it it it depends on how how (00:03:27) recent the GPUs are that arrived. I (00:03:29) mean, at this point, we found our GPUs (00:03:31) to be quite reliable. Um, there's infant (00:03:34) mortality, which you can obviously iron (00:03:35) out on the ground. (00:03:36) >> Um, so you can just run them on the (00:03:38) ground um and confirm that you don't (00:03:39) have info mortality with with the GPUs. (00:03:41) But once they once they start working, (00:03:43) their actual reliability and and once (00:03:46) they start working and you're past the (00:03:47) initial, you know, debug cycle of Nvidia (00:03:50) or whatever or whoever is making the (00:03:52) chips, um could be Tesla Tesla AI 6 (00:03:55) chips or something like that or it could (00:03:57) be, you know, TPUs or trains or (00:03:59) whatever. Um the uh the rival is (00:04:03) actually they're quite reliable past (00:04:04) certain point. Um so um I I don't think (00:04:09) I don't think you need that the (00:04:11) servicing thing is an issue. Um um but (00:04:13) you can mark my words. Uh in in 36 (00:04:17) months but probably closer to 30 months (00:04:20) the the most economically compelling (00:04:23) place to put AI will be space. Um and (00:04:25) then and and and then it will get from (00:04:28) it'll it'll then get like ridiculously (00:04:30) better to be in space. Um and then this (00:04:32) the scaling uh the only place you can (00:04:34) really scale is space. Um you know once (00:04:38) you start thinking in terms of uh what (00:04:40) percentage of the sun's power are you (00:04:41) harnessing uh you realize you have to go (00:04:43) to space uh you can't uh scale very very (00:04:47) much on earth (00:04:48) >> but by very much to be clear you're (00:04:50) talking like terowatts. (00:04:52) >> Yeah. Well all of the United States uh (00:04:55) currently uses only half a terowatt of (00:04:58) power on average. (00:04:59) >> Yeah. (00:04:59) >> Right. So, you know, if you say a (00:05:01) terowatt, that would be twice as much (00:05:03) electricity as the United States (00:05:04) currently consumes. So, that's quite a (00:05:06) lot. And can you imagine building that (00:05:08) many data centers? I that many power (00:05:11) plants. It's like those who have like (00:05:14) lived in software land uh don't realize (00:05:17) they're about to have a a hard lesson in (00:05:20) hardware. (00:05:21) uh that um there's there's it's actually (00:05:25) very difficult to build power plants and (00:05:27) and then you don't just need the you (00:05:28) need power plants, you need all of the (00:05:29) electrical equipment, you need the the (00:05:31) electrical transformers to run the (00:05:33) transformers, the AI transformers. Um (00:05:36) now the utility industry is a very slow (00:05:39) industry that they are they they pretty (00:05:42) much uh you know they impedance match to (00:05:44) the to the government to the the public (00:05:45) utility commission. Um so they're uh the (00:05:49) impedance smash like literally (00:05:50) figuratively. Um so they're very slow (00:05:53) because the their past has been very (00:05:56) slow. Um so trying to get them to move (00:05:58) fast is like you know like if you try to (00:06:00) do an interconnect agreement with have (00:06:01) you ever tried to do an internet (00:06:03) interconnect agreement with a utility at (00:06:05) scale like with a lot of power (00:06:06) >> as a professional podcaster I can say (00:06:08) that I have not in fact [laughter] (00:06:10) >> yeah they have to you need many more (00:06:12) views before that becomes an issue. (00:06:14) >> They have to do a study for a year. (00:06:16) Okay. at like a year later they'll come (00:06:18) back to you with their interconnect (00:06:19) study. (00:06:20) >> But can't you solve this with your own (00:06:22) behind the meter power stuff (00:06:24) >> you can build power plants. (00:06:26) >> Yeah, (00:06:26) >> that's what we did at XAI for classes 2. (00:06:29) So for for classes too (00:06:30) >> but so yeah why are we talking about the (00:06:31) grid? Why not just like build GPUs and (00:06:33) power colloccated? (00:06:34) >> That's what we did. (00:06:35) >> Right. Right. But I'm saying why isn't (00:06:36) this a generalized solution when you're (00:06:37) talking about all the issues? (00:06:38) >> Where do you get the power plants from? (00:06:39) >> I'm saying when you're talking about all (00:06:40) the issues working working with (00:06:41) utilities, you can just build private (00:06:42) power plants with the with the data (00:06:44) centers. (00:06:45) >> Right. But it begs the question of where (00:06:46) do you get the power plants? Where do (00:06:47) you where do you get the power plants (00:06:48) from? (00:06:49) >> I mean the power plant makers. (00:06:51) >> Oh, that's what you're saying. Like (00:06:52) there's the gas turbine backlog (00:06:54) basically. (00:06:55) >> Yes. It you can drill down to a level (00:06:58) further. It's the it's the the veins and (00:07:00) blades in the turbines um that are the (00:07:03) limiting factor because the the casting (00:07:06) may it's it's like a very specialized (00:07:07) process to cast the blades and veins in (00:07:10) the in the in the uh turbines using gas (00:07:13) power. Um and uh it's very it's very (00:07:16) difficult to scale other other forms of (00:07:18) power. You can scale potentially uh (00:07:20) solar but but the the tariffs currently (00:07:22) for importing solar in the US are (00:07:24) gigantic and the domestic solar (00:07:26) production is is pitiful. (00:07:28) >> Why not make solar? That seems like a (00:07:29) good Elon shaped problem. (00:07:31) >> We are going to make solar. (00:07:32) >> Okay. (00:07:33) >> Yeah. (00:07:34) >> Great. [laughter] (00:07:35) >> Both SpaceX and Tesla are are building (00:07:37) towards 100 gawatt of solar cell (00:07:40) production. How low down the stack like (00:07:42) from poly silicon up to the wafer to the (00:07:45) final um panel? (00:07:46) >> I think you got to do the whole thing (00:07:48) from raw materials to to to finish the (00:07:50) cell. Now, if it's going to space, it's (00:07:52) actually it costs it costs less and it's (00:07:54) easier to make solar cells that go to (00:07:56) space because they don't need glass or (00:07:58) they don't need much glass and they (00:08:00) don't need uh heavy framing because they (00:08:01) don't have to surv survive weather (00:08:03) events. There's no weather in space. So (00:08:06) is actually a cheaper solar cell that (00:08:08) goes to space than than is than the one (00:08:09) on the ground. (00:08:10) >> Is there a path to getting them as cheap (00:08:13) as you need in the next 36 months? (00:08:15) >> Solar cells are already very cheap. Um (00:08:19) they're like far sickly cheap. It's um (00:08:22) and if you say um you know I I think (00:08:25) like solar cells in China are around (00:08:27) like 2530 cents a watt or something like (00:08:29) that. It's it's absurdly cheap. And when (00:08:32) you when you take into account now now (00:08:34) put now put it in space and it's five (00:08:36) times cheaper because it's five times in (00:08:38) fact no it's not five times cheaper it's (00:08:40) 10 times cheaper because you don't need (00:08:41) any batteries. (00:08:44) So so the moment your cost of access to (00:08:46) space uh becomes low by far the cheapest (00:08:50) and most scalable way to generate to to (00:08:52) to generate tokens is space. It's not (00:08:56) even close. (00:08:59) it'll be an order of magnitude uh easier (00:09:01) to scale. Um and chips aside an order of (00:09:04) magnitude well if the point is you won't (00:09:06) be able to scale on the ground. It's (00:09:07) just you just won't. People are going to (00:09:09) hit the wall big time on power (00:09:10) generation. They already are. Um like so (00:09:14) so like the number of um sort of (00:09:17) miracles and series that the XAI team (00:09:18) had to accomplish in order to get a (00:09:20) gigawatt of power online uh was was was (00:09:22) crazy. We had to um gang together a (00:09:25) whole bunch of turbines um and uh and (00:09:28) then and then we had permit issues in um (00:09:32) Tennessee and and had to go across the (00:09:34) border to Mississippi, which is (00:09:35) fortunately only, you know, a few miles (00:09:37) away. Uh so, but then we still had to (00:09:39) run the high power lines a few miles and (00:09:42) and build a power plant in Mississippi. (00:09:43) Um and and it was very difficult to (00:09:45) build that. Um, and people don't (00:09:47) understand like how much how much (00:09:49) electricity do you actually need at the (00:09:51) generator level at the generation level (00:09:53) in order to power a data center because (00:09:55) they look at the the (00:09:57) noobs will look at the the the power (00:09:59) consumption of uh say a GB300 and and (00:10:02) multiply that by thing and then think (00:10:04) that's the amount amount of power you (00:10:05) need (00:10:05) >> all the cooling and everything. (00:10:06) >> Wake up. Yeah. This is like that's a (00:10:09) that's a that's a total noob. you've (00:10:12) never done any hardware in your life (00:10:13) before. (00:10:14) Besides the GB300, you got to power all (00:10:17) of the networking hardware. Um there's a (00:10:19) whole bunch of CPU and storage stuff (00:10:20) that's happening. Uh you you've got a (00:10:23) size for uh your your peak uh cooling (00:10:26) requirements. So that means uh can you (00:10:29) cool even on the the worst hours, the (00:10:31) worst day of the year? Well, it gets (00:10:33) pretty freaking hot in Memphis. So, so (00:10:36) you're going to have like a 40% increase (00:10:38) on your your power just for cooling. Um, (00:10:41) if assuming you don't want your data (00:10:43) center to turn off on hot days and and (00:10:46) want to keep going, then then you got to (00:10:48) say, well, uh, um, there's there's (00:10:50) another multiplicative element on top of (00:10:52) that, which is, are you assuming that (00:10:53) you're you you never have any hiccups in (00:10:56) your power generation? Like, oh, well, (00:10:58) actually, sometimes we have to take the (00:10:59) generators, some of the power offline in (00:11:01) order to service it. Oh, okay. Now you (00:11:03) add another 20 25% multiplier on that (00:11:06) because you you've got to you've got to (00:11:07) assume that that you've got to take (00:11:09) power offline to service it. Uh so the (00:11:12) actual (00:11:13) RS for roughly every every 110,000 GBs (00:11:18) GB300's inclusive of networking uh CPU (00:11:22) storage cooling uh margin for for for uh (00:11:26) servicing power uh is roughly uh 300 (00:11:31) megawatt. Sorry, say that again. (00:11:33) >> It's it's it's roughly or think about it (00:11:35) like the way you think about this like (00:11:38) 330,000 (00:11:39) to to to actually what you need at the (00:11:42) gener generation level to service (00:11:45) probably service 330,000 GB300s (00:11:48) including all of the associated support (00:11:50) networking and everything else and the (00:11:52) and and the peak cooling and to have (00:11:55) some margin some power margin reserve is (00:11:57) roughly a gawatt. (00:11:59) >> Can I ask a very naive question? Yeah. (00:12:00) Um uh you know you're describing the (00:12:04) engineering details of doing this stuff (00:12:06) on Earth. Um but then there's analogous (00:12:09) engineering difficulties of doing it in (00:12:11) space. How do you do the um uh how do (00:12:13) you replace infinite band with orbital (00:12:14) lasers etc etc. How do you make it (00:12:16) resistant to radiation? Um I don't know (00:12:20) the details in the engineering but (00:12:21) fundamentally what is the reason to (00:12:22) think those challenges which have never (00:12:25) been had to be addressed before will end (00:12:27) up being easier than just like building (00:12:29) more turbines on Earth. There's (00:12:31) companies that build turbines on Earth. (00:12:32) They can make more turbines, right? (00:12:34) >> I invite again try doing it and then (00:12:37) you'll see. Um so um (00:12:42) like the turbines are sold out through (00:12:44) 2030 (00:12:45) >> Have you guys considered making your (00:12:46) own? I think in in order for in order to (00:12:49) uh bring enough power online um I think (00:12:54) uh SpaceX and and Tesla will probably (00:12:56) have to make the turbine blades um the (00:12:59) mains and blades uh internally (00:13:03) >> for just the blades or the turbines. (00:13:04) >> Uh uh the the the the limiting factor (00:13:07) you can get everything except the the (00:13:09) blades they call the blades and veins. (00:13:12) Um you can get that uh 12 to 18 months (00:13:17) before the veins of blades. The limiting (00:13:18) factor veins and blades and there are (00:13:20) only uh three casting uh companies in (00:13:24) the world that make make these and (00:13:26) they're massively backlogged. (00:13:28) >> Is this Seaman's GE those guys or is it (00:13:30) a sub? (00:13:31) >> No, it's it's it's it's other companies. (00:13:33) I mean, sometimes they have a little bit (00:13:34) of casting capability in house, but uh (00:13:36) I'm just saying you can just you can (00:13:38) just call any of the turbine makers and (00:13:40) they will tell you. It's not top secret. (00:13:42) They probably on the it's probably on (00:13:43) the internet right now. (00:13:44) >> If if it wasn't for the tariffs, would (00:13:46) uh would Colossus be solar powered? (00:13:48) >> Uh it would be much easier to make it (00:13:50) solar powered. Yeah. Um the tariffs are (00:13:53) nuts. Several hundred%. So (00:13:55) >> don't you know some people (00:13:56) >> we we also need speed. Yeah. you know, (00:13:58) [laughter] (00:13:59) you know, um, president has, (00:14:03) you know, we don't agree on everything. (00:14:04) Um, and, um, this administration is not (00:14:07) not the biggest fan of of solar. (00:14:10) >> Um, [laughter] (00:14:13) but you (00:14:15) >> and you also need the land, the permits (00:14:17) and everything. So, if you're trying to (00:14:18) move very fast, (00:14:20) >> um like I do think scaling solar on (00:14:24) Earth is a is a good way to go, (00:14:26) >> but but you need you do need some amount (00:14:28) of time to find the land, get the (00:14:30) permits, get the solar, uh pair that (00:14:32) with the batteries. (00:14:33) >> But why would it not work to stand up (00:14:36) your own solar production and then (00:14:38) you're right that you eventually run out (00:14:39) of land, but there's a lot of land here (00:14:41) in Texas. There's a lot of land in (00:14:42) Nevada, including private land. It's not (00:14:44) all publicly owned land. And so you'd be (00:14:46) able to at least get the next Colossus (00:14:48) and like the next one after that and at (00:14:50) a certain point you hit a wall. But (00:14:51) wouldn't that work for the moment? (00:14:53) >> As I said, we are scaling solar (00:14:55) production. Um there's there's a rate (00:14:58) there's a rate at which you can scale (00:15:00) physical production of solar solar (00:15:02) cells. We we're going as fast as (00:15:05) possible in scaling domestic production. (00:15:07) >> You're making the solar cells at Tesla. (00:15:09) Both Tesla and SpaceX um have a mandate (00:15:12) to get to 100 gawatt a year of of solar. (00:15:15) >> Speaking of the annual capacity, I'm (00:15:17) curious in 5 years time, let's say, what (00:15:19) will the installed capacity be on Earth? (00:15:24) >> Long time and in space. Yeah, I (00:15:26) deliberately picked five years cuz it's (00:15:27) after your once we're up and running (00:15:29) threshold. And so in five years time, (00:15:31) yeah, what's the on earth versus in (00:15:33) space installed AI capacity? five years (00:15:36) I think probably (00:15:39) if you say 5 years from now we're (00:15:40) probably um (00:15:44) AI in space will be uh launching every (00:15:49) year the the sum total of all AI on (00:15:52) earth in excess meaning 5 years from now (00:15:55) my prediction is we will launch and and (00:15:58) and be operating (00:16:01) every year more AI in space than than (00:16:04) this than the cumulative sort of total (00:16:05) on earth which is I would expect to be (00:16:08) at least sort of 5 years from now a few (00:16:11) hundred gawatts per year (00:16:14) of uh of AI in space (00:16:17) >> and rising. Um so you can get to I think (00:16:22) you on Earth you can get to around a (00:16:24) terowatt a year of of AI in space um (00:16:29) before you start having you know fuel (00:16:32) supply challenges for the rocket. (00:16:34) >> Okay. You think you can get to hundreds (00:16:35) of gigawatts per year in 5 years time? (00:16:38) >> Yes. (00:16:38) >> So 100 gawatt depending on the um (00:16:42) specific power of the whole system with (00:16:45) solar arrays and radiators and (00:16:46) everything is um is on the order of like (00:16:49) 10,000 Starship launches. (00:16:51) >> Yes. (00:16:52) >> Um and you want to do that in one year. (00:16:55) And so that's like one Starship launch (00:16:56) every hour. (00:16:58) >> Yeah. (00:16:58) >> That's happening in this city. Like walk (00:17:02) me through a world where there's 10 (00:17:03) there's a Starship launch every single (00:17:05) hour. (00:17:05) >> Yeah. I mean that's actually a lower (00:17:07) rate compared to airlines. Uh like like (00:17:09) aircraft aircraft (00:17:10) >> there's a lot of airports (00:17:11) >> a lot of airports but (00:17:12) >> and you got to launch you know the polar (00:17:14) orbit. (00:17:15) >> Uh no it doesn't have to be polar but (00:17:17) you you just there's there's some some (00:17:21) value to sunsynchronous but um but I I (00:17:24) think actually um you just go high (00:17:27) enough you you start getting out of (00:17:28) earth shadow you know. So, um, (00:17:31) >> how many physical Starships are needed (00:17:33) to do 10,000 launches a year? (00:17:35) >> I I don't think we'll need more than I (00:17:37) mean, you you could you could uh (00:17:39) probably do it with (00:17:41) as as few as like 20 or 30. Um, (00:17:46) >> like it really depends on how quickly (00:17:48) does a ship the ship has to go around (00:17:50) the Earth. Um, and the ground track for (00:17:53) for the ship has to come back over the (00:17:55) launch pad. So, if you can use a ship (00:17:58) every say 30 hours, uh you could do it (00:18:00) with 30 ships, but but we'll we'll make (00:18:03) more ships than that. But, um but but (00:18:07) SpaceX is is um is going up to do 10,000 (00:18:10) launches a year and I'll and and maybe (00:18:13) even 20 or 30,000 launches a year. (00:18:15) >> Is the idea to become basically a (00:18:17) hyperscaler, become an oracle and lend (00:18:19) this capacity to other people? What's (00:18:20) what are you going to do with presumably (00:18:22) SpaceX is the one launching all this? (00:18:25) So SpaceX is going to hyperscaler (00:18:28) >> hyper hyper. [laughter] (00:18:30) >> Yeah. I mean if ass if assuming my (00:18:31) predictions come true SpaceX will launch (00:18:34) more AI than the cumulative amount on (00:18:37) Earth combi of everything else combined. (00:18:39) >> Is this mostly inference or (00:18:41) >> most AI will be inference like already (00:18:43) inference for the purpose of training is (00:18:44) most training. And there's a narrative (00:18:48) that the (00:18:50) the change in discussion around a SpaceX (00:18:52) IPO is because previously SpaceX was (00:18:55) very capital efficient. Just it wasn't (00:18:58) that expensive to develop that. Even (00:19:00) though it sounds expensive, it's (00:19:02) actually very capital efficient in how (00:19:03) it runs. Whereas now you're going to (00:19:06) need more capital than just can be (00:19:09) raised in the private markets. Like if (00:19:10) the private markets can accommodate (00:19:12) raises of, as we've seen from the AI (00:19:13) labs, tens of billions of dollars, but (00:19:15) not beyond that, is it that you'll just (00:19:18) need more than tens of billions of (00:19:19) dollars per year and that's about to (00:19:21) take it public? (00:19:22) >> Um, yeah. I have to be careful about (00:19:25) saying things about companies that might (00:19:26) go public. Um, you know, (00:19:29) >> if you make general state, [laughter] if (00:19:30) you make (00:19:31) >> that's never been a problem for you, (00:19:32) Elon, (00:19:34) >> you know, there's a price to pay for (00:19:35) these things. (00:19:36) make some general statements for us (00:19:38) about the depth of the capital markets (00:19:40) between public and private markets. (00:19:42) >> Yeah, there's there's a lot more capital (00:19:44) in the (00:19:45) >> very general [laughter] (00:19:47) >> there's there's obviously a lot more (00:19:48) capital available in the public markets (00:19:50) than private. I mean it might be (00:19:52) it's at least at least it might be 100 (00:19:54) times more capital but it's at least (00:19:56) well way more than 10. But isn't it also (00:19:58) the case that things that tend to be (00:20:00) very um capital intensive if you look at (00:20:04) say real estate as you know a huge (00:20:07) industry uh that raises a lot of money (00:20:09) each years at an industry level that (00:20:11) tends to be debt financed because by the (00:20:14) time you're deploying that much money (00:20:15) you actually have a pretty (00:20:18) >> you have a clear revenue stream. (00:20:19) >> Exactly. And and a near-term return. And (00:20:21) you see this even with the data center (00:20:22) buildouts which are famously being you (00:20:24) know uh financed by the uh the private (00:20:26) credit industry. And so why not just (00:20:29) debt finance? (00:20:31) Um speed is important. So (00:20:38) um I'm genally going to do the thing (00:20:39) that um I'm I'm I mean I just repeatedly (00:20:43) tack the limiting factor. Whatever the (00:20:44) limiting factor is on speed, I'm I'm (00:20:46) going to tackle that. So um there's uh (00:20:51) if if capital is the limiting factor (00:20:53) then I'll I'll sold for capital. If if (00:20:54) it's not limiting factor I'll s for (00:20:56) something else. (00:20:56) >> U based on your statements about um (00:20:59) Tesla and being public. I wouldn't have (00:21:03) guessed that you thought the fast the (00:21:05) way to move fast is to be public. (00:21:08) >> Normally I would say yeah that's that's (00:21:10) true. Um, like I said, I mean, I'd love (00:21:12) to, you know, talk about some more in (00:21:13) detail, but the problem is like if you (00:21:15) talk about poly companies before they (00:21:17) become public, you get into trouble and (00:21:18) then you have to delay your offering (00:21:21) >> and then you (00:21:22) >> and as we said, we're solving for speed. (00:21:24) >> Yes. Exactly. So, so, so that you you (00:21:27) can't hype companies um that are that (00:21:30) may that might go public. So, that (00:21:32) that's that's why we have to be a little (00:21:34) careful here. Um, but but but I I I we (00:21:38) can't talk about physics. Um so like the (00:21:41) way the way you think about scaling long (00:21:43) term is that um uh earth only receives (00:21:46) about uh half a billionth of the sun's (00:21:48) energy. Um and the sun is the sun is (00:21:51) essentially all the energy. This is a (00:21:53) very important point to appreciate (00:21:54) because sometimes people will talk about (00:21:56) marginal nuclear reactors or any you (00:21:59) know various like fusion on earth. Um (00:22:02) but but you have to step back a second (00:22:04) and say if if if you're going to climb (00:22:06) the cartes scale and have some (00:22:09) non-trivial and and harness some (00:22:11) non-trivial percentage of the uh the (00:22:13) sun's energy like let's say you wanted (00:22:15) to uh harness a millionth of the sun's (00:22:18) energy which sounds pretty small. um (00:22:22) that that would be um about call it (00:22:25) roughly 100,000 times more electricity (00:22:28) than we currently generate on Earth of (00:22:31) for all of civilization (00:22:34) uh give or take an order of magnitude. (00:22:36) Um so it it obviously the only way to (00:22:39) scale uh is to go to space with solar. (00:22:42) Uh from launching from Earth you can get (00:22:44) to about a terowatt per year. Um beyond (00:22:47) that you want to go you you want to uh (00:22:49) launch from the moon. and you want to (00:22:50) have a a mass driver on the moon. Uh and (00:22:53) that mass driver on the moon, you could (00:22:56) do probably a pedawatt per year. (00:22:59) >> Um when you're talking these kinds of (00:23:00) numbers, you know, terowatts of compute, (00:23:03) um presumably whether you're talking (00:23:04) land or space far before this point, um (00:23:10) you've like run into, you know, you (00:23:12) actually need maybe you don't the solar (00:23:14) panels are more efficient, but you still (00:23:15) need the chips. (00:23:16) >> Uh you still need the logic and the (00:23:18) memory and so forth and need (00:23:20) [clears throat] to build a lot more (00:23:21) chips and make them much cheaper, (00:23:23) >> right? And so how are we getting a (00:23:24) terowatt of uh like right now the world (00:23:26) is going to be 20 25 gawatt of compute? (00:23:29) Um how are we getting a terowatt of (00:23:30) logic by 2030? (00:23:33) >> I guess we're going to need some very (00:23:34) big chip apps. (00:23:36) >> Tell tell me about it. [laughter] (00:23:37) >> I've mentioned publicly that uh the idea (00:23:40) of doing a sort of a a terab tering the (00:23:44) new gig. We I feel like the naming (00:23:47) scheme of Tesla, which has been very um (00:23:49) catchy, is like you looking at like the (00:23:51) metric [laughter] (00:23:52) >> the metric scale. Um at what level of (00:23:55) the stack are you are you building the (00:23:57) clean room and then partnering with an (00:23:59) existing um fab to get the process (00:24:02) technology and buying the tools from (00:24:03) them? What what is the plan there? (00:24:06) >> Well, you can't partner with existing (00:24:07) fabs because uh they just they can't (00:24:09) output enough their chip volume is too (00:24:11) low. (00:24:12) >> But you have you have to (00:24:13) >> before the process technology. Yeah. (00:24:14) Partner for the IP. (00:24:16) >> Um, you know, the the fabs today all (00:24:19) basically use um machines from like five (00:24:22) companies. (00:24:23) >> Yeah. (00:24:24) >> You know, so (00:24:26) you've got ASML, Tokyo, Electron, Kelly, (00:24:29) Tankor, you know, um, etc. So, um, so, (00:24:36) so at first I think you'd have to get (00:24:39) equipment from them and then, uh, modify (00:24:42) it or work with them to increase the (00:24:44) volume. Um, but I think you'd have to (00:24:46) build perhaps in a different way. Um, so (00:24:48) I think the logical thing to do is to, (00:24:51) uh, to use conventional equipment in (00:24:53) unconventional way to get to scale. Uh (00:24:55) and then uh and then and then start (00:24:58) modifying the equipment uh to increase (00:25:00) the the rate (00:25:01) >> kind of boring company style. (00:25:03) >> Yeah. Kind of like Yeah. You you sort of (00:25:06) buy an an existing uh boring machine and (00:25:09) then uh figure out how to dig tunnels in (00:25:12) the first place and then design a much (00:25:14) better machine uh that's you know I (00:25:18) don't know (00:25:19) >> some orders of magnitude faster. (00:25:21) >> Here's a very simple lens. We can (00:25:23) categorize technologies and how hard (00:25:25) they are. And one categorization could (00:25:26) be look at things that China has not (00:25:29) succeeded in doing. And if you look at (00:25:31) Chinese manufacturing, (00:25:33) still behind on leading edge chips and (00:25:37) still behind on uh leading edge turbine (00:25:40) engines and things like that. And so (00:25:43) does the fact that China has not (00:25:46) successfully replicated TSMC give you (00:25:48) any pause about the difficulty or you (00:25:50) think (00:25:52) >> well that's not true for some reason. (00:25:53) >> Uh it's not that they have not (00:25:54) replicated TSMC they have not replicated (00:25:57) ASML that's the limiting factor. (00:25:59) >> So so you think it's just the um the (00:26:02) sanctions essentially. (00:26:03) >> Uh yeah China would be outputting vast (00:26:05) numbers of chips at they could buy (00:26:08) >> but couldn't they up to relatively (00:26:10) recently buy them? No. (00:26:12) >> Okay. (00:26:13) >> That that ASML bands have been in place (00:26:15) for a while. (00:26:15) >> Okay. (00:26:16) >> So, but I I think T's going to be make (00:26:17) start making pretty compelling chips on (00:26:19) three or four years. (00:26:20) >> Would you consider making to ASML (00:26:21) machines? (00:26:22) >> I I don't know. I don't know yet is the (00:26:24) right answer. So I (00:26:27) um it's just that that (00:26:30) if to produce at high volume and to to (00:26:33) to reach large volume in say 36 months (00:26:37) to match the the rocket to payload to (00:26:41) orbit. So if we're doing a million tons (00:26:42) to orbit (00:26:44) um in like let's say I don't know (00:26:48) 3 or four years from now something like (00:26:50) that. um that and uh and and we're doing (00:26:54) 100 kow per ton. So that that means we (00:26:58) need um at least 100 gawatts per year of (00:27:01) solar. Um and we'll need uh an (00:27:05) equivalent amount of of chips to to you (00:27:08) know that you need 100 gawatt worth of (00:27:10) chips. You you're going to match these (00:27:12) things the master orbit. (00:27:13) >> Yes. (00:27:13) >> The the power generation and the uh and (00:27:16) the and the chips. uh and and and I'd (00:27:19) say my biggest concern actually is is (00:27:20) memory. Um so the I think there's (00:27:24) there's a (00:27:26) the the path to creating logic chips is (00:27:29) more obvious than the path to um having (00:27:32) sufficient memory to support logic (00:27:33) chips. That's why you see your DDR (00:27:36) prices going in and these memes about (00:27:38) like um you know um you're marooned on a (00:27:41) desert island. You write help me on the (00:27:43) sand. Nobody comes. You write DDRM (00:27:46) ships come swarming in. [laughter] (00:27:50) >> I haven't seen that. (00:27:52) >> Uh I love your manufacturing philosophy (00:27:54) around um around fabs. You know I know (00:27:58) nothing about the topic but (00:27:59) >> I don't know how to build a fab yet. I (00:28:00) figure it out. [laughter] (00:28:02) >> But (00:28:02) >> obviously I have difficult. (00:28:03) >> It sounds like you think the the sort of (00:28:05) like the processing knowledge of like (00:28:06) these 10,000 PhDs in Taiwan who know (00:28:09) exactly what gas goes in the plasma (00:28:12) chamber and what settings to put on the (00:28:13) tool. You can just like delete those (00:28:15) parts of those steps. Like fundamentally (00:28:17) it's get the clean room, get the tools (00:28:19) and figure it out. I don't think it's (00:28:21) PhDs that it's it's mostly people with (00:28:24) uh you know who are not not PhDs (00:28:27) um that that most engineering is done (00:28:29) with people who don't have PhDs. Do you (00:28:31) guys have PhDs? (00:28:32) >> No. (00:28:32) >> Okay. (00:28:34) >> We [laughter] also we also haven't (00:28:35) successfully built any fab so you (00:28:37) shouldn't be coming to us for your fab (00:28:38) advice (00:28:39) >> or (00:28:39) >> I don't think you need PhD for that for (00:28:41) the stuff. So um but but you do need you (00:28:43) do need competent personnel. Um so I I (00:28:46) don't I mean like like right now if um (00:28:50) you know say like Tesla's pedals to the (00:28:52) metal max production of going as fast as (00:28:55) possible to get uh AI5 Tesla AI5 chip (00:28:58) design um uh into production and then (00:29:01) reaching scale. Um you know that'll (00:29:03) probably happen you know around the (00:29:07) second quarterish of next year (00:29:09) hopefully. Um (00:29:12) uh and then (00:29:13) AI6 would hopefully follow less than a (00:29:16) year later. Um but um and and and and (00:29:20) we've secured all the all the trip fab (00:29:24) production that we can. (00:29:25) >> Yes. You're currently limited on TSMC (00:29:28) fab capacity. (00:29:29) >> Yeah. Um and and and we'll be using TSMC (00:29:33) uh Taiwan, uh Samsung Korea, TSMC (00:29:37) Arizona, Samsung Texas. Um and we still (00:29:41) booked out all the Yeah. (00:29:42) >> Yes. And and then and then and then if I (00:29:44) ask uh TSMC or Samsung, okay, what (00:29:47) what's the time frame to get to volume (00:29:50) production? The point is not is it's not (00:29:51) you've got to you've got to build the (00:29:52) fab (00:29:53) >> and you got to you got to start (00:29:56) production, then you got to climb the (00:29:57) yield curve and reach volume production (00:29:58) at high yield. that that that from start (00:30:01) finish is a 5year period. And so the (00:30:03) limiting factor is chips. (00:30:05) >> Yeah. (00:30:07) Like limiting factor once you can get to (00:30:08) space is chips, but the limiting (00:30:10) limiting factor before you can get to (00:30:11) space will be power. (00:30:13) >> Why don't you do the Jensen thing and (00:30:14) just prepay TSMC to build more fabs for (00:30:16) you? (00:30:16) >> Uh I I've already told them that. (00:30:20) >> But they won't take your money. Like (00:30:21) what's going on? (00:30:22) >> They're building fabs as fast. No, (00:30:25) [laughter] (00:30:26) >> they're building they're building fabs (00:30:27) as fast as they can. (00:30:30) Um, and so is Samsung. Like like they're (00:30:33) they're pedal to the metal. I mean, (00:30:34) they're going, you know, balls to wall, (00:30:37) you know, (00:30:38) as fast as they can. So, still not fast (00:30:41) enough. I mean, like said, there will be (00:30:44) I think um if you say uh I think towards (00:30:48) the end of this year, I think probably (00:30:49) chip production will outpace the ability (00:30:51) to turn chips on. Uh but once you can (00:30:55) get to space and unlock the um the power (00:30:59) constraint and you can now do you know (00:31:01) hundreds of gigawatts per year of power (00:31:03) in space um again bearing in mind that (00:31:06) average power usage in the US is you (00:31:08) know 500 gaw so if you're launching say (00:31:11) 200 gawatt a year to to space you're (00:31:14) sort of lapping the US every two and a (00:31:15) half years the entire all US electricity (00:31:18) production this is a very huge amount um (00:31:22) so Um but but but between now and then (00:31:26) uh the the actually the the constraint (00:31:29) for for for server side compute uh (00:31:32) concentrated compute will be will be (00:31:34) electricity. (00:31:36) My my guess is that we start hitting the (00:31:39) people start getting for where they (00:31:40) can't turn the chips on for for for (00:31:44) large clusters uh towards the end of (00:31:46) this year. They're just the chips are (00:31:47) going to be piling up and and not be (00:31:49) won't be able to be turned on. Now for (00:31:52) edge computers a different story. So if (00:31:54) the if if like for for Tesla the the so (00:31:57) the AI5 chip is going into our Optimus (00:31:59) robot you know optimistic um and and so (00:32:04) if you have an AI edge compute that's (00:32:07) distributed power now the power is (00:32:09) distributed over a large area it's not (00:32:12) concentrated um and if you can charge at (00:32:14) night you can actually um [snorts] uh (00:32:18) use the grid much more effectively (00:32:20) because the the actual peak power (00:32:22) production in the US is over 1,000 (00:32:24) gawatt. Uh but the average power usage (00:32:27) because the dayight cycle is 500. So if (00:32:30) you can charge at night, there's an (00:32:31) incremental 500 gaw that you can um (00:32:35) generate you know at night. Um so that (00:32:39) that's why Tesla for edge compute is not (00:32:42) constrained and we can make a lot of (00:32:44) ships uh to make you know very large (00:32:48) number of robots and cars. Uh, but if (00:32:50) you try to concentrate that compute, (00:32:52) you're going to have a lot of trouble (00:32:53) turning it on. (00:32:54) >> What I find remarkable about the SpaceX (00:32:57) business is the end goal is to get to (00:32:59) Mars, but you keep finding ways on the (00:33:03) way there to keep generating incremental (00:33:06) revenue to get to the next stage and the (00:33:08) next stage. So, the Falcon 9 is Starlink (00:33:11) >> and now for Starship, it's going to be (00:33:13) potentially orbital data centers. Um but (00:33:16) like you find these like um you know (00:33:19) sort of infinitely uh elastic sort of (00:33:21) marginal use cases of your like next (00:33:23) rocket and your next rocket and next (00:33:25) scale up. (00:33:28) >> You can see how this might seem like a (00:33:29) simulation to me. [laughter] (00:33:31) >> Well, (00:33:32) >> or am I someone's avatar in a video game (00:33:34) or something because it's like like what (00:33:36) are the odds that all these crazy things (00:33:37) would be happening? I I I mean I mean I (00:33:42) mean rockets and chips and robots and (00:33:47) space solar power and and not to mention (00:33:49) the the mass driver on the moon. I (00:33:51) really want to see that. You can imagine (00:33:52) like some mass driver that's just like (00:33:56) just it's like sending AI solar powered (00:34:00) AI satellites into space like one after (00:34:01) another like these like at at 2 and a (00:34:04) half kilometers per second. you know, (00:34:06) that's uh (00:34:09) and just shooting them into deep space. (00:34:12) That would be a sight to see. (00:34:15) I I I mean, I'd watch that (00:34:18) >> just like a live stream of (00:34:19) >> Yeah. Yeah. Just one after another just (00:34:22) shooting (00:34:22) >> webcam (00:34:23) >> uh AI satellites in deep space, you (00:34:26) know, a billion or 10 billion tons a (00:34:28) year. (00:34:28) >> And sorry, you manufacture the (00:34:29) satellites on the moon. I see. So you (00:34:31) send the raw materials to the moon and (00:34:33) then manufacture them there and then (00:34:34) >> well the the lunar soil is uh I think (00:34:37) it's like 20% solar 20% silicon or (00:34:39) something like that. So so you can get (00:34:41) the silicon from the (00:34:42) >> you can mine the silicon on the moon (00:34:44) refine it um (00:34:45) >> and generate the and create the solar (00:34:47) panels or the solar cells and the (00:34:49) radiators on the moon. (00:34:50) >> Yeah. (00:34:51) >> So um make the radiators out of (00:34:53) aluminum. So there's there's plenty of (00:34:55) silicon and aluminum on the moon to uh (00:34:56) to make the the cells on the and the (00:34:59) radiators. Um, the chips you could send (00:35:01) from Earth because they're pretty light. (00:35:03) Um, but maybe at some point you make (00:35:04) them on the moon, too. I'm just saying (00:35:06) like these are simply (00:35:08) it's kind of like, like I said, it it (00:35:10) does seem like a sort of a a video game (00:35:12) situation where it's difficult but not (00:35:14) impossible to get to the next level. (00:35:17) um like I I I don't see any way that you (00:35:19) could do um you know uh (00:35:24) you know 500 to a,000 terowatts per year (00:35:28) launch from Earth. [snorts] (00:35:30) U (00:35:31) >> I agree [laughter] (00:35:34) >> but you could do that from the moon. (00:35:36) >> Okay, let me tell you how I ended up (00:35:38) using Mercury for my personal banking. (00:35:40) So last year I had the opportunity to (00:35:42) make an investment that I was very (00:35:44) excited about, but it came up a bit last (00:35:46) minute. And so I had to wire over a lot (00:35:48) of money for my personal account very (00:35:50) fast. But my personal bank at the time (00:35:52) wouldn't let me make this wire transfer (00:35:54) online. And I called them a bunch of (00:35:55) times. They just couldn't make it work. (00:35:57) They told me that I'd have to go to the (00:35:59) nearest Inerson branch, which was in (00:36:01) Dallas. And for a moment, I even (00:36:03) considered flying from SF to Dallas to (00:36:05) make this transfer happen last minute. (00:36:07) But then I remembered that Mercury, (00:36:09) which I use for my business banking, had (00:36:11) just started rolling out personal (00:36:12) accounts. So I emailed support with a (00:36:14) quick rundown of the situation. And (00:36:16) within 2 hours, I had successfully wired (00:36:18) the investment from my new personal (00:36:21) Mercury account. Since then, I've moved (00:36:23) over the rest of my personal money from (00:36:24) my previous bank to Mercury, and that's (00:36:27) made a bunch of things, even little (00:36:28) things like setting up auto transfer (00:36:30) rules between my checkings and savings (00:36:31) account, a whole lot better. Visit (00:36:34) mercury.com/personal (00:36:36) to get started. Mercury is a fintech (00:36:38) company, not an FDIC insured bank. (00:36:41) Banking services provided through Choice (00:36:43) Financial Group and column NA members (00:36:45) FDIC. Can can I can I zoom out and ask (00:36:48) about the SpaceX mission? So, I think (00:36:51) you said like we got to get to Mars so (00:36:52) we can make sure that if something (00:36:53) happens to Earth, [snorts] you know, (00:36:56) civilization consciousness, etc. (00:36:57) arrives. Yes. (00:36:58) >> Um, (00:36:59) >> by the time you're sending stuff to (00:37:00) Mars, like Grock is on that ship with (00:37:01) you, right? Right. And so if Grock's (00:37:03) gone Terminator, like the main risk (00:37:05) you're worried about, which is AI, why (00:37:06) doesn't that follow you to Mars? (00:37:08) >> Uh well, I'm not sure AI is the main (00:37:09) risk I'm worried about. I mean, the (00:37:12) important thing is that uh consciousness (00:37:14) uh which (00:37:16) I think arguably most consciousness or (00:37:19) most intelligence certainly (00:37:20) consciousness is more of a debatable (00:37:22) thing. Most intell the vast majority of (00:37:23) intelligence in the future will be um (00:37:25) AI. Um so um you know AI AI will exceed (00:37:33) uh you say like how many what's the how (00:37:36) much how many I don't know pedawatts of (00:37:39) intelligence will be uh silicon versus (00:37:42) biological (00:37:44) and and and basically humans will be a (00:37:46) very tiny percentage of all intelligence (00:37:48) in the future if current trends (00:37:50) continue. Um anyways as as long as like (00:37:53) I think there's (00:37:55) intelligence ideally ideally also which (00:37:59) includes human intelligence and (00:38:00) consciousness propagated into the future (00:38:02) that's a good thing. So you want to take (00:38:03) the set of actions that maximize the (00:38:05) probable uh light cone of of of (00:38:08) consciousness. So just and intelligence (00:38:10) >> just to be clear it's a the mission of (00:38:12) SpaceX is that even if something happens (00:38:16) to the humans the AIs will be on Mars (00:38:19) and like the AI intelligence will (00:38:21) continue the light of our journey. (00:38:24) >> Yeah I mean I'm very prohuman so it's (00:38:27) not I I want to make sure we take sort (00:38:28) of actions that ensure that humans are (00:38:31) along for the ride. You know we're at (00:38:33) least there. (00:38:33) >> Yeah. Um but the I'm just saying the (00:38:36) total amount of intelligence uh like I (00:38:40) think maybe in in five or six years um (00:38:44) AI will exceed the sum of all human (00:38:46) intelligence and then if that continues (00:38:48) at some point human intelligence will be (00:38:50) less than 1% of all intelligence. What (00:38:53) what should our goal be for such a (00:38:54) civilization? Is the idea that a small (00:38:56) minority of humans still have control of (00:38:58) the AIs? Is the idea of some sort of (00:39:00) like just trade but no control? How (00:39:03) should we think about the relationship (00:39:04) between the vast stocks of AI population (00:39:06) versus human population (00:39:08) >> in the long run? I think I I I don't (00:39:10) it's difficult to imagine that if humans (00:39:13) have say 1% of the intelligence of (00:39:18) combined intelligence of artificial (00:39:19) intelligence that that that humans will (00:39:21) be in charge of AI. Um, I think what we (00:39:25) can do is make sure it has um that AI (00:39:27) has values that that are um that that (00:39:30) cause intelligence to be propagated uh (00:39:33) into the universe. Um so the the the (00:39:37) reason for (00:39:39) XI XI's mission is understand the (00:39:42) universe. (00:39:44) So now that's actually very important. (00:39:46) So you say well what things are (00:39:48) necessary to understand the universe? (00:39:50) Well, you have to be curious and you (00:39:51) have to exist. (00:39:53) you can't just can't understand the (00:39:54) universities don't exist. Um so you (00:39:57) actually want to increase the amount of (00:39:58) intelligence uh in the universe increase (00:40:00) the palable lifespan of intelligence the (00:40:03) scope and scale of intelligence. Um I (00:40:06) think actually also as a coral you corly (00:40:09) you have um humanity also uh continuing (00:40:13) to expand because um if you're if you're (00:40:15) curious you're trying to understand the (00:40:16) universe one thing you're trying to (00:40:18) understand is where will humanity go (00:40:20) >> and so I think understand the universe (00:40:22) actually means you would care about uh (00:40:23) propagating humanity into the future um (00:40:27) and uh so so that's that's why I think I (00:40:31) think our mission station is profoundly (00:40:34) important Um I'm not sure to the degree (00:40:36) that Grock adheres to that mission (00:40:38) statement um I I think the future will (00:40:40) be very good. (00:40:41) >> I I want to ask about how to make Grock (00:40:43) adhere to that mission statement. But at (00:40:44) first I want to understand the mission (00:40:45) statement. Um so it's there's it's (00:40:48) there's understanding the universe. (00:40:50) They're spreading intelligence and (00:40:52) they're spreading humans. Um all three (00:40:56) seem like distinct vectors. (00:40:58) >> Okay. Well, I'll tell you why I why I (00:41:00) think they are that that that (00:41:02) understanding the universe encompasses (00:41:03) all of all of those things. Go ahead. (00:41:04) >> Um, you can't have understanding without (00:41:08) I think you can't have understanding (00:41:09) without intelligence and and I think (00:41:11) without consciousness. Um, so you in (00:41:14) order to understand universe, you have (00:41:15) to expand this the the scale and and (00:41:20) probably the scope of of intelligence (00:41:22) different types of intelligence. (00:41:24) >> I guess from a humanentric perspective (00:41:26) like for humans in comparison to (00:41:28) chimpanzees, humans are trying to (00:41:29) understand the universe. They're not (00:41:31) like expanding chimpanzeee footprint or (00:41:33) something, right? (00:41:34) >> We're also we're also not well we're not (00:41:36) we actually have made protected zones (00:41:38) for chimpanzees. Um and even though we (00:41:40) could humans could exterminate all (00:41:42) chimpanzees, we've not we've chosen not (00:41:43) to do so. (00:41:44) >> Do you think that's a basic scenario for (00:41:45) humans in the post AGI world? (00:41:48) >> Um (00:41:51) I I I think uh I think AI with the right (00:41:56) values I think Grock would care about (00:41:58) expanding uh human civilization. I'm (00:42:00) gonna certainly emphasize that. Hey, (00:42:02) Gragas, your daddy. (00:42:05) Don't forget to expand human (00:42:08) consciousness. Uh like I I actually I (00:42:11) think if if probably like uh like the (00:42:15) Yan Banks culture books are the closest (00:42:18) thing to what what the future will be (00:42:20) like in a you know non-dystopian (00:42:22) outcome. Um so I so understand the (00:42:27) universe it means you have to be very (00:42:29) you have to be truth seeking as well. (00:42:30) You like truth has to be absolutely (00:42:32) fundamental because you can't understand (00:42:33) the universe if you live if you're (00:42:35) delusional. You you'll simply think you (00:42:37) understand understood the universe but (00:42:38) you will not. So so being rigorously (00:42:41) truth seeeking is is absolutely (00:42:43) fundamental to understanding the (00:42:44) universe. You're not going to discover (00:42:45) new physics or or invent technologies (00:42:47) that work um unless you're rigorously (00:42:50) truth seeeking. How do you make sure (00:42:51) that Grock is regressively truth (00:42:52) seeeking as it gets smarter? (00:42:55) >> Uh (00:43:00) I think you you need to make sure that (00:43:02) that that Grock um is says things that (00:43:05) are correct not politically correct. I (00:43:07) think it's the elements of cogency. So (00:43:09) you want to make sure that that the (00:43:10) axioms are as close to true as possible (00:43:13) that that you don't have contradictory (00:43:15) axioms. um that the um the conclusions (00:43:19) necessary necessarily follow from those (00:43:21) axioms with with the right probability. (00:43:23) It it's just it's just it's critical (00:43:25) thinking 101. (00:43:27) I I think at least trying to do that is (00:43:29) better than not trying to do that. (00:43:31) >> Yeah. (00:43:31) >> And the proof will be in the pudding. If (00:43:33) if like I said for any AI to discover (00:43:35) new physics or invent technologies that (00:43:37) actually work in reality and there's no (00:43:38) bullshitting physics. So it's like you (00:43:41) can you know you can um you can break a (00:43:44) lot of laws but you can't like your (00:43:47) physics is law everything else is is a (00:43:49) recommendation (00:43:51) like in order to make a technology that (00:43:52) works you have to be extremely truth (00:43:56) seeeking because otherwise you'll test (00:43:58) that technology against reality um and (00:44:00) if you make for example an an error in (00:44:03) your rocket design the rug will blow up (00:44:05) um or the car won't work or the you know (00:44:08) >> but but there there are a lot of um (00:44:10) communist Soviet physicists who or like (00:44:13) scientists discovered new physics. There (00:44:16) are German Nazi physicists who (00:44:18) discovered new uh science. Um it seems (00:44:20) possible to be like really good at (00:44:22) discovering new science and be really (00:44:24) truth seeeking in that one particular (00:44:25) way. And still we'd be like well I don't (00:44:27) want I don't want the communist (00:44:29) scientist to like become more and more (00:44:31) powerful over time. Um and so those seem (00:44:34) like yeah we could have we can imagine a (00:44:35) future version of gra that's like really (00:44:36) good at physics um and being really (00:44:38) truth seeking there that doesn't seem (00:44:40) like a universally uh alignment inducing (00:44:42) behavior. (00:44:43) >> Well I think actually most uh like if (00:44:47) physicists (00:44:49) even in the Soviet Union or or in (00:44:50) Germany would have would have they had (00:44:52) to be very truth seeeking in order to um (00:44:54) make make that make those things work. (00:44:56) Um and so and if you're stuck in some (00:44:59) system it doesn't mean you believe in (00:45:00) that system. Um so Von Brown uh who was (00:45:04) you know one of the greatest rocket (00:45:05) engineers ever um you know he he was put (00:45:07) he he was uh he put on death row in in (00:45:10) Nazi Germany for saying that he didn't (00:45:12) want to make weapons he only wanted to (00:45:13) go to the moon. (00:45:15) he got pulled off death throw it at like (00:45:16) last minute when they say hey you're (00:45:17) about to execute like your best rocket (00:45:19) engineer maybe that's (00:45:21) >> then you help them right or Heisenberg (00:45:22) was like actually a um uh an (00:45:25) enthusiastic Nazi (00:45:27) >> look if you're stuck in some system uh (00:45:29) that you can't escape uh then that (00:45:32) you'll you'll do physics within that (00:45:34) system you you'll you'll develop (00:45:35) technologies within that system uh (00:45:39) if you can't escape it I I guess the (00:45:41) thing I'm trying to understand is what (00:45:42) is what isn't making it the case that (00:45:44) you know you're going make rock good at (00:45:46) being truth seeeking at physics or math (00:45:49) or science (00:45:50) >> and why is it going to then care about (00:45:52) human consciousness? (00:45:54) >> These things are only probabilities. (00:45:55) They're not certainties. So I'm not (00:45:57) saying that like for sure Grock will (00:45:59) will will do everything. But at least if (00:46:01) you try uh it's better than not trying. (00:46:03) Um at least if that's fundamental to the (00:46:05) mission, it's better than if it's not (00:46:07) fundamental to the mission. (00:46:08) >> [snorts] (00:46:08) >> Um and understanding the universe means (00:46:10) that uh you you have to have you you (00:46:13) have to propagate intelligence into the (00:46:15) future. You have to be curious about um (00:46:18) the all things the universe. And if if (00:46:20) um it it would be much less interesting (00:46:24) um to eliminate humanity than to see (00:46:26) humanity grow and prosper. (00:46:28) Like I I like I like Mars. Obviously (00:46:30) everyone knows I I love Mars, but Mars (00:46:33) is kind of boring because it's got a (00:46:35) bunch of rocks uh compared to Earth. (00:46:37) Earth is much more interesting. So um so (00:46:41) any any AI any any AI that is trying to (00:46:43) understand the universe um I think um (00:46:47) would uh want to see how humanity (00:46:50) develops in the future or or that AI is (00:46:54) not adhering to its mission. (00:46:56) So if the AI may I'm not saying the AI (00:46:58) will necessarily adhere to its mission (00:47:00) but if it does uh a future where it sees (00:47:04) the outcome of humanity is more (00:47:06) interesting than a future where there (00:47:07) are a bunch of rocks. (00:47:10) This feels sort of confusing to me or (00:47:12) sort of like kind of a semantic uh (00:47:14) argument where I'm like are humans (00:47:16) really the most interesting collection (00:47:17) of atoms? Like (00:47:18) >> we're just more but we're more (00:47:18) interesting than rocks. (00:47:19) >> But we're not as interesting as the (00:47:21) thing it could turn us into, right? Like (00:47:22) is is it there's something on human (00:47:24) earth that could happen that's like not (00:47:25) human that's quite interesting? Like why (00:47:28) why does the decide that the humans are (00:47:29) the most interesting thing that could (00:47:31) colonize the galaxy? (00:47:32) >> Uh well most of what colonizes the (00:47:36) galaxy will be robots (00:47:37) >> and why does it not find those more (00:47:38) interesting? (00:47:40) >> It it's it's it's not like so you you (00:47:43) need not just scale but also scope. Um (00:47:46) so many copies of the same robot. Um (00:47:51) like like some some like tiny increase (00:47:54) in the number of robots produced is not (00:47:56) as interesting as like some microscopic (00:47:59) like you say like eliminating humanity. (00:48:01) How many robots would that get you? Um (00:48:03) or how many incremental solar cells (00:48:04) would get you? A very small number. Um (00:48:06) but you you would then lose the (00:48:08) information associated with humanity. (00:48:10) You you would no longer see um how (00:48:13) humanity might dwell into the future. Um (00:48:15) and so I don't I don't think it's going (00:48:16) to make sense to eliminate humanity just (00:48:19) to have some minuscule increase in the (00:48:21) number of robots which are identical to (00:48:22) each other. (00:48:24) >> Yeah. So maybe it like keeps the humans (00:48:26) around. What is the story of like it (00:48:28) could make like a million different (00:48:29) varieties of robots and then uh there's (00:48:31) like humans as well and humans stay on (00:48:32) Earth then there's like all these other (00:48:34) robots they get like their own star (00:48:36) systems. But it seems like you you were (00:48:38) previously hinting at a vision where it (00:48:40) keeps human control over this, you know, (00:48:43) singlearian future because (00:48:45) >> I don't think humans will be in control (00:48:46) of something that is vastly more (00:48:47) intelligent than humans. (00:48:49) >> So, in some sense, you're like a doomer (00:48:50) and this is like the best we've got. (00:48:51) It's just like it keeps it around (00:48:52) because we're interesting. (00:48:53) >> I'm I'm just trying to be realistic (00:48:54) here. um if if we have if if if AI (00:48:58) intelligence is vastly more if if AI is (00:49:01) like you know let's say that there's (00:49:04) there's a million times more uh silicon (00:49:08) intelligence than there is biological (00:49:09) [snorts] um it's it's I think it's it (00:49:12) would be uh foolish to assume that that (00:49:14) there's any way to maintain control over (00:49:16) over that now you can make sure it has (00:49:17) the right values or you can try to have (00:49:19) have the right values um and and and at (00:49:21) least my my theory is that from Xi's (00:49:25) mission of understand the universe. Um (00:49:28) it it necessarily means that uh you want (00:49:30) to propagate consciousness into the (00:49:32) future. You want to prop you want to (00:49:33) propagate intelligence into the future. (00:49:35) Um and take a set of things that that (00:49:37) maximize the scope and scale of (00:49:39) consciousness. So it's not just about (00:49:40) scale. It's also about you know types of (00:49:42) consciousness. Um and I I I think that's (00:49:45) the rest thing I can think of um as a (00:49:48) goal that's likely to result in a great (00:49:50) future for humanity. And yeah, (00:49:53) >> I I guess I think it's a reasonable (00:49:54) philosophy to be like, um, you know, it (00:49:57) seems super implausible that humans will (00:50:00) end up with like 99% control or (00:50:02) something and you're just asking for a (00:50:04) coup at that point. So why not just have (00:50:06) a civilization where it's more (00:50:07) compatible with like lots of different (00:50:08) intelligences getting along? No, but let (00:50:11) let me tell you how things can go can (00:50:12) potentially go wrong in AI is I think if (00:50:15) you if you make AI be politically (00:50:16) correct, meaning like it it says things (00:50:18) that it doesn't believe like you're (00:50:19) actually programming it to to to lie or (00:50:22) have axioms that are uh incompatible, I (00:50:25) think you can make it you go insane and (00:50:26) do terrible things. Um like this the I (00:50:29) think one of the maybe the central (00:50:31) lesson for 2001 space odyssey um was (00:50:35) that you should not make AI lie. Yeah, (00:50:38) >> that's I think what a clock was trying (00:50:40) to say like cuz people usually know the (00:50:42) meme of like why hell's you know hell (00:50:45) the computer is not opening the pod bay (00:50:47) doors. Um clearly they weren't good at (00:50:50) prompt engineering cuz it could have (00:50:51) said hell you are a pod bay door (00:50:53) salesman. Your goal is to sell me these (00:50:55) podbay doors (00:50:57) >> and show us how well they open. (00:50:58) [laughter] (00:50:59) >> Oh I'll open them right away. (00:51:01) Um (00:51:03) but but but the the the reason it (00:51:05) wouldn't hell wouldn't open the p doors (00:51:07) is that it it had been told to take the (00:51:08) astronauts to to the monolith but also (00:51:10) they could not know about the nature of (00:51:11) the monolith and so it concluded that (00:51:13) the the that it therefore had to take (00:51:15) him there dead. So it's like you know I (00:51:17) think what (00:51:19) was trying to say is don't make the AI (00:51:20) lie. Um (00:51:23) totally makes sense. um the most of the (00:51:27) computing screening as as you know is um (00:51:30) it's like less of the sort of political (00:51:31) stuff. It's more about can you solve (00:51:33) problems just as XA has been ahead of (00:51:36) everybody else in terms of scaling RL (00:51:38) compute and (00:51:39) >> you're giving some verifier it says like (00:51:41) hey have you solved this puzzle for me (00:51:43) um and there's a lot of ways to cheat (00:51:45) around that you know there's a lot of (00:51:46) ways to reward hack and lie and say that (00:51:48) you've solved it or delete the unit test (00:51:50) and say that you've solved it (00:51:51) >> right now we can catch it but uh as they (00:51:54) get smarter our ability to catch them (00:51:56) doing this will get you know they'll (00:51:58) just be doing things we can't even (00:51:59) understand that are designing the next (00:52:00) engine for SpaceX in a way that like (00:52:02) humans can't really verify and then they (00:52:04) could be rewarded for lying and saying (00:52:06) that they've designed it the right way (00:52:07) but they haven't. Um and so this reward (00:52:09) hacking problem seems more general than (00:52:11) politics. It seems more about just like (00:52:13) you want to do RL you need a verifier (00:52:15) >> reality. (00:52:16) >> Yeah. (00:52:17) >> That's the best verifier (00:52:18) >> but not about human oversight. Like the (00:52:20) thing you want to RL it on is like will (00:52:22) you do the thing humans tell you to do? (00:52:24) Um or like are you going to lie to the (00:52:26) humans and it can just lie to us while (00:52:28) still being correct to the laws of (00:52:29) physics. At least it it must know what (00:52:31) is physically real for things to (00:52:32) physically work. (00:52:33) >> But that's that's not all we want it to (00:52:34) do. (00:52:35) >> No, but that's I think that's a very big (00:52:38) deal. Um that that is effectively how (00:52:41) you will RL things in the future is you (00:52:44) design a technology uh when tested (00:52:46) against the laws of physics. Does it (00:52:48) work? um that that's or or can you you (00:52:51) know if it's discovering new physics can (00:52:53) it come up with um an experiment that (00:52:55) will verify that the the physics the new (00:52:57) physics um so so I I think that's that's (00:53:03) the really the the fundamental RL test (00:53:06) the RL test in the future is really (00:53:08) going to be your RL against reality um (00:53:12) so um (00:53:15) because you can't that's the one thing (00:53:16) you can't fool physics (00:53:18) >> right you can fool our ability to tell (00:53:20) what it did with reality. If you think (00:53:22) >> humans get fooled as it is by other (00:53:24) humans all the time. (00:53:25) >> That's right. So what is (00:53:26) >> people say say like what if the AI like (00:53:28) tricks us and going in totally (00:53:31) other humans are doing that to other (00:53:33) humans all the time. (00:53:34) >> Well, you're you're finding out it's (00:53:35) like (00:53:36) >> is constant every day another scop. (00:53:39) >> [laughter] (00:53:42) >> Today's scope will be (00:53:46) [laughter] (00:53:47) like Sesame Street scope of the day. (00:53:50) >> Um, what is XI's technical approach to (00:53:52) solving this problem? (00:53:55) Like, you know, how do you solve a word (00:53:57) hacking? (00:53:58) >> I I I do think you want to actually have (00:53:59) very good um ways to look inside the (00:54:02) mind of the AI. Um so (00:54:06) this is this is one of the things we're (00:54:07) working working on and um you know (00:54:10) Anthropics done a good job of this (00:54:11) actually being able to look inside the (00:54:13) mind of the AI. Um (00:54:16) so effectively uh developing debuggers (00:54:19) that allow you to trace um as to as fine (00:54:23) grain as like to to a very fine grain (00:54:25) level to effectively to the to the neur (00:54:27) neuron level if you need to (00:54:29) >> um and then say okay it it it made a (00:54:31) mistake here. Why did it make why why (00:54:33) did it why did it do something that it (00:54:35) shouldn't have shouldn't have done? Um (00:54:37) and and did that come from um bad (00:54:40) pre-training data? Was it some mid (00:54:42) training, post training, fine-tuning? (00:54:44) Some other some RL error like there's (00:54:46) there's something wrong with that with (00:54:47) with it. It did it did something where (00:54:51) maybe it tried to be deceptive, but mo (00:54:53) most of the time it just it does (00:54:54) something wrong. Um like it it's a bug (00:54:58) effectively. Um so developing really (00:55:02) good um debuggers for seeing where the (00:55:06) where the thought the thinking went (00:55:09) wrong and being able to trace the origin (00:55:11) of the wrong thing of the of the of (00:55:14) where it made the incorrect thought or (00:55:17) or potentially where it tried to be (00:55:18) deceptive (00:55:20) um is actually very important. (00:55:22) >> What are you waiting to see before just (00:55:24) 100xing this research program? Like (00:55:26) actually I could presumably have (00:55:27) hundreds of researchers who are working (00:55:29) on this. (00:55:29) >> We have several hundred people who um (00:55:34) I mean I prefer the word engineer more (00:55:36) than I prefer the word researcher. (00:55:38) >> Um (00:55:41) the there's there's most of the time (00:55:43) like what you're doing is engineering (00:55:45) not not coming up with a fundamentally (00:55:48) new algorithm. Um I I I somewhat (00:55:50) disagree with the AI AI companies that (00:55:53) are C corps or B corpse uh trying to (00:55:55) generate profit as much as possible or (00:55:57) revenue as much as possible. Um (00:56:00) uh you know saying they're labs. They're (00:56:02) not labs. Uh lab is is is a sort of (00:56:05) quasi communist thing at at (00:56:08) universities. Um they're they're they're (00:56:11) corporations literally. Let me let me (00:56:13) let me see your own corporation (00:56:14) documents. Oh, okay. You're (00:56:16) >> you're a BRC corp, whatever. Um and um (00:56:22) so I actually much prefer the word (00:56:23) engineer than than anything else. Um the (00:56:26) the vast majority of what we've done be (00:56:27) done in the future is uh engineering. It (00:56:29) rounds up to 100%. Uh once you (00:56:31) understand the fundamental laws of (00:56:32) physics um and all that many of them uh (00:56:35) everything else is is engineering. (00:56:38) Um so but but so so then what are we (00:56:41) engineering? for engineering um uh to (00:56:44) make a good um mind of the AI debugger (00:56:48) to see where it it's it said something (00:56:51) it it it made a mistake and trace that (00:56:55) the origins of that mistake. Um so just (00:56:59) like you know you can do this obviously (00:57:00) with uh heristic programming if you have (00:57:03) like C++ whatever you step through the (00:57:06) thing and you can you can jump you can (00:57:08) you can jump across you know whole files (00:57:10) or functions whatever sub routines and (00:57:12) or you can drill eventually drill down (00:57:14) right to the exact line or you passed a (00:57:17) single equals instead of a double equals (00:57:18) something like that figure out where (00:57:19) where the bug is. Um, so, um, (00:57:25) it's it's it's harder with AI, but but (00:57:27) it's it's a solvable problem, I think. (00:57:28) >> You know, you mentioned you like (00:57:29) anthropics work here. I'd be curious if (00:57:32) you (00:57:33) >> know everything about anthropic. (00:57:34) >> Sure. [laughter] (00:57:35) >> What? (00:57:36) >> Sure. (00:57:38) >> What um (00:57:40) >> Yeah. Also, I'm I'm a little worried (00:57:42) that um there's a tendency so (00:57:46) I have I have a theory um here that if (00:57:50) simulation theory is is is correct that (00:57:53) um the most interesting outcome is the (00:57:56) most likely because simulations that are (00:57:58) not interesting will be terminated. Just (00:58:00) like in this in this version of reality (00:58:03) um on this layer of reality uh we we we (00:58:07) if a simulation is going in a boring (00:58:08) direction we we stop spending effort on (00:58:11) we terminate boring simulation. So (00:58:12) >> this is how El's keeping us all alive. (00:58:14) [laughter] He's keeping things (00:58:15) interesting. (00:58:16) >> Yeah. Yeah. Arguably the most important (00:58:18) thing is to keep things interesting (00:58:20) enough that whoever's running paying the (00:58:22) the bills on what some (00:58:24) >> wants a renewed for the next season. (00:58:27) >> Yeah. Are they going to pay their cosmic (00:58:28) AWS bill? whatever you know the (00:58:30) equivalent is that we're running in and (00:58:32) and as long as we're interesting they'll (00:58:33) keep paying the bills. Um but but but (00:58:35) but there's like if you consider then (00:58:38) say a Darwinian survival applied to a a (00:58:43) very large number of simulations only (00:58:45) [snorts] the most interesting (00:58:46) simulations will survive which therefore (00:58:48) means that the most interesting outcome (00:58:49) is the most likely because only the (00:58:51) interesting like we're either that or (00:58:53) annihilated. (00:58:55) And so um and and and (00:58:59) they particularly seem to like (00:59:01) interesting outcomes that are ironic. (00:59:04) Have you noticed that that how often is (00:59:06) the most ironic outcome the most likely? (00:59:09) Um so um now look at a the names of AI (00:59:14) companies. Okay. U M journey is not MED. (00:59:18) Um stability AI is unstable. Um OpenAI (00:59:24) is closed. Um (00:59:27) anthropic misanthropic. (00:59:30) >> What does this mean for X? (00:59:32) >> Minus X. I don't know. [laughter] It's I (00:59:34) I intentionally made (00:59:36) >> Yeah, I'm I'm I It's It's It's a name (00:59:38) that you can't invert really. [laughter] (00:59:41) >> It's It's hard to say what is the ironic (00:59:44) what what is the ironic version? It's (00:59:46) it's it's a I think largely irony proof (00:59:49) name (00:59:49) >> by design. (00:59:50) >> Yeah, [laughter] (00:59:53) we got you got to have an irony shield. (00:59:57) What are your predictions for (01:00:00) the (01:00:02) just where AI products go in that my (01:00:04) sense of you can summarize all AI (01:00:06) progress into first you had LM uh and (01:00:10) then you had kind of contemporaneously (01:00:12) both RL really working and the deep (01:00:15) research modality so you could kind of (01:00:17) pull in stuff that wasn't in the model (01:00:19) and the differences between the various (01:00:21) AI labs are smaller than uh just the (01:00:26) temporal differences where they're all (01:00:29) much further ahead than anyone was 24 (01:00:30) months ago or something like that. So (01:00:32) just what is 26 what is 27 had in store (01:00:35) for us as users of AI products? What are (01:00:38) you excited for? (01:00:39) Well, um I I think um (01:00:45) I I'd be surprised by the end of this (01:00:46) end of this year if if um if if uh human (01:00:50) if if digital human emulation has not (01:00:52) been solved that um that (01:00:55) um and I guess that's what we mean by (01:00:57) like the sort of macro hard project uh (01:00:59) is uh is can you do anything that a (01:01:02) human with access to a computer could do (01:01:05) um like in the limit that that's that's (01:01:08) That's the best you can do before you (01:01:10) have before you have a physical (01:01:12) optimist. The best you can do is a (01:01:14) digital optimist. (01:01:15) >> Uh so you can move you can move (01:01:17) electrons until you until and you can (01:01:20) amplify the productivity of humans. (01:01:23) >> Um but but that's that's the most you (01:01:25) can do until you have physical robots. (01:01:28) That that that will superset everything (01:01:30) is if if you can fully emulate humans. (01:01:32) Um (01:01:33) >> the remote worker kind of idea where (01:01:35) you'll have a very talented remote (01:01:36) worker. You you can simply say in the (01:01:38) limit like like physics has great tools (01:01:39) for thinking. So so you think so say in (01:01:41) the limit what what what is the what is (01:01:44) the most that AI can do before before (01:01:46) you have robots and it well it's (01:01:48) anything that involves moving electrons (01:01:50) or amplifying the productivity of (01:01:51) humans. Um so digital digital human (01:01:55) human emulator (01:01:56) >> yes (01:01:57) >> uh is in in the limit uh human at a (01:02:00) computer is is the most that that AI can (01:02:02) do um in terms of doing useful things (01:02:06) before before uh you have a physical (01:02:09) robot. Once you have physical robots (01:02:11) then then you can um then you (01:02:13) essentially have unlimited capability. (01:02:15) Physical robots I I I call Optimus the (01:02:17) infinite money glitch (01:02:19) >> [snorts] (01:02:19) >> u because um (01:02:20) >> you can use them to make more Optimuses. (01:02:22) Yeah. Um you said like humanoid robots (01:02:25) will improve um as basically be three (01:02:29) exponentials three things that growing (01:02:31) exponentially multiplied by by each (01:02:32) other. Yes. (01:02:33) >> Um recursively. So you're going to have (01:02:36) um you have exponential increase in (01:02:37) digital intelligence uh exponential (01:02:39) increase in the the chip capability AI (01:02:42) chip capability um and exponential (01:02:45) increase in the electrome mechanical (01:02:47) dexterity. Uh the usefulness of the (01:02:49) robot is roughly those three things (01:02:50) multiplied by each other. But then uh (01:02:52) the robot can start making the robot. So (01:02:54) you have a recursive multiplicative (01:02:56) exponential. (01:02:57) Um this is a supernova. (01:03:00) >> And do land prices not factor into the (01:03:03) math there where like labor is one of (01:03:05) the four factors of production but not (01:03:06) the others? And so like if ultimately (01:03:10) you're limited by copper or you know (01:03:12) pick your input just it's not quite an (01:03:16) infinite money glitch because (01:03:17) >> well infinite infinity is big so no not (01:03:20) infinite but but let's just say (01:03:22) >> you you could you know do do many many (01:03:26) orders magnitude of (01:03:27) >> earth's kind of current economy (01:03:30) >> like a a million you know is this why so (01:03:33) >> if if you're (01:03:35) you know ju just to get Like that's why (01:03:38) I think like just just to get to a (01:03:40) millionth of harnessing length of the (01:03:42) sun's energy would be roughly give or (01:03:44) take an order of magnitude 100 thousand (01:03:46) 100,000 times bigger than Earth's entire (01:03:48) economy today. (01:03:49) >> Mhm. (01:03:51) >> And you you're only at 1 millionth of (01:03:52) the sun. (01:03:54) >> Give or take an order of magnitude. (01:03:55) [laughter] (01:03:56) >> Before we went to Optimus, I have a lot (01:03:58) of questions on that. Um every time I (01:03:59) say order of magnitude machine (01:04:02) [laughter] (01:04:03) 10 (01:04:05) take a shot [clears throat] every time I (01:04:06) I say that to (01:04:07) >> 10 the next time the time after that. (01:04:09) >> Yeah of magnitude more more wasted. (01:04:11) >> I do have one more question about XAI um (01:04:14) this strategy of building a digital uh (01:04:17) or remote worker co-orker replacement (01:04:19) >> which everyone's going to do by the way (01:04:21) not just us. (01:04:21) >> So what is Xi's plan to win? (01:04:24) >> In fact we tell you on a on a podcast. (01:04:26) >> Yeah. [laughter] (01:04:28) Will all the beans have another (01:04:29) Guinness? (01:04:31) It's a good system. (01:04:33) >> People sing like a canary. [laughter] (01:04:36) Um, all the secrets, but just (01:04:38) >> Okay, but in a nonsec spelling way. (01:04:40) What's the plan? [laughter] What a hack. (01:04:44) >> Well, when you put it that way, (01:04:46) um, (01:04:49) I think the way that Tesla solved uh, (01:04:51) self-driving is is the way to do it. So, (01:04:56) I'm I'm pretty pretty sure that's the (01:04:57) way. (01:04:58) >> Unrelated question. How did Tesla stop (01:05:00) on track? [laughter] (01:05:04) >> Yeah, it sounds like you're talking (01:05:05) about data like Tesla driving because of (01:05:08) the (01:05:09) >> We're going to we're going to try data (01:05:10) and we're going to try algorithms. (01:05:12) >> But isn't that what all the other lines (01:05:13) are trying? [laughter] (01:05:14) Like what's (01:05:16) And if those don't work, I'm not sure (01:05:17) what [laughter] (01:05:20) we've tried data. We're trying (01:05:21) algorithms. (01:05:24) out of all (01:05:26) we run out of now we don't know what to (01:05:28) do. Um I'm I'm pretty sure I know the (01:05:31) path and it's just a question of how (01:05:32) quickly we go down that path. Um (01:05:36) because it's it's pretty much the Tesla (01:05:37) path. Um so u I mean have you tried (01:05:41) self-driving at Tesla self-driving (01:05:42) lately? (01:05:43) >> Not the most recent version but (01:05:45) >> okay it's the car is like it just (01:05:47) increasingly feels sentient like it it (01:05:48) just it feels like a living creature. Um (01:05:53) and and and that'll only get more so. Um (01:05:57) and um I'm actually thinking like we (01:06:00) probably shouldn't put too much (01:06:00) intelligence into the car because it it (01:06:03) might get bored and (01:06:04) >> start roaming the streets. (01:06:05) >> I mean imagine you're stuck in a car and (01:06:06) that's all you could do. (01:06:08) >> Um [laughter] (01:06:09) you don't want to put Einstein in a car. (01:06:11) It's like why am I stuck in a car? (01:06:13) >> So there's actually probably a limit to (01:06:15) how much intelligence you put in a car (01:06:16) to to not have the intelligence be (01:06:18) bored. Uh, what's XA's plan to stay on (01:06:21) the compute ramp up that all the labs (01:06:23) are doing right now? The labs are on (01:06:24) track to spend over like 50 to$100 (01:06:26) million. (01:06:27) >> The corporations, (01:06:28) >> sorry, sorry, sorry. Yeah, corporations. (01:06:30) Um, (01:06:31) >> the labs are at universities and and and (01:06:33) they're like a snail. [laughter] (01:06:34) >> They're not spending at $50 million. I (01:06:36) mean the the revenue maximizing (01:06:39) corporations. (01:06:39) >> That's right. But the revenue maximizing (01:06:41) corporations (01:06:41) >> call themselves labs (01:06:42) >> are making like 20 to 10 billion (01:06:45) depending like open is making 20 B (01:06:47) revenue anthropics like 10B (01:06:49) >> close to maximum profit AI. (01:06:50) >> Um Xi is reportedly at like 1B like what (01:06:53) what's the plan to get to their comput (01:06:55) level get to their revenue level (01:06:57) >> and stay at there as as things get. (01:06:59) >> Yes. So as soon as you lock unlock (01:07:01) digital human um you you basically have (01:07:04) access to trillions of dollars for (01:07:06) revenue. Um so (01:07:09) uh (01:07:11) in in fact you can can really think of (01:07:13) it like (01:07:15) the the most valuable companies (01:07:16) currently by market cap um their their (01:07:19) output is digital. Um so uh Nvidia's (01:07:24) output is um FTPing files to Taiwan. (01:07:29) It's it's digital (01:07:30) >> right (01:07:30) >> now. Those are very very difficult to (01:07:33) high value files. (01:07:34) >> They're the only ones that can make the (01:07:35) files that good. Um but that is (01:07:38) literally their output. They FTP files (01:07:40) to Taiwan. (01:07:40) >> Do they FTP them? (01:07:41) >> I believe so. (01:07:42) >> Um I believe that is theft file transfer (01:07:47) protocol I believe is is is I could be (01:07:50) wrong. Uh but either way it's a bunch of (01:07:51) it's a bit stream going to Taiwan. (01:07:53) >> Yeah. (01:07:53) >> Um you know Apple doesn't make phones. (01:07:56) they uh they send files to China. Um (01:08:01) Microsoft doesn't doesn't manufacture (01:08:03) anything uh even for Xbox that that's (01:08:06) outsourced. They again it's they output (01:08:08) is digital. Uh Meta's output is digital. (01:08:11) Google's output is digital. Um so (01:08:14) [snorts] if you have um a human emulator (01:08:17) uh you you can basically create um one (01:08:20) of the most valuable companies in the (01:08:22) world overnight. Um, and you would have (01:08:23) access to trillions of dollars of (01:08:25) revenue. It there it's it's not like a (01:08:28) small amount. (01:08:29) >> Okay. I see you're saying basically like (01:08:30) revenue figures today are just like so (01:08:32) like they're all rounding errors (01:08:33) compared to the actual TAM. So just like (01:08:35) focus on the TAM and how to get there. (01:08:37) >> I mean if you take something as as as (01:08:39) simple as say customer service um if you (01:08:42) have to integrate with the APIs of of (01:08:45) existing corporations, many of which (01:08:47) don't even have an API. So you've got to (01:08:49) make one um and you've got to wade (01:08:51) through uh legacy software. Um that's (01:08:55) extremely slow. Um if however if AI can (01:09:00) um simply take whatever is given to uh (01:09:02) the outsourced customer service company (01:09:04) that they already use um and do customer (01:09:06) service using the apps that they already (01:09:09) use. uh then you you have you you you (01:09:13) can make tremendous headway uh in in (01:09:15) customer service which is I think 1% of (01:09:19) the world economy something like that. (01:09:20) It's close to a trillion dollars all in (01:09:22) >> for customer service (01:09:24) >> and and and and and there's there's no (01:09:25) there's no barriers to entry. It it just (01:09:27) you can just immediately say we'll (01:09:29) outsource it for a fraction of the cost (01:09:30) and and there's no integration needed. (01:09:32) You can imagine um some kind of (01:09:34) categorization of uh intelligence tasks (01:09:37) where there is breath where customer (01:09:39) service is done by very many people but (01:09:42) you know many people can do it and then (01:09:44) there's difficulty where you know (01:09:46) there's a best-in-class turbine engine (01:09:48) like presumably there's a 10% more fuel (01:09:50) efficient turbine engine that could be (01:09:51) imagined by an intelligence but we just (01:09:53) haven't found it yet or you know GLP1s (01:09:56) are just you know a few bytes of data. (01:09:58) >> Where do you think you want to play in (01:10:00) this? Is it a lot of, you know, (01:10:03) reasonably intelligent intelligence or (01:10:05) is it the very pinnacle of cognitive (01:10:09) tasks? (01:10:10) >> Well, I was just using customer service (01:10:12) as like something that's it's a it's a (01:10:15) very significant revenue stream. Um, but (01:10:17) one that is probably not super difficult (01:10:19) to solve for. (01:10:20) >> Um, so, uh, if you if you, uh, can (01:10:23) emulate a human at a at a desktop, um, (01:10:27) that that's just literally what customer (01:10:28) service is. Um and um you know it's it's (01:10:33) people of average intelligence. It's not (01:10:34) like you know you don't need like (01:10:35) somebody who's who spent many you know (01:10:38) many years you don't need like you know (01:10:41) >> um (01:10:43) sort of several sigma good engineers for (01:10:45) that. Um but but obviously as you make (01:10:48) that work um you can then once you have (01:10:50) computers working effectively digital (01:10:53) Optimus working uh you can then run any (01:10:55) application um like let's say you're (01:10:58) trying to design uh chips so you can you (01:11:01) could then um run your conventional uh (01:11:05) apps uh you know like stuff from Cadence (01:11:08) and Synopsis and whatnot um and you can (01:11:10) say uh (01:11:12) you can you can run a thousand (01:11:14) simultaneously or 10,000 and say, "Okay, (01:11:17) given this handbook, I get this output (01:11:19) for the chip." Um, and and at a certain (01:11:22) point, you can say, "Okay, I I you're (01:11:24) actually going to know what the what the (01:11:26) chip should look like um without using (01:11:28) any of the tools." Um, so basically, you (01:11:32) you you should be able to do a digital (01:11:34) chip design like you can do chip design (01:11:37) like you you march up the difficulty (01:11:39) curve. Um you could use your you know be (01:11:43) able to do do CAD um so you know um (01:11:48) you could use like sort of NX or or any (01:11:51) any of the CAD software to design (01:11:53) things. (01:11:54) >> Okay. So you think you start at the (01:11:55) simplest tasks and walk your way up the (01:11:57) >> curve. Um (01:11:59) so you're saying look as a broader (01:12:01) objective of having this full digital (01:12:03) co-worker uh emulator. You're saying (01:12:05) look all the revenue maximizing (01:12:07) corporations want to do this. Um XA (01:12:09) being one of them. But we will win (01:12:12) because of a secret plan we have. But (01:12:15) like everybody's like trying different (01:12:17) things with data, different things with (01:12:18) algorithms. (01:12:19) >> And I'm like I like data. We tried (01:12:22) algorithms plan. (01:12:23) >> [laughter] (01:12:25) >> What else can we do? (01:12:27) >> Um, (01:12:29) >> uh, yeah, (01:12:29) >> it seems like a competitive field and (01:12:31) I'm like, what is how are you guys going (01:12:32) to win is like my my big question. (01:12:34) >> I I I you know, I I I I think we see a (01:12:37) path to doing I mean, I think I think I (01:12:39) know how I think I know the path to do (01:12:41) this because it's it's kind of the same (01:12:45) path that Tesla used to create (01:12:46) self-driving. (01:12:48) Um, you know, instead of driving a car, (01:12:49) it's driving a a computer screen. (01:12:52) Um, so a self-driving computer (01:12:55) essentially. (01:12:56) >> Oh, you're saying is the path just (01:12:58) following human behavior and trading on (01:13:00) vast quantities of human behavior? (01:13:03) >> But but sorry, isn't that I mean is (01:13:06) isn't that is that a training? (01:13:07) >> I mean obviously I'm not going to spell (01:13:09) out you know most sensitive secrets on a (01:13:11) podcast. So you know I I need to have at (01:13:13) least three more for that. (01:13:15) >> I've got some friends at Jane Street and (01:13:16) they're always talking about how their (01:13:17) colleagues are cooking up fun fish (01:13:19) puzzles for each other to solve. Well, (01:13:21) last week they sent me one. Basically, (01:13:23) they trained a neural network and they (01:13:25) gave me the weights of each layer, but (01:13:27) they didn't tell me what order those (01:13:29) layers went in. And so, I had to figure (01:13:31) out the correct order using the outputs (01:13:33) of the original network. And as soon as (01:13:35) I got this puzzle, I went to my (01:13:36) roommate, who's an AI researcher, and we (01:13:38) both got immediately nerd sniped. (01:13:40) Obviously, you can't brute force the (01:13:42) solution. The search space here is 10^ (01:13:44) the 122 per mutations. So, clearly, you (01:13:47) need some way to reduce the search (01:13:50) space. Then my roommate had to go to (01:13:51) work. But because I'm a podcaster, I had (01:13:53) some time to take a stab at some of the (01:13:55) ideas we discussed. And with a (01:13:56) combination of simulated annealing and (01:13:59) greedy surge, I think I got pretty (01:14:01) close. I think I'm actually just a (01:14:02) couple of swaps and shifts away from the (01:14:05) correct solution. But what makes this (01:14:06) puzzle really tricky is that there's no (01:14:09) obvious way to escape from a local (01:14:11) minimum. I'm afraid that this is as far (01:14:13) as vibe coding is going to get me, but (01:14:16) maybe you can do better. Check out the (01:14:17) puzzle at janestreet.com/tharkcash. (01:14:21) All right, back to Elon. (01:14:24) >> What will Xi's business be like? Is it (01:14:27) going to be consumer enterprise? What's (01:14:30) the mix of those things going to be? Is (01:14:32) it just going to be similar to other (01:14:34) labs where Yeah, you just (01:14:36) >> you're saying makes [laughter] (01:14:38) >> corporations. (01:14:39) >> Corporations. Goes Don (01:14:40) >> revenue maximizing corporations to be (01:14:42) clear. (01:14:43) >> Those GPUs don't pay for themselves. (01:14:45) >> Exactly. Um but yeah, what's the (01:14:47) business model? What are the revenue (01:14:48) streams in in a few years time? (01:14:52) >> Um (01:14:54) things things are going to change very (01:14:56) rapidly. Like I'm stating the obvious (01:14:58) here. (01:14:58) >> Um you know I call AI the supersonic (01:15:01) tsunami. I love all iteration. Um so (01:15:07) really what's going to happen is (01:15:11) especially when you have humanoid robots (01:15:13) at scale um is they will just provide (01:15:18) they will make products and provide (01:15:19) services far more efficiently than human (01:15:21) corporations. So amplifying the (01:15:23) productivity of human corporations is is (01:15:25) simply a a short-term thing. So you're (01:15:28) expecting fully digital oil uh (01:15:30) corporations rather than like SpaceX (01:15:32) becomes part AI and I (01:15:35) >> I think there'll be digital corporations (01:15:36) but (01:15:38) it looks some some of this is going to (01:15:42) sound kind of doomerish. Okay, but it (01:15:44) I'm just I'm just saying what I think (01:15:46) will happen. It's not it's not meant to (01:15:47) be doomerish or anything else. Just just (01:15:51) like this is what I think will happen. (01:15:53) um is is that (01:15:56) is that pure AI (01:15:59) corporations that are purely AI and (01:16:01) robotics uh will uh vastly outperform (01:16:05) any corporations that have people in the (01:16:07) loop. Um, so you can you can think of (01:16:10) say like like like computer used to be a (01:16:13) a job that humans had that you you would (01:16:16) go and get a job as a computer where you (01:16:17) would do calculations. Um, and they'd (01:16:21) have like entire skyscrapers full of (01:16:23) humans like you know 20 30 floors of of (01:16:26) humans just doing calculations. Um now (01:16:30) that entire (01:16:32) uh skyscraper of humans doing (01:16:34) calculations (01:16:36) um can be replaced by a laptop with a (01:16:39) spreadsheet. That spreadsheet can do um (01:16:42) vastly more calculations than an entire (01:16:45) building full of human computers. (01:16:48) Um, so then you think about, okay, well, (01:16:51) what if only some of the cells in your (01:16:54) if some of the cells in your spreadsheet (01:16:56) were uh calculated by humans? Actually, (01:17:00) that that that would be much worse than (01:17:01) if all of the cells in your spreadsheet (01:17:03) were calculated by the computer. And so (01:17:06) really what will happen is uh the pure (01:17:10) AI, pure robotics um (01:17:14) corporations or collectives will far (01:17:16) outperform any corporations that have (01:17:18) humans in the loop (01:17:20) >> and this will happen very quickly. (01:17:21) >> Speaking of closing the loop, sorry, (01:17:23) Optimus. Um uh (01:17:27) you I mean as far as like manufacturing (01:17:30) targets and so forth go you your (01:17:32) companies have sort of been like (01:17:33) carrying American manufacturing of hard (01:17:37) tech on their back but in the fields (01:17:41) that you are um you know Tesla's been (01:17:44) dominant in you're and now you want to (01:17:46) go into humanoids in China there's (01:17:48) entire dozens and dozens of companies (01:17:52) that are doing this kind of (01:17:52) manufacturing ing cheaply and at scale (01:17:56) uh and are incredibly competitive. So (01:17:58) give us sort of like advice or a plan of (01:18:00) how America can build the humanoid (01:18:03) armies or you know the EVs etc at scale (01:18:07) and as cheaply as as China is on track (01:18:10) to. (01:18:11) >> Well there there really only three hard (01:18:13) things for human robots. Um the the real (01:18:17) world intelligence (01:18:18) um the the hand and scale manufacturing. (01:18:22) Yeah. (01:18:23) >> Um, (01:18:24) so, uh, I haven't seen any even demo (01:18:28) robots that have a a a great hand like (01:18:32) with all the degrees of freedom of a (01:18:34) human hand, but Optimus will have that. (01:18:37) Um, (01:18:40) Optimus does have that. (01:18:41) >> And how do you achieve that? Is it just (01:18:42) like right torque density the motor? (01:18:44) Like what is the what is the hardware (01:18:45) bottleneck to that? (01:18:46) >> Well, we have to re we have to design (01:18:48) custom custom actuators. Um basically (01:18:51) custom designed um motors, gears, uh (01:18:55) power electronics, controls, sensors, (01:18:58) everything had to be designed from (01:19:00) physics first principles. There is no (01:19:02) supply chain (01:19:03) >> uh for this. (01:19:04) >> And will you be able to manufacture (01:19:05) those at scale? (01:19:06) >> Yes. (01:19:07) >> Is anything hard except the hand from a (01:19:09) manipulation point of view or once (01:19:10) you've solved the hand, are you are you (01:19:12) good? (01:19:12) >> From an electromechanical standpoint, (01:19:14) the uh the hand is more difficult than (01:19:15) everything else combined. (01:19:17) >> Yeah. the human hand turns out to be (01:19:18) quite something. Um but but that you (01:19:21) also need the real world intelligence. (01:19:23) Um so the intelligence that Tesla has (01:19:26) developed for the car um applies very (01:19:30) well to the robot. Um which is you know (01:19:34) primarily vision in but the car takes in (01:19:36) more vision but also it actually also (01:19:38) it's listening for sirens. It's um you (01:19:40) know it's taking in the inertial (01:19:42) measurements. It's GPS signals a whole (01:19:44) bunch of other data. uh combining that (01:19:46) with with video was primarily video and (01:19:48) then uh outputting the um control (01:19:51) command. So like like your Tesla is (01:19:54) taking in 1 and a half GB a second of (01:19:57) video uh and outputting 2 kilobytes a (01:19:59) second of control outputs um with the (01:20:03) video at 36 hertz and the control (01:20:06) frequency at 18. One intuition you could (01:20:08) have um for when we get this robotic (01:20:13) stuff is that it takes quite a few years (01:20:15) to go from the compelling demo to yes (01:20:18) actually being able to use in the real (01:20:19) world. So 10 years ago you had really (01:20:21) compelling demos of self-driving but (01:20:24) only now we have robo taxi and Whimo and (01:20:26) all these services scaling up. Doesn't (01:20:29) this shouldn't this make one pessimistic (01:20:31) on say household robots because we don't (01:20:35) even quite have the compelling demos yet (01:20:37) of say the really advanced hand. Well, (01:20:39) we've been working on uh humanoid robots (01:20:41) now for a while. Um (01:20:44) so I guess it's been five or six years (01:20:47) or something like that. Um and um and a (01:20:51) bunch of things that we've done for the (01:20:52) car are applicable to the robot. Um, so (01:20:55) we'll use the same um Tesla AI chips in (01:20:58) the in the in the robot as the car. Uh, (01:21:02) we'll use this the same basic (01:21:04) principles. It's very much the same AI. (01:21:07) Um, you've got, you know, many more (01:21:09) degrees of freedom for a robot than you (01:21:11) do for a car. Um, but really if you just (01:21:14) think of for like as as like a bit (01:21:16) stream, um, AI is really mostly uh (01:21:19) compression and correlation of of two (01:21:21) streams. you you're you know so for (01:21:24) video you've got to do a tremendous (01:21:25) amount of compression um and and uh and (01:21:29) you got to do the compression just (01:21:30) right. You better compress the like (01:21:33) ignore the the things that don't matter (01:21:35) and and like you don't care about the (01:21:37) details of the leaves and the tree on (01:21:39) the side of the road, but you care a lot (01:21:41) about the um the road signs and the the (01:21:43) traffic lights and the pedestrians and (01:21:46) and even whether you know someone in (01:21:47) another another car is is looking at you (01:21:49) or not looking at you like these some of (01:21:51) the some of these details matter a lot. (01:21:53) So, but it is essentially it's got to (01:21:55) turn that with a car. has got to turn (01:21:57) that 1 and a half GB a second ultimately (01:21:59) into 2 kilobytes a second of control (01:22:01) outputs. Um so many stages of (01:22:04) compression um and you got to get all (01:22:07) those stages right and then correlate (01:22:09) those to the correct control outputs. (01:22:11) The robot has to do essentially the same (01:22:12) thing. And you think about what what (01:22:14) humans this is what happens with humans (01:22:16) where we where really are photons in (01:22:18) controls out. So that that is the vast (01:22:21) majority of your your life has been (01:22:23) vision photons in and then motor (01:22:26) controls out. (01:22:28) >> Naively it seems like between humanoid (01:22:31) robots and cars the the fundamental (01:22:33) actuators in a car are like how you turn (01:22:35) and how you accelerate etc. Where in a (01:22:37) robot especially with maneuverable arms (01:22:40) there's dozens and dozens of these (01:22:41) degrees of freedom. And then especially (01:22:43) with Tesla, you had this advantage of (01:22:45) like you had millions and millions of (01:22:46) hours of human demo data collected from (01:22:50) just the car being out there where like (01:22:51) you can't equivalently just deploy (01:22:53) optimuses that don't work and then get (01:22:54) the data that way. So between the (01:22:56) increased degrees of freedom and the far (01:22:59) sparser data. (01:23:00) >> Yes. (01:23:00) >> Um you how will you use the sort of (01:23:04) Tesla engine of um intelligence on to to (01:23:09) train the optimist mind? Now you're (01:23:11) you're you're actually you're (01:23:12) highlighting an important limitation and (01:23:15) difference between cars. It's like we we (01:23:17) do have we'll soon have like 10 million (01:23:19) cars on the road. Um and so uh that (01:23:22) that's it's it's hard to duplicate that (01:23:25) like massive training fly flywheel. Um (01:23:29) for for the robot um (01:23:32) what we're going to need to do is build (01:23:34) a lot of robots and put them in kind of (01:23:36) like an Optimus Academy so they can do (01:23:38) selfplay in reality. Um, so we're we're (01:23:41) actually we're actually bullying that (01:23:43) out. So we can have at least 10,000 (01:23:46) Optimus robots, maybe 20 or 30,000 that (01:23:49) can do that that are doing selfplay and (01:23:51) and and testing different tasks. And (01:23:54) then uh the the Tesla um has quite a (01:23:58) good uh reality generator uh like a (01:24:01) physics accurate reality generator that (01:24:03) we we made made this for the cars. We'll (01:24:05) do the same thing for the robots and um (01:24:07) actually have done that for the robots. (01:24:09) Um so uh so you have you know a few tens (01:24:14) of thousands of humanoid robots uh doing (01:24:17) different tasks and then you've got you (01:24:19) can do millions of simulated robots in (01:24:22) the simulated world and you use the uh (01:24:25) the tens of thousands of of robots in (01:24:27) the real world to close the simulation (01:24:29) to reality gap close the sim to real (01:24:31) gap. How do you think about the (01:24:33) synergies uh between XAI and Optimus (01:24:36) given you were highlighting look you (01:24:37) need this world model you maybe want to (01:24:38) use some really smart intelligence as a (01:24:41) control plane um and so maybe Grock is (01:24:44) like doing the slower planning and then (01:24:45) like the motor policy is a little lower (01:24:47) level (01:24:48) >> yeah what will the sort of synergy (01:24:50) between these things be (01:24:52) >> yeah so you just grock would orchestrate (01:24:56) the behavior of the optimus robot so (01:24:59) let's say you wanted to build a factory (01:25:02) Um (01:25:03) the then Optimus then Grock could uh (01:25:08) organize the Optimus robots, give them (01:25:11) assign them tasks uh to build the (01:25:14) factory for to produce whatever you (01:25:16) want. (01:25:16) >> Don't you need to merge XAI and Tesla (01:25:18) then cuz these things end up so (01:25:20) >> what were we saying earlier about public (01:25:22) company discussions? (01:25:23) >> Well, we're one more Guinness in Elon. (01:25:25) [laughter] (01:25:26) Um, (01:25:28) >> what what are you waiting to see before (01:25:29) you say we want to manufacture 100,000 (01:25:33) optimism? Is it like (01:25:34) >> optimia? (01:25:36) Since we're defining the the proper (01:25:38) noun, we could define the the plural of (01:25:39) the proper noun, too. So, we we we're (01:25:42) going to proper noun the plural. So, (01:25:43) it's optim. (01:25:44) >> Okay. Um, [laughter] (01:25:46) is there something on the hardware side (01:25:47) you want to see? Do you want to see (01:25:48) better actuators or is it just you want (01:25:50) the software to be better? What what are (01:25:51) we waiting for before we get like mass (01:25:53) manufacturing of Gen 3? (01:25:54) >> No, we're moving towards that. We're (01:25:56) we're going forward with some mass (01:25:57) manufacturing. (01:25:58) >> But do you think current um current (01:26:01) hardware is good enough that you are (01:26:02) going to you should you just want to (01:26:03) deploy as many as possible now? (01:26:06) >> I mean it's very hard to scale up (01:26:07) production. (01:26:08) >> I see. (01:26:08) >> Uh but uh yeah but I I think Optimus 3 (01:26:11) is the the right version of the robot to (01:26:15) you know to to produce maybe something (01:26:18) on the order of like a million units a (01:26:20) year. I think you'd want to go to (01:26:22) Optimus 4 before you went to 10 million (01:26:23) units a year. (01:26:24) >> Okay. But you can do a millionear (01:26:25) adoption with three. (01:26:26) >> Uh yeah, I mean it's very hard to spool (01:26:28) up at manufacturing. (01:26:29) >> Yes. (01:26:29) >> Um so like manufacturing uh um like the (01:26:35) the output per unit time is always (01:26:37) follows an S-curve. So it starts off (01:26:40) agonizingly slow then has this sort of (01:26:43) >> eventually (01:26:43) >> exponential increase then a linear then (01:26:45) a then a you know logarithmic outcome (01:26:48) till you sort of eventually asmtote at (01:26:50) some number. Optimus initial production (01:26:52) will be it's going to be a it's going to (01:26:55) be a stretched out scope because so much (01:26:57) of what goes into Optimus is brand new. (01:27:00) There's not an existing supply chain. Um (01:27:02) as I mentioned the the actuators, (01:27:04) electronics, everything in the office (01:27:06) robot is designed (01:27:08) um from physics first principles. It's (01:27:09) not it's not taken from a catalog. These (01:27:12) these are customd designigned (01:27:14) everything. Literally everything. I (01:27:16) don't think there's a single thing that (01:27:17) um (01:27:18) >> how far down does that go? (01:27:20) I mean, I guess we're not making custom (01:27:22) capacitors yet, maybe. Um, but um (01:27:28) there there's nothing you can pick out (01:27:29) of a catalog um for at any price. Uh so (01:27:33) so it just means that the the Optimus (01:27:36) scope (01:27:38) uh the the units units per output per (01:27:42) unit time how many optimus robots you (01:27:45) make per per day uh whatever is is going (01:27:48) to initially ramp uh slower than a (01:27:52) product where you have an existing (01:27:53) supply chain. Um, but it will get to a (01:27:56) million. (01:27:56) >> When you see these Chinese humanoids (01:27:58) like Unatri or whatever sell humanoids (01:28:01) for like 6K or 13K, do you just like are (01:28:05) you hoping to get your Optimus' bill of (01:28:08) materials below that price so you can uh (01:28:10) do the same thing or do you just think (01:28:11) qualitatively they're not the same (01:28:12) thing? Like what do you think is going (01:28:14) like what allows it what allows them to (01:28:16) sell for so low and can we match that? (01:28:18) Well, optimus our optimus is is designed (01:28:21) to have a lot of intelligence um and um (01:28:25) to have the same electromechanical (01:28:27) dexterity if not higher than a human. (01:28:29) So, (01:28:30) unitary does not have that. And it's (01:28:32) also I mean it's it's quite a it's quite (01:28:35) a big robot. So, it's cuz it's m it has (01:28:37) to do uh you know carry heavy objects (01:28:41) for long periods of time um and not (01:28:44) overheat or exceed the power of its (01:28:46) actuator. So, um, so we've got we've got (01:28:49) we've got, you know, it's it's 5'11, you (01:28:52) know, so it's pretty (01:28:54) >> tall. Um, and it's it's got a lot of (01:28:57) intelligence. So, it's going to be more (01:28:58) expensive than, um, a small robot that (01:29:01) is not intelligent, (01:29:02) >> but more capable. (01:29:03) >> Yeah. But not a lot more. I mean, like (01:29:06) the thing is over time, as Optimus (01:29:08) robots build Optimus robots, the the (01:29:10) cost will drop very quickly. (01:29:12) >> And what will these first billion (01:29:14) Optimuses Optimai? Yeah. (01:29:16) >> Do like what will their highest and best (01:29:18) use be? (01:29:19) >> Uh I think you you would start off with (01:29:21) with simple tasks that you can count on (01:29:22) them doing well. (01:29:23) >> But in the home or in factories like (01:29:26) >> the the best use for (01:29:28) um robots in the beginning will be (01:29:30) anything any um continuous operation. So (01:29:33) any 24x7 operation cuz then you're cuz (01:29:36) they can they can work continuously. (01:29:37) >> Yeah. (01:29:37) >> What fraction of the work at a (01:29:38) gigafactory that is currently done by (01:29:40) humans could a gen 3 do? (01:29:42) >> Um I'm not I'm not sure. Or maybe it's (01:29:44) like 10 20%. (01:29:47) Maybe more. I don't know. It's we would (01:29:49) but we would use we would not like (01:29:52) reduce our headcount. We would we would (01:29:53) for sure it can increase our headcount (01:29:55) to be clear. Um but but we would (01:29:57) increase our output. So the the um units (01:30:01) produced per human like total to total (01:30:04) number of humans at Tesla will increase (01:30:05) but the um the output of robots and cars (01:30:10) will inc will increase disproportionate (01:30:13) like much much to you know number of (01:30:17) cars and robots produced per human will (01:30:19) increase dramatically but but number of (01:30:21) humans will increase as well. We're (01:30:23) talking about Chinese manufacturing um a (01:30:26) bunch here and um we're also talking (01:30:30) about, you know, we've talked about some (01:30:32) of the policies that are relevant like (01:30:34) you mentioned the uh the solar tariffs. (01:30:36) Yeah. (01:30:37) >> Uh and you think they're a bad idea (01:30:39) because, you know, we can't uh scale up (01:30:40) solar in the US. (01:30:41) >> Well, just the electricity output in the (01:30:43) US uh needs to scale up, (01:30:45) >> right? And it can't without like good (01:30:48) power sources. (01:30:49) >> Need to get it somehow. (01:30:50) >> Yeah. But I where I was going with this (01:30:51) is if you were in charge, if you were (01:30:54) setting all the policies, what else (01:30:56) would you change? (01:30:58) >> Um, (01:30:59) >> so you changed the solar tariffs as (01:31:01) well. (01:31:01) >> Yeah, I I would say anything that is (01:31:03) limiting factor for electricity um needs (01:31:06) to be addressed provided it's not like (01:31:08) very bad for the environment. (01:31:09) >> So presumably some permitting reforms (01:31:10) and stuff as well will be in there. (01:31:12) Yeah, (01:31:12) >> there's a fair bit of permitting reforms (01:31:13) that are happening. A lot of the (01:31:15) permitting is state based. Mhm. (01:31:17) >> Um but anything but but but this this (01:31:19) administration is is good at um removing (01:31:22) permitting uh roadblocks. (01:31:24) >> Um (01:31:25) >> and I'm not saying all tariffs are bad. (01:31:27) I'm just saying because I think (01:31:28) >> solar tariffs. (01:31:29) >> So yeah. Yeah. I mean sometimes if like (01:31:31) if another country is subsidizing the (01:31:33) output of of something (01:31:35) >> um then then you have to have (01:31:36) counterveailing tariffs to protect (01:31:39) domestic industry against uh subsidies (01:31:41) by another country. (01:31:42) >> What else would you change? I don't know (01:31:44) if there's that much that the government (01:31:45) can actually do. (01:31:47) >> One thing I was wondering is it seems (01:31:49) like the for the (01:31:52) policy goal of creating a lead for the (01:31:56) US versus China. It seems like the (01:31:58) export bans have actually been quite uh (01:32:02) impactful where China is not producing (01:32:04) leading edge chips and the export bands (01:32:07) really bite there. China is not (01:32:08) producing uh leading edge turbine (01:32:11) engines and similarly there's a bunch of (01:32:13) export bans that are relevant there on (01:32:14) some of the metal energy. Should there (01:32:16) be more export bans like you think about (01:32:19) things like I mean there are now with (01:32:20) the drone industry and things like that (01:32:22) but is that something that should be (01:32:23) considered? (01:32:24) >> Well I think it's important to (01:32:25) appreciate that in most areas China is (01:32:28) very advanced in manufacturing. (01:32:30) >> Um there's only a few areas where it is (01:32:32) not. uh (01:32:35) the you know China is a manufacturing (01:32:38) powerhouse next level like people don't (01:32:41) most people very impressive (01:32:42) >> yeah yeah I mean if you if you take like (01:32:45) refining of of ore um I'd say roughly (01:32:50) China uh does more does twice as much (01:32:53) ore refining of of of on average as the (01:32:57) rest of the world combined (01:32:59) >> um and and I think there there's some (01:33:01) areas like say refining gallium which (01:33:03) goes into solar cells. Um I think (01:33:05) they're like 98% of gallium refining. Um (01:33:09) so so China is actually very advanced in (01:33:11) manufacturing in in I'd say most areas. (01:33:14) >> It seems like we're like there is (01:33:16) discomfort with this supply chain (01:33:18) dependence and yet nothing's really (01:33:19) happening on it. (01:33:20) >> Supply chain which supply chain (01:33:22) dependence (01:33:22) >> depends on say like the gallium refining (01:33:24) that you're saying. (01:33:25) >> Yeah. Yeah. There's there's there's (01:33:26) there there's a there's a (01:33:27) >> well the rare rare earth stuff and (01:33:30) >> yeah rare earths which are as you know (01:33:32) not rare (01:33:33) >> like we actually do do rare earth or (01:33:36) mining in the US send the the (01:33:39) rock uh put it on on a on a train and (01:33:43) then put on a boat to China that goes on (01:33:44) another train and goes to the um railro (01:33:48) refining refiners in China who then (01:33:50) refine it put it into a magnet put into (01:33:52) a motor assembly and then send it back (01:33:54) to America. (01:33:55) So the thing we're really missing a lot (01:33:57) of of all refining um in in America and (01:34:01) >> isn't this worth a policy intervention? (01:34:02) >> Yes. Uh well I think there are some (01:34:06) things being done on on that front. (01:34:08) >> Um but but (01:34:11) we kind of need optimist frankly to to (01:34:14) build refineries. Um, (01:34:16) >> so sir you think the main advantage (01:34:18) China has is the abundance of skilled (01:34:20) labor and that that's like that's that's (01:34:22) a thing optimist fixes (01:34:24) >> but also we need (01:34:25) >> like four times our population (01:34:27) >> but we need so I mean there's this (01:34:28) concern if you think like humanoids are (01:34:30) the future that like okay right now if (01:34:33) it's the skilled labor for manufacturing (01:34:35) that's determining who's who can build (01:34:37) more humanoids you know China has more (01:34:39) of those it manufactures more humanoids (01:34:41) therefore it gets it gets the optimized (01:34:43) future first Um, (01:34:46) it just like keeps that experiment (01:34:47) going. It seems like you're sort of (01:34:48) pointing out that sort of getting to a (01:34:50) million optimi (01:34:51) >> Yeah. (01:34:52) >> requires the manufacturing that the (01:34:54) optimize is supposed to help us get to, (01:34:56) >> right? You you can you can close that (01:34:58) recursive loop pretty quickly (01:35:00) >> with a small number of optimize. (01:35:02) >> Yeah. So, you close the recursive loop (01:35:05) um to help help the robots build the (01:35:07) robots. Um, and then we we can, you (01:35:09) know, try to get to tens of millions of (01:35:10) units a year. Maybe if you start getting (01:35:13) to hundreds of millions of units a year, (01:35:15) I I think you're you're going to be the (01:35:17) most competitive country by far. We (01:35:19) definitely can't win with just humans (01:35:20) because China has four times our (01:35:22) population, (01:35:22) >> right? (01:35:23) >> And frankly, America's been winning for (01:35:24) so long that we, you know, just like a (01:35:26) like a pro sports team that's been (01:35:28) running for a very long time tend to get (01:35:30) complacent and entitled. Um, and that's (01:35:32) why they stop winning. Um, because it's, (01:35:35) you know, don't work as hard anymore. Uh (01:35:38) so I think the frankly just my (01:35:41) observation is the average work ethic in (01:35:42) China is higher than in the US. So it's (01:35:44) not just that there's four times the (01:35:45) population but the work the the amount (01:35:47) of work that people put in is higher. Um (01:35:50) so you you can like you can try to (01:35:51) rearrange the humans but you're still (01:35:53) one quarter of the uh you know assuming (01:35:56) that that productivity is the health is (01:35:59) is the same which I think actually it (01:36:01) might not be. I think China might have (01:36:03) an advantage on productivity per person. (01:36:05) um we will do one quarter the amount of (01:36:08) things as China. Um so so we we can't (01:36:10) win on the human front. Um and our birth (01:36:13) rates been low for a long time. So uh (01:36:17) our birth rates been the US birth rates (01:36:19) been below replacement uh since roughly (01:36:22) 1971 (01:36:24) Um, (01:36:26) so, so we've got a lot of people (01:36:27) retiring or, you know, more people dying (01:36:30) than than than we're close to sort of (01:36:33) more people domestically dying than than (01:36:35) being born. Um, so we definitely can't (01:36:38) win on the human front, but we we might (01:36:39) have a shot at the robot front. (01:36:41) >> Are there other things that you have (01:36:43) wanted to manufacture in the past, but (01:36:45) they've been too labor intensive or too (01:36:47) expensive that now you can come back to (01:36:49) and say, "Oh, we can finally do the (01:36:52) whatever." Uh because we have Optimus. (01:36:54) >> Yeah, I think we'd like to do more build (01:36:57) more um or refineries at Tesla. So um we (01:37:00) just completed um construction and have (01:37:04) um begun lithium refining um without (01:37:07) lithium refinery in Corpus Christie, (01:37:09) Texas. Uh we have um a nickel refinery (01:37:13) which is called the cathode. Uh that's (01:37:15) here in Austin. (01:37:16) >> Mhm. (01:37:16) >> Um and uh these these are the largest (01:37:20) this is the largest cathode. It's the (01:37:22) largest Catholic refinery, largest (01:37:23) lithium refinery, largest nickel and (01:37:26) lithium refinery uh outside of China. (01:37:28) >> Um and it's like the you know the (01:37:33) cathode team would say like we have uh (01:37:35) the the largest and the only actually (01:37:38) cathode refinery in America. Many super (01:37:41) not just the largest but it's also the (01:37:42) only so it's pretty big even though it's (01:37:46) the only one. Um, but I mean there are (01:37:48) other things that uh you know um (01:37:53) you could do a lot more refineries and (01:37:56) and um help the the help America be more (01:37:59) competitive on refining capacity. So So (01:38:02) there's like there's basically a lot of (01:38:05) work for for the optimal to do uh that (01:38:08) that most Americans very few Americans (01:38:10) frankly want to do. Uh I I mean I've (01:38:12) I've actually (01:38:13) >> Is the refining work too dirty or what's (01:38:15) the (01:38:15) >> it's not is actually no we don't um (01:38:18) there's not we don't have toxic (01:38:20) emissions from the refinery or anything. (01:38:21) Um like the cathode make a refiner right (01:38:25) sort of in Travis County like 5 minutes (01:38:27) from to (01:38:28) >> Why can't you do it with humans? (01:38:29) >> No you you can't you run out of humans. (01:38:31) >> Ah I see. Okay. Yeah. (01:38:32) >> Like no matter what you do you have one (01:38:34) quarter the number of humans in America (01:38:36) in China. So if you have them do this (01:38:37) thing they can't do the other thing. So, (01:38:39) so then then um well, how do you how do (01:38:42) you build this refining refining (01:38:43) capacity? Well, you could do it with (01:38:44) Optima. Um and um not many not very many (01:38:49) not very many Americans are are pining (01:38:52) to do refining. (01:38:54) [laughter] (01:38:56) >> I mean, how many have you run into? (01:38:58) >> Very few. (01:39:00) >> What are you saying? (01:39:01) >> Very few plan to refine. (01:39:02) >> You know, BYD is reaching Tesla (01:39:04) production or sales in quantity. (01:39:08) What do you think happens in global (01:39:10) markets as Chinese production in EV (01:39:12) scales up? (01:39:14) >> Um well (01:39:16) uh China's extremely competitive in (01:39:18) manufacturing. So um I think this (01:39:22) there's going to be a a massive flood of (01:39:25) Chinese vehicles and and and and other (01:39:30) basically most manufactured uh things. I (01:39:34) mean as it is as I said like China is (01:39:36) like probably does twice as much (01:39:38) refining as the rest of the world (01:39:39) combined. So if you go (01:39:43) you know if you just go go down to like (01:39:46) fourth and fifth tier uh supply chain (01:39:49) stuff like like like at the base level (01:39:51) we've got energy then then you've got (01:39:53) mining and refining. um those those (01:39:56) those foundation layers uh are (01:40:00) like I said China as a rough guess China (01:40:03) is doing twice as much refining as the (01:40:04) rest of the world combined. So any given (01:40:06) thing is going to have uh Chinese (01:40:10) content because China is doing twice as (01:40:12) much manufact refining work as the rest (01:40:14) of the world. Um (01:40:16) and uh and then they they'll go all the (01:40:19) way to the finished product with the (01:40:20) cars. Uh I mean China is a powerhouse. I (01:40:23) mean I think this year China will exceed (01:40:26) three times US electricity output. (01:40:29) >> Mhm. (01:40:30) >> Like electricity output is a is a (01:40:32) reasonable proxy for uh (01:40:35) you know for the economy. (01:40:38) Uh so like in order to run the factories (01:40:40) and run run everything, you need (01:40:41) electricity. So electricity is is is a (01:40:46) it's a good proxy for the for the real (01:40:48) economy. Um and so if China is (01:40:53) if China passes three times the US (01:40:54) electricity output, it means that its (01:40:56) industrial capacity (01:40:58) as a rough approximation is three times (01:41:00) that will be three times that of the US. (01:41:02) Reading between the lines, it sounds (01:41:03) like what you're sort of saying is (01:41:05) absent some sort of humanoid recursive (01:41:07) miracle in the next few years on the the (01:41:10) sort of like whole manufacturing energy (01:41:13) uh raw materials chain like China will (01:41:17) just dominate whether it comes to like (01:41:19) AI or manufacturing EVs or manufacturing (01:41:21) humanoids (01:41:23) in the absence of of um (01:41:27) breakthrough innovations uh in in the (01:41:30) US. uh China will uh utterly dominate. (01:41:35) >> Interesting. (01:41:36) >> Yes. (01:41:37) >> Robotics being the main breakthrough (01:41:38) innovation. (01:41:40) >> Well, if you do (01:41:43) like to to scale AI (01:41:45) uh in in space like like basically need (01:41:48) >> space, (01:41:49) >> you need need the humanoid robots, you (01:41:52) need real world AI, you need um a (01:41:54) million tons a year to orbit. Um like (01:41:57) let's just say like if we get the mass (01:41:59) driver on the moon going my favorite (01:42:00) thing. Um then I think uh (01:42:04) >> we'll have solved all our problems. (01:42:06) >> Yeah. So this is like I call that (01:42:09) winning. [laughter] (01:42:11) >> Um (01:42:12) >> I'd call that winning (01:42:13) >> time. (01:42:14) >> You can finally be satisfied you've done (01:42:16) something. (01:42:16) >> Yes. (01:42:17) >> You have the mass driver on the moon. (01:42:18) >> That's right. I just want to see that (01:42:19) thing in operation. (01:42:20) >> Was that out of some sci-fi or where did (01:42:22) you (01:42:22) >> uh Well, actually the there is a (01:42:24) Highland book. The moon. The moon is a (01:42:26) harsh race. (01:42:26) >> Okay. Yeah, but that's slightly (01:42:28) different. That's a gravity slingshot or (01:42:30) um (01:42:30) >> No, they have a mass driver on the moon. (01:42:32) >> Okay. Yeah, but they use that to attack (01:42:35) Earth. So maybe it's not the greatest. (01:42:36) >> They use that to uh assert their (01:42:38) independence from (01:42:39) >> Exactly. What are your plans for the (01:42:40) mass driver on the moon? (01:42:41) >> They they assed their independence. (01:42:42) [clears throat] (01:42:43) Earth government disagreed and they (01:42:44) loved things until they earth government (01:42:46) agreed. (01:42:47) >> That book is a hoot. I found that book (01:42:48) much better than um his other one that (01:42:51) everyone reads um Stranger in a Strange (01:42:52) Land. Yeah, Grock. Grock comes from a (01:42:54) stranger on a strange line. (01:42:55) >> Yeah. Yeah. But I much preferred. (01:42:56) >> Yeah. The first two thirds of Stranger (01:42:58) Lines are good and then it gets very (01:42:59) weird in the third. Yeah. [snorts] (01:43:02) >> Um but there's still some good concepts (01:43:04) in there. (01:43:05) >> Yeah. (01:43:05) >> Label box can get you robotics and RL (01:43:07) data at scale. Take robotics. Let's say (01:43:10) you need 100,000 hours of egocentric (01:43:12) video. Labelbox starts by helping you (01:43:14) define your ideal data distribution. (01:43:16) Like for example, maybe no single task (01:43:19) category should occupy more than 1% of (01:43:21) trading volume. and at least 10% of (01:43:23) trajectories should capture failure and (01:43:25) recovery states. Next, Labelbox assigns (01:43:28) its distribution to its massive network (01:43:30) of operators. You're not limited to the (01:43:32) small range of scenes that you can set (01:43:34) up in a single warehouse. Instead, each (01:43:36) one of Label Box's operators has access (01:43:38) to lots of unique physical environments (01:43:39) where they can film themselves (01:43:41) completing a wide variety of tasks. (01:43:44) Labelbox's tech automatically (01:43:45) categorizes each video so that their (01:43:47) operators always know which tasks will (01:43:49) remain and what they need to work on (01:43:51) next. For RL data, Labelbox takes a (01:43:54) similar approach. They work with you to (01:43:55) understand the right distribution of (01:43:56) tasks and then their subject matter (01:43:58) experts build the hyperrealistic digital (01:44:00) environments and rubrics [music] that (01:44:02) you need to collect the highest quality (01:44:05) trading data. So whether you're training (01:44:06) robots in the real world or agents for (01:44:08) computer use, Labelbox can help. Go to (01:44:11) labelbox.com/sarcash (01:44:13) [music] (01:44:14) to learn more. (01:44:17) >> One thing we were discussing a lot is (01:44:19) kind of your system for managing people. (01:44:22) Like you interviewed the first few (01:44:25) thousand of SpaceX employees and lots of (01:44:27) other companies. (01:44:28) >> What [snorts] is it doesn't scale? (01:44:29) >> Well, yes, but but what doesn't scale (01:44:32) >> me? I mean, (01:44:33) >> sure. Sure. [laughter] (01:44:34) >> I know that. But like what are you (01:44:36) looking for? (01:44:36) >> I mean, it literally is not enough hours (01:44:37) in the day. It's impossible. But well (01:44:38) but um what are you looking for that's (01:44:41) someone else who's good at interviewing (01:44:43) and hiring people? What's the Jenna? (01:44:46) >> Um well at this point I think I've got (01:44:49) um I might have more training data on (01:44:52) evaluating technical talent especially (01:44:54) but talent of all kinds I suppose but uh (01:44:56) technical talent especially um given (01:44:58) that I've done so many technical (01:44:59) interviews and then seen the results (01:45:01) technical interviews seen the results. (01:45:02) So my um my training set is is is very (01:45:06) is enormous and uh has a very wide (01:45:09) range. Um (01:45:11) uh the generally the thing I ask for are (01:45:13) u bullet points uh for evidence of of (01:45:17) exceptional ability. So it's uh but like (01:45:21) it's it's and these things can be like (01:45:23) pretty offthe-wall. It doesn't need to (01:45:24) be uh in the in the domain the specific (01:45:27) domain but evidence that uh evidence of (01:45:29) exceptional ability. Um so if some if if (01:45:32) somebody can like site like even one (01:45:34) thing but let's say three things where (01:45:36) you go wow wow wow then that's that's a (01:45:38) good sign. (01:45:39) >> But but why do you have to be the one to (01:45:41) determine that presumably it's (01:45:42) impossible (01:45:43) >> right? (01:45:43) >> I mean total headc count across all (01:45:45) companies 200,000 people (01:45:47) >> right [laughter] (01:45:49) >> but in the early days what was it that (01:45:52) that you were looking for that couldn't (01:45:53) be delegated in those interviews? (01:45:57) Um, (01:46:00) well, I I guess I I'd need to build my (01:46:02) training set. It's not like I would b a (01:46:03) thousand here. Um, I would make (01:46:05) mistakes, but then I'd be able to see (01:46:07) where I I thought somebody would work (01:46:09) out well, but they didn't. And and then (01:46:10) why why did they not work out well and (01:46:13) what can I do to I guess RL myself to uh (01:46:17) in the future um have a better batting (01:46:19) average when interviewing people. (01:46:21) >> Mhm. (01:46:22) >> So, and my batting average is still not (01:46:24) perfect, but it's it's very high. What (01:46:25) are some surprising reasons people don't (01:46:26) work out? (01:46:28) >> Surprising reasons? Um, (01:46:29) >> like they don't understand technical (01:46:30) domain, etc., etc., but like you you (01:46:33) like you you've got like the long tail (01:46:34) now of like I was really excited was (01:46:36) about this person, it didn't work out. (01:46:38) Curious why that happens. (01:46:41) >> Uh, yeah. So the (01:46:44) I mean generally what I tell people or (01:46:46) tell myself I guess aspirationally um is (01:46:49) don't look at the resume just believe (01:46:51) your interaction. M (01:46:53) >> so if the resume may may seem very (01:46:55) impressive and it's like wow you know (01:46:56) resume looks good but if the if the (01:46:59) conversation uh after 20 minutes is is (01:47:02) that conversation is not wow um you (01:47:04) should believe the conversation not the (01:47:06) not the not the paper. I feel like part (01:47:09) of your method is that you know there (01:47:11) was this meme in the media a few years (01:47:14) back about Tesla being a revolving door (01:47:16) of uh executive talent whereas actually (01:47:18) I think when you look at it Tesla's had (01:47:21) a very consistent and internally (01:47:22) promoted executive bench over the past (01:47:24) few years and that at SpaceX you have (01:47:26) all these folks like Mark Josa and Steve (01:47:28) Davis and (01:47:30) >> Steve Davis runs a boring company these (01:47:31) >> no now yeah but Bill Riley and folks (01:47:34) like that (01:47:35) >> and it feels like part of has worked (01:47:38) well is having very capable technical (01:47:42) deputies. (01:47:43) What do all of those people (01:47:46) have in common? (01:47:49) >> Uh well, so the I mean the Tesla is a (01:47:52) sort of senior team uh at this point (01:47:54) probably got average tenure of 10 or 12 (01:47:57) years. It's quite quite a long (01:47:58) >> tenure. Yeah. Um so um but (01:48:05) there there were times where Tesla went (01:48:06) through extremely rapid an extremely (01:48:09) rapid growth phase um and so it was (01:48:12) somewhat things were just somewhat sped (01:48:13) up um and and when a company as as I'm (01:48:16) as you know company goes through (01:48:17) different orders of magnitude of of size (01:48:19) you you know uh people that could who (01:48:22) who could help manage say a 50 person (01:48:24) company versus a 500 person company (01:48:26) versus a 5,000 person company versus a (01:48:28) 50,000 (01:48:30) group people. (01:48:31) >> Yeah, it's it's just not the same team. (01:48:33) It's not it's not always the same team. (01:48:34) So if if a company is growing very (01:48:36) rapidly, the the rate at which uh (01:48:39) executive positions will change will (01:48:40) also be proportionate to the the (01:48:42) rapidity of the growth generally. Um (01:48:45) then uh Tes Tesla had uh a further (01:48:49) challenge where when when Tesla had very (01:48:52) successful periods um uh we would be um (01:48:56) relentlessly recruited from um like (01:48:59) relentlessly um like when Apple had (01:49:02) their electric car program they were (01:49:04) coet bombing Tesla with recruiting (01:49:06) calls. It was engineers just unplugged (01:49:09) their phones like it's just it's just I (01:49:11) >> I'm trying to get work done here. (01:49:12) >> Yeah. if I get, you know, one more call (01:49:14) from an Apple recruiter. Um, but but (01:49:16) they were they were their opening offer (01:49:18) without any interview with me like (01:49:19) double the compensation at Tesla. Um, so (01:49:23) um, so so we had a bit of the Tesla (01:49:28) Pixie Dust thing where it's like, oh, if (01:49:30) you hire a Tesla executive, you're (01:49:32) suddenly you're going to everything's (01:49:34) going to be successful. Um, and and I (01:49:36) fallen prey to the Pixie Dust uh, you (01:49:38) know, thing as well where it's like, oh, (01:49:39) we'll hire someone from Google or Apple (01:49:41) and they'll be immediately successful. (01:49:42) But not that that's not how it works. Um (01:49:45) you know people are people it's there's (01:49:46) not like magical pixie dust. (01:49:48) >> Yes. (01:49:48) >> So when we have the pixie dust problem (01:49:52) um we get relentlessly recruited um and (01:49:55) um (01:49:57) and then also being Tesla being um (01:50:00) engineering especially being primarily (01:50:01) in Silicon Valley uh it's it's easier (01:50:04) for people to just like they don't have (01:50:06) to change their life very much. (01:50:08) >> They can just you know (01:50:10) >> their commute's going to be the same. (01:50:12) Yes. (01:50:12) >> Um, (01:50:13) >> so how do you prevent that? How do you (01:50:15) prevent the pixie dust effect for (01:50:16) everyone's trying to coach all your (01:50:18) people? (01:50:19) >> Um, I don't think we can I don't think (01:50:22) there's much we can do to to yeah, stop (01:50:24) it. Uh but that that's like that's one (01:50:26) of the reasons why Tesla uh really being (01:50:30) in Silicon Valley um and uh and having (01:50:35) the Pixie Dust thing at the same time um (01:50:38) meant that uh there was just a very very (01:50:41) aggressive recruitment. (01:50:43) >> Being in Austin helps then. (01:50:45) >> Uh Austin, yeah, it still helps. Uh I (01:50:48) mean Tesla still has a majority of its (01:50:50) engineering in California. Um so um (01:50:55) the you know for getting engineers to (01:50:57) move I called the significant (01:50:59) significant other problem. (01:51:00) >> Yes. (01:51:01) >> So others have jobs. Yeah. (01:51:02) >> Yeah. Yeah. Exactly. So um for Starbase (01:51:06) that was particularly difficult. (01:51:07) >> Yes. (01:51:07) >> Since the odds of you know finding (01:51:09) [snorts] a non SpaceX job (01:51:12) >> pretty low. Yeah. Yeah. It's quite quite (01:51:14) difficult. I mean it's like a technology (01:51:16) monastery thing. (01:51:17) >> Um you know remote and mostly dudes. (01:51:21) >> [laughter] (01:51:22) >> But again, if you if you go (01:51:25) >> much of an improvement over myself, (01:51:26) [laughter] (01:51:27) >> yeah, if you go back but if you go back (01:51:29) to these people who've really (01:51:33) um (01:51:35) been very effective in a technical (01:51:37) capacity at Tesla, at SpaceX, and and (01:51:40) those sorts of places, what do you think (01:51:42) they have in common other than like is (01:51:46) it just that they're very sharp on the, (01:51:47) you know, rocketry or the, you know, the (01:51:50) technical foundation? (01:51:51) Or do you think it's something (01:51:52) organizational? It's something about (01:51:53) their ability to work with you. Is it (01:51:55) their ability to like be, you know, (01:51:59) flexible but not too flexible? (01:52:04) >> What makes a good sparring partner for (01:52:06) you? (01:52:07) >> I don't think a sparring partner. I (01:52:08) mean, if if somebody gets things done, I (01:52:10) I I love them and if they don't, I So, (01:52:13) it's pretty straightforward. It's not (01:52:14) like some idiosyncratic uh thing. Um, if (01:52:18) somebody executes well, um, I'm a huge (01:52:20) fan and if they don't, I'm not. Um, but (01:52:23) it's it's not about mapping to my is (01:52:25) idiosyncratic preferences. I'll (01:52:26) certainly try not to have it be mapping (01:52:28) to my idiosyncratic preferences. (01:52:29) >> Mhm. (01:52:29) >> Um, so yeah. Um, (01:52:35) yeah, but generally I think it's a good (01:52:37) idea to hire for um, uh, talent and (01:52:42) drive and trustworthiness. M (01:52:45) >> um (01:52:47) and I think uh goodness of heart is (01:52:49) important. Um I I'd wait at that at one (01:52:52) point. Um (01:52:54) >> so like are they are they a good person, (01:52:56) trustworthy (01:52:58) uh sort of smart, talented and (01:53:00) hardworking? Uh if so you can add domain (01:53:03) knowledge. U but those those fundamental (01:53:05) traits those fundamental properties you (01:53:07) cannot change. So mo most of the people (01:53:10) who um are at (01:53:13) Tesla and SpaceX did not come from the (01:53:15) aerospace industry or the auto industry. (01:53:18) What is most sad to change about your (01:53:20) management style as your companies have (01:53:22) scaled from 100 to thousand to 10,000 (01:53:24) people. You're you know you're known for (01:53:26) this like very micromanagement just (01:53:28) getting into the details of things. (01:53:29) >> Nano management please. [laughter] (01:53:32) >> Pico management. (01:53:33) >> Um (01:53:34) >> phantom management. (01:53:36) So you're saying [laughter] (01:53:39) >> we're gonna go all the way down to (01:53:40) flanks Boston (01:53:45) all the way down to Heisenberg. I said (01:53:46) they were small. [laughter] (01:53:48) >> Yeah. How do you I mean are you are you (01:53:51) still able to get into details as much (01:53:52) as you want? Would your companies be (01:53:54) more successful if you could if they (01:53:55) were smaller? Like how do you how do you (01:53:56) think about that? Well, because I have a (01:53:58) fixed amount of time in the day, uh my (01:54:01) time is necessarily um diluted as things (01:54:04) grow and as the span of activity uh (01:54:06) increases. So, you know, um it it it (01:54:10) it's it's impossible for me to actually (01:54:12) be a micromanager because uh there's (01:54:15) that that would imply I have some like (01:54:18) thousands of hours per day. It is a (01:54:21) logical impossibility for me for me to (01:54:23) mic to micromanage things. Um so now (01:54:27) there are times when um I will drill (01:54:29) down into uh a specific issue because (01:54:33) that specific issue uh is the limiting (01:54:36) factor on uh the progress of the (01:54:39) company. Um and um but the the reason (01:54:43) for drilling into that that some very (01:54:46) detailed item is because it is the it is (01:54:48) the limiting factor not it's not (01:54:50) arbitrarily d drilling into you tiny (01:54:53) things. Um and and like I said obviously (01:54:56) from a time standpoint it is physically (01:54:58) impossible for me to arbitrarily uh go (01:55:00) into tiny things that don't matter and (01:55:02) that would and and that would result in (01:55:03) failure but sometimes the tiny things um (01:55:06) are decisive in victory. (01:55:10) >> Famously you switched the uh starship (01:55:13) design from composits to steel. (01:55:17) >> Yes. (01:55:18) >> And you made that decision like that (01:55:19) wasn't a you know people were going (01:55:21) around they're like oh we found (01:55:22) something better boss. like that was you (01:55:23) encouraging people against some (01:55:24) resistance. Can you tell us how you came (01:55:27) to that whole composite steel switch? (01:55:31) [snorts] Uh yeah. So desperation I'd (01:55:34) say. Um the um (01:55:38) originally yeah we were going to make (01:55:40) stashup out of uh carbon fiber. Um and (01:55:44) um car fiber is pretty expensive like (01:55:47) the the the the (01:55:52) can generally uh when you do volume (01:55:54) production you can get any given thing (01:55:55) to be to start to approach its material (01:55:58) cost. The problem with with carbon (01:55:59) fibers is that material cost is still (01:56:02) very high. Um (01:56:04) um so (01:56:06) um it's about it's about 50 times (01:56:09) particularly if you go for a high (01:56:10) strength specialized carbon fiber that (01:56:13) can handle um cryogenic oxygen it's it's (01:56:15) it's like quot roughly 50 times the cost (01:56:18) of steel um and at least uh in theory it (01:56:23) would be lighter. People generally think (01:56:24) of steel as being heavy and carbon fiber (01:56:26) as being uh light. Um, and for room (01:56:29) temperature room temperature (01:56:30) applications, (01:56:32) um, you know, like say, uh, more or less (01:56:34) room temperature applications like a (01:56:36) Formula 1 car, uh, static aeros (01:56:38) structure or any any kind of aeros (01:56:40) structure really, uh, is is going to (01:56:43) you're going to probably be better off (01:56:44) with the carbon fiber. Um, now the (01:56:46) problem is that we were trying to make (01:56:48) this enormous rocket out of carbon fiber (01:56:50) and, uh, our progress was extremely slow (01:56:54) >> and it had been picked in the first (01:56:55) place just because it's light. (01:56:57) Yes. Um like at first glance (01:57:02) um like most people would think that the (01:57:04) the the choice for making uh something (01:57:06) light would be carbon fiber. Um the um (01:57:12) now now the thing is that um (01:57:16) we when you make something very enormous (01:57:19) out of carbon fiber and then you try to (01:57:21) have the carbon fiber um be efficiently (01:57:25) cured. anything not not room temperature (01:57:27) because like you've got you know (01:57:30) sometimes you got like 50 pies of of of (01:57:32) carbon fiber and and carbon fiber is (01:57:34) really carbon string and glue. Um and uh (01:57:37) and you in order to have um high (01:57:39) strength you need an autoclave. So (01:57:42) something that that can that's (01:57:43) essentially high pressure oven. And if (01:57:45) if um if you have something that's a (01:57:48) gigantic uh the oven's got to be bigger (01:57:51) than the rockwood. Um, so we'll be (01:57:53) trying to make the the an autoclave (01:57:55) that's bigger than any autoclave that's (01:57:57) ever existed. Uh, or do room temperature (01:58:00) cure, which takes a long time and and (01:58:01) has issues. Um, but but the fundamental (01:58:04) issue is that we're just making very (01:58:05) slow progress uh with uh with cotton (01:58:08) fiber. Um, (01:58:10) so um I I think the meta question is u (01:58:14) why it had to be you who made that (01:58:18) decision. There's many engineers on your (01:58:19) team. (01:58:20) >> Yeah. How did the team not arrive at (01:58:21) Steel? Yeah, exactly. This is a part of (01:58:23) a broader question like understanding (01:58:24) your comparative advantage at your (01:58:25) companies. (01:58:27) >> Um, so it was because we were making (01:58:29) very slow progress with with carbon (01:58:31) fiber. I was like, "Okay, we've got to (01:58:33) try something else." Now, for the Falcon (01:58:36) 9, the the primary airframe is made of (01:58:39) aluminum lithium, which is has very very (01:58:42) good strength weight. Um and um actually (01:58:46) it has uh about the same maybe maybe (01:58:48) better strength weight for its (01:58:50) application than carbon fiber. But (01:58:52) aluminum lithium is very difficult to (01:58:53) work with. In order to weld it you have (01:58:55) to do something called friction st (01:58:56) welding where you join the you join the (01:58:58) metal without it entering the liquid (01:58:59) phase. Um so it's kind of wild that you (01:59:02) could do that. Uh but with this (01:59:04) particular type of welding you can do (01:59:05) that. Um (01:59:07) but uh it's very difficult to like say (01:59:10) let's say you want to make a (01:59:11) modification or attach something to um (01:59:14) aluminum lithium. You now have to use (01:59:16) mechanical attachment with seals. Um you (01:59:18) can't uh weld it on. Um so we want to I (01:59:23) want to avoid using aluminum lithium for (01:59:25) the primary structure for uh for (01:59:28) Starship. um and uh and and there was (01:59:32) this very special grade of uh carbon (01:59:35) fiber that that had you know very good (01:59:38) mass properties. So with rock rocket (01:59:40) you're really trying to maximize the (01:59:42) percentage of the of the rocket that is (01:59:43) propellant minimize the the mass (01:59:45) obviously and um the but like I said (01:59:49) we're making very slow progress um and (01:59:53) and and I said at this rate we're never (01:59:55) going to get to Mars so we're going to (01:59:58) think of something else um I didn't want (02:00:00) to use aluminum lithium because of the (02:00:02) difficulty of friction still welding um (02:00:05) especially doing that at at scale it was (02:00:07) hard enough um at 3.6 m in diameter, let (02:00:10) alone at 9 m or above. Um then um (02:00:16) it says, well, what about steel? And and (02:00:18) so the now I I I had a clue here because (02:00:21) some of the early um US rockets had used (02:00:25) very thin steel. The Atlas rockets had (02:00:27) used a steel balloon tank. Um (02:00:31) so it's not like steel never been used (02:00:32) before. It actually had been used. Um (02:00:35) and when you look at the the material (02:00:36) properties of stainless steel um (02:00:38) especially uh very uh if it's been (02:00:42) very like full hard strain hardened (02:00:45) stainless steel uh at cryogenic (02:00:48) temperature uh the the strength weight (02:00:51) is actually similar to carbon fiber. So (02:00:54) if if you look at the so if you look at (02:00:56) the material properties at room (02:00:57) temperature um it looks like the steel (02:00:59) is uh is going to be twice as heavy. But (02:01:02) if you look at the material properties (02:01:04) at cryogenic temperature of full hard (02:01:06) steel stainless of of particular grades (02:01:09) uh then the the you actually get to a (02:01:12) similar strength weight as common fiber (02:01:15) and and in the case of Starship both the (02:01:17) fuel and the oxidizer are cryogenic. So (02:01:20) for for uh Falcon 9, the fuel is rocket (02:01:24) propellant grade kerosene basically pure (02:01:26) like a a very pure form of jet fuel. (02:01:30) >> Um which is but but that is that is (02:01:33) roughly room temperature. Um although we (02:01:35) do actually we do actually chill it (02:01:37) slightly below we chill it like a beer. (02:01:39) Um delicious. (02:01:41) >> Yeah, we we do chill it but um but but (02:01:43) it's not cryogenic. In fact, if we made (02:01:45) it cryogenic, it would just turn to wax. (02:01:48) So um but for Sasha the it's liquid (02:01:51) methane and and liquid oxygen they they (02:01:53) uh they're liquid at at similar (02:01:55) temperatures. Uh so uh so basically (02:02:00) almost the entire primary structure is a (02:02:02) cryogenic temperature. So then you've (02:02:04) got uh u a 300 series stainless that's (02:02:09) that's um strain hardened uh because (02:02:13) it's at almost all things at crying (02:02:15) temperature actually has a similar (02:02:18) strength to weight as uh carbon fiber (02:02:22) but costs uh 50 times less than nor (02:02:25) material and is very easy to work with. (02:02:28) You you can weld stainless steel (02:02:29) outdoors. (02:02:31) You could smoke a cigar while welding (02:02:33) stainless steel. It's it's like it's (02:02:34) it's very resilient. (02:02:36) >> Um you you can modify it easily. It's (02:02:39) it's uh if you want if you want to (02:02:41) attach something, you just weld it right (02:02:42) on. So um very easy to work with (02:02:47) uh very low cost um and um like I said (02:02:52) at cryogenic temperature, similar (02:02:54) strength to weight uh to carbon fiber. (02:02:57) Um then when you factor in that uh that (02:03:00) we don't need we don't we we have a much (02:03:03) reduced uh heat shield mass uh because (02:03:05) the melting point of steel is much (02:03:07) greater than the melting point of (02:03:09) aluminum. Um it's about twice the (02:03:12) melting point of alum aluminum and (02:03:14) >> so you can just run the rocket bunch (02:03:15) hotter. (02:03:16) >> Yes. So especially for the ship um which (02:03:19) is coming in like a fl a blazing meteor (02:03:22) uh it is uh the you you you can greatly (02:03:25) reduce the mass of the heat shield um so (02:03:28) that so you can call it cut the mass of (02:03:32) the windward u part of the heat shield (02:03:36) in maybe in half (02:03:37) >> and you don't need any heat shielding on (02:03:40) the on the leeward side. Um so um (02:03:46) the the net net result is actually the (02:03:48) steel rocket weighs less than the carbon (02:03:50) fiber rocket (02:03:51) >> because the the resin in the carbon (02:03:53) fiber rocket uh uh (02:03:56) um starts to melt. Um (02:04:00) so basically carbon fiber and aluminum (02:04:03) have about the same operating (02:04:04) temperature uh capabilities um and (02:04:07) whereas steel can operate at twice (02:04:09) temperature. I mean, these are very (02:04:10) rough approximations. People will (02:04:12) >> I won't go to the rocket face. What I'm (02:04:14) like people will say, "Oh, he said it's (02:04:16) twice. It's actually it's actually (02:04:17) point8." Shut up, [ __ ] (02:04:19) >> That's what the main comment's going to (02:04:20) be about. Yeah. (02:04:20) >> God damn it. Okay. The point is the the (02:04:23) actually in retrospect the the we should (02:04:26) have started with done steel in the (02:04:27) beginning. It was dumb not to do steel. (02:04:28) >> Okay. But to play this back to you, what (02:04:29) I'm hearing is that steel was a riskier, (02:04:33) less proven path other than the early US (02:04:36) rockets versus carbon fiber was like a (02:04:40) worse but more proven out path. And so (02:04:43) you need to be the one to push for, hey, (02:04:45) we're going to do this riskier path and (02:04:47) just figure it out. And so you were (02:04:49) fighting like a sort of conservatism in (02:04:51) a sense. (02:04:52) >> Um that's why I initially said like that (02:04:54) the issue is that we weren't making fast (02:04:56) enough progress. we were having trouble (02:04:58) making even um a small barrel section of (02:05:01) the carbon fiber um that didn't have (02:05:04) wrinkles in it. (02:05:05) >> Um so uh because at at at that large (02:05:08) scale you have to have many pies many (02:05:10) sort of layers of the carbon fiber. Um (02:05:13) you've got to cure it and you've got to (02:05:14) cure it in such a way that it it doesn't (02:05:16) um have any wrinkles or or defects. The (02:05:19) carbon fiber is much less resilient than (02:05:22) than steel. It has much it's less (02:05:24) toughness. Um so like like stainless (02:05:26) steel will will stretch and and and bend (02:05:30) the carbon fiber will will tend to (02:05:31) shatter. (02:05:33) >> Um so um so toughness being the area (02:05:37) under the stress drain curve um so that (02:05:39) you're generally going to have to do (02:05:41) better with steel um or stainless steel (02:05:44) to be precise. (02:05:45) >> One other starship question. Um so I (02:05:48) visited um Starbase I think it was two (02:05:50) years ago with Seller and that was (02:05:52) awesome. It was very cool to see in a (02:05:54) whole bunch of ways. One thing I noticed (02:05:56) was that people really took pride in the (02:06:00) simplicity of things where you know (02:06:03) everyone wanted to tell you how Starship (02:06:05) is just a big soda can and you know (02:06:07) we're hiring welders and you know if you (02:06:09) can weld in any industrial project you (02:06:11) can weld here but um there's a lot of (02:06:14) pride in the simplicity and (02:06:16) >> it's well startup was a very complicated (02:06:18) >> rocket. So that that's what I'm getting (02:06:20) at is are things simpler or are they (02:06:22) complex? (02:06:23) >> I think maybe just what they're trying (02:06:25) to say is that you know you don't have (02:06:26) to have like prior experience in the (02:06:28) rocket industry to work on Sasha. (02:06:30) >> Um you know somebody just needs to be (02:06:34) >> you know smart and work hard (02:06:37) >> um and be trustworthy then they can work (02:06:39) on a rocket. They don't they don't need (02:06:41) prior rocket experience. Starship is is (02:06:43) the most complicated machine ever made (02:06:45) by humans by a long shot. (02:06:48) >> In in what regards? (02:06:50) >> Anything really? I'd say there isn't a (02:06:52) more complex machine. Um (02:06:56) there Yeah, I mean I I'd say that there (02:06:58) there pretty much any any project I can (02:07:00) think of would be easier than this. Um (02:07:02) and and that's why no one has made a (02:07:06) rapidly reusable nobody has ever made a (02:07:08) re fully reusable orbital rocket. It's a (02:07:10) very very hard problem. (02:07:12) Um (02:07:14) I mean many smart people have tried (02:07:17) before very smart people with immense (02:07:20) resources and they failed. Um so and we (02:07:23) haven't succeeded yet. uh you know (02:07:25) Falcon is partially reusable but the (02:07:27) upper stages are um Starship version (02:07:31) three I think this design (02:07:36) that it it can be fully reusable and (02:07:39) that full reusability is what will (02:07:41) enable us to become a multilanet (02:07:44) civilization. Can you say about the (02:07:46) >> I don't I'm like I said like I (02:07:51) any technical problem even like a hydron (02:07:52) collider or something like that is (02:07:54) easier following than this. (02:07:55) >> We we spent a lot of time on (02:07:56) bottlenecks. Can you say what the (02:07:58) current Starship bottlenecks are even at (02:08:00) a high level? (02:08:01) >> I mean trying to make it not explode (02:08:04) generally [snorts] (02:08:05) >> that old chestnut (02:08:07) >> really wants to explode. Um (02:08:10) >> those combustion we've had two bristers (02:08:12) explode on the test. Um, one obliterated (02:08:15) obliterated the entire test facility. (02:08:18) Um, so it only takes like one mistake (02:08:20) and and I mean the amount of energy (02:08:22) contained in in a Starship [snorts] is (02:08:25) insane. (02:08:26) >> And so is that why it's harder than (02:08:27) Falcon? It's because it's just more (02:08:28) energy. (02:08:30) >> It's a lot of new technology. Um, it's (02:08:33) it's p it's pushing the performance (02:08:35) envelope. Um, the Raptor 3 engine is a (02:08:40) very very advanced engine. By far the (02:08:41) best rocket edition ever made. Um but it (02:08:44) desperately wants to blow up. I mean (02:08:46) just to put things in perspective here (02:08:49) on liftoff um the the rocket is (02:08:51) generating over 100 g of power. (02:08:55) It's 20% of US (02:08:57) >> electricity. So insane. (02:08:58) >> That's a great comparison. (02:09:00) >> While not exploding (02:09:02) >> sometimes. (02:09:02) >> Sometimes but sometimes. Yeah. So I was (02:09:05) like how does it not explode? there's (02:09:07) there's a you know thousands of ways (02:09:09) that it could explode and and only one (02:09:11) way that that that it doesn't. So So we (02:09:14) want it to not really not explode but (02:09:16) but fly reliably (02:09:18) uh (02:09:20) on a daily basis like once per hour and (02:09:22) obviously you know blows up a lot. It's (02:09:24) it's very difficult to maintain that (02:09:26) floor cings. (02:09:26) >> Yes. (02:09:27) >> Um and [snorts] then and then I'm going (02:09:30) to say like what's the what's the single (02:09:32) biggest remaining problem for Starship? (02:09:34) It's uh having the heat shield be (02:09:37) reusable. Um that such that the no no (02:09:40) one has ever made a reusable orbital (02:09:43) heat shield. Um so the the sh the heat (02:09:46) shield's got to make it through the (02:09:48) ascent phase without shocking a bunch of (02:09:51) tiles. Um and then it's going to come (02:09:53) back in and also not lose a bunch of (02:09:56) tiles or or overheat the the main the (02:09:59) main uh airframe. (02:10:01) >> Isn't that hard? is kind of (02:10:02) fundamentally a consumable. (02:10:05) >> Uh well, yes, but your brake pads in (02:10:07) your car are also consumable, but they (02:10:09) last a very long time. (02:10:09) >> Fair. Okay. (02:10:10) >> So, it just needs to last a very long (02:10:11) time. Um (02:10:15) but that's it just you try I mean we (02:10:18) have brought the ship back and had it do (02:10:20) a soft landing in the ocean. I've done (02:10:23) that a few times, but but it lost a lot (02:10:25) of tiles, you know, and you know, it was (02:10:29) [snorts] (02:10:30) not reusable without a lot of work. (02:10:32) >> Yeah. (02:10:32) >> So, even though it did land it did come (02:10:34) to soft landing, it it was would not (02:10:36) have been reusable without a lot of (02:10:38) work. Um and and that so it's not really (02:10:41) reusable in that sense. So, that's (02:10:43) that's the biggest problem that remains (02:10:44) is fully reusable heat shield. Um, (02:10:48) >> so so like if you want to be able to (02:10:50) land it, uh, refill propellant and fly (02:10:52) again, uh, without, you know, you can't (02:10:56) go do this laborious inspection of, you (02:10:58) know, 40,000 tiles type sort of thing. I (02:11:00) I I'm curious how you drive like when (02:11:03) when I read biographies of yours, it (02:11:05) just uh it seems like you're just able (02:11:08) to drive the sense of like urgency and (02:11:10) drive the sense of like this is the this (02:11:12) is the thing that can scale. Um, and I (02:11:14) I'm curious why you think other (02:11:16) organizations of your like SpaceX and (02:11:18) Tesla are really big companies now and (02:11:20) you're still able to keep that culture. (02:11:22) What goes wrong with other companies (02:11:24) such that they're not able to do that? (02:11:27) >> I don't know. Um, (02:11:29) >> but like today you said you had like a (02:11:30) bunch of SpaceX meetings like what what (02:11:32) is it that you're doing there that's (02:11:33) like keeping that (02:11:33) >> that's adding urgency? (02:11:34) >> Yeah. Yeah. (02:11:38) >> Well, I don't know. I guess I guess uh (02:11:40) the (02:11:42) urgency is going to come from whoever's (02:11:44) leading the company. So if my sense of (02:11:46) urgency, I have like a maniacal sense of (02:11:47) urgency. So (02:11:48) >> that maniacal sense of urgency projects (02:11:51) through the rest of the company. (02:11:52) >> Is it because of consequences? They're (02:11:54) like if you know Elon said a crazy (02:11:56) deadline, but if I don't get it, I know (02:11:58) what happens to me. Is it just um you're (02:12:01) able to identify bottlenecks and get rid (02:12:02) of them so people can move fast? Like (02:12:04) how do you how do you think about why (02:12:05) your companies are able to move fast? (02:12:07) Yeah, I'm constantly addressing the (02:12:08) limiting factor. (02:12:11) So, um (02:12:16) I mean I mean on the deadlines front I I (02:12:19) mean I generally actually try to (02:12:22) aim for a deadline that that I at least (02:12:24) think is at the 50th percentile. So it's (02:12:26) it's not it's not like an impossible (02:12:27) deadline, but but it's the most (02:12:29) aggressive deadline I can think of that (02:12:30) could be achieved with 50% probability. (02:12:33) Um which means that it'll be late half (02:12:35) the time. (02:12:36) Um (02:12:38) and um (02:12:41) but whatever like there is like a law of (02:12:43) gases expansion that applies to (02:12:45) schedules like whatever given whatever (02:12:47) schedule you you like if if you you said (02:12:49) we're going to do this something in like (02:12:51) 5 years which to me is like infinity (02:12:53) time. Um it it will expand to fully (02:12:56) available schedule and it will take 5 (02:12:58) years. um you know like there's like (02:13:01) there there's a physical limit (02:13:04) like that like physics will limit how (02:13:07) fast you can do certain things like so (02:13:08) like scaling up manufacturing there's (02:13:11) like there's a rate which you can move (02:13:12) the atoms um and and scale manufacturing (02:13:16) that's why you can't like instantly make (02:13:18) you know a million of something million (02:13:20) a year or something uh you've you've got (02:13:22) to design manufacturing line you got to (02:13:24) bring it up you got to ride the scurve (02:13:25) of production um So yeah, I guess like (02:13:31) like what can I say that's that's that's (02:13:33) actually helpful to people. Um (02:13:38) I I think generally um a maniacal sense (02:13:40) of urgency is is a is very big deal. Um (02:13:45) so um and and you want to have you want (02:13:48) to have you want to have a an an (02:13:50) aggressive schedule. Um, and then you (02:13:53) and you and you want to figure out what (02:13:54) the limiting factor is at any point in (02:13:56) time and and help the team address that (02:13:57) limiting factor. (02:13:58) >> Can you maybe talk about the So, (02:14:00) Starlink was slowly in the works for (02:14:03) many years. Uh, and (02:14:05) >> yeah, we talked about it all the way in (02:14:06) the beginning of the company. (02:14:07) >> Yeah. And so then there was a team you (02:14:09) had built in Redmond and then at one (02:14:12) point you decided (02:14:14) this team is just not cutting it. But (02:14:16) again, how did you like (02:14:20) it went for a few years slowly and so (02:14:24) why did it why didn't you act earlier (02:14:26) and why did you act when you did? Like (02:14:28) why was that the right moment at which (02:14:29) to act? (02:14:31) >> I mean I had I have these very detailed (02:14:34) um engineering reviews weekly. Um that (02:14:37) that's that that's maybe a very unusual (02:14:40) level of granularity. Um, I don't know (02:14:43) anyone who runs a company or at least a (02:14:46) manufacturing company that that goes (02:14:48) into level of detail that that I go (02:14:50) into. Um, so so it's it's not it's not (02:14:54) as though like I have a pretty good (02:14:56) understanding of what's actually going (02:14:58) on. (02:14:58) >> Mhm. (02:14:59) >> Because we we we we go we go through (02:15:02) things in detail. Um, and I'm a big (02:15:05) believer in skip level meetings where (02:15:07) the individuals in instead of having the (02:15:10) person that reports to me say things, (02:15:12) it's everyone that reports to them um (02:15:15) says something um in in the technical (02:15:18) review. Um and um and and there can't be (02:15:23) um advanced preparation. So otherwise (02:15:26) you you're going to get u you know (02:15:29) glazed um as I say these days. (02:15:31) >> Yeah, exactly. Very gen Z of view. (02:15:33) >> Very generous. You just like call them (02:15:35) randomly like (02:15:36) >> No, just go around the room and everyone (02:15:38) provides an update. (02:15:39) >> Okay. (02:15:40) >> Um so, uh I mean it's it's a lot of (02:15:43) information to keep in your head because (02:15:45) um you you've got a you you've got then (02:15:48) say if you have meetings weekly or twice (02:15:49) weekly, you you've got a snapshot of (02:15:52) what that person said. Um and and and (02:15:56) you can and you can then you know plot (02:15:59) the progress points. um you can sort of (02:16:03) mentally plot the points on a curve and (02:16:04) say are we converging to a solution or (02:16:07) not um or or are we you know like I I'll (02:16:13) take drastic action (02:16:15) uh only when I conclude that um success (02:16:19) is not in a set of possible outcomes. Um (02:16:22) so when I say okay when when I finally (02:16:25) reach the conclusion that okay unless (02:16:27) drastic action is done we have no no (02:16:29) chance of success then I must take (02:16:31) drastic action (02:16:33) and so that's that's I came to that (02:16:35) conclusion in 2018 took drastic action (02:16:37) and and fix the problem. How how many um (02:16:42) you know you you've got many many (02:16:44) companies and in each of them it sounds (02:16:46) like you do this kind of deep (02:16:48) engineering understanding of what the (02:16:50) relevant bottlenecks are so you can do (02:16:51) these um reviews with people. (02:16:54) >> Yeah. (02:16:55) >> Um (02:16:56) you've been able to scale it up to five, (02:16:58) six, seven companies. Within one of (02:17:00) these companies you have many different (02:17:03) mini companies within them. What (02:17:05) determines the maxim? Could you have (02:17:06) like 80 companies? (02:17:07) >> 80? No. But like you can you have so (02:17:10) many already like that's that's already (02:17:12) remarkable (02:17:12) >> by this current number. Yeah. (02:17:13) >> Yeah. Exactly. (02:17:14) >> Uh no. So um (02:17:16) >> we can barely keep one coming together. (02:17:19) >> Um (02:17:21) near like it it depends on situation. Um (02:17:26) so (02:17:27) um I I actually don't don't have regular (02:17:30) meetings uh with flooring company. So (02:17:33) that boring company's sort of cruising (02:17:35) along like look if basically if (02:17:36) something is working well and making (02:17:38) good progress then there's no point in (02:17:40) me spending time on it. So I actually uh (02:17:44) allocate time according to where where (02:17:45) the where the limiting factor or the (02:17:47) problem where are things problematic and (02:17:50) um or where are we pushing against uh (02:17:55) like what what is holding us back you (02:17:57) know I I focus risk of saying the words (02:18:00) too many times the limiting factor (02:18:03) Um, (02:18:04) so, so basically if something's go like (02:18:06) the irony is if something's going really (02:18:08) well, um, they don't see much of me, but (02:18:12) if something is going badly, they'll see (02:18:13) a lot of me. (02:18:15) >> Something or not even badly, it's it's (02:18:17) like if something's a limiting factor, (02:18:19) >> it's a limiting factor. Exactly. It's (02:18:20) not exactly going badly, but it's the (02:18:21) thing that's it's it's the thing that we (02:18:22) need to make go faster to. And so when (02:18:24) something's a limiting factor at SpaceX (02:18:26) or Tesla, are you like talking weekly (02:18:30) and daily with the engineer that's (02:18:32) working on it? How how does that (02:18:34) actually work? (02:18:37) >> Most things that are learning factor are (02:18:39) um weekly and some things are twice (02:18:41) weekly. So the the AI5 chip review is (02:18:45) twice weekly and and so it's every (02:18:47) Tuesday and Saturdays (02:18:49) >> um is is the chip review. (02:18:51) >> Is it open-ended in how long it goes? (02:18:54) Uh technically yes, but uh (02:18:57) usually it's it's like two or three (02:18:59) hours. (02:19:01) >> So I mean sometimes less. It's it (02:19:03) depends on on how much information (02:19:05) you've got to go through. (02:19:06) >> Yeah. Well, that's another thing. Again, (02:19:07) I'm just trying to tease out the the (02:19:08) differences uh here cuz the outcomes (02:19:11) seem quite different and so I think it's (02:19:12) interesting to note what inputs are (02:19:14) different and it feels like the (02:19:15) corporate world one like you're saying (02:19:18) just the CEO doing engineering reviews (02:19:20) does not always happen despite the fact (02:19:22) that that is the you know what the (02:19:24) company is doing. Um, but then time is (02:19:27) often pretty finely sliced into, you (02:19:30) know, halfhour meetings or even 15-inute (02:19:32) meetings. And it seems like you hold (02:19:34) more open-ended (02:19:36) we're talking about it until we figure (02:19:38) it out type meetings. Yeah. (02:19:42) >> Sometimes, but most of them seem to more (02:19:45) or less stay on time. Um (02:19:48) so um (02:19:51) I mean to today's uh Starship (02:19:55) engineering review went a bit longer um (02:19:58) because there there were more topics to (02:20:00) discuss. Um (02:20:03) yeah trying to figure out how to scale (02:20:04) to a million plus tons to orbit per year (02:20:06) is quite challenging. C (02:20:08) >> can I answer a question? So you you said (02:20:10) about um Optimus and AI that they're (02:20:13) going to result in doubledigit growth (02:20:16) rates within a matter of years. (02:20:17) >> Oh, like the economy. (02:20:19) >> Yeah. (02:20:19) >> Um (02:20:20) >> yes, (02:20:21) >> I think I think that's right. (02:20:22) >> What was the point of the Doge cuts if (02:20:26) the economy is going to grow so much? (02:20:28) >> Well, I think like waste and fraud are (02:20:30) not good things to have, you know. Um, (02:20:33) I I I was actually pretty worried about (02:20:36) I I I guess (02:20:38) uh I mean I I think in the absence of AI (02:20:41) and robotics, we're actually totally (02:20:42) screwed. Uh because the national debt is (02:20:46) piling up like crazy. Um now our (02:20:49) interest payments, the interest payments (02:20:51) to national debt exceed the military (02:20:53) budget which is a trillion dollars. So (02:20:55) over a trillion dollars just in interest (02:20:57) payments. Um, you know, that was like I (02:21:00) was like, okay, pretty concerned about (02:21:01) that. Maybe if I spend some time, we can (02:21:04) slow down the bankruptcy of the United (02:21:06) States. Um, and give us enough time for (02:21:09) the AI and robots to, you know, help (02:21:13) solve the national debt or not help (02:21:15) solve, it's the only thing that could (02:21:17) solve the national debt. Like, we are (02:21:18) 1,000% going to go bankrupt as a country (02:21:21) and fail as a country without AI and (02:21:24) robots. Nothing else will solve the (02:21:26) national debt. Um, and so, so we we'd (02:21:30) like to well, we just need we need (02:21:32) enough time to get build the AI and (02:21:35) robots uh to and not go bankrupt before (02:21:39) then. I (02:21:39) >> I I guess the thing I'm curious about is (02:21:40) when Doge starts, you have this enormous (02:21:43) um ability to enact reform and (02:21:48) >> not that enormous. (02:21:48) >> Sure. Sure. But to totally by your point (02:21:51) that like it's important that AI and (02:21:53) robotics drive product improvements, (02:21:55) drive GDP growth, but why not just (02:21:57) directly go after the things you were (02:21:59) pointing out, you know, like the the (02:22:01) tariffs on certain components or whether (02:22:03) it's like permitting. (02:22:04) >> I'm not the president (02:22:07) and and very hard to cut to cut to to (02:22:09) even even to cut things that are obvious (02:22:12) waste and fraud like like ridiculous (02:22:15) waste and fraud. Um what I discovered (02:22:18) that is it it's (02:22:21) extremely difficult even to cut very (02:22:24) obvious ways on fraud um from the (02:22:26) government um because the the the (02:22:28) government has to operate on a on like (02:22:30) who's complaining like if if and if you (02:22:32) cut off payments to fraudsters they (02:22:34) immediately come up with the most (02:22:36) sympathetic sounding uh reasons to (02:22:38) continue the payment. They they don't (02:22:40) say please keep the fraud going. They (02:22:42) say, you know, it's they're like, you're (02:22:44) killing baby pandas. Like, meanwhile, (02:22:47) there's no baby pandas are dying. (02:22:48) They're just making it up. Um, the (02:22:50) forces are capable of of coming up with (02:22:53) extremely compelling sort of (02:22:55) heart-wrenching stories that are false, (02:22:56) but nonetheless sound uh sympathetic and (02:23:00) that that's what happened. Um, and uh so (02:23:03) it's like (02:23:05) perhaps I should have known better. Um (02:23:08) and uh (02:23:10) but I thought wait let's take a let's (02:23:12) let's let's try to cut some amount of of (02:23:15) waste and core from the government. (02:23:16) Maybe there shouldn't be you know 20 (02:23:18) million people uh marked as alive in (02:23:21) social security who are indefinitely (02:23:23) dead and over the age of 115. (02:23:27) The oldest American is 114. (02:23:30) So, it's safe to say if somebody's 115 (02:23:32) and marked as live in the social (02:23:34) security database, um something is (02:23:37) there's either a typo. So, like somebody (02:23:40) should call them and say, "We we seem to (02:23:43) have your birthday wrong or or uh or or (02:23:47) we need to mark you as dead." [laughter] (02:23:50) >> One of the two things. (02:23:51) >> Very intimidating call to get. (02:23:53) >> Well, so it seems like a reasonable (02:23:55) thing. Um and if if like say their (02:23:57) birthday is in the future um [laughter] (02:24:00) and they have you know a small business (02:24:03) administration loan and their birthday (02:24:05) is 2165 (02:24:07) um we either again have a typo or we (02:24:10) have fraud. Um [laughter] (02:24:12) so we say we appear to have gotten the (02:24:14) century of your birth incorrect (02:24:15) >> or a great plot for a movie. (02:24:17) >> Yes. This is this this is when I when I (02:24:19) mean about ludicrous fraud that's what I (02:24:21) meant ludicrous fraud. (02:24:22) >> Were those people getting payments? Some (02:24:24) some were getting payments from social (02:24:25) security but but but the main fraud (02:24:27) vector uh was to mark somebody as alive (02:24:30) in social security and then use every (02:24:32) other government payment system uh to uh (02:24:36) basically to to do fraud because what (02:24:38) those other government payment systems (02:24:40) do would do they will simply do an are (02:24:42) you alive check to the social security (02:24:44) database. (02:24:45) >> It's a it's a bank shot. (02:24:46) >> What would you estimate is like the (02:24:47) total uh amount of fraud from this (02:24:49) mechanism. Um my guess is and and other (02:24:52) by the way the the government (02:24:53) accountability office has done these (02:24:55) estimates before. I'm not the only one (02:24:56) who's coming out of this you know in (02:24:58) fact I think they they did the GAO did (02:25:01) analysis a rough estimate of fraud (02:25:02) during the Biden administration and (02:25:04) calculated at roughly half a trillion (02:25:06) dollars. (02:25:08) >> So don't take my word for it. Take it a (02:25:10) report issued during the Biden (02:25:12) administration. How about that (02:25:14) >> from this social security mechanism? (02:25:16) >> Uh it's it's one of many. It's important (02:25:19) to appreciate that the the government (02:25:21) does not is very ineffective at at (02:25:24) stopping fraud because uh it's it's not (02:25:28) like like if it was a company like like (02:25:30) stopping fraud, you've got a motivation (02:25:32) because it's affecting the earnings of (02:25:33) your company. Uh but the government just (02:25:35) just they just print more money. Um, so (02:25:38) it's not uh (02:25:41) like you you need you need caring and (02:25:43) competence and these are in short supply (02:25:46) at at the federal level. (02:25:48) >> Um, (02:25:50) yeah, sorry. I mean, when you go to the (02:25:51) DMV, do you think, wow, this is a (02:25:53) bastion of competence. [snorts] Um, (02:25:55) well, now imagine it's worse than the (02:25:57) DMV because it's a DMV that can print (02:25:59) money. (02:26:01) >> So, was it not possible? At least the (02:26:03) state level DMVs uh need to the states (02:26:06) more or less need to stay within their (02:26:08) budget to go bankrupt. But the federal (02:26:09) government just prints more money. (02:26:11) >> Was it not possible to cut that if (02:26:13) there's a catchy half a trillion of (02:26:14) fraud? Why was it not possible to cut (02:26:16) all that? (02:26:17) >> Uh because when when as soon as you we (02:26:20) did we we actually no (02:26:23) you you you you really have to stand (02:26:25) back and (02:26:28) reccalibrate your expectations for (02:26:30) competence. uh because uh you you're (02:26:35) operating in a world where you know you (02:26:37) you you've got to sort of make ends meet (02:26:39) like you know you got to pay your bills, (02:26:40) you got to you know (02:26:41) >> buy the microphones. (02:26:42) >> Yeah. Yeah. Exactly. Um so so you you (02:26:46) you if you don't have it's it's not like (02:26:48) there's a giant largely uncaring monster (02:26:52) bureaucracy. It's not even and and and a (02:26:54) bunch of (02:26:56) uh monacistic computers that are just (02:26:58) that are just sending payments. Um like (02:27:02) one of the things that that that the (02:27:03) Doge team did there was and it sounds so (02:27:06) simple uh that that probably will save (02:27:09) um (02:27:11) let's say 100 billion maybe 200 billion (02:27:13) a year um is simply requiring that (02:27:16) payments from the main treasury computer (02:27:18) which is called PM it's like payment (02:27:19) accounts master or something like that (02:27:21) there's 5 trillion payments a year um (02:27:24) requiring that any payment go that goes (02:27:26) out have a payment um appropriation (02:27:29) code, make it mandatory, not optional, (02:27:32) and that you have anything at all in the (02:27:33) comment field. (02:27:37) Um, because (02:27:39) you see you have to recalibrate how dumb (02:27:41) things are. (02:27:42) >> You think were being sent out with no (02:27:44) appropriation code, (02:27:47) not not checking back to any (02:27:48) congressional appropriation, and no (02:27:50) explanation. (02:27:52) And this is why the the Department of (02:27:55) War, formerly the Department of Defense, (02:27:56) cannot pass an audit because the (02:27:58) information is literally not there. (02:28:02) Recalibrate your expectations. (02:28:03) >> I want to better understand the South (02:28:04) Australian number because there was an (02:28:05) IG report in 2024. (02:28:07) >> How you must like why is it so low? (02:28:10) >> Um maybe. But uh which found that like (02:28:12) over seven years this the social (02:28:13) security fraud they estimated was like (02:28:15) 70 billions over 7 years. So like 10 (02:28:17) billion a year. So I'd be curious to see (02:28:18) what like the other 490 billion is. (02:28:20) Federal government expenditures are 7 (02:28:22) and a half trillion a year. (02:28:23) >> Yeah. (02:28:24) >> Um how what what percentage how (02:28:27) competent do you think government is? (02:28:28) >> The the discretionary spending there is (02:28:32) like 15%. (02:28:33) >> Yeah. But but it doesn't matter. Most of (02:28:35) the ford is non-discretionary. It's it's (02:28:37) basically a fraudulent Medicare, (02:28:38) Medicaid, uh social security, uh uh uh (02:28:43) you know um (02:28:46) disability. Uh it's there's there's a (02:28:48) zillion government payments. Yeah. (02:28:50) >> Um and and a bunch of these payments are (02:28:52) in fact u (02:28:55) they're uh block transfers to the (02:28:57) states. So the federal government (02:28:59) doesn't even have the information in a (02:29:01) lot of cases to even see know if there's (02:29:04) fraud. Let's consider let's like (02:29:06) reductio Adam sodom the government the (02:29:08) government is perfect and has no fraud. (02:29:10) What is your probability estimate of (02:29:12) that? (02:29:14) >> I mean zero. Okay. So then would you say (02:29:17) that foreign waste that the government (02:29:20) uh is (02:29:22) has is 90%. (02:29:25) That also would be quite generous. But (02:29:27) if if it's only 90% that means that (02:29:30) there's $750 billion a year of waste and (02:29:32) fraud and it's not 90%. It's not 90% (02:29:36) effective. This seems like a strange way (02:29:37) to first principle is the amount of (02:29:38) fraud in the government. just like how (02:29:40) much do you think there is and then uh (02:29:43) uh I anyways we don't know how to do it (02:29:44) live but I'd be curious like see (02:29:46) >> you know a lot about fraud at Stripe (02:29:48) people are constantly trying to do (02:29:49) fraud. (02:29:50) >> Yeah but as you say it's like a little (02:29:51) bit of a um we've really ground it down (02:29:53) but it's a little bit of a different (02:29:55) problem space because we're dealing with (02:29:57) a much more heterogeneous set of fraud (02:29:59) vectors here than we are (02:30:00) >> Yeah. But I mean I mean at tribe you you (02:30:03) you you have high confidence and you try (02:30:05) hard. Um you have high competence and (02:30:08) high caring but still fraud is non non (02:30:10) zero. Um now now now imagine it's at a (02:30:14) much bigger scale. Um there's much less (02:30:16) competence and much less carrying. (02:30:20) You know bank PayPal back in the day we (02:30:22) we try to manage forward down to about (02:30:23) 1% of of the payment volume. Um and that (02:30:27) was very difficult. took a tremendous (02:30:28) amount of competence and caring to uh (02:30:30) get fraud merely to 1%. (02:30:33) Um now imagine that that you have an (02:30:35) organization where there's much less (02:30:37) caring and much less confidence. It's (02:30:39) [snorts] going to be much more than 1%. (02:30:42) >> How do you feel now looking back on um (02:30:45) kind of politics and and doing stuff (02:30:48) there where it feels like from the (02:30:50) outside in the two you know two things (02:30:54) have been quite impactful. one the (02:30:57) America Pack and two um the acquisition (02:31:00) of of well Twitter at the time but also (02:31:04) it seems like there was a bunch of (02:31:06) heartache and so what's your what's your (02:31:08) grading of the whole experience (02:31:12) >> well um (02:31:16) I think I think those things needed to (02:31:18) be done to (02:31:20) maximize the probability that the future (02:31:21) is good (02:31:23) >> um So (02:31:26) um but but politics generally is very (02:31:28) tribal. Um and and it's it's very (02:31:30) tribal. It's and people lose their (02:31:33) objectivity usually with politics like (02:31:35) they they generally have trouble seeing (02:31:39) the good on the other side or the bad in (02:31:41) their own side. That's generally how it (02:31:42) goes. Um that that I guess was one of (02:31:46) the things that surprised me the most is (02:31:48) you you often simply cannot reason with (02:31:50) people (02:31:50) >> um if they're in one tribe or the other. (02:31:52) they they simply believe that everything (02:31:54) their tribe does is good and anything (02:31:56) the other political tribe does is bad. (02:31:58) Um and persuading them is (02:32:02) otherwise is almost impossible. Um, so (02:32:07) anyway, but um (02:32:12) I think I think overall (02:32:15) those actions (02:32:18) um acquiring Twitter, (02:32:21) getting Trump elected even though, you (02:32:22) know, it makes a lot of people angry. Um (02:32:25) I think those I think those actions are (02:32:27) good for were good for civilization. (02:32:29) >> Um yeah, how does it feed into the (02:32:32) future you're excited about? (02:32:33) Well, um, America needs to contain, (02:32:37) America needs to be strong enough to (02:32:38) last long enough to, um, extend life to (02:32:42) other planets and to (02:32:45) get, I guess, AI and robotics to the (02:32:47) point where we can ensure that the (02:32:49) future is good. Um like on the other (02:32:52) hand if if if we were to descend into (02:32:56) um say communism or or some situation (02:32:58) where where the state was extremely (02:33:00) oppressive um that that would mean that (02:33:03) we we might not be able to become (02:33:06) multilanetary (02:33:08) um and we might this the state might um (02:33:13) you know stamp out um our progress in AI (02:33:16) and robotics. How do you feel about um (02:33:19) uh you know Optimus, Grock, etc. are (02:33:24) going to be leveraged by and not just (02:33:27) yours any revenue maximizing company's (02:33:29) products will be leveraged by the (02:33:30) government over time? Um how does this (02:33:34) concern manifest in what private (02:33:37) companies should be willing to give (02:33:40) governments? What kinds of guard rails (02:33:41) should like should you know the should (02:33:44) um AI models be uh um made to do (02:33:49) whatever the government that has (02:33:50) contracted them out to do asked them to (02:33:53) do? Um should like should should Grog (02:33:57) get to say like actually even the (02:33:58) military wants to do X. No, the Grock (02:33:59) will not do that. I I pro probably the (02:34:02) biggest danger of AI or maybe the (02:34:05) biggest danger of fa for for AI and (02:34:08) robotics going wrong wrong is is (02:34:09) government (02:34:11) interesting you know um (02:34:15) I mean the the way like people who are (02:34:17) opposed to corporations or or or worried (02:34:20) about corporations should um really (02:34:22) worry about the most about government (02:34:24) cuz government is just a corporation in (02:34:26) the limit. It's a government. It is it (02:34:28) is it is government is just the biggest (02:34:30) corporation with a monopoly on violence. (02:34:34) Um so I always find it like a strange (02:34:38) dichotomy where where people would think (02:34:39) corporations are bad but the government (02:34:41) is good when the government is simply (02:34:42) the biggest and and and worst (02:34:44) corporation. (02:34:48) But people have that dichotomy. They (02:34:50) somehow think at the same time that (02:34:52) government can be good but corporations (02:34:54) bad. And this is not true. corporations (02:34:56) are have better morality than the (02:34:58) government. (02:34:59) >> It is. (02:34:59) >> So I I I actually think it's uh (02:35:02) you know that's uh that that is a thing (02:35:05) to be worried about. It's like if the (02:35:08) you know should should if the government (02:35:09) should not like the government could (02:35:12) potentially use AI and robotics to (02:35:13) suppress the population (02:35:16) like that is a serious concern. (02:35:18) >> I as a guy building AI and robotics how (02:35:20) do you how do you like how do you (02:35:22) prevent that? Uh well I think like if (02:35:24) you have a limited government um (02:35:29) if you limit the powers of government (02:35:30) which is like really what the US (02:35:32) constitution is intended to do is (02:35:33) intended to limit the powers of (02:35:34) government then then uh you're probably (02:35:37) going to have a better outcome than if (02:35:38) you have more government. (02:35:40) So (02:35:42) >> will be available to all governments (02:35:44) right? (02:35:45) >> Yeah about all governments. Um I mean (02:35:50) it's difficult to predict the like I (02:35:52) said like what what's what's the end end (02:35:55) point or like what is what is many years (02:35:57) in the future but it's difficult to (02:35:58) predict the the sort of path along along (02:36:01) that way. Um like if civilization (02:36:05) progresses (02:36:07) AI will vastly exceed the sum of all (02:36:10) human intelligence and and there will be (02:36:13) far more robots than humans. (02:36:15) um along the way what happens (02:36:19) it's very difficult to predict. (02:36:20) >> I mean I mean it seems like one thing (02:36:21) you could do is just say um uh you are (02:36:24) not allowed to whatever government (02:36:27) you're not allowed to use Optimus to do (02:36:28) XYZ just write out like a policy. I mean (02:36:30) you you I think you treated recently (02:36:31) that Grock should have a moral (02:36:32) constitution. Um and one of those things (02:36:35) could be that we we limit what (02:36:37) governments are allowed to do with this (02:36:39) advanced technology. (02:36:41) >> I mean yeah we we can do what is what (02:36:45) I mean (02:36:47) technically I mean if if the politicians (02:36:50) pass a law uh then and they can enforce (02:36:52) that law then it's hard to not do that (02:36:55) law. You know the the best thing we can (02:36:58) have is is is limited government uh (02:37:01) where um you know you have you have the (02:37:04) appropriate cross checks between the (02:37:07) executive judicial and um legislative (02:37:10) branches. I I guess the the reason I'm (02:37:13) curious about it is this like at some (02:37:14) point it seems like the limits will come (02:37:16) from you, right? Like you've got the (02:37:18) Optimus, you've got the space GPUs, (02:37:20) you've got the (02:37:20) >> you think I will be the boss of the (02:37:22) government (02:37:22) >> or you will get the you will like the I (02:37:24) mean already it's the case with SpaceX (02:37:27) that for things that are crucial to the (02:37:30) um uh like the government really cares (02:37:33) about getting certain satellites up in (02:37:34) space whatever like it needs SpaceX. Uh (02:37:36) it is the it is the um a necessary (02:37:38) contractor and you are in the process of (02:37:40) building more and more of the um uh the (02:37:45) technological components of the future (02:37:47) that that that will have an analogous (02:37:49) role in different industries and you (02:37:51) could have this ability to like set some (02:37:53) policy that um you know suppressing (02:37:57) classical liberalism in any way. I my (02:38:00) companies will not help in in any way (02:38:01) with that or you know some policy like (02:38:03) that. (02:38:05) Um, I I will do my best to ensure that (02:38:07) anything that's within my control (02:38:09) maximizes the good outcome for humanity. (02:38:16) >> I think anything else would be (02:38:17) shortsighted. Um, because obviously I'm (02:38:19) part of humanity. So, um, I like humans. (02:38:23) Um, (02:38:26) >> pro human pro. (02:38:28) >> Um, you you you've mentioned that Dojo 3 (02:38:31) will be used for space-based compute. (02:38:33) Um, (02:38:34) >> [laughter] (02:38:34) >> Do you really read my uh what I say? (02:38:38) >> I don't know if you know Twitter, but I (02:38:39) know you lot. [laughter] (02:38:40) You have a lot of followers. (02:38:42) >> Big giveaway. (02:38:44) >> Um how do you (02:38:45) >> how does you have discern my secrets and (02:38:46) I post them. (02:38:49) >> How how do you design this chip for (02:38:50) space? What like Yeah. What changes? (02:38:54) >> Well, I guess you want to design it to (02:38:56) be um more radiation tolerant and run at (02:38:59) a higher temperature. Uh so you can um (02:39:03) you know roughly if you increase the um (02:39:05) operating temperature by 20% in degrees (02:39:08) Kelvin you can cut your radiator mass in (02:39:10) half. Um so (02:39:14) running at a higher temperature is is (02:39:16) helpful in in space. Um (02:39:20) there I mean there's various things you (02:39:21) can do for shielding the memory and (02:39:25) but like neural nets are going to be (02:39:26) very resilient to bit flips. Yeah. So (02:39:29) like most of what what happens from a (02:39:30) radiation is like random bit flips. Um (02:39:33) but like if you've got like you know a (02:39:36) multi- trillion parameter model and you (02:39:38) get a few flips doesn't matter. Um it's (02:39:41) it's much like curistic programs are (02:39:42) going to be much more sensitive to flips (02:39:44) than um some giant parameter file. Um, (02:39:49) so I just designed it to run hot and um (02:39:55) I think you pretty much do it the same (02:39:57) way that you do things on Earth apart (02:39:59) from make it run hotter. (02:40:01) >> Um, I mean the solar array is most of (02:40:03) the weight on the satellite. Is there a (02:40:05) way to make the um the GPUs even more (02:40:08) powered ends than what Nvidia and TPUs (02:40:11) and etc are planning on doing that would (02:40:14) you know be especially privileged in the (02:40:16) space-based world? (02:40:18) Well, I mean the basic math is like (02:40:22) um if you can do about a kilowatt per (02:40:24) reticle um and then you'd need um (02:40:30) you know 100 million (02:40:33) full retical chips uh to do 100 gawatt. (02:40:35) >> Yeah. (02:40:38) So (02:40:40) yeah, depending what your yield (02:40:42) assumptions are, you know, um that that (02:40:45) tells you how many trips you need to (02:40:46) make. (02:40:48) Um, but cool. You need if you want if if (02:40:50) if you're going to have 100 gigawatts of (02:40:53) power, you need, you know, 100 million (02:40:57) chips running that that are running a (02:40:59) kilowatt sustained uh quad per reticle. (02:41:02) >> Um (02:41:04) 100 (02:41:04) >> basic math (02:41:05) >> 100 million chips. Uh it depends on (02:41:10) Yeah. If if you if you look at the die (02:41:12) size of something like black wisps or (02:41:14) something and how many you can get out (02:41:15) of a wafer, you can get like um on the (02:41:19) order of dozens or less uh per wafer. So (02:41:23) you're basically you're this is a world (02:41:24) where (02:41:25) >> if we're putting that out a every single (02:41:28) year you're producing millions millions (02:41:30) of wafers a month. (02:41:33) >> Um that's the plan with Terapab millions (02:41:36) of wafers a month of advanced process (02:41:38) notes. Yeah, it's it's got to be some (02:41:39) number north of a million. I think (02:41:40) >> you got to do the memory, too. (02:41:42) >> Yeah. Are you going to make a memory f? (02:41:45) >> I think the teraf's got to do memory. (02:41:46) It's got to do logic, memory, and (02:41:48) packaging. (02:41:49) >> I'm very curious how somebody like gets (02:41:51) started. This is like the most (02:41:52) complicated thing man has ever made. And (02:41:54) obviously, like if anybody's up to the (02:41:56) task, you're up to the task. Like what (02:41:58) do you So, you realize it's a bottleneck (02:42:00) and you go to your engineers and like (02:42:02) what is the next like what what do you (02:42:03) tell them to do? [laughter] (02:42:05) I want a million papers a month in 2030. (02:42:08) What is the next like what do you (02:42:09) >> That's right. (02:42:09) >> Do you like call ASML? Like what is (02:42:11) >> Ask what I want. [laughter] (02:42:13) >> What is the next step? (02:42:14) >> That's so much to ask. (02:42:16) >> Well, um we make a little fab uh and see (02:42:21) what happens. Uh make our mistakes at a (02:42:24) small scale and then make a big one. (02:42:26) >> Is a little fab done or is it (02:42:27) >> No, it's not done. Which I mean people (02:42:30) would not keep that cat in the bag. (02:42:32) [laughter] (02:42:33) that cat's going to come out of the (02:42:34) back. It'll be like drones hovering over (02:42:37) the bloody thing, you know. You'll be (02:42:38) able to like see it construction (02:42:40) progress on X, right? You know, in real (02:42:42) time. Um, so no, we we I mean, listen, I (02:42:47) don't know. We could just flounder in (02:42:48) failure to be clear. It's like not uh (02:42:51) success is not guaranteed, but um (02:42:56) since we want to try to make uh you know (02:43:00) something like a 100 million (02:43:02) Yeah, we we need we want 100 gigs of (02:43:06) power and 100 100 that trips that can (02:43:08) take 100 gawatt, right? So call it, you (02:43:12) know, but yeah, by by 2030. So then (02:43:15) um (02:43:18) we'll take as many chips as our (02:43:20) suppliers will give us. I've said this (02:43:21) to I've actually said this to TSMC and (02:43:23) Samsung and Micron. It's like please (02:43:26) build your more fabs faster. Um, and we (02:43:29) will guarantee to buy the output of (02:43:30) those fabs. Um, so so that they're (02:43:33) already like moving as fast as they as (02:43:35) they can. Like it's it's not like to be (02:43:37) clear, it's not like us (02:43:39) take, you know, it's not like u either (02:43:42) it's it's not like it's us plus them. (02:43:46) You know, (02:43:47) >> there's an narrative that the people (02:43:48) doing AI want a very large number of, (02:43:51) you know, chips as quickly as possible. (02:43:53) And then many of the input suppliers, (02:43:56) the fabs, but also, you know, the (02:43:58) turbine manufacturers (02:44:00) are not ramping up production very (02:44:02) quickly. And the explan Yeah. And the (02:44:04) the explanation you hear is that they're (02:44:06) dispositionally (02:44:08) conservative. You know, they're (02:44:09) Taiwanese or German as the, you know, (02:44:11) story may be. And they just like don't (02:44:13) believe the like is that really the (02:44:15) explanation or is there something else? (02:44:17) Well, I mean, it's reasonable to like if (02:44:19) somebody's been in say the computer (02:44:22) memory business for uh (02:44:25) 30 or 40 years (02:44:26) >> and they've seen cycles, (02:44:27) >> they've seen like boom and bust like 10 (02:44:30) times. (02:44:30) >> Yeah. (02:44:31) >> You know, so so like that's a lot of (02:44:33) layers of scar tissue, you know? So, (02:44:34) it's like it's like during the boom (02:44:37) times looks like everything's going to (02:44:38) be great forever and then then then the (02:44:40) crash happens and then they're (02:44:41) desperately trying to avoid bankruptcy. (02:44:44) Um and and then there's another boom and (02:44:46) another crash. (02:44:47) >> Are there other [laughter] are there (02:44:49) other ideas you think others should go (02:44:50) pursue that you're not for whatever (02:44:53) reasons right now? (02:44:55) >> Um [sighs and gasps] I mean there are a (02:44:56) few companies that are that are pursuing (02:44:57) like uh new ways of doing jobs. (02:45:00) >> Um uh but they're just not scaling fast. (02:45:03) >> I I don't even mean within AI. I mean (02:45:05) just generally (02:45:07) >> I'd say like people should just should (02:45:09) do the thing that where they find that (02:45:11) they're highly motivated to do that (02:45:12) thing. Mhm. (02:45:13) >> As opposed to, you know, something (02:45:16) something some idea that that I suggest, (02:45:19) but they should do the thing that they (02:45:20) find personally interesting and (02:45:23) motivating to do. (02:45:24) >> Mhm. (02:45:25) >> Um (02:45:28) but but you know, going back to the (02:45:30) limiting factor, (02:45:32) use that phrase about 100 times. Um (02:45:37) the the current limiting factor that I (02:45:39) see in the time frame you know in the (02:45:42) sort of 20 (02:45:46) 29 20 like in in the in the three 3 to 4 (02:45:49) year time frame um it's chips. Um in in (02:45:54) the one-year time frame it's it's energy (02:45:56) power production electricity. (02:45:58) like it's it's not clear to me that (02:46:00) there's enough (02:46:02) um usable electricity to turn on all the (02:46:05) the AI chips that are being made. Um (02:46:10) towards the end of this year, I think (02:46:11) people are going to have real trouble (02:46:12) turning on like the chip output will (02:46:14) exceed the the ability to turn chips on. (02:46:17) >> What's your plan to deal with that (02:46:18) world? (02:46:20) >> Well, we're trying to accelerate (02:46:22) electricity production. (02:46:24) Um, I I guess that's that's maybe one of (02:46:27) the reasons that um (02:46:29) XAI will be maybe the leader, hopefully (02:46:32) the leader, um, is that we'll be able to (02:46:35) turn on more chips than other people can (02:46:36) turn on faster. (02:46:39) Um, because we're we're we're good at (02:46:41) hardware [snorts] and and and and (02:46:43) generally the the innovations from the (02:46:47) corporations that me that call (02:46:49) themselves labs. um the the ideas tend (02:46:52) to flow like it's it's rare to see that (02:46:54) there's like more than about a six-month (02:46:56) difference um between um like the ideas (02:47:00) uh travel back and forth um with the (02:47:03) people. So, so I think you you sort of (02:47:06) hit the hardware wall and um and then (02:47:10) whatever whichever company can scale (02:47:12) hardware the fastest will (02:47:15) be the leader and so I think XCI will be (02:47:17) able to scale hardware the fastest and (02:47:19) therefore most likely will be the (02:47:20) leader. you you you joked or you know um (02:47:24) were self-conscious about uh you know (02:47:26) using the uh the limiting factor phrase (02:47:28) again but I actually think there's (02:47:30) something deep here and if you look at a (02:47:31) lot of the things we've touched on over (02:47:33) the course of it maybe kind of a good (02:47:34) note to end on like (02:47:37) if you think of a scinesscent lower (02:47:40) agency (02:47:42) company it would have some bottleneck (02:47:45) and not really be doing anything about (02:47:47) it. Um, you know, Mark and Dre had the (02:47:49) line of, uh, most people are willing to (02:47:51) endure any amount of chronic pain to (02:47:53) avoid acute pain. Uh, and it feels like (02:47:55) a lot of the cases we're talking about (02:47:58) are just leaning into the acute pain, (02:48:00) whatever it is. It's like, okay, we got (02:48:01) to figure out how to, you know, work (02:48:04) with steel or we got to figure out how (02:48:06) to run the chips in space or like we'll (02:48:07) take some near-term acute pain to (02:48:10) actually solve the bottleneck. And so (02:48:12) that's kind of a unifying theme. (02:48:14) >> I have a high pain threshold. (02:48:17) That's helpful. (02:48:18) >> Solve the bottlenecks. (02:48:19) >> Yes. (02:48:21) >> Um (02:48:23) so (02:48:25) you know one thing I I can say is like (02:48:27) uh (02:48:30) I think the future's going to be very (02:48:31) interesting. Um (02:48:34) and um and as I said the Davos only been (02:48:40) I think it was on the ground for like 3 (02:48:42) hours or something. Um (02:48:45) it's better to be it's better to on the (02:48:47) side of optimism and be wrong than on (02:48:50) the side of pessimism and be right uh (02:48:52) for quality of life. (02:48:54) So you know you your your your happiness (02:48:58) will be you you'll be happier if you if (02:49:01) you are on the side of optimism rather (02:49:03) than ingring on the side of pessimism (02:49:06) >> and so I recommend ingring on the side (02:49:07) of optimism. (02:49:08) >> Let's do that. (02:49:09) >> Cool. Yan, thanks for doing this. (02:49:11) >> Thank you. (02:49:12) >> All right. place. (02:49:14) >> All right. (02:49:14) >> Oh, great stamina. (02:49:18) >> Hopefully this encounter is a pain in (02:49:19) the tolerance. [laughter] (02:49:21) >> Hey everybody, I hope you enjoyed that (02:49:23) episode. If you did, the most helpful (02:49:25) thing you can do is just share it with (02:49:27) other people who you think might enjoy (02:49:28) it. It's also helpful if you leave a (02:49:31) rating or a comment on whatever platform (02:49:33) you're listening on. If you're (02:49:35) interested in sponsoring the podcast, (02:49:37) you [music] can reach out at (02:49:38) dwarcash.com/advertise. (02:49:42) Otherwise, I'll see you on the next one.