Elon Musk on DOGE, Optimus, Starlink Smartphones, Evolving with AI, Why the West is Imploding (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) [Music] (00:00:01) I believe Optimus is going to be the (00:00:03) greatest product ever created by (00:00:05) humanity. (00:00:06) >> Elon Musk and his XAI startup have built (00:00:09) the largest and most powerful artificial (00:00:11) intelligence training supercomputer in (00:00:14) the world. As far as I know, there's (00:00:16) only one person in the world who could (00:00:18) do that. You know, (00:00:19) >> this is an arms race of epic (00:00:21) proportions. (00:00:22) >> He's a big thinker. You guys went on Fox (00:00:25) the other day with the Doge team. You (00:00:27) saw Elon's face nodding while they were (00:00:29) speaking with a grin ear to ear. He was (00:00:32) proud. (00:00:32) >> XAI has acquired X in an old stock (00:00:35) transaction. (00:00:36) >> Tesla's first robo taxis are officially (00:00:38) on the road. (00:00:39) >> The company's board proposed a new (00:00:41) compensation package for the CEO that (00:00:43) could pay him just about a trillion (00:00:44) dollars in stock. (00:00:46) >> He gets nothing if he doesn't hit the (00:00:48) numbers. (00:00:49) SpaceX will buy wireless spectrum (00:00:52) licenses from Echoar for its Starlink (00:00:55) satellite network for about 17 billion. (00:00:59) >> 3 2 1. (00:01:05) [Music] (00:01:07) There's a (00:01:12) splash down. (00:01:16) How do you have time? this I I I never (00:01:18) understand you. (00:01:20) >> Yeah. Well, I do work a lot. (00:01:23) >> Ladies and gentlemen, please welcome (00:01:27) >> Elon Musk. (00:01:29) >> All right. Good. (00:01:32) >> All right. Where are you? (00:01:34) >> Alto. (00:01:34) >> You're in Palo Alto and um not (00:01:38) Washington DC. (00:01:39) >> I'm I'm at Tesla Global Engineering (00:01:41) Headquarters in Palo Alto. (00:01:43) >> Yeah. So, no more Washington DC. You're (00:01:46) back at work. You're focused. Yeah. (00:01:48) >> Uh, yeah. I haven't been to DC since (00:01:50) May. (00:01:51) >> Okay. Uh, (00:01:53) >> that was a that was a hell of a side (00:01:54) quest. (00:01:56) >> That was a good Any lessons from your (00:01:59) time in Washington DC? (00:02:01) >> Uh, (00:02:03) the government is basically unfixable. (00:02:10) >> Elon O only. (00:02:13) I support David's noble efforts and this (00:02:16) uh it's good to it's good to have (00:02:18) talented people in the administration uh (00:02:20) but at the end of the day um if you look (00:02:23) at our national debt which is uh (00:02:26) insanely high uh the interest payments (00:02:29) exceed the u defense department I guess (00:02:33) sorry war department uh budget (00:02:37) um and um Nikki Bryzy so if AI and (00:02:43) robots don't solve our national debt, (00:02:45) we're we're toast. (00:02:48) >> Which is a great segue. Um, Optimus is (00:02:52) um I think going to be the greatest (00:02:57) >> uh product in the history (00:03:00) of humanity. (00:03:01) What's the progress like and how much of (00:03:04) your how many of your cycles are going (00:03:06) specifically to Optimus? What's the (00:03:08) timeline? I think you're on version (00:03:10) three, maybe four. Tell us everything. (00:03:14) Uh well, (00:03:17) yeah, everything would take a long time. (00:03:20) >> We've got time. (00:03:23) >> Um (00:03:25) we're we're finalizing the design of (00:03:28) Optimus version 3 (00:03:31) and (00:03:33) uh that that really is going to be a (00:03:35) very remarkable robot. (00:03:37) Um (00:03:39) it will have the essentially the manual (00:03:42) dexterity of a human. So meaning a very (00:03:44) complex hand. (00:03:46) Um the a an AI mind that can navigate (00:03:50) and comprehend reality. (00:03:53) Um and it will be made in very high (00:03:56) volume. Uh those are the three things (00:03:58) that are missing. Like if you see any (00:04:00) other um robotics uh company, they're (00:04:03) missing those three things. Those are (00:04:05) the three really hard things. Um (00:04:10) and uh I I I spent actually at this (00:04:13) point (00:04:15) um (00:04:19) it it might be more of my mental cycles (00:04:22) than anything anything else any other (00:04:24) single thing on Optimus. Uh that's (00:04:27) that's that that's solving for (00:04:30) uh real world AI (00:04:32) uh all of the electro mechanical issues (00:04:35) of Optimus the the supply chain and (00:04:37) production challenges of it because we (00:04:39) have there is no supply chain that (00:04:41) exists for humanoid robots. So it has to (00:04:43) be we have to recreate it from scratch (00:04:45) um and which requires doing a lot of (00:04:47) vertical integration. (00:04:49) um (00:04:51) none of the actuators in Optimus um are (00:04:54) available from an existing supply chain. (00:04:58) Um so but I I think it is accurate to (00:05:01) say that if successful Optimus will be (00:05:04) the biggest product ever (00:05:06) >> and the cost of it at scale 2030 $40,000 (00:05:11) a robot. What what do you think the (00:05:14) first wave of them will cost? And yeah, (00:05:18) when will we be able to buy one to work (00:05:21) on the ranch? (00:05:24) >> I think that the the marginal cost of (00:05:26) production once you hit a million units (00:05:30) per year (00:05:31) uh is probably around the $20,000 range. (00:05:37) Uh it it it sort of depends on how much (00:05:39) you spend on the AI chip in the in the (00:05:43) robot. Um (00:05:46) and you need to achieve a lot of (00:05:48) efficiencies in the actuators. Uh there (00:05:51) are um 26 actuators per arm like 26 (00:05:56) electric like motors, gearboxes and (00:05:58) power electronics. (00:06:00) Um, (00:06:04) so, so, but but the the the AI chip will (00:06:07) be pretty expensive like that that might (00:06:09) be like55 or $6,000 of the of the bill (00:06:13) of materials, maybe more. Um, (00:06:16) and um (00:06:18) but but so I but I think at volume at a (00:06:21) million units a year, the the production (00:06:24) cost is probably on the order of (00:06:25) $20,000, maybe 25, something like that. (00:06:29) And um price will be as a function of (00:06:32) demand. (00:06:32) >> Elon um can you maybe explain to (00:06:35) everybody why the hand is so important (00:06:38) to get right and why you know the (00:06:40) actuator design is so unique and you (00:06:43) know why it's so difficult why nobody (00:06:45) makes it and why you have to start there (00:06:47) almost to build the rest of the the (00:06:49) robot properly. (00:06:52) Well, it turns out the human hands are (00:06:54) incredibly they've evolved to this to be (00:06:57) this incredibly sophisticated machine. (00:06:59) Like the your hand is an in actually a (00:07:03) remarkable thing. It's look look closely (00:07:06) at your hands (00:07:09) and and think of all the things you can (00:07:11) do with your hands, which is a lot. (00:07:13) >> I can think of many things. (00:07:15) >> Yeah, I was just thinking about (00:07:16) something. (00:07:17) >> You know, your hands are very versatile (00:07:19) instrument. Yeah, you can give him a (00:07:21) high five. (00:07:24) >> Very versatile. (00:07:26) Um, you know, you you you can swing a (00:07:28) baseball bat, you can thread needles, (00:07:31) you you can you put thread in a needle. (00:07:34) Uh, you can play the piano with violin. (00:07:37) Um, you know, you could disassemble or (00:07:40) assemble a car. The hands are incredibly (00:07:44) versatile instruments. Um and um most of (00:07:49) the muscles of of the hand are are (00:07:51) actually in the forearm. So your hand is (00:07:54) kind of like a like a like it's like a (00:07:56) puppet. Like it's mostly a puppet. The (00:07:58) mus the muscles are coming from the (00:07:59) forearm and they're pulling the tendons (00:08:02) uh which are you know (00:08:04) also human tendon designs or human human (00:08:06) tendon evolution is incredibly good. Um, (00:08:10) so you you've got this web of tendons. (00:08:13) You you you've got um I think I think (00:08:17) the the human hand is something like (00:08:19) depending on how you count it, 27 or 28 (00:08:21) degrees of freedom per you know in in (00:08:23) the hand. It's uh it's amazing. So (00:08:29) in order to create a robot that can uh (00:08:32) be a generalized uh humanoid, you you (00:08:35) must solve the hand the hands problem. (00:08:37) >> Yeah. We had uh we had (00:08:40) >> it's got hands, needs hands. (00:08:42) >> And so is it like uh when you were first (00:08:44) building Tesla where the supply chain (00:08:46) doesn't exist and now you have to go out (00:08:48) and find folks to work with and you know (00:08:50) build all this vertical integration, get (00:08:52) support. Is it is it literally like it's (00:08:54) just nowhere to be found and (00:08:56) >> you're going to have to build all of (00:08:57) this stuff up? (00:08:59) >> Yes, we we we could not actually buy the (00:09:01) actuators for any amount of money. they (00:09:03) simply didn't exist. Even though there (00:09:04) are, I don't know, 10 20,000 electric (00:09:07) motors out there of various sizes and (00:09:10) shapes. Um, we've had to design uh every (00:09:13) electric motor, gearbox um and and the (00:09:17) controlling electronics from scratch (00:09:18) basically from physics first principles. (00:09:20) >> The good news is you've got a lot of (00:09:22) experience with factories over the last (00:09:24) couple of decades. So, (00:09:27) >> how challenging is this versus Cybert (00:09:30) truck model Y (00:09:32) >> Model X (00:09:32) >> Gigafactory? You know, the Yeah. The (00:09:35) Fabra Egg known as the Model X. Yeah. (00:09:38) >> Right. (00:09:40) >> Um (00:09:42) it's hotter than any any of those (00:09:44) things. (00:09:44) >> Okay. (00:09:46) >> Yeah. (00:09:46) >> Much hotter significantly. Yeah. (00:09:49) Starship. (00:09:49) >> Yes. Well, more No, not Starship's (00:09:53) harder. (00:09:55) Okay. (00:09:56) >> So, somewhere between a Model X and a (00:09:58) Starship? (00:09:59) >> Yeah. (00:10:00) >> Is it is the What's harder, the hardware (00:10:03) or the software? (00:10:07) >> Right now, we're struggling with the (00:10:10) the final design of the hardware. Like I (00:10:13) said, it's really primarily the hand. (00:10:14) Not to just just dismiss the rest of the (00:10:16) robot. the rest of it's also uh (00:10:18) important but but the hands are the (00:10:21) hands inclusive of the forearm are a (00:10:23) majority of the engineering difficulty (00:10:25) of the entire robot and then let's (00:10:27) assume you get past the hardware (00:10:28) challenges how much do you sort of get (00:10:31) for free um based on all the progress (00:10:33) that's happening with LLMs will you know (00:10:36) will consumers just be able to interact (00:10:38) with this talk to the robot ask it to do (00:10:40) things it'll understand and sort of (00:10:42) >> Oh yeah (00:10:44) >> yeah no problem (00:10:45) >> you're spending a lot of time with any I (00:10:47) noticed online. (00:10:49) >> Not not that long. Um maybe I went a (00:10:51) little over the top from Bunning Grom (00:10:53) Imagine, but uh (00:10:54) >> well, but in all seriousness, those (00:10:58) characters and these robots that seems (00:11:01) to be, you know, like maybe they (00:11:03) >> you could get the embodiment of Annie, I (00:11:05) suppose. (00:11:06) >> Yeah. Why why the human form factor, (00:11:08) Elon? You could make something that's (00:11:10) maybe better than a human or maybe (00:11:12) simpler than a human to do specific (00:11:14) tasks and maybe better than a human to (00:11:15) do more things than a human can do. How (00:11:17) do you decide to make it just like a (00:11:19) human? (00:11:22) >> Well, if you wanted to do all the things (00:11:23) that a human can do, it turns out you (00:11:25) need a humanoid robot. (00:11:28) Um, so if you want to just do a subset, (00:11:31) it that's much easier. (00:11:33) Um but uh (00:11:37) it turns out humans evolved to this the (00:11:39) shape and capabilities that we we we (00:11:41) have. Um it it it for for good reasons. (00:11:46) Uh there actually is that there is (00:11:51) like there's value to having five, you (00:11:53) know, four fingers and the thumb. Um and (00:11:56) even the pinky actually is is quite (00:11:58) useful. Um toes are much more aggression (00:12:02) walk but but but the fingers (00:12:06) >> well also humans humans have designed (00:12:08) the world as well. So we designed it for (00:12:11) us. So (00:12:12) >> if you can make a humanoid robot it'll (00:12:14) be immediately backwards compatible with (00:12:15) what we've built the world for. (00:12:16) >> Precisely. (00:12:17) >> Elon there's another there's another (00:12:19) part of um the robot. So there's the (00:12:22) LLMs, there's the actuation in the (00:12:24) hands, but also there's the um the (00:12:29) silicon that runs it. And there was, you (00:12:32) know, Dojo, I think you you posted on X (00:12:36) AI5 and AI6, and it just seemed like you (00:12:38) were incredibly excited about the (00:12:41) direction in which the silicon layer was (00:12:43) also going. Can you tell us about that (00:12:45) and what that is and what what what what (00:12:48) are we what are we building here? What (00:12:50) is being built? Is it a complement to (00:12:52) everything that exists in the world? Is (00:12:54) it a potential long-term competitor? (00:12:55) What is it? (00:13:00) >> Um, yeah. So, (00:13:02) at at Tesla, we basically had two (00:13:05) different chip programs. One dojo and (00:13:08) one uh dojo on the training side and (00:13:11) then what we call, you know, AI for it, (00:13:14) which just our inference chip. um uh (00:13:18) that the AI Force is currently shipping (00:13:20) in all vehicles. Um and we're finaliz (00:13:23) finalizing the design of AI5 which will (00:13:25) be an immense jump from AI4. Um by some (00:13:29) metrics the improvement in AI5 will be (00:13:33) 40 times better than AI4. (00:13:36) >> Wow. (00:13:36) >> So 40% 40 times. Um and and uh this is (00:13:42) because we work so closely at a very (00:13:44) fine grade level on the AI software and (00:13:47) the AI hardware. So we know exactly (00:13:50) where the limiting factors are and and (00:13:54) um and so effectively the AI hardware (00:13:56) and software teams are co-designing the (00:13:57) chip. Um (00:14:00) >> so a 40x improvement in the silicon I (00:14:02) think then as it as everybody here in (00:14:05) the audience experiences it is that just (00:14:07) an almost like an order of magnitude (00:14:09) increase in the quality of FST and the (00:14:12) safety that you experience as a Tesla (00:14:14) driver and then the quality of the robot (00:14:16) like where does it all manifest when you (00:14:18) when you you know bring it up and (00:14:20) actually get it into production? (00:14:23) Yeah to be precise the 40x is on if you (00:14:26) say like compared to the worst (00:14:28) limitation on AI4 which is running the (00:14:32) softmax operation. (00:14:34) >> Yeah (00:14:34) >> we currently have to run softmax in (00:14:36) around 40 steps in emulation mode (00:14:39) whereas that'll be just be done in a few (00:14:42) steps uh natively in AI5. (00:14:45) Um AI5 trip will also be u easily handle (00:14:48) mixed precision um models. So you don't (00:14:52) have it'll dynamically handle mixed (00:14:54) precision. There's a bunch of sort of (00:14:56) technical stuff that AI will do a lot (00:14:58) better. Um in terms of of nominal sort (00:15:02) of uh raw compute, it's it's eight times (00:15:05) more compute. Um about nine times more (00:15:09) memory uh roughly five times more memory (00:15:12) bandwidth. Um so uh but because we're (00:15:16) addressing some core limitations in AI4, (00:15:19) you multiply that by that that 8x (00:15:22) computer improvement by another 5x (00:15:23) improvement because of of uh (00:15:26) optimization at a at a at a very fine (00:15:28) grain silicon level of things that are (00:15:30) currently suboptimal in AI4. That's (00:15:32) where you get the 40x improvement. (00:15:34) >> You had um I'll keep going, keep going. (00:15:36) Uh so now now that said I I'm I am (00:15:39) confident that the current uh chips uh (00:15:42) AI AI4 chips that are in the cars will (00:15:45) uh achieve self-driving safety that is (00:15:49) at least two to three times that of of (00:15:52) human and and maybe even 10x. Um and the (00:15:55) software that uh will be released for (00:15:57) that is is coming out over the next few (00:16:00) months. So version 14 will be the (00:16:04) biggest uh upgrade in Tesla software (00:16:06) since version 12. Um we are increasing (00:16:10) the uh parameter count by an order of (00:16:12) magnitude. Um the there's there's (00:16:17) there's a lot of uh reinforcement (00:16:19) learning that's been used. there's um (00:16:24) we we there there there's like you can (00:16:27) think of AI sort of as a way of (00:16:29) compressing reality and and and some of (00:16:31) those compression steps uh we uh were (00:16:35) too lossy and and we addressed the (00:16:37) lossiness in the compression steps. Um (00:16:40) so the these are all software updates (00:16:42) that'll that'll go out. So just over (00:16:43) there updates um your car is going to (00:16:46) feel like it is sentient by the end of (00:16:48) the year. (00:16:50) >> Yeah, it feels that way already to be (00:16:52) honest. Um I saw in the trades that you (00:16:56) spent about $17 billion on some spectrum (00:16:59) and that um (00:17:01) >> yeah um so some couch change um to (00:17:05) enable your satellites and the Starlink (00:17:08) network to connect directly with phones. (00:17:11) What will that look like in a year or (00:17:13) two? Are we going to drop our Verizon (00:17:15) account and just expand our Starlink (00:17:18) account? (00:17:20) >> Uh, thank you. (00:17:22) >> We're kind of hoping cuz Verizon kind of (00:17:24) sucks. (00:17:25) >> How How many of you want a Starlink (00:17:28) phone? (00:17:28) >> Who wants a Starlink phone? (00:17:32) >> Is it Is it technically possible? (00:17:34) >> I know you can't see it, but it's (00:17:35) everyone. (00:17:35) >> Yeah, All right, cool. Um so this is a (00:17:40) kind of a long-term thing. Uh it it will (00:17:44) allow (00:17:45) SpaceX to h uh deliver high bandwidth (00:17:51) connectivity directly from the (00:17:52) satellites to the phones. Um but uh (00:17:57) there are hardware changes that need to (00:17:59) happen in the phone. So the since these (00:18:01) frequencies are not supported in current (00:18:03) phones uh that the chipset has to be (00:18:06) modified to add these frequencies um and (00:18:10) that probably is a 2-year time frame. So (00:18:11) the phones that um are able to use the (00:18:14) spectrum that was acquired probably (00:18:16) start shipping in around 2 years. Um and (00:18:20) um and then we also need to build the (00:18:22) satellites that are going to communicate (00:18:24) on those frequencies. So, in parallel, (00:18:27) we're building the satellites and (00:18:28) working with the handset makers to add (00:18:31) these frequencies to the phones. Um, and (00:18:33) then the the satellites and the phones (00:18:36) will then handshake very well to achieve (00:18:38) high bandwidth connectivity. But the net (00:18:40) effect is that you should be able to (00:18:41) watch uh videos uh anywhere on your (00:18:44) phone. (00:18:46) >> Wow. (00:18:47) >> And it's going to be crazy. And what and (00:18:49) do these do these frequencies would they (00:18:51) work indoors inside buildings, you know, (00:18:53) like like your phone currently does? (00:18:55) Okay. (00:18:57) >> And so will you be able to have (00:18:58) basically like (00:18:59) >> if you if if you're in a building with a (00:19:00) with a like a a thick metal roof then (00:19:02) No. But um (00:19:03) >> no the the same types of of (00:19:06) >> Yeah. Normal normal homes. Yes. (00:19:08) >> Yes. (00:19:08) >> Elon is your vision for this that (00:19:10) instead of you know having an AT&T (00:19:11) account or and then roaming when you're (00:19:13) in the UK or you're in India. It's just (00:19:16) we could have one direct deal with (00:19:18) Starlink. It works all over the world (00:19:20) eventually. Not today but at some point. (00:19:23) Is that the end goal? That basically we (00:19:25) don't need a regional carrier. We have a (00:19:28) global carrier and that would be you. (00:19:30) >> Uh that that would be one of the (00:19:32) options. To be clear, we're not going to (00:19:33) put the other carriers out of business. (00:19:34) They're still going to be around cuz (00:19:36) they they own a lot of Spectrum. So, uh (00:19:39) there's uh but but yes, you you should (00:19:42) be able to have a Starlink uh like you (00:19:44) have like you have an AT&T or T-Mobile (00:19:47) or Verizon or whatever, you should be (00:19:48) you could have a you know account with (00:19:51) Starlink that uh works with your you (00:19:54) know Starlink uh antenna at home uh free (00:19:58) Wi-Fi as well as on your phone and um (00:20:02) yeah it would be a comprehensive (00:20:03) solution for high bandwidth at home and (00:20:06) for high bandwidth direct to sell. (00:20:08) >> Could you buy some carriers to have more (00:20:11) >> spectrum? (00:20:13) >> Maybe you could buy Verizon. (00:20:14) >> Not out of the question. I suppose it (00:20:16) that may happen. (00:20:18) >> Let's talk about um let's talk about (00:20:20) Starship. You just had a really what (00:20:23) appeared to be a phenomenal um launch. H (00:20:28) how close is it to, you know, being (00:20:31) predictable and ready to go in a (00:20:33) commercial setting? (00:20:37) I I I think we'll recover the ship next (00:20:40) year. Um we've got one more launch of (00:20:44) the um Starlink version two (00:20:48) uh uh stack that there's only one one (00:20:52) uh booster and ship left that's in the (00:20:54) version two uh design. Uh and then (00:20:58) thereafter it's it's version three which (00:21:00) is a gigantic upgrade cuz that's got (00:21:02) Raptor 3. Um, and pretty much everything (00:21:05) changes on the rocket with version 3. (00:21:07) Um, so version 3, you know, might have (00:21:10) some initial teething pains, uh, cuz (00:21:13) it's such a radical redesign. (00:21:15) Uh, but, uh, it's it's capable of over (00:21:20) 100 tons to orbit fully reusable. Um, (00:21:23) and I think it's I think I think um (00:21:26) unless we have unless we have some very (00:21:28) major setbacks, uh, SpaceX will (00:21:31) demonstrate uh, full reusability next (00:21:34) year uh, catching both the booster and (00:21:36) the ship um, and being able to deliver (00:21:39) over 100 tons to a useful orbit. (00:21:43) >> What does the best rocket in the world (00:21:45) do now in terms of tonnage to space? (00:21:50) Uh well in terms of (00:21:53) uh sort of commercial rockets there's (00:21:55) there's Falcon Heavy. (00:21:56) >> Yeah. (00:21:56) >> Uh which will do uh in (00:22:00) um with with side booster reuse uh will (00:22:04) do about 40 tons. (00:22:07) >> So this is five times bigger. Yeah. (00:22:11) >> Well two and a half times bigger in but (00:22:13) but Starship would be full reuse full (00:22:15) reusability. (00:22:16) >> Got it. Okay. So everything comes back (00:22:19) >> Elon after after the explosion that (00:22:22) happened um with the the the the failed (00:22:24) launch (00:22:26) >> um there was a lot of (00:22:28) >> sorry (00:22:29) >> which which failed (00:22:30) >> oh the more recent one the more recent (00:22:32) the starship with (00:22:33) >> the big boom yeah (00:22:34) >> the big boom on the base and and and (00:22:36) there was a lot of (00:22:37) >> there was a lot of proclamations that (00:22:39) there's going to be environmental and (00:22:41) FAA and all these other sorts the (00:22:43) recovery back to the launchpad again was (00:22:46) incredible. (00:22:47) fast. How did you get back so fast? Not (00:22:51) just technically and work-wise, but just (00:22:53) like regulatory clearance-wise because (00:22:55) they said there were going to be all (00:22:56) these questions and reviews and so on. (00:22:58) How how did you guys manage that? (00:23:01) >> Uh well, there were a lot of questions (00:23:02) and reviews. We got through them all. Um (00:23:05) and credit to the SpaceX team. They (00:23:06) worked incredibly hard and they uh got (00:23:10) the next trip and booster tested and on (00:23:13) the pad and and flown and um yeah, huge (00:23:17) credit to the SpaceX team. Very proud of (00:23:19) them for (00:23:20) >> doing doing such a job, a great job (00:23:22) recovering. (00:23:23) >> Um (00:23:24) I mean creating a fully reusable orbital (00:23:27) rocket is one of the hottest engineering (00:23:29) problems ever. (00:23:32) and certainly, you know, a candidate for (00:23:35) most difficult engineering project ever. (00:23:38) You know, it's on the podium at least. (00:23:40) Um, (00:23:42) so it's a that that's been the goal of (00:23:44) SpaceX from the beginning from 2002. Um, (00:23:48) and here we are 23 years later. So, it's (00:23:50) it's a long journey (00:23:53) and um (00:23:55) with with a a super talent like by far (00:23:57) the I think the most talented group of (00:23:59) rocket engineers that ever been (00:24:00) assembled. Um and uh and we're finally (00:24:05) next year I think we'll be able to (00:24:08) achieve full reusability. (00:24:10) >> Elon, what are the big um technical (00:24:13) blockers that you're focused on there (00:24:14) between now and that full reusability? (00:24:16) Are there some showstoppers where you're (00:24:19) just kind of literally just obsessing (00:24:22) over trying to figure out still or is it (00:24:24) more about getting through a laundry (00:24:28) list of your learnings and just (00:24:29) integrating it into the next launch? (00:24:33) >> Well, (00:24:36) that the (00:24:38) for for full reusability of the ship, (00:24:41) there's still a lot of work that remains (00:24:43) on the heat shield. So, no one's ever (00:24:45) made a fully reusable orbital heat (00:24:47) shield. Like the shuttle heat shield uh (00:24:50) had to go through nine months of repair (00:24:52) after every flight, (00:24:53) >> right? (00:24:54) >> Um so, no one has ever made a fully (00:24:57) reusable orbital heat shield. (00:24:59) >> And is that a material science problem (00:25:01) or is that an engineering problem or (00:25:03) both? (00:25:03) >> Uh yeah, I mean it's a material science (00:25:06) engineering problem. So, it's but we (00:25:09) really are uh looking at the fundamental (00:25:12) physics here. Um again physics first (00:25:15) principles and trying to figure out how (00:25:17) do we make something that (00:25:20) um is uh (00:25:23) you know can can withstand the heat is (00:25:26) very light doesn't transmit the heat to (00:25:29) the the primary sh (00:25:31) >> Yeah. (00:25:32) >> primary structure um and uh (00:25:38) whose integ (00:25:41) >> Yeah. (00:25:42) >> Um (00:25:43) Uh, and then as you ascend, if you hit (00:25:46) some rain, you know, the tiles don't (00:25:48) dissolve in rain. There's there's a lot (00:25:51) of different issues and and then you (00:25:54) really need to know that these tiles are (00:25:56) working. You can't uh, (00:25:59) you know, go through this laborious (00:26:00) inspection. So, it really needs to be (00:26:02) we're, you know, these these tens of (00:26:06) thousands of tiles all work and don't (00:26:10) need to be refurbished or checked one by (00:26:13) one as was the case with the shuttle. (00:26:15) >> Can we maybe um switch now? It's I mean, (00:26:18) who who else were you talked about (00:26:20) Tesla, then you go to SpaceX? Yeah. Now, (00:26:22) I' I'd like to ask you some questions (00:26:24) about Grock and um XAI. Um you want to (00:26:29) just give us an update? I think you you (00:26:30) kind of talked about where the nextG (00:26:32) model is and you said something (00:26:34) incredible. I still don't think people (00:26:35) really understand it which is you know (00:26:37) there's going to be a next training run (00:26:38) where you expect you know not to start (00:26:42) from the you know common web and common (00:26:45) crawl where you expected an enormous (00:26:48) amount of synthetic data. Just tell us (00:26:50) about how uh the evolution of Grock is (00:26:53) going and this innovation and why it's (00:26:55) so important. (00:26:59) Yeah. So we're we're running a lot of (00:27:02) using a lot of of inference compute and (00:27:04) um and reasoning to look at all of the (00:27:09) source data which is really the corpus (00:27:10) of human knowledge and then (00:27:14) uh thinking about each piece of (00:27:16) information and then adding mod adding (00:27:18) what's missing um and correcting (00:27:22) correcting mistakes and removing (00:27:25) falsehoods from the from that training (00:27:27) data. So it's it's it's like if you take (00:27:30) say Wikipedia as an example but this (00:27:32) really applies to to books, PDFs, (00:27:35) uh the websites, (00:27:37) uh every form of information. Um the (00:27:42) the Grock is using um heavy amounts of (00:27:46) inference compute to say to look at at (00:27:49) an example a Wikipedia page and say uh (00:27:52) what is true, partially true or false or (00:27:56) missing (00:27:58) uh in this page. Now rewrite the page to (00:28:02) in to correct the remove the falsehoods (00:28:07) uh (00:28:08) uh correct the half-truths and add the (00:28:11) missing context. (00:28:13) >> Well, Elon, by the way, could you just (00:28:15) publish that? Could we create like a (00:28:16) groipedia? I mean, that would (00:28:18) >> Yeah, especially for our bio pages, (00:28:19) which are a disaster. (00:28:21) >> Wikipedia is so biased and it's it's a (00:28:25) constant war. you know, if something (00:28:26) gets corrected, five minutes later, (00:28:28) there'll be an army of people trying to (00:28:30) >> I mean, it's become hyperartisan and (00:28:33) there's activists all over it. (00:28:35) >> So, if you do fix, for example, (00:28:37) Wikipedia as a source of truth, (00:28:40) >> it'd be great to publish that just so (00:28:42) the world has it. (00:28:45) >> All right, I'll talk talk about that. (00:28:47) So, talk to the team about that like (00:28:49) Groedia or whatever. This here's the (00:28:51) Groedia version. (00:28:53) >> It' be interesting. Yeah. and then just (00:28:55) have it out there for just a few minutes (00:28:56) >> where in terms of um people here like it (00:29:00) um in terms of training Gro 5 um you're (00:29:04) you're scaling up your supercluster in (00:29:07) Colossus in in Memphis (00:29:09) >> can yeah have a second one (00:29:12) >> yeah can could you give us an update on (00:29:13) that and then also as part of that um (00:29:16) where are we in the scaling laws um if (00:29:19) you scale a bigger cluster do you get a (00:29:21) more powerful AI model is there a point (00:29:23) of dimin diminishing returns or like how (00:29:26) much more compute if you throw twice as (00:29:28) much compute at it do you get a 10% (00:29:31) better model do you get 100% better (00:29:33) model like is it log linear what what I (00:29:36) guess how much more juice is there left (00:29:38) in scaling hardware do you think (00:29:42) >> I think I think there's a natural (00:29:44) logarithmic function associated with the (00:29:47) amount of compute so (00:29:49) uh then like say for argument sake like (00:29:52) 10x more compute will double the (00:29:53) intelligence. (00:29:55) Maybe that's that that might be a rough (00:29:58) rule of thumb, but you know, that still (00:30:00) means that, you know, you go from 100 IQ (00:30:02) to 200 IQ. Still pretty pretty big deal. (00:30:05) Um, (00:30:06) so I (00:30:10) and and I think I think we'll see (00:30:11) intelligence continue to scale all the (00:30:13) way up to where, you know, most of the (00:30:16) power of the sun is harnessed for (00:30:19) compute and then ultimately most of the (00:30:21) power of the galaxy, you know, sort of (00:30:23) cautev 2, cautev 3 scale uh compute. Um (00:30:28) so I guess once you think of artificial (00:30:31) intelligence not as sort of this (00:30:34) you know a destination that you reach (00:30:36) but really uh as part of the overall (00:30:39) escalation of intelligence (00:30:41) um that that that we are are aware of. (00:30:45) Um (00:30:46) you know human intelligence has also (00:30:48) scaled as you've have as the population (00:30:50) has increased um and we've been able to (00:30:54) store more and more information. uh (00:30:56) human intelligence has scaled. Now human (00:30:58) because of population declines and low (00:31:01) growth rate, human intelligence is is (00:31:04) somewhat plateauing um and will actually (00:31:05) decline. And (00:31:08) my guess is that (00:31:12) I I I I think that we might have AI (00:31:15) smarter than any single human at (00:31:16) anything as soon as next year. (00:31:20) >> Wow. (00:31:20) >> Um (00:31:21) >> Yeah. and and and then and then probably (00:31:23) within five like say 2030 probably AI is (00:31:27) smarter than the sum of all humans. (00:31:29) >> Do you think do you think humans are on (00:31:31) the decline because the AI is evolving? (00:31:34) Do you think there's this evolution of (00:31:36) the ecosystem on Earth that's underway (00:31:39) that we don't really understand the (00:31:40) structure of what's going on? But (00:31:47) >> maybe yeah, maybe we implicitly (00:31:49) know that it's coming. Um, (00:31:59) >> yeah. (00:32:00) >> I I I I mean I hope the birth rates turn (00:32:02) around. I'm a I'm a big proponent of (00:32:05) increased birth rate. Uh, obviously. (00:32:10) >> Well, you doing anything about it or no? (00:32:13) >> Yeah. I'm trying to set a good example. (00:32:16) You know, we had a big conversation at (00:32:18) this conference we didn't expect, which (00:32:20) is suicidal empathy, the West, (00:32:24) >> this um declining birth rate. Uh I (00:32:28) noticed you've been pretty active about (00:32:29) it (00:32:30) >> and open borders (00:32:31) >> and open borders is like let the (00:32:32) invaders in. Could all three of those be (00:32:34) the same thing? It (00:32:35) >> it seems like there's a number of (00:32:37) symptoms of the West being suicidal. The (00:32:40) most obvious one being the birth rate is (00:32:42) not a replacement level. So obviously if (00:32:43) that continues indefinitely then the (00:32:46) west will literally not reproduce enough (00:32:48) to replace itself. But there's other (00:32:50) things too. There's the fact that the (00:32:52) borders were totally opened to the point (00:32:54) where western culture the social fabric (00:32:57) start to come apart and you see this (00:32:59) especially in Europe where there um you (00:33:02) know the indigenous cultures of the UK (00:33:04) or France or Germany are starting to um (00:33:07) potentially be taken over by by cultures (00:33:09) of people who are brought in and aren't (00:33:11) assimilating. You have crime where, you (00:33:14) know, we have this case on social media (00:33:16) right now, this young woman, Ire Ina, (00:33:18) who's just (00:33:20) >> killed in a senseless way on a subway. (00:33:22) >> Uh, which is horrific enough in and of (00:33:24) itself, but then in addition to that, (00:33:26) the elite media just for whatever reason (00:33:28) just refused to cover it, like it didn't (00:33:30) exist. (00:33:31) >> Um, so you have this issue of crime (00:33:33) that's not being addressed or even (00:33:35) acknowledged (00:33:36) >> and no acknowledgement of this. like (00:33:38) it's almost like we're trying to deny (00:33:40) the reality of the spiral (00:33:43) >> and this Yeah. So you have the you have (00:33:46) all these data points um that seem to (00:33:48) suggest that um the west uh is suicidal (00:33:52) or doesn't you know doesn't seem to want (00:33:55) to defend itself or propagate itself. Um (00:33:58) look I think everyone in this room (00:33:59) thinks that um life is awesome right? I (00:34:02) mean it's (00:34:02) >> pretty great and I think (00:34:04) >> worth living. (00:34:05) >> Yeah. And when when Alex Karp was here (00:34:07) earlier today defending the West, that (00:34:09) got some of the loudest applause at the (00:34:11) conference. So, uh I guess we probably (00:34:13) don't really understand what's going on. (00:34:15) We don't really (00:34:16) >> Yeah. What's your take, Elon? Cuz you (00:34:17) you know, (00:34:18) >> what's your take on the suicide of the (00:34:19) West? (00:34:20) >> Yeah. (00:34:21) >> What's What's (00:34:22) >> I'm very worried about it. (00:34:23) >> Yeah. (00:34:24) >> I'm very worried about it. Um you know, (00:34:26) I think there's there's (00:34:29) just the actions of the West are (00:34:30) indistinguishable from suicide. (00:34:33) So, but it's (00:34:37) and look, at least in America, there's (00:34:40) there's there's generally a sense of (00:34:42) optimism, but when's the last time you (00:34:45) you talked to someone from Europe who (00:34:47) lives in Europe who's optimistic? (00:34:50) >> Not for a while. Yeah. (00:34:52) >> Decades, (00:34:52) >> like even one. (00:34:55) >> It's rare. So I I think unless people (00:34:59) have a sense of optimism and purpose (00:35:01) about the future, they (00:35:05) suicide might be just what happens. Um (00:35:08) like like like having a child is an act (00:35:10) of optimism about the future. So uh if (00:35:15) you're not optimistic, this Yeah. (00:35:21) So, so I think we need to maybe (00:35:24) give people a sense of optimism and (00:35:26) excitement about the future and and a (00:35:28) belief that the future will be better (00:35:29) than the past um and they'll be more (00:35:32) interested in having kids. (00:35:34) >> Did did religion play a role in the (00:35:36) past, Elon, to kind of plate and make (00:35:38) folks feel that way (00:35:40) >> when they (00:35:41) >> Yeah, I think so. uh (00:35:44) the nature abhores a vacuum and if you (00:35:47) take away religion then I think you (00:35:50) actually you you you get something in (00:35:53) its place which is actually worse than (00:35:56) what was there before I mean it's like (00:35:58) destructive basically you get you get (00:35:59) like the white work mind virus filling (00:36:01) filling the hole that religion used to (00:36:04) have taking the place of of of religion (00:36:08) you get these dystopian de facto (00:36:11) religions (00:36:12) um that uh that that are very very (00:36:16) self-destructive. (00:36:17) Um so (00:36:20) I I think perhaps (00:36:23) some some sort of re revival of religion (00:36:25) or at least what we need is is um some (00:36:28) coherent philosophy (00:36:31) that people can get excited about. Um (00:36:34) you know I mean for me it's a philosophy (00:36:36) of curiosity. I'm curious about the (00:36:38) nature of the universe and I want to go (00:36:40) out there and I want humanity to be out (00:36:42) there exploring the stars. Um maybe (00:36:46) meeting alien civilizations. Uh maybe in (00:36:49) some cases we we see the ruins of a long (00:36:51) dead alien civilization but they were (00:36:54) they were very strong for 10 million (00:36:55) years. Um you know the kind of stuff (00:36:57) that you see in Star Trek in in a (00:36:59) non-dystopian sci-fi book or or movie or (00:37:03) show. Um, and so I'm just I have I have (00:37:06) a philosophy of curiosity of of like I (00:37:08) just want to know what's going on. And (00:37:10) and in order to know what's going on, we (00:37:11) we must have u an an increase in the in (00:37:15) the scope and scale of consciousness, we (00:37:18) must we must expand uh consciousness. We (00:37:21) must grow. We must grow humanity and we (00:37:23) must extend humanity in order to (00:37:27) comprehend the and to to understand the (00:37:31) universe or even what what question (00:37:32) should we should ask about the answer (00:37:35) that is the universe. Um you know Doug (00:37:38) Douglas Adams (00:37:39) book the hitchhikers guide to the galaxy (00:37:41) is actually a a deep book on philosophy (00:37:44) disguised as humor. Um and what the (00:37:47) point he was trying to make in that book (00:37:49) was that u the questions are the really (00:37:52) the hard part. The answer is the (00:37:54) universe. Like the answer is everything (00:37:56) you see around you. But but but one of (00:37:59) the questions that we don't know to ask. (00:38:00) >> Yeah. (00:38:01) >> Um now now some of the questions I guess (00:38:03) I we I do know. I'd like to know is the (00:38:05) standard model of physics correct about (00:38:07) the origins of the universe? Are we (00:38:09) actually 13.8 billion years old? Um how (00:38:11) does the universe end? Does it end in a (00:38:13) heat death or in some other way? Um, you (00:38:16) know, (00:38:17) >> a black hole. (00:38:19) >> We might be. (00:38:20) >> Um, (00:38:22) >> Elon, can you talk about (00:38:24) >> the whole sort of simulation question? (00:38:25) Are we a simulation? Maybe. (00:38:27) >> Where does the uh where do you think we (00:38:28) find the answer first? In AI or in the (00:38:32) stars? (00:38:34) Because you're pursuing both obviously. (00:38:39) >> Yeah. (00:38:40) I I I don't know if if I I hope I hope (00:38:44) more people can get behind a (00:38:47) philosophy of curiosity. (00:38:49) >> Yeah. (00:38:50) >> Because I think it's very exciting. (00:38:54) >> Yeah. (00:38:54) >> Um and and and and inherently (00:38:57) optimistic. (00:38:59) Um you like because there's there's this (00:39:02) amazing sense of wonder (00:39:05) about the nature of the universe. And (00:39:07) when you just when you uncover some (00:39:09) secret in the universe, that's amazing. (00:39:12) And you're like a whole world of (00:39:15) understanding is opened up. I mean, we (00:39:18) we used to not even know where all the (00:39:20) continents were. Um, you know, used to (00:39:22) be like just the map would be there be (00:39:24) dragons and like all we know is that (00:39:26) when they sailed in that direction, they (00:39:28) didn't come back. (00:39:31) >> I mean, the moon base, (00:39:32) >> that's all that's all they knew. I I (00:39:35) kind of feel like the moon base or just (00:39:37) going to the moon for real this time (00:39:39) would be a big step in the right (00:39:40) direction. You still have the moon uh (00:39:43) planned. What's the status of that? Is (00:39:45) is that still on the agenda? (00:39:47) >> Yeah, I I think it I think having I (00:39:49) think we want to try to reach new (00:39:51) heights as a civilization. (00:39:53) >> Yeah. (00:39:53) >> So, I think it's it's fine to go to the (00:39:55) moon, but but we should go to the moon (00:39:57) in order to establish a lunar base, like (00:40:00) a a lunar research base. (00:40:02) >> Yeah. Um, I mean there are parts of the (00:40:04) moon that are perhaps older than parts (00:40:06) of of Earth. Um, and we we we might (00:40:10) understand more about the nature of the (00:40:11) universe if we had a science base on the (00:40:13) moon. (00:40:13) >> Um, that would be very cool. And then we (00:40:16) we obviously want to go beyond the moon (00:40:18) uh to Mars and uh build a (00:40:21) self-sustaining city on Mars. the I I I (00:40:24) do think that uh that that there is a (00:40:27) fork in the road of human destiny where (00:40:30) um if we can establish a self-sustaining (00:40:33) city on Mars with the the key test being (00:40:36) if the resupply shifts from Earth stop (00:40:38) coming for any reason does Mars continue (00:40:41) to to prosper or does it die out (00:40:44) >> at the point at which Mars is able to uh (00:40:47) prosper and grow on its own the probable (00:40:50) lifespan of consciousness is (00:40:51) dramatically greater. (00:40:53) because we are no longer dependent on (00:40:55) everything going right on Earth. You (00:40:58) know, there's there's always some (00:40:59) possibility of self annihilation on (00:41:01) Earth with the World War II or or a (00:41:03) supervirus or um or or a meteor like (00:41:07) extin, you know, that destroyed the (00:41:09) dinosaurs. We know from the fossil (00:41:10) record that there've been many mass mass (00:41:12) extinction events. So uh the question (00:41:16) that I sort of was wondering about is (00:41:19) will civilization will the (00:41:21) civilizational arc continue to ascend (00:41:23) such that we can make Mars self- (00:41:26) sustaining before the civil (00:41:28) civilizational ark descends (00:41:31) >> um because the the window of opportunity (00:41:34) to make life multilanetary (00:41:36) exists now for the first time in the 4 (00:41:38) and a half billion year history of (00:41:40) earth. (00:41:41) >> Yeah. Elon, let's assume that we get (00:41:43) there and you're there. (00:41:46) >> Um, you know, you'd be the elder (00:41:49) statesman. You'd have the moral (00:41:50) authority of Mars. (00:41:52) How do you run Mars? (00:41:57) >> But I just there's this point that I I (00:42:00) think I I want to just emphasize again (00:42:04) that that's that's it's more important (00:42:05) than the form of governance on Mars or (00:42:08) who's there in the early days. What (00:42:10) really matters is that Mars um is self- (00:42:15) sustaining that we are truly a (00:42:18) multilanet species and s such that we've (00:42:21) achieved planetary redundancy so that (00:42:23) that if if something and obviously we (00:42:26) should do everything possible to make (00:42:27) sure life on Earth is great but there's (00:42:29) always some risk that of an annihilation (00:42:31) event on Earth. (00:42:32) >> Yeah. (00:42:33) >> Um like I said self annihilation or some (00:42:35) natural disaster. Um (00:42:38) and uh and so the the probable lifespan (00:42:41) of consciousness increases dramatically (00:42:44) as soon as uh as soon as we are (00:42:47) multilanet species with the key test (00:42:49) being can Mars survive if the resupply (00:42:51) ships stop coming. So so getting like (00:42:54) the first missions to Mars are not that (00:42:56) important. The what matters is can you (00:42:58) get sufficient tonnage tonnage to Mars (00:43:01) such that Mars can prosper on its own. (00:43:04) Um, and that means it has to have all of (00:43:06) the ingredients of civilization. It it (00:43:08) it's not just that you need to build, (00:43:10) for example, a chip factory on Mars or (00:43:12) ship fab on Mars, but you you need the (00:43:14) ability to build. (00:43:16) >> Do you do you have a sense of the time (00:43:18) scale? Like, let's assume Starship is at (00:43:20) a state starting in, you know, 2026. (00:43:23) Then there's going to be a bunch of (00:43:24) testing. Obviously, there's going to be (00:43:25) a bunch of early testing. We only have (00:43:28) certain launch windows. So, there's a (00:43:29) bunch of time constraints. Is that is (00:43:32) this a 50-year thing in your mind? Is it (00:43:34) a 150 year thing? Is it something that (00:43:36) is for our generation or is it our (00:43:38) children's generation? Where do you see (00:43:40) that point if it's optimally possible? (00:43:43) You know, if things go and break our (00:43:45) way, (00:43:47) >> um I think it can be done in in 30 (00:43:49) years. (00:43:51) Um (00:43:51) >> wow. So if provided there's an (00:43:53) exponential increase in the in the (00:43:55) tonnage to Mars with each successive (00:43:58) Mars transfer window, which is every two (00:44:00) years. So every two years the the (00:44:03) planets align and you can you can (00:44:04) transfer to Mars. Um, so (00:44:08) I I think in (00:44:11) roughly 15, but maybe as few as 10, but (00:44:15) 10 to 15ish Mars transfer windows. If (00:44:18) you're um (00:44:22) seeing exponential increases in the (00:44:24) tonnage to Mars with each Mars transfer (00:44:26) window, then it should be possible to (00:44:28) make Mars self-sustaining um in in about (00:44:31) call it roughly 25 years. (00:44:33) >> Amazing. That's incredible. All right, (00:44:36) ladies and gentlemen, Elon Musk, we'll (00:44:39) see you when we're back in town. We miss (00:44:41) you. We'll (00:44:42) >> see you in person next time. (00:44:44) >> Thank you, brother. All (00:44:46) right.