Elon Musk - Where AI's Biggest Impact Will Be

um sh shifting gears to AI uh Peter was here earlier and he was talking about how so far the only company to really make money off AI is NVIDIA with the chips um do you have a sense yet of where you think the big applications will be from AI is it going to be an enabling self-driving is it going to be enabling robots is it transforming Industries I mean it's still I think early in terms of where the big business impact is going to be do you have a sense yet I I mean I think

I think the they the spending on AI probably runs ahead of I mean does run ahead of the revenue right now that's there's no question about that um but the rate of improvement of AI is faster than any technology I've ever seen by far and and and it it's I mean like the for example the touring test used to be a thing now you know your basic uh open- Source random llm writing on a freaking Raspberry Pi probably could uh you know be the during test um so there's I I I I think actually like

like the the good future of AI is one of immense proc Prosperity where there is an age of abundance no shortage of goods and services everyone can have whatever they want unless except for things we artificially Define to be scarce like some special artwork um but but anything that is a manufactured good or provided Service uh will I think with the adment of AI plus robotics that the cost of goods and services will be will Trend to zero like I'm not saying it be actually zero but it'll be it every everyone will be able to

have anything they want that that's the good future of course and you know in my view that's probably 80% likely so look on the bright side only 20% 20% probability of annihilation it's nothing um is is the 20% like what does that look like I don't know man I mean frankly I do have to go engage in some degree of of deliberate suspension of disbelief with respect to AI in order to sleep well um and even then um because I I I I think the actual issue the the most likely issue is like well how

do we find meaning in a world where AI can do everything we can do a bit better that that is that is perhaps the bigger challenge um although you know at this point I know more and more people who are retired and they seem to enjoy that life so uh but I think that that may be maybe there'll be some crisis of meaning like because the computer can do everything you can do B better so maybe that'll be a challenge um but but really uh you know you need you need the sort of end factors

you need the the autonomous cars and you need the sort of humanoid robots or your general purpose robots um but once you have purpose humanoid robots um and autonomous vehicles uh you really you you you can build anything um and and and this I think that there's no actual limit to the size of the economy I mean there obviously you know the mass of Earth you know like that be one limit um but the you know the the economy is is really just the average productivity per person times number of people that's the economy and

if you if you've got humanoid robots that can do you know where there's no real limit on the number of humanoid robots um and and they they can operate very intelligently then then there's no actual limit to the economy and it there's no meaningful limit to the economy you guys just turned on Colossus which is like the largest private compute cluster I guess of gpus anywhere is that it's it's the it's the most powerful supercomputer of any kind um which sort of speaks to what David said and kind of what Peter said which is a

lot of the kind of economic value so far of ai ai has entirely gone to Nvidia but there are people with Alternatives and you're actually one with an alternative now you have a very specific case because Dojo is really about images and large images huge video um yeah I mean the Tesla problem is different from the um you know the sort of llm problem uh the nature of the intelligence actually is actually and and the what matters in the AI is is different um to to the point you just made which is that in the

in T's case the context uh length is very long so we've got gigabytes of context context Windows yeah yeah you got you know sort of uh we just bringing it up kind of billions of tokens of context nighty amount of context because you've got um seven seven cameras and if if you've got several you know let's say got a minute of several high high def cameras then that's gigabytes so you need to compress so the Tesla problem is you got to compress a gigantic context um into the the pixels that are that actually matter um

and you know and and and condense that over a time and so you've got in both uh the time Dimension the space Dimension you've got to compress the pixels u in space and the pixels over in time um and and and then and then have that inference done on a tiny computer relatively speaking a small like you know a few hundred watt uh it's a Tesla designed AI inference computer uh which is probably still the best there isn't a better thing we could buy from suppliers so the Tesla designed AI inference computer that's in the

cars is better than anything we could buy from any supplier just by the way that's kind of a by the way the Tesla ai ai CH team is extremely good you guys in the design there was a technical paper and there was a deck that somebody on your team from Tesla published and it was stunning to me you designed your own transport control like layer over ethernet you like ah ethernet's not good enough for us you have this TT Coe or something and you're like oh we're just going to reinvent ethernet and like string these

chips it's pretty incredible stuff that's happening over there yeah um no the team the the Tesla chip design team is extremely extremely good um so um but is there a world where for example other people over time that need you know some sort of like video use case or image use case theoretically you know you'd say oh why not you know I have some extra Cycles over here so it which should kind of make you a competitor of Nvidia it's not intentionally per se but um yeah I mean the you know this training and inference

and we we do have the you know two those two projects at Tes we've got Dojo which is the the training computer uh and then um you know our inference chip which is in every every car inference computer um so and Dojo we've only had Dojo one Dojo 2 is um you know should be we should have Dojo 2 in volume towards the end of next year um and and that that that will be we think sort of comparable to uh the sort of a b200 typ type system a training system um and and um

you know so there's I guess there's some potential for for that to be used as a service um and like Dojo is is just kind of like I mean we're I guess I guess I have like some improved confidence in Dojo um but I think we won't really know how good Dojo is until probably version three like usually takes three major it itations on a technology for it to be to be excellent um and we'll only have the second major iteration next year um the third iteration I don't know maybe late you know 26 or

something like that how's the uh how's the Optimus project going I remember when we talked last um and you said this publicly that it's in doing some light testing inside the factory um so it's actually being useful what's the build of materials and when you know for something like that scale so when you start making it like you're making the model three now and there's a million of them coming off the factory line what would the they cost 20 30 $40,000 you think yeah I mean what I mean I've discovered really that you know anything

made in sufficient volume will ASM totically approach the cost of its of its uh materials so now there's there's I should say the there's some some things are constrained by the cost of intellectual property and like paying for patents and stuff so a lot of you know what what's in a a chip is like paying paying royalties um and depreciation of the chip faab so but the actual marginal cost of the chips is very low um so so so Optimus it obviously is humanoid robot it it is it weighs much less and it's much smaller

than a car um so the you could expect that in high volume uh and and I'd say you also probably need three three production versions of Optimus so you need to refine the design three at least three major times and and then you need to scale production to sort of the million unit plus per year level and I think at that point the cost the the you know the labor and materials on Optimus is probably not much more than $10,000 yeah and that's a decade long journey maybe basically think of it like Optimus will cost

less than um a a small car right so at at scale volume with three major iterations of technology and and so if a small car you know costs $225,000 you know it's it's it's probably like a I don't know $20,000 for for an Optimus for a humanoid robot that can be your your body like a combination of R2D2 and c3p a bit better I mean you know that's that's honestly I think people are going to get really attached to their humanoid robot because I mean like you look at sort of you watch Star Wars and

it's like R2D2 and C3 I love those guys um you know they're awesome um and their personality and and I mean and all all R2 could do is just beef at you couldn't couldn't speak English um see3 to translate the beeps you know so you're in year two of that if you did two or three years per iteration or something it's a decade long journey for this to hit some sort of scale and I I would say ma major iterations are less than two years so okay um it's probably on the order of five five

years yeah uh maybe six to get to a million units a year and at that price point everybody can afford one on planet Earth I mean it's going to be that one: one two to one what do you think ultimately if we're sitting here in 30 years the number of robots on the planet versus Humans yeah I think the number of robots will vastly exceed the number of humans vast I mean you have to say like who who would not want their robot buddy everyone wants a robot buddy um you know this is like it

especially if it can you know you know it can take care of your your take your dog for a walk it could you know mo mow the lawn it could watch your kids uh it could you know like it could it could teach your kids it could it could we could also send it to Mars absolutely a lot of robots to Mars to do the work needed to yeah make it a colonized planet for him and Mars is already the robot Planet there like a whole bunch of you know robots like Rovers and helicopter yes

only robots um so yeah the no I I think the the sort of useful humanoid robot opportunity is the single biggest opportunity ever um because if you assume like that I mean the I think the ratio of humanoid robots to humans is going to be at least 2 to one maybe 3 to one cuz everybody every everybody will want one and then there'll be a bunch of robots that you don't see that are making goods and services and you think it's a general one generalized robot that then learns how to do different tasks or yeah

hey um I mean we are a generalized yeah we're a generalized we're just made of meat ex uh we're a meat PB a generalized Meb yeah I mean operating my meat puppet you know so um yeah we are actually and by the way it turns out like as we're designing Optimus we're sort of learn more and more about why humans are shaped the way they're shaped and you know and why we have five fingers and why your little finger is smaller than you know your index finger um you know you know obviously why you have

opposable thumbs but also why for example your the muscles the major muscles that operate your hand are actually in your forearm and and your fingers are primarily operated like um your the muscles that actuate your fingers um are located the vast majority of the of of your finger strength is actually coming from your forearm um and your fingers are being operated by tendons little strings that that's and so the current version of the Optimus hand uh has the actuators in the hand and has only 11° of Freedom so it can't it's not as doesn't have

all the degrees of freedom of human hand which has depending on how you count it roughly 25 degrees of freedom um and uh and and and and it's also like not strong enough in certain ways because the actuators have to fit in the hand um so the Next Generation Optimus hand uh which we have in Prototype form uh the the actuators have moved to the forearm just like a human and they operate the the fingers through through cables just like a human hand and uh and then the next Generation had has 22 degrees of freedom

um which we think is enough to do almost anything that a human can do