If The Fed has sufficiently bent the curve on inflation then why take the big risk of breaking something? Major things could have gone catastrophically wrong for Humanity These Technologies are going to compound over the course of the decade at a 40% rate. Wouldn't take a big team to make it much much better. So then you have to ask, what are they waiting for? Honestly, we're at the beginning stages of multiple escorts that are compounding on top of each other. A lot of good faith people, lost money permission that is granted by the SEC blessing.

This product is going to create all kinds of credible arguments that like investment Banks can make Other Regulators around the world will look at this and say oh okay, this is permissible Now it does seem like companies are almost scrambling to figure out how to apply this. They had counterparty risk that they couldn't really assess because the government was like preventing kind of like legitimate counterparties to from coming to the table is Elon making a big mistake getting so political on X should he just shut up? Welcome back to another episode of the Meet Kevin show Today we have the honor and privilege of being here with Brett Whitten It's been 11 months since we last chatted. Thank you for inviting us back! Oh my pleasure, Thanks for coming! Did they discover AGI It's a funny question because no, uh and the definition of AGI I think is going to always be shifted to keep it from being discovered for some foreseeable future. I Think that um, you know, because of the board resolution, not because of the board resolution no, so not specific to open Ai and that's maybe a separate conversation instead.

I Think that people humans have like this idea that like AGI is a thing that we're going to achieve right and um, but any technology that gets sufficiently Advanced kind of becomes seemingly mundane to us and so then we set a new goal post for something that needs to be achieved. Uh, and I think we're going to be in a state of um, kind of like AI Systems like exceeding benchmarks that we never thought were possible to be exceeded. And then we as humans continuously defining different benchmarks that uh, they haven't yet achieved and saying well, that's something that is actual AGI or artificial general intelligence. Where do you see the biggest benefits? Is it uh, in in medicine or mostly I mean I know we've all got GPT now which is crazy because when we last met, that was just starting right? Uh, but I feel like at least my personal usage has sort of bell curved where it was like up and now it's kind of tapering off a little bit.

What's your experience been and what do you see for the future for AI I Think my own experience is like it's a great kind of search, then information retrieval Tool uh and and then also a kind of like certain written output basically like def frictioning uh for me. So it's like um I use it for um for instance, a GPT add-on to basically just be able to type extremely quickly and then it um kind of maintains my sentence structure but then like compresses it like not just fixing misspellings but also adding appropriate transitions between sentences and so so it it like really tracks kind of like the idea flow that I'm doing. and just like Uh leaves allows me to remain in flow in first draft mode rather than kind of like word picking mode kind of I like that what is that called uh, that's flow speed typist so it's a GPT a GPT with an open framework um I think that generally right now Enterprises are experimenting with hey, how can we get a return on investing in this technology and and they're trying all kinds of things it's obvious that some of kind of like is going to work and some is not um and then over the medium time frame I Think the right way to think about this is that um, this is really knowledge worker productivity enhancement that's going to occur. Um, so um, you know you could think of it as augmented intelligence, but it's really like how much can I do as a knowledge worker and how much quicker will I be able to do that? So like ways in which we've used it internally is it's like we have a database of 30,000 different rare diseases and um to previously like, say um, what Like how severe is this disease and what is the human health cost of this disease? Uh, and what's the addressable market for this disease? We might have to be like, Okay, we're going to have an intern like go through all of these things and try to figure out the health impact and how severe it is.

Now you can kind of, um, throw kind of the descriptions of those diseases at Uh at GPT Uh combined with some actually hand-coded data that you've done uh to then actually produce an estimated kind of severity and um, basically life impact that those diseases have. From that, you can, um, estimate. You know what is the human health impact of all rare diseases and then what's the market value in solving it. And so it's like an example of like something where it was possible to do if I wanted to go to the trouble of like trying to find like, you know, kind of interns to do it.

but then their first cut at it might be wrong and it would take a while and then you have to like. You just get into a faster iteration cycle of how do I take data and convert it into information. Uh, and so like at a aggregate level, we think that Um AI software will improve knowledge worker productivity by roughly 10x a little less by by 2030. and that like 50% of the knowledge working population will be using these tools actively to.

well, they still have jobs. Well of course because it's like you know, I think that there is a set of circumstances in where these systems just Spin and be able to operate themselves but for a very long time it's really going to be like augmented intelligence type products where it's um, you know there's a lot of um, intuition and human pattern recognition that even is guiding like what is the right spot to to guide this software where we might generate something of interest and so um, you know and you can imagine like how would that manifest. It's kind of like um, from the consumer side it might be like I get a much better understanding of of how good a product is for my Spe specific use case before I commit to buying it. So essentially it's kind of like the the interfaces between people trying to sell services and people receiving services are going to get much less full of friction.

Often like my wife will get something from Etsy and she'll receive and be like oh I didn't know it was this size as opposed to this size right? you know? So that's like a simple example of like where there's really like a a lack of information in the marketplace and that'll get cleaned up a lot and then kind of. All the entertainment products will get that much more compelling because of it. so we'll just get sucked into digital experiences more completely. But net If you go back to that expectation of productivity enhancement to Knowledge workers.

um, usually businesses pay uh, roughly 10% of the Um value they get out of a software for the software itself. W So if you do the calculation, there's roughly $30 trillion in knowledge worker wages by 2030. Outside of China If I improve their productivity by half of their productivity by almost 10x, you get to a number that you know is like 130 trillion in Um. Kind of like value that's generated to the economy and basically better knowledge work.

And then businesses will pay 10% of that. So $ 13 trillion in AI software Revenue by 2030, Do some more math and you can conclude well that suggests there's you know, 8090 trillion do worth of AI software market cap. Wow! by 2030, there's I Mean that's half of the entire US GDP more than half relative to today. Sure, just in artificial intelligence.

Uh, potential productivity generated Revenue Yes, yeah, exactly. But think, think about it on a global basis. if I'm paying Knowledge Workers $30 trillion and um, you know, in a world without AI, you know their output would be X Yeah, so I'm getting essentially almost 5x that output. Um, surely like I' it' be worth it to me.

To you know, increase their wages by 50% But instead of throwing it at Knowledge Workers I'm paying for kind of AI systems to basically enhance and you know, allow them to generate better forecasts. Allow us to think about how much Capital time and energy is wasted on technologies that simply aren't going to work right right? And so if if you have like a better way to tune and understand, take drug development like early in a drug development process that hey, kind of like this biological system is not going to like respond in the way I think it's going to If I plug in this molecule then you can forgo all of the expense in time of human trials. That would lead you to that conclusion: Hundreds of millions of dollars and savings GDP Then uh, you get to a product quicker. so GDP can grow faster essentially with output.

So yeah, that might represent a large number seemingly today, but your suggestion is output could explode through this, right? so at a at a very high level. And actually, this is like an important macroeconomic point if you look over the course of history. Um, it's not. Um, it's not a historical for us to cross a threshold when suddenly GDP growth changes in a structural way up.

as in um, you know, in like the the post Kind of like the turn of the 20th century. Um, we entered a new Global GDP growth regime because of the introduction of electrification and the telephone and the internal combustion engine and all The innovations that have happened like prior to that. kind of like GDP growth was maybe half as much and so like when you change the rate of growth. um, you actually end up with much, much bigger numbers on the back end.

So if if you look at the long run of History, we actually think over the course of this decade compounded annual real GDP growth will average um 7% from here forward. Wow, so it's It's actually a step change up and and to be clear, that's Us or global global. Okay, okay so like consensus if you look at like consensus forecasting agencies, the idea is that um, by 2030, the GDP Global will be around $130 trillion. This is 2021 and we think it'll be7 trillion.

So you go from uh expectations of $155,000 GDP per capita to $20,000 GDP per capita. And so that's on the basis of like, hey, if there's a like Given the rate at which these structural changes of GDP growth have happened historically, you could roughly conclude that. and then it's supported by the fact that if we look at the underlying technologies that we focus on, you can justify that forecast just based on our expectations for Robotics and Robo Taxis even leave out the AI Sof software knowledge work side. That's so interesting just because so many have, uh, argued.

oh, artificial intelligence is just a fat. It's just it's going to be here today, gone tomorrow. sort of thing. Uh, and your argument is what we've just done with AI is almost changed.

the rate of growth of our global economy. Uh, almost to a similar fashion as what electricity did for our Global economies do I understand that right? Yeah, I Think that and and actually like the way we think about the world is that there There are these major like um, signpost Technologies like major technological innovations. So the steam engine electrification uh and uh, and the internal combustion engine and that like at the the most kind of we've had. Enter the E Economic Marketplace at the same time is at the turn of the 20th century when there were three and right now we think there are five between energy storage, public blockchain AI Robotics and what we think of as multiomic sequencing.

all kind of hitting these credit stages of inflection at the same time and so then it's It's like consistent with history that that would Trigger or catalyze just a change in growth. I Mean if you have a you know, leave aside the knowledge work side which is a little intangible and I think a lot of that is going to acud to people's benefit without getting recognized into GDP statistics If we have a humanoid robot that you can buy that can basically make things to the same like quality or even lower quality but just working 24 hours a day as uh, human manufacturing worker um then like the amount of created stuff can explode you know, uh and like you all drove down here down the PCH like what a waste of your wonderful brain to have to be piloting a vehicle for that long. you know, in traffic like you'd rather be admiring the view or kind of like I don't know doing one of your crazy real estate deals something like that. Well like the a robo taxi will free you up to do that while costing basically the same per mile that you're already paying to drive yourself and so kind of like that in GDP statistics that'll appear in the amount you're paying for that service.

Uh, plus kind of like the other stuff that you're doing when you would otherwise be driving. so you're taking kind of non-paid wages which is you driving right now and turning it into a market service and plus freeing up your time to do some other either watch Netflix so you're paying somebody else or doing work that generates kind of like income for your consuming or generating. That's really interesting. Now you uh, have been tweeting about two seat uh Robo taxis and looking at how little usage there is in taxis.

uh above. basically two people in a taxi. Uh, and you're really into cars I mean are people going to be into two seat Robo Taxi? Does it matter how many seats there are I mean I Think from a the the general idea and this is uh, something that Sam Corus, our batteries analyst and I have been discussing a lot. Is that probably Optimum strategy for Tesla uh for uh, Robo taxi is a twodo vehicle.

might be four seats, but it's really like a two- seat vehicle. Um, because like if if you're if you can deliver Robo taxi service to your vehicles, you want to minimize your manufacturing cost and maximize your unit volume so you can basically maximize ride liquidity like maximize the number of places markets that can service and maximize your data intake rate. So it's not. You know it's not true that like a four-door vehicle, costs twice as much as a two-door vehicle, but it definitely costs more than a two-door vehicle.

All right. So um, it seems likely to me that the platform they develop for robo taxi is a two-door car. The the issue is if um, somebody doesn't believe that Robo taxi is possible and they find out Tesla's nextg $25,000 car is a two-door car. A traditional analysts will conclude well, there's no market for two-door cars like there's no Mass Mark You know it's not like the Honda Civic it's TW doors that's a four-door vehicle and they'll conclude that uh, you can't like this will actually kill the sales demand for the vehicle If you don't believe in Robo Taxi now one I Think that every Market that Tesla's entered into has been a market that didn't actually really previously exist as in when they launched the model X even like SUVs at that price point were not a thing right? and and actually they move people up Market into it.

There are kind of uh In In in expanded the category meaningfully. like there is like a neighborhood electric vehicle category that's lying nent. There's like smart cars that you know are fun except if you're driving around with a bunch of SUVs you might get like pancaked by somebody on the road. um and I suspect that kind of like even ex Robo Taxi Tesla launching a $25,000 vehicle if it has two doors will actually sell like quite well.

Um, but you know our expectation is Robotaxi Will commercialize basically this year or next year and then that that commercialization event will you know, transform the economics of Tesla and then it will. You know, um, mean that that kind of mass Market type vehicle is probably not even being sold into individuals in a material way. it's being acquired by operators who are trying to operate it as part of a robot taxi. Network And then with a two, two, even a two- seat vehicle, you can make you can meet more than 90% of the demand for kind of Robo Taxi and Ride haill.

Um, what's your concern with uh, potentially Herz complaining and now selling some of their used Tesla Fleet on the belief that or the realization that maintenance for these cars and the taxi Fleet or Uber Network or whatever rental network uh has been too expensive relative to Ice Vehicles Because they've had challenges finding Parts not as readily available, there are not as many service techs available to repair Cosmetics inside. Uh, these these Ubers or whatever to where uh Herz is now returning back to Ice Vehicle for their Fleet I Don't know if it's accurate to say they're returning back to Ice vehicles for their Fleet I Do think like if you the the price volatility in electric vehicles generally relative to internal combustion has been effectively higher in part because also, there's not the kind of like, um, call it buffer of selling into Uh dealers that then hold the inventory on lot and sell it at a discount without nationally discounting price. Uh, and so Herz in a position where they were acquiring a bunch of Robo taxis in 20 or a bunch of Teslas in 20121 Uh, and then being forced to like recognize well the value of these is dropped just because like The Upfront price has dropped, then feeds into their kind of financials in a way that they weren't expecting. So I I don't you know I I Think that there's a Um, there's some kind of like hey, you know they their financial statements have to recognize the fact that these vehicles have depreciated much more quickly than they thought they were going to.

Because if You look at kind of like the historical depreciation rate of EVS, it's actually they depreciate less quickly than internal combustion. and so, um, it's kind of like a the mapping to their financial reporting is definitely not what they expected and so then they have to have kind of like an explanation for that for Wall Street a PR move is what you're saying. Well, it's not PR I mean it's kind of like, listen, you know if if you bought or I bought like an electric vehicle in 2021 uh, relative to what I could get it at today I overpaid. Oh yeah, but the fact is like I it doesn't change like my cost per mile of driving that car.

It's kind of like when you buy a car you're pre-commit to basically a number of a bundle of miles at a specific cost per mile. And you know when you make the purchase decision, that's what you're committing to. you're not, you know, or I am not doing it saying. Oh, and then three years from now I'm going to sell the vehicle for X right, you know? and I'm and it's actually like on a cash basis I'm not at all sensitive to that.

Um, so whereas someone who's if you were strictly reporting that through like a financial statement that you're having to present to I don't know Wall Street or kind of like to your family, you know you'd be like ah, I made a bad move you know? um, but practically like I Think that the um even Tesla sold at kind of the peak purchase price. Those vehicles are continuing to improve year-over-year because they're delivering software updates to those vehicles that make them more useful and that's very different from kind of traditional vehicles. and you D have a Tesla right? Yeah, Uh, is that your black? X Yes, Nice, Yes. um so FSD how have you seen it evolve over the last year? Oh, it's amazing.

Like or for me it's a massive stress reducer and kind of like keeps me more alert even driving around. La Uh so also safety enhancer I Have kids, kids are in the back. kids need a snack if you're driving like trying to like, you know, yeah, give a snack back to the kids is like definitely net unsafe if you don't have some kind of like driving augmentation and it performs amazingly well in that and it's improved. it's gotten smoother it now crosses kind of like what I think of as the the Wife test where she's in the passenger seat we have it on.

she's not like being like you have to turn that off. this is making me motion sick. um and there's still a lot of work to do like I think that the um you know I use I basically use it at all times while driving and then there are intersections I come to where I'm like I don't bother to try to leave it on through the intersection even though it could probably make it through the intersection on its own. It's more like I'm making a left turn and a two-lane left and it's just like I I trust myself more than the system, you know? Uh and I do think that so sometime this month likely to in customers, they're going to release version 12, which is um, uh, full stack nurl Nets from top to bottom.

Uh. By doing that, they're eliminating 300,000 lines of of code out of the system and so it's like, well, what are they replacing that with They're replacing it with 3,000 lines of code and lots of data compressed by lots of compute. Uh. and so one way to think about the Improvement rate of the full self-driving system is prev viously like they would run into unique problems and then they would have to come up with a unique solution to each unique problem.

It's like how do I deal with a roundabout, you know? And so you need a software engineer to think about. Okay, what's the car going to go through as it goes through a roundabout? What are the if then statements about? Like when it can turn, How does it determine when to turn? Uh and and uh. And so every single Corner case required a clever software engineer to figure something out and kind of each incremental kind of reduction in the number of interventions you need unveils probably an order of magnitude more Corner cases. So you're always running up against this wall of like either I'm hiring 10x more people for each kind of like step up I make in kind of like the log error rate or um I need another solution.

So once you go full stack neuronet My intuition is that then every unique problem has the same solution. You throw more data and compute at all of the problems. Uh, And so it means that not only should there rate Improvement now just be governed by how much compute they can throw at the problem. Because they're not data constrained, it also means that their ability to forecast the Improvement rate should improve.

So, uh, Elon Musk has famously said basically, we're a year away from commercialization for maybe four or five years, right? Whereas kind of in our forecast we've been, we've slipped slightly. but it was 2023 for a while and now it's 2024. Like, as in, we really thought this was around the time frame where it would happen now I Think they'll have much better line of sight on this is the Improvement rate. This is how much additional Improvement rate we get to throwing more computational power at it.

This is how much cash we have to throw at computation. Uh, and so they can. Um, kind of with much more. Precision Say this is the thresh the place in which we'll cross a threshold in the specific geography say California where we can commercialize the product and so um, there will be.

They'll be ble to both business plan and message the street a lot more. Um, precisely on. like when Robo taxi is going to commercialize and what that will look like Uh, and then so from the analyst perspective, not us, but ones that, um, essentially rely on management to tell them what's going to happen and then map that to how they think financials are going to occur. Um, there will be like, um, an ability to begin to underwrite that into the stock.

Uh, and and and that'll you know it. It totally transforms the business model. If and when that occurs, there's no, um, you know there's not a comparable transition probably in any corporate entity's history. Um, to what that could do to cash flow generation for Tesla Do you expect early Robo taxi to have any level of Geo fencing? Well, yeah in that I mean not at the block by Block Level I would.

But yes, at the kind of like Geographic entity by Geographic entity level. as in, like you know, maybe it's just City of La or maybe it's state of California Or maybe you know they're going to have to get regulatory approval at least at the state level and maybe at the city level in order to launch. And so it's like Austin Texas probably lets it happen. Uh, and then there will be areas where you know it's as performant, but you're not yet allowed to commercialize it.

And then there will be areas where it's just not as performant because they don't have as much training data. Uh, and I think it's an open question. Um, like not just for Tesla But for all of the players, like what is the yeah rate expansion like? to what degree should they be concentrating on? you know, specific streets like Wh? Mo's Um strategy has been we're going to keep the thing off the highway, you know. And so we're g to and it means that for longer trips it's not.

You know you're having to sacrifice like double the trip length um, in order to ride in a robo taxi that's priced kind of the same as a traditional Uber and so then it doesn't actually make sense as a product that scales Um. and so I think that Tesla's approach is more aggressive, but ultimately more scalable. Where it's like, at least in the geographies we've trained on, this thing should be able to drive in any conditions. and we're not going to Blacklist any intersections or type of Um driving in part because I think a problem with a a more precisely Geo fenced system is then um, an intersection that you have whitelisted suddenly becomes an intersection it can't clear through.

Like it it like you know, the intersections under construction or like the structure of the intersection changes and so suddenly it acts more like an uncontrolled intersection rather than one with a traffic light. And so then you end up with like a really bulky and kind of like subject to lots of kind of hand tuning and changing uh system. And and it's likely both. bad consumer experience and then you have a huge kind of operational expense of just like figuring out like what the map looks like for you at any given time or any given day.

Well that was the whole I think uh, Crw problem was they had 1.5 workers per car driving on the road essentially for that operational control. I Think this is interesting because what what you've just described is essentially uh, you can underwrite. uh once Once Tesla starts providing that guidance you, you're you'll see the market more likely underwrite the value of Robo taxi from what's probably underwritten now of zero to whatever it might be. uh, more maybe so than full self-driving Is that your argument that people that there'll be more value really out of the robo taxi than trying to convince people to t you know, pick up full self driving.

The take rate seems like pluming. Yeah, I think that's A or it's not a sideshow, like it's material to their gross margins. but I don't think it actually. um yes as in like I have full self-driving It's an incredible luxury product for me.

as in, it's expensive and it delivers me a lot of value. In terms of my Driving, Experience is much better. and I can understand somebody buying a Tesla and being like $122,000 for what? even though the comparable like even if you're buying a BMW or whatever like comparable like just you know Lane following on Highway and Active Cruise Control is like sometimes like A6 $7,000 option. So actually Tesla's living delivering.

You know I would argue substantially more value per dollar. and it's something that people I can understand how people aren't opting into it because it's like I don't have $122,000 extra to spend and for what do you know what I mean? Well, a lot of people also have that trust Factor too. I Notic that, especially in uh, older Generations Frankly, uh, 40 Plus there a lot of people I just won't trust it. Have the Tesla but I just don't trust it, right? right? And so it really has to for people to kind of like buy that option.

They also have to have experienced it in some way. They have to know it's worthwhile, but all of that kind of disappears once you deliver Robo Tax or or or that Nuance doesn't matter anymore. Think about it this way: This is very crude math, so consider it heavy approximations, but roughly right. Okay, so um I sell you a model 3.

You should be a model 3 buyer. Uh, I sell it to you for $50,000 What is Tesla net an operating revenue for that sale? Maybe $5,000 Call it a 10% operating margin. It's not exactly precise, but roughly okay. so it's a one-time $5,000 operating.

Uh, profit for Tesla What if that model 3 becomes a robo taxi? Okay, so we think the model 3 could roughly do 100,000 miles per year. As a robo taxi, that might be an overestimate, but it's It's like that there's ways in which you can model it where you think that's kind of the utilization it could get to. Okay, and uh, at a dollar per mile, We think there's uh, a trillion miles addressable globally there. So let's say we're that the first trillion miles of Robo taxi driving is occurring.

so you can charge a dollar per mile for 100,000 miles. Okay, so then, uh, there's gross revenue somebody is paying. you know, $100,000 a year to ride in that vehicle. Now Tesla gets to extract a platform fee off of that Just like Uber Oryt do.

They should be able to do something more than Uber or lift because they're delivering this whole autonomous Driving Experience Say it's 50% Okay, so then that's $50,000 in Revenue to Tesla. If they're getting like software type margins on that, that's $25,000 in operating profit to Tesla. Okay, so I've gone from one time $5,000 in operating profit to annual for every vehicle that's capable and turned on $25,000 in operating profit per year into perpetuity. Okay, so from Tesla My gosh, like I've I've quintupled my operating profit per vehicle and turned it into an annual thing as opposed to the one time I sell it based on this very expensive Factor I've built to produce these things so the financials explode.

Cash flow. Okay, now think about it from your perspective though, and this is actually a really interesting and nuanced point is like you paid $50,000 for the car. Now it's generating $50,000 in net revenue To you. if Tesla's only charging a 50% platform margin and your cost to run that vehicle is.

You know you can estimate it in various ways, but it's on the order of 15 cents a mile or 20 cents a mile maybe. Okay, so that's $30,000 in profit to you who bought the the Tesla per year. So clearly you underpaid for that asset. At $50,000 buying an asset for $50,000 that generates $30,000 in income, it's worth much more than that.

So then like, what practically will happen is Tesla will actually collect more than 50% platform fee at that time cuz you shouldn't be getting like all of that, right? Or the cost per mile collapses as everybody's showing, right right? So so then there's the demand side. Like if there's a gazillion M If there's more than a trillion miles then being driven, then it collapses down to. ultimately we think around 50 cents is the equilibrium price going out. Um, but like between here and there Um, kind of like you can think of like buying Robo taxi now is you're buying the experience of having or buying FSD Now buying the experience of having the car help you drive for a while plus buying a modern day taxi Medallion Where it kind of like gives you kind of the opportunity to kind of extract value out of this vehicle or sell this vehicle who's going to maximize it as a robo taxi in the future? Um, and so like the I think the they'll have a lot of kind of like levers they can pull in terms of like how much did they charge for FSD versus what is their take rate in terms of platform fee? Um, you know, but ultimately it's I Think from a financials perspective, it's all going to flow through into an effective take rate against miles for them.

Yeah, that makes a lot of sense. The uh. So another thing that was fascinating that you mentioned is this idea that you can create a robot that can work even less like quality uh or or create a less quality output than a human but work 24 hours a day. So really there's this idea of even if it's just let's say unboxing boxes out of a freight truck or whatever.

Uh, neither of these these these robots that could work 24 hours a day or robot taxis neither are really underwritten into Tesla's valuation today. So where do you see Tesla going in the next decade even if they just started producing and selling the first robots to let's say Pepsi or whatever. like when does that that sort of trigger hit is it at the first sales and what does it do? I I think it's hard are one interesting thing that's happened to me as somebody that focuses on technology is I feel like my line of sight is actually getting compressed forward as in I think it's hard even going out a decade from 2024. I Think it's much harder than it was going out a decade from 2020.

Uh, because of the rate pace of change of artificial intelligence I will say that like relative to at the technology level. um, kind of. Our expectations for Robotics are the ones that have shifted most this year compared to last year. Um, because um, I or call it.

We had concern that there was like mechanical and actuator constraints that were going to prevent kind of robots from being super useful and so even if you had great software that you could attach to them, if if like, kind of like the ability to actually mechanically actuate these systems was the constraint, then that's like a hard Hardware problem which usually takes longer. And the evidence that we've seen there's an open source robot for it's like 20 or $30,000 that can you know, flip an egg and do all kinds of stuff in a teleoperated way. Uh, that. Uh, there's just a paper published on the demonstration.

certainly from Tesla's Optimus bot is much more, um, kind of mechanically faasil than I would have expected given they've been working on the project for relatively short. When they film that though, they're wearing the sort of uh, goggles and like it's the robots really just mimicking the person rather than the robot. Well, yeah, that could be or I I don't know for sure. But but that's also fine.

as in if if my concern was the governing constraint was like the being able to move fingers in a fluid way and being able to pick up an egg. and you know, um, if if it's teleoperated, then that indicates, well, that's not actually the constraint, then the constraint is the software side. and so then on the software side, it's like very clear that kind of AI is going to just any Benchmark you can set for it. saturate that Benchmark very quickly.

Uh, and so kind of like a way to think about it is is we think AI costs are declining 3x per year. So what in traditional compute takes two years to occur? Uh where every two years you get a hot cost having in AI happens every six months. Uh, and so one like Universal rule of technology is people overestimate how much is possible in the short term and underestimate how much is possible in the long term. Well if you consider long-term in traditional Computing like eight years in AI That's two as in so things that we think like the thing, the the time frame over which people underestimate performance I think is is getting compressed because of how quickly AI systems are advancing.

So your concern was that is the hardware even going to be able to catch up to how quickly the software is moving So you're very optimistic? Then from what you're seeing yes I mean I I think I think yes From the I it seems like hardware and and the two are interrelated and that you can imagine. you can have a very very precise Hardware system which is then very expensive. um and you can trade down in terms and and but it's very precise. It can go to the millimeter that you need, but that's not how your body works your body like if I'm going to pick up that object over there I Like my hand gets near it and then I look and see how close my hand is.

It's not like I like try to like close my eyes and grab the thing and so having good software allows you to trade into like less precise hardware and still achieve the same event. Um, but generally there was like a um um yeah. can you get actuators small enough and with Loosely precise enough that you can attach software to them and make them work. And now I'm more optimistic on that front.

Think about it this way. So this 20 to $30,000 robot that I talked about um Hardware is open source and they demonstrated that you can tell operate it to like crack an egg and flip an egg and and do all kinds of like everyday tasks. Well, maybe there's a bridge state in which um kind of. the AI soft Ware is good but not perfect and so you have a teleoperator on the back end who's helping the thing like you know.

Think about like the marginal cost of somebody to cook for you if you're paying Los Angeles salaries is a lot higher than if I'm paying somebody in Mexico to teleoperate that robot. Oh wow, uh and so you could. You could have the hardware infiltrate with like a kind of hybrid type model. um and and actually you know, provide return to people.

Um that uh even even if the software is not all the way there, that's fascinating. I mean you could have a an Indian Chef who's cooking you Indian food but they're in India yeah, possibly I mean and and and and then like. the the feedback loop in AI is like the software is going to improve on the basis of how much compute you throw at the problem and then how much data do you have on the problem you're trying to solve. Well having kind of like that hybrid type robot provides the data that then will improve the software so you don't need the Indian Chef anymore, right? And so um, they're training their own replacement Exactly.

but we're all training our own Replacements In some way, I Mean what are you doing with your kids I'm training my own Replacements they're like seven and 10 years old. That's a good point. That's a very good point. During your last interview with Kevin I was behind the camera on that one, but you were talking about the the power of AI and how much it's going to grow and you were expecting a Sevenfold increase in the like aggregate capabilities of AI systems in 2023.

How accurate do you think? were you on that prediction and then what do you expect going forward? I Mean it's hard. it's hard to measure. but I think I'm roughly accurate as in, like the cost declines have happened more quickly than we anticipated in a lot of ways. Um, and so um, okay I can't okay I know when that interview occurred.

Okay, so Gp4, for example, um, they released GPT 4.5 turbo or whatever. right? The that on a cost per token basis was at roughly a a third the the price. so that in itself was a third cost Decline and it was over 240 days not 365 days. And then if you adjust for latency, um as in like the tokens get produced more quickly out of it so same accuracy at a third the cost and uh and it was roughly three or four times faster so performance adjusted.

It was probably a 10x Improvement in terms of that specific AI system. so it was basically under uh on on a very clear like light like for like basis. Um, but there is so much evidence that these systems are just um, they are wildly under optimized right now like we we really are just figuring out how to use them. The fact that you can, you can ask an AI system a question in a slightly different way and get like a 10 or 20% Improvement in its answer rate is just indicative of how kind of undertuned the systems are like compar to like I was Was watching a video the other day of an internal combustion engine being manufactured like a V8 BMW V8 was like massively optimized you know, like so automated and precise.

and like you can tell that every single process that creates it the entire design has been to squeeze all of kind of like the energy input into horsepower output out of that system and the like. last iteration of something is always the best version of that thing because you you've just like you know, tightened every screw and in AI it's like all the screws are wildly loose and people are just trying to tighten one versus another being like Oh. My gosh, this works and this works and you combine it together and it works this much and so like at the architecture level, at the kind of data that we feed into it, level, at the applied to this class of problems level, at the kind of like how do we yeah query the systems level, there are still massive Improvement opportunities of available and so kind of yeah our our forecast is tuned really to the like: how expensive is it to train the raw underlying models and and we think that's declining 3x per year but there is so much on top of that that is improving system performance? Uh, and even like if if you think about the magic of Mors law was in part like if you can press the number of transistors on a chip then kind of like the number of transistors you get for the same amount of raw material doubles and because they're closer together like the cycle time of the chip can improve. So you got like a dual performance boost.

Similarly with AI as you are creating models that are um, the same performance on a smaller number of parameters, then it's less cost to actually infer that model and you generate material more quickly. So you get this double cost decline characteristic of like I can charge 3x less per token and I'm delivering. You know, three or four times more speed. uh and so uh it.

It yields this. You know, that's interesting too, because you're really not to cut you off there. But I mean this. It does seem like companies are almost scrambling to figure out how to apply this.

and uh, it's almost as if the AI in a weird way, is way ahead of us figuring out where to put it yet. Uh, and uh. then I Look at a company like Paler and I feel like they've always always been in AI that it just used to be called Big Data and now it's Ai And it seems to me they feel like one of the few very profitable AI companies whereas everybody else is just trying to figure it out. Maybe include open AI with Paler, what's your take on that or is who else really stands out in the AI uh space.

Yeah, I mean I I Think that there is broadly a set of companies that have, um, both, a product that can like benefit a lot from just deploying AI against it that, um, they will deliver productivity advances to employees that then they're going to be able to charge for in some way. So like, look at Zoom which by the way, right now, is like in the Russell 1000 value index like it has selling for 13 times right now. It's it. The the embedded expectation for that company is that you know it's basically X growth.

That's like, how much cash flow can we scrape out right? Five bucks I Think is the curing current EPs and it's expected to be there for the next like five years. So Zoom's marginal competitors are clearly like the Google suite and the Microsoft Suite right? Like Um, and I think it's There's plenty of compelling evidence that even companies that use those products internally still need Zoom for all external meetings because like, as a salesperson, the worst thing that can happen is like the meeting gets delayed by 10 minutes because people are trying to install software or they can't get logged in and stuff. Yeah, Yeah, yeah, exactly. Uh, And and and Zoom provides a frictionless experience and they are layering on AI capabilities onto their system right now.

They're not upcharging for those in a meaningful way, but Microsoft is if you look at like, at least on a list price basis. Microsoft Thinks they can almost double the price of Office 365 once they layer in AI systems. and if you look at at least what Google's announced, they also think they can double the price of Google office suite layering in AI. So so there there is either Zoom is going to become a much more compelling like you know cost tradeoff versus those products or Zoom also is going to be able to increase prices based on the productivity they deliver for AI So that might be their growth opportunity is is just provide more value, raise the price so to speak and that could be meeting summaries or or whatever.

Sure. Uh, I suppose. Isn't there some risk though, that there's almost this endless competition in these software software Suites that you get. you've got to get price compression at some point rather than price increase.

you know I mean it's funny, like over time, you know, uh, software franchises have proven to be incredibly durable. at least on the Enterprise side. I Like I think that enterprises tend to, um, be very conservative in shifting soft like you know you get a software get look at like Oracle or or you know any of these kind of software packages look at Salesforce People kind of, um, develop on top of the tools, not even officially, but also in in kind of implicit ways within their organization. That makes it very difficult to unseat the tool once it's developed in.

and if anything, I would think that kind of AI systems will increase the stickiness, not decrease the stickiness of a particular software system. um. so then Zoom also has like a potential call option where they're trying to develop a document um type system that will be AI facilitated on top of kind of the zoom Enterprise Suite And you can imagine, like, look at, like even think about the Microsoft versus Google competition in office software over time. like why did Google manage any penetration against Microsoft It's because when they designed docs and sheets and stuff, they designed it with a kind of cloud collaboration first mentality, right? And even today it's like very clear that Google Docs works better for collaboration than Microsoft where I don't know how many times I've had my dock desync and some like massive headache issue on the back end that actually cost me productivity? It doesn't deliver productivity and frustration? Factor Yeah and and so I think it's there.

It's really valid to say that whatever the kind of like office productivity Suite looks like if you were designing from scratch with the AI capabilities that we have today, you would design it in a different way. See the whole front end would look different right now. Excel has like even like toolbar and then another menu bar above that. Like, it's clear that there's lots of UI Cru that's been layered on to that software product over time as kind of compute paradigms have changed and so I think um, you know it's possible that you have another like a a real competitive office type Suite develop AI first that then penetrates the Enterprise and if and as that occurs, it massively expands kind of like the market opportunity for the likes of Zoom.

There's no guarantee that they get there, but it's like a call option embedded in a company that's being treated like okay. let's see how much cash flow we can scrape off this thing. Yeah, like you said, like a value stock right now. Exactly yeah right.

Uh, so what about um, uh um, You know there's so much hope that Elon is going to be able to turn into like an AWS For example, running a software data centers doing the AI compute. Uh, you've got uh, a lot of talk about Amazon Started as a bookstore and now 70% of ebit is from data centers. Uh, people don't even seem to underwrite Amazon anymore for the fact that it delivers of stuff overnight. Can that happen at a company like Tesla Sure.

I mean I I Think the transition from selling vehicles to like generating Revenue off Robo tax SE would look like that and then like I think that there's a valid Um perspective that essentially having all the vehicle data uh, feeding into their data center provides them a better Foundation model for things operating in the world. Tesla's bias will be to vertically integrate and kind of like just develop the robots themselves and not um, essentially sell kind of those models as a service. but it gives them strategic optionality to do it. So I think one way to think about companies is like what is the like optionality they're developing on top of the intangible assets that are acing on their balance sheet which don't appear in the financials.

Uh, and then what is the monetization value of that and they don't? You know you could ask Elon right now and he'll have a statement about the direction he thinks he's going. but uh one I think real value of having a A A A a founder-led franchise that is that is willing to take hard moves is if his prior conceived Notions about what the Strategic direction of the franchise um were uh changes he'll shift the franchise towards that. So if there's like a point at which it's like, well the Op Optimus Hardware is working. We're deploying AI models against it.

but actually the future requires there to be all different kinds of form factors of robots being manufactured and there's all different kinds of people making these things and we're better off. Kind of like selling access to the underlying model embedded within our data center to empower you know a whole variety of applications. that might be the way to maximize value. As an owner of the franchise, Ron Baron has this idea of Tesla inside that uh, you know the the FSD in the cars or maybe the brain of the robot could be the new Intel Inside uh is that roughly align? yeah I mean I think that there's a question of like do they at least in the vehicle level do they license the robot taxi technology out to other manufacturers and um I mean to be honest like yes, that will be Optimum if like other manufacturers are stepping up to the plate in terms of um, electric vehicle production sufficient to to meet the needs.

So um, you know if Tesla is ultimately constrained by the number of units they can get onto the road, then yes, they'll license the technology. Um, but you have to have like and and maybe in a geography like China, It makes a lot more sense where they're probably not going to be able to get full share of the economics of a robotaxi platform anyway, right? Um, just from a kind of like sensitivity to government perspective and kind of like how much um they'll rely upon being allowed to operate that service. Uh, and so their kind of optimal strategy might be to license it, collect some like smaller platform fee, and essentially deal in some of the local Chinese players. Um, we'll see I I I Think that um, if anything the last year has on the traditional autom manufacturer side demonstrated the danger of being a shareholder sensitive management team where it's like um, GM and Ford particularly GM Basically like almost throwing in the towel, but at least announcing that like this generation of EVS is not going to work for us and therefore we're kind of like pulling back and buying back some stock.

and kind of, you know it it. It's almost like announcing to the world they're going to run down their franchises. Uh, and um, because you know we think 70 million EVS are going to be sold by 2027. Uh, it's most of the market.

Uh, and um, so what is going to be the value of an internal Combustion Engine franchise? If you alongside everybody else who hasn't invested aggressively, have factories that are like producing kind of like you know, vehicles that people no longer want like it's unclear. Like to me, that's a recipe for bankruptcy and consolidation. Um, and so I I just don't know that they'll have that many people to license their software into where we be Optimum for Tesla to license versus kind of like just delivering to their own Fleet You know Brett Actually accurately predicted Grock last year when you were talking you had mentioned that Twitter has a very and I think it was still Twitter at the time. Uh, Twitter has a very unique data set and that you thought it was obvious that they would be coming up with their own Llm.

Um but my my question actually is there's a lot of money that's getting thrown at AI right now and I don't know how much of it is going to R&D and how much of it is going towards marketing, but we've seen some really amazing marketing. Uh, and I expect a lot of them aren't going to end up, you know, fulfilling that. So what do you look for when you're researching these companies? as far as what's truly good Ai and what's just really flashy marketing? Oh, I mean it's partly like using the tools and seeing if they're delivering value. Um, and I you know a lot of in talking to management teams like you end up with a real sense for how strategic It Is by talking through like what is your AI strategy and and how do you think about kind of your or um, kind of superpower that you can use to to deliver on that AI strategy and I agree.

Like with any technology cycle, it's suddenly like you have the you know blockchain IC te company Or you know people begin saying AI because they're like oh Wall Street wants to hear Ai and so they like attach AI to anything um, and you know, on the private and Venture side. It's like we've talked to a number of companies where it's clear that they develop technology. You know, prior there was more machine learning, traditional machine learning. Now that um, kind of the large language models have come out now, they start talking about their technology as if, well, this is AI and it's like okay, I mean I Guess that's within a loose definition of the concept.

It is. Um. and it's actually you're you're in a structurally disadvantaged position because you've sunk a lot of money into a system that is basically being surpassed by another set of systems. Uh, and so it's actually um, you know, uh, it it, uh it.

They would be dangerous spots to take from a allocators, but from an investor perspective because they they actually are being disrupted by rather than disrupting. Um, But when you're being like, it's universally true. When companies are being disrupted, they Co-op the terms and say we're already doing that like you know that's that's like always what happens. Uh and um, and you know that doesn't change like the trajectory of disruptive technology.

It's just indictive that uh, companies are getting into scramble mode and realizing this is where the puck is moving and that they're not well positioned for it. There's a small disruptive technology I wanted to ask you about Bitcoin Yeah, we've obviously just gotten uh Bitcoin ETF approvals and uh, you have a comparison uh of comparing the future value of Bitcoin to what happened after the exchange traded products for gold uh, were released. Uh, One of the concerns that I had that I wanted to ask you about that was it appeared the real takeoff and value of those gold ETFs occurred during the recession. Is it possible that might not be the best comparison then using gold to Bitcoin? No.

I mean I think it creates potential. Energy is the way to think about it. You're absolutely so like the gold ETF listed in the US in 2004. In over eight years, like the price of gold went up 4X Um, and um, but it was heavily catalyzed by the financial crisis, right? As in kind of like people, you know, the story of gold became a story people were paying attention to.

Uh in part because they're like, oh, the whole world is you know, going to go bankrupt and we need to own the asset that's going to survive afterwards, right? And um, the the Um gold as a call it a tool in the financial toolbox of investors. um Gra gained credibility because of that listing, right? As in like if you look after that, you know all the strategies and stuff. Began to put out papers on portfolio allocation and precious metals and how to think about the characteristic price movement of gold relative to other you know things, equities and bonds and everything else. This is a different asset and you should think about it as a, you know, low singled digit allocation in an efficient portfolio allocation.

Um, and so um, the I think similar. So and then there's an important difference between gold and Bitcoin which is the price of gold went for up Forex over over eight years. Uh, and that caused a lot more people to go out and find gold in the ground and dig it out. So like over the decade 20 2004 to 2014 like annual gold production went up 25% So regardless of what the Bitcoin price does, uh, you know, eight years from now, uh, it'll be basically like a quarter uh, annual Supply what it is today, roughly? Uh, and so um, you can you know there's reason to believe like so I think it's valid to say that that kind of the listing of the Bitcoin ETFs uh you know is a a A threshold and Tipping Point that is really momentous for Bitcoin as an asset class because now all of the traditional Financial ecosystem at least in the US is allowed to like, look at it and say hey, this is something I should consider in portfolio allocation.

So maybe that 6040 Bond uh, stock Bond portfolio turns into a 65 35 Something of that sense. Whatever the allocation is, the point is it can be allocated. Now the follow up question to that is theoretically it could have been allocated already versus with gold. you know you got a before the ETS you had to buy the gold, you had to store it somewhere whereas with Bitcoin we were still able to digitally buy it beforehand.

Yeah, you and I can like retail normal like people who are like hey, I'm interested in Bitcoin I'm going to open up a coinbase account we could buy it. A financial adviser would in many cases like no chance one think about it like if I'm a financial adviser I'm advising a client I basically have like access to their brokerage account right? So if they can't buy the asset through their brokerage account I might think it's the right thing to do. Um, the person I work for might say no, you're not allowed to advise that but the reason they would is they would say to protect client. But it's really because if I tell my client to hey, take some dollars out of this brokerage account, go open a Coinbase account and get some B coin.

Yeah there goes the AUM exactly that cost me money and by the way, hey people respond to incentives I'm it's it takes a lot of belief in Bitcoin to tell my to basically uh, hope that my client will do well by it and then appreciate me more even though I'm costing myself money by doing it whereas now the financial adviser can be like hey, this is a useful thing in your portfolio allocation you should, you know buy this spot Bitcoin ETF because it's low fee and it's an efficient exposure and it fits right in. Like if you're borrowing against your portfolio in any way, it fits right in it. If you're you know I still get credit for allocating you to it. You know it fits right in and so it's it's It's actually an important Bridge into the traditional all of the processes that happen in traditional financial services that previously Bitcoin was excluded from.

like imagine you're a financial adviser and you tell your client hey, you should buy Bitcoin It's a good thing to do and you send them to FTX as opposed to coinbase. Yeah, well then you might be subject to a loss like you know. So then you as a financial adviser do you really want to like try to underwrite Coinbase to help your you know client like custody Bitcoin Somewhere? No, you just want to be able to like, You know I Know that whoever Fidelity or Schwab is is an institution that's going to exist I know that there are rules in place even if Schwab goes bankrupt, that those assets will still be my clients. And so therefore I can trust anything that's on that platform as something that I can invest in.

That's fair. And you get Cpic at that point, Sipc insurance. so that's interesting then. So let's say we get a lot more institutional institutionally allocated Capital to Bitcoin.

That should be great for the underlying Bitcoin. Does it hurt Coinbase though? To see retail, which is the most profitable trading sector for Coinbase, somewhere around 73% of their revenue is uh uh, from Trading. about a third of that is from Bitcoin uh, and their most profitable segment being the retail consumer. If they shift from Coinbase Trading to an ETF.

now because it has Cpic coverage, it's easier. It's part of the brokerage account. Great is that going to devastate Coinbase's revenues? Well, one Coinbase is custodian on most of the Bitcoin spot ETF product. so they're generating Revenue on the money that goes into that albe it, it's different.

You know, a different um, kind of like take rate relative to the retail trading dollars. Two, the people who are in Coinbase are typically not people who are just wanting to own Bitcoin they are owning Bitcoin they're staking Ethereum. They are kind of investigating the entire kind of Smart Contracting protocol space and and so like Net. It's beneficial for Coinbase if the entire you know set of assets goes up in value and this, if anything, probably opens up and solidifies the foundation on which Smart Contracting Protocols are going to be built.

you know there's going to be I Think you know who knows what will happen with Ethereum in terms of any SEC listing, but certainly there's going to be any appreciation and Bitcoin will likely spill over into the other crypto assets and having like, a larger, uh, more diverse holder base of individuals who will frankly raise hell if Congress tries to do anything to these assets is net better for kind of like Coinbase's both political stance and the future prospects of like the call options embedded within their business model. Um, in in some ways I Think it's like there's a VIN diagram between the people who own on Coinbase and the people who are going to own within traditional Financial ecosystems and there's actually not as much overlap as you'd expect one. I Think interesting, call it byproduct. that is this in terms of financial: Innovations It's like typically, like big people get it first and then little people get it later, right? And a lot of the Innovations are.

Let's take big person pricing and convert it into little person pricing. Like securitization was Like big giant corporations could. Could you know borrow money in the bond market, but little individual homeowners couldn't. Well, if you securitize mortgage, um, mortgages, then it allows little consumer homeowners to actually get big corporate pricing on their bonds.

Here it's like retail has had access and now only now is institutional getting kind of dealt in in at least in a robust way. So you don't think, then it doesn't sound like a a coin Bas is a bearish move for you, nor does it sound like you think, uh, being bearish. On on, let's say Bitcoin where maybe it could stagnate like the price of gold over the last decade or so uh is in your forecast Instead, bullishness. Well, I mean, think about it.

There's like a couple ways to approach the problem, but like one way to think about it is we think that there's going to be roughly a quadrillion dollars in financial assets globally by 2030. Okay, uh, and that's not making anything any. There's no aggressiveness embedded in that forecast other than our GDP expectati