Dwarkesh Podcast
Dwarkesh Podcast
Austin Vernon - Energy Superabundance, Starship Missiles, & Finding Alpha

Austin Vernon - Energy Superabundance, Starship Missiles, & Finding Alpha

How energy superabundance will change the world, how Starship can be turned into a kinetic weapon, why nuclear is overrated, blockchains, batteries, flying cars, finding alpha, & much more!

Austin Vernon is an engineer working on a new method for carbon capture, and he has one of the most interesting blogs on the internet, where he writes about engineering, software, economics, and investing.

We discuss how energy superabundance will change the world, how Starship can be turned into a kinetic weapon, why nuclear is overrated, blockchains, batteries, flying cars, finding alpha, & much more!

Watch on YouTube. Listen on Apple Podcasts, Spotify, or any other podcast platform. Read the full transcript here.

Subscribe to find out about future episodes!

Follow Austin on Twitter. Follow me on Twitter for updates on future episodes.

Please share if you enjoyed this episode! Helps out a ton!



(0:00:00) - Intro

(0:01:53) - Starship as a Weapon

(0:19:24) - Software Productivity

(0:41:40) - Car Manufacturing

(0:57:39) - Carbon Capture

(1:16:53) - Energy Superabundance

(1:25:09) - Storage for Cheap Energy

(1:31:25) - Travel in Future

(1:33:27) - Future Cities

(1:39:58) - Flying Cars

(1:43:26) - Carbon Shortage

(1:48:03) - Nuclear

(2:12:44) - Solar

(2:14:44) - Alpha & Efficient Markets

(2:22:51) - Conclusion



Dwarkesh Patel (00:00:00):

Okay! Today, I have the pleasure of interviewing Austin Vernon who writes about engineering, software, economics, and investing on the internet, though not that much else is known about him. So Austin, do you want to give us a bit of info about your background? I know that the only thing the internet knows about you is this one little JPEG that you had to upload with your recent paper. But what about an identity reveal or I guess a little bit of a background reveal? Just to the extent that you're comfortable sharing.

Austin Vernon (00:00:29):

My degree is in chemical engineering and I’ve had a lifelong love for engineering as well as things like the Toyota Production System. I've also worked as a chemical engineer in a large processing facility where I've done a lot of petroleum engineering. I taught myself how to write software and now I'm working on more research and the early commercialization of CO2 electrolysis.

Dwarkesh Patel (00:00:59):

Okay yeah. I'm really interested in talking about all those things. The first question I have is from Alex Berger, who's the co-CEO of Open Philanthropy. When I asked on Twitter what I should ask you, he suggested that I should ask “Why so shady?” Famously you have kind of an anonymous personality, pseudonymous thing going on the internet. What's up with that?

Austin Vernon (00:01:25):

Yeah. I think he posted a tweet that said “I don't know who this guy is or if he's credible at all, but his stuff sure is interesting”. That really made me laugh. I thought that was hilarious. Fame just doesn't seem necessary, I think I'm fine with my ideas being well known and communicating, but I have less desire to be personally famous.

Starship as a Weapon

Dwarkesh Patel (00:01:52):

Gotcha, gotcha. I wanted to start off with a sexy topic, let's talk about using Starship as a kinetic weapon. I thought that was one of the more amusing posts you wrote. Do you want to talk more about how this would be possible?

Austin Vernon (00:02:08):

Well, I think the main thing with Starship is that you're taking a technology and you're making it about 100 times cheaper for cargo and 1000 times cheaper for people. When things like that happen that drastically, you're just looking at huge changes and it’s really hard to anticipate what some of those can be when the change is that drastic. I think there's a lot of moon-based, Mars-based stuff that doesn't really catch the general public's eye. They also have trouble imagining some of the point-to-point travel that could be possible. But when you start talking about it as a weapon, then I think it lets people know they should be paying attention to this technology. And we certainly do not want to be second or third getting it. We should make sure that we're going to be first.

Dwarkesh Patel (00:03:05):

Yeah. I think you mentioned this in the post, but as recently as the '90s, the cost of sending one kilogram to space was around $20,000. More recently, SpaceX has brought it to $2,000. Lots of interesting questions pop up when you ask, “What will be possible once we get it down to $200 per kilogram to send into orbit?” One of them could be about how we might manufacture these weapons that are not conventional ballistics. Do you want to talk about why this might be an advancement over conventional ballistic weapons?

Austin Vernon (00:03:37):

Well, regular conventional ballistic weapons are extremely expensive. This is more like a bomb truck. But usually we think of B52 as the bomb truck and this could be even cheaper than the B52, delivering just mass on target. When you think about how expensive it is to fly a B52 from Barksdale in Louisiana all the way across the world.. you can do it from south Texas or Florida with the Starship and get more emissions per day and the fuel ends up being. When you go orbital, it takes a lot to get to orbit. But then once you're in orbit, your fuel consumption's pretty good. So over long distances, it has a lot of advantage. That's why the point-to-point works for longer distances.

Austin Vernon (00:04:27):

There's really a sweet spot with these weapons where you want it to be pretty accurate, but you also want it to be cheap. You're seeing that problem with Russia right now as they have some fancy parade style weapons that are really expensive, like multi-billion dollar cruise missiles, but they're missing that $5,000 guided artillery shell or that $20,000 JDM that you can just pit massive. Or the multiple launch rocket system, guided rockets. They're really short on all those because I think they had just had a limited amount of chips they could get from the US into Russia to make these advanced weapons.

Austin Vernon (00:05:07):

But yeah, so the Starship gives you just a platform to deliver. You could put JDMs in a shroud, or you could just have the iron unguided kinetic projectiles, and it just becomes impossible for a ship to launch missiles to intercept yours if your cost is so low, you can just overwhelm them.

Dwarkesh Patel (00:05:29):

Okay. There are a few terms there that neither I nor the audience might know. So what is JDM? What is shroud? And why are chips a bottleneck here? Why can't it just be any micro-controller?

Austin Vernon (00:05:42):

So JDM is Joint Direct Attack Munition. So what we did is we took all our Vietnam surplus bonds and we put this little fin-kit on it and it costs like $20,000, which is cheap for a weapon because the actual bond costs, I don't know, $3,000. And then it turns it into a guided weapon that, before you were probably lucky to get within 500 meters of a target, now you can get it in with two meters. So the number of missions you have to do with your planes and all that goes down by orders of magnitude. So it's an absolutely huge advantage in logistics and in just how much firepower you can put on a target. And we didn't even have to make new bombs, we just put these kits on all our old bombs.

Austin Vernon (00:06:33):

Let's see.. Yeah the chips are a problem. There's this organization called RUSI. I think they're in the UK, but they've been tearing down all these Russian weapons they found in Ukraine and they all have American chips in them. So technically, they're not supposed to be able to get these chips. And yet, Russia can't make a lot of its own chips. And especially not the specialized kinds you might want for guided weapons. So they've been somehow smuggling in chips from Americans to make their advanced weapons

Dwarkesh Patel (00:07:03):

What is special about these? As far as I'm aware, the trade with China is still going on and we get a lot of our chips manufactured from Taiwan or China. So why can't they do the same?

Austin Vernon (00:07:14):

It's the whole integration. It's not just the specific chip, but the board. They're more like PLCs where you almost have wired-in programming and they come with this ability to do the guidance and all that stuff. It all kind of has to work together. I think that's the way I understand it. I don't know. Maybe I don't have a really good answer for that one, but they're hard to replicate is what matters.

Dwarkesh Patel (00:07:43):

Okay that's interesting. Yeah, I guess that has a lot of interesting downstream effects, because for example, India buys a lot of its weapons from Russia. So if Russia doesn't have access to these, then other countries that buy from Russia won't have access to these either.

Dwarkesh Patel (00:07:58):

You had an interesting speculation in the post where you suggested that you could just keep these kinetic weapons in orbit, in a sort of Damocles state really, almost literally. That sounds like an incredibly scary and risky scenario where you could have orbital decay and you could have these kinetic weapons falling from the sky and destroying cities. Do you think this is what it will look like or could look like in 10 to 20 years?

Austin Vernon (00:08:26):

Well, yeah, so the advantage of having weapons on orbit is you can hit targets faster. So if you're launching the rocket from Florida, you're looking at maybe 30 minutes to get there and the target can move away in that time. Whereas if you're on orbit, you can have them spaced out to where you're hitting within a few minutes. So that's the advantage there.

Austin Vernon (00:08:46):

You really have to have a two stage system I think for most, because if you have a really aerodynamic rod that's going to give you really good performance in the low atmosphere, it’ll end up going too fast and just burn up before it gets there. Tungsten's maybe the only thing that you could have that could go all the way through which is why I like the original concept of using these big tungsten rods the size of a telephone pole. But tungsten's pretty expensive. And the rod concept kind of limits what you can do.

Austin Vernon (00:09:28):

So a lot of these weapons will have, that's what I was talking about with the shroud, something that actually slows you down in the upper atmosphere. And then once you're at the velocity where you're not just going to melt, then you open it up and let it go. So if you actually had it fall from the sky, some may make it to the ground, but a lot would burn up. So a lot of the stuff that makes it to the ground is actually pretty light. It's stuff that can float and has a large surface area. Yeah, that's the whole thing with Starship. Or not Starship, but Starlink. All those satellites are meant to completely fall apart on de-orbit.

Dwarkesh Patel (00:10:09):

I see. One of the implications of that is that these may be less powerful than we might fear, because since kinetic energy is mass times velocity squared and there's an upper bound on the velocity (velocity being the component that grows the kinetic energy faster), then it suggests that you can upper bound the power these things will have. You know what I mean?

Austin Vernon (00:10:32):

Yeah, so even the tungsten rods. Sometimes people, they're not very good at physics, so they don't do the math. They think it's going to be a nuclear weapon, but it's really not. I think even the tungsten rod is like 10 tons of T&T or something. It's a big bomb, but it's not a super weapon.

Austin Vernon (00:10:54):

So I think I said in the post, it's about using advanced missiles where they're almost more defensive weapons so I can keep you from pitting your ship somewhere. Yeah I could try to bombard your cities, but I can't take ground with it. I can't even police sea lanes with it really. I'd still have to use regular ships if I had this air cover to go enforce the rules of the sea and stuff like that.

Dwarkesh Patel (00:11:23):

Yeah. You speculated in the post, I think, that you could load this up with shrapnel and then it could explode next to an incoming missile or an incoming aircraft. Could these get that accurate? Because that was surprising speculation to me.

Austin Vernon (00:11:43):

I think for ships, it's pretty... I was watching videos of how fast a ship can turn and stuff. If you're going to do an initial target on a ship to try to kill their radars, you'd want to do it above the ceiling of their missiles. So it's like, how much are they going to move between your release where you stop steering and that? The answer’s maybe 1000 feet. So that's pretty simple because you just shrapnel the area.

Austin Vernon (00:12:12):

Targeting aircraft, you would be steering all the way in. I'd say it's doable, but it'd be pretty hard. You'd actually maybe want to even go slower than you would with the ship attack. You'd need a specialized package to attack the aircraft, but if you have enough synthetic aperture radar and stuff like that, you could see these aircraft using satellites and then guide the bomb in the whole way. You could even load heat seeking missiles into a package that unfurls right next to them and launch conventional missiles too, probably. It’d be pretty hard to do some of this stuff, but they’re just the things you might be able to do if you put some effort into it.

Dwarkesh Patel (00:12:57):

Yeah. The reason I find this kind of speculation really interesting is because when you look at the modern weaponry that's used in conflicts, it just seems directly descendant from something you would've seen in World War II or something. If you think about how much warfare changed between 1900 and 1940, it's like, yeah, they're not even the same class of weapons anymore. So it's interesting to think about possibilities like these where the entire category of weapons has changed.

Austin Vernon (00:13:33):

You’re right and that's because our physical technology hasn't changed that much. So it really has just made more sense to put better electronics in the same tanks. We haven't learned enough about tanks to build a new physical tank that's way better, so we just keep upgrading our existing tanks with better electronics. They're much more powerful, they're more accurate. A lot of times, they have longer range weapons and better sensors. So the tank looks the same, but it maybe has several times more killing power. But the Ukraine war right now, they're using a lot of 40, 50 year old weapons so that especially looks like that.

Dwarkesh Patel (00:14:20):

Yeah. Which kind of worries you if you think about the stockpiles our own military has. I'm not well educated on the topic, but I imagine that we don't have the newest of the new thing. We probably have maintained versions of decades old technology.

Austin Vernon (00:14:35):

We spend so much, we've got relatively... This kind of gets into debate about how ready our military is. For certain situations, it's more ready than others. I'd say in general, most people talking about it have the incentive to downplay our capabilities because they want more defense spending. There's lots of reasons. So I think we're probably more capable than what you might see from some editorial in The Hill or whatever. Us just sending a few weapons over to Ukraine and seeing how successful they've been at using them, I think, shows a little bit of that.

Austin Vernon (00:15:18):

There's so much uncertainty when it comes to fighting, especially when you're talking about a naval engagement, where we don't just don't have that many ships in general… you can have some bad luck. So I think you always want to be a little bit wary. You don't want to get overconfident.

Dwarkesh Patel (00:15:37):

Yeah. And if the offensive tech we sent to Ukraine is potentially better than the defensive tech, it's very possible that even a ballistic missile that China or Russia could launch would sink a battleship and then kill the 2,000 or 1,000 whatever soldiers that are on board. Or I guess, I don't know, you think this opens up avenues for defensive tech as well?

Austin Vernon (00:16:03):

Yeah––generally the consensus is that defensive technology has improved much more recently than offensive technology. This whole strategy China has is something they call anti-access/area denial, A2/AD. That's basically just how missiles have gotten better because the sensors on missiles have gotten better. So they can keep our ships from getting close to them but they can't really challenge us in Hawaii or something. And it really goes both ways, I think people forget that. So yeah, it's hard for us to get close to China, but Taiwan has a lot of missiles with these new sensors as well. So I think it's probably tougher for China to do it close to Taiwan than most people would say.

Dwarkesh Patel (00:16:55):

Oh, interesting. Yeah, can you talk more about that? Because every time I read about this, people are saying that if China wanted to, they could knock out Taiwan's defenses in a short amount of time and take it over. Yeah, so can you talk about why that's not possible?

Austin Vernon (00:17:10):

Well, it might be, but I think it's a guess of the uncertainty [inaudible 00:17:14]. Taiwan has actually one of the largest defense budgets in the world and they've recently been upping it. I think they spend, I don't know, $25 billion a year and they added an extra $5 billion. And they've been buying a lot of anti-ship missiles, a lot of air defense missiles.. Stuff that Ukraine could only dream of. I think Ukraine's military budget was $2 billion and they have a professional army. And then the other thing is Taiwan’s an island, whereas Russia could just roll over the land border into Ukraine.

Austin Vernon (00:17:44):

There's just been very few successful amphibious landings in history. The most recent ones were all the Americans in World War II and Korea. So the challenge there is just... It's kind of on China to execute perfectly and do that. So if they had perfect execution, then possibly it would be feasible. But if their air defenses on their ships aren't quite as good as we think they could possibly be, then they could also end up with half their fleet underwater within 10 hours.

Dwarkesh Patel (00:18:20):

Interesting. And how has your view of Taiwan's defensive capabilities changed... How has the Ukraine conflict updated your opinion on what might happen?

Austin Vernon (00:18:29):

I didn't really know how much about it. And then I started looking at Wikipedia and stuff and all this stuff they're doing. Taiwan just has a lot of modern platforms like F16s with our anti-ship missiles. They actually have a lot of their own. They have indigenous fighter bombers, indigenous anti-ship missiles because they're worried we might not always sell them to them.

Austin Vernon (00:18:54):

They've even recently gotten these long range cruise missiles that could possibly target leadership in Beijing. So I think that makes it uncomfortable for the Chinese leadership. If you attack them, you're going to have to go live in a bunker. But again, I'm not a full-time military analyst or something, so there's a lot of uncertainty around what I'm saying. It's not a given that China's just going to roll over them.

Software Productivity

Dwarkesh Patel (00:19:22):

Okay. That's comforting to hear. Let's talk about an area where I have a little bit of a point of contact. I thought your blog post about software and the inability of it to increase productivity numbers, I thought that was super fascinating. So before I ask you questions about it, do you want to lay out the thesis there?

Austin Vernon (00:19:43):

Yeah. So if there's one post I kind of felt like I caught lightning in a bottle on, it's that one. Everything I wanted to put in, it just fit together perfectly, which is usually not the case.

Austin Vernon (00:19:55):

I think the idea is that the world's so complex and we really underestimate that complexity. If you're going to digitize processes and automate them and stuff, you have to capture all that complexity basically at the bit level, and that's extremely difficult. And then you also have diminishing returns where the easily automatable stuff goes first and then it's increasing corner cases to get to the end, so you just have to go through more and more code basically. We don't see runaway productivity growth from software because we're fighting all this increasing complexity.

Dwarkesh Patel (00:20:39):

Yeah. Have you heard of the waterbed theory of complexity by the way?

Austin Vernon (00:20:42):

I don't think so.

Dwarkesh Patel (00:20:44):

Okay. It's something that comes up in compiler design: the idea is that there's a fixed amount of complexity in a system. If you try to reduce it, what you'll end up doing is just you'll end up migrating the complexity elsewhere. I think an example that's used of this is when they try to program languages that are not type safe, something like Python. You can say, “oh, it's a less complex language”, but really, you've added complexity when, I don't know, two different types of numbers are interacting like a float and an int. As your program grows, that complexity exponentially grows along with all the things that could go wrong when you're making two things interact in a way that you were expecting not to. So yeah, the idea is you can just choose where to have your complexity, but you can't get rid of that complexity.

Austin Vernon (00:21:38):

I think that’s kind of an interesting thing when you start pairing it with management theory... when you add up all the factors, the most complex thing you're doing is high volume car manufacturing. And so we got a lot of innovations and organization from car manufacturers like the assembly line. Then you had Sloan at GM basically creating the way the modern corporation is run, then you have the Toyota Production System.

Austin Vernon (00:22:11):

But arguably now, creating software is actually the most complex thing we do. So there's all these kinds of squishy concepts that underlie things like the Toyota Production System that softwares had to learn and reimagine and adopt and you see that with Agile where, “oh, we can't have long release times. We need to be releasing every day,” which means we're limiting inventory there.

Austin Vernon (00:22:42):

There's a whole thing especially that's showing up in software that existed in carbon manufacturing where you're talking about reducing communication. So Jeff Bezos kind of now famously said, "I want to reduce communication," which is counterintuitive to a lot of people. This is age-old in car manufacturing where Toyota has these cards that go between workstations and they tell you what to do. So people normally think of them as limiting inventory, but it also tells the worker exactly what they're supposed to be doing at what pace, at what time. The assembly line is like that too. You just know what to do because you're standing there and there's a part here and it needs to go on there, and it comes by at the pace you're supposed to work at.

Austin Vernon (00:23:29):

It's so extreme that there's this famous paper, by List, Syverson and Levitt. They went to a car factory and studied how defects propagated in cars and stuff. Once a car factory gets up and running, it doesn't matter what workers you put in there, if workers are sick or you get new workers, the defect rate is the same. So all the knowledge is built into the manufacturing line.

Austin Vernon (00:23:59):

There’s these concepts around idiot-proofing and everything that are very similar to what you'll see. You had Uncle Bob on here. So Uncle Bob says only put one input into a function and stuff like that because you'll mix them up otherwise. The Japanese call it poka-yoke. You make it where you can't mess it up. And that's another way to reduce communication, and then software, of course you have APIs.

Austin Vernon (00:24:28):

So I'm really interested in this overall concept of reducing communication, and reducing how much cooperation and everything we need to run the economy.

Dwarkesh Patel (00:24:41):

Right. Right. Speaking of the Toyota Production System, one thing they do to reduce that defect rate is if there's a problem, all the workers in that chain are forced to go to the place where the defect problem is and fix it before doing anything else. The idea there is that this will give them context to understand what the problem was and how to make sure it doesn't happen again. It also prevents a build up of inventory in a way that keeps making these defects happen or just keeps accumulating inventory before the place that can fix the defects is able to take care of them.

Austin Vernon (00:25:17):

Right. Yeah, yeah. Exactly.

Dwarkesh Patel (00:25:19):

Yeah. But I think one interesting thing about software and complexity is that software is a place where complexity is the highest in our world right now but software gives you the choice to interface with the complexity you want to interface with. I guess that's just part of specialization in general, but you could say for example that a machine learning model is really complex, but ideally, you get to a place where that's the only kind of complexity you have to deal with. You're not having to deal with the complexity of “How is this program compiled? How are the libraries that I'm using? How are they built?” You can fine tune and work on the complexity you need to work on.

Dwarkesh Patel (00:26:05):

It's similar to app development. Byrne Hobart has this blog post about Stripe as solid state. The basic idea is that Stripe hides all the complexity of the financial system: it charges a higher fee, but you can just treat it as an abstraction of a tithe you have to pay, and it'll just take care of that entire process so you can focus on your comparative advantage.

Austin Vernon (00:26:29):

It's really actually very similar in car manufacturing and the Toyota Production System if you really get into it. It's very much the same conceptual framework. There's this whole idea in Toyota Production System, everyone works at the same pace, which you kind of talked about. But also, your work content is the same. There's no room for not standardizing a way you're going to do things. So everyone gets together and they're like, “All right, we're going to do this certain part. We're going to put it together this certain way at this little micro station. And it's going to be the same way every time.” That's part of how they're reducing the defect rates. If your assembly process is longer than what your time allotment is to stay in touch with the rest of the process, then you just keep breaking it down into smaller pieces. So through this, each person only has to know a very small part of it.

Austin Vernon (00:27:33):

The overall engineering team has all sorts of strategies and all sorts of tools to help them break up all these processes into very small parts and make it all hold together. It's still very, very hard, but it's kind of a lot of the same ideas because you're taking away the complexity of making a $30,000 car or 30,000 part car where everyone's just focusing on their one little part and they don't care what someone else is doing.

Dwarkesh Patel (00:28:06):

Yeah. But the interesting thing is that it seems like you need one person who knows how everything fits together. Because from what I remember, one of the tenets of the Toyota Production System was you need to have a global view. So, in that book, was it the machine or the other one, the Toyota Production System book? But anyways, they were talking about examples where people would try to optimize for local efficiencies. I think they especially pointed to Ford and GM for trying to do this where they would try to make machines run all the time. And locally, you could say that, “oh this machine or process is super efficient. It's always outputting stuff.” But it ignores how that added inventory or that process had a bad consequence for the whole system.

Dwarkesh Patel (00:28:50):

And so it's interesting if you look at a company like Tesla that’s able to do this really well. Tesla is run like a monarchy and this one guy has this total global view of how the entire process is supposed to run and where you have these inefficiencies.. You had some great examples of this in the blog post. I think one of the examples is this guy (the author) goes to this factory and he asks, "Is this an efficient factory?" And the guy's like, "Yeah, this is totally efficient. There's nothing we can do, adopting the Toyota way, to make this more efficient."

Dwarkesh Patel (00:29:22):

And so then he's like, "Okay, let me look." And he finds that they're treating steel in some way, and the main process does only take a couple of seconds, but some local manager decided that it would be more efficient to ship their parts out, to get the next stage of the process done somewhere else. So this is locally cheaper, but the result is that it takes weeks to get these parts shipped out and get them back. Which means that the actual time that the parts spend getting processed is 0.1% of the time, making the whole process super inefficient. So I don't know, it seems like the implication is you need a very monarchical structure, with one person who has a total view, in order to run such a system. Or am I getting that wrong?

Austin Vernon (00:30:12):

Not necessarily. I mean, you do have to make sure you're not optimizing locally, but I think it's the same. You have that same constraint in software, but I think a lot of times people are just running over it because processing has been getting so much cheaper. People are expensive, so if you could save development time, it just ends up the trade offs are different when you're talking about the tyranny of physical items and stuff like that, the constraints get a little more severe. But I think you have the same overall. You still have to fight local optimization, but the level you have to is probably different with physical goods.

Austin Vernon (00:30:55):

I was thinking about the smart grid situation from a software perspective, and there's this problem where, okay, I'm putting my solar farm here and it's impacting somewhere far away, and that's then creating these really high upgrade costs, that cost two or three times more than my solar farm. Well, the obvious thing would be, if you're doing software, is like you're going to break all these up into smaller sections, and then you wouldn't be impacting each other and all that, and you could work and focus on your own little thing.

Austin Vernon (00:31:29):

But the problem with that is if you're going to disconnect these areas of the grid, the equipment to do that is extremely expensive. It's not like I'm just going to hit a new tab and open a new file and start writing a new function. And not only that, but you still have to actually coordinate how this equipment is going to operate. So if you just let the grid flow as it does, everyone knows what's going to happen because they could just calculate the physics. If you start adding in all these checkpoints where humans are doing stuff, then you have to actually interface with the humans, and the amount of things that can happen really starts going up. So it's actually a really bad idea to try to cart all this stuff off, just because of the reality of the physical laws and the equipment you need and everything like that.

Dwarkesh Patel (00:32:22):

Okay. Interesting. And then I think you have a similar Coasean argument in your software post about why vertically integrating software is beneficial. Do you want to explain that thesis?

Austin Vernon (00:32:34):

Yeah. I think it actually gets to what we're talking about here, where it allows you to avoid the local optimization. Because a lot of times you're trying to build a software MVP, and you're tying together a few services… they don't do quite what you need, so if you try to scale that, it would just break. But if you're going to take a really complex process, like car manufacturing or retail distribution, or the home buying process or something, you really have to vertically integrate it to be able to create a decent end-to-end experience and avoid that local optimization.

Austin Vernon (00:33:20):

And it's just very hard otherwise, because you just can't coordinate effectively if you have 10 different vendors trying to do all the same thing. You end up in just constant vendor meetings, where you're trying to decide what the specs are or something instead of giving someone the authority, or giving a team the authority to just start building stuff. Then if you look at these companies, they have to implement these somewhat decentralized processes when they get too complex, but at least they have control over how they're interfacing with each other. Walmart, as the vendors, control their own stock. They don't tell the vendor, "We need X parts." It's just like, it's on you to make sure your shelf is stocked.

Dwarkesh Patel (00:34:07):

Yeah. Yeah. So what was really interesting to me about this part of the post was, I don't know, I guess I had heard of this vision of we're software setting, where everybody will have a software as a service company, and they'll all be interfacing with each other in some sort of cycle where they're all just calling each other's APIs. And yeah, basically everybody and their mother would have a SAAS company. The implication here was, from your argument, that given the necessity of integrating all those complexity vertically in a coherent way, then the winners in software should end up being a few big companies, right? They compete with each other, but still...

Austin Vernon (00:34:49):

I think that's especially true when you're talking about combining bits and apps. Maybe less true for pure software. The physical world is just so much more complex, and so the constraints it creates are pretty extreme, compared to like... you could maybe get away with more of everyone and their mom having an API in a pure software world.

Dwarkesh Patel (00:35:14):

Right. Yeah. I guess, you might think that even in the physical world, given that people really need to focus on their comparative advantage, they would just try to outsource the software parts to these APIs. But is there any scenario where the learning curve for people who are not in the firm can be fast enough that they can keep up with the complexity? Because there's huge gains for specialization and competition that go away if this is the world we're forced to live in. And then I guess we have a lot of counter examples, or I guess we have a lot of examples of what you're talking about. Like Apple is the biggest market cap in the world, right? And famously they're super vertically integrated. And yeah, obviously their thing is combining hardware and software. But yeah, is there any world in which it can keep that kind of benefit, but have it be within multiple firms?

Austin Vernon (00:36:10):

This is a post I've got on my list I want to write. The blockchain application, which excites me personally the most, is reimagining enterprise software. Because the things you're talking about, like hard typing and APIs are just basically built into some of these protocols. So I think it just really has a lot of exciting implications for how much you can decentralize software development. But the thing is, you can still do that within the firm. So I think I mentioned this, if the government's going to place all these rules on the edge of the firm, it makes transactions with other firms expensive. So a few internal transactions can be cheaper, because they're avoiding the government reporting and taxes and all that kind of stuff. So I think you'd have to think about how these technologies can reduce transaction costs overall and decentralize that, but also what are the costs between firms?

Dwarkesh Patel (00:37:22):

Yeah, it's really interesting if the costs are logistic, or if they're based on the knowledge that is housed, as you were talking about, within a factory or something. Because if it is just logistical and stuff, like you had to report any outside transactions, then it does imply that those technology blockchain could help. But if it is just that you need to be in the same office, and if you're not, then you're going to have a hard time keeping up with what the new requirements for the API are, then maybe it's that, yeah, maybe the inevitability is that you'll have these big firms that are able to vertically integrate.

Austin Vernon (00:37:59):

Yeah, for these big firms to survive, they have to be somewhat decentralized within them. So I think you have... you're going to the same place as just how are we viewing it, what's our perception? So even if it's a giant corporation, it's going to have very independent business units as opposed to something like a 1950s corporation.

Dwarkesh Patel (00:38:29):

Yeah. Byrne Hobart, by the way, has this really interesting post that you might enjoy reading while you're writing that post. It's type safe communications, and it's about that Bezos thing, about his strict style for how to communicate and how little to communicate. There's many examples in Amazon protocols where you have to... the only way you can put in this report, is in this place you had to give a number. You can't just say, "This is very likely," you had to say like, "We project X percent increase," or whatever. So it has to be a percent. And there's many other cases where they're strict about what type definition you can have in written reports or something. It has kind of the same consequence that type strict languages have, which is that you can keep track of what the value is through the entire chain of the flow of control.

Austin Vernon (00:39:22):

You've got to keep work content standardized.

Dwarkesh Patel (00:39:26):

So we've been hinting at the Coasean analysis to this. I think we just talked about it indirectly, but for the people who might not know, Coase has this paper called The Theory of Firms, and he's trying to explain why we have firms at all. Why not just have everybody compete in the open market for employment, for anything? Why do we have jobs? Why not just have... you can just hire a secretary by the day or something.

Dwarkesh Patel (00:39:51):

And the conclusion he comes to is that by having a firm you're reducing the transaction cost. So people will have the same knowledge about what needs to get done, obviously you're reducing the transaction cost of contracting, finding labor, blah, blah, blah. And so the conclusion it comes to is the more the transaction costs are reduced within people in a firm, as compared to the transaction cost between different firms, the bigger firms will get. So I guess that's why the implication of your argument was that there should be bigger tech firms, right?

Austin Vernon (00:40:27):

Yes, yes, definitely. Because they can basically decrease the transaction costs faster within, and then even at the limit, if you have large transaction costs outside the firm, between other firms that are artificially imposed, then it will make firms bigger.

Dwarkesh Patel (00:40:45):

What does the world look like in that scenario? So would it just be these Japanese companies, these huge conglomerates who are just... you rise through the ranks, from the age of 20 until you die? Is that what software will turn into?

Austin Vernon (00:40:59):

It could be. I mean, I think it will be lots of very large companies, unless there's some kind of change in inner firm transaction costs. And again, that could possibly come from blockchain like technology, but you probably also need better regulation to make that cheaper, and then you would have smaller firms. But again, in the end, it doesn't really matter. You'd be working in your little unit of the big bank of corporate, or whatever. So I don't know what that would look like on a personal level.

Car Manufacturing

Dwarkesh Patel (00:41:40):

Yeah. Okay. So speaking of these Japanese companies, let's talk about car manufacturing and everything involved there. Yeah, so we kind of hinted at a few elements of the Toyota way and production earlier, but do you want to give a brief overview of what that is, so we can compare it to potentially other systems?

Austin Vernon (00:42:02):

I think all these kinds of lean Toyota process systems, they do have a lot of similarities, where mostly you want to even-out your production, so you're producing very consistently, and you want to break it into small steps and you want to limit the amount of inventory you have in your system. When you do this, it makes it easy to see how the process is running and limit defects. And the ultimate is you're really trying to reduce defects, because they're very expensive. It's a little bit hard to summarize. I think that's my best shot at it there, quickly off the top of my head.

Dwarkesh Patel (00:42:49):

Yeah. The interesting thing about the Toyota system, so at least when the machine was released, is they talk about... that book was released I think the nineties, and they went to the history of Toyota, and one of the interesting things they talked about was there was a brief time where the company ran... I think, was this after World War II? But anyways, the company ran into some troubles. They needed to layoff people to not go bankrupt. They had much more debt on books than they had assets. So yeah, they wanted to layoff people, but obviously the people were not happy about this, so there were violent protests about this. And in fact I think the US written constitution gave strong protections to labor that they hadn't had before, which gave labor an even stronger hand here.

Dwarkesh Patel (00:43:42):

So anyway, Toyota came to this agreement with the unions that they'd be allowed to do this one time layoff to get the company on the right track, but afterwards they could never lay somebody off. Which would mean that a person who works at Toyota works there from the time they graduate college or high school till they die. Right? I don't know, that's super intense in a culture. I mean, in software, where you have the average tenure in a company's one year, the difference is so much.

Dwarkesh Patel (00:44:13):

And there's so many potential benefits here, I guess a lot of drawbacks too. But one is, obviously if you're talking in a time scale of 50 years, rather than one year, the incentives are more aligned between the company and the person. Because anything you could do in one year is not going to have a huge impact on your stock options in that amount of time. But if this company's your retirement plan, then you have a much stronger incentive to make sure that things at this company run well, which means you're probably optimizing for the company's long term cash flow yourself. And also, there's obviously benefits to having that knowledge built up in the firm from people who have been there for a long time. But yeah, that was an interesting difference. One of the interesting differences, at least.

Austin Vernon (00:45:00):

I mean, I think there's diminishing returns to how long your tenure's going to be. Maybe one year's too short, but there's a certain extent to where, if you grow faster than your role at the company, then it's time to switch. It's going to depend on the person, but maybe five years is a good number. And so if you're not getting promoted within the firm, then your human capital's being wasted, because you could go somewhere else and have more responsibility and perform better for them. Another interesting thing about that story, is almost all lean turnarounds, where they're like, we're going to implement something like Toyota production system, they come with no layoff promises. Because if you're going to increase productivity, that's when everyone's like, "Oh gosh, I'm going to get laid off." So instead you have to increase output and take more market share, is what you do.

Dwarkesh Patel (00:46:00):

It's kind of like burning your bridges, right? So this is the only way.

Austin Vernon (00:46:05):

The process really requires complete buy-in, because a lot of your ideas for how you're going to standardize work content come from your line workers, because that's what they're doing every day. So if you don't have their buy-in, then it's going to fail. So that's why it's really necessary to have those kinds of clauses.

Dwarkesh Patel (00:46:22):

Yeah. Yeah, that makes sense. I think it was in your post where you said, if somebody makes their process more efficient, and therefore they're getting more work allotted to them, then obviously they're going to stop doing that. Right? Which means that, I don't know, do you ought to give more downtime to your best workers or something or the people who are most creative in your company?

Austin Vernon (00:46:48):

I was just going to say, if you're a worker at a plant, then a lot of times for that level of employee, actually small rewards work pretty well. A lot of people on drilling rigs used to give the guys that met certain targets $100 Walmart gift cards. So sometimes small, it's a reward, new ideas, stuff like that works.

Austin Vernon (00:47:15):

But because the whole system has to grow together, if you just improve one part of the process, it may not help you. You have to be improving all the right processes so normally it's much more collaborative. There's some engineer that's looking at it and like, "All right, this is where we're struggling," or "We have our defects here." And then you go get together with that supervisor and the workers in that area, then you all figure out what improvements could be together. Because usually the people already know. This is like, you see a problem at the top, and you're just now realizing it. Then you go talk to the people doing the work, and they're like, "Oh yeah, I tried to tell you about that two weeks ago, man." And then you figure out a better process from there.

Dwarkesh Patel (00:47:58):

Based on your recommendation, and Steven Malina's recommendation, I recently read The Goal. And after reading the book, I'm much more understanding of the value that consultants bring to companies, potentially. Because before you could think, “What does a 21 year old, who just graduated college, know about manufacturing? What are they going to tell this plant that they didn't already know? How could they possibly be adding value?” And afterwards, it occurred to me that there's so many abstract concepts that are necessary to understand in order to be able to increase your throughput. So now I guess I can see how somebody who's generically smart but doesn't have that much industry knowledge might be able to contribute to a plan and value consultants could be bringing.

Austin Vernon (00:48:43):

I think this applies to consultants or young engineers. A lot of times you put young engineers just right in the thick of it, working in production or process right on the line, where you're talking to the workers the most. And there's several advantages to that. One, the engineer learns faster, because they're actually seeing the real process, and the other is there's easy opportunities for them to still have a positive impact on the business, because there's $100 bills laying on the ground just from going up and talking to your workers and learning about stuff and figuring out problems they might be having and finding out things like that that could help you lower cost. I think there's a lot of consultants that... I don't know how the industry goes, but I would guess there's... I know Accenture has 600,000 employees. I don't know if that many, but it's just a large number, and a lot are doing more basic tasks and there are some people that are doing the more high level stuff, but it's probably a lot less.

Dwarkesh Patel (00:49:51):

Yeah. Yeah. There was a quote from one of those books that said, "At Toyota we don't consider you an engineer unless you need to wash your hands before you can have lunch." Yeah. Okay. So in your blog post about car manufacturing, you talk about Tesla. But what was really interesting is that in a footnote, I think you mentioned that you bought Tesla stocks in 2014, which also might be interesting to talk about again when we go to the market and alpha part. But anyways. Okay. And then you talk about Tesla using something called metal manufacturing. So first of all, how did you know in 2014 that Tesla was headed here? And what is metal manufacturing and how does it differ from the Toyota production system?

Austin Vernon (00:50:42):

Yeah. So yeah, I just was goofing around and made that up. Someone actually emailed me and they were like, "Hey, what is this metal manufacturing? I want to learn more about this." It's like, "Well, sorry, I just kind of made that up, because I thought it sounded funny." But yeah, I think it's really the idea that there's this guy, Dimming, and he found a lot of the same ideas that Toyota ended up implementing, and Toyota respected his ideas a lot. America never really got fully on board with this in manufacturing. Of course it's software people that are coming and implementing this and manufacturing now which is like the real American way of doing things.

Austin Vernon (00:51:32):

Because when you look at these manufacturing processes, the best place to save money and optimize is before you ever build the process or the plant. It's very early on. So I think if there's a criticism of Toyota, it's that they're optimizing too late and they're not creative enough in their production technology and stuff. They're very conservative, and that's why they have hydrogen cars and not battery cars, even though they came out with the Prius, which was the first large sales hybrid.

Austin Vernon (00:52:12):

So yeah, I think what Tesla's doing with really just making Dimming's ideas our own and really just Americanizing it with like, "Oh, well, we want to cast this, because that would be easier." Well, we can't, because we don't have an alloy. "We'll invent the alloy." I love it. It's great. Mostly, I love Tesla because they do such... I agree with their engineering principles. So I didn't know that the company would come to be so valuable. It's just, I was just always reading their stock reports and stuff so I was like, "Well, at least I need to buy some stock so that I have a justification for spending all this time reading their 10 Ks."

Dwarkesh Patel (00:52:53):

I want to get a little bit more in detail about the exact difference here. So lean production, I guess, is they're able to produce their cars without defects and with matching demand or whatever. But what is it about their system that prevents them from making the kinds of innovations that Tesla is able to make?

Austin Vernon (00:53:16):

It's just too incremental. It's so hard to get these processes working. So the faster you change things, it becomes very, very difficult to change the whole system. So one of the advantages Tesla has is, well, if you're making electric cars, you have just a lot less parts. So that makes it easier. And once you start doing the really hard work of basically digitizing stuff, like they don't have speed limit dials, you start just removing parts from the thing and you can actually then start increasing your rate of change even faster.

Austin Vernon (00:53:55):

It makes it harder to get behind if you have these old dinosaur processes. But I think there's a YouTube channel called The Limiting Factor, and he actually went into the detail of numbers on what it costs for Tesla to do their giga-casting, which saves tons of parts and deletes zillions of thousands of robots from their process. If you already have an existing stamping line and all that, where you're just changing the dyes based on your model, then it doesn't make sense to switch to the casting. But if you're building new factories, like Tesla is, well, then it makes sense to do the casting and you can build new factories very cheaply and comparatively and much easier. So there's a little bit of... they just have lots of technical data, I guess you could say, in a software sense.

Dwarkesh Patel (00:54:47):

Yeah. That's super interesting. The analogy is actually quite... it's like, Microsoft has probably tens of thousands of software engineers who are just basically servicing its technical debt and making sure that the old systems run properly, whereas a new company like Tesla doesn't have to deal with that. The thing that's super interesting about Tesla is like, Tesla's market cap is way over a trillion, right? And then Toyota's is 300 billion. And Tesla is such a new company. The fact that you have this Toyota, which is legendary for its production system, and this company that's less than two decades old is worth many times more, it's kind of funny.

Austin Vernon (00:55:32):

Yeah. I would say that, in that measure, I don't like market cap. You need to use enterprise value. These old car companies have so much debt, that if you look at enterprise value, it's not so jarring. Literally, I don't know, I can't remember what GM's worth, like 40 billion or something, and then they have $120 billion in debt. So their enterprise value is five times more than their market cap.

Dwarkesh Patel (00:56:02):

What is enterprise value?

Austin Vernon (00:56:03):

Enterprise value is basically what is the value of the actual company before you have any claims on it. It's the market cap plus your debt. But basically, if you're the equity holder and the company gets sold, you have to pay the debt first. So you only get the value of what's left over after the debt. So that's why market cap is... when Tesla has very little debt and a lot of market cap, and then these other guys have a lot of debt with less market cap, it skews the comparison.

Dwarkesh Patel (00:56:34):

Yeah, and one of the interesting things, it's similar to your post on software, is that it seems like one of the interesting themes across your work is automating processes often leads to decreased eventual throughput, because you're probably adding capacity in a place that you're deciding excess capacity, and you're also making the money part of your operation less efficient by have it interface with this automated part. It sounds like there's a similar story there with car manufacturing, right?

Austin Vernon (00:57:08):

Yeah. I think if we tie it back into what we were talking about earlier, automation promotes local optimization and premature optimization. So a lot of times it's better to figure out, instead of automating a process to make a really hard to make part, you should just figure out how to make that part easy to make. Then after you do that, then it may not even make sense to automate it anymore. Or get rid of it all together, then you just delete all those robots.

Austin’s Carbon Capture Project

Dwarkesh Patel (00:57:37):

Yeah. Yeah, that's interesting. Okay. So let's talk about the project that you're working on right now, the CO2 electrolysis. Do you want to explain what this is, and what your current approach is? What is going on here?

Austin Vernon (00:57:55):

Yeah, so I think just overall, electrofuels right now are super underrated, because you're about to get hopefully some very cheap electricity from solar, or it could be, maybe, some land. If we get really lucky, possibly some nuclear, geothermal. It’ll just make sense to create liquid fuels, or natural gas, or something just from electricity and air, essentially.

Austin Vernon (00:58:25):

There's a whole spectrum of ways to do this, so O2 electrolysis is one of those. Basically, you take water, electricity, and CO2, and a catalyst. And then, you make more complex molecules, like carbon monoxide, or formic acid, or ethylene, or ethanol, or methane or methine. Those are all options. But it's important to point out that, right now, I think if you added up all the CO2 electrolyzers in the world, you'd be measuring their output and kilograms per day. We make millions of tons per day off of the products I just mentioned. So there's a massive scale up if it's going to have a wider impact.

Austin Vernon (00:59:15):

So there's some debate. I think the debate for the whole electrofuels sector is: How much are you going to do in the electrolyzer? One company whose approach I really like is Terraform Industries. They want to make methane, which is the main natural gas. But they're just making hydrogen in their electrolyzer, and then they capture the CO2 and then put it into a methanation reaction. So everything they're doing is already world scale, basically.

Austin Vernon (00:59:47):

We've had hydrogen electrolyzers power fertilizer plants, providing them with the Hydrogen that they need. Methanation happens in all ammonia plants and several other examples. It's well known, very old. Methanation is hydrogen CO2 combined to make water and methane. So their approach is more conservative, but if you do more in the electrolyzer, like I'm going to make the methane actually in the electrolyzer instead of adding this other process, you could potentially have a much simpler process that has less CapEx and scales downward better. Traditional chemical engineering heavily favors scaling. With the more Terraform processes, they're playing as absolutely ginormous factories. These can take a long time to build.

Austin Vernon (01:00:42):

So one of the things they're doing is: they're having to fight the complexity that creeps into chemical engineering every step of the way. Because if they don't, they'll end up with a plant that takes 10 years to build, and that's not their goal. It takes 10 years to build a new refinery, because they're so complex. So yeah, that's where I am. I'm more on the speculative edge, and it's not clear yet which products will be favorable for which approaches.

Dwarkesh Patel (01:01:15):

Okay, yeah. And you're building this out of your garage, correct?

Austin Vernon (01:01:19):

Yeah. So that's where electrolyzers... Everything with electric chemistry is a flat plate instead of a vessel, so it scales down. So I can have a pretty good idea of what my 100 square centimeter electrolyzer is going to do, if I make it quite a bit bigger. I have to worry about how my flow might interact in the larger one and make sure the mixing's good, but it's pretty straightforward because you're just making your flat plate a larger area. Whereas the scale, it is different from scaling a traditional chemical process.

Dwarkesh Patel (01:01:56):

I'm curious how cheap energy has to be before this is efficient. If you're turning it into methane or something like that, presumably for fuel, is the entire process energy positive? Or how cheap would energy, electricity you need to get before that's the case?

Austin Vernon (01:02:18):

The different products and different methods have different crossovers. So Terraform Industries, they're shooting for $10 a megawatt hour for electricity. But again, their process is simpler, a little less efficient than a lot of the other products. They also have better premiums, just worth more per ton than methane. So your crossover happens somewhere in between $10 and $20 a megawatt hour, which is... I mean, that's pretty... Right now, solar, it's maybe like $25. Maybe it's a little higher because payment prices have gone up in the last year, but I think the expectation is they'll come back down. And so, getting down to $15 where you start having crossovers for some of these products like ethanol or ethylene or methanol, it's not science fiction.

Dwarkesh Patel (01:03:08):

I think in Texas where I live, that's where it's at right? The cost of energy is 20 or something dollars per megawatt hour.

Austin Vernon (01:03:16):

Well, not this summer! But yeah, a lot of times in Texas, the wholesale prices are around $25 to $30.

Dwarkesh Patel (01:03:26):

Gotcha. Okay. Yeah. So a lot of the actual details you said about how this works went over my head. So what is a flat plate? I guess before you answer that question, can you just generally describe the approach? What is it? What are you doing to convert CO2 into these other compounds?

Austin Vernon (01:03:45):

Well, yeah, it literally just looks like an electrolyzer. You have two sides and anode and a cathode and they're just smushed together like this because of the electrical resistance. If you put them far apart, it makes it... uses up a lot of energy. So you smush them together as close as you can. And then, you're basically just trading electrons back and forth. On one side, you're turning CO2 into a more complex molecule, and on the other side, you're taking apart water. And so, when you take apart the water, it balances out the equation, balances out your electrons and everything like that. I probably need to work on that elevator pitch there, huh?

Dwarkesh Patel (01:04:31):

I guess what the basic idea is, you need to put power in to convert CO2 into these other compounds.

Austin Vernon (01:04:38):

The inputs are electricity, water, and CO2, and the output is usually oxygen and whatever chemical you're trying to create is, along with some side reactions.

Dwarkesh Patel (01:04:49):

And then, these chemicals you mentioned, I think ethanol, methane, formic acid, are these all just fuels or what are the other uses for them?

Austin Vernon (01:04:58):

A lot of people are taking a hybrid approach with carbon monoxide. So this would be like Twelve Co… They've raised a lot of money to do this and 100 employees or something. You can take that carbon monoxide and make hydrogen, and then you have to send gas to make liquid fuels. So they want to make all sorts of chemicals, but one of the main volume ones would be like jet fuel.

Austin Vernon (01:05:22):

Let's see Formic acid is, it's the little fry of all these. It is an additive in a lot of things like preserving hay for animals and stuff like that. Then, ethanol there's people that want to... There's this company that makes ethylene, which goes into plastics that makes polyethylene, which is the most produced plastic. Or you can burn it in your car, although I think ethanol is a terrible vehicle fuel. But then you can also just make ethylene straight in the electrolyzer. So there's many paths. So which path wins is an interesting race to see.

Dwarkesh Patel (01:06:13):

The ability to produce jet fuel is really interesting, because in your energy superabundance paper, you talk about... You would think that even if we can electrify everything in solar and when it becomes super cheap, that's not going to have an impact on the prices to go to space for example. But I don't know. If a process like this is possible, then it's some way to in financial terms, add liquidity. And then turn, basically, this cheap solar and wind into jet fuel through this indirect process. So the price to send stuff to space or cheap plane flights or whatever––all of that goes down as well.

Austin Vernon (01:06:52):

It basically sets a price ceiling on the price of oil. Whatever you can produce this for is the ceiling now, which is maybe the way I think about it.

Dwarkesh Patel (01:07:06):

Yeah. So do you want to talk a little bit about how your background led into this project? This is your full-time thing, right? I don't know if I read about that, but where did you get this idea and how long have you been pursuing it? And what's the progress and so on.

Austin Vernon (01:07:20):

I've always loved chemical engineering, and I love working at the big processing plant because it's like being a kid in a candy store. If I had extra time, I'd just walk around and look at the plant, like it’s so cool. But the plant where I worked at, their up time was 99.7%. So if you wanted to change anything or do anything new, it terrified everyone. That's how they earned their bonuses: run the plant a 100% uptime all the time. So that just wasn't a good fit for me. And also, so I always wanted my own chemical plant, but it's billions of dollars to build plants so that was a pretty big step. So I think this new technology of... there's a window where you might be able to build smaller plants until it optimizes to be hard to enter again.

Dwarkesh Patel (01:08:21):

And then, why will it become hard to enter again? What will happen?

Austin Vernon (01:08:27):

If someone figures out how to build a really cheap electrolyzer, and they just keep it as intellectual property, then it would be hard to rediscover that and compete with them.

Dwarkesh Patel (01:08:38):

And so, how long have you been working on this?

Austin Vernon (01:08:42):

Oh, not quite a year. But yeah, I actually got this idea to work on it from writing my blog. So when I wrote the heating fuel post, I didn't really know much about... There's another company in the space, Prometheus Fuels and I'm like, "Oh, this is an interesting idea." And then, I got talking to a guy named Brian Heligman, and he's like, "You should do this, but not what Prometheus is doing." And so, then I started looking at it and I liked it, so I've been working on it since.

Dwarkesh Patel (01:09:08):

Yeah. It's interesting because if energy does become as cheap as you suspect it might. If this process works, then yeah, this is a trillion dollar company probably, right? If you're going to get the patents and everything.

Austin Vernon (01:09:22):

I mean, maybe. With chemical plants, there's a certain limitation where your physical limitation is. There's only so many places that are good places for chemical plants. You start getting hit by transportation and all that. So, you can't just produce all the chemical for the entire world in Texas and transport it all around. It wouldn't work. So you're talking about a full, globe-spanning thing. At that point, if you're building factories all over the world, someone's going to figure out what your intellectual property is and all that. You would have to keep innovating to stay ahead of the competitors. I think that would limit your... Ultimately, it's a commodity. You're making commodity, so you don't have the same kind of defensibility that other sectors do.

Dwarkesh Patel (01:10:18):

I see. Yeah, yeah, yeah. Okay. There's not network effects I guess?

Austin Vernon (01:10:22):

Yeah. This is not quite consistent with what I just said about harder to enter. But I think what happens is the scale starts increasing as you go on. Even though this is easier to scale down, there's certain elements that are very much hard to scale and then the organization as well. Basically, you'll end up with early on a few competitors that continue to grow against each other and limit the margins. It'd be hard to be the fifth 30 years down the line.

Dwarkesh Patel (01:11:05):

What is the state of this project right now? Are you guys planning on starting a company, and what are the milestones you guys are shooting for?

Austin Vernon (01:11:14):

Right now, it is just me, but I have a family of engineers. We're all engineers, so it's loosely supported by them right now, by other people in my family as well, they're participating some. But yeah, basically, I just have to get... I've already done a lot of the theoretical design work at just a very cursory level to make sure it makes sense, and the cost will be reasonable and stuff like that. So now, it's working on the electrolizer to, basically, meet the targets you need for reliability and product concentration and energy costs. Also then just, is it manufacturable? Because right now, a lot of electrolyzers they use in the labs, they're literally smaller than a postage stamp, and they're very difficult to make.

Dwarkesh Patel (01:12:10):

Okay. I see. So you started working on this before or after you had quit your job?

Austin Vernon (01:12:16):

Oh, yeah. After. I quit my job five years ago or something. I was doing software stuff in between.

Dwarkesh Patel (01:12:21):

Oh, yeah? What did you work on?

Austin Vernon (01:12:24):

I worked on several products. I have one that’s an oil and gas data service that's somewhat successful. I’ve kept paying customers, but it's still relatively small.

Dwarkesh Patel (01:12:37):

Okay. I see. Yeah. So it seems like your blog is pretty recent, right? You started that about a year ago. What encouraged you to do that?

Austin Vernon (01:12:48):

Well, let's see. I was curious about cryptography in general, but specifically for blockchains. I wanted to be able to read the Bitcoin white paper and understand some of this IPFS. So I figured the best way to do it was... people talked about like, "Oh yeah, you should write, blah, blah, blah." So I did, "Well, I'll create an IPFS blog." I did that and learned a lot. It was not the most reliable blog when I was running it on my own Droplet and everything. So thankfully, I migrated to a service that has much more up time than my own server. So then, I wrote several posts to basically learn about it. I wrote posts about hash functions and private key cryptography. So then I could understand the white papers and what they're doing with the math and the cryptography. Eventually, I had this blog–– so my first non-crypto topic was on how to build a cheaper house or why it's difficult to reduce home construction costs. That made it on Hacker News and all that. It's like, "Oh, maybe actually people want to read this stuff," so I've just been writing since then in my spare time.

Dwarkesh Patel (01:14:18):

I actually interned for Protocol Labs, which is a place that built IPFS.

Austin Vernon (01:14:25):

Oh, yeah?

Dwarkesh Patel (01:14:26):

Yeah. So I got a chance to learn a lot about it and... about how file coin exactly works. That part threw me into a world for a while. But yeah, it's really interesting. I actually had a blog on IPFS. I mean, it was just a toy thing, not the one that I actually ended up writing on. The thing is, obviously at the moment, being it's like... nobody else is going to seat it for you, so you got to use a centralized service anyways, like Piñata, but it is a fun exercise.

Austin Vernon (01:14:57):

I was just running it off of Droplet and on DigitalOcean. If you use the direct content hash, it works pretty well, even if you're linking through your ENS name. But the problem is, of course, when I was first doing this, the fees on Ethereum were so high that I didn't want to change that link all the time. So I tried to use the pinning feature with IPNS because CloudFlare does the e-thought link. And then, they look up whatever your IPNS name is. And then, they tried to go find it.

Austin Vernon (01:15:35):

So the breaking point for me was how CloudFlare couldn't always find my server using IPNS. Now, the service I'm using is called Fleek. They basically go directly to the content hash, but on DNS it's cheap to change. You can change it in one minute. If Ethereum fees got lower, I might switch back to that, but I don't want to... Eventually, and I think it will be. What if it's like 1 cent transactions? Then it would be no big deal to just change the content hash every time you update your website.

Dwarkesh Patel (01:16:17):

What is the reason for having it on Ethereum?

Austin Vernon (01:16:21):

Just for fun.

Dwarkesh Patel (01:16:23):

It is inconvenient, I guess, if your content hash is changing every time you update the website, so you got to keep re-updating the actual... where people can find the site or use some other service to take care of it.

Austin Vernon (01:16:34):

I mean, yeah, if transactions are cheap, then you just have... You could automate it all, and it just cost you a little bit of money each time, and it'd be fine. But it was like $50, and I'm not going to pay $50 to post a blog post.

Energy Superabundance

Dwarkesh Patel (01:16:47):

Yeah. Yeah. And then, you find a typo. It's like, "Oh, gosh, I can't fix that." So you have a paper that you recently released with Eli Dourado on energy superabundance, and you have lots and lots of interesting speculation in there for what might be possible if energy gets a lot cheaper. I think we should just jump into it. On the big picture, as I'm sure you're aware, per capita energy used since the 1970s has not gone up. Before that, there's this thing called the Henry Adams Curve, where per capita energy use would increase 2% a year.

Dwarkesh Patel (01:17:22):

After 1970, that was no longer the case. Ironically enough, right after, the Department of Energy was created. But nonetheless, we've still had economic growth since the 1970s. I mean, it's been slower, but even though per capita energy hasn't increased per capita, GDP has increased. I think in the paper's abstract or the introduction, you talk about why increasing energy use is necessary for increasing economic growth. But doesn't that pattern suggest that you can still have decent economic growth without having to use energy, or have we just not come across the constraints yet?

Austin Vernon (01:17:57):

Hey, I mean, you just have diminishing returns. There's physical limits to how efficient things could be, and as you get closer to that efficiency limit, it's harder and harder and takes more and more effort. There's some diminishing returns there, where if you can just... A perfect example of what we were just talking about is oil's quite expensive and natural gas is expensive too. Oil’s easy to transport, you can produce it anywhere in the world and get anywhere else pretty cheaply. Natural gas is extremely expensive transport, but it's very useful fuel and for also making fertilizer. Not everyone has natural gas or the economic capability to extract natural gas using traditional processes. So if you have independent energy because you can just build these natural gas factories, where you're just using sunshine and water and air, then all of a sudden everyone has access to natural gas even if you weren't blessed with easily obtainable, natural gas reserves.

Austin Vernon (01:19:00):

I think that there's this whole story about the tyranny of geography here when it comes to energy. Because there are some countries that have extreme electricities per capita, like Iceland and Norway, where they have crazy amounts of hydro-power and people build aluminum plants there and stuff like that. Then you have places in Africa, where they have no coal, very little gas. They're just energy starved. Their transportation system sucks. You can't transport coal in.

Austin Vernon (01:19:39):

The hydropower is... There's only so much of it. It may not be close to where their cities are. So if you start adding solar to the mix for them, and some of these other technologies, it could really be an incredible increase in energy availability for them. I think we talked about that in the paper. We're looking at doubling rich-world use of energy, but it would be more like 10x more if you live in Africa.

Dwarkesh Patel (01:20:08):

Yeah. Yeah. So I wonder, if that's the case, then as energy becomes that abundant, will it just be other resources that become the bottleneck in terms of what our civilization needs... ? For example the actual resources that are necessary to build the factories and the raw materials? To what extent can even that be...

Austin Vernon (01:20:34):

I would argue the ultimate limit is like.. it is really human capital. What more abundant energy does is it allows you to redeploy human capital away from trying to figure out how to use scarce energy sources. Here's an example I love about trucking. I love trucks, not as big a fan of freight trains, but freight trains are extremely efficient. They get like 10 times more efficient than a truck or something.

Austin Vernon (01:21:15):

They use just very little fuel, but the flip sde is that the train doesn't come by all the time and they may not hold to the schedule. You have to aggregate your product with the other stuff or your raw materials, and it adds a lot of cost to your production. Toyota production system runs on trucks, not trains. The reason is that trucks are just extremely flexible. They come when you need it. They go when you need it. And even then, people still complain about truck drivers not showing up when you want them.

Austin Vernon (01:21:51):

When you have cheaper energy, this electrification and automation of trucking, you are going to shift a huge amount of goods from trains to trucks. And it's going to just have huge knock on effects all across the economy. It's more specialization. There's a lot of products that you're just limited to your suppliers because transportation's expensive. It reduces working capital, because a lot of times it takes longer on trains, similar stuff like smaller ships, more air freight. One thing that shocked me as Eli was telling me about how the elasticity of demand for air freight is just insane. When you even decrease the cost a little bit, demand goes to the roof.

Austin Vernon (01:22:37):

So I'm pretty sure that there'll be some kind of... You always think, "Oh, we can't do this with batteries," and then someone comes up with a more clever idea. So even if you have a 500 mile rein limit for your freight plane. The freight doesn't care if you have to stop every 500 miles to refuel or recharge. You can go over land on almost all these routes. You could go up through Japan and the Aleutian Islands, or you could go overland from China to Europe, charging just wherever's convenient.

Austin Vernon (01:23:15):

If that electric plane has half the operating cost of the jet plane, the amount of freight you're moving on airplanes will go way up, and it'll go down on ships. And then, everyone will be better off. Because right now, if you're a shipping company, you have real working capital problems, because your stuff sits on a boat for a month, and you've got to finance that and do all this stuff. And then, what if things change in the meantime? Like, "Oh, I don't really want that product anymore," so the air freight is just an absolute economic, just booster. So if you could make that cheaper, it's really exciting because it uses way more energy.

Dwarkesh Patel (01:24:01):

An analogy that had just occurred to me is, you could imagine that with computational power, if Moore's Law had stopped in 2005, we would still have a lot of interesting applications using compute and the effects of the computer would still have permeated society. But obviously, a lot of things that are possible today with computers, they just wouldn't have been tried or been possible in that kind of world.

Austin Vernon (01:24:27):

Okay. Yeah. I mean, all your engineers would be working on optimization instead of building new products.

Dwarkesh Patel (01:24:34):

Yeah. I think in J. Storr Hall's new book on... "Where's My Flying Car?" One of the points he makes is that GDP growth has been probably overstated because a lot of what constituted the growth has just been increasing the efficiency of existing machines to make them use less energy. Which still doesn't result in more total resources or goods or services being produced. But yeah, instead of making the laundry machine more efficient, you can just create a new kind of machine that may need to use more energy. So for this vision to come to pass, do you need energy to... Is it just enough that energy becomes super cheap or do you need advances in the ability to store that energy as well? So for example, lithium batteries are the bottleneck, does it matter if you can get energy super cheap, if you can't put them in appliances or cars or planes or whatever?

Austin Vernon (01:25:33):

I think the important thing to think about here is that our current energy is so expensive, especially electricity. With our energy resources, which are basically thermal, it's quite difficult to make electricity comparatively. And so, what we use electricity for is stuff we really want to use electricity for. It's hard to imagine that... We're not going to turn our air conditioner off. We're going to run it. And so, we're willing to pay a lot of money for that electricity to run our air conditioner. Whereas, if you look at really closely at a lot of the use cases that use tons of extra energy, they're much more flexible in how they use the energy. There's not a whole lot of storage evolved. If you're looking at growing crops or making methane for rocket fuel or making chemicals, you can design these processes to run when the energy's available. And so, the batteries are really going to be for keeping air conditioner on.. where you're willing to pay a lot of money. So I don't really see the batteries and storage as a limit.

Storage for Cheap Energy

Dwarkesh Patel (01:26:47):

Okay. If you had something like air freight, right? If that's the thing we're concerned about. Wouldn't you need some way to store that electricity for air freight, or maybe you can just convert it to jet fuel? Is that what you're saying?

Austin Vernon (01:27:03):

Yeah. I was thinking of more grid storage, but yeah, transportation's going to dominate battery demand. Grid storage is tiny in comparison, but I think you're basically getting to the point where we're making batteries out of dirt because that's how you scale it. So, if you're making batteries out of carbon and iron and phosphate, it becomes about how many battery factories you want to build. There's plenty of lithium, it's just you have to build the lithium mines. I don't really see any hard limits there eventually. Once you build all the factories, then yeah, you're pretty much ready to go.

Dwarkesh Patel (01:27:44):

Is the point you're making with the alternative batteries that even if they're worse than lithium batteries we'll have just so much energy that it doesn't matter? As in even if we lose a lot of it, that's fine, we'll just use whatever we can take. Or are you saying that we'll produce batteries with other chemistries that are as good as lithium batteries or better?

Austin Vernon (01:28:06):

Right now the shortage is really nickel. So in the very short term, lithium's kind of starting to be a shortage, but there's plenty of lithium. It won't be a problem. With lithium iron phosphate or whatever, there's a huge amount of substitution right now because it's avoiding nickel. It's not quite as good as some of the nickel chemistries, but for a lot of applications, it just doesn't matter like for a lot of cars and everything like that. And you're going to have the aircraft and stuff paying the premium for the high energy density batteries. Eventually  technologies could just use less and less materials because they're better batteries, like some of these concepts around solid state. I'm not sure if those will come to fruition and if they'll be really that much better when they do come, but I think there's lots of opportunities for substitution down the line.

Dwarkesh Patel (01:29:03):

What is solid state, by the way?

Austin Vernon (01:29:04):

Right now, all our batteries charge and discharge through the lithium ion going back and forth between the cathode and the anode. It travels through a liquid and the liquid is electrolyte, which means ions can travel through it. Solid electrolytes are a little more challenging so that's why we don't have them. You get rid of the liquid and it's just like the ion has to travel through a solid instead. The promise is it could be a much higher energy density and theoretically cheaper too, just because it weighs less and stuff, but there's all sorts of problems around how they degrade faster. Batteries have six different areas where you have to hit the requirements and if you miss one, then it's no good so they're kind of hard to improve in that sense.

Dwarkesh Patel (01:30:02):

I guess if the energy super abundance is going to come from solar and wind, obviously these are intermittent sources of energy, in that case, you would need there to be progress in the battery storage, that's contingent on that. Right?

Austin Vernon (01:30:16):

Yeah. I think that's what I mean, a lot of the extra energy uses that we talk about don't really require many batteries, if any batteries at all. The transportation, yes, you have batteries in them, but if you're going to have abundant nuclear electricity, or abundant geothermal electricity, you still have to build all those electric vehicles. You still need the batteries for that. So the extra batteries that solar and wind require over geothermal, I think could end up being pretty minimal. Maybe the way I think about it is if you can have solar, a farm that's going to give you $10 a megawatt of our electricity, you have to figure out how to utilize that. And if you do, then you'll be very rich and you'll beat the guy who's paying $40 a megawatt hour from the more expensive traditional generators.

Dwarkesh Patel (01:31:12):

Yeah. But before we get into which sources of energy are most promising, let's talk about some of the other applications of an energy support abundance. Obviously we talked a little bit about travel, but one thing that might be concerning with air travel, at least for passengers, is if the bottleneck step there is TSA and other regulations, to what extent will reducing the travel time or increasing flight speed or number of flights have an impact on how much time you're going to spend in an airport or in transit?

Austin Vernon (01:31:47):

Well, so right now, if you think about Airbus, they have this super jumbo thing. I can't remember what that plane’s number was, but none of the airlines really loved it because it's too big. It's too hard to get everyone loaded and unloaded and you really just hit this economies of scale. So the electric planes are likely to be just tiny in comparison, 10 passengers. It's easier to load and unload or you're going to fly out of smaller airports so you won't be going to this giant regional airport that just, has all the parking problems and all the security, you'll be driving to your neighborhood general aviation airport, where there's a small line to get through. And a lot of these small aircraft, under certain situations, even avoid some of the screening requirements, because they're just not as dangerous. If you have a small plane, there's only so much damage you can do with it. 

Dwarkesh Patel (01:32:42):

I did not know that. I got to start booking planes from the smaller airports or something to avoid the TSA.

Austin Vernon (01:32:48):

It's very nascent, but there's some business models that are coming down from the net jet style to a little more commercial. I think that they're trying to hit a price point that's similar to first class, but you get to avoid all the airport craziness. And I think, I'm just kind of a believer in if that existed, people would get angry enough that they would loosen up a lot of the rules. It seems impossible to change those rules now, but I think for the average person, it just costs them no time because most people don't even fly very much.

Travel in the Future

Dwarkesh Patel (01:33:26):

Yeah. Do you want to talk about your vision for what a city could look like if energy got a lot cheaper? In the paper you have all kinds of interesting projections about drones and electric deliveries and just the entire congestion of the 3D space, and I guess with tunnels as well. Well, what does the city look like with energy superabundance?

Austin Vernon (01:33:48):

Basically disaggregate the car to a certain extent, inner city car trips are less because flying's going to be cheaper and it's going to be more convenient to have the bots deliver your stuff. I love the tunnels because I don't like taking people's land so with tunnels you can run new roads and everything without imminent domaining and taking people's land away from them when they don't want to lose their land. That process is so, it makes people so angry when you take their land that it's very expensive to imminent domain people, because they will fight you until literally the sheriff has to show up and haul them away. If you can go around that with tunnels existing right away it just makes that societal cost of doing some of this stuff significantly less expensive and it's then just the engineering challenge.

Future Cities

Austin Vernon (01:34:54):

I think there's really an opportunity now there. Boring Company is the most famous, but recently I think there's another company that wants to do tunnels for electricity. They have this plasma boring machine concept–– it seems pretty crazy right now, but it's just one of those solutions where you're going to reduce the coordination cost across the whole economy and improve property rights so people should really try to build it.

Dwarkesh Patel (01:35:26):

You mentioned one of these machines in your blog post on tunneling and it was DS, SpaceX1, I forgot the name of it, but it's like this insane thing.

Austin Vernon (01:35:34):


Dwarkesh Patel (01:35:35):

Yeah, exactly. It's pretty big but apparently it's all electric, which is kind of insane, and it can just do it in one go. How is it getting the material out if you're just doing the tunneling in one step?

Austin Vernon (01:35:51):

The problem is that most of the tunneling is in soft soil and it's difficult to drill through soft soil because of the materials handling. When you first start drilling an oil well through this stuff, you actually have to limit your drilling speed and you don't even have to put any weight on the bit, just the pumping fluid around basically jets out the fluid. That's kind of what you're doing with the boring machine and the soft soil stuff. So managing the spoils, which is they have muck carts a lot of times, I think maybe it's basically like trying to do a conveyor belt, but you could also just make it a full liquid and pump it out. In the oil field we carry our cuttings in mud and we pump it.

Austin Vernon (01:36:36):

Then they have, the other big challenge is they have to keep the walls from caving in on them so that's... Current boring machines and soft soil spend enormous amounts of time erecting these tunnel supports that keep it from collapsing in themself. So it's kind of counterintuitive. It's actually dramatically faster to bore in hard rock than soft soil because in soft soil you spend so much non-productive time, whereas in hard rock, you're just blowing and going.

Dwarkesh Patel (01:37:09):

Interesting. Yeah. Okay. To get back to the cities, you mentioned something in the paper, Marchetti's constant, which is about how the amount of time people spend in transport per day is the same so if you just increase the amount of speed in which they can move with VTOLs (Vertical Takeoff and Landing) or other kinds of things, then they have a wider surface area in which they can explore. Right?

Austin Vernon (01:37:37):

Yeah. I don't know if physically the cities will look that much different, but their effective economic size will be much larger because you could live in Cedar Rapids and commute to Minneapolis with some of these technologies. Your city in Cedar Rapids still looks the same, but you don't have to work there. If you have a better job in Minneapolis, you could commute there three times a week or whatever it is, five days a week.

Dwarkesh Patel (01:38:11):

Yeah. It's super interesting. But does that imply, by the way, that if the commute time stays the same and people just get more spread out.. If energy becomes cheaper, then neighborhoods and cities kind of become this unwalkable mess out of a Jane Jacob's nightmare if the conglomeration goes away?

Austin Vernon (01:38:29):

I think it's actually the opposite. If you have tunnels and if you have some, these alternative methods to cars, then you use cars less. And I think in many cities, they never made sense for cars anyway, because they were built before cars. In New York city, you're never going to move everyone around on a car unless you build tunnels, you could then. But even then I think there's other technologies there that make a lot of sense. I think people like walkable so even though I live in a city that requires a car, some of the hottest neighborhoods are walkable neighborhoods where the neighborhood is walkable itself and then you just drive your car to wherever else you need. The car is hidden within the neighborhood.

Dwarkesh Patel (01:39:22):

Okay. It's interesting. I guess we'll see more segregation (not in the racial sense or anything) but in the sense that people will prefer to live in these walkable neighborhoods so they don't have any problem commuting to work using a VTOL or something. What you would end up seeing is these walkable neighborhoods and then industrial zones that are way far away, distance wise, but not that far away time wise.

Austin Vernon (01:39:46):

Right. And it's the same for if you want to live in a small town. Now it would be too far to commute to the city, but you could in the future. 

Flying Cars

Dwarkesh Patel (01:39:56):

Yeah. More choice. I see. What is holding back VTOLs? VTOL, by the way, is Vertical Take Off and Landing. The reason you need to go to an airport is because you need a large landing pad to take off and land. The hope is that if you could just vertically take off, then you would be able to lift off from your roof or something. Obviously we've had prototypes of this kind of stuff since the thirties. Why don't we have these widely available? Is energy the constraint, or is it something else?

Austin Vernon (01:40:26):

Well, I think in the past, theoretically liquid fuels were dense enough, but they're too complex, too expensive because when you're turning heat energy into mechanical energy, it's just a lot of weight and complexity comes with that. Some of these old concepts, you have all these engines and all that. And so if you electrify them, it really changes the game. Because it's not just batteries, it's the motors, it's the inverters, that are now getting dense enough and small enough to make sense. It takes time to get this stuff through the FAA, for better or worse. The technology hasn't been good enough long enough to get stuff through the FAA.

Austin Vernon (01:41:14):

And there are some limitations I think right now. A lot of people wouldn't use batteries and the batteries are just on the edge of good enough. You're going to have a 50 mile VTOL, not a couple hundred mile VTOL, but eventually my dream VTOL application is a nuclear power quadcopter that carries a container. So you can take the container directly from the factory in Vietnam or wherever, directly to the people who are using it or the warehouse in Arkansas or whatever.

Dwarkesh Patel (01:41:50):

Yeah. That would be interesting. Theoretically, you could have these drones that are carrying these huge payloads, weight wise.

Austin Vernon (01:41:58):

But you wouldn't necessarily want a large payload. You just want whatever the customer wants, you want to size your vehicle to deliver that payload, and that's the most efficient.

Dwarkesh Patel (01:42:12):

Oh, I see. Right. Because it doesn't need to be a shipping container or a shipping vessel where you just have it be huge. And then what does this mean for computing? If energy gets a lot cheaper, I guess if Bitcoin mining becomes, well, it doesn't necessarily become more profitable because other people's energy's cheaper too, but what are the other consequences? Is spinning up an AWS server just become trivial now and then building a deep learning model costs nothing in terms of GPU time? What impact does this have on computing?

Austin Vernon (01:42:48):

Yeah, I think the limitation would probably still be just chips for a while until you figure out a better production process for that. I think it'd be a while before it becomes energy. I think smartphones really worry about energy. There could be some interesting things with smartphones if you have a very power dense betavoltaic battery which is a nuclear battery that you don't have to worry about running down. But outside of smartphones, I'm not sure that energy is the limit for a lot of those computing.

Carbon Shortage

Dwarkesh Patel (01:43:26):

And one of the interesting things that you speculate about at the end of the paper is about a potential carbon shortage. I think in an email to Tyler Cowen, he published on his blog, you said by the end of this century, we'll have a carbon shortage because of the process you talked about earlier, the thing you're working on, if you can take CO2 out of the atmosphere. What is the probability that this ends up happening? Do you think it's more than 50% by the end of the century or is it just speculation?

Austin Vernon (01:43:58):

I think it's extremely high that it happens and it's harder to put the timeline on it. By the end of the century might be a little, I think I ran some numbers in there and if you 10X current plastic production, and then you're just land filling it all, I think it was a little over a hundred years to get, and you're assuming you're out, the rest of your carbon output is zero in that scenario. But it's probably pretty hard to do it in a hundred, by the end of the century without a lot of growth, but it's kind of the exponential thing that can get you where... I think some large number of the carbon emissions have happened in the last 20 years, and it was very small before 1950. You could kind of get surprised at the back half, the last 10 years it goes crazy. It makes it hard to predict.

Dwarkesh Patel (01:44:53):

By the way, in Will Mcaskill’s new book on longtermism, one of the things he speculates about is if society collapses and we need to restart, one of the things we'll need is coal or some other sort of dense, easy-to use-fuel. And the problem is we've been burning up easily accessible coal, like coal in places where we could just dig up and find it, and so one of the things he's concerned about is making sure we leave some easily accessible coal silos around so that in case society collapses, we can restart and use these to power up a second industrial revolution. I wonder if you could use a process like this with carbon sequestration to actually just build up these kinds of reserves. I don't know if a long term, if somebody's really interested in making sure we have that kind of resource. They could just use this process to... Is that possible? Or

Austin Vernon (01:45:51):

Actually there's a company called Charm Industrial, and they're basically doing that because they take trees and they do a process called fast pyrolysis. It's where you burn biomass without oxygen, anoxic environment. And it makes this bio-oil and then they're injecting the oil down into wells and selling carbon credit. So it's already happening, you could say.

Dwarkesh Patel (01:46:15):

Oh wow. And that’s easy to burn and stuff? 

Austin Vernon (01:46:21):

Yeah. If you just want to burn it for heat, it's okay, but it's hard to refine. There are a lot of people that tried to do bio-oil as an alternative for petroleum 20 years ago, Cleantech 1.0, and they all failed. So it makes me laugh that they're reimagining the process to sell what are right now very expensive carbon credits. But you could do something similar. Actually you could do something similar just to make straight carbon and stuff if you wanted to.

Dwarkesh Patel (01:46:50):

Okay. I see. The thing that I find interesting about this is that often in the case of global problems, people will early on identify that a thing is going to be a problem, but it often ends up being the case that they get the direction of the problem opposite. If you think about population, in the seventies, people were correct that global population was going to be a problem. The thing is, it seems like now the problem is going to be that the population might decline too fast, not that it's going to grow exponentially. And I think this is another example of this kind of thing where CO2 is going to be a problem either way. It just, I'm not sure if it's going to be a problem where we'll over reproduce it or we'll have shortages.

Austin Vernon (01:47:36):

Yeah. If you just think about the large scale, if you're going to be like Kardashev, whatever scale civilization, where you're using condensed amounts of energy, that's going to have side effects and you're going to have to figure out how to manage that one way or the other. One of those is eventually earth may just be a nature preserve and we all live in space or something.


Dwarkesh Patel (01:48:01):

Yeah. Okay. Let's talk about nuclear energy. It seems like you're much less optimistic about nuclear energy than you are about solar and wind. Do you want to explain why that's the case?

Austin Vernon (01:48:15):

Yeah. Well, especially solar more so than wind. Wind, I think it's limiting because it's transmission problems and again, if you want to build out huge amounts of wind like some of these zero carbon plans call for, you're going to have to take a lot of people's land to build transmission lines and stuff. Which really pisses people off and they fight hard and it becomes expensive. It's not like... The wind turbines are relatively easy to sight because you pay people and you'll actually see, they never put above ground power lines on the people's land, where they put the wind turbines. They're always underground so at least they get to the county right away. But when you get these giant transmission lines, like Grainbelt or something, they almost inevitably have to go across a lot of people's lands and you can't just stuff them all in county and state right aways because the pylons are so big.

Dwarkesh Patel (01:49:12):

Sorry, what is the pylon?

Austin Vernon (01:49:13):

The pylon is what holds the wire up, the tall tower. Yeah, solar is, it's much more flexible where it can go. I think when it comes to solar getting cheaper, the obstacles are just pretty simple. It's like, "Well gosh, it's expensive to put all this racking. Why don't we just lay the panels on the ground?”Gosh, this glass we're encasing with is getting expensive and we don't need it to last 80 years or 50 years. We can just put some plastic on it instead." We've gotten these, the actual materials and stuff so cheap and all the other labor and stuff is getting more expensive. Well, why don't we just add another layer and make more energy? Those are your solar solutions to get down to $10 a megawatt hour, and they're pretty straightforward.

Austin Vernon (01:49:57):

Whereas nuclear is like, "Well the light water reactor can't get us there. Let's instead cool our reactor with sodium, which catches fire when it, it explodes when it reacts with water and catches fire when it reacts with air." Or there's, you could cool it with liquid lead. That's an option. Helium, which leaks a lot. Or you could do molten salts that corrode everything. We don't really have anything like that, and so I think when you start looking at, this is for large reactors. So I think those solutions for very large reactors are pretty hard. It's pretty difficult. And there's a lot of reasons why. Why did we make these weird choices? Well, a lot of stuff just reacts poorly when you expose it to neutrons and stuff. They each have their own features that make them possibly good candidates.

Austin Vernon (01:50:57):

I actually think regulation is actually, it's oversold a little bit. And I think actually to the extent that if people were internally consistent, then they would see NRC as a regulatory success story because, the background on this is my wife's mom and step father are nuclear engineers that have worked at all levels of the nuclear power industry so I get to ask them the general questions and learn a lot about it, which is nice. It's very helpful for learning about it.

Austin Vernon (01:51:33):

But there's, back in the eighties, the nuclear power industry was in real trouble because their competitors in coal and natural gas got deregulated. Most of the cost of coal is the rail getting there, and the rail industry got deregulated and then the national gas industry got deregulated. So the cost of their alternatives was falling and they had to build more safety into their plants because of all these, it wasn't just three miles out. It was Browns Ferry. It was Rancho Heco, all these things that could have been really scary and, to a certain extent, we got a little bit lucky that we didn't have a worse disaster. They were just relatively limited accidents at their sites.

Austin Vernon (01:52:25):

There was actually a time where nuclear power plants were selling for less than what their fuel was worth in their plant, around there. What the industry did and what NRC did is they moved to probabilistic risk assessment, which is usually the gold standard. People are really happy that we use probabilistic risk assessment for commercial crew, with NASA and SpaceX. And they want the FDA to use more probability, more expected value. What this allowed was it's basically you're rolling up some of the rules and moving it into the risk assessment. Around 1980 nuclear power plants only ran about 60% of the time. They weren't very reliable. They had all sorts of unplanned outages, stuff like that. The safest mode of operation is just running as designed. The more consistent nuclear power is, the safer it is. The probabilistic risk assessment allows you to do repairs while you're running, which was kind of discouraged before. It'll be like if your main cooling pump is leaking, before you'd be like, "Oh, gosh, I hope we can make it," and then eventually it just fails and you shut down the reactor. And now it's like, "All right. Well, we have backups. The safest thing to do is actually repair it now while the plant's still running and then get it repaired and put it back online." Not only, to give you the idea of the safety standards that NRC has, I think for the nuclear plant taking damage is one time in 10,000 reactor years. And then for a large release it's one in 100,000 reactor years, and there's 93 operating reactors that––less than 93 sites. We should only see a three mile island under the current standards once every hundred years or so. A large release, like a Fukushima type situation, should only happen once every thousand years.

Austin Vernon (01:54:37):

But they had, just in a few years in the 1970s, the industry had three or four of these damage events at least. I don't know how many officially count, but probably at least three. The safety is incredible. And now the operating capacity is up to over 90% so the plants are just extremely reliable and it lowers their costs because their costs are so fixed. When you compare it to a country like France? They've had a lot of reliability problems with their nuclear fleet in the last couple years. This year, their capacity factor I think is barely over 60%. We have 90 gigawatts of nuclear. They have 60 gigawatts. That makes a huge difference for Europe that those plants aren't running full out. And it is really, you see a lot of charts about, "If Germany didn't shut down its reactors, what would the energy balance be?" But you don't see as many, "If the French could run their reactors like American reactors, what would the energy balance be?" I could go on about how that integrates into new plants, if you want to know about that.

Dwarkesh Patel (01:55:46):

Yeah, I do. Because the line I've always heard on this from my bubble is like, "Oh, they haven't approved a new plant. The NRC has not approved a new thing since, a new plant, since it was created." I guess they just approved the design for the new, small modular reactors, which I guess I'd love to hear your opinion on as well. I'm very curious to hear this perspective.

Austin Vernon (01:56:07):

Well, okay. So think about it, you're in the 1980s, you had new sources of fuel, you had new competitors, you also, by the end of the decade, increased the amount your nuclear power plants ran by a lot. So a lot of these new power plants that people were thinking about building were at existing sites, like an extra reactor at Watts Bar or whatever. And well, you basically just got like a buy two, get one free, by running your plant better. So you don't really need them as much. So all those contributed to just not making sense to build new nuclear power plants, because the existing fleet ran better, and more competitors, and electricity demand slowed down.

Austin Vernon (01:56:50):

Is it hard to get through NRC approval? Yes. That last one, the mini reactor you're talking about took 10 years or something. But when you think about a probabilistic risk assessment, no one ever says, "Well, gosh, NRC's current standards of a large release, which would basically happen one every thousand years... We're not arguing over that. We're just talking past each other, I guess, instead.

Austin Vernon (01:57:19):

So to me, that's a pretty reasonable risk level. If you're going to 10x your reactors,  that means almost certainly you'd have a Fukushima within your lifetime if you go with NRC standards. But it actually turns out that it's pretty cheap to do way better. A lot of the reason why the plants weren't built may not necessarily have been because of regulation, but because the market conditions changed. You had more competitors, and the coal with the gas, being deregulated. And then you also had increased production from the existing nuclear plant. So if you're going to build an extra nuclear plant or an extra reactor at an existing site, then you might not have needed to anymore because you got so much more production out of your existing plants. Stuff like shortening the fueling time and all-around improvements, paired with electricity demand flattening is what really made new plants not economic or not necessary.

Austin Vernon (01:58:27):

When we really think about the probabilistic risk assessment, it just takes a lot of engineering time to get it through. If you look at how hard it was for SpaceX to get Falcon 9 and Dragon through NASA's loss of crew risk calculations… it took years, it took hundreds of millions of dollars. So it's kind of funny that people see that as a success. Especially when the stakes were only a few lives for people that volunteered for danger. And then you have a nuclear power plant, where we're going through the same problematic risk assessment, and it could impact many more people's lives. That's not good enough. So I think it would make more sense to argue about the risk factor, how much risk we should take with the actual numbers, as opposed to just like, "Oh, I'm mad that we're not building nuclear power plants."

Austin Vernon (01:59:28):

Actually it becomes very inexpensive to improve the risk probabilities, because the old plants that we're running now have active safety systems, which means you have to maintain them and they have to work. So if you want to move the control rods back into the reactor, well, there's like a mechanism and a motor that does that that can fail. So when you're calculating your risk, you have to calculate, "Oh gosh, what if this motor fails? Or what if my control fails? Or what if I don't have a properly trained operator to do it?" And it's the same for the cooling systems. But this new generation of plants, they have passive safety systems where natural convection can cool a reactor in an emergency, or... The rods are more like a dead man switch, where if something happens, they just drop in from gravity. And so the new power plants, like this one that just got approved, or the one they're building down in Georgia, can be orders of magnitude safer than the running plants.

Austin Vernon (02:00:29):

It's not really a huge cost increase. You're just changing how you do these things. In fact, if you look at all the literature, they're actually supposed to be less complex and easier to build. But you're talking about a project that is so complicated, it takes thousands of workers, years and years to build, working every day. And it's like, if you're going to go through and do the engineering in great detail to prove that you're playing it safe under the problem risk assessment, it's going to take hundreds of thousands of hours of engineering time. That's why you see... I mean, why are investors willing to pay that at this point, after you build it because of the rank and cycle and all that, like is it going to generate economic power? The fact is it's not necessarily going to. I think one way to think about this is my father-in-law.

Austin Vernon (02:01:27):

He always says, when people ask him about why we're not doing more nuclear, he says, "Well, you have to think about the politics first, and economics second." Those are the important ones. So people are submitting designs to build plants that are big enough to impact lots of people's lives, even if that risk is very low.

Austin Vernon (02:01:47):

Some people still are bothered by that. But also, they're selling an easily substitutable commodity in most cases. I think a lot of times on the political side, if you can substitute nuclear power, people will. Even if it's coal or whatever, people don't really care that much about the emissions. They just care about their electricity turning on. And I think you see the opinion change very fast when nuclear power is no longer the substitute. All of a sudden Germany's like, "Well, we could turn our reactors back on." Or Japan, same way. They've had these reactors off for years, but now that there's an energy crunch, they're like, "Well, let's turn em back on."

Austin Vernon (02:02:25):

So I think the future for nuclear power, which would be a better future, is you create products that impact less people's lives, or have the potential to impact less people's lives, and also are not substitutable. I think that means small reactors. If you have a battery that can power your phone or have a little battery out in your garage that can power your house, these are hard to make. There's a lot of problems, especially in power density–– nuclear is very energy dense, but not necessarily power dense. So you have to do a lot of work on that to get there. But one of the most exciting examples of recent nuclear technology is these people at some national labs and NASA got together and created this KRUSTY reactor, and only to, it's only one kilowatt. So it's small. I think the thing weighs 400 kilograms. It fits in the room. But they got the whole project done in a couple years for less than $20 million. And it worked great. It's very safe, just because, partly because it's so small, but it has almost no moving parts. The whole thing has a Sterling engine on top, and that's the only moving part. 

Austin Vernon (02:03:19):

So it's really... And there's several startups now that are working on improving that technology and commercializing it. So that's the kind of nuclear stuff that when I talk about small nuclear, micro nuclear, is really exciting to me because it has so much potential. And when you start putting nuclear in that small form factor, there's no other energy source that can compete with it on energy density. So you can do things you could never do before, whereas selling to the grid in a large power plant is like, "Well, I can do that in lots of ways." And if you think through this lens, then you see the entire nuclear debate is the nuclear proponents trying to claim that nuclear is not substitutable, and that we should pay more, accept the risk, or whatever. And maybe we should, but it makes it hard to promote that technology. If you could have a phone that you never tried to charge, people would love that. It'd be like, "I don't care if that's nuclear. I just have a phone that never goes dead."

Dwarkesh Patel (02:04:41):

I guess the question is: is the lengthy and expensive process necessary for the probabilistic risk assessment? If there's a way you could just have the process not be more streamlined, and be as effective in evaluating the harm. I guess another thing is if we haven't seen... Zero people, or very few people have directly died from nuclear, right? So is it just that we've gotten lucky or you're saying that could have been way more, and we're just in a lucky timeline?

Austin Vernon (02:05:14):

I guess I'll go backwards a little bit here, on answering those questions. I think what people are responding to is even if Fukushima didn't have airborne radiation (that was very dangerous) people still got removed from their home. And there's a lot of costs for associated with that. It's hard for me to believe that if we had a similar thing in the US, there wouldn't be some type of mandatory evacuations that were really unpleasant. If you could get your power from coal or natural gas, without that risk, I mean, a lot of people would make that trade off.

Austin Vernon (02:05:54):

And I think the other thing with Fukushima is, as I understand it, because it was on the ocean with fast currents, they were able to use a lot of seawater to keep the reactor from getting too out of control. But they were just dumping a lot of the radioactive stuff into the ocean, but it was dispersing quickly, it wasn't a big deal. So if you're on a freshwater reservoir like most US power plants are, your risk equation might have been different there. I don't know enough about it to know if that really matters. But I think the main thing is because of the precautionary principle, people are still going to get removed from their homes and people don't like that.

Austin Vernon (02:06:39):

Let's see. I'm making it... I mean, you can always streamline processes, but the thing is, people are submitting designs that are extremely complex. So whether your design is ultra safe, or not safe at all, to do all the engineering to prove that costs about the same either way. So that's part of why these new plants are so much safer than the NRC standards. It's just not that hard to make them that much safer. And you're going to spend the same engineering resources no matter what, based on your plant complexity. So that's the difference why KRUSTY was able to go through so fast. Their thing is very simple. They don't have very many moving parts. There's only so many things that can go wrong with it. I think that's what's exciting to me about these other startups is they have the potential to get through faster, with less money, and then there's real markets and remote power space, military, where people are willing to pay the premium for these initial models.

Dwarkesh Patel (02:07:43):

Okay, I see. Okay. So you're not bearish on nuclear in the future, given the new designs with passive cooling and stuff like that. It was more like the old designs that you're pessimistic about, is that correct?

Austin Vernon (02:07:56):

Yeah. I mean, if you look at what the cost of electricity is going to be from that, if they ever build the reactors that just got proven or just got approved, it's quite expensive. I think usually it's around like $40 - $50 a megawatt hour, best case, but more likely it could be up to like $80 a megawatt hour. So they're not building it in deregulated power markets, because you lose money. But there are places where it could make sense. Some places, like in Europe, have very expensive electricity.

Dwarkesh Patel (02:08:31):

And Japan and Singapore, and there's a lot of other places that are...

Austin Vernon (02:08:34):

Yeah, yeah. So there could be some markets in there, but that technology then still has to compete with those places building solar panels or all these other technologies that you could do. Then there's the whole argument, well, nuclear can do this and that, but I think the people building the reactors clearly don't want to build them in deregulated power markets because it's not economic. That's why I'm excited about the small, because there's alternative markets other than selling this substitutable commodity that's very cheap.

Dwarkesh Patel (02:09:08):

Have you talked to Eli about this? What is his opinion?

Austin Vernon (02:09:12):

Yeah. So Eli finds out about these new startups that fit this bill and sends me the information on them because he knows I'm excited about it. His specialty is governmental affairs. So I'm sure there's still lots of opportunity to improve the process at NRC. Recently INPO, which is the industry group that's very much a German style industry group, it's very powerful, their goal with NRC was to reduce the nuclear rules by one third. And then you also have NRC writing new standards for gen-four reactors that's supposed to be done in a couple years, but Congress instructed them to do it. So there's lots of opportunity to try to improve the process, but it's very complex. I'll give one example, the Browns Ferry accident. The main thing that came out of that was, you can't have control cables for safety systems on, redundant safety systems on the same cable tray, because that cable tray catches on fire you lose both systems.

Austin Vernon (02:10:14):

So it's very, very expensive to run extra cable trays and all this cable separation. That's actually one of the problems that's delaying Vogtle and Georgia right now–– they had 500 issues of safety systems sharing the same cable tray. So they have to build all new cable trays and clean out the mess of the stuff they already built and redo it. Super expensive. So the NRC tried a pilot program where they did a performance base safety as opposed to just the strict cable separation rule, and I think Oconee was one of the power plants that tried it. It ended up being more expensive than just the simple rule. So the reality is often very complex. I think when you have these complex plants, it's just hard to do. So it can always be improved, but I think the small could end up greatly out competing the large because they have less complexity.

Dwarkesh Patel (02:11:18):

You had a small section in that piece about fusion where you were especially pessimistic about fusion. Well, what is your take on fusion?

Austin Vernon (02:11:28):

Well, it's kind of the same thing. I'm not pessimistic about fusion, I'm pessimistic about fusion technologies that heat up water to make steam, and run it through a steam turbine.

Dwarkesh Patel (02:11:37):

Because they're not efficient?

Austin Vernon (02:11:38):

It's just so expensive. Literally just pitting in the steam turbine and the condenser and all that kind of stuff you need for that, basically makes you uncompetitive on most deregular power markets.

Dwarkesh Patel (02:11:53):

Yeah. So I mean there's startups who have plans to do direct energy conversion. I don't know how feasible those plans are, but yeah, presumably you think those are... In those cases, do you think fusion could have a big future?

Austin Vernon (02:12:06):

Yeah, yeah. Again, I don't know too much about... The same as you, I don't know too much about their specific technology, but if you're pursuing a direct conversion technology, you actually have a chance of success. I think a lot of people I've talked to in the fusion space, they're like, "Well, I can make electricity for $50 a megawatt hour. And because I'm fusion, people should pay me $50." And it's like, well, not everyone may want to pay you $50.


Dwarkesh Patel (02:12:34):

Yeah. Yeah. I mean, it might involve an initial period of large subsidies that we had to give electric vehicles, and even solar, we had to give huge subsidies to solar in the beginning when we were at the beginning of the learning curve, so... That might be necessary though.

Austin Vernon (02:12:50):

Yeah. I mean, I really disagree with the subsidy solars had, actually. And I think it, just like if you actually look at the numbers, it proves the point. People say like, "Oh, because Germany did the feed in tariffs, it made solar cheap." So if you had a country that's 1% of the population, they spent a tiny portion of their GDP and that was enough to scale the technology, well, you should just let some other fool do that, and reap the benefits. So I would be supportive of taking away most of the subsidies for energy in general.

Dwarkesh Patel (02:13:21):

Just to make sure I understood that argument, you're saying that it's unlikely that the small subsidies that Germany gave were enough to actually make the difference? Is that what...

Austin Vernon (02:13:29):

Well, I'm just saying, if it took such a small amount of subsidy to do it, someone will be foolish enough to do that. In this case, it was Germany. They spent a lot of money doing that. They're not reaping the benefit from.

Dwarkesh Patel (02:13:43):

Yeah. It's not compatible with their environment, so... And their climate I mean, yeah.

Austin Vernon (02:13:48):

We benefited from them doing that. We still do spend some subsidies on solar. And I think they're very poorly designed. So it would be better just to get rid of them. But the thing with fusion, if you're just heating up water to make steam, it's that technology, there's no learning curve anymore for steam engines basically, because that technology's so mature. So that's why some people are looking at super critical CO2 cycles, because, well, maybe this could be a little cheaper than doing steam turbines. That's some possibility there. And there's some other technologies that, maybe someday you have thermoelectric generators and stuff like that. But I think the direct conversion technologies have just a massive advantage. Not only an initial cost, but in ongoing operating cost.

Alpha and Efficient Markets

Dwarkesh Patel (02:14:39):

Okay. Okay. There's one more topic I really want to talk about, which was... Yeah, you have an interesting post on where you can actually expect to find alpha, given that at least public markets are efficient? Do you want to explain the basic thesis of that post before I ask you specific questions about it?

Austin Vernon (02:14:58):

Yeah. If I was going to dumb that post down, I love Fama’s original paper where he lays out this efficient market hypothesis thesis and he is like, "There's multiple types of information." And so the first is, if you just have pricing data for stocks or whatever securities, you can be the smartest person in the world, and you're not going to make any money doing that, because it's just random.

Austin Vernon (02:15:22):

But if you start incorporating more information, what's in 10Ks and all that, if you're super, super smart, you might be able to make a little bit of money there. We see that with Renaissance Technologies, and you can debate about Warren Buffet and all that. But then there's the third category, which is the strong type information. And it's basically you have legally acquired private information, and you can make money that way and be significantly less smart. So if you want to just take Fama's paper, how do I make money? It's like, "Okay, well I should find legal ways to acquire this information, and then I don't have to be super genius to make money on it."

Dwarkesh Patel (02:16:11):

What I thought was really interesting in your post was you had this point about how one of the ways you can actually earn excess returns is through labor, right? Buffet, in the earliers at least, he would go into these factories and interrogate every single piece of operations and whatever. I thought it was an interesting twist on Piketty's thesis. So I don't know if you've seen his stuff, but he has this claim that, well, not only does capital earn more than... The gains to capital are higher than the gains to labor, but the more capital you have, the higher returns you can earn.

Dwarkesh Patel (02:16:49):

I guess Harvard has access to hedge funds that may be able to earn excess returns. Basically if you take this view, it's basically the inversion of Piketty, because over time as Buffet has gotten wealthier, his returns have gone down, because it's harder to invest the marginal dollar more effectively. As you said, with the medallion fund, yeah, they no longer accept outside money. Then the interesting thing about labor is, the reason that Buffet was able to earn those excess returns in the beginning, was because of the labor he put in. So the interesting thing is, capital is just fungible with other capital. So capital doesn't enjoy as high returns as really good labor, really smart labor, which is the opposite of the Piketty thesis.

Austin Vernon (02:17:33):

And I think, there was actually a paper, I think it was on marginal revolution a couple years back. So I'm pulling from my memory here, so I could be missing a little bit, but basically it studied all these businesses and what happened to the business after a founder unexpectedly died and the profit just... It looks like these are capital returns so many people would see them, but then the earnings just drop like a rock, because they lost some irreplaceable human capital there, and they didn't spend any time training them because they died unexpectedly.

Dwarkesh Patel (02:18:10):

Right, which also has an interesting implication for CEO pay, which is just that... Actually, okay. In the Marxist sense, what is pay? It's like, your pay is what it costs to replace you right? And if Steve Jobs was so irreplaceable that if he goes away, earnings are going to drop like a rock, and stock prices are going to drop like a rock. Actually, that means that he should get paid... That's how expensive it is to replace him. He may be irreplaceable. So it's actually worth whatever dozens of millions of dollars you're paying him.

Austin Vernon (02:18:43):

Yeah. Yeah. I'm generally a proponent for letting the market decide that.

Dwarkesh Patel (02:18:49):

Yeah. Yeah. Okay, and then another way you suggested was that firms could earn excess returns by developing a unique brand. So YCombinator is probably able to earn excess returns through normal venture capitalism because of their unique brand. Yeah, I thought that was really interesting. Do you want to talk to me more about that?

Austin Vernon (02:19:06):

I think it's just a lot of this intangible capital and labor are complements for regular capital, and I think you can see it too. If you build a brand around that, you're a good investor. You can raise money from other people and charge the money on it, more so than if you're just a no name. So I think there's lots of examples of that where building a brand or building relationships is extremely valuable, just like specific knowledge can conduct your returns. I mean it's a type of specific knowledge.

Dwarkesh Patel (02:19:45):

Well, what do you mean a specific knowledge?

Austin Vernon (02:19:48):

Well, I mean to build a brand like YCombinator, you have to understand what tech founders want. So they use that knowledge to create a place that's great to go do your startup.

Dwarkesh Patel (02:20:02):

Yeah, yeah, yeah. Interesting. Is the market for blogging efficient? So now there's actually financial rewards to blogging. The Effective Ideas blog prize, there's other kinds of grants like this. Recently they opened a contest where you can win many prizes where it seems like if you're really good at blogging, you could earn six figures, it seems. Given the regularity in size of these prizes, is this a market that we should expect to be efficient?

Austin Vernon (02:20:34):

I think it would be hard to measure. Given my own experience, I'm blogging for free, but the benefits I've gotten from learning about what I'm blogging about, and then the few connections I've made that have helped me with my projects I'm working on. There's huge returns. If my project's successful, these returns could be just almost immeasurable. So yeah, I would guess it's very hard to measure and probably inefficient in that more people could blog, because it's hard to predict the returns to what your blogging might have. But I guess if you're going to do these blog prizes, I don't know if the blog prize... Because the blog prizes are about specific topics, I don't know how much that helps the efficiency there.

Dwarkesh Patel (02:21:20):

Yeah. Yeah. Let's take that part of it out. Let's just talk about the factor you mentioned.. This is a regular thing you hear from people who write online, which is that the gains they get are huge and that's also the case in my case. So it's kind of interesting. Efficient Markets don’t.. just because the stock market is efficient, that doesn't mean that everybody will put their money into the stock market. That's not the implication. But the question is, given that you're writing something that's high quality, will it get noticed by the market? Will it get the attention and broadcasting that it deserves? And in my experience, actually, I guess this was the case. You mentioned that when some of your first posts ended up on Hacker News. So in that sense that market was efficient. But yeah, it seems to me that when somebody finds a good blogger, it's not hard for their initial post, or at least their subsequent post as they get better to gain an audience.

Austin Vernon (02:22:18):

Yeah, I do think that. And I don't know what the counterfactual is. We don't know about the people that didn't have posts go to Hacker News, so it could've easily been... I mean, I think that what the alternative for me is, I just would've blogged way less, if one of those early posts hadn't gotten more attention. So yeah. It's hard to know what the counterfactual is.. how many people have just abandoned blogs after writing three posts. They would've written one more. Maybe it would've been better.


Dwarkesh Patel (02:22:49):

Yeah. Yeah. Okay. Awesome. This has been a lot of fun. I think we're two hours over at this point, so thank you so much for your time!

Austin Vernon (02:22:59):

All right. Thank you.

Dwarkesh Patel (02:23:00):

I don't know if you have any other final thoughts or any other subjects that we should hit on or...

Austin Vernon (02:23:06):

No, I think we covered everything.

Dwarkesh Patel (02:23:09):

Okay, cool. Awesome. And then just people can find you at Austin Vernon dot...

Austin Vernon (02:23:15):

Dot site.

Dwarkesh Patel (02:23:16):

Okay. AustinVernon.site. And then your Twitter is?

Austin Vernon (02:23:20):

I think it's Vernon3Austin.

Dwarkesh Patel (02:23:22):

Okay. And it'll also be in the story description, but yeah. Okay. Yeah. Thanks so much for coming on, man. This was a lot of fun.

Austin Vernon (02:23:28):

All right. Thank you.

Dwarkesh Patel (02:23:31):

All right. I hope you enjoyed that episode with Austin Vernon. This one was a lot of fun. If you did like it, I would really appreciate it if you could share the podcast. Put the episode in your group chats, put it on Twitter, share it to friends who you think might enjoy it. That kind of stuff helps out more than you would imagine. I want to give a special shout out to Sonya Gupta for her help in prepping me for this episode. Austin is, as you can tell, really smart and technical, so I would not have been able to prepare for this episode without Sonya's help. She's got an amazing technical mind, and did a ton to help me out with preparing questions and doing research into the topics we talked about. I also want to thank my amazing editor and producer Graham Besolou for the work he puts into this podcast. All right. See you on the next one.