Today I'm chatting with Alex Imas, who  is Director of AGI Economics at Google  

DeepMind and Professor of Economics at the  University of Chicago, and Phil Trammell,  

who is Head of Economics at Epoch  and research scholar at Stanford. 

In general what I want to understand in this  interview is what economics tells us about  

what we can expect in a world with more  and more automation and more advanced AI. 

I want to understand what that tells us about  what will happen to wages and the labor share,  

what the best way to tax and redistribute  the wealth generated by AGI will be,  

and what kinds of things will be scarce. What is scarce tells you where  

the value will accrue. I want to start there. 

What are some plausible  candidates of what will be scarce? 

Something like the relational  sector, which is defined as  

services and goods where the fact that a human was  in the loop is part of the value of that product. 

Because humans are naturally scarce, if we have  automation where a lot of other things stop being  

scarce, we will still have scarcity in the things  that humans are involved in and in the loop for. 

I'm curious to understand whether humans  doing services for other humans can ever  

be a big part of the economy. Here's maybe one intuition pump. 

In a world where AI can physically do anything  humans can do, there's this whole machine  

economy where they're building factories and  doing research and coming up with new ideas. 

Humans may or may not be involved in the physical  production of those things, but probably not in  

the ultimate limit, if robotics is solved. If you don't care about humans being involved  

in that process, why would they be? But then there are these other things  

you point out where we actually do want  the ballerina or the barista to be a human. 

That's part of the value of  going to a cafe or a performance. 

But only humans have that preference. So there's this human economy where humans  

are doing services for each other, and part  of their wealth is flowing to other humans. 

But part of their wealth is also flowing out,  because they will want some of the automated goods  

this machine-only economy is creating. This is not a closed loop. 

A lot of things in the machine-only economy are a  closed loop because the machines don't care about  

getting the human barista to make them a coffee. Within that model, isn't it intrinsic that the  

human-only economy will become  a smaller and smaller share? 

I would like to pitch a  rephrasing of that question. 

My view is that the individual forecasts  economists like us would make, as individual  

forecasts, are not necessarily very useful. There was a blog post by Andrey Fradkin,  

Brian Jabarian, and Andrew Koh that came  out yesterday looking at economists'  

forecasts about the labor market. What they found is that there's a  

ton of disagreement in every single direction. What they advocate for, and I'm in agreement  

here, is that rather than thinking about  individual forecasts, we should be generating  

prediction markets where you get aggregate  forecasts and wisdom-of-the-crowd effects. 

The reason I think this is because we have  been famously terrible at forecasting. 

Let's go all the way back to 1820. This debate we've been having is  

actually 200 years old. David Ricardo is one of  

the classical economists, not neoclassical. When the Industrial Revolution started happening,  

he wrote a bunch of stuff saying, "This  is going to be great for everybody. 

Prices are going to come down." But then he turned around and said,  

"Wait, I can see all these jobs that are creating  value are going to be automated by these machines. 

This is going to be really bad. Everybody's going to become unemployed,  

and there's going to be political unrest." And if you look at Ricardo's predictions,  

they're actually right. All those jobs that made  

money in Ricardo's time got automated. If David Ricardo woke up and somebody told him all  

those jobs did get automated, and then asked him,  "What do you think the prime-age employment rate  

is in 2026?", I think he’d be surprised to be told  it was the highest it's ever been other than 2000. 

We have the highest number of employed people  that could potentially be employed since 2000. 

That was the peak and now it’s  the second peak basically. 

What David Ricardo ended up missing is that  you have these economics of structural change,  

where everything that got automated became cheap. People had more money to spend, and then they  

started spending it on services. This is the lump-of-labor fallacy. 

David Ricardo didn't consider  that new jobs would be created. 

But it's not obvious that  money would go to services. 

Why wouldn't it go to more automated  goods and something like that? 

I'm not using this anecdote to say this is  what's going to happen now and that we're  

going to have full employment. I'm using it to say it's  

really hard to make predictions. What may be a really useful tool that economists  

have is to instead start with a premise. Maybe we start today: labor share is zero. 

Labor share has gone down. What could possibly explain this? 

Let's write down an economic  model of what happened. 

Phil will talk about this later today. Or you can write down a model that asks,  

"What if labor share just stays the same? What can make that happen?" 

If you don't take anything else out of this  conversation from me: We don't have any data. 

I've been saying we need a  Manhattan Project for data. 

We don't have data on  consumer demand elasticities. 

We don't know what they are. We're not really tracking what  

jobs are getting created or destroyed. The O*NET database, with all of the tasks  

and different jobs, has been rarely  updated and is super low quality. 

What is really useful is to think about the  potential scenarios, map them out, and say what  

dimension of scarcity will generate each scenario. If there's full employment, we can talk  

about the relational sector. If the labor share collapses,  

we can talk about other sorts of scenarios. That will tell us what data  

we should be collecting. It's probably worth defining  

labor share and capital share real quick. The whole economy, the total sum of goods  

and services sold, is either paid out to  people in wages or it's paid out to capital,  

which is to say, there's rent on buildings and  shareholders of companies that get paid out. 

For many hundreds of years, ~60% of the economy  basically gets paid out to humans in wages,  

and the other 30-40% gets paid out to people who  own machines and land and claims on companies. 

The question is, if 60% is going to  wages right now, does that shrink as  

AIs get smarter and better? This is a Kaldor fact. 

We should stress this. It's incredibly surprising  that it's over 60% after the Industrial Revolution  

and all of the automation we've ever seen. Some people are worried it's an accounting  

error that it's been so constant. There's even a controversy right now. 

Some might say labor share has been  falling in the last 20 to 30 years. 

But there have been a lot of accounting  changes in the last 30 to 40 years. 

For example, Atkinson has a paper showing that if  you keep the accounting constant over the years,  

labor share hasn't even fallen ever. But it's not that surprising, right? 

Phil, you made this point that if  labor and capital are complements,  

you need both to do anything. It would make sense that you'd need  

to pay both of them to get something done. You have had stuff be completely automated. 

There's a sense in which nothing  has yet been completely automated. 

Look at the network-adjusted  factor shares of a good. 

Look down the supply chain and not just the  final step and how much of that is done by  

capital and labor, but what went into the  machines that can automate that final step. 

You'll find that labor is adding a  lot of value down the supply chain. 

Computer and electronic products  in the US have a very stable  

network-adjusted capital share of around 50%. It's not 100%. I do think there's this qualitative  

shift that I think we agree is coming, which  is that there will be at least some goods whose  

network-adjusted capital share goes to one. The whole supply chain can be automated,  

and there's no part in it that we care  intrinsically about having a human do. 

That will be a qualitative shift. Interestingly, the implications of that shift  

for the overall capital share are ambiguous. Let's say we've got two sectors:  

the human-intrinsic sector with the  ballerinas, and everything else. 

Right now, everything else has been  scarce because of the lack of labor in it. 

But if we fully automate the supply chains for  everything else, and we satiate in everything else  

really fast, then the quantity of everything  that's not a ballerina goes to infinity,  

but the marginal utility in that stuff goes  to zero faster than the quantity is rising. 

I also want to move away  from the ballerina example. 

The point I was trying to make in my  post—working backwards from a particular  

scenario—was that the ballerina and the  performer are the wrong reference class. 

Right now we have a lot of jobs  where you have different tasks. 

This is the task-based model of jobs. Take a doctor, what is their job? 

They're filling out insurance documents. They're going and calling  

different pharmaceutical companies. One of their tasks is to see the patient and talk  

to them, but that's not the main part of the job. You could have a job and a service or a good be a  

product of different types of tasks, and  you can automate a ton of those tasks. 

If the consumer is willing to pay more  for a product or service where every  

single task is automated except for that  one part where the doctor is delivering  

the diagnosis and providing support, we would  call that job part of the relational sector. 

People are willing to pay more for the  human to stay in the loop in the job. 

We don't have data to say, "Here  are relational jobs, here are not." 

You literally need to collect  data of the following sort. 

Do a conjoint analysis of your willingness  to pay for this service or good. 

Here's the counterfactual where  everything is produced by machine. 

Here's the counterfactual where this  one task is not produced by a machine. 

What is your willingness to pay? What is your elasticity for the  

human to not be in the loop? If I don't have that data,  

what prediction am I going to make in this story? Isn't there another point, which is that there  

are a lot of fully automated  goods that don't even exist yet? 

And you can't collect any data right now  about, say, how much people will want to  

keep buying more and more of some drug that makes  you healthier that is fully produced by the AIs. 

Absolutely. That's kind of Phil's point. You could have an increase in variety in  

capital where you don't get the satiation. You're increasing variety, so you're not hitting  

that diminishing marginal utility point where  most of your income is going to the human sector. 

If that increasing variety is fast enough,  and there is no such increasing variety in  

the human sector, then you can get all of the  relational goods you want, but it doesn't matter  

for labor share. It goes to zero. 

Phil, I liked your analogy to some  Mongolian economist sitting around  

in 1400 thinking about what will be scarce  and the limits of that kind of analysis. 

I think you should talk about that. Just look at the goods available to  

a Mongolian of the distant past. I'm no expert on this society,  

but I know that they didn't have  nearly the variety that we have now. 

Look at the jobs that were intrinsically  human, like being a singer. 

And then you look at the things that  were not intrinsically human, like  

the transportation services provided by their  horses or the different kinds of food they had. 

If they just held the varieties fixed in both  categories and asked, "What will happen once  

we have a lot more automation?", they might  have said, "We'll just satiate in horse-like  

transportation and in yogurt and in yurts. Those shares will all go to zero, and we'll be  

left spending all of our money on singers." But of course, that's not what happened. 

As we've accumulated more wealth and more advanced  machines, we've expanded the range of things other  

than singers to spend our money on, and the  share spent on singers has stayed negligible. 

Likewise, that's my central prediction about how  the future unfolds, though it could go either way. 

I was going to make a point  and I realize it's a fallacy,  

but the reason it's a fallacy is interesting. It's just hard to imagine a world where there  

are trillions upon trillions of robots, but  there's only some billion-odd humans, and the  

cumulative amount we're spending on robots is less  than what we're spending to pay Magnus Carlsen or— 

Financial advisors or doctors or tutors. Right, or podcasters or whatever. 

But then I realized why it's a fallacy. The number of transistors in the world has  

literally trillion-X’d, maybe quadrillion-X’d. Your colleague Chad Jones has a very interesting  

result about how the share of the economy  that is going towards paying for computing,  

paying for the transistors, has been decreasing. The point you made is that one way to think  

about Moore's law is… What sets price? Supply and demand. So not only are we producing  

more transistors more cheaply, but also the  value of the marginal transistor is decreasing. 

As you were saying, another  way of saying Moore's law is… 

I like the pessimistic framing of Moore's law:  every 18 months, the value of computation halves. 

We're running out of uses for computation  so fast that it's sustaining Moore's law. 

This is relevant to a conversation about AI where  maybe for the first time, this is no longer true. 

The famous fact here is that an H100 costs  more to rent now than it did three years ago,  

even though we have much superior technology  and much more compute in the world. 

Because as models get smarter, the  opportunity cost of compute gets higher. 

This is Phil's point about increasing variety. What we have done is increased the types of  

things that people demand from capital. Now all of a sudden you have a new  

variety that you could be using  capital for, and you jump back up. 

You could imagine we just never  satiate demand for compute. 

As long as that stays the case, the  share of the economy that is going  

towards compute would keep increasing. That's the big question. That is the  

ultimate question that we need to be looking at. What number of new uses are we finding for that  

compute where you have the demand for these uses? What I want to emphasize is that a lot of models  

in economics, especially in the space that we're  talking about, take demand as almost exogenous. 

They don't unpack the psychology  of what people actually want. 

What got me thinking about the idea of  the relational sector is work that I was  

doing on the fact that there does  seem to be this intrinsic value. 

It's not just because it's scarce; it's  because there's some intrinsic preference  

that people have for empathy, connection,  and interacting with another person. 

One of the experiments that  we ran involved an art print. 

We have an incentive-compatible  way of asking, "How much are  

you willing to pay for this art print?" People are actually paying real money for it. 

Then we say, "Look, there's  only one of those art prints,  

and it's either made by AI or by a person." These are between-subject conditions. With one,  

you get the effect that the person-produced art  print is valued much higher than the AI version. 

Then, in a set of other conditions, we  say there's 500 of these being produced. 

For the human-made one, the price goes down  a lot because it's no longer seen as making  

a connection with this one artist. With AI there's no difference. 

AI is already viewed as a commodity. We need to do a lot more research on this,  

but it seems that's the key difference  between this and something like a horse. 

A horse was an input into an output, where  you can replace the horse with something else. 

You only care about the output. The only way this relational story works—and  

this is what we need more data on—is if a human is  not a horse in the sense that they are providing  

value from the output, where if you replace  the human, the value of the output decreases. 

If that's not strong enough, and if it doesn't  hold for enough sectors or enough jobs,  

then this story doesn't work anymore. There aren't that many institutions  

that have thought as hard as Jane Street about  how to turn smart people into some of the most  

competent researchers and engineers in the world. This relies in part on an apprenticeship model,  

where new hires are paired with senior mentors. But Jane Street also runs a bunch of  

classroom-style lectures and hands-on bootcamps. These courses cover a range of  

topics and they go pretty deep. There's one lecture that focuses  

on reverse engineering systems with tools like  strace and gdb and another that teaches you how  

to profile code down to the cache hierarchy level. Importantly, Jane Street designs these courses not  

just to teach the relevant object level skills.  but also to impart the relevant tacit knowledge. 

For example, their week-long neural net  bootcamp starts with general theory,  

but then quickly progresses to how  to apply neural networks to trading. 

And here they cover the specific obstacles  that Jane Streeters tend to encounter and  

the workarounds they've come  up with to get around them. 

Jane Street takes this sort of  learning incredibly seriously. 

Every office has dedicated classroom space and  courses are prioritized as part of regular work. 

If you'd like to work at a place  like this, Jane Street is hiring. 

You can check out their open  roles at janestreet.com/dwarkesh. 

There's one possibility which Molly Kinder has  written about, this "Messy Middle" scenario. 

That possibility made me think about  whether it might be better to have—at  

least as far as wealth distribution and  redistribution go—a much faster AI takeoff. 

I want to ask you whether the following  possibility is at all likely, or if there's  

any set of assumptions that can make it so. AI makes it possible to automate jobs such  

that many people are losing their jobs, but it  doesn't create enough wealth, while the process  

of automation is happening, to basically pay  off the people who are getting laid off and  

create a Pareto improvement, where everybody's  getting better as a result of AI automation. 

Of course, there's a trivial  sense in which that must be true. 

Whatever money the company is saving by not  paying the humans instead of just paying  

the AIs, those resources still exist in the  economy and can just be paid out to people. 

But there's going to be some  allocative inefficiency. 

The government doesn't know exactly  who got laid off because of AI. 

There's a political problem. If the  Meta worker gets laid off first and  

they were making $200,000 a year, is there a  politically sustainable situation where you  

give them a $200,000 check a year when there  are many working people making much less? 

Do you find this scenario plausible, where AI is  automating a bunch of things, but there isn't as  

much wealth creation as there is automation? I think it's possible. To me, it does seem  

like a pretty narrow window. My guess is that if we have  

the technology to automate so many jobs that  it becomes a new kind of political problem,  

then the pie will also be growing really fast. Well, unless in all of those professions it's  

automating, it's just a hair more productive. So the cost of all the capital to replace all  

the software engineers is just a  hair less than the cost of what  

we've been paying the software engineers. Why is it implausible that a company can save  

money by laying off a bunch of software engineers? And in the long run, there's a Jevons paradox, and  

we can't anticipate in advance what we'd do with  more software, and surely there will be more uses. 

But in the short run, the effect is  just that a lot of people are laid off,  

and they still need to figure out how they can  use a million times more JavaScript tokens. 

Phil and I have been writing about these  things, and we have mathematical models  

in the back of these things. We don't have any political  

economy in any of our models. Andy Hall wrote a really nice  

blog post about the politics of AGI, and  he made a really interesting observation. 

If there's a 2% increase in unemployment,  the political winds completely change. 

Unemployment has a huge effect  on what happens politically. 

Referring to Molly's excellent essay, I think in  some ways one of the worst scenarios is a drip  

scenario because of the political economy piece. What you might see is people not really being  

unemployed en masse, but moving into  sectors that pay them less money. 

This is what happened with phone  operators between 1920 and 1940. 

Phone operators were completely automated, but it  took 20 years, even though the technology existed. 

There was this drip. It wasn't like  this giant sector just disappeared. 

There's a really nice QJE paper on this  showing that they got reabsorbed into  

the economy, but at lower salaries,  and they were mostly underemployed. 

That's the scenario Molly was writing about,  this messy middle where things aren't a disaster. 

We saw with COVID that the fiscal response  can move quickly if there's an emergency. 

An emergency is a quick uptick in  unemployment, which could even look like 2-3%. 

That becomes a national  emergency if it happens fast. 

The concern is that whatever you're saving on  those white-collar workers, if that's not growing  

the economy but just creating saved resources that  can be allocated elsewhere, is that enough to do  

a broad-based redistribution scheme? You have the money you've  

saved off a couple of people. Unless you can figure out exactly how to get it to  

them specifically, you have the problem of, "Can  I do a UBI off the money I saved by laying off…?" 

You're basically saying the  pie did not grow that much. 

You're just displacing a bunch of people, but  that didn't grow the technological frontier  

of what the economy can produce. Then there's a question of whether  

every time this has happened in history, the  technological frontier has expanded a bunch. 

I think that's the case. Simply in history,  

the technological frontier has expanded. I think Phil made the same point. 

It's hard to imagine that sort of scenario  where you are getting intelligence that's  

just enough to replace the software  engineer but still costs a lot of money. 

It's just a hair less expensive than  the software engineer, so you're not  

getting this abundance effect. Where is the redistribution going  

to happen because the pie didn't grow? This is very helpful. Many different  

things have to be true for this scenario to  come to pass, each of which seem unlikely. 

One, it has to be the case that it is  possible to automate entire white-collar jobs,  

but only in a piecemeal way. That is to say that you can  

only automate software engineers, but  that same program can't also automate  

an accountant and an analyst and whatever. My model of intelligence is such that—both  

the breadth of tasks it requires to  do something like software engineering  

and what intelligence is—if you can really  just lay off all the software engineers,  

you've got enough in the bucket there that you  could automate all kinds of white-collar work. 

There are huge amounts of potential savings that  have happened as a result of these layoffs, and  

also AI is going to be cheaper than human labor. If both of those things are true, this messy  

middle scenario where we literally don't  have the wealth to go around seems unlikely. 

Then the question is, what is the best  way to tax it and redistribute it? 

I have some thoughts. I think it's really  important to outline the costs and benefits. 

First, there's differential complexity  in implementing these things. 

Two, they differ in the timeline  of being actually helpful. 

Something like universal basic capital is not  going to generate returns for something that  

happens in six months. You probably are going  

to end up with a layer of things. Take a negative income tax, for example. 

You implement it, and the day it turns into  law, you already have this insurance that  

there's a floor where everybody gets a certain  amount of money, and if you earn more money,  

you get taxed more. But there are positives  

and negatives to a negative income tax. With UBI, for example, I worry a lot  

about the political economy implications. If people are just dependent on a check,  

it really matters who's in power. Right now, we're endowed with  

labor that can turn into income. When that is no longer the case and we  

are at the mercy of the elected official for  basic needs, that feels like a power-sharing  

arrangement that's really dangerous. But wouldn't that be true of any  

sort of government redistribution program? With something like universal basic capital,  

where you have an ownership share and property  rights for capital, you just have a share. 

You're a normal shareholder. You're just a normal person. 

But this goes back to the question of  indexing, because if indexing is hard,  

then universal basic capital is hard. That's the problem of universal basic  

capital: targeting. What do you target  

to put into people's portfolios? Like, what if Anthropic goes to zero, but some  

random robotics company takes all this over? Exactly. That's the risk  

of universal basic capital. With a negative income tax, you have the  

same sort of issues as with UBI, where somebody  comes into power and says, "We're not going  

to do that anymore," and people can't work, and  then you have the issue of the floor being gone. 

One concern with the wealth tax  is that there's no politically  

sustainable equilibrium at a 0.5% wealth tax. This happened with the income tax, of course. 

It starts low, it’s for war or something, and then  it slowly escalates until the marginal income tax  

rate in the US is on the order of 40%,  and in certain states, upwards of 50%. 

With a capital tax, is there a reason to  worry that it would distort investment? 

Would people just say, "Why would  I invest in Anthropic or Intel? 

The government is going to take larger and  larger shares of it and dilute my share." 

Hold on. It's worth separating how the revenue is  raised, what's taxed, and how it's distributed. 

It could be that the government hands  out shares of Anthropic to everyone by  

a broad-based tax and then buying Anthropic. Which would probably be the right thing to do. 

Hopefully, some populist proposal  doesn't interfere with that and  

expropriate some particular company  that everyone happens to know about. 

You're suggesting there could  be some sort of optimal tax. 

We're taxing externalities or we're taxing land. I guess we probably need to tax something other  

than just those two things. Or consumption. 

Ok, a consumption tax, like a European  value-added tax, allows the government  

to go buy a bunch of stocks, and then they  just distribute those stocks to everybody. 

That's David Autor's... That's not going to be  

that different from just redistributing the  stocks, but it will be a little different. 

That was the proposal for  Social Security, by the way. 

That was privatizing Social Security. It's worked so far, but there are questions  

about how long it's going to keep working. Privatizing Social Security was basically giving  

everybody a basket of stocks. People talk about whether there's a  

white-collar apocalypse already. Is there any evidence  

that suggests there is mass automation or  unemployment as a result of AI already? 

A lot of people are looking at it. 

This is an area where there's a lot of  eyes and a lot of data being produced. 

The Budget Lab over at Yale is  doing really good analysis on this. 

They just recently released a report, and you  really have to squint to see anything happening. 

If you want to take an approach across the entire  economy, even looking at software engineering,  

the most exposed sectors, there's  just not really anything going on. 

There might be a little bit of a signal about  junior developers getting jobs less than before. 

But that's a "less than before" rather than a  level shift, as in there's actually an increased  

demand for senior software engineers, if anything. If you look at the trend, for junior developers,  

it's a bit below trend. So you're saying the growth  

is slower than before, but there is still  growth even for entry-level software engineers. 

What do you think is going on with the anecdotal  evidence of graduating college students saying  

that they're finding it harder to find CS jobs? I think that's anecdotal evidence. 

You think it's always been hard to get jobs  for some people, and now it's getting turned  

into an AI narrative? Same with the layoffs,  

where it's probably just a normal layoff,  and they turned it into an AI layoff. 

You have to be careful with all of this. There are these public coordination devices. 

Let's say we get into a narrative where if  you're a firm and you're not laying people off,  

then you're seen as not adapting AI enough. Then you're going to just get a cascade  

effect of firms needing to keep up with the  Joneses in terms of starting to lay people off. 

That's super worrying, where the firm  might actually be worse off after the  

layoffs than before, but it's just doing  the layoffs to have the perception that,  

"Look, we're not behind the times. We're using AI." You probably heard  

these anecdotal stories of token counters, where  you have to maximize tokens and things like that. 

Right now, we don't really have any  evidence of a white-collar bloodbath. 

Is that surprising at all, given  all these things AI can do? 

This is a story as old as time. If you automate some complementary task,  

the overall bucket of things—the human labor which  complements the automation—will increase in value. 

One of the statistics that's really important  for that argument is elasticity of demand. 

Take the O-ring model of jobs. A job is a series of tasks. 

Let's say the AI automates nine out of ten tasks. One task is not automated. 

If that person can now focus in on that  task, the job will become more productive. 

If that translates into a price effect where  the product is actually cheaper, and if demand  

responds enough where it's being bought  more and the service is being used more,  

that could actually lead to more hiring. A lot of people on the internet have been  

making that argument very generally, saying,  "Look, if anything in the data, we're seeing  

an uptick in software engineering demand." Which suggests that at least for now, given the  

way that jobs work, it might be elastic enough. I think this elasticity of demand argument is  

incredibly important for a lot  of arguments that people make,  

or just a lot of labels that people use without  understanding what the underlying causation is. 

People often talk about Jevons paradox. This is the idea that as something gets cheaper,  

you will want so much more of it that the  total amount you spend on the thing increases. 

Famously, this happened to  coal in Britain ~200 years ago. 

But really this only happens if the  demand for something is highly elastic. 

There are many things for which  there is not super elastic demand. 

If oil, for example, gets super  cheap, it's not like magically— 

Or insulin. 

Exactly. It's not like magically there's going  to be so many more cars that now we're going to  

be using way more oil than before. At least not in the short run. 

Exactly. The long-run elasticity is  higher than short-run elasticity. 

But even in the long run, agriculture famously is  the example where we can produce way more food if  

we dedicated the same portion of the economy  that we dedicated to agriculture in the past. 

We're already producing more food regardless, but  we could produce even more if the same portion of  

the economy that was producing food 100  years ago was currently producing food. 

But you eat enough, and then you're done. The claim with software is that it is not  

some inherent property of markets that as it gets  cheaper, you'll just keep wanting more of it. 

The thing about software is this is a particular  kind of good where as it gets cheaper,  

we'll want more and more of it. It is also highly relevant,  

and you wrote an essay about this—a lot of this  podcast is me summarizing your essays back to you. 

There's this very viral scenario  planning about the future by Citrini,  

predicting that as a result of automation and  very powerful AI, there will be a recession. 

White-collar workers will get automated,  their salaries will no longer be available,  

and so there will be a slump. Do you want to recapitulate  

why this might be implausible? Part of it is plausible, part of it's not. 

The part that we started the  conversation with is the idea that  

there could be a lot of unemployment. If the speed of automation is quick,  

people could get laid off, and they  may not find work very quickly. 

We can quibble about the unemployment part of  the Citrini essay, but that's not the issue. 

The issue is that they talked  about negative economic growth. 

What I did in the piece, that Phil and  I had a back and forth on, was to say,  

let's start with the proposition that  there's negative economic growth. 

What conditions do you need in the  economy to get negative economic growth? 

It turns out the conditions are pretty improbable. One thing that you need is for the holders of  

capital, rich people basically… Basically what you  have in those sorts of scenarios is a reallocation  

of wealth and income from lower-income people who  are using their labor towards tech capital owners. 

So you need demand to be bounded, like a hard  bound, not even a soft diminishing sensitivity. 

You need for them to eventually  say, "I've had enough. 

I don't want to spend any more money." And for that money to not enter as investment. 

Then you can get negative growth. The crucial thing is, even if we don't want more  

shit, the world in which there's a singularity  and we don't want to invest more money is crazy. 

We're not saying, "Let's build more data centers. Let's build more fabs." Even though we have AGI,  

we're not investing in more data centers to run  the AGI and that's driving more economic growth. 

I sent the essay to Phil, and Phil wrote back  being like, "This is pretty dumb," like my essay. 

He said, "You're trying to say that there's  going to be negative economic growth,  

but these are very implausible conditions." And I was like, "That's the point of the essay. 

These are very implausible economic conditions." That's where scenario planning really shines. 

You have the Citrini essay, which was great that  it was written because it started a conversation. 

It's so intuitive, this idea that if there's  demand collapse, we can get the economy to shrink. 

You could get that with a depression. In the Depression,  

the technological frontier didn't expand. Here, the technological frontier is expanding. 

You actually have abundance. For  abundance to generate negative  

economic growth, that's really hard to get. Google recently announced Gemini Omni and its  

video editing capabilities are incredible. You can upload a video and then tell Omni  

to do things like change the background or  adjust the lighting or add or remove elements. 

All while keeping everything else consistent. But Omni isn't just a video editor. 

I got a chance to sit down with the  research and product team behind Omni  

and I learned that it's a preview of how  future Frontier models will be trained. 

It can take in any kind of input,  whether that's text or audio or video. 

And while it doesn't currently do so,  architecturally it's capable of just  

seamlessly outputting images or text. So it's really a bet on the multimodal  

data transfer hypothesis. The model becomes better at  

predicting one data type by seeing the others. For example, Omni is really good at accurately  

rendering text on video, even though Google didn't  specifically target that capability in this model. 

And Omni is the next step towards  more accurate world models. 

Because in order to predict the next  frame of a video, you have to have a deep  

understanding of physics and spatial dynamics. As Omni progresses, it'll be interesting to  

see whether it can close a Sim2Real gap. Because it's much harder to collect data in  

the real world than it is in simulation, robotics  progress has lagged other applications of AI. 

But if you have really good video models that can  simulate reality, maybe that stops being the case. 

In the meantime, if you want to try Omni,  you can check it out in the Gemini app at  

gemini.google or use it in Google's AI  Creative Studio, Flow, at flow.google. 

We were talking a second ago about why there  isn't more automation as a result of LLMs. 

One plausible mechanism could be, as you were  saying with the O-ring theory… O-ring theory  

refers to the fact that the Challenger shuttle  blew up because one component malfunctioned,  

and it destroyed the whole thing. Maybe that's a more general model of  

how goods are produced in the economy. You have to make sure everything is  

reliable and works well. So you can't automate  

an entire job to an AI right now. Even though it might be able to perform it  

at some probability, you need extreme reliability  in order for it to not destroy the finished good. 

This might explain why there's a lot less  automation now than there otherwise could be. 

But I think it works in the other  direction once AIs get advanced enough. 

Integrating humans into the production  flow of future goods will become difficult. 

Even beyond the arguments about how humans will  be more expensive or less capable, there will be  

whole production flows organized for AI labor. They're talking in neuralese. They're thinking  

many thousands of times faster. So even if there's some comparative  

advantage where it makes sense to hire a human,  there will be transaction costs and worries of  

reliability that will actually make it hard to  integrate humans into future production flows. 

That seems right to me. In particular, I just want  

to distinguish between the point that if  you automate nine-tenths of a job, people  

might shift over to the last tenth, but there  might be ten times more work demanded of them. 

Compare that to the model of O-ring  automation from Gans and Goldfarb recently. 

If you can only automate nine-tenths of  the job, but you do it to a lower standard  

of quality than the human could, you might  not want to automate even those nine-tenths. 

That's the thing that could totally port over. Symmetrically, it could be a reason why we don't  

use a human for one-tenth of the job anymore,  because a human just can't perform it to the  

level of quality that the AI can perform the  other parts of the job, or the level of speed. 

They end up pulling down the quality  or speed of the finished product. 

By the way, the model you're talking about seems  extremely plausible to me for why more lawyers,  

accountants, or even software  engineers are not automated. 

There are cases where there's a pretty good  probability that the thing worked as you expect,  

but the thing you're paying the lawyer  for is: "No, really, my company's not  

going to go under because—" You're also paying for a lot  

of regulation-type stuff. With lawyers particularly,  

you need some entity to back up the product. You need ownership of the product. 

You need somebody to be able to fire or  hire, and there are licensing issues. 

There's a lot of regulatory layers that are  also going to be keeping—even if there's  

no relational element—humans in the loop  that have nothing to do with the ability  

of the human to actually perform the service. All of these frictions on the political-type  

decisions that we are accustomed to  only trusting humans for—legislation,  

being a judge, being a juror, or all the  licensing that keeps certain professions  

human—that all strikes me as transitional. What we expect to come from a human and  

how we organize our politics has changed so  many times throughout history, from little  

hunter-gatherer bands to empires to whatnot. Once an AI-run political system is much more  

efficient than the alternatives, those will  probably tend to out-compete the others. 

Speaking of which, we've been talking about  what preferences humans currently have and  

what impact that has on what kinds of  goods will be scarce in the future. 

But of course, we'll have different  kinds of entities in the future: AIs. 

There was a time when there were no humans  on Earth, but evolution selected for agents  

that have specific drives and preferences  because those tend to survive the most,  

and those preferences now determine what a  hundred-trillion-dollar world economy produces. 

Why not expect the same  thing from AIs in the future? 

This is not even a world with catastrophic  misalignment, where they just kill everybody. 

But there will be evolution  of, even if not individual AIs,  

firms which have AIs as part of them. What will that evolution favor? 

It will probably favor firms or agents that grow. There's a selection argument that things  

which grow will be more prevalent. Maybe just based on that, you can make some  

predictions about what their preferences will be. Is the kind of entity which prefers to have  

human-intrinsic goods going to be the kind of  entity that accumulates resources the most? 

Probably not. Probably it saves more and  has unsatisfiable demand for whatever the  

relevant resource happens to be. Compute is an obvious one. 

Can we use that to make some  predictions about the non-human  

preferences that will be guiding the future? If there's an AI that has its own welfare,  

is fully autonomous, and is making its  own decisions that are welfare-relevant,  

to be honest, I have absolutely no prior  that it would prefer to deal with humans. 

There's no reason. But let me take  the other side of that argument. 

Humans' preferences to be interacting with one  another, to trust and empathize with other humans  

versus a simulated AI, I think it's a really  important question whether those will change. 

I've heard a lot of arguments saying, "Look,  right now we're just not used to the technology. 

What you’re thinking of as relational…  At some point, people are just going  

to see an AI therapist as a superior product,  

and they're not going to need the  empathy that the human is providing." 

I think this is actually a  really complicated question. 

Here's one argument for why it's not going  to go away, and it has to do with evolution. 

Let's say there are two types of people. One person doesn't really have this preference. 

They can just interact with an AI,  whatever can simulate it better. 

The other one has almost a moral emotion—using  Jonathan Haidt's framework—against offloading  

those sorts of social interactions to an AI. Which of those two people are going to reproduce,  

find a mate, all of these sorts of things? I think the answer is clear. 

It's the second one that has  the preference for other people. 

Depends on how the reproduction is happening. Fair. But if we're in the world where  

reproduction is still happening the way that it's  happening, I think… And this is a big question,  

I'm not making a prediction. You had David Reich on the show. 

His point on the last podcast was that  we're buzzing with natural selection. 

So even if you get some sort of indifference now,  you might get selection to point into an even  

stronger preference for other humans. Here's one way to think about it. 

How is the wealth of the richest  people in the world instantiated? 

We were having a call earlier, and you  made the point that their consumption  

is more geared towards relational goods. Like Mark Zuckerberg is hiring MMA instructors  

and dancers for his wife's birthday, and so forth. But most of his wealth is just stock in Meta. 

As a controlling shareholder, he could say,  "Meta, turn all this wealth into dividend income,  

and I will just spend that on consumption." Instead, he would rather have his wealth compound  

and have Meta build more data centers. So you don't even have to change  

humans for this to be the case. The humans who are wealthiest—and  

growing wealthier because their wealth  is compounding—just have this almost Nick  

Landian preference for accelerating capital. That does seem to suggest that this is an  

important determinant of what kinds  of things are produced in the future. 

There are two ways you could get the two kinds  of people, one of whom prefers a human therapist  

and one of whom is fine interacting with the AI. If they both satiate equally quickly in capital  

but the one who likes the human therapist also  just likes having some human-intrinsic services,  

then the marginal value of capital in the future,  compared to the marginal value of capital today,  

for each of them if they start out equally  rich, should be basically the same. 

There could be interactions and whatnot,  but basically, that should be the same. 

If what's driving the difference is that one  person just doesn't satiate in capital because  

they're engaged by the prospect of exploring  the universe and turning their head into a  

galaxy brain or whatever, and the other  one satiates, then the person who doesn't  

satiate in capital is going to, if they're  being rational, have a higher savings rate. 

So in the long run, they're going to have most  of the wealth, and the overall capital share  

will basically be the capital share of that  person's spending, which is going to be one. 

It's important that we're not  talking about a hypothetical future. 

Elon Musk is talking about  mass drivers on the moon. 

He's by far the wealthiest person in the world. Obviously, currently his investments are going  

towards humans as well as machines, but I don't  think he cares particularly that his future  

researchers and engineers are humans versus AI. And he manages to reproduce fast as well. 

So I just think it's worth  drawing that distinction. 

There are currently some rich people that don't  seem to satiate quickly in capital, and so maybe  

in the long run they'll save the most. That does seem right to me. 

I would also say, even if they do  reproduce more slowly biologically,  

that might just not matter that much in  the long run if they can live forever. 

The living forever is key. Again, we're scenario-building here.  

If you could live forever, a lot of  stuff changes for my story as well. 

To your point about rich people  not consuming a lot and investing,  

this will all depend on the returns to capital. Right now, the returns to data centers are super  

high, but if we get into a situation where people  are satiated with capital, then the returns  

to accumulating capital are going to be lower. Then these rich people are going to be consuming  

more, because the incentive to invest is smaller. Basically, you think about the general equilibrium  

of this sort of process… We have  gotten tremendously richer since 1820. 

Many more people are investing, but you're still  getting a consumption response which keeps people  

employed and labor share high. That's because— 

Hold on. Wait, not necessarily. I  

think you're probably making the same point. It could be that their investment has to be  

titrated through actual laborers who have  to do things for their investment to work. 

In the future, only the consumption  is human-mediated, right? 

Because the investment can  just be done by the robots. 

So we're in the scenario of how  you can keep high labor share. 

Let's take that scenario. In the scenario  with high labor share, for whatever reason,  

the returns to capital are going to be lower. That's right. To the earlier thing where we were  

saying why the messy middle is implausible,  I feel like we can do a similar thing here. 

For our returns to capital to be lower,  the growth rate has to be lower, right? 

It certainly has to be lower than what we're  expecting through the period of transformative AI. 

If there's explosive growth… Yes and no. The capital stock could  

grow quickly, but the price of capital goods  relative to consumption goods could be falling  

faster than the capital stock is growing. It's the difference between the potential  

frontier of technology and the  realized prices of these things. 

Because you have relative prices. So you're saying I could be putting  

my money towards earning 30% interest and  investing in data centers, or whatever. 

There will be something in the future, if the  growth rate is high, that earns high returns. 

Or, as a result of all these technological  breakthroughs, there's some cool product  

that I really want to buy right now, and  both of those will be compelling options. 

Yeah. It doesn't have to be a new product. It could be a human-intrinsic product. 

Although, if it's a human-intrinsic  product, we would want to have it much  

more in the future than we want it now,  because the thing it compares against is— 

We might want it the same as we want it now in  the sense that the marginal utility in a ballerina  

performance is exactly the same as now. But the marginal utility in a robot  

might just be a lot lower than now. So in units of robots, we want it  

a lot more than we want it now. Would the interest rate be 30%? 

It depends what you mean  by the real interest rate. 

It might be that every robot now  can turn into 100 robots next year. 

So in units of robots, the  interest rate's 10,000%. 

But if the price of robots  is falling really fast... 

Prices adjust. I think that's the whole point. 

Here prices are adjusting in this interesting  way that too many macro models don't allow for. 

What's happening is what would be called  investment-specific technical change. 

The price of capital is falling  relative to the price of consumption,  

instead of doing the standard macro thing  of saying there's just output, this chimera  

of a thing called output, which one for one  can be allocated to capital or consumption. 

That's not going to be true in this world. Every unit of capital next year is giving  

up way less consumption than  each unit of capital this year. 

One robot now turns into many robots next  year, but the number of ballerinas is the same. 

Again, we're going to go back to  the increasing varieties thing. 

If all of those extra robots next year  are actually different varieties of  

robots and I'm not getting satiated on those  robots, then it's a very different story. 

But now we're talking about the consumption world. For the investment side of things, there could  

be just some greedy titan of industry  who keeps wanting more and more robots. 

That alone would be enough to increase the  marginal value of robots and therefore decrease  

labor share? Yes. 

Why are we not expecting greedy  titans of industry to keep existing? 

Greedy titans of industry  historically have built libraries and— 

But that's because they die, and they're like— Oh, they all die. Everybody dies. 

Well, we'll see. Conditional on people dying… You  

had a guest on the show who said to understand  the future, you should think about the past. 

You could have new types of titans  being born whose entire reason for  

accumulating wealth is just to accumulate wealth. But a lot of the time, at least historically,  

the wealth accumulation process is part of  a large social interaction amongst peers,  

amongst the community, where you  want to be admired in some way. 

The stylized fact of titans of  industry is you accumulate the capital,  

and then you buy a bunch of stuff. I guess this is a historical question,  

but it does seem to me that in a lot of cases what  is happening is that as they near the end of their  

life, they either hand it off to their children,  who are worse stewards of capital than they are. 

They don't even manage to grow their wealth at the  rate the economy grows, much less faster than the  

economy grows, which their parents were doing. They're like, "Well, I care less about my  

children having it than me playing  this game of accumulating wealth. 

So I'm just going to give it to some trust." If people are living longer or if they can  

figure out some way to align their trust to  this wealth accumulation process… It just feels  

like the evolution here is so strong. You just need a couple of agents that  

think this way for this to be the dominant thing  determining the preferences of the whole economy,  

because this part is growing much faster  than the other parts of the economy. 

The part about satiation and diminishing  marginal utility keeps coming up,  

I think it's really important. If a person has an intrinsic  

preference for accumulation, that's just what they  want, then I think your story is totally right. 

But that's just not how preferences usually work. You have enough hedonics in your life,  

and then the social status… Rousseau wrote  about this, St. Augustine wrote about this. 

This is a basic part of preferences. Now, you guys are arguing about something else,  

where you could have such high concentration that  you could just have a couple of exceptions to the  

rule, and that's going to be enough. I have nothing to say about that. 

I think the claim is a little stronger, not just  that you could have some exceptions, but that  

historically and today we see the exceptions. They just haven't really taken over the  

economy historically because there have been  these dissipation shocks, as they're called. 

They've given it to their kids who squandered  it, or they put it in foundations which spent it. 

It's not really a shock, but… People might have liked to  

fill the universe with monuments to  themselves and live forever, very wealthy. 

It's a weird preference, but it's  not a hypothetical preference. 

I think that's the claim. But who knows what's going on in their heads? 

Even without the intrinsic preference  for accumulation, there are some  

instrumental reasons why some people might value  accumulation, which is also worth bringing up. 

There's a desire for political,  philosophical, or religious influence. 

People get into an arms race over what  society looks like and what people believe. 

Similarly but differently, because  it's not an arms race, there's just  

total utilitarian philanthropy. When I think about why it might be  

good to have a lot of wealth in the future as a  good classical utilitarian, to me, the value—or  

at least one way you could have an almost  unsatiating utility function in having wealth  

in the future—is to create new happy beings. They just add to the total welfare of the world. 

This idea goes at least as far back as Bostrom's  astronomical waste point, that we could put  

Dyson spheres around the stars and turn all the  energy into really happy simulations and whatnot. 

I think the particular greediness of this  optimizer doesn't matter, what they're greedy for. 

Forgetting about utilitarian philosophy or  whatever, a pure von Neumann probe has… I  

don't know, is this an accurate way to say it? They just have high marginal value for the random  

solar system they'll occupy because  that turns into more solar systems. 

A von Neumann probe is a thing that can exist. That's a very greedy optimizer. 

If we're talking about whether they'll dominate  the economy, maybe this is a technicality. 

But we only count final consumption  goods and investment goods as GDP. 

If there's just this phenomenon— How does a von Neumann probe show up in GDP? 

Exactly. If we recognize it as a person that  owns itself, and it's optimizing on the margin  

between spending a bit more on a baby von  Neumann probe that colonizes another star  

system or a ballerina or something, and it  just doesn't value the ballerina very much… 

When we're talking about AI beings,  it just completely depends on  

how we're doing the accounting there. What does the world look like in a world  

where von Neumann probes are possible? Is it possible labor share is high? 

I think it's possible the labor share  is high the way we usually count it. 

One of the biggest problems in RL right now  is credit assignment because you have these  

extremely long rollouts and you need  to know why they succeeded or failed. 

One of Cursor's researchers, Sasha Rush,  gave me a blackboard lecture on how they  

use targeted RL with textual feedback to deal  with this problem and train Composer 2.5. 

I filmed on my iPhone, so  apologies for the camera work. 

So we've generated this output. It's just a sequence of tokens. 

We're gonna send those sequence of tokens  to this model that's gonna read it,  

then it's gonna isolate a specific  turn that it says is problematic. 

Then we're just gonna do text manipulation. We're just gonna take that trajectory and we're  

literally just gonna smash in some extra tokens. After Cursor injects these hint tokens,  

they run another forward pass. The trajectory itself doesn't change,  

but the hint causes the model to assign  lower probability to the error tokens. 

Cursor then trains the original model  to match those probabilities, basically  

teaching it to downweight these specific mistakes. 

There's a lot more nuance that we  couldn't include in this mid-roll. 

If you want to watch the full  thing, I posted it on my Twitter. 

And if you want to try out Composer  2.5, head to cursor.com/dwarkesh. 

Do economists have any advice for countries  which are not in the AI production chain? 

If you're not either producing the AI models,  you're not producing the hardware that goes  

into AI models, if you're not Korea making HBM or  Taiwan with the fabs or the Netherlands with ASML. 

India or Nigeria, what should  they be doing right now? 

If you're talking to Modi  right now, what do you say? 

I think the biggest lack of resources  that we have allocated in the economics  

profession is thinking about middle-income  developing countries in the age of AI. 

This is something I fault myself with as well. There's not enough people thinking  

about this question. There are scenarios  

where you get AI technology being allocated and  dissipating to Nigeria and developing countries,  

leveling the playing field, essentially  giving them a level up as far as capabilities. 

But there's another world where, because they  don't have enough resources, they're not training  

the models, they don't have the hardware,  and they just completely get left behind. 

And because of automation, we can produce  commodities in developed countries now. 

Then we don't even have the consumer market. That world looks pretty bad. 

This seems to me like an extension  of the messy middle case. 

One of the ways in which the messy middle might  only be bad in a narrow range of scenarios isn't  

just that it would be easy to redistribute because  the pie would be bigger, but because the interest  

rate would be way higher, and/or, equivalently,  the price of everything except human-intrinsic  

goods would be falling really rapidly. They’re sort of two sides of the same coin. 

A little bit of savings would turn  into a lot of consumption next year. 

Things have to go really wrong for  us to just get over the threshold of  

capital being productive enough to automate  lots of work, but not be productive enough  

that the interest rate is high and the price  of capital-produced goods is falling a lot. 

Even without redistribution, a little  bit of savings will save a lot of people. 

You're saying if the developing countries  have some savings in the developed world,  

that will be enough to produce a  lot of surplus that they can then— 

They will now be able to consume  a lot using their savings. 

But the messy middle could be wider in this case. They're starting from such a lower level in terms  

of how much they have and how much it's  actually indexed to the global economy. 

I think it's important for them to get on it now. I don't have strong feelings about whether it  

should take the form of sovereign wealth funds  that invest in the right supply chains or  

just subsidies to their own  citizens to buy a little bit of— 

This is actually a crucial point. We were talking earlier about why  

the Rockefellers of the world, why their  descendants don't control everything,  

if our argument about the selection  of these greedy optimizers holds. 

One argument is just that it's  very hard to index the economy. 

Maybe they would've just decided to have  their heirs index the economy and have  

their wealth grow at the rate of economic growth,  and their heirs would be trillionaires by now. 

Before index funds existed, it was just very hard. A very small fraction of the economy,  

going back 100 years, accounts for  a majority of the value created now. 

If you missed those particular things,  your wealth would've just stagnated. 

Maybe there was a brief golden window  from the creation of index funds up  

until five years ago where you could  actually index the economy and have  

your wealth grow at the rate the economy grows. But now we're in this world with very concentrated  

returns, especially to private companies. As we were making the point in our blog post,  

this is capital that the average person  has disproportionately less access to. 

Most of their capital is having a  random house, at least in the US. 

Or a part of a house. Which, as we were saying,  

is capital that is uniquely ill-suited to be  complementary to the production of AI or the  

serving of AI or to robots. Or the kinds of goods that  

the rich will bid up the prices of. Exactly. What is the value of a house currently? 

It's that the land is close to other humans and  modulo relational stuff that is just not going to  

be the main factor of production in the future. This would be why a Georgist tax would not  

raise enough money for the sort of  programs that we will be discussing. 

Right. But stepping back, the point I was trying  to make is, if it gets harder to index the  

economy now, and that is the main way in which  normal people are supposed to—modulo some sort  

of universal basic income— In the developed world. 

—are supposed to have some  purchase on the wealth from AI. 

And it's also the way that developing  countries are supposed to have some  

purchase on the wealth gains from AI. But it's very hard. Does Nigeria own a  

lot of SK Hynix and Anthropic? I'm guessing not. It's not  

enough for them to just own the S&P 500. This brings up a really important point. 

Is AI going to be like  electricity or social media? 

Think about ConEd, or whatever  the electricity provider here is. 

It's a monopoly. It provides a  resource that everybody uses. 

But do we think about electricity as  creating a concentration of power? 

Does ConEd have this huge amount of political  power, social power, or something like that? 

No, because with electricity, a lot of  the downstream benefits actually came  

to the users of the electricity rather  than the actual entity producing it. 

On the other hand, with social  media, it was the opposite case. 

Social media was everywhere. Everybody uses  social media, but the rents went to the platform. 

That's a really interesting point. I don't endorse this take yet,  

I'm going to talk out loud. The more you think our economy is  

going to be run on AGI the way our economy  currently runs on electricity—that is,  

there's a broad fundamental transformation  of the entire economy—the more it looks like  

electricity… Every company in the S&P of the  future, if it's going to make it to the S&P 500,  

it is because it has leveraged AI. Exactly. And then you're indexed again. 

But then again, I guess if you just look  at how concentrated the S&P is over time,  

just these big tech companies much more  so… I guess this goes to a fundamental  

point that it's hard to reason about how much  of the gains from AI these individual private  

companies will be able to control. I think the open model thing is  

going to be a big point here. If we're indeed in a world where  

the open models are six months behind the  frontier—or nine months—then we'll hit AGI,  

we'll hit whatever, and in six months,  everybody has access to this resource. 

This goes to show you that every  question is connected to every other. 

That question about whether there's  runaway gains connects to questions  

about recursive self-improvement. Even if not recursive self-improvement,  

then continual learning, or online learning,  which lets a model learn on the job. 

So if it's deployed, it gets to learn more. These are just forecasting technical questions  

which then impact whether Uganda will  have any purchase on the returns of AGI. 

The reason I'm emphasizing the question  is I think both for the messy middle and  

for developing countries, a recommendation  that is often made naively is that you've  

got to do some kind of retraining. You've got to do some kind of jobs  

program, or you've got to have them  build data centers in your country. 

I think you guys are suggesting something  closer to just buying the index of AGI. 

That's probably a much cleaner strategy  and much more likely to succeed. 

These are the two scenarios. I think there is a world where  

it is concentrated, in which case it's  going to be really hard to index AGI. 

There is another world where it is electricity. Basically every company has access to AGI. 

So you just buy the index. Nigeria just needs to buy the index, and Nigeria  

has access to AGI because of the open models. Just to get back to the question about whether  

to go with retraining or trying to index. I would prioritize trying to index,  

just given how fast AI could hit the world. But I definitely wouldn't just rely on that. 

In the messy middle cases or the long-timeline  cases where we don't get anything like AGI  

all that soon, it would be leaving a lot  of value on the table if you could have  

retrained to be a bit better educated on  how to use the latest wave of computing. 

I don't think there's that  much of an either/or there. 

Maybe the reason to be pessimistic about this is  because one of the reasons a country is poor is  

that it has a bad education system, so becoming  the best in the world at retraining people at  

using AI doesn't seem like a particularly  promising strategy for that poor country. 

Although there are cases where, in developing  countries, you had this leapfrogging effect with,  

for example, mobile banking. It's much more prevalent in  

Nigeria than it is in Germany. Everybody is doing mobile banking. 

They have it on their phones, and they're  constantly doing this sort of thing. 

Again, I'm not putting probabilities on this,  but with a transformative technology like AI,  

you could get leapfrogging where you skip the step  in the middle and get really astronomical growth. 

Just about the ease of indexing, I  think it's definitely something to  

worry about a bit and keep an eye on. But as discussed in our own essay,  

and as other people have pointed out,  it's already not that hard to index. 

There's been a bit of an increase in  the privatization of returns, but still,  

well under 20% of the total market cap of  non-tiny companies in the US is private. 

Everyone thinks about OpenAI and Anthropic. If that's where all the wealth will accrue,  

then all these questions about whether open  models will stay only a little bit behind,  

those are important. But even they look like  

they're going public before too long, probably. The frictions that have been keeping companies  

from going public might themselves be  alleviated by AI a lot, just all of the  

disclosure requirements and whatnot. They want to get access to more  

potential investors, too. If I had to guess,  

I would guess that the long general trend of  lowering those frictions and making it easier  

for more and more people to index will continue,  despite the recent bump in the other direction. 

This actually makes me hope even more so than  before that the labs do get commoditized, or at  

the very least they go public as soon as possible. But hopefully they just get totally commoditized. 

I think AI will be much more popular and, more  importantly, will be much more likely to lead to  

broad increases in prosperity if it is as hard  to capture the gains of AI as it is to capture  

the gains of electrification. Exactly. There's no  

anti-electricity people out there. I mean electricity doesn't take your job, but— 

Well, it takes some people's jobs. It took some people's jobs, yeah. 

This is maybe tangential to the  conversation but I think narratives matter. 

There's this really negative narrative around  AI right now, but that's because people are  

not putting out the positive narrative. There’s a reason. It's more difficult  

to imagine a good thing that doesn't  exist than losing something that exists. 

It's much easier for somebody to go on a  podcast and say, "These jobs that you like,  

they're going away," than for somebody to  spin up a utopia which doesn't exist yet. 

I hope this isn't too out of left field, but  I would be remiss if I didn't point out one  

big cost of having commoditized frontier  AI models, which is the tech race dynamic. 

For safety purposes, you might want  fewer frontier companies so that each  

one has a buffer in case they want to  slow things down to make things safer. 

The way this relates to our point before  about the widespread access of the returns,  

is that I think there's a lot less of a  trade-off there than some people imagine. 

Some people think either frontier AI  gets commoditized and we all enjoy the  

benefits—but there might be some risk, because  the market's really competitive and cutthroat—or  

things are safer because there's a big  gap between the leader and the laggard. 

But that means the leaders  get fantastically wealthy? 

No. You could just have a relatively  big gap, but it's a public company,  

and ownership in it is widely distributed. More recently, I have been thinking that  

the risk of commodification—which  is that it diffuses the ability to  

use AI to harmful ends—is worth the benefit. I worry that having these concentrated labs  

not only makes it so that the surplus isn't  as widely distributed through society,  

but also creates a very tangible, clear  political target for the government. 

We saw this with the Defense Production  Act threat against Anthropic. 

If there wasn't one lab, or a couple of  labs, that are clearly ahead of others,  

this kind of threat would be much harder to make. 

Thank you guys for doing this. I feel like there's a lot of unresolved questions,  

but it is helpful to know what the first  branch is along all these important dimensions. 

Great. Thank you.