0:00

upstream listeners, today we're sharing my

0:02

conversation with Byrne Hobart, recorded earlier

0:04

this week. We explore the economic

0:06

consequences of Trump's tariffs, examine

0:08

open AI's latest innovations and their implications

0:10

for AI's future, and discuss

0:12

the potential transformation of white collar

0:14

work. Please enjoy. Last

0:31

week we didn't get to discuss the

0:33

market and the economy shifting rapidly

0:35

as a result of Trump's tariffs even

0:37

though we discussed them before at

0:40

a high level. What do

0:42

you make of the last couple weeks of

0:44

market activity and what if anything should

0:46

we learn from it in terms of what

0:48

it means going forward? Yeah, so I

0:50

think that part of what we learned is

0:52

that there's this weird You know,

0:54

I think one of the first ways to look

0:56

at this is to compare the first and second

0:58

Trump administrations because one the things that was going

1:01

on in the first Trump administration was that between

1:03

non -Trumpie Republicans and just Democrats, it was

1:05

really hard for Trump to do a lot

1:07

of the things that he kind of

1:09

impulsively wanted to do. And there

1:11

was also just not a ton

1:13

of long -term planning from the

1:16

Trump Republicans on exactly what they

1:18

wanted to do. So there

1:20

was just a natural level

1:22

of ambient chaos throughout that first

1:24

administration, and my assumption was

1:26

that we'd see pretty much the

1:28

same thing the second time.

1:30

But it turns out that a

1:32

lot of people spent the

1:34

years from about 2021 to early 2025

1:36

figuring out exactly what it is

1:38

that they'd want to do. And

1:40

so you had this spree of

1:43

executive orders right at the beginning,

1:45

and a lot of them were just getting

1:47

the high notes on things that Trump

1:49

had always said he wanted to do and

1:52

had really gotten around to doing and

1:54

kind of looked like, you know, just from

1:56

the neutral standpoint of does this administration

1:58

have its act together? Do they have coherent

2:00

goals that they're actually trying to pursue

2:02

in a way that makes sense? It

2:04

kind of looked like they really scaled

2:06

things up. And then the tariffs were

2:08

one of the first cases where it

2:10

was just this kind of last minute

2:13

sort of spontaneous decision. And

2:15

one of the reasons we know that is

2:17

that there were a couple blow -by -blow

2:19

articles in different publications. New York

2:21

Times, Wall Street Journal, I think Politico also

2:23

had one that just talks through, here's

2:25

how this process actually came about, that Trump

2:27

had other options. He went with the

2:30

most aggressive and kind of bluntest one. And,

2:32

you know, one of the things that

2:34

stood out there was just this administration is

2:36

suddenly as leaky as the first Trump

2:38

administration. So in some ways, it is like

2:40

this high watermark for state capacity Trumpism.

2:42

in that he did actually implement this tariff

2:45

thing that he'd been talking about for

2:47

decades. And at another

2:49

level, it was the last gas

2:51

of state capacity Trumpism because... that

2:53

some of the earlier Trump moves

2:55

seemed to work was that some

2:57

very smart heterodox right -wing person from

2:59

the think tank world was coming

3:01

up with exactly the plan that

3:03

they would use on immigration or

3:06

on, you know, how higher education

3:08

is funded or whatever. And then

3:10

they just implemented day one and

3:12

it's a shock and awe thing.

3:14

And then with tariffs, presumably

3:16

there were such glands, they

3:18

all apparently got torn up and

3:20

discarded. And now the administration

3:22

is trying to negotiate other deals.

3:25

But what it looks like is that

3:27

Trump does react when the market goes

3:29

down, if it goes down enough. And

3:31

he has enough people in his orbit

3:33

to tell him this is actually a

3:35

disaster and that the only reason the

3:37

market's not down more is that everyone

3:39

thinks that he'll back down. And one

3:41

of the problems is this actually creates

3:43

a really strong incentive for China in

3:45

particular to do things that make the

3:47

US equity market go down. So that's

3:49

not good. We now have this. this

3:51

situation where, you know, one country that

3:54

is much more centralized, much more top

3:56

down and has far weaker checks and

3:58

balances is able to deploy its economic

4:00

power in order to make things difficult

4:02

for the US. And the US just

4:04

doesn't have the same capacity to reciprocate,

4:06

although there is a lot that we

4:08

can do. So yeah, that that kind

4:10

of setup is just not good. And

4:12

then yeah, you look at the market

4:14

and you kind of see roughly what

4:16

you would expect to see given that

4:19

there was general bad news expected and

4:21

the specific bad news was worse than

4:23

expected where stocks go down volatility goes

4:25

up. And one of the things that

4:27

you saw at that time, which there

4:29

are a couple explanations for is that

4:32

treasury rates went up. And so In

4:34

general, since about 1998, it has been

4:36

very reliable that when stocks are down, treasuries

4:38

outperform, and that if you want to

4:40

have a more balanced, you know, a kind

4:42

of safe and not perfectly cyclical portfolio,

4:44

you have a mix of stocks and bonds,

4:46

and usually one of those trades is

4:48

winning when the other one isn't, which as

4:50

we were kind of reminded of, like

4:52

this was always theoretically true, but it was

4:54

good to get a bracing reminder of

4:56

this in 2022, that only works if inflation

4:58

is low. Because when inflation is low,

5:01

that means that when the economy slows down,

5:03

central banks just want to eject a

5:05

lot of liquidity to speed things up. And

5:07

that tends to push rates down. But

5:09

when inflation is high, they can't automatically

5:11

always do that. They have to consider

5:13

whether they're actually pushing price levels up

5:15

more than they're pushing growth up. So

5:18

it that weakens the correlation. And there

5:20

is a level at which tariffs are

5:22

just inflationary. They are a tax their

5:24

tax that shows up in terms of

5:26

higher prices paid by consumers. So straightforwardly

5:28

inflationary. If you offset that tax with

5:30

a tax cut, you now have higher

5:33

consumer prices and a little bit less

5:35

revenue for the federal government because of

5:37

the deadweight loss. And so to keep

5:39

US standards of living flat, you actually

5:41

do need to do some inflationary things.

5:43

But then the other piece is that

5:45

There's a narrative that the US dollar

5:48

will no longer be the reserve currency.

5:50

I think that to the extent that

5:52

that's true, it is just, it takes

5:54

way, way longer than frankly, than one

5:56

president's term to actually put that into

5:58

effect. But also that, that doesn't seem

6:01

to be the big driver of treasuries

6:03

moving. You know, we'll have to see

6:05

as, as different data from different countries

6:07

comes out, but it doesn't seem like

6:09

there are that many countries that are

6:11

just mass panic liquidating treasuries. Because again,

6:13

they seem to think, and they're probably

6:15

right, that a lot of this tariff

6:17

stuff will be reversed and that they

6:19

shouldn't make these long -term changes to their

6:21

plans just based on short -term tariff

6:23

considerations. But one of the things to

6:25

consider is that if Trump gets his

6:27

way, then countries that were

6:29

running a trade surplus with the US

6:31

will have to buy American made goods

6:33

to offset that. And what they bought

6:35

instead was treasuries. And so even if

6:37

Trump does actually negotiate the deal of

6:40

his dreams with all of these countries

6:42

and they're all buying American gas turbines

6:44

and medical equipment and planes and weapons

6:46

and LNG and soybeans and whatever else,

6:48

that means they're not buying our treasury

6:50

bonds. So in that case, treasuries should

6:52

go down. And also if Trump just

6:55

does a tariff and offsets some of

6:57

the impact with attacks, that's also inflationary.

6:59

So treasuries go down. The scenario

7:01

where treasuries actually outperform right now is

7:03

one where Tariffs are mostly completely or

7:05

completely unwound. We do mostly return to

7:07

the status quo, but we return to

7:09

the status quo having gone through this

7:11

very disruptive period where markets went down

7:14

and a lot of capital expenditure plans,

7:16

both in the corporate sector, the private

7:18

sector were deferred and everyone's a bit

7:20

more nervous. So they're saving a bit

7:22

more money. In other words, the,

7:24

if there's a recession, but it's kind

7:27

of a, it's a sort of separation

7:29

of powers recession where one way it

7:31

would happen, because I don't think it's

7:33

that likely that Trump just completely blanks

7:35

and says, Yeah, I accidentally made a

7:37

series of gigantic mistakes over the last

7:39

several weeks and they're all my fault.

7:41

The main way that that happens is

7:43

through Congress of the courts, which means

7:45

that in the end, treasuries are actually

7:47

this really interesting bet on the future

7:49

constitutional structure of the US. So if

7:52

you are a believer in a strong

7:54

executive branch, for better or for worse,

7:56

then you think that treasury yields go

7:58

up. And if you are a believer

8:00

in separation of powers, then you probably

8:02

expect a recession, but it's kind of

8:04

a dash to bullet recession and you

8:06

expect Treasury, Treasury yields to go down.

8:08

So that's where we are on that

8:10

side. Another piece of the Treasury market,

8:12

though, is that one of the big

8:14

trades that a number of large hedge

8:17

funds do is the Treasury basis trade,

8:19

which is basically betting on the gap

8:21

between Treasury futures and the actual bonds

8:23

themselves. And In some ways, this is

8:25

just the simplest possible trade because you

8:27

can look at the value of treasuries

8:29

applied by futures, you look at the

8:31

value of treasuries that you actually buy,

8:33

you look at how much you have

8:35

to pay to finance them, how much

8:37

capital you have to put up, and

8:39

you can calculate very straightforwardly what your

8:42

return is on buy the treasury, sell

8:44

the future, and then either exit that

8:46

trade at some point if the price

8:48

is converged or just when the futures

8:50

reach their expiration date, you have either

8:52

rolled onto the next set of futures

8:54

or you just exit the trade, like

8:56

you deliver your futures to close out

8:58

the short, or you deliver your bond

9:00

to close out the short futures position.

9:02

And this trade, it is kind of

9:04

like a, at one level, really, really

9:07

simple trade. At another level, you can

9:09

do some more complicated things with it

9:11

because not all bonds are created equal.

9:13

And there are different times when you

9:15

might adjust your position size, et cetera.

9:17

So there's like, there's room to do

9:19

interesting stuff there, but The main way

9:21

that it seems to work is that

9:23

if you are a gigantic hedge fund

9:25

and you do a lot of business

9:27

with every major bank and you are

9:29

constantly negotiating the terms of what you

9:32

borrow from them, what you

9:34

would often want to do is be able

9:36

to borrow larger amounts of money than

9:38

you really need and on better terms than

9:40

you really need, you know, have a

9:42

lot of essentially dry powder. And if you

9:44

have that dry powder, why not put

9:46

it into a very a trade that can

9:48

be very straightforward, like the basis trade

9:50

and just take advantage of that. So some

9:52

of what could be happening is that

9:54

some of the funds that have been doing

9:56

that, they have other ways they can

9:58

use their liquidity. So they are gradually reducing

10:00

their exposure to that trade. I don't

10:02

know for sure, but it is a plausible

10:04

explanation for how rates are moving right

10:06

now. So that is another piece. And then

10:08

we can finally move to what everyone

10:10

means when they talk about the market, which

10:12

is the stock market. So stocks had

10:14

some of their worst three -day stretches in

10:16

a very long time, their worst one -day

10:18

performances in a really long time. And

10:21

I thought the most interesting

10:23

thing there was anytime any kind

10:25

of asset moves a lot,

10:27

people will make jokes about how

10:29

the market's not efficient. And

10:31

these jokes have varying levels of

10:33

seriousness. And I think pretty

10:35

much anyone who's discussing market efficiency in

10:37

a serious way is discussing it as this

10:39

sort of directionally true thing, but not

10:41

as this completely ironclad principle where every price

10:43

is always and everywhere the exact value

10:45

of the asset in question. Because, you know,

10:48

you can always just create any fake

10:50

trade and immediately prove that that's not true.

10:52

Like you can log into your brokerage

10:54

account and look at something we traded company

10:56

wait till the market is closed. So

10:58

it has a wide bid as spread and

11:00

then You could if you felt like

11:02

it buy 100 shares at some price that

11:04

is well wildly off the market. And

11:06

then obviously that price is not the efficient

11:08

market price like that the valuation for

11:10

that company based on your 100 share trade

11:12

is not the true worth of that

11:14

company. So it's never strictly true but I

11:16

think. I think it is actually worth

11:18

talking about what market efficiency means in different

11:20

dimensions when you look at a really

11:22

volatile period like this. And especially when you

11:25

look at things like market rallying on

11:27

the rumor of a 90 day pause and

11:29

then market ripping on the fact of

11:31

a 90 day pause and then market being

11:33

choppy as that pause kind of gets

11:35

gets squished around a little bit in terms

11:37

of what it actually means and to

11:39

whom it applies. And as people

11:41

also digest the possibility that the

11:43

US will not in fact be able

11:45

to negotiate dozens of bilateral trade

11:48

agreements in a three month period, given

11:50

that those agreements typically take many

11:52

years of negotiation to actually get right

11:54

instead. And so the

11:56

way that I look at that volatility

11:58

and especially the volatility around things

12:00

like there's a rumor stocks rip, you

12:02

know, the value of the S &P

12:04

rises by several trillion dollars and

12:06

immediately comes back down. The way I'd

12:08

look at that is that when

12:10

the market is volatile, the price of

12:12

liquidity is higher. And one

12:14

of the measures of the price of liquidity

12:16

is how much stocks move when you

12:19

buy a given amount. So in a market

12:21

that is efficient with respect to prices,

12:23

you would not actually expect these huge swings

12:25

in volatility, or you'd expect smaller swings

12:27

in volatility because the world is more uncertain,

12:29

but it's not, it's not, you know,

12:31

trillions of dollars of intraday fluctuations more uncertain,

12:33

just. more uncertain than usual. But if

12:35

you think of volatility as something for which

12:37

there's a market and that volatility or

12:39

if you think of liquidity as something for

12:41

which there's a market and that liquidity

12:44

is expensive when the situation is really uncertain

12:46

and when whoever takes the other side

12:48

of your trade knows that there could be

12:50

some huge news story that they got

12:52

a fraction of a second too late and

12:54

that they try not to make a

12:56

bad trade. In a case like that, stocks

12:58

should actually move pretty massively on fairly

13:00

small, fairly insignificant news flow. So I wrote

13:02

a couple pieces kind of touching on

13:04

that from different angles over the last week

13:07

or two, including just one look at

13:09

kind of why the markets crash a lot

13:11

faster than they rise. Because if you

13:13

go back and look at the best single

13:15

day updates in the S &P 500's history,

13:17

They are not good times. They're not

13:19

times that you associate with, it was a

13:21

great time to own stocks. Like a

13:23

lot of the best days in the S

13:25

&P history are in the 1930s, early 1930s,

13:27

when everything was falling apart. Some of

13:30

them are in spring of 2020. Again, things

13:32

were falling apart somewhere in 2008. And

13:34

then you have one, I think from 1987,

13:36

a couple of days after the crash.

13:38

So pretty much anytime someone is fully invested

13:40

on one of the best days in

13:42

S &P history, it's because they were fully

13:44

invested on one of the worst days in

13:46

S &P history too. That's

13:49

just that is another reflection of

13:51

volatility gets really really high and that

13:53

means that whereas like if you

13:55

look at Periods where the market just

13:57

generally did well like the 1950s

13:59

You don't have a lot of crazy

14:01

days in part because just the

14:04

flow of news during good during a

14:06

good economy is Things are always

14:08

just slightly like things were always looking

14:10

good and they're always looking slightly

14:12

better like the the big positive news

14:14

1955 will be something like Ford

14:16

made a slightly larger, slightly more expensive

14:18

car and fortunately gas prices are

14:20

extremely stable and therefore people will buy

14:23

this slightly more expensive car. So

14:25

this is slightly good news, but

14:27

it's not world -changingly good. The world -changing

14:29

news is almost always really bad

14:31

news, especially for large publicly traded companies.

14:33

So if you look at something

14:35

like AI, it is good

14:37

news for just the economy and

14:39

for human flourishing in general that we

14:42

can offload a lot of intelligence

14:44

to machines. It's also really, really

14:46

bad news for any company where they

14:48

have like this implicit balance sheet position of

14:50

we have a lot of people together,

14:52

we have built all these tools for coordinating

14:54

them and getting them to do fairly

14:56

low value added white collar work and now

14:58

that work can be done. Orders of

15:01

magnitude more cheaply and more quickly and we

15:03

can check the work more easily, etc. So

15:05

it often means like it's good news

15:08

for the world and then. the profits

15:10

of the public companies that are a

15:12

large proportion of the market are sometimes

15:14

impaired. We're actually just very fortunate that

15:16

this is a case where there's a

15:18

labor saving technology that is primarily being

15:20

built and deployed by very large cap

15:22

companies that are very profitable and that

15:24

are able to convince their investors that

15:26

this will be profitable too. You could

15:28

totally imagine a case where there is

15:30

an AI driven bear market because we

15:32

just have a slightly different configuration of

15:34

what companies are large cap companies and

15:36

what companies aren't. We'll

15:39

continue our interview in a moment after a word

15:41

from our sponsors. Well,

15:58

if you act by May 31st,

16:00

Oracle Cloud Infrastructure can help you do

16:02

just that. OCI is

16:04

the next generation cloud design for

16:06

every workload, where you can run

16:08

any application, including any AI projects,

16:10

faster and more securely for less. In

16:13

fact, Oracle has a special promotion where you

16:15

can cut your cloud bill in half when you

16:17

switch to OCI. The savings are

16:19

real. On average, OCI costs

16:21

50 % less for compute, 70 %

16:24

less for storage, and 80 %

16:26

less for networking. Join

16:28

Modal, SkyGavs Animation, and

16:30

today's innovative AI tech companies

16:32

who upgrade OCI and saved. Offer

16:35

only for new US customers

16:37

with a minimum financial commitment.

16:39

See if you qualify for

16:41

half off at oracle .com slash

16:43

turpentine. That's oracle .com slash

16:45

turpentine. In

16:50

an age of AI and big data,

16:53

your personal information has become one of the

16:55

most valuable commodities. Data brokers are

16:57

the silent players in this new economy,

16:59

building detailed profiles of your digital life and

17:01

selling them to the highest bidder. But

17:03

data breaches happening frequently, these profiles

17:05

aren't just fueling spam, they're

17:07

enabling identity theft and putting your financial

17:10

security at risk. But what if

17:12

there's a way you could opt out of that system entirely? Incogny

17:14

is a service that ensures your personal data

17:17

remains private and secure. They

17:19

handle everything from sending removal requests to

17:21

managing any pushback from data brokers. What

17:24

I trust most about Incogny is how

17:26

it provides ongoing protection. Once you're

17:28

set up, it continuously monitors and ensures

17:30

your information stays off the market. Want

17:32

to extend that peace of mind to your loved ones?

17:35

Incogny's family and friends plan lets you

17:37

protect up to four additional people under

17:39

one subscription. Take control of

17:41

your digital privacy today. Use

17:43

code upstream at the link below

17:45

and get 60 % off an annual

17:47

plan. Again, that's in Cogni .com

17:50

slash Upstreet. It's

17:55

fascinating. There seems to be

17:57

some, you know, as euphemism, you

17:59

know, confused messaging over whether

18:02

the goal is to improve our

18:04

negotiating position in order to

18:06

have more free trade. Actually, you

18:08

know, tariffs in support of

18:10

having other countries block their preexisting

18:12

tariffs. which is a different

18:14

argument than saying, no, we actually

18:16

need these tariffs to last because

18:18

we either need the revenue or

18:21

we need to reindustrialize. And

18:23

so supporters of the sort of

18:25

policies are basically explaining both arguments,

18:27

even though they seem to be

18:29

somewhat contradictory. And that's perhaps,

18:31

I'm assuming that there are people in Trump's

18:33

camp and you alluded to a, or

18:35

linked to a Tanner Greer post about sort

18:37

of different factions in Trump's camp. who

18:40

think differently. Yeah. So I think one

18:42

piece of this is just that to the

18:44

extent that tariffs are actually negotiating tool. It

18:46

is just going to like anytime you do

18:48

that you you open yourself up to some

18:50

level of embarrassment because you if you are

18:52

a good negotiator you do not get what

18:54

you said you wanted. Like that is the

18:56

whole point of being a good negotiator is

18:58

that you got what you did secretly want

19:00

but the other side feels like you compromised

19:02

and that you you actually met them halfway.

19:05

So yeah to the extent that

19:07

Trump is actually doing a

19:09

really, really clever five -dimensional chess

19:11

thing here, then yeah, he's not going

19:13

to get all the things he says he wanted. And

19:16

I think that some of the messaging just, like

19:18

it is in Trump's interest regardless

19:20

of what the end goal is for

19:22

the stated, for the implied end

19:24

goal to be that the US is

19:26

manufacturing a lot more stuff and

19:29

that China in particular is manufacturing a

19:31

lot less. Now, there are

19:33

plenty of configurations of the world where

19:35

that doesn't happen or happens to only

19:37

a very small degree or the

19:39

US shifts what it manufactures but doesn't

19:41

actually increase manufacturing up with that much

19:43

in the aggregate and the US

19:45

still ends up better off. But some

19:47

of that, you know, some of these

19:49

goals, it is again, like if I'm

19:51

sure that there were really sophisticated

19:53

plans for particular parts of particular supply

19:55

chains that could realistically be moved to

19:58

the US and other parts that

20:00

could be moved to US aligned companies.

20:02

and other things that we could do

20:04

to sort of box China in

20:06

where we find different countries or collections

20:08

of countries. that can in the aggregate

20:10

mimic some of China's comparative advantages. And

20:13

in some cases we just won't like

20:15

it is you just can't really mimic the

20:17

density of the the social network that

20:19

is Shenzhen in addition to all of the

20:21

all of the equipment and all of

20:23

the people and all of the infrastructure. Like

20:25

that part just has to get built

20:27

over time. And if not everything is co

20:29

-located in the same city or the same

20:31

set of cities, then it is going

20:34

to run less efficiently. And in that case,

20:36

what we're doing is basically buying insurance. So

20:38

if we take a supply chain

20:40

that was all China and move it

20:42

to, it's 80 % not China. And

20:45

you have four or five different countries

20:47

that each have a big chunk of it,

20:49

but nobody has a majority of it.

20:51

then that's just a world where a lot

20:53

more intermediate goods spend a lot more

20:55

time on ships. And there a lot

20:57

more cases where people are trying to coordinate something,

20:59

but they all speak different languages. And

21:02

so there are miscommunications and things like

21:04

there, there is a lot of noise and

21:06

friction in that. And one difference between

21:08

that world and the, the continued reliance on

21:10

China status quo world is that in

21:12

the latter. You can have situations where the

21:14

US can no longer buy drones or

21:16

the US can no longer buy GPUs There

21:18

are there are a lot of things

21:21

you know the GPU supply chain like you

21:23

know We don't directly buy them from

21:25

China But we buy them from places that

21:27

are very close to China and at

21:29

places that are if you look at a

21:31

map in China that are part of

21:33

China so definitely There is definitely a case

21:35

for saying that if the US were

21:37

spending a little bit more on some of

21:39

its strategically important imports, but also

21:42

knew that those imports would still be

21:44

available if there were conflict with China,

21:46

that the US would be better off.

21:48

And also a conflict with China would

21:50

be less likely in a scenario where

21:52

China can't actually, can't immediately remove the

21:54

US's access to some critical inputs. But

21:57

yeah, we'll we'll see how the

21:59

how the tariff regime and how

22:02

the negotiations evolve to get us

22:04

either in that direction or maybe

22:06

back to the status quo or

22:08

maybe to some world where there

22:10

are more barriers to trade and

22:12

The US out of necessity is

22:14

manufacturing more but I think in

22:16

that world There's also the question

22:18

of which companies would actually want

22:20

to invest in the US and

22:22

we we do know that there

22:24

is a set of policies that

22:26

can get companies to invest in

22:28

the US because private sector spending

22:30

on fixed capital did actually rocket

22:33

over the last couple of years,

22:35

partly because of AI -induced demand,

22:37

but also because of this targeted

22:39

subsidies for particular strategic industries. Some

22:41

more strategic than others, but it

22:43

did work. And I think that

22:45

it's actually harder to credibly engage

22:47

in a strategy like that. If

22:49

you wanted something like the IRA

22:51

or CHIPS Act, but just targeting,

22:53

say, defense and dual use technologies

22:55

that could be applied to defense. I

22:58

think people would be telling you like

23:00

we want cash upfront before we build

23:02

our drone factory in Texas. We don't

23:04

want the promise of future subsidies because

23:06

you had just promised this great trade

23:08

rearrangement and then you unpromised it and

23:11

then repromise parts of it and so

23:13

on. So we just can't really underwrite

23:15

a long -term investment in the US

23:17

on the basis of the strategy that

23:19

changes day to day. There's

23:21

a level of subsidies where you can underwrite

23:23

that, but that's also a level of

23:25

subsidies where you are subsidizing grift. So if

23:28

you pay people upfront enough money that

23:30

it is worthwhile to build a drone factory

23:32

in the US, there are people who

23:34

are going to take the money and do

23:36

other things with it. And this is

23:38

actually a problem that China had too. They

23:40

had this massive scandal with chipfab in

23:42

Wuhan, actually, where a lot of the money

23:44

seems to have just disappeared, as well

23:46

as some of the people involved. And it's

23:49

unclear if they disappeared themselves or got

23:51

disappeared. But either way, when there is aggressive

23:53

industrial policy that tries to pay upfront

23:55

because there's some policy expectation and stability, it

23:57

does select for some fraud. All

24:00

in past week, had a fascinating debate

24:02

between sacks and Chamath in one

24:04

hand and Ezra Klein and Larry Summers.

24:06

On the other hand, it was

24:08

kind of, you know, sort of a

24:10

representative of a broader debate. It

24:12

seems like, you know, most people can

24:14

agree that this has been handled

24:16

not super well from an execution perspective,

24:18

but, you know, but people on

24:21

the writers saying, Hey, directionally,

24:23

something like this or some

24:25

people had to happen. because

24:27

the current situation was

24:29

untenable, and this is

24:31

short -term pain for

24:33

long -term gain, either as

24:35

it relates to our

24:37

need to reindustrialize, or

24:39

they make some economic arguments.

24:42

My guess is that the

24:44

geopolitical foreign policy arguments are

24:46

stronger than the economic arguments,

24:48

but I'm curious to hear

24:50

your reaction. curious to hear

24:52

your take on the idea

24:54

of... something like this directly

24:56

need to happen or did

24:58

Larry Summers position of, hey,

25:02

free trade is actually this sort of

25:04

error that we had was actually

25:06

very successful for us. But he would

25:08

dispute the Zeyhan idea that we

25:10

were sort of subsidizing an order that

25:12

didn't make sense for us anymore.

25:14

Well, I think that there are there

25:16

are definitely parts of the free

25:18

trade system that like absolutely, you know,

25:20

the free trade system works really

25:22

well for the US in terms of

25:24

GDP growth, and it worked really,

25:26

really well for a lot of the

25:28

developing world also in terms of

25:31

GDP growth and. The question is more

25:33

about what is like, is there

25:35

a long term goal where everything kind

25:37

of equilibriates and equilibriates at US

25:39

standards of living or equilibriates at some

25:41

other standard of living and are,

25:43

is the US willing to defer consumption

25:45

in order to preserve its, in

25:47

order to preserve industries that are high

25:49

value, high margin, have a high

25:51

upfront cost and a lot of uncertainty

25:53

and that probably just don't come

25:55

back if we lose them. And I

25:57

think that's That's the piece where

25:59

it gets trickier, where if you have

26:02

a planned economy, but with some

26:04

capitalist characteristics and the level of capitalism,

26:06

the economy can fluctuate opportunistically, but

26:08

the state is still able to direct

26:10

a lot of resources towards the

26:12

goals that it cares about, then they

26:14

can do catch up growth very

26:16

straightforwardly. And that was the story for

26:18

a lot of the East Asian

26:20

economic miracle countries like Japan and then

26:22

later Korea, much less Taiwan, but

26:24

they were kind of attached to Japan

26:26

in a way that let them

26:28

sort of free ride on some of

26:30

what Japan was doing. But part

26:33

of the approach in both Japan and

26:35

Korea was we're going to subsidize

26:37

industries that lower the cost and increase

26:39

the pace of iteration in higher

26:41

value added industries. So we know that

26:43

if we dump enough resources into

26:45

steel, we'll have really cheap steel, we'll

26:47

be able to make cars, we'll

26:49

be able to make low cost export

26:51

goods. utensils and just

26:53

like steel, you know, simple steel

26:55

components and simple, you know, simple industrial

26:57

components, but we can also move

26:59

up the value chain once we've built

27:01

that. And anytime they're doing that,

27:03

you can, you can have this period

27:05

where a US manufacturer who used

27:08

to produce goods for many different levels

27:10

of that value chain, you know,

27:12

some really low margin commoditized stuff, some

27:14

really high margin, very specialized stuff.

27:16

as they, as they're able to outsource

27:18

some of that or as they

27:20

stop producing some low margin stuff because

27:22

there is overseas competition, they actually

27:24

end up with a higher margin business

27:26

that has more, more competitive defensibility

27:28

on its own. But then they can

27:30

end up in a case where

27:32

it turns out that it was actually

27:34

valuable for them to have some

27:36

of this lower margin, lower value added

27:38

stuff, specifically because it just increased

27:40

their overall economic footprint, meant that more

27:42

people were working with the company,

27:44

the company was working with more suppliers,

27:46

it had more scale advantages, was

27:48

able to promote more talent internally. It

27:51

is really hard to have

27:53

this fluid way to promote

27:55

a lot of really smart

27:57

people within smart and effective

27:59

people within a given organization

28:01

if that organization's main job

28:03

is to slowly shrink into

28:05

its highest margin activities rather

28:07

than continuously growing at all

28:09

levels. So you can

28:11

kind of look at Amazon as

28:13

a sort of like taking Japanese industrial

28:15

policy and just doing it on

28:18

a private sector basis. within one

28:20

company where they do think very long

28:22

term, they are very happy to find something

28:24

where the only limitation is their willingness

28:26

to put a lot of capital into something

28:28

and lose money for a while. And

28:30

then they just continuously scale it and try

28:32

to ramp up the quality to whatever

28:34

the acceptable level is and then ramp up

28:36

the monetization continuously beyond that. And as

28:39

long as what they're doing after that is

28:41

reinvesting that in more of the same,

28:43

then they're able to sustain that growth

28:45

for a very long time. But as they do

28:47

that, they end up having a very strong negotiating

28:49

position with respect to many of the companies they

28:51

do business with. And it's just, you

28:53

know, over the last 20 years, it's always been

28:55

better in the long run to be Amazon than

28:57

to be someone who was doing business with Amazon.

29:00

In any case, where Amazon was the

29:02

main dependency and the bigger they

29:04

get, the more they are the main

29:06

dependency for other other companies. So

29:08

I think that that's, you know, that.

29:10

in that framework, you basically want

29:12

to have some way for either companies

29:14

or countries to subsidize some low

29:16

value added compliments to the high value

29:18

added stuff that they can export

29:20

competitively globally. And if you lose

29:23

that, it is just much more likely that

29:25

you end up eventually losing the high margin,

29:27

high return on equity things that were just

29:29

that last layer because someone else has done

29:31

the entire process that you did to get

29:33

up to that point. And they also know

29:35

what that point looks like. And they also

29:37

know what the prices for those products are.

29:39

And as they get closer to being able

29:41

to build them, they also know more about

29:43

the margins. So I think

29:46

for free trade to be this really,

29:48

for people to be very long

29:50

-term optimistic on free trade with

29:52

sort of protectionist characteristics on

29:54

the part of other countries, you

29:56

have to be optimistic that

29:58

the US will just have some

30:00

natural advantage in doing in

30:03

doing high margin things and that we don't

30:05

actually need a lot of expertise in the

30:07

lower margin stuff in order to achieve that.

30:10

And I think there's a kind of

30:12

coherent model of the world where

30:14

that is actually true and where as

30:16

the US does less manufacturing, we

30:18

get more and more specialized in other

30:20

things and because we're more specialized

30:22

in those things and we have critical

30:24

mass. It's hard for anyone to

30:26

compete with us. And it does seem

30:28

to be true for both consumer -facing

30:30

software and business -facing software. It is

30:32

just really hard for other countries

30:34

to build and scale companies the way

30:36

that you can in the US.

30:38

But that also means that those same

30:40

companies are vulnerable to a world

30:42

where they're tightly integrated hardware and software

30:44

bundles. And at that point, the

30:46

companies that are good at the capital

30:48

-intensive, operationally -messing world of hardware They

30:51

already know how to do that,

30:53

and U .S. companies would have

30:55

to re -learn it. Let's get deeper

30:57

into that. This is a big

30:59

bet on American dynamism. Say more

31:01

about what you foresee as to

31:03

our opportunities and challenges in sort

31:05

of reindustrializing and what are you

31:07

predicting or expecting in terms of

31:09

how it's going to play out.

31:11

Yeah, so I think the US,

31:14

we have like our general disadvantage in

31:16

all of this is very high labor

31:18

costs, and there's an ambiguous disadvantage in

31:20

terms of regulatory uncertainty where the old

31:22

regulatory uncertainty was micro level, which is

31:24

you want to build something, you have

31:27

no idea how long it will take,

31:29

you have no idea how much money

31:31

you will spend dealing with every single

31:33

holdout, and we give a lot of

31:35

people veto power and the power to

31:37

delay things. We have some

31:40

rules that I think were very

31:42

well intended, but are

31:44

open themselves up to being manipulated. So

31:46

I'm just a piece of the New York Times a couple

31:48

of months back talking about endangered species. And

31:51

there is this kind of dark model

31:53

where if you take a species

31:55

that's not endangered, and then you figure

31:57

out, you find two subpopulations, you

31:59

split it up into subpopulations based on

32:01

some unique trait. And you're able

32:03

to get each one of those classified

32:05

as a species well now you

32:07

have two species and they both have

32:09

smaller populations and so if you

32:11

divide them finally enough you can eventually

32:13

get your way you can define

32:15

your way to an endangered species and

32:17

then someone can't build the factory

32:19

of the dam or whatever they want

32:21

to build. And there there seems

32:23

to be some appeal to doing that.

32:25

So the US had that that

32:27

kind of that kind of obstacle to

32:29

reindustrializing that actually seems to be

32:31

going away under Trump. And it has

32:33

been seamlessly replaced with the macro

32:35

uncertainty of we just have no idea.

32:37

how US products will be priced

32:39

relative to products made elsewhere in the

32:41

world, because we just don't know

32:43

what the tariff situation will look like.

32:45

But I think there are some

32:47

categories. Like in drones, it is just

32:49

kind of weird that for military

32:51

purposes, the US does do drones. But

32:53

for law enforcement purposes, there's

32:55

a lot less of the US presence. And

32:57

that seems like a straightforward place where

32:59

we can look at what's going on in

33:01

Ukraine and say that drones are going

33:03

to be a pretty important part of the

33:05

world, we can look at just the

33:07

general history of military equipment and say that

33:09

it is good to have dual use

33:11

producers. It is good to have companies that

33:14

can, from year to year, make a

33:16

lot of money selling passenger planes and can

33:18

occasionally scale up their production of military

33:20

aircraft. Or we look at the US auto

33:22

industry, which got huge over the course

33:24

of the 1920s and say that it was

33:26

actually, from the perspective of a country

33:28

that needed to produce a lot of tanks

33:30

and jeeps, it was really, really convenient

33:32

that... we had a lot of people who

33:34

were in the business of putting tires

33:36

on things that move and getting them off

33:38

the assembly line as quickly and cheaply

33:40

as possible. So I think

33:42

that something like drones, drones makes a

33:45

lot of sense. I think that going

33:47

through the list of inputs for like

33:49

going through the list of raw materials

33:51

inputs is probably more of like a

33:53

check the box thing. And it

33:55

is still, it is annoying to me that China

33:57

is able to disrupt supply of rare because

33:59

They are, and I think this is

34:01

more widely known than it was a few

34:03

years ago, like they're not actually rare.

34:05

That's more a branding thing. What is rare

34:07

is the facilities to process them, and

34:09

China has invested in that. And that seems

34:11

like this is actually a case study

34:13

in why controlling your supply chain is strategic,

34:15

is that China can just tell the

34:18

US that you don't get any rare earths

34:20

for a while. There

34:22

have been a couple inputs like

34:24

that that got disrupted during Russia's invasion

34:26

of Ukraine. And it turned out

34:28

that a lot of the companies that

34:30

depended on those inputs were actually

34:32

reading the headlines and knew that this

34:34

was a risk and had done

34:36

some stockpiling beforehand and did think of

34:38

alternative sources. But on a country

34:40

level, it actually seems like a very

34:42

useful thing to do. So I

34:44

think that looking at it from the

34:46

perspective of we take our currently

34:48

locally competitive products. So planes, medical equipment,

34:50

LNG, and gas turbines,

34:52

and some other kind of

34:54

industrial capital goods. And we

34:56

start iterating through the supply chain for

34:58

those and look at one, what are

35:00

the cases where we depend a lot

35:03

on imports from one country, which is

35:05

usually going to be China, and then

35:07

look at where else we could depend

35:09

on imports from or whether or not

35:11

we could onshore that. And if it

35:13

is the case that on ensuring that

35:15

it requires some subsidies, but not immense

35:17

subsidies. And you want to

35:19

subsidize the factory so that the return

35:21

on investment goes from 5 % to 10

35:23

% and not you want to subsidize

35:25

it. The return goes from negative 50 %

35:27

to 10%. If we find cases where

35:29

it is that incremental subsidy that actually

35:31

makes it worth doing, then we can

35:34

look at what else is downstream from

35:36

that input and then look. Again, is

35:38

there a case where it is almost

35:40

profitable for the US to build this?

35:42

And if so, we might want to

35:44

subsidize it so we can build it.

35:46

And I think that exercise actually gets

35:48

continuously easier over time as you iterate

35:50

over it because you get more used

35:53

to underwriting this kind of thing and

35:55

you also just get more visibility into

35:57

the overall supply chain as more of

35:59

it is located within the US or

36:01

at least being carefully looked at from

36:03

kind of a national security standpoint. So

36:06

I would say that I think another

36:08

piece of this is just the US

36:10

annoyingly does actually have an advantage in

36:12

just natural resource extraction. We whether whether

36:14

it is drilling oil out of the

36:16

ground or growing corn out of the

36:18

ground, you know, things that come out

36:20

of the ground, we just we have

36:22

good ground for that. So there there's

36:24

there's a different set of supply chains

36:26

there where oil oil is a very

36:28

complicated industry with a lot of very

36:30

specialized equipment, some of that is manufactured

36:32

domestically, some of that is not. And

36:34

Again, knowing what the dependencies

36:37

are, knowing which countries could

36:39

be, which countries if they are blockaded mean

36:41

that the US is impaired in its

36:43

ability to supply a lot of oil and

36:45

natural gas. It's useful stuff to know.

36:47

On the shipping side, Trump has talked about

36:49

the US ship, bringing about the US

36:51

shipbuilding industry that would be very hard to

36:54

do. China makes a lot of ships

36:56

and it takes a long time to scale

36:58

such an industry. And the US does

37:00

have allies that also produce ships. So I

37:02

think if I were if I were

37:04

building out high level industrial policy, I would

37:06

probably say we should continue outsourcing that

37:09

to countries that are actually capable of doing

37:11

it. And there may be a category

37:13

of ships like maybe your maybe a fully

37:15

autonomous ship is just such a different

37:17

design question from a ship that is crude

37:19

that you could actually just build that

37:21

supply chain from the beginning in the US.

37:24

And I think that is a case where

37:26

we should be looking at What is

37:28

the next step that hasn't scaled yet and

37:30

trying to scale that domestically? Because it's

37:33

just a whole lot easier to maintain high

37:35

market share in something than to get

37:37

the market share from zero to non -zero,

37:39

especially if someone else has already as part

37:41

of their industrial policy subsidized the early

37:43

stages of it. I think people have been

37:45

debating about related to this are who's

37:47

negatively affected more? The US or China or

37:49

who has more leverage here? Some people

37:51

say that the country with the deficits is

37:53

the one that has more of the

37:56

leverage. Who can withstand? this

37:58

sort of where we're at right now

38:00

for longer. How do you think about

38:02

this? So I think that just in

38:04

terms of the accounting, it does look

38:06

like China has the advantage in the

38:08

sense of they are not any worse

38:10

off if they sell, if they build

38:12

the same amount of stuff and instead

38:14

of buying treasury bonds, they just don't,

38:17

you know, they just dump it in

38:19

the ocean or something. Like they still

38:21

have the same level of employment. They

38:23

still have the same scaling for their

38:25

manufacturing base. And, you know, it's kind

38:27

of dumb to do that because the

38:29

Treasury bonds are a claim on future

38:31

output, future dollar denominated economic output, but

38:33

it is something where it would not

38:36

day to day affect anyone's life in

38:38

China if they were not actually selling

38:40

these goods to the US and were

38:42

just building them and throwing them away.

38:45

On the other hand, longer term, the

38:47

U .S. does have more population growth

38:49

ahead of it. The U .S. does,

38:51

you know, you can debate a

38:53

lot of the nuances on the military

38:55

stuff, and certainly the alliances that

38:57

the U .S. has are more afraid

38:59

than before. But the U .S. seems

39:01

at least positioned to survive a military

39:04

conflict with China long enough to... the

39:06

parts of its industrial base that it

39:08

needs to actually be competitive. So, you know,

39:10

in the really extreme worst case scenarios,

39:12

I think the US does okay. It's just

39:14

a really dicey situation for a while. And

39:17

China is, you know,

39:19

there was always that debate over, will

39:21

they get rich before they get

39:23

old? And they are getting old and

39:25

they're no longer getting rich at

39:27

nearly the same pace. So in some

39:29

ways, China does need to wring

39:31

what it can out of the current

39:33

system before that system. just doesn't

39:35

really serve China's interests. And they did

39:37

take their time in opening up

39:40

their financial system enough that they could

39:42

kind of globalize their own efforts

39:44

to buy claims on foreign assets in

39:46

order to pay for the livelihoods

39:48

of people who are living in China

39:50

and are at retirement age and thus

39:52

no longer working, no longer contributing

39:54

directly to the economy. And there are

39:56

a lot of those people. So

39:58

I think that China's China's it kind

40:00

of has a wobbly sort of

40:02

dominance in that sense, where they they

40:04

in the short term rely less

40:06

on the US than than the US

40:08

relies on them in terms of

40:10

standard of living. But over longer periods,

40:12

I think they are just more

40:14

dependent on there being an end market

40:16

for the kinds of goods that

40:18

a middle income country could produce because

40:20

they're not a high income country.

40:22

And it doesn't seem like they're really

40:24

on the path to being a

40:26

high income country. I think that that

40:28

is another component of US industrial

40:30

policy is that we can basically make

40:32

a list of things that are done

40:34

by Chinese companies where if you

40:36

didn't know better, you would assume that

40:38

an American or European company would

40:40

be dominant in that sector. So

40:42

I think things like very popular

40:45

consumer facing social media apps that are

40:47

used outside of China or telecom

40:49

equipment, for example. And I

40:51

think in cases like that, having a

40:53

more explicitly defensive industrial policy and saying

40:55

that These are things where we, A,

40:58

want to make sure that there are

41:00

domestic producers, and B, want to limit

41:02

the ability of China to sell these

41:04

globally. I think it does

41:06

actually make sense. That does all it

41:08

makes sense. We'll continue our

41:10

interview in a moment after a word from

41:12

our sponsors. Hey, everyone. Eric

41:15

here. In this environment, founders

41:17

need to become profitable faster and do

41:19

more with smaller teams, especially when it comes

41:21

to engineering. That's

41:23

why Sean Lanahan started Squad. a

41:25

specialized global talent firm for top engineers

41:27

that will seamlessly integrate with your org.

41:31

Squad offers rigorously vetted top 1 %

41:33

talent that will actually work hard for

41:35

you every day. Their engineers

41:37

work in your time zone, follow your processes,

41:39

and use your tools. Squad

41:41

has front -end engineers excelling in TypeScript

41:43

and React, and Next .js ready to

41:45

onboard to your team today. For

41:47

back -end, Squad engineers are experts

41:49

at Node .js, Python, Java, and a

41:51

range of other languages and frameworks. While

41:54

it may cost more than the freelancer on Upwork,

41:56

billing you for 40 hours, but working only two,

41:59

Squad offers premium quality at a fraction of the

42:01

typical cost without the headache of assessing for

42:03

skills and culture fit. Squad takes

42:05

care of sourcing, legal compliance and local

42:07

HR for global talent. Increase

42:09

your velocity without amping up burn. Visit

42:12

chewsquad .com and mention Turpentine to

42:14

skip the wait list. I

42:18

think we've been talking a lot about tariffs. I want

42:20

to make sure we get to a couple other topics

42:22

here that I'm sure will return back

42:24

to it. Open AI,

42:26

there have been some developments with the company you've

42:28

written about them. We've written about the company

42:30

a few times in the past couple of weeks.

42:33

One of them was that Sam Altman has

42:35

hinted that they are in fact going to go

42:37

open source or release an open source version. Any

42:40

commentary on this or just broader

42:42

sort of open AI developments? Of

42:44

course, they had their fundraise. You

42:46

wrote about their cap table. What

42:49

else should we talk about as opposed to open AI?

42:51

Yeah. So there's an interesting, you know,

42:53

you can look at the economics of open source

42:55

software in a couple of different ways. And as

42:57

with a lot of things, you sort of have

42:59

to, if you vary your level of cynicism, you

43:01

get to very different answers because one version of

43:03

the open source model is we're going to build

43:05

something really, really great. And we will understand, we

43:07

will build something everyone wants to use. We will

43:10

understand it really well. And therefore we will make

43:12

our money from helping other people implement our wonderful

43:14

open source thing. And that

43:16

was kind of a popular model for people

43:18

who did want to build software and share

43:21

it and also pay their rent to pay

43:23

their mortgage. But it turned out to be

43:25

a model where you are limiting a lot

43:27

of your economic upside and you're also running

43:29

what ends up being a very headcount -centric kind

43:31

of organization. So if you are in the

43:33

business of helping companies use Linux and Linux

43:35

is all free and open source, so it's

43:38

hard to directly charge for it, but you

43:40

can charge for developer hours, installing Linux and

43:42

maintaining it and so on. then you tend

43:44

to be running a company that just has

43:46

a lot of people and only grows by

43:48

hiring more of them. And I think for

43:50

some open source people, this was actually kind

43:52

of utopia, which it was basically, we have

43:55

a guaranteed job for all of our favorite

43:57

developers who also love the software we built.

43:59

And we get to spend all of our

44:01

time thinking about how to make Linux better

44:03

and how it can solve more people's problems,

44:05

et cetera. So that kind

44:07

of model did work, but another model

44:09

that is a somewhat more cynical

44:11

one, but a lot more effective is

44:13

that you build a free product

44:15

that anyone can use and then you

44:17

build a very similar product that

44:19

is paid that has better features. And

44:22

you continuously advance the quality of

44:24

the free product such that it continues

44:26

to be widely deployed. You also

44:28

continue to advance the features of the

44:30

paid product and you always you

44:32

because you can see that you just

44:34

have a lot of information on

44:36

how the product is used. You can

44:38

kind of see or at least

44:40

infer what parts of it you can

44:42

charge for. And that model seems

44:44

to work pretty well. And

44:46

if more interactions between people are

44:48

mediated by AI, then

44:50

very cheap, open models where you're

44:52

basically paying a tiny markup on just

44:55

the cost of the hardware required

44:57

to run it, those models will end

44:59

up getting used very, very widely,

45:01

especially in low value added tasks like

45:03

customer service, but also scams. Well,

45:06

once they are, if there is a slightly

45:08

better version of that same model and a

45:10

model that is built in such a way

45:12

that it understands some of the limitations and

45:14

deficiencies of the open model, then

45:16

the proliferation of that open

45:19

model ends up being a subsidy

45:21

to the closed model. And

45:23

I think this probably works a lot

45:25

better in cases where it's a domain

45:27

-specific model because you just have a

45:29

lot more control over how it gets

45:32

used and how it gets deployed. So

45:34

if you imagine someone creates an

45:36

open -weight Hagelbot. And Hagelbot's job

45:38

is that if you are

45:40

selling some product on Facebook Marketplace,

45:42

Hagelbot will handle the negotiation

45:44

for you and help you get

45:46

the best price possible. Well,

45:49

in that case, if someone buys

45:51

Hagelbot Pro at Hagelbot Pro

45:53

specifically designed to out Hagel Hagelbot,

45:55

then the fact that everyone

45:57

uses Hagelbot means that Hagelbot Pro

45:59

has this built -in locked -in

46:01

market. I would be pretty ambitious

46:04

to do that in a totally open,

46:06

you know, general purpose model kind of

46:08

way. But it does, it does seem

46:10

like that's kind of the shape of

46:12

things where it is just a lot easier

46:14

to produce a huge stream of tokens

46:16

that are all customized for the recipient. And

46:18

that means there's a much higher premium

46:20

on being able to ingest and summarize

46:22

a lot of data and to do

46:24

so in a way that's helpful to one

46:26

specific person. So it kind of ties

46:28

into their other recent feature, which is

46:30

improving memory. And this is

46:33

just a great thing for lock -in

46:35

because once you, because whichever LLM

46:37

you have the most user hours or

46:39

the most user tokens exchanged with,

46:41

that LLM knows you better. And so

46:43

it's actually able to give better

46:45

answers to your specific questions than other

46:48

LLMs. So, for example, let's

46:50

say you sometimes use LLMs to figure

46:52

out what restaurant to go to when you're

46:54

visiting a new city. If you previously

46:56

mentioned to chat GPT that you're a vegan

46:58

or your carnivore or your

47:00

kosher or whatever, then then it

47:03

will always just know that and you

47:05

won't be explaining these details or if you

47:07

have other idiosyncratic references that it's slowly

47:09

picked up on, you won't be repeatedly explaining

47:11

those details. So you end up defaulting

47:13

to whichever LLM you've used the most and

47:15

then you are, of course, when you

47:17

default to that, you're giving it more data.

47:19

And if the data is actually proving

47:21

it, then you probably end up using it

47:23

more than you otherwise kind of actually

47:25

would have. And then I forget

47:27

who pointed this out on Twitter, but

47:29

someone pointed out recently that it is possible

47:31

that this can be tied in with

47:34

a login with open AI button. And at

47:36

that point, what that means is that

47:38

it can port a lot of your preferences

47:40

over to whatever new app you're using. So

47:43

if we have that example of

47:45

you've used open it, you've used

47:47

chat GPT a lot of times

47:49

to figure out where to eat

47:51

when you're visiting a particular city.

47:53

If you log into TripAdvisor with OpenAI,

47:55

then maybe either OpenAI shares a

47:58

bunch of references with TripAdvisor, or maybe

48:00

that's a privacy violation. What actually

48:02

happens is that when you interact with

48:04

TripAdvisor, you're actually interacting with an

48:06

OpenAI layer on top of TripAdvisor, where

48:08

TripAdvisor is just the kind of

48:10

dumb API play, and it's piping in

48:12

a bunch of data. the LLM is analyzing

48:14

that data for you and then it's giving

48:16

you your rank list based on your preferences and

48:19

TripAdvisor does not actually even know what the

48:21

ranking was and what the final output was. They

48:23

just provided the data. So it

48:25

does become a way to make LLMs

48:27

more of an interface and more of

48:29

a default way for interacting with other

48:31

services. And I think that for a

48:33

lot of companies, that's just, that's something

48:35

they have to think about is that

48:37

for now, If they have some service

48:39

where the idea is they can answer

48:41

complicated questions in some specific domain because

48:43

they have a lot of the data,

48:45

which does describe, you know, describe TripAdvisor.

48:47

That's also LinkedIn. That's also GitHub, et

48:49

cetera. They do want to

48:51

have a natural language interface because that is

48:53

a way that people are now looking

48:55

for that information. But they also probably want

48:57

to think about what their business model

48:59

looks like if they are purely providing. outputs

49:01

by API as their default and that

49:03

the actual front end for the site is

49:05

basically a demo version that you would

49:07

use to learn what TripAdvisor is so that

49:09

you can now use the data that

49:12

they have collected and filtered through your LLM

49:14

choice. So I think

49:16

that that will I'm sure some companies

49:18

will just make that transition pretty

49:20

seamlessly to pure purely providing data or

49:22

almost purely providing data to LLMs

49:24

rather than providing data to end users.

49:26

And then other companies will probably

49:28

just get run over and they will

49:30

die. But all the data

49:32

they gathered will live on forever in

49:34

the weights of all the models that

49:36

come after this. You wrote a blog

49:38

post I wanted to get into asking

49:40

if white collar workers will experience a

49:42

blue collar. economic recovery. Well, why don't

49:44

you unpack that? Yeah. So there was

49:46

this period in the early 20th century

49:48

where the US economy was growing and

49:50

it was growing because it was industrializing.

49:53

And then there's this period in roughly

49:55

the mid 20th century where the economy

49:57

is growing and manufacturing is not actually

49:59

growing with it. And then there's this

50:01

period starting pretty much 1990 or so

50:03

is about when the inflection happens in

50:05

the chart. So whatever the real forces

50:07

behind that work happened earlier. And that's

50:09

the period where the economy is growing

50:11

and manufacturing, at least in terms of

50:13

manufacturing employment is shrinking. And that is,

50:15

you know, actual manufacturing output still looks

50:17

pretty good. There are some weird price

50:19

level adjustments that, that end up being

50:21

made on that, that make those numbers

50:23

a little bit hard to compare over

50:25

long periods. But, you know, the US

50:27

does still make stuff, still makes a

50:29

lot of stuff. And China, China has

50:32

surpassed us, but it's a fairly recent

50:34

phenomenon. But the actual US manufacturing employment

50:36

has not really recovered from that peak.

50:38

And it does not seem likely to

50:40

ever recover from that peak. It will,

50:42

there will probably be bounces up and

50:44

down, but the US, it, the US

50:46

has more of a comparative advantage in

50:48

the kind of manufacturing where you have

50:50

a lot of really expensive equipment. You

50:52

have very specialized people who have some

50:54

combination of, you know, theoretical knowledge that

50:56

you learn from a textbook and practical

50:58

knowledge that you get from using machine.

51:00

According to its instruction manual, realizing the

51:02

instruction manual has missed in nuances that

51:04

you have been able to pick up

51:06

and then. you know, slowly, slowly

51:09

accumulating that kind of expertise. So

51:11

like we probably will just have this

51:13

increasingly small, more, more skilled up

51:15

manufacturing labor force. But one of my

51:17

questions on this was what happens?

51:19

What if white collar labor actually goes

51:21

through the same transition? Because when

51:23

we look at this industrial cycle and

51:25

how it affected employment was when

51:27

manufacturers were doing well, they would upgrade

51:29

a lot of their equipment. And

51:32

when they do that, that means that

51:34

they are basically increasing the amount of capital

51:36

versus the amount of labor as inputs

51:38

into what they do. And then when there's

51:40

a downturn, they can lay some people

51:42

off. They still have this equipment and they

51:44

can rehire some people, but. they

51:46

are still profitable operating at less

51:48

than full capacity, they probably

51:50

operate at less than full capacity. So

51:52

they can kind of take their time

51:54

on hiring people. They have these efficiency gains

51:57

from having enough equipment to service peak

51:59

demand. And then when they're at sub -peak

52:01

demand, they just don't have as many

52:03

people employed. And for a long

52:05

time, it's been hard to automate white collar

52:07

workers in the same way, but that

52:09

is getting easier. You could

52:11

imagine a lot of companies where they

52:13

maybe made a decision to spend a

52:16

lot on AI customer service tool or

52:18

on building a model out of their

52:20

building a model trained on their internal

52:22

source code or whatever the thing was.

52:24

And that they underwrote this based on

52:26

the level of demand for their

52:28

product that they saw. A few months

52:30

ago, that demand is no longer here,

52:32

but they have actually made the investment

52:35

already. So now their output per software

52:37

engineer, their output per customer service agent

52:39

is a lot higher. And they

52:41

may just find that they don't need

52:43

to staff up that much in the

52:45

next round. And also, as they're building

52:47

more of these specialized in -house tools,

52:49

you have to know more about how

52:51

an individual company operates in order to

52:54

function effectively there. So it actually becomes

52:56

more expensive to hire people. There's a

52:58

higher opportunity cost to interviewing them and

53:00

training them. And then there's also a

53:02

higher cost to letting them leave. So

53:04

you actually end up with this

53:06

more static workforce after a while. And

53:09

it seems like in some cases, what

53:11

the blue collar workforce did was they retired

53:13

earlier than they otherwise would have with

53:15

a lower standard of living. But they did

53:17

okay, just not as well as they

53:19

might have. And then in other cases, you

53:21

have people who downshifted from they had

53:23

a union job at the steel plant, now

53:26

there's no union and no steel plant.

53:28

So they spend the last couple of years

53:30

before retirement delivering pizzas or driving a

53:32

truck or something. And that is definitely not

53:34

what they wanted, but it is, you

53:36

know, they still need to work and it's

53:38

better than nothing. And I

53:41

think for a lot of workers, they

53:43

just end up in kind of a

53:45

less structured, less formal employment environment where

53:47

it is a lot more gig -based,

53:49

a lot more variable, etc. And

53:51

also the cost of that kind of

53:53

gig work has gone down because of LMS,

53:55

too. So you can have a 24 -7

53:57

live support chatbot that you just pay

53:59

for that speaks a natural language, can understand

54:02

natural language queries in a variety of

54:04

natural languages. And you can get some sleep.

54:06

You don't actually have to scale your

54:08

business to the point where you can have

54:10

full -time 24 -7 customer service. So you

54:12

can have a lot more just kind of

54:14

gig work in the white collar space. And

54:17

if the transaction costs for working

54:19

with a more random assortment of partner

54:21

companies goes down, again, because of

54:23

all of them, then maybe there is

54:25

kind of an afterlife for steady

54:27

white collar jobs that exist in that

54:29

more informal sector. And I'm still,

54:31

I'm not, I didn't write that because

54:33

I was convinced it would happen.

54:35

I wrote that because I was kind

54:37

of speculating that it is more

54:39

likely than it was years ago. Then

54:41

I still think there's a completely

54:43

coherent view of the world where the

54:45

higher productivity of software engineers who

54:47

are using LLMs means that we write

54:49

a whole lot more software. A

54:51

lot of that software interacts with other

54:53

software and there's just not indefinitely

54:56

forever more demand, but there is more

54:58

demand for a very long time.

55:00

But it's not the only outcome that

55:02

we can see. That makes sense

55:04

as an overview. You also

55:06

wrote about Harvard, and you

55:08

wrote about them borrowing. And I'm curious

55:10

just to ask you the broader question

55:12

of where are the Ivy League schools

55:14

in the state of them right now

55:16

with what Trump is planning to do

55:18

with their own sort of economic position

55:20

separate from Trump and sort of lost

55:22

popularity for a bunch of reasons, both

55:24

structural and political, and share what

55:26

value they offer in perspective. What

55:29

is the state of higher education?

55:31

You know, that's one where I think

55:33

I should actually punt on it

55:35

because what I would love to know

55:37

and don't have handy is, what

55:39

is their actual annual operating expense? What

55:41

is that operating expense? What is

55:43

their operating income, inclusive of and exclusive

55:45

of federal subsidies? And then how

55:48

big is the endowment and what's the

55:50

expected return on that? My,

55:52

my understanding, like the annoying thing about

55:54

talking about Harvard is that Harvard is like,

55:56

it is a great way to refer

55:58

to, it is a great bit of them

56:00

for just high status, high prestige school,

56:02

but actually a lot of the rules that

56:05

apply to those schools, a lot of

56:07

the generalizations that apply to them do not

56:09

apply as strong at Harvard because it

56:11

is just very far ahead of many of

56:13

its peers in terms of things like

56:15

having a very large and very well run

56:17

historically endowment fund. And having enough of

56:20

a brand name that if people are choosing

56:22

among, if someone has their choice of

56:24

elite schools, they will often choose Harvard above

56:26

the other options. So Harvard actually can

56:28

save itself from some of the difficult trade

56:30

-offs that other places have. But there's no

56:32

free lunch. You always just have a

56:34

different set of trade -offs because now you

56:37

are much more deeply symbolic. And you also

56:39

just have a bigger responsibility to actually

56:41

pick the right students, get them.

56:43

well -educated, get them well -networked, get them into

56:45

important jobs, and then make sure that

56:47

they actually do good things in those jobs.

56:50

So that's one that I want to

56:52

look into. It does seem like Harvard

56:54

just is probably, you know, they are

56:56

fine without federal funding. They would obviously

56:58

prefer to have more money than less,

57:00

but they will do okay. And I

57:02

think that's part of the purpose of

57:04

Endowment Fund is that you want to

57:07

be in a position where you don't

57:09

have to suddenly change your priorities because

57:11

somebody else did. And if you're Harvard,

57:13

you can just keep on going. And

57:15

then there are these other interesting open

57:17

questions on just what is the social

57:19

utility of having various investment vehicles that

57:21

have an indefinite time horizon and are

57:23

attached to institutions that are literally older

57:26

than the United States of America itself.

57:28

And I think it is probably prosocial.

57:30

It's probably useful to have at least

57:32

some organization where it literally makes sense

57:34

for someone to ask if we had

57:36

an investment where we think that the

57:38

payoff happens in a century, what is

57:40

the price at which it's worth doing

57:42

that today? because pretty much any other

57:45

organization just doesn't have that kind of

57:47

timeframe and shouldn't have that timeframe. Like

57:49

it's a silly exercise for someone at

57:51

KKR to ask about the 100 year

57:53

return on anything unless it is something

57:55

where it actually has cash flow between

57:57

now and that point 100 years in

57:59

the future. But at Harvard, they at

58:01

least can put a discount rate on

58:04

it and ask if it's worth doing

58:06

because they probably will be around once

58:08

the payoff actually hits. And then

58:10

the specific thing that I read about them

58:12

was just that partly in response to that

58:14

Shakir fund raising, the Shakir funding situation, they

58:16

were borrowing some money and they announced this

58:18

plan at a time when markets were pretty

58:20

choppy and they would have gotten much better

58:22

terms if they had decided to borrow a

58:24

few weeks earlier. And there was some speculation

58:26

online that, hey, maybe this means Harvard is

58:28

financially distressed. But I think what it actually

58:30

means is that Harvard is a relative safe

58:32

haven and they're just very, very unlikely to

58:34

default on their debts. But they're also a

58:36

large investor. They try to invest in riskier

58:38

assets. You know, they're trying to earn positive

58:40

carry if they borrow money. So they are

58:42

lending to people who are going to pay

58:44

them more than they're paying. And that probably

58:46

the rates that those people are paying have

58:48

gone up more than the rates Harvard is

58:50

paying. So in the event that Harvard does

58:52

need to move more money from its endowment

58:54

to the operating budget because that budget is

58:56

not going to be covered by the federal

58:58

government. it does actually make sense

59:00

for them to borrow money to do

59:02

that because the alternative is sell things that

59:05

are sell things at a more distressed

59:07

price and forgo some long term gains. So

59:09

again, if you if you had a

59:11

shorter outlook then indefinitely maybe you would choose

59:13

something different but If Harvard is an

59:15

exception to be around forever, and if the

59:17

Endowment Fund makes it very likely that

59:19

they will, in fact, be around forever, then

59:22

I think they are in the

59:24

privileged position to take those long -term

59:26

contrarian bets when they have the opportunity.

59:29

Speaking of privileged position, we'll close on this.

59:31

You know, some people are saying, hey,

59:33

is there going to be mass capital fly?

59:35

And we alluded to this earlier, but

59:37

you wrote a piece basically suggesting how, or

59:39

talking about how sticky the reserve currency

59:41

is. Yeah. And one of the elements of

59:43

that is just that anything that actually weekends,

59:46

the US status as the issuer of

59:48

reserve currency is so bad macroeconomically for

59:50

the rest of the world that people

59:52

get really cautious. They go into low

59:54

risk assets or just they try to

59:56

get liquidity in order to pay off

59:59

the payoff liabilities that they might otherwise

1:00:01

have difficulty paying off. And that often

1:00:03

means that it's a flight to dollars.

1:00:05

So When the US has a financial

1:00:07

crisis, dollar goes up because people have

1:00:09

borrowed in dollars and they're worried that

1:00:11

they won't be able to source dollars

1:00:13

by selling goods and services to Americans.

1:00:15

So they have to source dollars by

1:00:17

selling assets for dollars instead. So that

1:00:19

means that reserve currencies have this weird

1:00:22

self healing characteristic where anything that's bad

1:00:24

for the issuer is actually good for

1:00:26

the reserve currency in the short term.

1:00:28

And we also, because one of the

1:00:30

things associated with having a reserve currency

1:00:32

is having a large financial sector. And

1:00:34

those financial sectors are really good at

1:00:36

reallocating capital, but they're always kind of

1:00:38

built around whatever the previous status quo

1:00:40

was. And that means they are incredibly

1:00:42

good at snapping back to that previous

1:00:45

status quo. So financial markets are just

1:00:47

unbelievably good at pretending that the last

1:00:49

big market break did not happen and

1:00:51

that things that worked a year ago

1:00:53

and, you know, temporarily seemed to stop

1:00:55

working, but then slightly recovered. They're very

1:00:57

good at pretending that those things will

1:00:59

just work just fine indefinitely. So you

1:01:01

do have anything that hurts the US's

1:01:03

status as a reserve currency, that reserve

1:01:05

status tends to direct things back to

1:01:08

normal. And then anything that

1:01:10

helps recover that status, the US

1:01:12

financial markets tend to really push strongly

1:01:14

in that exact direction. That

1:01:16

plus just, it's hard

1:01:18

to overestimate how much the world

1:01:20

just thinks in terms of dollars and

1:01:22

how much of a global default

1:01:24

dollars are for a lot of transactions.

1:01:26

This is something that, you know,

1:01:28

I look for evidence that this is

1:01:30

not true and I will definitely

1:01:32

start paying attention if large multinationals have

1:01:35

debt that was denominated in dollars,

1:01:37

it's rolling over and now they're issuing

1:01:39

euro or yen bonds because those

1:01:41

are the better deal. Did they view

1:01:43

those as the better deal long

1:01:45

term? But my suspicion is that a

1:01:47

lot of dollar to nominated debt

1:01:49

will be rolled over in dollars and

1:01:51

that the financial system will remain

1:01:53

dollarized by default. And I think what

1:01:55

a lot of people are calculating

1:01:57

is that whatever Trump does, it is

1:01:59

very hard to undo

1:02:01

the dollar as a global default and

1:02:03

to completely unwind that effect over the

1:02:05

space of four years, and that the

1:02:07

closer gets to doing that, the less

1:02:10

likely it is that anything even remotely

1:02:12

Trump will be a viable policy direction

1:02:14

for the US. And so, you

1:02:16

know, you're sort think if you're

1:02:18

thinking about reserve status, you have to

1:02:20

ask yourself not just what could

1:02:22

be done to the US reserve currency

1:02:24

issuer over long periods, but literally

1:02:26

what could get down between now and

1:02:28

the midterms because midterms will not

1:02:31

be friendly to such strategies if those

1:02:33

strategies start to actually hurt US

1:02:35

standards of living. where that makes sense.

1:02:37

We're a time zone be mindful, but,

1:02:39

But Bern, as always, this is a

1:02:42

great, great wide -ranging discussion. Until next

1:02:44

time. Already. next time. Tornberg"]

1:02:50

Upstream with Eric Tornberg is a

1:02:52

show from Turpentine, the podcast network behind

1:02:54

Moment of Zen and Cognitive Revolution.

1:02:56

If you like the episode, please leave

1:02:59

a review in the Apple Store.