0:00

How do you navigate a warming world?

0:02

CFR Education is committed to

0:05

helping learners of all ages

0:07

understand climate change and its

0:09

role in global affairs. With

0:12

over 50 free multimedia learning

0:14

resources and accompanying teaching resources,

0:17

CFR education's content explains everything

0:19

from the fundamental science and

0:22

history of climate change to

0:24

the complex societal and political

0:26

concerns that the issue raises

0:29

today. Learn more at education.

0:31

CFR.org slash climate. And keep

0:34

an eye out for three new

0:36

self-guided courses on Corsera

0:38

coming in April for

0:41

Earth Day. Welcome to the

0:43

Interconnect. A new podcast series

0:46

from the Council on

0:48

Foreign Relations and the

0:50

Stanford Emerging Technology Review.

0:52

Each episode brings together

0:54

experts from critical fields

0:56

of emerging technology to

0:58

explore recent groundbreaking developments,

1:00

what's coming over the

1:02

horizon, and how the

1:04

implications for American innovation

1:06

leadership interconnect with the

1:08

fast-changing geopolitical environment. I'm

1:10

Martin Giles, and I'm

1:12

the managing editor of

1:14

the Stanford Emerging Technology Review.

1:16

In this episode, we'll be focusing

1:18

on semiconductors and computing. There are

1:21

certain classes of problems. that seemed

1:23

like the quantum computer could be

1:25

much better, but how broad that

1:27

space is, I think is currently

1:30

unknown. What happens when you do

1:32

not have access to the latest

1:34

media hardware? That's, I think, where

1:36

Deep Seek comes in. Joining me

1:39

to talk about these key domains

1:41

are Mark Horowitz, a member of

1:43

the Review's Faculty Council, and chair

1:46

of the Electrical Engineering Department

1:48

at Stanford University. and Sebastian

1:50

Elbaum, the technologist in residence

1:52

at the Council on Foreign

1:55

Relations and Professor of Computer

1:57

Science at the University of

1:59

Virginia. Thank you both for

2:01

being with us today. Thank you

2:04

very much Martin. Thank you, Martin.

2:06

I'd like to start by looking

2:08

at what's happening with the phenomena

2:10

known as Moore's Law, coined by

2:13

Intel co-founder Gordon Moore in 1965.

2:15

Now this law holds that roughly

2:17

every couple of years, a chip

2:19

that costs the same, will have

2:22

double the number of transistors on

2:24

it, boosting its processing power. And

2:26

this scaling's been so consistent that

2:28

we've come to expect that the

2:31

cost of computing will keep decreasing

2:33

over time, or at least not

2:35

significantly increase. there are some signs

2:38

that this trend may be coming

2:40

to an end if it isn't

2:42

already over. Mark, is that right?

2:44

And what's happening here? The situation

2:47

with Moore's Law is a little

2:49

bit complicated because many people associate

2:51

many different scaling trends and call

2:53

it Moore's Law. So what Gordon

2:56

Moore really said back in 1965

2:58

was that the number of transistors

3:00

that were the most economic to

3:02

produce on an integrated circuit seemed

3:05

to be increasing exponentially. That is,

3:07

every few years you would get

3:09

twice the number of transistors that

3:12

you had before. And the key

3:14

point that he made was that

3:16

this made those transistors cheaper because

3:18

with roughly the same area you

3:21

got many more transistors. and the

3:23

cost of producing that area didn't

3:25

grow that rapidly, so you got

3:27

more transistors for the same price.

3:30

And this was phenomenal because it

3:32

meant that we could expect increasing

3:34

computation or increasing data storage, all

3:36

for relatively constant dollars, which is

3:39

awesome. So what has happened over

3:41

time is that we have continued

3:43

to scale technology through Many many

3:46

orders of magnitude factors of 10,

3:48

and this law has sort of

3:50

held. Unfortunately, while we're still being

3:52

able to increase the number of

3:55

transistors, the cost scaling that was

3:57

associated with that increasing number of

3:59

transistors is dramatically changed. In the

4:01

recent technologies, even though we can

4:04

increase density, the cost per transistors

4:06

not been falling. Some claim it

4:08

rises a little bit. I'm not

4:10

a fab. I don't know the

4:13

exact numbers, but it's certainly not

4:15

decreasing exponentially like it used to.

4:17

Got it. When you say a

4:19

fab, you know, a fabric refers

4:22

to a fabrication plant as that

4:24

makes chips. A fab in the

4:26

lingo is basically a very expensive

4:29

building that fabricates integrated circuits and

4:31

the cost of building one of

4:33

these facilities is many billions of

4:35

dollars today. Got it. Sebastian, how

4:38

do you think about war's law

4:40

and do you agree with what

4:42

Mark just said? I mean, does

4:44

it look like it's coming to

4:47

an end? Yes, I think so.

4:49

I think so. I just find

4:51

it fascinating that someone made an

4:53

observation like these 60 years ago

4:56

and it still stands, right? I

4:58

mean, I'm amazed by that fact

5:00

alone. You know, in a sense,

5:03

it is interesting because this is

5:05

really an observation about scientific or

5:07

engineering progress. It's not about physical

5:09

progress, but lasting this long in

5:12

technology is pretty amazing. The other

5:14

thing that that makes me think

5:16

about is that sometimes the slow,

5:18

this more slow... has been not

5:21

just a prediction of what will

5:23

happen, but from the other side,

5:25

it has actually set the roadmaps

5:27

of progress or expectation of progress

5:30

for some of the industry. That

5:32

it was expected at every few

5:34

years, we will see some really

5:36

major gain in performance or in

5:39

cost. So in a sense, it

5:41

plays two roles. It is one

5:43

as a predictor, but it has

5:46

also shaped. how the expectations of

5:48

some of these companies were lining

5:50

up right now with It feels

5:52

like we're really hitting some physical

5:55

limits of what the low is

5:57

predicting. I mean, you know, it's

5:59

really hard to come up with

6:01

a wire that is thinner than

6:04

an atom, right? So as we're

6:06

trying to approximate that limit, we're

6:08

facing physical limitations that cannot be

6:10

easily overcome and the cause to

6:13

get close to that limit increases.

6:15

And so we kind of had

6:17

like a yin and yang thing

6:20

between hardware and software. You see

6:22

kind of advances in the hardware

6:24

drive advances in software, so we

6:26

get more capable smartphones, smarter cars,

6:29

and the price doesn't kind of

6:31

increase dramatically from generation to generation.

6:33

Is that going to stop? Are

6:35

we all going to end up

6:38

paying now a lot more for...

6:40

you know, electronics, is the military

6:42

going to pay a lot more

6:44

for capable weapons because of this

6:47

scaling coming to an end? So

6:49

what I would say is that

6:51

as the cost of computation fell,

6:53

it became economically viable or you

6:56

had an incentive to move things

6:58

into the computation domain if you

7:00

could, because it was cheaper than

7:03

the mechanical or other alternative. So

7:05

why do cars have so much

7:07

electronics in them? Well, for two

7:09

reasons. First of all, it made

7:12

the engines more efficient, made the

7:14

brakes safer with the anti-lock braking

7:16

system. You could build things using

7:18

electronics cheaper than you could build

7:21

them in some mechanical analog. And

7:23

the other reason that cars have

7:25

so much electronics is that you

7:27

need to build a better product

7:30

next year than the product you

7:32

had previously. And it was easier

7:34

to add features. in terms of

7:37

navigation systems or other user comforts

7:39

into the car using the electronics

7:41

because that gave you increased capability

7:43

at constant dollars. So that had

7:46

really motivated a big movement into

7:48

the information domain because it was

7:50

the thing that was improving quickest

7:52

and from an economic perspective. So

7:55

the first thing that we're going

7:57

to see is additional effort in

7:59

design and other things to be

8:01

a little bit more efficient in

8:04

how you use the computing because

8:06

the cost is not decreasing. And

8:08

what you'll find is for applications

8:11

that are in high demand. more

8:13

effort will be put in trying

8:15

to optimize the entire stack from

8:17

the application through the system's services

8:20

down to the hardware because we

8:22

can't rely on the hardware basically

8:24

getting cheaper. Got it. And this

8:26

has already happened and its decrease

8:29

in cost scaling has been around

8:31

for many years already and that's

8:33

what the large system vendors have

8:35

had to do. Yeah, I mean,

8:38

total agreement. I see the shift

8:40

that Mark has been referring to.

8:42

I mean, we have seen that

8:44

for a few years already. And,

8:47

you know, I think part of

8:49

the responsibility, you know, when Mark

8:51

mentions the development stack where you

8:54

have several layers at the bottom,

8:56

you have the hardware layers, as

8:58

you move to the top, you

9:00

have more software layers. I think

9:03

there's a lot of space on

9:05

the upper layers to make up

9:07

for things that we took for

9:09

granted for granted, that was very

9:12

generous and we didn't have to

9:14

pay attention to things that we

9:16

may have to in the future

9:18

in order to keep the overall

9:21

system cost constant. Now all this

9:23

is happening at the same time

9:25

that artificial intelligence is exploding. In

9:28

January we saw Chinese startup Deep

9:30

Seek rollout an AI model that

9:32

uses lower-cost chips and less data

9:34

than other major models but still

9:37

produces impressive results. Sebastian what's the

9:39

interplay here between what we've just

9:41

been talking about in terms of

9:43

the end of Moore's law and

9:46

what's happening in AI? Yeah, I

9:48

think there is a constant push

9:50

and pull here right on one

9:52

hand the amazing advances in AI

9:55

from the self-driving vehicles that use

9:57

AI to perceive the world around

9:59

them and to operate in very

10:01

constrained scenarios to the large language

10:04

models like the CHIGPTs that we

10:06

have seen in the last couple

10:08

of years, they could not have

10:11

happened without the high performance GPUs,

10:13

the graphical processing units that we

10:15

have access to. Now, what happens

10:17

when you do not have access

10:20

to the latest Envideo hardware? and

10:22

you still want to build these

10:24

large and powerful models. That's, I

10:26

think, where Deep Sea comes in.

10:29

Deep Sea is a Chinese company

10:31

that they release a couple of

10:33

really interesting models, V3 in December

10:35

and then R1. But I think

10:38

what is impressive about them is

10:40

that they offer the performance comparable

10:42

to the models of US companies,

10:45

but with training costs that are

10:47

about an order of magnitude lower.

10:49

And that's fairly significant. Now... The

10:51

paradox here is that some of

10:54

the gains in cost that the

10:56

Chinese models have have been caused

10:58

by the constraints under which those

11:00

models were developed. So the Chinese

11:03

engineers did not have access to

11:05

the latest and greatest invidious chips.

11:07

So they had to make up

11:09

for it. And in a sense,

11:12

this constraint led them to really

11:14

clever optimization on, you know, across

11:16

the board. the hard work, how

11:18

they manage to do low-level programming

11:21

on their processors, their ability to

11:23

train in specific ways, and the

11:25

way that they actually build this

11:28

model architecture. Yeah, and to add,

11:30

if you prevent a company from

11:32

using outside equipment and it's economically

11:34

imperative for them to have this

11:37

stuff, they're going to end up

11:39

basically building the round. And that's

11:41

going to... enable them in-house to

11:43

be better at doing what they

11:46

need to do because necessity is

11:48

the mother of invention as they

11:50

say. That's a good point Mark

11:52

and we saw Deep Seek's innovation

11:55

create a significant drop in the

11:57

stock market. although it since bounced

11:59

back, but it wiped hundreds of

12:02

billions of dollars off of invidious

12:04

stock market value, invidious the US

12:06

chip giant. Sebastianio, how lasting do

12:08

you think this change is going

12:11

to be? Well, I mean, the

12:13

disruption in the market was pretty

12:15

obvious to everyone, but invidia has

12:17

operated under the assumption that models

12:20

are going to grow in size

12:22

and complexity, and you need more

12:24

of that. And basically that underlying

12:26

assumptions has been challenge. with the

12:29

emergence of deep sick. What we've

12:31

been talking about so far is

12:33

classical computing, but there's also a

12:36

lot of work going on in

12:38

the field of quantum computing, which

12:40

seeks to harness these kind of

12:42

almost mystical phenomena from quantum physics

12:45

to create immensely powerful computers. Sebastian

12:47

has been a lot of hype

12:49

and hope around quantum and its

12:51

future potential. Looking ahead over the

12:54

horizon, what do you think we'll

12:56

see here? Well, you

12:58

know, on one hand, I don't

13:00

know, it's hard not to be

13:02

seduced by something that is so

13:05

sexy as subatomic particles solving hard

13:07

large problems. And my inner nerd

13:09

just gets excited about that no

13:11

matter what you do. But I

13:13

can see it particularly appealing for

13:15

large problems that can be paralyzed.

13:18

I mean, that's exactly the kind

13:20

of the problem space that these

13:22

type of machines can solve. You

13:24

know, currently... in spite of the

13:26

successes that are reported recently by

13:28

Google and so forth, they still

13:31

work at scales that are not

13:33

commercially appealing. You know, even the

13:35

benchmarks that they have, they don't

13:37

have the business element in them

13:39

yet. So you take it with

13:41

a grain of salt, but I

13:44

think that a lot of things

13:46

are converging in favor of this

13:48

technology. You know, you see a

13:50

lot of really good researchers working

13:52

on it. You see a lot

13:55

of investment. You see companies having...

13:57

fabrication facilities already producing some chips

13:59

with this technology. So I I

14:01

wouldn't discard that we're going to

14:03

see it. I'm not good at

14:05

predicting technology. I mean, I didn't

14:08

see that we would go from

14:10

GPT 2.0 that was kind of

14:12

a preschooler level to GPT4, which

14:14

is like a high schooler level

14:16

technology. I mean, if quantum can

14:18

imitate that rate of improvement in

14:21

four years, then we're going to

14:23

see some amazing things. But I

14:25

think we need to be cautious

14:27

that understand where the technology is

14:29

today and the problem that is

14:31

solving. The problems that are solving

14:34

technically are hard, they're fundamental, and

14:36

it's not at the point yet

14:38

that it is a technology that

14:40

can be easily transitioned into a

14:42

business problem. Great, and Mark, what's

14:44

your view of quantum's prospects? I

14:47

have dabbled in quantum computing, and

14:49

I've been tracking it, and I'll

14:51

say that it's very interesting space

14:53

because it fundamentally changes some assumptions

14:55

we have about... the complexity or

14:58

how hard it is to solve

15:00

some problems. Having said that, I

15:02

think the hype around quantum computing

15:04

is maybe a little frothy. Jensen

15:06

Wang and his invidious thing said

15:08

something about not being available or

15:11

something to think about in another

15:13

decade or so. Jensen being the

15:15

CEO of invidious. Yes. You know,

15:17

I think he's probably right. I

15:19

think that there's high uncertainty in

15:21

the quantum space. The one thing

15:24

I know for certain is it's

15:26

not a panacea. Many problems can't

15:28

be computed faster on a quantum

15:30

computer than on a conventional computer.

15:32

There are certain classes of problems

15:34

that seem like the quantum computer

15:37

could be much better, but how

15:39

broad that space is? I think

15:41

is currently unknown. What are some

15:43

of those examples, Mike? Can you

15:45

just cite a couple very quickly?

15:48

I mean, what would they be?

15:50

So the example that is the

15:52

cononic thing. is basically Shor's algorithm,

15:54

which is a way of factoring

15:56

numbers or calculating certain problems that

15:58

are very hard, that are the

16:01

basis of many of the crypto

16:03

systems that we use today. So

16:05

encryption. So that's the reason people

16:07

say, you know, a quantum computer

16:09

can break all our security. But

16:11

a quantum computer... to do that

16:14

would need to be a very

16:16

advanced quantum computer. We're nowhere close

16:18

to the scale of that quantum

16:20

computer yet. Got it. Got it.

16:22

People are working on things better,

16:24

but that's still. I mean, I've

16:27

seen it said, you know, maybe

16:29

it'd be very useful for things

16:31

like designing new drugs or proteins

16:33

and maybe tracking navigation systems helping

16:35

us get faster to where we

16:38

want to go, which I would

16:40

love. But it's still too too

16:42

far out. A little bit more

16:44

speculative. The new drugs have to

16:46

do with using quantum chemistry, which

16:48

is chemical reactions are basically atomic

16:51

reactions between different molecules. If we

16:53

can simulate that more accurately, we

16:55

should be able to do better

16:57

stuff. But again, the scale of

16:59

computer that you would need to

17:01

do those calculations from the people

17:04

I've talked to is still quite

17:06

advanced. Got it. Nothing close to

17:08

what we're doing right now. in

17:10

the area of optimization there's a

17:12

lot of discussion about how it

17:14

might be better for optimization problems

17:17

but again that is still to

17:19

be determined it seems promising but

17:21

there isn't a no there yeah

17:23

got it okay Let's turn now

17:25

to the interconnection between semiconductor tech

17:27

and geopolitics. The United States has

17:30

taken steps to restrict the export

17:32

to China of advanced chips, especially

17:34

the AI ones that we were

17:36

just discussing and chip making equipment.

17:38

But doesn't the rise of deep-seek

17:41

suggest that these restrictions have been

17:43

ineffective? Sebastian, what do you think?

17:45

It's hard to judge whether they

17:47

have been ineffective because there may

17:49

have been other many companies that

17:51

they have been slow down significantly

17:54

by the lack of chips. But

17:56

really, you know, Deep Seek also

17:58

raises questions. questions about the effectiveness

18:00

of any type of government control

18:02

on the distribution of chips. On

18:04

one hand, you could say, well,

18:07

look, we just need more controls.

18:09

The controls came too late, and

18:11

we really need to control the

18:13

chips because the software and the

18:15

models advantage seems to be eroding,

18:17

given the deep sea models performance.

18:20

But on the other hand, the

18:22

constraints that are imposed by those

18:24

government controls. kind of led to

18:26

the innovation. They planted the seeds

18:28

for the innovations that led to

18:31

the creation of deep sea. So

18:33

that's where we are kind of

18:35

in this conundrum. The US has

18:37

been trying to balance these very

18:39

tricky balance where you provide enough

18:41

access to hardware and chips to

18:44

remain the leading country and actually

18:46

diffusing the technology, but at the

18:48

same time, keep them one state

18:50

behind. so you can keep your

18:52

edge. That's a very, very hard

18:54

balance to achieve. And it can

18:57

be disrupted fairly easily as they

18:59

have shown. The other downside is

19:01

the fact that the United States

19:03

has been the benefit of huge

19:05

influx of very talented people to

19:07

come in, study at American universities,

19:10

and then do great things. And

19:12

I fear that in the shutting

19:14

down or of international trade and

19:16

suspicion about countries, that we will

19:18

disincentivize these bright people from coming

19:20

to the United States. And that

19:23

I think is directly shooting ourselves

19:25

in the flood. And I think,

19:27

you know, that's not good. If

19:29

you look at many of the

19:31

startups, and Vidia in particular, were

19:34

formed by people who came to

19:36

the United States. And I wouldn't

19:38

be surprised now if, you know,

19:40

lipstick is going to be like,

19:42

or maybe it already is, it's

19:44

a kind of a stark company

19:47

in China for the next few

19:49

months, and it's going to attract

19:51

a lot of talent. Before R1,

19:53

there was this other model called

19:55

V3. The paper on that that

19:57

described that model had more than

20:00

140 authors. And those are kind

20:02

of star engineers. After publishing that

20:04

paper, after releasing R1, my guess

20:06

is that this company is going

20:08

to have access to hundreds and

20:10

hundreds of top-notch talent joining in

20:13

the ranks. And I noticed there's

20:15

industry associations in the chip manufacturing

20:17

world are saying, look, we are

20:19

looking ahead and we are seeing

20:21

a scenario in which there will

20:24

be many, many thousands of hardware

20:26

jobs not filled. We just don't

20:28

have the talent pipeline domestically. What

20:30

should we be doing more to

20:32

address that human talent issue? It's

20:34

not just about the hardware and

20:37

the software, it's about the humans.

20:39

Yeah, absolutely. I mean, we... I

20:41

think restricted or closing the door

20:43

to talent from abroad is problematic.

20:45

At several levels, first of all,

20:47

we are, you know, like Marc

20:50

said, we need that talent. We

20:52

have the chips act that has

20:54

said, look, we're going to do

20:56

fabrication and design and a lot

20:58

of things at home and we're

21:00

going to invest money to do

21:03

it, but where are the engineers

21:05

going to be to actually do

21:07

the designs of those chips? These

21:09

are not jobs that you're going

21:11

to get people displaced from other

21:14

jobs in other areas and bring

21:16

them. to produce the hardware or

21:18

the AI models. You need specialized

21:20

people with a lot of training.

21:22

So, you know, in a sense,

21:24

I feel even further the market

21:27

express, but I think we're going

21:29

to be in agreement. It's like

21:31

you cannot stop the current pipelines,

21:33

but that's not enough. We actually

21:35

need to increase them in order

21:37

to fulfill the gaps that we're

21:40

going to have in the next

21:42

few years, both in the hardware

21:44

and in the models that we're

21:46

going to be creating. Got it.

21:48

And Ms. Bassin, you mentioned the

21:50

chipsac, so chips being an acronym

21:53

for creating helpful incentives to produce

21:55

semiconductors. And that allocated up to

21:57

53 billion. for investing in semiconductor

21:59

manufacturing in the US. We're concerned

22:01

because a lot of today a

22:03

lot of the manufacturing of advanced

22:06

chips is done in Taiwan and

22:08

obviously there are geopolitical risks around

22:10

that situation. So the idea of

22:12

chips, the chips act was to

22:14

re-onsure that man, some of that

22:17

manufacturing. I gathered at about 33

22:19

billion dollars in grants and five

22:21

and a half billion dollars in

22:23

loans have already been awarded across

22:25

48 projects in 23. States, but

22:27

is that enough? Do we need

22:30

more? Do we need to move

22:32

faster? Mark, what do you think?

22:34

I would say that there are

22:36

many aspects of integrated circuit design.

22:38

The investment so far has been

22:40

mostly on ensuring manufacturing and that

22:43

makes sense from the geopolitical risk

22:45

that you've mentioned. I think there

22:47

are many other aspects in terms

22:49

of the design and how do

22:51

you get... interesting new electronic systems

22:53

is very important. And I think

22:56

that that also is just a

22:58

very expensive thing that costs a

23:00

lot of money. In order to

23:02

make money, you have to build

23:04

product in it that people want.

23:07

So I do think that having

23:09

designed talent and ideas for building

23:11

interesting new electronic systems is very

23:13

important. And I think that that

23:15

also is an area that needs

23:17

some thought and investment. One of

23:20

the things that we do when

23:22

we have researchers, we measure things.

23:24

That's the only way that we

23:26

can assess if we're moving forward

23:28

or not. And one of the

23:30

things that a lot of these

23:33

acts have is that they have

23:35

concrete amount of funding that they're

23:37

going to give. And I think,

23:39

like you're saying, this one has

23:41

given most of it in 2024,

23:43

but the targets are not clear.

23:46

Where do we want to be

23:48

in terms of securing? production of

23:50

chips in five years or in

23:52

ten years. You know, the US

23:54

in the 90s had almost half

23:56

of the production of chips in

23:59

the world. Today it has close

24:01

to 10 and most of those are

24:03

not the most advanced ones. So

24:05

do we want to go back to

24:07

the 90s? Where are we sitting in

24:10

the bar? And I think to

24:12

me that's something that needs to

24:14

be part of the discussion because

24:16

that's going to drive the

24:18

answer to your question. Are we doing

24:20

enough? Are we doing fast enough? Well,

24:23

you know, what is the goal?

24:25

And that has not be stated

24:27

as clearly in any of these.

24:29

acts or executive orders. Got it.

24:31

One last dimension I wanted to

24:33

talk about, one last strategic dimension

24:35

I wanted to talk about is

24:37

the issue of energy. You know,

24:39

it's great having advanced chips, it's

24:41

great having wonderful computers, but we

24:43

got to power them. And I

24:45

was struck last year when Microsoft

24:47

announced that it was part of

24:49

an effort to restart the three

24:51

mile island nuclear plant. It's going

24:53

to sign a 20 year deal

24:55

to take energy from there to

24:57

power its data center. How

25:00

should America be thinking about

25:02

its energy strategy in relation

25:04

to advanced compute and what

25:06

we're trying to do with

25:09

semiconductors in the future? I

25:11

think that in the clamor for

25:13

machine learning and AI, people

25:15

are talking about just building

25:18

these enormous systems. I

25:20

don't think that that's

25:22

sustainable. While communities might want

25:24

a data center for jobs,

25:26

not very many communities want

25:28

to build power plants. And

25:30

while the data center companies

25:32

may be interested in nuclear

25:34

power, that technology and

25:36

that industry has extremely long lead

25:39

times. So I don't expect

25:41

nuclear power in the next decade

25:43

is going to really make a

25:45

big difference. So for the decade that's

25:47

coming up. We have to deal

25:50

with the energy infrastructure we have.

25:52

and power will constrain the

25:54

amount of computing we can do. It

25:56

used to be as part of Morse

25:58

law, computation... got more energy

26:01

efficient as well. So as you

26:03

scaled, you got not only more

26:05

computation for the same cost, you

26:07

got that computation for nearly the

26:09

same power. So it's great. You

26:11

got more for the same cost

26:13

and energy. The power thing has

26:15

changed a long time ago. We

26:17

lost that. And that fundamentally will

26:19

limit the computation that we can

26:21

do. And so we will need

26:23

to build more efficient algorithms, because

26:26

I do not think. You know,

26:28

if you just extrapolate, you will

26:30

extrapolate that the data centers will

26:32

take, you know, 50% of the

26:34

power at some point. I don't

26:36

think that's going to happen. Just

26:38

economically, I don't think it's viable.

26:40

And so... something else is going

26:42

to give. We're not going to

26:44

scale the computation as fast as

26:46

people say. That's great. We started

26:48

with Moore's Law. We ended with

26:51

Moore's Law. Perfect. We have about

26:53

a minute left and I just

26:55

want to do a quick lightning

26:57

round of questions. If you had

26:59

to pick one person from computing

27:01

history to have dinner with, who

27:03

would it be and why? Sebastian.

27:05

This is lightning round and I

27:07

feel like I'm not lighting. I'm

27:09

not lining. Mark. You want to

27:11

go first? Yeah, I'll do touring.

27:14

Alan Turing. Alan Turing, on why?

27:16

Turing was the person who was

27:18

a crypto breaker, did early work

27:20

on theory of computation, just a

27:22

really interesting character. I'd love to

27:24

just chat with him and understand

27:26

a little bit more about what

27:28

he thinks about. Got it. Sebastian,

27:30

any thoughts? Well, I think there

27:32

are a lot of Turing Awards

27:34

that I would like that I

27:36

would have liked to meet. Among

27:39

the latest round, just starting from

27:41

the most recent, I think if

27:43

I could meet... Hasabi and Hinton

27:45

on AI, I think that would

27:47

be a pleasure. So Demis Hasabi

27:49

and Jeff Hinton who just won

27:51

the Nobel Chemistry and Physics Prizes

27:53

for work involving AI, right? That's

27:55

right, that's right. Fabulous. What's one

27:57

useful app you'd really love to

27:59

see created that you don't have

28:01

on your phone right now? useful

28:04

app I would like is an

28:06

app that tells me when I'm

28:08

going to really make a really

28:10

stupid mistake that I do often

28:12

and just warns me a little

28:14

bit ahead of time. I don't

28:16

know how to create that app,

28:18

but I'd love to have it.

28:20

A personal error predictor. That's what

28:22

Mark is talking about. Personal error

28:24

predictor. Awesome. That's great. Unlimited storage

28:27

or unlimited processing power. Which would

28:29

you choose? And no, you can't

28:31

coordinate to share. Sebastian. Processing power.

28:33

And yeah, Mark. Yeah, I would

28:35

do processing power as well. Wow.

28:37

Okay, I'm going to have all

28:39

the storage. That's great. Unfortunately, we

28:41

don't have unlimited time. But thank

28:43

you so much for joining me

28:45

today and for this terrific conversation.

28:47

Thank you. It was great to

28:49

meet you Mark and thanks for

28:52

inviting me Martin. For

28:55

resources used in this episode

28:57

and more information, visit CFR.org

28:59

backslash the Interconnect and take

29:01

a look at the show

29:03

notes. If you have any

29:05

questions or suggestions, connect with

29:07

us at podcasts at CFR.org.

29:10

And to read the new

29:12

2025 Stanford Emerging Technology Review,

29:14

visit ceter. Stanford. edu. The

29:16

Interconnect is a production of

29:18

the Council on Foreign Relations

29:20

and the Stanford Emerging Technology

29:22

Review from the Hoover Institution

29:24

and the Stanford School of

29:26

Engineering. The opinions expressed on

29:28

the show are solely those

29:31

of the guests, not of

29:33

CFR, which takes no institutional

29:35

positions on matters of policy.

29:37

Nor do they reflect the

29:39

opinions of the Hoover Institution

29:41

or of Stanford School of

29:43

Engineering. This episode was produced

29:45

by Gabriel Sierra, Molly Makanani.

29:47

Sharnafali and Malaysia at Porter.

29:50

Our audio producer is Marcus

29:52

Zacharia. Special thanks

29:54

to our our

29:56

engineers engineers and and

29:58

Brian You can

30:00

subscribe to

30:02

the show on

30:04

on Apple Spotify, Spotify,

30:06

YouTube wherever you

30:08

get your

30:11

audio. For The The this

30:13

is Martin Giles. Thanks

30:15

for listening.