0:03

It's important to delineate the two different

0:06

camps of how AI really makes a

0:08

difference today. There is a camp

0:10

of use case where AI helps us do things that we already

0:12

do better, a little

0:14

bit faster, a little bit more efficiently. Then

0:17

there are these kind of transformational applications.

0:19

These are cases where we are doing something

0:22

that we previously couldn't, or

0:24

we've embarked on doing something in a different

0:26

way that we previously couldn't. Environmental

0:29

intelligence is a revolutionary technology with

0:31

the potential to transform a wide

0:33

range of sectors. For the

0:35

energy transition, the applicability of this

0:37

technology is broad, from methane monitoring

0:39

to integrating more renewables into the

0:41

power mix. It can also

0:44

be used to reduce emissions from food systems

0:46

and in hard to abate sectors like steel

0:48

and cement manufacturing. But the

0:50

amount of energy AI will require is

0:52

also a source of much interest, uncertainty,

0:54

and concern, coming on top of

0:57

the need for more and more electricity to help

0:59

decarbonize sectors from transportation to buildings.

1:02

So what are the high potential opportunities for using

1:04

AI to combat climate change, and what are the

1:06

risks? How will AI

1:08

exacerbate existing stress on the power sector?

1:11

And what are some of the opportunities to

1:13

lower costs and to increase efficiencies? This

1:19

is Columbia Energy Exchange, a weekly podcast

1:22

from the Center on Global Energy Policy

1:24

at Columbia University. I'm Jason

1:26

Bordoff. Today

1:35

on the show, David Sandelow and Alp

1:37

Kuchuklebier. David is the inaugural

1:39

fellow here at the Center on Global Energy Policy.

1:42

He founded and directs the Center's U.S.-China

1:45

program. Previously David served

1:47

at the U.S. Department of Energy and was a

1:49

senior fellow at the Brookings Institution. He

1:51

also served as an assistant secretary of state

1:53

and as a senior director on the staff

1:55

of the National Security Council. Alp

1:58

is the co-founder and chief scientist at the National Security Council. at Farrow

2:00

Labs. He's an adjunct professor of computer

2:03

science here at Columbia University, and he

2:05

leads the entrepreneurship efforts at Climate Change

2:07

AI. David and Alp help

2:09

coauthor the Roadmap on Artificial Intelligence

2:12

for Climate Change Mitigation, published last

2:14

year in Innovation for Cool

2:16

Earth Forum. So I brought them

2:18

on the show to discuss the report's findings and

2:20

the potential for AI to drive down emissions across

2:23

a range of sectors and its

2:25

range of applications for the energy system. I

2:28

hope you enjoy our conversation. David

2:31

Sandelow, Alp Kuchukalbir, thank you

2:33

for joining us on Columbia Energy Exchange. Great to

2:35

have you both with us to talk about your

2:37

latest piece of research. Thanks for inviting us, Jason.

2:39

Pleasure to be here. Okay,

2:42

so you wrote an important report

2:44

several months ago that the

2:46

New York Times wrote up and several others have been

2:48

referring to and has been of interest to many in

2:51

the policy world about this

2:53

broad topic. Everyone is talking about artificial

2:56

intelligence, what it means for the clean

2:58

energy transition, what it means for our

3:00

response to climate change, opportunities

3:03

and risks. And so that's what I want to

3:05

talk about now. David,

3:08

we were just at CIRA week in

3:10

Houston, biggest energy conference, about 10,000 people, and

3:13

it seemed like every panel was included

3:15

some reference to AI in multiple respects,

3:17

and we'll come to different pieces of

3:19

that. But as you know, because you got

3:21

this question often, one of

3:23

those was first, how much

3:25

electricity this is all going to take.

3:27

And so decarbonization means that we're going

3:30

to increase decarbonization

3:33

means we're going to increase electricity use for

3:35

things like transportation and heat. And then also

3:37

now we have this new dimension where data

3:40

centers and training machine learning models

3:42

potentially, estimates seem to be all

3:44

over the place, is going

3:46

to take an enormous amount of electricity. Tell us

3:48

what you think about what we know about that question

3:51

today, understanding there's surely a decent

3:53

amount of uncertainty, at least. Well,

3:55

you're absolutely right about two things, Jason. First, there's

3:57

a lot of attention to this topic. And second,

4:00

There's a lot of uncertainty, but we

4:02

do know something's We know that power

4:04

demand in the United States is projected

4:06

to increase in the next several years,

4:08

and a much faster rate than it's

4:11

increased in the past several decades. To.

4:13

Be specific: in December, twenty

4:15

twenty two Us utilities submitting their

4:18

five year low growth projections

4:20

to the for the Federal

4:22

Energy Regulatory Commission. Projected.

4:24

Two point six percent increase from the

4:26

next five years. By a year later,

4:29

December Twenty Twenty three their projections. it

4:31

almost doubled systems around five percent growth.

4:34

And. Many equity analysts and

4:37

ten other research shop sir are

4:39

doing studies on this topic. I

4:42

saw one. Researcher:

4:44

What? what? When research shop the

4:46

projected that. Get. To in the

4:49

next four years. Power.

4:51

Demand from generative A I would

4:53

increase and a seventy percent compound

4:56

annual growth rate since a global

4:58

figure. But

5:00

can they. Are there other figures

5:02

that are lower than that? Ten and

5:05

underscore? Two things. First, there is significant

5:07

uncertainty in this area. with we know

5:09

that. Ten. When of a

5:11

eyes. To man discuss can increase

5:14

significantly in the years ahead. And

5:17

send desk and have an impact on and power

5:19

demand. But but. We also

5:21

know that. The. Efficiency of the

5:23

hardware that use for generative a

5:25

eyes is can improve dramatically and

5:27

that algorithmic efficiency deficiency a software.

5:29

It's likely to improve as well.

5:32

So. There's. A lot of uncertainty in

5:34

the serious. Earth to

5:36

think. It's also worth noting that.

5:38

Day. I is. By no

5:40

means the only reason the power demand

5:43

is growing of United States are under

5:45

the by the administration. There is a

5:47

significant growth of. Manufacturing The

5:49

United States Researching a manufacturing. there's

5:52

also growth of Electric Vehicles Artist

5:54

and those sectors was other for

5:56

also driving the increase in powered

5:58

men and units. Our

6:00

plan but let let's set backs are going. Listening in

6:03

to be clear about what we're talking about says lot

6:05

of I feel like a lot of times these things

6:07

gets. All lumped together. and

6:09

there's data centers for cloud computing.

6:11

There's new tools of artificial intelligence

6:14

and the amount of energy it

6:16

takes the train Large language models

6:18

so again you you you were

6:20

one of the the coauthors long

6:22

with David who chaired this. This

6:26

this road map projects on artificial

6:28

intelligence for climate change mitigation. So

6:31

several listening when we. When. We

6:33

talk about artificial intelligence. What Are we talking about?

6:36

Status. Of citizens so I

6:38

think you know a I as captured

6:40

the imagination of of the public specially

6:43

with fields and of lords language bottles

6:45

and and said Cbt is how easy

6:47

it is using so that's kind of

6:50

really treated. As said that landscape as

6:52

right right for ten and misinterpreting are

6:54

you know what month despite mean So

6:57

when we think about what's ah the

6:59

recent rise of Ai is really talking

7:01

about it is putting pressure on our

7:04

data centers that large technology companies are.

7:06

Operating and add this type

7:08

of computation has been traditionally

7:11

serving things like V Com

7:13

or sector Answer it's in

7:16

general and other tenuous call

7:18

them a commodity ten applications

7:20

and that is where we

7:23

see are enormous amount of

7:25

gross been forecast Parallel to

7:28

this is also the amount

7:30

of computation that we are

7:33

looking at coming more from

7:35

the scientific. Community or through

7:37

the realm of supercomputers. Ops of these

7:39

types of supercomputing clusters are typically time

7:42

outside of what we're talking about. This

7:44

type of hardware has been used mostly

7:46

from by scientists in academia and government

7:48

wants to relieve some kind of continue

7:51

studying the climate and as on so

7:53

forth. So when we talk about a

7:55

i've really talking about any type of

7:58

computer applications that is a pro. The

8:00

meeting on what we think as far

8:02

as actions and to are planning and

8:04

other types of activities that we associate

8:06

with you Intelligence and so the most

8:09

recent incarnation is large Language models are

8:11

but A I is not limited to

8:13

are selling these models and in our

8:15

report we do talk about applications of

8:18

Ai across a variety of factors are

8:20

be to power sector manufacturing, agriculture out

8:22

where it's thought to start Swings models

8:24

makes in this kind of in a

8:26

big difference or so when we are

8:29

really kind of. Trying to explore what

8:31

the energy demand for a I look

8:33

like ah, that is part of uncertainty

8:35

went largely which ones are very new,

8:38

were exploring where they can be helpful

8:40

style, were trying to anticipate demands for

8:42

this particular type of A I and

8:44

all of us factoring into the uncertainty

8:46

in these are forecasts. And

8:49

just on the question I asked David, your

8:51

your take on to give people a sense

8:53

of how much electricity it really takes to

8:55

trade and some of these models were. as

8:57

you said, we're just getting started. I suspect

9:00

like with the Internet, me early days of

9:02

a digital technology revolution would barely scratched the

9:04

surface for the use cases that we will

9:06

find for tools like a I. So you

9:08

see some projections and are really quite staggering.

9:10

not just going from two percent to three

9:13

or four percent like David was talking about,

9:15

but dramatic growth and electricity use is that

9:17

is that? What's your. Best sense of where

9:19

we're headed right now. Yeah. I

9:22

know less. Concerned I

9:24

would say of seeing that kind

9:26

of exponential growth for of course

9:28

as possible doll tell you why

9:31

I'm in a people frequently think

9:33

about Crypto as an equivalent industry

9:35

that drew very quickly and his

9:37

energy demands has remained extremely high.

9:40

And that's by design. So the

9:42

crypto kind of efforts. this entire

9:44

activities the societal activity is driven

9:46

face also requiring computation, the whole

9:49

idea of mining currencies requires competition

9:51

speak to into the activity itself.

9:53

I'm a I, especially with

9:55

this enormous can largely English

9:57

model huge data Leno printing.

10:00

A form of Ai is new

10:02

and it is not fundamentally based

10:04

on competition. In what I mean

10:06

by that is me. see already

10:08

the academically, the scientific community working

10:10

towards reducing the energy required to

10:12

achieve similar outcomes. In a I

10:14

write this is it's a statement

10:17

I can't make and Crypto we

10:19

we. We haven't really found a

10:21

way of saying let's do the

10:23

Ifc if equivalent value from crypto

10:25

but at a lower and is

10:27

you footprint on is just to

10:29

fundamentally not possible. So where I am

10:31

hopeful is I'm seeing a lot of

10:33

software development that will reduce the amount

10:35

of energy required to see somewhere else

10:37

comes and seeing a lot of hardware

10:40

or development where hardware manufacturers are thinking

10:42

about new electronics or new chips and

10:44

new computing set ups or that will

10:46

reduce the amount of energy those required

10:48

to achieve rt somewhere else comes meaning

10:50

that flights the next year, every generation

10:52

of and video chips or whatever or

10:54

whoever is making these you get you

10:57

can get as much computational power are

10:59

more computational power. The same with the same

11:01

energy input, they just get more efficient overtime as

11:03

part of the trendier you're talking about. As I

11:05

said, Piazza is a think about how competition with

11:07

like in the nineteen seventies and eighties or we

11:10

can replicate if not for exceed the supercomputers of

11:12

those decades. and with the devices are pockets to

11:14

drop or died in the how many orders of

11:16

magnitude less energy than those that back of a

11:19

day. And just so everyone knows and I've come

11:21

to David the minute. But I think people have

11:23

been following this podcast or the Central Banerjee for

11:25

a while know David and his backgrounds. Of course

11:28

I'll see if your bios at the start. of

11:30

this podcast but sometimes from someone's title

11:32

you don't get a sense of it

11:34

just explain for people listening the work

11:36

that you do ah day to day

11:39

in and in computer science i were

11:41

to have some in the countryside apartment

11:43

at columbia an adjunct faculty it's am

11:45

i teach a class called the shimmering

11:47

and the climate and i research a

11:49

branch of machine learning that i like

11:51

to call explainable machine learning this is

11:54

a contrast to the tenant black boxes

11:56

in learning technology that has really catapulted

11:58

the or he kommersant accounts sectors

12:01

to drive the wonderful applications that we have,

12:03

large language models being one of them. Explainability

12:07

seeks to combat the issue that we

12:09

have now recognized, for example, with large

12:11

language models, the issue

12:13

of hallucination. The inability to

12:15

explain why the program is

12:17

delivering the output that it is delivering, given

12:20

a specific query. So

12:23

this is an exciting area that has

12:25

applications that I've been

12:27

exploring in the manufacturing sector in

12:29

my other hats that I wear, which

12:32

is as co-founder and chief scientist of

12:35

a company called Ferro Labs, which

12:38

builds factory optimization

12:40

software that's powered with AI. So

12:43

we sell software into the steel,

12:45

cement, and chemicals sectors, which

12:47

are by share the

12:49

largest emitting sectors in

12:51

heavy manufacturing. And

12:53

our software is being used to help

12:55

them reduce their energy consumption, reduce their

12:57

waste, and reduce the variability

13:00

of their production. David,

13:02

just coming back to you on some of

13:04

the numbers you gave and what you heard

13:06

Alp say about these offsetting

13:08

factors of significant growth in electricity potentially,

13:11

because we're going to have more and

13:13

more applications for AI. And on the

13:15

other hand, the technology gets better and

13:17

more efficient. And we

13:19

haven't seen electricity demand

13:21

growth in the US for quite some time.

13:24

It's been pretty flat. Getting

13:26

ready to meet that demand, the

13:29

kind of permitting of infrastructure, transmission

13:32

lines, which is so much a topic of discussion now.

13:35

And so we're trying to meet that growth

13:38

in demand at the same time that we're

13:40

putting more and more intermittent sources of electricity

13:42

onto the grid to meet the challenge of

13:45

decarbonization. How do you think about

13:47

our ability to handle and manage that right now?

13:49

What do we need to be doing from

13:52

a policy standpoint or otherwise to meet that

13:54

challenge? It's a very significant challenge,

13:56

Jason, and we need to pull together

13:59

all the forces. within

14:01

the policy world and the technical world to try

14:03

to meet it. I think AI

14:05

offers some helpful tools, interestingly. So

14:08

AI can help us, for example, optimize

14:10

production from solar and wind farms. That's

14:13

pretty well-established technology, actually, because AI is

14:15

very good at predicting patterns. And so

14:17

we need to deploy the technology to

14:19

help optimize the production from clean energy

14:22

using AI tools. But

14:25

AI can help us optimize as well.

14:27

It can help with optimal power flow

14:30

problems and its siting of transmission lines

14:32

and other issues like that. And

14:35

AI can help interpret in

14:38

interesting ways. There's, for

14:40

example, some companies have put the

14:43

databases from FERC orders and

14:45

NERC orders, Federal Energy Regulatory

14:47

Commission and Nuclear Energy Regulatory

14:49

Commission online. And

14:52

that allows querying of those

14:54

databases in ways that may

14:56

facilitate permitting over the years

14:58

ahead. So I think

15:00

we need to be ambitious and

15:02

creative about using AI tools to

15:04

help get over the challenges

15:06

that we're facing right now in managing the electric

15:09

grid and growing the electric grid to address these

15:11

problems. And

15:13

then things like basic permitting

15:15

reform, which has been almost

15:17

over the hump in the US Congress in

15:21

the past years. If we could possibly get

15:23

that over the hump, that would make a big difference as well. Yeah.

15:27

And so I want to come to those opportunities because that

15:29

is what was the focus of the report that you did.

15:31

Obviously, there's a lot of interest in how much additional electricity

15:34

demand these technologies will create. But as

15:36

you say, there's an opportunity to help

15:39

with integrating more low carbon sources of

15:41

electricity on the grid or

15:44

understanding other issues we need to understand like

15:46

emissions. So let's just talk about some

15:49

of those. I think

15:52

greenhouse gas emissions monitoring was the first

15:54

one you talked about. Maybe

15:57

start with David and then go to Al.

16:00

you could talk a little bit about we're

16:02

putting satellites up into space, we're trying to

16:04

better understand where emissions, methane leaks

16:07

are coming from, trying to get better granularity

16:09

and what greenhouse gas emissions actually look like

16:11

year to year. How will

16:13

AI change that and

16:16

how much of a difference will it give us a sense of

16:18

like magnitude? Like is it kind of helps a

16:20

little bit or is this a real sea change in our

16:23

ability to pay attention to emissions? I

16:25

think it's a sea change. I see it

16:27

all things but I think that we have

16:29

the ability right now to

16:31

understand greenhouse gas emissions in real time

16:33

in a way that historically we haven't.

16:35

So right

16:37

now vast amounts of data are being thrown

16:39

off of satellites, aerial

16:42

monitoring, drones, ground-based sensors

16:45

and that data is incredibly

16:47

useful but there's no way of interpreting and

16:49

understanding it without machine learning and AI tools.

16:52

Historically we've relied upon voluntary self-reporting to

16:55

some extent fossil fuel data consumption analysis

16:57

in order to understand the emissions that

16:59

are coming from individual plants. I

17:02

think in the decade ahead the combination of

17:04

this new sensor data

17:06

plus machine learning may give us the

17:08

opportunity to really understand

17:10

at a granular level what's happening in different

17:14

places and in fact this is already

17:16

making a difference in methane emissions policy

17:19

as I think you alluded to Jason. The

17:22

global methane pledge and the commitment to

17:24

reduce methane 30% by 2030 and all

17:26

the activity

17:28

that's going on around methane

17:31

policy right now really wouldn't be possible without

17:33

these machine learning and AI tools

17:35

and in the

17:38

years ahead with spectroscopy and other types

17:40

of tools I think we can actually

17:42

make a difference with

17:44

these machine learning and understanding where greenhouse gas

17:46

emissions are coming from and therefore in policy

17:50

development. But you

17:52

can help elaborate on that. We have a lot

17:54

more data coming in on Methane emissions

17:56

for example. And as David said, the tools

17:58

we've been using historically. The to

18:00

measure greenhouse gas emissions are imperfect.

18:04

What's different about a world where they I

18:06

tools them would have been true a decade

18:08

or so ago. to move to use all

18:10

of that data and make sure that we

18:12

have a clear understanding of what the emissions

18:14

picture looks like. An. It's A

18:16

it's a great pointers adjacent dataset. The ground

18:18

worked really well here so it's important to

18:21

delineate anger the two different chance of how

18:23

Ai really makes a difference. Today there is

18:25

a camp of use case where A I

18:27

helps us to things that we already do

18:29

a better. A little bit

18:31

faster, a little more efficiently we're we're

18:34

always talking about, let's say, ten to

18:36

twenty percent gain in one in whatever

18:38

were achieving. Then there are these kind

18:41

of transformation applications are these are cases

18:43

where we are doing something that we

18:45

previously couldn't. We've embark on doing something

18:48

in a different way that we previously

18:50

couldn't and greenhouse gas emissions monitoring falls

18:52

his the second attempt. A transformational camp

18:55

So. It. Doesn't

18:57

come as a surprise I'm sure if any of your listeners

18:59

out one. At. Property of Ai

19:01

is it's ability to ingest enormous

19:03

amounts of data and enlarge language.

19:05

Models are already articles in Britain

19:07

Today, and major news outlets that

19:10

accompanies that that develop are two

19:12

things. One is not having enough

19:14

information to feed them after having

19:16

said the entire internet to these

19:18

models sister that statement alone as

19:20

is extraordinary Up so. A

19:23

decade ago, we could write software that

19:25

had to be extremely ah that was

19:27

extremely expensive and difficult to build to

19:30

integrate data coming from the Sirius Satellite

19:32

systems, let alone interbreeding data from different

19:34

sources of data such as rose and

19:37

measurements on the ground. and putting all

19:39

that together into one picture into one

19:41

system that allows us to just objective

19:43

we say. here's the big picture up.

19:46

A I is ideally suited to be

19:48

able to do that at scales to

19:50

do that in a way that. Allows

19:53

us to really signs the

19:55

signal from the noise and

19:57

really gives this thrust into.

20:00

A One source of truth right? The

20:02

actual ground truth of where are we

20:05

have it seems how much are we

20:07

a meeting? when was it emitted in?

20:09

I'm really getting all stakeholders around us

20:12

to to agree and say like yes,

20:14

this is the source of truth That

20:16

was not possible with a I perform

20:18

and so I would argue that that

20:21

is the linchpin in taking all of

20:23

these extraordinary that. Hardware Advances

20:25

where we've developed the satellites, we developed

20:28

the sensors that we've collectively managed to

20:30

get them into orbit. Drones Crown Sensors

20:32

are to then say here's how we're

20:35

going to bring an altogether. So

20:38

does that as helpful frame that sort of

20:40

transformational or or call a marginal about am

20:42

ten to twenty percent change that as a

20:44

that back to be a big number two

20:47

right now. So those are the thought of

20:49

small same. It's just that framing is is

20:51

helpful. So from a standpoint of greenhouse gas

20:53

accounting David. How

20:56

does that change The the way we engage

20:58

with the challenge of climate change, the policy

21:00

response and international coordination does just give us

21:02

a higher degree of confidence and unearthing mean

21:04

your expertise. This time I don't think we're

21:07

going to some the learns and is not

21:09

the largest emitter. Maybe the numbers get revised

21:11

slightly? I'm or is it something like that

21:13

ability to really pinpoint where the methane leaks

21:15

are coming from? That in of simply wouldn't

21:17

have been possible with all the data coming

21:20

from satellites etc. Without this technology. How does

21:22

it? How does a change What we know

21:24

about the problem. Was. So full

21:26

of ten ten percent of greenhouse gas emissions

21:28

by good terms of we have a at

21:30

a ten percent difference at a depth of

21:33

that's a heck of a lots of or

21:35

I think it did makes big difference and

21:37

in for example messing messing regulation a me

21:39

or are we we have learned over the

21:41

course the past ten years what we didn't

21:43

know more than thirty to go for bar

21:46

for example of it's a huge amount of

21:48

methane emissions are coming from super emitting events

21:50

or from in a large releases and At

21:52

and that affects the policy development and and

21:54

the response. And they're simply. Central I think the

21:56

method control which is going to be fundamental to address

21:58

in the Bring Us Desktop. So that's

22:00

the thing one example, but you

22:03

know another area I think is

22:05

in research and development and innovation

22:07

and One area

22:09

that I'm particularly excited about in a

22:12

transformational standpoint is in materials innovation So

22:14

I recall I recall

22:17

that when I got to the US Department of Energy

22:19

in 2009 I remember

22:21

receiving a briefing from the staff

22:23

there saying that offshore wind

22:25

would probably never be feasible because the

22:28

Marine environments too corrosive and the steel that you'd

22:30

have to put out there couldn't really withstand the

22:33

corrosivity of the marine environment My

22:35

understanding is that as the

22:37

result of materials innovation super lightweight?

22:40

Materials and materials that withstand the

22:42

corrosivity marine environment. We've now moved dramatically forward

22:44

in terms of our ability to deploy offshore

22:47

wind artificial intelligence

22:49

can make an enormous a difference

22:51

in terms of accelerating the pace

22:53

of materials development and just

22:56

by way of example that when Thomas Edison was Impreting

22:59

the modern light bulb 150 years ago He

23:03

took months to take dozens of different types of

23:05

materials and run electric charges them to find out

23:07

what would happen Today with AI

23:09

tools we can simulate a million of those types

23:11

of interactions in a second And

23:14

that allows us to to both

23:16

doc unselect much more quickly and choose

23:18

among different materials find out what's best

23:20

But it also actually allows us to

23:22

test materials that don't actually exist right

23:24

now, but might exist using

23:27

simulation and chemical structural constraints

23:30

And then if you have a hypothetical

23:32

material seems to have good properties synthesize

23:34

that create it and move forward so

23:37

So I think this is going to affect research

23:39

development budgets and research and development agendas around the

23:42

world on clean energy Dramatically in the years ahead

23:44

now Maybe you could comment on that that point

23:46

David made about about materials and particularly with your

23:49

as you said your background is the Fero labs

23:51

talk a little more about how that relates

23:54

to these challenges For

23:57

example in hard to abate sectors Absolutely.

24:00

So I recently learned I want to

24:02

share this on the spot. Cast fat

24:04

or Thomas had this is seen as

24:07

sir Expression Nova series is one percent

24:09

a inspiration and nine nine percent perspiration

24:11

or is it feels talks about this

24:14

point but actually is attributed to Arcades

24:16

Sanborn I was Americana author, teacher and

24:18

and lecture it's miss attribute it to

24:21

or Edison. Nevertheless it does capture Edison's

24:23

method and so ah let's let's talk

24:25

about the the Materials Science how material

24:27

science a intersects with the energy transition.

24:30

Arthur We if we think about

24:32

our lithium ion battery technology. Ah

24:34

where we've come from the nineteen

24:36

seventies where we are initially identified

24:38

certain lithium based materials for and

24:41

you know anodes, cathodes, A on

24:43

electrolytes. Ah, we have improved on

24:45

this issue. See of are those

24:47

batteries in terms of the chemistry

24:49

the design over the that the

24:51

next fifty years. Ah what A

24:53

I really allows us to do

24:55

from a transformational perspective is to

24:57

say how do we do that

24:59

Cst year process. Of making schools

25:01

him ion batteries more efficient had me

25:03

compress that. The five years for sodium

25:06

ions authorities for a solid state batteries

25:08

for the next type of energy storage

25:10

and method that we're going start exploring.

25:13

How do we do this or rapidly

25:15

I'm that is are enormous, right? Wearing

25:17

a race against time here and the

25:20

ability to use a I to quick

25:22

me accelerate the progress of science is

25:24

is enormous sum. But since for me

25:27

so opportunities as you describe Jason do

25:29

not lie. So we ain't just

25:31

these can I had picked cherry

25:33

picked up sectors are there are

25:35

applications of these across the board.

25:37

so give an example in arms

25:39

deal which is considered earth hard

25:41

to beat kind of have you

25:43

any faction sector. I'm alone responsible

25:45

for anywhere between sixty percent of

25:47

the global carbon footprint. Some electrifying

25:50

skill or involves are using more

25:52

and more scrap metal as the

25:54

seed stock so steel manufacturing are

25:56

traditionally has been are designed to

25:58

take version material mind from. The

26:00

category of process knows a hike. while

26:02

the steel to electrify seal in a

26:04

very kind of straightforward way you want

26:07

to use old cars, owns appliances, a

26:09

scrap metal front rail or as your

26:11

feedstock use electricity, Smelt that and make

26:14

you feel the issue. There is every

26:16

batches steel that you meltzer today it's

26:18

a bunch of Honda's them as a

26:21

bunch of Chrysler's slightly different and so

26:23

what feel manufacturers you currently is that

26:25

the reduce the top the amount of

26:27

recycled feel that the use a chopper.

26:30

Twenty Five percent, Thirty percent Be so.

26:32

rely on the high quality pure ingredients

26:34

to make high quality steel. While was

26:36

a I actually coming in and giving

26:38

guidance the operators every five minutes, every

26:40

ten minutes for this, that's a steal.

26:43

The Sally need top rate your plants.

26:45

Dimitri get the high quality product or

26:47

we see seal manufacturers are United States

26:49

been able to push the boundaries of

26:51

how much we cycled Feel they're using

26:53

Fifty percent, Seventy five percent, eighty percent.

26:56

Even for high quality, it's a high

26:58

grade steel that they need to produce

27:00

for a variety of from applications that

27:02

transformation it right. So that sir, application

27:04

of Ai where you're doing something, you're

27:06

operating your factory in a way it's

27:08

the previews for you couldn't and without

27:11

the psychology you. Can't. Do the

27:13

real really interesting be so again for me

27:15

Back to what saw. Ten or

27:17

twenty percent. What makes a big

27:19

difference on what you might call transformational?

27:22

You talking to report about the impact

27:24

this just as on the power sector

27:26

more broadly. the ability to manage our

27:29

demands are energy efficiency, what it

27:31

means for renewables, stuff, weather better predictor

27:33

maintenance things like that, So talk a

27:35

little bit more about what. This.

27:37

Technology We start about how much more electricity

27:40

could use, but as you said earlier, David

27:42

I'm sort of in passing Man, it'll also

27:44

can help help with how we build this

27:47

grid and hopefully make a cleaner. Say more

27:49

about that and and is that. For

27:51

that. And twenty percent or is that transformational

27:53

my guns. And twenty percent a big number sauce is.

27:55

hop on a dumpling that but you know what I

27:58

mean. Like how big is? give us a sense. to

28:00

the magnitude of the impact. It's

28:02

very important. We didn't try to quantify

28:04

it in our report, but every stage

28:06

of the power sector can be significantly

28:08

affected and actually is already being affected

28:10

by this technology. I've already mentioned that

28:13

solar and wind farms use

28:15

AI today frequently in order to better predict

28:18

the solar and the variable solar and wind

28:20

resource and maximize output.

28:24

Any sighting of electricity generation

28:26

assets can benefit from AI

28:28

technologies in terms of both

28:31

weather and power demand in the area,

28:33

a variety of other different factors. Geothermal

28:38

power can benefit a lot from AI in

28:41

terms of understanding subsurface conditions and

28:43

it's definitely important for

28:45

the development of geothermal power.

28:49

One area there's a lot of interest in right

28:52

now is in nuclear power innovation, the ability of

28:54

AI to simulate

28:56

and do what's called digital twinning. Dramatically

29:00

accelerate the pace at which we understand

29:02

new nuclear technology. So I was having

29:04

to invest hundreds of millions of dollars

29:06

in building, or billions of dollars in

29:09

building new facilities. And

29:12

then there's transmission. On

29:15

the transmission side, AI can help with

29:17

what's called dynamic line rating and other

29:20

types of ways of improving the operation

29:22

of transmission lines. At

29:25

the end use stage, but AI

29:27

can certainly help with building energy

29:29

efficiency, better understanding patterns within a

29:31

building and energy use patterns. And

29:34

then with virtual power plants, right

29:37

now we have distributed resources

29:39

around the electric grid. Machine

29:42

learning AI tools are really fundamental to

29:45

using those and vehicle

29:47

degree technologies, which I think is a

29:49

hugely important area in VPPs

29:53

and virtual power plants are

29:55

gonna depend upon AI technology. So

29:57

basically all across the power sector, this technology...

30:00

can make a big difference. And

30:02

it's already starting to happen, but it's gonna, I

30:04

think, progress dramatically in the years ahead. Now,

30:07

anything you wanna add to that? And then if you

30:09

could kind of, for people for whom, when we talk

30:11

about AI, I may have used chat GPT for, or

30:13

something, but it's still somewhat abstract

30:15

concept, like maybe a concrete example

30:17

or two on the generation side or the demand

30:20

side to help people understand how this is really

30:22

gonna be deployed. Yeah,

30:24

so I think maybe that point is worth, you

30:26

know, digging a little bit deeper. And maybe I'll start with

30:29

a little bit of abstraction and give a little

30:31

bit of a more specific example as we go along,

30:33

right? So AI can do a variety of things, and

30:35

it's important to distinguish between them.

30:38

So one is just sifting through large

30:40

amounts of data. Think about this as

30:42

kind of pattern recognition. It's kind of

30:45

search, search done in a way that

30:48

you have messy data, you have a lot of data.

30:51

It's unstructured. The work

30:53

hasn't gone in. It physically cannot go in

30:55

to make it structured. AI can go through

30:57

this. And identify those patterns. Another

31:00

set of things that AI can do is forecasting,

31:02

predicting. Frequently these are used

31:06

interchangeably. Forecasting typically involves some notion

31:08

of time, thinking about a time

31:10

horizon over which you're forecasting. Predicting

31:12

might just be, you know, the

31:14

next hour. It might be just

31:16

simulating some sort of scenario and

31:18

predicting what would happen under that

31:20

scenario. That's another camp

31:22

of activities there. And then the

31:25

most complex, but arguably, when

31:28

applied well, the most high value add is

31:30

optimization. So this is when you're giving

31:32

a very complex problem

31:35

that is very difficult to

31:37

solve with traditional software methods.

31:40

AI frequently can either come up with

31:42

approximations so that they're fast, that

31:45

help solve the true

31:47

problem, the hard problem, exactly. Maybe an

31:49

approximate solution is fine. So

31:52

that's another example. So optimal power flow,

31:54

I would say, is one of those

31:56

cases where you have a... Let

32:00

me put it this way, you have a

32:02

problem space that would make the typical champions

32:05

of AI very afraid. So

32:08

the folks who are in the e-commerce and

32:10

the ad tech and the technology space, they

32:13

have some trepidation trying to solve problems like

32:15

these, really hard problems. You've

32:17

got physical constraints of how energy

32:19

is going to flow over a

32:22

particular grid, network, or topology, and

32:24

you need to satisfy physical requirements. You

32:27

can't just say, oh, chat,

32:29

GPD, you've made your best estimate of going,

32:31

what's good? No, certain equations of physics

32:33

need to be satisfied if you're going

32:35

to adopt this type

32:37

of technology. So AI is actually quite

32:39

good at once you kind of

32:42

iron out the wrinkles and really figure things out

32:44

of answering questions like, okay, if

32:46

I have variable supply or

32:49

I have some sort of un-modeled demand,

32:52

how do I use this existing network to

32:54

optimally route my power? David,

32:57

you mentioned how it can help with geothermal map

33:01

the subsurface to directional drilling more

33:03

accurately. I take it all of

33:05

those kind of are

33:07

consistent with the point that this technology

33:09

can be transformational to advance clean energy,

33:11

but I presume also oil gas as

33:13

well. No question

33:16

in the oil and gas industry is using it

33:18

pretty extensively today for all

33:20

kinds of purposes and

33:22

has been for a number of years. And you

33:25

also in the report, we don't

33:27

work quite as much on this at the

33:29

energy center, although Columbia does more broadly and

33:31

it's very important climate issue. What

33:34

impact this technology could have on food

33:36

systems globally? I was wondering, David, if

33:38

you could talk a little about that

33:40

and then Alp. You

33:42

got some very important applications here. So

33:45

first of all, AI

33:47

tools can help in addressing

33:49

the impact of the food system on

33:52

the climate for reducing greenhouse gas emissions,

33:55

the food system. For example, optimizing

33:57

the application of nitrous. oxide

34:00

in fields and fertilizers

34:03

that will emit nitrous

34:05

oxides from fields. And

34:09

also in developing new crops, new innovative crops that

34:11

may have better properties with some of the same

34:13

type of simulation tools that we were talking about

34:16

before. And then

34:18

absolutely AI tools can help in better predicting

34:20

weather patterns and other types of climate phenomenon

34:22

that will have an impact on the food

34:24

system and may damage the food system. So

34:27

anything to add to that? Again, kind of giving

34:30

people an example, maybe a use

34:32

case there? Absolutely. So

34:35

again, keeping our framework in mind, you've got

34:37

the kind of 10, 20% doing things better.

34:40

So this is how do I use less

34:43

fertilizer to achieve the same crop yields? How

34:46

do I use better forecasting to not waste

34:48

the amount of how I'm using my land?

34:50

How do I use my land a little

34:52

bit more optimally to reduce

34:54

waste and things like that? How

34:57

do I integrate data from different sensors? Now you're

34:59

no longer talking about satellites, but maybe you have

35:03

five ground sensors and two drones. How do

35:05

I get the data between those sources together?

35:09

Then you've got the transformational applications.

35:12

How do I build a more

35:15

drought resistant strain of a particular

35:17

grain? How

35:19

do I think about a

35:21

more heat resistant alternative of

35:23

a specific staple? These

35:26

are the more transformational, right? We can do this

35:28

today. We do it trial and error, it takes

35:30

time. How do we get ahead of that? When

35:34

you think about, I mean, this applies to everything we've been

35:36

talking about, but as you were

35:38

describing that, particularly for agriculture, you think

35:41

about the potential, but then you think

35:44

about the access that any individual landowner

35:47

has to that kind of information.

35:50

Well, you're describing sounds like pretty

35:52

high tech stuff, and you're talking

35:54

about, in some cases, large multinational

35:57

companies, but often... individuals,

36:00

families. And then when you think about around the

36:02

world, where agriculture is done, where

36:04

the emissions are coming from, kind of raises

36:06

the question of the accessibility

36:09

of these tools you're talking about. How

36:12

restricted will they be? How complex can

36:14

they just be on someone's mobile

36:16

device? Or is it much harder

36:18

and more expensive than that? And

36:21

what are the risks for kind

36:23

of how we think about what

36:26

this could mean? Everything you're talking about, food or

36:28

otherwise, could this be transformational,

36:30

say, in wealthy or developed countries and

36:32

leave others behind? That's a key

36:34

point. How long have you got, Jason? So we've

36:40

been talking about a number of

36:42

tremendously high potential

36:44

applications for artificial intelligence and addressing climate

36:46

change problems, energy system problems. None of

36:48

those outcomes are inevitable. There are barriers

36:50

to achieving all of them and their

36:52

risks, as you started to point out.

36:54

So just to talk about the barriers

36:56

for a minute, we

36:58

highlight two barriers

37:01

on our report as being probably the most

37:03

significant, people and data. And with

37:05

respect to people, none of this happens unless

37:07

we have people who were trained across a

37:10

whole range of disciplines. We

37:12

certainly need the computer scientists who can

37:14

develop the high, the algorithms

37:16

that do this type of work. But

37:18

we need a lot more than that, actually.

37:20

We need climate experts who understand enough about

37:23

AI to understand how their

37:25

field can benefit from the application of AI

37:27

tools and the same in

37:29

the energy system. And we just need

37:31

people generally to understand

37:33

how this type of work can integrate

37:35

into their institutions. Actually,

37:39

we recommend in our report

37:41

that every institution with a

37:43

role in climate change mitigation

37:45

have a top advisor for

37:47

AI to the CEO or to the

37:49

minister. And we're really pleased that last

37:51

week the Biden administration announced that every

37:54

federal agency will have a chief AI

37:56

officer. And that's exactly, I think, the type of

37:58

direction that makes the sense

38:00

in addressing the people issues around this. But

38:02

then there's also data issues, and

38:04

you were starting to get this at your question. Making

38:07

sure that there is available and

38:09

accessible data to prepare

38:13

these AI models, to train these AI

38:15

models is going to be incredibly important.

38:17

Then making sure that the results are accessible to people

38:20

of different types all around the world, is going to

38:22

be key as well. A lot to say about that.

38:25

Yeah. Alp, anything you want to add? Either

38:27

a particular question of exacerbating or

38:29

narrowing North-South divides, or more broadly,

38:31

some of the risks David talked

38:34

about. Yeah. Maybe start with

38:36

the risks and then on a higher note.

38:38

On the risk side, we're talking about almost

38:40

every segment of the economy, and we're talking

38:42

about applications that we can really transform how

38:44

we do things. So human

38:46

health and safety is typically a

38:48

concern that AI has not had, or the sector

38:50

in general has not been applied to problems where

38:53

that is a concern. Security,

38:56

applying technology like AI to

38:58

the grid, involves security concerns

39:00

that go beyond the

39:02

current applications of AI. All of this needs

39:04

to be top of mind. These are

39:07

legitimate risks of the adoption of technology

39:09

like AI at scale. You're talking about

39:11

cybersecurity risks. Correct. It can exacerbate those.

39:13

Okay. Correct. But on a positive note,

39:15

if you look at the barriers, there

39:17

is a world in which the exploration

39:20

of what is needed to

39:22

make this technology productive is

39:25

done in wealthier nations that allow

39:28

developing nations to leapfrog, and

39:30

not waste the time that's needed to explore

39:33

and develop this technology on their

39:35

own. So if best practices are developed in

39:37

terms of the application of AI, let's say,

39:39

to agriculture, could that help

39:42

accelerate the development of developing nations

39:44

to get to those results quicker?

39:47

Alp, when you, again, thinking about what will

39:49

have a smaller or larger difference, I mean,

39:51

we've talked about a number of things just

39:53

in the last 30 minutes. If

39:56

you had to spotlight, And

39:59

there are. There. Are many

40:01

potential impacts on clean energy on arms

40:03

and climate change, but there are really

40:05

one or two that kind of strike

40:07

you as being order of magnitude larger

40:09

than the rest is is that true

40:11

or not And and if so, just

40:13

help people listening understand what might be

40:15

the biggest thing in this. In

40:18

Obe. somewhat. Extensive

40:20

list of potential impacts and opportunities

40:22

you identify. The. Here's how

40:24

I think about it. so if you

40:26

think about that tenor of cult marginal

40:29

the tenth when he presents a kind

40:31

of benefits ah I just apply that

40:33

to the size of the pie. so

40:35

a slice of the pie that the

40:37

sector occupies. So in that regard I

40:40

see an hour and manufacturing as being

40:42

the to sector is that will benefit

40:44

net numbers the most from technology that

40:46

be no works right now and we

40:48

just need to get out. We need

40:51

the right incentives, we need to write

40:53

a bureaucracy. We need the rights, supports

40:55

mechanisms to deploy this technology at

40:57

scale and is sold. Me read

40:59

those benefits and that is posted

41:01

on the transformational sides father's higher

41:03

uncertainty and I don't know if

41:06

material size innovation the lead to

41:08

way our sarbanes that will make

41:10

carbon capture tenth and super or

41:12

ten times more effective or that

41:14

will make car sodium ion battery

41:16

technology hundred times more efficient in

41:18

the next five years. but is

41:20

it does. These.

41:22

Will have use impacts

41:24

so these transformational applications

41:26

of a has higher

41:28

uncertainty but potentially kids.

41:31

Have a higher impact if they're successful To. Is.

41:33

That david how you see a to be

41:36

these are all it's exciting opportunities but can

41:38

to save as ones that psyche was the

41:40

biggest your back with my days after that

41:42

would opt of said wife had to guess

41:44

there's potential for tremendous transformational benefits and them

41:46

materials innovation space we don't have that will

41:48

happen fi uncertainty ah but type of a

41:50

very high reward of the does and spot

41:52

but I just underscore the pointed these. So.

41:54

Called incremental. the improvements we've

41:56

been talking about wraps the jews in the

41:59

context of of climate energy policy. If we're talking

42:01

5 to 10 gigatons of production zero, that makes a

42:03

big, big difference. And I just,

42:05

let me play devil's advocate and just, you know, push

42:08

you on some of this because the report

42:10

lays out so many exciting and promising areas

42:12

where this could make a transformational difference. And

42:15

in preparing for this, I kind of went back

42:17

and tried, I had a vague recollection, it

42:19

was a long time ago, but

42:22

you can go back to the

42:24

early days of the internet and

42:26

find Xerox researchers saying, we'll never use paper

42:29

again. You can find myriad

42:31

reports saying internet and digital technology

42:34

will increase efficiency and reduce emissions

42:36

by allowing for telecommuting. Of course,

42:38

global energy demand's risen about 50%

42:41

in the last 25 years. There

42:44

were reports from World Resources Institute and

42:46

others predicting we would democratize access to

42:48

information and build awareness around the world

42:50

for strong environmental action. You find all

42:52

of these things that's like, here's the

42:55

opportunity, here's all the things it could do. And

42:57

in retrospect, it didn't necessarily have

43:00

the impact. Why do you think

43:02

AI might be different or will it? I'm

43:04

so glad you raised this point Jason, because

43:06

none of these results are inevitable. And

43:09

there's enormous uncertainty. We're at the, first we're

43:11

at the beginning stages of transformational

43:14

technology having an impact in the

43:16

world. So the directions are quite

43:18

uncertain, but I think it

43:20

underscores the need for policy,

43:22

that policy guidance is hugely

43:24

important. You

43:27

know, innovation can happen

43:29

in a variety of different areas. Innovation

43:31

can happen with respect to technologies that

43:34

are not good for the planet or innovation can happen.

43:36

So all of these that are good for the planet.

43:39

And so we need policy that helps guide us.

43:41

And that's why it's so important in this area

43:43

that we have governments step in

43:45

and do things like bring

43:47

together the communities that are working on

43:49

this topic in order to better understand

43:51

each other, help to develop training programs,

43:54

help to support research in areas that are going

43:56

to make a difference, help to standardize data

43:59

and make data more accessible. accessible, help

44:02

to address bias issues that we've touched upon in

44:04

this. All these things are kind of important. The

44:06

role of policy in this area is

44:08

absolutely central if we're going to get the results that we

44:10

hoped for. I'll be even doing this

44:12

for a long time. Is that question fair? Do

44:14

we suffer from optimism bias with new technology? And

44:17

is that potentially applicable here? And what

44:19

needs to be done to make

44:22

sure we realize some of these opportunities? Yeah,

44:24

I completely agree. So, yeah, techno-optimism,

44:27

if that's the term to use

44:29

here, is potentially, you know, that

44:31

doesn't really achieve the end goal.

44:34

But at the end of the day, artificial

44:36

intelligence, just like many of these things that

44:38

you mentioned, Jason, like the Internet, is the

44:40

general purpose technology, right? And so

44:43

recognizing that not buying too much into the

44:45

hype, trying to sift through the noise to

44:47

find a signal is important. And

44:50

I agree with David here that we

44:52

need incentive structures to really guide

44:55

towards the outcome that we are looking for

44:57

here. David, can you say

44:59

more about what your policy recommendations would be

45:01

to achieve what you just described? Yes, thank

45:03

you, Jason. We start with some institutional recommendations,

45:05

like the one I mentioned, which is that

45:08

every institution should be paying attention to this

45:10

with people at the top. Then

45:13

government should use this convening power. It's one of the,

45:15

I think, lowest cost abilities of

45:17

government is to bring different communities together.

45:20

And that's very much needed here.

45:22

For example, in this area, bringing together climate experts

45:25

and AI experts, focusing

45:27

research and development dollars on applications

45:30

that will have a difference in

45:32

this area. There's been a tendency

45:34

in some areas to focus on

45:36

the next breakthrough in AI innovation.

45:39

We recommend focusing research development dollars on

45:41

how AI can be applied for

45:44

benefits. A

45:46

big area, which is beyond

45:48

in some ways the scope of AI,

45:50

but hugely important in this area are

45:52

utility incentives. Utilities often have

45:55

incentives that cut against improving

45:57

energy efficiency, for example, against

45:59

investing. in certain clean technologies, it's

46:02

important to align utility incentives here with

46:04

the outcomes that we want. And

46:07

then we recommend as well international

46:09

cooperation in this area. Institutions

46:12

like the UN Framework Commission

46:14

on Climate Change, the Clean Energy Ministerial

46:16

and others can provide a platform for

46:19

sharing information in this area that can

46:21

be very helpful globally. Al,

46:23

what are you most worried about? And

46:26

risk might be we don't realize opportunities, that's

46:28

sort of a missed opportunity. But in terms

46:30

of the things that could be

46:32

even worse than that, what are you

46:34

worried about with how this

46:37

technology might get deployed

46:39

or misused and what risks should

46:41

we be paying attention to, particularly

46:43

in the energy and climate space?

46:46

Yeah, the biggest technical risk I'll add

46:48

to kind of David's list here is

46:51

bias. And bias

46:53

means something specific in the AI

46:55

community. It means the data that

46:57

is used to

46:59

train the machine learning and AI

47:01

software systems do not reflect the

47:03

full picture of what we're trying to

47:06

solve. So it's, quote,

47:09

easy to fall into that trap of

47:11

saying, hey, we found a way to

47:13

predict climate patterns, forecast

47:15

patterns really, really well, but it only

47:17

works for the northern hemisphere. And

47:20

we emit the bias of not having good

47:22

enough data for, let's say, another part of

47:24

the world. Similarly,

47:26

crops that get developed that only work

47:28

with specific soils, that those

47:30

soils are not available. We don't take the full

47:33

picture into account. There's increasing

47:35

awareness in this. I

47:37

think the applications

47:39

of AI on human data

47:42

with privacy and

47:44

various other inequalities have created

47:47

awareness around this. But

47:50

it's a little bit less prominent when

47:52

we think about applications in these

47:54

novel sectors. So. And

47:56

what's the solution when you do, in fact, have a much larger

47:58

data set for? you know, one

48:01

application or one area than another. Yeah,

48:03

twofold. One is technical. One is

48:05

much more about awareness. Awareness is the education, you

48:07

know, just people need to be aware that this

48:09

is a risk. We need to

48:11

have guidelines and frameworks

48:13

in place to understand whether

48:16

the bias risk has been mitigated

48:18

in any AI technology that we

48:21

are developing and then ideally, you

48:23

know, deploying. On the

48:25

technical side, from the research

48:27

community, there is very active work towards

48:30

quantifying. This bias detecting

48:32

it, better methods to be

48:34

able to understand whether there are gaps in

48:37

how an AI system has been built. These

48:39

are all tools that will make this process a little bit

48:41

easier. David, you mentioned

48:43

international coordination and cooperation on this. And

48:46

of course, you have deep

48:49

expertise in China and how it's

48:51

approaching the clean energy revolution and

48:53

climate change. The

48:56

US and China announced a new bilateral

48:58

channel for consultation on AI in

49:00

November of last year. Do

49:03

you see this as an area of potential

49:05

cooperation? Do you see what is happening? How

49:07

do you view the role of China and

49:09

AI? Is this

49:12

going to exacerbate tensions that are already quite

49:14

high between the US and China? Technology

49:16

cooperation is probably the principal area

49:18

of tension in the bilateral relationship right

49:21

now, along with Taiwan. I think the

49:24

disputes over semiconductor gypsum have

49:28

risen to the top of the agenda in the

49:30

bilateral relationship. So I think genuine

49:33

cooperation on AI in the

49:35

US-China relationship is going to

49:37

be challenging. I think communication

49:39

is incredibly important. And as

49:43

you said, there's a channel right now for that discussion.

49:46

I think that's incredibly important. I

49:48

think I hope it will grow and continue. China's

49:50

capabilities in this area are very, very significant.

49:53

Enormous amount of the peer review literature

49:55

globally is coming from China. There's

49:58

a lot of technical development happening. China

50:00

in this area right now.

50:02

The world will be a better place

50:05

if we can find a way to maintain

50:07

our open, maintain at least

50:09

some amount of communication and

50:12

cooperate, but it's not going to be easy given

50:14

broader geopolitical tensions in the years ahead. Al,

50:17

not just with China, but more broadly,

50:19

you're a computer scientist, academic, cooperating with

50:21

people all over the world. How

50:24

do you see the impact of potential

50:26

geopolitical tensions on our ability to do that

50:28

and work together? Yeah, I

50:30

think one example is how Europe is approaching

50:32

the kind of challenge versus the US. Europe

50:35

is much more of a stick and the

50:37

US taking much more of a carrot kind

50:39

of approach in incentivizing how

50:41

this technology gets adopted. And I

50:43

think even that is an interesting

50:45

divide, right? Where technology

50:48

providers in Europe have to

50:50

really think about how does

50:54

development fit into the regulatory

50:56

environment in Europe. GDPR

50:58

is much more on the personal

51:01

privacy side, but the

51:03

carbon border adjustment mechanism, so on and so forth,

51:05

it's a very different landscape. And so that

51:08

coupled with the kind of bias,

51:10

again, now a different type of

51:14

bias of how far ahead the United States

51:16

is relative to Europe in the development

51:19

of AI technology is

51:21

leading to even tension among kind

51:23

of friendly and allied nation states.

51:27

We're just about out of time, but I just want to ask

51:29

each of you in closing, we've talked about there's

51:32

a huge amount of interest on the potential for

51:34

significant electricity demand from this technology. We've talked

51:36

about some of the opportunities to

51:39

lower costs, improve efficiency. What

51:41

is most misunderstood, unappreciated?

51:44

What's coming around the corner no one's talking about?

51:46

That in the course of your work, you

51:49

would highlight for people people should

51:52

be more aware of and paying closer attention

51:54

to in the broad space of AI

51:57

and the energy transition and

51:59

climate. Matt, maybe I'll start with you.

52:02

I think the solutionation problem is

52:04

exacerbating, hopefully making it clear that

52:06

these black boxes are not

52:09

suitable for adoption into these hard

52:11

to debate sectors, these high risk

52:14

sectors. Do you want to just

52:16

explain for everyone what you're referring

52:18

to? Absolutely. If you take your

52:21

favorite ridesharing app and it tells you, hey,

52:24

it's going to take 17 minutes to get

52:26

to your destination, you don't really need an

52:28

explanation for that. Or if your favorite media

52:30

app gives you a recommendation of a television

52:32

show to watch, you don't really need an explanation, you

52:34

don't benefit from it. If

52:36

your ridehailing app gives you a confidence

52:38

interval, it says you're going to get

52:41

to your destination 17 minutes plus or

52:43

minus five, that's not valuable to you,

52:45

you just want to get there as quickly as possible. These

52:48

are all just applications of machine

52:50

learning and AI in where black

52:52

boxes are okay. We

52:54

don't need explanations, they're low

52:56

risk applications. When

52:58

we talk about the power sector, we talk about

53:00

manufacturing, we talk about agriculture, the

53:03

risks are too high for just adopting these

53:05

types of just black mystery, black boxes. Explainability,

53:08

tackling the problem of hallucinations, being

53:10

able to understand when an AI

53:12

system can be trusted and plugged

53:15

into a workflow is essential. David,

53:18

same thing, question for you. Just in

53:20

your sense, what's most misunderstood

53:22

or people have the least awareness of

53:25

with this technology? I

53:27

found Jason in the past couple of months

53:29

when I've had conversations with people about this,

53:32

but I would say roughly 80 percent

53:34

of the commentary and questions I get

53:36

are all about how AI is going

53:38

to drive up power demand and cause

53:40

problems. I

53:43

think the larger picture, the

53:45

potential for AI to deliver

53:47

enormous benefits here, is

53:49

just not getting the attention that I think

53:51

it deserves. I hope that we'll

53:53

see much more of that dialogue going forward. As we

53:55

said in this discussion, those benefits

53:58

are not inevitable. They

54:00

are absolutely achievable if we pay attention to

54:02

them, if we have the right policy framework,

54:04

and we have people dedicated to working on

54:06

them. Al, could you go beer? David

54:09

Sandolo, thank you so much for your work

54:11

on this report, and thanks for sharing your

54:13

insights with us here on Columbia Energy Exchange

54:15

today. I appreciate it. Thanks for having us,

54:18

Jason. Thank you, Jason. Pleasure to be

54:23

here. Thank you again, David and Al, and thank

54:25

you for listening to this week's episode of Columbia

54:27

Energy Exchange. The show is brought

54:29

to you by the Center on Global Energy Policy

54:31

at Columbia University School of International and Public Affairs.

54:34

The show is hosted by me, Jason Bordoff, and

54:36

by Bill Lovelace. The show is produced by Aaron

54:38

Hardick from Latitude Studios. Additional

54:41

support from Paul DeBarr, Lily Lee, Caroline

54:43

Pittman, Victoria Prado, and Q. Lee. Roy

54:46

Campanella engineered the show. For

54:49

more information about the podcast or the Center

54:51

on Global Energy Policy, please visit us

54:53

online at energypolicy.columbia.edu or follow

54:55

us on social media at

54:59

ColumbiaUEnergy. And please,

55:01

if you feel inclined, give us a rating on

55:03

Apple Podcasts. It really helps us out. Thanks

55:06

again for listening. We'll see you next week.