0:00

Welcome to MIT Technology

0:03

Review. My name is My name is

0:05

Matt I'm our our Editor -in -Chief. Every

0:07

Every week we'll bring you a fascinating new,

0:10

in -depth story for the

0:12

leading edge of science

0:14

and technology, covering topics like

0:16

covering topics climate, energy, robotics,

0:18

and more. robotics, this week's story. Here's

0:20

I hope you enjoy it. story. I hope you

0:22

enjoy name is David

0:25

Rotman. I'm the editor -at

0:27

-large at MIT Technology

0:29

Review. The article you're

0:31

about to hear asks the

0:33

question. question, can generative AI lead to

0:36

to renewed economic prosperity. Can

0:38

it Can it make the country and

0:40

our economy richer? I talked to I

0:42

talk to economists, on experts

0:45

on manufacturing, scientists, and inventors

0:47

of AI to find out to find

0:49

out the challenges and the potential of

0:51

the technology to make us more

0:53

efficient and productive. and productive. Here's

0:55

a look at the future of AI of AI

0:57

how it could improve our working lives. lives.

1:01

Narrated by Listen to Listen

1:03

to more of the best

1:05

articles from the the biggest publishers

1:07

on the NOAA app on the Noah app,

1:09

or at.com When Chad Siverson

1:12

When Chad Bureau of Labor the U .S.

1:14

Bureau of Labor Statistics' looking days

1:16

looking for the latest data

1:18

on productivity, he does he does so

1:20

with a sense of optimism that

1:22

he hasn't felt in ages. in

1:24

ages. The The numbers for the last

1:26

year or so have been generally

1:28

strong for various financial and business

1:30

reasons, reasons, from the early days of

1:32

the pandemic. And though

1:34

the quarterly numbers are notoriously

1:37

noisy and inconsistent, the

1:39

University of Chicago of is

1:41

scrutinizing the data to spot

1:43

any early clues that clues

1:45

economic growth has begun. begun.

1:48

Any Any effect on

1:50

the current statistics,

1:52

he says, will

1:55

likely be quite

1:57

small and won't

1:59

be be -changing, so

2:01

he's not surprised

2:04

that signs of

2:06

impact haven't been

2:08

detected yet. detected yet.

2:10

watching closely with the hope that

2:13

over the next few years AI

2:15

could help reverse a two-decade slump

2:17

in productivity growth that is undermining

2:19

much of the economy. If that

2:22

does happen, Siberson says, then it

2:24

is world-changing. The newest versions of

2:26

generative AI are bedazzling, with lifelike

2:28

videos, seemingly expert-sounding pros, and other

2:31

all-to-human-like behaviors. Business leaders are fretting

2:33

over how to reinvent their companies

2:35

as billions flow into startups, and

2:37

the big AI companies are creating

2:40

ever more powerful models. Predictions abound

2:42

on how chat-GPT and the growing

2:44

list of large language models will

2:46

transform the way we work and

2:49

organize our lives, providing instant advice

2:51

on everything from financial investments to

2:53

where to spend your next vacation

2:55

and how to get there. But

2:58

for economists like Siverson... The most

3:00

critical question around our obsession with

3:02

AI is how the fledgling technology

3:04

will or won't boost overall productivity,

3:07

and if it does, how long

3:09

it will take. Think of it

3:11

as the bottom line to the

3:13

AI hype machine. Can the technology

3:16

lead to renewed prosperity after years

3:18

of stagnant economic growth? Productivity growth

3:20

is how countries become richer. Technically,

3:22

labor productivity is a measure of

3:25

how a worker produces on average.

3:27

Innovation and technology advances account for

3:29

most of its growth. As workers

3:32

and businesses can make more stuff

3:34

and offer more services, wages and

3:36

profits go up, at least in

3:38

theory, and if the benefits are

3:41

shared fairly. The economy expands, and

3:43

governments can invest more and get

3:45

closer to balancing their budgets. For

3:47

most of us, it feels like

3:50

progress. It's why? Until the last

3:52

few decades, most Americans believed their

3:54

standard of living and financial opportunities

3:56

would be greater than those of

3:59

their parents and grandparents. But when

4:01

productivity growth is flat, or nearly

4:03

flat. The pie is no longer

4:05

growing. Even a 1% annual slowdown

4:08

or speed up can spell the

4:10

difference between a struggling economy and

4:12

a flourishing one. In the late

4:14

1990s and early 2000s, U.S. labor

4:17

productivity grew at a healthy rate

4:19

of nearly 3% a year as

4:21

the Internet age took off. It

4:23

grew even faster, well over 3%

4:26

in the booming years after World

4:28

War II. But since about 2005...

4:30

Productivity growth in most advanced economies

4:32

has been dismal. There are various

4:35

possible culprits to blame, but there

4:37

is a common theme. The seemingly

4:39

brilliant technologies invented over the last

4:41

two decades, from the iPhone to

4:44

ubiquitous search engines to all-consuming social

4:46

media, have grabbed our attention yet

4:48

failed to deliver large-scale economic prosperity.

4:50

In 2016, I wrote an article

4:53

titled Dear Silicon Valley. forget flying

4:55

cars, give us economic growth. I

4:57

argued that while Big Tech was

4:59

making breakthrough after breakthrough, it was

5:02

largely ignoring desperately needed innovations in

5:04

essential industrial sectors, such as manufacturing

5:06

and materials. In some ways, it

5:08

made perfect financial sense. Why invest

5:11

in these mature, risky businesses when

5:13

a successful social media startup could

5:15

net billions? But such choices came

5:17

with a cost in sluggish productivity

5:20

growth, while a few in Silicon

5:22

Valley and elsewhere became fabulously wealthy.

5:24

At least some of the political

5:26

chaos and social unrest experienced in

5:29

a number of advanced economies over

5:31

the last few decades can be

5:33

blamed on the failure of technology

5:36

to increase financial opportunities for many

5:38

workers and businesses and expand vital

5:40

sectors of the economy across different

5:42

regions. Some preach patience. The breakthroughs

5:45

will take time to work through

5:47

the economy, but once they do

5:49

watch out, that's probably true. But

5:51

so... far, the The

5:54

result is a

5:56

deeply divided country,

5:58

where the the techno

6:00

and immense wealth

6:03

oozing from Silicon

6:05

Valley seem relevant

6:07

to only a

6:09

few. only a few. It's still still

6:12

too early to know how things

6:14

will shake out out time around. around,

6:16

whether AI is truly a once -in

6:18

-a -century breakthrough that will spur a

6:20

return to financial good times. good times,

6:23

whether it will do

6:25

little to create real real

6:27

prosperity. prosperity. Put way, way. will it

6:29

be the the harnessing of electricity and

6:31

the invention of the electric motor,

6:33

which led to an industrial boom, or

6:36

more like like and social

6:38

media, media, which have consumed our

6:40

collective consciousness without bringing significant

6:42

economic growth? growth? For AI, For AI,

6:45

particularly models, models, to have a greater

6:47

economic impact than other digital advances

6:49

over the last few decades, we will

6:51

we will need to use the

6:54

technology to transform productivity across the

6:56

economy, even in how we generate

6:58

new ideas. It's a huge It's a

7:00

huge undertaking, and won't happen overnight, but

7:02

but we're at a critical inflection point.

7:04

point. Do Do we start down that

7:06

path to broadly increased prosperity? Or

7:09

do the creators of today's today's

7:11

continue to ignore the vast

7:13

potential of the technology to truly

7:15

improve our lives? our lives?

7:17

A series of A series of

7:19

studies of the last year show

7:21

how AI can can boost productivity

7:24

for people doing various jobs. jobs.

7:26

Economists at Stanford and MIT MIT

7:28

found that those working in call

7:30

call are 14 % more productive

7:32

when using AI conversational assistance. Notably,

7:35

there was a 35 %

7:37

improvement in the performance of

7:39

inexperienced and low -skilled workers. Another

7:41

Another study showed that software engineers

7:43

could code twice as fast

7:45

with the technology's help. Last

7:47

Last year, Goldman Sachs calculated

7:50

that generative AI would likely

7:52

boost overall productivity growth by

7:54

1 .5 percentage points every year

7:56

in developed countries and increase

7:58

global GDP by by... trillion over

8:01

10 years, and some predict

8:03

that the effects will appear

8:05

soon. Anton Kornik, an economist

8:07

at the University of Virginia,

8:09

says the added growth has

8:11

not yet shown up in

8:13

the productivity numbers because it

8:15

takes time for generative AI

8:17

to diffuse throughout the economy,

8:19

but he predicts a 1%

8:21

to 1.5% boost to U.S.

8:23

productivity by next year. and

8:25

if there continue to be

8:27

breakthroughs in generative AI models,

8:30

think chatGPT5, the eventual impact

8:32

could be significantly higher, says

8:34

Coranac. Not everyone is so

8:36

bullish. Darren Asamaglu, an MIT

8:38

economist, says his calculations are

8:40

a corrective against those who

8:42

say that within five years

8:44

the entire US economy is

8:46

going to be transformed. As

8:48

he sees it, generative AI,

8:50

could be a big deal.

8:52

We don't know yet. But

8:54

if it is, we're not

8:56

going to see transformative effects

8:58

within 10 years. It's too

9:00

soon. It will take time.

9:03

In April, Asamoglu posted a

9:05

paper predicting that generative AI's

9:07

impact on total factor productivity,

9:09

TFP, the portion that specifically

9:11

reflects the contribution from innovation

9:13

and new technologies, will be

9:15

around 0.6% in total over

9:17

10 years, far less than

9:19

Goldman Sachs and others expect.

9:21

For decades, TFP growth has

9:23

been sluggish, and he sees

9:25

generative AI doing little to

9:27

significantly reverse the trend, at

9:29

least in the short term.

9:31

Asimoglu says he expects relatively

9:33

modest productivity gains from generative

9:36

AI because its big-tech creators

9:38

have largely had a narrow

9:40

focus on using AI to

9:42

replace people with automation and

9:44

enable online monetization of search

9:46

and social media. To have

9:48

a greater impact on productivity,

9:50

he argues, AI needs to

9:52

be useful for a far

9:54

broader portion of the workforce

9:56

and relevant for more parts

9:58

of the economy. Critically

10:01

it needs to be used to

10:03

create new types of jobs, not

10:05

just to replace workers. Asamaglu argues

10:07

that generative that generative AI could

10:09

be used to expand the

10:11

capabilities of workers for for example, supplying

10:13

supplying real -time data and reliable

10:15

information for many types of

10:17

jobs. of jobs. Think of an intelligent

10:19

AI agent, but but one versed

10:22

on the intricacies of, say, factory factory

10:24

floor production. Yet he

10:26

writes, gains gains will remain unless there

10:28

is a there is a fundamental

10:30

reorientation of the tech industry,

10:32

including a major major change in

10:34

the architecture of the most

10:36

common generative AI models. to

10:38

think that It's tempting to think that

10:40

perhaps it's simply a matter of tweaking

10:42

today's large foundation models with the appropriate

10:45

data to make them widely useful for

10:47

various industries. But in fact, we

10:49

we will need to rethink the models and

10:51

how how they can be more effectively

10:53

deployed in a far broader range of

10:55

uses. Take manufacturing. For manufacturing. years,

10:57

it was one of many years, it was

11:00

one of the important sources of

11:02

productivity gains in the U .S. economy. accounts

11:04

for still of for much of the

11:06

country's R And recent And recent increases

11:08

in automation of the use of

11:10

industrial robots might suggest that

11:12

manufacturing is becoming more productive. But

11:14

that has not been the case. the case. For

11:16

For somewhat mysterious reasons, productivity

11:19

in U .S. manufacturing has been

11:21

a disaster since about 2005, which

11:23

which has played an outsize

11:25

role in the overall productivity slowdown. The

11:27

promise of The promise of

11:29

generative AI and reviving productivity is

11:31

that it could help integrate

11:33

everything from initial materials and design

11:35

choices to real -time data from

11:37

sensors embedded in production equipment. Multimodal

11:41

capabilities could allow a factory worker to

11:43

to say snap a picture of a

11:45

problem and ask the AI model for

11:47

a solution based on the image. the

11:49

image. The company's operating manual, any

11:51

relevant regulatory guidelines, and

11:54

vast amounts of real -time data from

11:56

the machinery. from the machinery. vision,

11:58

at least. at least. The reality

12:00

is that efforts to deploy

12:02

today's foundation models in design and

12:04

manufacturing are in their very

12:07

early days. Use of AI so

12:09

far has been limited to

12:11

narrow domains, says Fez Ahmed, an

12:13

MIT mechanical engineer specializing in

12:15

machine learning, think scheduling maintenance on

12:18

the basis of data from

12:20

a particular piece of equipment. In

12:22

contrast, generative AI models could,

12:24

in theory, be broadly useful for

12:26

everything from improving initial designs

12:28

with real data to monitoring the

12:31

steps of a production process

12:33

to analyzing performance data on the

12:35

factory floor. In

12:37

a paper released in

12:39

March, a team of MIT

12:41

economists and mechanical engineers, including

12:44

Osamoglu and Ahmed, identified numerous

12:46

opportunities for generative AI in

12:48

design and manufacturing before

12:50

concluding that current generative AI

12:52

solutions cannot accomplish these goals

12:54

due to several key deficiencies.

12:56

Chief among the shortcomings of

12:59

chatGPT and other AI

13:01

models are their ability to

13:03

supply reliable information, their lack

13:05

of relevant domain knowledge, and

13:07

their unawareness of industry standards

13:09

requirements. The models

13:11

are also ill -designed to handle

13:14

the spatial problems on manufacturing

13:16

floors and the various types of

13:18

data created by production equipment,

13:20

including old machinery. The biggest difficulty

13:22

is that existing generative AI

13:24

models lack the appropriate data, says

13:26

Ahmed. They are trained on

13:29

data scraped from the internet, and

13:31

it's a lot more about

13:33

cats and dogs and multi -media

13:35

content rather than how do you

13:37

actually operate a lathe machine,

13:39

he says. The reason these models

13:41

perform relatively poorly on manufacturing

13:43

tasks is that they've never seen

13:46

manufacturing tasks. Gaining

13:48

access to such data is tricky

13:50

because much of it is proprietary. Some

13:52

people are really scared that a

13:54

model will take my data and run

13:56

away with it, he says. A

13:59

related problem is that manuf... Factoring requires

14:01

precision, and and often to

14:03

strict industry or government

14:05

guidelines. If the

14:07

systems are not precise are not trustworthy,

14:10

people are less likely to use them, he

14:12

says. use them, And it's a chicken and

14:14

egg problem. egg Because the models are

14:16

not precise are not there is no data. there

14:18

is no data. MIT researchers called for

14:20

a next generation of AI

14:22

models that would be tailored to

14:24

manufacturing. But But here's a problem.

14:27

Creating a manufacturing relevant relevant AI

14:29

takes advantage of the power

14:31

of models will require close

14:33

collaboration between industry and AI

14:35

companies. and AI And that's something

14:38

still in its nascent stage. nascent

14:40

The lack of progress so far, says

14:42

says Chandra, managing managing of of

14:44

research for industry at Microsoft

14:46

Research, is not not because people

14:48

are not interested or they don't

14:51

see the business value. value. The

14:53

hold-up is finding ways to secure

14:55

the data and make sure

14:57

it is in a useful form

14:59

and provides relevant answers to specific

15:01

manufacturing questions. questions. Microsoft is

15:03

pursuing several strategies. One is asking the asking

15:05

the foundation model to base

15:07

its answers on a company's proprietary

15:09

data, data, say say a company's

15:11

operations manual and production data. data.

15:14

A far far more difficult but appealing

15:16

alternative is fine the the underlying

15:18

architecture of the model to

15:20

better suit manufacturing. Yet

15:22

another approach, so -called small language

15:24

models, which which also can be

15:26

trained specifically on the data

15:28

from a company. a company. they

15:31

are smaller than foundation models like

15:33

GPT -4, GPT-4, they need less

15:35

computational power power can be more

15:37

targeted to specific manufacturing tasks. But

15:40

this is all this is this point,

15:42

says point, says Have we solved it? we

15:44

Not yet. Not yet. Using AI to

15:47

to boost scientific discovery and

15:49

innovation could have the greatest

15:51

overall productivity impact over the

15:53

long over the long term. Economists have

15:55

long long new ideas as the

15:57

source of long -term growth, growth,

15:59

and hope is... that new AI tools

16:01

could turbocharge the search for them.

16:03

While improving the efficiency of say

16:05

a call center worker could mean

16:07

a one-time jump in productivity in

16:10

that business, using AI to improve

16:12

the process of inventing new technologies

16:14

and business practices to create useful

16:16

new ideas could lead to an

16:18

enduring increase in the rate of

16:21

economic growth as it reshapes the

16:23

innovation process and the way research

16:25

is done. There are

16:27

already tantalizing clues to AI's

16:29

potential. Most notably, Google Deep

16:31

Mind, which defines its mission

16:34

as solving some of the

16:36

hardest scientific and engineering challenges

16:38

of our time, says more

16:40

than 2 million users have

16:42

accessed its deep learning AI

16:44

system to predict protein folding.

16:46

Many drugs target a particular

16:48

protein, and knowing the 3D

16:50

structure of such proteins, something

16:53

that traditionally takes painstaking lab

16:55

analysis, could be an invaluable

16:57

step in creating new medicines.

16:59

In May, Google released Alpha

17:01

Fold 3, claiming it predicts

17:03

the structure and interactions of

17:05

all of life's molecules to

17:07

help identify how various biomolecules

17:09

alter each other, providing an

17:11

even more powerful guide for

17:14

finding new drugs. Creators

17:16

of AI models, including Deep Mind

17:18

and Microsoft Research, are also working

17:20

on other problems in biology, genomics,

17:22

and materials science. The hope is

17:24

that generative AI could help scientists

17:26

glean key information from the vast

17:28

data sets common in these fields,

17:30

making it easier and faster to,

17:33

say, discover new drugs and materials.

17:35

We badly need such a boost.

17:37

A few years ago, a team

17:39

of leading economists wrote a paper

17:41

called, Our Ideas Getting Harder to

17:43

Find, and found that it takes

17:45

more and more researchers and money

17:47

to find the kinds of new

17:49

ideas that are key to sustaining

17:51

technology advances. The problem, in technical

17:53

terms, is that research productivity.

17:56

The output of

17:58

ideas, given the

18:00

number of scientists,

18:02

is falling rapidly. falling

18:04

In other words, yes, ideas are getting

18:06

harder to find. harder generally kept

18:08

up by adding more researchers

18:10

and investing more in R &D,

18:13

but overall US but productivity itself

18:15

is in a deep decline. is

18:17

in a deep to uphold Moore's

18:19

law, Law. which predicts that the number

18:21

of transistors on a chip will

18:23

double will every two years. every two

18:26

The semiconductor industry needs 18 times

18:28

more 18 than it had in the

18:30

early 1970s. 1970s. Likewise, it

18:32

takes far more scientists to come up with

18:34

roughly the same number of new drugs than

18:36

it did a few decades ago. did a few

18:38

decades ago. John Van Rienen, a a professor

18:40

at the London School of Economics and

18:42

one of the authors of the paper. the

18:45

paper. knows it's still too early to

18:47

see any real change in the productivity

18:49

data from AI, data but he says he

18:51

says, the is that it can make some

18:53

difference. Alpha fold is a

18:55

a poster child for how AI

18:57

can change science, he says. the And

18:59

the question is this this can

19:01

go from anecdotes to something more

19:03

systematic. The ambition is The

19:05

ambition is not only to

19:07

supply various tools that will make

19:09

the lives of scientists easier easier,

19:11

automated literature research. research. But for AI

19:14

itself to come up with

19:16

original and useful scientific ideas

19:18

ideas would otherwise evade researchers. In

19:20

that In that vision, up new dreams

19:22

up new compounds that are

19:24

more effective and safer than

19:26

existing drugs and astonishing materials

19:28

that expand the possibilities of

19:30

computation and clean energy. energy. The

19:32

goal goal is especially compelling the

19:34

the universe of potential

19:36

molecules is virtually unlimited. unlimited. Navigating

19:39

such a nearly infinite space and exploring

19:41

the vast number of possibilities

19:44

is what machine learning is especially

19:46

good at. good at. But don't

19:48

don't hold your breath for AI's Thomas

19:50

Edison moment. Though the scientific

19:52

scientific popularity of Alpha Fold has

19:54

raised expectations for the potential of

19:57

AI, of AI, it is still very

19:59

early days in turn. the research into

20:01

actual products, whether new drugs

20:03

or novel materials. In

20:05

a recent analysis, a team

20:07

of MIT scientists put it this

20:09

way. Generative AI has undoubtedly

20:12

broadened and accelerated the early stages

20:14

of chemical design. However,

20:16

real world successes take place further

20:18

downstream, where the impact of AI

20:20

has been limited so far. In

20:23

fact, the process of turning

20:25

the intriguing scientific advances in using

20:27

AI into actual useful stuff

20:30

is still very much in its

20:32

infancy. Perhaps

20:34

nowhere is the excitement over AI's

20:36

potential to transform research greater

20:38

than in the often neglected field

20:40

of materials discovery. The

20:42

world desperately needs better materials. We

20:45

need them for cheaper and more

20:47

powerful batteries and solar cells, and

20:49

for new types of catalysts that

20:51

would make cleaner industrial processes possible.

20:54

and we need practical high

20:56

-temperature superconductors to revolutionize how

20:58

we transport electricity. So

21:01

in DeepMind said it had

21:03

used deep learning to discover

21:06

some 2 .2 million inorganic

21:08

crystals, including some 380 ,000 predicted

21:10

to be stable and promising

21:12

candidates for actual synthesis. The

21:14

report was greeted with great

21:16

excitement, especially in the AI

21:19

community. A materials

21:21

Revolution. It seemed like

21:23

a goldmine of new stuff,

21:25

an order of magnitude expansion in

21:27

stable materials known to humanity,

21:29

wrote the DeepMind researchers in Nature.

21:32

The DeepMind database, called Genome,

21:34

an acronym for Graph

21:36

Networks for Materials Exploration, is

21:38

equivalent to 800 years

21:40

of knowledge, according to the

21:42

company's media release. But

21:45

in the months after the

21:47

paper, some researchers disputed the hype.

21:50

Materials scientists at the University of

21:52

California, Santa Barbara published a

21:54

paper in which they reported finding

21:56

scant evidence that any of

21:58

the structures in the deep mind database

22:00

the trifecta of

22:03

novelty, credibility, and

22:05

utility. For some For some

22:07

tasked with finding new materials, the

22:09

the huge databases of possible inorganic

22:11

crystals, many of which may not

22:13

be stable enough to actually exist,

22:16

seems like a distraction. a If

22:18

you If us us with 400,000 new materials,

22:20

and and we don't even know

22:22

which one of those are realistic, then

22:25

we don't know which one of

22:27

those will be good for a battery

22:29

for a battery or whatever you want to

22:31

make them. you want to this information is

22:33

not useful, says not useful, says a chemist

22:35

at Princeton at co -wrote a paper describing

22:37

the challenges of using automation and of

22:40

in materials discovery and synthesis. discovery and

22:42

To be clear, be this doesn't

22:44

mean that AI won't prove to

22:46

be important in material science and

22:48

chemistry. and Even critics say they are

22:50

excited by the long the possibilities. possibilities.

22:52

But the hint at just how

22:54

early we are in using AI

22:57

to tackle the daunting task of

22:59

discovery and making it it a reliable

23:01

tool for finding new compounds that

23:03

are better than existing ones. It's

23:05

extremely It's extremely expensive and

23:08

time to make and test any

23:10

possible new material. What What

23:12

industrial researchers really need are

23:14

reliable clues pointing to materials that

23:16

are predictably stable, can

23:18

be synthesized, and likely

23:20

have intriguing properties, including being

23:22

cheap to make. make. The genome

23:24

database database probably includes

23:26

interesting compounds, says its -mind

23:29

scientific creators, but

23:31

they acknowledge it's only a preliminary

23:33

step in showing how AI could help

23:35

in how AI could help in Much

23:37

work remains to broaden its

23:39

usefulness. broaden its usefulness. Eken Doge

23:42

a Google scientist and co -author of

23:44

The Nature Nature Paper, describes the work

23:46

it reports as as an in predicting

23:48

a large number of possible inorganic

23:51

crystals that are stable, based

23:53

on quantum based on calculations calculations

23:55

at where atomic motion

23:57

comes to a standstill. a

23:59

standstill. Such predictions could be useful

24:01

for those running computational simulations of

24:04

new materials, a very early stage

24:06

of materials discovery. But, he says,

24:08

machine learning has not yet been

24:11

used to predict crystals that are

24:13

stable at room temperature. After that

24:15

is achieved comes the goal of

24:17

using AI to predict how structures

24:20

can be synthesized in the lab,

24:22

and eventually how to make them

24:24

at larger scale. All that must

24:27

be done before machine learning can

24:29

really transform the lengthy and expensive

24:31

process of coming up with new

24:34

materials, he says. For those hoping

24:36

that AI models could boost economic

24:38

productivity by transforming science, one lesson

24:41

is clear. Be patient. Such scientific

24:43

advances could well have an impact

24:45

one day, but it will take

24:48

time, likely measured in decades. As

24:51

senior vice president for research,

24:53

technology, and society at Google,

24:55

James Manukkah is unsurprisingly enthusiastic

24:57

about the huge potential for

24:59

AI to transform the economy,

25:01

but he is far from

25:03

an unabashed cheerleader, mindful of

25:05

the lessons gleaned from his

25:07

years of studying how technologies

25:10

affect productivity. Before joining Google

25:12

in 2022, Manuka spent several

25:14

decades as a consultant, a

25:16

researcher, and finally chairman of

25:18

the McKinsey Global Institute, the

25:20

economic research arm of the

25:22

consulting giant. At McKinsey he

25:24

became a leading authority on

25:26

the link between technology and

25:28

economic growth, and he counts

25:30

Robert Solo. the MIT economist

25:32

who won the 1987 Nobel

25:34

Prize for explaining how technological

25:36

advances are the main source

25:39

of productivity growth as an

25:41

early mentor. Among the lessons

25:43

from Solow, who died late

25:45

last year at the age

25:47

of 99, is that even

25:49

powerful technologies can take time

25:51

to affect economic growth. In

25:53

1987, Solow equipped, you can

25:55

see the computer age everywhere,

25:57

but in the productivity statistics.

25:59

At the time At the time,

26:01

information technology was undergoing a

26:04

revolution, most visible with the introduction

26:06

of the personal computer. personal computer.

26:08

Yet as measured by economists,

26:10

by was sluggish. This became

26:13

known as as paradox. It

26:15

It wasn't until the late 1990s, decades

26:17

decades after the birth of the

26:19

computer age, that that productivity growth began

26:21

to finally pick up. pick up.

26:23

History has taught Manuka to

26:26

be circumspect in predicting how and

26:28

when the overall economy will

26:30

feel the impact of impact of generative AI. I

26:32

don't have a a time he says. says. The The

26:34

estimates of productivity gains are generally

26:36

spectacularly large, large, but when it comes

26:38

to a question of a frame, I

26:41

say it depends. I say it depends.

26:43

Specifically, says it depends on

26:45

what economists call the pace of

26:47

diffusion, basically how quickly

26:49

users take up the technology

26:51

both within sectors and across

26:53

sectors. sectors. It also hinges on the ability

26:55

of various users, especially businesses in

26:57

the largest sector of the economy,

27:00

of to to functions and

27:02

tasks and processes, and to capitalize

27:04

on the technology the to make

27:06

their operations and workers more

27:08

productive. workers more those pieces, we'll

27:10

be stuck in we'll be stuck in

27:13

solo says land, says Manika. Tech

27:15

can can do whatever tech wants,

27:17

and it doesn't really matter from

27:19

a labor productivity standpoint, he

27:21

says, he since its is relatively small.

27:23

We have to have to have changes happen

27:25

in the largest sectors we we can

27:28

start to see productivity gains at an

27:30

economy level. level. Late Late last

27:32

year, Manuka co -wrote a piece in Foreign

27:34

Affairs with Michael Spence, winner

27:36

of the 2001 Nobel Prize in in

27:39

called called The Coming AI Revolution,

27:41

Can Artificial Intelligence

27:43

Reverse the Productivity

27:45

Slowdown? In it, the

27:47

authors offered a decidedly optimistic, though

27:50

cautious, answer. the beginning of the

27:52

By the beginning of the next decade,

27:54

AI could to a could become a leading

27:56

driver of global they wrote, they wrote, it

27:58

has the it has the potential. to affect

28:00

just about every aspect of human

28:02

and economic activity. They added, if

28:04

these innovations can be harnessed, AI

28:07

could reverse the long-term declines in

28:09

productivity growth that many advanced economies

28:11

now face. But it's a big

28:13

if, they acknowledged, saying it won't

28:15

happen on its own and will

28:18

require positive policies that foster AI's

28:20

most productive uses. The call for

28:22

policies is a recognition of the

28:24

immense task ahead. and an acknowledgement

28:26

that even giant AI companies like

28:29

Google can't do it alone. It

28:31

will take widespread investments in infrastructure

28:33

and additional innovations by governments and

28:35

businesses. Companies ranging from small startups

28:37

to large corporations will need to

28:40

take the foundation models such as

28:42

Google's Gemini and tailor them for

28:44

their own applications in their own

28:46

environments in their own domains, says

28:48

Manuka. In a few cases, he

28:51

says... Google has done some of

28:53

the tailoring because it's kind of

28:55

interesting to us. For example, Google

28:57

released Med Gemini and May using

28:59

the multimodal abilities of its foundation

29:02

model to help in a wide

29:04

range of medical tasks, including making

29:06

diagnostic decisions based on imaging, videos

29:08

of surgeries, and information in electronic

29:10

health records. Now, says Manuka, it's

29:13

up to health care practitioners and

29:15

researchers to think how to apply

29:17

this, because we're not in the

29:19

health care business in that way.

29:21

But, he says, it is giving

29:24

them a running start. But therein

29:26

lies the great challenge going forward

29:28

if AI is to transform the

29:30

economy. Despite the fanfare around generative

29:32

AI and the billions of dollars

29:35

flowing to startups around the technology,

29:37

The speed of its diffusion into

29:39

the business world is not all

29:41

that encouraging. According to a survey

29:43

of thousands of businesses by the

29:46

US Census Bureau, released in March,

29:48

the proportion of firms using AI

29:50

rose from about 3.7% .7 %

29:52

in September 2023

29:54

to 5 .4 %

29:56

this February, reach around

29:59

and it is

30:01

expected to reach

30:03

around 6 .6 %

30:05

by the end

30:07

of the year.

30:10

like of this uptake has come in

30:12

sectors like finance and technology. Industries

30:15

like construction and manufacturing

30:17

are virtually untouched. main reason

30:19

the main reason for the lack of interest. what

30:22

what most companies see as the

30:24

inapplicability of AI to their

30:26

business. business. For For many companies, particularly

30:28

small ones, it still takes a

30:30

huge leap of faith to bet

30:32

on AI and invest invest the money and

30:34

time it takes to reorganize business

30:36

functions around it. it. In In

30:39

addition to not seeing any value in the

30:41

technology, lots of business leaders

30:43

have ongoing questions over the reliability

30:45

of the generative AI models. AI

30:47

models. are one thing in the

30:49

chat the chat but quite something else

30:52

on the manufacturing floor or in

30:54

a hospital in a hospital ER. They They also

30:56

have concerns over data privacy and the

30:58

security of proprietary information. information.

31:01

models more tailored to the needs of

31:03

various businesses, it's likely that

31:05

many will stay on the sidelines. sidelines.

31:07

Meanwhile, Meanwhile, Silicon Valley

31:09

and are Tech are obsessed with

31:11

intelligent agents with videos videos created

31:14

by AI. Individual Individual and corporate fortunes

31:16

are being amassed on the

31:18

promise of turbocharging smartphones and and

31:20

searches. searches. As in the early much

31:22

of the much of the rest

31:24

of the economy is being left

31:26

out. are not They're not either from

31:28

the financial rewards of the technology

31:31

or from its ability to expand

31:33

large sectors and make them more

31:35

productive. make them more Maybe it's

31:37

too much to expect big tech

31:39

to change Tech to change care about

31:41

using its massive power to benefit

31:43

sectors such as manufacturing. such as manufacturing.

31:46

tech does what it does. it does.

31:48

And it it won't be easy

31:50

for AI companies to rethink their

31:52

huge foundation models for such real -world

31:54

problems. They will They will need to

31:57

engage with industry experts from a

31:59

wide variety of sectors. and and respond

32:01

to their needs. But the But is

32:03

reality is AI companies are the companies

32:05

are the only the with

32:07

the vast computational power to

32:09

run today's foundation models the the

32:11

talent to invent the next

32:13

generations of the technology. So,

32:16

or like it or not, in

32:18

dominating the field on have taken

32:20

on the responsibility for its

32:22

broad applicability, whether they

32:24

will shoulder that responsibility

32:26

for all our benefit, or

32:28

once again ignore it the

32:30

the siren of wealth accumulation

32:32

will eventually reveal itself, perhaps

32:35

initially in those often

32:37

in those often nearly numbers from

32:39

the U .S. Bureau

32:41

of Labor Statistics website. of

32:43

Labor Statistics website. You were are

32:45

listening to to MIT Technology where

32:48

David Rotman writes, How

32:50

how to fine-tune AI for Prosperity. This

32:52

This article was published on

32:54

the 20th of August, of

32:57

August, and was read read by

32:59

Michael Satow for Noah.