0:00

This is Intelligence

0:02

Matters, sponsored by Ginkgo

0:04

Biosecurity, whose cutting -edge

0:07

biological intelligence platform

0:09

empowers national security decision

0:11

-makers to identify and

0:13

respond to emerging

0:15

biological threats. Alexander

0:20

Wang is the founder and CEO

0:22

of Scale AI, a leading

0:24

AI data foundry, driving

0:26

some of the most significant

0:28

advancements in artificial intelligence

0:31

today, spanning autonomous vehicles, defense

0:33

applications, and generative AI. Alex

0:36

started Scale AI in 2016

0:38

as a 19 -year -old MIT student.

0:40

Under his leadership, the company

0:43

has grown into a market leader,

0:45

serving clients from the U .S.

0:47

Department of Defense to Microsoft,

0:49

Meta, and Open AI. Alex

0:51

joins us today to discuss his journey as

0:53

a pioneer in AI, the

0:55

evolving role of data in driving

0:57

AI innovation and the implications

0:59

of AI for national security. We'll

1:01

be right back with Alexander

1:03

Wang. Biosecurity

1:12

matters. What's biosecurity? It's

1:14

the mission -critical industry safeguarding

1:16

against 21st -century biological

1:18

threats. Biology doesn't respect

1:20

borders. Ginkgo Biosecurity monitors

1:22

critical assets and produces actionable

1:24

biointelligence for national security

1:27

decision -makers confronting biological threats,

1:29

whether they are natural, accidental,

1:32

or malicious. Ginkgo's technology

1:34

tracks deadly pathogens worldwide, and

1:36

our globally -scaled platform, the first of

1:38

its kind, detects outbreaks in near

1:40

real -time, deploys AI -powered threat

1:42

assessments, and identifies the source

1:44

of a threat, mitigating outbreaks before

1:46

they become national emergencies. When

1:49

lives and livelihoods are on the

1:51

line, every second counts. Ginkgo Biosecurity,

1:53

the biosecurity partner for an

1:55

uncertain new era. Learn more

1:57

at biosecuritymatters .com. Welcome

2:00

to Intelligence Matters. Hey, listen, before

2:02

we start, let's sort of level

2:04

set with everybody. And I've heard

2:06

you in the past sort of

2:08

talk about the three pillars of

2:10

AI. And then of course, scale,

2:12

AI, your company fits into one

2:15

of those. So can you just

2:17

lay that out for everyone so

2:19

make sure we're all sort of

2:21

on the same sheet as we

2:23

go along? Yeah. So artificial intelligence

2:25

or AI has been a pretty

2:27

modern, well, this field itself exists

2:29

for a long time, but modern

2:32

AI using neural networks and sort

2:34

of the new tech trend has

2:36

been basically since 2011 or 2012.

2:38

And it relies on three fundamental

2:40

pillars. Data, compute, and algorithms. So

2:42

data is the raw material for

2:44

intelligence. Data is what the AI

2:46

needs to learn from. That's where

2:48

scale plays and that's where we

2:51

play. Compute are the large, all

2:54

the computational power, all the supercomputers,

2:56

all of the very advanced GPUs

2:58

and chips that need to go

3:00

into actually crunching all the data

3:02

as well as processing all the

3:05

information for AI systems. That's where

3:07

NVIDIA and others have been incredibly

3:09

pivotal. And then there's the models,

3:11

which are the algorithmic techniques used

3:13

to actually sort of take compute

3:15

and data together and turn them

3:17

into intelligence. And those are developed

3:20

by the major AI labs that

3:22

everybody knows, OpenAI, Anthropic, Google, DeepMind,

3:24

Meta, and other major companies. And

3:26

so these three ingredients are three

3:28

pillars, really are the basis for

3:30

everything in AI is built on

3:33

top of these three pillars, models,

3:35

data, and compute. And one of

3:37

the reasons why we've seen such

3:39

incredibly fast progress in AI over

3:41

the past few years is that

3:43

there's been really meaningful innovation across

3:45

every single one of these pillars.

3:48

So computational power that we've invested

3:50

far more, built much larger data

3:52

centers than we've ever built in

3:54

the past, as well as the

3:56

chips... themselves have gotten better due

3:58

to Moore's law. So computers advanced

4:01

pretty dramatically. On the algorithmic side,

4:03

there was breakthroughs. In 2017, we

4:05

had the transformer model, which is

4:07

a pretty meaningful breakthrough. And progress

4:09

basically has continued recently the 01

4:11

style of breakthroughs. So we've continued

4:14

to have sort of like algorithmic

4:16

breakthroughs. And then data has also

4:18

continued to sort of advance pretty

4:20

dramatically from a technological perspective. You

4:22

know, we started out by leveraging

4:24

all of the data off the

4:26

internet. We kind of ran out

4:29

of all that data. And now we're

4:31

sort of onto more complex and more

4:33

advanced data types, you know, complex reasoning

4:35

data, agent data, other forms of data.

4:37

And so all of the progress that

4:39

you see up here on AI is

4:41

built, is fundamentally driven by progress within

4:43

each of the pillars. So

4:45

Alex, let's sort of just talk broadly

4:47

here about the state of AI.

4:50

And just recently, I saw this chart,

4:52

this graph, it's by the tech

4:54

firm Gartner, and they named it after

4:56

themselves, the Gartner hype cycle. You

4:58

might have seen this graph. And for

5:00

our listeners, it sort of shows

5:02

a steep ramp up to what they

5:04

call peak of inflated expectations. Then

5:06

it leads to something called a trough

5:09

of disillusionment. It finally leads back

5:11

up to the plateau of productivity. So

5:13

I know it's a little tongue -in

5:15

-cheek, but if you're thinking of that

5:17

model, because I'm sure you've seen it,

5:19

so where are we today with AI?

5:21

Are we at inflated expectations or have

5:23

we not even gotten there yet? Or

5:26

are we past it? We

5:28

are probably somewhere in between,

5:30

yeah, we're in this early

5:32

phase. We're not quite at

5:34

the trough, because I think

5:37

expectations are obviously still quite

5:39

high. And there's an argument

5:41

that AI as a technology

5:43

may skip the trough of

5:46

disillusionment because the progress within

5:48

the field is just so

5:50

fast. I mean, it's

5:53

very, usually new disruptive

5:55

technology sort of toil

5:57

in relative obscurity for some time.

6:00

but AI has become the number

6:02

one focus for the largest companies

6:04

in the world, largest countries in

6:06

the world. I mean, it is

6:08

by many metrics the most critical

6:10

technology of today. And so there's

6:12

a theory that the sort of

6:14

just the sheer amount of investment

6:17

smart people, and then by force

6:19

of will, will kind of skip

6:21

a trough of disillusionment. But it's

6:23

undeniable that the expectations are very

6:25

high right now. Given that as

6:27

background, Alex, let's take a look

6:29

at, we're looking at the AI

6:32

landscape, but let's sort of look

6:34

at it in particular at how

6:36

it may reshape national security, right?

6:38

Because that's typically what we talk

6:40

about in the podcast here. And

6:42

I realize, you know, looking out

6:44

10 years is sort of crazy

6:47

because it's hard to predict what's

6:49

going to happen in, you know,

6:51

five years. But if you look

6:53

out into that sort of five

6:55

year time period, and you're thinking

6:57

about the national security challenges that,

6:59

you know, all of us can

7:02

sort of come up with, you

7:04

know, how do you broadly see

7:06

AI reshaping that? So I grew

7:08

up in Los Alamos, New Mexico,

7:10

the birthplace of the atomic bomb,

7:12

the home of the Manhattan Project,

7:14

both my parents worked as physicists

7:17

at the national lab. And so

7:19

I grew up, you know, really

7:21

deeply well studied on the sort

7:23

of story of, of the atomic

7:25

bomb, the sort of core scientific

7:27

discovery, and then how that played

7:29

out in national security in the

7:32

world thereafter. And nuclear fission is

7:34

a very interesting parallel because nuclear

7:36

fission as a, as a technology,

7:38

or as a scientific discovery is

7:40

obviously this incredibly brilliant scientific discovery.

7:42

And there's, there were sort of

7:44

two, there was a fork in

7:46

the road of two applications. One

7:49

is, you know, cheap energy for

7:51

the world. And the other, the

7:53

other direction was for, for weapons,

7:55

for bombs. And I think what

7:57

we've seen over the past, you

7:59

know, 80 years or so is

8:01

that the, the path to, to

8:04

commercialization of of nuclear energy has

8:06

been somewhat embroiled due to there

8:08

being various sorts of risks and

8:10

the regulatory path's not been very

8:12

clear and environmental concerns, all of

8:14

that. But the application towards weaponry

8:16

and national security has been immense.

8:19

Both the United States and Russia

8:21

and other countries, we all have

8:23

large nuclear arsenals. There were, we

8:26

certainly have more than enough

8:28

nuclear weapons in the

8:30

world to sort of create

8:32

a nuclear armageddon and

8:34

make the earth inhospitable. And

8:36

so I look at

8:38

this, what had

8:40

happened and then obviously

8:43

the sort of the world,

8:46

the national security community had

8:48

to sort of, had basically been

8:50

working quite hard to ensure that we

8:52

never to use those nuclear weapons

8:54

for fear of sort of the escalation

8:56

and the sort of damage to

8:58

the world and to humanity. And I

9:00

look at that sort of arc or

9:03

that story and I think

9:05

you have to take the lesson

9:07

and think about what's gonna

9:09

happen to AI. So artificial intelligence

9:11

is dual use technology. There's

9:14

an application towards the economy and

9:16

towards commercial applications of the

9:18

technology and then there's undeniably military

9:20

and national security applications of

9:22

the technology. And it's very clear

9:24

that the United States, we

9:27

have our eyes on these use

9:29

cases, but our adversaries do

9:31

as well. The application of AI

9:33

and autonomy in Ukraine, for

9:35

example, is undeniable by both

9:38

sides, by both the Ukrainians

9:40

and the Russians. And

9:43

so it's this technology

9:45

that I expect undeniably is

9:47

going to be quite

9:49

central to the national security

9:51

strategies of every country globally.

9:53

Like it's clearly going to

9:56

be where most countries and most

9:58

militaries and intelligence communities sort

10:00

of look to gain advantage over

10:02

competitors or over adversaries. And

10:04

the applications are going to be

10:06

pretty widespread. Like as a

10:08

technology, it's much more general purpose

10:11

than nuclear fission was. So

10:13

nuclear fission, you build warheads or

10:15

you build bombs, but AI,

10:17

you can use it for offensive

10:19

cyber hacking or offensive and

10:21

defensive cyber. You can use it

10:24

for autonomous drones and autonomous

10:26

weaponry. You can use it as

10:28

a tool for bio weaponry.

10:30

That's certainly like a concerning potential

10:32

vector. You can use it

10:34

for all your back office functions

10:36

and make it, you know,

10:39

so you're dramatically better planning. The

10:41

application towards intelligence is pretty

10:43

clear, like this process of converting

10:45

data to intelligence is something

10:47

that AI is potentially going to

10:49

be very good at. And

10:51

so the applications are endless. But

10:54

what I think is undeniable

10:56

is that as a new technology

10:58

with so much potential, it

11:00

is going to be a cornerstone

11:02

element to national security strategies

11:04

globally and will be one of

11:07

the major elements that will

11:09

dictate which militaries or which countries

11:11

have advantages over other countries.

11:13

So this is an interesting parallel

11:15

that you drew on here

11:17

a minute ago, Alex. And let

11:19

me ask, you know, maybe

11:22

what do you see as some

11:24

of the most pressing risks

11:26

posed by our increased reliance on

11:28

AI? And if I think

11:30

about your parallel, you know, one

11:32

of the key themes of

11:35

the last Cold War was this

11:37

mutual assured destruction, you know,

11:39

which, you know, today, maybe people

11:41

think it's crazy, but, you

11:43

know, it was a pillar, let's

11:45

say of that time period.

11:47

Is, do you see any parallel

11:50

there? Maybe not that extreme,

11:52

of course, but can we draw

11:54

that out a little more?

11:56

Yeah, totally. I think that deterrence

11:58

theory in an AI paradigm

12:00

is, you know, this is, this

12:02

is, the central thing to look at. And one of

12:05

the things that I, you know, I've always thought about that

12:07

as important is like, how would you, the

12:09

issue is we actually, we don't know,

12:11

we don't have all the answers about

12:14

AI yet. We don't know what

12:16

are the exact properties of

12:18

the technology when we get towards

12:20

super intelligence or very advanced AI

12:22

systems. So things that we don't

12:24

know. One, we don't know if

12:27

AI will require. a huge amount

12:29

of computational power to build and

12:31

even to use, or if you're

12:33

going to be able to use

12:35

a relatively small amount of computing

12:38

power to leverage the capabilities. Maybe

12:40

you need a lot of computational

12:42

power to produce advanced AI, but

12:44

then you need relatively little to

12:46

actually utilize it. That is very

12:48

wide-ranging implications, because if AI is

12:50

something that only the biggest countries can

12:53

utilize, then that is a much

12:55

safer world than if anybody

12:57

in the world can get

12:59

a handful of GPUs and

13:01

utilize the technology. Another big

13:04

question that we don't know

13:06

is, is it going to

13:08

be possible to build safeguards

13:10

into the models themselves to

13:12

prevent misuse or use for

13:14

harmful areas like, you know,

13:16

bio weaponry or building nuclear

13:18

capabilities or building other forms

13:20

of weapons? Or is that

13:22

something that is like technically

13:24

impossible to safeguard and therefore

13:26

adversaries are going to be

13:28

able to remove those safeguards

13:30

very easily? Terrorists are going

13:32

to remove those safeguards very

13:34

easily and then utilize it for

13:36

those dangerous capabilities. So

13:38

there's all these questions about what

13:41

are the end-state facts about AI as

13:43

a technology? And depending on those end-state

13:45

facts, you have very different dynamics.

13:48

So one version of the world is

13:50

that there's something very close to

13:52

mutually sure destruction. You know, this

13:55

version of the world looks something

13:57

like, you know, let's say advanced

13:59

AI. both takes huge amounts of

14:01

computational power to produce, to

14:04

develop, as well as huge

14:06

amounts of computational power to

14:08

utilize. In this scenario, you

14:10

know, the United States and

14:12

China, potentially Russia, and other

14:14

countries will be leaders, and

14:16

I think there will be

14:18

a sort of natural... there

14:20

will be some level of

14:22

natural stability because you know

14:24

the leaders all know the

14:26

the power of the technology

14:29

by and large they're not

14:31

really trying they know the

14:33

other leaders are are are

14:35

quite powerful they're you know

14:37

and so there's there's some

14:39

level sort of mutually assured

14:41

destruction or some level of

14:43

of stability. And and every

14:46

country has access to powerful

14:48

AI, every terrorist organization has

14:50

access to powerful AI, and

14:52

in that case, it's a

14:54

much more complicated world, because even

14:56

if you have, even if you're the

14:59

United States or you're one of these

15:01

major countries, it's hard to stamp out

15:03

all the potential misuse or all the

15:06

potential sort of chaos ensued by the

15:08

technology. The last point I'll bring up

15:10

is this question of first strike. So.

15:13

Mutually sure destruction is a good

15:15

paradigm because if you launch a

15:17

nuke, as soon as other people

15:19

know that you're launching a leader,

15:22

they can launch they can launch

15:24

their nukes. And so it's very,

15:26

very, very difficult to have first

15:28

strike to be able to prevent

15:30

sort of this mutually acceptable, or

15:32

unattributable first strike. Yeah, yeah, sorry,

15:34

undetectable first strike. The issue is AI

15:37

as a technology doesn't seem

15:39

to have those characteristics

15:41

characteristics. Like it seems. potentially possible

15:43

you can utilize artificial intelligence and in

15:45

a you know as a as a

15:48

weapon in a negative way without it

15:50

being as attributable or detectable in the

15:52

same way and then two there's an

15:54

argument that oh if my AI is

15:57

a hundred times smarter than my adversary's

15:59

AI they won't be able to

16:01

retaliate. They won't even be able

16:03

to retaliate. Like my AI will

16:06

just sort of, you know, swiftly

16:08

dominate my adversary's AI. So long

16:10

story short, I think that it's

16:12

a very complicated picture. And it

16:14

is not obvious to me that

16:17

we necessarily have the same sort

16:19

of deterrence theory of mutually sure

16:21

destruction or whatnot, and we may

16:23

not, we can't know. I think

16:26

until we learn more about the

16:28

technology. It's really interesting parallels, Alex.

16:30

And even in your first parallel, the,

16:32

you know, it's the purview of, you

16:34

know, great powers kind of thing. You

16:36

know, even the nuclear weapons, that monopoly

16:39

didn't last all that long. If you

16:41

look at it on a, you know,

16:43

time scale, right? Yeah. And so, you

16:45

know, these kind of things tend to,

16:47

tend to leak out. So what do

16:49

you, as we just started to talk

16:51

about some of the risks? How should

16:54

we think about mitigating some

16:56

of the risks of adversarial

16:58

AI? You know, deep fakes, some of the

17:00

malicious use you talked about. So it

17:02

seems like it's this constant

17:04

trying to stay ahead of somebody else's

17:06

AI, right? So you make your AI

17:09

better, theirs becomes better, you know, and

17:11

so on and so forth. So it's less

17:13

of a race than it is a, I guess because

17:15

a race has an end, and this

17:17

doesn't at least appear to me to

17:20

have an end. Yeah, it

17:22

is a, it is a, it

17:24

is an industry with constant one-opsmanship.

17:26

I mean, it's very relevant. So

17:28

today, literally today, the, so Open

17:31

AI two months ago released their

17:33

O1 model, which was the sort

17:35

of like this very advanced reasoning

17:38

model. And then two months later,

17:40

the very first replication of the

17:42

technology came out of Deep Seek,

17:44

which is a Chinese startup. and

17:47

Deep Seek is choosing to open

17:49

source it globally or open source

17:51

the technology. So this is

17:53

notable for two main reasons.

17:56

The first is that one

17:58

China is very clearly competent

18:00

and competitive, and was the

18:02

first public replication of the

18:04

most advanced technology in the

18:07

United States, which is, you don't

18:09

have to take that data point

18:11

very seriously. And the second is,

18:14

this is now open-sourced. And just

18:16

as we're talking about before, like,

18:18

you know, the proliferation environment, it

18:21

seems like there's a real proliferation

18:23

sort of prerogative or mandate driven

18:25

by the sort of private market

18:27

incentives. And so, On the answer

18:30

of like risks, well, I think

18:32

that my framework on this is

18:34

so they're clearly in the

18:36

abstract are ways in which AI can

18:39

be incredibly damaging and

18:41

incredibly harmful to humanity. And

18:44

so the key is that

18:46

we have very accurate measurement,

18:48

you know. AI is actually pretty

18:50

close to being very dangerous for

18:53

bi- weaponry, but maybe it's nowhere

18:55

close to being useful for nuclear

18:57

weapons. For example, or what are

18:59

the areas in which from a

19:01

practical standpoint, AI is actually poses

19:03

any real risk? And we need to

19:06

be extremely good at this measurement process,

19:08

this test and evaluation process for all

19:10

of the AIs that we see pop

19:12

up globally so that we can... we

19:15

can properly tackle the risks. Because what's

19:17

been clear is that, you know, the

19:19

sort of very generic concern that like,

19:22

you know, powerful AIs are really dangerous,

19:24

and so we need to be really

19:26

careful. Like the general argument is pretty

19:28

uncompelling because I think people can use

19:30

these systems and see how flawed they

19:33

are, but we need specific conversations. We

19:35

need a specific conversation that, you know,

19:37

these AI models are pretty close to,

19:40

you know, some dangerous capability, and so

19:42

we need to manage it. We've been

19:44

working at scale. We're proud to work

19:47

pretty closely with the Pentagon and the

19:49

DOD within their chief digital and artificial

19:51

intelligence office to actually produce sort of

19:54

a test and evaluation framework for the

19:56

use of elements in national security, in

19:58

defense, and military. to be able

20:00

to have accurate and reasonable measurements

20:03

of the sort of of the

20:05

risks associated with the technology. We're

20:07

in the process of scaling industrial

20:10

commercial use as well so we've

20:12

produced our seal evaluations where we'll

20:14

evaluate in kind of the same

20:17

way as consumer reports maybe does

20:19

for most other other products or

20:21

technologies, a public evaluation of all

20:24

the models, and there's sort of

20:26

various capabilities, including in areas like

20:28

agent safety or other dangerous capabilities.

20:30

But I think that we need to

20:33

keep going on this, and I certainly

20:35

encourage other people aid us in this

20:37

process of actually having very, very

20:40

accurate measurement to mitigate risks. So,

20:42

you know, this sort of leads

20:44

into this point about how the

20:46

US can maintain a competitive advantage.

20:48

in AI, we're the world leaders today,

20:50

right? And you can tell me I'm

20:53

wrong about that, but I think

20:55

we are. So developing AI, but

20:57

also ensuring responsible use, right? It's

20:59

not a Wild West scenario, which,

21:02

you know, maybe the case in

21:04

other countries. So how do we

21:06

think we strike that balance?

21:08

We are lucky that, so yeah,

21:10

the US is the leader. indisputably,

21:12

and we have, and this is

21:15

due to the benefit of our

21:17

frankly our incredible system, a lot

21:19

of AI research was was funded

21:21

by the private sector, has been

21:23

funded for decades by the public

21:25

sector, and that industry has attracted

21:27

huge amounts of commercial capital as

21:30

well as the best and brightest

21:32

people in the world, and we've

21:34

been able to produce and create,

21:36

you know, the largest data centers

21:38

of the world are being built in

21:40

the United States, the smartest people in

21:43

the United States, the most advanced data

21:45

for these systems is being produced in

21:47

the United States. So we have every

21:49

reason to maintain the advantage, but competitors

21:52

are quite close on the chase. So,

21:54

you know, as I mentioned, Deep Sea

21:56

replicated opening eyes models, Ali Baba, a

21:58

Chinese company released a new. model Quinn

22:01

which by some metrics is

22:03

actually the top LLLM in

22:06

the world so according to

22:08

Hugging Faces Open LLL leaderboard

22:11

is actually outperforms many

22:13

Western counterparts and so

22:15

fundamentally I think we the United

22:17

States is in a we have

22:19

a very tight-type to walk which

22:22

is that on the one hand

22:24

we have global adversaries who are

22:26

catching up and we need to

22:28

stay ahead of. That's undeniable. The

22:30

United States needs to stay ahead.

22:33

But on the other hand, because

22:35

we are the leaders, it is

22:37

up to a certain extent to

22:39

really set the right guard rails

22:41

and set the right standards for

22:43

how this technology should develop in

22:45

a safe manner. And luckily, because

22:47

so much of the progress in

22:49

other countries is due to almost

22:52

pure replication of what the United

22:54

States does, we develop in a

22:56

direction and build regulation or are

22:58

thoughtful about building in a direction

23:00

that mitigate some of the risks,

23:02

then we actually can be in a

23:04

pretty good spot as a country. and

23:06

set the world up for this new

23:08

technology in a good way. You know,

23:11

no specific answer there, but I think

23:13

it's an incredibly hard question, and I

23:15

don't think the answer is necessarily accelerate

23:17

at all costs, but it's probably to

23:20

leverage our leadership position to thoughtfully set

23:22

the standards for the world. So that's

23:24

a great lead into the next question

23:26

I was going to raise, Alex, and

23:29

that's the role of government regulation,

23:31

for lack of a better way to put

23:33

it. So what role should the

23:35

US government play in play in?

23:38

in regulating AI development, but also

23:40

making sure that we stay

23:42

ahead as you just described.

23:45

Clearly there's a role here. Where

23:47

do you think the left and

23:49

right boundaries are? I think

23:51

the most critical piece, there's

23:54

a few pieces that I view as

23:56

like, you know, hands down, absolutely

23:58

critical. which are one,

24:01

ensuring integration into our national

24:03

security community. So we have

24:05

all this incredible commercial technology.

24:08

It needs to get integrated

24:10

into the DOD and IC very, very

24:12

rapidly. That's like, you know, kind

24:14

of immensely, immensely important. And I

24:16

think the second is also ensuring

24:19

that we have, kind of as I

24:21

mentioned, we have the right measurement

24:23

mechanisms and schemes to ensure that

24:25

we. are cognizant of when AI

24:28

capabilities are at a point where

24:30

they create real societal level risks

24:32

or around, you know, buy a

24:34

weapon re, nuclear weapon re, you

24:36

know, stuff like that. I think

24:38

those are relatively, those are very

24:40

clear. I think that the, the

24:42

additional, you know, if you get further from

24:45

that, I mean, I think there's. You

24:47

can have a lot of debate

24:49

about what regulations are needed or

24:51

important or what's important after that.

24:53

I think one perspective, which I

24:55

really I admire and I appreciate,

24:58

is, hey, we have to ensure

25:00

that we stay ahead. So, you

25:02

know, there's the use of regulation

25:04

to accelerate the ability to build

25:06

out large data centers to enable

25:09

sort of a Manhattan project like.

25:11

build out of capacity in the

25:13

United States, that certainly seems important.

25:15

Other things are in ensuring that,

25:17

you know, I do think the guardrails

25:19

that we set around the technology

25:21

are going to be important as

25:24

it continues becoming more powerful,

25:26

so we need a thoughtful

25:28

approach to both. But the

25:30

national security stuff and the

25:32

measurement of national security related

25:34

capabilities are table stakes in

25:36

my view. We're going to take

25:38

a quick break and we'll be back with

25:40

Alex Wang. Beacon Global Strategies is

25:43

the premier national security advisory

25:45

firm. Beacon works side by

25:47

side with leading companies to

25:50

help them understand national security

25:52

policy, geopolitical risk, global technology

25:54

policy, and federal procurement trends.

25:57

Beacon's insight gives business leaders

25:59

the... Decision Advantage, founded in

26:01

2013, Beacon develops and supports

26:04

the execution of bespoke strategies

26:06

to mitigate business risk, drive

26:09

growth, and navigate a complex

26:11

geopolitical environment. With a bipartisan

26:13

team and decades of experience,

26:16

Beacon provides a global perspective

26:18

to help clients tackle their

26:20

toughest challenges. Alex, what do you

26:23

see as the role here for

26:25

international cooperation in order to

26:27

help sort of develop some of

26:29

these guidelines? there seems to be

26:31

a place here for like-minded nations,

26:33

right? It's not everyone's

26:36

going to participate, but

26:38

how do you see that

26:40

developing? Absolutely critical. You know,

26:42

I think that the entire

26:44

world I think is going

26:46

to be safer, stronger, friendlier

26:48

if, you know, the United

26:50

States is in the driver's

26:52

seat in ensuring that there's

26:54

international cooperation and that

26:56

we, you know, this is... The

26:58

stakes here are so high fundamentally.

27:00

I think I was listening, I

27:02

was reading, listening to a podcast

27:05

by a sort of war historian who

27:07

said that, you know, wars are won

27:09

by great alliances more than anything else.

27:11

I don't know specifically that's true, but

27:13

I think in this case it's pretty

27:15

true, which is that the AI is a large

27:18

scale, human level, humanity level, infrastructure,

27:20

and scientific project, and as we

27:22

embark on it, we want to

27:25

ensure that we have the... the

27:27

right levels of international cooperation. So,

27:29

relevantly, today in San Francisco, there

27:32

is a convening of the global

27:34

AI safety institutes. So, many countries

27:36

now, you know, the United States,

27:38

the UK, Japan, many countries have

27:41

established AI safety institutes and are

27:43

convening and sort of collaborating on

27:45

ensuring that each of these countries

27:48

is sort of the right guardrails in place

27:50

to ensure that. we mitigate the sort of

27:52

like negative risks. I think that's a great

27:55

step in the right direction. I think that

27:57

there's probably more to do. There's probably deeper

27:59

collaboration. at Iron Heart levels of

28:01

government coalitions, and so I

28:04

think it's absolutely necessary.

28:06

So Alex, something triggered when you so

28:08

said humanity level, and so you don't

28:10

see a AI winter, I guess, coming

28:12

again, right? I mean, that's been the

28:14

history, look, when I was getting my

28:17

graduate degree in computer science in the

28:19

Dark Ages, AI was just around the

28:21

corner. There was claim was that you

28:23

could just almost touch and feel it,

28:26

and of course. you know, nothing happened

28:28

for a long time. So you don't

28:30

see this win, you know, AI winter,

28:32

you just see this progressing, you know,

28:35

the slope of the curve, you know,

28:37

we could probably debate what that would

28:39

look like, but you don't really see

28:41

a downturn, right? It's a complicated

28:43

question because the answer is yes

28:45

and no. I think pure scaling,

28:48

which is what has been the

28:50

sort of technical calling card of

28:52

AI for the past few years,

28:54

which is all we need is more

28:56

and bigger data centers and more data

28:58

and you know the thing will just

29:00

get smarter. I think this sort of

29:03

like pure scaling approach is hitting

29:05

limits. There's some winter associated with

29:07

that approach but at the same

29:09

time there's totally new approaches like

29:11

post training, test time compute and

29:14

and other new technical approaches which

29:16

seem to have very high ceilings

29:18

as a technology. So there's... One

29:20

way to put this is going

29:22

back to the pillar analogy at

29:24

the start. One of the beauties

29:27

of AI is because there's these

29:29

three pillars, compute data and algorithms.

29:31

If you hit roadblocks in some

29:33

of them, let's say we hit

29:35

a roadblock in computer, a roadblock

29:37

in data, then you still can

29:39

work on the others. And so

29:41

as long as one of these

29:43

pillars is humming and driving progress,

29:46

we're going to see continual progress

29:48

in AI systems. One of the

29:50

big issues, especially when you think, well,

29:52

I guess in a lot of applications,

29:54

but when you think about national security

29:56

for AI, that sort of three things,

29:59

transparency, accountability, explainability. Right and I can

30:01

we get to I mean right now you've

30:03

got you know, and maybe you believe

30:05

this too There AI does things that people

30:07

can't quite explain You know

30:09

how did it do that? You know how

30:11

did it come to that conclusion which you

30:13

know creates it? You know challenging a lot

30:15

of fields, but in particular national security So

30:18

how do you see that shaping? Yeah

30:22

So what one of the things that

30:24

I think is is really important

30:26

for national security use cases is is

30:28

a fundamental level of You know

30:30

transparency and explainability so in all

30:32

of the Systems that

30:35

we've built for the national security

30:37

community like the DoD itself and

30:39

others we have had an approach

30:41

which Which all

30:43

claims need to be backed up

30:45

by Grounded and verifiable information and those

30:47

need to be those sources need

30:49

to be sort of cited and sourced

30:51

and Tied with all the claims

30:53

because obviously if you don't have this

30:56

verifiability then you know you we

30:58

have You know we have bigger we

31:00

have bigger problems And so so

31:02

that's been pretty critical and core to

31:04

our approach I think has enabled

31:06

our system scaled on of into to

31:08

have much greater impact than some

31:10

other AI systems Where these guarantees weren't

31:12

there looking forward. I think

31:15

the technology is actually moving

31:17

in direction of Greater transparency and

31:19

and explainability, you know

31:21

with the new Thinking models

31:23

or the new reasoning models you can

31:25

actually just you know The model tells

31:27

you what it's thinking and there have

31:29

been cases where we've found the models

31:31

will even say when it's trying to

31:33

be deceptive Like it'll tell you when

31:35

it's trying to be deceptive Yeah, it

31:38

admits it so so the industry is

31:40

moving in a direction of like greater

31:42

transparency or greater explainability which is good,

31:44

but I think many people's greatest fear

31:46

and I probably share this is We

31:48

hook AI systems up to very powerful

31:50

weapon systems and they make decisions that

31:52

are totally unexplainable Inexplicable and I think we're

31:54

gonna avoid that for what it's worth. I

31:56

don't think that's like a real near -term

31:58

risk, but think we need

32:01

more progress in this direction,

32:03

for sure. Yeah, that's probably not

32:05

a near -term risk for the

32:07

US, but others may be

32:09

more willing to allow that to

32:11

happen, right? 100%, yeah. So

32:15

since you sit on top of

32:17

this big data pillar and

32:20

the fact that these

32:22

huge amounts of data have

32:25

now been ingested, how

32:27

do we ensure security and

32:29

privacy of sensitive information, particularly

32:33

when it has, I mean, there's personal privacy,

32:36

but particularly when it has national security implications,

32:38

right? You need to have as much data

32:40

as possible, the more the better. Something like

32:42

you just explained a few moments ago. Is

32:45

there a way we can ensure the security

32:47

of that information? Yes,

32:50

this is possible. I think

32:52

that there's, the AI industry has

32:54

gotten potentially a bad reputation

32:56

because in some cases, some of

32:58

the most prominent actors are

33:00

a little bit loose with how

33:02

they work with data. But

33:04

I think there's no fundamental reason

33:07

why AI and rigorous standards

33:09

around data security and privacy are

33:11

fundamentally at odds. So for

33:13

example, in our work with government

33:15

work, we're authorized at Fedoramp

33:17

High. We comply with all the

33:19

necessary data security authorizations, and

33:21

then we employ these pretty deeply

33:23

in our methods for how

33:25

we've developed this technology for national

33:28

security use cases. So I

33:30

think it's important, and I think

33:32

probably what's more important than

33:34

anything is that we ensure that

33:36

for all national security purposes,

33:38

we're doing things right and properly.

33:40

And I think that's very

33:42

much feasible and possible. So

33:45

Alex, as we sort of

33:48

just, again, sort of look a

33:50

little bit forward and the

33:52

implications of AI for geopolitics sort

33:54

of call it that. What

33:56

are the emerging, we just talked

33:58

about this inference capability. But what

34:00

are the emerging technologies that

34:02

you believe will have the AI

34:04

emerging technologies that you think will have

34:07

the greatest impact on national security? I

34:10

think that there's a few

34:12

tranches. I think the first

34:14

is obviously AI and then

34:16

AI applied to every field.

34:18

So AI applied to cyber

34:20

capabilities, AI applied to greater levels

34:23

of autonomy, AI applied to

34:25

biology and chemistry and medicine,

34:27

AI applied to advanced manufacturing.

34:29

So you can peer down

34:31

each of these threads and

34:33

see just sort of like, you

34:35

know, I expect sort of

34:37

10x or greater improvements

34:39

in each of these

34:41

areas, which is going to

34:43

be pretty incredible. But

34:45

I would say specifically, if

34:48

you think from a

34:50

military standpoint, the use

34:52

of AI that is

34:54

quite fundamentally disruptive

34:56

is in autonomy.

35:00

And, you know,

35:02

the replacement

35:04

of manned systems to

35:06

unmanned systems to autonomous systems

35:08

to a treatable autonomous systems,

35:10

like that entire chain or

35:13

sort of like progression pattern,

35:15

we're seeing basically play out

35:17

in Ukraine. Or like, you

35:19

know, some folks have mentioned

35:21

to me that basically the level

35:23

of iteration and the fully

35:25

autonomous systems are going to be

35:28

deployed either by the Ukrainians or

35:30

the Russians are moving just incredibly

35:32

quickly on the drone level.

35:34

This flips military doctrine on its

35:36

head a little bit because you

35:38

have a very different calculus,

35:40

you have a very different reality,

35:42

and all of a sudden you

35:44

get to a point where

35:46

the, you know, warfare becomes to

35:48

some extent somewhat definitely difficult to

35:50

fully perceive and comprehend

35:53

and control. And so,

35:55

autonomy is clearly one

35:58

major theme, which I think is of... And

36:00

then I think from an

36:02

intelligence standpoint, I think that the

36:04

world has to grapple with

36:06

this sort of like new reality,

36:08

certainly, of open source intel.

36:10

And there's just so much information

36:12

out there. And how do

36:14

we think about all that? I

36:16

mean, I think these clearly

36:18

seem like the big themes that

36:20

seem quite important. So that's

36:22

a great point, Alex. And if

36:24

you think about it, not

36:26

only from an intelligence, but if

36:28

you're an analyst almost of

36:30

any kind looking at data, whether

36:32

it's geopolitical data or whatever,

36:34

it's becoming impossible to think about

36:36

doing that without AI right

36:38

next to you helping you. The

36:41

amount of information is so vast

36:43

that there's just no way for a

36:46

human being to go through it

36:48

anymore. All

36:51

right, let me ask you

36:53

for a prediction. How far

36:56

away are we from artificial

36:58

general intelligence? It's

37:03

the challenge is so hard to

37:05

define exactly what it is. But

37:08

I do think we're sort of

37:10

some small number of years, let's

37:12

say under five years from very

37:14

powerful AI systems that will vastly

37:16

outperform humans in many areas and

37:18

still be somehow worse than humans

37:21

in a host of other areas.

37:23

I mean, in some sense, we

37:25

already have them today, but I

37:27

think that trend will continue and

37:29

over the next few years, we're

37:31

going to see progressively, basically we'll

37:33

see systems that are better than

37:36

humans at more things while still

37:38

being worse than humans at others.

37:40

Sorry, Alex. Last question. Are you

37:42

as worried as some AI luminaries

37:44

who've worn against a

37:48

almost apocalyptic future where AI takes over?

37:50

I mean, if you listen to them,

37:52

that's sort of the path that they're

37:54

going down. Where do you come out

37:56

on that? I

37:59

think that Instead, if you look at

38:01

it, I think that there's a, you

38:06

know, with any discontinuous

38:08

technology, there's obviously always a

38:10

fair amount of risk, but

38:12

the most complicated question is

38:14

whether or not AI develops

38:16

the ability to self -improve. And

38:19

this, I think,

38:21

is the core, you

38:23

know, I think when

38:25

people have, you know,

38:27

some called doomer or, you know,

38:29

very concerned about AI risks and

38:31

others aren't, I think the core

38:33

question is, do they believe whether

38:35

or not AI can self -improve? Because

38:38

in a world where AI can't

38:40

really self -improve, then I think risks

38:42

are very manageable, right? Like, you

38:44

know, we're developing a technology, we'll

38:46

know what the limits are. If

38:48

we get to any point that's

38:50

concerning, we'll be able to, you

38:52

know, stop it. That's a governable

38:54

technology. If instead you get to

38:56

a point where AI is meaningfully

38:58

self -improving, which is clearly plausible,

39:00

but we have not seen

39:02

yet in a very in

39:05

a very real way.

39:07

But if you get that,

39:09

then you have this technology, this,

39:11

effectively this form of like digital

39:13

life that evolves at a faster rate

39:15

than humans do, and, you know,

39:17

then you might get worried that natural

39:20

selection might benefit the sort of

39:22

AI systems versus the human systems. So

39:24

I think that really is like

39:26

where the question boils down to. That's

39:28

kind of how I frame it.

39:30

If we see real self -improvement out

39:33

of AI models, I

39:35

think we ought to really take a

39:37

step back and think quite hard about the

39:39

technology. In a paradigm where we still

39:41

don't, then I think we need to be

39:43

more worried about misuse and be more

39:45

worried about, okay, how will that actually utilize

39:47

this technology and what does that mean

39:50

for humans? Alex,

39:52

that was a great discussion. Thanks so much. I mean,

39:54

there was a lot to think about there, and

39:56

thanks for taking the time to be on the podcast.

39:59

Can't thank you so much for me. me. I I

40:01

agree, Intelligence Matters. Matters. That was Alex Wang.

40:03

I'm was Alex Wang. I'm

40:05

Andy join us Please join us

40:07

next week for another episode

40:09

of Intelligence Matters. And you

40:12

can always reach us at always

40:14

reach us at gmail .com. pod@gmail.com.

40:21

Intelligence Matters is produced

40:23

by Steve Dorsey assistance from

40:25

from Ashley Barry. Matters is

40:27

a production of of Beacon Global

40:30

Strategies.