Latest / China Considered / China's AI Breakthrough: DeepSeek vs. American Dominance with Amy Zegart | China Considered | Hoover Institution

Transcript

China's AI Breakthrough: DeepSeek vs. American Dominance with Amy Zegart | China Considered | Hoover Institution · China Considered

0:08
[MUSIC] >> Elizabeth Economy: Welcome to
0:09
China Considered, a podcast that brings fresh insight and
0:12
informed discussion to one of the most consequential issues of our time,
0:15
how China's changing and changing the world.
0:18
I'm Liz Economy Hargrove, Senior Fellow and Co-director of the program on
0:21
the US-China and the World at the Hoover Institution at Stanford University.
0:26
Today I have with me my good friend and colleague Dr. Amy Zegart.
0:29
She's the Morris Arnold and Nona Jean Cox Senior Fellow here at Hoover and
0:33
also teaches in Stanford's Political Science department.
0:36
She's a specialist on all things technology and national security.
0:41
And today, we're gonna talk about one of the seminal issues of our time,
0:45
the DeepSeek moment.
0:46
Welcome, Amy.
0:47
It's great to have you here.
0:48
>> Amy Zegart: It is so nice to be with you, Liz.
0:50
No place I'd rather be.
0:52
>> Elizabeth Economy: Okay, so we're gonna talk about DeepSeek.
0:55
Let's just start with where were you and
0:58
what was your reaction when you first heard about DeepSeek?
1:02
What about this DeepSeek moment?
1:05
>> Amy Zegart: So DeepSeek, as you know, created a deep freak.
1:08
And so in the beginning, I was watching the press coverage and thinking,
1:12
this can't possibly be the full story.
1:15
And a couple weeks after, I was in a meeting at the AI Institute at Stanford
1:19
with a bunch of computer scientists and others around the table.
1:24
And it was a remarkable conversation because what I realized was, first of all,
1:28
a lot of the media coverage was wrong.
1:31
And secondly, DeepSeek was not a surprise to any of the faculty in AI at Stanford.
1:37
It was a surprise to markets, it was a surprise to investors.
1:41
It was a surprise to the CIA.
1:43
It was not a surprise to people in computer science because they've been
1:46
following DeepSeek for a couple of years.
1:49
>> Elizabeth Economy: So what made DeepSeek?
1:51
What made the DeepSeek moment?
1:53
Why was this such a big deal?
1:55
Even if, and I agree with you,
1:57
because I happen to be at a meeting in Southern California with a lot of
2:00
sort of representatives from AI companies here in the United States.
2:04
And while I was surprised and some of the other China folk around the table
2:08
were surprised, they were not at all surprised.
2:10
Exactly what you said.
2:11
They'd been tracking DeepSeek for a while, but clearly for
2:15
the majority of Americans and for the US Government, it was a big surprise.
2:19
What made this so significant?
2:21
>> Amy Zegart: So I think it's important that it was a surprise, number one.
2:25
So the fact that our government was taken by surprise should concern all of us.
2:30
So, as you know,
2:31
I'm really worried about the broader category of strategic technical surprise.
2:35
How do we anticipate what's happening in technology, when could it come confer
2:39
advantage to the Chinese that we might not be able to catch up on.
2:43
And so this was a near miss in terms of that.
2:46
But your question, why was this a big deal?
2:48
And it is a big deal.
2:50
And I think there are really four reasons why it's a big deal.
2:53
The first is the DeepSeek moment said, it's game on.
2:58
The thinking, you know China much better than I do.
3:00
I think the thinking in Beijing, certainly the thinking in Washington,
3:04
was the US is ahead, really far ahead in large language models.
3:07
And turns out that's not the case.
3:09
The second to the uh-huh part of this DeepSeek moment was it
3:14
clarified what race we're in.
3:17
So we often talk in technology about we're in a race with China.
3:20
Well, what's the race?
3:22
I think until DeepSeek, most people thought the race was to invent.
3:26
Who's gonna invent the latest model first?
3:29
Well, what DeepSeek revealed was the race is to adopt, not invent.
3:33
So DeepSeek didn't invent.
3:34
It wasn't a major engineering milestone.
3:38
But what they did was basically replicate the performance of
3:42
the best OpenAI model without access to anything that OpenAI did.
3:47
So they knew that this thing existed and they could recreate it with less money,
3:52
less compute,
3:53
less frontier capabilities than the best companies in the United States.
3:58
So it's a race to adopt globally AI platforms.
4:02
And DeepSeek is cheaper, it's free in the third world.
4:06
And so the US companies are racing to be first to invent.
4:09
And maybe that's not the race we should be worried about.
4:12
The third thing is, I think, and we can get into this,
4:15
that it seems like a very nerdy debate, but
4:17
it's a really important debate about open models versus closed models.
4:22
And what DeepSeek did was say really suggest the future is open.
4:28
And by that I mean they published what they did, their weights were open, which
4:33
enables other people to modify, accelerate and replicate what they're doing.
4:39
Most US companies, Meta accepted, take a closed model approach.
4:43
They don't publish what they do.
4:46
They don't share this information.
4:47
They think they're creating a moat.
4:49
I think what DeepSeek is suggesting is that is a failing strategy.
4:52
So for the United States, this is a big deal.
4:54
And then last, I think it reveals a lot about talent.
4:57
And I know we're gonna talk more about talent, but
5:00
who are these DeepSeek researchers that created this really important moment?
5:05
The answer is they were not people who were trained in the United States.
5:09
>> Elizabeth Economy: Right and you've done some really novel research into that
5:12
last point that you just made.
5:14
And I do wanna talk about it, but let me just pick up on the third point about
5:18
the sort of open versus sort of protected models, right,
5:22
where they share, you don't share.
5:24
I mean, it would seem to me then that if DeepSeek basically
5:27
has laid bare everything they did, how they got there,
5:31
why couldn't be the case that some very young, talented AI researchers and
5:35
computer scientists here in the United States could take that and
5:39
just replicate it and improve upon it, and then we're back in the game again?
5:45
>> Amy Zegart: Well, I think that's already happening, right?
5:47
Researchers are already going to town on the DeepSeek model and
5:50
this is a learning from learning.
5:52
So the Chinese are learning from the Americans,
5:54
the Americans are learning from the Chinese.
5:57
But I think we're in a bizarro world where the US Used to be in favor
6:02
of open research.
6:04
Universities publish things openly.
6:05
We think we accelerate innovation faster that way.
6:09
We have open international collaborations.
6:11
And now it's the Americans that are saying, actually,
6:13
we don't want anyone to know what we're doing.
6:15
We're gonna operate in secret.
6:17
We won't tell our employees can't say what they're doing.
6:19
We're not publishing what we're finding.
6:21
And the Chinese are saying that they're open and accelerating research.
6:26
And so within computer science departments, where are researchers
6:31
turning to use models to develop their capabilities in AI China?
6:36
Why?
6:37
Cuz the models are open and they can play with them.
6:40
>> Elizabeth Economy: Okay, we have Meta.
6:42
So I'm just gonna push a little bit on this point, I really understand it.
6:48
We do have Meta, which is open, right?
6:50
And which does share.
6:53
Isn't it okay to have some of our companies doing approaching this one way,
6:57
some approaching another way?
6:59
And I mean, shouldn't we be a little protective of our IP?
7:02
I mean, I think across the full sort of range of technologies and
7:07
over time, where the United States has been all too willing to share,
7:11
or sometimes China has just appropriated the technology,
7:15
it feels as though there's a reason behind what these companies are doing.
7:21
>> Amy Zegart: Absolutely and you bring up a good point.
7:23
And we should distinguish between, there are different levels of openness.
7:29
So DeepSeek did not reveal its data.
7:31
We don't know what data it trained on and there's all sorts of consternation about.
7:35
Well, really, did they just train on data from other models in the United States.
7:38
They released the model weights, not the data they use.
7:41
That's a big difference for people in the field.
7:44
And there is this difference that Meta is the only major company that has sort of
7:49
an open weight model.
7:51
As you know well, there are real concerns, legitimate concerns about, number one,
7:56
how do we make sure these models don't do things that are really unsafe for
8:00
everybody?
8:01
And the horse leaves the barn and we can't get the horse back in the barn.
8:05
And that's a really important set of concerns.
8:08
And number two, what you raise is why would we just let all this technology and
8:12
IP walk out the door for the world?
8:14
Like, shouldn't we protect it, like patents, etc.
8:17
But the DeepSeek moment really changed my mind on a lot of these things.
8:22
I think, as you know better than anyone, Liz,
8:25
this is a technology that naturally diffuses.
8:28
You can't keep it to yourself.
8:30
It's always going to get out because it's not like nuclear material where you
8:35
can control it.
8:36
And much of the capability to develop these models is right inside people's
8:41
heads.
8:41
It's the ultimate portable weapon.
8:43
So this technology isn't born classified.
8:46
You can't keep it.
8:47
And so if that's the case, then shouldn't we be racing faster to be ahead and
8:52
set the standards and understand the guardrails and try to set norms?
8:58
So my thinking really has changed because of DeepSeek.
9:01
I used to think much more that we should lean more toward pausing,
9:06
protecting, and now I actually think that is a misguided strategy.
9:11
And I'm really concerned that US Companies are so focused on beating
9:14
their American competitors, they're going to lose to their Chinese competitors.
9:19
>> Elizabeth Economy: Well, that is a really important point.
9:22
I mean, I would guess that the DeepSeek moment has sort of profoundly,
9:27
I think, shaped your thinking.
9:29
My guess is that it's probably shaped thinking of some people at OpenAI and
9:33
Anthropic and other places about what they need to be doing differently, but
9:37
maybe not.
9:38
And I think one of the big issues that you touched on in
9:43
the fourth point was the talent issue and DeepSeek.
9:47
The sort of the narrative around DeepSeek was that its model was developed by
9:52
a very small team of young scientists and developers, virtually all of whom,
9:57
as you said, were educated exclusively in Chinese universities.
10:02
I think your research shows part of that's true, part of that's not true, but
10:05
has huge implications, which I wanna get into.
10:08
But talk a little bit about what you found in your research, and how you went
10:13
about it, because I think really it's a unique lens into this company.
10:18
>> Amy Zegart: So let me just take a step back and say,
10:20
why do we care about where the talent comes from?
10:23
I think that knowledge power is far more important in today's
10:27
geopolitical competition than it's ever been before.
10:31
So, yes, military hard power matters.
10:33
Yes, soft power of our values still matters, but increasingly,
10:37
because technology is so central to economic competition and security,
10:42
competition, knowledge, power, the ability to innovate is really important.
10:47
And one key component of that is where the talent goes.
10:51
So, you know they're often these Pentagon maps that show, how many nuclear missiles
10:56
does this country have or how many tanks does that country have.
10:59
I would like to understand that map for AI researchers around the world.
11:05
How many cutting edge AI researchers are in Beijing versus in Silicon Valley.
11:10
So that's the idea behind it.
11:12
And then, so what we did was I asked this amazing research assistant who you now
11:16
have hired full time, named Everson Johnston.
11:20
Can you just figure out what you can find about all the people who are listed on
11:24
these DeepSeek papers?
11:26
And there wasn't just one paper, which is what the media focused on,
11:29
there were five.
11:30
Let's look at all five papers this company has ever produced.
11:34
Gather up all the information you can about the 223 people that were listed as
11:38
contributors in some way.
11:41
She did an amazing job.
11:43
She used AI to study AI.
11:46
So she developed her own AI scraping tools to look across the Internet at
11:51
everything she could find about the authors.
11:54
And of the 223 total authors that contributed to any of those papers,
11:59
she found incredible data about 201.
12:02
So a lot.
12:03
>> Elizabeth Economy: Yeah.
12:03
>> Amy Zegart: And then what we did is we looked at where did these people live,
12:08
where did they train, where did they work, and where did they go to school.
12:13
And what can we infer from the patterns that we found?
12:18
We found a couple of really interesting things.
12:21
One is that 98% of them had significant training in China.
12:27
This is not the story we often hear, which is China sends all of its best and
12:30
brightest in the United States.
12:31
We give them all the great things that they learn, and
12:34
then they go back to China.
12:35
What we also found, to me, this is the most disconcerting thing we found more
12:41
than half of those researchers in DeepSeek have trained nowhere outside of China.
12:46
They have spent their entire life inside China,
12:50
which tells me China has an incredible and robust domestic talent pipeline.
12:57
We do not have that kind of pipeline in near the numbers that China does.
13:02
So we are asymmetrically vulnerable.
13:05
If you think about this as a supply chain of talent, or
13:08
we rely much more on foreign talent in STEM than China increasingly does,
13:13
so they can grow their own in a way that we can't large population also.
13:17
So this suggests that we need to compete for
13:20
global talent in a much more vigorous way than we've done, not just grow our own.
13:24
We have to do all of the above.
13:26
I would say two more things that we found.
13:30
One is that of the 49 DeepSeek researchers that spent time in the United States,
13:35
most of them only came for a year.
13:38
They didn't spend a lot of time here.
13:40
They spent one year here and
13:42
they went to 65 different institutions in 26 different states.
13:46
This is a very widespread, unusual.
13:50
I would have expected a concentration in CS departments at Stanford and Cal and
13:55
MIT, right, the three top departments.
13:58
That's not what we found.
13:59
Whether this is delivered or not, we don't know.
14:01
But it's an interesting pattern and
14:03
a short pattern of people coming to the United States and then leaving.
14:06
>> Elizabeth Economy: Well, can I just stop you there for a second?
14:08
I mean, presumably not all of them can get into Stanford, at MIT and Caltech, right.
14:14
So it's maybe not surprising that some went elsewhere.
14:18
I mean, were they all studying computer science?
14:22
>> Amy Zegart: No. >> Elizabeth Economy: Were they not.
14:24
>> Amy Zegart: So we found some of them were in medical programs,
14:26
some of them were in bioengineering programs.
14:28
So a much more widespread set of programs that they went to,
14:32
not just computer science departments.
14:34
>> Elizabeth Economy: Right, and I think that was one of the things that
14:37
the founder of DeepSeek mentioned, right, when he was talking about his strategy was
14:41
that he was bringing together people from sort of disparate fields.
14:45
This was not simply a group of computer scientists.
14:49
So I think that's reflective.
14:51
What you found, I think, is reflective of some element of his strategy and
14:55
who he brought in.
14:56
>> Amy Zegart: And so I'm glad you raised that point, Liz,
14:59
because the mantra has been they're copying US companies.
15:02
DeepSeek didn't copy US companies.
15:04
DeepSeek copied US universities, right?
15:07
And that's the model of bringing people together across fields and
15:11
having a lot of young people involved, right?
15:13
And so the last thing that we found was they may be young,
15:17
the DeepSeek talent, but they ain't green, right?
15:20
So you look at citation metrics, right?
15:23
So Alexander Wang, who is just hired by Meta, right,
15:28
an AI superstar is 28 years old.
15:32
>> Elizabeth Economy: Yeah. >> Amy Zegart: He's young, but
15:33
he is not green.
15:34
Well, DeepSeek showed a similar pattern.
15:36
So we looked at the citation counts of the researchers, and
15:39
we compared the median and the average to citation counts to the extent we could at
15:44
OpenAI in one of their papers.
15:45
And what we found was this was a group that was actually pretty highly cited.
15:51
They had a lot of research under their belt.
15:52
They were young but they were not inexperienced.
15:55
And in fact, the sort of median level of citation was higher
16:00
relative to the average than OpenAI.
16:03
What does all that mean?
16:04
It means they relied on fewer superstars than the OpenAI paper we looked at, and
16:09
their.
16:10
Their average bench was better.
16:13
>> Elizabeth Economy: That is important.
16:15
I also thought, if I recall correctly, though,
16:18
that it didn't skew quite as young as we might have thought, right?
16:22
I mean, I have to say, when I read about it, just from the press reports,
16:27
that I was envisioning this, you know, small garage type environment with 20,
16:34
15, 20 young people ages, 18 to 25.
16:38
But it seemed to me that a number of the most senior people were actually quite
16:41
a bit older than that, is that fair?
16:44
>> Amy Zegart: I think so, yes.
16:45
I mean, I have to look specifically at the data, but I also think there's a tendency,
16:49
as you know, in Washington to think 25 is really young, but in the tech field,
16:54
25 is not young.
16:55
>> Elizabeth Economy: Fair enough, fair enough, fair enough [LAUGH].
16:57
It's true, it's true.
17:00
All right.
17:00
So you see the sort of a different model that they've developed, homegrown talent,
17:06
obviously, one of the takeaways, which you've already mentioned for
17:10
the US is that we're gonna have to work harder and
17:14
run faster to continue to attract more talent from outside the United States.
17:19
Same time, we should probably be doing a better job of growing our own talent here.
17:24
Talk to me about sort of what you would see as the lessons that the United States
17:28
should take away from how DeepSeek has approached this, if there are any.
17:32
Maybe it's just too different and we can't really adapt our model to their model but
17:36
maybe there are some takeaways.
17:39
>> Amy Zegart: I actually think the biggest takeaway is we need
17:41
to remember our own model.
17:43
So what China is doing is replicating the best of the US model, and
17:47
we are headed in exactly the opposite direction.
17:51
What is China doing right now?
17:53
You can talk about this better than I can.
17:56
It is an S and T, Science and Technology strategy that says we wanna build
18:01
in fundamental research, we wanna empower our research universities,
18:06
we wanna invest in big, hairy basic research questions,
18:10
not just things that have a commercial application today.
18:15
And we're gonna play the long game.
18:16
That is exactly the US Model that led us to be the innovation
18:21
superpower of the world since World War II.
18:24
But what are we doing?
18:26
We're reducing funding to universities, we're reducing support for
18:30
that fundamental research that asks big, hairy questions, and
18:35
we're putting more money into applications.
18:38
So it's wonderful that venture capitalists are investing in all sorts of companies,
18:43
but those are designed to create a return on investment today or soon.
18:47
Whereas basic research funded by the Federal government is patient
18:52
capital designed across generations to develop insights that can
18:56
then lead to breakthroughs.
18:58
The example I always use is Google.
19:01
We all use Google, but very few people realize that Google really emerged from
19:06
Federal funding of university research in something called digital libraries.
19:13
The Chinese model is actually the American model, and
19:16
we've lost sight of the American model.
19:18
Just one final thing, which is if we look at R&D as a percentage of GDP,
19:22
the Federal government spending on research and development,
19:26
it's a third of what it was at its peak in the 60s.
19:29
And China is spending at a rate that is six times faster than we are.
19:33
So they're gonna eclipse us within a few years.
19:37
>> Elizabeth Economy: Yeah, I think they're already at 2.7% of GDP and
19:41
they're increasing by 0.7% per year or something.
19:44
And we're at, I don't know, 3.5% maybe still,
19:48
I don't know depending on what the current administration is doing.
19:52
So what do you think explains,
19:55
this seems kind of obvious that we have this model that has led us to this point.
20:01
What do you think explains the sort of decision by the current administration to.
20:06
To change course at this moment when, as you've outlined,
20:10
we're kind of in the race of our life?
20:14
>> Amy Zegart: I don't know.
20:15
I'm not inside the brains of the folks in the administration.
20:19
I think some of it, frankly, is the own goals made by universities.
20:23
This is a tough moment and
20:25
a moment of reckoning for higher education on a lot of fronts.
20:28
We've made ourselves opportune targets.
20:32
The complaints about universities having a monoculture are true.
20:37
The concern about not being able to have alternative points of view,
20:41
those are true.
20:42
And so universities have not done what they should do to create
20:46
the environment and the mission that we say we want to do.
20:49
That's, I think, point one.
20:52
I think point two is we have done a poor job of explaining the innovation
20:57
model that the United States has that's made us so great.
21:02
I am amazed at how little understanding there is of the role of
21:07
basic research in Silicon Valley.
21:10
I have had major tech executives say to me,
21:13
what does it matter if we fund fundamental research in universities?
21:18
So I think there's an excitement about venture capital, and I love venture
21:23
capital as much as the next person, but all investment is not created equal.
21:28
And all research is not created equal.
21:31
And so I think universities need to do a much better job of
21:35
articulating what it is that we bring to the table.
21:39
One of the big new initiatives we're doing in the tech policy accelerator at Hoover
21:43
is what we're calling Lab to Launch, which is exactly that.
21:46
It's gathering the data and telling the story so
21:49
that taxpayers understand the return on investment of this fundamental research,
21:54
because we're not just drinking lattes and sitting around in the summer.
21:59
But amazingly, my own relatives think that's what we do.
22:02
And so we need to do a much better job of actually gathering the fact base and
22:06
presenting it in a compelling way of what it is that we do in universities and
22:10
the relationship to the private sector, how those things have to go hand in hand.
22:15
>> Elizabeth Economy: Yeah, I think I remember reading at one point that of
22:19
the top 100 sort of big inventions that have come out of the United States,
22:23
technological inventions like the Internet, for example,
22:27
at least half of them were funded initially, at least through government,
22:32
were funded with government support.
22:36
So it shouldn't be that difficult to look back through history.
22:38
But I think it's great that you're trying to have more contemporaneous
22:43
examples of how university basic research has led to or continues to lead
22:48
to sort of breakthrough inventions that have important applications for
22:53
bettering our society, bettering our economy.
22:56
Do you have one in mind that you could share something that you've just been
23:00
working on in this Lab to Launch?
23:03
>> Amy Zegart: So there are a few, I think,
23:04
examples that most people don't realize.
23:06
So I talked about Google.
23:08
The one that I spoke about when we went to Washington, which is a personal one for
23:12
me, is the MRI machine, right?
23:14
So most of us have had an MRI fortunately or unfortunately, at some time or another.
23:19
The MRI saved my dad's life, right?
23:22
Gave my father more than 20 years of longer life,
23:25
the ultimate gift because it detected the kidney cancer that was killing him.
23:31
He got that MRI in 2001.
23:33
The machine was really commercialized in the 1970s.
23:37
The technology that enabled that machine to be commercialized was
23:41
started in the 1940s when my father was born.
23:44
So, that's a very sort of personal example of one person's life extended by
23:50
this harebrained fundamental research that no one could imagine it was going to then
23:56
lead to this breakthrough machine that we all sort of take for granted today.
24:01
So that's one cryptography that protects us to the extent that our
24:06
data is protected on the Internet.
24:09
Stemmed from decades of research and
24:12
pure math with no sense that it might be applicable today.
24:16
The COVID mRNA vaccine.
24:18
Decades of research in universities before the baton was passed to the pharmaceutical
24:22
industry.
24:24
I do not want people listening to come away saying that I don't think the private
24:29
sector has an incredibly important role to play.
24:32
I do but I think the two go hand in hand.
24:35
And I just say one other thing, Liz,
24:37
which is that emerging technologies often don't emerge, right?
24:41
And so part of the bargain with universities is that a lot of the stuff
24:45
we're working on doesn't pan out.
24:49
That's all part of the innovation ecosystem.
24:52
Most emerging technologies don't emerge.
24:54
So we have to be patient about looking at paths that
24:58
don't actually become pathways to success.
25:03
That's how this business works.
25:05
>> Elizabeth Economy: I think that's really smart.
25:07
So when you went to Washington, can I just ask,
25:10
how was it received by the policymakers?
25:13
Do you think that you sort of informed and elevated the thinking?
25:18
I don't know whether in Congress or in the administration.
25:22
>> Amy Zegart: So I'd like to say the answer is yes,
25:24
but there's a little bit of selection bias, right?
25:27
The folks that most wanna meet with us in Washington are those who think this is
25:31
important.
25:32
So I will say we went with colleagues of mine in science and engineering,
25:37
some of whom hadn't done this before, and they were very pleasantly surprised
25:42
at the bipartisan support that we got when we went around Congress and
25:47
when we went in the executive branch, the nonpartisan support.
25:52
So I think we met with Senator Rounds and Senator Booker.
25:58
Senator Rounds, the Chair of the AI Caucus, Cory Booker, Stanford alum.
26:02
And they were singing from the same songbook.
26:05
This is a Paul Revere moment, we need to come together on this.
26:09
So I think there's a lot more bipartisan support for these technological issues,
26:14
particularly given the competition with China today, than most people might think.
26:21
>> Elizabeth Economy: When I was thinking about the field of AI in particular,
26:26
I just was realizing that just last year, when you look at
26:30
the most recent Nobel Prizes that were awarded in both physics and
26:36
chemistry, they were both tied to advances in AI, and
26:40
three were awarded to US Scientists then, one from the UK and one from Canada.
26:49
So maybe our system still has some sort of positive elements to it.
26:56
Although, I will say that at least one of the scientists was 91 years old.
27:00
And so clearly, as a product of a more traditional period of development.
27:05
But do you see that there's reason for hope here?
27:10
>> Amy Zegart: Absolutely.
27:11
So there's magic in Silicon Valley and in the United States and
27:16
how many foreign leaders come to this area and say,
27:20
I wanna create a Silicon Valley in my country and it can't happen?
27:26
And I think there are key elements to that magic.
27:29
One is we wanna be a place where the world's best and
27:32
brightest wanna continue to come.
27:34
I hope that that mood in Washington changes.
27:37
But I think people want to live in freedom, I think fundamentally, and
27:42
I think they want to be in a place that rewards hard work.
27:45
And historically, the United States has been then, and
27:47
I don't think it's too far gone for us to have that back.
27:50
I also think we have in research universities this thirst for
27:57
exploration and that is its own kind of secret sauce that we have here.
28:05
Not for the country necessarily, we do it because we do it, right?
28:10
It's not top down driven.
28:11
It's not because the Chinese Communist Party is telling us to do it.
28:15
We let people do their thing and we're more likely to have creative,
28:19
innovative pathways because we do that.
28:22
And then of course, you and I are around young people in the university.
28:25
Whenever I look at the future and I talk to my students, I can't help but
28:29
be inspired.
28:31
And especially when we think about tech and how much of our technological
28:35
breakthroughs are happening by younger and younger people,
28:38
it gives me great hope actually for the future, that it's not too far gone.
28:43
We just have to get out of our own way.
28:45
>> Elizabeth Economy: Yeah.
28:45
So I mean, the administration potentially to its credit has put out
28:50
a number of initiative, initiatives that seemed as though they seem
28:55
as though they're designed to advance AI education and diffusion.
29:00
They have this AI education task force, for example.
29:04
I mean, if you look at what this administration is doing in the space,
29:10
do you see things that they're doing that suggest that
29:15
they're committed in a serious way to advancing sort of
29:20
frontier research in the US and diffusion of AI?
29:24
I mean, is there a different kind,
29:26
maybe different from what the Biden administration was doing?
29:30
But do you sense that there is a real drive within the Trump
29:34
administration in this area?
29:36
>> Amy Zegart: I do and I think it's important to look beyond just the Trump
29:40
administration.
29:41
We talk about the decline in investment and fundamental research.
29:45
We look at the Federal government not supporting universities.
29:49
This is a long time coming.
29:51
This spans across administrations, Democratic and Republican, for many,
29:55
many years.
29:56
So our innovation ecosystem has been eroding for a very long time.
30:00
That erosion is accelerating under the current administration but
30:03
it certainly didn't start with the current administration.
30:07
And then you look at sort of the narrative and some of the initiatives coming out of
30:11
the Trump administration, there's a lot to be excited about, right?
30:14
The framing of AI opportunity, not just AI safety, I think is right.
30:19
We need to look at the opportunity, not just the risks.
30:22
And the Biden administration was understandably,
30:24
really concerned about the risks and started sort of a risk first approach.
30:28
But things have changed and the DeepSeek moment is one of them.
30:32
And we need to lean more forward into the opportunity side.
30:35
And that started with the Vice President's speech about AI opportunity.
30:40
You mentioned the AI Executive Order and Education.
30:45
Education is one of the most promising areas to me for
30:48
AI, because AI is the ultimate patient tutor, right?
30:52
Where a human gets frustrated when a student asks for the fifth time.
30:56
I don't get it, how do you do that math problem?
30:58
AI will not get mad at you.
31:00
>> Elizabeth Economy: [LAUGH] It'll never get tired of explaining it, that's true.
31:02
>> Amy Zegart: And we'll come up with a different way for as long as it takes.
31:05
So I think there's tremendous opportunity to enhance what is an abysmal
31:10
education record in our country and K12, if we are smart about adopting AI.
31:16
So I'm encouraged by many things that the administration is doing.
31:20
I'm discouraged by some things that the administration is doing, but
31:24
I think that's often the case that you have.
31:26
It's a complicated terrain and
31:28
people have legitimately different views about how to navigate it.
31:32
>> Elizabeth Economy: Yeah. I mean,
31:33
I think as long as there's some funding, right?
31:36
Some actual funding behind these initiatives, I think that gives me hope.
31:40
I think if we're gonna rely on the private
31:42
sector to pick up all of what had been some significant government funding for
31:47
basic research, I think that's gonna be problematic.
31:51
But let me give you a chance then to offer two or
31:54
three suggestions to the administration,
31:57
if you had the opportunity to design the sort of AI strategy for
32:01
the United States, perhaps incorporating some of the talent element of this,
32:07
what would be your sort of top two or three recommendations?
32:12
>> Amy Zegart: This is my queen of the world if I could-
32:15
>> Elizabeth Economy: Totally,
32:15
I do think you are queen of the world.
32:17
>> Amy Zegart: [LAUGH] I think you're queen of the world.
32:19
>> Elizabeth Economy: We're on the same page.
32:20
>> Amy Zegart: We can be more together.
32:22
>> Elizabeth Economy: Go meet the queen, Queen Amy.
32:23
Go to it.
32:25
>> Amy Zegart: So starting on the talent side,
32:27
I think K12 education is a national security crisis.
32:31
Now, lots of people have talked about K12 for a long time.
32:34
I think that crisis is urgent.
32:36
We rank 34th in the world in Math according to the latest test, and
32:41
we're going down.
32:43
And if you look at the top performers in Math, we have 50% as a percentage
32:48
of our population of top performers in Math compared to Canada.
32:52
We have one-third that the percentage of the South Koreans.
32:57
If we want to educate the workforce of tomorrow, it's got to be done today.
33:02
So, number one, focus on K12 education.
33:05
And I often joke to folks who are in the education space,
33:08
the only thing worse than teaching to the test is not teaching to the test [LAUGH].
33:13
So we need some standards, we need some metrics, and
33:16
we need to measure performance against them.
33:18
So that's number one.
33:20
Number two, capacity building.
33:23
So the two most important enablers for AI and
33:27
other areas of scientific innovation are talent and compute power.
33:33
And right now, compute is dominated by a handful of companies.
33:38
So the example I always give is, last year, Princeton University had to dip into
33:42
its endowment to buy 300 advanced Nvidia chips.
33:46
Meta, the same year, bought 350,000 of the same chips.
33:51
So I'm not saying Princeton has to compete with Meta, but
33:55
we need to make national compute a critical infrastructure for
34:00
more organizations and universities to be able to do frontier science.
34:06
>> Elizabeth Economy: Go ahead, sorry.
34:07
>> Amy Zegart: There's a bill in Congress, a bipartisan bill, to do just that, but
34:12
it has stalled.
34:13
It was proposed in the Trump administration,
34:15
it was supported in the Biden administration.
34:17
It has still not become law.
34:19
Compute is like the highway system of the 1950s.
34:22
It has economic and national security implications that are enormous.
34:27
So I would double down on making compute available for
34:30
more organizations in the United States.
34:33
And the last piece, I would do allies and partners, right?
34:38
We have this amazing set of allies and partners.
34:42
We can do much more together with our talent, with our compute,
34:46
with our data, than we can do alone.
34:49
And so I would fast track more programs to harness the data,
34:54
talent and compute capabilities of our allies and partners.
34:59
>> Elizabeth Economy: So that sounds like a recipe for success but
35:01
I can't resist just asking you on the compute side of things.
35:06
Number one, does DeepSeek tell us that you can can do more with less advanced chips?
35:12
Does that open up some opportunities?
35:14
And number two, this does sound like an area where a company like Nvidia could
35:19
do some public-private partnership with universities, right, to make
35:24
some of those advanced chips available to universities at a reduced cost, perhaps.
35:30
I don't know, what do you think about those two thoughts?
35:32
>> Amy Zegart: I would say yes and yes.
35:34
So one of the interesting things about DeepSeek,
35:36
when you talk to sort of PhD students in computer science, they were so excited.
35:41
Why were they so excited?
35:43
Exactly the point you raised, so the minimum viable compute to be able to do
35:48
really cool research in this area was less than people thought it was.
35:53
It reduces that gap between the hyperscalers and
35:56
what others on the frontier can do but that gap ain't nothing.
36:00
It's still a pretty big gap.
36:02
So that's exciting that the gap has narrowed in terms of being able to
36:07
compete on a wider scale.
36:09
Public private partnerships are gonna be really important for these companies and
36:14
it's not just Nvidia.
36:15
I think companies in general in this moment need to be thinking
36:20
much differently about not.
36:22
They shouldn't be just reacting to the moment.
36:24
They are shaping the international order.
36:27
They are shaping the geopolitical battleground.
36:30
And they need to think about being proactive and strategic, not just for
36:34
their shareholders, but for the nation.
36:36
And so I think what you suggested is one component of that but
36:40
it's not just Nvidia.
36:41
I think many companies need to think about how can we contribute in
36:46
a way to build scale in our nation to compete against China.
36:52
>> Elizabeth Economy: Yeah, I will say,
36:53
when I was at the Commerce Department in the Biden administration,
36:58
Secretary Raimondo was very proactive about reaching out to companies
37:04
to get them to work with the administration on various initiatives,
37:09
like digital skilling, for example, globally, right?
37:15
So to partner in ways that helped to reinforce the positive message about
37:20
the United States on the global stage and to help realize US national objectives
37:25
on things like supply chains, for example, on critical minerals.
37:31
So I think there were many ways in which private companies, in fact, did step up.
37:35
But I think what you're suggesting is that they should be developing their own
37:40
visions and be more proactive about it, not necessarily wait for the US Government
37:45
to come to them and say, please partner with us on these initiatives
37:49
that help realize our sort of broader national economic and security objectives.
37:54
>> Amy Zegart: I mean, I think Secretary Raimondo, your former boss,
37:58
did a fantastic job at bringing the private sector together.
38:03
But we also often leave out the third P, well, actually,
38:07
it's not even a P, public private academic partnership.
38:13
So it's not just industry and government, it's industry, government, and academia.
38:18
We can't leave out the university piece of this.
38:21
And I think that's where we really need to focus more,
38:24
is all three of those legs of the stool need to be supported and come together.
38:29
>> Elizabeth Economy: Okay, Amy, I can't thank you enough, not only for your sort
38:33
of excellent analytical presentation and helping us understand the DeepSeek moment.
38:39
But also just for your optimism and your ideas for
38:42
what we need to do to sort of move this forward and move together,
38:47
I think, really, as a nation forward to advance ourselves in this
38:52
really important strategic arena of Artificial Intelligence.
38:56
So thank you for all the work that you're doing.
39:01
>> Amy Zegart: Thank you for having me.
39:02
I'm just trying to keep up with you and all the great work you're doing.
39:05
>> Elizabeth Economy: [LAUGH] All right, if you enjoyed this podcast and
39:08
want to hear more recent discourse and debate on China, I encourage you to
39:12
subscribe to China Considered via The Hoover Institution YouTube channel or
39:16
podcast platform of your choice.
39:18
Next on China Considered,
39:19
I interviewed George Washington University Professor David Shambaugh on
39:23
his fascinating new book about how China lost the United States.
39:26
[MUSIC]