00:00:00
[Music]
00:00:02
I honestly believe that the AI
00:00:04
revolution is underhyped.
00:00:07
Now, why is this all important?
00:00:09
Eric Schmidt is here. He's the former
00:00:10
Google executive chairman and CEO.
00:00:12
These agents are going to be really
00:00:14
powerful, and they'll start to work
00:00:15
together. We're soon going to be able to
00:00:17
have computers running on their own,
00:00:19
deciding what they want to do. Now we
00:00:21
have the arrival of a new nonhuman
00:00:23
intelligence which is likely to have
00:00:26
better reasoning skills than humans can
00:00:29
have.
00:00:29
So if you were emperor of the world for
00:00:32
1 hour, the most important thing I do is
00:00:34
make sure that the west wins.
00:00:36
Ladies and gentlemen, please welcome
00:00:40
Eric Schmidt.
00:00:42
[Music]
00:00:44
Hi.
00:00:46
[Music]
00:00:47
Hi. Looking good. Good to see you.
00:00:51
Good to see you.
00:00:52
Good to see you.
00:00:53
You're looking, Eric. You're looking.
00:00:55
Very nice.
00:00:57
Oh my god. David, good to see you. It's
00:00:59
like
00:00:59
David Sax is here as well.
00:01:00
It's like a It's like a reunion of all
00:01:02
of our former companies. David, why did
00:01:04
you quit after all?
00:01:06
My old
00:01:07
my old boss.
00:01:09
It was
00:01:10
What was it like working with Young
00:01:12
Freedberg? Take us back.
00:01:14
Can I Can I tell a story? We have to
00:01:16
come down to Orange County and they're
00:01:17
like, "Hey, we're going to take the
00:01:19
plane and it was Eric's plane." We get
00:01:21
on the plane and then he goes up and
00:01:23
flies the plane. I'm in the back of the
00:01:24
plane by myself.
00:01:24
Was it King Air?
00:01:25
I'm like, "The CEO of Google's flying me
00:01:27
down to Orange County." It was
00:01:28
incredible. That was my first time
00:01:30
actually hanging out with Eric.
00:01:31
It was It was my Gulfream.
00:01:33
That's right.
00:01:34
Um, he was way too smart.
00:01:37
Way too smart.
00:01:37
Way too smart.
00:01:38
Was he focused? Did he contribute? Did
00:01:40
he move the needle?
00:01:41
But he was very smart. Okay, that's kind
00:01:44
of our consensus off the pot as well.
00:01:47
Look, you guys know this guy well. He's
00:01:49
really that smart. So, he taught me more
00:01:52
stuff than most of any of the employees
00:01:54
at Google and then you left.
00:01:56
Well, tell us what you've been doing.
00:01:57
So, um
00:01:58
No, no, wait. Before that, I got to ask
00:02:00
you this question. There was a recently
00:02:02
deleted video from Stanford.
00:02:04
Oh, no.
00:02:04
You had a moment of clarity where you
00:02:07
said, "Hey, you know, like at Google,
00:02:09
people are like too much work life
00:02:11
balance. They need to commit. They need
00:02:13
to work harder. We had Sergey at the
00:02:15
last event. He's going back to work. So
00:02:17
Sergey got the message.
00:02:22
Predicting Sergey's behavior is
00:02:24
something I can fail at. I tried for 20
00:02:26
years. Um I am not in favor of um
00:02:31
essentially working at home. I And the
00:02:34
reason I mean many of you guys all work
00:02:36
at home to some degree, but your careers
00:02:38
are already established. But think about
00:02:40
a 20-some who has to learn how the world
00:02:43
works and you know they they come out of
00:02:45
Berkeley or Dartmouth and they're very
00:02:47
well educated. When I think about how
00:02:49
much I learned when I was at Sun just
00:02:51
listening to these elder people who were
00:02:54
5 or 10 years older than I was argue
00:02:56
with each other in person. I don't how
00:03:00
do you recreate that in this new thing?
00:03:02
And and I'm in favor of work life
00:03:04
balance and that's why people work for
00:03:06
the government. Um, sorry. Um,
00:03:10
strays.
00:03:11
Sorry, sorry, sorry. Um, if you're going
00:03:14
to be in tech and you're going to win,
00:03:16
you're going to have to make some
00:03:17
tradeoffs. And you're remember, we're up
00:03:20
against the Chinese. The Chinese work
00:03:23
life balance consists of 996, which is
00:03:25
9:00 a.m. to 9:00 p.m. 6 days a week. By
00:03:28
the way, the Chinese have clarified that
00:03:30
this is illegal. However, they all do
00:03:33
it.
00:03:34
That's who you're competing against. I
00:03:35
brought I brought everybody back to
00:03:37
office. It's so much better.
00:03:38
So, it let's just pick up on that theme.
00:03:40
So,
00:03:41
you don't need to defend the government.
00:03:42
No, no. Believe me, I don't I don't see
00:03:44
the need to. I'm an unpaid part-time
00:03:46
adviser to the government. So, uh but uh
00:03:48
but we are in this high-tech competition
00:03:51
with China. They obviously care about
00:03:53
AI, too. They're trying to race ahead.
00:03:56
How do you uh I I understand that you
00:03:59
recently ma made a trip there. How do
00:04:00
you um handicap this this this
00:04:02
competition? Well, you and I just talked
00:04:04
about this as part of your as your
00:04:06
incredibly important work in the White
00:04:08
House. Um, I had thought that China and
00:04:11
the United States were competing at the
00:04:13
peer level in AI and that the good work
00:04:17
that you have done and your predecessors
00:04:19
did to restrict chips were slowing them
00:04:21
down. They're really doing something
00:04:23
more different than I thought. They're
00:04:25
not pursuing crazy AGI strategies partly
00:04:29
because the hardware limitations that
00:04:31
you've put in place, but partly because
00:04:33
the depth of their capital markets don't
00:04:35
exist. They can't raise based on a wing
00:04:38
and a prayer $100 million or maybe an
00:04:41
equivalent to to build the data centers.
00:04:42
They just can't do it. And so the result
00:04:45
is they're very focused on taking AI and
00:04:48
applying it to everything. And so the
00:04:51
concern I have is that while we're
00:04:53
pursuing AGI, which is incredibly
00:04:54
interesting and we should talk about and
00:04:56
all of us will be affected by this, we
00:04:58
better also be competing with the
00:05:00
Chinese in day-to-day stuff. Consumer
00:05:03
apps, this is something you understand
00:05:04
very well, Chimath, uh, consumer apps,
00:05:06
uh, robots and so forth and so on. I saw
00:05:08
all the the Shanghai robotics companies
00:05:11
and these guys are attempting to do in
00:05:13
robots what they've successfully done
00:05:15
with electric vehicles, right? and
00:05:18
they're re they their work ethic is
00:05:20
incredible. They're wellunded. It's not
00:05:22
the crazy valuations that we have in
00:05:25
America. They can't raise the capital,
00:05:27
but they can win across that. The other
00:05:29
thing the Chinese are doing, and I want
00:05:31
to emphasize this is a major
00:05:32
geopolitical issue, is that my own
00:05:35
background is open source. In the
00:05:36
audience, you all know open source means
00:05:38
open code. Open weight open weights
00:05:41
means open training data.
00:05:43
China is competing with open weights and
00:05:45
open training data. And the US is
00:05:47
largely and majority focused on closed
00:05:49
weights, closed data. That means that
00:05:52
the majority of the world, think of it
00:05:54
as the belt and road initiative, are
00:05:56
going to use Chinese models and not
00:05:57
American models. Now, I happen to think
00:06:00
the West and democracies are correct.
00:06:02
And I'd much rather have the
00:06:04
proliferation of large language models
00:06:06
and that learning be done based on
00:06:07
Western values. Eric, we had a a major
00:06:09
open- source initiative um with Meta,
00:06:13
you know, incredible balance sheet,
00:06:15
tremendous technical firepower, but they
00:06:17
seem to have misexecuted and now are
00:06:19
taking a step back and reformulating
00:06:21
something to your point that looks a
00:06:23
little bit more closed source. It's not
00:06:25
clear. You know, Alex Wang's a good
00:06:27
friend. Uh he's come in, he's taken
00:06:29
over. He's obviously incredibly uh
00:06:30
incredibly capable. I would not hold
00:06:33
keep um I I would not say that they're
00:06:35
going fully closed. And I think also
00:06:37
they got screwed up because the deepseek
00:06:40
people uh R1 did such a good job, right?
00:06:44
If you look at the reasoning model in
00:06:46
deepseek and in particular their ability
00:06:48
to do reinforcement learning forward and
00:06:50
back, forward and back and forward and
00:06:51
back. This is a major achievement and it
00:06:54
appears that they're doing it with less
00:06:56
precision than numeric precision than
00:06:59
the American models. As a bit of
00:07:01
technical things, uh there's something
00:07:03
called FP64, FP32, FP16. The American
00:07:07
models are typically using 16 bit
00:07:09
precision for their training. The
00:07:11
Chinese are pushing eight and now even
00:07:13
four.
00:07:14
Is there is there something that um the
00:07:17
American you know bigger companies need
00:07:19
to be doing in open source so that we
00:07:21
can actually combat this?
00:07:22
Well, a number of the large companies
00:07:23
have said that they want to be leaders
00:07:25
in open source as well. Um Sam Alman
00:07:27
indicated that the smallest version of
00:07:29
the 03 model would be released I believe
00:07:32
open weights and they have done so and
00:07:35
he told me anyway that this model is
00:07:38
much smaller than 10 the 26. It's much
00:07:41
easier to train and it will fit or or
00:07:43
can fit on your phone. So one path is to
00:07:47
say that we'll have these supercomputers
00:07:50
doing AGI which will always be
00:07:52
incredibly expensive and so forth. But
00:07:54
we also have to watch to make sure that
00:07:55
the proliferation of these models for
00:07:57
handheld devices is under American
00:08:00
control whether it's OpenAI or Meta or
00:08:03
Gemini or what have you. recently you uh
00:08:06
took over uh relativity space and I
00:08:09
think for for the people that don't know
00:08:10
this is a business that effectively
00:08:13
whose ambition is to compete with
00:08:14
SpaceX.
00:08:15
I think you were the first investor or
00:08:17
the earliest investor in it.
00:08:19
I was
00:08:19
I'm sorry.
00:08:20
It's okay.
00:08:21
You lost some money in the first what
00:08:23
happened? Did you get crammed down?
00:08:24
No, no, no. I mean all of us did.
00:08:26
All of I mean look
00:08:27
I mean I've been very happily uh you
00:08:30
know an investor in SpaceX and Swarm and
00:08:32
Starlink. relativity was
00:08:34
and by the way swarm is a big deal
00:08:35
created
00:08:36
swarm was a
00:08:36
so thank you
00:08:37
yeah swarm has been a really great
00:08:39
success for for them and I think for the
00:08:41
world um but what I was going to ask you
00:08:43
is walk us through the evolution of the
00:08:46
space market why you decided of all the
00:08:48
companies you have the capital base to
00:08:50
kind of put your money anywhere why did
00:08:53
you pick that why did you pick that
00:08:54
business why now
00:08:56
rockets are really cool and they're
00:08:58
really hard I had I'm as you know I'm a
00:09:01
pilot and I know lots about jets And I
00:09:03
had assumed that rockets were as mature
00:09:05
as jet engines. They're not. It is an
00:09:08
art and a science. These things are very
00:09:11
hard to do. The amounts of power, I
00:09:13
mean, in our case, the rocket is 4
00:09:15
million pounds of thrust. Um, you have
00:09:18
to hold the thing down to test it. And
00:09:20
you can't even hold it with metal
00:09:21
things. You have to have other things to
00:09:23
hold it down as well. There's so much
00:09:24
force otherwise it will take off. Um,
00:09:26
another interesting thing about rockets
00:09:28
is that a rough number is that 2% of the
00:09:30
weight of the rocket is the payload, 18%
00:09:33
is roughly the rocket, and 80% is the
00:09:36
propellant. And my reaction as a new
00:09:38
person is, you're telling me you can't
00:09:40
do any better. And the physicists say
00:09:42
after 60 years of physics, that's the
00:09:44
best we can do to get out of the
00:09:46
gravitation of of the Earth. And so, I
00:09:49
think rockets are interesting and
00:09:50
they're challenging. There's always an
00:09:52
opportunity for competition. um in
00:09:54
relativity space's area, it's
00:09:56
essentially a LEO com a LEO competitor.
00:09:59
So low Earth orbit satellites, that sort
00:10:01
of thing. The order book is full. We
00:10:03
just have to launch the rocket and and
00:10:05
this entry into space happened and I'm
00:10:08
not sure how well known this is. So you
00:10:10
can go as far as you want to go into
00:10:11
this, but you've done a lot as well in
00:10:13
next generation warfare as well. Do you
00:10:15
want us just to talk about that and how
00:10:18
you ended up there and what role that
00:10:20
plays and just give us a landscape?
00:10:22
Maybe David asked about the China
00:10:23
question, but it's they're all kind of
00:10:24
almost interrelated.
00:10:25
Well, at first place, I'm a software
00:10:27
person, not a hardware person. I I
00:10:28
explain to people that hardware people
00:10:30
go to different schools than software
00:10:31
people. Um, and they think slightly
00:10:34
differently. So, I'm always at a
00:10:35
limitation in these new industries. Um,
00:10:37
I had worked for the Secretary of
00:10:39
Defense and have a top secret clearance
00:10:41
and all that. I was given a medal, etc.,
00:10:43
uh, for trying to help the Pentagon
00:10:45
reorganize itself. And when the Ukraine
00:10:48
war started, I was watching and I
00:10:50
thought, well, here's an opportunity to
00:10:52
see a country that has no navy u and no
00:10:55
uh and no air force how they do this
00:10:58
with automation. And indeed, it has been
00:11:00
a spectacular success as a matter of
00:11:03
innovation. Um and outnumbered 3 to one
00:11:06
with huge differences in um kinetic
00:11:10
strength, weapons, mobilization, and so
00:11:13
forth and so on. Ukraine has held on
00:11:15
really quite well. Um and what's
00:11:17
happening now is you're seeing
00:11:19
essentially the birth of a completely
00:11:20
new military national security
00:11:22
structure. Um one way to think about it
00:11:25
is that we all u so first place and I've
00:11:28
I've seen it live and I will tell you
00:11:29
that real war is much worse than the
00:11:32
worst movies you have ever seen about
00:11:34
war. And that's all I'll say. It's
00:11:36
really horrific and it's to be avoided
00:11:37
at all cost. Um and then right for
00:11:41
obvious reasons and the the
00:11:45
and I love all these people say well
00:11:46
we'll well you know the wararmongering
00:11:48
talk be careful what you wish for
00:11:51
because the other side gets a vote when
00:11:53
I started working and and trying to
00:11:55
understand what Ukraine was doing uh
00:11:57
Russia was pushed back and they've come
00:12:00
back with a very very strong second and
00:12:03
third round so the enemy gets a vote in
00:12:05
this situation uh but to but to go on um
00:12:08
The rough way in which war will evolve
00:12:12
is first things will have to be very
00:12:14
very mobile and very much not in fixed
00:12:17
places. This takes out most of the
00:12:20
military infrastructure that exists in
00:12:22
the world. Um things like tanks um of
00:12:25
which we're now building a whole bunch
00:12:27
more even stronger tanks here in America
00:12:29
don't make any sense in a world where a
00:12:31
2 kg payload from a a well-armed drone
00:12:35
can destroy the tank. It's called the
00:12:37
kill ratio. And that drone costs retail
00:12:40
$5,000, $4,000. The tank, the American
00:12:42
tank costs $30 million. You can see the
00:12:46
you can send an awful lot of those
00:12:47
drones to destroy those tanks. Um, the
00:12:50
likely evolution goes something like
00:12:52
this. So, first, people learn that
00:12:54
drones are like rifles and like
00:12:56
artillery. So, it's more efficient to
00:12:59
use drones now than to use mortars,
00:13:01
grenades, artillery. That's clear. If
00:13:04
you just look at the economics,
00:13:05
economics in terms of cost or
00:13:07
effectiveness as it's called. Um the
00:13:10
next thing that happens is that both
00:13:12
sides develop drone capabilities which
00:13:15
what you're seeing now and each then
00:13:17
becomes a war of drone against drone. So
00:13:20
you have drone against anti- drone. And
00:13:22
so then the shift moves to how do you
00:13:24
detect the enemy drone and how do you
00:13:26
destroy it before it destroys you. So
00:13:28
the doctrine ultimately is the drones
00:13:30
are forward and the people are behind.
00:13:34
And I've seen operations in for example
00:13:36
sitting in Kev where the Ukrainians are
00:13:38
commanding things over Starlink I might
00:13:40
add um in the distance in the distant
00:13:43
war and they're very very effective. So
00:13:45
we've solved the latency problems, we've
00:13:47
solved the timing problems and so forth
00:13:49
in that area. The ultimate state is very
00:13:52
interesting and I don't think anyone has
00:13:54
foreseen this. If you go back to our
00:13:56
conversation about RL and planning,
00:13:59
which is what you're seeing with AI,
00:14:01
let's say that that we're on one side
00:14:03
and we have a million drones and there's
00:14:05
another side over here that has another
00:14:06
million drones. Each side will use
00:14:09
reinforcement learning AI strategies to
00:14:11
do battle plans, but neither side can
00:14:14
figure out what the other side's battle
00:14:16
plan is. And therefore, the deterrence
00:14:19
against attacking each other will be
00:14:20
very high. Today, the way military
00:14:23
planners operate is that they count
00:14:24
weapons. They say, "Well, you have this
00:14:26
many and I have this many and you can do
00:14:28
this kind of a maneuver and so forth."
00:14:30
But in an AI world where you're doing
00:14:31
reinforcement learning, you can't count
00:14:34
what the other side is planning. You
00:14:35
can't see it. You don't know it. And I
00:14:37
believe that that will deter what I view
00:14:39
as one of the most horrendous things
00:14:41
ever done by humans, which is war.
00:14:43
Because unless there's a perfect balance
00:14:45
between either side, there will be some
00:14:47
mutual destruction of the drone supply
00:14:49
like there would be with any artillery
00:14:50
stock in traditional warfare and
00:14:53
whoever's left ends up winning. Like
00:14:55
they're just
00:14:56
Well, it's very important to understand
00:14:57
that there's no winners in war. Um, by
00:15:00
the time you have a drone battle of the
00:15:02
scale I'm describing, the entire
00:15:04
infrastructure of your side will be
00:15:06
destroyed. The entire infrastructure of
00:15:07
the other side will be destroyed. These
00:15:09
are lose-lose scenarios. Isn't there
00:15:11
like an an equilibrium though that that
00:15:13
can also create where there because of
00:15:16
that mutually assured destruction
00:15:18
there's a det or is that
00:15:19
well I'm arguing that it's it's not a
00:15:21
deterrence
00:15:23
right
00:15:23
that as deterrence can be understood as
00:15:25
I want to hit you which I don't but I
00:15:28
want to hit you so much but that if I do
00:15:30
that the penalty is the penalty is
00:15:32
greater than the value of me hitting you
00:15:34
right
00:15:35
and that's how det that's how
00:15:36
but that seems like an that seems like a
00:15:38
great um advantage antage and upside of
00:15:42
this move to sort of drones and
00:15:44
automation that we don't have today.
00:15:46
Well, I there are many advantages to
00:15:48
moving to drones and automation. One,
00:15:49
they're much much cheaper, right?
00:15:51
They're much much cheaper.
00:15:53
Yeah.
00:15:53
And two, and two, you can stockpile
00:15:56
algorithms. You can essentially learn
00:15:58
and learn and learn. And remember, you
00:16:00
can also build training data, right,
00:16:02
that's synthetic, so you can be even
00:16:04
better than the others. The final
00:16:05
question I've been asked by our military
00:16:08
is what's the role of the of a
00:16:11
traditional land army? And I wish I
00:16:14
could say that all of these human
00:16:17
behaviors can occur without humans being
00:16:19
at at risk. I don't think so. I think
00:16:22
that the way um robot war essentially
00:16:26
drone war will occur is there will be
00:16:28
these destructive waves, but eventually
00:16:30
humans are going to have to cross a
00:16:32
line. they're going to have to
00:16:33
after we've depleted them. So, you're
00:16:35
investing in this drone technology and
00:16:37
then do you think
00:16:39
Optimus and humanoid robots are the
00:16:41
next, you know, volley in this um new
00:16:45
warfare. It's going to be a long time
00:16:48
before humanoid robots we which is what
00:16:50
we see in the movies all day, right? Be
00:16:52
a very long time before we see that. Uh
00:16:55
what you're going to see is very very
00:16:58
fast mobility solutions, right? airly
00:17:00
airbased solutions and also hypersonics
00:17:02
hypersonics also things underwater
00:17:05
there's a lot of that going on it's a
00:17:06
different domain um if you look at the
00:17:09
um the muro and some other boats that
00:17:12
the Ukrainians used they have
00:17:14
essentially used USVs to destroy the uh
00:17:17
Russian fleet in the Black Sea this was
00:17:20
crucial for them because they needed to
00:17:22
be able to export the grain from Odessa
00:17:25
around and it's like 6% or 10% of their
00:17:28
economy It's a very big deal and they
00:17:30
did that with drones.
00:17:31
Eric, it seems like there's this
00:17:33
overarching worldview that you have,
00:17:36
meaning you have this view on AI.
00:17:40
There's all the stuff you're doing now
00:17:41
in drones, in warfare, in rocketry. It
00:17:44
all converges quite honestly because in
00:17:46
the in the next five or 10 years, these
00:17:48
things will all come to pass. What is
00:17:50
the like how do you view the world? Like
00:17:53
what is the role of America? What is
00:17:55
your role as a as a capitalist, as a
00:17:57
technologist, as like a statesman?
00:18:00
I want America to win,
00:18:04
right?
00:18:07
Uh I am here because the American dream,
00:18:11
the people who invested in me, in my
00:18:13
case, Berkeley and so forth, people took
00:18:15
a chance on me. I want the next
00:18:17
generation to have that. I also want you
00:18:20
all to remember I was just in in uh as
00:18:23
part of the World War II surrender
00:18:24
ceremony in in Honolulu and they talked
00:18:27
about fighting tyranny, right? We forget
00:18:30
that our ancestors or greatgrandparents
00:18:32
or whatever fought the Great War to keep
00:18:35
liberalism and democracy alive. I want
00:18:38
us to do that. How do we do that as
00:18:40
Americans? We use our strengths. What
00:18:43
are our strengths? We're chaotic,
00:18:46
confusing, loud, you know, but we're
00:18:48
clever. Uh we allocate capital smartly.
00:18:51
We have very deep financial markets. We
00:18:54
have this enormous industrial base of
00:18:56
universities and entrepreneurs which are
00:18:58
represented here. We should celebrate
00:19:00
this. We should stoke it. We should make
00:19:01
it go faster and faster. I spent lots of
00:19:04
time in Europe because of the Ukraine
00:19:06
stuff. They are so envious of us. When
00:19:08
you're in Asia, they are envious of us.
00:19:11
Don't screw it up, guys. That's what I
00:19:13
want to work on.
00:19:14
Can Can I Can I just ask you outside of
00:19:19
this external conflict? We had um a
00:19:22
conversation with Alex Karp today and we
00:19:25
actually had Tucker Carlson here
00:19:26
yesterday and some of the dialogue was
00:19:29
around the I I don't know if the right
00:19:31
term is the erosion of the west that
00:19:33
there may be social issues that are
00:19:34
brewing in the west that may be hurting
00:19:37
us from the inside. How much do you
00:19:38
observe or spend time on these issues?
00:19:41
And the metric that often is cited now
00:19:43
is declining birth rates in the west.
00:19:45
And that our population, and we're gonna
00:19:47
talk with Elon in a few minutes about
00:19:48
this. Um, oh, sorry.
00:19:51
Oh, we just ruined my bad. My surprise.
00:19:53
Oops.
00:19:53
Sorry. Sorry.
00:19:54
There's your surprise guest.
00:19:55
Sorry. Sorry. Sorry. Sorry.
00:19:57
Um, slip. Uh, but um,
00:20:02
Elon is a good friend and he's
00:20:04
addressing this issue of population
00:20:06
directly himself.
00:20:11
problem solve for it. Good for you.
00:20:13
Is it is it is it a reflection of
00:20:15
something going on? There's a rise of
00:20:16
Mandani getting elected in New York. Uh
00:20:18
some of the historic values of the West
00:20:22
seem to be, you know, kind of under a
00:20:25
state of transformation. Right now,
00:20:27
one metric of the success of a society
00:20:29
is its ability to reproduce. And so, I
00:20:32
think this is a legitimate concern of
00:20:33
the West. It's much worse in Asia. The
00:20:36
um the Chinese number is about 1.0 0 for
00:20:39
two parents. In Korea, it's now down to
00:20:42
78 for two. So, it's really important to
00:20:46
recognize that we as humans are
00:20:48
collectively choosing to depopulate. And
00:20:51
the numbers are staggering, right? And
00:20:54
imagine a situation where instead of
00:20:56
having growth, you have shrinkage. And
00:20:58
furthermore, they're getting older. And
00:21:00
so, as a business, all of a sudden, your
00:21:03
revenue is declining. And there's
00:21:05
nothing you can do because you can't
00:21:06
innovate with fewer and fewer customers.
00:21:08
So if you just put it in a business
00:21:10
context, ignoring the moral issues which
00:21:12
are all very real, it's just bad, right?
00:21:14
So we have to solve that problem. I
00:21:16
happen to be in favor broadly of
00:21:17
immigration because I think immigration
00:21:19
helps us solve that problem. But as a
00:21:21
global mechanism, we have to address
00:21:23
that. Um, in any case, from my
00:21:26
perspective, you're going to have these
00:21:28
issues, but America is organized around
00:21:32
the concept of American exceptionalism.
00:21:35
And as long as we understand that the
00:21:37
way we make progress is we invest in the
00:21:40
right people, in the right businesses,
00:21:42
we have a a strong capital market, we
00:21:45
invest in the infrastructure that they
00:21:46
need, um, we'll be fine. That is my
00:21:49
actual opinion. Can can we go back to um
00:21:52
AI for a second?
00:21:54
So um Eric, I think you can help us get
00:21:56
to a let's call it a a bipartisan
00:21:59
understanding of these issues. I think
00:22:00
you you you think really clearly about
00:22:02
this. Um you know, in in the wake of
00:22:06
Chad GBT launching at the end of 2022, I
00:22:10
think the discourse was really dominated
00:22:12
in 2023 and 24 by this idea of AGI and
00:22:15
that AGI was imminent. And I think it
00:22:18
created almost like a panicky atmosphere
00:22:20
in Washington among policy makers and
00:22:23
you saw things like we got to restrict
00:22:25
open source because you know then China
00:22:27
will get it and um and this is before
00:22:29
Deepseek launched and then we saw that
00:22:31
actually they're ahead of us on open
00:22:32
source but it feels like there's been um
00:22:35
a pullback a little bit from the AGI
00:22:37
narrative which I think I think it's
00:22:39
actually a good thing. I think it's more
00:22:40
conducive to calm rational policym.
00:22:43
What's your perception of AGI right now?
00:22:45
Where where are we on that whole train?
00:22:48
So So um so first place, the speech that
00:22:51
the president delivered about a month
00:22:53
ago about AI strategy, which I think you
00:22:56
probably wouldn't say it, but you kind
00:22:57
of wrote it for him, was exactly right.
00:23:00
Right. So thank you.
00:23:02
David collaborated with an amazing
00:23:04
leader who we all respect and admire so
00:23:07
much, Eric.
00:23:10
Yes. Uh, so nevertheless,
00:23:13
saying I wrote it was was way too
00:23:15
strong. I mean, actually, but anyway,
00:23:17
if you didn't if you didn't write it,
00:23:18
then it must have been your twin. But in
00:23:20
any case, um, the you you got you got
00:23:23
the emphasis right, which was that
00:23:25
investment in research, investment in
00:23:28
the kind of stuff that we do is really,
00:23:30
really important.
00:23:32
I don't agree with you on this on this
00:23:34
AGI thing because there's this group
00:23:36
which I call the San Francisco um
00:23:39
narrative because they all live in San
00:23:41
Francisco and their narrative goes
00:23:43
something like this. Um today we're
00:23:46
doing agents uh the agentic revolution
00:23:48
will change businesses which I agree
00:23:49
with. Um that what happens is the
00:23:52
systems will become recursively
00:23:54
self-intelligent
00:23:56
with recursive self-improvement as it's
00:23:59
called. If you have a scale-free problem
00:24:01
and a scale-free problem for example is
00:24:03
programming or math where you can just
00:24:05
keep doing it you get these enormous
00:24:08
fast gains if you buy enough hardware do
00:24:11
enough software so forth and so on that
00:24:13
is still underway.
00:24:15
The collective of that says that in the
00:24:18
next three-ish years they believe that
00:24:21
we will get forms of super intelligence
00:24:24
and the way they define it is basically
00:24:27
a savant a chemist so a physics soant a
00:24:30
mathematician soant I don't agree with
00:24:32
the three years but I do agree that
00:24:34
it'll be maybe six or seven years
00:24:36
but if it's a savant in you know a
00:24:39
particular area is that general
00:24:41
intelligence
00:24:42
it's not general intelligence yet
00:24:44
general internal intelligence is when it
00:24:46
can set its own objective function.
00:24:48
Right?
00:24:49
And there's no evidence of that.
00:24:51
There's no evidence right now of the
00:24:52
ability to set your own objective
00:24:54
function. Um the the the thinking and
00:24:56
I'm writing a paper on this so I've been
00:24:58
studying it is that the the the
00:25:01
technical problem is non-stationerity of
00:25:04
mathematical proofs. And what you're
00:25:07
doing is you're trying to solve against
00:25:09
objective function but the objective
00:25:10
function keeps changing which is how
00:25:12
humans operate. your goal changes every
00:25:14
day. Whereas computers have trouble with
00:25:16
that. As a math problem, we don't have
00:25:18
an algorithm yet for LLMs that can do
00:25:22
that. People are working on it. Um and
00:25:24
the the test will be can you basically
00:25:28
um using the information available in
00:25:30
1902, can you derive the same thing that
00:25:33
Einstein did with special relativity
00:25:36
followed by general relativity? We
00:25:38
cannot do that today. Um and most people
00:25:42
believe that the way this will be solved
00:25:44
is through analogy. So the theory of
00:25:47
great geniuses is that they understand
00:25:50
one area of extremely well and they're
00:25:52
so brilliant the lady or man can then
00:25:55
take their ideas and apply it to a
00:25:57
completely different domain. If we can
00:26:00
solve that problem then I think it's
00:26:02
over. Then we get to AGI and then it's a
00:26:06
whole different world.
00:26:08
I think one of the reasons why it's hard
00:26:10
to replace a human and you know JK and I
00:26:14
debate this is that humans are end to
00:26:16
end. You know we can do the whole job.
00:26:18
You have sort of a complete
00:26:19
understanding. You can pivot very
00:26:21
easily. AI at least as we know it today
00:26:24
is not end to end. It has to be
00:26:25
prompted. You get an answer. That answer
00:26:27
has to be validated. Then you have to
00:26:29
ask a new question because it never
00:26:30
gives you exactly what you want. You
00:26:32
have to apply more context. You have to
00:26:34
go through an iterative loop. Finally,
00:26:35
you get to an answer that has business
00:26:37
value. The way biology puts it is that
00:26:40
AI is not end to end. It's middle to
00:26:42
middle. Humans are end to end. And so,
00:26:43
as a result of that, instead of AI
00:26:47
replacing all of us, AI will be very
00:26:49
synergistic with humans because we can
00:26:52
define the objective function. We do the
00:26:54
prompting and we work with it to iterate
00:26:56
and it does a lot of the work in the
00:26:58
middle. Um, that seems to me like a very
00:27:01
optimistic, less duoristic take on it.
00:27:04
What you just said is exactly what's
00:27:06
going to happen for the next few years
00:27:08
that each of us will have assistance
00:27:11
which on our command and our prompting
00:27:13
will be incredibly helpful to whatever
00:27:16
problem we have you know personal uh you
00:27:18
have people who are using these things
00:27:19
for relationship advice for you know
00:27:21
talking to their kids I mean it's all
00:27:22
crazy stuff um but the fact of the
00:27:25
matter is that's it the to me the real
00:27:27
question is when does it cross over to
00:27:29
having its own valition its own ability
00:27:31
to seek information and solve new
00:27:33
problems s that's a different animal.
00:27:35
But have we seen any evidence of
00:27:36
recursive self-improvement yet?
00:27:38
Um, not yet. I'm I'm I've funded a
00:27:41
number of startups which claim to be
00:27:43
close to it, but of course these are
00:27:45
startups and you never know, which tells
00:27:48
me it's 5 10 years cutting numbers.
00:27:50
What do you think Google's doing on this
00:27:51
front?
00:27:52
Um, well, I'm not at Google anymore. Um,
00:27:55
every issue of Gemini is top of the
00:27:58
leaderboard. So 2.5 just overcame
00:28:02
everybody and I'm sure there's another
00:28:03
one coming. Um Demis is working really
00:28:06
hard on this question about um
00:28:08
scientific discovery. So that's a p that
00:28:12
is a path to getting to AGI.
00:28:15
Eric, um we appreciate the work you're
00:28:17
doing. Uh we appreciate you being here
00:28:19
with us. We appreciate what you've done,
00:28:21
the impact you've had on Silicon Valley,
00:28:23
uh as society. Yeah. No, but it's it's
00:28:26
really been
00:28:26
I am so happy to be part of this. You
00:28:29
created this incredible community and
00:28:31
there's all of these smart people that
00:28:33
spend all their time listening to you.
00:28:37
Very concerning.
00:28:40
Wow. Eric
00:28:45
[Music]
00:28:48
very
00:28:49
Thanks, Eric. Appreciate you. Cheers.
00:28:53
All right.
