1
00:00:05,440 --> 00:00:07,837
Every day for the last 3 years, we have

2
00:00:07,920 --> 00:00:12,477
seen dramatic images from the war
in Ukraine. Periodically, we also see how

3
00:00:12,560 --> 00:00:16,797
Ukraine is defending itself
by developing new tactics, but also new

4
00:00:16,880 --> 00:00:20,717
tools. Drones, in particular,
are frequently pointed out as the

5
00:00:20,800 --> 00:00:23,757
protagonist of the new warfare. We had a

6
00:00:23,840 --> 00:00:26,157
live demonstration of this technology at

7
00:00:26,240 --> 00:00:29,277
the Ukrainian pavilion during CES. The

8
00:00:29,360 --> 00:00:33,917
drones on display were very cheap
to produce in mass quantity but also very

9
00:00:34,000 --> 00:00:36,317
deadly. >> This is also
military drone and we are

10
00:00:36,400 --> 00:00:40,557
right now looking know how to use it
in a in a civil in the civil sphere.

11
00:00:40,640 --> 00:00:42,477
>> It's actually a kamicazi drone.

12
00:00:42,560 --> 00:00:45,757
>> Oh wow. So the idea
of this drone was to

13
00:00:45,840 --> 00:00:50,637
develop super super cheap and effective
comma because you need like enormous

14
00:00:50,720 --> 00:00:53,917
amount of them you know on the front line
because it flies only one way >> right

15
00:00:54,000 --> 00:00:55,917
>> and it's really really cheap and it can

16
00:00:56,000 --> 00:00:59,117
fly up to 100 km with 4 kilo of payload

17
00:00:59,200 --> 00:01:01,117
>> how cheap

18
00:01:01,200 --> 00:01:05,997
>> in Ukraine less than 200 $2,000 uh in

19
00:01:06,080 --> 00:01:07,437
Ukraine abroad it would be more

20
00:01:07,520 --> 00:01:09,437
expensive obviously because you know we

21
00:01:09,520 --> 00:01:13,437
have some like our country like help um

22
00:01:13,520 --> 00:01:16,957
helps companies like us in we're
not paying taxes and so on.

23
00:01:17,040 --> 00:01:18,797
>> Everybody will be a bit more expensive

24
00:01:18,880 --> 00:01:25,000
but still not 20,000 like
we see like for example

25
00:01:27,440 --> 00:01:29,917
>> in fact Eric Schmid is suggesting that

26
00:01:30,000 --> 00:01:31,837
every government should consider

27
00:01:31,920 --> 00:01:34,637
investing in alternative modern weapons

28
00:01:34,720 --> 00:01:37,277
like drones alongside the traditional

29
00:01:37,360 --> 00:01:39,357
war machines. Take a listen. [music]

30
00:01:39,440 --> 00:01:43,997
is take 10% and put it on new stuff.

31
00:01:44,080 --> 00:01:46,957
Spend 90% on all the stuff that you

32
00:01:47,040 --> 00:01:50,237
think you should buy and run a trial.

33
00:01:50,320 --> 00:01:54,077
Literally test whether that 10% is

34
00:01:54,160 --> 00:01:58,797
better than the 90%.
And for example, for every tank you buy,

35
00:01:58,880 --> 00:02:02,557
build a [clears throat] unmanned ground
vehicle. For every airplane you buy,

36
00:02:02,640 --> 00:02:04,957
build a thousand drones or whatever,

37
00:02:05,040 --> 00:02:07,037
whatever the numbers are appropriate and

38
00:02:07,120 --> 00:02:10,157
run the test. uh

39
00:02:10,240 --> 00:02:14,717
you should be skeptical in this in your
organization. If you don't have tech

40
00:02:14,800 --> 00:02:17,037
people like these people embedded in

41
00:02:17,120 --> 00:02:21,917
every layer of in in this case the
chancellory and the mods and so forth of

42
00:02:22,000 --> 00:02:23,917
the countries you're going to miss out

43
00:02:24,000 --> 00:02:27,277
on a once in aundundredyear transition.

44
00:02:27,360 --> 00:02:31,517
This transition is analogous to the
invention of the tank 100 years ago. And

45
00:02:31,600 --> 00:02:33,997
by the way, tanks are not a good place

46
00:02:34,080 --> 00:02:38,237
to be in right now. 2 kg drone costs

47
00:02:38,320 --> 00:02:42,157
what $5,000 takes out 5 million to30

48
00:02:42,240 --> 00:02:44,797
million in the case of US tanks. Boom.

49
00:02:44,880 --> 00:02:48,237
Just like that. And they're
very hard to defend against.

50
00:02:48,320 --> 00:02:50,637
>> Eric Schmidt was a guest during one of

51
00:02:50,720 --> 00:02:55,357
the satellite events of the Munich
Security Forum. And he has no doubt that

52
00:02:55,440 --> 00:02:59,757
what we are seeing in Ukraine has many
elements of the warfare of the future

53
00:02:59,840 --> 00:03:03,757
wherever this may happen. >> [music]
>> Um, I spent lots of time in Ukraine

54
00:03:03,840 --> 00:03:06,957
trying to understand the future of war
and I think I I understand what's going

55
00:03:07,040 --> 00:03:10,477
to happen. Um, and I'm not talking about
Ukraine specifically. I'm talking about

56
00:03:10,560 --> 00:03:13,437
combat. Um, all of us are familiar with

57
00:03:13,520 --> 00:03:16,637
tanks, boats, and air force jets of one

58
00:03:16,720 --> 00:03:21,357
kind or another. All of you command
some combination of them. In every case,

59
00:03:21,440 --> 00:03:25,197
there's going to be an automation
equivalent of that. Instead of buying

60
00:03:25,280 --> 00:03:27,677
tanks, you'll be buying unmanounded

61
00:03:27,760 --> 00:03:32,397
ground vehicles. And I'll describe why
in a sec. Instead of buying big ships,

62
00:03:32,480 --> 00:03:34,077
you'll be buying the equivalent of drone

63
00:03:34,160 --> 00:03:37,597
ships automated in various ways. And in

64
00:03:37,680 --> 00:03:42,077
terms of t and in terms of air force,
you'll have drones. And the drones will

65
00:03:42,160 --> 00:03:44,077
be attack drones, but the drones will

66
00:03:44,160 --> 00:03:47,437
also be able to defend themselves. This

67
00:03:47,520 --> 00:03:50,557
is the future of war. And of course the

68
00:03:50,640 --> 00:03:55,037
existing systems be around but the quicker
you can make this transition the

69
00:03:55,120 --> 00:03:59,037
smarter and more sophisticated your
national security approaches will be.

70
00:03:59,120 --> 00:04:01,117
The elements that are composed of this

71
00:04:01,200 --> 00:04:03,837
are first abundance. I come from a

72
00:04:03,920 --> 00:04:06,157
background of abundance. You all spend a

73
00:04:06,240 --> 00:04:08,877
time of scarcity and ex exquisite

74
00:04:08,960 --> 00:04:13,677
expensive things. Uh in my view that's
wrong. Sorry to be obnoxious but I'm an

75
00:04:13,760 --> 00:04:17,037
American. um it's much
better to say how do we

76
00:04:17,120 --> 00:04:19,037
solve this problem with abundance with

77
00:04:19,120 --> 00:04:21,117
inexpensive systems that work well in

78
00:04:21,200 --> 00:04:24,317
swarms and that technology now works. So

79
00:04:24,400 --> 00:04:26,237
the first thing is drive the prices

80
00:04:26,320 --> 00:04:28,557
down. And when I say down, I mean down

81
00:04:28,640 --> 00:04:32,237
by a factor of 10, not 10%. And you do

82
00:04:32,320 --> 00:04:34,077
that by using product lines of one kind

83
00:04:34,160 --> 00:04:37,037
or another. You use um consumer product

84
00:04:37,120 --> 00:04:39,357
lines, you use contract manufacturers.

85
00:04:39,440 --> 00:04:42,957
You don't use the traditional military
procurement. The second thing that you

86
00:04:43,040 --> 00:04:44,477
do

87
00:04:44,560 --> 00:04:49,357
is um think of it as autom automation,

88
00:04:49,440 --> 00:04:51,917
right? Autonomy. Um, so for example,

89
00:04:52,000 --> 00:04:56,477
with modern vision systems, you can have
automatic target detection, and so far

90
00:04:56,560 --> 00:05:00,797
the systems that I've watched in Ukraine
have had zero collateral damage. These

91
00:05:00,880 --> 00:05:03,837
things are accurate enough that they're
more accurate than humans, which is

92
00:05:03,920 --> 00:05:06,477
important in in com in combat and in

93
00:05:06,560 --> 00:05:08,957
national security. And the third one is

94
00:05:09,040 --> 00:05:12,797
AI dominance. And AI dominance can be

95
00:05:12,880 --> 00:05:16,557
interpreted as many things. It can be
interpreted as vision systems. What I'm

96
00:05:16,640 --> 00:05:19,357
talking about is in strategy and

97
00:05:19,440 --> 00:05:21,757
algorithmic warfare using all of these

98
00:05:21,840 --> 00:05:26,877
tools is the future. It's how war will
be done. And obviously having spent a

99
00:05:26,960 --> 00:05:28,717
fair amount of time watching this real

100
00:05:28,800 --> 00:05:31,437
war, these are to be avoided, right? So

101
00:05:31,520 --> 00:05:34,157
I'm not in any way endorsing this. But

102
00:05:34,240 --> 00:05:36,397
the the the core point I'm trying to say

103
00:05:36,480 --> 00:05:39,517
is that this is happening very very

104
00:05:39,600 --> 00:05:43,197
quickly in Ukraine and all of you will

105
00:05:43,280 --> 00:05:45,757
either help or be dragged into it or

106
00:05:45,840 --> 00:05:48,637
will fight it because you don't like it.

107
00:05:48,720 --> 00:05:50,557
But the fact of the matter is this is

108
00:05:50,640 --> 00:05:53,517
the future and I'm convinced of it. Wars

109
00:05:53,600 --> 00:05:56,957
are won and lost through a whole bunch

110
00:05:57,040 --> 00:05:59,517
of players. There's no single reason why

111
00:05:59,600 --> 00:06:02,477
you win or lose. But every successful

112
00:06:02,560 --> 00:06:07,117
war plan will fundamentally have AI
as its strategic base and as its targeting

113
00:06:07,200 --> 00:06:11,677
base. The only question is how much
scale do you have in order to do this?

114
00:06:11,760 --> 00:06:15,997
By the way, you have to have lots of
training data, right? So the key with AI

115
00:06:16,080 --> 00:06:20,717
based things is that you actually have
to the system has to learn. Um and thank

116
00:06:20,800 --> 00:06:24,477
goodness there aren't that many wars,
but things like how do you identify

117
00:06:24,560 --> 00:06:28,637
targets? What does the enemy do? How do
they respond? All of these are playing

118
00:06:28,720 --> 00:06:30,477
out right now between Russia and

119
00:06:30,560 --> 00:06:34,157
Ukraine. In AI development,

120
00:06:34,240 --> 00:06:38,637
the fastest gains right now are
occurring in coding and you see this

121
00:06:38,720 --> 00:06:41,037
with Cloud Code and Augment and

122
00:06:41,120 --> 00:06:44,397
Microsoft Copilot and so forth. Um, my

123
00:06:44,480 --> 00:06:48,717
category of jobs, the job that I did
50 years ago is going away. It's rather

124
00:06:48,800 --> 00:06:53,277
disturbing. Um, but the fact of the
matter is that's where all the initial

125
00:06:53,360 --> 00:06:55,437
gains are right now for a whole bunch of

126
00:06:55,520 --> 00:07:00,317
reasons. If you can write code, you can
also write cyber attacks. So, one of the

127
00:07:00,400 --> 00:07:03,197
things that's going to happen is you're
gonna have an enormous number of zeroday

128
00:07:03,280 --> 00:07:05,597
cyber attacks coming out of this because

129
00:07:05,680 --> 00:07:08,237
of this innovation cycle. So, having

130
00:07:08,320 --> 00:07:13,197
said that, AI and AI servers that are
used in the military are legitimate

131
00:07:13,280 --> 00:07:16,797
military targets and therefore need
to be hardened and protected. They need to

132
00:07:16,880 --> 00:07:21,437
be hardened physically. They need to have
reliable power. But also the models

133
00:07:21,520 --> 00:07:25,277
uh there are things called adversarial
attacks where you can go in and take the

134
00:07:25,360 --> 00:07:29,037
model which is from your perspective
is just a bunch of numbers and change them

135
00:07:29,120 --> 00:07:31,757
enough that the AI does the wrong thing.

136
00:07:31,840 --> 00:07:36,317
So protecting them cyber security cyber
attacks are all incredibly important. I

137
00:07:36,400 --> 00:07:38,237
think this audience knows this is a

138
00:07:38,320 --> 00:07:40,477
growth business. Unfortunately when I

139
00:07:40,560 --> 00:07:44,397
talk about AI globally I always talk

140
00:07:44,480 --> 00:07:47,597
about what the US is doing and what

141
00:07:47,680 --> 00:07:52,717
China is doing. And the simple summary
is the US, thank God that the American

142
00:07:52,800 --> 00:07:55,997
financial system will give these people
billions and billions of dollars with no

143
00:07:56,080 --> 00:07:57,917
credible strategy for for making money

144
00:07:58,000 --> 00:07:59,997
at it. Um, Germans don't do that.

145
00:08:00,080 --> 00:08:03,117
Europeans don't do that. Um, but the

146
00:08:03,200 --> 00:08:07,837
fact of the matter is this enormous wave.
It's $2 trillion of capital going

147
00:08:07,920 --> 00:08:12,077
into the data centers in America and
around the world for those companies. Um

148
00:08:12,160 --> 00:08:13,917
the estimate for power needed in the

149
00:08:14,000 --> 00:08:19,357
United States is 90 gawatts
of additional power and as reference if

150
00:08:19,440 --> 00:08:23,677
you're not familiar with this the nuclear
power plant is around 1.5 gawatt

151
00:08:23,760 --> 00:08:27,917
gives you a sense of the scale involved
and this is these are the plans in China

152
00:08:28,000 --> 00:08:32,157
you have the same thing except that
the Chinese are using a technique called

153
00:08:32,240 --> 00:08:37,357
open source they're releasing all of the
weights and models um which is important

154
00:08:37,440 --> 00:08:41,517
strategically because it means the
majority of the world the third world if

155
00:08:41,600 --> 00:08:44,957
you will the other countries that we
don't typically visit are going to be

156
00:08:45,040 --> 00:08:47,597
using Chinese models not American models

157
00:08:47,680 --> 00:08:49,997
whereas Europe Europe has a couple of

158
00:08:50,080 --> 00:08:54,797
leading champions Mstral is one I'm an
investor for example but Europe does not

159
00:08:54,880 --> 00:08:58,077
have the scale does not have the data
centers does not have the financing to

160
00:08:58,160 --> 00:09:01,037
do it this is a tragedy Europe should

161
00:09:01,120 --> 00:09:03,117
organize in my view and I've said this

162
00:09:03,200 --> 00:09:05,437
and I'll say it again your own

163
00:09:05,520 --> 00:09:08,717
open-source models and they should be

164
00:09:08,800 --> 00:09:13,437
done thank you Um, it makes no sense for

165
00:09:13,520 --> 00:09:18,317
you to to allow your intelligence
systems and I mean human intelligence,

166
00:09:18,400 --> 00:09:19,837
not military intelligence, to be

167
00:09:19,920 --> 00:09:23,917
controlled by the US or China, right?

168
00:09:24,000 --> 00:09:27,917
Europe has its own special thing going
and I happen to like Europe and I want

169
00:09:28,000 --> 00:09:30,957
you guys to be successful and I don't
want you to be overrun by the Chinese or

170
00:09:31,040 --> 00:09:35,197
the US. You're going to have to do this
and furthermore it's going to cost you

171
00:09:35,280 --> 00:09:37,277
10 to20 billion.

172
00:09:37,360 --> 00:09:40,877
That's what they cost.

173
00:09:40,960 --> 00:09:45,197
Why open source? >> Um,
open source is the quickest way to

174
00:09:45,280 --> 00:09:47,437
catch up

175
00:09:47,520 --> 00:09:51,037
and having hired tens of thousands of

176
00:09:51,120 --> 00:09:56,557
people in Europe, the engineers
in Europe are fantastic, right? So, you

177
00:09:56,640 --> 00:09:59,597
need a way of organizing them. You need
a way to move quickly and open source is

178
00:09:59,680 --> 00:10:03,677
the easiest way to do it. My specific
proposal is that Brussels or equivalent

179
00:10:03,760 --> 00:10:06,717
organize a group of 10 computer
scientists who are obviously European

180
00:10:06,800 --> 00:10:10,557
and align with the British because
they're kind of near you um and figure

181
00:10:10,640 --> 00:10:12,717
it out. It can be done fairly quickly.

182
00:10:12,800 --> 00:10:17,437
>> The importance of open-source models is
often mentioned in this type of meetings

183
00:10:17,520 --> 00:10:21,437
also from the leading authorities
in this field. Take a listen. Let's face

184
00:10:21,520 --> 00:10:24,557
it, the single most important thing that
has happened last year, if you were to

185
00:10:24,640 --> 00:10:27,677
ask me the one single most important

186
00:10:27,760 --> 00:10:32,237
event last year and how it has activated

187
00:10:32,320 --> 00:10:34,717
AI researchers here in this region, it's

188
00:10:34,800 --> 00:10:37,277
actually Llama 2. It's an open- source

189
00:10:37,360 --> 00:10:40,157
model. All were possible because of some

190
00:10:40,240 --> 00:10:42,797
of these open source languages. And so I

191
00:10:42,880 --> 00:10:45,117
think that democratizing

192
00:10:45,200 --> 00:10:47,917
activating every region, activating

193
00:10:48,000 --> 00:10:52,957
every country to join the AI advance is

194
00:10:53,040 --> 00:10:55,597
probably one of the most important thing.

195
00:10:55,680 --> 00:11:00,157
>> For more content about the impact
of the Yama model and the importance of

196
00:11:00,240 --> 00:11:04,877
opensource models in general, you can
watch the video at the link above. But

197
00:11:04,960 --> 00:11:08,237
let's carry on with the future of warfare.

198
00:11:08,320 --> 00:11:10,557
What I'm going to say is that in a

199
00:11:10,640 --> 00:11:13,117
national security perspective,

200
00:11:13,200 --> 00:11:15,837
first the money is showing up which is

201
00:11:15,920 --> 00:11:18,317
great and that money needs to be applied

202
00:11:18,400 --> 00:11:21,357
to the right areas specifically the

203
00:11:21,440 --> 00:11:25,037
algorithms lowcost a factors technically

204
00:11:25,120 --> 00:11:28,557
drones drone industry you have a number
of national champions here you have a in

205
00:11:28,640 --> 00:11:32,237
Germany you have a company in Helsing
called Helsing that's quite good um

206
00:11:32,320 --> 00:11:33,757
there are a number of such examples and

207
00:11:33,840 --> 00:11:37,917
we need a lot more of them the um I most

208
00:11:38,000 --> 00:11:43,117
interested in not the problem of volume
which everyone is focused on because I

209
00:11:43,200 --> 00:11:46,157
think the volume stuff will
ultimately be sorted out but rather the

210
00:11:46,240 --> 00:11:48,157
intelligence layer and I'll give you an

211
00:11:48,240 --> 00:11:50,157
example

212
00:11:50,240 --> 00:11:53,597
um the way AI is evolving is we start

213
00:11:53,680 --> 00:11:56,797
with um the AI that you all know which

214
00:11:56,880 --> 00:12:00,957
is sort of chat GPT language models
things like that and then we move to

215
00:12:01,040 --> 00:12:05,277
strategy and planning which is where we
are now with strategy and planning you

216
00:12:05,360 --> 00:12:07,437
can begin to do steps Think of it as

217
00:12:07,520 --> 00:12:12,317
workflow. And when you can do that,
this is what militaries do. This is what

218
00:12:12,400 --> 00:12:16,957
businesses do. This is what universities
do. Europe should lead in the agentic

219
00:12:17,040 --> 00:12:20,797
revolution where you build agents
to solve the problems that your citizens,

220
00:12:20,880 --> 00:12:23,037
your military, your universities have.

221
00:12:23,120 --> 00:12:26,877
This is the year of the agentic
revolution. And that stuff is happening

222
00:12:26,960 --> 00:12:30,797
very fast. I'm talking about every month.
You saw last week all the

223
00:12:30,880 --> 00:12:34,557
publicity about a fake world of agents
that were beginning to learn how to

224
00:12:34,640 --> 00:12:37,277
interact with each other. And everyone
was surprised that they didn't

225
00:12:37,360 --> 00:12:41,037
immediately try to assassinate each other.
Well, that's presumably in game

226
00:12:41,120 --> 00:12:43,357
three of the simulation. But the

227
00:12:43,440 --> 00:12:46,397
important point is that the agent model

228
00:12:46,480 --> 00:12:48,477
is going to drive huge revenue in

229
00:12:48,560 --> 00:12:51,437
America. It should drive huge revenue in

230
00:12:51,520 --> 00:12:53,837
Europe. Along with that, start thinking

231
00:12:53,920 --> 00:12:58,637
about military uses. There's all
the supply chain stuff that's it. And you

232
00:12:58,720 --> 00:13:02,717
all spend so much money in my view
wasted it on these complicated systems

233
00:13:02,800 --> 00:13:06,237
just to have agents that figure
this stuff out and move it around. It's

234
00:13:06,320 --> 00:13:10,237
possible now using a technique called
MCP which is used throughout the

235
00:13:10,320 --> 00:13:14,077
industry model context protocol.
You basically take all your databases and

236
00:13:14,160 --> 00:13:17,677
you start asking it questions and it
figures it out. I mean these are huge

237
00:13:17,760 --> 00:13:20,077
breakthroughs architecturally and Europe

238
00:13:20,160 --> 00:13:22,557
is not leading in these areas. They will

239
00:13:22,640 --> 00:13:24,797
make you much more effective in whatever

240
00:13:24,880 --> 00:13:27,677
your objective is military, business,

241
00:13:27,760 --> 00:13:30,717
profit, entertainment, education. The

242
00:13:30,800 --> 00:13:33,517
the key next step after that is what

243
00:13:33,600 --> 00:13:37,197
agents do you build for war. So most

244
00:13:37,280 --> 00:13:42,397
people when they talk about AI talk
about it in the context of what in

245
00:13:42,480 --> 00:13:46,397
America is called project maven, which
is essentially vision and the project

246
00:13:46,480 --> 00:13:48,237
maven thing was very controversial.

247
00:13:48,320 --> 00:13:50,077
Google started it with the Pentagon.

248
00:13:50,160 --> 00:13:54,877
Google canceled it. Um, I was on the
Pentagon side, which is a mess. Um, and

249
00:13:54,960 --> 00:13:58,797
then eventually it's been taken
over largely by a palunteer.

250
00:13:58,880 --> 00:14:00,557
That's not the kind of AI I'm talking

251
00:14:00,640 --> 00:14:04,157
about. It's an enabler for reasoning.

252
00:14:04,240 --> 00:14:07,917
And every general I've met talks
about how they want to have battlefield

253
00:14:08,000 --> 00:14:11,277
awareness and they want to be able
to see everything. Well, here's what I

254
00:14:11,360 --> 00:14:13,837
want. When I sit in a command center and

255
00:14:13,920 --> 00:14:17,517
I watch the commanders who are very very

256
00:14:17,600 --> 00:14:19,757
committed to doing the right thing, I

257
00:14:19,840 --> 00:14:22,717
want the computer to give them a plan, a

258
00:14:22,800 --> 00:14:26,477
battle plan that we can prove will win

259
00:14:26,560 --> 00:14:30,957
in a series of iterations. There's
a technique called reinforcement learning

260
00:14:31,040 --> 00:14:35,757
which was invented um some years ago.
It was heavily used by Deep Mind winning

261
00:14:35,840 --> 00:14:39,677
the go game if you remember that about
10 years ago. And reinforcement learning

262
00:14:39,760 --> 00:14:44,317
allows you to sort of prove what the
optimal paths are. So here's an example.

263
00:14:44,400 --> 00:14:47,997
This would never happen, but you
and I are opponents. You have you have

264
00:14:48,080 --> 00:14:50,317
algorithms and I have algorithms. So now

265
00:14:50,400 --> 00:14:53,197
you and you and I are going to have a um

266
00:14:53,280 --> 00:14:56,477
and we're we're mad at each other. What

267
00:14:56,560 --> 00:14:59,277
is the deterrence that keeps you and me

268
00:14:59,360 --> 00:15:01,997
from fighting each other? Unlike

269
00:15:02,080 --> 00:15:06,797
traditional weapons,
you can't count the algorithms.

270
00:15:06,880 --> 00:15:08,637
If your algorithms are better than mine,

271
00:15:08,720 --> 00:15:11,357
and I suspect they would be, I better be

272
00:15:11,440 --> 00:15:14,157
very careful before I do something

273
00:15:14,240 --> 00:15:16,237
because you get a vote and you get a

274
00:15:16,320 --> 00:15:19,837
vote to to win. I spent a lot of time

275
00:15:19,920 --> 00:15:21,997
trying to figure out what this fight

276
00:15:22,080 --> 00:15:26,157
looks like. I'm sure within two or three

277
00:15:26,240 --> 00:15:29,517
years, algorithmic warfare as a category

278
00:15:29,600 --> 00:15:34,637
will exist. What I don't believe is
that the algorithms will be the same. And

279
00:15:34,720 --> 00:15:38,157
I'll give you an example.
In one of the more recent battles in

280
00:15:38,240 --> 00:15:42,557
Ukraine, and I won't go into the details,
um it was pretty clear that the

281
00:15:42,640 --> 00:15:47,437
Russians were focused on land and the
Ukrainians were focused on the people on

282
00:15:47,520 --> 00:15:48,877
the other side, right? These are

283
00:15:48,960 --> 00:15:52,157
different military strategies.

284
00:15:52,240 --> 00:15:55,197
It's not possible to model strategies

285
00:15:55,280 --> 00:15:58,077
where you have different goals, right?

286
00:15:58,160 --> 00:16:01,997
It's too hard. So the deterrence is if

287
00:16:02,080 --> 00:16:06,717
you and I have different strategies,
I don't know what you're really going

288
00:16:06,800 --> 00:16:11,517
after and therefore I can't model what
you're going to do. You become harder to

289
00:16:11,600 --> 00:16:15,597
predict with less
prediction capability. I'm

290
00:16:15,680 --> 00:16:18,957
less likely to start a war. So deterrence,

291
00:16:19,040 --> 00:16:22,317
>> it is a form of deterrence. And and the

292
00:16:22,400 --> 00:16:24,237
reason I'm hammering on this is that

293
00:16:24,320 --> 00:16:26,317
almost everybody learned from the

294
00:16:26,400 --> 00:16:29,997
nuclear stuff. Sorry to as an aside. So

295
00:16:30,080 --> 00:16:31,757
Henry Ken and I were best best friends.

296
00:16:31,840 --> 00:16:36,557
We wrote two books together and uh
and of course I miss him every day and he

297
00:16:36,640 --> 00:16:40,317
spent a lot of time here. He was by the
way here at the original MSSE way back

298
00:16:40,400 --> 00:16:42,877
when when I think none of us were here.

299
00:16:42,960 --> 00:16:45,437
Um and but anyway, Henry would would

300
00:16:45,520 --> 00:16:50,077
tell all these stories about how he did
deterrence and how the deterrence in the

301
00:16:50,160 --> 00:16:55,197
1950s was erected and it was fundamentally
because you could count things.

302
00:16:55,280 --> 00:16:59,117
I'm talking about a world where you
don't know. You can't model what the

303
00:16:59,200 --> 00:17:01,197
algorithms are going to do. Algorithmic

304
00:17:01,280 --> 00:17:03,997
warfare should be extremely

305
00:17:04,080 --> 00:17:08,557
self-deterring because of that issue.

306
00:17:08,640 --> 00:17:13,117
>> Some of these early capabilities
were already on display at the Ukrainian

307
00:17:13,200 --> 00:17:15,437
booth during CES.

308
00:17:15,520 --> 00:17:17,997
>> Run uh you know
the sophisticated things

309
00:17:18,080 --> 00:17:21,677
like like uh AI models. Then you just

310
00:17:21,760 --> 00:17:26,317
fly over your fields and the system
automatically makes a photos detect what

311
00:17:26,400 --> 00:17:29,917
what are you going you know to detect
and then just collects all the

312
00:17:30,000 --> 00:17:33,597
statistics calculates all the statistics
that you want and it depicts here on

313
00:17:33,680 --> 00:17:35,997
your on your screen.
>> There's wishful thinking and then

314
00:17:36,080 --> 00:17:39,997
there's reality. The wishful thinking is
that there's going to be a new algorithm

315
00:17:40,080 --> 00:17:41,597
that replaces the transformer

316
00:17:41,680 --> 00:17:43,197
architecture that Google invented in

317
00:17:43,280 --> 00:17:47,517
2017 that uses less hardware. That is a

318
00:17:47,600 --> 00:17:50,717
great idea. It's it's not obvious how

319
00:17:50,800 --> 00:17:53,677
and when that will occur. If you look at

320
00:17:53,760 --> 00:17:55,837
let's use Google as an example. Google

321
00:17:55,920 --> 00:17:58,557
is a highly efficient answering machine.

322
00:17:58,640 --> 00:18:03,437
The computational costs of doing the same
thing with any large language model

323
00:18:03,520 --> 00:18:07,437
are many orders of magnitude more
expensive which is why these things take

324
00:18:07,520 --> 00:18:11,837
so much time and the value is there.
And the part of the reason is that they do

325
00:18:11,920 --> 00:18:15,677
something called test time planning
where they are actually learning as

326
00:18:15,760 --> 00:18:20,637
they're going. They literally you you
you feed it what it knows and it knows

327
00:18:20,720 --> 00:18:24,077
something and then you feed it back
in again. It's called it's essentially a

328
00:18:24,160 --> 00:18:26,237
recursive process. The important thing

329
00:18:26,320 --> 00:18:28,877
about all of this is it uses an enormous

330
00:18:28,960 --> 00:18:31,277
amount of energy. Um as I mentioned

331
00:18:31,360 --> 00:18:35,517
before in the United States the um it's

332
00:18:35,600 --> 00:18:38,877
pretty clear that every real estate
developer in the United States has

333
00:18:38,960 --> 00:18:42,477
become obsessed with making more
money from data center builds than from

334
00:18:42,560 --> 00:18:44,877
building building on farmland. And the

335
00:18:44,960 --> 00:18:47,757
US has lots of extra land. So you're

336
00:18:47,840 --> 00:18:52,637
going to see lots and lots of that. And
this particular company um is working in

337
00:18:52,720 --> 00:18:54,637
an area called the Perian Basin which

338
00:18:54,720 --> 00:18:57,037
has a great deal of natural gas. In

339
00:18:57,120 --> 00:18:59,677
looking at it for Europe, Europe has uh

340
00:18:59,760 --> 00:19:04,957
France has excess nuclear capacity
as far as I can tell and the the only

341
00:19:05,040 --> 00:19:08,717
problem is the prices in France
for energy are high. I was recently

342
00:19:08,800 --> 00:19:12,077
contacted by somebody who's figured out
a way to build a gigawatt in Bulgaria

343
00:19:12,160 --> 00:19:16,397
which I thought was interesting.
My guess for Europeans because the energy

344
00:19:16,480 --> 00:19:18,557
costs are so high is that people are

345
00:19:18,640 --> 00:19:20,397
here. The training will be done

346
00:19:20,480 --> 00:19:23,757
somewhere else. This is a a national

347
00:19:23,840 --> 00:19:27,837
security issue for Europe as a strategic

348
00:19:27,920 --> 00:19:30,637
issue. If I were doing things for the

349
00:19:30,720 --> 00:19:33,597
military in Europe, I'd want my training

350
00:19:33,680 --> 00:19:36,157
to be done in a European NATO type

351
00:19:36,240 --> 00:19:38,877
country. European values, European laws,

352
00:19:38,960 --> 00:19:43,917
European privacy laws. And that's hence
why I keep going on about Europe needs a

353
00:19:44,000 --> 00:19:47,517
strategy. You need some form of open
source model and you need to do the

354
00:19:47,600 --> 00:19:51,277
training in Europe.
>> So you're building this facility where

355
00:19:51,360 --> 00:19:55,837
natural gas is abundant. So you're in
the right place for data centers. Middle

356
00:19:55,920 --> 00:19:59,277
East would seem to be a logical place
to put some of them, but not the ones for

357
00:19:59,360 --> 00:20:03,117
military use. >> Yeah,
I think I think that for the

358
00:20:03,200 --> 00:20:06,797
militaries will ultimately figure out
that controlling that value chain is

359
00:20:06,880 --> 00:20:11,437
really important. You want to have pretty
much vertical control over your stuff.

360
00:20:11,520 --> 00:20:13,517
>> Now before this event, we followed the

361
00:20:13,600 --> 00:20:16,397
World Economic Forum in Davos and it was

362
00:20:16,480 --> 00:20:21,117
evident that even among the leaders in AI,
there was no consensus about the

363
00:20:21,200 --> 00:20:23,357
timeline [music] for AI to become a

364
00:20:23,440 --> 00:20:25,917
truly revolutionary. Eric Schmidt

365
00:20:26,000 --> 00:20:29,517
confirms the current situation.

366
00:20:29,600 --> 00:20:33,277
The real narrative is that America

367
00:20:33,360 --> 00:20:37,197
because of money, capital scale,

368
00:20:37,280 --> 00:20:39,997
craziness, people like me touting a

369
00:20:40,080 --> 00:20:41,517
future in two years where everything

370
00:20:41,600 --> 00:20:44,477
will be different. The AI boom in the US

371
00:20:44,560 --> 00:20:48,077
is like a religion, right? And I call it

372
00:20:48,160 --> 00:20:52,317
the San Francisco consensus. And um the
people in San Francisco that I know and

373
00:20:52,400 --> 00:20:54,317
I spent lots of time with them all

374
00:20:54,400 --> 00:20:57,837
believe within 2 or 3 years

375
00:20:57,920 --> 00:20:59,437
something very interesting is going to

376
00:20:59,520 --> 00:21:03,597
happen. And one of my friends said h and

377
00:21:03,680 --> 00:21:06,477
he goes then we don't know what will

378
00:21:06,560 --> 00:21:09,437
happen. And by the way this is the way

379
00:21:09,520 --> 00:21:12,557
the pitch goes. When these systems these

380
00:21:12,640 --> 00:21:14,317
systems are smart and they can do

381
00:21:14,400 --> 00:21:17,437
planning and so forth, there's a point

382
00:21:17,520 --> 00:21:22,237
at which the the system
can begin to build itself.

383
00:21:22,320 --> 00:21:24,397
It can take off on its own. The

384
00:21:24,480 --> 00:21:27,277
technical term is called rein uh re

385
00:21:27,360 --> 00:21:32,157
essentially recursive self-improvement
recursively getting smarter and smarter.

386
00:21:32,240 --> 00:21:35,757
Now, if you think about it,
if the computer evolves faster than we do,

387
00:21:35,840 --> 00:21:40,317
we're in trouble. And then everyone goes,
"Oh my god, we're in trouble." Now

388
00:21:40,400 --> 00:21:41,997
I personally don't think it'll be within

389
00:21:42,080 --> 00:21:45,197
2 years but the the but I do believe

390
00:21:45,280 --> 00:21:48,077
that over a 5 or 10 year period the rise

391
00:21:48,160 --> 00:21:51,117
of this foreign intelligence and how it

392
00:21:51,200 --> 00:21:54,557
gets hard harnessed is going to be huge.

393
00:21:54,640 --> 00:21:57,837
My view about Europe since we're
in Europe is Europe needs to be a player

394
00:21:57,920 --> 00:22:01,917
there and I don't see Europe doing
it right now because of the lack of

395
00:22:02,000 --> 00:22:04,637
hardware and companies and so forth and

396
00:22:04,720 --> 00:22:06,957
uh Europe is in big danger to be left

397
00:22:07,040 --> 00:22:10,637
out of this and China for example does

398
00:22:10,720 --> 00:22:14,557
understand this and the Chinese are
working very very hard using a different

399
00:22:14,640 --> 00:22:19,277
approach. They do more computation
at the edge. The computers are not as

400
00:22:19,360 --> 00:22:23,677
powerful because the US won't allow
them to get those chips. Um, and that's

401
00:22:23,760 --> 00:22:28,397
there's a complicated discussion about
all of that. I should mention that ASML,

402
00:22:28,480 --> 00:22:32,077
which is a European company,
is incredibly impressive and incredibly

403
00:22:32,160 --> 00:22:36,237
important. If Europe could just create
more ASMLs, everyone would be good,

404
00:22:36,320 --> 00:22:39,037
right? It's it's a true monopoly on a

405
00:22:39,120 --> 00:22:41,437
technology that is essential for the AI

406
00:22:41,520 --> 00:22:43,837
revolution. The these transitions take

407
00:22:43,920 --> 00:22:46,717
decades, right? Did anyone here arrive

408
00:22:46,800 --> 00:22:50,557
in a driverless car here in Munich?

409
00:22:50,640 --> 00:22:55,837
This is the car capital of the world,
isn't it? Germany, right? By the way, in

410
00:22:55,920 --> 00:22:58,957
America, most people don't arrive
in self-driving cars, too. And they were

411
00:22:59,040 --> 00:23:02,237
invented in the basically the late '9s.

412
00:23:02,320 --> 00:23:05,117
So, what is that 30 years, right? It

413
00:23:05,200 --> 00:23:07,277
just takes a long time. The industry

414
00:23:07,360 --> 00:23:10,237
uses the phrase technology overhang to

415
00:23:10,320 --> 00:23:12,317
refer to the fact that the rate of

416
00:23:12,400 --> 00:23:15,197
diffusion or adoption of the techn tech

417
00:23:15,280 --> 00:23:18,077
the stuff we're talking about is slow.

418
00:23:18,160 --> 00:23:21,917
It's slow for many reasons. Sometimes
it doesn't work. Sometimes you don't have

419
00:23:22,000 --> 00:23:25,437
the right people on the other side.
But at least in the case of national

420
00:23:25,520 --> 00:23:27,757
security, you have an obligation to

421
00:23:27,840 --> 00:23:30,077
protect your countries. You have to, in

422
00:23:30,160 --> 00:23:33,437
my view, take the jump, take the risk,

423
00:23:33,520 --> 00:23:38,077
and at least run the test. Right? To me,
it's an it's a moral imperative. It's a

424
00:23:38,160 --> 00:23:43,277
national imperative. And is exemplified
in the war. um with respect to to the

425
00:23:43,360 --> 00:23:47,197
rate at which this stuff is going
to happen, a lot of it depends on how easy

426
00:23:47,280 --> 00:23:51,277
it will be for normal people to use it,
right? So, for example, Google and I

427
00:23:51,360 --> 00:23:55,997
obviously have a pro Google bias.
Um many of its Gemini answers are really

428
00:23:56,080 --> 00:23:57,677
better than the search results, right?

429
00:23:57,760 --> 00:24:01,277
And Gemini is really good.
It's currently the best performing model,

430
00:24:01,360 --> 00:24:03,117
right? So, it's according to the

431
00:24:03,200 --> 00:24:06,637
rankings. So, good for them. Uh so the

432
00:24:06,720 --> 00:24:10,397
the key thing to understand here
is you have this brutal competition for

433
00:24:10,480 --> 00:24:14,077
innovation which is being funded
by trillions of dollars. We all benefit

434
00:24:14,160 --> 00:24:18,157
from that competition. Let's see how far
it goes. Specific things that are going

435
00:24:18,240 --> 00:24:21,037
to happen are agents will be used by

436
00:24:21,120 --> 00:24:24,077
businesses to improve customer service

437
00:24:24,160 --> 00:24:28,477
in there's evidence that customer
service jobs for example are declining

438
00:24:28,560 --> 00:24:32,637
because you can automate those.
A general rule about automation is that

439
00:24:32,720 --> 00:24:37,277
highly um highly
repetitive jobs and highly

440
00:24:37,360 --> 00:24:39,117
dangerous jobs are automated out first.

441
00:24:39,200 --> 00:24:40,237
And I think that's what you'll see.

442
00:24:40,320 --> 00:24:42,340
[music]