1
00:00:02,720 --> 00:00:16,400
Speaker 1: Bloomberg Audio Studios, Podcasts, Radio News.

2
00:00:18,079 --> 00:00:21,759
Speaker 2: Hello and welcome to another episode of the Odd Lots Podcast.

3
00:00:21,880 --> 00:00:24,280
Speaker 3: I'm Joeisenthal and I'm Tracy Alloway.

4
00:00:24,640 --> 00:00:27,440
Speaker 2: So Tracy recording this March twenty fourth, And of course

5
00:00:27,480 --> 00:00:29,800
Speaker 2: almost all of our episodes.

6
00:00:29,440 --> 00:00:31,680
Speaker 4: Lately have been about the war in Iran.

7
00:00:32,280 --> 00:00:34,879
Speaker 2: But what's interesting, or what's a little weird is that

8
00:00:35,240 --> 00:00:38,520
Speaker 2: just prior to the war, literally days or maybe hours,

9
00:00:38,880 --> 00:00:42,320
Speaker 2: the biggest story in the world was actually about defense

10
00:00:42,400 --> 00:00:44,760
Speaker 2: and you know the DoD that's right.

11
00:00:44,800 --> 00:00:48,239
Speaker 3: So you are referring to anthrop a. Yeah, and it's

12
00:00:48,520 --> 00:00:51,280
Speaker 3: disagreement to put it mildly in the Department of War.

13
00:00:51,680 --> 00:00:52,440
Speaker 4: Yeah, exactly.

14
00:00:52,479 --> 00:00:55,160
Speaker 2: This was the biggest story going right up on the

15
00:00:55,200 --> 00:00:57,920
Speaker 2: eve of the war in Iran. And of course obviously

16
00:00:58,040 --> 00:01:02,240
Speaker 2: there was this contract topics technology was used by the

17
00:01:02,240 --> 00:01:05,280
Speaker 2: Defense Department. So it was not a disagreement about the

18
00:01:05,360 --> 00:01:08,920
Speaker 2: use of AI per se in war, but the question

19
00:01:09,000 --> 00:01:12,440
Speaker 2: of the degree to which AI could be used for

20
00:01:12,520 --> 00:01:15,360
Speaker 2: autonomous weapons systems on their own without a human in the.

21
00:01:15,360 --> 00:01:18,720
Speaker 5: Loop and surveillance and surveillance this was another key element.

22
00:01:18,840 --> 00:01:22,479
Speaker 3: But you're right, so we've heard this expression autonomous weapons, yeah,

23
00:01:22,520 --> 00:01:25,640
Speaker 3: pop up more and more especially in recent days. And

24
00:01:25,680 --> 00:01:28,880
Speaker 3: I've a lot of questions over what exactly that means,

25
00:01:29,000 --> 00:01:31,880
Speaker 3: because my impression is the US military certainly has been

26
00:01:31,959 --> 00:01:33,320
Speaker 3: using AI for some time.

27
00:01:33,440 --> 00:01:35,480
Speaker 5: Yes, and so we're really talking about.

28
00:01:35,560 --> 00:01:40,360
Speaker 3: Degrees here of autonomy, right, And so if you think

29
00:01:40,360 --> 00:01:43,319
Speaker 3: about an autonomous weapon, I think your mind could go

30
00:01:43,560 --> 00:01:46,760
Speaker 3: fully terminator and you know there's like a murder robot

31
00:01:46,800 --> 00:01:50,680
Speaker 3: out there that's making its own decisions on which.

32
00:01:50,600 --> 00:01:52,720
Speaker 5: People or places to target.

33
00:01:53,200 --> 00:01:55,880
Speaker 3: And then you get levels below that, right where AI

34
00:01:56,040 --> 00:01:59,920
Speaker 3: is kind of helping humans to come up with strategic decisions.

35
00:02:00,160 --> 00:02:00,360
Speaker 4: Right.

36
00:02:00,480 --> 00:02:03,000
Speaker 2: So, if there is a missile coming and you have

37
00:02:03,040 --> 00:02:06,200
Speaker 2: a missile defense system, I don't think you want a

38
00:02:06,280 --> 00:02:08,400
Speaker 2: human in the loop. Is like, Okay, here are the

39
00:02:08,440 --> 00:02:11,200
Speaker 2: coordinates X, y Z that we think it's going to hit.

40
00:02:11,639 --> 00:02:13,360
Speaker 2: At this point in time, we think the missile will

41
00:02:13,360 --> 00:02:14,880
Speaker 2: be here, Are you cool with firing it?

42
00:02:15,040 --> 00:02:16,240
Speaker 4: I think everyone's.

43
00:02:15,840 --> 00:02:18,960
Speaker 2: Probably okay with that level of autonomy. But I have

44
00:02:19,000 --> 00:02:22,799
Speaker 2: a feeling that to your point exactly a lot of

45
00:02:22,840 --> 00:02:26,320
Speaker 2: this discussion, and maybe it's core to what Anthropic and

46
00:02:26,360 --> 00:02:28,680
Speaker 2: the Department of Defense who are disagreeing on, I have

47
00:02:28,720 --> 00:02:30,720
Speaker 2: a feeling a lot of this is going to come

48
00:02:30,760 --> 00:02:33,880
Speaker 2: down to definitions. My guess is that there is not

49
00:02:34,200 --> 00:02:38,440
Speaker 2: one shared agreement of this is an autonomous weapon system

50
00:02:38,720 --> 00:02:39,560
Speaker 2: and this one is.

51
00:02:39,520 --> 00:02:42,440
Speaker 3: Not absolutely And of course there are also questions over

52
00:02:42,600 --> 00:02:47,440
Speaker 3: exactly how places like the Department of Defense, not only

53
00:02:47,440 --> 00:02:51,120
Speaker 3: how they define it, but once they have those definitions,

54
00:02:51,240 --> 00:02:55,120
Speaker 3: whether or not certain companies trust them. Yes, so stick

55
00:02:55,160 --> 00:02:57,400
Speaker 3: to those policies, because the US will say, well, our

56
00:02:57,520 --> 00:03:01,400
Speaker 3: policy is not to survey are citizens at the moment,

57
00:03:01,520 --> 00:03:03,720
Speaker 3: So if you're anthropic, you don't need to worry about that.

58
00:03:04,080 --> 00:03:07,600
Speaker 3: Clearly anthropic feels otherwiseer say they do. So there are

59
00:03:07,600 --> 00:03:11,799
Speaker 3: all these really interesting thematic questions that pop up from all.

60
00:03:11,639 --> 00:03:12,840
Speaker 4: Of this totally.

61
00:03:12,880 --> 00:03:15,320
Speaker 2: And then there's the question of Okay, here's a technology

62
00:03:15,360 --> 00:03:17,639
Speaker 2: and the government says, we believe that we can use

63
00:03:17,680 --> 00:03:19,720
Speaker 2: this to make the country safe. For what, You're not

64
00:03:19,760 --> 00:03:22,200
Speaker 2: going to let us do it like private corporation. There's

65
00:03:22,240 --> 00:03:25,000
Speaker 2: some very interesting questions about the role of corporate power

66
00:03:25,080 --> 00:03:26,800
Speaker 2: visa vi the government and so forth.

67
00:03:26,840 --> 00:03:29,359
Speaker 4: Anyway, this is something that has.

68
00:03:29,240 --> 00:03:32,120
Speaker 2: Become even more timely in the reports, even in the

69
00:03:32,160 --> 00:03:35,640
Speaker 2: early days of the Iran War, about these AI systems

70
00:03:35,640 --> 00:03:38,480
Speaker 2: having been used perhaps in target selection, but we don't

71
00:03:38,520 --> 00:03:40,200
Speaker 2: really know. None of the reporting is like that.

72
00:03:40,440 --> 00:03:40,720
Speaker 6: Clear.

73
00:03:40,800 --> 00:03:44,400
Speaker 2: I don't think that they're going out and advertising this strike.

74
00:03:44,600 --> 00:03:46,440
Speaker 5: This is exactly how we're using AI.

75
00:03:46,600 --> 00:03:48,680
Speaker 2: This is exactly how we're using AI, and so forth.

76
00:03:48,760 --> 00:03:51,920
Speaker 2: But this is obviously a huge debate and war aside.

77
00:03:51,920 --> 00:03:54,760
Speaker 2: It's totally going to grow, and just as an AI

78
00:03:54,880 --> 00:03:57,840
Speaker 2: is going to grow, it seems in so many different areas. Anyway,

79
00:03:57,840 --> 00:04:00,280
Speaker 2: I'm really excited to say, we really do have the

80
00:04:00,320 --> 00:04:02,680
Speaker 2: perfect guest, someone who's been writing and thinking about this

81
00:04:02,760 --> 00:04:05,440
Speaker 2: stuff for a long time. When we talked to an

82
00:04:05,480 --> 00:04:08,520
Speaker 2: AI expert, I marked the dividing line AND's like, where

83
00:04:08,640 --> 00:04:12,120
Speaker 2: you talking about AI prior to when Chad Gipt was released,

84
00:04:12,160 --> 00:04:14,280
Speaker 2: as like, I take a little bit more seriously the

85
00:04:14,280 --> 00:04:17,200
Speaker 2: people who are in this space prior to November twenty

86
00:04:17,320 --> 00:04:20,120
Speaker 2: twenty two. Anyway, I'm very excited to say we're gonna

87
00:04:20,120 --> 00:04:22,440
Speaker 2: be speaking with Paul Shari. He's the executive vice president

88
00:04:22,520 --> 00:04:25,120
Speaker 2: at the Center for a New American Security and he's

89
00:04:25,160 --> 00:04:27,520
Speaker 2: the author of two books related to this, one as

90
00:04:27,680 --> 00:04:30,240
Speaker 2: the most recent for Battlegrounds Power in the Age of

91
00:04:30,320 --> 00:04:33,040
Speaker 2: Artificial Intelligence, and then prior to that, he is the

92
00:04:33,080 --> 00:04:36,159
Speaker 2: author of Army of non autonomous weapons and the future

93
00:04:36,240 --> 00:04:39,080
Speaker 2: of war. He was previously in the Office of the

94
00:04:39,120 --> 00:04:42,760
Speaker 2: Secretary of Defense, also a previous Army ranger, so truly

95
00:04:42,800 --> 00:04:45,080
Speaker 2: the perfect guest. So, Paul, thank you so much for

96
00:04:45,279 --> 00:04:46,640
Speaker 2: coming on odlots.

97
00:04:47,240 --> 00:04:49,440
Speaker 6: Oh thank you for having me. Very excited to be here.

98
00:04:49,800 --> 00:04:50,760
Speaker 4: What do you we start?

99
00:04:50,880 --> 00:04:53,880
Speaker 2: I mentioned I had a feeling that maybe the definition

100
00:04:53,960 --> 00:04:57,360
Speaker 2: of an autonomous weapon is a contested one. But if

101
00:04:57,400 --> 00:04:59,640
Speaker 2: I say to you, what's an autonomous weapon, what's an

102
00:04:59,640 --> 00:05:00,559
Speaker 2: autonom this weapon?

103
00:05:01,520 --> 00:05:03,280
Speaker 6: So I think you're right from the beginning that there

104
00:05:03,320 --> 00:05:07,320
Speaker 6: is not a unified definition that everyone agrees on. The

105
00:05:07,360 --> 00:05:11,080
Speaker 6: Defense Department has their definition that's written in their policy.

106
00:05:11,480 --> 00:05:14,800
Speaker 6: I think conceptually, I think the distinction really is a

107
00:05:14,839 --> 00:05:18,599
Speaker 6: weapon that is choosing its own targets on the battlefield,

108
00:05:18,800 --> 00:05:20,920
Speaker 6: and it's not where we are today. Right now, today

109
00:05:21,000 --> 00:05:24,240
Speaker 6: people are choosing those targets. But it is kind of

110
00:05:24,240 --> 00:05:26,839
Speaker 6: a spectrum because we do have examples of weapons that

111
00:05:26,920 --> 00:05:30,919
Speaker 6: have some measure of autonomy. A good analogy might be

112
00:05:30,920 --> 00:05:35,160
Speaker 6: self driving cars, where conceptually, like, okay, a self driving

113
00:05:35,200 --> 00:05:36,720
Speaker 6: car would be where the AI is driving the car,

114
00:05:36,960 --> 00:05:39,640
Speaker 6: but then you get into an actual car today and

115
00:05:39,680 --> 00:05:43,200
Speaker 6: a lot of them have intelligence cruise control, automatic breaking,

116
00:05:43,760 --> 00:05:47,040
Speaker 6: automated self parking, They have all these like automated features

117
00:05:47,040 --> 00:05:50,120
Speaker 6: that are kind of creeping you in this direction where

118
00:05:50,160 --> 00:05:53,400
Speaker 6: the AI is taking over more and more control for

119
00:05:53,520 --> 00:05:56,240
Speaker 6: what the vehicle can do. And it's actually pretty similar

120
00:05:56,240 --> 00:05:57,560
Speaker 6: thing in the military space as well.

121
00:05:58,080 --> 00:06:00,880
Speaker 3: So Joe mentioned that had we been having this conversation

122
00:06:00,960 --> 00:06:03,920
Speaker 3: even a month ago now, it probably would have had

123
00:06:04,160 --> 00:06:09,760
Speaker 3: fewer concrete examples of AI enabled weaponry, let's say, or strategy.

124
00:06:10,320 --> 00:06:14,240
Speaker 3: When the Pentagon talks about its advanced AI tools that

125
00:06:14,279 --> 00:06:18,360
Speaker 3: it's deploying for the Iran conflict, what are some examples

126
00:06:18,400 --> 00:06:21,760
Speaker 3: that you're seeing right now that are different to say,

127
00:06:22,080 --> 00:06:25,000
Speaker 3: maybe just a year ago when we had another Iran conflict.

128
00:06:26,440 --> 00:06:29,080
Speaker 6: Right, So, there's a couple ways in which the Pentagon

129
00:06:29,279 --> 00:06:33,200
Speaker 6: is using AI right now. One is narrow AI systems

130
00:06:33,200 --> 00:06:36,440
Speaker 6: that have around for over a decade now that do

131
00:06:36,760 --> 00:06:41,400
Speaker 6: image classification. For example, So this was the military's original

132
00:06:41,400 --> 00:06:45,520
Speaker 6: project Maven almost a decade ago, where they took machine

133
00:06:45,560 --> 00:06:49,320
Speaker 6: learning image classifiers to sift through drone video feeds and

134
00:06:49,400 --> 00:06:53,080
Speaker 6: saddle images to identify objects. Okay, here's a building, here's

135
00:06:53,120 --> 00:06:56,839
Speaker 6: a person, here's a vehicle, that's pretty mature technology. Now,

136
00:06:56,960 --> 00:06:58,880
Speaker 6: what's come out in just the last couple of weeks

137
00:06:59,000 --> 00:07:02,400
Speaker 6: that's really quite interesting is that in the midst of

138
00:07:02,440 --> 00:07:06,160
Speaker 6: this huge, messy public breakup between Anthropic and the Pentagon,

139
00:07:06,600 --> 00:07:10,000
Speaker 6: we found out that in fact, Anthropics AI tools are

140
00:07:10,040 --> 00:07:13,840
Speaker 6: being used by the US military to help plan the

141
00:07:13,880 --> 00:07:17,640
Speaker 6: war against Iran. That's obviously a different kind of AI tool,

142
00:07:18,040 --> 00:07:22,120
Speaker 6: AI Large language models, AI being used to write code

143
00:07:22,200 --> 00:07:25,040
Speaker 6: AI agents, and that's being used in a different ways,

144
00:07:25,160 --> 00:07:29,640
Speaker 6: really helping intel analysts sift through just the massive amounts

145
00:07:29,640 --> 00:07:33,080
Speaker 6: of data that the US military has. And so you

146
00:07:33,080 --> 00:07:36,160
Speaker 6: can imagine the problem that the military is facing right

147
00:07:36,200 --> 00:07:39,760
Speaker 6: now when they're looking at targets in Iran. US military

148
00:07:39,760 --> 00:07:44,560
Speaker 6: has flown over six thousand sorties against Iran. The Iranian

149
00:07:44,600 --> 00:07:47,480
Speaker 6: military architecture is degraded in a lot of ways. US

150
00:07:47,520 --> 00:07:50,880
Speaker 6: militaries are bombed a lot of targets. There are mobile targets,

151
00:07:51,160 --> 00:07:56,360
Speaker 6: senior Iranian commanders, mobile missile launchers and air defense systems

152
00:07:56,480 --> 00:07:59,600
Speaker 6: and drone launchers. US militter has got to bring off

153
00:07:59,640 --> 00:08:02,720
Speaker 6: it in f together and find out where are these

154
00:08:02,720 --> 00:08:05,520
Speaker 6: targets right now and where is there an aircraft that

155
00:08:05,640 --> 00:08:07,720
Speaker 6: is the right bombs on it to take these targets out.

156
00:08:08,040 --> 00:08:10,600
Speaker 6: And that's how AI is being used to help basically

157
00:08:11,040 --> 00:08:12,960
Speaker 6: process and understand all that information.

158
00:08:13,240 --> 00:08:16,720
Speaker 2: When I think about the description that you gave for that,

159
00:08:17,240 --> 00:08:19,200
Speaker 2: I sometimes think like, could it be that?

160
00:08:19,760 --> 00:08:19,880
Speaker 1: Now?

161
00:08:19,920 --> 00:08:23,120
Speaker 2: I don't think that using anthropics technology mean so that

162
00:08:23,120 --> 00:08:26,520
Speaker 2: they go into Claude dot AI and say, give us

163
00:08:26,520 --> 00:08:31,000
Speaker 2: a list of suitable targets for sorties. But could it

164
00:08:31,040 --> 00:08:33,120
Speaker 2: be something like that? But I'm sure there's a different

165
00:08:33,120 --> 00:08:36,839
Speaker 2: interface and so forth. But is that a completely ridiculous

166
00:08:36,880 --> 00:08:40,800
Speaker 2: way of essentially framing the service that AI is providing

167
00:08:40,880 --> 00:08:41,320
Speaker 2: right now?

168
00:08:42,640 --> 00:08:45,800
Speaker 6: So the way that these AI tools are being integrated

169
00:08:45,960 --> 00:08:49,559
Speaker 6: are through a existing system called the Maven Smart System okay,

170
00:08:49,559 --> 00:08:52,880
Speaker 6: which is built by Palenteer, and that fuses all this

171
00:08:53,000 --> 00:08:56,600
Speaker 6: data together. So you basically have an existing architecture for

172
00:08:56,800 --> 00:09:00,680
Speaker 6: data management for Intel analysts that the Milita already has

173
00:09:01,040 --> 00:09:03,120
Speaker 6: that brings together all these different forms of data. You

174
00:09:03,200 --> 00:09:08,280
Speaker 6: might have satellite imagery, geolocation, data, signals, intelligence, other forms

175
00:09:08,280 --> 00:09:12,960
Speaker 6: of information. That's pretty great for Intel analysts, but that's

176
00:09:13,000 --> 00:09:16,520
Speaker 6: also really unwieldy because how does a human understand all

177
00:09:16,559 --> 00:09:20,120
Speaker 6: that data and process it? And that's where the large

178
00:09:20,160 --> 00:09:22,880
Speaker 6: language model tools, whether it's Clawed or other companies can

179
00:09:22,880 --> 00:09:26,960
Speaker 6: be valuable. Is that could be a way for a

180
00:09:27,080 --> 00:09:31,040
Speaker 6: human to interact with that data, to basically task a

181
00:09:31,120 --> 00:09:33,280
Speaker 6: large language model to say, Okay, here's a bunch of

182
00:09:33,400 --> 00:09:36,160
Speaker 6: data I'm giving you. I want you to look for

183
00:09:36,720 --> 00:09:39,320
Speaker 6: intersections and things, right. I want you to look for

184
00:09:39,360 --> 00:09:44,160
Speaker 6: a place where we have satellite imagery and some other

185
00:09:44,280 --> 00:09:47,960
Speaker 6: forms of intelligence that can help identify the location of

186
00:09:48,440 --> 00:09:51,800
Speaker 6: you know, some missile launcher for example, and then humans

187
00:09:51,840 --> 00:09:55,040
Speaker 6: can look at that and help one just like find

188
00:09:55,200 --> 00:09:58,920
Speaker 6: where are all these targets, and then it helpfully planning too.

189
00:09:58,960 --> 00:10:01,320
Speaker 6: A human could say, okay, here's this list of potential

190
00:10:01,360 --> 00:10:04,199
Speaker 6: targets that I have. Now they're scattered all over run

191
00:10:04,240 --> 00:10:06,440
Speaker 6: around is a really big country. I want to map

192
00:10:06,480 --> 00:10:10,760
Speaker 6: these two locations for US aircraft at different bases across

193
00:10:10,760 --> 00:10:14,400
Speaker 6: the region. What are available aircraft and what are available

194
00:10:14,600 --> 00:10:17,760
Speaker 6: munitions on those aircraft? The commused to take out those

195
00:10:17,760 --> 00:10:21,280
Speaker 6: targets to help build a strike package. So like, the

196
00:10:21,320 --> 00:10:24,800
Speaker 6: AI is definitely being used to help understand the battlespace

197
00:10:24,800 --> 00:10:27,800
Speaker 6: and to plan operations. But in I would say ways

198
00:10:27,800 --> 00:10:31,559
Speaker 6: that are pretty narrowly directed by people It's not quite

199
00:10:31,559 --> 00:10:33,680
Speaker 6: as simple as like, dump all this data into a

200
00:10:33,679 --> 00:10:37,000
Speaker 6: context winder for LM and say, oh, AI figure it out. Okay,

201
00:10:37,000 --> 00:10:40,040
Speaker 6: people are asking the AI some really specific questions, So.

202
00:10:40,280 --> 00:10:43,880
Speaker 3: I'm thinking how to phrase this question diplomatically, But I

203
00:10:43,920 --> 00:10:47,480
Speaker 3: get that the difference between fully autonomous weapons is, you know,

204
00:10:47,520 --> 00:10:51,640
Speaker 3: the human as a decision maker in the current setup,

205
00:10:51,800 --> 00:10:55,520
Speaker 3: how meaningful is the human actually? Like, what's your sense

206
00:10:55,559 --> 00:10:58,360
Speaker 3: of it? Because I'm imagining if you're an intelligence officer

207
00:10:58,400 --> 00:11:01,920
Speaker 3: and you're getting realms and reams data from Iran and

208
00:11:01,960 --> 00:11:04,880
Speaker 3: you ask the AI to pick out certain patterns or

209
00:11:04,960 --> 00:11:09,440
Speaker 3: identify potential strategic targets, how much due diligence are you

210
00:11:09,480 --> 00:11:12,560
Speaker 3: actually doing on what that model spits out. Because of course,

211
00:11:12,640 --> 00:11:16,960
Speaker 3: the tendency when a lot of people use LLM certainly

212
00:11:17,080 --> 00:11:19,440
Speaker 3: is just you accept what it shows you on the screen.

213
00:11:19,559 --> 00:11:22,360
Speaker 2: And just to add on to Tracy's question, because there's

214
00:11:22,440 --> 00:11:24,199
Speaker 2: on to go, which is then in the early.

215
00:11:24,000 --> 00:11:26,120
Speaker 4: Days of the war, we hit that school.

216
00:11:26,360 --> 00:11:29,520
Speaker 2: Yeah, and I'm reading a New York Times report and

217
00:11:29,559 --> 00:11:33,880
Speaker 2: that was the result of quote outdated data provided by

218
00:11:33,880 --> 00:11:37,319
Speaker 2: the Defense Intelligence Agency. Now we don't know exactly what

219
00:11:37,360 --> 00:11:41,120
Speaker 2: that means, but Okay, various outputs come out, then what happens,

220
00:11:41,200 --> 00:11:44,280
Speaker 2: like how much is the human layer currently in terms

221
00:11:44,280 --> 00:11:46,959
Speaker 2: of Okay, here are targets, here are ships that are

222
00:11:46,960 --> 00:11:50,200
Speaker 2: on a battleship. This could be plausible. What do you

223
00:11:50,360 --> 00:11:53,000
Speaker 2: think or what do you know about the level of

224
00:11:53,120 --> 00:11:57,000
Speaker 2: human decision making that happens between some output and then

225
00:11:57,040 --> 00:11:59,760
Speaker 2: the ultimate call for a strike on whatever it is.

226
00:12:00,920 --> 00:12:02,079
Speaker 6: Yeah, I mean, I think, first of all, I think

227
00:12:02,080 --> 00:12:04,480
Speaker 6: it's a really important question because it is one of

228
00:12:04,520 --> 00:12:08,040
Speaker 6: the possible failure modes, if you will, of AI and

229
00:12:08,080 --> 00:12:10,199
Speaker 6: how we use it, because you can end up in

230
00:12:10,240 --> 00:12:13,480
Speaker 6: a place where humans are nominally in the loop and

231
00:12:13,520 --> 00:12:15,800
Speaker 6: you could say, well, it's not an autonous weapon humans

232
00:12:15,840 --> 00:12:18,640
Speaker 6: making these decisions. But if the human is not meaningfully

233
00:12:18,679 --> 00:12:22,040
Speaker 6: engaged and they're just kind of rubber stamping some kind

234
00:12:22,040 --> 00:12:24,520
Speaker 6: of decision, and that's not really what we're looking for.

235
00:12:24,600 --> 00:12:28,040
Speaker 6: So I think that's it's been a longstanding concern for

236
00:12:28,120 --> 00:12:31,319
Speaker 6: many years about people worried about autonous weapons. I think

237
00:12:31,320 --> 00:12:34,400
Speaker 6: that's a very real risk with how AI is used now.

238
00:12:34,440 --> 00:12:37,480
Speaker 6: Based on my understanding of how the AI technology is

239
00:12:37,559 --> 00:12:41,000
Speaker 6: used in Maven today and based on what I've seen

240
00:12:41,040 --> 00:12:43,760
Speaker 6: of demonstrations of it, because I have seen some demonstrations

241
00:12:43,800 --> 00:12:46,800
Speaker 6: of this in action. I think humans are pretty involved

242
00:12:46,880 --> 00:12:50,600
Speaker 6: right now in terms of actually looking at the output

243
00:12:50,679 --> 00:12:54,920
Speaker 6: from AI, giving pretty specific guidance to the AI systems.

244
00:12:55,360 --> 00:12:59,440
Speaker 6: I do think there is an underlying challenge that the

245
00:12:59,559 --> 00:13:02,920
Speaker 6: strike on the school highlights, which is, when you're talking

246
00:13:03,000 --> 00:13:06,640
Speaker 6: about thousands and thousands of targets, what's the degree of

247
00:13:06,800 --> 00:13:11,120
Speaker 6: vetting that's gone into all of that information, both in

248
00:13:11,160 --> 00:13:14,360
Speaker 6: the run up to the war, which in this case,

249
00:13:14,400 --> 00:13:16,839
Speaker 6: that school was a fixed object, and so that's likely

250
00:13:16,960 --> 00:13:20,000
Speaker 6: something that should have clearly been much more vetted prior

251
00:13:20,120 --> 00:13:23,760
Speaker 6: to the we're kicking off. That someone could have identified

252
00:13:24,400 --> 00:13:27,760
Speaker 6: that that building that was struck had actually been at

253
00:13:27,800 --> 00:13:30,120
Speaker 6: one point in time part of an Iranian military compound.

254
00:13:30,480 --> 00:13:33,920
Speaker 6: But you could see based on public available satellite imagery

255
00:13:34,320 --> 00:13:36,000
Speaker 6: that it had been moved out of that compound some

256
00:13:36,080 --> 00:13:38,560
Speaker 6: time ago and it had been converted to a school,

257
00:13:39,000 --> 00:13:41,680
Speaker 6: and it would appear, based on what's been reported in

258
00:13:41,720 --> 00:13:44,880
Speaker 6: the Times, that that information had never been updated in

259
00:13:44,920 --> 00:13:49,079
Speaker 6: this DEIA targeting database. Now, I would hope that we'll

260
00:13:49,120 --> 00:13:51,600
Speaker 6: get more information in the future and some investigation about

261
00:13:51,679 --> 00:13:54,079
Speaker 6: exactly where that went wrong. But I think that does

262
00:13:54,160 --> 00:13:57,200
Speaker 6: speak to this underlying challenge of how good is the

263
00:13:57,360 --> 00:14:00,679
Speaker 6: data going into this AI system and how thoroughly are

264
00:14:00,720 --> 00:14:03,760
Speaker 6: people vetting it? And again, in principle, AI might be

265
00:14:03,840 --> 00:14:05,800
Speaker 6: able to help you with those things, but you got

266
00:14:05,840 --> 00:14:07,520
Speaker 6: to use it the right way and people still have

267
00:14:07,559 --> 00:14:09,520
Speaker 6: to be meaningfully engaging THECC.

268
00:14:25,760 --> 00:14:28,200
Speaker 2: Why don't we back up for a second, tell us

269
00:14:28,200 --> 00:14:31,600
Speaker 2: about the work that you've done in this area, really

270
00:14:31,640 --> 00:14:34,440
Speaker 2: several years ahead of the curve, and talk about this stuff,

271
00:14:34,480 --> 00:14:36,600
Speaker 2: planning for this stuff. Give us a little bit of

272
00:14:36,680 --> 00:14:40,120
Speaker 2: sort of your background and what got you on this

273
00:14:40,240 --> 00:14:43,320
Speaker 2: train again several years before Chad Shept.

274
00:14:44,880 --> 00:14:49,120
Speaker 6: Yeah, so really over a decade ago now, around say

275
00:14:49,440 --> 00:14:53,320
Speaker 6: twenty eleven, I let an effort inside the Pentagon when

276
00:14:53,320 --> 00:14:55,720
Speaker 6: I worked at the Office the Sectary of Defense on

277
00:14:56,360 --> 00:14:59,920
Speaker 6: developing the Pentagon's policy on the role of a ton

278
00:15:00,040 --> 00:15:02,400
Speaker 6: I meet in weapons, the one that's still in effect

279
00:15:02,440 --> 00:15:06,960
Speaker 6: today in fact, and that was really part of at

280
00:15:06,960 --> 00:15:10,120
Speaker 6: the time. We weren't at all where the military is

281
00:15:10,160 --> 00:15:12,760
Speaker 6: now in terms of integrating AI tools. I mean, these

282
00:15:12,800 --> 00:15:15,200
Speaker 6: types of large language models just did exist at the time,

283
00:15:15,920 --> 00:15:19,680
Speaker 6: but the military had kind of woken up to what

284
00:15:19,720 --> 00:15:23,600
Speaker 6: I would call this accidental robotics revolution during the wars

285
00:15:23,600 --> 00:15:26,680
Speaker 6: in the Rock and Afghanistan, where the military had deployed

286
00:15:26,720 --> 00:15:30,720
Speaker 6: thousands of air and ground robots, drones in the air

287
00:15:30,760 --> 00:15:34,160
Speaker 6: and ground robots for diffusing bombs, and the military was

288
00:15:34,200 --> 00:15:36,640
Speaker 6: starting to think through where is this going in the future,

289
00:15:37,160 --> 00:15:39,040
Speaker 6: And one of the things that everyone could see would

290
00:15:39,080 --> 00:15:42,320
Speaker 6: be valuable would be having more autonomy in these systems,

291
00:15:42,360 --> 00:15:45,400
Speaker 6: the ability to not be tortally reliant on a human

292
00:15:45,480 --> 00:15:48,160
Speaker 6: remotely controlling them, which was really the case at the time,

293
00:15:48,760 --> 00:15:51,640
Speaker 6: but that raised all these obviously the any questions about

294
00:15:51,640 --> 00:15:54,960
Speaker 6: like well, how much autonomy should they have and what

295
00:15:55,040 --> 00:15:57,800
Speaker 6: are the legal and ethical implications of that, And that

296
00:15:57,920 --> 00:16:00,600
Speaker 6: was actually the topic of a lot of discussion among

297
00:16:00,720 --> 00:16:04,080
Speaker 6: people in the military at the time and in the

298
00:16:04,080 --> 00:16:06,640
Speaker 6: Pentagon for people working on these issues, and so that

299
00:16:06,760 --> 00:16:10,720
Speaker 6: ultimately led to that policy directive that's still in place

300
00:16:10,720 --> 00:16:13,040
Speaker 6: on the role of autonomy and weapons. And then when

301
00:16:13,040 --> 00:16:15,680
Speaker 6: I left the government, I continued to work on this topic.

302
00:16:15,840 --> 00:16:20,080
Speaker 6: Is we've seen discussions internationally through the United Nations as

303
00:16:20,080 --> 00:16:23,560
Speaker 6: we've seen the technology evolve in really amazing ways but

304
00:16:23,640 --> 00:16:26,160
Speaker 6: also ones out of risks with artificial intelligence.

305
00:16:26,760 --> 00:16:28,760
Speaker 5: So when you were doing that job.

306
00:16:28,800 --> 00:16:31,000
Speaker 3: I get that you're on the policy side, but did

307
00:16:31,000 --> 00:16:34,240
Speaker 3: you ever see anything on the contractor side similar to

308
00:16:34,240 --> 00:16:36,800
Speaker 3: what we're seeing with Anthropic right now. Was there ever

309
00:16:36,840 --> 00:16:40,680
Speaker 3: a contractor who said, actually, no, I'm really uncomfortable with

310
00:16:40,720 --> 00:16:44,160
Speaker 3: the way that the department wants to use this particular tech.

311
00:16:45,040 --> 00:16:48,640
Speaker 6: Not at that time. Now, a few years later, after

312
00:16:48,720 --> 00:16:52,240
Speaker 6: the US military launched Project mavin there was a big

313
00:16:52,480 --> 00:16:55,440
Speaker 6: dust up when it came out publicly that Google had

314
00:16:55,480 --> 00:16:57,280
Speaker 6: been a part of Product Maven, and a number of

315
00:16:57,280 --> 00:17:01,080
Speaker 6: Google employees signed and open letter protesting and Google eventually

316
00:17:01,120 --> 00:17:04,480
Speaker 6: discontinue their work on Project Mavane. And you know, it's

317
00:17:04,520 --> 00:17:08,200
Speaker 6: not exact replica here of what's going on, but there's

318
00:17:08,200 --> 00:17:12,040
Speaker 6: certainly some similarities in terms of a disconnect between how

319
00:17:12,080 --> 00:17:15,000
Speaker 6: some people in the AI community are thinking about how

320
00:17:15,080 --> 00:17:17,439
Speaker 6: their technology ought to be used in war and how

321
00:17:17,480 --> 00:17:19,880
Speaker 6: the military is thinking about it. And I think part

322
00:17:19,880 --> 00:17:23,240
Speaker 6: of that's like there's this underlying challenge of AI is

323
00:17:23,359 --> 00:17:27,000
Speaker 6: really different than a lot of traditional military technologies because

324
00:17:27,000 --> 00:17:29,200
Speaker 6: it's coming out of the commercial sector. So in a way,

325
00:17:29,200 --> 00:17:31,800
Speaker 6: it's kind of like the opposite of stealth technology that

326
00:17:31,960 --> 00:17:34,600
Speaker 6: was invented in secret defense labs and doesn't have a

327
00:17:34,640 --> 00:17:38,520
Speaker 6: lot of commercial applications. AI is all of these different applications.

328
00:17:38,560 --> 00:17:41,280
Speaker 6: It's not being invented by the military. The military is

329
00:17:41,280 --> 00:17:44,240
Speaker 6: having to be import it in and there are a

330
00:17:44,280 --> 00:17:48,399
Speaker 6: lot of debates about how AI should be used, you know,

331
00:17:48,440 --> 00:17:50,320
Speaker 6: in the military and more broadly in society.

332
00:17:50,520 --> 00:17:53,920
Speaker 3: Actually, on that note, I think this is really interesting

333
00:17:54,000 --> 00:17:56,320
Speaker 3: and definitely a pivotal point in I guess, the history

334
00:17:56,320 --> 00:18:00,360
Speaker 3: of the military industrial complex. But why can't the US government,

335
00:18:00,600 --> 00:18:04,399
Speaker 3: with all its resources, actually develop AI in house and

336
00:18:04,480 --> 00:18:07,719
Speaker 3: just avoid the seeming complication of having to deal with

337
00:18:07,760 --> 00:18:08,800
Speaker 3: a commercial enterprise.

338
00:18:09,520 --> 00:18:13,840
Speaker 6: Part ofly, it doesn't have the technical skills the AI

339
00:18:13,920 --> 00:18:19,399
Speaker 6: scientists and engineers are. Really there's a fierce competition for

340
00:18:19,440 --> 00:18:22,359
Speaker 6: talent in the AI space, and so the military just

341
00:18:22,359 --> 00:18:25,400
Speaker 6: like can't buy that talent they don't have it. And

342
00:18:25,600 --> 00:18:27,720
Speaker 6: the government spends a lot of money, hundreds of billions

343
00:18:27,760 --> 00:18:31,560
Speaker 6: of dollars annually on defense. But we've seen actually in

344
00:18:31,560 --> 00:18:34,399
Speaker 6: the last few years that private enterprise is able to

345
00:18:34,400 --> 00:18:38,040
Speaker 6: mobile massive amounts of capital towards building data centers, to

346
00:18:38,320 --> 00:18:42,720
Speaker 6: training AI models. And partly because the commercial applications for

347
00:18:42,720 --> 00:18:47,159
Speaker 6: this technology are much bigger than the defense applications, and

348
00:18:47,240 --> 00:18:50,639
Speaker 6: so for a lot of these tech companies there's some

349
00:18:51,400 --> 00:18:54,280
Speaker 6: at least maybe not in this particular instance, but in

350
00:18:54,280 --> 00:18:57,200
Speaker 6: the past there could offer be some prestige associated with saying, oh,

351
00:18:57,520 --> 00:19:01,280
Speaker 6: you know, the Air Force uses our AI system or

352
00:19:01,320 --> 00:19:04,680
Speaker 6: the Navy users our technology, but the defense sector is

353
00:19:04,680 --> 00:19:07,320
Speaker 6: actually kind of small for them as a customer. I mean,

354
00:19:07,359 --> 00:19:10,000
Speaker 6: the dollar amount that's been talked about publicly for the

355
00:19:10,000 --> 00:19:12,840
Speaker 6: Anthropic contract, it's two hundred million dollars. That's not a

356
00:19:12,880 --> 00:19:15,320
Speaker 6: lot of money for these AI companies. Yeah, and so

357
00:19:15,400 --> 00:19:18,640
Speaker 6: I think that actually we've seen the defense sectors struggle

358
00:19:18,720 --> 00:19:21,480
Speaker 6: to just keep pace with the amount of investment that's

359
00:19:21,480 --> 00:19:22,280
Speaker 6: needed in the space.

360
00:19:22,960 --> 00:19:26,040
Speaker 2: Cherioucy, I think it's a good question. And then you remember, well,

361
00:19:26,200 --> 00:19:29,600
Speaker 2: the government couldn't build a good healthcare website to sign up.

362
00:19:29,560 --> 00:19:30,400
Speaker 4: For health insurance.

363
00:19:30,640 --> 00:19:32,600
Speaker 2: And I hate to bring that up because it's old,

364
00:19:33,000 --> 00:19:34,960
Speaker 2: but it's true. Right, So it's like, are they gonna

365
00:19:35,000 --> 00:19:38,399
Speaker 2: build a world class LLM or can a government build

366
00:19:38,440 --> 00:19:42,720
Speaker 2: a good employment insurance website. We've done multiple episodes the

367
00:19:43,160 --> 00:19:45,680
Speaker 2: answer continues to be not the case. I do find

368
00:19:45,680 --> 00:19:49,680
Speaker 2: it fascinating. However, your point about there is this novelty.

369
00:19:50,200 --> 00:19:53,960
Speaker 2: It is impossible to imagine and say Lockheed Martin inventing

370
00:19:54,000 --> 00:19:56,000
Speaker 2: get technology and they's like, no, you can't use it.

371
00:19:56,040 --> 00:19:59,840
Speaker 2: Because Lockheed Martin's entire raise on death Right is building.

372
00:19:59,600 --> 00:20:00,639
Speaker 4: Technolog for the government.

373
00:20:00,680 --> 00:20:03,639
Speaker 2: It is inconceivable what that would be. But it is

374
00:20:03,920 --> 00:20:07,720
Speaker 2: sort of novel when you're getting these defend technologies. And

375
00:20:07,800 --> 00:20:09,920
Speaker 2: you know the Google was also an example of Google

376
00:20:09,920 --> 00:20:13,840
Speaker 2: obviously had technology that did not originally serve a purpose

377
00:20:14,080 --> 00:20:17,960
Speaker 2: of defense. We saw the we remember the employee revolt.

378
00:20:18,240 --> 00:20:22,359
Speaker 2: Let's talk more about that disagreement though, between Anthropic and

379
00:20:22,440 --> 00:20:26,360
Speaker 2: the Department of Defense. In your mind, where does Peter

380
00:20:26,800 --> 00:20:30,760
Speaker 2: heg Seth want to go with this technology? And is

381
00:20:30,800 --> 00:20:34,080
Speaker 2: that deviate from some of the policies and the directives

382
00:20:34,080 --> 00:20:36,919
Speaker 2: that you were working on when you were when you

383
00:20:36,960 --> 00:20:37,879
Speaker 2: were working on this stuff.

384
00:20:38,720 --> 00:20:42,439
Speaker 6: So it's kind of crazy about this whole dispute is

385
00:20:42,440 --> 00:20:46,360
Speaker 6: particularly an issue of autonomous weapons. Literally everyone I've spoken

386
00:20:46,440 --> 00:20:50,000
Speaker 6: with has said that there's no intention by the military

387
00:20:50,359 --> 00:20:53,800
Speaker 6: to use AI to make fully autonomous weapons today. Okay,

388
00:20:53,840 --> 00:20:56,280
Speaker 6: I think anybody that's actually worked with a large language

389
00:20:56,280 --> 00:20:58,840
Speaker 6: model with any kind of chat bought, whether it's Claude

390
00:20:58,920 --> 00:21:02,600
Speaker 6: or Gemini or chat Chie knows that if you use

391
00:21:02,640 --> 00:21:05,159
Speaker 6: these to write an email, you need to double check it. Like,

392
00:21:05,200 --> 00:21:08,200
Speaker 6: in no way, shape or form are they reliable enough

393
00:21:08,240 --> 00:21:10,520
Speaker 6: to make life and death decisions. I don't think the

394
00:21:10,520 --> 00:21:14,160
Speaker 6: military actually wants to do that. What's at dispute here

395
00:21:14,240 --> 00:21:17,520
Speaker 6: is a more fundamental disagreement about well, who sets the

396
00:21:17,640 --> 00:21:20,840
Speaker 6: rules and so the origins of this really was that

397
00:21:21,240 --> 00:21:23,960
Speaker 6: when the Pentagon came out with a new strategy for

398
00:21:24,040 --> 00:21:27,240
Speaker 6: AI in January, one of the things in their strategy

399
00:21:27,320 --> 00:21:30,320
Speaker 6: was that going forward, they wanted their contracts with AI

400
00:21:30,359 --> 00:21:33,800
Speaker 6: companies to allow the military to use their AD tools

401
00:21:33,800 --> 00:21:37,760
Speaker 6: for any lawful use. Basically, look, anything that's legal, we

402
00:21:37,760 --> 00:21:39,880
Speaker 6: want the ability to do it. And that has conflicted

403
00:21:39,920 --> 00:21:41,439
Speaker 6: with how a lot of these tech companies have been

404
00:21:41,440 --> 00:21:44,320
Speaker 6: thinking about their AI tools. They're very nervous, many of

405
00:21:44,320 --> 00:21:47,880
Speaker 6: these companies about harms from AI. They're conscious of these risks,

406
00:21:48,080 --> 00:21:50,000
Speaker 6: and so a lot of them have various use policies

407
00:21:50,080 --> 00:21:52,879
Speaker 6: in place. You can't use AI, you know, to launch

408
00:21:52,920 --> 00:21:56,200
Speaker 6: offensive cyber attacks for example. That's kind of thing that

409
00:21:56,359 --> 00:21:59,159
Speaker 6: actually like the government might want to do. And so

410
00:21:59,840 --> 00:22:02,480
Speaker 6: this is a This was really the rub with the

411
00:22:02,480 --> 00:22:06,440
Speaker 6: government was like who sets the rules rather than necessarily

412
00:22:06,640 --> 00:22:09,800
Speaker 6: like a near term question of fully autonomous weapons.

413
00:22:09,960 --> 00:22:13,600
Speaker 3: So what we've already seen is Anthropic has this disagreement

414
00:22:13,920 --> 00:22:18,439
Speaker 3: with the government, and then OpenAI steps in and raises

415
00:22:18,480 --> 00:22:20,639
Speaker 3: its hand and says, okay, Anthropic doesn't want to do it,

416
00:22:20,640 --> 00:22:23,480
Speaker 3: We'll do it happily. Does this just leave us in

417
00:22:23,520 --> 00:22:26,680
Speaker 3: a situation where it's sort of a race to the bottom, Right,

418
00:22:26,720 --> 00:22:29,840
Speaker 3: It's like the lab with maybe the least amount of

419
00:22:29,920 --> 00:22:34,080
Speaker 3: safety concern or the least amount of reputational concern is

420
00:22:34,119 --> 00:22:36,520
Speaker 3: able to do this, and so we still wind up

421
00:22:36,560 --> 00:22:39,399
Speaker 3: in a situation where the government is using AI.

422
00:22:40,680 --> 00:22:43,280
Speaker 6: Well, I think what's unfortunate here is that when you

423
00:22:43,280 --> 00:22:46,320
Speaker 6: think about would be optimal for the government one, I

424
00:22:46,359 --> 00:22:47,959
Speaker 6: think it would be ideal for the government to have

425
00:22:47,960 --> 00:22:51,600
Speaker 6: access to this technology and to have access to all

426
00:22:51,680 --> 00:22:55,080
Speaker 6: of the best in class models available, because they are

427
00:22:55,160 --> 00:22:58,920
Speaker 6: good at slightly different things sometimes, and it's much healthier

428
00:22:58,920 --> 00:23:01,359
Speaker 6: for the government to have access to a number of

429
00:23:01,400 --> 00:23:04,760
Speaker 6: different providers so that there is healthy competition in the market.

430
00:23:04,800 --> 00:23:07,120
Speaker 6: You don't get locked in with one vendor. But also,

431
00:23:07,720 --> 00:23:10,800
Speaker 6: if the AI scientists are saying, hey, it's not reliable

432
00:23:10,800 --> 00:23:13,600
Speaker 6: for this, you to listen like that seems like a

433
00:23:13,680 --> 00:23:16,520
Speaker 6: thing you want to hear them like out about right,

434
00:23:16,680 --> 00:23:19,560
Speaker 6: And so I think like in order to use AI

435
00:23:19,600 --> 00:23:23,400
Speaker 6: and ways that actually are effective for the US military,

436
00:23:23,640 --> 00:23:26,680
Speaker 6: we've got to have a healthy dialogue between the AI

437
00:23:26,760 --> 00:23:29,639
Speaker 6: community and people in the military profession. But what the

438
00:23:29,680 --> 00:23:33,919
Speaker 6: technology can and cannot do. And I think it's unfortunate

439
00:23:33,960 --> 00:23:35,960
Speaker 6: that we've seen that dialogue break down in such a

440
00:23:36,000 --> 00:23:37,320
Speaker 6: dramatic way over the speed.

441
00:23:37,880 --> 00:23:40,359
Speaker 3: Just going back to the idea, who actually makes the

442
00:23:40,440 --> 00:23:43,920
Speaker 3: rules you mentioned earlier that you know you can't use

443
00:23:44,400 --> 00:23:47,600
Speaker 3: claude to hack into to illegally hack into a system.

444
00:23:47,600 --> 00:23:49,879
Speaker 3: Supposedly it is unable to do that. It has like

445
00:23:49,920 --> 00:23:52,760
Speaker 3: a kill switch within itself that prevents it from doing that.

446
00:23:53,000 --> 00:23:55,560
Speaker 3: If you're anthropic, could you not just hardcode some of

447
00:23:55,600 --> 00:23:58,320
Speaker 3: these systems and say you're not going to be able

448
00:23:58,320 --> 00:24:03,080
Speaker 3: to be used for domestic surveillance of Americans or for

449
00:24:03,359 --> 00:24:04,120
Speaker 3: war crimes.

450
00:24:05,000 --> 00:24:06,879
Speaker 6: So yeah, So this is where it gets a little

451
00:24:07,000 --> 00:24:09,440
Speaker 6: more technical. It has to do with some of the

452
00:24:09,480 --> 00:24:13,080
Speaker 6: ways in which the companies may be providing their technology

453
00:24:13,119 --> 00:24:15,160
Speaker 6: to the government. So there's a couple of different ways

454
00:24:15,160 --> 00:24:17,439
Speaker 6: in which an AI company can put safeguards in place

455
00:24:17,800 --> 00:24:21,320
Speaker 6: to make sure that their models not being abused. One

456
00:24:21,400 --> 00:24:24,639
Speaker 6: is training the model itself to refuse certain requests. So

457
00:24:24,680 --> 00:24:26,639
Speaker 6: if you ask the model to do something, it's just

458
00:24:26,640 --> 00:24:28,280
Speaker 6: going to say, like, I'm not going to do that.

459
00:24:28,280 --> 00:24:30,640
Speaker 6: That's not consistent with the guidance that I've been given,

460
00:24:30,640 --> 00:24:32,680
Speaker 6: and the model has been trained to do that response.

461
00:24:33,480 --> 00:24:37,800
Speaker 6: Another way is that the company can put classifiers on

462
00:24:37,840 --> 00:24:41,160
Speaker 6: the input and or the output of a model, where

463
00:24:41,800 --> 00:24:44,400
Speaker 6: the model might give you an answer, but then there's

464
00:24:44,440 --> 00:24:47,320
Speaker 6: like another aisystem that's checking that answer or checking what

465
00:24:47,359 --> 00:24:50,000
Speaker 6: you ask of it and saying, well, that's not acceptable.

466
00:24:50,280 --> 00:24:52,800
Speaker 6: And then a third and I've want into that actually

467
00:24:52,840 --> 00:24:55,240
Speaker 6: myself in my own research because the nature of the

468
00:24:55,240 --> 00:24:57,520
Speaker 6: things that I work on a security things, and I've

469
00:24:57,520 --> 00:25:01,480
Speaker 6: had situations where I asked Claude help me understand this issue.

470
00:25:01,640 --> 00:25:05,120
Speaker 6: Claud actually generates response and then it gets deleted. It's

471
00:25:05,119 --> 00:25:07,439
Speaker 6: really interesting to see. And then the other way an

472
00:25:07,440 --> 00:25:10,840
Speaker 6: anthropic has actually talked about this in response to countering

473
00:25:11,359 --> 00:25:15,960
Speaker 6: some use of Claude by Chinese hackers who are using

474
00:25:15,960 --> 00:25:21,720
Speaker 6: it for cyber attacks, is the company monitors use through

475
00:25:22,280 --> 00:25:25,320
Speaker 6: that people are doing, and so people are doing things

476
00:25:25,359 --> 00:25:28,080
Speaker 6: that seem suspicious. Maybe they're logging in from an IP

477
00:25:28,200 --> 00:25:30,920
Speaker 6: address that's known to be associated with cyber criminals or

478
00:25:30,960 --> 00:25:33,840
Speaker 6: a hacking group that to try and find ways to

479
00:25:33,880 --> 00:25:36,560
Speaker 6: get around some of these protections. The company can also

480
00:25:36,600 --> 00:25:38,440
Speaker 6: find ways to try to catch that, and so there's

481
00:25:38,440 --> 00:25:41,040
Speaker 6: a couple different ways to do it that might not

482
00:25:41,240 --> 00:25:45,000
Speaker 6: all be in place. If you're thinking about military use

483
00:25:45,040 --> 00:25:48,600
Speaker 6: where if depending on how that relationship is structure between

484
00:25:48,600 --> 00:25:51,320
Speaker 6: the company. If the model is, for example, that hosted

485
00:25:51,520 --> 00:25:55,040
Speaker 6: on a different cloud infrastructure and the military is direct

486
00:25:55,040 --> 00:25:57,359
Speaker 6: access to it, the company may not have the same

487
00:25:57,600 --> 00:26:01,879
Speaker 6: ways to actually shape what or not the technology is

488
00:26:01,920 --> 00:26:05,040
Speaker 6: being used according to their principles, which is partly why

489
00:26:05,040 --> 00:26:07,560
Speaker 6: the contract details do matter of like what is the

490
00:26:07,600 --> 00:26:10,640
Speaker 6: agreement between the company the government about what the military

491
00:26:10,640 --> 00:26:28,520
Speaker 6: can and cannot use the technology trace.

492
00:26:28,600 --> 00:26:31,040
Speaker 2: I think your point about like this sort of seemingly

493
00:26:31,680 --> 00:26:34,560
Speaker 2: safety or safety race at the bottom is very real

494
00:26:34,600 --> 00:26:36,640
Speaker 2: and it's one that I think about a lot. When

495
00:26:37,400 --> 00:26:41,760
Speaker 2: LMS or AI was basically just synonymous with open AI,

496
00:26:42,160 --> 00:26:44,080
Speaker 2: they could set the pace of development.

497
00:26:43,760 --> 00:26:44,600
Speaker 4: Right, they could do it.

498
00:26:44,800 --> 00:26:47,960
Speaker 2: As soon as this became a hyper competitive space where

499
00:26:47,960 --> 00:26:50,239
Speaker 2: you have open AI and you have anthropic, and you

500
00:26:50,240 --> 00:26:54,120
Speaker 2: have Gemini and a thousand open source AI models out

501
00:26:54,160 --> 00:26:58,560
Speaker 2: of China, et cetera. The tempo of release has really heightened,

502
00:26:58,600 --> 00:27:01,680
Speaker 2: and the degree which it feels like they have no

503
00:27:01,800 --> 00:27:05,560
Speaker 2: choice but to accelerate just for the commercial imperative feels

504
00:27:05,560 --> 00:27:07,960
Speaker 2: like a very real dynamic in which like, I don't

505
00:27:08,000 --> 00:27:10,600
Speaker 2: know where that leaves AI safety well totally.

506
00:27:10,640 --> 00:27:13,560
Speaker 3: And also you mentioned China. Then it's not just domestic

507
00:27:13,600 --> 00:27:17,119
Speaker 3: competition became, you know, open AI versus anthropic, it's a

508
00:27:17,160 --> 00:27:20,000
Speaker 3: competition between international actors where it's like, Okay, well, the

509
00:27:20,080 --> 00:27:22,800
Speaker 3: US might want to have safeguards on technology or say

510
00:27:22,800 --> 00:27:25,760
Speaker 3: that it does, but maybe Russia or China don't care.

511
00:27:26,080 --> 00:27:28,160
Speaker 2: Right tell you you know, it's funny all you mentioned

512
00:27:28,200 --> 00:27:30,280
Speaker 2: where you like see the output for one second and

513
00:27:30,280 --> 00:27:33,560
Speaker 2: then deletes, Like when deep Seak came out doing some

514
00:27:33,680 --> 00:27:37,679
Speaker 2: experiments about like figuring out its censorship and I was

515
00:27:37,720 --> 00:27:40,600
Speaker 2: trying to do some adversarial prompting and I was like,

516
00:27:41,080 --> 00:27:43,880
Speaker 2: historians like to talk about a period in the twentieth

517
00:27:43,880 --> 00:27:48,480
Speaker 2: century where it failed. Attempt at extreme rapid industrialization happened

518
00:27:48,480 --> 00:27:50,399
Speaker 2: and it led to famine. And then you see the

519
00:27:50,480 --> 00:27:53,760
Speaker 2: output and it said Okay, what happened in the twentieth century?

520
00:27:53,960 --> 00:27:56,280
Speaker 2: Where did this famine? Well, there's something called the Great

521
00:27:56,359 --> 00:27:59,040
Speaker 2: Leap Forward, and then immediately it just as soon as

522
00:27:59,080 --> 00:27:59,640
Speaker 2: the chain of.

523
00:27:59,560 --> 00:28:01,840
Speaker 4: Thought hit the Great Leap Forward, it just like disappeared.

524
00:28:01,880 --> 00:28:04,159
Speaker 2: So I'm always like very amused by like when the

525
00:28:04,240 --> 00:28:07,439
Speaker 2: system recognizes that the system has gone too far. Anyway,

526
00:28:08,080 --> 00:28:12,480
Speaker 2: we've been talking about quote large language models, but actually

527
00:28:12,520 --> 00:28:15,760
Speaker 2: AI is beyond large language models, including the image stuff,

528
00:28:15,760 --> 00:28:18,520
Speaker 2: and that actually lllm's at this point. It's a very

529
00:28:18,560 --> 00:28:22,040
Speaker 2: twenty twenty three sort of term. And when and I

530
00:28:22,040 --> 00:28:24,000
Speaker 2: think this is important because when we get to the

531
00:28:24,040 --> 00:28:28,720
Speaker 2: intersection of AI and robotics and so forth, or AI

532
00:28:28,920 --> 00:28:32,119
Speaker 2: and target, we're talking about something but beyond large language model,

533
00:28:32,320 --> 00:28:34,640
Speaker 2: but we might still be talking about generative AI.

534
00:28:35,200 --> 00:28:36,719
Speaker 4: Where do you see this going?

535
00:28:36,960 --> 00:28:41,720
Speaker 2: And what are the weapon systems that aren't currently You said, currently,

536
00:28:41,760 --> 00:28:45,520
Speaker 2: no one is actually talking about true autonomous weapons. But

537
00:28:45,560 --> 00:28:47,840
Speaker 2: if that were the case, then there wouldn't be a controversy.

538
00:28:47,880 --> 00:28:50,000
Speaker 2: There was, so there was clearly something just beyond the

539
00:28:50,000 --> 00:28:52,920
Speaker 2: horizon that could come into the picture of a true

540
00:28:52,960 --> 00:28:56,200
Speaker 2: autonomous weapon system where the technology is building towards that.

541
00:28:56,680 --> 00:28:58,400
Speaker 2: If this weren't the case, there would be no dispute,

542
00:28:58,440 --> 00:29:01,240
Speaker 2: you wouldn't have two books written about this subject. So

543
00:29:01,280 --> 00:29:05,200
Speaker 2: what are these weapons systems that would classify as autonomous

544
00:29:05,240 --> 00:29:09,120
Speaker 2: weapons today that the technology is building towards right now?

545
00:29:10,480 --> 00:29:13,040
Speaker 6: Yeah, I mean I certainly the trends are taking us there.

546
00:29:13,040 --> 00:29:14,640
Speaker 6: And one of the things that you see in the

547
00:29:14,680 --> 00:29:17,120
Speaker 6: Pentagon's position in this issue, for example, is they want

548
00:29:17,120 --> 00:29:19,920
Speaker 6: to preserve that option. Yeah, for there's certainly not interested

549
00:29:19,920 --> 00:29:22,360
Speaker 6: in tying their hands. I think you could see that

550
00:29:22,480 --> 00:29:26,160
Speaker 6: evolving in a couple ways. One trend we're clearly seeing

551
00:29:26,280 --> 00:29:30,280
Speaker 6: with the largest and most capable AI systems is they're

552
00:29:30,320 --> 00:29:33,440
Speaker 6: increasingly multimodal. They're bringing in lots of different forms of data,

553
00:29:33,480 --> 00:29:35,760
Speaker 6: of course, and they're increasing the general purpose they could

554
00:29:35,760 --> 00:29:38,080
Speaker 6: just like do a variety of different kinds of things.

555
00:29:38,080 --> 00:29:40,440
Speaker 6: They become more capable of that, and so that's like

556
00:29:40,600 --> 00:29:45,680
Speaker 6: one way in which you could see AI being used

557
00:29:46,120 --> 00:29:50,040
Speaker 6: in ways that might sort of slowly pulls humans out

558
00:29:50,040 --> 00:29:53,120
Speaker 6: of the loop, where instead of person giving an AI

559
00:29:53,240 --> 00:29:56,680
Speaker 6: like really narrow tasks to do in a planning process,

560
00:29:56,880 --> 00:29:58,720
Speaker 6: maybe the AI is able to take on more, bring

561
00:29:58,840 --> 00:30:02,280
Speaker 6: in more data, take on more like sophisticated longer term tasks,

562
00:30:02,280 --> 00:30:04,320
Speaker 6: and we're certainly seeing this in other areas like coding,

563
00:30:04,720 --> 00:30:07,000
Speaker 6: where the task length that an AI system could do

564
00:30:07,120 --> 00:30:11,080
Speaker 6: is growing exponentially over time. Another sort of way that

565
00:30:11,120 --> 00:30:13,480
Speaker 6: we might see this look is just we see a

566
00:30:13,520 --> 00:30:17,440
Speaker 6: network of AI agents that are interacting with different pieces

567
00:30:17,480 --> 00:30:20,560
Speaker 6: of data, doing different types of things, and that the

568
00:30:20,600 --> 00:30:24,200
Speaker 6: net effect of that is that maybe humans are again

569
00:30:24,240 --> 00:30:27,160
Speaker 6: sort of like nominally looking at these targets but not

570
00:30:27,200 --> 00:30:30,560
Speaker 6: actually approving them in some meaningful way. And then there's

571
00:30:30,640 --> 00:30:32,400
Speaker 6: like a more separate I would almost think of like

572
00:30:32,400 --> 00:30:35,680
Speaker 6: an embodied form of AI in robotics, which could be

573
00:30:35,720 --> 00:30:40,200
Speaker 6: a drone or munition or robotic system that has some

574
00:30:40,280 --> 00:30:45,040
Speaker 6: kind of onboard autonomy that might be partly a distilled

575
00:30:45,280 --> 00:30:47,480
Speaker 6: model so that it can be operating at the edge

576
00:30:47,520 --> 00:30:51,680
Speaker 6: on lower computing on this actual munition or drone, or

577
00:30:51,800 --> 00:30:54,720
Speaker 6: it might be some hybrid system that has partly machine

578
00:30:54,760 --> 00:30:57,920
Speaker 6: learning but also just a lot of hand coded code.

579
00:30:58,240 --> 00:31:01,240
Speaker 6: It's more like expert level system that's going out into

580
00:31:01,320 --> 00:31:04,560
Speaker 6: the battle space and hunting targets directly and attacking them.

581
00:31:04,640 --> 00:31:06,840
Speaker 6: So something kind of like the low cost drones that

582
00:31:06,840 --> 00:31:10,760
Speaker 6: we're seeing Iran launch, but once the can loiter and

583
00:31:11,040 --> 00:31:12,720
Speaker 6: identify targets that attack them.

584
00:31:12,880 --> 00:31:16,440
Speaker 2: Whether that doesn't exist today. We don't have drones loiter

585
00:31:16,920 --> 00:31:19,959
Speaker 2: that are just hanging out there. And then when something flat,

586
00:31:20,080 --> 00:31:21,920
Speaker 2: then there's a system is like this looks like a

587
00:31:21,960 --> 00:31:24,920
Speaker 2: target attacks That actually doesn't exist currently as far as you.

588
00:31:24,880 --> 00:31:29,760
Speaker 6: Know, well, I mean, they're not in widespread use. So

589
00:31:29,800 --> 00:31:33,200
Speaker 6: there have been some narrow examples, I would say historically

590
00:31:33,480 --> 00:31:38,000
Speaker 6: dating back to the eighties, in fact of lording munitions

591
00:31:38,400 --> 00:31:41,320
Speaker 6: that could search over wider and would queue off of radars,

592
00:31:41,720 --> 00:31:44,600
Speaker 6: and so radars are what the military would call a

593
00:31:44,640 --> 00:31:48,360
Speaker 6: cooperative target that when they're emitting in the electromaenetic spectrum,

594
00:31:48,680 --> 00:31:50,680
Speaker 6: if you know the signature of the radar you're looking for,

595
00:31:50,840 --> 00:31:52,120
Speaker 6: you could see it. You can just home in on

596
00:31:52,120 --> 00:31:54,600
Speaker 6: that radar. Now, if they turn off, that's different than

597
00:31:54,920 --> 00:31:57,480
Speaker 6: kidden and they're harder to find. But there have been

598
00:31:57,520 --> 00:32:01,440
Speaker 6: some examples. A system that the US Navy had in

599
00:32:01,440 --> 00:32:04,560
Speaker 6: the eighties called the Tomahawk anti shit missile, not actually

600
00:32:04,640 --> 00:32:07,480
Speaker 6: the same Tomahawk cruise missile that the military is using now,

601
00:32:07,720 --> 00:32:10,400
Speaker 6: a different one that was designed to fly a search

602
00:32:10,480 --> 00:32:13,760
Speaker 6: pattern and hunt Soviet chips. There was an Israeli system

603
00:32:13,760 --> 00:32:16,120
Speaker 6: called the Harpy drone that was designed to go after

604
00:32:16,240 --> 00:32:18,440
Speaker 6: radars that would loorder for a period of time. But

605
00:32:18,520 --> 00:32:21,479
Speaker 6: these lording musicians have never really been in widespread use

606
00:32:21,520 --> 00:32:22,280
Speaker 6: by militaries.

607
00:32:22,440 --> 00:32:24,560
Speaker 3: We got to invent one of those, like high pitched

608
00:32:24,640 --> 00:32:28,800
Speaker 3: alarms to deter the loitering drones from hanging out outside targets.

609
00:32:28,800 --> 00:32:30,479
Speaker 3: I guess I mean we have electrical jammings.

610
00:32:30,560 --> 00:32:31,400
Speaker 5: Yeah, that does exist.

611
00:32:31,720 --> 00:32:34,480
Speaker 3: Okay, So when I think about as we move towards

612
00:32:34,640 --> 00:32:39,080
Speaker 3: more autonomous weaponry, I think about bots basically interacting with

613
00:32:39,160 --> 00:32:41,560
Speaker 3: bots at that point, and then I think back to

614
00:32:41,640 --> 00:32:44,800
Speaker 3: previous examples of bots interacting with bots, and there are

615
00:32:45,160 --> 00:32:47,640
Speaker 3: numerous ones where things tend to go off the rails.

616
00:32:47,720 --> 00:32:49,640
Speaker 4: They just start debating the meaning of life, right or.

617
00:32:49,640 --> 00:32:52,080
Speaker 3: They start talking in like a language that no one

618
00:32:52,160 --> 00:32:53,680
Speaker 3: understands accept them.

619
00:32:53,520 --> 00:32:54,200
Speaker 5: Stuff like that.

620
00:32:54,840 --> 00:33:01,120
Speaker 3: Does the possibility of undesired escalation go up the more

621
00:33:01,200 --> 00:33:03,480
Speaker 3: we move towards fully autonomous weaponry.

622
00:33:05,280 --> 00:33:08,080
Speaker 6: I think that is a very serious risk. And so

623
00:33:08,800 --> 00:33:11,800
Speaker 6: the mental model that I have for this are things

624
00:33:11,840 --> 00:33:16,000
Speaker 6: like flash crashes that we've seen in financial markets due

625
00:33:16,000 --> 00:33:20,320
Speaker 6: to the interactions of different algorithms that are executing trades

626
00:33:20,640 --> 00:33:23,400
Speaker 6: where you get these emergent properties of how the algorithms

627
00:33:23,440 --> 00:33:26,680
Speaker 6: might interact in the market, and it's a competitive environment.

628
00:33:27,000 --> 00:33:28,800
Speaker 6: Companies aren't going to share the details of what their

629
00:33:28,800 --> 00:33:32,040
Speaker 6: algorithms are doing, and you can do strange behaviors. Now,

630
00:33:32,080 --> 00:33:34,960
Speaker 6: the way that financial markets have dealt with this problem

631
00:33:35,120 --> 00:33:38,120
Speaker 6: is regulators have installed circuit breakers that take stocks offline

632
00:33:38,120 --> 00:33:40,560
Speaker 6: if the price moves too quickly. There's no referee to

633
00:33:40,560 --> 00:33:42,800
Speaker 6: call time out in war, and so I think that's like,

634
00:33:42,880 --> 00:33:46,520
Speaker 6: particularly in cyberspace, one can envision a future where that

635
00:33:46,680 --> 00:33:48,840
Speaker 6: is a risk, where things are happening at machine speed

636
00:33:49,400 --> 00:33:53,720
Speaker 6: and you have autonomous offensive cyber operations. You need to

637
00:33:53,760 --> 00:33:57,239
Speaker 6: defend against that. You need some measure of autonomy on

638
00:33:57,280 --> 00:34:00,840
Speaker 6: the defensive side to defend it. Machines space need and

639
00:34:00,880 --> 00:34:04,080
Speaker 6: you could get situations where you get weird interactions that

640
00:34:04,160 --> 00:34:06,840
Speaker 6: might escalate the conflict, or it could also happen between

641
00:34:06,920 --> 00:34:11,640
Speaker 6: drones interacting in some kind of crisis situation. Now, a

642
00:34:11,640 --> 00:34:14,000
Speaker 6: situation where like there's a big shooting war in the

643
00:34:14,080 --> 00:34:16,640
Speaker 6: way people are already attacking, there might be less of

644
00:34:16,680 --> 00:34:20,320
Speaker 6: a concern, although you still could worry about escalation geographically

645
00:34:20,480 --> 00:34:23,600
Speaker 6: against bringing new countries into a conflict, or maybe attacking

646
00:34:24,040 --> 00:34:27,200
Speaker 6: really sensitive sites that are tied to nuclear command and

647
00:34:27,200 --> 00:34:29,480
Speaker 6: control that you'd rather not go after. So I think

648
00:34:29,480 --> 00:34:31,520
Speaker 6: that's a very real risk when we think about how

649
00:34:31,560 --> 00:34:33,320
Speaker 6: this technology might be employed going forward.

650
00:34:33,800 --> 00:34:39,719
Speaker 2: What about AI in really difficult ethical questions strikes where

651
00:34:39,719 --> 00:34:42,399
Speaker 2: we know that civilians, for example, are going to be killed,

652
00:34:42,440 --> 00:34:45,440
Speaker 2: which that happens all the time in war and presumably

653
00:34:45,840 --> 00:34:50,080
Speaker 2: tries to be minimized, but war planners will find some

654
00:34:51,040 --> 00:34:56,040
Speaker 2: level of acceptable they called collateral damage. Is AI playing

655
00:34:56,160 --> 00:34:58,799
Speaker 2: a role or do you expect it to play a

656
00:34:59,000 --> 00:35:02,120
Speaker 2: role in some of these strikes?

657
00:35:02,160 --> 00:35:03,360
Speaker 4: That may be great areas.

658
00:35:04,400 --> 00:35:06,400
Speaker 6: I think you could envision in ways that a I

659
00:35:06,440 --> 00:35:09,160
Speaker 6: would be used that would make warfare more precise and

660
00:35:09,200 --> 00:35:11,200
Speaker 6: more human and ethical, and ways that it could be

661
00:35:11,280 --> 00:35:13,239
Speaker 6: used that would not and what can be opposite. So,

662
00:35:13,280 --> 00:35:17,840
Speaker 6: for example, if you had an AI system that could

663
00:35:17,880 --> 00:35:21,719
Speaker 6: look over all this targeting data and then identify if

664
00:35:21,719 --> 00:35:24,759
Speaker 6: a strike is within a certain distance using you know,

665
00:35:24,880 --> 00:35:29,080
Speaker 6: musis of a certain size of protected targets, whether it's

666
00:35:29,120 --> 00:35:34,319
Speaker 6: schools or hospitals or critical civilian infrastructure, and say hey, well,

667
00:35:34,560 --> 00:35:38,120
Speaker 6: like warning here, you should not carry out the strike

668
00:35:38,320 --> 00:35:41,800
Speaker 6: or it's a higher level of approval, or maybe you

669
00:35:41,800 --> 00:35:44,520
Speaker 6: should use small or more precise munitions. That would be

670
00:35:44,520 --> 00:35:47,239
Speaker 6: a really beneficial use of AI, particularly when you're talking

671
00:35:47,239 --> 00:35:50,080
Speaker 6: about a military campaign that hits a lot of targets

672
00:35:50,080 --> 00:35:52,080
Speaker 6: in a short period of time, That could be really

673
00:35:52,120 --> 00:35:56,200
Speaker 6: valuable and may reduce civilian casualties. You know, the risk

674
00:35:56,280 --> 00:35:58,160
Speaker 6: of all of this is you could end up in

675
00:35:58,160 --> 00:36:03,080
Speaker 6: a world where humans are just less engaged in this process, right,

676
00:36:03,160 --> 00:36:05,880
Speaker 6: and so there's both mistakes that humans miss or humans

677
00:36:05,920 --> 00:36:09,040
Speaker 6: just don't feel as morally responsible, which I think is

678
00:36:09,080 --> 00:36:13,200
Speaker 6: like a really tricky thing to think about morally because

679
00:36:14,000 --> 00:36:17,560
Speaker 6: on the one hand, as a democratic society we make

680
00:36:17,600 --> 00:36:19,759
Speaker 6: a decision as a nation to go to war. It's

681
00:36:19,760 --> 00:36:21,680
Speaker 6: a very small number of people that have to carry

682
00:36:21,680 --> 00:36:24,879
Speaker 6: that burden. And if someone if you could say, well, look,

683
00:36:25,000 --> 00:36:28,800
Speaker 6: what's the benefit to someone having like PTSD years after

684
00:36:28,840 --> 00:36:31,600
Speaker 6: a conflict that they're haunted mess having happened, that doesn't

685
00:36:31,600 --> 00:36:34,360
Speaker 6: seem great. Maybe we could reduce that. On the other hand,

686
00:36:34,560 --> 00:36:37,520
Speaker 6: if we thought of war and nobody felt morally responsible

687
00:36:37,560 --> 00:36:40,040
Speaker 6: for the killing that occurred, that doesn't seem good either,

688
00:36:40,320 --> 00:36:43,760
Speaker 6: and that could lead to more suffering and civilian casualties

689
00:36:43,760 --> 00:36:46,320
Speaker 6: and war. So I think that's certainly a concern we

690
00:36:46,360 --> 00:36:47,799
Speaker 6: think about how to use the technology.

691
00:36:48,080 --> 00:36:51,680
Speaker 3: Yeah, this is very enders game coded, right, where you

692
00:36:51,760 --> 00:36:54,120
Speaker 3: have someone who's basically playing like a video game and

693
00:36:54,160 --> 00:36:56,960
Speaker 3: wiping out entire civilizations and they think it's just a

694
00:36:57,040 --> 00:36:59,120
Speaker 3: video game, just an exercise, but it turns out it's

695
00:36:59,160 --> 00:37:02,160
Speaker 3: actual warfare. And we're seeing some degree of that in

696
00:37:02,200 --> 00:37:05,520
Speaker 3: the way that the Department of War is portraying this

697
00:37:05,600 --> 00:37:09,880
Speaker 3: conflict so far. It's very video game, yes.

698
00:37:09,200 --> 00:37:10,480
Speaker 4: Especially in public presentation.

699
00:37:10,760 --> 00:37:13,960
Speaker 2: Yeah, literal animated gifts of video games exactly.

700
00:37:14,320 --> 00:37:18,240
Speaker 3: So, Paul, you mentioned something you mentioned the word circuit breaker,

701
00:37:18,760 --> 00:37:22,200
Speaker 3: and circuit breakers are nice things to have in markets.

702
00:37:22,280 --> 00:37:24,319
Speaker 3: I think they'd be even nicer things to have in

703
00:37:24,520 --> 00:37:27,600
Speaker 3: armed conflict and war. Is there any possibility that you

704
00:37:27,680 --> 00:37:32,360
Speaker 3: could design something like that for a major conflict.

705
00:37:32,680 --> 00:37:34,719
Speaker 6: I think it's possible, like at a technical level, to

706
00:37:34,719 --> 00:37:37,200
Speaker 6: figure out how you would do that and where you

707
00:37:37,239 --> 00:37:39,759
Speaker 6: put protections on your side in the military, and what

708
00:37:39,800 --> 00:37:43,080
Speaker 6: you would do with even maybe cooperatively with an enemy.

709
00:37:43,320 --> 00:37:45,560
Speaker 6: The challenge is how do you avoid what we were

710
00:37:45,560 --> 00:37:48,040
Speaker 6: talking about earlier, a race to the bottom on safety. Yeah,

711
00:37:48,080 --> 00:37:50,719
Speaker 6: and right, we're seeing this in the private sector between

712
00:37:50,719 --> 00:37:52,640
Speaker 6: the AA companies as the Russian to get products out

713
00:37:52,680 --> 00:37:56,360
Speaker 6: the market. I think it's especially hard in the military space,

714
00:37:56,360 --> 00:38:00,680
Speaker 6: where countries are investing in the military because they're worried

715
00:38:00,719 --> 00:38:03,359
Speaker 6: about what some other adversary might do and then want

716
00:38:03,400 --> 00:38:05,719
Speaker 6: to get a leg up on them. And so it's

717
00:38:05,760 --> 00:38:09,920
Speaker 6: not that cooperation in the midst of conflict never happens.

718
00:38:09,960 --> 00:38:15,120
Speaker 6: It does, and countries have agreed to take certain weapons

719
00:38:15,239 --> 00:38:19,400
Speaker 6: off the table chemical and biological weapons, for example. It

720
00:38:19,400 --> 00:38:22,840
Speaker 6: doesn't mean that they're never used, but most civilized countries

721
00:38:22,840 --> 00:38:25,240
Speaker 6: that were not going to use them. But those examples

722
00:38:25,239 --> 00:38:28,120
Speaker 6: are pretty rare and it's pretty hard to do, and

723
00:38:28,200 --> 00:38:30,920
Speaker 6: some think that dynamic is the really challenging one. It's like,

724
00:38:30,960 --> 00:38:34,280
Speaker 6: how do you find ways to cooperate with your enemies

725
00:38:34,600 --> 00:38:36,480
Speaker 6: to avoid some of the biggest dangers here.

726
00:38:37,200 --> 00:38:39,799
Speaker 2: So I think there's the last question for me. You know,

727
00:38:39,880 --> 00:38:43,160
Speaker 2: you mentioned that drones are kind of robot and there

728
00:38:43,160 --> 00:38:46,320
Speaker 2: are other robots that have been existence in either national

729
00:38:46,320 --> 00:38:50,680
Speaker 2: security or police work for a while. I think there

730
00:38:50,719 --> 00:38:53,400
Speaker 2: are robots on the subway sometimes that seem to be

731
00:38:53,520 --> 00:38:55,480
Speaker 2: really yeah, but I don't think they really.

732
00:38:55,280 --> 00:38:57,279
Speaker 3: See the robots at the grocery store and they end

733
00:38:57,360 --> 00:38:59,480
Speaker 3: up like chasing me while I'm trying to buy like

734
00:38:59,680 --> 00:39:01,920
Speaker 3: rare or something. Yeah, the ones that sweep the floors

735
00:39:01,960 --> 00:39:02,360
Speaker 3: and stuff.

736
00:39:02,440 --> 00:39:03,480
Speaker 4: Yeah, there's the robots.

737
00:39:03,600 --> 00:39:06,480
Speaker 2: Yeah, I think there is a Eric Adams did a

738
00:39:06,520 --> 00:39:08,600
Speaker 2: contract with some company that was doing like subway for

739
00:39:08,760 --> 00:39:10,880
Speaker 2: cards or something like that. But these are really different

740
00:39:11,280 --> 00:39:13,759
Speaker 2: ai as we talk about it in robots are two

741
00:39:13,760 --> 00:39:17,520
Speaker 2: different technological trees, but they are going to merge, and

742
00:39:17,560 --> 00:39:21,400
Speaker 2: there's the possibility of their ultimate merger. Do you foresee

743
00:39:21,440 --> 00:39:25,839
Speaker 2: a world in which essentially we don't have human soldiers

744
00:39:26,080 --> 00:39:30,520
Speaker 2: and wars are fought with who has the most advanced

745
00:39:30,560 --> 00:39:33,839
Speaker 2: autonomous robots? We know China is investing a lot in

746
00:39:33,920 --> 00:39:37,560
Speaker 2: humanoid robotics. Do you foresee a world in which that

747
00:39:37,760 --> 00:39:41,279
Speaker 2: is the nature of a ground invasion as you and

748
00:39:41,520 --> 00:39:43,840
Speaker 2: it happens with robots or various other sorts, talk to

749
00:39:43,920 --> 00:39:45,319
Speaker 2: us about how far that could go.

750
00:39:46,520 --> 00:39:48,919
Speaker 6: Yeah, So, I mean, look, I think where we see

751
00:39:49,000 --> 00:39:52,600
Speaker 6: robots whose more and were for absolutely the long arc

752
00:39:52,680 --> 00:39:56,160
Speaker 6: of technology in war, from the first time someone picked

753
00:39:56,200 --> 00:39:58,160
Speaker 6: up a rocket through it at somebody else has been

754
00:39:58,680 --> 00:40:04,120
Speaker 6: towards great distance between adversaries moving up through bows and

755
00:40:04,239 --> 00:40:07,239
Speaker 6: arrows and rifles and intercontinental ballistic missiles. And I think

756
00:40:07,520 --> 00:40:09,880
Speaker 6: robotics will be the next evolution of this trend of

757
00:40:09,960 --> 00:40:12,879
Speaker 6: finding ways to find the enemy strike the enemy without

758
00:40:12,880 --> 00:40:15,760
Speaker 6: putting yourself at risk. And there's certainly a role for

759
00:40:16,280 --> 00:40:19,040
Speaker 6: robotics out on the battlefield. I think a vision of

760
00:40:19,200 --> 00:40:23,000
Speaker 6: like future wars of just robots fighting robots, there's no

761
00:40:23,080 --> 00:40:26,040
Speaker 6: humans in wealth. That's not realistic for a couple of reasons.

762
00:40:26,080 --> 00:40:29,080
Speaker 6: One is I think militaries are going to need people

763
00:40:29,239 --> 00:40:34,440
Speaker 6: relatively forward deployed to execute command and control for robotic systems.

764
00:40:35,200 --> 00:40:37,880
Speaker 6: The US military right now can fly drones remotely from

765
00:40:37,920 --> 00:40:42,279
Speaker 6: the United States in a relatively uncontested environment against more

766
00:40:42,320 --> 00:40:45,759
Speaker 6: sophisticated adversaries who could jam your communications like and we see,

767
00:40:45,800 --> 00:40:47,680
Speaker 6: for example, like a lot of jamming on the front

768
00:40:47,719 --> 00:40:50,399
Speaker 6: lines in Ukraine. That's one of the ways you should

769
00:40:50,400 --> 00:40:53,439
Speaker 6: go after these drones. Then you need people close by

770
00:40:53,920 --> 00:40:58,080
Speaker 6: because it is easier to have shorter range protected communications.

771
00:40:58,200 --> 00:41:00,000
Speaker 6: When you go to a longer distances, that's just hard

772
00:41:00,200 --> 00:41:02,439
Speaker 6: to do. So I think given be people relatively close

773
00:41:02,480 --> 00:41:04,840
Speaker 6: for that reason. I think if you want to control territory,

774
00:41:05,080 --> 00:41:07,680
Speaker 6: you have to put people there eventually to get out

775
00:41:07,680 --> 00:41:10,200
Speaker 6: of a vehicle and walk around and control it. But

776
00:41:10,239 --> 00:41:12,200
Speaker 6: I think the other reason is maybe a little dark,

777
00:41:12,239 --> 00:41:15,680
Speaker 6: which I think realistically, in order for wars to end,

778
00:41:16,360 --> 00:41:19,120
Speaker 6: there will have to be some human price that's paid.

779
00:41:19,160 --> 00:41:22,440
Speaker 6: I think that's an unfortunate reality that if it's just

780
00:41:22,680 --> 00:41:26,239
Speaker 6: machines that are being destroyed, that we may not get

781
00:41:26,280 --> 00:41:28,960
Speaker 6: to the place where one side or the other is

782
00:41:29,040 --> 00:41:32,759
Speaker 6: willing to suit for peace. And I think, unfortunately, war

783
00:41:32,840 --> 00:41:36,880
Speaker 6: is likely to involve people and human costs for a

784
00:41:36,960 --> 00:41:37,640
Speaker 6: very long time.

785
00:41:38,400 --> 00:41:40,319
Speaker 3: I have one more question as well, and I guess

786
00:41:40,360 --> 00:41:42,319
Speaker 3: it's a thought experiment, but if we think back to

787
00:41:42,600 --> 00:41:46,239
Speaker 3: sort of pivotal moments in military history and their intersection

788
00:41:46,440 --> 00:41:49,400
Speaker 3: with technology, one of them that comes up is the

789
00:41:49,480 --> 00:41:53,279
Speaker 3: Russian officer who decided not to press the button in

790
00:41:53,360 --> 00:41:56,879
Speaker 3: response to the US and thereby, you know, supposedly save

791
00:41:56,960 --> 00:42:01,960
Speaker 3: the world from nuclear disaster. Happen in a fully autonomous

792
00:42:02,000 --> 00:42:04,560
Speaker 3: military environment nowadays, I.

793
00:42:04,560 --> 00:42:06,880
Speaker 6: Mean today, it would still happen because there's people involved. Right,

794
00:42:06,920 --> 00:42:11,200
Speaker 6: So this incident standis la Petrov, He's sitting at a

795
00:42:11,280 --> 00:42:14,960
Speaker 6: terminal and it gets this warning that there's a ballistic

796
00:42:15,000 --> 00:42:17,920
Speaker 6: missile launched from the United states against the Soviet Union,

797
00:42:18,320 --> 00:42:21,839
Speaker 6: and then another missile, another five missiles coming in. And

798
00:42:22,239 --> 00:42:25,279
Speaker 6: the thing that's interesting about this is when Petrov talked

799
00:42:25,280 --> 00:42:27,759
Speaker 6: about it afterwards, and we could hear what he said

800
00:42:27,800 --> 00:42:30,400
Speaker 6: because we all lived because he made the right decision.

801
00:42:30,400 --> 00:42:33,480
Speaker 6: Here is he talked about how he said he had

802
00:42:33,480 --> 00:42:35,880
Speaker 6: a funny feeling in his gut and that he knew

803
00:42:35,920 --> 00:42:38,279
Speaker 6: that the Russian system, the Russians had just deployed, or

804
00:42:38,280 --> 00:42:41,359
Speaker 6: said the Soviets rather just deployed a new satellite based

805
00:42:41,400 --> 00:42:44,960
Speaker 6: early warning system to detect us ICBM launches. That it

806
00:42:45,000 --> 00:42:46,360
Speaker 6: was new, and he knew that a lot of the

807
00:42:46,400 --> 00:42:49,000
Speaker 6: Soviet technology doesn't work that great at first, so he

808
00:42:49,080 --> 00:42:51,680
Speaker 6: was skeptical of it. Turns out it was in fact faulty.

809
00:42:51,760 --> 00:42:54,359
Speaker 6: He was detecting the reflection of sunlight off the top

810
00:42:54,400 --> 00:42:57,320
Speaker 6: of clouds and the system was identifying that as a

811
00:42:57,640 --> 00:43:01,160
Speaker 6: missile launch, and that's what it was. And he went

812
00:43:01,280 --> 00:43:04,359
Speaker 6: and then called the early warning radar stations and staid,

813
00:43:04,360 --> 00:43:06,239
Speaker 6: are you seeing his missiles come over the horizon and said, no,

814
00:43:06,320 --> 00:43:08,640
Speaker 6: there's not missiles. So he reported up the chain that

815
00:43:08,640 --> 00:43:12,040
Speaker 6: the system was malfunctioning. I think the scary question here

816
00:43:12,120 --> 00:43:14,040
Speaker 6: is like if that was an AI, what would the

817
00:43:14,080 --> 00:43:17,000
Speaker 6: AI have done? And it's kind of like whatever it

818
00:43:17,040 --> 00:43:20,520
Speaker 6: was programmed to wherever it was trained to do. And

819
00:43:21,360 --> 00:43:25,040
Speaker 6: obviously we're seeing more general purpose AI systems like large

820
00:43:25,080 --> 00:43:29,000
Speaker 6: language models, have the ability to bring together more information,

821
00:43:29,239 --> 00:43:32,960
Speaker 6: to understand better context, to have just like a more

822
00:43:33,000 --> 00:43:35,759
Speaker 6: contextual understanding of the questions that you're asking of it.

823
00:43:36,280 --> 00:43:39,360
Speaker 6: But it still doesn't know the stakes of a conference.

824
00:43:39,400 --> 00:43:42,160
Speaker 6: It still doesn't know, like in some visceral level, what

825
00:43:42,200 --> 00:43:44,800
Speaker 6: the consequences are. And so I think that's a strong

826
00:43:44,880 --> 00:43:48,160
Speaker 6: compelling reason why we need to have humans involved these decisions.

827
00:43:48,600 --> 00:43:51,359
Speaker 6: Even as the AI becomes more capable, there's still gonna

828
00:43:51,360 --> 00:43:53,960
Speaker 6: be things we want humans to do because humans understand

829
00:43:53,960 --> 00:43:54,680
Speaker 6: why it matters.

830
00:43:55,719 --> 00:43:59,080
Speaker 2: I started the conversation by mentioning that you know, it's

831
00:43:59,160 --> 00:44:02,600
Speaker 2: not very common traversial to say, have an anti missile

832
00:44:02,640 --> 00:44:06,080
Speaker 2: system fire a missile when there's one coming in, but

833
00:44:06,440 --> 00:44:08,040
Speaker 2: that could be wrong, and you want to make sure

834
00:44:08,040 --> 00:44:10,600
Speaker 2: that it is in fact a missile and not a

835
00:44:10,640 --> 00:44:13,520
Speaker 2: civilian air jet or something like that. Be So even

836
00:44:13,680 --> 00:44:16,880
Speaker 2: there where it seems like a canonical example, if you

837
00:44:16,920 --> 00:44:19,759
Speaker 2: just want to have the missile system go off, you

838
00:44:19,760 --> 00:44:23,080
Speaker 2: would want to have human safeguards and human oversight and

839
00:44:23,160 --> 00:44:27,840
Speaker 2: human understanding of the system such that it is in

840
00:44:28,000 --> 00:44:32,879
Speaker 2: fact shooting down missile anyway, Paul Sharie, fascinating conversation. Really

841
00:44:32,920 --> 00:44:35,040
Speaker 2: appreciate you coming out on the up lots and talking

842
00:44:35,040 --> 00:44:35,560
Speaker 2: about your work.

843
00:44:36,360 --> 00:44:37,720
Speaker 6: Thank you. Really joyed the discussion.

844
00:44:38,239 --> 00:44:41,239
Speaker 5: Thanks so much, Paul. That was depressing and fascinating.

845
00:44:41,360 --> 00:44:44,160
Speaker 4: Yeah at the time, Yeah, no, it was great. It

846
00:44:44,200 --> 00:44:45,000
Speaker 4: was Thank you so much.

847
00:44:45,400 --> 00:44:46,359
Speaker 6: Yeah, Facebook cover.

848
00:44:59,080 --> 00:44:59,680
Speaker 4: I kind of get.

849
00:44:59,640 --> 00:45:01,880
Speaker 2: Choked up at the end thinking about that decision that

850
00:45:02,080 --> 00:45:05,000
Speaker 2: saved humanity at the end, And.

851
00:45:05,000 --> 00:45:06,680
Speaker 5: It's actually it's a crazy story.

852
00:45:06,719 --> 00:45:09,000
Speaker 2: It's a crazy story. It's one of those stories that like,

853
00:45:09,080 --> 00:45:10,720
Speaker 2: why doesn't everybody know.

854
00:45:10,920 --> 00:45:12,479
Speaker 4: That that person's name?

855
00:45:12,640 --> 00:45:14,799
Speaker 2: I mean, when you think about how many people, I

856
00:45:14,840 --> 00:45:15,960
Speaker 2: couldn't remember it either.

857
00:45:16,280 --> 00:45:18,480
Speaker 3: But shout out to another podcast if you want to

858
00:45:18,520 --> 00:45:21,759
Speaker 3: learn more about this. Dan Carlin's Hardcore History has at

859
00:45:21,880 --> 00:45:27,680
Speaker 3: least one, possibly two episodes on narrow aversions of nuclear disaster.

860
00:45:27,920 --> 00:45:31,200
Speaker 3: So very good to listen to, if not terrifying.

861
00:45:31,320 --> 00:45:34,040
Speaker 2: You know, there's another point in that exact story that

862
00:45:34,120 --> 00:45:36,600
Speaker 2: I think is really interesting and it's something I've been

863
00:45:36,680 --> 00:45:40,680
Speaker 2: thinking about a lot across AI because There's something similar

864
00:45:40,719 --> 00:45:45,200
Speaker 2: about humans and AI, which is that there is definitely

865
00:45:45,320 --> 00:45:48,799
Speaker 2: a gap between what we know and what we can articulate.

866
00:45:49,040 --> 00:45:51,600
Speaker 2: And this is certainly true with AI. Right, so the

867
00:45:51,640 --> 00:45:54,480
Speaker 2: bot makes some decision, or it determines something, it does

868
00:45:54,520 --> 00:45:56,440
Speaker 2: not mean it's going to be able to spit out

869
00:45:56,480 --> 00:45:58,960
Speaker 2: in words, how it arrived at that decision.

870
00:45:59,160 --> 00:46:01,080
Speaker 4: But that's true for humans as well.

871
00:46:01,120 --> 00:46:04,200
Speaker 2: And so the idea that like, okay, maybe we do

872
00:46:04,239 --> 00:46:07,560
Speaker 2: get funny feelings about or take you know, like, again,

873
00:46:08,000 --> 00:46:11,439
Speaker 2: we're still pretty good at determining the difference between AI

874
00:46:11,480 --> 00:46:15,280
Speaker 2: generated texts and human generated text can we but often

875
00:46:15,680 --> 00:46:17,799
Speaker 2: I mean we're still like often could get it right,

876
00:46:18,719 --> 00:46:21,560
Speaker 2: but could we write down exactly what we saw that

877
00:46:21,600 --> 00:46:25,959
Speaker 2: we understood? There is that gap, and when we're talking

878
00:46:25,960 --> 00:46:31,040
Speaker 2: about life or death decisions being made, it is scary

879
00:46:31,160 --> 00:46:34,480
Speaker 2: to think about that role of instinct that we can't

880
00:46:34,600 --> 00:46:37,480
Speaker 2: articulate having been taken out of the decision. Well.

881
00:46:37,520 --> 00:46:40,480
Speaker 3: I think also technology is very good at pattern recognition,

882
00:46:40,760 --> 00:46:44,319
Speaker 3: right and responding to patterns and preset paths. It's been

883
00:46:44,360 --> 00:46:47,600
Speaker 3: programmed to do certain things, and I think in a

884
00:46:47,640 --> 00:46:50,719
Speaker 3: war environment, that's one of the most uncertain environments that

885
00:46:50,719 --> 00:46:53,399
Speaker 3: you can possibly imagine, and so you have to think

886
00:46:53,480 --> 00:46:57,400
Speaker 3: that there should be some element of flexibility in your response,

887
00:46:57,440 --> 00:47:00,600
Speaker 3: But I don't know how you actually encode that into

888
00:47:00,640 --> 00:47:03,800
Speaker 3: a thing that like runs on rigid numbers and lines

889
00:47:04,160 --> 00:47:06,360
Speaker 3: lines of code. The other thing I was thinking was

890
00:47:06,560 --> 00:47:09,960
Speaker 3: the anthropic situation and just how new that is from

891
00:47:10,000 --> 00:47:12,440
Speaker 3: a sort of military history perspective, in the sense that

892
00:47:12,520 --> 00:47:16,879
Speaker 3: here we have this really important pivotal piece of technology

893
00:47:16,920 --> 00:47:21,120
Speaker 3: that hasn't come out of like actual military demand. Right

894
00:47:21,280 --> 00:47:24,520
Speaker 3: to Paul's point, it's a commercial product. Its commercial uses

895
00:47:24,560 --> 00:47:27,920
Speaker 3: are arguably a lot more profitable than its military ones,

896
00:47:28,160 --> 00:47:31,799
Speaker 3: and so seeing that now interact with the Pentagon and

897
00:47:31,880 --> 00:47:34,080
Speaker 3: the Department of War really interesting.

898
00:47:34,200 --> 00:47:35,279
Speaker 5: It's been flipped right.

899
00:47:35,560 --> 00:47:35,840
Speaker 4: Yeah.

900
00:47:35,880 --> 00:47:39,040
Speaker 2: The closest example actually that comes to mind, there is

901
00:47:39,080 --> 00:47:42,760
Speaker 2: one example that's in fairly recent history, and it's starlink

902
00:47:42,840 --> 00:47:44,920
Speaker 2: oh yeah co And of course that was developed for

903
00:47:44,920 --> 00:47:47,960
Speaker 2: commercial internet purposes, but it played a role in Ukraine

904
00:47:48,280 --> 00:47:51,280
Speaker 2: and so forth, and at one point, if I recall,

905
00:47:51,680 --> 00:47:54,839
Speaker 2: there was a tension point about the degree to which

906
00:47:54,880 --> 00:47:58,040
Speaker 2: the Ukrainians could use and so I do think that

907
00:47:58,320 --> 00:48:00,759
Speaker 2: is sort of an interesting parallel here.

908
00:48:00,960 --> 00:48:02,480
Speaker 4: The other thing that we didn't.

909
00:48:02,239 --> 00:48:04,600
Speaker 2: Get to this and I this is gonna be a

910
00:48:04,640 --> 00:48:07,439
Speaker 2: little cynical, but I think it's right, which is that

911
00:48:08,200 --> 00:48:12,600
Speaker 2: there's another element I believe to the Anthropic situation, which is,

912
00:48:12,680 --> 00:48:15,040
Speaker 2: like Anthropic is the last.

913
00:48:14,800 --> 00:48:16,000
Speaker 4: Big lib tech company.

914
00:48:16,239 --> 00:48:19,239
Speaker 2: We're perceived as such, right, and we know that there's

915
00:48:19,280 --> 00:48:22,120
Speaker 2: been this fairly sort of right word turn in Silicon

916
00:48:22,200 --> 00:48:25,920
Speaker 2: Valley over the years. And I don't think like Anthropic

917
00:48:26,040 --> 00:48:28,120
Speaker 2: is like totally part of that. I think there's still

918
00:48:28,200 --> 00:48:31,600
Speaker 2: sort of lib coded. I also think it's incidentally why

919
00:48:31,640 --> 00:48:33,319
Speaker 2: a bunch of people who probably like work in media

920
00:48:33,360 --> 00:48:35,400
Speaker 2: are like end up using Claude even though they're all

921
00:48:35,440 --> 00:48:37,880
Speaker 2: kind of the same. Like, I do think there's something

922
00:48:37,920 --> 00:48:41,080
Speaker 2: there and that they have this thing they say, we're

923
00:48:41,120 --> 00:48:41,880
Speaker 2: not going to ever.

924
00:48:41,880 --> 00:48:44,720
Speaker 4: Have ads, and we know that.

925
00:48:44,600 --> 00:48:47,319
Speaker 2: Like Andreas and Horowitz, Mark Andresen just talked about ads

926
00:48:47,360 --> 00:48:50,600
Speaker 2: are good. Ads democratize the internet. Ads enabled the Internet

927
00:48:50,640 --> 00:48:53,080
Speaker 2: to be spread to everyone. There are some other politics

928
00:48:53,080 --> 00:48:56,800
Speaker 2: at play. Because again, like from my understanding, and it

929
00:48:56,840 --> 00:48:58,480
Speaker 2: would take a lawyer, it's like, I don't think that

930
00:48:58,560 --> 00:49:02,040
Speaker 2: the agreement that open as Ey sign was that different

931
00:49:03,080 --> 00:49:06,040
Speaker 2: than what the agreement that Anthropic had. There is probably

932
00:49:06,040 --> 00:49:08,000
Speaker 2: a little bit of difference. I just think there's some

933
00:49:08,080 --> 00:49:13,720
Speaker 2: other politics and play here. Per to Paul's point, maybe

934
00:49:13,760 --> 00:49:19,600
Speaker 2: nobody is talking about currently autonomous weapons right now fully autonomous,

935
00:49:19,680 --> 00:49:23,000
Speaker 2: but it can't be long, and I think this is

936
00:49:23,040 --> 00:49:24,040
Speaker 2: going to be a real tension.

937
00:49:24,120 --> 00:49:24,959
Speaker 4: Sooner resan later.

938
00:49:25,040 --> 00:49:26,480
Speaker 3: Can I say one thing and I'm going to be

939
00:49:26,560 --> 00:49:30,000
Speaker 3: slightly facetious, but also not Can you be slightly facetious.

940
00:49:30,040 --> 00:49:30,560
Speaker 4: I'm going to be.

941
00:49:30,480 --> 00:49:33,160
Speaker 3: Facetious, which is I have a solution to modern warfare.

942
00:49:33,760 --> 00:49:34,399
Speaker 4: Don't do it?

943
00:49:35,960 --> 00:49:38,319
Speaker 5: Okay, but no for real?

944
00:49:38,680 --> 00:49:41,480
Speaker 3: Yeah, if we're just gonna have bots fighting points and

945
00:49:41,520 --> 00:49:43,440
Speaker 3: it's going to cost a lot of money and result

946
00:49:43,480 --> 00:49:49,000
Speaker 3: in people's deaths, every country should have to build its biggest, best,

947
00:49:49,239 --> 00:49:52,719
Speaker 3: most technologically advanced robot and just have them fight it

948
00:49:52,760 --> 00:49:56,400
Speaker 3: out gladiatorial. And my twist is to Paul's point about

949
00:49:56,400 --> 00:50:00,840
Speaker 3: war always having to be painful in some way, Everyone

950
00:50:00,960 --> 00:50:04,240
Speaker 3: in that particular society has to be engaged and dedicate

951
00:50:04,320 --> 00:50:08,640
Speaker 3: some amount of time or money to building that particular robot,

952
00:50:09,040 --> 00:50:11,720
Speaker 3: and you just have to iterate on the robot forever

953
00:50:11,880 --> 00:50:14,759
Speaker 3: until you feel comfortable to have them fight, and that

954
00:50:14,840 --> 00:50:17,840
Speaker 3: way everyone shares the pain, but without the loss of

955
00:50:17,920 --> 00:50:19,320
Speaker 3: human life. Am I high.

956
00:50:19,400 --> 00:50:20,640
Speaker 5: I don't think so well.

957
00:50:20,840 --> 00:50:22,279
Speaker 4: I think you should write a book. No, I don't

958
00:50:22,280 --> 00:50:23,520
Speaker 4: think you should write a sci fi book.

959
00:50:23,640 --> 00:50:24,839
Speaker 5: All right, shall we leave it there.

960
00:50:24,920 --> 00:50:25,640
Speaker 4: Let's leave it there.

961
00:50:25,800 --> 00:50:27,960
Speaker 3: This has been another episode of the Authoughts podcast.

962
00:50:28,040 --> 00:50:31,080
Speaker 5: I'm Tracy Alloway. You can follow me at Tracy Alloway.

963
00:50:30,800 --> 00:50:33,560
Speaker 2: And I'm Joe Wisenthal. You can follow me at the Stalwart.

964
00:50:33,680 --> 00:50:37,120
Speaker 2: Follow our guest Paul Shari. He's at Paul Underscore Shari.

965
00:50:37,440 --> 00:50:40,600
Speaker 2: Follow our producers Carmen Rodriguez at Carmen armat dash, Ol

966
00:50:40,600 --> 00:50:42,000
Speaker 2: Bennett at Dashbot.

967
00:50:41,680 --> 00:50:43,520
Speaker 4: And kill Brooks at Kelbrooks.

968
00:50:43,800 --> 00:50:46,080
Speaker 2: For more Odd Lots content, go to Bloomberg dot com

969
00:50:46,080 --> 00:50:48,239
Speaker 2: slash odd Lots, where the daily newsletter and all of

970
00:50:48,280 --> 00:50:50,279
Speaker 2: our episodes, and you can chaut about all of these

971
00:50:50,320 --> 00:50:53,439
Speaker 2: topics twenty four to seven in our discord Discord dot

972
00:50:53,520 --> 00:50:54,880
Speaker 2: gg slash onlines.

973
00:50:54,920 --> 00:50:56,719
Speaker 3: And if you enjoy all thoughts, if you like it

974
00:50:56,760 --> 00:50:59,399
Speaker 3: when we talk about the future of autonomous weapons, then

975
00:50:59,400 --> 00:51:02,640
Speaker 3: please leave us a positive review on your favorite podcast platform.

976
00:51:02,760 --> 00:51:05,239
Speaker 3: And remember, if you are a Bloomberg subscriber, you can

977
00:51:05,280 --> 00:51:07,960
Speaker 3: listen to all of our episodes absolutely ad free.

978
00:51:08,080 --> 00:51:10,600
Speaker 5: All you need to do is find the Bloomberg.

979
00:51:10,120 --> 00:51:12,799
Speaker 3: Channel on Apple Podcasts and follow the instructions there.

980
00:51:13,200 --> 00:51:14,040
Speaker 5: Thanks for listening.