1
00:00:00,040 --> 00:00:08,719
Speaker 1: Bloomberg Audio Studios, Podcasts, radio News. I'm Monica Ricks in

2
00:00:08,760 --> 00:00:11,799
Speaker 1: the Bloomberg newsroom in New York with a special conversation.

3
00:00:12,240 --> 00:00:15,640
Speaker 1: Bloomberg Tech host ed Ludlow sat down within Video CEO

4
00:00:15,840 --> 00:00:19,239
Speaker 1: Jensen Huang to discuss the tech giant's latest investments in

5
00:00:19,280 --> 00:00:22,480
Speaker 1: South Korea, which includes teaming up with sk Group to

6
00:00:22,560 --> 00:00:25,799
Speaker 1: build new data centers and investing a billion dollars in

7
00:00:25,880 --> 00:00:29,400
Speaker 1: internet and cloud service provider Neighbor. Let's listen into a

8
00:00:29,400 --> 00:00:32,600
Speaker 1: portion of their conversation, coming not long after the Nvidia

9
00:00:32,680 --> 00:00:35,200
Speaker 1: CEO appeared at a Korean AI summit.

10
00:00:35,760 --> 00:00:37,360
Speaker 2: Jens and I think we just start with the basics,

11
00:00:37,479 --> 00:00:41,839
Speaker 2: like career is incredibly important to the AI built out globally,

12
00:00:42,120 --> 00:00:44,720
Speaker 2: we'll get intohigh bandwidth memory. But just from this summit,

13
00:00:45,280 --> 00:00:47,800
Speaker 2: from the president being here, what is the takeaway? What

14
00:00:47,880 --> 00:00:48,920
Speaker 2: is it you're trying to achieve?

15
00:00:49,640 --> 00:00:51,839
Speaker 3: Well, we're announcing a whole bunch of partnerships with them.

16
00:00:51,840 --> 00:00:53,960
Speaker 3: This is the golden ages for Korea. As you know,

17
00:00:54,360 --> 00:00:59,200
Speaker 3: there's semiconductor businesses booming, their industrial businesses booming. You know,

18
00:00:59,240 --> 00:01:01,320
Speaker 3: this is a country that has the ability to help

19
00:01:01,360 --> 00:01:05,000
Speaker 3: the world build out the AI infrastructure that they're incredibly

20
00:01:05,319 --> 00:01:11,040
Speaker 3: adapt at adopting new technologies, and it's a really technologically

21
00:01:11,880 --> 00:01:15,600
Speaker 3: forward leaning society and they love using AI. AI has

22
00:01:15,880 --> 00:01:19,640
Speaker 3: has really diffused throughout their society and and their industry,

23
00:01:19,680 --> 00:01:22,160
Speaker 3: and so so this is a great time for them.

24
00:01:22,160 --> 00:01:26,399
Speaker 3: We're announcing several things. We announce a big partnership with

25
00:01:27,080 --> 00:01:31,720
Speaker 3: SK Group where our companies are going to enter into

26
00:01:31,800 --> 00:01:36,120
Speaker 3: a business partnership where we do over five hundred billion

27
00:01:36,200 --> 00:01:39,760
Speaker 3: dollars of business with each other, whether it's consump consumption

28
00:01:39,959 --> 00:01:44,720
Speaker 3: and purchasing of memories or selling AI supercomputers to them

29
00:01:44,720 --> 00:01:48,120
Speaker 3: as they scale out to gigawatts of AI factories. There's

30
00:01:48,120 --> 00:01:51,400
Speaker 3: a whole bunch of other announcements. We're investing a billion

31
00:01:51,480 --> 00:01:55,600
Speaker 3: dollars in neighbor to help the they're the Korea's leading

32
00:01:55,640 --> 00:01:58,480
Speaker 3: AI cloud. They're going to scale up in Korea up

33
00:01:58,480 --> 00:02:00,360
Speaker 3: to two they're going to scale up to in mega

34
00:02:00,400 --> 00:02:03,120
Speaker 3: watts I think it is, and they're going to expand

35
00:02:03,560 --> 00:02:06,200
Speaker 3: across the world. And so we have a whole bunch

36
00:02:06,200 --> 00:02:07,880
Speaker 3: of announcements that we're making today.

37
00:02:07,720 --> 00:02:11,160
Speaker 2: With the expanded SK relationship. There's also sort of more

38
00:02:11,200 --> 00:02:14,840
Speaker 2: direct involvement with video on the roadmap for HBM future

39
00:02:14,840 --> 00:02:19,000
Speaker 2: generations of HBM talk about that. You know, I remember

40
00:02:19,000 --> 00:02:21,320
Speaker 2: you being on stage earlier this year saying five years

41
00:02:21,320 --> 00:02:23,519
Speaker 2: ago we told our supply chain what was going to happen,

42
00:02:24,000 --> 00:02:27,040
Speaker 2: and it did happen, and you gave some credit to

43
00:02:27,080 --> 00:02:30,280
Speaker 2: the memory makers in going with you on that journey.

44
00:02:30,639 --> 00:02:33,120
Speaker 2: But clearly you want to be involved in the direction

45
00:02:33,160 --> 00:02:36,000
Speaker 2: of travel for future generations of HBM.

46
00:02:36,560 --> 00:02:38,800
Speaker 3: Yeah, we're working together on of course, we started with

47
00:02:38,919 --> 00:02:42,720
Speaker 3: HBM two, worked on HBM three, three E, four, four E,

48
00:02:42,880 --> 00:02:45,960
Speaker 3: and then beyond, and so we've got a whole roadmap

49
00:02:46,000 --> 00:02:49,880
Speaker 3: of memories that we're working on together. It is also

50
00:02:49,919 --> 00:02:53,400
Speaker 3: the case that the semiconductor industry has really changed, and

51
00:02:53,440 --> 00:02:57,000
Speaker 3: the reason for that because we used to build computers

52
00:02:57,360 --> 00:02:59,640
Speaker 3: for people to use, and we're going to still continue

53
00:02:59,680 --> 00:03:03,800
Speaker 3: to build incredible computers. These are now eight processing ais

54
00:03:03,800 --> 00:03:07,760
Speaker 3: for humans to collaborate with. But in the future we

55
00:03:07,800 --> 00:03:10,799
Speaker 3: also have AI agents and robots and they're going to

56
00:03:10,840 --> 00:03:13,560
Speaker 3: be using computers. So instead of just a billion people

57
00:03:13,639 --> 00:03:16,560
Speaker 3: using computers, we're going to have one hundred billion agents

58
00:03:16,600 --> 00:03:20,639
Speaker 3: and billions of robots all using computers. The computer industry,

59
00:03:20,680 --> 00:03:23,000
Speaker 3: the chip that's built on top of the chip industry

60
00:03:23,240 --> 00:03:25,520
Speaker 3: surely is not big enough. And so this is one

61
00:03:25,560 --> 00:03:30,279
Speaker 3: of the realizations of the semiconductor industry that now computers

62
00:03:30,560 --> 00:03:32,880
Speaker 3: are built not just for people to use, but computers

63
00:03:32,919 --> 00:03:35,960
Speaker 3: are being built for computers to use. My guess is

64
00:03:36,000 --> 00:03:40,440
Speaker 3: that that the semiconductor industry is probably going to have

65
00:03:40,480 --> 00:03:43,520
Speaker 3: to be ten times larger than it is today over

66
00:03:43,560 --> 00:03:47,360
Speaker 3: the next decade or so. And so working with our

67
00:03:47,400 --> 00:03:50,400
Speaker 3: partners in Korea and around the world to scale up

68
00:03:50,400 --> 00:03:52,920
Speaker 3: the supply chain and the semiconductors so that we're prepared

69
00:03:52,960 --> 00:03:54,440
Speaker 3: for this AI future. It's really important.

70
00:03:54,480 --> 00:03:55,920
Speaker 2: I've had the obvious need to ask you about this

71
00:03:55,960 --> 00:03:57,640
Speaker 2: more than once this year, but how much do you

72
00:03:57,720 --> 00:04:00,480
Speaker 2: need the Korean economy to kind of get going to

73
00:04:00,560 --> 00:04:04,280
Speaker 2: increase the supply of HBM bits for Nvidia based systems

74
00:04:04,280 --> 00:04:05,000
Speaker 2: wherever they are.

75
00:04:05,760 --> 00:04:09,960
Speaker 3: Well, we don't have enough bits. We're constrained in HBO memories,

76
00:04:10,360 --> 00:04:13,240
Speaker 3: LPDD our memories. We're constrained, and just about every part

77
00:04:13,280 --> 00:04:16,480
Speaker 3: of the supply chain. We're even constrained now with land

78
00:04:16,600 --> 00:04:20,160
Speaker 3: and power and construction workers to set up the data centers.

79
00:04:20,400 --> 00:04:23,400
Speaker 3: I think this is one of the areas that is

80
00:04:23,440 --> 00:04:25,800
Speaker 3: going to make sure that we continued to build out

81
00:04:25,800 --> 00:04:28,479
Speaker 3: in a throttled way, you know, for a decade. And

82
00:04:28,520 --> 00:04:31,840
Speaker 3: the reason for that is because of these infrastructure Unlike electronics,

83
00:04:32,080 --> 00:04:35,040
Speaker 3: electronic devices like PCs and phones and things like that,

84
00:04:35,360 --> 00:04:38,000
Speaker 3: it's really really hard to scale up land power and

85
00:04:38,040 --> 00:04:41,320
Speaker 3: shell and so all of the supply chain just really

86
00:04:41,320 --> 00:04:43,840
Speaker 3: needs to get built out over the years. I think

87
00:04:43,880 --> 00:04:46,400
Speaker 3: we have the ability as an industry to double each year,

88
00:04:46,839 --> 00:04:48,880
Speaker 3: but we're going to have a hard time growing much faster.

89
00:04:48,960 --> 00:04:51,560
Speaker 2: In that the five hundred billions a number is large.

90
00:04:51,760 --> 00:04:53,120
Speaker 2: Would you just took a little bit more about what

91
00:04:53,160 --> 00:04:55,720
Speaker 2: it encompasses. We've gone over a lot on your commitment

92
00:04:55,760 --> 00:04:58,560
Speaker 2: to the US in terms of spending. Is that in

93
00:04:58,680 --> 00:05:01,880
Speaker 2: Vidia spending in the in economy or it's it's s

94
00:05:02,000 --> 00:05:05,760
Speaker 2: K fronting capital expenditures just a little bit more detail.

95
00:05:06,040 --> 00:05:09,000
Speaker 3: We're gonna we're gonna be purchasing memories from them for

96
00:05:09,480 --> 00:05:12,280
Speaker 3: many years to come. And as you know, we buy,

97
00:05:12,320 --> 00:05:14,479
Speaker 3: we build a lot of computers. In order to build

98
00:05:14,520 --> 00:05:17,240
Speaker 3: a trillion dollars worth of a roommate systems, you're gonna

99
00:05:17,240 --> 00:05:18,640
Speaker 3: have to buy a lot of system memories to go

100
00:05:18,760 --> 00:05:21,200
Speaker 3: with it. And so we have we have large purchase

101
00:05:22,320 --> 00:05:27,760
Speaker 3: agreements and large purchase intentions with s K Heinex. Meanwhile,

102
00:05:27,920 --> 00:05:30,440
Speaker 3: s K Telecom is going to become an AI cloud.

103
00:05:30,800 --> 00:05:33,440
Speaker 3: They're going, we're starting to build already, They're going they're

104
00:05:33,480 --> 00:05:36,320
Speaker 3: intending to build up to two gigawatts in the near future.

105
00:05:36,920 --> 00:05:40,000
Speaker 3: And uh and in that in that agreement, we will

106
00:05:40,000 --> 00:05:43,520
Speaker 3: be selling AI supercomputers to them. So between us, we're gonna,

107
00:05:43,560 --> 00:05:46,040
Speaker 3: we're gonna, we're gonna do half a trillion dollars worth

108
00:05:46,040 --> 00:05:48,320
Speaker 3: of business. Over half a trillion dollars worth a business.

109
00:05:48,400 --> 00:05:50,600
Speaker 2: I was able to sit down with SK Group chair

110
00:05:50,720 --> 00:05:53,560
Speaker 2: and JK One very recently for about forty minutes, and

111
00:05:54,240 --> 00:05:55,719
Speaker 2: at the end of the conversation we got to what

112
00:05:55,839 --> 00:05:58,640
Speaker 2: is the difference in approach, the academic difference and approach

113
00:05:58,640 --> 00:06:01,840
Speaker 2: on AI between the United States in China. And his

114
00:06:01,960 --> 00:06:04,640
Speaker 2: view on it was that China is very focused on

115
00:06:04,800 --> 00:06:08,000
Speaker 2: lowering the dollar PA token. In America were still focused

116
00:06:08,000 --> 00:06:10,960
Speaker 2: on the quality of tokens. I wonder what you think

117
00:06:11,000 --> 00:06:11,240
Speaker 2: of that.

118
00:06:13,839 --> 00:06:16,760
Speaker 3: The goal of the goal of AI is to produce

119
00:06:16,760 --> 00:06:20,200
Speaker 3: an intelligent smart answer. Now you could approach it in

120
00:06:20,240 --> 00:06:22,000
Speaker 3: a couple of different ways. You could, of course make

121
00:06:22,080 --> 00:06:24,320
Speaker 3: all of the token smart and smarter, and as a

122
00:06:24,320 --> 00:06:27,880
Speaker 3: result result in using less tokens to do so, you

123
00:06:27,920 --> 00:06:31,919
Speaker 3: could also produce AIS that are much more efficient, and

124
00:06:31,960 --> 00:06:35,360
Speaker 3: maybe you can think longer, explore more options and as

125
00:06:35,360 --> 00:06:38,159
Speaker 3: a result result as a result, produces a smart answer.

126
00:06:38,400 --> 00:06:39,960
Speaker 3: There are a couple, there are many different ways to

127
00:06:41,480 --> 00:06:46,279
Speaker 3: reach intelligence and deliver smart answers. In the end, really,

128
00:06:46,279 --> 00:06:47,640
Speaker 3: I think you have to take a step back and

129
00:06:47,680 --> 00:06:52,760
Speaker 3: just realize that both countries has extraordinary AI researchers and

130
00:06:52,800 --> 00:06:58,160
Speaker 3: whatever whatever conditions and whatever resources that they have, amazing

131
00:06:58,160 --> 00:07:00,880
Speaker 3: people will find great answers. And so you're going to

132
00:07:00,960 --> 00:07:03,120
Speaker 3: find You're going to you know, my expectation is that

133
00:07:03,279 --> 00:07:05,640
Speaker 3: China and the United States will continue to advance AI.

134
00:07:06,400 --> 00:07:09,159
Speaker 3: The conditions are different, the resources are different, their constraints

135
00:07:09,200 --> 00:07:11,880
Speaker 3: are different, but they're all they're you know, these amazing

136
00:07:11,880 --> 00:07:15,720
Speaker 3: researchers will find answers and and I think that that

137
00:07:16,440 --> 00:07:19,840
Speaker 3: in the case of China, they're producing more AI researchers

138
00:07:20,320 --> 00:07:23,680
Speaker 3: than probably all of the world has, you know, in

139
00:07:23,720 --> 00:07:28,520
Speaker 3: any given year, and so they're producing. If manufacturing intelligence

140
00:07:29,280 --> 00:07:32,640
Speaker 3: is important, they manufacture the most most important version of it,

141
00:07:32,680 --> 00:07:35,480
Speaker 3: which is the researchers. And so so this is a

142
00:07:35,720 --> 00:07:38,360
Speaker 3: this is an area a country that's going to produce

143
00:07:38,480 --> 00:07:42,440
Speaker 3: excellent AI technology. We how to keep and keep continue

144
00:07:42,480 --> 00:07:45,040
Speaker 3: to learn from them, work with them. As you know,

145
00:07:45,080 --> 00:07:47,320
Speaker 3: you're here in Silicon Valley, right here in San Francisco.

146
00:07:47,920 --> 00:07:50,320
Speaker 3: The number of AI researchers here that came from China,

147
00:07:50,320 --> 00:07:53,800
Speaker 3: that are Chinese is really quite significant. And so, uh,

148
00:07:53,840 --> 00:07:56,000
Speaker 3: you know, we're really fortunate to have them here, and

149
00:07:56,000 --> 00:07:58,640
Speaker 3: and uh, you know, we just got to keep on racing.

150
00:07:58,800 --> 00:08:02,720
Speaker 2: You made your first post on x I did and

151
00:08:02,880 --> 00:08:06,720
Speaker 2: you did so by sharing a letter signed by many

152
00:08:06,720 --> 00:08:10,640
Speaker 2: of your peers American companies to talk about the importance

153
00:08:10,640 --> 00:08:14,200
Speaker 2: of open models to America, to to the industry, to

154
00:08:14,280 --> 00:08:16,840
Speaker 2: the development of AI. And in the letter it's pretty

155
00:08:16,880 --> 00:08:19,520
Speaker 2: well explained. You know your rationale, But what what was

156
00:08:19,560 --> 00:08:23,239
Speaker 2: the catalyst for now? Why did you and saty Inodella

157
00:08:23,360 --> 00:08:26,000
Speaker 2: and others need to do that in this moment?

158
00:08:26,320 --> 00:08:31,160
Speaker 3: What we sense we sense that there's there's a growing

159
00:08:34,920 --> 00:08:40,400
Speaker 3: sentiment that that and the wrong sentiment for open models.

160
00:08:41,440 --> 00:08:45,000
Speaker 3: It's really important to realize that open models is essential

161
00:08:45,280 --> 00:08:49,319
Speaker 3: for safety. Open models is essential for security, for cybersecurity.

162
00:08:49,360 --> 00:08:51,520
Speaker 3: Open models are essential for innovation. It's and as a

163
00:08:51,559 --> 00:08:58,280
Speaker 3: server for startups. It's necessary for sovereignty company sovereignty. I

164
00:08:58,360 --> 00:09:01,800
Speaker 3: see a future where the world uses tons of closed models,

165
00:09:02,200 --> 00:09:06,240
Speaker 3: and I encourage everybody, including my company, to use open

166
00:09:06,320 --> 00:09:12,079
Speaker 3: AI and cloud and cursor and cognition and perplexity, use

167
00:09:12,200 --> 00:09:16,240
Speaker 3: everything that you can because it's out of the cloud,

168
00:09:16,280 --> 00:09:20,520
Speaker 3: because it's just easier, and you build only what you must,

169
00:09:21,720 --> 00:09:23,560
Speaker 3: and so in order to build what you must, you

170
00:09:23,600 --> 00:09:25,440
Speaker 3: need to have open models to do that with. And

171
00:09:25,520 --> 00:09:27,920
Speaker 3: the area is where we must. Maybe it's because we

172
00:09:27,960 --> 00:09:31,440
Speaker 3: have expertise that we simply cannot afford to share. This

173
00:09:31,480 --> 00:09:34,160
Speaker 3: is our company's alpha, our company's intelligence, and we have

174
00:09:34,200 --> 00:09:36,720
Speaker 3: to make sure we keep that proprietary. Maybe it's because

175
00:09:36,720 --> 00:09:40,480
Speaker 3: our company works in an industry that's regulated, and therefore

176
00:09:40,720 --> 00:09:45,080
Speaker 3: we simply can't pass along the service level agreement and

177
00:09:45,120 --> 00:09:47,240
Speaker 3: we have to make sure that we can deliver fully

178
00:09:47,520 --> 00:09:49,679
Speaker 3: on the service and the promise that we sign up for.

179
00:09:50,000 --> 00:09:53,959
Speaker 3: Maybe it's something to do with sovereignty that you simply

180
00:09:54,720 --> 00:09:57,720
Speaker 3: in a particular country you have to have your own AI,

181
00:09:57,800 --> 00:10:00,559
Speaker 3: you have to control your own AI. Whatever those reasons are,

182
00:10:00,960 --> 00:10:03,440
Speaker 3: it could be cost reasons, but I think I think

183
00:10:03,440 --> 00:10:07,280
Speaker 3: that largely I would recommend people build their own AIS,

184
00:10:07,960 --> 00:10:10,079
Speaker 3: especially when they need to control it for whatever reason.

185
00:10:10,200 --> 00:10:12,600
Speaker 3: And so I think the future is going to have

186
00:10:12,679 --> 00:10:15,000
Speaker 3: lots and lots of use of AI that's closed and

187
00:10:15,720 --> 00:10:19,160
Speaker 3: AI that's open that you can build your own AI. Now,

188
00:10:19,200 --> 00:10:23,680
Speaker 3: one of the things that people, you know, misunderstand about

189
00:10:24,240 --> 00:10:28,240
Speaker 3: these open models is yes, you can host it yourself,

190
00:10:28,760 --> 00:10:30,800
Speaker 3: but you can build your own computer, but most people

191
00:10:30,840 --> 00:10:34,120
Speaker 3: use computers in a cloud. Frankly, I think close models

192
00:10:34,160 --> 00:10:36,640
Speaker 3: are cheaper. You know, if you don't have to build yourself,

193
00:10:36,720 --> 00:10:38,280
Speaker 3: if you don't have to train it yourself. It costs

194
00:10:38,280 --> 00:10:42,360
Speaker 3: a lot of expertise to fine tune and maintain and

195
00:10:42,440 --> 00:10:44,880
Speaker 3: guard rail and keep it safe and evaluate it and

196
00:10:45,320 --> 00:10:47,640
Speaker 3: of course even build computers to host it. So there's

197
00:10:47,679 --> 00:10:50,800
Speaker 3: nothing cheap about doing that. The reason why you need

198
00:10:50,880 --> 00:10:54,160
Speaker 3: open open models is because you need to have control,

199
00:10:54,360 --> 00:10:57,000
Speaker 3: because you need to adapt something for your own very

200
00:10:57,000 --> 00:11:00,840
Speaker 3: specialized use cases. And so I I think there's a

201
00:11:00,880 --> 00:11:03,960
Speaker 3: lot of misunderstanding about about closed versus open. We felt

202
00:11:03,960 --> 00:11:06,080
Speaker 3: that it was important for people to understand that there's

203
00:11:06,120 --> 00:11:06,800
Speaker 3: a world for.

204
00:11:06,840 --> 00:11:11,080
Speaker 2: Both weighted versus open source as well. There is a distinction.

205
00:11:12,520 --> 00:11:16,400
Speaker 3: I open open weighted as much as much as open

206
00:11:16,440 --> 00:11:18,960
Speaker 3: as that you can the more open. It is in

207
00:11:19,000 --> 00:11:21,920
Speaker 3: the way that we work. We put the weights out there,

208
00:11:22,000 --> 00:11:24,920
Speaker 3: we also teach people how to train the model from

209
00:11:24,960 --> 00:11:28,240
Speaker 3: the data that we also open source. And the reason

210
00:11:28,360 --> 00:11:31,360
Speaker 3: that is we want to enable you to completely reproduce

211
00:11:32,080 --> 00:11:35,160
Speaker 3: the AI model that we've open weighted, and so that

212
00:11:35,240 --> 00:11:37,560
Speaker 3: ability by us teaching you how to do that, you

213
00:11:37,600 --> 00:11:39,880
Speaker 3: can then do it for yourself. You know, I think,

214
00:11:40,040 --> 00:11:43,960
Speaker 3: I think the the idea that the world is going

215
00:11:44,000 --> 00:11:45,880
Speaker 3: to be one or the other is just completely wrong,

216
00:11:46,400 --> 00:11:50,679
Speaker 3: and and the idea that open models is somehow unsafe

217
00:11:50,760 --> 00:11:53,559
Speaker 3: is also fundamentally wrong, and so we just want to

218
00:11:53,559 --> 00:11:54,720
Speaker 3: make sure that people understand.

219
00:11:54,800 --> 00:11:57,160
Speaker 2: To finish our conversation, you know, the two big case

220
00:11:57,200 --> 00:12:00,320
Speaker 2: studies where the release of can communicate three which on

221
00:12:00,360 --> 00:12:03,160
Speaker 2: an open way to basis releases fully July twenty seventh,

222
00:12:03,679 --> 00:12:07,760
Speaker 2: and then the case study of two open AI models

223
00:12:07,800 --> 00:12:12,800
Speaker 2: mistakenly accessing hugging faces systems and hugging Face trying to

224
00:12:12,920 --> 00:12:15,480
Speaker 2: use an open model in its defense where the guard

225
00:12:15,559 --> 00:12:18,040
Speaker 2: rails were a factor. Would you just reflect on those

226
00:12:18,040 --> 00:12:21,080
Speaker 2: two I know that you've been asked about them. They

227
00:12:21,120 --> 00:12:23,119
Speaker 2: seem to be like really big moments in AI.

228
00:12:22,920 --> 00:12:26,720
Speaker 3: Over those are perfectly perfect canonical examples. Just because something

229
00:12:26,840 --> 00:12:31,480
Speaker 3: is closed doesn't necessarily therefore make it safe or secure.

230
00:12:31,760 --> 00:12:35,280
Speaker 3: It is possible for a model to be jailbroken, as

231
00:12:35,280 --> 00:12:37,800
Speaker 3: possible for a model to be if you will, stolen.

232
00:12:37,840 --> 00:12:40,360
Speaker 3: It could be possible that that somehow has leaked from

233
00:12:40,360 --> 00:12:44,319
Speaker 3: the inside. It's possible that the guardrails or the sandboxes

234
00:12:44,640 --> 00:12:48,800
Speaker 3: of an AI closed AI model wasn't properly engineered, and

235
00:12:48,840 --> 00:12:54,320
Speaker 3: as a result, it was able to attack another company

236
00:12:54,360 --> 00:12:57,120
Speaker 3: in some way. And so just because something is closed,

237
00:12:57,120 --> 00:13:00,440
Speaker 3: and just because something is proprietary, doesn't necessary really make

238
00:13:00,480 --> 00:13:04,560
Speaker 3: it secure and safe. Of course, thank goodness, we have

239
00:13:04,720 --> 00:13:08,840
Speaker 3: two companies, well, I guess more than that several companies

240
00:13:08,880 --> 00:13:12,840
Speaker 3: that build closed AI models, and these are extraordinary technology companies,

241
00:13:12,880 --> 00:13:14,559
Speaker 3: and they're doing their best to keep it safe and

242
00:13:14,640 --> 00:13:18,200
Speaker 3: keep it secure. But it is also the cainetical case

243
00:13:18,840 --> 00:13:22,000
Speaker 3: that single points of failure is where we have the

244
00:13:22,040 --> 00:13:25,880
Speaker 3: greatest vulnerability. We cannot have single points of failure as

245
00:13:25,880 --> 00:13:30,119
Speaker 3: an industry, as a world, we should have distributed, massively

246
00:13:30,160 --> 00:13:33,760
Speaker 3: distributed self defense. And so in the case of in

247
00:13:33,800 --> 00:13:36,160
Speaker 3: the case of the example you just mentioned, hugging Face

248
00:13:36,600 --> 00:13:39,959
Speaker 3: thankfully was able to access an open model and I

249
00:13:40,000 --> 00:13:43,000
Speaker 3: think they used GLM five point two. As my understanding,

250
00:13:44,160 --> 00:13:47,400
Speaker 3: they couldn't get a proprietary model, they could not get

251
00:13:47,440 --> 00:13:50,880
Speaker 3: a closed model to help them figure out what happened.

252
00:13:51,840 --> 00:13:54,079
Speaker 3: But this is exactly the reason why you want to

253
00:13:54,120 --> 00:13:56,920
Speaker 3: have open models, because in that case, they use GLM

254
00:13:56,920 --> 00:14:00,319
Speaker 3: five point two to identify where the vulnerability was, where

255
00:14:00,320 --> 00:14:03,800
Speaker 3: the penetration was, and we're able to quickly identify them

256
00:14:03,800 --> 00:14:06,360
Speaker 3: and patch it up. And so this is a perfect

257
00:14:06,440 --> 00:14:11,199
Speaker 3: example of self defense that's necessary, as a perfect example

258
00:14:11,480 --> 00:14:15,640
Speaker 3: of diversity of AI technology being necessary, and so a

259
00:14:15,720 --> 00:14:19,520
Speaker 3: perfect example of why open models and open capabilities for

260
00:14:19,640 --> 00:14:22,080
Speaker 3: self defense is really important.

261
00:14:22,240 --> 00:14:25,359
Speaker 1: That's in Nvidia CEO Jensen Huang in a special conversation

262
00:14:25,480 --> 00:14:28,320
Speaker 1: with Bloomberg Tech host ed Ludlow. You can watch the

263
00:14:28,400 --> 00:14:31,520
Speaker 1: full interview now at Bloomberg dot com, slash Videos and

264
00:14:31,600 --> 00:14:34,560
Speaker 1: on the Bloomberg Business app. You can also listen by

265
00:14:34,560 --> 00:14:38,640
Speaker 1: subscribing to the Bloomberg Tech podcast feed. I'm Monica Ricks,

266
00:14:38,960 --> 00:14:41,119
Speaker 1: thanks for listening. This is Bloomberg