1
00:00:02,520 --> 00:00:05,439
Speaker 1: Bloomberg Audio Studios, podcasts.

2
00:00:05,760 --> 00:00:10,479
Speaker 2: Radio news man Deep Singh joins us, Right now, did

3
00:00:10,520 --> 00:00:14,280
Speaker 2: you know what deep seek was before twelve or eighteen

4
00:00:14,400 --> 00:00:15,040
Speaker 2: hours ago?

5
00:00:15,400 --> 00:00:18,160
Speaker 3: I did, yes, And so what did you know about

6
00:00:18,200 --> 00:00:19,280
Speaker 3: them two days ago?

7
00:00:20,079 --> 00:00:23,160
Speaker 1: That it is a competitor in the LM space. They've

8
00:00:23,160 --> 00:00:26,520
Speaker 1: been training their own model and it's used mostly in

9
00:00:26,560 --> 00:00:27,800
Speaker 1: the East Asian region.

10
00:00:27,960 --> 00:00:30,560
Speaker 2: Okay, what I learned, thank you zero head for this

11
00:00:30,600 --> 00:00:35,720
Speaker 2: from Morgan Brown at dropbox is basically they've said, we're

12
00:00:35,720 --> 00:00:38,199
Speaker 2: not going to be as perfect as pristine as the

13
00:00:38,280 --> 00:00:42,040
Speaker 2: others out to thirty two decimal points. We're just going

14
00:00:42,080 --> 00:00:47,560
Speaker 2: to get it done out to point eight decimal points.

15
00:00:47,680 --> 00:00:49,239
Speaker 2: Is that what this is all about is they're not

16
00:00:49,240 --> 00:00:50,120
Speaker 2: going to be perfect.

17
00:00:50,479 --> 00:00:52,760
Speaker 1: I think there's more to it than just you know,

18
00:00:53,000 --> 00:00:57,240
Speaker 1: using the floating point fewer floating point operations than open

19
00:00:57,280 --> 00:01:01,279
Speaker 1: AI and Entropic and others. And in this case, they

20
00:01:01,800 --> 00:01:06,679
Speaker 1: basically focused on hardware efficiency, using the hardware in the

21
00:01:06,720 --> 00:01:10,399
Speaker 1: most efficient fashion. They had the benefit of all these

22
00:01:10,600 --> 00:01:14,160
Speaker 1: lms being out there. They use meta Llama as a

23
00:01:14,200 --> 00:01:17,880
Speaker 1: reference point. It's another open source model, and they figured

24
00:01:17,880 --> 00:01:21,040
Speaker 1: out a way to do this most efficiently than everyone else.

25
00:01:21,080 --> 00:01:24,360
Speaker 1: I mean, everyone is focused on scale right now, they

26
00:01:24,560 --> 00:01:26,240
Speaker 1: focused on hardware efficiency.

27
00:01:26,240 --> 00:01:27,720
Speaker 3: Call the Lama meta.

28
00:01:27,760 --> 00:01:31,800
Speaker 2: It's Facebook is the grandson of the original Lama on

29
00:01:31,840 --> 00:01:32,360
Speaker 2: Saturday Night.

30
00:01:32,680 --> 00:01:36,280
Speaker 4: There you go, right, very good man, Deep You've got

31
00:01:36,280 --> 00:01:39,440
Speaker 4: an analyst Bloomberg Intelligence in Hong Kong, Robert Lee. I'm

32
00:01:39,480 --> 00:01:42,480
Speaker 4: reading his research literally as we speak. And for folks

33
00:01:42,880 --> 00:01:45,600
Speaker 4: on the Bloomberg terminal, go to big and check out

34
00:01:45,600 --> 00:01:46,920
Speaker 4: if you want to need to figure out what's going

35
00:01:46,959 --> 00:01:48,960
Speaker 4: on with Deep seek and what China's doing with AI.

36
00:01:49,520 --> 00:01:53,920
Speaker 4: Bloomerg Intelligence has got the data and the research there.

37
00:01:55,080 --> 00:01:59,520
Speaker 4: Can China support a real competitive AI environment? It seems

38
00:01:59,520 --> 00:02:04,000
Speaker 4: like they're government is more restricting some of the technology there.

39
00:02:04,360 --> 00:02:08,080
Speaker 1: Well, so now we know the export controls probably weren't

40
00:02:08,200 --> 00:02:11,480
Speaker 1: as effective as they were supposed to be here, you know,

41
00:02:11,560 --> 00:02:14,440
Speaker 1: in terms of restricting the access to Nvidia latest chips,

42
00:02:14,800 --> 00:02:17,799
Speaker 1: and some people are saying they did use you know,

43
00:02:17,960 --> 00:02:21,680
Speaker 1: the latest h one hundred chips, albeit they were not

44
00:02:21,800 --> 00:02:25,000
Speaker 1: of the same scale as were available to Open AI

45
00:02:25,160 --> 00:02:28,959
Speaker 1: and the others. But look, we don't know the extent

46
00:02:29,120 --> 00:02:32,280
Speaker 1: of the hardware that was used for training. All I

47
00:02:32,320 --> 00:02:34,680
Speaker 1: can say, you know, based on this the fact that

48
00:02:34,720 --> 00:02:37,760
Speaker 1: they have a comparable model in performance to the others.

49
00:02:38,240 --> 00:02:42,279
Speaker 1: It shows that they clearly have a combination of algorithms

50
00:02:42,280 --> 00:02:45,239
Speaker 1: and compute to compete with the others.

51
00:02:45,320 --> 00:02:48,240
Speaker 2: Most of the financials are the revenue growth models that

52
00:02:48,320 --> 00:02:52,040
Speaker 2: you and b I have. Of all these fancy AI people,

53
00:02:52,160 --> 00:02:53,440
Speaker 2: are they now at risk?

54
00:02:53,760 --> 00:02:54,359
Speaker 3: Absolutely?

55
00:02:54,360 --> 00:02:57,120
Speaker 1: I mean, look at open ais oh one pro model

56
00:02:57,160 --> 00:03:01,560
Speaker 1: they're charging two hundred dollars per month because it's the

57
00:03:01,560 --> 00:03:04,520
Speaker 1: best model they have to offer. The fact that you

58
00:03:04,600 --> 00:03:07,639
Speaker 1: have an open source model that is comparable to the

59
00:03:07,760 --> 00:03:09,600
Speaker 1: model that Opening Eyes chargeable.

60
00:03:09,840 --> 00:03:11,720
Speaker 3: You're saying it's comparable.

61
00:03:11,400 --> 00:03:13,480
Speaker 1: I mean it is, yeah, I mean look at the

62
00:03:13,520 --> 00:03:16,799
Speaker 1: benchmarks and that's why you know all these models have

63
00:03:16,919 --> 00:03:20,480
Speaker 1: common benchmarks. This is within one to two percentage points

64
00:03:20,560 --> 00:03:21,400
Speaker 1: of that benchmark.

65
00:03:21,480 --> 00:03:23,120
Speaker 3: I know you don't do buy hole cell, but are

66
00:03:23,120 --> 00:03:24,760
Speaker 3: we going to sell here? Are we going to see

67
00:03:24,800 --> 00:03:27,560
Speaker 3: Wall Street put a cell on an video? I think

68
00:03:27,639 --> 00:03:28,080
Speaker 3: predict that.

69
00:03:28,800 --> 00:03:31,040
Speaker 1: I mean right now, the fact that Meta raised their

70
00:03:31,120 --> 00:03:35,040
Speaker 1: Capex on Friday to me earning season is where we

71
00:03:35,080 --> 00:03:37,600
Speaker 1: will find out what all these companies end up doing.

72
00:03:37,680 --> 00:03:40,560
Speaker 1: So you know, all this has happened in the last

73
00:03:40,920 --> 00:03:44,080
Speaker 1: forty eight hours, where Meta raised their CAPEX Deep Sea

74
00:03:44,200 --> 00:03:46,880
Speaker 1: came out it suddenly everyone is going crazy.

75
00:03:46,960 --> 00:03:50,200
Speaker 2: Paul the President lined up with the son of Japan

76
00:03:50,720 --> 00:03:52,760
Speaker 2: to do a ginormous tech deal as well.

77
00:03:52,920 --> 00:03:54,440
Speaker 3: You okay, there, you look at you.

78
00:03:55,000 --> 00:03:59,080
Speaker 4: I'm doing social Here are you doing social words? Doing

79
00:03:59,120 --> 00:04:02,120
Speaker 4: that out the social I'm saying, take a look at that. So, Mandy,

80
00:04:02,200 --> 00:04:05,680
Speaker 4: what do we do here? As we think about AI's

81
00:04:05,720 --> 00:04:09,600
Speaker 4: obviously from technology perspective, it has been the story for

82
00:04:09,640 --> 00:04:12,680
Speaker 4: the last two years at least. How do you think

83
00:04:12,680 --> 00:04:15,600
Speaker 4: about it now? Has anything changed in the last twenty

84
00:04:15,640 --> 00:04:16,359
Speaker 4: four hours for you?

85
00:04:16,960 --> 00:04:20,679
Speaker 1: A lot of software companies will feel, you know, they

86
00:04:20,800 --> 00:04:23,479
Speaker 1: can do their own AI now and not having to

87
00:04:23,560 --> 00:04:28,120
Speaker 1: rely on the big hyperscalers because the narrative was if

88
00:04:28,160 --> 00:04:32,120
Speaker 1: you can't spend upfront on Capex, then you don't have

89
00:04:32,200 --> 00:04:34,640
Speaker 1: a play. And it comes to the foundational model there

90
00:04:35,000 --> 00:04:38,240
Speaker 1: and so suddenly everyone feels empowered that they can do

91
00:04:38,320 --> 00:04:40,960
Speaker 1: their own AI with this development, I.

92
00:04:40,920 --> 00:04:44,039
Speaker 4: See Nvidio, Broadcom, some of these big chip makers that

93
00:04:44,080 --> 00:04:46,000
Speaker 4: have had such a huge run on the IP side,

94
00:04:46,120 --> 00:04:49,760
Speaker 4: they're down ten to eleven twelve percent here today? Is

95
00:04:49,839 --> 00:04:52,680
Speaker 4: that realistic? Is that does that seem reasonable to you?

96
00:04:53,080 --> 00:04:55,640
Speaker 1: I mean with semi companies we know the risk is

97
00:04:55,680 --> 00:04:59,159
Speaker 1: always that cyclical element. And if we are calling that

98
00:04:59,279 --> 00:05:01,640
Speaker 1: this is the the top of the cycle in terms

99
00:05:01,680 --> 00:05:05,479
Speaker 1: of semi demand and estimates aren't going up anymore, then

100
00:05:05,600 --> 00:05:07,800
Speaker 1: you know you will see this sort of stock reaction.

101
00:05:07,880 --> 00:05:09,920
Speaker 1: But I go back to my point about Meta raising

102
00:05:09,960 --> 00:05:13,279
Speaker 1: their capex projects target at it one hundred billion.

103
00:05:13,279 --> 00:05:15,960
Speaker 5: Of that, we haven't talked to you about that fifty

104
00:05:16,000 --> 00:05:20,080
Speaker 5: billion they were Meta Cappex was twenty five billion in

105
00:05:20,160 --> 00:05:23,480
Speaker 5: twenty nineteen, then it was fifty billion for this year.

106
00:05:23,920 --> 00:05:26,680
Speaker 4: Now they're saying to go on a sixty five I

107
00:05:26,760 --> 00:05:28,600
Speaker 4: feel like it when I look at those numbers, it

108
00:05:28,600 --> 00:05:32,240
Speaker 4: feels like drunken sale or time. What Did they have

109
00:05:32,320 --> 00:05:34,760
Speaker 4: a real strategy behind that spending or are they just

110
00:05:34,760 --> 00:05:36,960
Speaker 4: saying we need to be in this game, We need

111
00:05:37,000 --> 00:05:39,520
Speaker 4: to spend whatever we need to spend, typical Meta type

112
00:05:39,520 --> 00:05:40,000
Speaker 4: of spending.

113
00:05:40,200 --> 00:05:42,960
Speaker 1: Yeah, and I think there is that aspect where you

114
00:05:43,120 --> 00:05:45,559
Speaker 1: want to be that front end player when it comes

115
00:05:45,560 --> 00:05:48,440
Speaker 1: to the lllms. But in the case of Meta, the challenge,

116
00:05:48,880 --> 00:05:50,880
Speaker 1: in addition to the fact that you know deep sea

117
00:05:51,000 --> 00:05:53,440
Speaker 1: model is out there, is they don't have a cloud business.

118
00:05:53,520 --> 00:05:55,800
Speaker 1: You look at Microsoft, you look at Amazon, you look

119
00:05:55,839 --> 00:05:58,279
Speaker 1: at Google, they don't they have a cloud business to

120
00:05:58,400 --> 00:06:01,680
Speaker 1: monetize those GPUs. Even if let's say deep Sea came

121
00:06:01,760 --> 00:06:05,080
Speaker 1: up with a better model, Microsoft can use that capacity

122
00:06:05,080 --> 00:06:06,440
Speaker 1: for inferencing on the cloud.

123
00:06:06,600 --> 00:06:10,000
Speaker 3: They can generate cloud revenue. Mandy, what are the meetings

124
00:06:10,200 --> 00:06:12,880
Speaker 3: like in Silicon Valley? Now?

125
00:06:13,279 --> 00:06:15,600
Speaker 2: I mean, these guys never get up before nine am,

126
00:06:15,839 --> 00:06:18,160
Speaker 2: But today they're gonna get up at six am their time.

127
00:06:18,400 --> 00:06:21,760
Speaker 2: They're gonna be rock at nine am surveillance time. What

128
00:06:21,800 --> 00:06:24,599
Speaker 2: are the meetings gonna be like at Google? What's the

129
00:06:24,640 --> 00:06:26,440
Speaker 2: meeting gonna be like for Zuckerberg?

130
00:06:27,160 --> 00:06:30,279
Speaker 1: Oh? So I think again, Given these companies are reporting

131
00:06:30,320 --> 00:06:33,400
Speaker 1: earnings next week right now, they have to figure out

132
00:06:33,839 --> 00:06:35,960
Speaker 1: what is it that they relate to the investors in

133
00:06:36,040 --> 00:06:39,520
Speaker 1: terms of the scaling laws. Like up until now, the

134
00:06:39,560 --> 00:06:43,159
Speaker 1: biggest debate was will the scaling laws hold in twenty

135
00:06:43,200 --> 00:06:47,280
Speaker 1: twenty five? And based on this development, we don't talk

136
00:06:47,320 --> 00:06:48,200
Speaker 1: about scaling laws.

137
00:06:48,279 --> 00:06:51,159
Speaker 3: Do they replicate what deep seek is doing?

138
00:06:51,839 --> 00:06:55,080
Speaker 1: Absolutely? I think they will use some of the innovations

139
00:06:55,120 --> 00:06:57,400
Speaker 1: that deep Seek has showed in their paper, and given

140
00:06:57,440 --> 00:07:00,720
Speaker 1: they have open source it. My bed is everyone cares

141
00:07:00,720 --> 00:07:04,520
Speaker 1: about hardware efficiency, even if Microsoft is spending eighty billion

142
00:07:04,560 --> 00:07:07,640
Speaker 1: on AI Capex. They want to use that cap ex efficiency.

143
00:07:07,720 --> 00:07:11,240
Speaker 2: The smartest thing I've heard of from the technologist Paul Sweeney,

144
00:07:11,400 --> 00:07:13,880
Speaker 2: your Ted talk on this right, your Ted talk.

145
00:07:14,120 --> 00:07:16,680
Speaker 3: You almost wore a tie to your Ted thing. Exactly.

146
00:07:17,080 --> 00:07:18,520
Speaker 3: They're drunken sailors.

147
00:07:18,680 --> 00:07:21,080
Speaker 2: I mean, that's the smartest thing they've heard. Do you

148
00:07:21,120 --> 00:07:22,880
Speaker 2: see capex responsibility?

149
00:07:23,360 --> 00:07:24,240
Speaker 3: No? I think.

150
00:07:24,920 --> 00:07:29,160
Speaker 1: I mean to my mind, all this Capex can be used,

151
00:07:29,240 --> 00:07:32,000
Speaker 1: you know, for the future, and it's quite fungible. It's

152
00:07:32,040 --> 00:07:34,880
Speaker 1: not like laying a fiber network and oh it's gonna

153
00:07:34,920 --> 00:07:39,880
Speaker 1: go waste. Compute is always used in some form. Now, yes,

154
00:07:39,960 --> 00:07:43,040
Speaker 1: they may have overpaid for the GPUs. You could argue

155
00:07:43,560 --> 00:07:45,920
Speaker 1: was it worth paying thirty thousand dollars for a GPU?

156
00:07:46,080 --> 00:07:48,840
Speaker 1: Probably not, but still this compute is still useful. Man.

157
00:07:48,880 --> 00:07:51,120
Speaker 2: Deep single those folks. We're gonna go around equities, bond's

158
00:07:51,160 --> 00:07:53,400
Speaker 2: currencies come out of these again. We're off four percent.

159
00:07:53,840 --> 00:07:56,960
Speaker 3: No, sit down, you're not done yet. You know where

160
00:07:57,040 --> 00:07:57,520
Speaker 3: you're going.

161
00:07:58,240 --> 00:07:59,320
Speaker 4: Two segments you should.

162
00:07:59,080 --> 00:08:01,720
Speaker 3: See as an englishakfast on Monday. It's like this huge

163
00:08:01,760 --> 00:08:02,800
Speaker 3: kind of a thing.

164
00:08:02,960 --> 00:08:06,640
Speaker 2: The bond market yields in big time, which you'd expect

165
00:08:06,720 --> 00:08:10,120
Speaker 2: here the screening yield four point one eight percent in

166
00:08:10,320 --> 00:08:14,520
Speaker 2: nine basis points, the tenure yield in eleven basis points.

167
00:08:14,560 --> 00:08:17,360
Speaker 2: First thing, I looked at the real yield crushes in

168
00:08:17,520 --> 00:08:20,400
Speaker 2: ten basis points from a two point two zero to

169
00:08:20,400 --> 00:08:23,000
Speaker 2: two point one zero some of the angst there in

170
00:08:23,040 --> 00:08:23,680
Speaker 2: the market.

171
00:08:23,880 --> 00:08:25,560
Speaker 3: And what I look at, and it's not a good

172
00:08:25,600 --> 00:08:26,240
Speaker 3: number yet.

173
00:08:26,120 --> 00:08:29,000
Speaker 2: With the Vicks twenty one point five four, not thirty,

174
00:08:29,080 --> 00:08:31,040
Speaker 2: but from fifteen to twenty one.

175
00:08:31,040 --> 00:08:33,719
Speaker 4: That gets your attention, poll Man, Deep is a concern here.

176
00:08:34,120 --> 00:08:37,760
Speaker 4: The deep Seak product could be as good an AI

177
00:08:37,880 --> 00:08:40,920
Speaker 4: solution as what some of the Western companies are providing,

178
00:08:41,280 --> 00:08:43,880
Speaker 4: but at a lower cost. Is that the bottom line here?

179
00:08:44,000 --> 00:08:46,400
Speaker 1: Yeah? I mean look at the app stores that deep

180
00:08:46,400 --> 00:08:49,679
Speaker 1: Seak is the top appro right now here in the US,

181
00:08:49,760 --> 00:08:53,800
Speaker 1: and everyone is using it multitude chat.

182
00:08:54,360 --> 00:08:56,160
Speaker 3: No, I don't mean ititer up, but this is important.

183
00:08:56,559 --> 00:09:00,959
Speaker 2: Lesa Matao's not using it, Tom Keene is not using it.

184
00:09:01,480 --> 00:09:04,720
Speaker 2: An adult like you or Joe Wisenthal who's knee deep

185
00:09:04,720 --> 00:09:05,360
Speaker 2: into this stuff.

186
00:09:05,400 --> 00:09:09,280
Speaker 3: Good morning, Joe, Joey. Joe's not up yet, Okay, Joe

187
00:09:09,280 --> 00:09:12,960
Speaker 3: Wisenthal is going to bring up deep seek and bring up.

188
00:09:13,040 --> 00:09:16,400
Speaker 2: Chet GPT of one of the different managed levels. Can

189
00:09:16,440 --> 00:09:17,400
Speaker 2: you tell the difference?

190
00:09:18,559 --> 00:09:21,800
Speaker 1: Look, I mean there is an element of personalization that

191
00:09:21,920 --> 00:09:24,880
Speaker 1: these apps pro wide when it comes to asking the

192
00:09:24,920 --> 00:09:27,560
Speaker 1: type of questions you're asking. But if you have a

193
00:09:27,600 --> 00:09:30,400
Speaker 1: generic query, deep seek is going to give you an

194
00:09:30,440 --> 00:09:32,240
Speaker 1: answer that's comparable to Chat GPT.

195
00:09:32,679 --> 00:09:36,839
Speaker 4: I'm looking Tom, tell me you're on deep Seek Now.

196
00:09:36,880 --> 00:09:41,640
Speaker 4: I google the top app number one, deep Seek, number

197
00:09:41,640 --> 00:09:47,400
Speaker 4: two Chat GPT, and number three Paramount plus for land

198
00:09:47,480 --> 00:09:50,160
Speaker 4: Man and for Yellowstone. That's what I'm talking about.

199
00:09:49,880 --> 00:09:51,960
Speaker 2: That we had a land Man weekd at home. I said,

200
00:09:51,960 --> 00:09:53,559
Speaker 2: Sweeney's liked Deep into.

201
00:09:53,400 --> 00:09:54,120
Speaker 4: The number seven.

202
00:09:54,200 --> 00:09:57,600
Speaker 3: Fox Sports get one more in here, one more in here,

203
00:09:57,640 --> 00:09:59,080
Speaker 3: because man, Deep's coming back at eight.

204
00:09:59,280 --> 00:10:02,520
Speaker 4: What do you need? What do you think you're gonna

205
00:10:02,520 --> 00:10:04,720
Speaker 4: hear from some of these technology companies that next week

206
00:10:04,760 --> 00:10:07,680
Speaker 4: as they talk about China, as they talk about this,

207
00:10:08,360 --> 00:10:09,480
Speaker 4: what is a deep Seek thing?

208
00:10:09,840 --> 00:10:10,040
Speaker 3: Yeah?

209
00:10:10,080 --> 00:10:13,920
Speaker 1: I mean the whole aspect around CAPEX and scaling laws

210
00:10:13,960 --> 00:10:17,560
Speaker 1: and what to do with the latest GPUs. Are you

211
00:10:17,800 --> 00:10:22,200
Speaker 1: spending more on training or inferencing? Everything is on the

212
00:10:22,240 --> 00:10:24,479
Speaker 1: table in terms of hardware efficiency.

213
00:10:24,679 --> 00:10:27,520
Speaker 2: I have no idea what inferences is a clinic for

214
00:10:27,600 --> 00:10:30,240
Speaker 2: the man deep sing of Bloomberg intelligence