1
00:00:00,480 --> 00:00:03,600
Speaker 1: Bryan Bridge Tech Companies while they've been up and down

2
00:00:03,640 --> 00:00:05,640
Speaker 1: like a Yo yo this week since the Deep Seek

3
00:00:05,720 --> 00:00:09,880
Speaker 1: thing happened. Interesting timing because Microsoft and Meta have reported

4
00:00:09,960 --> 00:00:13,480
Speaker 1: their annual results today. Sam Dicky from Fisher Funds is

5
00:00:13,480 --> 00:00:16,200
Speaker 1: with us. Sam. Gooday, good evening, right, how are you

6
00:00:16,200 --> 00:00:19,280
Speaker 1: going good? Thank you? Quick throwback? What was the straw

7
00:00:19,320 --> 00:00:20,599
Speaker 1: that broke the camel's back here?

8
00:00:22,040 --> 00:00:24,440
Speaker 2: Yeah, that's an interesting question because it wasn't all new news,

9
00:00:24,480 --> 00:00:28,320
Speaker 2: so that deep Seak has been openly using these innovative

10
00:00:28,360 --> 00:00:31,320
Speaker 2: techniques to drive down the cost to train these AI

11
00:00:31,400 --> 00:00:34,199
Speaker 2: models over the last year or so. But it was

12
00:00:34,240 --> 00:00:36,000
Speaker 2: kind a threefold. It was the release of its most

13
00:00:36,040 --> 00:00:39,280
Speaker 2: recent model R one on twenty fifth January, which showed

14
00:00:39,520 --> 00:00:42,080
Speaker 2: the significantly improved results, which you guys have spoken about

15
00:00:42,120 --> 00:00:44,640
Speaker 2: this week, plus the talking heads picked it up at

16
00:00:44,720 --> 00:00:48,720
Speaker 2: Davos later that week. Plus critically, the deep Sea Gap

17
00:00:48,760 --> 00:00:51,560
Speaker 2: went from number thirty one on the Apple lapp store

18
00:00:51,600 --> 00:00:54,000
Speaker 2: to number one over the weekend, so that what really

19
00:00:54,080 --> 00:00:56,840
Speaker 2: was the straw that broke the camel's back? The consumer

20
00:00:57,040 --> 00:00:57,960
Speaker 2: voted with their feet.

21
00:00:59,680 --> 00:01:03,400
Speaker 1: So the significant breakthrough, obviously is that they managed to

22
00:01:03,440 --> 00:01:06,839
Speaker 1: do this some well, a lot of questions been asked

23
00:01:06,840 --> 00:01:09,520
Speaker 1: about how much US tech was used to make this,

24
00:01:10,120 --> 00:01:14,160
Speaker 1: how palatable the platform will be for American users or

25
00:01:14,200 --> 00:01:16,280
Speaker 1: America you know, users around the rest of the world

26
00:01:16,319 --> 00:01:17,279
Speaker 1: outside of China.

27
00:01:18,680 --> 00:01:20,160
Speaker 2: That's right. I mean one of the breakthrough is it

28
00:01:20,200 --> 00:01:22,760
Speaker 2: is important to talk about one of the breakthroughs. There

29
00:01:22,800 --> 00:01:25,959
Speaker 2: were many, but the simplest one to understand, the one

30
00:01:25,959 --> 00:01:29,759
Speaker 2: that stood out was pure reinforcement learning. So in playing English,

31
00:01:29,840 --> 00:01:33,080
Speaker 2: pure reinforcement learning is quite simple. So don't teach one

32
00:01:33,120 --> 00:01:35,679
Speaker 2: of these AI models to solve every puzzle on Earth

33
00:01:35,760 --> 00:01:38,360
Speaker 2: using every bit of data on earth, like how most

34
00:01:38,360 --> 00:01:41,959
Speaker 2: of the US large language models have been trained. Just

35
00:01:42,120 --> 00:01:44,080
Speaker 2: reward it if it does a good job with a

36
00:01:44,080 --> 00:01:46,760
Speaker 2: digital pad on the back, so a plus one for

37
00:01:46,800 --> 00:01:48,880
Speaker 2: the right answer and a minus one for the wrong answer.

38
00:01:48,920 --> 00:01:52,160
Speaker 2: And this sort of trial and error way of learning

39
00:01:52,160 --> 00:01:54,440
Speaker 2: has proved to be far more efficient and cost effective

40
00:01:54,440 --> 00:01:56,280
Speaker 2: as the model, just like we do, loves to reward

41
00:01:56,320 --> 00:01:58,560
Speaker 2: and kind of sort them out. In terms of the

42
00:01:58,880 --> 00:02:01,440
Speaker 2: controversy and people saying they don't believe the six million

43
00:02:01,480 --> 00:02:05,400
Speaker 2: dollar cost to train the model, and the controversy around

44
00:02:05,440 --> 00:02:10,680
Speaker 2: open AI saying that deep Seat piggybacked off, it's existing

45
00:02:11,400 --> 00:02:14,160
Speaker 2: chat GPT models. We don't know. We just know that

46
00:02:14,200 --> 00:02:17,200
Speaker 2: the arms race and the chipboard that you and I

47
00:02:17,240 --> 00:02:19,799
Speaker 2: spoke about late last years is really hotting up.

48
00:02:20,320 --> 00:02:23,320
Speaker 1: So basically what you're talking about, they're the pure reinforcement

49
00:02:23,400 --> 00:02:28,520
Speaker 1: is like positive reinforcement. How did the big, biggest, you know,

50
00:02:28,639 --> 00:02:31,880
Speaker 1: AI companies in the world missed this, Well, that's right,

51
00:02:31,919 --> 00:02:32,960
Speaker 1: they were certainly aware of it.

52
00:02:33,000 --> 00:02:36,880
Speaker 2: So Google, for example, used pure reinforcement learning. So the

53
00:02:36,919 --> 00:02:39,080
Speaker 2: word pure, by the way, is a lot of the

54
00:02:39,120 --> 00:02:43,800
Speaker 2: companies use human feedback reinforcement learning learning, but this is

55
00:02:43,880 --> 00:02:48,400
Speaker 2: sort of digital reinforcement learning with no human feedback or oversight.

56
00:02:49,320 --> 00:02:51,600
Speaker 2: Google used that technique with one of its much older

57
00:02:51,639 --> 00:02:55,440
Speaker 2: models called Alpha Go in twenty seventeen. But this pure

58
00:02:55,480 --> 00:02:59,880
Speaker 2: reinforcement learning technique historically had shortcomings and western which we

59
00:02:59,880 --> 00:03:01,320
Speaker 2: can go into it in a minute. But wesn AI

60
00:03:01,360 --> 00:03:04,000
Speaker 2: companies either weren't quick enough to solve those shortcomings or

61
00:03:04,320 --> 00:03:07,160
Speaker 2: probably weren't incentivized to solve them. Given it was an

62
00:03:07,160 --> 00:03:11,880
Speaker 2: AI boom, capital was freely available, and they had access

63
00:03:11,919 --> 00:03:15,360
Speaker 2: to cutting edge AI chips, which the Chinese did not.

64
00:03:16,360 --> 00:03:19,919
Speaker 2: And it seems like perversely the US restricting China from

65
00:03:20,200 --> 00:03:23,880
Speaker 2: accessing these cutting edge computer chips to train the AI

66
00:03:23,960 --> 00:03:27,760
Speaker 2: models forced deep Seek's hand, force them into a corner

67
00:03:27,760 --> 00:03:29,080
Speaker 2: and force them to get there quicker.

68
00:03:31,040 --> 00:03:34,840
Speaker 1: So perverse indeed, so can you just because I'm still

69
00:03:34,840 --> 00:03:37,800
Speaker 1: struggling a bit with pure reinforcement earning. Is this a

70
00:03:37,920 --> 00:03:41,720
Speaker 1: person is using the AI app and the machine gets

71
00:03:41,760 --> 00:03:44,080
Speaker 1: an answer right and you take, yes, you've got that right,

72
00:03:44,120 --> 00:03:46,840
Speaker 1: and then understands that it's doing something right or is

73
00:03:46,840 --> 00:03:48,160
Speaker 1: this a different process?

74
00:03:48,800 --> 00:03:51,200
Speaker 2: No, So that that's the inferencing, that's when we're using that.

75
00:03:51,280 --> 00:03:53,680
Speaker 2: This is the pre training, so the training of the model.

76
00:03:53,760 --> 00:03:57,680
Speaker 2: So really simplistically, you know that the models have been

77
00:03:57,720 --> 00:04:00,680
Speaker 2: getting bigger and bigger, this is excluding deep Seek, and

78
00:04:00,720 --> 00:04:03,800
Speaker 2: they've been getting trained on you know, most of the

79
00:04:03,840 --> 00:04:06,000
Speaker 2: text on the internet in the world, and a lot

80
00:04:06,040 --> 00:04:08,000
Speaker 2: of the photos and a lot of the numbers and

81
00:04:08,000 --> 00:04:11,040
Speaker 2: a lot of the videos, and they've been getting taught

82
00:04:11,080 --> 00:04:14,720
Speaker 2: to solve every puzzle on Earth using every bit of data.

83
00:04:15,400 --> 00:04:19,600
Speaker 2: But this, this deep Seek technique was, like I said,

84
00:04:19,600 --> 00:04:21,599
Speaker 2: like getting a digital pad on the back when it

85
00:04:21,839 --> 00:04:23,960
Speaker 2: got an answer right when it was training the model.

86
00:04:24,000 --> 00:04:26,520
Speaker 2: So not when we're influencing the model. I kind of

87
00:04:26,520 --> 00:04:29,560
Speaker 2: think of it as like, instead of studying for a

88
00:04:29,560 --> 00:04:32,400
Speaker 2: maths exam by going through all the mass textbooks, you

89
00:04:32,680 --> 00:04:34,880
Speaker 2: just do practice exams and get better and better and

90
00:04:34,880 --> 00:04:37,240
Speaker 2: better marks until eventually you're getting sort of ninety eight

91
00:04:37,240 --> 00:04:40,200
Speaker 2: percent the practice exams, but you don't bother studying in

92
00:04:40,240 --> 00:04:41,400
Speaker 2: any of the source material.

93
00:04:41,960 --> 00:04:44,880
Speaker 1: Wow, and then you're just reinforcing, like teaching your dog

94
00:04:44,960 --> 00:04:49,440
Speaker 1: how to set with a piece of every time. Okay,

95
00:04:49,720 --> 00:04:52,279
Speaker 1: so let's talk Microsoft and MESA. How did their results

96
00:04:52,320 --> 00:04:54,040
Speaker 1: look today? What are they saying about all of this?

97
00:04:55,480 --> 00:04:58,720
Speaker 2: Yeah, so they Microsoft said that this happens every technology

98
00:04:58,760 --> 00:05:02,360
Speaker 2: technology cycle, and we see sort of ten x improvements

99
00:05:02,440 --> 00:05:05,279
Speaker 2: and costs and efficiency. In other words, costs come down

100
00:05:05,320 --> 00:05:09,640
Speaker 2: by ninety percent in most tech cycles. And they also

101
00:05:09,680 --> 00:05:12,279
Speaker 2: said Deep Seak have have had some real innovations, so

102
00:05:12,279 --> 00:05:15,400
Speaker 2: they're underwriting the innovations, some of this reinforcement learning and

103
00:05:15,440 --> 00:05:17,880
Speaker 2: some of the other things we haven't discussed that will

104
00:05:18,040 --> 00:05:21,720
Speaker 2: likely get copied almost immediately. Microsoft sophomore or less said

105
00:05:21,760 --> 00:05:25,120
Speaker 2: that verbatim, and they also said the big beneficiaries of

106
00:05:25,160 --> 00:05:29,400
Speaker 2: these normal sort of big step changes and efficiency are

107
00:05:29,480 --> 00:05:32,560
Speaker 2: us Ryan, so the customers as prices come down now

108
00:05:32,800 --> 00:05:36,880
Speaker 2: Meta and Zuck he sounded a bit flat footed, so

109
00:05:36,960 --> 00:05:40,239
Speaker 2: he said, deep seek to a number of novel things.

110
00:05:40,960 --> 00:05:45,360
Speaker 2: We are still digesting, but Meta will copy them. It's

111
00:05:45,520 --> 00:05:47,839
Speaker 2: too early to have a strong opinion on what this

112
00:05:47,920 --> 00:05:51,440
Speaker 2: means for capital spending levels. So some pretty interesting comments

113
00:05:51,440 --> 00:05:52,480
Speaker 2: from big tech there.

114
00:05:52,480 --> 00:05:55,200
Speaker 1: Especially when you consider how much they're planning. I think,

115
00:05:55,320 --> 00:05:57,320
Speaker 1: was it eighty seven billion dollars or something they were

116
00:05:57,360 --> 00:05:58,599
Speaker 1: planning on spending on AI?

117
00:05:59,560 --> 00:06:03,520
Speaker 2: So well, Microsoft, Yeah, Microsoft's eighty billion dollars of capex total,

118
00:06:03,560 --> 00:06:05,760
Speaker 2: but yes, a lot of that will be data sinkers

119
00:06:06,160 --> 00:06:10,360
Speaker 2: and AI chips, and met Is underwrote a sort of

120
00:06:10,440 --> 00:06:13,799
Speaker 2: sixty billion dollar check a few days ago, even.

121
00:06:13,640 --> 00:06:16,680
Speaker 1: Before its results, because it'll surely have a huge effect

122
00:06:16,760 --> 00:06:18,040
Speaker 1: on where you put that right.

123
00:06:18,960 --> 00:06:21,320
Speaker 2: Well, that that's why I do think it's really interesting

124
00:06:21,320 --> 00:06:24,840
Speaker 2: that Zuckerberg's saying we are still digesting and it's too

125
00:06:24,880 --> 00:06:26,440
Speaker 2: early to have a strong opinion on what this means

126
00:06:26,440 --> 00:06:29,320
Speaker 2: for capital spinning levels, despite the fact that two days

127
00:06:29,400 --> 00:06:31,040
Speaker 2: or two or three days ago, in the last week,

128
00:06:31,200 --> 00:06:33,640
Speaker 2: he underwrote a sixty billion dollar CAPEX.

129
00:06:33,279 --> 00:06:37,560
Speaker 1: Program exactly right, So what are the implications for investors? See,

130
00:06:37,560 --> 00:06:38,920
Speaker 1: how should investors think about this?

131
00:06:39,600 --> 00:06:42,080
Speaker 2: Well, I think we do need to answer sort of

132
00:06:42,080 --> 00:06:44,159
Speaker 2: four or five critical questions which we don't have the

133
00:06:44,200 --> 00:06:46,440
Speaker 2: answer to yet. So firstly, that the model has only

134
00:06:46,440 --> 00:06:47,720
Speaker 2: been out for it less than a week. This is

135
00:06:47,720 --> 00:06:49,919
Speaker 2: a very simple one. So let's see other reviews go

136
00:06:50,040 --> 00:06:51,760
Speaker 2: from all these brand new customers. You know, it went

137
00:06:51,800 --> 00:06:54,040
Speaker 2: from number thirty one on the air collapse sort at one.

138
00:06:54,440 --> 00:06:57,440
Speaker 2: I've tested it, Ryan, and it's always jammed and tells

139
00:06:57,480 --> 00:07:00,159
Speaker 2: me to try again later. So let's see whether the

140
00:07:00,200 --> 00:07:03,680
Speaker 2: consumers are loving the model. The second one is does

141
00:07:03,720 --> 00:07:05,840
Speaker 2: this mean that companies like in Video and AMD you

142
00:07:05,880 --> 00:07:07,880
Speaker 2: guys have spoken about this earlier in the week, have

143
00:07:08,000 --> 00:07:10,800
Speaker 2: to cut the prices of their cutting edge accelerated compute

144
00:07:10,880 --> 00:07:13,840
Speaker 2: chips and related to that, the third question is if so,

145
00:07:14,000 --> 00:07:16,520
Speaker 2: will there be a big enough spike in demand like

146
00:07:16,560 --> 00:07:20,360
Speaker 2: we've seen over previous technological cycles to offset the price

147
00:07:20,400 --> 00:07:22,320
Speaker 2: custs Like we saw that with flat screen TVs, we

148
00:07:22,360 --> 00:07:26,600
Speaker 2: saw producers get way more efficient at producing them. Prices

149
00:07:26,600 --> 00:07:30,520
Speaker 2: of flat screen TVs collapsed and demand skyrocketed. And the

150
00:07:30,560 --> 00:07:32,640
Speaker 2: final thing is what does all this mean for power

151
00:07:32,680 --> 00:07:36,240
Speaker 2: demand which was exploding for AI data centers. But as

152
00:07:36,280 --> 00:07:38,720
Speaker 2: far as investors go, the market is shot first and

153
00:07:38,760 --> 00:07:42,160
Speaker 2: we'll answer those questions we discussed later, but it's going

154
00:07:42,200 --> 00:07:44,160
Speaker 2: to take time to get answers to those questions, and

155
00:07:44,360 --> 00:07:48,040
Speaker 2: if Zuckerberg is still digesting, as he said a few

156
00:07:48,080 --> 00:07:49,880
Speaker 2: hours ago, it's hard for the average investor to have

157
00:07:49,880 --> 00:07:52,120
Speaker 2: clarity yet. But one final point I'll leave you with

158
00:07:52,200 --> 00:07:54,880
Speaker 2: Ryan is I think we can expect more volatility, not

159
00:07:54,920 --> 00:07:57,280
Speaker 2: just in stop prices, but in how the China versus

160
00:07:57,440 --> 00:08:01,920
Speaker 2: us chipwoar. We discussed last year ratchet's up, so we

161
00:08:02,000 --> 00:08:03,320
Speaker 2: really do need to buckle up, I.

162
00:08:03,200 --> 00:08:06,000
Speaker 1: Think, Sam. Thanks for that, Sam, Dicky Fisher Funds. Great

163
00:08:06,040 --> 00:08:06,400
Speaker 1: to have you on.

164
00:08:06,400 --> 00:08:10,160
Speaker 2: As always, for more from Heather Duplessy Allen Drive, listen

165
00:08:10,240 --> 00:08:13,280
Speaker 2: live to news talks it'd be from four pm weekdays,

166
00:08:13,360 --> 00:08:15,520
Speaker 2: or follow the podcast on iHeartRadio