1
00:00:00,120 --> 00:00:03,960
Speaker 1: This is Bloomberg Business Week with Carol Messer and Tim

2
00:00:04,000 --> 00:00:06,280
Speaker 1: Stenebek on Bloomberg Radio.

3
00:00:06,840 --> 00:00:10,080
Speaker 2: All Right, when someone says AI, they can be referring

4
00:00:10,119 --> 00:00:11,719
Speaker 2: to a lot of different things, you know of that, right,

5
00:00:11,720 --> 00:00:14,320
Speaker 2: They can be talking about chatbots such as chat Gibt

6
00:00:15,040 --> 00:00:18,599
Speaker 2: or Google Bard, which are examples of LM's or large

7
00:00:18,720 --> 00:00:21,120
Speaker 2: language models. We talked. We talked about this a little

8
00:00:21,120 --> 00:00:21,960
Speaker 2: bit with Mendeep earlier.

9
00:00:22,079 --> 00:00:24,640
Speaker 1: Exactly. Okay, what about when it comes to AGI, artificial

10
00:00:24,640 --> 00:00:27,280
Speaker 1: general intelligence. It's the holy grail of AI. It's not

11
00:00:27,400 --> 00:00:29,680
Speaker 1: around yet, but everyone is working on it. What that

12
00:00:29,760 --> 00:00:32,160
Speaker 1: means is that AI can perform as well or better

13
00:00:32,200 --> 00:00:34,839
Speaker 1: than humans on most tasks.

14
00:00:35,000 --> 00:00:37,280
Speaker 2: All right, so machine learning is something else. It's the

15
00:00:37,320 --> 00:00:39,680
Speaker 2: focus of our next guest, new book. Eric Siegel is

16
00:00:39,720 --> 00:00:42,519
Speaker 2: a consultant and former Columbia University professor. He's also the

17
00:00:42,560 --> 00:00:45,879
Speaker 2: author of a new book. It is entitled The AI Playbook,

18
00:00:45,920 --> 00:00:48,960
Speaker 2: Mastering the Rare Art of Machine Learning Deployment. It is

19
00:00:49,120 --> 00:00:53,159
Speaker 2: out today. Eric joins us on Zoom from the Bay Area. Eric, congratulations,

20
00:00:53,600 --> 00:00:56,200
Speaker 2: Let's start with the basics, because I think sometimes we

21
00:00:56,280 --> 00:01:00,720
Speaker 2: all throw around phrases assuming everybody gets it and they

22
00:01:00,720 --> 00:01:03,840
Speaker 2: don't necessarily. How do you define machine learning?

23
00:01:05,400 --> 00:01:09,319
Speaker 3: Well, thanks Carol. Machine learning is technology that learns from

24
00:01:09,520 --> 00:01:13,480
Speaker 3: experience in order to make predictions in order to target

25
00:01:13,800 --> 00:01:17,000
Speaker 3: and improve large scale operations. So who to target to

26
00:01:17,360 --> 00:01:21,640
Speaker 3: for marketing based on predicting who's going to buy which transaction?

27
00:01:21,720 --> 00:01:24,360
Speaker 3: To audit for fraud based on predicting which is going

28
00:01:24,440 --> 00:01:27,240
Speaker 3: to be turned out to be fraudulent. Who to approve

29
00:01:27,280 --> 00:01:29,840
Speaker 3: for credit application based on who's going to be the

30
00:01:29,880 --> 00:01:35,399
Speaker 3: most reliable debtor, etc. Etc. Which satellite to investigate for

31
00:01:35,440 --> 00:01:38,000
Speaker 3: potentially running out of a battery based on whether it's

32
00:01:38,000 --> 00:01:41,440
Speaker 3: going to where to drill for oil? Pretty much this

33
00:01:41,520 --> 00:01:43,280
Speaker 3: is the type of AI, and you could call it

34
00:01:43,360 --> 00:01:47,800
Speaker 3: predictive AI or predictive analytics to differentiate it from generative AI.

35
00:01:47,880 --> 00:01:50,680
Speaker 3: This is the type of AI you turn to when

36
00:01:50,720 --> 00:01:54,040
Speaker 3: you want to improve pretty much any and all of

37
00:01:54,080 --> 00:01:58,960
Speaker 3: your large scale operations, which are which consists of many decisions,

38
00:01:59,040 --> 00:02:02,720
Speaker 3: and prediction is the holy grail for improving decisions.

39
00:02:02,760 --> 00:02:07,040
Speaker 1: Would you consider the way that content is surfaced on

40
00:02:07,880 --> 00:02:12,239
Speaker 1: a social media platform like x slash, Twitter or Instagram?

41
00:02:12,440 --> 00:02:14,120
Speaker 1: Would you consider that machine learning?

42
00:02:16,000 --> 00:02:16,960
Speaker 3: Oh? I see? Yeah?

43
00:02:17,120 --> 00:02:19,480
Speaker 1: As far as the ordering of your feed, yeah, like

44
00:02:19,520 --> 00:02:20,880
Speaker 1: if I log in, you know, if I start to

45
00:02:20,880 --> 00:02:22,800
Speaker 1: follow someone new, Like if I follow someone new on

46
00:02:22,840 --> 00:02:25,800
Speaker 1: Instagram and then I log in Instagram again, that like

47
00:02:25,800 --> 00:02:27,640
Speaker 1: an old post from that person is going to be

48
00:02:27,680 --> 00:02:29,839
Speaker 1: like the first thing that comes up because Instagram thinks

49
00:02:29,880 --> 00:02:30,960
Speaker 1: I'm now interested in that.

50
00:02:31,840 --> 00:02:33,720
Speaker 3: Right, that's a prediction task. And the same thing with

51
00:02:33,760 --> 00:02:36,600
Speaker 3: the ordering of your Facebook feed the default feed, assuming

52
00:02:36,639 --> 00:02:38,639
Speaker 3: you leave it at that, and the same thing in

53
00:02:38,720 --> 00:02:41,520
Speaker 3: the ordering of your Google search results. It's all based

54
00:02:41,560 --> 00:02:44,760
Speaker 3: on predictive models. That's what machine learning generates from data.

55
00:02:44,840 --> 00:02:48,160
Speaker 3: Is a model that captures the patterns, that's the discoveries

56
00:02:48,160 --> 00:02:50,560
Speaker 3: it's made from data. That helps it predict. Predict is

57
00:02:50,560 --> 00:02:52,760
Speaker 3: the action. So you're predicting in order to say which

58
00:02:52,760 --> 00:02:54,679
Speaker 3: of these content ten items is going to be of

59
00:02:54,800 --> 00:02:57,880
Speaker 3: most interest or most relevant, whether it's for Internet search

60
00:02:57,960 --> 00:03:01,040
Speaker 3: like with Google, or the ordering of your news feed,

61
00:03:01,120 --> 00:03:04,200
Speaker 3: or the ordering of your search results for properties on Airbnb.

62
00:03:04,720 --> 00:03:08,880
Speaker 2: All right, whiteboard it for me. So AI big umbrella

63
00:03:09,040 --> 00:03:12,520
Speaker 2: machine learning a form of AI. How do we stack

64
00:03:12,600 --> 00:03:15,880
Speaker 2: up and make sense in this environment of because we

65
00:03:15,919 --> 00:03:18,280
Speaker 2: do throw around AI, which has been around for a

66
00:03:18,320 --> 00:03:21,560
Speaker 2: long time, but now we're talking about you know, generative AI.

67
00:03:22,440 --> 00:03:24,520
Speaker 2: Give me the whiteboard on it. You're teaching a class, like,

68
00:03:24,800 --> 00:03:26,400
Speaker 2: how do you lay it all out in terms of

69
00:03:26,560 --> 00:03:29,280
Speaker 2: AI what it means or machine learning how it'll play

70
00:03:29,280 --> 00:03:29,639
Speaker 2: a role.

71
00:03:31,080 --> 00:03:34,800
Speaker 3: Well, the typical hierarchy is that machine learning is part

72
00:03:34,800 --> 00:03:37,560
Speaker 3: of AI, and that's a more bigger umbrella term. But

73
00:03:37,800 --> 00:03:40,440
Speaker 3: my opinion varies from a lot of the mainstream, although

74
00:03:40,440 --> 00:03:43,240
Speaker 3: more people are jumping on board that AI is really

75
00:03:43,280 --> 00:03:46,440
Speaker 3: an amorphous, ill defined term. We're trying to ascribe the

76
00:03:46,520 --> 00:03:50,120
Speaker 3: word intelligence to a machine. That's quite problematic to nail down.

77
00:03:50,280 --> 00:03:53,320
Speaker 3: But if you don't define something well, you can't pursue it.

78
00:03:53,360 --> 00:03:56,920
Speaker 3: For engineering, AI is the story we hear about machine

79
00:03:56,960 --> 00:03:59,240
Speaker 3: learning is the technology that we have, and machine learning

80
00:03:59,320 --> 00:04:01,480
Speaker 3: in all those ways. I just describe where you're predicting

81
00:04:01,640 --> 00:04:05,200
Speaker 3: for each individual customer healthcare client, as far as their

82
00:04:05,200 --> 00:04:08,080
Speaker 3: disease progression, or where to drill for oil which I'll

83
00:04:08,160 --> 00:04:11,080
Speaker 3: like to investigate in which transaction to audit on that

84
00:04:11,200 --> 00:04:15,080
Speaker 3: individual level. It's the same core technology that drives generative

85
00:04:15,080 --> 00:04:19,440
Speaker 3: AI for its ability to generate first drafts of writing

86
00:04:19,800 --> 00:04:22,960
Speaker 3: or code, or of images, and in those cases what

87
00:04:22,960 --> 00:04:25,440
Speaker 3: it's doing is predicting what should the next word be Okay,

88
00:04:25,440 --> 00:04:27,320
Speaker 3: well it's actually the next token, but it's on that

89
00:04:27,400 --> 00:04:30,040
Speaker 3: level of detail. What should the next word be? How

90
00:04:30,040 --> 00:04:33,360
Speaker 3: should I change this individual pixel in an iteration? As

91
00:04:33,400 --> 00:04:36,799
Speaker 3: I'm rendering this image, I being the computer in this case.

92
00:04:37,680 --> 00:04:40,880
Speaker 3: It's the same core technology learning from data to predict.

93
00:04:41,320 --> 00:04:44,560
Speaker 1: You know, it's funny, Carol, when you're so just a

94
00:04:44,600 --> 00:04:48,039
Speaker 1: little behind the scenes. Eric, It's like, we use this

95
00:04:48,080 --> 00:04:50,520
Speaker 1: Google doc to prepare the show, and we both work

96
00:04:50,520 --> 00:04:53,039
Speaker 1: on it and work on different things. And Google now

97
00:04:53,160 --> 00:04:56,560
Speaker 1: has predictive texts, And when I was writing this one

98
00:04:57,839 --> 00:05:01,200
Speaker 1: for Eric's intro, it actually got some stuff wrong, which

99
00:05:01,200 --> 00:05:02,960
Speaker 1: I thought was really funny because here we are talking

100
00:05:02,960 --> 00:05:05,560
Speaker 1: about the technology that it's getting wrong. When I'm like,

101
00:05:05,640 --> 00:05:07,839
Speaker 1: you know, how do you writing a question in the doc?

102
00:05:07,920 --> 00:05:08,839
Speaker 1: I mean, have you noticed that?

103
00:05:09,000 --> 00:05:11,120
Speaker 2: Yeah? No, you're absolutely right if you're very careful, because

104
00:05:11,120 --> 00:05:13,640
Speaker 2: it jumps ahead right if you don't catch it. So, yeah,

105
00:05:13,640 --> 00:05:15,560
Speaker 2: how do we make sure all this stuff that were

106
00:05:16,040 --> 00:05:19,480
Speaker 2: it feels like increasingly moving towards relying on make sure

107
00:05:19,520 --> 00:05:23,520
Speaker 2: whether it's machine learning, that it's accurate and predicting right

108
00:05:23,560 --> 00:05:27,240
Speaker 2: things or the right outcomes or the smart outcomes. How

109
00:05:27,240 --> 00:05:28,640
Speaker 2: do we make sure I know. Is it just the

110
00:05:28,720 --> 00:05:30,719
Speaker 2: data sets that go into it? Is it that simple

111
00:05:31,520 --> 00:05:32,560
Speaker 2: and that complicated? No?

112
00:05:32,640 --> 00:05:37,560
Speaker 3: I mean no, no, because we're not headed with definitiveness

113
00:05:37,720 --> 00:05:42,520
Speaker 3: towards reliability when it comes to that type of generated text. Look,

114
00:05:42,800 --> 00:05:46,200
Speaker 3: these models are so seemingly human like. They're amazing. I

115
00:05:46,279 --> 00:05:48,880
Speaker 3: spent six years of my career in the Natural Language

116
00:05:48,880 --> 00:05:51,719
Speaker 3: Processing Research Group at Columbia, and I never thought I

117
00:05:51,720 --> 00:05:54,680
Speaker 3: would see what we're seeing today. It's so amazing. The

118
00:05:54,720 --> 00:05:58,320
Speaker 3: way it creates often cohesive content, can talk about anything,

119
00:05:58,680 --> 00:06:01,800
Speaker 3: use expressions the humans use, and because it's trained over

120
00:06:01,880 --> 00:06:06,200
Speaker 3: so much data and the actual modeling itself is so advanced. However,

121
00:06:06,800 --> 00:06:08,960
Speaker 3: what it's trained to do is essentially on that per

122
00:06:09,040 --> 00:06:11,480
Speaker 3: word level of detail, which really gives it a human

123
00:06:11,640 --> 00:06:14,680
Speaker 3: like aura. But that doesn't mean that it was developed

124
00:06:15,240 --> 00:06:19,440
Speaker 3: to pursue higher order human goals like being correct. That's

125
00:06:19,480 --> 00:06:22,800
Speaker 3: a whole nother thing. The fact that it's seemingly human

126
00:06:22,920 --> 00:06:26,760
Speaker 3: like doesn't mean it's a step towards general human behavior.

127
00:06:27,240 --> 00:06:31,280
Speaker 3: So you know you earlier Tim mentioned artificial general intelligence.

128
00:06:31,360 --> 00:06:33,039
Speaker 3: I'm actually sort of a disbeliever in that.

129
00:06:33,080 --> 00:06:35,520
Speaker 1: I don't think it's Why are you a disbeliever.

130
00:06:35,720 --> 00:06:39,279
Speaker 3: Yeah, I don't think it's technically impossible that someday, But

131
00:06:39,400 --> 00:06:41,840
Speaker 3: I do not believe that any of the advancements, as

132
00:06:41,880 --> 00:06:46,680
Speaker 3: impressive and valuable as they are, actually represent a concrete

133
00:06:46,720 --> 00:06:51,560
Speaker 3: step towards general human level capabilities. Where the machine is basically,

134
00:06:51,720 --> 00:06:54,240
Speaker 3: let's call it what it is in the story, it's

135
00:06:54,240 --> 00:06:57,880
Speaker 3: an artificial human. You can onboard them like an employee

136
00:06:57,920 --> 00:07:00,240
Speaker 3: at like a human employee, and let it rip. They

137
00:07:00,240 --> 00:07:03,360
Speaker 3: can run a fortune five hundred company, whatever it is.

138
00:07:03,680 --> 00:07:07,080
Speaker 3: That is a science fiction fantasy. I do not believe

139
00:07:07,400 --> 00:07:09,360
Speaker 3: that we're taking concrete steps in that direction.

140
00:07:09,520 --> 00:07:11,280
Speaker 1: So you're not concerned about the rise.

141
00:07:11,120 --> 00:07:16,000
Speaker 3: Of the bots, right, And that's they call it critic hype,

142
00:07:16,280 --> 00:07:18,000
Speaker 3: right where you say, hey, look this stuff is so

143
00:07:18,160 --> 00:07:21,160
Speaker 3: good it could kill all of us. It's really just

144
00:07:21,200 --> 00:07:25,480
Speaker 3: another way to sort of mismanage expectations. And there's a

145
00:07:25,560 --> 00:07:29,520
Speaker 3: variety of reasons why people do that. Some genuinely believe it.

146
00:07:29,840 --> 00:07:33,880
Speaker 3: I'm trying to calm the world down a little bit here.

147
00:07:34,080 --> 00:07:38,320
Speaker 3: The stuff is extremely valuable and in what it can

148
00:07:38,360 --> 00:07:42,360
Speaker 3: do today, and the story that it's becoming human like

149
00:07:42,800 --> 00:07:45,280
Speaker 3: is it over sells? In other words, it's hype and

150
00:07:45,320 --> 00:07:50,360
Speaker 3: that gap between what's real and what's plausible from the

151
00:07:50,440 --> 00:07:54,720
Speaker 3: stories is bad. And it's with mismanaged expectations that's when

152
00:07:54,760 --> 00:07:59,000
Speaker 3: you have the downfall, the disappointment. The disillusion meant also

153
00:07:59,400 --> 00:08:02,160
Speaker 3: in the more extreme case called an AI winter, and

154
00:08:02,480 --> 00:08:04,480
Speaker 3: the problem there is you throw the baby out with

155
00:08:04,520 --> 00:08:07,960
Speaker 3: the bathwater. You throw the value of generative AI for

156
00:08:08,040 --> 00:08:11,600
Speaker 3: first drafts and predictive AI, which by the way, is

157
00:08:11,640 --> 00:08:15,640
Speaker 3: still a much bigger industry right now, out with the bathwater.

158
00:08:16,360 --> 00:08:16,560
Speaker 1: You know.

159
00:08:16,640 --> 00:08:18,640
Speaker 2: Our producer Paul Brennan said, Yep, you're gonna want to

160
00:08:18,640 --> 00:08:20,120
Speaker 2: talk to this guy for a long time. We have

161
00:08:20,200 --> 00:08:22,600
Speaker 2: unfortunately run out of time, so promise you will come

162
00:08:22,600 --> 00:08:24,680
Speaker 2: back soon because I feel like this is a conversation

163
00:08:24,800 --> 00:08:28,680
Speaker 2: we need to continue. Just making so much sense, Eric,

164
00:08:28,720 --> 00:08:31,400
Speaker 2: thank you so much. Eric Siegel, he's a consultant, former

165
00:08:31,400 --> 00:08:33,760
Speaker 2: Columbia University professor. He's got a new book at the

166
00:08:33,760 --> 00:08:38,360
Speaker 2: AI Playbook, Mastering the rare art of Machine Learning Deployment.