1
00:00:06,059 --> 00:00:09,199
Speaker 1: Welcome to COVID Time podcast series on markets and economies

2
00:00:09,199 --> 00:00:12,640
Speaker 1: from DBS Group Research. I'm Tamur Bek, chief economist. Welcome

3
00:00:12,640 --> 00:00:15,729
Speaker 1: you to our 174th episode.

4
00:00:16,440 --> 00:00:22,790
Speaker 1: Today is an issue that has been vexing policymakers, riveting employers,

5
00:00:23,079 --> 00:00:26,200
Speaker 1: and stressing out employees, which is AI and the future

6
00:00:26,200 --> 00:00:29,620
Speaker 1: of jobs. My guest is Leslie Teo, senior director of

7
00:00:29,620 --> 00:00:32,799
Speaker 1: AI products at AI Singapore. Leslie and his team lead

8
00:00:32,799 --> 00:00:36,519
Speaker 1: the initiative Sea Lion, which stands for Southeast Asian Languages

9
00:00:36,520 --> 00:00:39,520
Speaker 1: in One Network. Sea Lion is the region's open source

10
00:00:39,520 --> 00:00:42,519
Speaker 1: large language model family spanning 11 languages.

11
00:00:42,889 --> 00:00:46,750
Speaker 1: Leslie is one of the 40 members at the UN's

12
00:00:46,750 --> 00:00:50,439
Speaker 1: Independent International Scientific Panel on AI. Prior to AI Singapore,

13
00:00:50,520 --> 00:00:55,069
Speaker 1: Leslie worked at CRAB, GIC, and IMF in various leadership roles.

14
00:00:55,529 --> 00:00:58,250
Speaker 1: I have had the pleasure of knowing Leslie for a

15
00:00:58,250 --> 00:01:02,090
Speaker 1: couple of decades with overlapping professional journeys through Washington DC

16
00:01:02,090 --> 00:01:05,930
Speaker 1: and Singapore. Recently we collaborated on a couple of articles

17
00:01:05,930 --> 00:01:09,250
Speaker 1: for Business Times on redefining jobs and expertise in the

18
00:01:09,250 --> 00:01:11,250
Speaker 1: AI era. We'll talk about that.

19
00:01:11,339 --> 00:01:15,059
Speaker 1: Effort and more in today's podcast. Lesito, a warm welcome

20
00:01:15,059 --> 00:01:15,220
Speaker 1: to

21
00:01:15,220 --> 00:01:18,300
Speaker 2: COVID time. Thank you, Tai. It's taken 5 years, but

22
00:01:18,300 --> 00:01:21,099
Speaker 1: yes, I, I cannot, I cannot believe that you have

23
00:01:21,099 --> 00:01:23,860
Speaker 1: not been on COVID time before and I'm on episode 174,

24
00:01:23,970 --> 00:01:26,209
Speaker 1: but this is not gonna be the last time either. Leslie,

25
00:01:26,260 --> 00:01:27,260
Speaker 1: I assure you of that.

26
00:01:27,769 --> 00:01:31,760
Speaker 1: Uh, Leslie, um, let's talk about our collaboration. It began

27
00:01:31,760 --> 00:01:33,360
Speaker 1: a few months ago over a lunch.

28
00:01:33,529 --> 00:01:37,369
Speaker 2: Yes, and, uh, I think we said, can we actually

29
00:01:37,370 --> 00:01:41,629
Speaker 2: use AI and let's experiment to write a publishable editorial

30
00:01:42,269 --> 00:01:45,369
Speaker 2: in in in in uh in, you know, uh, in

31
00:01:45,370 --> 00:01:50,330
Speaker 2: old school uh uh uh publication and uh it took

32
00:01:50,330 --> 00:01:52,930
Speaker 2: a while. Uh, we had lots of tos and fro's

33
00:01:52,930 --> 00:01:54,709
Speaker 2: among ourselves, but we also had

34
00:01:55,110 --> 00:01:58,150
Speaker 2: Uh, lots of discussions with the editors to get it

35
00:01:58,150 --> 00:02:00,910
Speaker 2: finally published, and that was published last week. So

36
00:02:00,910 --> 00:02:02,550
Speaker 1: let's, let's talk about the journey. By the way, I

37
00:02:02,550 --> 00:02:04,470
Speaker 1: will make sure to put a link to that article

38
00:02:04,470 --> 00:02:06,419
Speaker 1: in the show notes so that, uh, those who are

39
00:02:06,419 --> 00:02:09,149
Speaker 1: listening and watching can, uh, read the two articles. So,

40
00:02:09,309 --> 00:02:11,029
Speaker 1: and I keep on saying two articles because there is

41
00:02:11,029 --> 00:02:14,070
Speaker 1: a core article, and then the editors themselves told us

42
00:02:14,070 --> 00:02:16,850
Speaker 1: to have a companion article about how we went about

43
00:02:17,070 --> 00:02:17,550
Speaker 1: writing it.

44
00:02:17,940 --> 00:02:20,699
Speaker 1: Let's see, let's talk about that process. Yes.

45
00:02:21,000 --> 00:02:24,630
Speaker 2: Um, so maybe for me, uh, I think at the lunch, uh,

46
00:02:24,639 --> 00:02:27,679
Speaker 2: just to give the audience, um, a sense, um,

47
00:02:28,529 --> 00:02:31,809
Speaker 2: Yes, we cared about AI, we cared about jobs, but

48
00:02:31,809 --> 00:02:34,570
Speaker 2: we wanted to experiment to see what AI could really

49
00:02:34,570 --> 00:02:39,630
Speaker 2: do for a real job, not just the hype. And

50
00:02:39,770 --> 00:02:42,529
Speaker 2: I don't know about you, um, but my bottom line

51
00:02:42,529 --> 00:02:48,210
Speaker 2: is that it really did simplify and actually magnified and

52
00:02:48,210 --> 00:02:50,910
Speaker 2: made more effective certain parts of our work processes.

53
00:02:51,940 --> 00:02:52,779
Speaker 2: But it still doesn't

54
00:02:52,779 --> 00:02:53,059
Speaker 2: cut it.

55
00:02:53,179 --> 00:02:55,899
Speaker 1: Right, so before even going to the not cutting it part,

56
00:02:55,970 --> 00:02:59,220
Speaker 1: let's talk a little bit about the heavy lifting that

57
00:02:59,220 --> 00:03:01,019
Speaker 1: we made the AI do. Obviously,

58
00:03:01,139 --> 00:03:04,440
Speaker 2: one of the heavy lifting was to figure out what

59
00:03:04,699 --> 00:03:08,660
Speaker 2: is actually being said, uh, to summarize and search for articles.

60
00:03:08,740 --> 00:03:09,850
Speaker 2: AI is really good at that

61
00:03:09,850 --> 00:03:11,258
Speaker 1: actually. I think 100 articles.

62
00:03:11,779 --> 00:03:15,020
Speaker 2: More than 200 articles, right? And also, um, because we

63
00:03:15,020 --> 00:03:18,039
Speaker 2: wanted to make sure the AI understood.

64
00:03:18,869 --> 00:03:24,429
Speaker 2: Basic economic facts and labor economics. We also made sure

65
00:03:24,429 --> 00:03:27,589
Speaker 2: it read all the seminal papers, you know, Bryn Thompson

66
00:03:27,830 --> 00:03:32,690
Speaker 2: or David Arthur's papers, all the latest research and canonical

67
00:03:32,690 --> 00:03:37,309
Speaker 2: research in economics on labor and technological progress. And I

68
00:03:37,309 --> 00:03:40,550
Speaker 2: would say AI did really well. It summarized things, it

69
00:03:40,550 --> 00:03:45,050
Speaker 2: could communicate really well about the economic theory. In fact, um,

70
00:03:45,110 --> 00:03:46,850
Speaker 2: one of the models even built

71
00:03:47,270 --> 00:03:50,789
Speaker 2: An economic model for us, uh, uh, for how the

72
00:03:50,789 --> 00:03:53,070
Speaker 2: labor market would, could evolve in

73
00:03:53,070 --> 00:03:55,750
Speaker 1: Singapore. But you know this stuff already. You are what

74
00:03:55,750 --> 00:03:59,550
Speaker 1: we call is a subject matter expert. So, did you

75
00:03:59,550 --> 00:04:01,949
Speaker 1: actually learn anything from that process or it was just

76
00:04:01,949 --> 00:04:03,050
Speaker 1: telling you things you already know?

77
00:04:03,229 --> 00:04:06,600
Speaker 2: Uh, no, it confirms one of the key things about,

78
00:04:06,990 --> 00:04:09,789
Speaker 2: you know, how we can use AI, right? Um,

79
00:04:10,559 --> 00:04:14,520
Speaker 2: AI has many capabilities and we might be

80
00:04:15,300 --> 00:04:19,179
Speaker 2: Assuming that it has these capabilities to the same extent

81
00:04:19,178 --> 00:04:21,649
Speaker 2: in terms of how well it does, uh, for example,

82
00:04:21,700 --> 00:04:26,779
Speaker 2: in coding, uh, for example, in, um, drafting reports. But

83
00:04:26,779 --> 00:04:30,420
Speaker 2: the truth is, AI is still at the stage where

84
00:04:30,420 --> 00:04:33,320
Speaker 2: capabilities vary depending on

85
00:04:34,130 --> 00:04:36,450
Speaker 2: The kinds of things you expect the AI to do.

86
00:04:37,238 --> 00:04:40,320
Speaker 2: Um, actually Anthropic had a, had a report this week

87
00:04:40,320 --> 00:04:45,040
Speaker 2: where they talked about, you know, they mapped these capabilities

88
00:04:45,040 --> 00:04:49,799
Speaker 2: against human activity and you know, I'd say roughly 50%

89
00:04:50,279 --> 00:04:54,558
Speaker 2: um of the capabilities are at 90% or 95%, but

90
00:04:54,559 --> 00:04:59,519
Speaker 2: many are still way below, you know, 50%. So,

91
00:05:00,010 --> 00:05:04,410
Speaker 2: AI was good, basically a good research intern that gave

92
00:05:04,410 --> 00:05:08,510
Speaker 2: us the facts that could remember what theory says today,

93
00:05:08,730 --> 00:05:11,230
Speaker 2: that could even build a quantitative model.

94
00:05:11,959 --> 00:05:15,540
Speaker 2: Um, but that, it kind of stopped there, right? Um.

95
00:05:16,420 --> 00:05:18,820
Speaker 2: In fact, one of the more troubling things, and again

96
00:05:18,820 --> 00:05:22,089
Speaker 2: we know this, is that once we have the facts,

97
00:05:22,260 --> 00:05:25,899
Speaker 2: we could make AI make any almost any argument we

98
00:05:25,899 --> 00:05:30,470
Speaker 2: wanted it to make, right? Um, uh, if we said that,

99
00:05:30,820 --> 00:05:36,070
Speaker 2: let's write an article that would be so pessimistic that, um,

100
00:05:36,299 --> 00:05:38,700
Speaker 2: you know, in 12 months' time, we will be dealing

101
00:05:38,700 --> 00:05:42,779
Speaker 2: with massive unemployment, it probably could do that. Uh, we

102
00:05:42,779 --> 00:05:46,000
Speaker 2: could also get the AI to write an article that

103
00:05:46,660 --> 00:05:49,420
Speaker 2: said, you know, in 12 months' time we will be

104
00:05:49,420 --> 00:05:54,839
Speaker 2: living in utopia. So, I, I think one of the shortfalls,

105
00:05:54,850 --> 00:05:57,859
Speaker 2: I would say and uh and actually this was the

106
00:05:57,859 --> 00:06:00,619
Speaker 2: point of the article, um, is that

107
00:06:01,450 --> 00:06:05,049
Speaker 2: Yes, AI could do these things, but it lacked an

108
00:06:05,049 --> 00:06:08,570
Speaker 2: ability to make judgments and it was unanchored.

109
00:06:08,690 --> 00:06:12,859
Speaker 1: Right, and those are two very important issues, the inability

110
00:06:12,859 --> 00:06:17,570
Speaker 1: to judge with human consideration, human context in mind, and

111
00:06:17,570 --> 00:06:18,950
Speaker 1: the anchoring aspect.

112
00:06:19,260 --> 00:06:22,750
Speaker 1: That you and I can massage it to our own

113
00:06:22,750 --> 00:06:27,549
Speaker 1: needs very easily. Leslie, I have, uh, several young relatives

114
00:06:27,549 --> 00:06:30,789
Speaker 1: in the US and in India who use AI on

115
00:06:30,790 --> 00:06:33,989
Speaker 1: a daily basis as their personal psychologist, and I asked

116
00:06:33,988 --> 00:06:35,230
Speaker 1: both of them that, you know, why do you do that?

117
00:06:35,269 --> 00:06:38,149
Speaker 1: Is it because it gives us positive reinforcement. So it's

118
00:06:38,149 --> 00:06:40,808
Speaker 1: almost like a drug. It will, it will soothe you

119
00:06:41,029 --> 00:06:43,470
Speaker 1: and it would help you shape the world's reality the

120
00:06:43,470 --> 00:06:45,190
Speaker 1: way you would like it to be done.

121
00:06:45,540 --> 00:06:46,420
Speaker 1: Uh, so a bit of a

122
00:06:46,420 --> 00:06:50,920
Speaker 2: drug. Yeah, and actually, uh, let's, let's also be clear, um,

123
00:06:51,459 --> 00:06:55,089
Speaker 2: sometimes we confuse a product like chat GPT or cloud

124
00:06:55,420 --> 00:06:59,320
Speaker 2: with the model, and today, models and products are starting

125
00:06:59,320 --> 00:07:00,320
Speaker 2: to diverge.

126
00:07:02,928 --> 00:07:04,909
Speaker 2: If we are a responsible AI builder.

127
00:07:06,250 --> 00:07:08,529
Speaker 2: It's a bit hard for you to build general models

128
00:07:08,529 --> 00:07:11,869
Speaker 2: that are, you know, kind of safe and appropriate in

129
00:07:11,880 --> 00:07:14,369
Speaker 2: in general and and fit for all.

130
00:07:15,480 --> 00:07:17,910
Speaker 2: It actually has to happen at the product level.

131
00:07:19,369 --> 00:07:23,890
Speaker 2: And it's not anyone's fault, but right now there's huge

132
00:07:23,890 --> 00:07:29,109
Speaker 2: competition to make your AI models basically attractive and useful.

133
00:07:29,529 --> 00:07:32,630
Speaker 2: So you can imagine this actually it happened with um TikTok,

134
00:07:32,929 --> 00:07:36,470
Speaker 2: it happened with social media, it happened with the web, right?

135
00:07:36,850 --> 00:07:37,329
Speaker 2: I

136
00:07:38,619 --> 00:07:42,010
Speaker 2: No one has the wrong intention, but I want to

137
00:07:42,010 --> 00:07:46,299
Speaker 2: make sure my model and my daily active users are engaged.

138
00:07:46,540 --> 00:07:48,459
Speaker 2: Cause that's how I'm gonna raise money, that's how I'm

139
00:07:48,459 --> 00:07:49,459
Speaker 2: gonna get advertising.

140
00:07:50,359 --> 00:07:53,660
Speaker 2: So I'm gonna build, I'm gonna use this as the algorithm, right?

141
00:07:54,000 --> 00:07:59,339
Speaker 2: The more people use it, the more it will reinforce this. Hence,

142
00:07:59,690 --> 00:08:02,980
Speaker 2: maybe the product is not designed to be a psychologist,

143
00:08:03,519 --> 00:08:08,239
Speaker 2: but it is designed to keep you hooked. And uh

144
00:08:08,239 --> 00:08:11,679
Speaker 2: this is part of I think understanding. I'm a big

145
00:08:11,679 --> 00:08:15,980
Speaker 2: believer that this technology is wonderful and has great potential.

146
00:08:16,329 --> 00:08:20,570
Speaker 2: But we all must understand the limitations and the risks, right?

147
00:08:20,850 --> 00:08:24,649
Speaker 2: And in this case, it's about understanding that this product

148
00:08:24,649 --> 00:08:27,429
Speaker 2: is doing something that actually it's not designed to do

149
00:08:27,809 --> 00:08:31,170
Speaker 2: and it may have negative consequences. Indeed.

150
00:08:31,369 --> 00:08:33,489
Speaker 1: So Leslie, since you brought up models, let's talk about

151
00:08:33,489 --> 00:08:36,210
Speaker 1: again the use of multiple models in the in the

152
00:08:36,210 --> 00:08:39,280
Speaker 1: writing process. What was the intuition behind that? Well,

153
00:08:39,570 --> 00:08:43,829
Speaker 2: because actually you want different perspectives on different models are

154
00:08:43,840 --> 00:08:45,109
Speaker 2: are trained differently.

155
00:08:45,599 --> 00:08:50,159
Speaker 2: Uh, so we have models that are less censored, more censored, uh,

156
00:08:50,239 --> 00:08:53,239
Speaker 2: models that are safer, models that are better at coding,

157
00:08:53,400 --> 00:08:57,359
Speaker 2: models that are better at communication. Um, I, I would

158
00:08:57,359 --> 00:09:00,559
Speaker 2: add though, um, you know, when we started writing maybe

159
00:09:00,559 --> 00:09:03,239
Speaker 2: in November, I think, kind of like in November, on

160
00:09:03,239 --> 00:09:04,559
Speaker 2: and off, um,

161
00:09:06,669 --> 00:09:10,569
Speaker 2: Using multiple models was, you know, par for the course. Um,

162
00:09:10,669 --> 00:09:13,969
Speaker 2: over that time in the last 3 months, it became

163
00:09:13,969 --> 00:09:17,189
Speaker 2: completely agentic, right? So the way we did things actually

164
00:09:17,190 --> 00:09:21,469
Speaker 2: changed dramatically in just 3 months. And it was no longer,

165
00:09:21,789 --> 00:09:24,190
Speaker 2: you know, I had, it's almost like I had 5 colleagues,

166
00:09:24,340 --> 00:09:26,229
Speaker 2: they all have different opinions, I worked with them.

167
00:09:27,059 --> 00:09:29,289
Speaker 2: At the end of the process, actually I'm just talking

168
00:09:29,289 --> 00:09:33,348
Speaker 2: to the 5 agents and they're doing the, they're doing

169
00:09:33,349 --> 00:09:37,150
Speaker 2: the hard lifting and I'm just coordinating. This is a

170
00:09:37,150 --> 00:09:39,710
Speaker 2: massive change in just 3 months. Yes,

171
00:09:39,820 --> 00:09:39,950
Speaker 1: it

172
00:09:39,950 --> 00:09:44,330
Speaker 1: is. Uh, and the sort of differences that we see,

173
00:09:44,510 --> 00:09:47,940
Speaker 1: you know, from model to model. Uh, again, I'm going

174
00:09:47,940 --> 00:09:50,939
Speaker 1: back to the point that as a subject matter expert

175
00:09:50,940 --> 00:09:55,069
Speaker 1: on the interaction or the intersection of labor and technology.

176
00:09:55,840 --> 00:09:58,919
Speaker 1: Did it surprise you? So let me sort of add

177
00:09:58,919 --> 00:10:01,280
Speaker 1: to that question, did any of these models give you

178
00:10:01,280 --> 00:10:03,359
Speaker 1: surprising responses versus the others?

179
00:10:04,330 --> 00:10:04,969
Speaker 2: Um,

180
00:10:06,909 --> 00:10:07,650
Speaker 2: I think

181
00:10:09,229 --> 00:10:12,570
Speaker 2: You can begin to see that when

182
00:10:13,390 --> 00:10:17,289
Speaker 2: So what's surprising is, yes they can summarize, we know that.

183
00:10:17,630 --> 00:10:22,909
Speaker 2: They can um put together, you know, lots of stories

184
00:10:22,909 --> 00:10:27,069
Speaker 2: that seem plausible. But I do begin to see uh

185
00:10:27,070 --> 00:10:31,830
Speaker 2: something else which is in putting together the, you know,

186
00:10:32,109 --> 00:10:35,650
Speaker 2: sort of putting together different models and and uh it's

187
00:10:35,650 --> 00:10:40,020
Speaker 2: starting to uh exhibit very low levels of cognition I

188
00:10:40,030 --> 00:10:42,569
Speaker 2: I would say, right? Maybe not original?

189
00:10:43,190 --> 00:10:47,309
Speaker 2: Um, but clearly it knew enough to build a labor

190
00:10:47,309 --> 00:10:51,750
Speaker 2: market model for Singapore, right? And the numbers it, you know,

191
00:10:51,830 --> 00:10:56,500
Speaker 2: the calibration was actually quite realistic in my view. Uh, so,

192
00:10:56,840 --> 00:10:59,030
Speaker 2: you know, again in the vastness of what's on the

193
00:10:59,030 --> 00:11:03,390
Speaker 2: public internet together with some fine tuning, um, or post-training.

194
00:11:04,190 --> 00:11:06,750
Speaker 2: We do see some of these improvements, and I think

195
00:11:06,750 --> 00:11:08,030
Speaker 2: it will only get better and better.

196
00:11:08,190 --> 00:11:10,429
Speaker 1: Yeah, no, uh, Leslie, I'll share with you one anecdote.

197
00:11:10,510 --> 00:11:14,968
Speaker 1: So last year I asked, uh, OpenAI to basically run

198
00:11:14,969 --> 00:11:19,190
Speaker 1: a 10 equation open economy computational generaling could be a model.

199
00:11:19,229 --> 00:11:20,189
Speaker 1: I may have actually mentioned this to you.

200
00:11:20,320 --> 00:11:23,959
Speaker 1: The past, uh, and it, it worked for hours. I

201
00:11:23,960 --> 00:11:27,090
Speaker 1: felt bad for the number of tokens that was being spent, uh,

202
00:11:27,119 --> 00:11:30,440
Speaker 1: and initially it seemed like a jaw dropping output that

203
00:11:30,440 --> 00:11:34,159
Speaker 1: it did have a tremendous amount of understanding of the

204
00:11:34,159 --> 00:11:38,119
Speaker 1: existing literature out there and brought in financial sector and

205
00:11:38,119 --> 00:11:41,119
Speaker 1: labor sector and capital flows everything into the model, things

206
00:11:41,119 --> 00:11:43,500
Speaker 1: that we would like to see in an open economy model.

207
00:11:43,940 --> 00:11:46,299
Speaker 1: Uh, but once I took a hard look at it,

208
00:11:46,500 --> 00:11:48,219
Speaker 1: especially the code that it provided me so that I

209
00:11:48,219 --> 00:11:50,780
Speaker 1: could start calibrating on my own, I realized that there

210
00:11:50,780 --> 00:11:54,179
Speaker 1: was a limit of very high level work, a model

211
00:11:54,179 --> 00:11:55,739
Speaker 1: I can that can do it. I mean, maybe it's

212
00:11:55,739 --> 00:11:58,260
Speaker 1: unfair to ask it to do basically PhD level economics

213
00:11:58,260 --> 00:12:02,000
Speaker 1: in the first go, but I suppose as time goes by, uh,

214
00:12:02,260 --> 00:12:04,419
Speaker 1: it might even be able to do that. But my

215
00:12:04,419 --> 00:12:07,299
Speaker 1: biggest takeaway from that, Leslie, was that some of my

216
00:12:07,299 --> 00:12:09,218
Speaker 1: younger economists, let alone an intern.

217
00:12:09,669 --> 00:12:12,510
Speaker 1: would actually succumb to the mistakes of the model. They

218
00:12:12,510 --> 00:12:12,950
Speaker 1: can't tell.

219
00:12:13,229 --> 00:12:15,819
Speaker 2: Exactly. Yeah, and actually in a way that was one

220
00:12:15,820 --> 00:12:18,510
Speaker 2: of the points of our paper, right? Uh we were

221
00:12:18,510 --> 00:12:22,409
Speaker 2: trying to say that the tool is getting more powerful.

222
00:12:23,750 --> 00:12:27,710
Speaker 2: What is it that is about the human that is crucial?

223
00:12:28,130 --> 00:12:31,330
Speaker 2: And I was kinda inspired by this um this this

224
00:12:31,330 --> 00:12:34,770
Speaker 2: thought um with a book called Reshuffle, right? Um, and

225
00:12:34,770 --> 00:12:37,609
Speaker 2: the author actually spent many years in Singapore. And the

226
00:12:37,609 --> 00:12:39,150
Speaker 2: idea is that when

227
00:12:39,840 --> 00:12:42,460
Speaker 2: You know, many of our processes, we have a value chain.

228
00:12:43,109 --> 00:12:44,919
Speaker 2: Cognition is going down to 0.

229
00:12:45,840 --> 00:12:48,650
Speaker 2: But actually the value chain, it's not always, you know,

230
00:12:48,799 --> 00:12:52,039
Speaker 2: you don't solve a problem or make money by just

231
00:12:52,039 --> 00:12:54,559
Speaker 2: a model, it's the whole business value chain.

232
00:12:55,440 --> 00:12:59,599
Speaker 2: What becomes the constraints? And in our paper, I, I

233
00:12:59,599 --> 00:13:02,598
Speaker 2: like to think we were trying to communicate a positive

234
00:13:02,599 --> 00:13:03,849
Speaker 2: message to say that

235
00:13:05,049 --> 00:13:09,049
Speaker 2: It's things like judgment, it's things like emotional intelligence, it's

236
00:13:09,049 --> 00:13:11,789
Speaker 2: things like leadership that will become

237
00:13:13,049 --> 00:13:18,270
Speaker 2: The constraint and people will pay for the constraint, right?

238
00:13:18,979 --> 00:13:21,489
Speaker 1: Absolutely. Uh, so Leslie, I mean, I wanna talk to

239
00:13:21,489 --> 00:13:24,049
Speaker 1: you about, uh, education and AI later, but just since

240
00:13:24,049 --> 00:13:26,130
Speaker 1: you brought this up, so my son, you know, sometimes

241
00:13:26,130 --> 00:13:27,809
Speaker 1: asks me that, you know, if the AI knows everything,

242
00:13:27,890 --> 00:13:29,450
Speaker 1: what's the point of going to school and learning.

243
00:13:29,909 --> 00:13:32,359
Speaker 1: And my knee jerk response to him has been that

244
00:13:32,359 --> 00:13:34,320
Speaker 1: we need to understand the world if we're going to

245
00:13:34,320 --> 00:13:38,039
Speaker 1: be worthy inhabitants of this world. Uh, yes, I don't

246
00:13:38,039 --> 00:13:41,159
Speaker 1: know how a calculator works and therefore the complex sum

247
00:13:41,159 --> 00:13:42,880
Speaker 1: it does. I just take the advantage of that. I

248
00:13:42,880 --> 00:13:45,750
Speaker 1: don't know exactly how it's done, but if I didn't

249
00:13:45,750 --> 00:13:48,280
Speaker 1: know anything about the way the world works, whether it

250
00:13:48,280 --> 00:13:51,919
Speaker 1: is weather patterns or history of human humanity.

251
00:13:52,549 --> 00:13:55,210
Speaker 1: I would not be a very worthy or a useful

252
00:13:55,830 --> 00:13:57,989
Speaker 1: member of the society. I would just wait for the

253
00:13:57,989 --> 00:13:59,419
Speaker 1: machine to tell me stuff and I would not know

254
00:13:59,419 --> 00:14:01,179
Speaker 1: how to discern the right from the wrong. Right.

255
00:14:01,309 --> 00:14:03,949
Speaker 2: And actually that's also a point of our paper, right,

256
00:14:04,070 --> 00:14:08,150
Speaker 2: which is that um we can use AI to take

257
00:14:08,150 --> 00:14:10,429
Speaker 2: the simple path or we can use AI to actually

258
00:14:10,429 --> 00:14:13,659
Speaker 2: challenge ourselves. And I actually maybe uh we'll talk, I

259
00:14:13,669 --> 00:14:16,109
Speaker 2: I I hope we talk a little bit about AI

260
00:14:16,109 --> 00:14:16,929
Speaker 2: and jobs, but

261
00:14:17,280 --> 00:14:19,520
Speaker 2: Uh, the, the one of the key points that I'd

262
00:14:19,520 --> 00:14:22,380
Speaker 2: like to make here is this, that AI itself

263
00:14:23,250 --> 00:14:25,349
Speaker 2: It's going to take a lot of

264
00:14:27,020 --> 00:14:31,070
Speaker 2: How we do things away, right? And it's gonna simplify,

265
00:14:31,530 --> 00:14:36,210
Speaker 2: but it doesn't change the why and the what. And

266
00:14:36,210 --> 00:14:39,090
Speaker 2: I like to think that's, you know, really where we

267
00:14:39,090 --> 00:14:43,400
Speaker 2: make a a big big leap of, you know, value

268
00:14:43,400 --> 00:14:46,119
Speaker 2: and in in we make a difference that the AI

269
00:14:46,119 --> 00:14:47,669
Speaker 2: cannot get to yet.

270
00:14:48,109 --> 00:14:51,190
Speaker 2: OK, maybe one day it will, um, but not where

271
00:14:51,190 --> 00:14:51,830
Speaker 2: the technology

272
00:14:51,830 --> 00:14:52,349
Speaker 2: is today.

273
00:14:52,510 --> 00:14:54,909
Speaker 1: Because, you know, Sam Altman, I think, famously said last

274
00:14:54,909 --> 00:14:58,039
Speaker 1: year that there will be a billion dollars company with

275
00:14:58,039 --> 00:15:01,530
Speaker 1: an employee of one or a 1 trillion trillionaire coming

276
00:15:01,530 --> 00:15:04,890
Speaker 1: from the AI space. Uh, but my sort of response

277
00:15:04,890 --> 00:15:07,229
Speaker 1: to an observation like that, which sounds rather outrageous and

278
00:15:07,229 --> 00:15:09,340
Speaker 1: scary because if there are no employees, what's the point of,

279
00:15:09,349 --> 00:15:11,950
Speaker 1: you know, having a labor force is that that.

280
00:15:12,520 --> 00:15:15,400
Speaker 1: A person of one selling a billion dollars worth of

281
00:15:15,400 --> 00:15:18,599
Speaker 1: products has to still have buyers of that product. And

282
00:15:18,599 --> 00:15:20,239
Speaker 1: that would not be a person of one, that would

283
00:15:20,239 --> 00:15:22,159
Speaker 1: be a large number of ends required for that.

284
00:15:22,359 --> 00:15:25,299
Speaker 2: And it's a joke, um, you know, that Eric Brimhofton

285
00:15:25,840 --> 00:15:29,309
Speaker 2: told at the last big hype which was data science,

286
00:15:29,440 --> 00:15:33,159
Speaker 2: which is, you know, when, uh, when, uh, when I

287
00:15:33,159 --> 00:15:36,109
Speaker 2: think it's Robert McNamara, I could get the name wrong,

288
00:15:36,400 --> 00:15:38,159
Speaker 2: says to the Ford, um,

289
00:15:38,690 --> 00:15:42,369
Speaker 2: Uh, union chief, right? Everything will be robots. And the

290
00:15:42,369 --> 00:15:44,330
Speaker 2: union chief says, yeah, and robots are going to buy

291
00:15:44,330 --> 00:15:44,530
Speaker 2: your

292
00:15:44,530 --> 00:15:44,809
Speaker 2: cars.

293
00:15:44,969 --> 00:15:48,729
Speaker 1: Leslie, the first three words of this article that we

294
00:15:48,729 --> 00:15:51,849
Speaker 1: wrote was the Star Trek reference to boldly go. I

295
00:15:51,849 --> 00:15:52,750
Speaker 1: didn't come up with that.

296
00:15:53,289 --> 00:15:54,570
Speaker 1: Did you come up with that or the AI came

297
00:15:54,570 --> 00:15:57,210
Speaker 1: up with that? Alright, thank you for that

298
00:15:57,210 --> 00:16:01,010
Speaker 2: deliberately, right, because I am old enough, um, maybe I'm

299
00:16:01,010 --> 00:16:03,369
Speaker 2: a bit older than you, but I'm old enough to

300
00:16:03,369 --> 00:16:04,330
Speaker 2: remember Star Trek.

301
00:16:04,729 --> 00:16:06,739
Speaker 1: I remember every single iteration of Star Trek

302
00:16:07,250 --> 00:16:08,229
Speaker 2: and um.

303
00:16:08,729 --> 00:16:12,130
Speaker 2: For me, Star Trek represents a positive view of the

304
00:16:12,130 --> 00:16:15,729
Speaker 2: future before we had the Matrix and Terminator. That's right.

305
00:16:15,929 --> 00:16:18,450
Speaker 2: And the whole point about this and and Star Trek

306
00:16:18,450 --> 00:16:22,690
Speaker 2: to Golibo was to go and explore and I think

307
00:16:22,690 --> 00:16:26,169
Speaker 2: that's actually what we need to do. Um, there will

308
00:16:26,169 --> 00:16:27,150
Speaker 2: be a lot of

309
00:16:28,400 --> 00:16:33,289
Speaker 2: Job destruction, there will be a lot of changes, but fundamentally,

310
00:16:33,330 --> 00:16:34,159
Speaker 2: it's a positive thing.

311
00:16:35,520 --> 00:16:39,330
Speaker 2: But our mindsets have to change in order to embrace that, right?

312
00:16:39,479 --> 00:16:43,070
Speaker 1: And, and there is, therein lies the issue of policy,

313
00:16:43,080 --> 00:16:44,719
Speaker 1: which I'd like to come to later, but let's go

314
00:16:44,719 --> 00:16:46,719
Speaker 1: to the heart of this article, not the process of

315
00:16:46,719 --> 00:16:50,340
Speaker 1: how we wrote it, but the AI and jobs aspect. Uh,

316
00:16:50,599 --> 00:16:54,969
Speaker 1: what if there is less valuable human work than we thought,

317
00:16:55,330 --> 00:16:58,039
Speaker 1: and then what do we actually do? I mean, if

318
00:16:58,039 --> 00:16:59,760
Speaker 1: a lot of the stuff that we did actually is

319
00:16:59,760 --> 00:17:01,960
Speaker 1: no longer that valuable because the machine can do it

320
00:17:01,960 --> 00:17:02,780
Speaker 1: faster and better.

321
00:17:03,349 --> 00:17:04,280
Speaker 1: What's left for us?

322
00:17:04,520 --> 00:17:07,169
Speaker 2: For me, again, I, I don't.

323
00:17:08,619 --> 00:17:12,418
Speaker 2: I, I'm asked this question a lot and to be honest, um, my,

324
00:17:12,538 --> 00:17:17,659
Speaker 2: my view is the future is not preordained and how

325
00:17:17,659 --> 00:17:22,158
Speaker 2: we define a technology. So, uh, actually Eric Brynhausen wrote

326
00:17:22,159 --> 00:17:24,738
Speaker 2: about this and he said that, you know, he called

327
00:17:24,739 --> 00:17:29,098
Speaker 2: it the Turing paradox or trap, right? That we envisage

328
00:17:29,098 --> 00:17:30,979
Speaker 2: AI to replace us.

329
00:17:32,150 --> 00:17:34,630
Speaker 2: But as economists we know that if it's a labor

330
00:17:34,630 --> 00:17:38,389
Speaker 2: saving technology that complements us, then we are less worried.

331
00:17:38,430 --> 00:17:41,910
Speaker 2: In fact, that has been the story of technology, right? Yes,

332
00:17:42,030 --> 00:17:45,069
Speaker 2: there is some replacement, in fact, replacement of things we

333
00:17:45,069 --> 00:17:45,829
Speaker 2: don't want to do.

334
00:17:47,000 --> 00:17:50,890
Speaker 2: At the same time, the complementarity is very important, right? So,

335
00:17:51,280 --> 00:17:53,919
Speaker 2: very important concept to me. The other important concept to

336
00:17:53,920 --> 00:17:56,140
Speaker 2: me is the difference between a job

337
00:17:56,819 --> 00:18:01,540
Speaker 2: And a task. And um AI is going to replace

338
00:18:01,540 --> 00:18:05,218
Speaker 2: a lot of tasks, but it doesn't have to replace jobs.

339
00:18:05,420 --> 00:18:09,119
Speaker 2: And a great example of that is someone I respect greatly.

340
00:18:09,380 --> 00:18:12,609
Speaker 2: You know, Jeff Hinton, father of, you know, many of

341
00:18:12,609 --> 00:18:15,699
Speaker 2: the things we do, famously said in 2015 there will

342
00:18:15,699 --> 00:18:19,880
Speaker 2: be no radiologists in 5 years' time. I recall that. Yes?

343
00:18:21,489 --> 00:18:26,550
Speaker 2: There are more radiologists today. Yes. And why? It's because

344
00:18:26,689 --> 00:18:30,050
Speaker 2: what radiologists do is not what you think the AI does, right?

345
00:18:30,170 --> 00:18:33,369
Speaker 2: In fact, um, there's also another recent study, uh, in

346
00:18:33,369 --> 00:18:36,099
Speaker 2: Sweden that is a RCT, a randomized uh you know,

347
00:18:36,930 --> 00:18:39,209
Speaker 2: very few of these studies exist, as you know, especially

348
00:18:39,209 --> 00:18:43,290
Speaker 2: in economics, and they found that AI can improve productivity

349
00:18:43,290 --> 00:18:47,229
Speaker 2: by 44%, right? Uh, significant number and reduce errors.

350
00:18:48,050 --> 00:18:50,979
Speaker 2: Doesn't have to reduce the number of jobs, right? So

351
00:18:51,130 --> 00:18:57,349
Speaker 2: I think these are important things for us to to recall. Um,

352
00:18:57,449 --> 00:19:00,530
Speaker 2: the issue of the 1 billion company, actually it's also

353
00:19:00,530 --> 00:19:04,729
Speaker 2: important to think there will be some processes where agents

354
00:19:04,729 --> 00:19:05,650
Speaker 2: can do completely.

355
00:19:06,699 --> 00:19:09,420
Speaker 2: And yes, this will be the $1 billion company with

356
00:19:09,420 --> 00:19:13,859
Speaker 2: 1 person. It's the best market structure to so so

357
00:19:13,859 --> 00:19:15,180
Speaker 2: uh you know, solve this problem.

358
00:19:16,560 --> 00:19:20,839
Speaker 2: But I can imagine a world where that will coexist

359
00:19:20,839 --> 00:19:24,879
Speaker 2: with solutions where yes there's AI but we need a

360
00:19:24,880 --> 00:19:25,680
Speaker 2: lot of humans.

361
00:19:26,459 --> 00:19:30,099
Speaker 2: And I give you two life examples that make me

362
00:19:30,099 --> 00:19:33,250
Speaker 2: optimistic and tell me, tell me about the

363
00:19:34,439 --> 00:19:37,199
Speaker 2: Both about who we are, what we can do.

364
00:19:38,359 --> 00:19:43,140
Speaker 2: One is a food example, McDonald's. McDonald's is a crazily

365
00:19:43,140 --> 00:19:47,560
Speaker 2: efficient food operation. It's global, it has scale.

366
00:19:48,939 --> 00:19:51,300
Speaker 2: But none of us want to eat McDonald's everyday all

367
00:19:51,300 --> 00:19:53,978
Speaker 2: the time, right? And if you go to Japan, we

368
00:19:53,979 --> 00:19:57,339
Speaker 2: have all these mom and pop shops and and some

369
00:19:57,339 --> 00:19:59,979
Speaker 2: of them have Michelin stars, they, they have 10 seats.

370
00:20:00,939 --> 00:20:04,209
Speaker 2: And society is willing to to keep that going, right?

371
00:20:04,339 --> 00:20:09,780
Speaker 2: It's not by any means efficient, right? But it brings value.

372
00:20:10,920 --> 00:20:12,199
Speaker 2: I love photography.

373
00:20:13,060 --> 00:20:16,260
Speaker 2: One of the cameras I love is a Leica, and

374
00:20:16,260 --> 00:20:18,660
Speaker 2: the Leica I love is a rangefinder.

375
00:20:20,020 --> 00:20:25,520
Speaker 2: You pay premium to have a camera that's completely manual.

376
00:20:27,479 --> 00:20:31,339
Speaker 2: And it's very hard to find uh some of these

377
00:20:31,339 --> 00:20:35,489
Speaker 2: Leica cameras, um, because people buy them and

378
00:20:36,319 --> 00:20:39,989
Speaker 2: In a world with Sony's and you know, uh,

379
00:20:40,949 --> 00:20:44,510
Speaker 2: Cameras that are automatic and fast and quick. People are

380
00:20:44,510 --> 00:20:49,579
Speaker 2: willing to spend even more money on something that's manual. Uh,

381
00:20:49,670 --> 00:20:52,709
Speaker 2: so again on, uh, creative, you know, I think creative

382
00:20:52,709 --> 00:20:57,750
Speaker 2: industry photography will suffer, it will be disrupted by AI.

383
00:20:58,069 --> 00:21:02,949
Speaker 2: But I can well imagine human photographers taking certain photos

384
00:21:02,949 --> 00:21:03,949
Speaker 2: will be paid premium.

385
00:21:04,569 --> 00:21:09,489
Speaker 1: Yeah, absolutely, and the desire for humans to sort of

386
00:21:09,489 --> 00:21:13,448
Speaker 1: find satisfaction in craft, uh, that's been going on for

387
00:21:13,449 --> 00:21:16,609
Speaker 1: a very long time and through various waves of automation.

388
00:21:17,109 --> 00:21:20,760
Speaker 1: Uh, actual films for cameras. Let's say you go to forums,

389
00:21:20,920 --> 00:21:24,459
Speaker 1: they're everywhere now, yeah, uh, vinyl records, you know, I mean,

390
00:21:25,199 --> 00:21:26,439
Speaker 1: we thought that it was all gonna be in the

391
00:21:26,439 --> 00:21:29,550
Speaker 1: dustbin of history, but these things make comebacks, and vinyl

392
00:21:29,550 --> 00:21:32,859
Speaker 1: records is no longer a niche industry. It sells a lot. Yes,

393
00:21:33,280 --> 00:21:35,479
Speaker 2: and you're right. I mean, I can remember the day

394
00:21:35,479 --> 00:21:38,829
Speaker 2: when I wanted to get rid of digit of physical form, right?

395
00:21:38,880 --> 00:21:42,760
Speaker 2: It was the goal was to be digital, right? And

396
00:21:42,760 --> 00:21:44,619
Speaker 2: now we're going back to

397
00:21:44,880 --> 00:21:45,290
Speaker 1: physical

398
00:21:45,290 --> 00:21:45,520
Speaker 2: form.

399
00:21:45,739 --> 00:21:46,079
Speaker 1: Absolutely.

400
00:21:46,829 --> 00:21:51,199
Speaker 1: Um, let's, uh, contextualize the future of jobs and the

401
00:21:51,199 --> 00:21:55,040
Speaker 1: risk to tasks as opposed to jobs, uh, in the

402
00:21:55,040 --> 00:21:58,000
Speaker 1: context of ASEAN. Uh, what's your sense? I mean, I

403
00:21:58,000 --> 00:22:01,199
Speaker 1: would probably ask you to add your macroeconomist hat, you know,

404
00:22:01,260 --> 00:22:03,829
Speaker 1: I hope you have not abandoned it entirely, Leslie, uh,

405
00:22:04,079 --> 00:22:07,198
Speaker 1: to answer this question that is ASEAN ready to embrace

406
00:22:07,199 --> 00:22:08,770
Speaker 1: this disruption that's coming? Yeah.

407
00:22:09,650 --> 00:22:11,810
Speaker 2: Maybe before we do that, I, I think I want

408
00:22:11,810 --> 00:22:14,180
Speaker 2: to set and feel free to disagree with me. So,

409
00:22:14,449 --> 00:22:17,369
Speaker 2: so as economists, uh, generally we are, I mean, and

410
00:22:17,369 --> 00:22:21,949
Speaker 2: we observe, technology is the greatest, most important, um, improver

411
00:22:21,949 --> 00:22:22,209
Speaker 2: of

412
00:22:22,209 --> 00:22:25,670
Speaker 2: humanity.

413
00:22:25,770 --> 00:22:29,569
Speaker 2: Productivity has been the reason why we have uh seen

414
00:22:29,569 --> 00:22:34,329
Speaker 2: changes in the last 300 years that are just simply amazing. Uh, secondly,

415
00:22:34,489 --> 00:22:36,959
Speaker 2: we tend to think that at the end of the day,

416
00:22:36,969 --> 00:22:39,089
Speaker 2: over the long term, uh,

417
00:22:39,880 --> 00:22:43,209
Speaker 2: Jobs will be created, and this is what the evidence shows.

418
00:22:43,790 --> 00:22:46,270
Speaker 2: But I want to be more nuanced, um, and the

419
00:22:46,270 --> 00:22:49,510
Speaker 2: nuance comes from reading, uh, I guess in all of

420
00:22:49,510 --> 00:22:53,040
Speaker 2: this I've been reading a lot of history, uh, so

421
00:22:53,040 --> 00:22:56,770
Speaker 2: Carl Benefit Fry writes about this, but there's a literature

422
00:22:57,310 --> 00:23:00,909
Speaker 2: to point out that actually in the industrial revolution 1.0.

423
00:23:01,880 --> 00:23:06,760
Speaker 2: It took 70 to 90 years for wages, real wages

424
00:23:06,760 --> 00:23:08,599
Speaker 2: to survive.

425
00:23:09,619 --> 00:23:12,660
Speaker 2: The initial job destruction.

426
00:23:13,449 --> 00:23:15,869
Speaker 2: And to be very blunt, you know, in a way,

427
00:23:15,930 --> 00:23:20,889
Speaker 2: the industrial revolution happened in England because the political class

428
00:23:20,890 --> 00:23:25,069
Speaker 2: was willing to suppress the workers, right? They shot them,

429
00:23:25,410 --> 00:23:26,489
Speaker 2: they put them in prison.

430
00:23:27,810 --> 00:23:31,209
Speaker 2: It only, it's only when we banned child labor, we

431
00:23:31,209 --> 00:23:35,889
Speaker 2: had other institutions that we have to balance. So basically

432
00:23:35,890 --> 00:23:38,479
Speaker 2: I I guess I would say that the

433
00:23:39,550 --> 00:23:43,859
Speaker 2: The truth is, um, and, and this is made by

434
00:23:44,189 --> 00:23:48,270
Speaker 2: Simon Johnson in a book that wasn't very well received.

435
00:23:48,469 --> 00:23:53,369
Speaker 2: I think progress and uh he wrote it with Darren, um,

436
00:23:53,380 --> 00:23:58,400
Speaker 2: but also echoed in in other in other authors that

437
00:23:58,780 --> 00:24:02,579
Speaker 2: we need policies to address the fact that there will

438
00:24:02,579 --> 00:24:05,579
Speaker 2: be change, right? There will be jobs lost.

439
00:24:07,170 --> 00:24:09,290
Speaker 2: The question is what do we do with those who

440
00:24:09,290 --> 00:24:11,550
Speaker 2: are losing the jobs and how do we create the

441
00:24:11,560 --> 00:24:17,760
Speaker 2: the new jobs, right? And that's about having open innovation, uh, systems,

442
00:24:18,010 --> 00:24:21,569
Speaker 2: creating new things. So you think about this, then you

443
00:24:21,569 --> 00:24:24,209
Speaker 2: think about what's happening in ASEAN, um.

444
00:24:25,030 --> 00:24:29,469
Speaker 2: Um, I observed, I'm actually very worried about white-collar workers.

445
00:24:30,260 --> 00:24:33,599
Speaker 2: And indeed in anthropic study uh in the study recently,

446
00:24:33,939 --> 00:24:37,339
Speaker 2: the good news was, you know, AI can do 993,

447
00:24:37,380 --> 00:24:42,569
Speaker 2: you know, 93% capability, but most companies were only at 33%, right?

448
00:24:42,660 --> 00:24:46,040
Speaker 2: So there was a, it seemed like people were still

449
00:24:46,540 --> 00:24:48,579
Speaker 2: in in the mainstay. But one of the other things

450
00:24:48,579 --> 00:24:52,780
Speaker 2: they observed is that um people were more willing to

451
00:24:52,780 --> 00:24:53,479
Speaker 2: replace AI.

452
00:24:54,500 --> 00:24:58,750
Speaker 2: With or people who had higher degrees were more likely

453
00:24:58,750 --> 00:25:00,010
Speaker 2: to be replaced by AI.

454
00:25:01,050 --> 00:25:04,260
Speaker 2: In more knowledge work and they tended to be females.

455
00:25:05,189 --> 00:25:07,839
Speaker 2: Um, I think the female is a bit of a

456
00:25:07,839 --> 00:25:11,069
Speaker 2: noise because of the labor market in the US. Females

457
00:25:11,069 --> 00:25:15,239
Speaker 2: tend to have higher degrees for the last couple of years. Um, but,

458
00:25:15,479 --> 00:25:18,959
Speaker 2: but basically in ASEAN, who am I worried the most? Actually, I,

459
00:25:19,079 --> 00:25:22,359
Speaker 2: I'm personally not worried about Indonesia or Philippines.

460
00:25:23,050 --> 00:25:26,739
Speaker 2: Because wages are low. I'm worried about Singapore because we

461
00:25:26,739 --> 00:25:32,680
Speaker 2: are eminently replaceable in many, many business operations, uh, many, many, uh,

462
00:25:33,439 --> 00:25:33,969
Speaker 2: wages are

463
00:25:33,969 --> 00:25:34,319
Speaker 2: high.

464
00:25:34,540 --> 00:25:37,659
Speaker 1: So whenever you are trying to manage your cost structure,

465
00:25:37,819 --> 00:25:40,319
Speaker 1: you look at the chunky part and if per capita

466
00:25:40,660 --> 00:25:43,540
Speaker 1: wages are substantially higher, immediately your eye just go there

467
00:25:43,540 --> 00:25:46,169
Speaker 1: and say, how can I sort of either cut it

468
00:25:46,170 --> 00:25:50,739
Speaker 1: or maintain top line growth while not adding to that cost.

469
00:25:51,109 --> 00:25:52,958
Speaker 1: Uh, so there are two risks. One is, if I

470
00:25:52,959 --> 00:25:55,119
Speaker 1: want to cut it, I fire people. Or the second

471
00:25:55,119 --> 00:25:57,739
Speaker 1: is that if I keep growing, I don't hire anybody else.

472
00:25:58,280 --> 00:26:01,520
Speaker 1: So then what happens to the aspirational workers who want

473
00:26:01,520 --> 00:26:03,180
Speaker 1: to come into the workforce? So,

474
00:26:03,520 --> 00:26:06,680
Speaker 2: um, before we talk about possible solutions, uh, I, I

475
00:26:06,680 --> 00:26:08,599
Speaker 2: wanna observe it, I mean, I'm having lunch with some

476
00:26:08,599 --> 00:26:10,050
Speaker 2: friends today and

477
00:26:10,709 --> 00:26:11,579
Speaker 2: Yeah, they were saying

478
00:26:12,390 --> 00:26:15,380
Speaker 2: You know, do I pay for a $4000 intern or

479
00:26:15,380 --> 00:26:19,250
Speaker 2: uh $5000 entry job? Or do I just get clock code?

480
00:26:20,469 --> 00:26:24,489
Speaker 2: And by the way, um, it's clock code max $200 right?

481
00:26:24,869 --> 00:26:28,229
Speaker 2: And with that I can actually have 10 or 100

482
00:26:28,229 --> 00:26:32,229
Speaker 2: or X number, I mean I will have to pay more. Um, uh,

483
00:26:32,239 --> 00:26:34,708
Speaker 2: 11 thing that might not be so evident is actually

484
00:26:34,709 --> 00:26:38,469
Speaker 2: when you do the clock code uh subscriptions or even GPT.

485
00:26:39,239 --> 00:26:43,180
Speaker 2: If you were to see the pricing relative to APIs, um,

486
00:26:43,359 --> 00:26:47,550
Speaker 2: it's almost 10 times different. So, so basically cloud and Anthropic, uh,

487
00:26:47,560 --> 00:26:52,300
Speaker 2: Anthropic and OpenAI are subsidizing. They're giving you a service

488
00:26:52,969 --> 00:26:55,560
Speaker 2: for $100 that is probably costing them $50,000.

489
00:26:55,790 --> 00:27:00,910
Speaker 2: dollars of compute. Right? So this is distorting the market, right? Um,

490
00:27:01,089 --> 00:27:04,250
Speaker 2: but uh but uh again, you know, if I were,

491
00:27:04,489 --> 00:27:08,169
Speaker 2: so for a business, actually I don't blame businesses. Businesses

492
00:27:08,170 --> 00:27:13,150
Speaker 2: will have to optimize for the P&L, right? That's their job. Um,

493
00:27:13,569 --> 00:27:17,250
Speaker 2: so who's left in this picture? Um, this is why

494
00:27:17,250 --> 00:27:20,069
Speaker 2: in our article I thought it was very important that

495
00:27:20,410 --> 00:27:20,839
Speaker 2: uh

496
00:27:21,449 --> 00:27:25,069
Speaker 2: Before we talk about government solutions, actually our solution, we

497
00:27:25,369 --> 00:27:29,130
Speaker 2: mean the individual, and I'm a big big um encourager

498
00:27:29,130 --> 00:27:32,688
Speaker 2: of getting people to use the tools to experiment, to

499
00:27:32,689 --> 00:27:35,489
Speaker 2: understand where it fails. And it's it's important to do

500
00:27:35,489 --> 00:27:39,010
Speaker 2: it all the time because I have many conversations with

501
00:27:39,010 --> 00:27:40,849
Speaker 2: people who say, oh, I used it 6 months ago

502
00:27:40,849 --> 00:27:43,660
Speaker 2: and they hallucinates. Well, 6 months ago it's very different

503
00:27:43,660 --> 00:27:46,109
Speaker 2: than today, right? And it keeps changing.

504
00:27:46,469 --> 00:27:49,300
Speaker 2: And the way we use it actually keeps changing, like,

505
00:27:49,439 --> 00:27:51,520
Speaker 2: you know, as we wrote the paper we went for

506
00:27:51,520 --> 00:27:56,520
Speaker 2: multiple multiple LLMs to agents. Um, now I'm all into

507
00:27:56,520 --> 00:27:59,319
Speaker 2: open claw, right? Which you may or may not, it's

508
00:27:59,319 --> 00:28:03,760
Speaker 2: a very geeky thing and it's just amazing in 3

509
00:28:03,760 --> 00:28:07,739
Speaker 2: months I don't know what else would be the next iteration, right? Um,

510
00:28:08,079 --> 00:28:09,170
Speaker 2: so you

511
00:28:09,530 --> 00:28:12,540
Speaker 2: As an individual, we have to do this, and some

512
00:28:12,540 --> 00:28:15,280
Speaker 2: of the best ways to learn are actually very fast-moving.

513
00:28:15,420 --> 00:28:16,800
Speaker 2: It's not taking a course.

514
00:28:17,910 --> 00:28:20,239
Speaker 2: Whoever is teaching you on a course is out of

515
00:28:20,239 --> 00:28:23,160
Speaker 2: date by definition. That's right. It's actually YouTube.

516
00:28:24,079 --> 00:28:24,969
Speaker 2: Discourse?

517
00:28:26,390 --> 00:28:32,000
Speaker 2: Very, very agile, um, means of communication, right? That's where

518
00:28:32,000 --> 00:28:33,380
Speaker 2: the action is happening.

519
00:28:33,839 --> 00:28:37,599
Speaker 1: But uh Leslie, again talking about ASEAN workforce, you know,

520
00:28:37,680 --> 00:28:40,119
Speaker 1: we're not talking about people sitting in Silicon Valley, we're

521
00:28:40,119 --> 00:28:42,670
Speaker 1: not talking about people with a tech background. There is

522
00:28:42,670 --> 00:28:46,119
Speaker 1: usually some degree of stigma, some degree of fear associated

523
00:28:46,119 --> 00:28:50,550
Speaker 1: with new tech, and the standards that you are sort of,

524
00:28:50,640 --> 00:28:52,660
Speaker 1: you know, asking for is in people.

525
00:28:52,969 --> 00:28:55,609
Speaker 1: Sort of go seek out knowledge or try new things

526
00:28:55,609 --> 00:28:58,689
Speaker 1: all the time. I don't think that characterizes most of humanity.

527
00:28:58,810 --> 00:29:01,349
Speaker 1: Most of humanity has been raised to get a job,

528
00:29:01,890 --> 00:29:04,530
Speaker 1: do a 9 to 5 thing, a task, and then

529
00:29:04,530 --> 00:29:08,209
Speaker 1: come home and then focus on things that make them happy, uh,

530
00:29:08,449 --> 00:29:13,969
Speaker 1: whereas you're basically talking about a lifelong learning process, uh,

531
00:29:14,280 --> 00:29:17,780
Speaker 1: sort of completely breaking down the way we learn things,

532
00:29:17,890 --> 00:29:20,250
Speaker 1: the entire pedagogical model breaking down.

533
00:29:20,859 --> 00:29:24,380
Speaker 1: Is the society, ASEAN or anywhere else, is capable of

534
00:29:24,380 --> 00:29:25,729
Speaker 1: absorbing such a big shift?

535
00:29:26,060 --> 00:29:29,849
Speaker 2: Answer is probably uh no and very difficult, um, but

536
00:29:29,849 --> 00:29:34,420
Speaker 2: it's also why I believe the most important thing is leadership, right? Uh, I,

537
00:29:34,479 --> 00:29:37,780
Speaker 2: I think individual leadership, you have to take a take

538
00:29:37,780 --> 00:29:44,439
Speaker 2: control and own your your uh journey. Uh but company leadership, uh,

539
00:29:44,619 --> 00:29:48,420
Speaker 2: many times in conversations the last few years, people are saying, oh,

540
00:29:48,459 --> 00:29:49,949
Speaker 2: I I got to implement AI.

541
00:29:50,680 --> 00:29:53,880
Speaker 2: I want to implement, what should I do? And my

542
00:29:53,880 --> 00:29:56,349
Speaker 2: general observation is that you're thinking about it from a

543
00:29:56,349 --> 00:29:57,500
Speaker 2: technical point of view.

544
00:29:58,290 --> 00:30:01,729
Speaker 2: You're forgetting the most important thing actually that AI raises

545
00:30:01,729 --> 00:30:05,810
Speaker 2: is not the technology, it's the change in how you

546
00:30:05,810 --> 00:30:09,449
Speaker 2: do things, which is a leadership problem and ultimately at

547
00:30:09,449 --> 00:30:13,329
Speaker 2: the country level. Um, but again the good news is

548
00:30:13,329 --> 00:30:16,349
Speaker 2: in my view for much of ASEAN

549
00:30:17,260 --> 00:30:20,540
Speaker 2: It's probably less dire than you might imagine, so again

550
00:30:20,540 --> 00:30:24,199
Speaker 2: some data, right? Um, can't remember who did this, but

551
00:30:24,390 --> 00:30:27,619
Speaker 2: they were looking at BPOs, right, in Philippines, right, as

552
00:30:27,619 --> 00:30:34,579
Speaker 2: you know, look at the number, it's 3%. Now it

553
00:30:34,579 --> 00:30:38,260
Speaker 2: might be calm before the storm, but my guess is

554
00:30:38,260 --> 00:30:40,859
Speaker 2: that costs are low enough. The last mile is a

555
00:30:40,859 --> 00:30:42,500
Speaker 2: bigger problem than you might imagine.

556
00:30:43,209 --> 00:30:48,099
Speaker 2: And this general technology cannot, cannot help, right? By itself, right?

557
00:30:48,369 --> 00:30:52,619
Speaker 2: It's still not there. Um, but low wages are a powerful,

558
00:30:52,760 --> 00:30:55,530
Speaker 2: I mean, we, we are economists, we know that, yeah,

559
00:30:55,609 --> 00:30:58,770
Speaker 2: the technology will be very shiny and cool, but if

560
00:30:58,770 --> 00:31:02,729
Speaker 2: the price is too high, humans are still a viable alternative.

561
00:31:02,849 --> 00:31:04,430
Speaker 2: I mean this is why self-driving cars

562
00:31:05,000 --> 00:31:07,410
Speaker 2: Probably makes sense in Singapore, but do they make sense

563
00:31:07,410 --> 00:31:10,729
Speaker 2: in Jakarta? Do they make sense in Kuala Lumpur? I mean,

564
00:31:10,810 --> 00:31:14,250
Speaker 2: aside from the technical difficulties of of of it, it's

565
00:31:14,250 --> 00:31:18,189
Speaker 2: just an economic issue. Doesn't mean this is continuing forever.

566
00:31:18,770 --> 00:31:19,239
Speaker 2: Um

567
00:31:20,050 --> 00:31:22,859
Speaker 2: What I do worry about, however, is the following. I,

568
00:31:22,969 --> 00:31:23,910
Speaker 2: I would say that

569
00:31:24,689 --> 00:31:27,150
Speaker 2: It's a challenge, it's an opportunity.

570
00:31:28,589 --> 00:31:31,719
Speaker 2: What AI also does, and we see this in the data,

571
00:31:31,959 --> 00:31:33,560
Speaker 2: is that it levels the playing field.

572
00:31:35,270 --> 00:31:38,270
Speaker 2: So actually I'll flip the question. Another reason why I

573
00:31:38,270 --> 00:31:43,400
Speaker 2: am kind of uh uh worried for Singapore or any

574
00:31:43,400 --> 00:31:44,209
Speaker 2: high wage

575
00:31:45,000 --> 00:31:47,589
Speaker 2: Country is that.

576
00:31:49,739 --> 00:31:54,880
Speaker 2: The average developer in Vietnam and Malaysia in the past

577
00:31:55,099 --> 00:31:58,979
Speaker 2: would be paid, let's say 1/5, 1/3 the price but

578
00:31:58,979 --> 00:32:03,660
Speaker 2: could never reach the quality, could never write English. That's

579
00:32:03,660 --> 00:32:04,579
Speaker 2: all we raised.

580
00:32:05,109 --> 00:32:09,189
Speaker 2: So now we have equal capability at 1/3 and 1/2

581
00:32:09,189 --> 00:32:13,729
Speaker 2: the price. Yeah. This is going to be a challenge.

582
00:32:13,989 --> 00:32:16,469
Speaker 1: I heard this in Indonesia a couple of years ago

583
00:32:16,469 --> 00:32:18,400
Speaker 1: and at that time it was almost a joke, but

584
00:32:18,400 --> 00:32:20,709
Speaker 1: uh now it doesn't sound funny. It actually sounds substantive,

585
00:32:20,750 --> 00:32:21,380
Speaker 1: which was that

586
00:32:22,359 --> 00:32:25,709
Speaker 1: In the region, unlike the Filipinos or the Malaysians or Singaporeans,

587
00:32:25,780 --> 00:32:30,260
Speaker 1: Indonesians actually have fairly poor English capability, and the point

588
00:32:30,260 --> 00:32:31,989
Speaker 1: that that time was being raised is that AI is

589
00:32:31,989 --> 00:32:35,160
Speaker 1: gonna solve that, that even those who don't actually speak

590
00:32:35,160 --> 00:32:37,479
Speaker 1: English very well would now use AI to actually write

591
00:32:37,479 --> 00:32:41,750
Speaker 1: and probably use the voice commands, uh, to come up with, uh, uh,

592
00:32:41,760 --> 00:32:45,920
Speaker 1: audio output that is seamless, articulate, and therefore that gap

593
00:32:45,920 --> 00:32:46,959
Speaker 1: will sort of disappear.

594
00:32:47,319 --> 00:32:49,569
Speaker 1: Uh, yeah, I can now that I'm seeing these sort

595
00:32:49,569 --> 00:32:52,359
Speaker 1: of simultaneous translation tools and so on, which are getting

596
00:32:52,359 --> 00:32:54,949
Speaker 1: better by the day, yep, it's definitely possible.

597
00:32:55,250 --> 00:32:59,520
Speaker 2: I mean, to be honest, I'm guessing 75% of your

598
00:32:59,520 --> 00:33:04,060
Speaker 2: email is all generated by sorry, the email you is

599
00:33:04,479 --> 00:33:06,520
Speaker 2: generated by AI in one form or another,

600
00:33:06,660 --> 00:33:07,160
Speaker 1: right? And also, you know,

601
00:33:07,239 --> 00:33:08,719
Speaker 1: when I look at social media these days, some of

602
00:33:08,719 --> 00:33:10,949
Speaker 1: social media posts have become very long. I wonder who

603
00:33:10,949 --> 00:33:11,280
Speaker 1: has the time.

604
00:33:11,410 --> 00:33:12,459
Speaker 1: To write all these things on the

605
00:33:12,459 --> 00:33:14,020
Speaker 2: LinkedIn, you know, they all sound alike.

606
00:33:14,130 --> 00:33:14,369
Speaker 1: That's

607
00:33:14,369 --> 00:33:16,640
Speaker 1: right. That's right. Yeah. So the, the AI slop is

608
00:33:16,640 --> 00:33:19,010
Speaker 1: certainly there. So maybe we'll pay a premium for actual

609
00:33:19,010 --> 00:33:20,369
Speaker 1: human slop as opposed to AI

610
00:33:20,369 --> 00:33:20,800
Speaker 1: sloth.

611
00:33:21,050 --> 00:33:25,349
Speaker 2: Well, curation, curation will be very important and trust and credibility, right?

612
00:33:25,520 --> 00:33:28,410
Speaker 1: Correct. So let's just, uh, you know, talk about this

613
00:33:28,410 --> 00:33:31,130
Speaker 1: in the context of Singapore a little more. Uh, Singapore

614
00:33:31,130 --> 00:33:34,209
Speaker 1: has a few policy initiatives in place. Skills Future has

615
00:33:34,209 --> 00:33:35,489
Speaker 1: been around for a long time.

616
00:33:36,270 --> 00:33:39,989
Speaker 1: Singaporeans don't have um universal basic income, but they have

617
00:33:39,989 --> 00:33:43,790
Speaker 1: CPF which is somewhat approximating it in that direction. So

618
00:33:43,790 --> 00:33:46,869
Speaker 1: would that make Singapore somewhat better positioned to manage the

619
00:33:46,869 --> 00:33:49,310
Speaker 2: transition? Yeah, I think we have many things that make

620
00:33:49,310 --> 00:33:52,630
Speaker 2: us in a better, better state, right? Relative to many,

621
00:33:52,709 --> 00:33:55,030
Speaker 2: many other countries. Uh, certainly,

622
00:33:55,790 --> 00:33:58,589
Speaker 2: First of all, we are more educated. So even though

623
00:33:58,589 --> 00:34:00,589
Speaker 2: it hits white-collar workers, to me,

624
00:34:02,479 --> 00:34:06,290
Speaker 2: I think we we can handle it. Um, my biggest

625
00:34:06,680 --> 00:34:09,638
Speaker 2: and my pet if you like, uh, everything I've been

626
00:34:09,639 --> 00:34:13,679
Speaker 2: doing since I left um like GIC and now especially

627
00:34:13,679 --> 00:34:18,120
Speaker 2: doing AI Singapore, it's really about confident mindset, right? We

628
00:34:18,120 --> 00:34:22,129
Speaker 2: like to do the Singapore special which I characterize as

629
00:34:22,790 --> 00:34:27,699
Speaker 2: Um, because we are perfectionists, we always say something is wrong.

630
00:34:28,939 --> 00:34:32,280
Speaker 2: I would like us to think that we've matured enough

631
00:34:32,280 --> 00:34:36,000
Speaker 2: that we can say what's possible, right? And so for me,

632
00:34:37,320 --> 00:34:41,909
Speaker 2: No other country, uh, except maybe Norway, Qatar, uh, a

633
00:34:41,909 --> 00:34:45,718
Speaker 2: few oil-producing countries, um, has the resources we do.

634
00:34:46,370 --> 00:34:49,370
Speaker 2: And our educational literacy levels are very high.

635
00:34:50,810 --> 00:34:53,209
Speaker 2: Yes, we might have been doing road learning for a

636
00:34:53,209 --> 00:34:55,570
Speaker 2: long time and now we have to learn a different

637
00:34:55,570 --> 00:35:00,299
Speaker 2: way of doing things, but it's eminently poss possible. Um

638
00:35:01,100 --> 00:35:02,388
Speaker 2: I do think that

639
00:35:05,600 --> 00:35:08,859
Speaker 2: One part that is very challenging for all of us

640
00:35:08,860 --> 00:35:09,360
Speaker 2: is

641
00:35:10,729 --> 00:35:13,750
Speaker 2: We don't know, we don't know a lot of things.

642
00:35:14,169 --> 00:35:17,189
Speaker 2: It's very hard for us to admit that, right?

643
00:35:19,679 --> 00:35:22,800
Speaker 2: One year ago, if I went to Silicon Valley, nobody,

644
00:35:23,050 --> 00:35:25,639
Speaker 2: I mean people were talking about entropic, but it wasn't

645
00:35:25,639 --> 00:35:28,870
Speaker 2: the hype that we have today. Everyone was talking about

646
00:35:28,870 --> 00:35:33,860
Speaker 2: cursor and and and cognition or relet, right, or lovable.

647
00:35:34,929 --> 00:35:37,760
Speaker 2: In one year that has changed dramatically and now these

648
00:35:37,760 --> 00:35:38,989
Speaker 2: companies have an existential

649
00:35:40,149 --> 00:35:44,020
Speaker 2: Right. So in one year from now, we don't know

650
00:35:44,020 --> 00:35:47,639
Speaker 2: because it's improving so fast where it would go. Um,

651
00:35:47,699 --> 00:35:50,459
Speaker 2: but this requires us to be agile in a way,

652
00:35:50,550 --> 00:35:54,850
Speaker 2: to ride the wave, not to be overwhelmed by it, um,

653
00:35:54,860 --> 00:35:56,679
Speaker 2: but to also be willing to say

654
00:35:57,860 --> 00:35:59,399
Speaker 2: Our goals are very clear.

655
00:36:01,350 --> 00:36:04,669
Speaker 2: But the method will have to change, and what works,

656
00:36:05,219 --> 00:36:10,649
Speaker 2: actually we're not very sure. Um, so practical things like

657
00:36:10,649 --> 00:36:13,870
Speaker 2: I think we should give as much training as possible. Uh,

658
00:36:13,949 --> 00:36:18,459
Speaker 2: the caveat is is the training correct and effective. Uh,

659
00:36:18,729 --> 00:36:21,489
Speaker 2: but we may have to do other things. So for example,

660
00:36:21,530 --> 00:36:24,489
Speaker 2: I observed one thing, I, I hesitate to do too

661
00:36:24,489 --> 00:36:27,969
Speaker 2: much policy cause it's not my, um, it's not my

662
00:36:27,969 --> 00:36:29,989
Speaker 2: area anymore, um but

663
00:36:30,830 --> 00:36:35,070
Speaker 2: You know, if we're talking about junior uh entry-level workers.

664
00:36:36,610 --> 00:36:40,060
Speaker 2: To me, it's unfair to ask companies to bear the

665
00:36:40,060 --> 00:36:41,139
Speaker 2: burden of training them.

666
00:36:41,889 --> 00:36:46,009
Speaker 2: Especially if AI will be taking them. Maybe we can

667
00:36:46,010 --> 00:36:48,839
Speaker 2: make some moral argument, but companies will have to do

668
00:36:48,840 --> 00:36:49,888
Speaker 2: what companies have to do.

669
00:36:50,800 --> 00:36:55,270
Speaker 2: It's unfair to to ask the institutions of higher learning

670
00:36:55,520 --> 00:36:59,840
Speaker 2: to train them because I shells also need reform in

671
00:36:59,840 --> 00:37:02,459
Speaker 2: this new world, and they're not set up for that.

672
00:37:03,040 --> 00:37:05,679
Speaker 2: So in a way, um, it may be, and we

673
00:37:05,679 --> 00:37:08,879
Speaker 2: are well placed to do this, government will have to

674
00:37:08,879 --> 00:37:11,319
Speaker 2: be doing this for a few years, right? And I'm

675
00:37:11,320 --> 00:37:14,939
Speaker 2: a generally a market-friendly, not very

676
00:37:15,989 --> 00:37:18,879
Speaker 2: Don't like too much intervention, uh, kind of guy.

677
00:37:19,639 --> 00:37:24,120
Speaker 2: Um, but it might be necessary. In other words, everyone

678
00:37:24,120 --> 00:37:28,580
Speaker 2: who goes to university or polytechnic will get an opportunity

679
00:37:29,080 --> 00:37:32,350
Speaker 2: to be trained so that an employer won't say I

680
00:37:32,350 --> 00:37:35,520
Speaker 2: choose an AI over you because you've got no no experience.

681
00:37:35,679 --> 00:37:40,199
Speaker 1: Leslie, when you and I sort of returned to Singapore

682
00:37:40,199 --> 00:37:43,479
Speaker 1: from our career in Washington DC, we're traveling back 15

683
00:37:43,479 --> 00:37:44,879
Speaker 1: years now or even longer.

684
00:37:45,280 --> 00:37:48,810
Speaker 1: Um, one thing that we used to see regularly featured

685
00:37:48,810 --> 00:37:52,770
Speaker 1: in media in government discussion was the disappointing rate of

686
00:37:52,770 --> 00:37:56,719
Speaker 1: productivity growth in Singapore, and it just became a repeated refrain.

687
00:37:56,820 --> 00:37:59,709
Speaker 1: All the effort that was going in, it was not really, uh,

688
00:37:59,969 --> 00:38:02,879
Speaker 1: leading to anything. But now, um, I have been looking

689
00:38:02,879 --> 00:38:06,370
Speaker 1: at Singapore's TFP data of the last few years. Things

690
00:38:06,370 --> 00:38:06,969
Speaker 1: are turning.

691
00:38:07,479 --> 00:38:10,399
Speaker 1: Uh, so now we are looking at 2%+ productivity growth

692
00:38:10,399 --> 00:38:13,010
Speaker 1: in the last couple of years. Uh, could be for

693
00:38:13,010 --> 00:38:14,850
Speaker 1: a variety of reasons. I'm not just saying it's because

694
00:38:14,850 --> 00:38:17,530
Speaker 1: of AI, but it's an encouraging trend and particularly when

695
00:38:17,530 --> 00:38:20,549
Speaker 1: we look at the ICT sector, very strong productivity growth.

696
00:38:21,239 --> 00:38:23,199
Speaker 1: So let me just ask you this question, not just

697
00:38:23,199 --> 00:38:24,840
Speaker 1: in the context of Singapore, but also in the context

698
00:38:24,840 --> 00:38:27,760
Speaker 1: of the US. Do we have decent data where it

699
00:38:27,760 --> 00:38:30,158
Speaker 1: is pointing in the direction that it is the technology

700
00:38:30,159 --> 00:38:32,270
Speaker 1: that is leading to an improvement in productivity?

701
00:38:32,429 --> 00:38:34,560
Speaker 2: Again, it's been a, you know, I, I don't do

702
00:38:34,560 --> 00:38:37,520
Speaker 2: this as a day job, but I noticed that we're

703
00:38:37,520 --> 00:38:40,340
Speaker 2: getting more and more data that seem to be saying

704
00:38:40,715 --> 00:38:45,395
Speaker 2: That, you know, the solo paradox is actually resolving faster

705
00:38:45,395 --> 00:38:46,205
Speaker 2: than we expected.

706
00:38:47,225 --> 00:38:49,625
Speaker 1: everywhere. Yeah,

707
00:38:49,715 --> 00:38:51,955
Speaker 2: and uh I just mentioned to you there's this new

708
00:38:51,955 --> 00:38:56,145
Speaker 2: paper by uh Ima that did, you know, we we

709
00:38:56,145 --> 00:38:58,584
Speaker 2: always saw, we had a lot of micro evidence of

710
00:38:58,584 --> 00:39:02,034
Speaker 2: the productivity gain from AI but the macro numbers were

711
00:39:02,034 --> 00:39:03,675
Speaker 2: always a bit messy.

712
00:39:05,000 --> 00:39:09,399
Speaker 2: Well, in this paper, they're starting to see uh productivity

713
00:39:09,399 --> 00:39:13,279
Speaker 2: being reflected in the macro numbers. And I mean again,

714
00:39:13,360 --> 00:39:14,489
Speaker 2: if we believe that

715
00:39:15,699 --> 00:39:20,259
Speaker 2: AI is going to improve our efficiency and productivity.

716
00:39:21,520 --> 00:39:23,879
Speaker 2: This is generally good news. I think the caution is

717
00:39:23,879 --> 00:39:27,830
Speaker 2: that um especially when it comes to US data, I mean,

718
00:39:27,989 --> 00:39:29,679
Speaker 2: a couple of things are happening. The structure of the

719
00:39:29,679 --> 00:39:33,520
Speaker 2: world economy is changing, uh we we have uh the

720
00:39:33,520 --> 00:39:37,709
Speaker 2: impact of tariffs, and then we have cuts in data

721
00:39:37,709 --> 00:39:38,178
Speaker 2: collection

722
00:39:38,179 --> 00:39:42,429
Speaker 1: and labor force is shrinking because of cyclical and structural reasons, yes.

723
00:39:42,600 --> 00:39:42,909
Speaker 2: Yeah.

724
00:39:43,639 --> 00:39:46,560
Speaker 2: But at least, you know, we're starting to see that number.

725
00:39:46,800 --> 00:39:49,659
Speaker 2: And by the way, don't uh want to also remind

726
00:39:49,659 --> 00:39:54,639
Speaker 2: um the audience that that this pattern of um you know,

727
00:39:54,800 --> 00:39:58,879
Speaker 2: massive investment, no ROI and it takes time.

728
00:39:59,659 --> 00:40:03,080
Speaker 2: It's not unique to AI. There's a lady called Carletta Perez,

729
00:40:03,300 --> 00:40:06,479
Speaker 2: and her work is not well reflected in part because, um,

730
00:40:06,489 --> 00:40:09,300
Speaker 2: not usually quoted because she's more of a, you know,

731
00:40:09,419 --> 00:40:12,719
Speaker 2: not a pure numbers general economist, but more of a social.

732
00:40:13,340 --> 00:40:17,219
Speaker 2: Uh, anthropologist, economist, I, I, but she documented it happening

733
00:40:17,219 --> 00:40:20,879
Speaker 2: with canals, with railroads, uh, of course the internet boom, right?

734
00:40:21,100 --> 00:40:23,139
Speaker 2: So we are in a phase where there's a massive

735
00:40:23,139 --> 00:40:26,500
Speaker 2: amount of investment, but that massive amount of investment, it's

736
00:40:26,500 --> 00:40:29,689
Speaker 2: not necessarily good news for investors, but that massive amount

737
00:40:29,689 --> 00:40:34,100
Speaker 2: of investment will actually lead to massive amount of supply

738
00:40:34,100 --> 00:40:37,340
Speaker 2: at a cheap price, which will lead to more and

739
00:40:37,340 --> 00:40:37,899
Speaker 2: more innovation.

740
00:40:38,189 --> 00:40:38,600
Speaker 2: Right over time,

741
00:40:39,050 --> 00:40:42,050
Speaker 1: absolutely, I'm actually very grateful to the innovators out of

742
00:40:42,050 --> 00:40:44,969
Speaker 1: China who are basically making everything open source and as consumers,

743
00:40:45,010 --> 00:40:46,928
Speaker 1: you know, that's just, you know, big win-win for us,

744
00:40:47,050 --> 00:40:49,870
Speaker 1: not necessarily for all the billions or other trillions that

745
00:40:49,870 --> 00:40:53,929
Speaker 1: are going into the building the infrastructure of AI. So

746
00:40:53,929 --> 00:40:57,089
Speaker 1: related to that issue is the return, return, right? Yes,

747
00:40:57,500 --> 00:40:59,889
Speaker 1: a lot of input going in, and as a result

748
00:40:59,889 --> 00:41:01,529
Speaker 1: of this a lot of input going in, we're seeing

749
00:41:01,530 --> 00:41:02,659
Speaker 1: an improvement in productivity.

750
00:41:03,010 --> 00:41:05,040
Speaker 1: But it's too much input going in.

751
00:41:05,290 --> 00:41:07,569
Speaker 2: Again, if you take the pattern of history, the answer

752
00:41:07,570 --> 00:41:11,810
Speaker 2: is yes. There will be too much. It's necessary. I

753
00:41:11,810 --> 00:41:15,388
Speaker 2: don't say it's a bug, it's a feature. It takes

754
00:41:15,729 --> 00:41:18,929
Speaker 2: the human exuberance to overdo it for us to get

755
00:41:18,929 --> 00:41:19,750
Speaker 2: the next phase.

756
00:41:20,290 --> 00:41:23,370
Speaker 1: OK. That's a very helpful way of looking at it, uh, Rusty,

757
00:41:23,479 --> 00:41:26,179
Speaker 1: but I think it is based on historical regularity.

758
00:41:26,330 --> 00:41:28,830
Speaker 2: That's one, even if you just look at the US,

759
00:41:29,010 --> 00:41:31,689
Speaker 2: but think about China. All these people are building models

760
00:41:31,689 --> 00:41:32,350
Speaker 2: for free.

761
00:41:33,629 --> 00:41:36,439
Speaker 2: I mean, what deludes them to do so, but in

762
00:41:36,439 --> 00:41:38,350
Speaker 2: the meantime, we benefit. Correct,

763
00:41:38,639 --> 00:41:41,040
Speaker 1: correct. Uh, I mean with even with EVs we've seen

764
00:41:41,040 --> 00:41:41,739
Speaker 1: the same thing.

765
00:41:42,520 --> 00:41:46,320
Speaker 1: The chance of any one company making it is approximately 0%,

766
00:41:46,520 --> 00:41:49,340
Speaker 1: but one company will make it out of the 100%, yes.

767
00:41:49,530 --> 00:41:51,888
Speaker 2: But, but just because it's your program, just note, I'm

768
00:41:51,889 --> 00:41:54,129
Speaker 2: making an AI point. I'm not making the point about

769
00:41:54,129 --> 00:41:58,209
Speaker 2: oversupply and global trade, I think and China macro, which

770
00:41:58,209 --> 00:41:58,919
Speaker 2: is a different issue.

771
00:41:59,169 --> 00:41:59,408
Speaker 1: That's

772
00:41:59,409 --> 00:42:01,449
Speaker 1: a whole different, we can, we can bring that back

773
00:42:01,449 --> 00:42:04,310
Speaker 1: in some other day, uh, Leslie, the excess capacity story.

774
00:42:04,780 --> 00:42:07,110
Speaker 1: OK. So now, uh, I want to talk a little

775
00:42:07,110 --> 00:42:10,110
Speaker 1: bit about the work that you're doing, uh, in AI Singapore,

776
00:42:10,389 --> 00:42:13,110
Speaker 1: especially this Sea Lion, which is the name of my car,

777
00:42:13,189 --> 00:42:16,810
Speaker 1: but also the name of your GPT. Um, so can

778
00:42:17,110 --> 00:42:19,239
Speaker 1: regional AI, I mean, give me the sort of the

779
00:42:19,239 --> 00:42:21,979
Speaker 1: understanding or the logic behind having a regional AI and

780
00:42:21,979 --> 00:42:24,350
Speaker 1: how that sort of competes with those multi-billion dollar models

781
00:42:24,350 --> 00:42:24,549
Speaker 1: out

782
00:42:24,550 --> 00:42:24,929
Speaker 1: there.

783
00:42:25,189 --> 00:42:27,590
Speaker 2: And actually it's a really good question. It's a question

784
00:42:27,590 --> 00:42:30,620
Speaker 2: we ask ourselves all the time, uh, a question I

785
00:42:30,620 --> 00:42:33,109
Speaker 2: have to ask and answer. Um.

786
00:42:33,830 --> 00:42:37,909
Speaker 2: It's important to realize why we're doing this and the

787
00:42:37,909 --> 00:42:40,469
Speaker 2: answer is there are a couple of reas uh what

788
00:42:40,469 --> 00:42:44,069
Speaker 2: it is, right? Um, we're not trying to compete with

789
00:42:44,070 --> 00:42:49,609
Speaker 2: the big guys, right? But big guys try to build

790
00:42:49,610 --> 00:42:55,239
Speaker 2: things to meet general needs and they're not always specific.

791
00:42:55,639 --> 00:42:59,189
Speaker 2: I give you a real common example today. If you

792
00:42:59,189 --> 00:43:01,388
Speaker 2: use some of these closed models and you say speak

793
00:43:01,389 --> 00:43:02,770
Speaker 2: in English uh Singlish,

794
00:43:03,679 --> 00:43:09,340
Speaker 2: It's cringe in Singlish. No Singaporean would speak Singlish that way, right?

795
00:43:09,679 --> 00:43:12,469
Speaker 2: It's a data problem. So the first thing about sea

796
00:43:12,469 --> 00:43:16,320
Speaker 2: lion is actually not the models, it's the data. The

797
00:43:16,320 --> 00:43:20,290
Speaker 2: second is evaluation. How do we know what is good?

798
00:43:20,600 --> 00:43:23,219
Speaker 2: And to be honest, when we started off and today,

799
00:43:23,939 --> 00:43:27,860
Speaker 2: Uh, in, in areas like voice, we have no good

800
00:43:27,860 --> 00:43:32,139
Speaker 2: benchmark of what is good because again researchers or companies

801
00:43:32,139 --> 00:43:35,299
Speaker 2: will do the English, they will do Chinese. Well they

802
00:43:35,300 --> 00:43:38,138
Speaker 2: do proper Basa and worse yet Malay.

803
00:43:38,889 --> 00:43:44,729
Speaker 2: There are very few really good consistent, comprehensive Malay benchmarks, right?

804
00:43:45,010 --> 00:43:48,929
Speaker 2: So if I cannot even articulate what is good, how

805
00:43:48,929 --> 00:43:53,439
Speaker 2: do I build a better model? So to us, data model,

806
00:43:53,610 --> 00:43:56,969
Speaker 2: these are the unsung heroes in a way, regardless of

807
00:43:56,969 --> 00:43:59,489
Speaker 2: how good or where we go with models, we will

808
00:43:59,489 --> 00:44:03,250
Speaker 2: always need them. And lastly is the models, um,

809
00:44:03,600 --> 00:44:06,379
Speaker 2: Yes, we have to do them because it proves credibility,

810
00:44:06,479 --> 00:44:09,340
Speaker 2: it brings everything together, and we open source it for

811
00:44:09,340 --> 00:44:10,040
Speaker 2: people to use.

812
00:44:10,770 --> 00:44:13,330
Speaker 2: Um, but it's also a network. Uh, I like to

813
00:44:13,330 --> 00:44:16,250
Speaker 2: think we're doing this in the open. We have collaborators

814
00:44:16,250 --> 00:44:21,020
Speaker 2: in Indonesia, in Thailand, in every ASEAN country actually, and

815
00:44:22,270 --> 00:44:26,350
Speaker 2: We are trying to build things that are public comments

816
00:44:26,709 --> 00:44:30,090
Speaker 2: so that companies can build on top and create the

817
00:44:30,100 --> 00:44:33,909
Speaker 2: the value. And it's uh not a new idea, it's

818
00:44:33,909 --> 00:44:38,550
Speaker 2: an open-source idea, it's an idea that you you think

819
00:44:38,550 --> 00:44:40,709
Speaker 2: about when you don't have a lot of resources.

820
00:44:41,149 --> 00:44:43,310
Speaker 2: And I I'm proud of a couple of things. One

821
00:44:43,310 --> 00:44:46,610
Speaker 2: is because of what we do, we get the attention

822
00:44:46,610 --> 00:44:50,060
Speaker 2: of big tech like Google and Nvidia and they're willing

823
00:44:50,060 --> 00:44:52,570
Speaker 2: to put resources into making

824
00:44:53,689 --> 00:44:56,969
Speaker 2: Yes, they're making the data better, they're making the benchmark better.

825
00:44:57,250 --> 00:44:59,929
Speaker 2: They're making their models better, so Gemini gets better.

826
00:45:00,750 --> 00:45:05,310
Speaker 2: Uh, but they're also helping build expertise in the region, right?

827
00:45:05,590 --> 00:45:08,469
Speaker 2: And um this is happening with all the big tech

828
00:45:08,469 --> 00:45:13,469
Speaker 2: actually and uh rather uniquely in Singapore, we can bring

829
00:45:13,469 --> 00:45:18,270
Speaker 2: Chinese and and uh US big tech, right? So um

830
00:45:18,270 --> 00:45:18,790
Speaker 2: this is

831
00:45:19,560 --> 00:45:22,310
Speaker 2: Anchors who we are, it reflects who we are as

832
00:45:22,310 --> 00:45:26,550
Speaker 2: a as a hub, right? Um, that's really, so ultimately,

833
00:45:26,840 --> 00:45:30,290
Speaker 2: Sea Lion is not about trying to beat the big model,

834
00:45:30,679 --> 00:45:35,679
Speaker 2: it's trying to build certain capabilities, anchor it here, fill

835
00:45:35,679 --> 00:45:36,600
Speaker 2: certain gaps.

836
00:45:37,610 --> 00:45:38,929
Speaker 2: But keep it in the open.

837
00:45:39,209 --> 00:45:41,589
Speaker 1: That's uh very promising. I'm very excited actually. I have

838
00:45:41,810 --> 00:45:44,569
Speaker 1: not downloaded Sea Lion. I'm guilty as charged, and that'll

839
00:45:44,570 --> 00:45:47,830
Speaker 1: be the thing to do for me this evening, uh, Leslie. Um,

840
00:45:48,330 --> 00:45:50,850
Speaker 1: I work at a bank. You spent a very large

841
00:45:50,850 --> 00:45:53,810
Speaker 1: part of your career working in financial institutions. Let's talk

842
00:45:53,810 --> 00:45:57,250
Speaker 1: about the financial sector and the risk it faces of

843
00:45:57,250 --> 00:46:00,129
Speaker 1: disruption or challenges coming from these, uh, models.

844
00:46:00,229 --> 00:46:04,149
Speaker 2: Actually, I'm, I'm less worried for the following reason.

845
00:46:05,899 --> 00:46:06,419
Speaker 2: Um,

846
00:46:08,500 --> 00:46:11,979
Speaker 2: It's a highly regulated industry and to be honest, there

847
00:46:11,979 --> 00:46:15,580
Speaker 2: is a compact between regulators and the banks. And I'm

848
00:46:15,580 --> 00:46:17,859
Speaker 2: talking about banks here. Um

849
00:46:19,209 --> 00:46:22,250
Speaker 2: It's not so easy for an incumbent, as we've seen

850
00:46:22,250 --> 00:46:26,169
Speaker 2: this with digital banks and many others, to overcome the

851
00:46:26,169 --> 00:46:30,989
Speaker 2: franchise and the regulatory uh sand sandbox. And no matter,

852
00:46:31,080 --> 00:46:33,969
Speaker 2: you know, to be very frank, um, I feel like

853
00:46:33,969 --> 00:46:37,009
Speaker 2: no bank in Singapore has a good chatbot. Customer service

854
00:46:37,010 --> 00:46:39,109
Speaker 2: is not great, but

855
00:46:39,610 --> 00:46:42,850
Speaker 2: You have no choice, right? And you're not likely to

856
00:46:42,850 --> 00:46:47,659
Speaker 2: put your savings, especially the bulk of your savings in

857
00:46:47,659 --> 00:46:51,610
Speaker 2: some new institution that has a two-year track record, no

858
00:46:51,610 --> 00:46:53,330
Speaker 2: matter how good the customer experience.

859
00:46:54,149 --> 00:46:57,909
Speaker 2: Um, you, you might use it for transactions, but not

860
00:46:57,909 --> 00:47:01,750
Speaker 2: much more, right? So, I think the disruption is really

861
00:47:01,750 --> 00:47:05,860
Speaker 2: not banks and new banks. Actually, the disruption will be

862
00:47:05,860 --> 00:47:09,989
Speaker 2: with other banks and the disruption will be how banks

863
00:47:09,989 --> 00:47:12,949
Speaker 2: happen within the banks. I think there are many, many

864
00:47:12,949 --> 00:47:18,090
Speaker 2: tasks that banks so we could be running banks, uh,

865
00:47:18,510 --> 00:47:21,540
Speaker 2: you know, maybe not great for labor force but with

866
00:47:21,540 --> 00:47:22,689
Speaker 2: a lot less people.

867
00:47:23,669 --> 00:47:25,908
Speaker 2: And a lot more efficiency, right?

868
00:47:26,550 --> 00:47:27,149
Speaker 2: Um,

869
00:47:28,560 --> 00:47:31,179
Speaker 2: Because a lot of compliance work.

870
00:47:32,360 --> 00:47:33,500
Speaker 2: AML CFT

871
00:47:34,689 --> 00:47:35,500
Speaker 2: Scams

872
00:47:36,810 --> 00:47:41,840
Speaker 2: These are heavy on the documentation, they're heavy on the checking,

873
00:47:42,239 --> 00:47:47,389
Speaker 2: they're eminently AIable, right? In compli you know, in general compliance, uh,

874
00:47:47,399 --> 00:47:50,959
Speaker 2: risk management to some extent. I still think getting the

875
00:47:50,959 --> 00:47:51,658
Speaker 2: customer

876
00:47:52,659 --> 00:47:56,658
Speaker 2: Really making good credit decisions, um, AI will help, but

877
00:47:56,659 --> 00:47:59,979
Speaker 2: humans are probably gonna still do a, a lot of it.

878
00:48:00,580 --> 00:48:03,419
Speaker 1: So, uh, let's see, I think the one issue is

879
00:48:03,419 --> 00:48:07,840
Speaker 1: the availability of data across various privacy barriers, uh, which

880
00:48:07,840 --> 00:48:10,580
Speaker 1: still sort of keeps some of the compliance work or

881
00:48:10,580 --> 00:48:12,580
Speaker 1: some of the KOYC work that takes place in the

882
00:48:12,580 --> 00:48:15,679
Speaker 1: manual sphere because we just don't have databases that allow

883
00:48:15,679 --> 00:48:18,659
Speaker 1: us to just, you know, send the AI through it and,

884
00:48:18,699 --> 00:48:20,658
Speaker 1: and come up with the solution that we could, right?

885
00:48:20,979 --> 00:48:24,219
Speaker 1: Uh, I, I see that in particularly as sort of

886
00:48:24,219 --> 00:48:28,300
Speaker 1: the customer becomes wealthier, things become more manual. Uh, customer

887
00:48:28,300 --> 00:48:32,939
Speaker 1: himself prefers or herself prefer high touch RMs and senior

888
00:48:32,939 --> 00:48:35,520
Speaker 1: management going and having a chit chat or people like

889
00:48:35,520 --> 00:48:39,000
Speaker 1: me giving them presentations. The desire for that seems undiminished.

890
00:48:39,179 --> 00:48:42,570
Speaker 2: I, I, and I will validate that, right? Because you

891
00:48:42,570 --> 00:48:46,000
Speaker 2: want that, right? But that doesn't mean, you know, like, uh,

892
00:48:46,020 --> 00:48:46,719
Speaker 2: so to be.

893
00:48:47,379 --> 00:48:50,179
Speaker 2: Uh, to, to put some ideas on the table, you

894
00:48:50,179 --> 00:48:54,060
Speaker 2: have to write reports. I mean those will be automated, right? Um,

895
00:48:54,260 --> 00:48:56,870
Speaker 2: before you see the client, you're gonna get a profile

896
00:48:56,870 --> 00:49:00,739
Speaker 2: and you're gonna get a personalized what you can say

897
00:49:00,739 --> 00:49:02,860
Speaker 2: and what you should say for that client. That's all

898
00:49:02,860 --> 00:49:06,540
Speaker 2: driven by AI, right? So actually you, it's back to

899
00:49:06,540 --> 00:49:09,459
Speaker 2: our article. The parts of the job that hopefully you

900
00:49:09,459 --> 00:49:13,939
Speaker 2: feel more satisfaction, which is actually talking to people, listening

901
00:49:13,939 --> 00:49:16,500
Speaker 2: to them, understanding what their real problems are.

902
00:49:17,379 --> 00:49:20,819
Speaker 2: will be where you find and create value. The parts

903
00:49:20,820 --> 00:49:23,299
Speaker 2: of their job like creating an asset allocation.

904
00:49:23,929 --> 00:49:25,810
Speaker 2: Well, AI can do that, right? In fact, I can

905
00:49:25,810 --> 00:49:30,129
Speaker 2: do 30 of them and it can do the pitch

906
00:49:30,129 --> 00:49:33,949
Speaker 2: deck for you and yeah. So, unfortunately, what does happen

907
00:49:33,949 --> 00:49:37,209
Speaker 2: um I probably is we used to hire interns to

908
00:49:37,209 --> 00:49:39,429
Speaker 2: do this. They're not gonna be there.

909
00:49:39,689 --> 00:49:42,560
Speaker 1: Right? Well, therein lies the problem that if you don't

910
00:49:42,560 --> 00:49:45,209
Speaker 1: learn about the institution as an intern, how do you

911
00:49:45,209 --> 00:49:47,489
Speaker 1: get in sort of you know admitted as a young

912
00:49:47,489 --> 00:49:49,050
Speaker 1: professional and if you're not a young professional, how do

913
00:49:49,050 --> 00:49:50,530
Speaker 1: you ever manage middle management, right?

914
00:49:51,000 --> 00:49:52,879
Speaker 1: So, so that's something that we have to sort of

915
00:49:52,879 --> 00:49:53,840
Speaker 1: think through. Yeah.

916
00:49:54,040 --> 00:49:57,060
Speaker 2: Well, um, in our paper, what did we say? I,

917
00:49:57,110 --> 00:49:59,580
Speaker 2: I kind of think this way, right? Ah, traditional company,

918
00:49:59,909 --> 00:50:02,399
Speaker 2: and especially, let's take a back, you you are are

919
00:50:02,679 --> 00:50:04,909
Speaker 2: very sensitive, right? You're not gonna put an intern in

920
00:50:04,909 --> 00:50:08,399
Speaker 2: front of a client, right? You're not. Uh, prob problem

921
00:50:08,399 --> 00:50:11,560
Speaker 2: is you if an intern can't do certain things, you

922
00:50:11,560 --> 00:50:14,739
Speaker 2: also don't wanna hire an intern now as we just mentioned. So,

923
00:50:15,879 --> 00:50:18,799
Speaker 2: What could change is the following. Actually, I, I need

924
00:50:18,800 --> 00:50:22,520
Speaker 2: to make the intern to 70% of TMO in 3 months.

925
00:50:22,959 --> 00:50:23,750
Speaker 2: How do I do that?

926
00:50:25,310 --> 00:50:28,229
Speaker 2: He or she is going to interact with hundreds of

927
00:50:28,229 --> 00:50:33,389
Speaker 2: AI agents who are potential customers, right? And be faced

928
00:50:33,669 --> 00:50:36,060
Speaker 2: with many, you know, it's not about doing the deck anymore.

929
00:50:36,350 --> 00:50:39,949
Speaker 2: It's actually doing what you do in a in a

930
00:50:39,949 --> 00:50:43,090
Speaker 2: in a more structured, so we call this flight simulators, right?

931
00:50:43,699 --> 00:50:45,870
Speaker 2: In medical, we call it, you know, doing the rounds

932
00:50:45,870 --> 00:50:48,529
Speaker 2: in legal we say this is what happens in class, right?

933
00:50:49,030 --> 00:50:50,989
Speaker 2: So I think we will go back to a lot

934
00:50:50,989 --> 00:50:53,550
Speaker 2: more of that or could go back to more of that.

935
00:50:54,030 --> 00:50:56,979
Speaker 1: Um, why don't we just go back to the sea

936
00:50:56,979 --> 00:51:00,729
Speaker 1: lion question for a second. Um, there are certain ubiquitous

937
00:51:00,729 --> 00:51:05,310
Speaker 1: aspects of ASEAN, so Islamic finance, for example, or the

938
00:51:05,310 --> 00:51:07,860
Speaker 1: way certain loan pools work in this part of the world,

939
00:51:07,909 --> 00:51:09,189
Speaker 1: which is very different from the way it does in

940
00:51:09,189 --> 00:51:11,069
Speaker 1: the US and so on. So are you also trying

941
00:51:11,070 --> 00:51:11,750
Speaker 1: to capture those

942
00:51:11,750 --> 00:51:12,320
Speaker 1: nuances?

943
00:51:12,449 --> 00:51:16,069
Speaker 2: So it's not, it's nuances. So maybe I, I want

944
00:51:16,070 --> 00:51:18,569
Speaker 2: to um specifically say

945
00:51:19,159 --> 00:51:22,360
Speaker 2: So our main goal is to build capacity, but actually

946
00:51:22,360 --> 00:51:25,739
Speaker 2: the gap we're trying to fill that is real, is you,

947
00:51:25,879 --> 00:51:30,199
Speaker 2: as you know, just pointed out, local knowledge or regional knowledge.

948
00:51:30,560 --> 00:51:34,419
Speaker 2: So finance is one, in medical, we get questions from doctors,

949
00:51:35,159 --> 00:51:39,399
Speaker 2: actually how do I communicate with uh again religion?

950
00:51:40,159 --> 00:51:44,939
Speaker 2: And context is very important. What food can you eat, right? Uh,

951
00:51:45,040 --> 00:51:46,959
Speaker 2: AI can tell you to eat pork if you are

952
00:51:46,959 --> 00:51:49,239
Speaker 2: in Jakarta, right, and you're Muslim.

953
00:51:49,989 --> 00:51:55,320
Speaker 2: Um, what happens about, you know, very sensitive private practices

954
00:51:55,419 --> 00:51:58,729
Speaker 2: which differ culture to culture, right? Or even how you

955
00:51:58,729 --> 00:52:02,810
Speaker 2: greet people can be very different. So, sea lions trying

956
00:52:02,810 --> 00:52:07,419
Speaker 2: to meet that that need and it's an expression of

957
00:52:07,419 --> 00:52:07,879
Speaker 2: saying

958
00:52:08,919 --> 00:52:13,159
Speaker 2: If we have agents that know this, then people are

959
00:52:13,159 --> 00:52:16,399
Speaker 2: more willing to use the AI. It feels more, uh,

960
00:52:16,600 --> 00:52:18,040
Speaker 2: they feel more comfortable.

961
00:52:18,399 --> 00:52:24,319
Speaker 1: Fascinating. Um, Leslie, in the final part of this issue, uh, discussion,

962
00:52:24,399 --> 00:52:26,840
Speaker 1: I would really want to bring in something, uh, first

963
00:52:26,840 --> 00:52:29,198
Speaker 1: of all, congratulations, being one of the 40 members of

964
00:52:29,199 --> 00:52:32,409
Speaker 1: the UN's independent International scientific panel on AI.

965
00:52:33,370 --> 00:52:34,530
Speaker 1: Tell us a bit about that.

966
00:52:34,850 --> 00:52:35,129
Speaker 2: It's a

967
00:52:35,129 --> 00:52:37,389
Speaker 2: very unique thing. I actually didn't know much about them,

968
00:52:37,610 --> 00:52:40,770
Speaker 2: but um it came about from a long process of

969
00:52:40,770 --> 00:52:44,569
Speaker 2: trying to deliberate how um what role can the UN

970
00:52:44,570 --> 00:52:50,449
Speaker 2: play in the global AI debate, right? And the recommendation

971
00:52:50,449 --> 00:52:55,429
Speaker 2: was to create a panel that's independent and scientific. Um,

972
00:52:55,449 --> 00:52:57,908
Speaker 2: it's actually unique, it's not happened before.

973
00:52:58,719 --> 00:53:02,689
Speaker 2: Uh, there is the IPCC as um closer analog, but

974
00:53:02,689 --> 00:53:08,469
Speaker 2: that was also a process that had government involved. Yes. Um,

975
00:53:08,530 --> 00:53:11,889
Speaker 2: but you will recall actually IPCC was very important in

976
00:53:11,889 --> 00:53:16,659
Speaker 2: setting certain consensus, right? And actually that's what the committee

977
00:53:16,659 --> 00:53:20,620
Speaker 2: is set up to do, uh, to provide independent

978
00:53:21,540 --> 00:53:26,959
Speaker 2: And scientifically evidence-based, you know, and, and actually let me

979
00:53:26,959 --> 00:53:30,000
Speaker 2: elaborate on that. You know, science

980
00:53:30,840 --> 00:53:32,840
Speaker 2: You know, if something is not researched, you don't have

981
00:53:32,840 --> 00:53:35,580
Speaker 2: an answer and many things are not researched in this area.

982
00:53:36,120 --> 00:53:39,389
Speaker 2: But I think it's the discipline of science, meaning being

983
00:53:39,389 --> 00:53:41,699
Speaker 2: as rigorous as possible, evidence-based.

984
00:53:43,060 --> 00:53:46,560
Speaker 2: And perhaps to both look at the evidence, but also

985
00:53:47,530 --> 00:53:48,959
Speaker 2: future areas for research.

986
00:53:49,719 --> 00:53:51,919
Speaker 2: And the goal was, you know, first I thought it

987
00:53:51,919 --> 00:53:55,879
Speaker 2: was mostly about governance, but to be frank from what

988
00:53:55,879 --> 00:54:00,439
Speaker 2: I've um seen so far, um, it's a very encompassing

989
00:54:00,439 --> 00:54:04,520
Speaker 2: set of issues. And again, the idea is to, you know, can,

990
00:54:04,639 --> 00:54:08,719
Speaker 2: in a world that's very, you know, split up with

991
00:54:08,719 --> 00:54:09,959
Speaker 2: many different views,

992
00:54:10,669 --> 00:54:13,149
Speaker 2: Uh, in fact, you can even question the relevance of

993
00:54:13,149 --> 00:54:16,260
Speaker 2: this committee. Uh, the US actually voted against it.

994
00:54:16,510 --> 00:54:17,570
Speaker 1: How am I not surprised?

995
00:54:18,439 --> 00:54:24,550
Speaker 2: 117 countries approved, 2 abstained and 22 said no, and

996
00:54:24,550 --> 00:54:26,429
Speaker 2: one of them was the US, um.

997
00:54:27,439 --> 00:54:30,239
Speaker 2: But I'd like to think there is an opportunity to

998
00:54:30,239 --> 00:54:34,339
Speaker 2: kind of say, you know, this is objective, this is independent,

999
00:54:34,679 --> 00:54:38,439
Speaker 2: this is for the good of humanity, and from there,

1000
00:54:38,719 --> 00:54:40,179
Speaker 2: then we can build good policy.

1001
00:54:40,399 --> 00:54:42,040
Speaker 1: So where does the work stand now?

1002
00:54:42,239 --> 00:54:46,270
Speaker 2: Um, so it just was established this month and we

1003
00:54:46,270 --> 00:54:47,359
Speaker 2: have a March, yes.

1004
00:54:47,580 --> 00:54:49,319
Speaker 1: Oh wow, OK, just now. It's for 3

1005
00:54:49,320 --> 00:54:52,800
Speaker 2: years. Yeah, it's um for 3 years and it has

1006
00:54:52,800 --> 00:54:55,080
Speaker 2: its first report due in July, not a lot of time.

1007
00:54:55,659 --> 00:54:58,389
Speaker 2: Uh, so we'll, we'll see what happens, um.

1008
00:54:59,239 --> 00:55:03,439
Speaker 2: Maybe I would end with with this uh thought um

1009
00:55:03,439 --> 00:55:04,389
Speaker 2: on on AI.

1010
00:55:05,350 --> 00:55:08,270
Speaker 2: I think you and I are the generation, we were

1011
00:55:08,270 --> 00:55:11,799
Speaker 2: preachers of open markets and free trade. I don't know

1012
00:55:11,800 --> 00:55:14,310
Speaker 2: if people still admit to that, to be honest.

1013
00:55:14,429 --> 00:55:15,080
Speaker 1: I'm not saying anything,

1014
00:55:16,469 --> 00:55:16,810
Speaker 2: but

1015
00:55:18,459 --> 00:55:23,580
Speaker 2: We were both intellectually in school, believers of that, but

1016
00:55:23,580 --> 00:55:26,659
Speaker 2: we preached it at the IMF, right? And we negotiated

1017
00:55:26,659 --> 00:55:30,560
Speaker 2: with countries to make sure this has happened. I like

1018
00:55:30,560 --> 00:55:30,929
Speaker 2: to

1019
00:55:32,199 --> 00:55:36,540
Speaker 2: You know, retrospectively, I think that we as economists didn't

1020
00:55:36,540 --> 00:55:40,520
Speaker 2: think hard about the consequences. I mean David Arter's work

1021
00:55:40,520 --> 00:55:42,820
Speaker 2: on the China impact, for example.

1022
00:55:43,699 --> 00:55:44,159
Speaker 2: And

1023
00:55:45,540 --> 00:55:48,259
Speaker 2: You know, again, economists will say things like, oh yeah,

1024
00:55:48,340 --> 00:55:51,419
Speaker 2: of course jobs will be destroyed and jobs will be created.

1025
00:55:51,739 --> 00:55:55,399
Speaker 2: If you look at consultants, WAF says 70 million jobs destroyed,

1026
00:55:55,860 --> 00:56:00,600
Speaker 2: 170 million jobs created. But we as economists don't say how, right?

1027
00:56:01,580 --> 00:56:04,100
Speaker 2: For me, um, part of Sea Lion, part of what

1028
00:56:04,100 --> 00:56:07,899
Speaker 2: we do, is actually trying to understand the how, right?

1029
00:56:08,179 --> 00:56:09,000
Speaker 2: And the how

1030
00:56:10,510 --> 00:56:14,949
Speaker 2: Because I do believe open markets and free trade was

1031
00:56:14,949 --> 00:56:18,350
Speaker 2: the AI of our times. And it did deliver, right?

1032
00:56:18,409 --> 00:56:22,120
Speaker 2: It brought billions of people out of poverty. But it

1033
00:56:22,120 --> 00:56:24,609
Speaker 1: also brought in populism and Donald Trump.

1034
00:56:25,270 --> 00:56:28,629
Speaker 2: And that's because we didn't understand or think about the

1035
00:56:28,629 --> 00:56:31,439
Speaker 2: other parts of it, right? Which is not just about

1036
00:56:31,439 --> 00:56:34,669
Speaker 2: people being left behind but social consequences and so on.

1037
00:56:35,110 --> 00:56:35,459
Speaker 2: So

1038
00:56:36,209 --> 00:56:39,529
Speaker 2: The AI panel for me is a way of trying

1039
00:56:39,530 --> 00:56:43,649
Speaker 2: to say, how can we think hard? Yes, the technology

1040
00:56:43,649 --> 00:56:44,429
Speaker 2: is powerful.

1041
00:56:45,360 --> 00:56:48,279
Speaker 2: But how can we do it in a way that

1042
00:56:48,280 --> 00:56:50,580
Speaker 2: will not repeat the mistakes of globalization?

1043
00:56:51,000 --> 00:56:54,080
Speaker 1: Uh, Leslie, I think that is just a very, very

1044
00:56:54,080 --> 00:56:56,919
Speaker 1: apt note to end this conversation. Thank you so much

1045
00:56:56,919 --> 00:56:58,239
Speaker 1: for your time and insights. No, thank

1046
00:56:58,239 --> 00:56:59,159
Speaker 2: you. It's a pleasure.

1047
00:56:59,489 --> 00:57:01,340
Speaker 1: It's finally, it's better late

1048
00:57:01,340 --> 00:57:03,629
Speaker 2: than never. Yeah, and, and just so the audience knows,

1049
00:57:03,649 --> 00:57:05,790
Speaker 2: it took so long, not because Timo didn't want to

1050
00:57:05,790 --> 00:57:07,800
Speaker 2: invite me, it's because I was avoiding him for.

1051
00:57:09,850 --> 00:57:11,959
Speaker 1: OK, thanks for setting the record straight. Thanks.

1052
00:57:12,064 --> 00:57:14,745
Speaker 1: To our listeners as well. Kobe Time was produced by

1053
00:57:14,745 --> 00:57:17,385
Speaker 1: Ken Delbridge at Spice Studios. Violet Lee and Daisy S

1054
00:57:17,385 --> 00:57:21,485
Speaker 1: Sherma provided additional assistance. This podcast is for information only

1055
00:57:21,625 --> 00:57:26,024
Speaker 1: and does not constitute any trade recommendation. All 174 episodes

1056
00:57:26,024 --> 00:57:28,425
Speaker 1: of this podcast are available on YouTube as well as

1057
00:57:28,425 --> 00:57:32,554
Speaker 1: on on major podcast platforms including Apple, Google, and Spotify.

1058
00:57:32,864 --> 00:57:35,864
Speaker 1: As for our research publications webinars, you can find them

1059
00:57:35,864 --> 00:57:38,804
Speaker 1: by Googling DBS Research Library. Have a great day.