1
00:00:00,720 --> 00:00:02,960
Speaker 1: This week on the Business of Tech, powered by two

2
00:00:02,960 --> 00:00:07,720
Speaker 1: Degrees business visiting, machine learning and AI expert Jessica Zosa

3
00:00:08,000 --> 00:00:12,039
Speaker 1: four joins us to talk large language models, multi lingual

4
00:00:12,200 --> 00:00:15,880
Speaker 1: AI and how it can actually get more transparency and

5
00:00:15,960 --> 00:00:20,360
Speaker 1: accountability as AI starts to infiltrate nearly every aspect of

6
00:00:20,400 --> 00:00:21,080
Speaker 1: modern life.

7
00:00:21,200 --> 00:00:25,600
Speaker 2: Fundamentally, you can know if it makes a positive outcome

8
00:00:25,680 --> 00:00:28,680
Speaker 2: on your day to day workflow. You know, if it's

9
00:00:28,680 --> 00:00:32,600
Speaker 2: not making your life easier and it's making things more frustrating,

10
00:00:32,920 --> 00:00:35,839
Speaker 2: then you know this the inclusion of this technology may

11
00:00:35,880 --> 00:00:36,680
Speaker 2: be a problem.

12
00:00:37,159 --> 00:00:40,440
Speaker 3: Plus, Peter heads to Pallo Alto and catches up with

13
00:00:40,479 --> 00:00:43,920
Speaker 3: Alex Cho, the man who heads up HP's thirty five

14
00:00:44,200 --> 00:00:48,880
Speaker 3: billion dollar PC business, about our crappy relationship with work

15
00:00:49,120 --> 00:00:52,120
Speaker 3: and what HP is trying to do about it, funnily enough,

16
00:00:52,360 --> 00:00:53,440
Speaker 3: by leveraging AI.

17
00:00:53,720 --> 00:00:56,760
Speaker 1: All of that and more on episode seventy off The

18
00:00:56,800 --> 00:00:57,800
Speaker 1: Business of Tech.

19
00:00:58,360 --> 00:01:00,920
Speaker 3: Pete, you've spent a couple of weeks in the including

20
00:01:00,920 --> 00:01:04,039
Speaker 3: in Silicon Valley. What was the vibe like? I remember

21
00:01:04,080 --> 00:01:06,520
Speaker 3: I was there a couple of years ago and it

22
00:01:06,600 --> 00:01:10,000
Speaker 3: was pretty depressed and everyone was saying COVID's killed everything,

23
00:01:10,200 --> 00:01:12,240
Speaker 3: and it's just not what it once was. Have we

24
00:01:12,280 --> 00:01:13,720
Speaker 3: seen a bit of a rebound you reckon?

25
00:01:14,440 --> 00:01:17,840
Speaker 1: Well, definitely. San Francisco had been cleaned up, and this

26
00:01:18,080 --> 00:01:22,080
Speaker 1: is around the dream Force conference, that one of the

27
00:01:22,080 --> 00:01:25,800
Speaker 1: biggest tech conferences that's held there. That Salesforce, the big

28
00:01:26,080 --> 00:01:30,720
Speaker 1: customer relationship management software provider. But I think the vibe

29
00:01:30,760 --> 00:01:33,440
Speaker 1: between Silicon Valley where I spent about a week and

30
00:01:33,480 --> 00:01:37,400
Speaker 1: then San Francisco, I mean everyone is jumping on AI bandwagon,

31
00:01:37,440 --> 00:01:41,720
Speaker 1: from Salesforce to HP who I visited in Palo Alto.

32
00:01:42,640 --> 00:01:45,560
Speaker 1: They were all trying to monetize AI. When you drive

33
00:01:45,680 --> 00:01:50,200
Speaker 1: down the freeway, all the signs are about AI. And

34
00:01:50,240 --> 00:01:53,240
Speaker 1: I actually spent a few trips in one of these

35
00:01:53,440 --> 00:01:56,520
Speaker 1: driverless cars from Weimo that's all about AI.

36
00:01:56,920 --> 00:01:58,840
Speaker 3: I read your piece and the listener about that was

37
00:01:58,920 --> 00:01:59,400
Speaker 3: quite interesting.

38
00:01:59,440 --> 00:02:01,960
Speaker 1: It's incredible that one, because you know, we've heard so

39
00:02:02,080 --> 00:02:04,000
Speaker 1: much about it over the last decade. That was the

40
00:02:04,000 --> 00:02:07,240
Speaker 1: most impressive thing really I saw, because it's the culmination

41
00:02:07,520 --> 00:02:10,800
Speaker 1: of all of this intensive work to make it safe

42
00:02:10,840 --> 00:02:12,760
Speaker 1: and reliable and it's actually paid.

43
00:02:12,520 --> 00:02:15,880
Speaker 3: Off exciting times for a real signal of the future,

44
00:02:15,919 --> 00:02:19,040
Speaker 3: those self driving cars. But it's not just Salesforce that

45
00:02:19,080 --> 00:02:22,160
Speaker 3: you went over to San Francisco. You it also attended

46
00:02:22,400 --> 00:02:23,720
Speaker 3: the HP conference.

47
00:02:24,080 --> 00:02:29,280
Speaker 1: Yeah, the bymarket share, the biggest PC seller in New Zealand,

48
00:02:29,560 --> 00:02:33,160
Speaker 1: and you know, a company that is fundamentally on that

49
00:02:33,280 --> 00:02:35,560
Speaker 1: pathway to AI as well. We talked a lot about

50
00:02:35,600 --> 00:02:41,440
Speaker 1: the Microsoft Surface books. They're using those ARM processes to

51
00:02:41,520 --> 00:02:45,040
Speaker 1: run AI on the laptop, the so called neural processor.

52
00:02:45,280 --> 00:02:48,000
Speaker 1: So HP is doing exactly the same thing. They've just

53
00:02:48,080 --> 00:02:52,680
Speaker 1: launched a bunch of new devices, so they want everyone

54
00:02:52,720 --> 00:02:55,560
Speaker 1: to adopt AI. They've got their own AI assistant built

55
00:02:55,600 --> 00:02:58,560
Speaker 1: in to these, so you know, famously you can press

56
00:02:58,639 --> 00:03:02,400
Speaker 1: the co pilot but to get Microsoft's AI. But HP

57
00:03:02,800 --> 00:03:05,600
Speaker 1: has its own one on there, and interestingly, you can

58
00:03:05,639 --> 00:03:09,400
Speaker 1: load your own documents into that, into a library and

59
00:03:09,480 --> 00:03:11,720
Speaker 1: use the AI to sort of interrogate them and give

60
00:03:11,760 --> 00:03:15,400
Speaker 1: you information and insights about them. So we're starting to

61
00:03:15,480 --> 00:03:21,320
Speaker 1: see this, I think competition between who is going to

62
00:03:21,320 --> 00:03:23,840
Speaker 1: provide the AI experience that you're going to use on

63
00:03:23,880 --> 00:03:26,040
Speaker 1: a day to day basis. Is it going to be Microsoft,

64
00:03:26,120 --> 00:03:28,040
Speaker 1: the biggest software maker in the world. Is it going

65
00:03:28,120 --> 00:03:31,040
Speaker 1: to be an HP or a Dell doing their version

66
00:03:31,080 --> 00:03:33,359
Speaker 1: of it, or are you just going to go out

67
00:03:33,400 --> 00:03:35,000
Speaker 1: to the web increasingly to do this.

68
00:03:35,240 --> 00:03:39,520
Speaker 3: Google's notebook LM is fascinating. It's incredible and if you

69
00:03:39,520 --> 00:03:42,800
Speaker 3: haven't played with that, and getting the podcast Deep Dive

70
00:03:43,160 --> 00:03:45,520
Speaker 3: generated in there, which is kind of terrifying, but at

71
00:03:45,520 --> 00:03:47,720
Speaker 3: the same time, I don't know what they're doing, but

72
00:03:47,720 --> 00:03:49,760
Speaker 3: it's doing some clever stuff. I threw some payments in

73
00:03:49,760 --> 00:03:53,160
Speaker 3: New Zealand documents in their long long documents, said give

74
00:03:53,160 --> 00:03:57,560
Speaker 3: me a podcast about this, the future of the payment system,

75
00:03:57,600 --> 00:03:58,840
Speaker 3: and it was like, oh, yeah, they're talking about this,

76
00:03:58,880 --> 00:04:01,680
Speaker 3: they're talking about this, but what they're talking about is cryptocurrencies.

77
00:04:01,680 --> 00:04:04,119
Speaker 3: And it's not strange how there's not that much conversation

78
00:04:04,120 --> 00:04:06,440
Speaker 3: around cryptocurrencies in there, And I was like, nobody asked

79
00:04:06,440 --> 00:04:08,280
Speaker 3: you to say that. It's quite incredible.

80
00:04:09,360 --> 00:04:12,680
Speaker 1: Yeah, And we saw lots of sort of applications. They're

81
00:04:12,800 --> 00:04:14,880
Speaker 1: a little bit like with Microsoft, you know, that give

82
00:04:14,920 --> 00:04:17,080
Speaker 1: you a hint of what's to come, but it just

83
00:04:17,320 --> 00:04:21,080
Speaker 1: still sort of feels like maybe a year away before

84
00:04:21,279 --> 00:04:26,040
Speaker 1: it's actually worth buying one of these machines for those purposes, especially.

85
00:04:25,720 --> 00:04:27,480
Speaker 3: If the web based stuff is going to continue to

86
00:04:27,520 --> 00:04:28,200
Speaker 3: get so good.

87
00:04:28,520 --> 00:04:29,680
Speaker 1: Yeah, yeah, exactly.

88
00:04:29,760 --> 00:04:33,240
Speaker 3: And you spoke to one of HP's people over there

89
00:04:33,320 --> 00:04:35,479
Speaker 3: Alex Chow, who is Alex Chow.

90
00:04:36,480 --> 00:04:40,520
Speaker 1: He's one of the top executives at HP. He runs

91
00:04:40,560 --> 00:04:44,200
Speaker 1: the PC division, which is worth about US thirty five billion,

92
00:04:44,320 --> 00:04:48,480
Speaker 1: is huge business, and I talked to him about AI

93
00:04:48,520 --> 00:04:52,000
Speaker 1: obviously being central to it. But every year HP also

94
00:04:52,040 --> 00:04:55,720
Speaker 1: does this thing called the Work Relationship Index, and the

95
00:04:56,640 --> 00:05:00,960
Speaker 1: key takeaway from this year's index is that around the

96
00:05:01,000 --> 00:05:03,599
Speaker 1: world only twenty eight percent of people said they have

97
00:05:03,680 --> 00:05:07,839
Speaker 1: a healthy relationship with work. And it sort of taps

98
00:05:07,839 --> 00:05:12,520
Speaker 1: into that argument we're having again about return to office,

99
00:05:12,720 --> 00:05:16,320
Speaker 1: this mandate from Nikola Willis and other companies saying get

100
00:05:16,320 --> 00:05:19,680
Speaker 1: back into the office. Well, you know what is driving that?

101
00:05:19,920 --> 00:05:24,680
Speaker 1: Is that to prop up CBDs and restaurants and hospitality

102
00:05:24,839 --> 00:05:29,520
Speaker 1: or is it actually in to try and serve the

103
00:05:29,560 --> 00:05:32,960
Speaker 1: purpose of building a better culture. And this research suggests

104
00:05:32,960 --> 00:05:36,560
Speaker 1: that it doesn't matter where you work to build that culture,

105
00:05:36,600 --> 00:05:41,520
Speaker 1: it's things like trust and leadership, being provided the right

106
00:05:41,600 --> 00:05:45,560
Speaker 1: tools and the right training tailored to your specific needs,

107
00:05:45,640 --> 00:05:50,000
Speaker 1: not just one size fits all. So that's HP's angle

108
00:05:50,040 --> 00:05:52,520
Speaker 1: on that is, Okay, we need to build these devices

109
00:05:52,560 --> 00:05:56,159
Speaker 1: and AI services in a way that meet the specific

110
00:05:56,240 --> 00:05:58,560
Speaker 1: needs off a worker rather than just giving a fleet

111
00:05:58,560 --> 00:06:01,680
Speaker 1: of laptops to a opinion saying everyone gets the same thing.

112
00:06:02,440 --> 00:06:04,920
Speaker 3: Cool. So let's sell your interview with Alex cho where

113
00:06:04,960 --> 00:06:06,920
Speaker 3: he unpacks that report with you a little bit and

114
00:06:06,960 --> 00:06:09,000
Speaker 3: talks more broadly about HP's devices.

115
00:06:09,960 --> 00:06:12,799
Speaker 1: Alex, thanks so much for being on the business of tech.

116
00:06:13,080 --> 00:06:16,480
Speaker 1: Were in Palo Alto at the headquarters of Hewlett Packard,

117
00:06:16,720 --> 00:06:22,120
Speaker 1: literally stones throw from the original offices of Hewlett and Packard,

118
00:06:22,600 --> 00:06:26,200
Speaker 1: preserved in their sort of nineteen sixties glory. So pretty

119
00:06:26,440 --> 00:06:29,120
Speaker 1: cool to be here. And a lot of stuff's been

120
00:06:29,200 --> 00:06:33,560
Speaker 1: launched today at Hewlett Packard. Obviously AIPCS will get to those,

121
00:06:33,600 --> 00:06:36,160
Speaker 1: but one of the things you also launched was this

122
00:06:36,400 --> 00:06:39,880
Speaker 1: work Relationship Index that you publish every year. It's into

123
00:06:39,880 --> 00:06:43,560
Speaker 1: its second year now and quite some sort of quite

124
00:06:43,560 --> 00:06:47,479
Speaker 1: concerning data in there. This is nearly sixteen thousand people

125
00:06:47,520 --> 00:06:51,920
Speaker 1: surveyed around the world, and our relationship with work is

126
00:06:51,960 --> 00:06:54,000
Speaker 1: not very good. I think something like twenty eight percent

127
00:06:54,040 --> 00:06:57,240
Speaker 1: of those who were surveyed actually said I have a

128
00:06:57,279 --> 00:07:01,760
Speaker 1: healthy relationship with work. Took us through that and what

129
00:07:02,200 --> 00:07:05,760
Speaker 1: that means for you and how that informs what HP

130
00:07:05,880 --> 00:07:09,320
Speaker 1: does in terms of buillding devices for the modern workforce.

131
00:07:09,520 --> 00:07:12,440
Speaker 4: It's sobering data. Maybe let's call it for what it is.

132
00:07:12,480 --> 00:07:15,720
Speaker 4: Twenty eight percent on average of employees feel like they

133
00:07:15,760 --> 00:07:18,800
Speaker 4: have a healthy work relationship index, So we call it

134
00:07:18,840 --> 00:07:21,560
Speaker 4: for what it is. It's our starting point, and then

135
00:07:21,600 --> 00:07:25,000
Speaker 4: we dig underneath that more and understand what are the

136
00:07:25,120 --> 00:07:27,640
Speaker 4: areas that are driving that, and what are areas that

137
00:07:27,960 --> 00:07:31,440
Speaker 4: give us insights to how we can contribute. And there's

138
00:07:31,480 --> 00:07:33,360
Speaker 4: a couple of key ones that we have found. And

139
00:07:33,400 --> 00:07:36,000
Speaker 4: then number one is, you know, in the top three

140
00:07:36,320 --> 00:07:40,680
Speaker 4: areas that have the highest relationship to improving that. Number

141
00:07:40,720 --> 00:07:43,120
Speaker 4: one is giving them more skills, They give them more

142
00:07:43,120 --> 00:07:48,320
Speaker 4: confidence for their job. Number two is that as their

143
00:07:48,400 --> 00:07:52,440
Speaker 4: companies give them the type of technologies and tools for

144
00:07:52,520 --> 00:07:56,240
Speaker 4: them to be more successful while working in a hybrid environment,

145
00:07:57,080 --> 00:08:00,960
Speaker 4: they find is very impactful and heard. Actually, the other

146
00:08:01,040 --> 00:08:04,600
Speaker 4: thing that was really interesting for this year's results in

147
00:08:04,600 --> 00:08:08,600
Speaker 4: particular was the fact that for those who are using

148
00:08:08,720 --> 00:08:13,440
Speaker 4: AI on a regular basis, their relationship in next is

149
00:08:13,520 --> 00:08:16,560
Speaker 4: up to eleven points higher. So we take some of

150
00:08:16,560 --> 00:08:19,840
Speaker 4: that and there's a lot more insights there, and as

151
00:08:19,880 --> 00:08:23,920
Speaker 4: we share today, we think we're really uniquely positioned. We're

152
00:08:24,080 --> 00:08:29,040
Speaker 4: certainly passionate about transforming the future of work. The advent

153
00:08:29,120 --> 00:08:33,120
Speaker 4: of hybrid and AI as mega trends that we can

154
00:08:33,200 --> 00:08:37,280
Speaker 4: embrace in our products and solutions and service offerings. We

155
00:08:37,360 --> 00:08:39,840
Speaker 4: think gives us a wonderful opportunity.

156
00:08:39,960 --> 00:08:43,560
Speaker 1: Yeah. Look, this research arrives at a time In New Zealand.

157
00:08:43,559 --> 00:08:47,760
Speaker 1: Literally this week, our politicians have been literally ordering our

158
00:08:47,800 --> 00:08:50,439
Speaker 1: public servants back into the office, saying you've got to

159
00:08:50,440 --> 00:08:52,240
Speaker 1: be back in five days a week unless there's a

160
00:08:52,360 --> 00:08:55,320
Speaker 1: very good reason not. We saw Amazon issue a similar

161
00:08:55,360 --> 00:08:58,600
Speaker 1: sort of memo a couple of weeks ago. What I'm

162
00:08:58,640 --> 00:09:02,240
Speaker 1: seeing from urie search is it's maybe not the right

163
00:09:02,800 --> 00:09:04,880
Speaker 1: lens to look at it through. It's not really where

164
00:09:04,920 --> 00:09:08,680
Speaker 1: you're working, it's how you're working. They want personalization off

165
00:09:08,760 --> 00:09:11,680
Speaker 1: their workspace. The skills that they're offered needs to be

166
00:09:11,720 --> 00:09:16,360
Speaker 1: personalized to their needs. So really the whole remote versus

167
00:09:16,440 --> 00:09:19,400
Speaker 1: office thing isn't really a productive conversation.

168
00:09:19,600 --> 00:09:23,720
Speaker 4: Our best description and interpretation of what employees are seeing

169
00:09:23,760 --> 00:09:27,760
Speaker 4: is that they want flexibility. Now, flexibility is a broad term,

170
00:09:27,840 --> 00:09:32,000
Speaker 4: so how a company decides to enable and provide flexibility

171
00:09:32,040 --> 00:09:34,800
Speaker 4: I think companies are working through. But what they're seeing

172
00:09:34,840 --> 00:09:37,280
Speaker 4: is they would like the flexibility to have the option

173
00:09:37,400 --> 00:09:39,160
Speaker 4: of being able to work at home when they need

174
00:09:39,200 --> 00:09:41,960
Speaker 4: to do so, and the ability obviously when they're in

175
00:09:41,960 --> 00:09:44,240
Speaker 4: the office to be as productive as they can. In fact,

176
00:09:44,280 --> 00:09:47,160
Speaker 4: one of the interesting insights is that employees are feeling

177
00:09:47,240 --> 00:09:49,480
Speaker 4: like when they're at home they got used to being productive.

178
00:09:49,520 --> 00:09:51,480
Speaker 4: They go into the office and suddenly it takes them

179
00:09:51,480 --> 00:09:53,520
Speaker 4: a little bit of time to get into the zone

180
00:09:53,559 --> 00:09:56,280
Speaker 4: of work. So we have a rich opportunity as HP

181
00:09:56,600 --> 00:09:59,040
Speaker 4: to develop solutions for the office.

182
00:09:59,280 --> 00:10:02,120
Speaker 1: But also the interesting that the sort of the AI

183
00:10:02,240 --> 00:10:06,160
Speaker 1: findings here that people feel they're more productive when they

184
00:10:06,160 --> 00:10:09,640
Speaker 1: have access to AI. We do seem to have, particularly

185
00:10:09,640 --> 00:10:11,520
Speaker 1: I think in my part of the world, a little

186
00:10:11,520 --> 00:10:14,000
Speaker 1: bit of a digital divide when it comes to how

187
00:10:14,080 --> 00:10:17,600
Speaker 1: AI is being spread through organizations. It has expensive you know,

188
00:10:17,640 --> 00:10:20,320
Speaker 1: co pilots. You know in New Zealand's about thirty seven

189
00:10:20,360 --> 00:10:23,040
Speaker 1: dollars New Zealand a month. So what we're finding is

190
00:10:23,480 --> 00:10:26,760
Speaker 1: CEOs are saying, okay, marketing, you can use that. It's

191
00:10:26,800 --> 00:10:30,559
Speaker 1: not uniform across the organization. But the research suggests if

192
00:10:30,600 --> 00:10:33,400
Speaker 1: that carries on, we will see this lack of confidence

193
00:10:33,440 --> 00:10:36,600
Speaker 1: developing in the section of the workforce that doesn't have

194
00:10:36,640 --> 00:10:38,000
Speaker 1: access to these tools. Yeah.

195
00:10:38,080 --> 00:10:40,120
Speaker 4: So one of the things that we think is important

196
00:10:40,160 --> 00:10:43,200
Speaker 4: for us to develop in enable if you've heard some

197
00:10:43,280 --> 00:10:46,800
Speaker 4: of our languages, we want to give people access to

198
00:10:46,880 --> 00:10:51,199
Speaker 4: EI and that is one of the key motivators of

199
00:10:51,240 --> 00:10:55,560
Speaker 4: our aapcs. It's about giving people access to running AI

200
00:10:55,840 --> 00:10:59,960
Speaker 4: locally on his device. Now you have opportunities and there's

201
00:11:00,120 --> 00:11:03,000
Speaker 4: a lot of value in accessing AI in the cloud.

202
00:11:03,240 --> 00:11:07,000
Speaker 4: Not going to value, but there will be added benefit

203
00:11:07,120 --> 00:11:09,840
Speaker 4: or a set of benefits run running HEI locally in

204
00:11:09,880 --> 00:11:13,280
Speaker 4: your device. One of them is cost, and so that's

205
00:11:13,320 --> 00:11:15,480
Speaker 4: why we think there's a unique value for the AAPC.

206
00:11:15,720 --> 00:11:18,240
Speaker 4: By the way, the other is that for many things

207
00:11:18,280 --> 00:11:21,000
Speaker 4: it will be faster, right. And then third is that

208
00:11:21,920 --> 00:11:24,840
Speaker 4: for things that people care about in terms of their

209
00:11:24,920 --> 00:11:28,680
Speaker 4: private or local or corporate data, they keep that locally

210
00:11:28,760 --> 00:11:31,680
Speaker 4: versus sending into the cloud to run their AI.

211
00:11:31,880 --> 00:11:34,240
Speaker 1: Yeah, and it really has been the year of the AIPC.

212
00:11:34,400 --> 00:11:37,960
Speaker 1: We've seen Microsoft with it's Copilot plus PCs and the

213
00:11:37,960 --> 00:11:42,000
Speaker 1: rest of the industry with aipcs, and we've seen in

214
00:11:42,080 --> 00:11:46,120
Speaker 1: the demos here today the battery life improvement, the speed

215
00:11:46,120 --> 00:11:49,920
Speaker 1: improvement by having that neural processor on the device. HP

216
00:11:50,080 --> 00:11:52,920
Speaker 1: also has I think it's called an AI companion. This

217
00:11:53,040 --> 00:11:55,480
Speaker 1: is sort of similar to Copilot, where you can ask

218
00:11:55,520 --> 00:11:59,720
Speaker 1: it questions and it will access the information on your

219
00:11:59,720 --> 00:12:02,040
Speaker 1: device as well. We have sort of had a proliferation

220
00:12:02,320 --> 00:12:05,199
Speaker 1: of these services. Interesting in your views on how this

221
00:12:05,280 --> 00:12:07,040
Speaker 1: is going to shake out. For instance, some of the

222
00:12:07,080 --> 00:12:09,760
Speaker 1: devices that you're selling, you've got Copilot on a you've

223
00:12:09,760 --> 00:12:13,440
Speaker 1: also got AI Companion. Do these things work in tandem

224
00:12:13,520 --> 00:12:15,040
Speaker 1: or are you in sort of in competition?

225
00:12:15,880 --> 00:12:19,520
Speaker 4: We think is very complementary. So first thing is starting

226
00:12:19,520 --> 00:12:23,200
Speaker 4: with our last generation of device last it's recent this year,

227
00:12:23,840 --> 00:12:27,120
Speaker 4: is we've enabled Copilot and Copilot in the cloud on

228
00:12:27,120 --> 00:12:29,640
Speaker 4: our devices, meaning there's a key and you can access

229
00:12:29,640 --> 00:12:32,080
Speaker 4: in the cloud. So a lot of value for customers

230
00:12:32,120 --> 00:12:35,200
Speaker 4: and they're increasingly seeing the value of that. Secondly, what

231
00:12:35,760 --> 00:12:40,439
Speaker 4: is increasingly enabled through our Copilot plus PCs is enabling

232
00:12:40,600 --> 00:12:43,840
Speaker 4: customers to take advantage of the NPU and models on

233
00:12:43,880 --> 00:12:46,679
Speaker 4: the device that allow you to do things like what

234
00:12:46,720 --> 00:12:50,439
Speaker 4: we showed live translation, the ability to create using AI

235
00:12:50,559 --> 00:12:54,280
Speaker 4: locally on the device, So that is incremental value. And

236
00:12:54,320 --> 00:12:56,640
Speaker 4: then what we offer as well is we have our

237
00:12:56,679 --> 00:12:59,640
Speaker 4: AI Companion that does really two things. One is it

238
00:12:59,679 --> 00:13:03,240
Speaker 4: helps optimize the device real time based on how you

239
00:13:03,360 --> 00:13:05,439
Speaker 4: use it. But the other thing is what we showed

240
00:13:05,640 --> 00:13:10,120
Speaker 4: allows you to have your own library of files that

241
00:13:10,200 --> 00:13:12,280
Speaker 4: you can run AI models again, so that you have

242
00:13:12,440 --> 00:13:15,559
Speaker 4: that research assistant capability. But locally, we think it's very

243
00:13:15,559 --> 00:13:19,000
Speaker 4: complementary to the broader suite of tools that are out there.

244
00:13:19,080 --> 00:13:22,319
Speaker 1: Tell us about some of these devices that you launch today.

245
00:13:22,360 --> 00:13:26,079
Speaker 1: There's the Omnibook, the new Omnibook Yes, which is obviously

246
00:13:26,080 --> 00:13:29,880
Speaker 1: an AIPC very powerful Yes, Yes, the Omnibook Ultra flip,

247
00:13:30,600 --> 00:13:34,440
Speaker 1: and also on the printing side as well, AI print

248
00:13:34,559 --> 00:13:36,480
Speaker 1: features as well. Tell us about those. Yeah.

249
00:13:36,800 --> 00:13:39,720
Speaker 4: One of the things that we found is that while

250
00:13:39,880 --> 00:13:42,720
Speaker 4: we very much are continued to innovate on the printer,

251
00:13:43,320 --> 00:13:47,520
Speaker 4: it is the output that people have challenged at times,

252
00:13:47,800 --> 00:13:51,400
Speaker 4: and we've had the opportunity to innovate and offer AI

253
00:13:51,520 --> 00:13:54,920
Speaker 4: utilities so that you get the right output that you're

254
00:13:54,960 --> 00:13:57,320
Speaker 4: looking for. And we know what a good print job

255
00:13:57,360 --> 00:13:59,160
Speaker 4: looks like, we know what that should look like, and

256
00:13:59,200 --> 00:14:01,840
Speaker 4: so we're able to pass that into our EI tool

257
00:14:01,880 --> 00:14:04,480
Speaker 4: so that the output that you get is what you

258
00:14:04,520 --> 00:14:05,480
Speaker 4: had intended to get.

259
00:14:05,600 --> 00:14:08,600
Speaker 1: Yeah, Instead of printing out an EXCEL spreadsheet and having

260
00:14:08,600 --> 00:14:09,400
Speaker 1: fifteen pages.

261
00:14:09,480 --> 00:14:12,959
Speaker 4: Yes, that's happened to me several times. Yeah, so I'm

262
00:14:13,120 --> 00:14:13,920
Speaker 4: a huge fan.

263
00:14:14,040 --> 00:14:16,880
Speaker 1: So I think a lot of people so tangible and practical. Yeah.

264
00:14:17,640 --> 00:14:20,640
Speaker 1: So in terms of when you think we're going to

265
00:14:20,640 --> 00:14:24,400
Speaker 1: see a critical mass of applications that the average maybe

266
00:14:24,440 --> 00:14:27,800
Speaker 1: consumer will buy one of these machines and see a

267
00:14:27,880 --> 00:14:31,120
Speaker 1: dramatic different difference. Because at the moment, most consumers are

268
00:14:31,200 --> 00:14:34,800
Speaker 1: using AI and the cloud, chat, GPT or copilot web

269
00:14:34,880 --> 00:14:37,320
Speaker 1: version or something like that, and probably the performance is okay.

270
00:14:37,600 --> 00:14:40,040
Speaker 1: But when are we going to see that shift when

271
00:14:40,560 --> 00:14:44,920
Speaker 1: consumers and people in the enterprise environments are going, I've

272
00:14:44,920 --> 00:14:47,000
Speaker 1: got to have one of these machines because the performance

273
00:14:47,040 --> 00:14:47,720
Speaker 1: is so much better.

274
00:14:47,920 --> 00:14:50,520
Speaker 4: Almost all of the ivs we talk to are in

275
00:14:50,560 --> 00:14:54,360
Speaker 4: the process of embedding and capitalizing on AI models and

276
00:14:54,400 --> 00:14:58,640
Speaker 4: their applications. They're not yet available. We expect that to

277
00:14:58,720 --> 00:15:01,680
Speaker 4: continue to increase throughout the year, and as that happens,

278
00:15:01,680 --> 00:15:04,120
Speaker 4: we think there will only be more appetite around them.

279
00:15:04,520 --> 00:15:07,440
Speaker 4: That being the case, the time is now. Time is

280
00:15:07,440 --> 00:15:10,840
Speaker 4: now because a if you're in a corporate enterprise space,

281
00:15:11,280 --> 00:15:13,600
Speaker 4: you have many devices you bought during COVID and with

282
00:15:13,720 --> 00:15:17,680
Speaker 4: Windows eleven. You have a great platform to migrate to.

283
00:15:18,240 --> 00:15:21,800
Speaker 4: Time is now because we now have the capability on

284
00:15:21,840 --> 00:15:25,360
Speaker 4: these devices to run these ISV applications. As they become

285
00:15:25,400 --> 00:15:28,160
Speaker 4: available in so many ways, you'll be future proved. Time

286
00:15:28,240 --> 00:15:31,080
Speaker 4: is now because even more than AI, what do most

287
00:15:31,080 --> 00:15:35,080
Speaker 4: employees do. They're on teams called zoom calls, and the

288
00:15:35,120 --> 00:15:38,400
Speaker 4: device that they have wasn't really designed or optimized for them.

289
00:15:38,840 --> 00:15:42,359
Speaker 4: We have the integration of our latest audio and video technology,

290
00:15:42,800 --> 00:15:45,520
Speaker 4: a lot that we've actually gleaned from our poly acquisition

291
00:15:46,080 --> 00:15:49,200
Speaker 4: embedded in the device to give you the best in

292
00:15:49,320 --> 00:15:53,080
Speaker 4: audio and video, which employees are capitalizing on every day.

293
00:15:53,080 --> 00:15:55,080
Speaker 4: So we actually think the time is now and as

294
00:15:55,120 --> 00:15:57,960
Speaker 4: applications come, they're going to actually get a richer set

295
00:15:57,960 --> 00:15:59,600
Speaker 4: of things that they can do on the device. The

296
00:15:59,680 --> 00:16:03,440
Speaker 4: value of the piecs it's so much more now because

297
00:16:03,480 --> 00:16:06,440
Speaker 4: again it's not just faster, but you're starting to do

298
00:16:06,560 --> 00:16:09,600
Speaker 4: new things. That's why we think this is not just

299
00:16:09,720 --> 00:16:12,800
Speaker 4: another year. This is an inflection point when you can

300
00:16:12,960 --> 00:16:17,360
Speaker 4: do new things that you couldn't do before. You couldn't

301
00:16:17,360 --> 00:16:20,440
Speaker 4: do it an order on an existing device, and these

302
00:16:20,480 --> 00:16:24,120
Speaker 4: new things allow you to be more productive, to create

303
00:16:24,200 --> 00:16:27,720
Speaker 4: better to connect better show up better. Then we're really

304
00:16:27,800 --> 00:16:30,080
Speaker 4: changing the day in the life of the worker. That's

305
00:16:30,080 --> 00:16:31,520
Speaker 4: how it ties to our first topic.

306
00:16:32,600 --> 00:16:33,880
Speaker 5: In a world where.

307
00:16:33,640 --> 00:16:38,680
Speaker 4: The work relationship index is pretty low and part of

308
00:16:38,720 --> 00:16:42,720
Speaker 4: that primary drivers for increasing that is confidence and having

309
00:16:42,760 --> 00:16:45,640
Speaker 4: skills and doing the job, and AI is one of

310
00:16:45,640 --> 00:16:48,760
Speaker 4: the key skills that the more they use they improve

311
00:16:48,800 --> 00:16:52,280
Speaker 4: their index. We think bringing device solutions allow people to

312
00:16:52,320 --> 00:16:54,600
Speaker 4: access AI in the most powerful ways.

313
00:16:54,560 --> 00:16:56,960
Speaker 1: Will help them. Yeah, you've got something like seven of

314
00:16:57,040 --> 00:17:01,560
Speaker 1: your already thousand employees around the world world. What have

315
00:17:01,680 --> 00:17:04,639
Speaker 1: you learned from your approach to hybrid work, which I

316
00:17:04,640 --> 00:17:08,800
Speaker 1: think is pretty flexible by industry standards that you apply

317
00:17:09,040 --> 00:17:12,160
Speaker 1: to developing these devices that you're selling all over the world.

318
00:17:12,359 --> 00:17:16,520
Speaker 4: Yeah, many First is it makes a really difference how

319
00:17:16,560 --> 00:17:19,840
Speaker 4: you show up and how well you're heard, which is interesting.

320
00:17:19,880 --> 00:17:23,119
Speaker 4: We've thought that the way you show up and camera

321
00:17:23,200 --> 00:17:26,439
Speaker 4: technology will be the leading part of that experience and

322
00:17:26,480 --> 00:17:30,680
Speaker 4: it's very important. But in your next meeting, see how

323
00:17:30,760 --> 00:17:33,440
Speaker 4: much and how well you pay attention to someone it seems.

324
00:17:33,200 --> 00:17:34,800
Speaker 1: Kind of far away, I can't really hear.

325
00:17:34,720 --> 00:17:38,119
Speaker 4: Them, versus someone who's right near and present. You find

326
00:17:38,119 --> 00:17:41,119
Speaker 4: you pay attention more so, what we're finding is the

327
00:17:41,240 --> 00:17:44,280
Speaker 4: quality of the audio and video makes a huge difference.

328
00:17:44,760 --> 00:17:47,000
Speaker 4: That's why we're investing that space and that's why we're

329
00:17:47,000 --> 00:17:50,359
Speaker 4: making even better with AI. The other thing we've learned

330
00:17:50,400 --> 00:17:55,520
Speaker 4: a lot is that people have become accustomed to the

331
00:17:55,560 --> 00:17:59,000
Speaker 4: productivity that you have when you have a large display,

332
00:17:59,480 --> 00:18:02,640
Speaker 4: multiple displays, or wide display that you can have both

333
00:18:02,680 --> 00:18:05,960
Speaker 4: on because they like the ability of that productivity which

334
00:18:06,000 --> 00:18:10,000
Speaker 4: they have learned so to speak, in COVID, and they

335
00:18:10,040 --> 00:18:11,080
Speaker 4: really want to continue.

336
00:18:11,520 --> 00:18:13,880
Speaker 3: The third thing is what we found.

337
00:18:13,600 --> 00:18:17,720
Speaker 4: From our hybrid work environment is employees want to have

338
00:18:17,760 --> 00:18:22,639
Speaker 4: the flexibility to be working multiple locations. We want to

339
00:18:22,680 --> 00:18:25,400
Speaker 4: make sure though that they are productive and that they

340
00:18:25,400 --> 00:18:29,920
Speaker 4: can get from arrival to productive as fast as they can.

341
00:18:30,280 --> 00:18:33,720
Speaker 4: And enabling that through the combination of our PCs plus

342
00:18:33,720 --> 00:18:38,480
Speaker 4: our docking solutions, plus our displays, plus having peripherals that

343
00:18:38,600 --> 00:18:42,000
Speaker 4: quickly connect. They seem like minor, but those friction points

344
00:18:42,040 --> 00:18:43,840
Speaker 4: make a big deal, so that you feel like wherever

345
00:18:43,920 --> 00:18:47,560
Speaker 4: you go from arrival to work can be fast, so

346
00:18:48,080 --> 00:18:52,440
Speaker 4: quick productive. Make sure that you show up, you can

347
00:18:52,480 --> 00:18:55,040
Speaker 4: be seen, you can see others, you can hear, and

348
00:18:55,080 --> 00:18:58,320
Speaker 4: then as well that you're productive with a full rig

349
00:18:59,119 --> 00:19:01,640
Speaker 4: a whole product any set up some of the learnings

350
00:19:01,640 --> 00:19:03,040
Speaker 4: that we found for our own employees.

351
00:19:03,200 --> 00:19:06,359
Speaker 1: Very good, well, good luck with the latest range that

352
00:19:06,400 --> 00:19:09,719
Speaker 1: you've launched today, and thanks for coming on the business

353
00:19:09,720 --> 00:19:10,080
Speaker 1: of Tech.

354
00:19:10,359 --> 00:19:10,879
Speaker 5: You're welcome.

355
00:19:17,920 --> 00:19:20,400
Speaker 1: So you know, Alex chose job in the next year

356
00:19:20,400 --> 00:19:24,280
Speaker 1: really is to make AIPCS a big success. The PC

357
00:19:24,480 --> 00:19:28,000
Speaker 1: industry has sort of slumped a bit following the COVID boom,

358
00:19:28,640 --> 00:19:31,920
Speaker 1: so PC makers, like HP and others, they really need

359
00:19:31,960 --> 00:19:35,480
Speaker 1: to ensure that people and particularly business leaders see AI

360
00:19:35,560 --> 00:19:37,879
Speaker 1: features is driving a refresh of their devices.

361
00:19:38,200 --> 00:19:41,560
Speaker 3: But that requires trust and confidence that AI can not

362
00:19:41,680 --> 00:19:46,120
Speaker 3: only deliver accurate results without leaking company data, but also

363
00:19:46,200 --> 00:19:51,359
Speaker 3: that the productivity boosts using these AI devices will actually materialize.

364
00:19:51,920 --> 00:19:56,600
Speaker 1: Jessica Zosa Ford is an up and coming data scientist

365
00:19:56,760 --> 00:20:00,680
Speaker 1: machine learning expert thinking deeply about all of this. She's

366
00:20:00,720 --> 00:20:04,280
Speaker 1: a machine learning researcher focused on the empirical study of

367
00:20:04,320 --> 00:20:08,439
Speaker 1: deep learning models to improve their reliability in high stakes

368
00:20:08,480 --> 00:20:11,439
Speaker 1: domains such as healthcare. She was out here as a

369
00:20:11,440 --> 00:20:14,679
Speaker 1: guest of the US Embassy a few weeks ago, keynoting

370
00:20:15,080 --> 00:20:19,399
Speaker 1: at the AI summit where the guv GPT chatbot was launched.

371
00:20:19,080 --> 00:20:21,800
Speaker 3: She's got a very interesting background, involved in a lot

372
00:20:21,840 --> 00:20:26,160
Speaker 3: of open source products such as Project Jupiter, and she's

373
00:20:26,200 --> 00:20:28,560
Speaker 3: worked with Facebook on its deep learning efforts.

374
00:20:28,800 --> 00:20:32,280
Speaker 1: She also helped develop Bloom one hundred and seventy six

375
00:20:32,400 --> 00:20:38,480
Speaker 1: billion parameter open access multiling language model. In fact, language

376
00:20:38,520 --> 00:20:41,960
Speaker 1: is really what Jessica is passionate about and trying to

377
00:20:41,960 --> 00:20:44,720
Speaker 1: make sure that languages other than English get a decent

378
00:20:44,800 --> 00:20:46,359
Speaker 1: treatment in the world of AI.

379
00:20:46,600 --> 00:20:49,320
Speaker 3: So, Jessica was here in New Zealand talking to businesses,

380
00:20:49,440 --> 00:20:53,080
Speaker 3: researchers and the government's cross party group on AI and

381
00:20:53,320 --> 00:20:55,639
Speaker 3: Peter you caught up with her in Wellington at data

382
00:20:55,640 --> 00:20:59,040
Speaker 3: COM's headquarters in the Capitol, and here's your interview with

383
00:20:59,160 --> 00:20:59,920
Speaker 3: Jessica Zozo.

384
00:21:00,200 --> 00:21:00,480
Speaker 4: Forward.

385
00:21:04,320 --> 00:21:07,760
Speaker 1: Jessica, thanks so much for being on the business of tech.

386
00:21:08,520 --> 00:21:11,480
Speaker 1: You're a machine learning expert working on some sort of

387
00:21:11,520 --> 00:21:17,040
Speaker 1: cutting edge areas of machine learning and artificial intelligence. You're

388
00:21:17,040 --> 00:21:21,000
Speaker 1: doing your PhD at Brown University, and you've worked with

389
00:21:21,040 --> 00:21:25,960
Speaker 1: the likes of Facebook on open source large language models

390
00:21:26,520 --> 00:21:29,840
Speaker 1: other open source projects. What's brought you to New Zealand

391
00:21:29,880 --> 00:21:30,680
Speaker 1: at this point in time?

392
00:21:30,800 --> 00:21:33,320
Speaker 2: So I was fortunate enough to be invited by the

393
00:21:33,560 --> 00:21:39,159
Speaker 2: US Embassy in Wellington to come and meet a technologists

394
00:21:39,240 --> 00:21:43,600
Speaker 2: and business people and educators and people all over New

395
00:21:43,720 --> 00:21:46,680
Speaker 2: Zealand to kind of just share the kinds of things

396
00:21:46,760 --> 00:21:48,680
Speaker 2: I've been thinking about as a researcher and kind of

397
00:21:48,760 --> 00:21:51,040
Speaker 2: learn more about the things that are happening in New Zealand.

398
00:21:51,080 --> 00:21:53,040
Speaker 2: And so far things have been really, really great. I'm

399
00:21:53,080 --> 00:21:54,480
Speaker 2: really having an amazing time.

400
00:21:54,640 --> 00:21:57,000
Speaker 1: I mean, what an incredible time to be doing your

401
00:21:57,359 --> 00:22:02,199
Speaker 1: PhD in machine learning. It's almost perfectly timed for what

402
00:22:02,359 --> 00:22:04,320
Speaker 1: is going on in the world with this technology, which

403
00:22:04,359 --> 00:22:06,199
Speaker 1: has been around for a long time, but with the

404
00:22:06,280 --> 00:22:11,000
Speaker 1: advents of large language models generative AI, it's really poised

405
00:22:11,000 --> 00:22:11,960
Speaker 1: just to explode, isn't it.

406
00:22:12,240 --> 00:22:15,159
Speaker 2: Yeah, I know, it's been a really fun time. You know,

407
00:22:15,280 --> 00:22:20,399
Speaker 2: like before starting my PhD, chat GBT was not available

408
00:22:20,440 --> 00:22:23,360
Speaker 2: to the public, and now, you know, it's much easier

409
00:22:23,359 --> 00:22:26,000
Speaker 2: to have these conversations about what do I work on,

410
00:22:26,280 --> 00:22:27,320
Speaker 2: what do I care about?

411
00:22:27,480 --> 00:22:29,639
Speaker 5: And you know what do I hope people.

412
00:22:29,400 --> 00:22:31,720
Speaker 2: Are aware about this technology because now we have real

413
00:22:31,760 --> 00:22:34,520
Speaker 2: world these cases that people can really interact with and understand.

414
00:22:34,560 --> 00:22:37,560
Speaker 1: One thing I see that you are really interested in

415
00:22:37,640 --> 00:22:42,720
Speaker 1: care about is language and how actual languages are treated

416
00:22:42,760 --> 00:22:46,359
Speaker 1: by large language models. And the chatbots that sort of

417
00:22:46,400 --> 00:22:49,119
Speaker 1: are the front face off them. And just earlier, I

418
00:22:49,160 --> 00:22:50,879
Speaker 1: was just trying because I know you've done a paper

419
00:22:50,880 --> 00:22:53,720
Speaker 1: you're interested in Southeast Asian languages in particular and how

420
00:22:53,720 --> 00:22:56,920
Speaker 1: they're treated by these large language models, and I was

421
00:22:57,000 --> 00:22:59,840
Speaker 1: just sort of translating a bit of English into Vietnamese

422
00:23:00,080 --> 00:23:05,320
Speaker 1: to Malaysian, comparing Perplexity to Gemini to chet GPT, And

423
00:23:05,520 --> 00:23:07,480
Speaker 1: I think it's fair say they did a reasonable job

424
00:23:07,720 --> 00:23:11,280
Speaker 1: of translating at least from English to one language and back.

425
00:23:11,359 --> 00:23:15,639
Speaker 1: But they all had slightly different answers to the same

426
00:23:16,840 --> 00:23:19,720
Speaker 1: paragraph of texts that I gave them, so they're all

427
00:23:19,920 --> 00:23:23,600
Speaker 1: obviously working from slightly different models. But why is it

428
00:23:23,680 --> 00:23:29,359
Speaker 1: so challenging to deal with multi languages, multi lingual setting

429
00:23:29,920 --> 00:23:31,320
Speaker 1: by these large language models.

430
00:23:31,359 --> 00:23:34,240
Speaker 2: I mean, I think fundamentally a lot of the challenges

431
00:23:34,320 --> 00:23:35,720
Speaker 2: have to do with the fact that we have taken

432
00:23:35,800 --> 00:23:39,080
Speaker 2: a English first approach to modeling, in that since English

433
00:23:39,200 --> 00:23:41,520
Speaker 2: is the number one language are presented on the Internet,

434
00:23:41,560 --> 00:23:44,160
Speaker 2: it's very easy for us to think about modeling from

435
00:23:44,240 --> 00:23:47,639
Speaker 2: an English language context because we have the data. So

436
00:23:48,040 --> 00:23:51,040
Speaker 2: you know, many of these models are built with the

437
00:23:51,040 --> 00:23:52,760
Speaker 2: assumption that we will try to give it as much

438
00:23:52,840 --> 00:23:56,879
Speaker 2: data as physically possible given the kinds of systems that

439
00:23:56,920 --> 00:23:59,359
Speaker 2: we have, and we crime it all in in the model,

440
00:23:59,400 --> 00:24:03,080
Speaker 2: and as a result, we are now optimized towards questions

441
00:24:03,119 --> 00:24:07,159
Speaker 2: around scaling and scalability. Now granted, you know, like not

442
00:24:07,320 --> 00:24:10,560
Speaker 2: all languages have the same amount of data available.

443
00:24:10,160 --> 00:24:10,800
Speaker 5: To the public.

444
00:24:10,880 --> 00:24:13,200
Speaker 2: You know, languages that are closer in terms of amount

445
00:24:13,200 --> 00:24:16,800
Speaker 2: of scale, say Arabic or Chinese, or say Spanish. You know,

446
00:24:16,840 --> 00:24:19,520
Speaker 2: the kinds of performance we say is pretty good. You know,

447
00:24:19,760 --> 00:24:23,440
Speaker 2: there was a recent Apple announcement in terms of their

448
00:24:23,520 --> 00:24:26,800
Speaker 2: uses of generative AI, and you see the languages they're

449
00:24:26,800 --> 00:24:30,040
Speaker 2: really saying they are larger languages such as Spanish. So yeah,

450
00:24:30,040 --> 00:24:33,439
Speaker 2: I mean, like that's definitely something we see in terms

451
00:24:33,440 --> 00:24:36,280
Speaker 2: of the language quality. But yes, we go down in

452
00:24:36,400 --> 00:24:38,560
Speaker 2: terms of the size of number of speakers of the

453
00:24:38,840 --> 00:24:41,920
Speaker 2: or the availability of online data, then we have questions

454
00:24:41,920 --> 00:24:44,920
Speaker 2: more and more about the quality of the outputs. So yeah,

455
00:24:44,960 --> 00:24:47,800
Speaker 2: I mean, given it's really just kind of a downstream

456
00:24:47,840 --> 00:24:51,000
Speaker 2: consequence of the kinds of technical choices we've made in

457
00:24:51,119 --> 00:24:52,960
Speaker 2: terms of how we formulated this problem.

458
00:24:53,320 --> 00:24:55,240
Speaker 1: Yeah, and when you come to New Zealand, we have

459
00:24:55,680 --> 00:24:58,119
Speaker 1: our own Night of Language here Maori, and there's a

460
00:24:58,160 --> 00:25:00,399
Speaker 1: revival of that, a lot of people very passionate about

461
00:25:00,520 --> 00:25:03,960
Speaker 1: learning it. You're not going to get great translation out

462
00:25:03,960 --> 00:25:06,040
Speaker 1: of I guess out of a chat GPT. What's the

463
00:25:06,080 --> 00:25:08,320
Speaker 1: answer to that? Do we have to work with them

464
00:25:08,440 --> 00:25:09,680
Speaker 1: or build our own models?

465
00:25:09,800 --> 00:25:11,639
Speaker 2: Yeah, I mean, I think there's a lot of excitement

466
00:25:11,760 --> 00:25:14,520
Speaker 2: with regards to the kind of work that's being done

467
00:25:14,880 --> 00:25:20,399
Speaker 2: in New Zealand about the ethical and technical questions in

468
00:25:20,480 --> 00:25:23,440
Speaker 2: modeling today or memory. So like I think it's I

469
00:25:23,520 --> 00:25:26,080
Speaker 2: think that there are in terms of the kinds of

470
00:25:26,160 --> 00:25:29,880
Speaker 2: answers that will be generated, in terms of the ethics

471
00:25:29,880 --> 00:25:32,399
Speaker 2: and the technology. I think it's best to kind of

472
00:25:32,400 --> 00:25:35,080
Speaker 2: see what's going to be happening in the future from

473
00:25:35,400 --> 00:25:40,199
Speaker 2: you know, language experts, in community leaders and technologists within

474
00:25:40,600 --> 00:25:43,720
Speaker 2: the country. And honestly that's part of what has driven

475
00:25:43,800 --> 00:25:46,840
Speaker 2: me to come here because a lot of the assumptions

476
00:25:46,880 --> 00:25:50,959
Speaker 2: we've historically meet about lower resource languages is that you know,

477
00:25:51,320 --> 00:25:55,639
Speaker 2: the people who speak those languages don't necessarily may not

478
00:25:55,640 --> 00:25:59,840
Speaker 2: necessarily have the kinds of access to technology that people

479
00:25:59,880 --> 00:26:01,399
Speaker 2: are actually do have in New Zealand.

480
00:26:01,480 --> 00:26:03,160
Speaker 5: So you know, people are familiar.

481
00:26:02,680 --> 00:26:07,000
Speaker 2: With things about Siri and Alexa and chat GPT, so

482
00:26:07,080 --> 00:26:09,600
Speaker 2: they do know about this, But then their question is, Okay,

483
00:26:09,640 --> 00:26:11,600
Speaker 2: so now that we have this familiarity, what do we

484
00:26:11,600 --> 00:26:13,920
Speaker 2: want to do with our own technologies. So as a result,

485
00:26:13,960 --> 00:26:15,800
Speaker 2: a lot of the kinds of assumptions we've made as

486
00:26:16,119 --> 00:26:21,040
Speaker 2: international AI researchers working on other languages are kind of

487
00:26:21,119 --> 00:26:24,720
Speaker 2: violated in when we think about the today O Maori context.

488
00:26:24,760 --> 00:26:26,840
Speaker 2: And so that's kind of what gets me really excited

489
00:26:27,040 --> 00:26:29,199
Speaker 2: is you know, we get to approach these challenges with

490
00:26:29,240 --> 00:26:29,600
Speaker 2: a new.

491
00:26:29,520 --> 00:26:30,119
Speaker 5: Set of eyes.

492
00:26:30,200 --> 00:26:32,600
Speaker 1: Yeah. I mean, we've got a supercomputer sitting in this

493
00:26:32,760 --> 00:26:36,000
Speaker 1: in the city. We've got incredible work going on by

494
00:26:36,119 --> 00:26:40,639
Speaker 1: MARI researchers around MARI data sovereignty, and we even have

495
00:26:40,680 --> 00:26:43,080
Speaker 1: the likes of Microsoft in that have done a lot

496
00:26:43,080 --> 00:26:47,840
Speaker 1: of work on getting TODAYO right in their products. So

497
00:26:47,880 --> 00:26:49,639
Speaker 1: we've got a lot going for us to sort of

498
00:26:49,640 --> 00:26:50,520
Speaker 1: address this issue.

499
00:26:50,600 --> 00:26:52,320
Speaker 2: Yeah, I know, I think I think that there I

500
00:26:52,400 --> 00:26:55,760
Speaker 2: think there's a lot of really exciting implications with regards

501
00:26:55,760 --> 00:26:58,880
Speaker 2: to how do we deal with the ethical questions around

502
00:26:59,000 --> 00:27:04,680
Speaker 2: the use of data and the curation of data and

503
00:27:05,040 --> 00:27:09,680
Speaker 2: the protection of people's knowledge and values.

504
00:27:09,880 --> 00:27:10,960
Speaker 5: I think it's really exciting.

505
00:27:11,000 --> 00:27:13,760
Speaker 1: You've arrived in New Zealand at quite an interesting time

506
00:27:13,800 --> 00:27:17,560
Speaker 1: in our journey on artificial intelligence. You attended the AI

507
00:27:17,680 --> 00:27:23,960
Speaker 1: Summit where the Minister of Digitizing Government, Judith Collins announced

508
00:27:24,040 --> 00:27:27,840
Speaker 1: gov gpt, which is sort of an entry level chatbot

509
00:27:27,960 --> 00:27:32,840
Speaker 1: that they'll be piloting for businesses and their interaction with government.

510
00:27:33,320 --> 00:27:36,280
Speaker 1: So that's a positive thing. What's your impression really off

511
00:27:36,400 --> 00:27:39,240
Speaker 1: the New Zealand landscape. You've been through Australia recently, you

512
00:27:39,320 --> 00:27:41,680
Speaker 1: hear in New Zealand. What's your sense when you talk

513
00:27:41,720 --> 00:27:44,879
Speaker 1: to people in government and business about how mature our

514
00:27:44,880 --> 00:27:45,640
Speaker 1: approach is here?

515
00:27:45,800 --> 00:27:48,280
Speaker 2: I mean, I think that people are very aware of

516
00:27:48,320 --> 00:27:52,199
Speaker 2: what's happening, both happening here locally and abroad. So to me,

517
00:27:52,600 --> 00:27:57,160
Speaker 2: I've not necessarily had any sort of challenges in terms

518
00:27:57,200 --> 00:27:59,480
Speaker 2: of making sure that we're on the same page about

519
00:27:59,560 --> 00:28:02,639
Speaker 2: understand and the technology you're seeing what's on the horizon.

520
00:28:02,800 --> 00:28:06,680
Speaker 2: So to me, it's been really easy kind of having

521
00:28:06,760 --> 00:28:10,240
Speaker 2: these technical, deep technical conversations and honestly, I've just been

522
00:28:10,280 --> 00:28:13,439
Speaker 2: really excited about the kinds of insights people have been

523
00:28:13,440 --> 00:28:14,080
Speaker 2: sharing with me.

524
00:28:14,200 --> 00:28:14,960
Speaker 5: So it's been great.

525
00:28:15,119 --> 00:28:17,840
Speaker 1: Yeah, we don't have an executive order or something that

526
00:28:18,160 --> 00:28:20,840
Speaker 1: the Biden administration put out on AI. We don't have

527
00:28:20,880 --> 00:28:23,720
Speaker 1: an AI Act, we don't really have legislation tailored to

528
00:28:23,800 --> 00:28:28,040
Speaker 1: this yet. Judith Collins, the Minister, has said she wants

529
00:28:28,119 --> 00:28:33,560
Speaker 1: to take a light touch, risk based and proportional approach

530
00:28:33,800 --> 00:28:38,800
Speaker 1: to this, using existing legislation where necessary. So that's fine,

531
00:28:38,800 --> 00:28:40,880
Speaker 1: but I guess we are now getting into the nuts

532
00:28:40,880 --> 00:28:43,520
Speaker 1: and bolts. If you are implementing this, either as a

533
00:28:43,560 --> 00:28:46,120
Speaker 1: government department or as a business in New Zealand, you

534
00:28:46,200 --> 00:28:49,120
Speaker 1: really have to have those guide rails in place. So

535
00:28:49,160 --> 00:28:53,560
Speaker 1: this gets us into the ethics of AI and interested

536
00:28:53,560 --> 00:28:55,880
Speaker 1: in your reflections on what you're seeing perhaps in other

537
00:28:55,920 --> 00:28:58,920
Speaker 1: markets like the US, about how large businesses in the

538
00:28:59,040 --> 00:29:02,520
Speaker 1: US and is sort of grappling with this exactly how

539
00:29:02,880 --> 00:29:04,080
Speaker 1: that is going to work in practice.

540
00:29:04,160 --> 00:29:06,880
Speaker 2: Yeah, I mean I think it's it's a a broad

541
00:29:06,960 --> 00:29:09,920
Speaker 2: based challenge. I don't think anyone has really gotten it right,

542
00:29:10,000 --> 00:29:12,040
Speaker 2: to be honest, I think we're all kind of learning

543
00:29:12,120 --> 00:29:14,800
Speaker 2: at the same time. So to me, you know, I

544
00:29:14,840 --> 00:29:17,560
Speaker 2: don't have an immediate use case and being like, oh, yeah,

545
00:29:17,560 --> 00:29:18,200
Speaker 2: we should.

546
00:29:17,960 --> 00:29:19,400
Speaker 5: All look to this organization.

547
00:29:19,480 --> 00:29:22,080
Speaker 2: They've done a really great job, because you know, not

548
00:29:22,360 --> 00:29:26,800
Speaker 2: only is the people trying to implement these systems and

549
00:29:26,800 --> 00:29:30,400
Speaker 2: deploy them, you know, dealing with the challenges and within

550
00:29:30,440 --> 00:29:33,520
Speaker 2: their organization. But the technology is just changing so quickly

551
00:29:33,960 --> 00:29:36,080
Speaker 2: that even if I were to say, oh, like this

552
00:29:36,240 --> 00:29:39,760
Speaker 2: organization has done it great, you know, some company might

553
00:29:39,800 --> 00:29:41,959
Speaker 2: come out and say, oh, we've released the new LM,

554
00:29:42,360 --> 00:29:43,800
Speaker 2: and then we're all going to have to you know,

555
00:29:43,880 --> 00:29:45,480
Speaker 2: reckon with oh do we want to switch over to

556
00:29:45,480 --> 00:29:47,360
Speaker 2: that model? And how are we going to build our

557
00:29:47,400 --> 00:29:50,719
Speaker 2: new systems around that system. So to me, it's you know,

558
00:29:50,760 --> 00:29:55,640
Speaker 2: it's honestly a very dynamic set of questions, and I

559
00:29:55,680 --> 00:29:58,360
Speaker 2: don't think we all we all have the answers.

560
00:29:58,400 --> 00:29:59,000
Speaker 5: That's sad.

561
00:29:59,080 --> 00:30:02,440
Speaker 2: I think the thing to keep in mind as a

562
00:30:02,960 --> 00:30:06,960
Speaker 2: business person or someone who's thinking about implementing AI or

563
00:30:07,080 --> 00:30:11,080
Speaker 2: using AI is to really trust in your own domain

564
00:30:11,240 --> 00:30:16,280
Speaker 2: expertise and to allow your domain expertise to lead that process. So,

565
00:30:16,800 --> 00:30:19,440
Speaker 2: you know, while there is a lot of people who

566
00:30:19,440 --> 00:30:22,480
Speaker 2: are saying, oh, you really need to implement AI, you

567
00:30:22,560 --> 00:30:24,000
Speaker 2: need to really get on top of it.

568
00:30:24,320 --> 00:30:24,800
Speaker 5: I think you.

569
00:30:25,080 --> 00:30:27,719
Speaker 2: I think it's also very important to say, you know,

570
00:30:27,760 --> 00:30:30,560
Speaker 2: what does your workflow look like, what are your goals?

571
00:30:30,600 --> 00:30:34,520
Speaker 2: How do you measure success and to what extent would

572
00:30:34,640 --> 00:30:38,400
Speaker 2: adding a particular use case of AI change your workflow,

573
00:30:39,120 --> 00:30:42,160
Speaker 2: because you know, like I work a lot with radiologists

574
00:30:42,160 --> 00:30:46,640
Speaker 2: in the United States, and there isn't necessarily a certainty

575
00:30:46,880 --> 00:30:48,640
Speaker 2: as to, despite the fact that there's a lot of

576
00:30:48,680 --> 00:30:53,880
Speaker 2: excitement about radiology AI, as to what the meaningful difference

577
00:30:53,920 --> 00:30:56,040
Speaker 2: is in terms of how they end up doing their job,

578
00:30:56,160 --> 00:30:59,560
Speaker 2: because there are lots of considerations that radiologists have with

579
00:30:59,680 --> 00:31:03,120
Speaker 2: regard to making sure that they're making the right decisions,

580
00:31:03,160 --> 00:31:05,680
Speaker 2: and like having this information coming in and saying like, oh,

581
00:31:05,680 --> 00:31:07,800
Speaker 2: you should think about it this way they have to

582
00:31:08,000 --> 00:31:10,240
Speaker 2: and then come back and say, okay, well, I understand

583
00:31:10,240 --> 00:31:12,680
Speaker 2: that this is what the model is telling me, but

584
00:31:12,880 --> 00:31:15,560
Speaker 2: ultimately I have to have the final say. I have

585
00:31:15,600 --> 00:31:17,360
Speaker 2: to be the decision maker, and it has to be

586
00:31:17,440 --> 00:31:20,560
Speaker 2: on me who is signing that radiology report, doctor so

587
00:31:20,720 --> 00:31:23,520
Speaker 2: and so MD, and this is my authority. And so,

588
00:31:24,000 --> 00:31:26,360
Speaker 2: you know, I don't necessarily want to be in a

589
00:31:26,400 --> 00:31:29,920
Speaker 2: relationship with this model in which I am ceding authority

590
00:31:29,960 --> 00:31:34,240
Speaker 2: to that model, because ultimately the clinical environment in which

591
00:31:34,240 --> 00:31:37,200
Speaker 2: I am in still has me as being the ultimate

592
00:31:37,280 --> 00:31:40,960
Speaker 2: responsible stakeholder. So to me, you know, it really then

593
00:31:41,040 --> 00:31:44,080
Speaker 2: depends on the kinds of use cases and to really

594
00:31:44,120 --> 00:31:48,520
Speaker 2: have a clear understanding of your own workflow. So for me,

595
00:31:48,720 --> 00:31:51,360
Speaker 2: like the kinds of ways that I use AI is

596
00:31:51,400 --> 00:31:54,240
Speaker 2: oftentimes in ways in which I want to be efficient

597
00:31:54,280 --> 00:31:57,440
Speaker 2: and the stakes are relatively low, So you know, maybe

598
00:31:57,480 --> 00:32:00,560
Speaker 2: I want to summarize, you know, summarize it a set

599
00:32:00,600 --> 00:32:03,440
Speaker 2: of notes or a meeting or something like that, or

600
00:32:03,880 --> 00:32:07,280
Speaker 2: you know, I need a quick reply sent to somebody

601
00:32:07,320 --> 00:32:10,000
Speaker 2: else or something like that. Those sorts of quick things

602
00:32:10,040 --> 00:32:11,720
Speaker 2: I think makes a lot of sense when it comes

603
00:32:11,760 --> 00:32:15,160
Speaker 2: to like larger sorts of questions, and you know, even

604
00:32:15,160 --> 00:32:19,320
Speaker 2: something like legislation. I think that really like understanding the domain,

605
00:32:19,480 --> 00:32:22,680
Speaker 2: understanding the problem and being very specific and bringing as

606
00:32:22,760 --> 00:32:26,400
Speaker 2: much domain knowledge and expertise to the problem is something

607
00:32:26,440 --> 00:32:29,600
Speaker 2: that we should be top of mind and not necessarily get.

608
00:32:29,400 --> 00:32:30,880
Speaker 5: Distracted by the oh this is.

609
00:32:31,120 --> 00:32:34,719
Speaker 2: You know LAMA three. It has so many model parameters,

610
00:32:34,760 --> 00:32:37,040
Speaker 2: it's so big, it can do anything, you know.

611
00:32:37,120 --> 00:32:38,320
Speaker 5: I think we just really need to.

612
00:32:39,520 --> 00:32:44,120
Speaker 2: Keep our problem focused and our goals in mind and

613
00:32:44,160 --> 00:32:47,600
Speaker 2: really instead decide if we are going to be changing

614
00:32:47,600 --> 00:32:50,040
Speaker 2: our workflow, what makes the most sense to make our

615
00:32:50,280 --> 00:32:53,080
Speaker 2: day to day tasks better and then target those tasks

616
00:32:53,160 --> 00:32:53,840
Speaker 2: very deliberately.

617
00:32:54,040 --> 00:32:58,440
Speaker 1: Yeah, I mean healthcare is there are lots of very

618
00:32:58,440 --> 00:33:00,800
Speaker 1: good use cases in there. We've got New Zealand company

619
00:33:00,880 --> 00:33:04,800
Speaker 1: was just acquired recently called Volpower, which is using AI

620
00:33:05,000 --> 00:33:10,880
Speaker 1: to identify tumors or suspicious sort of growths on mammogram images.

621
00:33:11,640 --> 00:33:14,920
Speaker 1: So great technology just been sold automating the process still

622
00:33:14,920 --> 00:33:19,800
Speaker 1: has some human oversight, but very accurate there. So every country,

623
00:33:19,960 --> 00:33:23,760
Speaker 1: our health systems are struggling in the post COVID era,

624
00:33:25,120 --> 00:33:28,680
Speaker 1: hard to find enough people long waiting lists. So I

625
00:33:28,680 --> 00:33:31,360
Speaker 1: guess that's an area where people like yourself, you know,

626
00:33:31,680 --> 00:33:35,000
Speaker 1: machine learning experts AI experts are in great demand to

627
00:33:35,120 --> 00:33:37,920
Speaker 1: actually make sure that this is done in a way

628
00:33:37,960 --> 00:33:40,480
Speaker 1: that is accurate without bias and misinformation.

629
00:33:40,720 --> 00:33:41,080
Speaker 5: Yeah.

630
00:33:41,200 --> 00:33:43,360
Speaker 2: No, I think there's a really lot of really interesting

631
00:33:43,720 --> 00:33:45,280
Speaker 2: use cases within healthcare.

632
00:33:45,400 --> 00:33:47,320
Speaker 5: I think, you know, and I really.

633
00:33:47,080 --> 00:33:51,920
Speaker 2: Am excited about the technologies that allow for meaningful partnership

634
00:33:51,960 --> 00:33:53,240
Speaker 2: with clinicians.

635
00:33:54,040 --> 00:34:00,280
Speaker 1: You mentioned Lamma III, which is Facebook's open source large

636
00:34:00,360 --> 00:34:03,040
Speaker 1: language model, and I was really surprised when they sort

637
00:34:03,040 --> 00:34:06,600
Speaker 1: of went big on this and released previous iterations of this. Here,

638
00:34:06,640 --> 00:34:10,319
Speaker 1: you've got quite a proprietary company that holds onto a

639
00:34:10,400 --> 00:34:14,840
Speaker 1: lot of data, is quite commercially aggressive, going we're not

640
00:34:14,880 --> 00:34:16,920
Speaker 1: going to go to chat GPT route, We're going to

641
00:34:17,040 --> 00:34:20,120
Speaker 1: open sources. Yeah, and you've worked with them, interested in

642
00:34:20,160 --> 00:34:26,320
Speaker 1: your view on that and the strategy around going open source.

643
00:34:26,680 --> 00:34:28,520
Speaker 2: I think that there's been a lot of really exciting

644
00:34:28,560 --> 00:34:32,799
Speaker 2: work happening out of Facebook AI research in terms of

645
00:34:32,960 --> 00:34:40,359
Speaker 2: providing tooling and sharing knowledge through open publication. So you know,

646
00:34:40,440 --> 00:34:42,399
Speaker 2: a lot of the tools that I use every day

647
00:34:42,920 --> 00:34:45,560
Speaker 2: is the open source library PyTorch, which was invented at

648
00:34:45,600 --> 00:34:49,080
Speaker 2: Facebook and is now just available to public and many

649
00:34:49,120 --> 00:34:52,080
Speaker 2: people use it today day to day to build AA models.

650
00:34:52,080 --> 00:34:54,600
Speaker 2: It's what my typical framework that I use, and so

651
00:34:54,680 --> 00:34:58,080
Speaker 2: I think it's just part of a larger tradition that

652
00:34:58,120 --> 00:35:02,360
Speaker 2: has happened within Facebook with regards words to open source

653
00:35:02,440 --> 00:35:06,440
Speaker 2: and regards to taking a very science focused approach to

654
00:35:06,480 --> 00:35:08,320
Speaker 2: the publication of research.

655
00:35:08,000 --> 00:35:10,239
Speaker 1: And sort of hand in hand with that, if you

656
00:35:10,320 --> 00:35:14,520
Speaker 1: can gain access to these models at low costs, you

657
00:35:14,520 --> 00:35:17,080
Speaker 1: still obviously have to train them and run them, which

658
00:35:17,160 --> 00:35:20,600
Speaker 1: is intensive in terms of computer capacity. But it sort

659
00:35:20,600 --> 00:35:23,400
Speaker 1: of has given rise to discussions in maybe some smaller

660
00:35:23,400 --> 00:35:27,080
Speaker 1: countries like New Zealand around sovereign AI and the potential

661
00:35:27,120 --> 00:35:30,240
Speaker 1: for a country like our tairo to have its own

662
00:35:30,800 --> 00:35:33,279
Speaker 1: large language model or foundational model, is that something you

663
00:35:33,320 --> 00:35:34,080
Speaker 1: think is realistic.

664
00:35:34,239 --> 00:35:36,160
Speaker 2: Yeah, I mean, I think that I've seen this happen

665
00:35:36,200 --> 00:35:40,640
Speaker 2: at least in other countries where there are very specific

666
00:35:40,719 --> 00:35:42,920
Speaker 2: use cases because of the fact that the national language

667
00:35:42,960 --> 00:35:43,640
Speaker 2: may not be English.

668
00:35:43,680 --> 00:35:45,720
Speaker 5: So we see this, for example, happening in Korea.

669
00:35:45,800 --> 00:35:50,879
Speaker 2: We see what is happening in the UAE and in

670
00:35:51,400 --> 00:35:54,440
Speaker 2: other Gulf states, and so those cases, it makes a

671
00:35:54,480 --> 00:35:56,520
Speaker 2: lot of sense to be able to say, like, we

672
00:35:56,560 --> 00:35:59,440
Speaker 2: want to be able to build a model lingual model

673
00:35:59,560 --> 00:36:05,960
Speaker 2: because we don't necessarily want the technical concerns around the

674
00:36:06,320 --> 00:36:10,920
Speaker 2: data being over in English overwhelming the capabilities of the

675
00:36:11,080 --> 00:36:17,000
Speaker 2: language model, because sometimes there are negative impacts of multi linguality.

676
00:36:17,200 --> 00:36:21,280
Speaker 2: So I definitely understand that, least anecdotally.

677
00:36:20,600 --> 00:36:22,400
Speaker 5: From what I've seen in the States.

678
00:36:22,120 --> 00:36:24,640
Speaker 2: There is a lot of interest in fine tuning open

679
00:36:24,680 --> 00:36:28,880
Speaker 2: source models or even closed source models on a particular domain.

680
00:36:29,440 --> 00:36:31,759
Speaker 2: Then again, at the same time, I do think that

681
00:36:31,760 --> 00:36:34,720
Speaker 2: there's a lot of international attention given to large language

682
00:36:34,760 --> 00:36:37,640
Speaker 2: models that are published and available to the public. So

683
00:36:37,719 --> 00:36:40,680
Speaker 2: anytime you publish a large you publish a large language

684
00:36:40,680 --> 00:36:42,840
Speaker 2: model and you put it up on hugging face, it gets.

685
00:36:42,640 --> 00:36:43,360
Speaker 5: A lot of attention.

686
00:36:43,760 --> 00:36:46,760
Speaker 2: So you know, for example, my currently most cited paper

687
00:36:47,080 --> 00:36:49,920
Speaker 2: is the paper that I participated in in building Bloom,

688
00:36:49,920 --> 00:36:52,920
Speaker 2: which is a large multi lingual language model, and so

689
00:36:53,000 --> 00:36:54,600
Speaker 2: that does tend to get a lot of attention. And

690
00:36:54,640 --> 00:36:57,520
Speaker 2: so for I guess from a standpoint of being able

691
00:36:57,560 --> 00:36:59,520
Speaker 2: to say, you know, we have the skill sets in

692
00:37:00,080 --> 00:37:02,480
Speaker 2: and knowledge in a particular place to say that we

693
00:37:02,520 --> 00:37:04,719
Speaker 2: know how to train our own language models because and

694
00:37:04,760 --> 00:37:07,680
Speaker 2: now we understand this process very deeply. You know, I

695
00:37:07,680 --> 00:37:09,120
Speaker 2: think it makes a lot of sense to have a

696
00:37:09,160 --> 00:37:11,280
Speaker 2: lot of interest in building one's own lms.

697
00:37:11,560 --> 00:37:12,320
Speaker 5: At the same.

698
00:37:12,120 --> 00:37:15,239
Speaker 2: Time, you know, there is certain advantages to being able

699
00:37:15,320 --> 00:37:17,000
Speaker 2: to just say, you know what, like let's just take

700
00:37:17,040 --> 00:37:19,279
Speaker 2: a really powerful pre train model, fine tuned it a

701
00:37:19,320 --> 00:37:21,000
Speaker 2: little bit, and now we can tailor it to a

702
00:37:21,040 --> 00:37:24,040
Speaker 2: particular task. And so I see the benefits also towards

703
00:37:24,120 --> 00:37:26,479
Speaker 2: using it in other use cases as well. So in

704
00:37:26,520 --> 00:37:29,240
Speaker 2: that case, you know, it really is up to business

705
00:37:29,239 --> 00:37:32,560
Speaker 2: people and research communities in New Zealand to really focus

706
00:37:32,600 --> 00:37:35,560
Speaker 2: on what their ultimate needs are and to build the

707
00:37:35,600 --> 00:37:37,239
Speaker 2: technology that makes THEEMO sense for them.

708
00:37:37,360 --> 00:37:40,919
Speaker 1: Sure in terms of where we're at when it comes

709
00:37:40,960 --> 00:37:45,680
Speaker 1: to the accuracy of large language models, the journey that

710
00:37:45,719 --> 00:37:48,000
Speaker 1: the average user has been on over the last couple

711
00:37:48,040 --> 00:37:49,400
Speaker 1: of years, I might have seen a little bit of

712
00:37:49,880 --> 00:37:53,839
Speaker 1: hallucination out of chat GPT or Bang famously had some

713
00:37:54,120 --> 00:37:58,720
Speaker 1: very weird conversations when chat GPT was integrated into being Search.

714
00:37:59,160 --> 00:38:00,520
Speaker 1: We seem to have come a lot way in the

715
00:38:00,600 --> 00:38:04,600
Speaker 1: last couple of years. What's your assessment about how fine

716
00:38:04,640 --> 00:38:09,279
Speaker 1: tuned these models have become to reduce the likelihood of

717
00:38:09,280 --> 00:38:11,360
Speaker 1: them spinning out misinformation and bias.

718
00:38:11,600 --> 00:38:13,880
Speaker 2: I mean, I think unfortunately, there are a lot of

719
00:38:14,239 --> 00:38:18,480
Speaker 2: technical challenges that still remain with regards to ensuring the

720
00:38:18,480 --> 00:38:21,880
Speaker 2: integrity of information because ultimately we do not know the

721
00:38:22,000 --> 00:38:25,319
Speaker 2: quality of the information that is being input into these

722
00:38:25,360 --> 00:38:28,840
Speaker 2: models at the time of training. So a great example

723
00:38:28,880 --> 00:38:32,800
Speaker 2: of this was somebody had posted that they had queries

724
00:38:32,960 --> 00:38:35,880
Speaker 2: Gemini how many rocks they should eat a day, and

725
00:38:35,920 --> 00:38:38,040
Speaker 2: they said you should eat one rock a day because

726
00:38:38,080 --> 00:38:40,000
Speaker 2: they pulled an article from the Onion that said you

727
00:38:40,000 --> 00:38:43,040
Speaker 2: should eat one were not a day. Now, ultimately this

728
00:38:43,080 --> 00:38:44,840
Speaker 2: has a lot to do with the fact that Gemini

729
00:38:44,880 --> 00:38:47,480
Speaker 2: does not know that the Onion is not a.

730
00:38:47,800 --> 00:38:49,920
Speaker 5: Serious source of news.

731
00:38:49,960 --> 00:38:52,880
Speaker 2: And so, you know, if you don't know if a

732
00:38:52,960 --> 00:38:56,640
Speaker 2: document is of good quality, you're going to model it

733
00:38:56,920 --> 00:38:58,960
Speaker 2: agnostically and just assume that it is.

734
00:38:59,360 --> 00:39:02,360
Speaker 5: And so with out a real clear understanding.

735
00:39:01,840 --> 00:39:05,360
Speaker 2: Of the factuality or quality of different kinds of information,

736
00:39:05,400 --> 00:39:08,080
Speaker 2: which it doesn't necessarily have, you're still going to have

737
00:39:08,120 --> 00:39:10,560
Speaker 2: some of these challenges coming down the pike. And so

738
00:39:11,200 --> 00:39:14,560
Speaker 2: to me, you know, it's ultimately a long term challenge

739
00:39:14,560 --> 00:39:16,600
Speaker 2: that we are going to have to reckon with. And

740
00:39:17,160 --> 00:39:19,200
Speaker 2: some of the hope I think that some people have

741
00:39:19,520 --> 00:39:23,279
Speaker 2: is with greater information, greater transparency. This is the great

742
00:39:23,440 --> 00:39:25,640
Speaker 2: the great example of this that just happened recently was

743
00:39:25,680 --> 00:39:29,880
Speaker 2: the Instagram post by Taylor Swift in which she clarifies

744
00:39:29,920 --> 00:39:32,719
Speaker 2: her own political leanings and what her plans are for

745
00:39:32,840 --> 00:39:34,759
Speaker 2: the upcoming US elections.

746
00:39:35,120 --> 00:39:35,319
Speaker 3: You know.

747
00:39:35,400 --> 00:39:38,840
Speaker 2: And so I think, you know, greater information, greater transparency,

748
00:39:39,080 --> 00:39:41,239
Speaker 2: and even you know, things that are happening on the

749
00:39:41,280 --> 00:39:47,560
Speaker 2: technology side about maybe cryptographic verification of information or with

750
00:39:47,600 --> 00:39:51,960
Speaker 2: regards to watermarking, I think will be technologies that will

751
00:39:51,960 --> 00:39:55,280
Speaker 2: continue will grow in importance as we continue to develop

752
00:39:55,320 --> 00:39:56,080
Speaker 2: these technologies.

753
00:39:56,280 --> 00:39:58,279
Speaker 1: Yeah, I think when it comes to the government's little

754
00:39:58,320 --> 00:40:04,840
Speaker 1: experiment gov G. They're using a framework called retrieval augmented operation,

755
00:40:05,280 --> 00:40:09,640
Speaker 1: which maybe you can explain, but as I understand is

756
00:40:09,760 --> 00:40:13,279
Speaker 1: it will cross reference official sources of government information to

757
00:40:13,320 --> 00:40:15,560
Speaker 1: make sure that you're not going to get a hallucin,

758
00:40:15,600 --> 00:40:18,000
Speaker 1: it's not going to make up information on the spot.

759
00:40:18,280 --> 00:40:24,759
Speaker 2: Yeah, yeah, Yeah, that kind of retrieval augmented output is

760
00:40:25,320 --> 00:40:28,480
Speaker 2: very exciting. That's the kind of thing that we see

761
00:40:29,080 --> 00:40:31,839
Speaker 2: even with the kinds of output about Google Search. Unfortunately,

762
00:40:31,880 --> 00:40:34,120
Speaker 2: they have a much broader set of documents they've been

763
00:40:34,120 --> 00:40:36,520
Speaker 2: looking through, and not say a narrow set of documents

764
00:40:36,520 --> 00:40:40,600
Speaker 2: like specific government reports that are officially known to be

765
00:40:40,719 --> 00:40:42,280
Speaker 2: of a certain level of quality.

766
00:40:42,440 --> 00:40:44,279
Speaker 1: So there's clearly a lot of work that still needs

767
00:40:44,320 --> 00:40:46,920
Speaker 1: to be done on that front. And at the other

768
00:40:47,040 --> 00:40:50,239
Speaker 1: end of it, for instance, we've just had in Australia

769
00:40:50,520 --> 00:40:54,799
Speaker 1: they've floated some mandatory guard rails for AI which are

770
00:40:54,800 --> 00:40:58,600
Speaker 1: out for public consultation, so the government may be legislating

771
00:40:58,600 --> 00:41:01,000
Speaker 1: on the basis of those, and it seems to have

772
00:41:01,040 --> 00:41:05,840
Speaker 1: some quite pragmatic stuff like record keeping so actually show

773
00:41:07,160 --> 00:41:10,759
Speaker 1: allow third parties access to compliance and that sort of thing.

774
00:41:11,080 --> 00:41:15,640
Speaker 1: Undertaking conformity assessments, audits, that sort of stuff which at

775
00:41:15,640 --> 00:41:18,839
Speaker 1: the moment are really not happening to be honest, at least,

776
00:41:18,880 --> 00:41:22,520
Speaker 1: there's no external scrutiny of these, and frankly probably a

777
00:41:22,560 --> 00:41:25,160
Speaker 1: lot of organizations that are using these large language models,

778
00:41:25,200 --> 00:41:27,960
Speaker 1: and the clue what goes on in that black box?

779
00:41:28,080 --> 00:41:31,320
Speaker 1: How do we get some standardization and some real best

780
00:41:31,400 --> 00:41:35,600
Speaker 1: practice around those sorts of checks and balances to make

781
00:41:35,640 --> 00:41:38,560
Speaker 1: sure that these systems are running as intended?

782
00:41:38,960 --> 00:41:39,800
Speaker 5: Yeah, I mean, I think.

783
00:41:39,640 --> 00:41:42,160
Speaker 2: That there's a lot of interest in audits. I think

784
00:41:42,160 --> 00:41:48,440
Speaker 2: that ultimately again focus on use cases and downstream applications.

785
00:41:48,520 --> 00:41:51,480
Speaker 2: Is really the level of granularity that I would like

786
00:41:51,560 --> 00:41:56,960
Speaker 2: to focus on validation and testing to be geared towards,

787
00:41:56,960 --> 00:41:59,440
Speaker 2: as opposed to taking a more general AI as a

788
00:41:59,520 --> 00:42:02,600
Speaker 2: technology G level approach, though, you know, I think this

789
00:42:02,680 --> 00:42:05,560
Speaker 2: really will be application and use case specific, you know.

790
00:42:05,719 --> 00:42:08,640
Speaker 2: At the same time, like, while there's a lot of

791
00:42:08,680 --> 00:42:12,560
Speaker 2: excitement around audits, I am a little bit concerned about

792
00:42:12,960 --> 00:42:15,480
Speaker 2: how do we know for certain the quality of the

793
00:42:15,480 --> 00:42:18,920
Speaker 2: conclusions we make because these systems are very, very complicated,

794
00:42:19,320 --> 00:42:22,120
Speaker 2: and I don't necessarily want to get us into a

795
00:42:22,200 --> 00:42:26,440
Speaker 2: kind of framework in which we are limiting innovation because

796
00:42:26,440 --> 00:42:29,440
Speaker 2: of the fact that you know a you know, astray

797
00:42:29,880 --> 00:42:33,440
Speaker 2: output at one time, because the model may have you know,

798
00:42:33,520 --> 00:42:37,799
Speaker 2: there is a degree of stochasticity of the kinds of outputs.

799
00:42:38,000 --> 00:42:42,040
Speaker 2: But if you happen to randomly pick the article that

800
00:42:42,120 --> 00:42:45,680
Speaker 2: says you should go eat rocks in order to say

801
00:42:45,800 --> 00:42:47,759
Speaker 2: you know whether or not you should eat rocks, and

802
00:42:47,800 --> 00:42:50,319
Speaker 2: that one time you tell the auditor go eat a rock,

803
00:42:50,680 --> 00:42:53,200
Speaker 2: you know, that ends up making the conclusion, oh, this

804
00:42:53,280 --> 00:42:55,720
Speaker 2: thing is not safe, as opposed to say, oh, well,

805
00:42:55,960 --> 00:42:58,759
Speaker 2: you know, ninety nine percent of the time we are

806
00:42:58,800 --> 00:43:01,640
Speaker 2: not outputting you should go lead rocks. So to me,

807
00:43:01,800 --> 00:43:04,800
Speaker 2: it's I guess I have a lot of concerns about

808
00:43:05,000 --> 00:43:07,200
Speaker 2: what do we really mean when we get to the

809
00:43:07,320 --> 00:43:13,200
Speaker 2: specific business of doing audits and evaluation, is you know,

810
00:43:13,280 --> 00:43:15,359
Speaker 2: how do we make these conclusions and how do we

811
00:43:15,719 --> 00:43:21,399
Speaker 2: make sure that these conclusions are based on meaningful assessments

812
00:43:21,440 --> 00:43:26,839
Speaker 2: of the underlying dynamics of the model, Because ultimately there

813
00:43:26,840 --> 00:43:30,120
Speaker 2: are so many pieces and partners of the model that

814
00:43:30,239 --> 00:43:37,080
Speaker 2: are not deterministic that you need to basically make conclusions

815
00:43:37,120 --> 00:43:41,360
Speaker 2: based on some sort of statistical testing or variedability. But

816
00:43:41,480 --> 00:43:44,839
Speaker 2: those sorts of conclusions are not necessarily clear. I think

817
00:43:45,080 --> 00:43:47,560
Speaker 2: historically what we've seen in terms of audits has been

818
00:43:47,640 --> 00:43:51,640
Speaker 2: much more of a shaming campaign in which, you know,

819
00:43:51,760 --> 00:43:55,920
Speaker 2: say a journalist has tested out a system has found

820
00:43:55,920 --> 00:44:00,359
Speaker 2: that there is some sort of legally problematic out come

821
00:44:00,400 --> 00:44:03,280
Speaker 2: that occurs. So, for example, like we look at something

822
00:44:03,440 --> 00:44:05,400
Speaker 2: like the kinds of stuff that's been happening out of

823
00:44:05,440 --> 00:44:09,440
Speaker 2: pro Publica, there's a really exciting yeah, really exciting we're

824
00:44:09,480 --> 00:44:12,200
Speaker 2: coming out of pro Publica. But you know, at the

825
00:44:12,239 --> 00:44:16,600
Speaker 2: same time, that's a one time view. It's not necessarily

826
00:44:17,120 --> 00:44:21,560
Speaker 2: a broader understanding of a much more complex system like

827
00:44:22,040 --> 00:44:25,439
Speaker 2: say uh GPT four or something like that, in which

828
00:44:25,440 --> 00:44:29,560
Speaker 2: there are many different use cases, many different users, and

829
00:44:29,640 --> 00:44:32,000
Speaker 2: there's lots and lots of data involved. So to me,

830
00:44:32,760 --> 00:44:35,600
Speaker 2: you know, given that I work on the evaluation of

831
00:44:35,640 --> 00:44:38,560
Speaker 2: AI systems as an academic, and I have a hard

832
00:44:38,600 --> 00:44:42,319
Speaker 2: time wrapping around my head around to what extent we

833
00:44:42,400 --> 00:44:45,239
Speaker 2: should be happy with the outputs for a specific task,

834
00:44:45,320 --> 00:44:49,720
Speaker 2: say even English language summarization of documents. I'm concerned about

835
00:44:49,920 --> 00:44:53,000
Speaker 2: how other people coming in will make these conclusions, because

836
00:44:53,000 --> 00:44:55,960
Speaker 2: even with my level of expertise, I think it's a

837
00:44:56,200 --> 00:44:57,560
Speaker 2: challenging problem.

838
00:44:57,600 --> 00:45:00,520
Speaker 1: So it sounds like you said, focus on the use cases.

839
00:45:00,520 --> 00:45:03,640
Speaker 1: So if you apply your deep domain knowledge to whether

840
00:45:03,680 --> 00:45:09,200
Speaker 1: it's government policy formation or radiograms in hospitals or something

841
00:45:09,239 --> 00:45:11,399
Speaker 1: like that, you're much more likely to have a good

842
00:45:11,400 --> 00:45:15,640
Speaker 1: handle on whether the output of this AI system is

843
00:45:15,680 --> 00:45:16,600
Speaker 1: actually accurate.

844
00:45:17,000 --> 00:45:20,960
Speaker 2: Yeah, And I think you know, more fundamentally, you can

845
00:45:21,400 --> 00:45:25,120
Speaker 2: know if it makes a positive outcome on your day

846
00:45:25,160 --> 00:45:27,919
Speaker 2: to day workflow. You know, if it's not making your

847
00:45:27,960 --> 00:45:31,920
Speaker 2: life easier and it's making things more frustrating, then you

848
00:45:31,960 --> 00:45:35,120
Speaker 2: know this the inclusion of this technology may be a problem.

849
00:45:35,360 --> 00:45:37,799
Speaker 2: I mean, one of the things I'm thinking about is

850
00:45:38,280 --> 00:45:40,719
Speaker 2: recently in the United States there was a whole outage

851
00:45:40,920 --> 00:45:42,960
Speaker 2: of the airline system.

852
00:45:43,000 --> 00:45:44,000
Speaker 5: I don't know if you saw the.

853
00:45:43,920 --> 00:45:46,120
Speaker 1: Crowdstrikes struck here as well.

854
00:45:46,239 --> 00:45:49,200
Speaker 2: Yeah, and so CrowdStrike was a huge problem. And you know,

855
00:45:49,280 --> 00:45:52,600
Speaker 2: we build all of these systems on top of technology

856
00:45:52,920 --> 00:45:55,799
Speaker 2: and just assume that everything will be fine, and then

857
00:45:56,040 --> 00:45:58,200
Speaker 2: and we then become reliant on it, and then we

858
00:45:58,239 --> 00:46:00,960
Speaker 2: don't know necessarily how to function without it. And so

859
00:46:01,080 --> 00:46:03,360
Speaker 2: to me, you know, even though that's like a you know,

860
00:46:03,400 --> 00:46:07,040
Speaker 2: a relatively you know, non AI example, I think that

861
00:46:07,120 --> 00:46:10,040
Speaker 2: as a as a researcher who builds a model. It's

862
00:46:10,040 --> 00:46:13,040
Speaker 2: one thing to build the model and demonstrate it on

863
00:46:13,080 --> 00:46:15,960
Speaker 2: a benchmark, and it's another thing to have it ongoing

864
00:46:16,320 --> 00:46:19,520
Speaker 2: in a system as part of people's everyday workflow, in

865
00:46:19,560 --> 00:46:21,840
Speaker 2: which they are having to interact with it, you know,

866
00:46:22,280 --> 00:46:24,920
Speaker 2: like hundreds, if not millions of times a day. And

867
00:46:25,000 --> 00:46:30,080
Speaker 2: so to me, you know, like, ultimately, I think the

868
00:46:30,200 --> 00:46:33,080
Speaker 2: fundamental benchmark we need to go back to is, you know,

869
00:46:34,000 --> 00:46:36,279
Speaker 2: does it help us do our jobs better? Does it

870
00:46:36,320 --> 00:46:39,600
Speaker 2: help us achieve our goals? And that's ultimately something that

871
00:46:39,880 --> 00:46:42,040
Speaker 2: you know, domain experts have to answer for us.

872
00:46:42,160 --> 00:46:44,759
Speaker 1: Yeah, and I think, you know, we have a productivity

873
00:46:44,760 --> 00:46:48,480
Speaker 1: problem in New Zealand. We're less productive, we work longer

874
00:46:48,520 --> 00:46:51,279
Speaker 1: hours and earn less than than other nations. So we've

875
00:46:51,280 --> 00:46:53,239
Speaker 1: got an issue there and AI is seen as part

876
00:46:53,280 --> 00:46:55,759
Speaker 1: of the answer to that. And we're already saying in

877
00:46:55,880 --> 00:46:59,440
Speaker 1: offices all over this country where people have access to copilot,

878
00:47:00,040 --> 00:47:04,080
Speaker 1: grudgery of creating a PowerPoint presentation or trying to write

879
00:47:04,080 --> 00:47:07,320
Speaker 1: a document, it's starting to go away because the stuff

880
00:47:07,360 --> 00:47:10,120
Speaker 1: actually really works for that. But what are what are

881
00:47:10,120 --> 00:47:14,080
Speaker 1: the use cases that really excite you as a machine

882
00:47:14,160 --> 00:47:16,279
Speaker 1: learning expert. You're working in the hell sick there. What

883
00:47:16,320 --> 00:47:19,400
Speaker 1: are the ones where you go, Wow, this will fundamentally

884
00:47:19,520 --> 00:47:22,319
Speaker 1: change people's lives in a very positive way.

885
00:47:22,600 --> 00:47:25,200
Speaker 2: The example that I want to give back to you

886
00:47:25,480 --> 00:47:28,600
Speaker 2: is the kinds of things happening out of Papareo, you know,

887
00:47:28,800 --> 00:47:33,680
Speaker 2: like Tahiku Media that just was named in the time.

888
00:47:33,760 --> 00:47:36,320
Speaker 2: It's time a y one hundred and so I think

889
00:47:36,680 --> 00:47:40,040
Speaker 2: you know that they have a model there they've been

890
00:47:40,160 --> 00:47:45,960
Speaker 2: building on Deo Maori language pronunciation in which they're fine tuning,

891
00:47:46,239 --> 00:47:49,480
Speaker 2: you know, open source language models in order to improve

892
00:47:49,920 --> 00:47:50,600
Speaker 2: language learning.

893
00:47:50,640 --> 00:47:52,440
Speaker 5: I think that's a really exciting application.

894
00:47:52,800 --> 00:47:56,200
Speaker 2: I also think that you know that that the kinds

895
00:47:56,200 --> 00:47:59,640
Speaker 2: of things that are happening with regards to increasing the

896
00:47:59,640 --> 00:48:04,359
Speaker 2: ability and the evaluation of large language models and other

897
00:48:04,480 --> 00:48:07,120
Speaker 2: languages is something that excites me. This is what I've

898
00:48:07,120 --> 00:48:10,919
Speaker 2: been working on most recently, and also on evaluating them

899
00:48:11,040 --> 00:48:14,440
Speaker 2: for the cultural norms and cultural biases that we need

900
00:48:14,480 --> 00:48:15,280
Speaker 2: to be concerned about.

901
00:48:15,320 --> 00:48:16,800
Speaker 5: So those are things that I'm really excited.

902
00:48:17,239 --> 00:48:21,240
Speaker 1: And just finally, we're not that far from a big

903
00:48:21,320 --> 00:48:27,920
Speaker 1: US election presidential election. We saw the candidates recently on

904
00:48:28,000 --> 00:48:30,960
Speaker 1: stage debating and around that there was a sort of

905
00:48:31,000 --> 00:48:35,080
Speaker 1: a lot of sort of AI driven misinformation and images

906
00:48:35,080 --> 00:48:38,359
Speaker 1: of Taylor Swift, you know, supporting Donald Trump and the like.

907
00:48:38,840 --> 00:48:40,959
Speaker 1: So it is playing a little bit of a role

908
00:48:41,320 --> 00:48:46,359
Speaker 1: in attempts to manipulate democracy. But what's your view on

909
00:48:47,080 --> 00:48:50,160
Speaker 1: the impact in the mid to long term that AI

910
00:48:50,280 --> 00:48:53,600
Speaker 1: is going to have fundamentally and how we elect people,

911
00:48:53,760 --> 00:48:57,239
Speaker 1: how the voting system works, how we run a democracy.

912
00:48:57,400 --> 00:49:02,759
Speaker 2: I mean, my hope is that people take seriously their

913
00:49:03,000 --> 00:49:08,279
Speaker 2: role in informing AI policy and play a bigger role

914
00:49:08,480 --> 00:49:12,799
Speaker 2: in having conversations with their elected leaders in telling them

915
00:49:12,840 --> 00:49:17,239
Speaker 2: about the kinds of relationships with AI they want to have.

916
00:49:17,360 --> 00:49:21,400
Speaker 2: Because without public input, we are not going to be

917
00:49:21,440 --> 00:49:26,520
Speaker 2: able to know how our technology will have a meaningfully

918
00:49:26,560 --> 00:49:29,640
Speaker 2: positive impact on people's lives. So my hope is that,

919
00:49:30,120 --> 00:49:34,000
Speaker 2: you know, as we engage with this technology, we start

920
00:49:34,080 --> 00:49:37,040
Speaker 2: to understand what we want out of it so that

921
00:49:37,080 --> 00:49:40,919
Speaker 2: we can inform our elected leaders, and through that then

922
00:49:41,080 --> 00:49:43,239
Speaker 2: my hope is that then in the long run, we

923
00:49:43,360 --> 00:49:48,279
Speaker 2: can encourage researchers then to continue their efforts in and

924
00:49:48,880 --> 00:49:52,920
Speaker 2: funding of work that might be used to mitigate a mitigate,

925
00:49:53,040 --> 00:49:57,920
Speaker 2: mitigate disinformation and disinformation, because there are efforts to be

926
00:49:58,760 --> 00:50:02,319
Speaker 2: there can be used to either watermark the kinds of

927
00:50:02,640 --> 00:50:05,600
Speaker 2: content that comes out of language models as being specifically

928
00:50:05,600 --> 00:50:11,280
Speaker 2: AI generated, or to cryptographically validate the authenticity of a document.

929
00:50:11,680 --> 00:50:14,880
Speaker 2: But this has gotten much less attention than say, you know,

930
00:50:15,080 --> 00:50:17,920
Speaker 2: building bigger and bigger models. And so my hope is that,

931
00:50:18,040 --> 00:50:21,160
Speaker 2: you know, if we emphasize that this is important, we

932
00:50:21,239 --> 00:50:24,440
Speaker 2: can encourage the further development.

933
00:50:24,000 --> 00:50:24,960
Speaker 5: Of these technologies.

934
00:50:25,080 --> 00:50:27,600
Speaker 1: Yeah, so that's all really important, and I guess it's

935
00:50:27,640 --> 00:50:31,080
Speaker 1: also the institutions of society have to really get on board.

936
00:50:31,120 --> 00:50:35,120
Speaker 1: For instance, you can have GROWK, which allows the faces

937
00:50:35,160 --> 00:50:38,160
Speaker 1: of public figures to be to be put on deep

938
00:50:38,440 --> 00:50:41,080
Speaker 1: deep fakes. Others don't allow that. So we've got to

939
00:50:41,120 --> 00:50:44,120
Speaker 1: get some standardization and everyone on the same page, and

940
00:50:44,120 --> 00:50:46,800
Speaker 1: if necessary, with a bit of a stick, a regulatory

941
00:50:46,880 --> 00:50:49,480
Speaker 1: stick as well, to say this is what is appropriate

942
00:50:49,520 --> 00:50:51,160
Speaker 1: if we want to have a healthy democracy.

943
00:50:51,440 --> 00:50:53,799
Speaker 2: Yeah, and I think it's really up to members of

944
00:50:53,960 --> 00:50:58,719
Speaker 2: various democracies around the world to have these conversations, you know,

945
00:50:59,080 --> 00:51:02,680
Speaker 2: both among them and with their elected officials about what

946
00:51:02,719 --> 00:51:06,760
Speaker 2: their their hopes were, you know, a trustworthy AERI system

947
00:51:06,840 --> 00:51:07,920
Speaker 2: might be in the future.

948
00:51:14,880 --> 00:51:17,839
Speaker 3: I think she must have been pretty excited when you know,

949
00:51:18,000 --> 00:51:21,440
Speaker 3: she was diving on machine learning and AI for her

950
00:51:21,680 --> 00:51:25,960
Speaker 3: studies and suddenly the world exploded around her. So she's

951
00:51:25,960 --> 00:51:28,960
Speaker 3: obviously pretty must be pretty stoked about that. And I

952
00:51:29,000 --> 00:51:32,640
Speaker 3: think the focus on languages is fascinating and talking about

953
00:51:32,640 --> 00:51:36,560
Speaker 3: it in relation to today Amori and how she said that,

954
00:51:36,600 --> 00:51:38,200
Speaker 3: you know a lot of the stuff that's happening here

955
00:51:38,280 --> 00:51:41,680
Speaker 3: is kind of being looked to from other jurisdictions as well.

956
00:51:41,719 --> 00:51:43,919
Speaker 3: I thought that was exciting to hear.

957
00:51:44,280 --> 00:51:49,719
Speaker 1: Yeah, we do have companies like Hiku Media, which has

958
00:51:49,760 --> 00:51:52,840
Speaker 1: done all that great work we've talked about before, building

959
00:51:52,880 --> 00:51:56,919
Speaker 1: this sort of AI model for Terrey Amari and doing

960
00:51:56,960 --> 00:51:59,680
Speaker 1: it with ninety two percent accuracy. This is speech to

961
00:51:59,719 --> 00:52:04,200
Speaker 1: tear next translation, so they've been developing that. So I

962
00:52:04,200 --> 00:52:06,680
Speaker 1: think Jessica sort of looked at that and went wow,

963
00:52:07,000 --> 00:52:10,400
Speaker 1: because you know, I heard this incredible statistic. There is

964
00:52:11,719 --> 00:52:16,759
Speaker 1: about three thousand languages that are at risk of basically disappearing.

965
00:52:18,080 --> 00:52:22,000
Speaker 1: That's according to ENESCO. So every year it's a little

966
00:52:22,040 --> 00:52:25,319
Speaker 1: bit like you hear the endangered species and species at

967
00:52:25,360 --> 00:52:27,520
Speaker 1: risk of going extinct, it's the same with languages. We

968
00:52:27,560 --> 00:52:30,320
Speaker 1: have a lot of languages, thousands of languages around the world,

969
00:52:31,200 --> 00:52:36,040
Speaker 1: and the ones that unfortunately are less popularly spoken are

970
00:52:36,080 --> 00:52:38,839
Speaker 1: at risk of dying out. So Jessica has been looking

971
00:52:38,880 --> 00:52:44,000
Speaker 1: particularly at Southeast Asian languages, those ones that don't get

972
00:52:44,040 --> 00:52:48,440
Speaker 1: the exposure in large language models, and going what do

973
00:52:48,480 --> 00:52:50,800
Speaker 1: we need to do? And part of it is technical

974
00:52:51,600 --> 00:52:56,239
Speaker 1: making it easier for people to build models that incorporate

975
00:52:56,640 --> 00:52:59,160
Speaker 1: those sources, but also the quality of the data that

976
00:52:59,280 --> 00:53:02,960
Speaker 1: is put into them, and and the testing and the

977
00:53:03,120 --> 00:53:07,160
Speaker 1: checking off the accuracy. So it's great that she's you know,

978
00:53:07,200 --> 00:53:09,400
Speaker 1: she set out on this quest and through building that

979
00:53:09,560 --> 00:53:14,680
Speaker 1: multi lingual model herself, to actually try and preserve some

980
00:53:14,719 --> 00:53:17,719
Speaker 1: of these languages in the AI world so that those

981
00:53:17,760 --> 00:53:23,800
Speaker 1: people and everyone else can actually experience those languages through AI.

982
00:53:23,960 --> 00:53:26,879
Speaker 1: And here actually here you know, through a multimodal AI.

983
00:53:27,040 --> 00:53:28,080
Speaker 1: Hear what they sound like?

984
00:53:28,360 --> 00:53:33,680
Speaker 3: Hmm yeah, yeah, I mean it's the modern resurgence of

985
00:53:33,719 --> 00:53:37,359
Speaker 3: like trying to recapture languages that might be disappearing has

986
00:53:37,719 --> 00:53:40,919
Speaker 3: been kind of an underswell that's happening for a long time.

987
00:53:40,960 --> 00:53:44,799
Speaker 3: But I think the ability for AI to store these

988
00:53:44,840 --> 00:53:49,600
Speaker 3: things digitally and recreate them in a more active way

989
00:53:49,920 --> 00:53:53,360
Speaker 3: rather than just being you know, a dictionary for example,

990
00:53:53,360 --> 00:53:56,160
Speaker 3: which doesn't have that same level of depth. I think

991
00:53:56,160 --> 00:53:57,760
Speaker 3: that's super exciting.

992
00:53:57,800 --> 00:54:01,279
Speaker 1: Yeah, and just really nice to having come from, you know,

993
00:54:01,480 --> 00:54:05,839
Speaker 1: from Silicon Valley. Everyone is just trying to monetize how

994
00:54:05,880 --> 00:54:08,480
Speaker 1: do I monetize AI? How do I build it into

995
00:54:08,520 --> 00:54:12,280
Speaker 1: my platform and ten x my business and my customer's

996
00:54:12,360 --> 00:54:16,239
Speaker 1: business and all that to actually then talk to researchers

997
00:54:16,280 --> 00:54:18,759
Speaker 1: who are fundamentally more interested and how do we make

998
00:54:18,800 --> 00:54:23,960
Speaker 1: this accurate and robust and something that people should be

999
00:54:24,120 --> 00:54:27,360
Speaker 1: using and not going to cause more harm than good.

1000
00:54:27,840 --> 00:54:31,080
Speaker 1: So it's just great that you know, there is a

1001
00:54:31,160 --> 00:54:33,839
Speaker 1: big pool of people in the US and elsewhere who

1002
00:54:33,840 --> 00:54:37,960
Speaker 1: are actually thinking deeply about this. The question is to

1003
00:54:38,320 --> 00:54:44,799
Speaker 1: what extent that their warnings and their approach are adopted

1004
00:54:45,080 --> 00:54:48,040
Speaker 1: into the products that are being put out. And at

1005
00:54:48,040 --> 00:54:50,680
Speaker 1: the moment, I think the industry is well ahead of them.

1006
00:54:50,719 --> 00:54:52,960
Speaker 1: We're getting to the point now where all of the

1007
00:54:52,960 --> 00:54:57,040
Speaker 1: big developments are coming out of private companies like open

1008
00:54:57,120 --> 00:55:01,960
Speaker 1: AI and Microsoft and AWS. Right, other than the Stanfords

1009
00:55:01,960 --> 00:55:05,319
Speaker 1: and the Brown universities, they're still they still pay play

1010
00:55:05,360 --> 00:55:09,120
Speaker 1: an important part. But that makes me nervous that you know,

1011
00:55:09,160 --> 00:55:12,399
Speaker 1: this really the most powerful technology off the twenty first

1012
00:55:12,440 --> 00:55:15,640
Speaker 1: century has really been driven by private enterprise.

1013
00:55:16,400 --> 00:55:20,440
Speaker 3: Yeah, I definitely, I definitely feel you there. It is

1014
00:55:20,440 --> 00:55:23,280
Speaker 3: a bit of a concern how much power that these

1015
00:55:23,440 --> 00:55:29,080
Speaker 3: companies have over over the technology that's fundamentally ground shifting,

1016
00:55:29,280 --> 00:55:34,399
Speaker 3: especially when you have almost weekly news articles about how

1017
00:55:34,480 --> 00:55:40,319
Speaker 3: open ai is unstable, or somebody's gone, or they want

1018
00:55:40,320 --> 00:55:42,319
Speaker 3: to get rid of the nonprofit oversight and all of

1019
00:55:42,320 --> 00:55:45,400
Speaker 3: these kind of yeah, worrying signs.

1020
00:55:45,640 --> 00:55:50,160
Speaker 1: They are worrying signs. Anyway, great to catch up with

1021
00:55:50,520 --> 00:55:54,799
Speaker 1: Jessica Zosa Ford and Alex Cho. Thanks to both of

1022
00:55:54,840 --> 00:55:57,240
Speaker 1: them for coming on this week's episode of the Business

1023
00:55:57,239 --> 00:55:59,640
Speaker 1: of Tech. More on both of them in the show notes.

1024
00:55:59,680 --> 00:56:02,759
Speaker 1: Go to business Desk, dot Co, dot m Z for

1025
00:56:02,880 --> 00:56:06,759
Speaker 1: those and all seventy episodes of the podcast now, which

1026
00:56:06,800 --> 00:56:10,680
Speaker 1: you can stream there via iHeartRadio or your favorite podcast

1027
00:56:10,760 --> 00:56:11,640
Speaker 1: app yep.

1028
00:56:11,719 --> 00:56:14,040
Speaker 3: Drop us a line with suggestions for future guests, and

1029
00:56:14,080 --> 00:56:17,040
Speaker 3: you can email me on Ben at business Desk, dot co,

1030
00:56:17,320 --> 00:56:20,000
Speaker 3: dot and z or fine both of us on LinkedIn and.

1031
00:56:20,520 --> 00:56:24,200
Speaker 1: X and another episode coming your way next Thursday.

1032
00:56:24,440 --> 00:56:26,040
Speaker 3: Until then, have a great week.