1
00:00:02,800 --> 00:00:05,200
Speaker 1: Cura and welcome back to the Business of Tech Powered

2
00:00:05,240 --> 00:00:08,520
Speaker 1: by Two Degrees. I'm Peter Griffin, and today we're looking

3
00:00:08,520 --> 00:00:11,400
Speaker 1: at what I think is one of the most consequential

4
00:00:11,520 --> 00:00:15,440
Speaker 1: questions facing New Zealand's digital future. How do we build

5
00:00:15,520 --> 00:00:20,640
Speaker 1: real autonomy over the tools, data and infrastructure that increasingly

6
00:00:20,720 --> 00:00:25,360
Speaker 1: run our economy and society. The recent national Artificial Intelligence

7
00:00:25,400 --> 00:00:28,680
Speaker 1: Strategy released by the government, well, it had a pretty

8
00:00:28,720 --> 00:00:33,120
Speaker 1: tepid response from industry and from academics. It was seen

9
00:00:33,159 --> 00:00:37,120
Speaker 1: as underwhelming, missing critical areas, and with little in the

10
00:00:37,120 --> 00:00:40,559
Speaker 1: way of firm goals or targets to guide the way forward.

11
00:00:41,000 --> 00:00:44,920
Speaker 1: There was no mention of sovereign AI, a dedicated effort

12
00:00:45,000 --> 00:00:48,239
Speaker 1: to build our own AI capability, so we aren't reliant

13
00:00:48,360 --> 00:00:53,680
Speaker 1: on overseas platforms and technologies. Other countries have prioritized this.

14
00:00:54,360 --> 00:00:58,040
Speaker 1: The Swiss, for instance, have just released a large language

15
00:00:58,040 --> 00:01:01,800
Speaker 1: model developed on their own supercomput by their own scientists.

16
00:01:02,200 --> 00:01:05,800
Speaker 1: But they've also open sourced the model so that anyone

17
00:01:05,959 --> 00:01:09,080
Speaker 1: can use it for free. The government may not have

18
00:01:09,240 --> 00:01:12,319
Speaker 1: mentioned it or be thinking about it, but leveraging open

19
00:01:12,400 --> 00:01:15,720
Speaker 1: source AI is probably the best way for us to

20
00:01:15,800 --> 00:01:19,360
Speaker 1: avoid being reliant on models from the likes of open

21
00:01:19,480 --> 00:01:24,640
Speaker 1: AI chat GPT debuted last week too decidedly mixed reviews.

22
00:01:25,240 --> 00:01:28,119
Speaker 1: So Don Christy returns to the Business of Tech this

23
00:01:28,240 --> 00:01:32,479
Speaker 1: week to discuss open source AI. Donn as co founder

24
00:01:32,560 --> 00:01:36,080
Speaker 1: of Catalyst It and Catalyst Cloud, which is built on

25
00:01:36,160 --> 00:01:39,880
Speaker 1: the open Stack open source platform. He's got some interesting

26
00:01:39,920 --> 00:01:44,000
Speaker 1: ideas on what sovereign AI should actually mean here. He

27
00:01:44,080 --> 00:01:48,240
Speaker 1: sees it as an issue of self determination, choosing architectures,

28
00:01:48,480 --> 00:01:52,440
Speaker 1: models and governance that keep New Zealand in control of

29
00:01:52,520 --> 00:01:56,240
Speaker 1: critical systems and sensitive data. We get into the nuts

30
00:01:56,280 --> 00:01:59,160
Speaker 1: and bolts of open source AI. Y open models and

31
00:01:59,200 --> 00:02:03,400
Speaker 1: transparent training data sets matter how locally owned cloud can

32
00:02:03,440 --> 00:02:07,040
Speaker 1: deliver practical AI today without the need to spend hundreds

33
00:02:07,080 --> 00:02:11,240
Speaker 1: of millions on high powered hips from Nvidia. We also

34
00:02:11,280 --> 00:02:16,000
Speaker 1: tackle the thorny parts political bias and models, data exposure

35
00:02:16,480 --> 00:02:20,359
Speaker 1: under foreign jurisdictions, and the risk of defaulting to bundled

36
00:02:20,520 --> 00:02:24,240
Speaker 1: AI just because it might be included in the license agreement.

37
00:02:24,520 --> 00:02:26,680
Speaker 1: All that and more on the Business of Tech, So

38
00:02:26,960 --> 00:02:28,480
Speaker 1: stay tuned for the interview at dawn.

39
00:02:37,160 --> 00:02:38,359
Speaker 2: Don Chrissy, how are you doing.

40
00:02:38,760 --> 00:02:40,080
Speaker 3: I'm doing very well, thank you.

41
00:02:40,240 --> 00:02:43,320
Speaker 4: Yeah, there's lots of exciting things happening, and thank you

42
00:02:43,400 --> 00:02:44,440
Speaker 4: very much for having me back.

43
00:02:44,520 --> 00:02:47,800
Speaker 1: Well, it's great to be back at Catalyst I Catalyst Cloud.

44
00:02:47,840 --> 00:02:51,560
Speaker 1: Gloomy day in Wellington. It feels like, you know, winter

45
00:02:51,720 --> 00:02:56,040
Speaker 1: is dragging along, but a lot of really cool stuff happening.

46
00:02:56,080 --> 00:02:59,000
Speaker 1: We just saw this morning, for instance, the public release

47
00:02:59,040 --> 00:03:03,000
Speaker 1: of chat GP a little bit delayed. It's been sort

48
00:03:03,040 --> 00:03:04,560
Speaker 1: of hyped up. Have you had to play with it.

49
00:03:04,560 --> 00:03:05,440
Speaker 3: Yet at the moment.

50
00:03:06,000 --> 00:03:10,239
Speaker 4: My tool that we use for corporate services, Catalyst, that

51
00:03:11,040 --> 00:03:15,800
Speaker 4: tends to be claued AI. We find its tone much

52
00:03:15,800 --> 00:03:18,880
Speaker 4: more aligned with Catalyst tone, and it's we find it

53
00:03:19,080 --> 00:03:21,960
Speaker 4: safer tool to use as well. It's more thoughtful if

54
00:03:21,960 --> 00:03:24,400
Speaker 4: you can humanize it that way in terms of the

55
00:03:24,440 --> 00:03:25,480
Speaker 4: responses it gives.

56
00:03:25,880 --> 00:03:29,919
Speaker 1: As Sam Orman has said, incremental gains. Everything's just getting faster,

57
00:03:30,120 --> 00:03:35,400
Speaker 1: more efficient, better subject matter expertise, better reasoning powers. And

58
00:03:35,440 --> 00:03:37,520
Speaker 1: I guess that's the journey we're on. We've got all

59
00:03:37,520 --> 00:03:41,600
Speaker 1: these big players who have billions of dollars of investment

60
00:03:41,640 --> 00:03:45,360
Speaker 1: now high stakes because they need to return on investment,

61
00:03:45,360 --> 00:03:47,160
Speaker 1: and that's really what we're going to talk about. In

62
00:03:47,200 --> 00:03:50,040
Speaker 1: the context of this conversation, there's been a lot of

63
00:03:50,040 --> 00:03:54,000
Speaker 1: discussion about so called sovereign AI and the role that

64
00:03:54,080 --> 00:03:56,400
Speaker 1: open source can play in it.

65
00:03:56,440 --> 00:03:58,440
Speaker 2: And you've got a lot of expertise in that area.

66
00:03:58,880 --> 00:04:02,240
Speaker 1: But just to maybe give this conversation a bit of context,

67
00:04:02,240 --> 00:04:04,880
Speaker 1: we had a few weeks back the government put out

68
00:04:04,920 --> 00:04:10,120
Speaker 1: a national Artificial Intelligence strategy for New Zealand and frankly,

69
00:04:10,200 --> 00:04:13,160
Speaker 1: it was not well received. The tone of commentary on

70
00:04:13,560 --> 00:04:18,320
Speaker 1: LinkedIn from experts in the field very disappointed, and it's

71
00:04:18,400 --> 00:04:21,200
Speaker 1: seen as more of a vision statement really than a strategy,

72
00:04:21,240 --> 00:04:24,840
Speaker 1: and certainly not backed up with any significant new investments

73
00:04:24,960 --> 00:04:28,160
Speaker 1: or initiatives around AI. What was your take on it

74
00:04:28,200 --> 00:04:28,800
Speaker 1: when you read it?

75
00:04:28,880 --> 00:04:31,560
Speaker 4: My take was probably more nuanced than that. My take

76
00:04:31,680 --> 00:04:35,039
Speaker 4: is that the government is making a start. When I

77
00:04:35,120 --> 00:04:38,080
Speaker 4: looked at that strategy, I thought it was quite generic

78
00:04:38,120 --> 00:04:40,839
Speaker 4: in its applications, So if you were a small business

79
00:04:41,520 --> 00:04:45,080
Speaker 4: for a manufacturing company, it would give you some idea

80
00:04:45,120 --> 00:04:47,600
Speaker 4: of what the technologies could be used for in your

81
00:04:47,680 --> 00:04:51,719
Speaker 4: day to day work. And my understanding, certainly in talking

82
00:04:51,760 --> 00:04:55,240
Speaker 4: to people in government, is that this is a start

83
00:04:55,440 --> 00:04:59,080
Speaker 4: and they really want to hear from New Zealand organizations

84
00:04:59,520 --> 00:05:02,080
Speaker 4: that are doing amazing stuff with AI, they really do,

85
00:05:02,800 --> 00:05:07,120
Speaker 4: and then from that they can develop a strategy, funding

86
00:05:07,240 --> 00:05:10,799
Speaker 4: and so on. So you know, we partner with companies

87
00:05:10,839 --> 00:05:14,760
Speaker 4: like Tehikomedia who have been doing some amazing work in

88
00:05:14,839 --> 00:05:16,960
Speaker 4: AI in the Maori language.

89
00:05:16,520 --> 00:05:18,799
Speaker 3: Space for decades.

90
00:05:19,480 --> 00:05:22,160
Speaker 4: David Bribner, who founded You Imagine, I don't know if

91
00:05:22,160 --> 00:05:25,360
Speaker 4: you've heard of them. Incredible work that they do, you know,

92
00:05:26,120 --> 00:05:31,720
Speaker 4: from safety on construction sites to quality control on EV

93
00:05:31,920 --> 00:05:35,880
Speaker 4: manufacturing plants in Europe, applying the AI technologies in so

94
00:05:35,960 --> 00:05:39,840
Speaker 4: many different ways. So it's important for New Zealanders not

95
00:05:40,000 --> 00:05:44,599
Speaker 4: to allow our narrative to be smothered by the narratives

96
00:05:44,640 --> 00:05:47,240
Speaker 4: that's coming largely out of the US, and not to

97
00:05:48,120 --> 00:05:51,560
Speaker 4: feel that we are helpless actors because we don't have

98
00:05:51,600 --> 00:05:53,920
Speaker 4: five hundred billion dollars to build a data center.

99
00:05:54,200 --> 00:05:57,599
Speaker 1: Yeah, but having I totally agree with that. But the

100
00:05:57,640 --> 00:06:00,839
Speaker 1: tone of the document really sort of it could almost

101
00:06:00,839 --> 00:06:02,320
Speaker 1: have been written by the multinationals.

102
00:06:02,360 --> 00:06:02,520
Speaker 3: You know.

103
00:06:02,560 --> 00:06:05,560
Speaker 1: It's basically like it's all about adoption. The priorities get

104
00:06:05,800 --> 00:06:09,680
Speaker 1: everyone using it, but we don't really care what they're using.

105
00:06:09,960 --> 00:06:13,960
Speaker 1: And the obvious choice at the moments is these multinationals,

106
00:06:14,040 --> 00:06:18,040
Speaker 1: it's as you're using open AI. There was no mention

107
00:06:18,160 --> 00:06:20,880
Speaker 1: in their off sovereign AI, you know, building our own

108
00:06:21,000 --> 00:06:26,320
Speaker 1: capability ourselves. There was reference to data sovereignty and interesting

109
00:06:26,360 --> 00:06:29,960
Speaker 1: your views on what sovereign AI is and actually should be.

110
00:06:30,120 --> 00:06:33,640
Speaker 1: Is it full stack development off our own AI from

111
00:06:33,680 --> 00:06:37,400
Speaker 1: hardware through to the models that run on it, or

112
00:06:38,000 --> 00:06:42,040
Speaker 1: can we do bits of that that allow us to

113
00:06:42,120 --> 00:06:45,080
Speaker 1: have enough autonomy to steer our own path on AI.

114
00:06:45,360 --> 00:06:48,000
Speaker 4: So I mean sovereignty means different things to different people,

115
00:06:48,000 --> 00:06:51,760
Speaker 4: particularly in the New Zealand context. There's jurisdictional sovereignty and

116
00:06:51,760 --> 00:06:56,440
Speaker 4: then there's langoti danga sovereignty and self determination. So from

117
00:06:56,480 --> 00:07:00,280
Speaker 4: my perspective, in the open source narrative, we talk about

118
00:07:00,320 --> 00:07:04,760
Speaker 4: self determination a lot, and that can apply to individuals, organizations,

119
00:07:04,839 --> 00:07:10,800
Speaker 4: communities and countries. So the question then is how do

120
00:07:10,960 --> 00:07:15,760
Speaker 4: these technologies remove self determination or how do they build

121
00:07:15,760 --> 00:07:20,280
Speaker 4: that self determination. I was at the largest open source

122
00:07:20,320 --> 00:07:24,840
Speaker 4: conference in Australia New Zealand in January, the same day

123
00:07:25,360 --> 00:07:28,520
Speaker 4: that Trump's inauguration took place, and Elon Musk was doing

124
00:07:29,000 --> 00:07:33,560
Speaker 4: Nazi salutes. All those people tech billionaires that were on

125
00:07:33,600 --> 00:07:36,280
Speaker 4: stage with Donald Trump. They built their empires out of

126
00:07:36,320 --> 00:07:39,480
Speaker 4: open source. My message to the government into New Zealand

127
00:07:39,520 --> 00:07:42,520
Speaker 4: and This speaks to the strategy is cut out the middleman.

128
00:07:42,720 --> 00:07:44,920
Speaker 4: The technologies are there, you don't have to build it

129
00:07:44,960 --> 00:07:47,760
Speaker 4: from scratch, as you were saying, is one option. The

130
00:07:47,840 --> 00:07:50,360
Speaker 4: technologies are there. We've done this with the Linux, We've

131
00:07:50,360 --> 00:07:52,960
Speaker 4: done this with open stack in the cloud space. And

132
00:07:53,040 --> 00:07:57,360
Speaker 4: so as the open source models start maturing, suddenly the

133
00:07:57,440 --> 00:08:02,880
Speaker 4: opportunities to build self determination within New Zealand, however you

134
00:08:02,960 --> 00:08:06,160
Speaker 4: want to think about, that will explode. And if we're

135
00:08:06,200 --> 00:08:10,000
Speaker 4: thinking about strategically and what the government could be doing,

136
00:08:10,720 --> 00:08:12,960
Speaker 4: we should think about making sure we don't close off

137
00:08:13,000 --> 00:08:16,640
Speaker 4: those opportunities, and that we encourage Kiwi businesses that are

138
00:08:16,640 --> 00:08:21,320
Speaker 4: investing in those opportunities. But with patronage making sure that

139
00:08:21,360 --> 00:08:23,760
Speaker 4: the government and corporates and so on use it.

140
00:08:24,520 --> 00:08:27,760
Speaker 1: And two with investment, we certainly haven't cut off those opportunities.

141
00:08:27,760 --> 00:08:30,320
Speaker 1: We're just not really doing anything to pursue them. I

142
00:08:30,360 --> 00:08:33,640
Speaker 1: think is the frustration that tech community, the business community

143
00:08:33,760 --> 00:08:36,120
Speaker 1: is to some extent feels. But as you said, there

144
00:08:36,120 --> 00:08:40,000
Speaker 1: are some great open source large language models out there now.

145
00:08:40,080 --> 00:08:45,360
Speaker 1: Obviously Meta interestingly has been a leader, although their latest

146
00:08:45,400 --> 00:08:48,600
Speaker 1: version disappointed in terms of its benchmarking. But the one

147
00:08:48,600 --> 00:08:51,240
Speaker 1: there's a lot of excitement around is the Swiss large

148
00:08:51,320 --> 00:08:55,080
Speaker 1: language model developed on the Alps supercomputer. They have a

149
00:08:55,120 --> 00:09:00,000
Speaker 1: massive supercomputer in Switzerland. Ten thousand Invidia GPUs twenty million

150
00:09:00,679 --> 00:09:05,920
Speaker 1: GPU hours annually is available to them. Eight hundred researchers

151
00:09:05,960 --> 00:09:10,000
Speaker 1: across Switzerland are involved in this government funded effort to

152
00:09:10,080 --> 00:09:14,040
Speaker 1: build an open source large language model that has been released.

153
00:09:14,400 --> 00:09:16,320
Speaker 1: I don't know if you've looked at it interested in

154
00:09:16,360 --> 00:09:18,280
Speaker 1: your perspective on how we could use that.

155
00:09:18,600 --> 00:09:21,360
Speaker 4: So what I see is that initiative, whether it's a

156
00:09:21,400 --> 00:09:24,520
Speaker 4: Swiss one or another one, will be something that basically

157
00:09:24,520 --> 00:09:27,600
Speaker 4: the rest of the world coalesces around and further developed.

158
00:09:27,640 --> 00:09:30,400
Speaker 4: We saw it happen with open stacks. That was a

159
00:09:30,440 --> 00:09:33,839
Speaker 4: cloud platform, open source that came out of a collaboration

160
00:09:33,960 --> 00:09:37,400
Speaker 4: between CERN and NASA who were having to store and

161
00:09:37,480 --> 00:09:40,480
Speaker 4: process the largest data sets that the world had ever seen,

162
00:09:41,040 --> 00:09:43,760
Speaker 4: and they released their open stack product. And that's what

163
00:09:43,800 --> 00:09:47,200
Speaker 4: we're using Catalyst Cloud. It was Hyperscale before hyperscale was

164
00:09:47,200 --> 00:09:50,400
Speaker 4: invented as a term, and it became the largest IT

165
00:09:50,840 --> 00:09:54,679
Speaker 4: project in the world with contributions from anyone that wasn't

166
00:09:54,920 --> 00:09:58,160
Speaker 4: AWS or Microsoft basically, and we'll see that happen with

167
00:09:58,840 --> 00:10:03,040
Speaker 4: examples like the Swiss model, and they will iterate, and

168
00:10:03,080 --> 00:10:05,200
Speaker 4: I'll keep repeating, and we will be able to take

169
00:10:05,240 --> 00:10:09,839
Speaker 4: those models and where appropriate, bring them to our own servers,

170
00:10:10,000 --> 00:10:13,400
Speaker 4: our own GPUs, and build on them and build in

171
00:10:13,520 --> 00:10:17,360
Speaker 4: local context and build in local solutions, and be able

172
00:10:17,360 --> 00:10:21,360
Speaker 4: to provide that at a very individual level to organizations

173
00:10:21,559 --> 00:10:26,520
Speaker 4: that want to keep their self determination very close to hand,

174
00:10:27,520 --> 00:10:32,000
Speaker 4: at an incredibly affordable price. So that cuts out the

175
00:10:32,040 --> 00:10:36,400
Speaker 4: need and kind of undermines this whole narrative that you

176
00:10:36,440 --> 00:10:39,199
Speaker 4: have to have five hundred billion dollars to build a

177
00:10:39,280 --> 00:10:42,200
Speaker 4: data center in Texas, where you need to reopen a

178
00:10:43,360 --> 00:10:46,600
Speaker 4: blown up nuclear power station on one mile island or whatever.

179
00:10:47,320 --> 00:10:51,080
Speaker 4: Those narratives are quite deliberately set up to smother the

180
00:10:51,200 --> 00:10:53,880
Speaker 4: idea that you can do AI independently.

181
00:10:54,000 --> 00:10:56,240
Speaker 1: So it's obviously cost the Swiss tens, if not one,

182
00:10:56,280 --> 00:10:59,400
Speaker 1: hundreds of millions of dollars to build this, right, So

183
00:10:59,440 --> 00:11:02,800
Speaker 1: it's a significo investment. But in terms of what is

184
00:11:02,880 --> 00:11:05,760
Speaker 1: required for us to take that open source large language

185
00:11:05,760 --> 00:11:08,800
Speaker 1: model and put it on Catalyst Cloud or some local

186
00:11:08,800 --> 00:11:12,960
Speaker 1: infrastructure and data commerce Spark or the newest supercomputer if

187
00:11:13,000 --> 00:11:16,080
Speaker 1: it's technically capable of doing it. What are we talking

188
00:11:16,120 --> 00:11:19,600
Speaker 1: about in terms of the technical requirements.

189
00:11:19,040 --> 00:11:19,400
Speaker 3: I don't know.

190
00:11:19,520 --> 00:11:23,440
Speaker 4: I mean the AI research soundboxes that we've set up,

191
00:11:23,720 --> 00:11:25,959
Speaker 4: and we've got one that we use internally for R

192
00:11:26,000 --> 00:11:27,840
Speaker 4: and D. I think we've got about a dozen different

193
00:11:27,840 --> 00:11:30,600
Speaker 4: open source models that people tool around with, and they're

194
00:11:30,600 --> 00:11:34,479
Speaker 4: basically those ones are running on one GPU with a

195
00:11:34,520 --> 00:11:37,760
Speaker 4: forty Mega whatever's a VRAM. It's not the best performance,

196
00:11:37,840 --> 00:11:42,120
Speaker 4: but we've just modernized our GPUs on Catalyst Cloud. We're

197
00:11:42,120 --> 00:11:47,120
Speaker 4: getting great performance running these models. We've got clients using

198
00:11:47,200 --> 00:11:51,319
Speaker 4: them and ready to launch products on them. Using large

199
00:11:51,400 --> 00:11:56,319
Speaker 4: language models with literally just a handful of GPUs, You've

200
00:11:56,320 --> 00:11:59,520
Speaker 4: got people like to Hickamedia who are investing for New

201
00:11:59,600 --> 00:12:03,360
Speaker 4: Zealand amounts of money, but nothing like you know we're

202
00:12:03,400 --> 00:12:07,840
Speaker 4: talking about overseas and doing amazing things with it. So again,

203
00:12:07,920 --> 00:12:13,280
Speaker 4: it's it's finding out what people need to use AI

204
00:12:13,400 --> 00:12:19,760
Speaker 4: for and providing a contextual solution using these freely available technologies.

205
00:12:19,760 --> 00:12:22,200
Speaker 4: Whether it's a Swiss models, or whether it's small language

206
00:12:22,200 --> 00:12:27,319
Speaker 4: models quite specific, or whether it's some other AI technology

207
00:12:27,360 --> 00:12:30,000
Speaker 4: that's been around for a long time. There's a lot

208
00:12:30,040 --> 00:12:32,760
Speaker 4: of sort of machine learning capability that we've used for

209
00:12:32,800 --> 00:12:34,080
Speaker 4: our clients for decades.

210
00:12:34,200 --> 00:12:38,160
Speaker 1: If it's simply as good as the Claudes and the

211
00:12:38,760 --> 00:12:43,800
Speaker 1: open aiyes, the fact that we are fine tuning at ourselves.

212
00:12:43,800 --> 00:12:47,920
Speaker 1: It's open source. There's this concept emerging off, you know,

213
00:12:48,679 --> 00:12:52,760
Speaker 1: federated AI. So at a national level it could be

214
00:12:52,760 --> 00:12:56,800
Speaker 1: the Australians, the Swiss and all of that. We contribute

215
00:12:57,040 --> 00:13:01,920
Speaker 1: many Maybe it would be AI expertise research capabilities as

216
00:13:02,000 --> 00:13:05,560
Speaker 1: well as our investment in hardware and software to an

217
00:13:05,600 --> 00:13:08,760
Speaker 1: effort that countries at a national level can take advantage

218
00:13:08,760 --> 00:13:11,160
Speaker 1: of so we can create something. Sure we'll still have

219
00:13:11,240 --> 00:13:15,000
Speaker 1: the hyperscales here and we'll be using those to some extent,

220
00:13:15,280 --> 00:13:19,480
Speaker 1: but the default option could be the national AI that

221
00:13:19,520 --> 00:13:20,679
Speaker 1: we've created ourselves.

222
00:13:20,840 --> 00:13:23,920
Speaker 4: Yeah, this is where concepts like matarangamori are really helpful,

223
00:13:24,440 --> 00:13:27,280
Speaker 4: because if you think about the governance of knowledge and

224
00:13:27,400 --> 00:13:32,080
Speaker 4: the concepts that some knowledge is there and been there

225
00:13:32,120 --> 00:13:35,840
Speaker 4: for hundreds of years to be shared, some knowledge is

226
00:13:35,880 --> 00:13:38,760
Speaker 4: there to be shared in specific contexts, and some knowledge

227
00:13:38,800 --> 00:13:41,440
Speaker 4: is not to be shared. And if you have those

228
00:13:41,440 --> 00:13:46,160
Speaker 4: sort of governance contexts clearly understood, then you can absolutely

229
00:13:46,200 --> 00:13:49,520
Speaker 4: do what you've described. You can say, well, here's our information,

230
00:13:49,640 --> 00:13:52,960
Speaker 4: here's our knowledge, here's our data. We can contribute this

231
00:13:53,240 --> 00:13:56,080
Speaker 4: to the national infrastructure, but it's got to stay national

232
00:13:56,240 --> 00:13:58,560
Speaker 4: or no, no, we can actually contribute that to a

233
00:13:58,559 --> 00:14:02,600
Speaker 4: global knowledge base. But as as we start getting closer

234
00:14:02,600 --> 00:14:06,319
Speaker 4: and closer to our own issues of self determination, then

235
00:14:06,360 --> 00:14:06,959
Speaker 4: we want.

236
00:14:06,760 --> 00:14:09,240
Speaker 3: To be able to.

237
00:14:08,360 --> 00:14:13,040
Speaker 4: Use the technology on this data, but only for ourselves.

238
00:14:13,640 --> 00:14:18,160
Speaker 4: And those are the sort of nuances and models that

239
00:14:18,760 --> 00:14:22,520
Speaker 4: you know, open source software and open source large language

240
00:14:22,520 --> 00:14:26,120
Speaker 4: models or any language model can enable. And in the

241
00:14:26,120 --> 00:14:30,560
Speaker 4: Swiss context, they're talking about the algorithms being open, the

242
00:14:30,680 --> 00:14:33,400
Speaker 4: data sets that they're using for training being.

243
00:14:33,400 --> 00:14:34,080
Speaker 3: Clear and.

244
00:14:35,920 --> 00:14:36,560
Speaker 2: Doesn't do that.

245
00:14:36,880 --> 00:14:38,240
Speaker 3: No, what deepca.

246
00:14:37,960 --> 00:14:41,080
Speaker 4: Has trained itself on open LAMA and then added stuff

247
00:14:41,120 --> 00:14:41,640
Speaker 4: on top.

248
00:14:41,440 --> 00:14:43,720
Speaker 3: Of that, and that's how open source works as well.

249
00:14:44,040 --> 00:14:45,000
Speaker 3: There is nothing wrong with.

250
00:14:46,440 --> 00:14:47,240
Speaker 2: What others have done.

251
00:14:47,520 --> 00:14:48,200
Speaker 3: It's fine.

252
00:14:48,760 --> 00:14:51,840
Speaker 4: And you know, the scary thing about some of those

253
00:14:52,000 --> 00:14:58,320
Speaker 4: US models now is the US presidency wanting to have

254
00:14:58,480 --> 00:15:03,600
Speaker 4: control over what information comes out of those models, So

255
00:15:03,720 --> 00:15:06,360
Speaker 4: even if they're open source, they will have biases in

256
00:15:06,400 --> 00:15:08,080
Speaker 4: there that will be politically driven.

257
00:15:08,160 --> 00:15:11,360
Speaker 1: The other issue, which you've spoken about repeatedly when it

258
00:15:11,400 --> 00:15:15,520
Speaker 1: comes to the changing political climate in the US. The

259
00:15:15,600 --> 00:15:20,320
Speaker 1: challenging of legal precedents on a whole host of issues

260
00:15:20,520 --> 00:15:23,960
Speaker 1: is the data sovereignty issue. And we had an executive

261
00:15:23,960 --> 00:15:28,400
Speaker 1: from Microsoft France, Anton Carneo. He said recently in front

262
00:15:28,440 --> 00:15:30,440
Speaker 1: of the French Senate when they ask them, can you

263
00:15:30,520 --> 00:15:34,240
Speaker 1: guarantee sovereignty of data for your customers in France? And

264
00:15:34,240 --> 00:15:36,840
Speaker 1: he said, look, I can't, no, because there is the

265
00:15:36,840 --> 00:15:41,080
Speaker 1: Cloud Act introduced in twenty eighteen. Technically the US government

266
00:15:41,120 --> 00:15:44,840
Speaker 1: could ask for this information that hasn't happened, and Microsoft

267
00:15:44,920 --> 00:15:46,440
Speaker 1: takes it very seriously.

268
00:15:46,480 --> 00:15:51,880
Speaker 4: Well that we know of, because under other situations, the

269
00:15:52,000 --> 00:15:56,720
Speaker 4: US government can ask for data on foreign nationals that

270
00:15:56,800 --> 00:15:59,000
Speaker 4: they don't have to notify those requests about.

271
00:15:59,120 --> 00:16:03,800
Speaker 3: Right, That's a cute bit that they kind of don't highlight.

272
00:16:04,080 --> 00:16:07,640
Speaker 4: Yeah, and the fact that Microsoft pulled some of the

273
00:16:07,640 --> 00:16:11,960
Speaker 4: capability for the International Criminal Court in Europe just shows

274
00:16:12,080 --> 00:16:13,560
Speaker 4: how long that reach is.

275
00:16:13,920 --> 00:16:14,720
Speaker 2: Yeah.

276
00:16:14,800 --> 00:16:19,200
Speaker 1: So that argument remains that you've made quite strongly in

277
00:16:19,280 --> 00:16:22,600
Speaker 1: terms of data sovereignty. If you really want data sovereignty,

278
00:16:22,640 --> 00:16:27,520
Speaker 1: it's got to be on sovereign infrastructure, local cloud providers essentially.

279
00:16:27,280 --> 00:16:28,560
Speaker 3: Yeah, and locally owned.

280
00:16:28,600 --> 00:16:33,120
Speaker 4: I mean, you know, it doesn't matter if your Amazon Microsoft, Oracle, Google,

281
00:16:33,440 --> 00:16:36,480
Speaker 4: You are still a US company just because Larry Allison

282
00:16:36,600 --> 00:16:40,760
Speaker 4: hired Russell Cootz. He didn't suddenly become Team New Zealand, right,

283
00:16:41,440 --> 00:16:46,680
Speaker 4: And so it's about understanding those risks. I'm not going

284
00:16:46,760 --> 00:16:48,720
Speaker 4: to say don't use them, of course not. You know,

285
00:16:48,920 --> 00:16:51,560
Speaker 4: that's that ship of saled and we're not going to

286
00:16:51,560 --> 00:16:55,560
Speaker 4: cut our noses off. But we do need to support choice,

287
00:16:55,600 --> 00:16:58,520
Speaker 4: and we need to support resilience, and we need to

288
00:16:58,720 --> 00:17:04,640
Speaker 4: be aware that our supply chains are no longer as

289
00:17:05,080 --> 00:17:08,720
Speaker 4: secure and allied to our interests as we might have

290
00:17:08,760 --> 00:17:10,920
Speaker 4: thought there were five or ten years ago.

291
00:17:11,240 --> 00:17:14,160
Speaker 2: So we've got a strategy. There is some good stuff

292
00:17:14,200 --> 00:17:14,560
Speaker 2: going on.

293
00:17:14,600 --> 00:17:16,679
Speaker 1: We've got a new research institute that's going to have

294
00:17:16,760 --> 00:17:20,000
Speaker 1: AI capability as part of it, so hopefully that will

295
00:17:20,040 --> 00:17:22,520
Speaker 1: inject some money into the research community. Where should our

296
00:17:22,560 --> 00:17:25,160
Speaker 1: priorities lie do you think, I mean, should it be?

297
00:17:25,520 --> 00:17:28,200
Speaker 1: Should we be looking very seriously at this point at

298
00:17:28,240 --> 00:17:30,960
Speaker 1: open source large language models, running them on our own

299
00:17:31,000 --> 00:17:34,199
Speaker 1: infrastructure here, fine tuning them. Is that something that is

300
00:17:34,240 --> 00:17:35,680
Speaker 1: going to shift a needle for us?

301
00:17:36,040 --> 00:17:38,840
Speaker 4: What we have New Zealand is seen as a trusted

302
00:17:39,280 --> 00:17:43,240
Speaker 4: leader in many areas. You know, that's why our produce

303
00:17:43,280 --> 00:17:47,320
Speaker 4: sells so well. You know, we're certified that we feed

304
00:17:47,320 --> 00:17:49,600
Speaker 4: our charas on grass all that sort of stuff, and

305
00:17:50,080 --> 00:17:53,879
Speaker 4: that trust should enable us to be world leaders in

306
00:17:53,920 --> 00:17:57,800
Speaker 4: the digital space and particularly in AI. And so this

307
00:17:57,880 --> 00:18:02,320
Speaker 4: is a huge opportunity, so adless of the ethics of

308
00:18:02,440 --> 00:18:08,520
Speaker 4: sovereign AI. From a value perspective, New Zealand is a

309
00:18:08,560 --> 00:18:11,800
Speaker 4: place now where indigenous populations are beginning to ask us

310
00:18:11,880 --> 00:18:15,240
Speaker 4: to look after their data because in their own jurisdictions

311
00:18:15,480 --> 00:18:19,720
Speaker 4: they're under threat and that's again because of Maori leadership

312
00:18:19,880 --> 00:18:23,840
Speaker 4: and t Tariti and things like that, and so we

313
00:18:23,840 --> 00:18:26,080
Speaker 4: should be able to build on that and I just

314
00:18:26,160 --> 00:18:29,240
Speaker 4: don't want us to miss out on that opportunity. One

315
00:18:29,280 --> 00:18:32,320
Speaker 4: of the biggest opportunities, the most important ones from a

316
00:18:32,359 --> 00:18:35,560
Speaker 4: New Zealand economic perspective, would be the use of AI

317
00:18:35,800 --> 00:18:41,720
Speaker 4: in agriculture and giving our farmers the tools to create,

318
00:18:41,920 --> 00:18:45,840
Speaker 4: produce and look after their land in ways that leaves

319
00:18:45,840 --> 00:18:50,159
Speaker 4: them in control of their livelihoods and their businesses that

320
00:18:50,240 --> 00:18:52,600
Speaker 4: you just don't You won't get if all your data

321
00:18:52,640 --> 00:18:55,640
Speaker 4: is going to John Deere in Montana or Wisconsin, wherever

322
00:18:55,720 --> 00:18:59,560
Speaker 4: they headquarters, because as soon as you start allowing that

323
00:18:59,640 --> 00:19:01,679
Speaker 4: to happen. And it does happen because John Dea collects

324
00:19:01,680 --> 00:19:03,920
Speaker 4: a lot of data from its machinery and it does

325
00:19:03,960 --> 00:19:07,480
Speaker 4: a brilliant job processing it. But then you're in this situation, well,

326
00:19:07,480 --> 00:19:10,000
Speaker 4: who owns a farm, you know? Is it John Deere

327
00:19:10,119 --> 00:19:11,720
Speaker 4: or is it the farmer? And I think from an

328
00:19:12,000 --> 00:19:18,080
Speaker 4: national economic perspective, thinking about where those solutions are from

329
00:19:18,240 --> 00:19:22,200
Speaker 4: major exporters is a big opportunity for us and again

330
00:19:22,320 --> 00:19:24,639
Speaker 4: something we can lead the world on. And again there

331
00:19:24,640 --> 00:19:29,200
Speaker 4: are open source projects out there. At Multicore World, one

332
00:19:29,200 --> 00:19:32,400
Speaker 4: of the US universities, interesting enough, was talking about an

333
00:19:32,400 --> 00:19:37,919
Speaker 4: open source Agricultural AI project that gives farmers the power

334
00:19:38,400 --> 00:19:42,119
Speaker 4: in Sub Saharan Africa or wherever to analyze their crops,

335
00:19:42,160 --> 00:19:45,280
Speaker 4: their land or whatever and decide on its health or

336
00:19:45,280 --> 00:19:47,719
Speaker 4: what inputs are required and when and so on. So

337
00:19:47,760 --> 00:19:49,840
Speaker 4: those are the sort of opportunities that we should be

338
00:19:49,840 --> 00:19:52,920
Speaker 4: looking to bring to New Zealand and create this sort

339
00:19:52,960 --> 00:19:56,920
Speaker 4: of independent capability rather than just saying no, you can't

340
00:19:56,960 --> 00:20:00,399
Speaker 4: have it because it's too risky. Yeah, or she just

341
00:20:00,480 --> 00:20:02,960
Speaker 4: leave it to John Deere or open AI.

342
00:20:03,200 --> 00:20:05,600
Speaker 1: That's right, it's it's to just leave it to them.

343
00:20:05,880 --> 00:20:08,520
Speaker 1: And that is sort of our default position. I think

344
00:20:08,520 --> 00:20:10,920
Speaker 1: in government procurement, for instance, it's like, well, we've got

345
00:20:11,040 --> 00:20:13,960
Speaker 1: you know, embedded, we're all using Microsoft three sixty five.

346
00:20:14,000 --> 00:20:16,000
Speaker 1: What have you got on the AI front. We'll just

347
00:20:16,080 --> 00:20:19,000
Speaker 1: bundle that into our licensing. That's sort of been the

348
00:20:19,000 --> 00:20:19,840
Speaker 1: way we approach this.

349
00:20:20,160 --> 00:20:22,440
Speaker 4: Well, and then that allows Microsoft to put their fees

350
00:20:22,520 --> 00:20:24,919
Speaker 4: up twenty percent in the year yes or whatever it

351
00:20:25,080 --> 00:20:28,040
Speaker 4: was exactly, you know, So again it makes no sense

352
00:20:28,040 --> 00:20:31,080
Speaker 4: from an economic perspective. And it's not to say that

353
00:20:31,080 --> 00:20:33,000
Speaker 4: their products are not good, it's just to say this

354
00:20:33,119 --> 00:20:37,320
Speaker 4: is stupid. You want to modularize your approach to technology.

355
00:20:37,400 --> 00:20:40,840
Speaker 4: You want to make sure that your your layer interoperability

356
00:20:41,160 --> 00:20:44,600
Speaker 4: using open standards, you know, so you can plug in

357
00:20:44,640 --> 00:20:47,440
Speaker 4: the right solution and you're not just relying on this

358
00:20:47,840 --> 00:20:51,440
Speaker 4: homogeneous stack, which always leaves you with one answer.

359
00:20:51,760 --> 00:20:55,600
Speaker 1: So, if the minister had come out with the strategy

360
00:20:55,920 --> 00:20:59,439
Speaker 1: and said we've got fifty one hundred million dollars earmarked

361
00:20:59,480 --> 00:21:03,280
Speaker 1: for a for the national good, what would you have

362
00:21:03,280 --> 00:21:03,800
Speaker 1: spent it on?

363
00:21:04,000 --> 00:21:08,360
Speaker 4: Well, I've talked about agricultural opportunity. I would look at

364
00:21:08,520 --> 00:21:12,760
Speaker 4: things through the lenses of some of our maori businesses

365
00:21:12,840 --> 00:21:16,320
Speaker 4: and what they're doing, because again that gives New Zealand

366
00:21:16,560 --> 00:21:21,120
Speaker 4: a very unique perspective on technology, and it's a perspective

367
00:21:21,160 --> 00:21:23,239
Speaker 4: that actually the world wants. So if you look at

368
00:21:23,240 --> 00:21:25,280
Speaker 4: what Europe is doing, it's quite similar in a way

369
00:21:25,280 --> 00:21:28,679
Speaker 4: in terms of their concerns. The world is desperate for

370
00:21:28,800 --> 00:21:35,359
Speaker 4: solutions that allow agency and self determination to be sustained.

371
00:21:36,200 --> 00:21:39,479
Speaker 4: And if you can't do that, then your society starts

372
00:21:39,520 --> 00:21:43,199
Speaker 4: losing trust in its institutions, your democracies start failing, and

373
00:21:43,240 --> 00:21:45,879
Speaker 4: all these sort of bad things start happening. And I

374
00:21:45,920 --> 00:21:47,760
Speaker 4: think New Zealand's in a prime place. So those are

375
00:21:47,800 --> 00:21:50,240
Speaker 4: two areas I would look at, and then I would

376
00:21:50,280 --> 00:21:52,760
Speaker 4: look at the companies, and of course, you know, i'd

377
00:21:52,880 --> 00:21:56,120
Speaker 4: put catalysts down as one that can enable these things

378
00:21:56,119 --> 00:22:01,480
Speaker 4: to happen. Catalysts, dragonflies another one. Nicholson's is another one.

379
00:22:02,080 --> 00:22:05,000
Speaker 4: You know, there are lots out there, a surety they've

380
00:22:05,040 --> 00:22:08,000
Speaker 4: brought an AI engine to their own cloud tendancies to

381
00:22:08,680 --> 00:22:12,320
Speaker 4: do a lot more test automation and QA automation. People

382
00:22:12,359 --> 00:22:14,720
Speaker 4: aren't sitting back just with their arms folded in New

383
00:22:14,800 --> 00:22:18,000
Speaker 4: Zealand doing nothing. They're developing stuff. But we need the

384
00:22:18,040 --> 00:22:21,520
Speaker 4: opportunity to sell it and if government isn't willing to

385
00:22:21,600 --> 00:22:25,720
Speaker 4: experiment and to engage with us, then these opportunities won't

386
00:22:25,760 --> 00:22:31,200
Speaker 4: be realized. The Parliamentary Council Office did a great suite

387
00:22:31,240 --> 00:22:33,960
Speaker 4: of proof of concepts at the start of this year

388
00:22:34,200 --> 00:22:36,280
Speaker 4: and we were involved in a couple of them. When

389
00:22:36,320 --> 00:22:40,360
Speaker 4: with Dragonfly, who I mentioned around legislation and using AI

390
00:22:40,400 --> 00:22:45,439
Speaker 4: to summarize legislation, to allow politicians to query legislation to

391
00:22:45,520 --> 00:22:47,000
Speaker 4: determine what it might mean.

392
00:22:47,040 --> 00:22:48,600
Speaker 3: And certainly these are just case studies.

393
00:22:48,680 --> 00:22:51,280
Speaker 4: The more you have government agencies ready to do that

394
00:22:51,400 --> 00:22:55,240
Speaker 4: kind of experimentation or clusters of them, the more opportunity

395
00:22:55,280 --> 00:22:57,760
Speaker 4: they'll be to discover what we can do here in

396
00:22:57,800 --> 00:22:58,760
Speaker 4: New Zealand.

397
00:22:58,480 --> 00:22:59,400
Speaker 2: And it's been a bit slow.

398
00:22:59,560 --> 00:23:03,560
Speaker 1: We had famously we had gov GPT out of Callahan,

399
00:23:04,000 --> 00:23:06,600
Speaker 1: which unfortunately has been shut down now.

400
00:23:06,880 --> 00:23:10,040
Speaker 4: The thing about that as a supplier, it just used

401
00:23:10,119 --> 00:23:12,800
Speaker 4: co Parlat or one of them GPT, I'm not sure

402
00:23:12,800 --> 00:23:15,359
Speaker 4: which one. I couldn't quite understand where they were taking

403
00:23:15,359 --> 00:23:16,080
Speaker 4: that initiative.

404
00:23:16,359 --> 00:23:18,160
Speaker 1: I think it was supposed to be just a customer

405
00:23:18,200 --> 00:23:22,760
Speaker 1: service chatbot if you wanted to navigate government departments. There

406
00:23:22,800 --> 00:23:25,080
Speaker 1: was the in affording government departments, so it was all

407
00:23:25,200 --> 00:23:27,480
Speaker 1: very safe. You weren't going to get hallucinations, you weren't

408
00:23:27,480 --> 00:23:30,320
Speaker 1: going to get inaccurate information. It would serve up exactly

409
00:23:30,560 --> 00:23:33,280
Speaker 1: so something that you would expect now just to be

410
00:23:33,359 --> 00:23:38,280
Speaker 1: built into a customer service delivery for government services. But

411
00:23:38,560 --> 00:23:40,920
Speaker 1: you know, whereas the innovation out of you know, we've

412
00:23:41,119 --> 00:23:44,840
Speaker 1: had the merger of AG Research and land Care Research,

413
00:23:44,960 --> 00:23:49,600
Speaker 1: so there's potential there in this new Biosciences Research division

414
00:23:49,640 --> 00:23:52,639
Speaker 1: that we've got to do some really cool AI stuff

415
00:23:52,680 --> 00:23:57,160
Speaker 1: there in the health sector and very sensitive data, very

416
00:23:57,280 --> 00:24:01,720
Speaker 1: particular issues we're grappling with in New Zealand around public health.

417
00:24:02,480 --> 00:24:05,639
Speaker 1: Where is the real drive and effort to coordinate across

418
00:24:05,680 --> 00:24:08,600
Speaker 1: all these great health tech companies yep, get some momentum

419
00:24:08,640 --> 00:24:12,800
Speaker 1: going the high level stuff yep, tic tick OECD principles,

420
00:24:12,880 --> 00:24:16,240
Speaker 1: great guardrails for government departments. We've sort of done all

421
00:24:16,240 --> 00:24:19,240
Speaker 1: of that, But the actual what's going to drive the

422
00:24:19,320 --> 00:24:21,919
Speaker 1: really innovative stuff that was what was missing.

423
00:24:22,080 --> 00:24:26,800
Speaker 4: It is applying context, local context in local needs and

424
00:24:26,880 --> 00:24:30,520
Speaker 4: not just defaulting to the homogeneous de fault that's coming

425
00:24:30,520 --> 00:24:34,159
Speaker 4: out of the US and the risks of any of

426
00:24:34,320 --> 00:24:38,320
Speaker 4: over reliance on any overseas platforms is always there, but

427
00:24:38,359 --> 00:24:41,160
Speaker 4: it's just got heightened. You know, how many people are

428
00:24:41,160 --> 00:24:44,119
Speaker 4: not traveling through the US because some of their health

429
00:24:44,200 --> 00:24:48,240
Speaker 4: data might have landed in the wrong hands. There are

430
00:24:48,240 --> 00:24:50,240
Speaker 4: states now that will throw you in prison if you've

431
00:24:50,240 --> 00:24:53,840
Speaker 4: had a miscarriage. That doesn't feel like a safe place

432
00:24:54,400 --> 00:24:56,920
Speaker 4: to hold our health data. And yet it's there. This

433
00:24:56,960 --> 00:25:00,440
Speaker 4: isn't so like trying to cry wolf. It's trying to say, well,

434
00:25:00,480 --> 00:25:02,800
Speaker 4: you know, if we addressed that issue and we came

435
00:25:02,880 --> 00:25:05,119
Speaker 4: up with a solution, we could sell that.

436
00:25:05,960 --> 00:25:09,159
Speaker 1: And this is very topical what you say there, because

437
00:25:09,600 --> 00:25:14,240
Speaker 1: the Trump administration has just now some massive project to

438
00:25:15,320 --> 00:25:19,760
Speaker 1: harmonize data across the health sector, including in conjunction with

439
00:25:19,800 --> 00:25:22,560
Speaker 1: the big tech giants. So all that Apple data. Apple

440
00:25:22,560 --> 00:25:26,000
Speaker 1: has been very proactive about gathering data from Apple watches

441
00:25:26,640 --> 00:25:31,200
Speaker 1: in the iPhone itself. It's integrated GPS into that, very

442
00:25:31,320 --> 00:25:34,800
Speaker 1: progressive stuff. All of that data being shared with Medicaid

443
00:25:34,880 --> 00:25:38,120
Speaker 1: and all of that. That is the end goal now

444
00:25:38,119 --> 00:25:41,080
Speaker 1: for the Trump administration. Sounds great in terms of getting

445
00:25:41,080 --> 00:25:44,400
Speaker 1: more visibility into your healthcare, but what are the implications

446
00:25:44,440 --> 00:25:44,720
Speaker 1: of that.

447
00:25:44,840 --> 00:25:47,840
Speaker 4: Well, until the Elon Musk's though Group comes in and

448
00:25:48,280 --> 00:25:51,560
Speaker 4: forces its way into that data, regardless of what legislation

449
00:25:51,760 --> 00:25:54,560
Speaker 4: and international treaties or anything like that said.

450
00:25:55,080 --> 00:25:56,400
Speaker 3: They just forced their way in.

451
00:25:56,920 --> 00:25:59,760
Speaker 1: So the impetus there, I think is as strong to

452
00:25:59,800 --> 00:26:02,880
Speaker 1: do some sort of sovereign AI for these really sensitive things.

453
00:26:03,320 --> 00:26:04,880
Speaker 1: What can we do in the next couple of years,

454
00:26:04,920 --> 00:26:07,879
Speaker 1: do you think, given the trajectory that the government is

455
00:26:07,920 --> 00:26:10,800
Speaker 1: putting us on, what are some wins that we could chase?

456
00:26:11,000 --> 00:26:13,200
Speaker 4: I think do more of these things like the Parliamentary

457
00:26:13,200 --> 00:26:17,160
Speaker 4: Council Office has been doing, and it is thinking about

458
00:26:17,200 --> 00:26:21,000
Speaker 4: the governance and the policy very in principles. That's quite

459
00:26:21,040 --> 00:26:24,400
Speaker 4: straightforward and simple. But think about it. Get them, get

460
00:26:24,440 --> 00:26:26,280
Speaker 4: them out there, get them done. As I said, the

461
00:26:26,320 --> 00:26:29,280
Speaker 4: Marii lens is an enabler apart from it's the right

462
00:26:29,280 --> 00:26:32,240
Speaker 4: thing to do, because Titi, it's actually an enabler and

463
00:26:32,280 --> 00:26:36,719
Speaker 4: it really I think sometimes people don't recognize that it

464
00:26:36,840 --> 00:26:40,760
Speaker 4: makes the job of all New Zealand organizations easier, the

465
00:26:40,800 --> 00:26:43,000
Speaker 4: fact that this part of our society has put so

466
00:26:43,080 --> 00:26:46,440
Speaker 4: much thinking into it because it applies broadly right that

467
00:26:46,520 --> 00:26:49,320
Speaker 4: the thinking and then the work that's been done in

468
00:26:49,359 --> 00:26:52,359
Speaker 4: that space, and it also helps give New Zealand a

469
00:26:52,400 --> 00:26:55,639
Speaker 4: competitive advantage. So those are two things, and the others

470
00:26:56,000 --> 00:27:01,359
Speaker 4: is just invest in support, findingives, find the companies that

471
00:27:01,400 --> 00:27:04,639
Speaker 4: are doing interesting things, create a space for them to

472
00:27:04,640 --> 00:27:10,399
Speaker 4: come into. It's quite hard for New Zealand companies to

473
00:27:10,520 --> 00:27:14,200
Speaker 4: get an individual voice or even a collective voice at

474
00:27:14,200 --> 00:27:16,680
Speaker 4: the table, and it's sometimes hard for the government to

475
00:27:16,720 --> 00:27:19,560
Speaker 4: reach them because there's so many of us and relative

476
00:27:19,640 --> 00:27:23,679
Speaker 4: to you know us type companies, we're small and we

477
00:27:23,720 --> 00:27:26,800
Speaker 4: don't have their lobbying powers and buckets.

478
00:27:26,960 --> 00:27:28,320
Speaker 2: Has the government listened to you?

479
00:27:28,359 --> 00:27:32,000
Speaker 1: Have they sought your advice on the National AI strategy,

480
00:27:32,040 --> 00:27:32,960
Speaker 1: for instance.

481
00:27:32,720 --> 00:27:34,280
Speaker 3: Not before it was published.

482
00:27:34,320 --> 00:27:36,560
Speaker 4: On the other hand, I've been fortunate enough to have

483
00:27:36,600 --> 00:27:39,440
Speaker 4: a meeting with the Minister sing Riti recently and some

484
00:27:39,520 --> 00:27:41,480
Speaker 4: of his team, which is fantastic and.

485
00:27:41,600 --> 00:27:44,679
Speaker 3: Like it was very, very engaged in the topic.

486
00:27:45,080 --> 00:27:48,000
Speaker 4: The other thing I would encourage New Zealand businesses like

487
00:27:48,119 --> 00:27:52,600
Speaker 4: us to do is to leverage the fact of you know,

488
00:27:52,720 --> 00:27:56,040
Speaker 4: the degrees of separation in New Zealand being very small,

489
00:27:56,920 --> 00:28:01,800
Speaker 4: and keep using the contacts that you might have with politicians,

490
00:28:01,840 --> 00:28:06,000
Speaker 4: with government people within your communities to build that New

491
00:28:06,080 --> 00:28:06,840
Speaker 4: Zealand narrative.

492
00:28:06,920 --> 00:28:11,000
Speaker 1: Okay, well, and what's ahead for catalysts in the AI space.

493
00:28:11,040 --> 00:28:13,359
Speaker 1: You're working with Thelexa ta Hiku media and that some

494
00:28:13,440 --> 00:28:18,639
Speaker 1: innovative stuff there. Are you seeing more demand for applications?

495
00:28:18,640 --> 00:28:21,960
Speaker 1: For instance, the whole AI agentic AI thing is that

496
00:28:22,080 --> 00:28:22,960
Speaker 1: on your radar.

497
00:28:23,160 --> 00:28:25,560
Speaker 4: Various aspects of it at all on our radia of course,

498
00:28:25,600 --> 00:28:28,840
Speaker 4: because you know, in order to stay relevant to our clients,

499
00:28:28,840 --> 00:28:31,879
Speaker 4: to our staff, to our country, we have to be

500
00:28:31,920 --> 00:28:35,000
Speaker 4: on top of everything. I think for us, we're just

501
00:28:35,040 --> 00:28:36,760
Speaker 4: having a lot of fun with the fact that we

502
00:28:36,800 --> 00:28:41,080
Speaker 4: have a cloud company and so we can kind of

503
00:28:41,120 --> 00:28:44,360
Speaker 4: go to town on a bunch of technologies, and you know,

504
00:28:44,640 --> 00:28:48,360
Speaker 4: we're pushing capability into some of our product sets, which

505
00:28:48,360 --> 00:28:52,480
Speaker 4: are largely in the education space, but also trying to

506
00:28:52,680 --> 00:28:57,680
Speaker 4: find ways of working with our clients to help them

507
00:28:58,120 --> 00:29:01,160
Speaker 4: realize their own product sets and outputs and so on.

508
00:29:01,320 --> 00:29:02,200
Speaker 3: And it's a lot of fun.

509
00:29:02,440 --> 00:29:04,720
Speaker 4: One of the really good bits of advice I got

510
00:29:04,720 --> 00:29:07,800
Speaker 4: from another CEO and another digital company was to build

511
00:29:07,840 --> 00:29:10,880
Speaker 4: an AI practice group in the company right before you

512
00:29:11,000 --> 00:29:15,000
Speaker 4: even started rushing into it. And we did that not

513
00:29:15,160 --> 00:29:19,040
Speaker 4: just with engineers, but with project managers, with designers, you know,

514
00:29:19,120 --> 00:29:22,920
Speaker 4: so across the company to play with tools to play

515
00:29:22,920 --> 00:29:25,240
Speaker 4: with them in a safe environment to you know, kind

516
00:29:25,240 --> 00:29:30,160
Speaker 4: of develop principles, and that has driven understanding and acceptance

517
00:29:30,360 --> 00:29:33,400
Speaker 4: and it's been incredibly useful and the people that have

518
00:29:33,440 --> 00:29:36,880
Speaker 4: been involved have been great. So that's an approach companies

519
00:29:36,920 --> 00:29:39,440
Speaker 4: can take. I set up something called the AI Discovery

520
00:29:39,520 --> 00:29:42,360
Speaker 4: Lab which people can register for, and it just does

521
00:29:42,400 --> 00:29:43,160
Speaker 4: a little bit what you.

522
00:29:43,120 --> 00:29:44,400
Speaker 3: Were saying you do.

523
00:29:44,640 --> 00:29:47,480
Speaker 4: It allows you to sort of use a couple of

524
00:29:47,760 --> 00:29:50,920
Speaker 4: language models on Catalyst Cloud, put in a prompt, get

525
00:29:50,920 --> 00:29:53,680
Speaker 4: two answers back, and then reflect on those answers and

526
00:29:53,720 --> 00:29:56,440
Speaker 4: reflect on the fact that if those answers aren't perfect,

527
00:29:56,480 --> 00:29:59,760
Speaker 4: surely that means that a human with some capability and

528
00:29:59,800 --> 00:30:06,200
Speaker 4: not needs to be involved in checking them, in creating

529
00:30:06,200 --> 00:30:08,960
Speaker 4: a narrative that absolutely suits what they're trying to do,

530
00:30:09,240 --> 00:30:11,800
Speaker 4: and so on. So it's kind of about reflection and

531
00:30:11,920 --> 00:30:13,200
Speaker 4: agency and so on.

532
00:30:13,240 --> 00:30:15,800
Speaker 1: What exactly is involved when you take a large language

533
00:30:15,800 --> 00:30:19,480
Speaker 1: model and as they say, fine tune it for your needs?

534
00:30:19,680 --> 00:30:20,440
Speaker 2: Is that quite.

535
00:30:20,200 --> 00:30:25,200
Speaker 1: A there's process in computer processing involved in that, but

536
00:30:25,200 --> 00:30:27,480
Speaker 1: there's also human oversight of that as well. Right, Is

537
00:30:27,480 --> 00:30:29,920
Speaker 1: that a complicated process to undertake?

538
00:30:30,440 --> 00:30:32,840
Speaker 4: What I know is that you can basically bring your

539
00:30:33,080 --> 00:30:37,600
Speaker 4: data sets and use tools called rags and there are

540
00:30:37,640 --> 00:30:41,240
Speaker 4: other tools as well and sort of add that training

541
00:30:41,280 --> 00:30:44,000
Speaker 4: and context to the large language model, and that gives

542
00:30:44,040 --> 00:30:47,960
Speaker 4: you a much more accurate response to specific query. So

543
00:30:48,080 --> 00:30:51,640
Speaker 4: often those models fail when you get very contextual and

544
00:30:51,760 --> 00:30:55,720
Speaker 4: very specific. You know, according to chat GPT a year ago,

545
00:30:55,800 --> 00:30:58,200
Speaker 4: I was a great musician, got a few other things right,

546
00:30:58,240 --> 00:31:01,480
Speaker 4: but it was wrong about that assure you. So when

547
00:31:01,520 --> 00:31:04,080
Speaker 4: you get down into those sort of contextual things and things.

548
00:31:04,280 --> 00:31:06,160
Speaker 4: But there are other ways of doing that as well,

549
00:31:06,160 --> 00:31:10,520
Speaker 4: with small language models and so on, potentially more effective

550
00:31:10,640 --> 00:31:12,920
Speaker 4: in the long run in terms of how costly it

551
00:31:12,960 --> 00:31:15,080
Speaker 4: is to keep retraining them and refreshing them.

552
00:31:15,160 --> 00:31:19,520
Speaker 1: Okay, well, it sounds like you're talking to the minister,

553
00:31:19,560 --> 00:31:21,960
Speaker 1: which is great as one of the big open source

554
00:31:22,200 --> 00:31:25,880
Speaker 1: platforms in the country. That's excellent to see in terms

555
00:31:25,880 --> 00:31:30,080
Speaker 1: of the pace of development. It seems as though there's

556
00:31:30,200 --> 00:31:32,760
Speaker 1: some discussion about whether we've hit some HRD limits on

557
00:31:32,760 --> 00:31:35,160
Speaker 1: on AI. And you look at the money that's still

558
00:31:35,200 --> 00:31:38,800
Speaker 1: going in to this, you know, billions Microsoft and it's

559
00:31:38,880 --> 00:31:42,280
Speaker 1: quarterly results recently up to spend again for the next

560
00:31:42,400 --> 00:31:44,440
Speaker 1: quarter thirty billion in one quarter.

561
00:31:44,600 --> 00:31:45,360
Speaker 2: One company.

562
00:31:46,120 --> 00:31:49,360
Speaker 1: Is this a bubble or is this a genuine effort

563
00:31:49,440 --> 00:31:51,880
Speaker 1: to build the infrastructure of the future.

564
00:31:52,120 --> 00:31:55,040
Speaker 4: It's I think it's a bit of both. It's possibly

565
00:31:55,080 --> 00:31:58,720
Speaker 4: more bubble. What those companies are facing is a shortage

566
00:31:58,720 --> 00:32:01,840
Speaker 4: of what they call tokens in other words, words, they've

567
00:32:01,880 --> 00:32:05,080
Speaker 4: run out. That's why they're stealing people's data. That's why

568
00:32:05,120 --> 00:32:09,120
Speaker 4: they're trying to push into knowledge sets and data sets

569
00:32:09,120 --> 00:32:11,560
Speaker 4: that they have no rights to push into. That's why

570
00:32:11,560 --> 00:32:14,800
Speaker 4: the in breach of copyright all these things, right, because

571
00:32:14,840 --> 00:32:17,840
Speaker 4: they've run out of data to train their algorithms on.

572
00:32:18,440 --> 00:32:22,280
Speaker 4: Your data is far more valuable to those AI companies

573
00:32:22,320 --> 00:32:25,560
Speaker 4: than it ever has been in the past. That's why

574
00:32:25,600 --> 00:32:28,640
Speaker 4: they're pushing into your phones. You don't have the choice

575
00:32:28,680 --> 00:32:30,640
Speaker 4: of whether or not to switch the AI off and

576
00:32:30,760 --> 00:32:34,680
Speaker 4: WhatsApp because they're basically trying to collect data to retrain

577
00:32:34,760 --> 00:32:38,960
Speaker 4: their models. I think that's pretty unethical and evil, and

578
00:32:39,000 --> 00:32:41,880
Speaker 4: that's why they're hitting limits. Yeah, it's nutty as well,

579
00:32:41,920 --> 00:32:43,880
Speaker 4: because you know, you sort of think, well, how much

580
00:32:43,880 --> 00:32:46,640
Speaker 4: more do we need of those kind of models when

581
00:32:46,680 --> 00:32:51,880
Speaker 4: there's all these other opportunities to be exploring, you know,

582
00:32:52,080 --> 00:32:56,000
Speaker 4: using those sort of technologies. Yeah, so yeah, I don't know.

583
00:32:56,080 --> 00:32:59,080
Speaker 1: And the other constraint obviously they're coming up against is talent.

584
00:32:59,280 --> 00:33:02,200
Speaker 2: You know, when willing to spend two hundred and fifty.

585
00:33:02,040 --> 00:33:06,240
Speaker 1: Million dollars on one, albeit very talented, twenty four year

586
00:33:06,280 --> 00:33:10,959
Speaker 1: old AI developer, they're clearly desperate for smart people who

587
00:33:11,000 --> 00:33:13,040
Speaker 1: are going to lead to the next big shifts.

588
00:33:13,560 --> 00:33:14,440
Speaker 3: I don't understand that.

589
00:33:14,520 --> 00:33:16,720
Speaker 4: I think the other thing just to keep in mind

590
00:33:16,840 --> 00:33:20,000
Speaker 4: is that GPU and you look at in videos market

591
00:33:20,080 --> 00:33:24,560
Speaker 4: cap it's it's stupendous at the moment. But GPU technology

592
00:33:24,600 --> 00:33:27,600
Speaker 4: is old technology. There's equally old technology.

593
00:33:28,000 --> 00:33:28,160
Speaker 3: You know.

594
00:33:28,200 --> 00:33:30,480
Speaker 4: They are current to GPUs that runs in your phones

595
00:33:30,720 --> 00:33:33,320
Speaker 4: that we were using on the SKA project to stream

596
00:33:33,400 --> 00:33:36,880
Speaker 4: data out of telescopes and to basically do some very

597
00:33:36,960 --> 00:33:41,640
Speaker 4: quick data analysis on and so low powered parallel processes

598
00:33:42,360 --> 00:33:45,240
Speaker 4: I think are going to be the future of AI.

599
00:33:45,600 --> 00:33:50,040
Speaker 4: Everyone at the moment is buying GPUs, which are expensive

600
00:33:50,240 --> 00:33:54,560
Speaker 4: in many ways, they're terrible for our planet in terms

601
00:33:54,560 --> 00:33:57,680
Speaker 4: of their energy usage. And everyone is doubling down on

602
00:33:57,720 --> 00:34:01,160
Speaker 4: this GPU space, and the people that get the breakthroughs

603
00:34:01,160 --> 00:34:04,440
Speaker 4: on low power parallel processes are going to win.

604
00:34:04,800 --> 00:34:07,960
Speaker 1: And we do have some smart people on parallel processing

605
00:34:07,960 --> 00:34:10,880
Speaker 1: here in New Zealand as well. Yeah, maybe that's the

606
00:34:10,960 --> 00:34:12,960
Speaker 1: other if we are going to do anything hardware related,

607
00:34:13,000 --> 00:34:16,040
Speaker 1: it's what is the alternative to GPUs.

608
00:34:15,680 --> 00:34:18,480
Speaker 3: Yep, completely Yeah.

609
00:34:17,719 --> 00:34:21,480
Speaker 1: Well thanks Don as always a very enlightening conversation. Good

610
00:34:21,560 --> 00:34:25,759
Speaker 1: luck with things at Catalyst, and hopefully you'll have some

611
00:34:26,160 --> 00:34:28,120
Speaker 1: input into these plans.

612
00:34:28,120 --> 00:34:29,960
Speaker 2: It is a starting point, as in the marriages.

613
00:34:30,360 --> 00:34:32,799
Speaker 1: I think there's a lot of scope for improvements and

614
00:34:32,840 --> 00:34:33,600
Speaker 1: expanding on it.

615
00:34:33,800 --> 00:34:36,959
Speaker 4: Yeah, and Peter, I hope this makes the recording. Thank

616
00:34:37,000 --> 00:34:41,480
Speaker 4: you you doing God's work with all this, you know

617
00:34:41,560 --> 00:34:43,600
Speaker 4: work that you're doing the technology space for New Zealand.

618
00:34:43,680 --> 00:34:45,640
Speaker 1: So thank you, Thank you, and thanks to everyone out

619
00:34:45,640 --> 00:34:48,440
Speaker 1: there for listening to these are important issues close to

620
00:34:48,640 --> 00:34:49,360
Speaker 1: light anyway.

621
00:34:49,719 --> 00:34:50,200
Speaker 2: Thanks Don.

622
00:34:55,960 --> 00:34:59,239
Speaker 1: Thanks to Don Christy for candid, practical tour of what

623
00:34:59,440 --> 00:35:03,000
Speaker 1: sovereign can look like when it's grounded in open source,

624
00:35:03,320 --> 00:35:07,799
Speaker 1: clear governance and locally controlled infrastructure. A few takeaways from

625
00:35:07,800 --> 00:35:12,440
Speaker 1: me major led data governance principles are an enabler for

626
00:35:12,520 --> 00:35:16,040
Speaker 1: trustworthy systems. A lot of work has been done in

627
00:35:16,040 --> 00:35:19,120
Speaker 1: the space where world leading in it. We can apply

628
00:35:19,440 --> 00:35:24,080
Speaker 1: them to trustworthy AI in conjunction with open source efforts

629
00:35:24,160 --> 00:35:27,640
Speaker 1: like what the Swiss are doing. Sovereignty is a design choice,

630
00:35:28,040 --> 00:35:33,120
Speaker 1: not a dollar amount. Federated open models give us leverage

631
00:35:33,160 --> 00:35:38,760
Speaker 1: and procurement that prizes interoperability over commercial bundles. Don rightly

632
00:35:38,840 --> 00:35:42,439
Speaker 1: points to AI in agriculture being an area of opportunity

633
00:35:42,480 --> 00:35:45,480
Speaker 1: for us. Just look at Halter's use of AI in

634
00:35:45,520 --> 00:35:49,640
Speaker 1: managing cowherds, and there's a need to keep experimenting in

635
00:35:49,719 --> 00:35:53,080
Speaker 1: government and in the private sector with open source AI

636
00:35:53,200 --> 00:35:55,440
Speaker 1: at the heart of it. That's it for this week's

637
00:35:55,480 --> 00:35:59,560
Speaker 1: Business of Tech. If this sparked ideas, share the episode

638
00:35:59,600 --> 00:36:03,760
Speaker 1: with a COG and government, a founder building with open models,

639
00:36:03,840 --> 00:36:07,120
Speaker 1: or a CIO rethinking their AI stack. And if you're

640
00:36:07,160 --> 00:36:10,719
Speaker 1: working on a New Zealand sovereign AI pilot, especially in

641
00:36:10,800 --> 00:36:13,080
Speaker 1: health or agriculture, I'd love to hear about it for

642
00:36:13,080 --> 00:36:16,480
Speaker 1: a future show. I'm Peter Griffin. Thanks so much for listening.

643
00:36:16,520 --> 00:36:19,200
Speaker 1: I'll catch you next week with another episode of the

644
00:36:19,239 --> 00:36:29,919
Speaker 1: Business of Tech Matowa