1
00:00:03,560 --> 00:00:06,840
Speaker 1: Within five years, open source AI will have raised the

2
00:00:06,920 --> 00:00:11,040
Speaker 1: GDP in the world's poorest countries. That is a premise

3
00:00:11,080 --> 00:00:16,720
Speaker 1: for today's conversation. I'm Asimazar. Welcome to the Exponentially podcast.

4
00:00:19,120 --> 00:00:22,960
Speaker 1: Recent reports from the banks and consultancies have suggested that

5
00:00:23,320 --> 00:00:27,520
Speaker 1: new breakthroughs in AI, particularly generative AI, could add trillions

6
00:00:27,560 --> 00:00:31,040
Speaker 1: of dollars to global GDP. But the most advanced of

7
00:00:31,080 --> 00:00:34,640
Speaker 1: these models are built in the West, on expensive supercomputers

8
00:00:34,920 --> 00:00:38,959
Speaker 1: and trained using English language data sets. So how can

9
00:00:39,000 --> 00:00:42,760
Speaker 1: the global South, with their young populations, tighter finances and

10
00:00:42,840 --> 00:00:48,600
Speaker 1: shake your infrastructure share in this productivity boom. Today's guest

11
00:00:48,680 --> 00:00:52,680
Speaker 1: is m Ad Mustark, a Jordanian born Bangladeshi immigrant to

12
00:00:52,760 --> 00:00:56,680
Speaker 1: Britain who's a founder and CEO of Stability AI. It's

13
00:00:56,720 --> 00:00:59,400
Speaker 1: a firm that has accelerated to a billion dollar valuation

14
00:00:59,560 --> 00:01:03,240
Speaker 1: in less than three years. Emad has been vocal about

15
00:01:03,280 --> 00:01:06,520
Speaker 1: the potential that open source AI offers to the poorest

16
00:01:06,520 --> 00:01:09,720
Speaker 1: in the world, but there have been serious criticisms leveled

17
00:01:09,720 --> 00:01:13,560
Speaker 1: against him, most prominently in a recent story by Forbes magazine.

18
00:01:13,680 --> 00:01:17,200
Speaker 1: These include questions about taking credit for the company's technology,

19
00:01:17,400 --> 00:01:19,520
Speaker 1: about some of the partnerships the firm is meant to

20
00:01:19,560 --> 00:01:24,280
Speaker 1: have and also governance practices at Stability AI, where Mustak

21
00:01:24,400 --> 00:01:29,680
Speaker 1: remains CEO. MD has publicly rejected these as grossness characterizations,

22
00:01:29,880 --> 00:01:32,520
Speaker 1: and a debate over the story and others critical of

23
00:01:32,520 --> 00:01:36,400
Speaker 1: Stability AI continues to play out on social media and elsewhere.

24
00:01:37,319 --> 00:01:39,640
Speaker 1: But as long as investors who have poured hundreds of

25
00:01:39,640 --> 00:01:42,880
Speaker 1: millions of dollars into Stability AI continue to stick by

26
00:01:42,920 --> 00:01:45,360
Speaker 1: a mad and as long as a firm continues to

27
00:01:45,480 --> 00:01:48,840
Speaker 1: innovate and bring out new AI models, he will remain

28
00:01:49,160 --> 00:01:58,640
Speaker 1: a powerful force in the expanding AI universe. What is

29
00:01:58,880 --> 00:02:02,640
Speaker 1: generative AI and how is it different from all of

30
00:02:02,640 --> 00:02:04,800
Speaker 1: the AI systems that we saw previously.

31
00:02:05,240 --> 00:02:07,640
Speaker 2: Generator AI is a new type of AI that started

32
00:02:07,680 --> 00:02:10,560
Speaker 2: in twenty seventeen. There was a seminal paper called Attention

33
00:02:10,720 --> 00:02:13,320
Speaker 2: is all you Need because not all data is the same.

34
00:02:13,360 --> 00:02:16,320
Speaker 2: You pay attention to what's important. So classical AI was

35
00:02:16,320 --> 00:02:18,400
Speaker 2: built on this concept of big data. So you had

36
00:02:18,440 --> 00:02:21,399
Speaker 2: all this data of Facebook and Google and they used

37
00:02:21,440 --> 00:02:24,200
Speaker 2: it to sell you coconut shampoo, but it couldn't go

38
00:02:24,280 --> 00:02:27,320
Speaker 2: outside its boundaries. So it's like a very logical The

39
00:02:27,360 --> 00:02:30,400
Speaker 2: future is like the past, stable kind of environment this

40
00:02:30,480 --> 00:02:32,679
Speaker 2: new type of AI said pay attention to the important

41
00:02:32,720 --> 00:02:35,600
Speaker 2: parts of data to compress it. So as people listening

42
00:02:35,639 --> 00:02:37,760
Speaker 2: to this, we'll see they might take away a few points,

43
00:02:37,760 --> 00:02:40,640
Speaker 2: they're not going to remember our entire conversation. That's what

44
00:02:40,720 --> 00:02:42,880
Speaker 2: human mind does. You've got the very logical part that

45
00:02:42,960 --> 00:02:46,880
Speaker 2: can memorize stuff, and you've got the part that builds principles, stories,

46
00:02:47,360 --> 00:02:48,720
Speaker 2: frameworks for understanding.

47
00:02:49,280 --> 00:02:51,920
Speaker 1: And this paper attention is all you need in a

48
00:02:51,960 --> 00:02:55,679
Speaker 1: sense analogize that for a computer system exactly.

49
00:02:56,000 --> 00:02:57,560
Speaker 2: It was the first one that said, this is how

50
00:02:57,600 --> 00:03:01,160
Speaker 2: you show it at scale, and let's simplify it down

51
00:03:01,240 --> 00:03:05,440
Speaker 2: to a problem of better data and bigger computers. So

52
00:03:05,560 --> 00:03:08,960
Speaker 2: using gigantic supercomputers, you can take these big data sets

53
00:03:08,960 --> 00:03:11,440
Speaker 2: of text, images and others and can press it down

54
00:03:11,480 --> 00:03:15,639
Speaker 2: to just a few gigabytes of file that learns principles,

55
00:03:15,680 --> 00:03:16,359
Speaker 2: not facts.

56
00:03:16,880 --> 00:03:18,120
Speaker 3: And so this was the missing.

57
00:03:17,880 --> 00:03:21,079
Speaker 2: Piece in AI, and that's why using these systems feels

58
00:03:21,080 --> 00:03:23,360
Speaker 2: actually quite surprisingly human.

59
00:03:23,200 --> 00:03:26,000
Speaker 1: Surprisingly human. But how do we get to the generative

60
00:03:26,000 --> 00:03:26,720
Speaker 1: part of all of that.

61
00:03:26,840 --> 00:03:30,080
Speaker 2: The generative part is that you put a prompt in

62
00:03:30,240 --> 00:03:32,400
Speaker 2: or some words in, and then it gives you something back.

63
00:03:32,440 --> 00:03:36,040
Speaker 2: It generates the outputs and the outputs are not always

64
00:03:36,080 --> 00:03:38,800
Speaker 2: even the same because it has principles as a base.

65
00:03:39,000 --> 00:03:41,840
Speaker 1: So in the same sense that if you and I

66
00:03:41,920 --> 00:03:46,200
Speaker 1: meet one day, the way the conversation plays out could

67
00:03:46,200 --> 00:03:49,119
Speaker 1: be quite different to the next time we meet, because

68
00:03:49,160 --> 00:03:53,120
Speaker 1: we have principles of socialization and of behavior and of

69
00:03:53,160 --> 00:03:56,080
Speaker 1: how well we know each other, and those get applied

70
00:03:57,080 --> 00:03:59,400
Speaker 1: in real time at that moment each time we shake hands.

71
00:03:59,480 --> 00:04:02,400
Speaker 2: In real time, it's a file with just a bunch

72
00:04:02,440 --> 00:04:04,240
Speaker 2: of like it's called neural nets.

73
00:04:04,240 --> 00:04:04,600
Speaker 3: Weights.

74
00:04:05,000 --> 00:04:07,920
Speaker 2: Words go in and get shaken out like a pinball

75
00:04:07,920 --> 00:04:10,400
Speaker 2: slot machine, and then the output comes. But the output

76
00:04:10,440 --> 00:04:13,160
Speaker 2: input can be a painting of a cup in the

77
00:04:13,200 --> 00:04:16,240
Speaker 2: style of Vermere, and then it understands the nature of

78
00:04:16,279 --> 00:04:19,760
Speaker 2: painting cup Vermire. But cup has so many different meanings,

79
00:04:19,839 --> 00:04:21,640
Speaker 2: can mean this cup or cup your hands, or cup

80
00:04:21,680 --> 00:04:24,320
Speaker 2: your ears, or a world cup, and it understands those

81
00:04:24,320 --> 00:04:29,839
Speaker 2: things in place because it's been trained on images and text. Similarly,

82
00:04:29,920 --> 00:04:32,640
Speaker 2: a lot of these language models they've been trained on sentences,

83
00:04:33,000 --> 00:04:35,160
Speaker 2: so they look at the context of the sentence and

84
00:04:35,200 --> 00:04:38,520
Speaker 2: they say, what's coming next, you know, like that game

85
00:04:38,560 --> 00:04:41,200
Speaker 2: of improvisation where you start with the sentence and then

86
00:04:41,240 --> 00:04:45,160
Speaker 2: you provide something in the next one, and on we go.

87
00:04:45,240 --> 00:04:48,159
Speaker 1: And it's moving so so quickly. So last summer I

88
00:04:48,400 --> 00:04:52,480
Speaker 1: was away in Tanzania on a safari and there's really

89
00:04:52,480 --> 00:04:55,120
Speaker 1: no mobile signal. We were away from about two and

90
00:04:55,120 --> 00:04:58,200
Speaker 1: a half weeks and when I came back, the Internet

91
00:04:58,480 --> 00:05:03,440
Speaker 1: was full of stable diffusion, stable diffusion, stable difusion, stable diffusion.

92
00:05:03,680 --> 00:05:06,680
Speaker 1: And that is one of your your generative products. What

93
00:05:06,920 --> 00:05:10,839
Speaker 1: it does it takes text and it produces images. So

94
00:05:10,880 --> 00:05:16,240
Speaker 1: I can say I want an image of a badger

95
00:05:16,560 --> 00:05:21,480
Speaker 1: playing football on a bicycle, and stable difusion will will

96
00:05:21,480 --> 00:05:24,039
Speaker 1: produce that image. It could also produce more useful, commercially

97
00:05:24,160 --> 00:05:29,400
Speaker 1: interesting images. Recently, you've brought out a new generative product

98
00:05:29,480 --> 00:05:34,839
Speaker 1: which is stable LM, which looks a lot like the

99
00:05:34,839 --> 00:05:37,520
Speaker 1: these text models that we've seen are running around totage.

100
00:05:37,520 --> 00:05:40,200
Speaker 2: Stable lemmas are language model suites, and what we do

101
00:05:40,279 --> 00:05:42,240
Speaker 2: is a bit different from a lot of the other companies,

102
00:05:42,279 --> 00:05:44,880
Speaker 2: and that a lot of the focus and the breakthroughs

103
00:05:44,880 --> 00:05:46,880
Speaker 2: were because a lot of these research labs, the open

104
00:05:47,000 --> 00:05:49,520
Speaker 2: eyes and thropics of the world deep minds, have this

105
00:05:49,600 --> 00:05:52,720
Speaker 2: focus on AGI artificial general intelligence.

106
00:05:52,720 --> 00:05:54,160
Speaker 3: Can you be an AI that can do anything?

107
00:05:54,360 --> 00:05:54,600
Speaker 1: Yes?

108
00:05:54,760 --> 00:05:56,720
Speaker 3: It turns out maybe you can so.

109
00:05:57,000 --> 00:05:59,120
Speaker 1: Making a machine that's in our own image in a sense,

110
00:05:59,120 --> 00:06:00,560
Speaker 1: that's what AGI sound.

111
00:06:00,680 --> 00:06:04,200
Speaker 2: More than that, it's a general intelligence. So it's the

112
00:06:04,880 --> 00:06:06,599
Speaker 2: kid that made you look bad at school because he

113
00:06:06,680 --> 00:06:09,800
Speaker 2: was good at everything. You know, the top performer like

114
00:06:10,200 --> 00:06:13,040
Speaker 2: GPT four now can pass the bar exam, the medical

115
00:06:13,080 --> 00:06:16,279
Speaker 2: licensing exam, the GIRE. It's probably going to Stanford, you know,

116
00:06:16,520 --> 00:06:21,119
Speaker 2: and next floor. But then our take was that's great,

117
00:06:21,279 --> 00:06:24,800
Speaker 2: you can have these amazing giant general models. What will

118
00:06:24,880 --> 00:06:27,880
Speaker 2: be better is to mimic humanity where our companies are

119
00:06:27,880 --> 00:06:30,960
Speaker 2: not one generalist doing everything. What if we made it

120
00:06:31,000 --> 00:06:33,599
Speaker 2: so you could bring your own data to the models

121
00:06:33,640 --> 00:06:36,360
Speaker 2: and have lots of specialists working together and have those

122
00:06:36,360 --> 00:06:37,719
Speaker 2: models working for you that you own.

123
00:06:37,960 --> 00:06:38,720
Speaker 3: Right, What if you.

124
00:06:38,680 --> 00:06:41,840
Speaker 2: Had open data, open models and allowed it to be

125
00:06:41,920 --> 00:06:44,880
Speaker 2: customized and specialized, so rather than relying on one models

126
00:06:44,880 --> 00:06:48,240
Speaker 2: to do everything, instead you optimize them.

127
00:06:48,600 --> 00:06:48,800
Speaker 3: Right.

128
00:06:49,000 --> 00:06:50,960
Speaker 1: So, but the idea is still the same. The idea

129
00:06:51,040 --> 00:06:54,919
Speaker 1: is still constructing a model that is generative. I can

130
00:06:55,000 --> 00:06:57,320
Speaker 1: give it some text and it can produce something that

131
00:06:57,440 --> 00:07:02,200
Speaker 1: is commercially useful or in emotionally useful. So working software

132
00:07:02,240 --> 00:07:06,359
Speaker 1: code or a working invitation to a meeting, or you know,

133
00:07:06,440 --> 00:07:08,960
Speaker 1: things that will save us time and help us perhaps

134
00:07:09,000 --> 00:07:10,360
Speaker 1: be a bit more creative.

135
00:07:10,480 --> 00:07:12,640
Speaker 2: So again, like a talented graduate, and they can learn

136
00:07:12,760 --> 00:07:14,960
Speaker 2: very quickly from a few examples. So they have this

137
00:07:15,040 --> 00:07:18,000
Speaker 2: base of generalized knowledge. They've been through kindergarten, high school,

138
00:07:18,040 --> 00:07:21,720
Speaker 2: and university, but they're not specialized yet. You can train

139
00:07:21,800 --> 00:07:23,680
Speaker 2: them yourself, or you can just show them some examples

140
00:07:23,720 --> 00:07:27,000
Speaker 2: and they learn very quickly. Unlike classical AI models that

141
00:07:27,040 --> 00:07:29,040
Speaker 2: you had to train on the whole data set.

142
00:07:29,240 --> 00:07:33,040
Speaker 3: They weren't good at at aaptation like the badger riding

143
00:07:33,080 --> 00:07:33,440
Speaker 3: a bike.

144
00:07:33,920 --> 00:07:36,920
Speaker 2: That's not really a normal thing, you know, And so

145
00:07:37,120 --> 00:07:39,920
Speaker 2: that has been the province of well, just us right.

146
00:07:39,960 --> 00:07:41,880
Speaker 2: We were able to take these concepts together, whereas a

147
00:07:41,880 --> 00:07:45,600
Speaker 2: computer could never merge together concepts. Now we've got that

148
00:07:45,720 --> 00:07:49,360
Speaker 2: missing link of being able to take concepts, merge them together,

149
00:07:49,880 --> 00:07:58,480
Speaker 2: understand some of these hidden meanings.

150
00:07:59,480 --> 00:08:03,440
Speaker 1: I'm curier about what you think the economic impact of

151
00:08:03,480 --> 00:08:05,440
Speaker 1: all of this will be. I mean, there have been

152
00:08:05,800 --> 00:08:09,520
Speaker 1: any number of papers coming out from the investment banks

153
00:08:09,520 --> 00:08:12,640
Speaker 1: and the research houses and economists in the last few months.

154
00:08:12,680 --> 00:08:16,360
Speaker 1: I think Golden Sas had a report that suggested that

155
00:08:16,360 --> 00:08:20,240
Speaker 1: that with one of their scenarios, they could see fast

156
00:08:20,280 --> 00:08:23,520
Speaker 1: implementation of generative AI across the world, leading to a

157
00:08:23,560 --> 00:08:27,600
Speaker 1: seven percent increase in global GDP in about ten years.

158
00:08:28,240 --> 00:08:30,240
Speaker 1: What's your sense of what it could do economically.

159
00:08:30,320 --> 00:08:33,280
Speaker 2: I think it's the biggest thing since the Gutenberg Press,

160
00:08:33,320 --> 00:08:34,439
Speaker 2: maybe even fire.

161
00:08:34,559 --> 00:08:38,200
Speaker 1: The Gutenberg Press six hundred years ago, or fire two

162
00:08:38,280 --> 00:08:39,040
Speaker 1: million years ago.

163
00:08:39,120 --> 00:08:42,400
Speaker 2: I think that humans are driven by stories. It's what

164
00:08:42,480 --> 00:08:45,840
Speaker 2: allowed us to form tribes and then money and things

165
00:08:45,880 --> 00:08:48,040
Speaker 2: like that are stories. The press allowed us to write

166
00:08:48,040 --> 00:08:50,760
Speaker 2: down the stories. But it's very lossy. Again, you me

167
00:08:50,960 --> 00:08:54,120
Speaker 2: everyone listening like you're looking through your things right now.

168
00:08:54,600 --> 00:08:57,319
Speaker 2: We do power points, we write things, but it doesn't

169
00:08:57,360 --> 00:09:02,320
Speaker 2: capture the richness of humanity. Our organizations are built on

170
00:09:02,440 --> 00:09:05,440
Speaker 2: layers of text, which is painful, and that's why it

171
00:09:05,520 --> 00:09:08,200
Speaker 2: turns us into cogs and the machine, shall we say.

172
00:09:08,240 --> 00:09:10,200
Speaker 1: I mean there's something that's quite powerful about the models.

173
00:09:10,240 --> 00:09:11,960
Speaker 1: I think that you are getting to here, which is

174
00:09:11,960 --> 00:09:16,360
Speaker 1: that you can feed them text and through that the

175
00:09:16,400 --> 00:09:21,760
Speaker 1: machination of the billions of different switches and cogs you

176
00:09:21,800 --> 00:09:25,319
Speaker 1: guys call them parameters. Yes, in the system, it starts

177
00:09:25,360 --> 00:09:31,800
Speaker 1: to find those underlying relationships that we know, probably deeply

178
00:09:31,800 --> 00:09:34,280
Speaker 1: in our brains, but don't express. What we do is

179
00:09:34,320 --> 00:09:37,080
Speaker 1: we express words, one word at a time, and they

180
00:09:37,080 --> 00:09:38,920
Speaker 1: look at all these words and they're able to find

181
00:09:39,320 --> 00:09:44,199
Speaker 1: some representations of reality that actually humans use but can't

182
00:09:44,520 --> 00:09:45,360
Speaker 1: touch and describe.

183
00:09:45,440 --> 00:09:47,240
Speaker 2: I think there's that part of it. Another part of

184
00:09:47,280 --> 00:09:51,120
Speaker 2: it is just being able again. AI is about information classification.

185
00:09:51,800 --> 00:09:54,199
Speaker 2: When you're writing, the hardest thing is to write something

186
00:09:54,880 --> 00:09:57,600
Speaker 2: terse and compress. It's a bit easier to write big,

187
00:09:57,640 --> 00:09:59,240
Speaker 2: but it's still difficult. The easiest thing for us to

188
00:09:59,280 --> 00:09:59,800
Speaker 2: do is talk.

189
00:10:00,280 --> 00:10:02,959
Speaker 1: Now there's the old added. You know, I couldn't send

190
00:10:02,960 --> 00:10:05,080
Speaker 1: you a short letter, so I've sent you a long one.

191
00:10:05,280 --> 00:10:09,160
Speaker 2: Right now, anyone anywhere can create any image soon, any

192
00:10:09,360 --> 00:10:13,080
Speaker 2: PowerPoint slide almost instantly that looks beautiful. So the fact

193
00:10:13,080 --> 00:10:15,400
Speaker 2: that it understands concepts is a big deal because the

194
00:10:15,440 --> 00:10:17,800
Speaker 2: barriers to information flow are reduced, so in motion can

195
00:10:17,840 --> 00:10:20,079
Speaker 2: flow better around. Our organizations are systems.

196
00:10:20,200 --> 00:10:24,480
Speaker 1: As you eliminate barriers to information flow, you're taking friction

197
00:10:25,000 --> 00:10:27,960
Speaker 1: out of systems. You're taking friction out of daily life,

198
00:10:28,040 --> 00:10:31,360
Speaker 1: You're taking friction out of business processes, You're taking friction

199
00:10:31,360 --> 00:10:34,760
Speaker 1: out of the economy and so we would hope to

200
00:10:34,880 --> 00:10:40,120
Speaker 1: see improvements in productivity and with that improvements in prosperity.

201
00:10:40,240 --> 00:10:42,800
Speaker 2: All of finance can be broken down into two things,

202
00:10:42,960 --> 00:10:48,480
Speaker 2: securitization and leverage. Securitization is a representation of a asset

203
00:10:48,520 --> 00:10:52,000
Speaker 2: of some sort. It's money, the trust of the American government,

204
00:10:52,480 --> 00:10:54,640
Speaker 2: it is a bond, it is a property deed, something

205
00:10:54,679 --> 00:10:55,000
Speaker 2: like that.

206
00:10:56,000 --> 00:10:57,720
Speaker 3: But you can only have so much information on that.

207
00:10:58,400 --> 00:11:00,440
Speaker 2: You and I, we have our credit score based on

208
00:11:00,480 --> 00:11:02,800
Speaker 2: the information of who we are, what we do, our

209
00:11:02,880 --> 00:11:03,960
Speaker 2: functional identities.

210
00:11:04,320 --> 00:11:06,120
Speaker 3: Most of the world is invisible.

211
00:11:05,880 --> 00:11:08,560
Speaker 1: The global South. This is people who are under banked,

212
00:11:08,679 --> 00:11:11,679
Speaker 1: people who perhaps don't have formal IDs and so on.

213
00:11:11,760 --> 00:11:14,480
Speaker 2: You need identity and you need information to allow for banking.

214
00:11:15,360 --> 00:11:17,959
Speaker 2: You need that for finance, and our financial systems are

215
00:11:18,040 --> 00:11:20,959
Speaker 2: quite slow. As you get increases in information flow, you

216
00:11:21,000 --> 00:11:24,199
Speaker 2: get increases in prosperity because you can direct assets to

217
00:11:24,240 --> 00:11:26,760
Speaker 2: where they're needed. You can direct resources to where they're needed.

218
00:11:27,200 --> 00:11:29,800
Speaker 2: It's like I always tell people in the team roadmaps,

219
00:11:29,800 --> 00:11:32,240
Speaker 2: are they a resource constrained? Their story constrained? Because if

220
00:11:32,280 --> 00:11:34,520
Speaker 2: it's a good idea. As as a leader, I hope

221
00:11:34,559 --> 00:11:36,160
Speaker 2: I will find you the resources. But you have to

222
00:11:36,160 --> 00:11:36,640
Speaker 2: convince me.

223
00:11:36,679 --> 00:11:40,920
Speaker 1: First, I can imagine these models in rich, advanced economies

224
00:11:40,960 --> 00:11:42,880
Speaker 1: where there's a huge service sector. There are lots of

225
00:11:42,880 --> 00:11:46,120
Speaker 1: people who sit behind computers typing away creating spreadsheets and

226
00:11:46,120 --> 00:11:50,400
Speaker 1: PowerPoint slides. You can imagine these models helping economies like that.

227
00:11:51,200 --> 00:11:55,439
Speaker 1: But how can we see them helping the Global South

228
00:11:55,559 --> 00:11:57,280
Speaker 1: or poorer, less advanced economies.

229
00:11:57,400 --> 00:11:59,199
Speaker 2: One of the reasons these models have got everywhere is

230
00:11:59,240 --> 00:12:01,880
Speaker 2: they've become good enough, fast enough, and cheap enough. Stuff

231
00:12:01,920 --> 00:12:04,680
Speaker 2: that used to cost dollars, tens of dolls, hundreds of dollars,

232
00:12:04,679 --> 00:12:06,960
Speaker 2: thousands of dollars you can now do with a few

233
00:12:07,000 --> 00:12:10,280
Speaker 2: simple prompts now. I think it will remove a lot

234
00:12:10,320 --> 00:12:12,920
Speaker 2: of the basic tasks and make people more productive as

235
00:12:12,960 --> 00:12:15,000
Speaker 2: opposed to leading to mascul unemployment.

236
00:12:15,080 --> 00:12:18,160
Speaker 1: And we're not seeing demand for coders drop off right.

237
00:12:18,240 --> 00:12:21,400
Speaker 1: Coders can still get work pretty quickly as they need it.

238
00:12:21,400 --> 00:12:24,240
Speaker 2: Because you will write better code and again a productivity increase.

239
00:12:24,520 --> 00:12:27,079
Speaker 2: Smart Phones can take amazing pictures, but there are more

240
00:12:27,480 --> 00:12:30,640
Speaker 2: employed photographers in the world now than ever. You know, again,

241
00:12:30,679 --> 00:12:34,000
Speaker 2: we adapt, we improve, we use the technology. However, the

242
00:12:34,000 --> 00:12:37,280
Speaker 2: Global South is a very interesting promonm We had mobile phones,

243
00:12:37,320 --> 00:12:38,880
Speaker 2: you remember it used to be for the rich, only

244
00:12:38,920 --> 00:12:39,880
Speaker 2: these big, big things.

245
00:12:40,200 --> 00:12:43,199
Speaker 1: And now in the global South there are mobile phones everywhere.

246
00:12:42,960 --> 00:12:43,120
Speaker 3: There.

247
00:12:43,440 --> 00:12:47,319
Speaker 2: They leapt over the PC to mobile. Yeah, they leapt

248
00:12:47,360 --> 00:12:50,360
Speaker 2: over to instant payments. Whereas we took a while to

249
00:12:50,360 --> 00:12:52,680
Speaker 2: catch up. In certain of western countries, you still haven't

250
00:12:52,679 --> 00:12:54,720
Speaker 2: caught up to instant payments. I think what will happen

251
00:12:54,800 --> 00:12:56,760
Speaker 2: is these models become good enough, fast enough and cheap

252
00:12:56,840 --> 00:12:58,640
Speaker 2: enough just within the next few years that they will

253
00:12:58,720 --> 00:13:01,679
Speaker 2: leap forward to intelligence augmentation.

254
00:13:01,559 --> 00:13:11,400
Speaker 1: To the benefit of these emerging markets. Yes, so let's

255
00:13:11,400 --> 00:13:13,560
Speaker 1: think about where we've got to. We've got this very

256
00:13:13,679 --> 00:13:18,680
Speaker 1: powerful technology that we've characterized. It's extremely helpful in many

257
00:13:18,679 --> 00:13:21,280
Speaker 1: many different ways. But it is the case that these

258
00:13:21,280 --> 00:13:26,960
Speaker 1: systems are extremely expensive to build. To train, as it's called,

259
00:13:27,000 --> 00:13:30,240
Speaker 1: they require lots and lots of data. The British government

260
00:13:30,240 --> 00:13:33,280
Speaker 1: has allocated more than a billion dollars to build a

261
00:13:33,400 --> 00:13:36,719
Speaker 1: supercomputer just to train these models. The rumors are that

262
00:13:37,200 --> 00:13:40,959
Speaker 1: the GPT four model from open Ai cost hundreds of

263
00:13:41,000 --> 00:13:45,440
Speaker 1: millions of dollars to train. But they're also trained on

264
00:13:46,200 --> 00:13:48,400
Speaker 1: pretty much everything that you can find on the internet,

265
00:13:48,840 --> 00:13:53,720
Speaker 1: a large part of which will be Western American English language,

266
00:13:53,720 --> 00:13:56,960
Speaker 1: a strong cultural buias. So it sounds like that not

267
00:13:57,040 --> 00:14:01,240
Speaker 1: only can poorer countries not afford this, but even if

268
00:14:01,280 --> 00:14:05,240
Speaker 1: they could, the technologies wouldn't necessarily be be suitable for

269
00:14:05,320 --> 00:14:11,199
Speaker 1: the economic requirements or the cultural requirements of Tanzania or Bangladesh.

270
00:14:11,400 --> 00:14:13,160
Speaker 2: Yeah, I think this is a real problem. I think

271
00:14:13,200 --> 00:14:16,800
Speaker 2: the quality of data we're feeding to these incredible models

272
00:14:17,160 --> 00:14:19,640
Speaker 2: is poor. It's scraped from the whole Internet. We need

273
00:14:19,680 --> 00:14:23,280
Speaker 2: better data, We need that as infrastructure. There is a

274
00:14:23,360 --> 00:14:25,680
Speaker 2: monetary equation if we need giant supercomputers, but more is

275
00:14:25,720 --> 00:14:28,240
Speaker 2: a question of talent and expertise. It's complicated to build

276
00:14:28,280 --> 00:14:30,480
Speaker 2: these things. This is one of the reasons again kind

277
00:14:30,480 --> 00:14:33,120
Speaker 2: of we had stability to do an open version and

278
00:14:33,160 --> 00:14:36,000
Speaker 2: build these data sets for each country on an open basis.

279
00:14:36,160 --> 00:14:38,360
Speaker 1: So what do you actually mean by by an open

280
00:14:38,400 --> 00:14:42,520
Speaker 1: model and how does that solve the computational problem.

281
00:14:42,640 --> 00:14:45,000
Speaker 2: We got the giant supercomputers, and then we got we

282
00:14:45,120 --> 00:14:48,520
Speaker 2: stability ability, yes, and then we made it available, and

283
00:14:48,560 --> 00:14:50,760
Speaker 2: then we released an open source so people could take

284
00:14:50,760 --> 00:14:52,440
Speaker 2: these models as a base and then extend them.

285
00:14:52,560 --> 00:14:57,040
Speaker 1: But I'm familiar with open source in software. It's a

286
00:14:57,120 --> 00:15:00,280
Speaker 1: whereas with closed source. If you're getting a Microsoft word

287
00:15:00,560 --> 00:15:04,920
Speaker 1: you buy it from Microsoft, and you can't inspect the

288
00:15:04,960 --> 00:15:07,800
Speaker 1: source code the instructions that make it run. You just

289
00:15:07,920 --> 00:15:10,960
Speaker 1: run it, so you are fundamentally consumer of it. But

290
00:15:11,160 --> 00:15:15,200
Speaker 1: with an open source project like Libra Office, which is

291
00:15:15,200 --> 00:15:18,720
Speaker 1: an open source office product, you just download the code.

292
00:15:18,760 --> 00:15:21,120
Speaker 1: You can look at the code, you can inspect the code,

293
00:15:21,200 --> 00:15:23,760
Speaker 1: you can modify the code, and you can tailor it

294
00:15:23,760 --> 00:15:26,680
Speaker 1: to your own requirement. So that's open source in software.

295
00:15:27,240 --> 00:15:29,760
Speaker 1: What's open source in a model?

296
00:15:29,960 --> 00:15:32,000
Speaker 2: So in a model, you can inspect the code, you

297
00:15:32,000 --> 00:15:34,480
Speaker 2: can expect the data sets, and the model weights themselves

298
00:15:34,560 --> 00:15:37,680
Speaker 2: are freely available as a fresh trained graduate, as it were.

299
00:15:37,880 --> 00:15:41,840
Speaker 2: And by releasing this openly so you could take it

300
00:15:41,880 --> 00:15:45,000
Speaker 2: and adapt it, it's a massive development boom where you

301
00:15:45,000 --> 00:15:46,040
Speaker 2: start seeing it everywhere.

302
00:15:46,240 --> 00:15:48,680
Speaker 1: Help me understand the mechanics of all this, because I

303
00:15:48,720 --> 00:15:55,560
Speaker 1: think it's important. Stability has its own machine learning AI supercomputer.

304
00:15:56,040 --> 00:15:59,440
Speaker 1: So you run up the cost of training these models

305
00:15:59,440 --> 00:16:02,960
Speaker 1: for the first time. Yeah, you then release them as models,

306
00:16:03,200 --> 00:16:07,920
Speaker 1: data and weights which any developer can take and use.

307
00:16:07,960 --> 00:16:10,960
Speaker 1: And when the developer runs them, they run them on

308
00:16:11,040 --> 00:16:15,480
Speaker 1: their own computing hardware, and then they're paying for that.

309
00:16:15,760 --> 00:16:17,440
Speaker 3: Yes, in some sense, they don't pay for it.

310
00:16:17,440 --> 00:16:19,520
Speaker 2: But if you want enterprise support, then you work with

311
00:16:19,600 --> 00:16:22,240
Speaker 2: us and our partners, or if you want customized versions,

312
00:16:22,280 --> 00:16:25,520
Speaker 2: because it's our view that you enterprise will want their

313
00:16:25,560 --> 00:16:28,360
Speaker 2: own versions with their own data sets underlying it. Every

314
00:16:28,400 --> 00:16:31,360
Speaker 2: country will want their own version because this is the

315
00:16:31,400 --> 00:16:34,520
Speaker 2: next generation of infrastructure. The actual comparison is five G.

316
00:16:35,520 --> 00:16:39,000
Speaker 2: This is five gv phonology works right, Yes, this is

317
00:16:39,040 --> 00:16:41,160
Speaker 2: five G for creativity as it were, it's five G

318
00:16:41,280 --> 00:16:45,040
Speaker 2: for information flow. And our trillion dollars has been spent

319
00:16:45,080 --> 00:16:45,520
Speaker 2: on five G.

320
00:16:46,480 --> 00:16:49,880
Speaker 1: Right, so we can spend a lot more on AI

321
00:16:49,960 --> 00:16:54,160
Speaker 1: systems across our economies. If we come back to your

322
00:16:54,600 --> 00:16:58,520
Speaker 1: open source models, it strikes me that one of the

323
00:16:58,520 --> 00:17:02,720
Speaker 1: things that you can do with them is you could

324
00:17:02,760 --> 00:17:05,639
Speaker 1: make them very culturally relevant. And I think back to

325
00:17:06,320 --> 00:17:09,080
Speaker 1: this idea that that sort of Western values get exported

326
00:17:09,160 --> 00:17:13,440
Speaker 1: for every country regardless. Back in the day, when you

327
00:17:13,600 --> 00:17:17,000
Speaker 1: register for Facebook, it would ask you what your marital

328
00:17:17,000 --> 00:17:19,720
Speaker 1: status was, and it was sort of single, divorced, married

329
00:17:19,800 --> 00:17:20,840
Speaker 1: or it's complicated.

330
00:17:21,080 --> 00:17:21,320
Speaker 3: Yes.

331
00:17:21,520 --> 00:17:23,199
Speaker 1: And my mum, who was in her late seventies at

332
00:17:23,240 --> 00:17:26,359
Speaker 1: the time, is registering for Facebook and is on the

333
00:17:26,359 --> 00:17:28,199
Speaker 1: phone to me game, what does it's complicated mean?

334
00:17:28,240 --> 00:17:28,760
Speaker 3: Because it just.

335
00:17:28,720 --> 00:17:32,760
Speaker 1: Didn't exist within her sort of mental space. And it

336
00:17:32,760 --> 00:17:36,080
Speaker 1: seems like given how important AI is going to be

337
00:17:36,119 --> 00:17:38,240
Speaker 1: as infrastructure, I mean, it is going to be the

338
00:17:38,320 --> 00:17:41,680
Speaker 1: layer between me and the services that I access as

339
00:17:41,720 --> 00:17:45,600
Speaker 1: a consumer, as a citizen or as an employee. It's

340
00:17:45,640 --> 00:17:50,720
Speaker 1: a really critical gatekeeper. So is that part of your vision?

341
00:17:50,800 --> 00:17:53,399
Speaker 1: An Indian version of chat GPT, a Brazilian version of

342
00:17:53,480 --> 00:17:55,879
Speaker 1: chat GPT, in Indonesian version of chat GPT.

343
00:17:56,440 --> 00:17:56,680
Speaker 3: Yes.

344
00:17:56,760 --> 00:18:00,760
Speaker 2: My vision is that every person, company, country, culture has

345
00:18:00,800 --> 00:18:04,440
Speaker 2: their own models that they themselves build and have the

346
00:18:04,520 --> 00:18:08,800
Speaker 2: data sets for, because this is vital infrastructure to represent themselves,

347
00:18:08,880 --> 00:18:10,440
Speaker 2: to extend their abilities.

348
00:18:10,800 --> 00:18:12,719
Speaker 1: How much of what you're saying is theory? Do you

349
00:18:12,760 --> 00:18:18,879
Speaker 1: actually have national models being built across the global South?

350
00:18:19,480 --> 00:18:21,280
Speaker 2: A lot of this stuff is still in the research

351
00:18:21,320 --> 00:18:24,040
Speaker 2: phase and now is only just entering the engineering phase,

352
00:18:24,480 --> 00:18:25,920
Speaker 2: and so there's a lot we still don't know about

353
00:18:25,920 --> 00:18:27,680
Speaker 2: these models. But they're good enough, fast enough, and cheap

354
00:18:27,800 --> 00:18:28,320
Speaker 2: enough to do it.

355
00:18:28,640 --> 00:18:31,320
Speaker 1: Okay, even if the models are cheaper, you still have

356
00:18:31,400 --> 00:18:35,480
Speaker 1: to get the relevant data because you know, the Internet

357
00:18:35,520 --> 00:18:40,119
Speaker 1: doesn't necessarily have lots of information about Pakistani culture in

358
00:18:40,160 --> 00:18:43,040
Speaker 1: Pakistani broadcasts in Pakistani media exactly.

359
00:18:43,160 --> 00:18:45,280
Speaker 2: And so what you need to have is you need

360
00:18:45,320 --> 00:18:50,680
Speaker 2: to have Pakistani newspapers, Pakistani broadcast and then have Pakistani's

361
00:18:50,720 --> 00:18:53,720
Speaker 2: come together to build better data sets that teach these models.

362
00:18:54,080 --> 00:18:56,840
Speaker 2: We know the technology now, but we lack the data

363
00:18:57,320 --> 00:18:59,800
Speaker 2: and so that is the key blocking point.

364
00:18:59,600 --> 00:19:03,240
Speaker 1: To get access to the data as is needed.

365
00:19:03,480 --> 00:19:06,200
Speaker 2: No to enable it so that as these data sets

366
00:19:06,240 --> 00:19:08,719
Speaker 2: are built, the models can then come from there and

367
00:19:08,760 --> 00:19:11,600
Speaker 2: then people can build on those models for their own people.

368
00:19:11,880 --> 00:19:15,040
Speaker 1: It almost sounds philanthropic, right, So what is your model

369
00:19:15,359 --> 00:19:17,960
Speaker 1: for making money from these open source models that you

370
00:19:18,000 --> 00:19:20,480
Speaker 1: are effectively giving away after all your hard work.

371
00:19:21,000 --> 00:19:25,240
Speaker 2: The whole of the infrastructure here in London in the

372
00:19:25,240 --> 00:19:27,960
Speaker 2: West is all based on open source. And the model

373
00:19:27,960 --> 00:19:29,879
Speaker 2: for open source is that you have an open version

374
00:19:30,280 --> 00:19:33,600
Speaker 2: that anyone can use, they can start experimenting with, and

375
00:19:33,640 --> 00:19:37,360
Speaker 2: then there are variant enterprise versions that you provide full

376
00:19:37,400 --> 00:19:41,600
Speaker 2: support around other services and facilities integration, and you're making

377
00:19:41,680 --> 00:19:44,280
Speaker 2: money that way with these models. We have our open

378
00:19:44,320 --> 00:19:47,440
Speaker 2: models based on open data and we have open models

379
00:19:47,480 --> 00:19:51,000
Speaker 2: based on licensed data from our partners. Because when you

380
00:19:51,040 --> 00:19:54,280
Speaker 2: talk to regulated industries and others they want models they own.

381
00:19:54,440 --> 00:19:57,520
Speaker 2: They don't want to send their data away to other people,

382
00:19:57,880 --> 00:19:59,960
Speaker 2: and they want to know every single piece of data

383
00:20:00,080 --> 00:20:01,600
Speaker 2: in there, and they want it to be the best

384
00:20:01,680 --> 00:20:02,159
Speaker 2: data you know.

385
00:20:02,320 --> 00:20:05,919
Speaker 1: Essentially, your average young developer or a small startup can

386
00:20:06,520 --> 00:20:09,200
Speaker 1: download your models and use them, but if there's an issue,

387
00:20:09,400 --> 00:20:11,320
Speaker 1: there's not a lot of support. But if you're a

388
00:20:11,320 --> 00:20:14,960
Speaker 1: big company, you're a national government, you might enter into

389
00:20:15,000 --> 00:20:18,640
Speaker 1: a more detailed contract where there is support and advice

390
00:20:18,960 --> 00:20:22,240
Speaker 1: and potentially even data yes coming through. We know that

391
00:20:22,280 --> 00:20:26,440
Speaker 1: AI is a very powerful technology, and Stability has taken

392
00:20:26,480 --> 00:20:29,520
Speaker 1: a different path to other firms by going through an

393
00:20:29,560 --> 00:20:34,840
Speaker 1: open source approach, which could democratize the technology, making it culturally, linguistically,

394
00:20:35,320 --> 00:20:41,760
Speaker 1: locally relevant for any nation, any business, any region, any individual.

395
00:20:42,320 --> 00:20:46,119
Speaker 1: But you're up against firms like Microsoft and open AI

396
00:20:46,600 --> 00:20:50,280
Speaker 1: and Google and deep Mind and others. How do you

397
00:20:50,280 --> 00:20:51,080
Speaker 1: plan to compete?

398
00:20:51,880 --> 00:20:54,760
Speaker 2: I think there's a cab career on addressable market. Our

399
00:20:54,760 --> 00:20:57,040
Speaker 2: addressable market is all the private data in the world.

400
00:20:57,800 --> 00:21:00,280
Speaker 2: Data you can't send anyone but your personal data or

401
00:21:00,359 --> 00:21:03,920
Speaker 2: enterprise data, financial regulated data, and so our models will

402
00:21:03,920 --> 00:21:07,399
Speaker 2: go in and transform that into knowledge and we'll have

403
00:21:07,400 --> 00:21:09,800
Speaker 2: a hybrid AI think, we've got our models on your

404
00:21:09,840 --> 00:21:11,879
Speaker 2: private data, and we standardize all of that, make it

405
00:21:12,000 --> 00:21:14,800
Speaker 2: very predictable, loads of support, and then you use these

406
00:21:14,800 --> 00:21:17,800
Speaker 2: proprietary systems for the best outcomes. You'll have your own

407
00:21:17,840 --> 00:21:21,280
Speaker 2: graduates that you hire, and you'll hire from McKinsey, and.

408
00:21:21,280 --> 00:21:22,200
Speaker 1: You'll put them together.

409
00:21:22,240 --> 00:21:23,040
Speaker 3: You'll put them together.

410
00:21:23,320 --> 00:21:28,200
Speaker 1: But I'm curious because other companies are taking a closed

411
00:21:28,359 --> 00:21:33,399
Speaker 1: source approach to proprietary data. So there are companies like

412
00:21:33,560 --> 00:21:39,680
Speaker 1: cohere and Anthropic who will build a powerful generative AI

413
00:21:39,760 --> 00:21:43,920
Speaker 1: model just on a company's own private data. They're competing

414
00:21:43,920 --> 00:21:45,280
Speaker 1: with you as well, right, and so why is your

415
00:21:45,280 --> 00:21:46,119
Speaker 1: approach better than that?

416
00:21:46,440 --> 00:21:49,159
Speaker 2: They will not give that company ownership of that model.

417
00:21:50,680 --> 00:21:53,440
Speaker 2: They will not share the detail of every single piece

418
00:21:53,440 --> 00:21:54,600
Speaker 2: of data that's in that model.

419
00:21:55,040 --> 00:21:57,879
Speaker 1: But that's the case today with lots of the technology

420
00:21:57,880 --> 00:21:59,800
Speaker 1: that we use. You know, when I'm running my e

421
00:22:00,000 --> 00:22:04,280
Speaker 1: commerce application on the cloud, I don't know the details

422
00:22:04,320 --> 00:22:07,520
Speaker 1: of every configuration of the servers that I'm renting from

423
00:22:08,040 --> 00:22:11,160
Speaker 1: Amazon or Microsoft as are so businesses are used to that.

424
00:22:11,080 --> 00:22:13,960
Speaker 2: One hundred percent, and you have data that you can

425
00:22:14,000 --> 00:22:16,639
Speaker 2: share with people, but there is a core of regulated

426
00:22:16,680 --> 00:22:19,440
Speaker 2: data and other things hip a compliant data medical data

427
00:22:19,480 --> 00:22:21,240
Speaker 2: that you cannot send to other companies, and you have

428
00:22:21,240 --> 00:22:25,119
Speaker 2: to build your own systems for inside regulated environments. The

429
00:22:25,160 --> 00:22:28,480
Speaker 2: feedback we've got from regulated entities and again from policymakers

430
00:22:28,520 --> 00:22:32,160
Speaker 2: and others is open, transparent models, even if it's licensed

431
00:22:32,200 --> 00:22:34,440
Speaker 2: data are something that we would like a lot, and

432
00:22:34,440 --> 00:22:36,520
Speaker 2: we would like to own this technology if it's going

433
00:22:36,560 --> 00:22:37,920
Speaker 2: to be vital infrastructure.

434
00:22:38,280 --> 00:22:41,600
Speaker 1: Right, we're in London. Now there's a deep bench of

435
00:22:41,640 --> 00:22:46,000
Speaker 1: AI skills. How do we expand this and democratize it

436
00:22:46,320 --> 00:22:51,080
Speaker 1: out to countries where there's just less talent in these

437
00:22:51,160 --> 00:22:52,080
Speaker 1: breakthrough areas.

438
00:22:52,240 --> 00:22:54,280
Speaker 3: I think there's talent, it just hasn't been accessed.

439
00:22:54,320 --> 00:22:56,640
Speaker 2: And these models are very interesting and you can use

440
00:22:56,680 --> 00:22:59,879
Speaker 2: AI to help you develop applications of the AI.

441
00:23:00,119 --> 00:23:01,760
Speaker 3: Quite a funny kind of recursion there.

442
00:23:01,680 --> 00:23:06,040
Speaker 1: Right, So you effectively will start to support places where

443
00:23:06,040 --> 00:23:09,280
Speaker 1: perhaps the workforce doesn't have the depth of San Francisco's

444
00:23:09,320 --> 00:23:11,199
Speaker 1: AI talent with the tools themselves.

445
00:23:11,320 --> 00:23:14,399
Speaker 2: Yes, because the AI models are pre computed, the actual

446
00:23:14,480 --> 00:23:17,960
Speaker 2: running of the AI is very computationally non intensive. The

447
00:23:18,040 --> 00:23:21,399
Speaker 2: creation of the AI is ridiculously intensive, right, So you've

448
00:23:21,400 --> 00:23:23,360
Speaker 2: got all the energy at the start rather than the end.

449
00:23:23,960 --> 00:23:27,919
Speaker 1: So that takes us back to what stability will do.

450
00:23:28,000 --> 00:23:31,679
Speaker 1: Stability will take the cost upfront, and then it'll find

451
00:23:32,320 --> 00:23:35,880
Speaker 1: rich companies, rich nations, rich clients to tailor the models,

452
00:23:36,520 --> 00:23:39,879
Speaker 1: which allows you to continue to make the base foundation model,

453
00:23:40,080 --> 00:23:44,960
Speaker 1: which can then be given as open source to anyone

454
00:23:45,040 --> 00:23:45,760
Speaker 1: else who yes.

455
00:23:45,840 --> 00:23:48,080
Speaker 2: It stimulates demand, and then as people go up and

456
00:23:48,119 --> 00:23:50,439
Speaker 2: they need the support, they come to us and our partners.

457
00:23:50,480 --> 00:23:53,240
Speaker 2: As any customization, they come to us on our partners, right,

458
00:23:53,280 --> 00:23:56,840
Speaker 2: And it's for private data open models, and then other

459
00:23:56,920 --> 00:23:59,959
Speaker 2: models are for data that you either are semi private

460
00:24:00,080 --> 00:24:02,080
Speaker 2: or you don't mind, and you combine the two so

461
00:24:02,119 --> 00:24:03,399
Speaker 2: you have models of both types.

462
00:24:04,440 --> 00:24:08,120
Speaker 1: There are, of course concerns with the safety of AI systems,

463
00:24:08,160 --> 00:24:12,480
Speaker 1: and one argument is that with closed models that are controlled,

464
00:24:12,880 --> 00:24:15,359
Speaker 1: there's a lot more safety because if I'm accessing it

465
00:24:15,440 --> 00:24:19,240
Speaker 1: over the web, the organization that's running the AI system

466
00:24:19,320 --> 00:24:21,680
Speaker 1: can can stop me if I'm trying to do something

467
00:24:21,720 --> 00:24:24,720
Speaker 1: bad with it. And with open source models, of course

468
00:24:24,720 --> 00:24:27,359
Speaker 1: they're just available for anyone to download, so the cat

469
00:24:27,440 --> 00:24:29,840
Speaker 1: is literally out of the bag many many times over,

470
00:24:29,960 --> 00:24:35,360
Speaker 1: millions of times over. Is your approach less safe than

471
00:24:35,800 --> 00:24:36,680
Speaker 1: the closed approach?

472
00:24:37,280 --> 00:24:39,600
Speaker 2: I think it's more safe. There's a reason that our

473
00:24:39,640 --> 00:24:42,720
Speaker 2: infrastructure is based on open source databases and servers and

474
00:24:42,760 --> 00:24:45,560
Speaker 2: others because it can be checked, it can be tested,

475
00:24:45,600 --> 00:24:48,760
Speaker 2: and it can be fully audited and battle tested. Our

476
00:24:48,800 --> 00:24:51,000
Speaker 2: approach to stability is to create the standard around this,

477
00:24:51,080 --> 00:24:54,640
Speaker 2: so there aren't thousands of different models. There is an

478
00:24:54,760 --> 00:24:58,159
Speaker 2: entity in a partnership and an ecosystem that standardizes around

479
00:24:58,160 --> 00:25:02,919
Speaker 2: this principles line safety, water marking, and other things, so

480
00:25:02,960 --> 00:25:04,040
Speaker 2: it becomes predictable.

481
00:25:04,440 --> 00:25:06,600
Speaker 1: But you've put the models out for any bad actor,

482
00:25:06,760 --> 00:25:13,000
Speaker 1: any hacker, any annoyed employee to build something difficult with.

483
00:25:13,480 --> 00:25:15,280
Speaker 2: We weren't the ones that came up with open models.

484
00:25:15,280 --> 00:25:19,479
Speaker 2: We're standardizing it. We're supporting open innovation for detection and

485
00:25:19,560 --> 00:25:20,959
Speaker 2: prevention as well as creation.

486
00:25:21,320 --> 00:25:24,520
Speaker 1: But it does sound like bad actors will end up

487
00:25:24,560 --> 00:25:26,800
Speaker 1: having a little bit of a field day, which creates

488
00:25:26,880 --> 00:25:31,080
Speaker 1: I suspect an enormous opportunity for an AI driven security

489
00:25:31,119 --> 00:25:32,840
Speaker 1: and resilience industry.

490
00:25:33,000 --> 00:25:36,240
Speaker 2: The reality is that we're stronger together when things are open,

491
00:25:36,320 --> 00:25:40,600
Speaker 2: and open is required for all the private, regulated and

492
00:25:40,640 --> 00:25:43,280
Speaker 2: other data out there. If you don't have open systems,

493
00:25:43,359 --> 00:25:46,080
Speaker 2: then you will only have proprietary entities, and they become

494
00:25:46,119 --> 00:25:48,560
Speaker 2: the choke points on the Internet, and that's far more

495
00:25:48,640 --> 00:25:51,879
Speaker 2: dangerous than the other side. Open is there anyway, But

496
00:25:51,920 --> 00:25:54,280
Speaker 2: like I said, let's standardize it, let's make it safer,

497
00:25:54,320 --> 00:25:57,800
Speaker 2: and let's work together to combat the bad as opposed

498
00:25:57,840 --> 00:26:00,320
Speaker 2: to leaving it with a few unelected giant companies.

499
00:26:00,640 --> 00:26:04,880
Speaker 1: Part of the story of technology is that technology has

500
00:26:04,960 --> 00:26:10,439
Speaker 1: been exported from one place really for everyone else to use.

501
00:26:10,560 --> 00:26:13,200
Speaker 1: I think one exciting opportunity now is the idea that

502
00:26:14,200 --> 00:26:17,000
Speaker 1: the people on whom this technology is going to operate

503
00:26:17,040 --> 00:26:20,560
Speaker 1: could potentially build their own Now, many of those people

504
00:26:20,640 --> 00:26:23,000
Speaker 1: are going to be in the Global South, and the

505
00:26:23,040 --> 00:26:26,639
Speaker 1: premise of our conversation is that within five years, safe

506
00:26:26,680 --> 00:26:30,960
Speaker 1: and open source generative AI could make communic for contribution

507
00:26:31,480 --> 00:26:36,760
Speaker 1: to increase the GDP of those world's poorest nations. How

508
00:26:36,920 --> 00:26:39,480
Speaker 1: likely is it that this vision could become reality? I

509
00:26:39,560 --> 00:26:43,240
Speaker 1: think it's incredibly likely. The desired talent and passion to

510
00:26:43,320 --> 00:26:47,080
Speaker 1: adopt technology like this is huge within the Global South,

511
00:26:47,280 --> 00:26:49,280
Speaker 1: and it is where it can have the most impact

512
00:26:49,320 --> 00:26:51,800
Speaker 1: the highest ROI. So I think they'll take the building

513
00:26:51,840 --> 00:26:54,760
Speaker 1: blocks that we and others provide and they'll build some

514
00:26:54,840 --> 00:26:59,000
Speaker 1: amazing things to activate their potential. Mada, it's a great vision.

515
00:26:59,280 --> 00:27:01,399
Speaker 1: Thank you so much, my pleasure, thank you for having me.

516
00:27:08,920 --> 00:27:12,440
Speaker 1: Reflecting on my conversation with Emmad, I'm reminded that much

517
00:27:12,480 --> 00:27:15,200
Speaker 1: of the software that powers the Internet today, used by

518
00:27:15,240 --> 00:27:18,840
Speaker 1: billions of US, is actually open source. It's proven to

519
00:27:18,880 --> 00:27:24,320
Speaker 1: be resilient, stable, and importantly affordable. The open approach is

520
00:27:24,440 --> 00:27:28,800
Speaker 1: one reason why the Internet is today ubiquitous, So why

521
00:27:28,800 --> 00:27:31,840
Speaker 1: wouldn't that be true for generative AI? And if the

522
00:27:31,880 --> 00:27:35,440
Speaker 1: technology can live up to its promise of improving productivity,

523
00:27:35,800 --> 00:27:38,720
Speaker 1: wouldn't the open approach make it more widely accessible to

524
00:27:38,760 --> 00:27:41,399
Speaker 1: the poorest countries in the world. That seems to make

525
00:27:41,480 --> 00:27:53,720
Speaker 1: sense to me. Thanks for listening to the Exponentially podcast.

526
00:27:53,920 --> 00:27:56,919
Speaker 1: If you enjoy the show, please leave a review or rating.

527
00:27:57,080 --> 00:28:00,960
Speaker 1: It really does help others find us. Any podcast is

528
00:28:01,000 --> 00:28:04,920
Speaker 1: presented by me Azeem Azar. The sound designer is Will Horrocks.

529
00:28:05,200 --> 00:28:08,120
Speaker 1: The research was led by Chloe Ippah and music composed

530
00:28:08,119 --> 00:28:11,560
Speaker 1: by Emily Green and John Zarcone. The show is produced

531
00:28:11,560 --> 00:28:15,719
Speaker 1: by Frederick Cassella, Maria Garrilov and me Azeem Azar. Special

532
00:28:15,760 --> 00:28:19,159
Speaker 1: thanks to Sage Bauman, Jeff Grocott and Magnus Henrikson. The

533
00:28:19,200 --> 00:28:23,280
Speaker 1: executive producers are Andrew Barden, Adam Kamiski, and Kyle Kramer.

534
00:28:23,560 --> 00:28:27,239
Speaker 1: David Rivella is the managing editor. Exponentially was created by

535
00:28:27,240 --> 00:28:29,199
Speaker 1: Frederick Cassella and is an e to the pie I

536
00:28:29,280 --> 00:28:33,400
Speaker 1: plus one Limited production in association with Bloomberg LC