1
00:00:15,356 --> 00:00:31,836
Speaker 1: Pushkin. I'm Jacob Goldstein. This is what's your problem? And

2
00:00:31,876 --> 00:00:35,276
Speaker 1: my guest today is Josh Tieringel. He's a staff writer

3
00:00:35,356 --> 00:00:37,716
Speaker 1: at The Atlantic, and he just wrote a new book

4
00:00:37,756 --> 00:00:41,236
Speaker 1: called AI for Good And the subtitle of the book

5
00:00:41,316 --> 00:00:44,796
Speaker 1: explains the problem that Josh is interested in. The subtitle

6
00:00:44,916 --> 00:00:48,716
Speaker 1: is how real people are using artificial intelligence to fix

7
00:00:48,796 --> 00:00:51,956
Speaker 1: things that matter. As you'll hear, I talked with Josh

8
00:00:51,996 --> 00:00:55,956
Speaker 1: about several of these real people. Professor trying to understand

9
00:00:55,996 --> 00:00:59,436
Speaker 1: her nonverbal son, a general who ran the logistics side

10
00:00:59,436 --> 00:01:03,156
Speaker 1: of Operation Warp Speed, a hospital CEO trying to reduce

11
00:01:03,196 --> 00:01:06,196
Speaker 1: the rate of deadly infections. And one thing just to

12
00:01:06,236 --> 00:01:08,356
Speaker 1: note here is you know, and I guess this is

13
00:01:08,356 --> 00:01:11,076
Speaker 1: not surprising in the state of AI. Some of these

14
00:01:11,116 --> 00:01:13,356
Speaker 1: stories are not yet finished. They're kind of stories that

15
00:01:13,396 --> 00:01:16,516
Speaker 1: are still in the middle. But before we got into

16
00:01:16,516 --> 00:01:20,276
Speaker 1: the specific stories, I asked Josh what common threads he

17
00:01:20,356 --> 00:01:23,036
Speaker 1: saw across all these people he wrote about, who are

18
00:01:23,076 --> 00:01:25,756
Speaker 1: you trying to use AI to do good things?

19
00:01:26,436 --> 00:01:30,876
Speaker 2: They're largely stubborn. Uh huh, They're really stubborn. And you know,

20
00:01:30,916 --> 00:01:33,556
Speaker 2: one of the things that I admire. I'm very impatient

21
00:01:33,596 --> 00:01:36,356
Speaker 2: and when I see a problem. Oftentimes I have what

22
00:01:36,436 --> 00:01:38,916
Speaker 2: I think is a fairly normal human impulse, which is like,

23
00:01:39,276 --> 00:01:42,236
Speaker 2: wipe the mismatched puzzle pieces off the table and get

24
00:01:42,236 --> 00:01:46,036
Speaker 2: a different puzzle. Right, It's frustrating to dig into systems

25
00:01:46,436 --> 00:01:50,916
Speaker 2: to learn technology, and these people, they don't get frustrated,

26
00:01:51,276 --> 00:01:53,796
Speaker 2: not to the point of destroying the system. In fact,

27
00:01:54,716 --> 00:01:57,316
Speaker 2: most of them have no experience with AI or technology.

28
00:01:57,716 --> 00:02:00,636
Speaker 2: They care about medicine, they care about teaching, They care

29
00:02:00,676 --> 00:02:03,476
Speaker 2: about solving a really hard problem, and so they just

30
00:02:03,556 --> 00:02:07,276
Speaker 2: keep looking for solutions. And when something seems promising, their

31
00:02:07,316 --> 00:02:10,476
Speaker 2: ability to concentrate and learn, earn the technology of that

32
00:02:10,636 --> 00:02:14,636
Speaker 2: promise and apply it to their problem is through the roof.

33
00:02:15,196 --> 00:02:18,276
Speaker 2: And so I just kept being awed by the fact

34
00:02:18,276 --> 00:02:22,916
Speaker 2: that they were curious, they're patient, They were impatient at

35
00:02:22,956 --> 00:02:26,356
Speaker 2: exactly the right moments in the sense that things weren't moving.

36
00:02:26,716 --> 00:02:29,876
Speaker 2: They knew how to deploy either their political capital or

37
00:02:29,876 --> 00:02:34,036
Speaker 2: their temper in ways to get things moving. And you know,

38
00:02:34,076 --> 00:02:35,996
Speaker 2: I found time and time again that when I would

39
00:02:36,036 --> 00:02:38,796
Speaker 2: talk to other people who were witnessing these phenomenon, whether

40
00:02:38,836 --> 00:02:41,716
Speaker 2: it was in hospitals or in government, we would all

41
00:02:41,756 --> 00:02:43,996
Speaker 2: just kind of be like, Yeah, these are the ones

42
00:02:44,036 --> 00:02:46,996
Speaker 2: making it happen. They're kind of Samurai, right. They care

43
00:02:47,076 --> 00:02:50,396
Speaker 2: about their problem more than other people care about standing

44
00:02:50,396 --> 00:02:52,916
Speaker 2: in the way of the solution, and they get it done.

45
00:02:53,156 --> 00:02:54,476
Speaker 1: Tell me about Christy Johnson.

46
00:02:55,076 --> 00:03:00,316
Speaker 2: Yeah, So, Christy Johnson is certainly one of the most

47
00:03:00,396 --> 00:03:05,796
Speaker 2: inspiring interesting people I've met in my journalism career. So, I,

48
00:03:05,876 --> 00:03:08,436
Speaker 2: like a lot of people, was thinking about AI and

49
00:03:08,516 --> 00:03:11,796
Speaker 2: human connect and reading all the headlines about the sort

50
00:03:11,796 --> 00:03:16,596
Speaker 2: of parasocial crisis in which people were dating, chat GPT,

51
00:03:16,916 --> 00:03:20,276
Speaker 2: falling in love with Claude, realizing that this was mimicking

52
00:03:20,356 --> 00:03:23,076
Speaker 2: a lot of the worst elements of social media, seeing

53
00:03:23,116 --> 00:03:26,916
Speaker 2: the studies and getting very, very depressed about what AI

54
00:03:27,036 --> 00:03:29,596
Speaker 2: would mean for human connection. So I was having that

55
00:03:29,676 --> 00:03:33,036
Speaker 2: conversation with Rosin Picard, who runs the MIT Media Lab,

56
00:03:33,076 --> 00:03:34,876
Speaker 2: and she basically said, look, you got to meet Christy,

57
00:03:35,676 --> 00:03:39,396
Speaker 2: and so she introduced me. Christy Johnson was a physicist

58
00:03:40,276 --> 00:03:47,076
Speaker 2: and a brilliant student. She met another physicist in a program.

59
00:03:47,516 --> 00:03:49,996
Speaker 2: He is an astrophysicist and actually was on the team

60
00:03:50,636 --> 00:03:53,076
Speaker 2: that took the first picture of a black hole. The

61
00:03:53,116 --> 00:03:55,236
Speaker 2: two of them fell in the deepest form of nerd

62
00:03:55,276 --> 00:03:59,036
Speaker 2: love you could possibly imagine. I mean even the two

63
00:03:59,036 --> 00:04:01,196
Speaker 2: of them laugh at just how dorky their love letters

64
00:04:01,236 --> 00:04:03,996
Speaker 2: are to each other, all about physics and math and

65
00:04:04,516 --> 00:04:07,956
Speaker 2: the world. They get married and they have a kid,

66
00:04:08,036 --> 00:04:12,116
Speaker 2: and about a couple of weeks in she's like, there's

67
00:04:12,116 --> 00:04:14,596
Speaker 2: something wrong with my son. And she's kind of gas

68
00:04:14,716 --> 00:04:17,276
Speaker 2: lit by the medical profession for a year just hearing no,

69
00:04:17,356 --> 00:04:21,636
Speaker 2: he's fine, Stop worrying, enjoy your kid. Finally they get

70
00:04:21,636 --> 00:04:25,476
Speaker 2: some testing done and he has a genetic deficiency. And

71
00:04:26,476 --> 00:04:29,876
Speaker 2: the other six children like him in the world are

72
00:04:29,916 --> 00:04:35,556
Speaker 2: all severely autistic, epileptic, and completely nonverbal, and so this

73
00:04:35,596 --> 00:04:39,876
Speaker 2: is the kind of diagnosis that changes everybody's lives. Within

74
00:04:39,916 --> 00:04:44,636
Speaker 2: the family. They begin to grieve a little bit for

75
00:04:44,676 --> 00:04:47,636
Speaker 2: their expectations for their son, but they're also scientists, and

76
00:04:47,996 --> 00:04:51,396
Speaker 2: Christy and Michael both are like, we should understand this,

77
00:04:51,516 --> 00:04:54,516
Speaker 2: we should do something about it. A couple of years

78
00:04:54,516 --> 00:04:57,676
Speaker 2: ago past and Christy is like, no longer all that

79
00:04:57,716 --> 00:05:01,036
Speaker 2: interested in physics. She's very interested in her son. And

80
00:05:01,076 --> 00:05:03,396
Speaker 2: so what she set out to do was solve the

81
00:05:03,436 --> 00:05:08,236
Speaker 2: problem of the more than one million nonverbal Americans who

82
00:05:08,556 --> 00:05:11,436
Speaker 2: are stuck to their parents and their caregivers. That's it

83
00:05:11,756 --> 00:05:15,556
Speaker 2: for everybody in that universe, the universe shrinks.

84
00:05:16,036 --> 00:05:18,596
Speaker 1: Just to be clear, what is her dream? What is

85
00:05:18,596 --> 00:05:20,156
Speaker 1: the outcome she hopes to achieve.

86
00:05:20,636 --> 00:05:23,156
Speaker 2: Her dream is that her son and other people like

87
00:05:23,236 --> 00:05:25,836
Speaker 2: him can interact with people they don't know all that

88
00:05:25,876 --> 00:05:29,636
Speaker 2: well and share what it is they want and need.

89
00:05:29,796 --> 00:05:35,236
Speaker 1: So a machine will turn their nonverbal vocalizations into words. Yeah,

90
00:05:35,396 --> 00:05:38,316
Speaker 1: verbal people can understand that other people can respond to

91
00:05:38,476 --> 00:05:40,876
Speaker 1: and interact with. And so she made a bet. She

92
00:05:40,956 --> 00:05:42,996
Speaker 1: basically thought, look, I see where AI.

93
00:05:42,836 --> 00:05:45,756
Speaker 2: And machine learning are right now. My hunch is that

94
00:05:45,796 --> 00:05:48,276
Speaker 2: they will get better and better. But what I need

95
00:05:48,476 --> 00:05:52,476
Speaker 2: is data. And so she created a scientific protocol to

96
00:05:52,596 --> 00:05:57,316
Speaker 2: get vocalizations from these kids, and she used every tool available.

97
00:05:57,396 --> 00:05:59,676
Speaker 2: So she went on YouTube, she went on social media.

98
00:05:59,796 --> 00:06:01,956
Speaker 2: She said, this is me and my son. This is

99
00:06:01,956 --> 00:06:04,956
Speaker 2: what happens when he's hungry. Listen to that vocalization, and

100
00:06:04,956 --> 00:06:08,836
Speaker 2: now I'm recording it, and she began to generate a

101
00:06:08,876 --> 00:06:10,956
Speaker 2: little bit of momentum. She went to the MIT Media

102
00:06:10,996 --> 00:06:13,716
Speaker 2: Lab and spoke to Rosbaccard one night, kind of cold

103
00:06:13,796 --> 00:06:18,596
Speaker 2: calder and Roz realized, like, this woman has no qualifications

104
00:06:18,636 --> 00:06:21,476
Speaker 2: to be at the media lab, but my god, like,

105
00:06:21,876 --> 00:06:26,276
Speaker 2: what an incredible human she got there, she did her PhD.

106
00:06:27,076 --> 00:06:29,516
Speaker 2: And she proved that if you can gather enough of

107
00:06:29,556 --> 00:06:33,236
Speaker 2: these vocalizations as audio files, you can begin to attack

108
00:06:33,276 --> 00:06:36,676
Speaker 2: the problem. And so what she learned largely through using AIS.

109
00:06:36,716 --> 00:06:41,716
Speaker 2: At first you have to multiply the files, so a

110
00:06:41,956 --> 00:06:47,116
Speaker 2: system like chat, GPT or Claude is trained on potentially

111
00:06:47,156 --> 00:06:50,156
Speaker 2: trillions of data points in order to become as fluent

112
00:06:50,196 --> 00:06:53,996
Speaker 2: as it is. She had a couple thousand audio files.

113
00:06:54,316 --> 00:06:56,996
Speaker 2: So the first thing she did is get her graduate

114
00:06:57,036 --> 00:07:00,236
Speaker 2: students to figure out how to synthesize more data, how

115
00:07:00,236 --> 00:07:04,516
Speaker 2: to take those data files, those original sounds, and grow

116
00:07:04,556 --> 00:07:05,916
Speaker 2: them and multiply them.

117
00:07:05,996 --> 00:07:08,316
Speaker 1: It's like what they're trying to do with self driving cars,

118
00:07:08,436 --> 00:07:12,196
Speaker 1: right exactly, like drive on pretend roads essentially exactly.

119
00:07:12,316 --> 00:07:15,396
Speaker 2: It's called synthetic data. And AI is really good at

120
00:07:15,396 --> 00:07:16,316
Speaker 2: making synthetic data.

121
00:07:16,316 --> 00:07:18,316
Speaker 1: And they're even they're working on that for text, right

122
00:07:18,356 --> 00:07:20,996
Speaker 1: because now the current models basically are trained on the

123
00:07:20,996 --> 00:07:23,596
Speaker 1: whole Internet, so like there's no more real text for

124
00:07:23,636 --> 00:07:25,636
Speaker 1: them to read, so they need synthetic texts.

125
00:07:25,836 --> 00:07:28,116
Speaker 2: They're out of stuff, right, And so if you want

126
00:07:28,116 --> 00:07:30,196
Speaker 2: the training to advance, you have to synthesize a lot

127
00:07:30,236 --> 00:07:32,756
Speaker 2: of this data, and she's continuing to grow the organic

128
00:07:32,756 --> 00:07:33,556
Speaker 2: files as well.

129
00:07:34,196 --> 00:07:35,596
Speaker 1: Tell me about ROSCO.

130
00:07:36,836 --> 00:07:42,716
Speaker 2: Yeah, So ROSCO is a protocol and basically, you can't

131
00:07:42,996 --> 00:07:46,476
Speaker 2: do what Christy is trying to do without science because

132
00:07:46,996 --> 00:07:49,396
Speaker 2: you need some sort of rigorous idea of what you're

133
00:07:49,436 --> 00:07:53,676
Speaker 2: testing for or else. The data that you're recording is meaningless.

134
00:07:53,676 --> 00:07:54,636
Speaker 1: It's not really data.

135
00:07:55,036 --> 00:07:58,036
Speaker 2: It's not really data, it just sounds. And so working

136
00:07:58,036 --> 00:08:00,436
Speaker 2: with a grad student, what she figured out is, Okay,

137
00:08:00,516 --> 00:08:04,876
Speaker 2: we don't need all the data. All the data is unprocessable.

138
00:08:05,316 --> 00:08:07,596
Speaker 2: We need the right kinds of data for this task.

139
00:08:07,676 --> 00:08:10,596
Speaker 2: And so what they did is create this protocol where

140
00:08:10,596 --> 00:08:13,956
Speaker 2: they would run families through a certain kind of test

141
00:08:14,076 --> 00:08:16,476
Speaker 2: and it was split timed so that it was standardized,

142
00:08:16,516 --> 00:08:20,996
Speaker 2: and they would say, for two minutes, give your child

143
00:08:21,316 --> 00:08:24,076
Speaker 2: a YouTube video because we want to hear. And you know,

144
00:08:24,116 --> 00:08:26,476
Speaker 2: in the autism community, YouTube is great. It cues up

145
00:08:26,516 --> 00:08:29,156
Speaker 2: the same video as many times as you want. People

146
00:08:29,196 --> 00:08:31,756
Speaker 2: always have their favorite video, she said, whatever their favorite

147
00:08:31,796 --> 00:08:34,916
Speaker 2: video is queued up. At the one forty five mark,

148
00:08:34,916 --> 00:08:39,076
Speaker 2: they would introduce a beach ball and buffering.

149
00:08:38,996 --> 00:08:41,636
Speaker 1: Beach ball meaning the spinning wheel that says like pause,

150
00:08:41,676 --> 00:08:42,396
Speaker 1: it's loading.

151
00:08:42,596 --> 00:08:45,316
Speaker 2: Yeah, they would pause the video because what they wanted

152
00:08:45,356 --> 00:08:50,076
Speaker 2: to do was get a vocalization of dissatisfaction and frustration,

153
00:08:50,316 --> 00:08:52,036
Speaker 2: and so that was a way to do it. And

154
00:08:52,116 --> 00:08:55,156
Speaker 2: they would say, okay, this is a five minute snack period.

155
00:08:55,516 --> 00:08:57,156
Speaker 2: We want you to make sure that you are giving

156
00:08:57,196 --> 00:09:00,196
Speaker 2: them their favorite sack, and we want to hear what

157
00:09:00,196 --> 00:09:02,476
Speaker 2: that is like. And so they would do these things

158
00:09:02,516 --> 00:09:05,156
Speaker 2: that are really relevant to her ultimate goal, which is

159
00:09:05,156 --> 00:09:08,076
Speaker 2: making sure that these kids could be out there in

160
00:09:08,116 --> 00:09:10,796
Speaker 2: the world talking to other people and having needs met.

161
00:09:11,596 --> 00:09:14,916
Speaker 2: And so it was a very scientific protocol. And then

162
00:09:14,956 --> 00:09:18,356
Speaker 2: she would send it to families along with a phone

163
00:09:19,316 --> 00:09:22,156
Speaker 2: and a camera and basically say, record it from these

164
00:09:22,236 --> 00:09:24,596
Speaker 2: various angles, send it back to us, so that the

165
00:09:24,716 --> 00:09:28,276
Speaker 2: data is standardized. It's not just random vocalizations that they're

166
00:09:28,276 --> 00:09:32,316
Speaker 2: figuring out. And because the population is so specific and

167
00:09:32,316 --> 00:09:35,196
Speaker 2: because the keepers of the knowledge are so specific to

168
00:09:35,276 --> 00:09:38,396
Speaker 2: parents and caregivers, it took her a while to figure out,

169
00:09:38,476 --> 00:09:40,076
Speaker 2: that's how you actually want to get the data.

170
00:09:40,636 --> 00:09:44,916
Speaker 1: Huh. Yeah, it sounds like a lot of work. It

171
00:09:44,956 --> 00:09:47,476
Speaker 1: sounds hard to scale.

172
00:09:47,636 --> 00:09:50,036
Speaker 2: Yeah, And it would be hard to scale if not

173
00:09:50,076 --> 00:09:52,716
Speaker 2: for synthetic data, and if not for all of the

174
00:09:52,756 --> 00:09:54,556
Speaker 2: advances on the other side with translation.

175
00:09:54,756 --> 00:09:57,476
Speaker 1: Yeah, so you don't have to scale that much, is

176
00:09:57,516 --> 00:09:58,276
Speaker 1: what those things.

177
00:09:58,196 --> 00:10:00,076
Speaker 2: Well, you don't have to scale it nearly as much, right,

178
00:10:00,116 --> 00:10:02,676
Speaker 2: And what would have been a project that would probably

179
00:10:02,756 --> 00:10:07,476
Speaker 2: have happened over three scientists lifetimes yea, is now maybe

180
00:10:07,596 --> 00:10:10,756
Speaker 2: only over half of her career. And that's a real

181
00:10:10,876 --> 00:10:13,756
Speaker 2: form of progress. And so even though Christy doesn't have

182
00:10:13,836 --> 00:10:16,916
Speaker 2: the kinds of funding that Google or open Ai have,

183
00:10:17,516 --> 00:10:19,716
Speaker 2: she's able to kind of surf in their wake, and

184
00:10:20,116 --> 00:10:23,476
Speaker 2: they've made these progressions on things like zero shot translation,

185
00:10:23,596 --> 00:10:26,076
Speaker 2: with some people may have heard of. You know, historically

186
00:10:26,116 --> 00:10:29,476
Speaker 2: we would translate things by taking one set of language,

187
00:10:30,516 --> 00:10:33,836
Speaker 2: decoding each word and rewriting it in a different language,

188
00:10:34,036 --> 00:10:37,276
Speaker 2: and up until about ten years ago, that's what machine

189
00:10:37,276 --> 00:10:41,396
Speaker 2: translation was. Zero shot translation is actually a translation at

190
00:10:41,436 --> 00:10:45,796
Speaker 2: such a deep level that you're not translating language to language,

191
00:10:45,836 --> 00:10:50,796
Speaker 2: you're basically translating every language to every language simultaneously. And

192
00:10:50,836 --> 00:10:52,916
Speaker 2: then we got to a certain point recently where we

193
00:10:52,956 --> 00:10:56,356
Speaker 2: can translate based on audio waves. And so if you

194
00:10:56,356 --> 00:11:01,756
Speaker 2: can translate audio waves, it becomes much easier to translate

195
00:11:01,796 --> 00:11:05,756
Speaker 2: vocalizations from people who don't use words at all, and

196
00:11:05,796 --> 00:11:08,556
Speaker 2: so she you know, you mentioned a lot of these

197
00:11:08,596 --> 00:11:11,476
Speaker 2: stories aren't done. This is the least done story, right,

198
00:11:11,756 --> 00:11:16,716
Speaker 2: And you know what's interesting is that technology has already

199
00:11:16,796 --> 00:11:19,996
Speaker 2: changed the lives of lots of autistic people, because you know,

200
00:11:20,076 --> 00:11:22,956
Speaker 2: she showed me her phone.

201
00:11:23,436 --> 00:11:23,636
Speaker 1: Right.

202
00:11:23,716 --> 00:11:26,996
Speaker 2: So we were having dinner on our porch, and her son, Felix,

203
00:11:27,036 --> 00:11:31,916
Speaker 2: who's lovely sweet in many ways, a normal American teenager,

204
00:11:31,996 --> 00:11:34,236
Speaker 2: just picked up the phone and went off to go

205
00:11:34,276 --> 00:11:36,276
Speaker 2: do something. And then he came back and he dropped

206
00:11:36,276 --> 00:11:38,876
Speaker 2: off the phone, and she and her husband showed me

207
00:11:38,956 --> 00:11:42,636
Speaker 2: photos that he had taken of what he's fascinated by.

208
00:11:43,316 --> 00:11:45,916
Speaker 2: And so some of them were hairbrush bristles that were

209
00:11:45,916 --> 00:11:49,636
Speaker 2: actually these these kind of beautiful plastic forests. He'd used

210
00:11:49,636 --> 00:11:52,996
Speaker 2: a filter and he'd taken them and hundreds of these photos.

211
00:11:53,436 --> 00:11:56,396
Speaker 2: And before that, he'd taken photos of windows and was

212
00:11:56,476 --> 00:11:59,196
Speaker 2: really interested in the way windows around their neighborhood worked,

213
00:11:59,796 --> 00:12:02,836
Speaker 2: or castles or sketches of castles. And she said, like,

214
00:12:02,956 --> 00:12:07,356
Speaker 2: before the iPhone, when he's staring at something, I would

215
00:12:07,396 --> 00:12:10,196
Speaker 2: have no idea what he would have been daring. But

216
00:12:10,276 --> 00:12:13,356
Speaker 2: here's this really rudimentary piece of technology, at least now

217
00:12:13,996 --> 00:12:16,996
Speaker 2: and through that I'm able to understand what he's looking at.

218
00:12:17,156 --> 00:12:21,116
Speaker 2: I don't necessarily understand why he's fascinated by it, but

219
00:12:21,276 --> 00:12:22,916
Speaker 2: I know that he's interested. And I can make his

220
00:12:22,996 --> 00:12:25,836
Speaker 2: life richer by exposing him to more of it, and

221
00:12:26,036 --> 00:12:28,196
Speaker 2: I can frustrate him less by making sure there's always

222
00:12:28,236 --> 00:12:32,196
Speaker 2: a hairbrush around. And so those kinds of things I

223
00:12:32,236 --> 00:12:35,516
Speaker 2: have to say, like they make me more optimistic that

224
00:12:35,596 --> 00:12:38,996
Speaker 2: if we can wrestle this technology, like you know, tech

225
00:12:39,036 --> 00:12:41,676
Speaker 2: doesn't always have to be bad, Like we're not faded

226
00:12:41,756 --> 00:12:44,996
Speaker 2: to the world sucking because technology is coming out, but

227
00:12:45,076 --> 00:12:46,716
Speaker 2: we do have to figure out how we want it

228
00:12:46,756 --> 00:12:50,676
Speaker 2: to work. And so I left Christy feeling like she's

229
00:12:50,716 --> 00:12:53,876
Speaker 2: going to get there, and I know that that's not guaranteed,

230
00:12:54,436 --> 00:12:58,036
Speaker 2: but she is so stubborn, she's so motivated by love

231
00:12:58,356 --> 00:13:01,196
Speaker 2: and by problem solving, and the tech is moving so

232
00:13:01,396 --> 00:13:04,516
Speaker 2: fast that she's been at this ten years, right, and

233
00:13:04,596 --> 00:13:07,236
Speaker 2: ten years she's only gotten so far. But the next

234
00:13:07,236 --> 00:13:09,996
Speaker 2: ten years are going to move so much faster, and

235
00:13:10,036 --> 00:13:11,916
Speaker 2: more people are going to come and attack the problem.

236
00:13:11,956 --> 00:13:15,156
Speaker 2: And it wouldn't surprise me if ten years from now

237
00:13:15,476 --> 00:13:16,956
Speaker 2: she's successful.

238
00:13:18,236 --> 00:13:22,396
Speaker 1: So another story you tell in your book is about

239
00:13:22,436 --> 00:13:26,556
Speaker 1: Operation Warp Speed, and in particular it's about the logistics

240
00:13:26,596 --> 00:13:29,836
Speaker 1: of getting vaccines out to the public. And you know,

241
00:13:29,956 --> 00:13:33,556
Speaker 1: I knew a moderate amount about the vaccine development side,

242
00:13:33,916 --> 00:13:36,076
Speaker 1: but I really didn't know anything about the getting the

243
00:13:36,156 --> 00:13:39,036
Speaker 1: vaccine out to the world side. And I certainly didn't

244
00:13:39,076 --> 00:13:42,516
Speaker 1: know that it was an AI story. So tell me

245
00:13:42,556 --> 00:13:44,956
Speaker 1: about Operation Warp Speed, and in particular about this guy

246
00:13:44,996 --> 00:13:48,356
Speaker 1: you write about, this general named Gus Purna.

247
00:13:49,516 --> 00:13:53,276
Speaker 2: Look, I was writing, and I was hungry for understanding

248
00:13:53,516 --> 00:13:56,316
Speaker 2: and trying to figure out how exactly machine learning and

249
00:13:56,436 --> 00:14:00,116
Speaker 2: artificial intelligence integrate into society and ways are helpful, and

250
00:14:00,156 --> 00:14:03,796
Speaker 2: so I kind of tripped across this video of Gus Perna,

251
00:14:03,956 --> 00:14:07,996
Speaker 2: who oversaw Operation Warp Speed, who is not a technologist

252
00:14:08,076 --> 00:14:11,956
Speaker 2: by any means, He's a logistician. He oversees, generally, you know,

253
00:14:12,516 --> 00:14:15,636
Speaker 2: or did before he retired, the delivery of munitions and

254
00:14:15,716 --> 00:14:18,556
Speaker 2: uniforms and food so that we can fight wars. And

255
00:14:18,596 --> 00:14:21,036
Speaker 2: he was placed in charge of Operation Warp Speed and

256
00:14:21,076 --> 00:14:23,436
Speaker 2: the delivery of it so that you know, hopefully Big

257
00:14:23,476 --> 00:14:27,916
Speaker 2: Pharma can can execute on all these vaccines. But then

258
00:14:27,956 --> 00:14:30,556
Speaker 2: what how do they get into plastic vials and how

259
00:14:30,556 --> 00:14:33,236
Speaker 2: do those vials get delivered to everything from a CBS

260
00:14:33,756 --> 00:14:36,756
Speaker 2: to a doctor's office to in rural areas, sometimes a

261
00:14:36,796 --> 00:14:41,716
Speaker 2: general store that has no digital footprint. And so he

262
00:14:41,756 --> 00:14:46,236
Speaker 2: shows up in Washington and discovers that there's no plan,

263
00:14:46,356 --> 00:14:50,556
Speaker 2: like he is the plan. And so in Washington, like

264
00:14:50,596 --> 00:14:54,596
Speaker 2: it or not, everything starts with consultants everything, and so

265
00:14:54,676 --> 00:14:56,876
Speaker 2: people come in and they pitch him on things, and again,

266
00:14:57,036 --> 00:15:00,156
Speaker 2: not a technical guy, but he keeps hearing these crazy

267
00:15:00,196 --> 00:15:03,036
Speaker 2: pitches over the course of an hour. Somebody's like, we're

268
00:15:03,036 --> 00:15:06,076
Speaker 2: going to put a medical blockchain together and in a

269
00:15:06,076 --> 00:15:09,476
Speaker 2: couple of years, and he's like, no. Somebody else pitched

270
00:15:09,516 --> 00:15:12,036
Speaker 2: like hardware, We'll just put a hardware device in every

271
00:15:12,076 --> 00:15:15,356
Speaker 2: doctor's and he knows enough to say no. And finally

272
00:15:15,356 --> 00:15:18,756
Speaker 2: he meets with these two reasonably seeing people who are like, look,

273
00:15:19,036 --> 00:15:21,556
Speaker 2: you have a data problem. This can all be solved

274
00:15:21,676 --> 00:15:24,676
Speaker 2: if we get the right number of data streams flowing

275
00:15:24,676 --> 00:15:27,596
Speaker 2: into the right interface so that you can actually see

276
00:15:27,956 --> 00:15:33,196
Speaker 2: everything from plastic vials, metal needles, refrigerated trucks because remember,

277
00:15:33,796 --> 00:15:37,636
Speaker 2: half of the vaccines needed to be really cold refrigerated

278
00:15:37,676 --> 00:15:41,396
Speaker 2: for delivery. To track which states might want Johnson and

279
00:15:41,476 --> 00:15:44,556
Speaker 2: Johnson versus like you have to build a whole civilization

280
00:15:44,636 --> 00:15:46,036
Speaker 2: to do this stuff. And they said, look, you can

281
00:15:46,036 --> 00:15:50,956
Speaker 2: build a civilization in data, probably won't cost that much,

282
00:15:51,276 --> 00:15:54,436
Speaker 2: but we will need you to solve the human problem,

283
00:15:54,756 --> 00:15:57,076
Speaker 2: which is getting these various agencies to play ball, so

284
00:15:57,116 --> 00:15:59,356
Speaker 2: we can actually get to the data, clean it up,

285
00:15:59,476 --> 00:16:01,596
Speaker 2: and feed it to you. And so that's a machine

286
00:16:01,636 --> 00:16:04,476
Speaker 2: learning story. And when I went down the rabbit hole

287
00:16:04,516 --> 00:16:07,276
Speaker 2: of how it actually happened, it was really instructive because

288
00:16:07,356 --> 00:16:11,556
Speaker 2: obviously the end state was a huge success, underrated.

289
00:16:11,916 --> 00:16:15,076
Speaker 1: Every chance I get, I want to scream about how underrated,

290
00:16:15,076 --> 00:16:18,916
Speaker 1: worse feed is weirdly politically orphaned, right, Like the Republicans

291
00:16:19,316 --> 00:16:21,876
Speaker 1: decided they don't like vaccines even though they did it.

292
00:16:21,876 --> 00:16:24,076
Speaker 1: It was like, as far as I'm concerned, the great

293
00:16:24,076 --> 00:16:27,116
Speaker 1: success of Trump's first administration, and then because it was Trump,

294
00:16:27,156 --> 00:16:28,956
Speaker 1: the Democrats can't say it was great.

295
00:16:28,756 --> 00:16:31,796
Speaker 2: Right exactly, And so it's wandering out there, this great

296
00:16:31,836 --> 00:16:34,916
Speaker 2: American success story that just got buried by the stupidity

297
00:16:34,916 --> 00:16:38,356
Speaker 2: of our politics. But we did it, and we wouldn't

298
00:16:38,356 --> 00:16:41,236
Speaker 2: have been able to do it without really successful machine

299
00:16:41,316 --> 00:16:41,996
Speaker 2: learning and AI.

300
00:16:42,716 --> 00:16:45,116
Speaker 1: So how did it work? Like why was that an

301
00:16:45,116 --> 00:16:46,556
Speaker 1: AI problem? And what did they actually do.

302
00:16:46,916 --> 00:16:49,396
Speaker 2: So it was a data problem, but also an actionable

303
00:16:49,476 --> 00:16:53,956
Speaker 2: data problem. And so there's a company called Pounteer, very

304
00:16:54,116 --> 00:16:57,156
Speaker 2: very provocative company. A lot of people respond poorly to

305
00:16:57,236 --> 00:17:01,516
Speaker 2: its name, and it has this kind of mystical aura

306
00:17:01,556 --> 00:17:03,916
Speaker 2: to it, in part because it was founded by Peter

307
00:17:03,996 --> 00:17:07,516
Speaker 2: Chiel and co funded by the CIA, and in part

308
00:17:07,516 --> 00:17:11,556
Speaker 2: because it's named for the stones in the Tolkien trilogy.

309
00:17:12,036 --> 00:17:13,956
Speaker 2: But really all they do is plumbing.

310
00:17:14,356 --> 00:17:18,516
Speaker 1: I am not reflexively anti tech or anti company. I

311
00:17:18,556 --> 00:17:21,676
Speaker 1: have some amount of wariness toward Palenteer, I guess because

312
00:17:21,716 --> 00:17:26,156
Speaker 1: I associate them with domestic surveillance at some level, like

313
00:17:26,196 --> 00:17:28,396
Speaker 1: that piece of it I mean, but I don't know

314
00:17:28,436 --> 00:17:30,316
Speaker 1: it that well. I've never interviewed somebody from Palenter. I

315
00:17:30,316 --> 00:17:32,716
Speaker 1: haven't read a book about them. So, like, tell me

316
00:17:32,716 --> 00:17:36,596
Speaker 1: about Palenteer as a company. I guess, first of all,

317
00:17:36,596 --> 00:17:39,916
Speaker 1: like the side that makes people weary, and then I'm

318
00:17:39,916 --> 00:17:42,716
Speaker 1: curious more generally how they operate, what differentiates them, what

319
00:17:42,756 --> 00:17:44,236
Speaker 1: they actually do as a tech company.

320
00:17:44,556 --> 00:17:48,156
Speaker 2: Yeah. So as a company, they're created in two thousand

321
00:17:48,156 --> 00:17:50,796
Speaker 2: and three by two friends who have no business being friends.

322
00:17:50,796 --> 00:17:53,156
Speaker 2: One is Peter Thiel, who I think most of your

323
00:17:53,196 --> 00:17:57,876
Speaker 2: listeners really know about. He's a sort of largely libertarian Republican,

324
00:17:58,156 --> 00:18:02,516
Speaker 2: big Trump donor, has very controversial stances on pretty much

325
00:18:02,556 --> 00:18:03,836
Speaker 2: every possible.

326
00:18:03,436 --> 00:18:07,236
Speaker 1: Issue, including the Antichrist, weirdly including the anti Christ.

327
00:18:07,316 --> 00:18:10,036
Speaker 2: He really finds a way. His friend is a guy

328
00:18:10,116 --> 00:18:12,876
Speaker 2: named Alex Karp, who is kind of his polar opposite.

329
00:18:12,916 --> 00:18:17,956
Speaker 2: He's New York born, he's half black, half Jewish, largely

330
00:18:17,956 --> 00:18:21,716
Speaker 2: identifies as a socialist. A big donor to Hillary Clinton

331
00:18:21,756 --> 00:18:25,876
Speaker 2: and Kamala Harris. They met in law school. They argued

332
00:18:25,876 --> 00:18:28,836
Speaker 2: all the time. They agreed about the need for this

333
00:18:28,996 --> 00:18:31,476
Speaker 2: kind of technology, and the only other thing they agree

334
00:18:31,476 --> 00:18:34,436
Speaker 2: on is that in general, the world is better when

335
00:18:34,476 --> 00:18:38,316
Speaker 2: America is a strong actor and American values like freedom

336
00:18:38,596 --> 00:18:40,876
Speaker 2: are or propagated throughout the world. So that's what they

337
00:18:40,916 --> 00:18:41,316
Speaker 2: agree on.

338
00:18:41,636 --> 00:18:45,436
Speaker 1: Do you think that the liberal antagonism toward Palenteer is

339
00:18:45,476 --> 00:18:47,476
Speaker 1: misplaced or do you think it's valid in some ways?

340
00:18:48,436 --> 00:18:52,316
Speaker 2: It's incredibly complicated because it has now grown to include

341
00:18:52,356 --> 00:18:55,756
Speaker 2: not just surveillance, but Alex Karp's position on Israel, which

342
00:18:55,796 --> 00:18:58,636
Speaker 2: is that Israel should be supported by Palanteer products, just

343
00:18:58,676 --> 00:19:02,036
Speaker 2: as Ukraine should be. And so there's lots of different

344
00:19:02,276 --> 00:19:05,516
Speaker 2: valances where people can find something they don't like about Palanteer.

345
00:19:05,756 --> 00:19:07,716
Speaker 2: And by the way, I respect all of them.

346
00:19:08,036 --> 00:19:12,396
Speaker 1: Yeah, and so just as a tech company, what is

347
00:19:12,436 --> 00:19:15,796
Speaker 1: Palanteer's thing, What do they do? And in particular, what

348
00:19:15,876 --> 00:19:18,796
Speaker 1: did they do in Operation Warp Speed to get vaccines

349
00:19:18,836 --> 00:19:19,556
Speaker 1: out to the public.

350
00:19:20,076 --> 00:19:22,156
Speaker 2: I spoke to their chief architect and he's like, look,

351
00:19:22,236 --> 00:19:25,436
Speaker 2: we're the mole people of Silicon Valley. We take these

352
00:19:25,516 --> 00:19:28,956
Speaker 2: data streams, we lay them out, we clean them, We

353
00:19:29,116 --> 00:19:32,756
Speaker 2: integrate the data from various different places, so like, for instance,

354
00:19:32,756 --> 00:19:35,716
Speaker 2: a CVS and a right aid and a mom and

355
00:19:35,756 --> 00:19:38,916
Speaker 2: pop place and a trucking company, and we put them

356
00:19:38,916 --> 00:19:42,196
Speaker 2: together so that on an iPad or a laptop someone

357
00:19:42,276 --> 00:19:44,396
Speaker 2: could play the pandemic like a video game.

358
00:19:44,596 --> 00:19:47,236
Speaker 1: Right, play it like a video game, meaning there's like

359
00:19:48,076 --> 00:19:50,956
Speaker 1: a crazy dashboard, so you can see here's where the

360
00:19:50,956 --> 00:19:53,596
Speaker 1: people are, and here's where the vials are, and here's

361
00:19:53,596 --> 00:19:55,436
Speaker 1: where the vaccine is, and here's where the trucks are,

362
00:19:55,436 --> 00:19:57,356
Speaker 1: and we need to get these trucks to this vaccine

363
00:19:57,396 --> 00:19:59,476
Speaker 1: to these people. That's the exact video game you're playing.

364
00:19:59,716 --> 00:20:03,716
Speaker 2: Yeah, and constantly up to date and constantly accurate, and

365
00:20:04,076 --> 00:20:06,676
Speaker 2: literally an administrator can sit there and just move things

366
00:20:06,716 --> 00:20:08,916
Speaker 2: around and know that they're actually going to the right place.

367
00:20:09,196 --> 00:20:12,756
Speaker 1: Yes, a wildly hard problem. Like it's not sexy, like

368
00:20:12,956 --> 00:20:16,636
Speaker 1: developing a vaccine, and we've heard a lot about mRNA vaccines,

369
00:20:16,636 --> 00:20:19,796
Speaker 1: but like figuring out where the actual vaccine is is

370
00:20:20,676 --> 00:20:21,956
Speaker 1: totally non trivial health.

371
00:20:22,156 --> 00:20:24,756
Speaker 2: Yeah, it's a huge deal. And so there's two problems

372
00:20:24,796 --> 00:20:27,476
Speaker 2: you're solving right. One is the technical problem, and that

373
00:20:27,556 --> 00:20:31,196
Speaker 2: involves a lot of really dull, really unglamorous work, which

374
00:20:31,236 --> 00:20:34,436
Speaker 2: is like trying to figure out why this pipe you

375
00:20:34,476 --> 00:20:36,956
Speaker 2: know it's because it is pipe coming from a distribution

376
00:20:37,156 --> 00:20:40,156
Speaker 2: center is broken, and so AI can attack that.

377
00:20:40,436 --> 00:20:42,316
Speaker 1: When you say pipe, do you mean the pipe through

378
00:20:42,316 --> 00:20:44,276
Speaker 1: which data is flowing? What do you mean pipe in

379
00:20:44,316 --> 00:20:44,956
Speaker 1: this instance?

380
00:20:45,236 --> 00:20:49,156
Speaker 2: So, to build a data pipeline is basically a work

381
00:20:49,196 --> 00:20:52,356
Speaker 2: of code. But everybody's got lots of different code, and

382
00:20:52,436 --> 00:20:54,476
Speaker 2: so when your data pipelines are built in many many

383
00:20:54,516 --> 00:20:58,876
Speaker 2: different codes to many different standards, they have to be standardized.

384
00:20:59,036 --> 00:21:01,396
Speaker 2: You have to be able to see, oh, okay, this

385
00:21:01,516 --> 00:21:04,476
Speaker 2: data coming from CBS matches up to the data coming

386
00:21:04,516 --> 00:21:08,876
Speaker 2: from the trucking company. It's clean, it's modern. Most of

387
00:21:08,916 --> 00:21:11,596
Speaker 2: the that they were dealing with in fact was not

388
00:21:11,716 --> 00:21:15,156
Speaker 2: modern in any way. It was created a decade ago

389
00:21:16,116 --> 00:21:20,036
Speaker 2: using old code and often had bad information, and so

390
00:21:20,156 --> 00:21:22,476
Speaker 2: you need to standardize it, you need to clean it.

391
00:21:22,796 --> 00:21:25,116
Speaker 2: And then where it really gets fun is you can

392
00:21:25,276 --> 00:21:30,236
Speaker 2: use those pieces of information to recommend actions to note

393
00:21:30,276 --> 00:21:33,556
Speaker 2: that like, oh, you know you're overly distributing in one area,

394
00:21:33,636 --> 00:21:37,316
Speaker 2: you could be moving here. In fact, this rural area,

395
00:21:37,476 --> 00:21:41,316
Speaker 2: for instance, might really want the one shot doses, whereas

396
00:21:42,116 --> 00:21:43,556
Speaker 2: an urban area might want too.

397
00:21:43,876 --> 00:21:47,876
Speaker 1: And those suggestions might be coming from the machine learning model,

398
00:21:47,876 --> 00:21:48,916
Speaker 1: from the AI basic.

399
00:21:48,836 --> 00:21:52,236
Speaker 2: Correct, so that while you know, the very overtaxed humans

400
00:21:52,236 --> 00:21:56,196
Speaker 2: dealing with political problems and interagency problems are on the phone,

401
00:21:56,756 --> 00:21:58,876
Speaker 2: the AI is just kind of nudging them toward like

402
00:21:58,996 --> 00:22:02,116
Speaker 2: easier ways to make this whole project happen. So like

403
00:22:02,156 --> 00:22:05,116
Speaker 2: the technical is one problem, and then something that's kind

404
00:22:05,116 --> 00:22:08,516
Speaker 2: of a light motif of your show is the humans

405
00:22:08,556 --> 00:22:12,276
Speaker 2: are another. Because we create our own systems and there

406
00:22:12,316 --> 00:22:15,876
Speaker 2: is nothing more complicated than the federal bureaucracy. And I

407
00:22:15,916 --> 00:22:19,116
Speaker 2: mean that oftentimes with admiration and sometimes not with admiration.

408
00:22:19,676 --> 00:22:24,116
Speaker 2: And so inside of the CDC are many different agencies,

409
00:22:24,516 --> 00:22:28,156
Speaker 2: and a lot of them were even in COVID had

410
00:22:28,276 --> 00:22:29,196
Speaker 2: very sharp elbows.

411
00:22:29,276 --> 00:22:29,636
Speaker 1: Yeah.

412
00:22:29,716 --> 00:22:32,876
Speaker 2: They didn't want the Department of Defense, which was nominally

413
00:22:32,916 --> 00:22:36,476
Speaker 2: overseeing this effort, getting involved. They thought, no, this is

414
00:22:36,516 --> 00:22:40,516
Speaker 2: our territory. What do you possibly know about this? And

415
00:22:40,596 --> 00:22:44,076
Speaker 2: General Perna was so important because he really didn't know

416
00:22:44,116 --> 00:22:46,956
Speaker 2: that much about technology. He really didn't know that much

417
00:22:46,956 --> 00:22:51,036
Speaker 2: about vaccines. What he knew about was getting stuff done.

418
00:22:51,636 --> 00:22:54,036
Speaker 1: Yeah. You tell the story of a moment when he's

419
00:22:54,276 --> 00:22:58,236
Speaker 1: arguing with the governor about whether some amount of vaccine

420
00:22:58,276 --> 00:23:01,356
Speaker 1: has been delivered to the governor state. Tell me, tell

421
00:23:01,356 --> 00:23:01,796
Speaker 1: me about that.

422
00:23:02,316 --> 00:23:06,116
Speaker 2: Yeah. So, once the vaccines started to get delivered, this

423
00:23:06,156 --> 00:23:08,716
Speaker 2: is where the video game element, the dashboard Elmont really

424
00:23:08,716 --> 00:23:12,156
Speaker 2: became an important because they could see trucks pulling up

425
00:23:12,356 --> 00:23:16,236
Speaker 2: to a registered refrigeration point or warehouse point in every state.

426
00:23:16,916 --> 00:23:19,716
Speaker 2: And there were states that were really eager to get

427
00:23:19,716 --> 00:23:23,076
Speaker 2: the vaccines, and the governors felt that their entire political

428
00:23:23,076 --> 00:23:26,076
Speaker 2: futures depended on getting it and distributing it and being

429
00:23:26,116 --> 00:23:28,996
Speaker 2: able to tell their citizens like, hey, Massachusetts, we got it,

430
00:23:29,316 --> 00:23:31,756
Speaker 2: we got it. You can go to these places. Remember,

431
00:23:31,756 --> 00:23:34,676
Speaker 2: there are people, you know, frantically refreshing their browsers to

432
00:23:34,716 --> 00:23:37,436
Speaker 2: try and find weights up to three four weeks, and

433
00:23:37,476 --> 00:23:41,236
Speaker 2: so somebody would say, hey, you promised me this dosage.

434
00:23:41,316 --> 00:23:43,436
Speaker 2: I don't have it. And he said, oh, but you do.

435
00:23:43,876 --> 00:23:46,236
Speaker 2: Here's who signed for it. Go get them. I'm going

436
00:23:46,316 --> 00:23:49,236
Speaker 2: to stay right here. You go get them. And so

437
00:23:49,356 --> 00:23:51,796
Speaker 2: what he, you know, also did and he had plenty

438
00:23:51,796 --> 00:23:54,276
Speaker 2: of people working for him who help with this. These

439
00:23:54,316 --> 00:23:57,916
Speaker 2: systems need accountability. And what I think, what I saw

440
00:23:57,996 --> 00:24:00,356
Speaker 2: time and again in reporting on the book, is that

441
00:24:00,396 --> 00:24:05,756
Speaker 2: the accountability is human based. We respond to human beings saying, actually,

442
00:24:05,756 --> 00:24:08,196
Speaker 2: I'm calling your bluff because I can see the answer here.

443
00:24:08,356 --> 00:24:08,676
Speaker 1: Uh huh.

444
00:24:09,196 --> 00:24:11,476
Speaker 2: Actually, I'm gonna follow up with you because you said

445
00:24:11,476 --> 00:24:13,196
Speaker 2: you would do this pilot with us and you didn't.

446
00:24:13,636 --> 00:24:15,756
Speaker 2: And if you don't have the right human in that chair,

447
00:24:15,916 --> 00:24:19,636
Speaker 2: with the right temperament, the right force, the right seductive charm,

448
00:24:19,676 --> 00:24:22,516
Speaker 2: and that's what's called for, it's going to fail. And

449
00:24:22,556 --> 00:24:26,316
Speaker 2: so I can't say enough. Like my conversations with Perna,

450
00:24:26,436 --> 00:24:28,276
Speaker 2: I just came away thinking we should put him in

451
00:24:28,356 --> 00:24:29,356
Speaker 2: charge of a lot of stuff.

452
00:24:33,596 --> 00:24:46,036
Speaker 1: We'll be back in just a minute. So let's talk

453
00:24:46,036 --> 00:24:48,596
Speaker 1: about the Cleveland clinic. Your book has start of divided

454
00:24:48,636 --> 00:24:51,596
Speaker 1: into sections. There's a warp speed section. There's a healthcare section.

455
00:24:52,156 --> 00:24:56,076
Speaker 1: Healthcare sections basically the Cleveland Clinic. Why did you choose

456
00:24:56,276 --> 00:24:58,596
Speaker 1: for your healthcare section the Cleveland Clinic?

457
00:25:00,036 --> 00:25:03,196
Speaker 2: Mostly because they asked me to come in, and so

458
00:25:03,676 --> 00:25:04,036
Speaker 2: I was.

459
00:25:04,036 --> 00:25:07,076
Speaker 1: Really clicking phrase kings under the street light. Yeah.

460
00:25:07,076 --> 00:25:09,676
Speaker 2: And also, listen, healthcare is a universe, right, So what

461
00:25:09,756 --> 00:25:11,916
Speaker 2: I really wanted to explore is as much of that

462
00:25:12,036 --> 00:25:14,316
Speaker 2: universe as possible. Cleveland Clinic is one of the top

463
00:25:14,396 --> 00:25:17,756
Speaker 2: systems in the world and the United States, and they're

464
00:25:17,796 --> 00:25:21,276
Speaker 2: run by a really interesting guy named doctor Thomas Mihaliovich.

465
00:25:21,836 --> 00:25:23,596
Speaker 2: And so what he told me at the beginning is

466
00:25:24,076 --> 00:25:27,036
Speaker 2: I want you to wander around because what we do

467
00:25:27,116 --> 00:25:28,556
Speaker 2: with AI is going to be different than what a

468
00:25:28,556 --> 00:25:31,436
Speaker 2: lot of other people do. First of all, we actually

469
00:25:31,636 --> 00:25:34,396
Speaker 2: have one major priority, which is the treatment of our patients.

470
00:25:34,996 --> 00:25:37,236
Speaker 2: And so that means that when technology comes to the

471
00:25:37,236 --> 00:25:40,916
Speaker 2: Cleveland Clinic, it's not run by technologists. It's run by doctors.

472
00:25:41,476 --> 00:25:43,756
Speaker 2: And that is going to eliminate certain partners right off

473
00:25:43,796 --> 00:25:47,556
Speaker 2: the bat, because most software folks believe that software is

474
00:25:47,596 --> 00:25:50,076
Speaker 2: the most powerful force in the world, and they don't

475
00:25:50,076 --> 00:25:53,556
Speaker 2: want to explain themselves or work within some sort of

476
00:25:53,596 --> 00:25:56,996
Speaker 2: old fashioned medical system. I spent a fair amount of

477
00:25:57,036 --> 00:25:59,836
Speaker 2: time with the chief technical officer, who had come from

478
00:25:59,916 --> 00:26:03,076
Speaker 2: outside of healthcare and was recruited specifically from outside of it,

479
00:26:03,636 --> 00:26:07,476
Speaker 2: and he obviously knew tons about technology, but knew very

480
00:26:07,476 --> 00:26:08,476
Speaker 2: little about healthcare.

481
00:26:09,196 --> 00:26:11,516
Speaker 1: I'm somewhat frustrated in your book.

482
00:26:12,316 --> 00:26:15,756
Speaker 2: Yeah, he is frustrated because he's used to being in charge,

483
00:26:16,236 --> 00:26:19,236
Speaker 2: and what he learned in his first couple of years

484
00:26:19,356 --> 00:26:21,956
Speaker 2: is like, I'm not in charge.

485
00:26:21,956 --> 00:26:25,436
Speaker 1: And like every doctor is in charge in a weird way, right,

486
00:26:25,516 --> 00:26:28,876
Speaker 1: one of the maybe distinctive things about healthcare. And I

487
00:26:28,876 --> 00:26:30,836
Speaker 1: know doctors don't see it that way, but doctors have

488
00:26:30,916 --> 00:26:33,076
Speaker 1: a lot of autonomy and a lot of them don't

489
00:26:33,116 --> 00:26:34,756
Speaker 1: want AI.

490
00:26:35,196 --> 00:26:37,076
Speaker 2: They have a ton of autonomy, but they also have

491
00:26:37,156 --> 00:26:40,116
Speaker 2: a ton of authority. Doctor. You know, medicine is a

492
00:26:40,196 --> 00:26:43,076
Speaker 2: hierarchical business, and you want it to be that way

493
00:26:43,156 --> 00:26:45,716
Speaker 2: when you're the patient. You want someone who actually says,

494
00:26:45,876 --> 00:26:47,796
Speaker 2: I see all the data, I have all the inputs,

495
00:26:48,436 --> 00:26:50,636
Speaker 2: here's what we're going to go do, and then distributes

496
00:26:50,676 --> 00:26:54,036
Speaker 2: the tasks back. Technology doesn't work all that well that way,

497
00:26:54,116 --> 00:26:57,956
Speaker 2: and so ro heat Chandra came in and he's basically like.

498
00:26:58,116 --> 00:27:00,516
Speaker 1: This is this is the chief technology this is yeah,

499
00:27:00,516 --> 00:27:02,116
Speaker 1: this CTO, and.

500
00:27:02,076 --> 00:27:04,876
Speaker 2: He's like, I'm supposed to help turn this place around

501
00:27:04,876 --> 00:27:08,356
Speaker 2: with technology. The business is terrible. Our margins. You know,

502
00:27:08,436 --> 00:27:10,916
Speaker 2: Clinton Clinic is actually one of the more successful nonprofit

503
00:27:10,956 --> 00:27:14,596
Speaker 2: hospital systems, and it's got it two percent margin, which

504
00:27:14,756 --> 00:27:18,116
Speaker 2: most people would find horrifying. And so he goes at

505
00:27:18,156 --> 00:27:20,836
Speaker 2: it like a technologist, and time and again he runs

506
00:27:20,836 --> 00:27:24,996
Speaker 2: into this wall, which is doctor saying well, I don't

507
00:27:24,996 --> 00:27:26,876
Speaker 2: want to do it that way. Yeah, and sometimes with

508
00:27:26,956 --> 00:27:30,316
Speaker 2: good reason and sometimes not. But again, you can't just

509
00:27:30,356 --> 00:27:32,796
Speaker 2: solve the technical problem. If you want AI to actually

510
00:27:32,836 --> 00:27:34,836
Speaker 2: work for good, you have to solve both the technical

511
00:27:35,076 --> 00:27:37,476
Speaker 2: and the structural problem of what you're trying to fix.

512
00:27:38,436 --> 00:27:41,516
Speaker 1: So actually, at the Cleveland Clinic, you encountered this company

513
00:27:41,596 --> 00:27:45,196
Speaker 1: called Baesian Health, And as it happens, I interviewed the

514
00:27:45,236 --> 00:27:47,596
Speaker 1: founder of that company, Suci Saria, a few years ago

515
00:27:47,676 --> 00:27:51,396
Speaker 1: on this show, and they built this system where they

516
00:27:51,516 --> 00:27:56,036
Speaker 1: use AI to flag when patients are likely developing sepsis.

517
00:27:56,756 --> 00:27:58,156
Speaker 1: And you know, that was a few years ago that

518
00:27:58,196 --> 00:28:00,516
Speaker 1: I talked to her, and so I'm curious, how's it going,

519
00:28:00,516 --> 00:28:02,996
Speaker 1: How is it working in the world. At the Cleveland clinic.

520
00:28:03,596 --> 00:28:07,356
Speaker 2: Yeah, so's she's brilliant. And I talked to her as well,

521
00:28:07,396 --> 00:28:10,796
Speaker 2: and from a technical perspective of her ability to understand

522
00:28:10,796 --> 00:28:13,476
Speaker 2: not just what sepsis is, but again to create a

523
00:28:13,516 --> 00:28:16,476
Speaker 2: bunch of data streams and integrate them and create weights

524
00:28:16,556 --> 00:28:20,676
Speaker 2: between the streams, to create a flag for doctors to say, hey,

525
00:28:20,916 --> 00:28:24,036
Speaker 2: this might be sepsis. And just as a reminder, sepsis

526
00:28:24,116 --> 00:28:28,116
Speaker 2: kills more people every year than breast cancer, prostate cancer,

527
00:28:28,196 --> 00:28:30,996
Speaker 2: and the opioid crisis. Yeah, three hundred and fifty thousand

528
00:28:30,996 --> 00:28:33,636
Speaker 2: Americans a year, and it happens in hospitals. And so

529
00:28:33,716 --> 00:28:36,876
Speaker 2: the clinic, which loses in twenty twenty one, I think

530
00:28:37,116 --> 00:28:40,356
Speaker 2: lost about three thousand patients to sepsis, said well, we can't.

531
00:28:40,716 --> 00:28:43,356
Speaker 2: This is crazy. It may be the industry standard, but

532
00:28:43,396 --> 00:28:46,356
Speaker 2: it's crazy. So they brought in beision. I will tell

533
00:28:46,396 --> 00:28:48,956
Speaker 2: you that the founder, when first approached by the CTO

534
00:28:48,956 --> 00:28:51,916
Speaker 2: of the Cleveland clinic, said yeah, I'm not sure we're interested.

535
00:28:53,076 --> 00:28:55,356
Speaker 2: She said, you guys are a really big system. Your

536
00:28:55,396 --> 00:28:58,276
Speaker 2: doctors are going to be resistant. You're kind of a pain.

537
00:28:59,116 --> 00:29:01,236
Speaker 2: Prove to me that you actually want to implement this

538
00:29:01,316 --> 00:29:04,196
Speaker 2: in your hospital. And the CTO, who has a similar

539
00:29:04,236 --> 00:29:08,476
Speaker 2: attitude about medicine. Was actually like, oh, I see a

540
00:29:08,516 --> 00:29:11,596
Speaker 2: kindred spirit. I'm going to talk up the fact that

541
00:29:11,636 --> 00:29:13,356
Speaker 2: our doctors are committed and we're going to do this.

542
00:29:13,996 --> 00:29:17,356
Speaker 2: And so they went into Cleveland Clinic and they really

543
00:29:17,436 --> 00:29:19,636
Speaker 2: piloted for a year, and it was a hard year.

544
00:29:19,836 --> 00:29:21,956
Speaker 2: They had to figure out how it works, not just

545
00:29:22,276 --> 00:29:26,356
Speaker 2: in community hospitals in an emergency room in other places,

546
00:29:26,356 --> 00:29:29,756
Speaker 2: but like in the ICU and the Cleveland Clinic. ICU

547
00:29:29,876 --> 00:29:35,156
Speaker 2: is different even than most other healthcare ICUs because it's

548
00:29:35,196 --> 00:29:38,516
Speaker 2: a world famous institution. People come in from all over

549
00:29:38,556 --> 00:29:42,316
Speaker 2: the world, and a lot of sepsis looks exactly like

550
00:29:42,436 --> 00:29:45,876
Speaker 2: a cardiac infection, looks exactly like sixteen other things that

551
00:29:45,916 --> 00:29:48,356
Speaker 2: they're all flagging for. And so if you're an ICU

552
00:29:48,356 --> 00:29:51,596
Speaker 2: and doctor or nurse, your job is basically responding to

553
00:29:51,636 --> 00:29:55,836
Speaker 2: beeps coming from machinery near your patient. It can be

554
00:29:55,876 --> 00:29:58,356
Speaker 2: a huge pain in the ass. And so when a

555
00:29:58,396 --> 00:30:01,996
Speaker 2: sepsis flag isn't properly calibrated, when it flags too much

556
00:30:02,036 --> 00:30:04,716
Speaker 2: or flags too little, it's actually not doing its job.

557
00:30:05,036 --> 00:30:08,196
Speaker 2: So even though everyone may agree on the premise, which is, hey,

558
00:30:08,236 --> 00:30:10,956
Speaker 2: we want to reduce more metality, from sepsis because it's

559
00:30:11,236 --> 00:30:15,076
Speaker 2: twenty twenty six or whatever. You know, you think about

560
00:30:15,116 --> 00:30:17,916
Speaker 2: your own job, You think about the own annoyances you

561
00:30:17,996 --> 00:30:20,676
Speaker 2: have in getting your job done, and even though it's

562
00:30:20,756 --> 00:30:24,316
Speaker 2: life or death, those annoyances mount. And so the software

563
00:30:24,436 --> 00:30:27,156
Speaker 2: has to work within the way doctors work, or else

564
00:30:27,196 --> 00:30:28,516
Speaker 2: it's not going to get used well.

565
00:30:28,556 --> 00:30:31,396
Speaker 1: And I mean, just to be clear, it was a

566
00:30:31,396 --> 00:30:34,116
Speaker 1: little bit unclear to me in the book how good

567
00:30:34,196 --> 00:30:37,036
Speaker 1: the software is in that context, like.

568
00:30:38,516 --> 00:30:39,396
Speaker 2: Is it good enough?

569
00:30:39,516 --> 00:30:42,076
Speaker 1: Is part of the problem that there's too many false positives,

570
00:30:42,116 --> 00:30:43,956
Speaker 1: which is a real problem.

571
00:30:44,076 --> 00:30:47,556
Speaker 2: Not outside of the ICU. Okay, the ICU is the

572
00:30:47,836 --> 00:30:51,556
Speaker 2: is the last frontier because everything is so complicated in

573
00:30:51,596 --> 00:30:52,156
Speaker 2: an ICU.

574
00:30:52,316 --> 00:30:57,276
Speaker 1: Yeah, it's the hardest diagnosis essentially exactly. So outside of

575
00:30:57,316 --> 00:31:01,196
Speaker 1: the ICU, is there like empirical benefit, are fewer people

576
00:31:01,276 --> 00:31:02,276
Speaker 1: dying of sepsis?

577
00:31:02,596 --> 00:31:06,476
Speaker 2: Well? The clinic reported a forty one percent drop in

578
00:31:06,556 --> 00:31:10,316
Speaker 2: mortality over the course of a year using the software. Now,

579
00:31:10,396 --> 00:31:13,396
Speaker 2: some of it was because they had drawn everyone's attention

580
00:31:13,476 --> 00:31:16,556
Speaker 2: to the reduction of sepsis. It's the Hawthorne effect when

581
00:31:16,596 --> 00:31:19,596
Speaker 2: you say we're going to do this, and everyone's attention shifts.

582
00:31:20,236 --> 00:31:23,316
Speaker 2: But people are realizing that that sepsis is a very

583
00:31:23,396 --> 00:31:26,156
Speaker 2: perfect for AI kind of problem because it has all

584
00:31:26,236 --> 00:31:30,676
Speaker 2: of these inputs that are sometimes conflicting and so much

585
00:31:30,716 --> 00:31:34,876
Speaker 2: data that it can overwhelm a caregiver. And AI doesn't

586
00:31:34,876 --> 00:31:37,596
Speaker 2: get overwhelmed by too much data. It's really great at

587
00:31:37,636 --> 00:31:41,756
Speaker 2: processing data at scale, very very quickly. What it often

588
00:31:42,676 --> 00:31:45,796
Speaker 2: isn't as great at is figuring out how to tell people.

589
00:31:45,516 --> 00:31:50,556
Speaker 1: About it, especially probabilistically. Right that like correct, the AI

590
00:31:50,676 --> 00:31:53,956
Speaker 1: is very good at pattern matching and making a probabilistic assessment,

591
00:31:55,676 --> 00:31:59,196
Speaker 1: but that last piece seems like the hard to figure

592
00:31:59,196 --> 00:32:00,276
Speaker 1: out for humans piece.

593
00:32:00,756 --> 00:32:04,036
Speaker 2: Yeah, And so what Beijin did, much to their credit,

594
00:32:04,236 --> 00:32:07,076
Speaker 2: is they heard the caregivers and they said, look, it's

595
00:32:07,076 --> 00:32:09,436
Speaker 2: not good enough just to send a flag. I get

596
00:32:09,476 --> 00:32:13,196
Speaker 2: flags all day, Tell me why explain why this flagged

597
00:32:13,236 --> 00:32:16,676
Speaker 2: on this patient. And what Cleveland Clinic did in the ICU,

598
00:32:16,756 --> 00:32:20,796
Speaker 2: particularly because they're very aware that this was the hardest place,

599
00:32:21,236 --> 00:32:24,956
Speaker 2: they basically just set up a workstation. They called it

600
00:32:24,996 --> 00:32:27,356
Speaker 2: a command center. Command center is a little glorious for

601
00:32:27,396 --> 00:32:27,876
Speaker 2: what I saw.

602
00:32:27,996 --> 00:32:28,596
Speaker 1: It was a PC.

603
00:32:29,276 --> 00:32:32,116
Speaker 2: There's a lovely woman named Dana operating it. She would

604
00:32:32,156 --> 00:32:34,596
Speaker 2: look at the patients in the ICU. She would see

605
00:32:34,596 --> 00:32:36,796
Speaker 2: the flag, she would hover over the flag, and the

606
00:32:36,796 --> 00:32:41,636
Speaker 2: flag would say I flagged this because I'm noticing X, Y,

607
00:32:41,676 --> 00:32:45,156
Speaker 2: and Z, and so she would say, Okay, Well, what

608
00:32:45,236 --> 00:32:48,396
Speaker 2: I'm seeing when I go look at the patient is

609
00:32:48,676 --> 00:32:51,356
Speaker 2: they're actually okay, and I can see why the AI

610
00:32:51,516 --> 00:32:54,196
Speaker 2: might be confused about this. I'll keep an eye on it.

611
00:32:54,676 --> 00:32:57,756
Speaker 2: Other times it was very clear that like, yep, you're

612
00:32:57,796 --> 00:33:00,476
Speaker 2: seeing what we're seeing. This is validating what we're seeing.

613
00:33:00,556 --> 00:33:02,916
Speaker 2: And occasionally it would flag something and she would go

614
00:33:02,916 --> 00:33:05,196
Speaker 2: in and say to the caregivers, we have to do

615
00:33:05,236 --> 00:33:05,996
Speaker 2: a check right now.

616
00:33:07,916 --> 00:33:11,036
Speaker 1: When you step back, having learned what you've learned from

617
00:33:11,076 --> 00:33:14,036
Speaker 1: the Cleveland Clinic, when you think about healthcare and AI

618
00:33:14,676 --> 00:33:17,196
Speaker 1: in the next few years, what do you think about I.

619
00:33:17,116 --> 00:33:19,996
Speaker 2: Think that healthcare and AI are going to really beat

620
00:33:20,076 --> 00:33:23,236
Speaker 2: like peanut butter and chocolate. I think it's the best

621
00:33:23,356 --> 00:33:26,396
Speaker 2: place for AI to come into our lives and have

622
00:33:26,636 --> 00:33:29,636
Speaker 2: rational uses that make our health care better and that

623
00:33:29,676 --> 00:33:33,156
Speaker 2: make our costs lower. It is going to take change.

624
00:33:33,196 --> 00:33:35,796
Speaker 2: And so one of the most interesting people I met

625
00:33:35,796 --> 00:33:38,116
Speaker 2: at the clinic is this woman named doctor Rita Pappus,

626
00:33:38,476 --> 00:33:42,916
Speaker 2: who's a hospitalist and hospitalists oversee the administration of the hospital,

627
00:33:42,956 --> 00:33:45,476
Speaker 2: but from a medical point of view. And she started

628
00:33:45,516 --> 00:33:48,196
Speaker 2: as a nurse and in her thirties went back to

629
00:33:48,236 --> 00:33:51,876
Speaker 2: medical school, and so she's seen everything a hospital has

630
00:33:51,916 --> 00:33:54,596
Speaker 2: to offer, including every type of human in a hospital.

631
00:33:55,236 --> 00:33:57,236
Speaker 2: And one of the things she told me is like, look,

632
00:33:58,676 --> 00:34:02,556
Speaker 2: I know that roheat the CTO is finding lots of

633
00:34:02,596 --> 00:34:06,436
Speaker 2: resistance across the hospital, and guess what, I'm not the

634
00:34:06,516 --> 00:34:09,396
Speaker 2: least bit surprised. Doctors are trained a very particular way

635
00:34:10,036 --> 00:34:13,716
Speaker 2: and ultimately, like long term, two decades out, if we

636
00:34:13,756 --> 00:34:16,676
Speaker 2: want AI to take hold in hospitals, we actually going

637
00:34:16,716 --> 00:34:19,716
Speaker 2: to need different kinds of people to go into medicine

638
00:34:20,356 --> 00:34:24,756
Speaker 2: who are less lowercase C conservative, less grounded in their ways,

639
00:34:25,196 --> 00:34:31,316
Speaker 2: more collaborative, and more willing to trust that technology can

640
00:34:31,316 --> 00:34:33,276
Speaker 2: do some of the job. The people that she sees

641
00:34:33,356 --> 00:34:36,476
Speaker 2: going into the profession right now after hundreds of years

642
00:34:36,476 --> 00:34:40,076
Speaker 2: of selection, are very particular type. And so she didn't

643
00:34:40,116 --> 00:34:43,796
Speaker 2: say this in a value judgment. She's very objective about it,

644
00:34:43,836 --> 00:34:45,676
Speaker 2: but she said, look, that's what it's going to take.

645
00:34:45,756 --> 00:34:48,276
Speaker 2: We've already proven that it can work for our bottom line,

646
00:34:48,316 --> 00:34:50,556
Speaker 2: that it can work for patients if we're going to

647
00:34:50,596 --> 00:34:54,156
Speaker 2: scale it, education for doctors is going to have to

648
00:34:54,236 --> 00:34:54,716
Speaker 2: change too.

649
00:34:55,596 --> 00:34:59,796
Speaker 1: So at the end of the book you have this epilogue,

650
00:34:59,836 --> 00:35:02,876
Speaker 1: that's the one part of the book where you're basically like,

651
00:35:04,036 --> 00:35:07,516
Speaker 1: here's what you should do. You being a person who's

652
00:35:07,516 --> 00:35:09,796
Speaker 1: reading the book, who cares about the world it's interested

653
00:35:09,836 --> 00:35:13,876
Speaker 1: in AI, what should I do? What's your advice?

654
00:35:14,676 --> 00:35:18,396
Speaker 2: Well, I think, look, there's two pieces of advice here, right.

655
00:35:18,836 --> 00:35:22,076
Speaker 2: The first is it, if you're the kind of person

656
00:35:22,156 --> 00:35:25,996
Speaker 2: who cares about living in a better world, you're going

657
00:35:26,076 --> 00:35:28,356
Speaker 2: to have to use these tools to understand what they're

658
00:35:28,396 --> 00:35:33,436
Speaker 2: capable of. And that responsibility is really frustrating to people. Right,

659
00:35:33,516 --> 00:35:36,396
Speaker 2: We've now listened to technologists tell us that we have

660
00:35:36,436 --> 00:35:38,676
Speaker 2: to change our lives for whatever their product is for

661
00:35:38,716 --> 00:35:43,076
Speaker 2: the last twenty years, and it's annoying and they get richer,

662
00:35:43,756 --> 00:35:46,676
Speaker 2: and we spend more time adapting, and you haven't seen

663
00:35:46,676 --> 00:35:49,196
Speaker 2: a lot of public good come from it. But I'm

664
00:35:49,236 --> 00:35:52,836
Speaker 2: pretty convinced that in the right situations, these tools can

665
00:35:52,916 --> 00:35:56,636
Speaker 2: help us, can help their republic, can help education, medicine.

666
00:35:58,196 --> 00:36:01,436
Speaker 2: But if we leave that to the technologists, we're going

667
00:36:01,476 --> 00:36:03,556
Speaker 2: to get a bad result. So we need to figure

668
00:36:03,556 --> 00:36:06,036
Speaker 2: out what good they can play in our lives and

669
00:36:06,076 --> 00:36:09,236
Speaker 2: what kind of good we want to advocate for in

670
00:36:09,236 --> 00:36:12,916
Speaker 2: the public space. And if it's not hard, there are

671
00:36:12,956 --> 00:36:15,156
Speaker 2: lots of ways to understand what this stuff does. You

672
00:36:15,156 --> 00:36:16,676
Speaker 2: can do it in a couple minutes, in a day,

673
00:36:17,196 --> 00:36:22,036
Speaker 2: repeat it. It's a really important component to being a

674
00:36:22,036 --> 00:36:24,396
Speaker 2: citizen in the twenty first century. And it's a lot

675
00:36:24,396 --> 00:36:27,156
Speaker 2: of what the Pope wrote in his encyclical, which is

676
00:36:27,876 --> 00:36:30,236
Speaker 2: we cannot wish these away, but we do have to

677
00:36:30,276 --> 00:36:31,916
Speaker 2: make them more human. And the way to make them

678
00:36:31,916 --> 00:36:34,916
Speaker 2: more human is to be insisting that they be more human.

679
00:36:34,956 --> 00:36:35,916
Speaker 2: So you have to use them.

680
00:36:36,716 --> 00:36:39,916
Speaker 1: When you say make them more human, what does that mean.

681
00:36:40,156 --> 00:36:42,276
Speaker 2: Left to their own devices. This is going to be

682
00:36:42,396 --> 00:36:45,996
Speaker 2: enterprise software. It's going to be sold into the Department

683
00:36:46,036 --> 00:36:49,356
Speaker 2: of Defense. It's going to be used to eliminate jobs.

684
00:36:49,796 --> 00:36:52,436
Speaker 2: It's going to be used to make corporations more efficient.

685
00:36:53,236 --> 00:36:55,956
Speaker 2: And I think what we want to advocate for is, no,

686
00:36:56,276 --> 00:37:00,956
Speaker 2: we want these tools to make many, many people's lives better.

687
00:37:01,476 --> 00:37:04,076
Speaker 2: It's not to say that, you know, capitalism can't continue

688
00:37:04,116 --> 00:37:07,156
Speaker 2: churning on, but like you have to advocate for what

689
00:37:07,196 --> 00:37:09,316
Speaker 2: you want, and so in the same way, that mister

690
00:37:09,436 --> 00:37:13,476
Speaker 2: Johnson's use of AI is sitting way downstream of what

691
00:37:13,556 --> 00:37:16,636
Speaker 2: Google and open aiy ultimately want. We want to move

692
00:37:16,676 --> 00:37:19,276
Speaker 2: those things upstream, and so the only way to do

693
00:37:19,316 --> 00:37:21,516
Speaker 2: it is to know what the tools can do and

694
00:37:21,516 --> 00:37:24,756
Speaker 2: then insist that we get that kind of utility because

695
00:37:26,116 --> 00:37:28,996
Speaker 2: it's just not on the radar for these companies. So

696
00:37:29,036 --> 00:37:32,716
Speaker 2: that's one component of it, and the other, frankly, is

697
00:37:32,716 --> 00:37:36,076
Speaker 2: we probably need greater regulation for the negatives to prevent

698
00:37:36,116 --> 00:37:36,836
Speaker 2: them from happening.

699
00:37:37,476 --> 00:37:38,676
Speaker 1: Like what's good regulation?

700
00:37:39,996 --> 00:37:43,476
Speaker 2: The best model I've heard so far is really, you know,

701
00:37:43,596 --> 00:37:49,316
Speaker 2: kind of using nuclear power and nuclear weaponry as a model. Right,

702
00:37:49,636 --> 00:37:52,196
Speaker 2: that was the last thing we created that has this

703
00:37:52,396 --> 00:37:56,796
Speaker 2: kind of scale and power. The International Atomic Energy Commission

704
00:37:56,836 --> 00:37:59,396
Speaker 2: was established to make sure that we knew exactly how

705
00:37:59,516 --> 00:38:01,756
Speaker 2: much nuclear material there was in the world yea, and

706
00:38:01,756 --> 00:38:03,756
Speaker 2: who had control of it and what it was being

707
00:38:03,876 --> 00:38:04,156
Speaker 2: used for.

708
00:38:04,236 --> 00:38:06,996
Speaker 1: I mean, like, how does that map to this? I

709
00:38:07,036 --> 00:38:09,596
Speaker 1: get that it says big deal read flo but what

710
00:38:09,596 --> 00:38:10,636
Speaker 1: does it actually mean.

711
00:38:10,916 --> 00:38:14,036
Speaker 2: It means keeping an inventory of all of these models

712
00:38:14,036 --> 00:38:17,436
Speaker 2: and their capabilities, knowing how much energy they're using, knowing

713
00:38:17,436 --> 00:38:21,316
Speaker 2: how many chips they have, knowing how many floating point operations,

714
00:38:21,556 --> 00:38:24,676
Speaker 2: which is the technical term for what these models are processing.

715
00:38:25,236 --> 00:38:27,756
Speaker 2: Each model is doing. You could set a limit on

716
00:38:27,756 --> 00:38:31,436
Speaker 2: the floating point operations and say beyond this model, you

717
00:38:31,476 --> 00:38:33,676
Speaker 2: can't make it, at least for now.

718
00:38:33,956 --> 00:38:37,916
Speaker 1: So you're talking about like a global speed limit on

719
00:38:38,036 --> 00:38:39,796
Speaker 1: the development of bigger models.

720
00:38:40,076 --> 00:38:42,956
Speaker 2: Say, I think it's probably the simplest way to go

721
00:38:43,036 --> 00:38:47,876
Speaker 2: about regulating how AI works in the world and giving

722
00:38:48,076 --> 00:38:50,956
Speaker 2: governments and societies time to reckon with the consequences of

723
00:38:51,036 --> 00:38:51,916
Speaker 2: each new model.

724
00:38:52,116 --> 00:38:55,596
Speaker 1: And you think it might happen, I think.

725
00:38:55,516 --> 00:38:58,476
Speaker 2: Something akin to it is going to happen because I

726
00:38:58,716 --> 00:39:03,516
Speaker 2: think politically actors are realizing in twenty twenty six, AI

727
00:39:03,636 --> 00:39:05,036
Speaker 2: is likely to be one of the two or three

728
00:39:05,076 --> 00:39:08,076
Speaker 2: biggest issues in the election, and by twenty twenty eight

729
00:39:08,716 --> 00:39:12,156
Speaker 2: it's like to be the biggest. Look, it's America, and

730
00:39:12,156 --> 00:39:16,436
Speaker 2: in the end, America tends to regulate after the crisis.

731
00:39:16,756 --> 00:39:20,636
Speaker 2: That's historically what we do. If I had to beat

732
00:39:20,676 --> 00:39:22,716
Speaker 2: I would say that's probably what will happen this time.

733
00:39:22,916 --> 00:39:25,396
Speaker 2: Is something bad will happen and then we will all

734
00:39:25,436 --> 00:39:27,916
Speaker 2: be looking at the ashes of it and figuring out

735
00:39:27,916 --> 00:39:30,396
Speaker 2: what to do. I hope that's not the case.

736
00:39:35,356 --> 00:39:42,396
Speaker 1: We'll be back in a minute with the lightning round.

737
00:39:46,716 --> 00:39:47,916
Speaker 1: Let's finish with the lightning round.

738
00:39:48,596 --> 00:39:48,796
Speaker 2: Yeah.

739
00:39:49,956 --> 00:39:51,796
Speaker 1: Best Baltimore musician of all time.

740
00:39:53,796 --> 00:39:56,076
Speaker 2: Tupac Shakur did not know.

741
00:39:57,476 --> 00:39:57,916
Speaker 1: Baltimore.

742
00:39:57,956 --> 00:40:00,636
Speaker 2: A student of Baltimore School for the Arts and the

743
00:40:00,676 --> 00:40:04,116
Speaker 2: Baltimore Orioles recently gave out a Tupac Shakur bobblehead and

744
00:40:04,156 --> 00:40:05,516
Speaker 2: the stadium was full.

745
00:40:05,716 --> 00:40:06,316
Speaker 1: Were you there?

746
00:40:06,756 --> 00:40:08,596
Speaker 2: I was not, but I contemplated.

747
00:40:09,116 --> 00:40:11,596
Speaker 1: We were on the ground crew of the Orioles. Yeah.

748
00:40:11,676 --> 00:40:13,236
Speaker 1: In high school, what'd you learn?

749
00:40:15,196 --> 00:40:17,596
Speaker 2: I learned that the most disgusting thing in the world

750
00:40:18,076 --> 00:40:22,516
Speaker 2: is a tobacco ball combined with bazooka, and that most

751
00:40:22,596 --> 00:40:26,036
Speaker 2: relief pitchers would form that in their mouths, get the

752
00:40:26,116 --> 00:40:29,396
Speaker 2: juices out, spit it out into the air, and then

753
00:40:29,516 --> 00:40:30,956
Speaker 2: kick it into the outfield.

754
00:40:31,756 --> 00:40:35,956
Speaker 1: Incredible. What's your walk on song?

755
00:40:38,596 --> 00:40:40,196
Speaker 2: I really love the Rocky theme?

756
00:40:40,556 --> 00:40:42,876
Speaker 1: Amazing a cliche for a reason.

757
00:40:43,196 --> 00:40:46,036
Speaker 2: It absolutely is like, if you are working out and

758
00:40:46,076 --> 00:40:50,156
Speaker 2: you need to go that last quarter mile, just put

759
00:40:50,196 --> 00:40:50,436
Speaker 2: it on.

760
00:40:50,596 --> 00:40:53,196
Speaker 1: I really respect that you went to that instead of

761
00:40:53,236 --> 00:40:59,756
Speaker 1: some obscure hipster pick so right that you oversaw Times

762
00:40:59,796 --> 00:41:02,876
Speaker 1: Person of the Year when you were at time I did.

763
00:41:02,956 --> 00:41:05,196
Speaker 2: Yeah, what's a tip.

764
00:41:05,236 --> 00:41:06,316
Speaker 1: If I want to be Times.

765
00:41:06,076 --> 00:41:09,596
Speaker 2: Person of the Year, you know, it's really year dependent.

766
00:41:09,836 --> 00:41:14,316
Speaker 3: So I, like a fool, thought we would just evaluate

767
00:41:14,356 --> 00:41:17,356
Speaker 3: it based on the most important person of the year,

768
00:41:17,956 --> 00:41:20,916
Speaker 3: and that is in that case, it turns out it's like, oh,

769
00:41:20,956 --> 00:41:23,236
Speaker 3: but last year we did a bad person, so this.

770
00:41:23,276 --> 00:41:25,276
Speaker 2: Year we need to do a good person.

771
00:41:25,476 --> 00:41:28,556
Speaker 1: Well that makes some sense. There's a mix, it does.

772
00:41:28,436 --> 00:41:31,596
Speaker 2: You want to mix, but it is not the pure

773
00:41:31,716 --> 00:41:32,996
Speaker 2: thing that I was raised to believe.

774
00:41:32,996 --> 00:41:36,716
Speaker 1: It is. Were you there? There was a year I

775
00:41:36,836 --> 00:41:39,036
Speaker 1: remember exactly what it was. I don't know exactly when

776
00:41:39,036 --> 00:41:41,076
Speaker 1: you were there when they put like something reflective on

777
00:41:41,116 --> 00:41:42,676
Speaker 1: the cover in the Person of the Are was you?

778
00:41:43,076 --> 00:41:43,516
Speaker 2: Was that you?

779
00:41:44,116 --> 00:41:44,276
Speaker 1: Oh?

780
00:41:44,356 --> 00:41:49,396
Speaker 2: I was there? It was not me. I thought this

781
00:41:49,596 --> 00:41:52,796
Speaker 2: is pretty silly. You know, we could just make the

782
00:41:52,876 --> 00:41:54,756
Speaker 2: YouTube guys, the people of the year.

783
00:41:54,716 --> 00:41:58,036
Speaker 1: Failure is an orphan success as a thousand fathers failures. Yeah,

784
00:41:58,276 --> 00:41:59,276
Speaker 1: that was happening right now.

785
00:41:59,476 --> 00:42:01,596
Speaker 2: So I was no, no, no, I genuinely did not

786
00:42:01,676 --> 00:42:02,316
Speaker 2: oversee that one.

787
00:42:02,436 --> 00:42:04,676
Speaker 1: Is there nobody who would say I thought that was

788
00:42:04,676 --> 00:42:05,316
Speaker 1: a good idea.

789
00:42:05,476 --> 00:42:08,636
Speaker 2: I was asked forty eight hours before it went to

790
00:42:08,636 --> 00:42:11,476
Speaker 2: print to write the back page, and so I wrote

791
00:42:11,476 --> 00:42:15,196
Speaker 2: a backpage that, as I recall, was largely about Andy

792
00:42:15,196 --> 00:42:18,036
Speaker 2: Warhol and the notion of you know, fifteen minutes and

793
00:42:18,076 --> 00:42:21,836
Speaker 2: how that works in a distributed sort of technological I

794
00:42:21,836 --> 00:42:23,236
Speaker 2: don't want to go back and look at it or

795
00:42:23,276 --> 00:42:25,596
Speaker 2: defend it, but I will tell you it was not

796
00:42:25,756 --> 00:42:28,556
Speaker 2: my call to put the reflective view on the cover.

797
00:42:29,196 --> 00:42:31,796
Speaker 1: You know, everybody hated it at the time, and I

798
00:42:31,836 --> 00:42:34,756
Speaker 1: get that. I mean, for anybody to care at all

799
00:42:34,836 --> 00:42:37,556
Speaker 1: now it would be amazing. But it also, like you

800
00:42:37,596 --> 00:42:39,836
Speaker 1: could argue it's sort of the fifteen minutes of famous,

801
00:42:39,876 --> 00:42:42,316
Speaker 1: like the foreshadowing the creator economy or something, if you

802
00:42:42,356 --> 00:42:42,756
Speaker 1: wanted that.

803
00:42:43,636 --> 00:42:46,116
Speaker 2: I think others there have been others that have held

804
00:42:46,196 --> 00:42:46,636
Speaker 2: up worse.

805
00:42:47,356 --> 00:42:51,516
Speaker 1: If you were running a media company right now, what

806
00:42:51,556 --> 00:42:52,236
Speaker 1: would your play be.

807
00:42:54,036 --> 00:42:59,396
Speaker 2: I would really retreat to original reporting and high quality presentation,

808
00:43:00,116 --> 00:43:02,876
Speaker 2: because I think what you're really competing is right now

809
00:43:02,996 --> 00:43:05,196
Speaker 2: is a lot of noise and a lot of slop

810
00:43:05,356 --> 00:43:07,756
Speaker 2: and a lot of volume. And when you see that

811
00:43:07,796 --> 00:43:12,236
Speaker 2: going on, treat to quality and trust your taste because

812
00:43:12,316 --> 00:43:15,236
Speaker 2: the individual the taste of individuals is still pretty interesting,

813
00:43:16,036 --> 00:43:19,156
Speaker 2: I think. But I wouldn't try to compete everywhere. I

814
00:43:19,196 --> 00:43:21,956
Speaker 2: would try to compete on the most interesting stories and

815
00:43:22,116 --> 00:43:24,876
Speaker 2: make them, you know, keep it simple, make them as

816
00:43:24,876 --> 00:43:26,636
Speaker 2: good as you can possibly make them.

817
00:43:27,076 --> 00:43:28,956
Speaker 1: Thanks Mian, congratulations on the book.

818
00:43:29,036 --> 00:43:29,756
Speaker 2: Thank you so much.

819
00:43:38,236 --> 00:43:41,076
Speaker 1: Josh Tarrengel is the author of the book AI for Good,

820
00:43:41,596 --> 00:43:45,236
Speaker 1: How real people are using artificial intelligence to fix things

821
00:43:45,276 --> 00:43:49,196
Speaker 1: that matter. Please let us know what you think of

822
00:43:49,236 --> 00:43:51,156
Speaker 1: the show, which you want to hear more of, which

823
00:43:51,196 --> 00:43:55,116
Speaker 1: you want to hear less of particular guest ideas. You

824
00:43:55,156 --> 00:43:58,436
Speaker 1: can email us at problem at pushkin dot fm. I

825
00:43:58,476 --> 00:44:00,996
Speaker 1: read all the emails. You can also find me on

826
00:44:01,716 --> 00:44:05,996
Speaker 1: x run LinkedIn. Really do appreciate all the messages that

827
00:44:06,036 --> 00:44:08,716
Speaker 1: we get. Today's show was produced by Gabriel Hunter Chang

828
00:44:08,716 --> 00:44:11,756
Speaker 1: and The Menino. It was engineered by Hansdale She and

829
00:44:11,996 --> 00:44:15,316
Speaker 1: edited by Lydia Jane Kott. I'm Jacob Goldstein and we'll

830
00:44:15,316 --> 00:44:24,516
Speaker 1: be back next week with another episode of What's Your Problem.