1
00:00:09,960 --> 00:00:12,640
Speaker 1: I've lived in LA for a decade, and this whole time,

2
00:00:12,760 --> 00:00:15,680
Speaker 1: I haven't owned a car. When I tell people that,

3
00:00:15,840 --> 00:00:19,279
Speaker 1: they usually look at me weird. And yes, riding the

4
00:00:19,320 --> 00:00:22,360
Speaker 1: bus and walking and using a bike is less convenient,

5
00:00:22,800 --> 00:00:25,800
Speaker 1: but at this point I'm used to it. But sometimes

6
00:00:26,000 --> 00:00:28,000
Speaker 1: I do wonder if I should just give in and

7
00:00:28,120 --> 00:00:31,720
Speaker 1: buy a car like everyone else. So to help me decide,

8
00:00:31,840 --> 00:00:33,879
Speaker 1: I did what a lot of people do recently when

9
00:00:33,920 --> 00:00:38,680
Speaker 1: they're weighing options. I asked AI. I opened up Claude

10
00:00:38,680 --> 00:00:42,240
Speaker 1: dot AI and I input all my current transportation costs.

11
00:00:42,280 --> 00:00:45,160
Speaker 1: I put in my bus fares, my bike, the cost

12
00:00:45,240 --> 00:00:48,159
Speaker 1: of my occasional uber and I asked it to compare

13
00:00:48,200 --> 00:00:51,280
Speaker 1: that to the average costs of car ownership in Los Angeles,

14
00:00:51,400 --> 00:00:56,240
Speaker 1: so parking, gas, insurance, repairs, all that sort of thing.

15
00:00:56,360 --> 00:00:58,279
Speaker 1: And I asked it to give me the pros and

16
00:00:58,360 --> 00:01:02,320
Speaker 1: cons on either side, and it did. The big con

17
00:01:02,600 --> 00:01:06,119
Speaker 1: is convenience, which I already knew. And on the pro side,

18
00:01:06,200 --> 00:01:08,759
Speaker 1: it said that I was saving thousands of dollars per year,

19
00:01:09,440 --> 00:01:12,120
Speaker 1: but it added one extra thing. It said that I

20
00:01:12,120 --> 00:01:14,680
Speaker 1: could have the nice feeling of knowing that I was

21
00:01:14,720 --> 00:01:18,360
Speaker 1: also being eco friendly. And I thought, hold on, wait

22
00:01:18,400 --> 00:01:22,840
Speaker 1: a second, eco friendly, I just spent half an hour

23
00:01:22,920 --> 00:01:26,240
Speaker 1: running scenarios through an LM, which I know is built

24
00:01:26,280 --> 00:01:29,160
Speaker 1: off the back of a massive amount of computing, which

25
00:01:29,280 --> 00:01:32,600
Speaker 1: in turn means a massive amount of energy. So am

26
00:01:32,600 --> 00:01:35,880
Speaker 1: I actually helping the environment here? Or am I hurting it?

27
00:01:37,360 --> 00:01:40,120
Speaker 1: So this week I set out to answer what seems

28
00:01:40,160 --> 00:01:45,560
Speaker 1: like a pretty simple question, how bad is AI's environmental impact? Really?

29
00:01:46,160 --> 00:01:49,840
Speaker 1: And yes, before you ask, I did consider asking Claude

30
00:01:49,840 --> 00:01:53,720
Speaker 1: and maybe chat GBT about AI's own impact on the environment.

31
00:01:53,960 --> 00:01:56,480
Speaker 1: But then I figured, you know what, maybe this is

32
00:01:56,480 --> 00:02:00,160
Speaker 1: a question I should ask actual human beings. And I

33
00:02:00,160 --> 00:02:02,240
Speaker 1: found a couple of people who've been studying this stuff

34
00:02:02,280 --> 00:02:04,160
Speaker 1: for a while to help me parse all of this.

35
00:02:05,400 --> 00:02:08,120
Speaker 1: Is there a way that I can compare? Is my

36
00:02:08,240 --> 00:02:13,120
Speaker 1: AI usage worse or better than car usage, or worse

37
00:02:13,200 --> 00:02:16,280
Speaker 1: or better than my impact on the environment from eating

38
00:02:16,320 --> 00:02:18,799
Speaker 1: meat or something like that? Are we able to make

39
00:02:18,840 --> 00:02:20,000
Speaker 1: those kinds of comparisons.

40
00:02:20,200 --> 00:02:22,880
Speaker 2: That's what carbon footprints were kind of invented for, so

41
00:02:22,960 --> 00:02:25,720
Speaker 2: you can make this type of comparison. If you're driving

42
00:02:25,720 --> 00:02:27,799
Speaker 2: in fossil fuel based car, you know exactly how much

43
00:02:27,840 --> 00:02:30,320
Speaker 2: gas you're using and what that might mean in terms

44
00:02:30,320 --> 00:02:34,640
Speaker 2: of carbon emissions. That's pretty straightforward. It's much harder to

45
00:02:34,720 --> 00:02:35,640
Speaker 2: do this for AI.

46
00:02:39,400 --> 00:02:48,440
Speaker 1: I'm afraid from Kaleidoscope and iHeart podcasts, this is kill switch.

47
00:02:49,800 --> 00:02:51,360
Speaker 1: I'm Jackster Thomas.

48
00:02:52,040 --> 00:02:52,640
Speaker 2: I'm sorry.

49
00:03:21,400 --> 00:03:24,400
Speaker 1: If you use social media, you've probably seen people being

50
00:03:24,400 --> 00:03:27,760
Speaker 1: criticized for using AI, and depending on who you hang

51
00:03:27,800 --> 00:03:30,720
Speaker 1: out with, that criticism can be kind of different in

52
00:03:30,800 --> 00:03:33,680
Speaker 1: how it shows up. When I see someone post an

53
00:03:33,680 --> 00:03:37,280
Speaker 1: AI generated image or some AI generated text and there's

54
00:03:37,360 --> 00:03:40,480
Speaker 1: angry comments in the comment section, it's usually one of

55
00:03:40,560 --> 00:03:44,720
Speaker 1: two types. The first is people saying that it's disrespectful

56
00:03:45,200 --> 00:03:49,760
Speaker 1: that by posting AI generated poetry or drawings you're devaluing

57
00:03:49,800 --> 00:03:52,680
Speaker 1: the original artists who didn't consent to having their work

58
00:03:52,720 --> 00:03:55,280
Speaker 1: fed into an l l M. That one's pretty easy

59
00:03:55,360 --> 00:03:57,800
Speaker 1: to understand, even if you don't agree with it or

60
00:03:57,880 --> 00:04:00,320
Speaker 1: think it's a big deal. The other comment I see

61
00:04:00,320 --> 00:04:03,280
Speaker 1: a lot of is people saying that using AI is

62
00:04:03,320 --> 00:04:07,160
Speaker 1: destroying the environment. Figuring out whether that's a big deal

63
00:04:07,280 --> 00:04:09,640
Speaker 1: or not is a little bit less straightforward.

64
00:04:09,960 --> 00:04:12,000
Speaker 2: What I tried to do in my research, I tried

65
00:04:12,000 --> 00:04:16,200
Speaker 2: to keep track of how the global electricity consumption of

66
00:04:16,320 --> 00:04:18,320
Speaker 2: AI is developing.

67
00:04:18,080 --> 00:04:20,919
Speaker 1: Alex Deviries is the founder of digit Economists and a

68
00:04:20,960 --> 00:04:24,560
Speaker 1: PhD candidate at the VU Amsterdam. He's been researching the

69
00:04:24,600 --> 00:04:27,359
Speaker 1: sustainability of new technologies for about a decade.

70
00:04:27,640 --> 00:04:29,760
Speaker 2: The way I do that is by looking at how

71
00:04:29,800 --> 00:04:33,640
Speaker 2: many machines of specialized AI's devices are being produced by

72
00:04:33,680 --> 00:04:37,400
Speaker 2: the AI hardware supply chain, and then considering that power

73
00:04:37,440 --> 00:04:41,040
Speaker 2: consumption profile how much power is now being consumed by

74
00:04:41,080 --> 00:04:44,479
Speaker 2: all of these devices. Which is a very imperfect way

75
00:04:44,480 --> 00:04:46,200
Speaker 2: of keeping track of this, but it's kind of like

76
00:04:46,240 --> 00:04:48,680
Speaker 2: the only rule you have available at the moment.

77
00:04:48,520 --> 00:04:50,800
Speaker 1: And even Alex is having a hard time keeping up

78
00:04:50,839 --> 00:04:53,240
Speaker 1: with this. What I called him. He was in the

79
00:04:53,240 --> 00:04:56,400
Speaker 1: middle of putting together new research. Back in twenty twenty three.

80
00:04:56,760 --> 00:04:59,679
Speaker 1: His data showed that by twenty twenty seven, new AI

81
00:04:59,680 --> 00:05:02,360
Speaker 1: service sold could use the same amount of energy annually

82
00:05:02,560 --> 00:05:05,720
Speaker 1: as the yearly energy consumption of a country like Argentina

83
00:05:06,040 --> 00:05:10,799
Speaker 1: or the Netherlands. But things have accelerated. His current research

84
00:05:10,839 --> 00:05:13,479
Speaker 1: shows that it won't take until twenty twenty seven for

85
00:05:13,520 --> 00:05:16,360
Speaker 1: that to happen. At this rate, we're going to hit

86
00:05:16,360 --> 00:05:17,840
Speaker 1: that mark sometime this year.

87
00:05:18,120 --> 00:05:21,719
Speaker 2: Simply because now the amount of devices that's being produced

88
00:05:21,760 --> 00:05:24,600
Speaker 2: by the AI hardware supply chain is way higher than

89
00:05:24,640 --> 00:05:25,960
Speaker 2: it was two years ago.

90
00:05:26,200 --> 00:05:31,760
Speaker 1: So it's even exceeding your pretty bleak estimations that you

91
00:05:31,800 --> 00:05:32,480
Speaker 1: made a while ago.

92
00:05:32,839 --> 00:05:35,320
Speaker 2: Oh yeah, it's just that the hype is so big,

93
00:05:35,320 --> 00:05:37,080
Speaker 2: and then the mod for this type of hardware is

94
00:05:37,120 --> 00:05:40,599
Speaker 2: so big that the numbers are going up much faster

95
00:05:40,680 --> 00:05:42,960
Speaker 2: than could be anticipated just two years ago.

96
00:05:43,080 --> 00:05:45,720
Speaker 1: But hold on, before we get too much further, let's

97
00:05:45,800 --> 00:05:49,200
Speaker 1: just clarify what we're even talking about when we say AI.

98
00:05:50,160 --> 00:05:52,839
Speaker 1: If you could break it down for me, how does

99
00:05:53,880 --> 00:05:57,080
Speaker 1: artificial intelligence use natural resources?

100
00:05:57,400 --> 00:06:01,480
Speaker 3: Heah, it's a general umbrella term that includes many different things.

101
00:06:01,839 --> 00:06:04,680
Speaker 3: But right now, if you're talking to a person on

102
00:06:04,720 --> 00:06:08,159
Speaker 3: the streets random like when they say AI, they're referring

103
00:06:08,200 --> 00:06:11,600
Speaker 3: to large languine models, or maybe you meage generation models.

104
00:06:11,680 --> 00:06:14,320
Speaker 3: So these are the generaryty AI models.

105
00:06:14,400 --> 00:06:17,840
Speaker 1: Chale Wren is an associate professor of electrical and computer

106
00:06:17,960 --> 00:06:21,640
Speaker 1: engineering at the University of California, Riverside, and he's kind

107
00:06:21,640 --> 00:06:25,359
Speaker 1: of a colleague of mine. Our fields are completely different.

108
00:06:25,720 --> 00:06:27,719
Speaker 1: But a couple of years ago I taught a class

109
00:06:27,720 --> 00:06:30,360
Speaker 1: in the building right next to his. I'd had no

110
00:06:30,520 --> 00:06:33,560
Speaker 1: idea that on the same campus there was an expert

111
00:06:33,560 --> 00:06:36,840
Speaker 1: who'd been researching the environmental impact of generative AI the

112
00:06:36,880 --> 00:06:40,640
Speaker 1: whole time, and I thought, perfect, this guy's kind of

113
00:06:40,640 --> 00:06:43,240
Speaker 1: a colleague, so I can stop doing all this research

114
00:06:43,279 --> 00:06:46,640
Speaker 1: on my own and just go ask him. Can you

115
00:06:46,680 --> 00:06:51,640
Speaker 1: give me an idea of how, say, car usage compares

116
00:06:51,760 --> 00:06:54,480
Speaker 1: to usage of an AI model.

117
00:06:54,920 --> 00:06:58,640
Speaker 3: I would say having a large language or medium sized

118
00:06:58,760 --> 00:07:05,000
Speaker 3: language model. Right, roughly ten short emails could be consuming

119
00:07:05,440 --> 00:07:08,640
Speaker 3: a quarter of the electric kill what hour energy, So

120
00:07:08,720 --> 00:07:12,000
Speaker 3: that's roughly enough to drive a test in a Model three.

121
00:07:11,880 --> 00:07:14,960
Speaker 2: For one mile, or as Alex puts it, Chat GPT

122
00:07:15,120 --> 00:07:17,480
Speaker 2: must be running on like something like five hundred mega

123
00:07:17,520 --> 00:07:20,320
Speaker 2: what hours a day, which is enough to power a

124
00:07:20,360 --> 00:07:21,400
Speaker 2: small city.

125
00:07:21,960 --> 00:07:26,360
Speaker 1: Basically, chat GPT's overall daily energy use, it's about the

126
00:07:26,400 --> 00:07:30,000
Speaker 1: same as powering every home, every grocery store, every street

127
00:07:30,040 --> 00:07:33,160
Speaker 1: light in a small city like San Luis, Opistpo in

128
00:07:33,200 --> 00:07:37,160
Speaker 1: California or Ithaca in upstate New York. But what does

129
00:07:37,160 --> 00:07:40,320
Speaker 1: that actually mean for me and you? How much energy

130
00:07:40,360 --> 00:07:43,480
Speaker 1: does it take to just ask chat GBT one question

131
00:07:44,280 --> 00:07:44,600
Speaker 1: on a.

132
00:07:44,560 --> 00:07:48,440
Speaker 2: Per interaction basis? It's actually not that much. You're talking

133
00:07:48,480 --> 00:07:51,800
Speaker 2: about something like three one hours maybe per interaction that's

134
00:07:51,840 --> 00:07:54,080
Speaker 2: something like a low lumin led build that you have

135
00:07:54,160 --> 00:07:56,600
Speaker 2: running for one hour. It's not a lot of power,

136
00:07:56,760 --> 00:08:00,160
Speaker 2: but it's it's nevertheless significantly more than a standard Google.

137
00:08:00,240 --> 00:08:03,160
Speaker 1: Search in one second. Just as an aside here, we

138
00:08:03,280 --> 00:08:05,720
Speaker 1: usually don't think of something like a Google Search as

139
00:08:05,800 --> 00:08:08,800
Speaker 1: using electricity. I mean, your phone or your computer is

140
00:08:08,840 --> 00:08:11,320
Speaker 1: already on, so what does it matter if you're typing

141
00:08:11,320 --> 00:08:13,960
Speaker 1: stuff into it or not. But on the other end

142
00:08:14,000 --> 00:08:17,160
Speaker 1: of that Google search you typed in, they're servers and

143
00:08:17,280 --> 00:08:20,280
Speaker 1: those are using energy. So as we keep going in

144
00:08:20,360 --> 00:08:23,640
Speaker 1: this episode, maybe think about that that on your end

145
00:08:23,920 --> 00:08:27,480
Speaker 1: you're not seeing any energy used or environmental effects, but

146
00:08:27,760 --> 00:08:32,040
Speaker 1: doing a Google search, watching a video, or even downloading

147
00:08:32,080 --> 00:08:35,120
Speaker 1: this podcast that does use some amount of energy.

148
00:08:35,559 --> 00:08:39,479
Speaker 2: Even Google's CEO at some point commented, like, hey, interacting

149
00:08:39,600 --> 00:08:42,960
Speaker 2: with these large language models, it takes ten times more

150
00:08:43,040 --> 00:08:45,640
Speaker 2: power than the standard Google search. So and that would

151
00:08:45,720 --> 00:08:47,880
Speaker 2: mean that if you're talking about three one hours for

152
00:08:48,000 --> 00:08:51,000
Speaker 2: interaction in a large language model, for a standard Google search,

153
00:08:51,040 --> 00:08:53,120
Speaker 2: it would be like zero point three one hours, which

154
00:08:53,160 --> 00:08:55,120
Speaker 2: is a very very tiny amount.

155
00:08:56,360 --> 00:08:58,880
Speaker 1: Just to explain here, a what hour is a unit

156
00:08:58,920 --> 00:09:01,559
Speaker 1: that tells you how much energy device uses over time.

157
00:09:02,120 --> 00:09:04,720
Speaker 1: For example, a sixty watt light bulb running for an

158
00:09:04,760 --> 00:09:09,079
Speaker 1: hour uses sixty watt hours. A single Google search uses

159
00:09:09,120 --> 00:09:12,240
Speaker 1: about zero point three watt hours. That's enough to power

160
00:09:12,280 --> 00:09:16,360
Speaker 1: that same light bulb for around eighteen seconds. But now

161
00:09:16,400 --> 00:09:19,280
Speaker 1: there's that AI add on that comes stacked by default

162
00:09:19,360 --> 00:09:22,200
Speaker 1: on top of every Google search, which takes that number

163
00:09:22,280 --> 00:09:25,600
Speaker 1: up ten x, up to three full watt hours per search.

164
00:09:26,320 --> 00:09:28,960
Speaker 1: That's a little different now you're running that same light

165
00:09:29,000 --> 00:09:32,559
Speaker 1: bulb for three full minutes and then but.

166
00:09:32,640 --> 00:09:35,800
Speaker 2: It's of course in the number of interactions where these

167
00:09:36,120 --> 00:09:38,520
Speaker 2: numbers start to stack up quickly, Because if you're talking

168
00:09:38,559 --> 00:09:42,400
Speaker 2: about Google Skill, you're talking about nine billion interactions a day,

169
00:09:42,679 --> 00:09:46,440
Speaker 2: going as three one hours per interaction. Then, interestingly, the

170
00:09:46,480 --> 00:09:50,040
Speaker 2: whole company Google would require as much power as Ireland

171
00:09:50,120 --> 00:09:51,840
Speaker 2: just to serve as a search engine.

172
00:09:51,880 --> 00:09:55,480
Speaker 1: If that was the case, Wow, using as much power

173
00:09:55,520 --> 00:09:58,959
Speaker 1: as a small country sounds wild, but if we think

174
00:09:59,000 --> 00:10:02,760
Speaker 1: about it, it kind of makes sense. We've default ten

175
00:10:02,880 --> 00:10:07,560
Speaker 1: xed our energy use overnight across nine billion searches a day.

176
00:10:08,080 --> 00:10:10,600
Speaker 1: That energy use is going to add up pretty fast.

177
00:10:11,120 --> 00:10:13,880
Speaker 1: But there's another thing to consider when we talk about

178
00:10:13,920 --> 00:10:17,840
Speaker 1: AI's energy use. The difference between training the model or

179
00:10:17,920 --> 00:10:19,600
Speaker 1: giving it a bunch of data to teach it how

180
00:10:19,600 --> 00:10:22,840
Speaker 1: to work, and using it like when you ask it

181
00:10:22,920 --> 00:10:25,640
Speaker 1: to write a cover letter or I ask it if

182
00:10:25,720 --> 00:10:28,240
Speaker 1: I should buy a car. When we talk about AI

183
00:10:28,440 --> 00:10:30,840
Speaker 1: and the energy consumption that can go into AI, there's

184
00:10:30,840 --> 00:10:34,640
Speaker 1: different phases, right. There's the training phase. There's me actually

185
00:10:34,640 --> 00:10:38,280
Speaker 1: sitting down and asking you an agent, a question. Can

186
00:10:38,320 --> 00:10:39,320
Speaker 1: you break that down for me?

187
00:10:39,600 --> 00:10:43,080
Speaker 3: The training part we call it learning, so based on

188
00:10:43,120 --> 00:10:45,920
Speaker 3: the data we try to optimize the parameters so that

189
00:10:46,200 --> 00:10:49,120
Speaker 3: when we see some new queries from the users, we

190
00:10:49,120 --> 00:10:51,679
Speaker 3: can give you as equate an answer as possible. And

191
00:10:51,920 --> 00:10:54,960
Speaker 3: training is really one time. Of course, later we're going

192
00:10:55,000 --> 00:10:58,520
Speaker 3: to do some update fine tuning. Inference is when the

193
00:10:58,600 --> 00:11:01,880
Speaker 3: users actually interact with the model, and depending on the

194
00:11:01,880 --> 00:11:04,480
Speaker 3: popularity of the model, but once it gets trained, it

195
00:11:04,559 --> 00:11:07,480
Speaker 3: could be used by many hundreds of millions of times

196
00:11:07,880 --> 00:11:10,120
Speaker 3: or even billions of times if you train the large

197
00:11:10,200 --> 00:11:13,440
Speaker 3: lunguine model like LAMA three point one. According to the

198
00:11:13,520 --> 00:11:16,600
Speaker 3: data released by Meta Training, a large lan grain model

199
00:11:16,760 --> 00:11:20,200
Speaker 3: like that air pllutant we gamerated through the training will

200
00:11:20,240 --> 00:11:24,200
Speaker 3: be roughly equivalent to more than ten thousand round trips

201
00:11:24,200 --> 00:11:25,720
Speaker 3: by cart LA and New.

202
00:11:25,679 --> 00:11:29,000
Speaker 1: York City ten thousand round trips by car. Yeah, so

203
00:11:29,040 --> 00:11:34,040
Speaker 1: that sounds bad, that sounds like a lot. But is

204
00:11:34,600 --> 00:11:36,920
Speaker 1: that a one time and it's just the one time.

205
00:11:36,960 --> 00:11:37,640
Speaker 3: It's a one time.

206
00:11:39,000 --> 00:11:42,400
Speaker 1: Let's clear something up here. That number ten thousand round

207
00:11:42,480 --> 00:11:45,200
Speaker 1: trips from LA to New York by car. It's not

208
00:11:45,240 --> 00:11:49,360
Speaker 1: just about carbon it's about air pollution, specifically things like

209
00:11:49,440 --> 00:11:52,840
Speaker 1: nitrogen oxides and fine particles that come from power plants

210
00:11:52,880 --> 00:11:56,160
Speaker 1: and can get deep into your lungs. This isn't theoretical.

211
00:11:56,360 --> 00:12:00,240
Speaker 1: This is stuff that raises risks of diseases like cancer,

212
00:12:00,280 --> 00:12:02,439
Speaker 1: and it doesn't just affect people next to the place

213
00:12:02,440 --> 00:12:07,080
Speaker 1: where all those computers are. Pollution travels and it lingers.

214
00:12:07,720 --> 00:12:11,160
Speaker 1: So what Challet's talking about here isn't just numbers. His

215
00:12:11,240 --> 00:12:14,160
Speaker 1: calculations are showing that training a single model the size

216
00:12:14,160 --> 00:12:17,480
Speaker 1: of Metaslama three point one can produce that level of

217
00:12:17,480 --> 00:12:21,839
Speaker 1: pollution on its own. So yes, training these models is

218
00:12:21,880 --> 00:12:25,000
Speaker 1: a one time hit, but it's a big one. If

219
00:12:25,040 --> 00:12:29,120
Speaker 1: we're talking just about energy usage, using an LM to say,

220
00:12:29,240 --> 00:12:32,480
Speaker 1: write ten emails might be like driving an electric vehicle

221
00:12:32,559 --> 00:12:35,240
Speaker 1: for a mile. And since an electric vehicle was maybe

222
00:12:35,280 --> 00:12:38,120
Speaker 1: three times more efficient than a gas vehicle. Figure, those

223
00:12:38,120 --> 00:12:40,440
Speaker 1: ten emails might get you a quarter to a third

224
00:12:40,440 --> 00:12:43,000
Speaker 1: of a mile in a regular car. And yeah, maybe

225
00:12:43,040 --> 00:12:45,720
Speaker 1: these are relevant numbers for me and my decision about

226
00:12:45,720 --> 00:12:48,960
Speaker 1: whether or not my AI usage is counterbalanced by me

227
00:12:49,120 --> 00:12:52,400
Speaker 1: not having a car. But these numbers are just estimates,

228
00:12:52,480 --> 00:12:54,960
Speaker 1: and we are going to get to that. But the

229
00:12:55,000 --> 00:12:58,280
Speaker 1: bigger issue here is that running those data centers doesn't

230
00:12:58,320 --> 00:13:02,160
Speaker 1: just use electricity. And this is where Chalet's research comes

231
00:13:02,160 --> 00:13:06,040
Speaker 1: in because we've heard about AI's carbon footprint, but what

232
00:13:06,160 --> 00:13:09,720
Speaker 1: about its water footprint, which could be a much bigger

233
00:13:09,760 --> 00:13:13,520
Speaker 1: concern for us living here on earth. That's after the break.

234
00:13:26,480 --> 00:13:29,760
Speaker 1: So you had a study come out last year called

235
00:13:30,000 --> 00:13:33,760
Speaker 1: making AI Less Thirsty, uncovering and addressing the secret water

236
00:13:33,800 --> 00:13:36,520
Speaker 1: footprint of AI models. What made you want to look

237
00:13:36,559 --> 00:13:36,800
Speaker 1: at this?

238
00:13:37,400 --> 00:13:39,760
Speaker 3: Maybe that was due to my childhood experience. I spent

239
00:13:39,840 --> 00:13:42,080
Speaker 3: a couple of years in a small town in China

240
00:13:42,640 --> 00:13:45,520
Speaker 3: where we only had water access for half an hour

241
00:13:45,600 --> 00:13:48,120
Speaker 3: each day, so we just had to think about how

242
00:13:48,160 --> 00:13:51,560
Speaker 3: to use water wisely and to every possible means to

243
00:13:51,600 --> 00:13:55,280
Speaker 3: save water. Then in twenty thirteen, oh I saw this issue.

244
00:13:55,559 --> 00:13:57,640
Speaker 3: I want you to find out more about it. What

245
00:13:57,720 --> 00:14:01,000
Speaker 3: about the water consumption and nobody new at that time.

246
00:14:01,320 --> 00:14:04,160
Speaker 1: A big environmental impact we don't talk about as often

247
00:14:04,200 --> 00:14:08,000
Speaker 1: its carbon emissions is water usage. And the impact that

248
00:14:08,040 --> 00:14:11,080
Speaker 1: water usage has on all of us depends on where

249
00:14:11,120 --> 00:14:14,600
Speaker 1: that water comes from and where it goes when it

250
00:14:14,600 --> 00:14:17,240
Speaker 1: comes to AI. A main use of water is to

251
00:14:17,240 --> 00:14:19,920
Speaker 1: cool down the data centers, which, as we know, use

252
00:14:20,040 --> 00:14:22,200
Speaker 1: a lot of energy. This is how they make sure

253
00:14:22,240 --> 00:14:23,440
Speaker 1: that they don't overheat.

254
00:14:23,720 --> 00:14:27,720
Speaker 3: To prevent syrup from overheating, usually we use water evaporation,

255
00:14:28,040 --> 00:14:30,760
Speaker 3: and that's a very efficient way to move the heat,

256
00:14:30,880 --> 00:14:33,360
Speaker 3: to dissipate the heat to the environment, and this water

257
00:14:33,400 --> 00:14:36,840
Speaker 3: evaporation could be in the cooling towers. That is essentially

258
00:14:37,000 --> 00:14:38,600
Speaker 3: evaporating water twenty four to seven.

259
00:14:38,960 --> 00:14:41,920
Speaker 1: When water evaporates from a data center's cooling system, it

260
00:14:42,040 --> 00:14:45,120
Speaker 1: goes out into the air and is basically considered gone,

261
00:14:45,360 --> 00:14:47,920
Speaker 1: at least from the local supply. You might be thinking

262
00:14:47,920 --> 00:14:49,920
Speaker 1: of the water that you use when you take a shower.

263
00:14:50,120 --> 00:14:53,160
Speaker 1: How that water goes down the drain, It gets treated

264
00:14:53,520 --> 00:14:56,680
Speaker 1: and it can be reused. But evaporated water rises up

265
00:14:56,680 --> 00:14:59,920
Speaker 1: into the atmosphere and you can't reuse it. It can

266
00:15:00,160 --> 00:15:03,680
Speaker 1: eventually come back down as rain, but that takes a while.

267
00:15:04,320 --> 00:15:07,600
Speaker 3: Some tech companies they can use over twenty billion liters

268
00:15:07,600 --> 00:15:08,480
Speaker 3: of water each year.

269
00:15:09,400 --> 00:15:11,720
Speaker 1: Twenty billion, Yeah, that.

270
00:15:11,720 --> 00:15:15,040
Speaker 3: Number basically is the same as some major beverage companies

271
00:15:15,120 --> 00:15:18,720
Speaker 3: annual water consumption, the water they put into their product,

272
00:15:18,960 --> 00:15:22,320
Speaker 3: basically the water we drink from a bottled water. Those

273
00:15:22,360 --> 00:15:25,080
Speaker 3: are the water consumption for the beverage industry. So in

274
00:15:25,120 --> 00:15:29,520
Speaker 3: some sense, this AI is turning these tech companies into

275
00:15:29,960 --> 00:15:32,320
Speaker 3: a beverage company in term of water consumption.

276
00:15:32,880 --> 00:15:37,840
Speaker 1: Nobody's drinking that bottled water or those sodas is just evaporating.

277
00:15:38,520 --> 00:15:39,840
Speaker 3: Yes, yes, yes.

278
00:15:40,080 --> 00:15:43,080
Speaker 1: One important thing here is that when Challte's talking about water,

279
00:15:43,360 --> 00:15:46,960
Speaker 1: he's talking about a specific kind of water. For example,

280
00:15:47,080 --> 00:15:49,600
Speaker 1: you might have heard that for every kilogram of beef,

281
00:15:50,040 --> 00:15:54,240
Speaker 1: it needs fifteen thousand liters of water. But ninety percent

282
00:15:54,280 --> 00:15:57,560
Speaker 1: of that water is what's called green water. That's water

283
00:15:57,600 --> 00:16:00,760
Speaker 1: that's naturally stored in soil and used by plants like

284
00:16:00,960 --> 00:16:03,880
Speaker 1: rain water. It doesn't have to be clean enough for

285
00:16:03,920 --> 00:16:06,520
Speaker 1: people to drink it. It would be nice if data

286
00:16:06,520 --> 00:16:09,280
Speaker 1: centers could use that, but that's not really practical for

287
00:16:09,320 --> 00:16:12,960
Speaker 1: their usage. They rely on what's called blue water, the

288
00:16:13,000 --> 00:16:16,520
Speaker 1: stuff that's clean enough for humans to drink. So when

289
00:16:16,600 --> 00:16:19,640
Speaker 1: Chalet is comparing a tech company's usage of water to

290
00:16:19,720 --> 00:16:23,200
Speaker 1: say like Pepsi's global use of water, this is a

291
00:16:23,240 --> 00:16:28,320
Speaker 1: pretty direct comparison. Use the phrase when you're evaluating GPT three,

292
00:16:28,880 --> 00:16:32,000
Speaker 1: that GBT three needs to drink a certain amount of water.

293
00:16:32,120 --> 00:16:35,840
Speaker 3: A rap flight ten to fifty queries for five hundred

294
00:16:35,840 --> 00:16:38,120
Speaker 3: million digits of water, so basically a bottle water.

295
00:16:38,360 --> 00:16:41,280
Speaker 1: Let's pause on that number for a second. Ten to

296
00:16:41,280 --> 00:16:43,640
Speaker 1: fifty queries the kind of thing you might do in

297
00:16:43,640 --> 00:16:47,760
Speaker 1: a single session using chat GPT that could drink half

298
00:16:47,800 --> 00:16:50,800
Speaker 1: a liter of water. I'm pretty sure that me going

299
00:16:50,840 --> 00:16:53,520
Speaker 1: back and forth about buying a car, I probably used

300
00:16:53,560 --> 00:16:58,080
Speaker 1: about a leader, and that's using conservative estimates. Challeat and

301
00:16:58,120 --> 00:17:01,280
Speaker 1: his team. We're focusing on GPT three, which was released

302
00:17:01,320 --> 00:17:05,520
Speaker 1: back in twenty twenty. Even five years later, OpenAI hasn't

303
00:17:05,520 --> 00:17:08,359
Speaker 1: released all the details researchers would need to give us

304
00:17:08,359 --> 00:17:12,040
Speaker 1: a clear picture of its environmental impact. Do the companies

305
00:17:12,080 --> 00:17:13,400
Speaker 1: know how much water that they're using?

306
00:17:13,480 --> 00:17:15,600
Speaker 3: Of course I can't really speak on their behalf, but

307
00:17:15,680 --> 00:17:18,800
Speaker 3: I think they do. They could figure out the water

308
00:17:18,960 --> 00:17:22,240
Speaker 3: consumption easily because they know their energy they know their

309
00:17:22,240 --> 00:17:24,719
Speaker 3: water efficiency of the courting system, they know where they

310
00:17:24,760 --> 00:17:27,600
Speaker 3: build the data centers, so they have the information, but

311
00:17:27,720 --> 00:17:29,399
Speaker 3: we're not seeing their own disclosure.

312
00:17:29,800 --> 00:17:31,960
Speaker 1: By this point you might be picking up on a

313
00:17:32,000 --> 00:17:35,720
Speaker 1: recurring theme here. Putting a specific number on the impact

314
00:17:35,720 --> 00:17:39,679
Speaker 1: of AI is basically impossible, and it's not because the

315
00:17:39,760 --> 00:17:40,919
Speaker 1: math is too difficult.

316
00:17:41,080 --> 00:17:43,720
Speaker 2: The thing is the tech companies are also refusing to

317
00:17:43,800 --> 00:17:46,119
Speaker 2: tell us exactly what's going on. So if you take

318
00:17:46,160 --> 00:17:48,840
Speaker 2: Google's environmental report, it will show you the numbers are

319
00:17:48,880 --> 00:17:52,000
Speaker 2: bad because in twenty three they show that their carbon

320
00:17:52,040 --> 00:17:55,440
Speaker 2: emations were up like fifty percent compared to five years before,

321
00:17:55,560 --> 00:17:58,439
Speaker 2: and they were pointing to AI as the main culprit.

322
00:17:58,480 --> 00:18:02,760
Speaker 2: They were saying, Okay, data center infrastructure is adding to

323
00:18:02,800 --> 00:18:06,159
Speaker 2: our combon emissions, we're using more electricity. And at the

324
00:18:06,160 --> 00:18:10,920
Speaker 2: same time they just don't specify exactly what's going on

325
00:18:11,000 --> 00:18:14,480
Speaker 2: with regard to AI. They say that making distinctions is

326
00:18:14,520 --> 00:18:17,399
Speaker 2: not meaningful at all, even though weirdly, Google was the

327
00:18:17,480 --> 00:18:21,240
Speaker 2: company that just three years ago was in fact making

328
00:18:21,240 --> 00:18:24,200
Speaker 2: this distinction. They were disclosing the ten to fifteen percent

329
00:18:24,240 --> 00:18:27,720
Speaker 2: of their total energy costs were related to artificial intelligence.

330
00:18:27,800 --> 00:18:29,640
Speaker 2: Now they stop doing that they don't want to tell

331
00:18:29,720 --> 00:18:30,320
Speaker 2: us anymore.

332
00:18:30,560 --> 00:18:33,800
Speaker 1: All of a sudden, they it seems like something changed there.

333
00:18:33,800 --> 00:18:35,240
Speaker 1: What do you think changed?

334
00:18:35,480 --> 00:18:37,479
Speaker 2: The numbers got big, that's what's changed.

335
00:18:38,800 --> 00:18:41,760
Speaker 1: Okay, not to spoil the end here, but it looks

336
00:18:41,760 --> 00:18:43,439
Speaker 1: like I'm not going to get a direct answer to

337
00:18:43,480 --> 00:18:46,200
Speaker 1: my question. But at least I have something of a ballpark,

338
00:18:46,320 --> 00:18:49,159
Speaker 1: even if it's a conservative one. And I also know

339
00:18:49,240 --> 00:18:53,120
Speaker 1: that we're using AI every day for everything. We might

340
00:18:53,160 --> 00:18:56,600
Speaker 1: not know the exact environmental impact of AI, but we

341
00:18:56,680 --> 00:19:00,159
Speaker 1: do know that it's increasing, So what do we do

342
00:19:00,200 --> 00:19:12,880
Speaker 1: about it? That's after the break. So in this episode,

343
00:19:13,000 --> 00:19:16,159
Speaker 1: we've been having some trouble figuring out the exact environmental

344
00:19:16,160 --> 00:19:19,680
Speaker 1: costs of AI. But this is a pretty common problem.

345
00:19:19,960 --> 00:19:22,280
Speaker 1: I mean, my friend Matthew Galt wrote up an article

346
00:19:22,280 --> 00:19:26,320
Speaker 1: at four or four Media explaining that the Government Accountability Office,

347
00:19:26,600 --> 00:19:29,919
Speaker 1: which is a nonpartisan group that answers to Congress, is

348
00:19:29,960 --> 00:19:33,359
Speaker 1: struggling with the exact same thing. They came up with

349
00:19:33,480 --> 00:19:35,960
Speaker 1: roughly the same numbers that we talked about earlier. They

350
00:19:36,000 --> 00:19:39,119
Speaker 1: put together a forty seven page report that acknowledges that

351
00:19:39,359 --> 00:19:44,920
Speaker 1: even after interviewing agency officials, researchers, experts, they're still left

352
00:19:44,960 --> 00:19:48,880
Speaker 1: with having to do estimates because, as they said, quote,

353
00:19:49,040 --> 00:19:53,520
Speaker 1: generative AI uses significant energy and water resources, but companies

354
00:19:53,600 --> 00:19:57,840
Speaker 1: are generally not reporting details of these uses. So even

355
00:19:57,880 --> 00:20:02,000
Speaker 1: the US government has no idea exactly how much carbon

356
00:20:02,040 --> 00:20:04,800
Speaker 1: we're pumping out or how much water we're pouring into

357
00:20:04,840 --> 00:20:09,200
Speaker 1: the sand. And this is an issue because when researchers

358
00:20:09,240 --> 00:20:12,680
Speaker 1: like Chalet and Alex were first looking into AI's environmental impact,

359
00:20:13,200 --> 00:20:17,160
Speaker 1: the biggest concern was training. That's the one time process

360
00:20:17,200 --> 00:20:20,160
Speaker 1: of feeding those massive data sets in the powerful machines.

361
00:20:20,640 --> 00:20:24,200
Speaker 1: That's what was making headlines for energy use. But then

362
00:20:24,440 --> 00:20:29,120
Speaker 1: came chat GBT three and suddenly people weren't just training models,

363
00:20:29,480 --> 00:20:33,719
Speaker 1: they were using them all the time, and that shift

364
00:20:34,000 --> 00:20:35,200
Speaker 1: changed everything.

365
00:20:35,840 --> 00:20:38,480
Speaker 2: As an end user, you can't even manage it properly

366
00:20:38,560 --> 00:20:41,160
Speaker 2: because the companies are not telling you. So it's not

367
00:20:41,280 --> 00:20:44,800
Speaker 2: like when you're interacting with chut GPT that judge GPT

368
00:20:44,920 --> 00:20:47,760
Speaker 2: is gonna tell you, Okay, be aware, now the carbon

369
00:20:47,800 --> 00:20:51,840
Speaker 2: footprint of this conversation has already exceeded this amount. Open

370
00:20:51,880 --> 00:20:53,840
Speaker 2: AI knows this kind of stuff. They could tell you,

371
00:20:53,880 --> 00:20:56,679
Speaker 2: but they won't, And then other people are left trying

372
00:20:56,800 --> 00:20:58,919
Speaker 2: to make some kind of customer to figure out what

373
00:20:59,040 --> 00:21:01,160
Speaker 2: might be going on. We also see that they are

374
00:21:01,240 --> 00:21:03,880
Speaker 2: kind of downplaying the impact of what they're doing here.

375
00:21:03,920 --> 00:21:07,480
Speaker 2: I mean, we see their environmental reports or disasters. The

376
00:21:07,560 --> 00:21:09,920
Speaker 2: carbon emitions are shooting up, and the only thing they're

377
00:21:09,920 --> 00:21:12,480
Speaker 2: saying is like, Okay, don't worry about it. AI will

378
00:21:12,520 --> 00:21:14,480
Speaker 2: solve this in a couple of years from now.

379
00:21:14,720 --> 00:21:17,560
Speaker 1: So the thing that's causing the problem is going to

380
00:21:17,600 --> 00:21:18,439
Speaker 1: solve the problem.

381
00:21:18,440 --> 00:21:22,480
Speaker 2: Also, yeah, that's the excuse they're using. AI is going

382
00:21:22,520 --> 00:21:24,920
Speaker 2: to solve it. It's bad right now, but everything will

383
00:21:24,920 --> 00:21:26,879
Speaker 2: be better in a couple of years, trust us. But

384
00:21:27,119 --> 00:21:31,320
Speaker 2: it's one hundred percent wishful thinking. And to be honest,

385
00:21:31,320 --> 00:21:34,880
Speaker 2: if you look at the whole history of technological developments,

386
00:21:34,960 --> 00:21:37,320
Speaker 2: even if we do end up realizing a lot of

387
00:21:37,359 --> 00:21:40,879
Speaker 2: efficiency gains with AI, this is definitely not a given.

388
00:21:40,960 --> 00:21:43,639
Speaker 2: It doesn't mean that our resource uses in total is

389
00:21:43,680 --> 00:21:46,439
Speaker 2: going to go down. This is the infamous Jevins paradox.

390
00:21:46,920 --> 00:21:49,680
Speaker 1: Jevins paradox is a concept that comes up a lot

391
00:21:49,680 --> 00:21:54,359
Speaker 1: in AI recently. Basically, in the Industrial Revolution, cold powered

392
00:21:54,440 --> 00:21:58,399
Speaker 1: engines started to get more efficient, and some people assume that, Okay,

393
00:21:58,400 --> 00:21:59,960
Speaker 1: this is going to mean that now we're going to

394
00:22:00,200 --> 00:22:04,600
Speaker 1: use less coal overall, but an economist named William Jevins said, no,

395
00:22:05,040 --> 00:22:07,840
Speaker 1: this is going to have the opposite effect. As coal

396
00:22:07,880 --> 00:22:12,520
Speaker 1: powered energy gets cheaper, demand will increase, and total consumption

397
00:22:12,600 --> 00:22:16,320
Speaker 1: of coal won't go down, it'll go up. He was right,

398
00:22:16,600 --> 00:22:18,840
Speaker 1: and that effect seems to keep repeating.

399
00:22:19,520 --> 00:22:22,639
Speaker 2: Despite all the efficiency gains that we had. We're not

400
00:22:22,800 --> 00:22:26,320
Speaker 2: saving on resources, we are using more resources.

401
00:22:26,440 --> 00:22:29,400
Speaker 1: And essentially you're saying here is even if we are

402
00:22:29,600 --> 00:22:32,119
Speaker 1: able to make AI more efficient, we're just going to

403
00:22:32,240 --> 00:22:35,520
Speaker 1: use it more, and so any efficiency gains are going

404
00:22:35,600 --> 00:22:37,520
Speaker 1: to be offset by the fact that we're just constantly

405
00:22:37,600 --> 00:22:38,840
Speaker 1: using this more and more and more.

406
00:22:39,200 --> 00:22:42,440
Speaker 2: One thing that's extra annoying with AI is that there's

407
00:22:42,520 --> 00:22:45,440
Speaker 2: also this bigger is better dynamic going on, whereas if

408
00:22:45,480 --> 00:22:48,080
Speaker 2: you make the models bigger, you'll actually end up with

409
00:22:48,160 --> 00:22:50,720
Speaker 2: a better performing model, but it just means that your

410
00:22:50,760 --> 00:22:53,320
Speaker 2: efficiency gains are completely negated all the time.

411
00:22:53,760 --> 00:22:58,680
Speaker 1: Every chat, every prompt, every AI generated jibbli image adds up.

412
00:22:58,960 --> 00:23:03,000
Speaker 1: We just don't see that impact directly, So let's all

413
00:23:03,040 --> 00:23:07,320
Speaker 1: just stop using AI. Right, Well, that's probably not realistic

414
00:23:07,320 --> 00:23:10,760
Speaker 1: at this point, and that's not necessarily what everyone's recommending.

415
00:23:10,880 --> 00:23:13,000
Speaker 3: So I work on optimization, and I think this is

416
00:23:13,040 --> 00:23:15,320
Speaker 3: a problem. We can optimize it, we can make it better,

417
00:23:15,359 --> 00:23:18,080
Speaker 3: reduce the cost, and there are a lot of opportunities,

418
00:23:18,520 --> 00:23:21,800
Speaker 3: so we should definitely not panic. I hope the model

419
00:23:21,840 --> 00:23:24,520
Speaker 3: developer can disclosee that cost to the users. I will

420
00:23:24,520 --> 00:23:26,680
Speaker 3: figure out should I use it now or should I

421
00:23:26,760 --> 00:23:27,240
Speaker 3: use it later.

422
00:23:27,840 --> 00:23:30,760
Speaker 1: Let's say that I log in the chat GBT and

423
00:23:30,840 --> 00:23:34,840
Speaker 1: it says this query is going to use this much energy,

424
00:23:35,359 --> 00:23:38,520
Speaker 1: this much carbon, and this much water. And if I

425
00:23:38,560 --> 00:23:41,720
Speaker 1: have that information up front, then I, the user might

426
00:23:41,760 --> 00:23:45,040
Speaker 1: decide maybe I don't need to have it summarize the

427
00:23:45,200 --> 00:23:47,320
Speaker 1: entirety of the collective works of Shakespeare today.

428
00:23:47,720 --> 00:23:47,960
Speaker 2: Yeah.

429
00:23:48,119 --> 00:23:51,520
Speaker 3: Maybe. Or they could tell you if you do it later,

430
00:23:51,640 --> 00:23:54,760
Speaker 3: in one hour or in the evening, the cost will

431
00:23:54,800 --> 00:23:57,520
Speaker 3: be different. And do you figure it out? Do you

432
00:23:57,560 --> 00:23:59,119
Speaker 3: want to do it now or do it later.

433
00:24:00,280 --> 00:24:03,240
Speaker 1: What Shelley is proposing here is that developers could build

434
00:24:03,240 --> 00:24:05,960
Speaker 1: in a system that would alert users that their query

435
00:24:06,040 --> 00:24:08,400
Speaker 1: is coming in at a high impact time of day,

436
00:24:08,920 --> 00:24:11,120
Speaker 1: and it could suggest that there might be a better

437
00:24:11,160 --> 00:24:14,760
Speaker 1: time to make that request when data centers have lower usage.

438
00:24:14,880 --> 00:24:19,040
Speaker 1: They can use optimization techniques to reduce energy consumption. This

439
00:24:19,160 --> 00:24:23,520
Speaker 1: concept isn't totally new. Google flights shows carbon emissions estimates

440
00:24:23,560 --> 00:24:26,080
Speaker 1: for flights and it will show you which option has

441
00:24:26,080 --> 00:24:29,399
Speaker 1: the least impact. So something like this for AI is

442
00:24:29,440 --> 00:24:34,000
Speaker 1: definitely possible, but I'm not totally convinced people would actually care.

443
00:24:34,680 --> 00:24:36,720
Speaker 1: The last time I booked a flight, I saw the

444
00:24:36,720 --> 00:24:40,040
Speaker 1: most carbon friendly option, but I didn't pick it because

445
00:24:40,080 --> 00:24:42,000
Speaker 1: it had a long layover. I didn't want to deal

446
00:24:42,040 --> 00:24:46,040
Speaker 1: with that. Putting the responsibility on users can sound good

447
00:24:46,040 --> 00:24:48,679
Speaker 1: in theory, but the flip side of that is it

448
00:24:48,680 --> 00:24:50,920
Speaker 1: can just be a way for companies to avoid doing

449
00:24:50,920 --> 00:24:55,720
Speaker 1: anything themselves. So should this responsibility really fall on us?

450
00:24:56,240 --> 00:24:58,840
Speaker 1: I mean, sure, you could decide to skip the chatbot

451
00:24:58,920 --> 00:25:01,679
Speaker 1: and take notes by hand, and that only really works

452
00:25:01,760 --> 00:25:04,520
Speaker 1: if you know what the trade off actually is, and

453
00:25:04,640 --> 00:25:08,440
Speaker 1: right now we don't, because the companies building these tools

454
00:25:08,560 --> 00:25:10,719
Speaker 1: aren't giving us the data that we would need to

455
00:25:10,760 --> 00:25:14,359
Speaker 1: make informed decisions in the first place. So maybe the

456
00:25:14,400 --> 00:25:19,080
Speaker 1: responsibility should fall elsewhere. Like policy makers, Shelley is already

457
00:25:19,080 --> 00:25:21,280
Speaker 1: thinking about what this could look like and how much

458
00:25:21,280 --> 00:25:22,280
Speaker 1: of a difference it could make.

459
00:25:22,359 --> 00:25:26,199
Speaker 3: We're informing the policy makers, hopefully when they make decisions

460
00:25:26,440 --> 00:25:30,960
Speaker 3: they could take into account this public health burden, water consumption,

461
00:25:31,200 --> 00:25:35,000
Speaker 3: power strain on their infrastructures. These are the cost the

462
00:25:35,040 --> 00:25:38,480
Speaker 3: local people will be paying for the companies. I think,

463
00:25:38,760 --> 00:25:42,440
Speaker 3: especially for those big techs, they already have the systems

464
00:25:42,680 --> 00:25:45,879
Speaker 3: ready to do this type of optimization. They are doing

465
00:25:45,920 --> 00:25:48,640
Speaker 3: it for carbon orware computing. And we use the math

466
00:25:48,680 --> 00:25:51,679
Speaker 3: as location as an example. If they factor in the

467
00:25:51,720 --> 00:25:55,359
Speaker 3: public health burden into their decision making, for example, where

468
00:25:55,400 --> 00:25:57,840
Speaker 3: they route to their workload, they can reduce the public

469
00:25:57,880 --> 00:26:01,639
Speaker 3: health cost by about twenty five percent really and reduce

470
00:26:01,680 --> 00:26:05,520
Speaker 3: the energy bo by about two percent and also cut

471
00:26:05,640 --> 00:26:08,520
Speaker 3: the carbon by about one point three percent.

472
00:26:09,240 --> 00:26:12,280
Speaker 1: So just by being more intentional where they route digital traffic,

473
00:26:12,560 --> 00:26:16,040
Speaker 1: a company like Meta could reduce detrimental impacts on public

474
00:26:16,080 --> 00:26:19,000
Speaker 1: health and they'd be saving some cash at the same time.

475
00:26:19,680 --> 00:26:23,359
Speaker 1: This is called geographic load balancing, and for the user

476
00:26:23,600 --> 00:26:27,560
Speaker 1: it's totally seamless. You log in your feed loads, you

477
00:26:27,600 --> 00:26:30,960
Speaker 1: don't notice anything, but behind the scenes, your request is

478
00:26:31,000 --> 00:26:35,280
Speaker 1: going somewhere where it's cleaner, cheaper, and less harmful to process.

479
00:26:36,000 --> 00:26:39,240
Speaker 1: Even beyond where companies route traffic, they can also consider

480
00:26:39,280 --> 00:26:42,680
Speaker 1: where they build the data centers from a public health perspective.

481
00:26:43,040 --> 00:26:45,560
Speaker 3: When they built data centers in the future, they can

482
00:26:45,720 --> 00:26:48,840
Speaker 3: take into account this of factors because the decision that

483
00:26:48,880 --> 00:26:51,719
Speaker 3: we make today will be impacting the public health, the

484
00:26:51,760 --> 00:26:54,920
Speaker 3: water consumption, the power infrastructure for many years to come.

485
00:26:55,800 --> 00:26:58,520
Speaker 1: Shelle is thinking about the future and research on future

486
00:26:58,520 --> 00:27:02,679
Speaker 1: optimization is a big deal because the AI boom is

487
00:27:02,840 --> 00:27:06,440
Speaker 1: already here. Big tech companies are projected to spend three

488
00:27:06,520 --> 00:27:10,040
Speaker 1: hundred and twenty billion dollars on AI technology and data

489
00:27:10,080 --> 00:27:12,880
Speaker 1: centers this year, which is nearly one hundred billion more

490
00:27:12,920 --> 00:27:16,160
Speaker 1: than last year. So where we put these data centers

491
00:27:16,200 --> 00:27:19,720
Speaker 1: and where we route the traffic really matters.

492
00:27:19,720 --> 00:27:23,560
Speaker 3: Something that I was not expecting to be widespread because

493
00:27:23,560 --> 00:27:26,359
Speaker 3: I was thinking, if I leave, let's say, five miles

494
00:27:26,359 --> 00:27:28,879
Speaker 3: away from a data center or power plan, I wouldn't

495
00:27:28,880 --> 00:27:32,720
Speaker 3: be affected. That was wrong. These air pollutants are what

496
00:27:32,760 --> 00:27:37,280
Speaker 3: EPA defines as cross state air pollutants. They do travel

497
00:27:37,400 --> 00:27:39,919
Speaker 3: hundreds of miles along with a wind. We're going to

498
00:27:40,040 --> 00:27:45,040
Speaker 3: have a significant impact just by strategically placing the data

499
00:27:45,040 --> 00:27:46,240
Speaker 3: centers for the public health.

500
00:27:46,680 --> 00:27:49,120
Speaker 1: What that really highlights is something that we don't think

501
00:27:49,119 --> 00:27:52,480
Speaker 1: about with tech infrastructure. It doesn't just impact the people

502
00:27:52,480 --> 00:27:56,280
Speaker 1: who live next door. When air pollution travels hundreds of miles,

503
00:27:56,600 --> 00:27:59,919
Speaker 1: it turns these data centers into regional issues, not just

504
00:28:00,160 --> 00:28:03,320
Speaker 1: local ones. I'll give you an example right here. As

505
00:28:03,320 --> 00:28:05,440
Speaker 1: we're working on this episode, I saw this article in

506
00:28:05,480 --> 00:28:07,240
Speaker 1: Politico and I just want to read you the first

507
00:28:07,320 --> 00:28:12,760
Speaker 1: sentence quote Elon Musk's artificial intelligence company is belching smog

508
00:28:12,840 --> 00:28:16,760
Speaker 1: forming pollution into an area of South Memphis that already

509
00:28:16,840 --> 00:28:21,240
Speaker 1: leads the state in emergency department visits for ASTHMA end quote.

510
00:28:21,640 --> 00:28:23,679
Speaker 1: That's probably enough to give you the idea. But just

511
00:28:23,720 --> 00:28:27,480
Speaker 1: to explain more, XAI, which is the company behind groc

512
00:28:27,600 --> 00:28:30,040
Speaker 1: which is the AI chatbot that you use on Twitter,

513
00:28:30,400 --> 00:28:33,720
Speaker 1: set up shop in Memphis with enough methane gas turbines

514
00:28:33,800 --> 00:28:38,120
Speaker 1: to power two hundred and eighty thousand homes. The company

515
00:28:38,200 --> 00:28:41,320
Speaker 1: didn't get the required air pollution permits. They're run without

516
00:28:41,320 --> 00:28:45,000
Speaker 1: the emission controls that federal law usually requires, and in

517
00:28:45,120 --> 00:28:48,640
Speaker 1: under a year of operation, XAI is now one of

518
00:28:48,680 --> 00:28:52,400
Speaker 1: the largest emitters of smog producing nitrogen oxides in the

519
00:28:52,560 --> 00:28:56,600
Speaker 1: entire county. And this facility is located near predominantly black

520
00:28:56,640 --> 00:29:00,600
Speaker 1: neighborhoods that are already dealing with high levels of indust pollution.

521
00:29:01,480 --> 00:29:06,160
Speaker 1: These inequalities already existed, and tech development is not making

522
00:29:06,160 --> 00:29:10,440
Speaker 1: it better, it's making it worse. It is often like this.

523
00:29:12,080 --> 00:29:14,880
Speaker 1: There are absolutely people who are feeling the impacts of

524
00:29:14,920 --> 00:29:18,720
Speaker 1: this right now, and there's people who will feel it

525
00:29:18,760 --> 00:29:22,960
Speaker 1: in the future. Maybe somebody will write an article about them,

526
00:29:23,120 --> 00:29:26,400
Speaker 1: maybe not. So I was hoping that I could use

527
00:29:26,440 --> 00:29:30,040
Speaker 1: this podcast to solve all my personal problems. But apparently

528
00:29:30,080 --> 00:29:32,680
Speaker 1: we're over one here, because when I started working on

529
00:29:32,720 --> 00:29:35,680
Speaker 1: this episode, I was thinking that this section right here,

530
00:29:35,800 --> 00:29:38,800
Speaker 1: the outro is where I'd say, Wow, now I know

531
00:29:39,040 --> 00:29:41,760
Speaker 1: exactly what impact my use of AI is having on

532
00:29:41,800 --> 00:29:47,120
Speaker 1: the planet. But I don't. And that's pretty annoying because,

533
00:29:47,560 --> 00:29:49,280
Speaker 1: and I guess this is as close to an answer

534
00:29:49,320 --> 00:29:52,160
Speaker 1: as we're going to get. It's not really about how

535
00:29:52,200 --> 00:29:55,920
Speaker 1: often I personally decide to use Chat, GPT or Gemini

536
00:29:56,000 --> 00:29:59,320
Speaker 1: or Claude or whatever. It's about what happens when companies

537
00:29:59,320 --> 00:30:03,640
Speaker 1: build systems that are this powerful but also this resource hungry,

538
00:30:04,040 --> 00:30:06,760
Speaker 1: and they refuse to tell us what it really costs.

539
00:30:07,440 --> 00:30:10,120
Speaker 1: And I think we deserve to know, not just so

540
00:30:10,160 --> 00:30:12,880
Speaker 1: that we can make individual choices about how often to

541
00:30:13,000 --> 00:30:16,080
Speaker 1: use Chat, GIBT or Gemini or whatever, but so that

542
00:30:16,080 --> 00:30:19,560
Speaker 1: we can hold the right people accountable, because if AI

543
00:30:19,720 --> 00:30:21,920
Speaker 1: is really going to change the future like they say

544
00:30:21,920 --> 00:30:24,800
Speaker 1: it will, we shouldn't know how much that future costs.

545
00:30:35,080 --> 00:30:37,880
Speaker 1: Thank you so much for listening to kill Switch. If

546
00:30:37,920 --> 00:30:40,400
Speaker 1: you got any ideas or thoughts about the show, you

547
00:30:40,440 --> 00:30:43,880
Speaker 1: could hit us at kill Switch at Kaleidoscope dot NYC,

548
00:30:44,240 --> 00:30:46,560
Speaker 1: or you could hit me at dex Digi that's d

549
00:30:46,720 --> 00:30:49,880
Speaker 1: e x d ig I on Instagram or on Blue

550
00:30:49,920 --> 00:30:52,360
Speaker 1: Sky if that's more your thing. And if you like

551
00:30:52,440 --> 00:30:56,120
Speaker 1: this episode, if you're on Apple Podcasts or Spotify, take

552
00:30:56,160 --> 00:30:58,840
Speaker 1: your phone out your pocket and leave us a review.

553
00:30:59,080 --> 00:31:01,720
Speaker 1: It really helps people find the show, and in turn,

554
00:31:01,920 --> 00:31:04,440
Speaker 1: that helps us keep doing our thing. Kill Switch is

555
00:31:04,440 --> 00:31:08,000
Speaker 1: hosted by Me Dexter Thomas. It's produced by sen Ozaki,

556
00:31:08,400 --> 00:31:11,840
Speaker 1: darl Luk Potts, and Kate Osborne. Our theme song is

557
00:31:11,880 --> 00:31:14,800
Speaker 1: by me and Kyle Murdoch, and Kyle also mixed the

558
00:31:14,840 --> 00:31:20,160
Speaker 1: show from Kaleidoscope. Our executive producers are Ozma Lashin, Mungesh Hatigadour,

559
00:31:20,560 --> 00:31:24,920
Speaker 1: and Kate Osborne from iHeart. Our executive producers are Katrina

560
00:31:25,000 --> 00:31:27,680
Speaker 1: Norville and Nikki e. Tour. See you on the next

561
00:31:27,720 --> 00:31:29,680
Speaker 1: one