1 00:00:00,160 --> 00:00:03,760 Speaker 1: Jeff Adams as a computer scientist. In two thousand and eleven, 2 00:00:03,880 --> 00:00:07,120 Speaker 1: he was at a really tiny startup, just thirteen employees, 3 00:00:07,440 --> 00:00:10,360 Speaker 1: called the App. They were working on technology that could 4 00:00:10,400 --> 00:00:15,080 Speaker 1: transcribe voicemail into text messages. And then one day Jeff's 5 00:00:15,080 --> 00:00:19,239 Speaker 1: fairy godmother showed up in the form of executives from Amazon. 6 00:00:19,720 --> 00:00:22,000 Speaker 1: We went to a show and some people from Amazon 7 00:00:22,160 --> 00:00:24,640 Speaker 1: came and talked to us, and they might be interested 8 00:00:24,720 --> 00:00:28,360 Speaker 1: in possibly even acquiring us. And I thought, that's absurd. 9 00:00:28,480 --> 00:00:32,120 Speaker 1: What could Amazon possibly be interested in speech for? Maybe 10 00:00:32,440 --> 00:00:35,600 Speaker 1: they want you to be able to, you know, call 11 00:00:35,680 --> 00:00:38,600 Speaker 1: up on the phone and order a book. Jeff figured 12 00:00:38,600 --> 00:00:41,040 Speaker 1: the talks wouldn't amount too much, but over the course 13 00:00:41,080 --> 00:00:45,000 Speaker 1: of the next few months they became serious. Naturally, he 14 00:00:45,080 --> 00:00:48,280 Speaker 1: wanted to know why Amazon was interested. They said, you know, 15 00:00:48,360 --> 00:00:51,000 Speaker 1: don't ask, it's our business. And I thought, well, it's 16 00:00:51,000 --> 00:00:53,000 Speaker 1: gonna be my business too, and you might want to 17 00:00:53,040 --> 00:00:58,280 Speaker 1: know whether whether we can do it um, But they wouldn't. 18 00:00:58,280 --> 00:01:00,160 Speaker 1: They wouldn't say anything. They said, I'm sorry, were not 19 00:01:00,200 --> 00:01:02,639 Speaker 1: at liberty to discuss any of this. There was one 20 00:01:02,720 --> 00:01:05,120 Speaker 1: hint that Amazon would give, so one of the things 21 00:01:05,200 --> 00:01:07,600 Speaker 1: they did let us know was that the guy who 22 00:01:07,600 --> 00:01:10,440 Speaker 1: was kind of running this was Greg Hart, who we 23 00:01:10,560 --> 00:01:14,080 Speaker 1: knew was the right hand person of Jeff Bezos. So 24 00:01:14,160 --> 00:01:16,280 Speaker 1: they go to all the meetings with them. They are 25 00:01:16,319 --> 00:01:20,360 Speaker 1: like the confidante, the consilior or whatever. Uh. And so 26 00:01:20,400 --> 00:01:23,800 Speaker 1: we knew, okay, something this does have visibility at the 27 00:01:23,840 --> 00:01:27,520 Speaker 1: highest levels, and Bezos must be behind it in in 28 00:01:27,680 --> 00:01:31,000 Speaker 1: some form. As the deal was being finalized, Jeff's team 29 00:01:31,040 --> 00:01:34,760 Speaker 1: traveled to Florence, Italy, to a computer speech conference along 30 00:01:34,800 --> 00:01:38,479 Speaker 1: with several Amazon managers. They even stayed in a villa together, 31 00:01:38,959 --> 00:01:41,399 Speaker 1: but the Amazon people didn't want to be seen with 32 00:01:41,440 --> 00:01:45,320 Speaker 1: the voice transcription people during the conference. From day to 33 00:01:45,400 --> 00:01:49,120 Speaker 1: day in the conference, we had to let we we 34 00:01:49,200 --> 00:01:51,800 Speaker 1: I rented a van and we drove down into town 35 00:01:51,880 --> 00:01:54,240 Speaker 1: and to the conference together. I had to let them 36 00:01:54,240 --> 00:01:57,160 Speaker 1: out around the block, around the corner. It's like being 37 00:01:57,200 --> 00:01:59,920 Speaker 1: a teenager and trying to hide the person you're dating 38 00:02:00,160 --> 00:02:02,280 Speaker 1: from your parents. First of all, they didn't want anyone 39 00:02:02,280 --> 00:02:04,160 Speaker 1: to know that they were from Amazon. They didn't want 40 00:02:04,160 --> 00:02:06,320 Speaker 1: anyone to see them with us, so we had to 41 00:02:06,440 --> 00:02:08,480 Speaker 1: avoid them at the conference. We would like, you know, 42 00:02:08,560 --> 00:02:12,280 Speaker 1: look at each other and kind of smile from across 43 00:02:12,360 --> 00:02:15,000 Speaker 1: the room or across the courtyard or whatever. But we 44 00:02:15,000 --> 00:02:18,160 Speaker 1: couldn't sit together, we couldn't talk together, or whatever. That 45 00:02:18,280 --> 00:02:22,440 Speaker 1: fall the deal was done. Amazon buys the speech company, 46 00:02:22,520 --> 00:02:25,480 Speaker 1: and Jeff is still living in mystery. We had been 47 00:02:25,520 --> 00:02:28,200 Speaker 1: employees of Amazon for like a week, but they had 48 00:02:28,240 --> 00:02:31,040 Speaker 1: still not told us any anything. They said, no, we 49 00:02:31,080 --> 00:02:34,000 Speaker 1: have to tell you in a closed room. So we 50 00:02:34,040 --> 00:02:36,360 Speaker 1: all came out to Seattle. We all got in the 51 00:02:36,400 --> 00:02:41,360 Speaker 1: room together. They closed the door, locked the door, put 52 00:02:41,680 --> 00:02:46,200 Speaker 1: paper over the window in the door, closed the exterior windows. Uh. 53 00:02:46,440 --> 00:02:51,320 Speaker 1: It was very secret, very mysterious, and hush, hush. Jeff 54 00:02:51,360 --> 00:02:55,520 Speaker 1: and his team leaned forward. This was the moment they said, 55 00:02:55,919 --> 00:02:59,519 Speaker 1: imagine something the size of a coke can that's sitting 56 00:02:59,560 --> 00:03:02,000 Speaker 1: on your able and we're going to sell this for 57 00:03:04,720 --> 00:03:06,760 Speaker 1: uh and people will be able to talk to it. 58 00:03:06,840 --> 00:03:10,200 Speaker 1: And we thought, you know, we can't do this. The 59 00:03:10,280 --> 00:03:13,360 Speaker 1: technology isn't there yet, it doesn't exist yet. And of 60 00:03:13,400 --> 00:03:17,359 Speaker 1: course they did do it by inventing the technology, because 61 00:03:17,360 --> 00:03:22,440 Speaker 1: the product they're talking about becomes Alexa, Amazon's AI virtual assistant, 62 00:03:22,720 --> 00:03:26,000 Speaker 1: and that twenty co can is the Echo smart speaker. 63 00:03:26,360 --> 00:03:29,640 Speaker 1: Though it ends up selling for more than twenty Alexa 64 00:03:29,800 --> 00:03:33,480 Speaker 1: is one of Amazon's first forays into artificial intelligence, and 65 00:03:33,520 --> 00:03:37,080 Speaker 1: it would bring Amazon into people's living rooms, further weaving 66 00:03:37,120 --> 00:03:43,760 Speaker 1: it into their lives. You're listening to Foundering. I'm your host, 67 00:03:43,960 --> 00:03:47,320 Speaker 1: Brad Stone. In this episode, we're going to tell the 68 00:03:47,400 --> 00:03:51,080 Speaker 1: story of the creation of Alexa. How Jeff Bezos almost 69 00:03:51,160 --> 00:03:55,440 Speaker 1: singlehandedly conceived the idea for a voice activated computer and 70 00:03:55,640 --> 00:04:00,400 Speaker 1: drove the device's creation. Alexa helped to solidify Amazon's image 71 00:04:00,400 --> 00:04:04,560 Speaker 1: as an innovator and in the process challenge conventional notions 72 00:04:04,560 --> 00:04:07,840 Speaker 1: of privacy. Will tell you more after a quick break. 73 00:04:19,080 --> 00:04:22,000 Speaker 1: In the last episode, we talked about how Amazon survived 74 00:04:22,000 --> 00:04:25,000 Speaker 1: the dot com bust, how Jeff Bezos said to colleagues, 75 00:04:25,240 --> 00:04:27,320 Speaker 1: the only way out of this is to invent her 76 00:04:27,320 --> 00:04:31,080 Speaker 1: way out. During those next few years, Bezos launched Amazon 77 00:04:31,160 --> 00:04:35,159 Speaker 1: Prime and the Kindle. But there's another innovation from this era, 78 00:04:35,560 --> 00:04:38,880 Speaker 1: one that generates huge profits for Amazon and sets the 79 00:04:38,960 --> 00:04:42,680 Speaker 1: stage for the creation of its pioneering voice assistant. It's 80 00:04:42,720 --> 00:04:47,560 Speaker 1: Amazon's cloud service called AWS, or Amazon Web Services. It's 81 00:04:47,600 --> 00:04:51,440 Speaker 1: the most difficult of Bezos's inventions for lay people to understand, 82 00:04:52,080 --> 00:04:54,799 Speaker 1: because as much as it may already seem like Amazon 83 00:04:54,920 --> 00:04:58,279 Speaker 1: is everywhere a massive store, a network of warehouses, and 84 00:04:58,320 --> 00:05:01,039 Speaker 1: the fleet of delivery people. There's one more way in 85 00:05:01,040 --> 00:05:05,960 Speaker 1: which Amazon is ubiquitous indispensable. Amazon is literally holding up 86 00:05:06,040 --> 00:05:09,520 Speaker 1: much of the Internet today. AWS is a highly profitable, 87 00:05:09,640 --> 00:05:14,320 Speaker 1: fifty billion dollar annual business for Amazon. When you watch Netflix, 88 00:05:14,480 --> 00:05:18,440 Speaker 1: you're often streaming video from Amazon. When you send a Snapchat, 89 00:05:18,640 --> 00:05:23,120 Speaker 1: you're using Amazon servers. By two thousand and ten, Bezos 90 00:05:23,240 --> 00:05:25,760 Speaker 1: was asking everyone in the company, what are you doing 91 00:05:25,800 --> 00:05:30,039 Speaker 1: for AWS? You wanted Amazon to deepen and exploit the 92 00:05:30,080 --> 00:05:33,000 Speaker 1: early lead it already had in running massive, cutting edge 93 00:05:33,040 --> 00:05:36,320 Speaker 1: data centers. Around this time, he was at lunch with 94 00:05:36,360 --> 00:05:39,800 Speaker 1: Greg Hart, the guy described as his consiliary, and the 95 00:05:39,880 --> 00:05:43,200 Speaker 1: conversation turned how Google was starting to let people search 96 00:05:43,640 --> 00:05:47,640 Speaker 1: just by talking into their smartphones. Here's Greg and and 97 00:05:47,680 --> 00:05:49,680 Speaker 1: I said, look at how convenient is And I just 98 00:05:49,800 --> 00:05:53,640 Speaker 1: did a Google voice search on pizza and your me um. 99 00:05:53,760 --> 00:05:57,000 Speaker 1: And you know, it's just so much faster than actually typing, 100 00:05:57,360 --> 00:05:59,080 Speaker 1: you know, with your fingers or your thumbs pizza and 101 00:05:59,120 --> 00:06:01,920 Speaker 1: your me on your phone waiting for results to come back. Now, 102 00:06:02,120 --> 00:06:04,960 Speaker 1: Bezos has long been a believer in voice. Here he 103 00:06:05,080 --> 00:06:07,160 Speaker 1: is all the way back in the year two thousand 104 00:06:07,560 --> 00:06:12,040 Speaker 1: talking to Charlie Rose, I believe that for mobile commerce, 105 00:06:12,440 --> 00:06:14,640 Speaker 1: the thing that's going to be the biggest part of 106 00:06:14,640 --> 00:06:17,919 Speaker 1: that is voice, and I think there's a lot that 107 00:06:18,000 --> 00:06:20,640 Speaker 1: can happen in the In the short term, it'll be 108 00:06:20,720 --> 00:06:23,520 Speaker 1: kind of a stilted special purpose language for talking to 109 00:06:23,560 --> 00:06:25,839 Speaker 1: Amazon dot Com. But in the long term it could 110 00:06:25,839 --> 00:06:29,400 Speaker 1: even be natural language processing. I think that is a 111 00:06:29,480 --> 00:06:33,240 Speaker 1: real mind bender. Now. A decade later, it seemed like 112 00:06:33,279 --> 00:06:36,360 Speaker 1: this early prediction might come true. A few weeks after 113 00:06:36,400 --> 00:06:39,680 Speaker 1: their fateful lunch, Bazo sent an email to Greg that read, 114 00:06:40,240 --> 00:06:43,240 Speaker 1: we should build a twenty device with its brains in 115 00:06:43,240 --> 00:06:47,200 Speaker 1: the cloud that's completely controlled by your voice. The central 116 00:06:47,240 --> 00:06:50,279 Speaker 1: insight here was that the product could be inexpensive because 117 00:06:50,279 --> 00:06:53,400 Speaker 1: it's brains resided in Amazon's data centers and could be 118 00:06:53,480 --> 00:06:57,000 Speaker 1: constantly improved, So if you buy it, the product would 119 00:06:57,000 --> 00:07:00,159 Speaker 1: be upgrading itself and you wouldn't even know it. A 120 00:07:00,200 --> 00:07:03,440 Speaker 1: few weeks later, Greg started recruiting people inside the company 121 00:07:03,480 --> 00:07:07,200 Speaker 1: for the project. He proceeded with total secrecy. The only 122 00:07:07,240 --> 00:07:09,760 Speaker 1: thing he told colleagues was that it had the potential 123 00:07:09,800 --> 00:07:12,200 Speaker 1: to be bigger than the Kindle, and at the time 124 00:07:12,240 --> 00:07:15,440 Speaker 1: that seemed like laughable that that we could create something 125 00:07:15,480 --> 00:07:17,560 Speaker 1: that could be bigger than Kindle, because Kindle, you know, 126 00:07:17,560 --> 00:07:22,480 Speaker 1: at that time was now for years into its life, 127 00:07:23,760 --> 00:07:27,480 Speaker 1: dominating the e reader and e ink market and was 128 00:07:27,560 --> 00:07:30,400 Speaker 1: changing the way that people thought about meeting. Next, Greg 129 00:07:30,400 --> 00:07:33,960 Speaker 1: acquired YAP, the speech recognition company where Jeff Adams worked. 130 00:07:34,480 --> 00:07:36,480 Speaker 1: That team didn't think he knew what he was doing either. 131 00:07:37,280 --> 00:07:39,520 Speaker 1: Amazon wanted people to be able to talk to a 132 00:07:39,600 --> 00:07:43,160 Speaker 1: device across a room and haven't understand them. That's called 133 00:07:43,200 --> 00:07:47,320 Speaker 1: far field speech recognition, and the technology for it didn't exist. 134 00:07:47,720 --> 00:07:52,960 Speaker 1: Here's Jeff Adams again. The big problem is speech recognition 135 00:07:53,000 --> 00:07:56,840 Speaker 1: at the time really relied on a close talking microphone. 136 00:07:56,880 --> 00:07:59,400 Speaker 1: You needed to capture the speech close to the person's mouth. 137 00:07:59,800 --> 00:08:01,880 Speaker 1: And they were talking about, Oh, I'm going to be 138 00:08:01,880 --> 00:08:04,080 Speaker 1: in the garage and this thing is gonna coach me 139 00:08:04,160 --> 00:08:07,320 Speaker 1: through changing my car's oil, and it's gonna be over 140 00:08:07,360 --> 00:08:09,000 Speaker 1: on the other side of the garage, and I'm gonna 141 00:08:09,040 --> 00:08:11,480 Speaker 1: be shouting things to it, and it's gonna be shouting 142 00:08:11,480 --> 00:08:15,520 Speaker 1: instructions back or whatever. And we thought, you can't do that. 143 00:08:15,560 --> 00:08:19,200 Speaker 1: There's there's too many reflective surfaces in the room. They're 144 00:08:19,200 --> 00:08:22,480 Speaker 1: gonna mess up the audio. Basically, if you're shouting across 145 00:08:22,520 --> 00:08:25,600 Speaker 1: the room, that introduces lots of echo and the computer 146 00:08:25,680 --> 00:08:28,920 Speaker 1: gets confused. So I realized that this far field issue 147 00:08:28,960 --> 00:08:30,520 Speaker 1: was going to be a problem. I didn't want to 148 00:08:30,560 --> 00:08:33,240 Speaker 1: say anything in in front of the group. I didn't wanna, 149 00:08:33,320 --> 00:08:38,480 Speaker 1: you know, um, appear unsupportive. But afterwards I tracked down 150 00:08:38,480 --> 00:08:41,400 Speaker 1: Greg Hart in the hallway and I said, Greg, we're 151 00:08:41,400 --> 00:08:43,560 Speaker 1: excited about this, but I think you should know that 152 00:08:43,920 --> 00:08:47,400 Speaker 1: what you want to do, the technology isn't there yet. 153 00:08:47,440 --> 00:08:50,720 Speaker 1: It doesn't exist yet. We don't we we don't have 154 00:08:50,960 --> 00:08:54,720 Speaker 1: that technology to solve this far field speech problem. And 155 00:08:55,000 --> 00:08:58,960 Speaker 1: he was unflapped. He said, uh, I appreciate that, thank 156 00:08:59,000 --> 00:09:01,679 Speaker 1: you for telling me, But solve it. We are an Amazon, 157 00:09:02,000 --> 00:09:04,839 Speaker 1: We've got resources, hire as many people as you need, 158 00:09:04,880 --> 00:09:07,679 Speaker 1: to take as long as you want, but you know, 159 00:09:07,720 --> 00:09:11,360 Speaker 1: solve the problem. The team gave itself the code name Doppler, 160 00:09:11,600 --> 00:09:14,319 Speaker 1: as in the Doppler effect, which describes the way a 161 00:09:14,400 --> 00:09:18,240 Speaker 1: sound wave moves with respect to a listener. They hotly 162 00:09:18,280 --> 00:09:21,079 Speaker 1: debated everything from what to name the device, to what 163 00:09:21,120 --> 00:09:23,679 Speaker 1: it should do, to how to market that to the public. 164 00:09:24,520 --> 00:09:27,920 Speaker 1: At first, they met with Bezos once a month. He 165 00:09:27,920 --> 00:09:31,960 Speaker 1: would get deeply involved in the technology. UM, not just 166 00:09:32,160 --> 00:09:35,280 Speaker 1: the business or the product, but deeply involved in the technology. 167 00:09:35,360 --> 00:09:37,400 Speaker 1: He I would say, he stayed very close to the 168 00:09:37,440 --> 00:09:41,200 Speaker 1: project throughout. As a project progressed, those meetings would increase 169 00:09:41,440 --> 00:09:44,880 Speaker 1: in frequency UM and UH. And by the end we 170 00:09:44,880 --> 00:09:47,400 Speaker 1: were meeting with him, you know, leading up to launch, 171 00:09:47,440 --> 00:09:49,440 Speaker 1: I would say, there probably wasn't a day that went 172 00:09:49,480 --> 00:09:51,960 Speaker 1: by that we didn't have at least one meeting with 173 00:09:52,040 --> 00:09:55,000 Speaker 1: Jeff UM. And you know, sometimes there were multiple meetings 174 00:09:55,000 --> 00:09:59,880 Speaker 1: with Jeff in a single day. UM. That is both 175 00:10:00,280 --> 00:10:03,320 Speaker 1: a blessing and a curse. The blessing came in the 176 00:10:03,360 --> 00:10:06,360 Speaker 1: form of money. The team could spend whatever they needed 177 00:10:06,400 --> 00:10:09,760 Speaker 1: to break through the technical obstacles. This is a big 178 00:10:09,800 --> 00:10:13,120 Speaker 1: project that Jeff is personally invested in. It was his 179 00:10:13,240 --> 00:10:17,120 Speaker 1: idea originally. He can be very demanding. UM, he can 180 00:10:17,160 --> 00:10:20,680 Speaker 1: push teams UM. But at the same time, like being 181 00:10:20,720 --> 00:10:24,240 Speaker 1: able to go through that experience and have his brain 182 00:10:24,400 --> 00:10:31,360 Speaker 1: on your side is an immensely powerful opportunity and experience 183 00:10:31,480 --> 00:10:34,520 Speaker 1: if you let it, like if you get really frustrated 184 00:10:34,520 --> 00:10:36,200 Speaker 1: by it. And at times some of the people on 185 00:10:36,280 --> 00:10:39,600 Speaker 1: my product team would say, it feels like Jeff is 186 00:10:39,600 --> 00:10:43,600 Speaker 1: the product manager for Alexa. Bezo set the vision for Alexa. 187 00:10:44,160 --> 00:10:46,760 Speaker 1: He wanted it to be the Star Trek computer. He 188 00:10:46,840 --> 00:10:51,600 Speaker 1: pictured a versatile, conversational machine that could respond to any question. 189 00:10:52,160 --> 00:10:54,000 Speaker 1: You should be able to sit in your living room 190 00:10:54,240 --> 00:10:58,280 Speaker 1: and ask Alexa anything. It sounds simple, but it's not 191 00:10:58,800 --> 00:11:00,840 Speaker 1: because they needed a lot of data to train the 192 00:11:00,880 --> 00:11:04,760 Speaker 1: AI algorithms. For example, for Alexa to tell you the weather, 193 00:11:05,120 --> 00:11:07,559 Speaker 1: they would have to understand the phrasing of your question 194 00:11:07,800 --> 00:11:10,880 Speaker 1: and your dialect from across a noisy room, and sort 195 00:11:10,960 --> 00:11:14,240 Speaker 1: through databases for the right answer. This would take Alexa 196 00:11:14,320 --> 00:11:17,679 Speaker 1: a step further than Syrior Google Voice, which only worked 197 00:11:17,679 --> 00:11:20,880 Speaker 1: when you spoke directly into your phone. But the Alexa 198 00:11:20,960 --> 00:11:23,520 Speaker 1: team just didn't have the data to get the AI smart. 199 00:11:24,000 --> 00:11:27,959 Speaker 1: The early prototypes worked so badly that even Amazon employees 200 00:11:28,000 --> 00:11:31,280 Speaker 1: didn't really want to test them. Bassos was getting impatient. 201 00:11:31,720 --> 00:11:34,480 Speaker 1: He actually walked out of a number of internal meetings 202 00:11:34,480 --> 00:11:37,800 Speaker 1: and frustration. Then Greg and his colleague struck upon an 203 00:11:37,840 --> 00:11:43,240 Speaker 1: idea they called Project Amped. Basically, rents houses or apartments 204 00:11:43,280 --> 00:11:46,600 Speaker 1: in cities all over the US, and we would put 205 00:11:46,640 --> 00:11:51,120 Speaker 1: devices in those apartments. They were all camouflaged, and they 206 00:11:51,160 --> 00:11:54,080 Speaker 1: were not just Amazon devices. There were other companies devices 207 00:11:54,160 --> 00:11:57,120 Speaker 1: there as well, UM, some of which were visible, some 208 00:11:57,200 --> 00:12:01,080 Speaker 1: of which were not visible. And characteristic of all Amazon projects, 209 00:12:01,120 --> 00:12:05,720 Speaker 1: secrecy was the priority. Employees were careful to conceal Amazon's 210 00:12:05,720 --> 00:12:08,079 Speaker 1: identity when they set up the room, and that was 211 00:12:08,120 --> 00:12:10,400 Speaker 1: all about obfuscation, you know, we didn't want people to 212 00:12:10,480 --> 00:12:13,600 Speaker 1: understand what company we were working for. UM, and so 213 00:12:13,760 --> 00:12:15,800 Speaker 1: you could see an xbox, you could see this, you 214 00:12:15,840 --> 00:12:18,199 Speaker 1: could see that UM. And then we would have all 215 00:12:18,200 --> 00:12:21,440 Speaker 1: these Alexa devices hidden throughout the room or echo devices, 216 00:12:21,960 --> 00:12:25,680 Speaker 1: prototype devices, and then they brought in testers, thousands of 217 00:12:25,720 --> 00:12:28,640 Speaker 1: people paid an hourly wage, coming in at all hours 218 00:12:28,640 --> 00:12:30,640 Speaker 1: of the day and days of the week to train 219 00:12:30,679 --> 00:12:34,920 Speaker 1: the machines. Then we would bring in we would recruit participants, 220 00:12:35,320 --> 00:12:38,600 Speaker 1: have people with different accents, you know, male, female, different 221 00:12:38,600 --> 00:12:41,520 Speaker 1: ages who would come in and we would ask them 222 00:12:41,600 --> 00:12:45,640 Speaker 1: to do a mixture of reading scripted things and then 223 00:12:45,720 --> 00:12:48,319 Speaker 1: also talking in a much more off the cuff all 224 00:12:48,320 --> 00:12:51,600 Speaker 1: the cart fashion to ask for things that people would 225 00:12:51,640 --> 00:12:55,920 Speaker 1: be we hoped asking Alexa when we launched. Amazon conducted 226 00:12:55,920 --> 00:12:58,800 Speaker 1: the data gathering effort with such stealth that neighbors started 227 00:12:58,840 --> 00:13:02,280 Speaker 1: to get suspicious. House that we rented in Boston that 228 00:13:03,440 --> 00:13:05,400 Speaker 1: the neighbors thought that because we had a lot of 229 00:13:05,400 --> 00:13:08,000 Speaker 1: cars showing up and individuals getting out on their own 230 00:13:08,040 --> 00:13:11,240 Speaker 1: and coming in and spending I think maybe an hour, 231 00:13:11,920 --> 00:13:13,680 Speaker 1: and so there's a lot of sort of you know, 232 00:13:13,720 --> 00:13:16,679 Speaker 1: transient in and out traffic, and neighbors thought that maybe 233 00:13:16,679 --> 00:13:19,760 Speaker 1: there was you know, a drug running ring or something 234 00:13:19,800 --> 00:13:22,240 Speaker 1: else going on, and so the police actually showed up 235 00:13:22,480 --> 00:13:26,680 Speaker 1: despite the attention from police. Bezos loved the inventiveness of 236 00:13:26,679 --> 00:13:30,960 Speaker 1: the data gathering program. When we first took AMPED to him, 237 00:13:32,240 --> 00:13:36,959 Speaker 1: his response was effectively like, now you're talking, like let's 238 00:13:37,000 --> 00:13:40,120 Speaker 1: do this, um, Like tell me if you want more money, 239 00:13:40,440 --> 00:13:43,720 Speaker 1: Greg says. The project AMPED ran in thirteen cities and 240 00:13:43,760 --> 00:13:47,080 Speaker 1: included over ten people. The devices were placed all over 241 00:13:47,120 --> 00:13:49,400 Speaker 1: the room, and so we were trying to capture a 242 00:13:49,440 --> 00:13:53,960 Speaker 1: massive amount of acoustic data about, you know, how noise 243 00:13:54,040 --> 00:13:56,760 Speaker 1: performs in a room in all kinds of different rooms, 244 00:13:56,760 --> 00:14:00,160 Speaker 1: in bedrooms and living rooms, in kitchens, in bathrooms. The 245 00:14:00,200 --> 00:14:04,920 Speaker 1: result Amazon basically solved the far field voice problem. It 246 00:14:05,000 --> 00:14:07,959 Speaker 1: took six months for this company to solve a problem 247 00:14:08,000 --> 00:14:11,960 Speaker 1: that had stumped speech scientists for decades. But since the 248 00:14:12,000 --> 00:14:15,120 Speaker 1: project was top secret, they weren't able to tell anyone, 249 00:14:15,440 --> 00:14:19,640 Speaker 1: not even the speech science community, about their big technological breakthrough. 250 00:14:20,560 --> 00:14:24,119 Speaker 1: Here's Ahmed Boozid. He had been working in voice technology 251 00:14:24,200 --> 00:14:27,320 Speaker 1: for almost twenty years when Amazon tried to recruit him 252 00:14:27,400 --> 00:14:30,720 Speaker 1: in early two fifteen. I turned him down. I said, 253 00:14:30,840 --> 00:14:33,720 Speaker 1: you're not showing me anything. You have never done anything 254 00:14:33,760 --> 00:14:36,880 Speaker 1: in voice, and you're telling me that you are going 255 00:14:36,920 --> 00:14:39,640 Speaker 1: to any uly to to go from DC and go 256 00:14:39,720 --> 00:14:42,720 Speaker 1: to Seattle and they'll let you. Just trust us. I know, no, 257 00:14:42,920 --> 00:14:44,800 Speaker 1: I don't think so. I mean, I've seen the Kindle 258 00:14:44,880 --> 00:14:48,160 Speaker 1: and it's fine, it's okay, but Kindall is like, you know, 259 00:14:49,040 --> 00:14:50,800 Speaker 1: it's a toy compared to what you guys are trying 260 00:14:50,800 --> 00:14:54,000 Speaker 1: to do. So anyway, I said no twice. Eventually they 261 00:14:54,040 --> 00:14:57,480 Speaker 1: flew Ahmed out to Amazon headquarters to see Alexa in person. 262 00:14:57,880 --> 00:15:00,400 Speaker 1: So it was, you know, on the table, and so 263 00:15:00,440 --> 00:15:02,880 Speaker 1: the first thing I did is I asked it for 264 00:15:02,880 --> 00:15:04,960 Speaker 1: for music, and he plays some Miles Davis and it 265 00:15:05,080 --> 00:15:09,400 Speaker 1: did I'm like, okay, cool, And then I suppos I said, 266 00:15:09,440 --> 00:15:12,320 Speaker 1: you know what time is it? Man stopping and told 267 00:15:12,320 --> 00:15:14,280 Speaker 1: me the time, and then it continued playing Miles datas, 268 00:15:14,320 --> 00:15:17,640 Speaker 1: which was great. Interacting with Alexa was almost a moving 269 00:15:17,720 --> 00:15:21,360 Speaker 1: moment for Ahmed. He felt like decades of scientific research 270 00:15:21,400 --> 00:15:24,880 Speaker 1: had been realized in this little device, and I, you know, 271 00:15:24,960 --> 00:15:27,600 Speaker 1: I was like, Okay, this is amazing, This is amazing. 272 00:15:28,480 --> 00:15:31,080 Speaker 1: This is clearly an important moment in technology. And this 273 00:15:31,120 --> 00:15:32,720 Speaker 1: is what I've been doing all my life, right, solving 274 00:15:32,720 --> 00:15:35,360 Speaker 1: the problem problems. See, you know, we people who are 275 00:15:35,400 --> 00:15:38,680 Speaker 1: in the speech were are always saying this expression, you know, 276 00:15:38,720 --> 00:15:41,120 Speaker 1: speeches around the corner. We've been saying it's since the 277 00:15:41,120 --> 00:15:44,120 Speaker 1: mid nineties, species around the corner, because we do believe 278 00:15:44,160 --> 00:15:47,000 Speaker 1: that if you do speech well, a lot of stuff 279 00:15:47,920 --> 00:15:52,120 Speaker 1: becomes easy to do for people. Ahmed was particularly impressed 280 00:15:52,120 --> 00:15:56,240 Speaker 1: by two innovations that Amazon made. One that they solved 281 00:15:56,240 --> 00:16:00,120 Speaker 1: the far field speech issue and secondly a let so 282 00:16:00,200 --> 00:16:03,760 Speaker 1: it was fast to speed right. The fact that it's 283 00:16:03,800 --> 00:16:05,720 Speaker 1: just I was, I was just amazed, Like, how the 284 00:16:05,720 --> 00:16:09,680 Speaker 1: hell does come back within within like two seconds? I mean, 285 00:16:09,880 --> 00:16:12,200 Speaker 1: if I if I type, you know, if I want 286 00:16:12,200 --> 00:16:14,640 Speaker 1: to launch a page on my browser, or it takes 287 00:16:14,680 --> 00:16:17,840 Speaker 1: like sometimes like three four or five seconds. Right, How 288 00:16:17,920 --> 00:16:19,520 Speaker 1: is how the hell is this thing doing all of 289 00:16:19,520 --> 00:16:23,520 Speaker 1: these things, getting my voice going to the cloud, processing it, 290 00:16:24,040 --> 00:16:27,640 Speaker 1: coming back talking back within two seconds. It's like a 291 00:16:27,680 --> 00:16:30,360 Speaker 1: magic even for someone like me who was been like 292 00:16:30,440 --> 00:16:33,640 Speaker 1: you should be jaded by by by then, Right, okay, 293 00:16:33,680 --> 00:16:36,200 Speaker 1: I know exactly how it's happened. I was astonished. Right, 294 00:16:36,440 --> 00:16:39,080 Speaker 1: how the hell did they do that? We'll be right back. 295 00:16:50,080 --> 00:16:53,400 Speaker 1: I want to take a moment to address Amazon's obsessive secrecy. 296 00:16:54,040 --> 00:16:57,360 Speaker 1: Many companies are secretive, but tech companies today have the 297 00:16:57,400 --> 00:16:59,800 Speaker 1: eyes of the world on them, so they are taking 298 00:17:00,000 --> 00:17:03,600 Speaker 1: corporate secrecy to a whole other level. Like that story. 299 00:17:03,680 --> 00:17:07,679 Speaker 1: Jeff Adams, the speech scientist, told that the Amazon executives 300 00:17:07,840 --> 00:17:10,600 Speaker 1: were so secretive when the acquired a company that they 301 00:17:10,600 --> 00:17:14,160 Speaker 1: wouldn't even be seen together at a conference. That's typical 302 00:17:14,160 --> 00:17:18,040 Speaker 1: of Amazon. This intense drive for secrecy comes straight from 303 00:17:18,119 --> 00:17:21,480 Speaker 1: Jeff Bezos. He wants to tightly control the messaging around 304 00:17:21,480 --> 00:17:25,639 Speaker 1: Amazon's new products. The idea is that complete secrecy pays 305 00:17:25,720 --> 00:17:29,600 Speaker 1: off with a surprising, almost magical reveal once the product 306 00:17:29,680 --> 00:17:34,119 Speaker 1: is launched. That man keeping Alexa under wraps until launch, 307 00:17:34,560 --> 00:17:38,480 Speaker 1: preserving the details of how it actually worked, and choosing 308 00:17:38,680 --> 00:17:42,520 Speaker 1: the perfect voice here's Greg Hart, how do you define 309 00:17:42,560 --> 00:17:46,000 Speaker 1: the characteristics that you want the voice to have? And 310 00:17:46,040 --> 00:17:48,120 Speaker 1: so there was sort of a brief that we wrote 311 00:17:48,160 --> 00:17:51,639 Speaker 1: up about the qualities you wanted the person to be knowledgeable. 312 00:17:51,680 --> 00:17:54,560 Speaker 1: We very quickly early on, decided the first voice would 313 00:17:54,560 --> 00:17:57,199 Speaker 1: be female. We knew there would be additional voices, but 314 00:17:57,280 --> 00:17:59,800 Speaker 1: we felt that the first voice should be female in 315 00:18:00,160 --> 00:18:04,400 Speaker 1: because yeah, so that's the logical question why In part 316 00:18:04,480 --> 00:18:07,439 Speaker 1: because the we knew that the device would be in 317 00:18:07,560 --> 00:18:12,040 Speaker 1: the kitchen, and we felt that a female voice would 318 00:18:12,080 --> 00:18:15,680 Speaker 1: be more open and inviting and warm than a male 319 00:18:15,800 --> 00:18:18,719 Speaker 1: voice would be in that environment, and more appropriate in 320 00:18:18,760 --> 00:18:22,720 Speaker 1: that environment. Not because of any sexist things, but just 321 00:18:22,760 --> 00:18:25,480 Speaker 1: because of the fact that we knew, um that we 322 00:18:25,600 --> 00:18:28,920 Speaker 1: knew the way that, um, that's the right way to 323 00:18:28,960 --> 00:18:34,959 Speaker 1: say this. We had seen evidence that people respond differently 324 00:18:35,080 --> 00:18:39,040 Speaker 1: to male computer voices than to female computer voices, and 325 00:18:39,040 --> 00:18:42,760 Speaker 1: they respond more positively to female computer voices, and we 326 00:18:42,880 --> 00:18:45,520 Speaker 1: wanted because the device was in the home, we wanted 327 00:18:45,560 --> 00:18:48,919 Speaker 1: it to be a device that everybody would respond positively to. 328 00:18:49,400 --> 00:18:52,280 Speaker 1: Putting a female voice in the kitchen, so to speak, 329 00:18:52,440 --> 00:18:55,879 Speaker 1: would turn out to be a somewhat controversial choice, and 330 00:18:55,960 --> 00:18:59,520 Speaker 1: Amazon isn't the only company to do this. Serie. Google 331 00:18:59,600 --> 00:19:04,439 Speaker 1: Voice and Microsoft Cortana are all women by default, So 332 00:19:04,600 --> 00:19:07,640 Speaker 1: Amazon moved forward in their search for a voice. They 333 00:19:07,680 --> 00:19:10,640 Speaker 1: contracted with the same studio that had developed the voice 334 00:19:10,640 --> 00:19:13,520 Speaker 1: of Sirie for Apple. Okay, I set up your meeting 335 00:19:13,520 --> 00:19:16,480 Speaker 1: with David tomorrow. Shall I schedule it? Who is voiced 336 00:19:16,480 --> 00:19:19,960 Speaker 1: by Susan Bennett, a career voice actress. Hi, my name 337 00:19:20,000 --> 00:19:22,800 Speaker 1: is Susan Bennett, and I have a voice actor and 338 00:19:23,440 --> 00:19:27,240 Speaker 1: the original voice of Sirie. So to find their Alexa, 339 00:19:27,520 --> 00:19:30,760 Speaker 1: the studio had half a dozen female voice actresses read 340 00:19:30,800 --> 00:19:35,320 Speaker 1: for hours. They read entire books and random articles, and 341 00:19:35,400 --> 00:19:39,600 Speaker 1: finally Greg Hart and Jeff Bezos picked one woman. Her 342 00:19:39,680 --> 00:19:44,520 Speaker 1: identity hasn't been revealed in all this time. It's amazing 343 00:19:44,560 --> 00:19:47,600 Speaker 1: to me as an Amazon reporter that Amazon has still 344 00:19:47,640 --> 00:19:49,800 Speaker 1: been able to keep the voice of Alexa a secret. 345 00:19:50,200 --> 00:19:54,520 Speaker 1: In I started canvassing voice over actors, asking if they 346 00:19:54,600 --> 00:19:58,639 Speaker 1: knew the identity of Alexa. No one knew. Finally, I 347 00:19:58,680 --> 00:20:00,919 Speaker 1: got a tip from someone who had worked with the 348 00:20:00,960 --> 00:20:04,800 Speaker 1: studio that Amazon contracted with, and they said that Alexa 349 00:20:04,880 --> 00:20:07,600 Speaker 1: was voiced by an actress and singer from Boulder, Colorado 350 00:20:07,880 --> 00:20:12,080 Speaker 1: named Nina Raleigh reached out to Raleigh numerous times, she 351 00:20:12,119 --> 00:20:14,960 Speaker 1: wouldn't confirm that she's the voice of Alexa, but she 352 00:20:15,040 --> 00:20:18,879 Speaker 1: didn't deny it either. Eventually I confirmed with enough people 353 00:20:18,880 --> 00:20:23,000 Speaker 1: that I feel confident she's the one. She's Alexa. Here's 354 00:20:23,000 --> 00:20:24,960 Speaker 1: a clip from her website and add that she did 355 00:20:25,000 --> 00:20:29,000 Speaker 1: for Time Warner Cable. Thanks for choosing Time Warner. Now 356 00:20:29,000 --> 00:20:31,760 Speaker 1: that you've ordered your installation, let's set up a time 357 00:20:31,800 --> 00:20:35,760 Speaker 1: to get a crew over to your place. And here's Alexa. 358 00:20:36,240 --> 00:20:39,800 Speaker 1: Time Warner Cable, also simply known as Time Warner, was 359 00:20:39,800 --> 00:20:42,960 Speaker 1: an American cable television company. It was ranked the second 360 00:20:43,040 --> 00:20:45,720 Speaker 1: largest cable company in the United States. But imagine what 361 00:20:45,800 --> 00:20:48,720 Speaker 1: it might be like to be Nina Raleigh. Your voice 362 00:20:48,760 --> 00:20:52,080 Speaker 1: is piped into millions of homes every day, each new 363 00:20:52,119 --> 00:20:55,439 Speaker 1: iteration of Alexa, each update for Amazon, you have to 364 00:20:55,480 --> 00:20:59,040 Speaker 1: record something new, Like when Amazon Fresh released a product 365 00:20:59,040 --> 00:21:02,879 Speaker 1: called the single Burger. She had to be available single 366 00:21:02,920 --> 00:21:05,719 Speaker 1: Cow Burger, a beef burger made with meat from just 367 00:21:05,800 --> 00:21:09,560 Speaker 1: a single cow. She's chained to Alexa. It sounds like 368 00:21:09,600 --> 00:21:14,080 Speaker 1: a life of obscurity and loneliness. Compare that to Siri 369 00:21:14,240 --> 00:21:18,159 Speaker 1: and Susan Bennett. She's spoken on CNN and countless talk shows. 370 00:21:18,520 --> 00:21:20,240 Speaker 1: She's even been able to use her fame as the 371 00:21:20,320 --> 00:21:25,040 Speaker 1: voice of Sirie as leverage for many more opportunities. Greg 372 00:21:25,080 --> 00:21:28,240 Speaker 1: saw the burden that Amazon secrecy put on the human 373 00:21:28,359 --> 00:21:33,520 Speaker 1: behind the voice of Alexa. I never met her um 374 00:21:34,560 --> 00:21:37,160 Speaker 1: and I don't even know that I might have. I'm 375 00:21:37,160 --> 00:21:39,480 Speaker 1: not have spoken with her once, but I never met her. 376 00:21:40,040 --> 00:21:43,600 Speaker 1: And it would be interesting to be in her shoes now, 377 00:21:43,680 --> 00:21:46,840 Speaker 1: because it's on the one hand, it's it's incredible that 378 00:21:46,920 --> 00:21:50,040 Speaker 1: this thing that you contributed to is now so ubiquitous. 379 00:21:50,119 --> 00:21:52,480 Speaker 1: On the other hand, I would think it would be, 380 00:21:52,640 --> 00:21:56,960 Speaker 1: you know, maybe a little bit unsettling. Amazon introduced the 381 00:21:57,000 --> 00:22:01,160 Speaker 1: Echo in November two, fourteen were re strived from Amazon. 382 00:22:01,359 --> 00:22:04,080 Speaker 1: I didn't know what it was. With a YouTube video, 383 00:22:04,520 --> 00:22:12,720 Speaker 1: Alexa play rock music, rock music, alexis stop. We want 384 00:22:12,720 --> 00:22:16,000 Speaker 1: to try Alexa, what time is it? The time is three. 385 00:22:17,119 --> 00:22:19,159 Speaker 1: You actually don't have to yell at it. Okay. It 386 00:22:19,240 --> 00:22:21,760 Speaker 1: uses far Fueld technology so it can hear you from 387 00:22:21,840 --> 00:22:24,880 Speaker 1: anywhere in the room, So I can just hear you anywhere. Yes, 388 00:22:25,600 --> 00:22:29,280 Speaker 1: that promise was enough. A responsive computer that can tell 389 00:22:29,320 --> 00:22:32,840 Speaker 1: the time or the weather, play music, and answer questions 390 00:22:32,840 --> 00:22:36,000 Speaker 1: from across the room. Tens of thousands of people joined 391 00:22:36,000 --> 00:22:39,159 Speaker 1: a waiting list to receive a device. Here's Bezos at 392 00:22:39,200 --> 00:22:42,800 Speaker 1: the Recode Technology conference in two thousand sixteen saying that 393 00:22:42,840 --> 00:22:47,000 Speaker 1: the future is Alexa. He's speaking to Walt Mossburg. But 394 00:22:47,080 --> 00:22:49,879 Speaker 1: it has been a dream ever since, you know, people started, 395 00:22:49,920 --> 00:22:52,080 Speaker 1: you know, in the early days of science fiction to 396 00:22:52,160 --> 00:22:55,280 Speaker 1: have a computer that you can talk to. So are 397 00:22:55,320 --> 00:22:58,520 Speaker 1: you deeply committed to this becoming a huge part of 398 00:22:58,560 --> 00:23:02,119 Speaker 1: your business and what you Absolutely We've been working on it. 399 00:23:02,200 --> 00:23:04,320 Speaker 1: You know, we worked on it. We have more than 400 00:23:04,359 --> 00:23:09,160 Speaker 1: a thousand people dedicated us to Alexa in the Echo ecosystem, 401 00:23:09,240 --> 00:23:12,320 Speaker 1: and it's a and there's so much more to come. 402 00:23:12,760 --> 00:23:15,960 Speaker 1: Baz also is deploying his playbook. For experiments that produced 403 00:23:16,000 --> 00:23:20,399 Speaker 1: promising sparks, he poured gasoline on them. Amazon ramped up 404 00:23:20,480 --> 00:23:24,600 Speaker 1: hiring the Alexa team, balloon to ten thou employees, and 405 00:23:24,680 --> 00:23:28,480 Speaker 1: Bezos paider on ten million dollars for the company's first 406 00:23:28,600 --> 00:23:32,280 Speaker 1: ever Super Bowl at starring Alec Baldwin and Missy Elliott. 407 00:23:33,920 --> 00:23:38,560 Speaker 1: Alex to stop. How you do that? It's my Amazon 408 00:23:38,680 --> 00:23:44,119 Speaker 1: Echo like extreme music, order things and watch this. I 409 00:23:44,240 --> 00:23:50,719 Speaker 1: like to turn on the lights. Wow. Inside the company, 410 00:23:50,880 --> 00:23:53,560 Speaker 1: employee has noticed that he seemed to take real joy 411 00:23:53,680 --> 00:23:57,280 Speaker 1: in the work here's Ahmed Jeff, you know, the CEO 412 00:23:57,600 --> 00:24:01,920 Speaker 1: and the guy in charge, uh was I would say 413 00:24:01,960 --> 00:24:04,480 Speaker 1: he was obsessed with Alexa. How many he had this 414 00:24:04,600 --> 00:24:07,320 Speaker 1: saying that I am happiest. He used to say this, 415 00:24:07,440 --> 00:24:10,720 Speaker 1: I'm happy is when I'm working on Alexa UM and 416 00:24:10,760 --> 00:24:13,159 Speaker 1: so he you know, it was his favorite time of 417 00:24:13,200 --> 00:24:15,399 Speaker 1: the day. Used to meet with ALEX a team and 418 00:24:15,440 --> 00:24:18,720 Speaker 1: work and work on ALEX. So he was directly involved 419 00:24:18,960 --> 00:24:23,359 Speaker 1: in the early days. Basis was frequently asked about privacy. 420 00:24:23,520 --> 00:24:26,600 Speaker 1: At the two thousand sixteen Code Conference, he promised to 421 00:24:26,680 --> 00:24:30,399 Speaker 1: be a good steward of the sensitive personal data that 422 00:24:30,440 --> 00:24:33,080 Speaker 1: Alexa was sure to pick up. This is going to 423 00:24:33,160 --> 00:24:36,399 Speaker 1: get much deeper into our lives. So so what are 424 00:24:36,440 --> 00:24:38,560 Speaker 1: what are the privacy I think that if you take 425 00:24:38,600 --> 00:24:44,000 Speaker 1: the totality of you know, privacy UM and our ability 426 00:24:44,040 --> 00:24:46,879 Speaker 1: to store large amounts of information to use it in 427 00:24:46,920 --> 00:24:49,960 Speaker 1: ways that customers actually do want us to use it. 428 00:24:50,080 --> 00:24:52,560 Speaker 1: So there are benefits. And I think one of the 429 00:24:52,560 --> 00:24:54,719 Speaker 1: things that you have to do is when you collect 430 00:24:54,720 --> 00:24:57,240 Speaker 1: and store data, you have to be clear about what 431 00:24:57,280 --> 00:25:00,359 Speaker 1: you're doing. You have to and not just you know, 432 00:25:00,920 --> 00:25:05,000 Speaker 1: subsection seventeen, paragraph three clearly as you can see at 433 00:25:05,000 --> 00:25:07,720 Speaker 1: our privacy policy we were allowed to do that. Would 434 00:25:07,720 --> 00:25:10,480 Speaker 1: you have to figure out ways to be kind of 435 00:25:10,520 --> 00:25:14,560 Speaker 1: obviously clear? He stood on stage and promised that the 436 00:25:14,600 --> 00:25:19,040 Speaker 1: privacy policy would be obviously clear. What he wasn't saying 437 00:25:19,520 --> 00:25:22,120 Speaker 1: was that Alexa was so smart in part because there 438 00:25:22,119 --> 00:25:25,960 Speaker 1: were real life humans listening through the machines, helping to 439 00:25:26,080 --> 00:25:29,399 Speaker 1: craft many of Alexa's answers and to fix its errors 440 00:25:29,440 --> 00:25:32,040 Speaker 1: by listening to what a subset of users said. The 441 00:25:32,160 --> 00:25:35,440 Speaker 1: company kept Alexa owners in the dark about how this 442 00:25:35,600 --> 00:25:39,520 Speaker 1: aspect of their devices worked. Ruthie Hope Slattess was one 443 00:25:39,560 --> 00:25:42,760 Speaker 1: of those people being paid to listen. Several years ago, 444 00:25:43,080 --> 00:25:45,840 Speaker 1: she saw an ad from a temp agency seeking someone 445 00:25:45,880 --> 00:25:49,280 Speaker 1: with an English or journalism degree. They offered twelve dollars 446 00:25:49,280 --> 00:25:52,920 Speaker 1: an hour to transcribe audio recordings. There was a kind 447 00:25:52,920 --> 00:25:55,760 Speaker 1: of a generic add on prex list. I don't think 448 00:25:55,840 --> 00:25:59,960 Speaker 1: it mentioned Amazon at all. She applied, She passed US 449 00:26:00,080 --> 00:26:03,480 Speaker 1: curity check and a grammar test, and then she was 450 00:26:03,560 --> 00:26:06,720 Speaker 1: let in on her task. She would listen to conversations 451 00:26:06,800 --> 00:26:09,600 Speaker 1: picked up by the echoes microphones and type them up 452 00:26:09,800 --> 00:26:13,320 Speaker 1: and then feed the information back in the Amazon system. 453 00:26:13,359 --> 00:26:15,359 Speaker 1: She said it seemed like a good job. At first, 454 00:26:15,680 --> 00:26:19,280 Speaker 1: we heard the customer using it in you know, ordering 455 00:26:19,320 --> 00:26:22,080 Speaker 1: flour and asking what time it was and asking to 456 00:26:22,119 --> 00:26:24,879 Speaker 1: be told to joke and so forth, and it seemed 457 00:26:25,040 --> 00:26:28,080 Speaker 1: pretty cool. I mean a little invasive, for sure, but 458 00:26:28,080 --> 00:26:32,080 Speaker 1: but pretty cool overall. Ruthie was there to make Alexa's 459 00:26:32,119 --> 00:26:37,040 Speaker 1: responses seem more nuanced, more intuitive, more human. This was 460 00:26:37,119 --> 00:26:39,600 Speaker 1: right around the time Amazon started selling the Echo to 461 00:26:39,720 --> 00:26:43,960 Speaker 1: customers on a limited basis. She assumed at first that 462 00:26:44,040 --> 00:26:46,639 Speaker 1: the folks she was listening to had signed up to 463 00:26:46,680 --> 00:26:51,200 Speaker 1: help improve Amazon speech recognition software, and knew that someone 464 00:26:51,400 --> 00:26:55,080 Speaker 1: might be listening. Who were all of these all of 465 00:26:55,119 --> 00:26:58,560 Speaker 1: these many many voices and people were they? Various people 466 00:26:58,640 --> 00:27:02,920 Speaker 1: work for Amazon who agreed to bring this home where 467 00:27:02,960 --> 00:27:07,080 Speaker 1: they you know? But then if they were, would they 468 00:27:07,160 --> 00:27:11,440 Speaker 1: really be ordering sex toys, you know, um and talking 469 00:27:11,480 --> 00:27:14,159 Speaker 1: dirty to it and all of the grotesque things that 470 00:27:14,200 --> 00:27:18,240 Speaker 1: we occasionally heard. Almost immediately, it was clear to Ruthie 471 00:27:18,240 --> 00:27:21,000 Speaker 1: that people liked a toy with Alexa in ways that 472 00:27:21,080 --> 00:27:24,520 Speaker 1: Amazon maybe didn't intend. What can we ask her? How 473 00:27:24,560 --> 00:27:27,080 Speaker 1: how will she respond if we abuse her, if we 474 00:27:27,160 --> 00:27:30,480 Speaker 1: talked dirty to her, if we asked her to marry us? 475 00:27:30,560 --> 00:27:33,399 Speaker 1: If we you know, that sort of thing. Most of 476 00:27:33,440 --> 00:27:40,080 Speaker 1: the sexual stuff seemed innocuous enough. Percent of it wasn't disturbing. 477 00:27:40,280 --> 00:27:43,600 Speaker 1: You know, it was sometimes humorous and sometimes you know, 478 00:27:43,720 --> 00:27:48,159 Speaker 1: kind of silly or weird or entertaining even you know, 479 00:27:48,240 --> 00:27:50,439 Speaker 1: like somebody trying to figure out what kind of dildo 480 00:27:50,520 --> 00:27:53,520 Speaker 1: toward her or something like that, you know, but sometimes 481 00:27:53,720 --> 00:27:57,600 Speaker 1: it did get disturbing if it was a man who 482 00:27:57,920 --> 00:28:03,320 Speaker 1: was talking to her, like he would talk in an 483 00:28:03,359 --> 00:28:06,320 Speaker 1: abusive way to a woman, Like you could hear it 484 00:28:06,359 --> 00:28:09,120 Speaker 1: in his voice, And she wondered if men were being 485 00:28:09,160 --> 00:28:13,040 Speaker 1: so rude in part because Alexa was female, and being 486 00:28:13,080 --> 00:28:16,440 Speaker 1: a woman, even an ai woman, meant that Alexa would 487 00:28:16,440 --> 00:28:20,320 Speaker 1: be on the receiving end of misogyny. Even children did this. 488 00:28:20,720 --> 00:28:24,760 Speaker 1: There were a lot of kids who would talk abusively 489 00:28:24,840 --> 00:28:28,760 Speaker 1: to her, and it felt as though they were exercising 490 00:28:29,240 --> 00:28:32,240 Speaker 1: some sort of anger that they had towards their parents 491 00:28:32,320 --> 00:28:35,679 Speaker 1: or their teacher or something like that. There was a 492 00:28:35,800 --> 00:28:40,080 Speaker 1: lot of psychology that I thought was fascinating. Ruthie felt 493 00:28:40,120 --> 00:28:42,640 Speaker 1: like the way people spoke to Alexa was so private, 494 00:28:42,920 --> 00:28:46,200 Speaker 1: so intimate. They would never ever speak like this if 495 00:28:46,200 --> 00:28:49,160 Speaker 1: they knew someone was listening. There in their private home. 496 00:28:49,360 --> 00:28:51,760 Speaker 1: They think that no one will ever hear the words 497 00:28:51,800 --> 00:28:54,400 Speaker 1: that are coming out of their mouth. Ever, they do 498 00:28:54,720 --> 00:28:57,640 Speaker 1: know that they're talking to a robot, but at the 499 00:28:57,680 --> 00:29:01,880 Speaker 1: same time they're speaking to a row that as those 500 00:29:02,000 --> 00:29:06,080 Speaker 1: essentient being. Ruthie's job was a secret. It wouldn't become 501 00:29:06,120 --> 00:29:09,600 Speaker 1: public until years later that people like her were helping 502 00:29:09,640 --> 00:29:14,120 Speaker 1: to improve the software by analyzing real Alexa recordings. My 503 00:29:14,280 --> 00:29:18,160 Speaker 1: colleagues first interviewed Ruthie in two thousand nineteen, and after 504 00:29:18,200 --> 00:29:21,120 Speaker 1: we published a story describing what she and her colleagues did, 505 00:29:21,520 --> 00:29:26,720 Speaker 1: Amazon acknowledged the listening program. Technically, Alexa's terms of use 506 00:29:26,800 --> 00:29:30,200 Speaker 1: gave Amazon wide latitude to duke basically whatever it wanted 507 00:29:30,200 --> 00:29:33,040 Speaker 1: with the recordings, as long as the company was using 508 00:29:33,080 --> 00:29:37,240 Speaker 1: that audio to improve the software. But many customers fell duped. 509 00:29:37,520 --> 00:29:39,720 Speaker 1: They thought that when they spoke to their Alexa device, 510 00:29:39,960 --> 00:29:43,680 Speaker 1: they were only dealing with clever software, and Bezos had 511 00:29:43,720 --> 00:29:46,920 Speaker 1: gone back on his word his precise promise to be 512 00:29:47,040 --> 00:29:50,440 Speaker 1: transparent with how he used as customers data. It seemed 513 00:29:50,440 --> 00:29:54,720 Speaker 1: that while he highly prioritized Amazon's corporate privacy it's secrecy, 514 00:29:55,040 --> 00:29:57,959 Speaker 1: he was being careless with the privacy of his customers, 515 00:29:58,760 --> 00:30:01,680 Speaker 1: and in two thousand eight teen, Alexa had a huge 516 00:30:01,720 --> 00:30:05,960 Speaker 1: privacy mishap. Here's my colleague Prea and Nod describing the 517 00:30:06,000 --> 00:30:10,640 Speaker 1: infamous incident. What happened there is a family in Portland's 518 00:30:10,640 --> 00:30:14,880 Speaker 1: said that their echo randomly sent recorded conversations to a 519 00:30:15,000 --> 00:30:18,120 Speaker 1: contact of their's, a contact of the man in the houses. 520 00:30:18,680 --> 00:30:21,200 Speaker 1: And what happened at the time was Amazon said this 521 00:30:21,240 --> 00:30:25,400 Speaker 1: is how it worked. They said that Alexa interpreted something 522 00:30:26,000 --> 00:30:29,240 Speaker 1: those people were saying in their background conversation as Alexa, 523 00:30:29,360 --> 00:30:32,440 Speaker 1: so it woke up. Then it misunderstood a different phrase 524 00:30:32,480 --> 00:30:36,320 Speaker 1: in their background conversation as send that message. It understood 525 00:30:36,320 --> 00:30:39,800 Speaker 1: a different phrase in their background conversation. After that, the 526 00:30:39,840 --> 00:30:43,520 Speaker 1: device asked to whom and thought that they responded with 527 00:30:43,560 --> 00:30:47,040 Speaker 1: a name, So it ended up sending these conversations to 528 00:30:47,360 --> 00:30:49,920 Speaker 1: one of their contacts through this crazy chain of events. 529 00:30:49,960 --> 00:30:53,120 Speaker 1: It's like a horror movie. After the couple's friend reached 530 00:30:53,160 --> 00:30:55,479 Speaker 1: out to them and said, hey, I think you're Alexa 531 00:30:55,520 --> 00:30:58,320 Speaker 1: has been hacked, it turned out they weren't hacked. This 532 00:30:58,400 --> 00:31:01,800 Speaker 1: was Alexa mishearing. But at the end of the day, 533 00:31:01,840 --> 00:31:05,000 Speaker 1: this kind of scenario where Alexa could misinterpret a series 534 00:31:05,000 --> 00:31:08,400 Speaker 1: of things people were saying in their own home, interpret 535 00:31:08,440 --> 00:31:12,440 Speaker 1: them as requests, and send these people's recordings to someone else. 536 00:31:12,960 --> 00:31:17,120 Speaker 1: If this kind of unlikely scenario can happen, and did happen, 537 00:31:17,680 --> 00:31:20,800 Speaker 1: then who's to say it can't happen again? What else 538 00:31:20,800 --> 00:31:25,400 Speaker 1: can Alexa mix up? Even so, people were undeterred. Over 539 00:31:25,440 --> 00:31:28,560 Speaker 1: the next few years, Amazon would sell tens of millions 540 00:31:28,560 --> 00:31:32,560 Speaker 1: of Alexa devices and inspire imitators from Google and Apple. 541 00:31:33,080 --> 00:31:37,080 Speaker 1: It's popularity cemented bezos notion of himself as an inventor 542 00:31:37,480 --> 00:31:40,760 Speaker 1: and the general public's perception of Amazon as an innovator. 543 00:31:41,480 --> 00:31:44,320 Speaker 1: But Alexa hasn't quite met the goals its creators had 544 00:31:44,360 --> 00:31:47,560 Speaker 1: for it. Most people only use it as a kitchen timer, 545 00:31:47,880 --> 00:31:50,719 Speaker 1: a music player, a source for the occasional weather report. 546 00:31:51,280 --> 00:31:55,400 Speaker 1: It's never been conversational in the way Bezos imagined no 547 00:31:55,440 --> 00:31:58,120 Speaker 1: one would call at the Star Trek computer, but that 548 00:31:58,240 --> 00:32:02,720 Speaker 1: hasn't stopped consumers and estimated one in three US households 549 00:32:02,880 --> 00:32:06,480 Speaker 1: currently have a smart speaker. In the years since Alexa launched, 550 00:32:06,680 --> 00:32:10,040 Speaker 1: people have invited more items that integrate AI assistance into 551 00:32:10,040 --> 00:32:14,600 Speaker 1: their homes. That includes stuff like nest thermostats, ring cameras, 552 00:32:14,640 --> 00:32:18,760 Speaker 1: and Sono speakers. Whenever a story about an Alexa privacy 553 00:32:18,800 --> 00:32:22,320 Speaker 1: breach gets published, it makes a big splash, but it 554 00:32:22,400 --> 00:32:26,320 Speaker 1: hardly ever changes anything. Because Alexa marked a turning point, 555 00:32:26,840 --> 00:32:30,520 Speaker 1: It was another doorway into a more surveiled world and 556 00:32:30,600 --> 00:32:34,520 Speaker 1: placed Amazon alongside other big tech companies like Google and 557 00:32:34,560 --> 00:32:40,440 Speaker 1: Facebook at the hot controversial center of the battle over privacy. Finally, 558 00:32:40,760 --> 00:32:43,960 Speaker 1: Alexa turned to Amazon into a company its customers interacted 559 00:32:44,000 --> 00:32:47,400 Speaker 1: with almost every single day, not just once a week 560 00:32:47,640 --> 00:32:49,440 Speaker 1: or a few times a month when they wanted to 561 00:32:49,480 --> 00:32:53,000 Speaker 1: buy something online. And there was another way Bezos would 562 00:32:53,000 --> 00:32:56,600 Speaker 1: accomplish the same goal by producing his own TV shows 563 00:32:56,640 --> 00:33:00,840 Speaker 1: and movies. Bezos is risky expansion into hollyw would that's 564 00:33:00,920 --> 00:33:14,040 Speaker 1: next time on Foundering The Amazon story. Foundering is hosted 565 00:33:14,080 --> 00:33:17,880 Speaker 1: by me brad Stone. Sean wen Is. Our executive producer. 566 00:33:18,360 --> 00:33:21,920 Speaker 1: Pria Nad and Matt Day contributed reporting to this episode. 567 00:33:22,360 --> 00:33:26,680 Speaker 1: Raymondo is our audio engineer, Molly Nugent as our associate producer, 568 00:33:27,280 --> 00:33:30,680 Speaker 1: Mark Million and Manner May Robin a Jello and Molly 569 00:33:30,760 --> 00:33:34,320 Speaker 1: shoots our story editors. Francesca Levi is the head of 570 00:33:34,320 --> 00:33:37,760 Speaker 1: Bloomberg Podcast. Be sure to subscribe and if you like 571 00:33:37,840 --> 00:33:41,680 Speaker 1: our show, leave a review. Most importantly, tell your friends 572 00:33:42,000 --> 00:33:42,880 Speaker 1: see you next time.