The Purple Eye
Events like last week’s Amazon Web Services 2018 Developer Day are an opportunity to witness the knowledge economy’s most vital dance: when sleek corporate power courts the slovenly engineer class. I arrived early and,as I waited in the foyer of the Melbourne Conference and Exhibition Centre, I thought how much it was like watching two river systems meet: engineers streamed in, wearing t-shirts with javascript code or UNIX commands and blinking astigmatically in the natural light to meet kiosks that sported the over-caffeinated, heavy branding of Amazon Web Services (AWS), New Relic and the National Australia Bank. This immense foyer has one wall where passages lead to giant lecture theatres and conference halls while the other is a steel girder ribcage looking onto a bank of the Yarra. It was a catered affair and tuxedoed catering staff carried trays of biscuits and danishes for breakfast threaded through the kiosks discretely.
Stereotypes are cheap but sometimes too illuminating to ignore. In the room was a deficit of social confidence. No-one networked. They clustered to those familiar like shy children or raised the protective barrier of a laptop or stared down at their phones. I watched one engineer in a C:\DOS\RUN RUN\DOS\RUN t-shirt and gortex jacket, his lanyard untidily threaded through his collar, approach a AWS DeepLens kiosk. The AWS representative wearing a black shirt that said only “GO BUILD” smiled brightly. The dev asked a question and, only when it was clear they were shared a common language, he relaxed.
Once coffee and tea had been drunk and danishes consumed the tempo of the hubbub rose. The foyer was full now. A bell went off. Apart from one hi-viz jacketed security guard all the ushers were women in black tights and black t-shirts. They all had ear-pieces and those with long hair had it tied back. At the bell they moved like dark arrows, raising index fingers to ears to round up the developers and shunt them into an immense lecture theatre where the keynote speaker was about to begin.
Inside I managed to secure a seat 2 rows from the front. A tall screen had DEVDAY Australia projected onto the back of it. I looked back to see a sea of faces all trained ahead. According to the Melbourne Conference & Exhibition Centre, this room, the main Plenary room, can seat 5,564 and there were not many empty seats.
Soon the lights went down and a hush fell. There was a thumping bass and a screen came alive with a night sky. The AWS Logo appeared and moved towards the audience. Suddenly there was a computer generated cityscape - like from a video games a decade ago. As the camera raced through the streets of this city, statistics about how many services AWS had, how many clients, then glowing testimonials from major clients. Finally the words GO BUILD came toward the screen - rising up to become 3D, then the screen went off and the lights turned on to thumping bass a disembodied voice said “Welcome Head of Emerging Technologies for Asia-Pacific Olivier Klein”. On stage came a balding slim, European looking man in a suit, the thumping bass kept playing then soon receded. “Hello hello” he said in a slight Swiss accent. The message, arriving with fat wads of cash and slick production values, was clear: YOU the engineer are a rockstar. YOUR code is now cool.
Klein paced around stage, headset on, extolling the benefits of microservice architecture, decoupling of nodes, polyglot lambda functions. It was like a mega-church, he was preaching to the converted, in the languages they understood best. He was soon joined onstage by Jeff Barr, “VP & Chief Evangelist, Amazon Web Services"Jeff Barr had a smiling benevolent face of a family man that might give sentimental speeches at Christmas after too much eggnog. Except Jeff Barr was hip because Jeff Bar had purple hair. They sat awkwardly on two high chairs and faced one another. Klein punted Barr some puff questions beginning with "What’s up with that purple hair there?” Barr told a story about incentivising his developers. If they made a release by a date, he would dye his hair purple. “I love that interesting culture. That’s how you push for a delivery date!” Klein said laughing.
It was just before lunch in a breakout room with a talk on Machine learning that things began to take a more ominous course. This talk was held in a smaller room and was packed with some devs standing in the aisles. Machine Learning was the ability to use complicated algorithms to ‘teach’ a machine to notice patterns. The presenter was a tall Australian woman whose name seems to have vanished from the record. She had a broad accent and an engineer’s hunch. She was being helped out by Michael a smiling but silent assistant that spent most of his time sitting in the front row. She began explaining Amazon Rekognition. With Rekognition you could locate a individual’s face within a stream of video. On a giant screen, she used the example of AWS CEO Andy Jassy walking on stage. With a few lines of code she was able to locate the times in the video when Andy Jassy was visible. It wasn’t perfect - there was a section where he was in the shadows, invisible to the computer, but the potential was there.
Next was another service called Amazon Comprehend. She used an example of two reviews for a pair shoes - tall purple pumps. “You can use this to monitor negative sentiment,” she explained. It was a strangely authoritarian way of putting it. A few lines of code here showed the computer was 78% certain the review was negative. The margin of error was because there was some positive sentiment within the review, which began by the reviewer explaining how excited he or she was to open the box.
She then described Amazon Transcribe - the use-case given was to transcribe audio to text for television captions. At Amazon there is a heavy emphasis on chaining services together - so with Amazon Transcribe you could record the text from the audio of a video feed, then with Comprehend you could monitor it for sentiment. Finally with Rekognition you could attach the text and sentiment to a particular person if you had their features stored on file. It was a toolkit that would have a modern authoritarian state salivating. And it was all runable on a modern laptop with an internet connection, a microphone and camera.
As a Machine Learning Luddite there was a lot to take in. Historically, anonymity and privacy had relied on a wide gap between what was computer readable and the chaos stream of real-life. Machine learning hoped to bridge that gap. This meant that the awesome power that computers had to index, search and sort could be applied to raw video and audio feed. If I set up the requisite surveillance items - with cameras and microphones (hidden if I chose) I could be alerted when a particular individual said something negative. I could register who they were talking to and even who or what they were talking about. This information could also be stored in perpetuity and manipulated at will by me whoever I was - an over-vigilant employer, an unscrupulous business-person or a jilted spouse.
I noticed that a strange silence had fallen over the crowd. The presenter’s voice continued to boom out. It was difficult to work out whether it was the uneasy quiet of citizens and workers wondering how this would impact their workplace and public spaces, or the awestruck silence of engineers impressed by an immense achievement and mushrooming potential. I looked around. Some faces were underlit by laptop screens, their expressions were inscrutable.
I struggled to think of legitimate use cases but could only think of the money governments had poured into surveillance systems like the NSA’s Prism that Edward Snowden had exposed. Now these developers had the same processing power at their disposal with a few lines of code.
The presentation moved on and the presenter began talking about Amazon SageMaker. Comprehend, Rekognition and Transcribe were niche services, while SageMaker allowed access to the algorithms that underpinned these application. With it you could teach a computer to recognise a pattern. In this demonstration the presenter built a program that could determine if workers were wearing a hardhat on a worksite. The presenter explained that she had to take 5000 photos to begin to get accurate readings. This was considered a small set. These were separated into 4 different categories - a person wearing a hardhat designated ‘compliant’, a person without a hardhat designated ‘not compliant’, a person where the head could not be seen and a shot without a person. These last two were denoted as ‘unsure’. “I took photos around the AWS office” on the screen came a bunch of people in an office, some with hardhats some without. “We had a lot of fun”, some striking ridiculous poses. A titter rippled through the audience.
The presenter then went through how she had coded this. The 5000 photos were converted to a particular format then uploaded to an Amazon storage bucket. The very complex and intelligent machine learning algorithms required could be simply imported via an AWS library. The library would point to where the images were stored, given some parameters and run. Because Amazon has such colossal server farms you can lease the run-time required to run these algorithms. It’s relatively cheap and means individuals have near state level processing power at their disposal. It took a few seconds to run the algorithm.
“Ok Michael get up here in front of the Deep Lens Camera” She gestured to a camera on a tripod I had thought was for recording the talk. Michael stepped onstage and collected a white hardhat from a table on the side of the stage. The feed from the camera was projected onto the screen, massive. Text at the top of the screen said “Not compliant” in red. When Michael put on the hardhat the text changed to “Compliant”, then he took it off again. “Not compliant” appeared again. The crowd applauded enthusiastically. Michael beamed.
Back outside engineers were talking excitedly as they lined up to get wraps and chicken salad that had been doled out on trays for lunch. I felt a chill. That giant “Not Compliant” seemed like something from Robocop. It was trivial to log and alert any instance detected on non-compliance. What other rules could be enforced in this manner? Dress code violations, certain people talking with other people? I recalled the novel The Circle by David Eggers which describe how a Facebook style social media giant called The Circle uses images and video to rise above a democratic state. Democracy is undermined when people constantly monitor each other and any dissent is crushed with the exposure of the dissenter’s real or doctored past.
Even though the dystopia described by Eggers is a little exaggerated, this tool seemed like the first step toward a darker destination. It also seemed typical for the docile engineer caste to be wowed by the technological achievements without broadly considering the social implications. To be nerdy was to have narrow interests, tunnel vision, not understand how things fit in. Yet the bulk of the responsibility should rest with Amazon, like other tech giants, which never seemed to consider whether something should be invented or unleashed onto the general public. In the boardroom the justification was probably reduced to the fact that there existed a market. Marketing could use the most palatable and noble use-case (they were already sprouting the ability to prevent sex-trafficking) even if it had a thousand other darker use cases existed. Amazon could then hide behind it’s Terms and Conditions without ensuring they were being abided by. And with that a surveillance engine that once would have cost millions and many operatives would be available on an individual’s budget, with a few lines of code. And all the purple hair dye in the world could not obscure this fact.