WEBVTT - What is Project Maven? 0:00:04.160 --> 0:00:07.160 Get in touch with technology with tech Stuff from how 0:00:07.240 --> 0:00:13.600 stuff works dot com. Hey there, and welcome to tech Stuff. 0:00:13.640 --> 0:00:16.720 I'm your host, Jonathan Strickland. I'm an executive producer at 0:00:16.720 --> 0:00:19.880 how stuff Works and I love all things tech. Although 0:00:20.040 --> 0:00:24.560 this particular topic maybe a little less than usual because 0:00:24.600 --> 0:00:29.080 it gets pretty harry. Seeing May two thousand eighteen, News 0:00:29.120 --> 0:00:32.440 broke that more than a dozen Google employees had handed 0:00:32.440 --> 0:00:36.320 in their resignations over the company's involvement in a program 0:00:36.360 --> 0:00:40.599 called Project Maven. So what the heck is Project Maven 0:00:40.720 --> 0:00:44.440 and why did those employees leave? And why have an 0:00:44.560 --> 0:00:48.560 estimated four thousand Google employees put their names on petitions 0:00:48.600 --> 0:00:51.960 to end the company's involvement with the project. It's time 0:00:52.000 --> 0:00:58.520 to dive into a really serious current topic. Project Maven 0:00:58.840 --> 0:01:03.080 is a large technology project overseen by the United States 0:01:03.200 --> 0:01:08.320 Defense Department with a specific focus on bringing artificial intelligence 0:01:08.480 --> 0:01:14.600 or AI applications into military functions and campaigns. The argument 0:01:14.600 --> 0:01:17.640 for Project may even was that AI as a field 0:01:17.959 --> 0:01:22.479 has been advancing for years, with particularly impressive advancements made 0:01:22.480 --> 0:01:25.319 in the last couple of years alone, and yet the 0:01:25.319 --> 0:01:29.600 military has lagged behind. As Air Force Lieutenant General Jack 0:01:29.640 --> 0:01:34.280 sent Shanahan wrote in the Bulletin of the Atomic Scientists 0:01:34.280 --> 0:01:39.959 back in November. Quote. The US military still performs many 0:01:40.000 --> 0:01:42.760 activities in a style that would be familiar to the 0:01:42.800 --> 0:01:46.360 military of World War Two end quote. The argument was 0:01:46.400 --> 0:01:49.520 that something needed to change to bring these processes into 0:01:49.520 --> 0:01:53.080 the twenty first century. To that effect, on April two 0:01:53.120 --> 0:01:56.880 thousand seventeen, Robert Work, who then was the Deputy Secretary 0:01:56.920 --> 0:01:59.400 of the Defense Department, released a memo calling for the 0:01:59.440 --> 0:02:03.080 establishment of a new team called Project Maven. Well that's 0:02:03.120 --> 0:02:05.240 the nickname. The other name for the team was the 0:02:05.280 --> 0:02:10.120 Algorithmic Warfare Cross Functional Team. The memos opening paragraph says this, 0:02:10.320 --> 0:02:14.240 and I quote, as numerous studies have made clear, the 0:02:14.280 --> 0:02:18.440 Department of Defense must integrate artificial intelligence and machine learning 0:02:18.440 --> 0:02:23.520 more effectively across operations to maintain advantages over increasingly capable 0:02:23.560 --> 0:02:27.640 adversaries and competitors. Although we have taken tentative steps to 0:02:27.760 --> 0:02:32.080 explore the potential of artificial intelligence, big data, and deep learning, 0:02:32.400 --> 0:02:34.760 I remain convinced that we need to do much more 0:02:34.919 --> 0:02:38.600 and move much faster across d O D that's Department 0:02:38.639 --> 0:02:42.600 of Defense to take advantage of recent and future advances 0:02:42.639 --> 0:02:47.359 in these critical areas end quote. On its face, this 0:02:47.440 --> 0:02:51.720 sounds pretty reasonable, or at least understandable. After all, AI 0:02:51.840 --> 0:02:55.720 has the potential to do enormous good or harm depending 0:02:55.800 --> 0:03:00.160 upon its design and implementation. To not pursue AI in 0:03:00.200 --> 0:03:03.280 the realm of military applications seems like it would be 0:03:03.400 --> 0:03:07.960 a bad idea. Other nations and militaries are certainly exploring 0:03:08.000 --> 0:03:11.720 such options, and the landscape of warfare continues to change 0:03:11.760 --> 0:03:15.720 and become more complex. Artificial or augmented intelligence would be 0:03:15.720 --> 0:03:18.160 really handy in such a world. Being able to work 0:03:18.200 --> 0:03:23.000 with sophisticated programs to identify targets, gather intelligence, and form 0:03:23.080 --> 0:03:27.079 strategies could potentially win a conflict, save lives, or it 0:03:27.160 --> 0:03:29.920 might even allow for a non violent method to resolve 0:03:29.919 --> 0:03:33.200 a situation, which in my mind, tends to be the 0:03:33.200 --> 0:03:36.800 best of all options. The memos third paragraph details what 0:03:36.920 --> 0:03:40.080 this group's focus will be, and I quote yet again, 0:03:40.480 --> 0:03:44.200 the a w c f t S first task is 0:03:44.240 --> 0:03:49.520 to field technology to augment or automate processing, exploitation and 0:03:49.600 --> 0:03:54.680 dissemination p e D for tactical Unmanned aerial system u 0:03:54.720 --> 0:03:58.680 a S and mid altitude full motion video fm V 0:03:59.320 --> 0:04:03.240 in support of the Defeat ISIS campaign. This will help 0:04:03.280 --> 0:04:07.320 to reduce the human factors burden of FMV analysis, increase 0:04:07.400 --> 0:04:11.960 actionable intelligence, and enhance military decision making. A w c 0:04:12.120 --> 0:04:17.159 FT will number one organize a data labeling effort and develop, acquire, 0:04:17.320 --> 0:04:21.400 and or modify algorithms to accomplish key tasks. Number two 0:04:21.680 --> 0:04:25.920 identify required computational resources and identify a path to fielding 0:04:25.960 --> 0:04:30.600 that infrastructure. And number three integrate algorithmic based technology with 0:04:30.640 --> 0:04:34.280 programs of record in nine D day sprints. Now, remember 0:04:34.279 --> 0:04:38.440 a sprint when we talked about agile frameworks is a 0:04:38.960 --> 0:04:42.440 essentially a period of time in which a project takes 0:04:42.480 --> 0:04:46.080 place and you're expected to do numerous updates throughout that 0:04:46.160 --> 0:04:49.840 period and have something that is implementable by the end 0:04:49.960 --> 0:04:53.320 of the sprint. Now, in other words, the earliest task 0:04:53.400 --> 0:04:56.200 for this new project was to work on programs that 0:04:56.200 --> 0:04:59.560 would allow unmanned aerial systems which we commonly refer to 0:04:59.680 --> 0:05:04.640 as drones, to analyze full motion video with object detection 0:05:04.720 --> 0:05:08.520 and classification in an effort to identify members of ISIS, 0:05:09.440 --> 0:05:14.000 areas of interest, equipment and weapons that bad actors might 0:05:14.040 --> 0:05:16.680 have at their disposal. The memo continues on to talk 0:05:16.720 --> 0:05:19.760 about the use of machine learning and automation and efforts 0:05:19.839 --> 0:05:25.560 of improving intelligence, surveillance and reconnaissance missions. The memo established 0:05:25.600 --> 0:05:29.880 May one, twenty seventeen, as the date of the first 0:05:29.960 --> 0:05:33.240 meeting of the a w c F T OR Project 0:05:33.240 --> 0:05:36.679 MAVEN to give reports directly to the Administrator of the program, 0:05:36.680 --> 0:05:41.120 who would be Robert Work himself. Lieutenant General Shanahan was 0:05:41.320 --> 0:05:44.440 named director of the project, which originally had only six 0:05:44.560 --> 0:05:47.760 members on it, and this marked an aggressive strategy to 0:05:47.839 --> 0:05:51.240 implement these technologies in the realm of the combat theater, 0:05:51.640 --> 0:05:54.919 and to do so very quickly. Now that's not to 0:05:54.960 --> 0:05:57.440 say that the Department of Defense was a stranger to 0:05:57.480 --> 0:06:01.000 cutting edge technology. Far from it. In nineteen fifty eight, 0:06:01.000 --> 0:06:05.400 President Dwight D. Eisenhower formed the Advanced Research Projects Agency 0:06:05.640 --> 0:06:09.360 or ARPA, which would change names to the Defense Advanced 0:06:09.400 --> 0:06:13.680 Research Projects Agency or DARPA in nineteen seventy two. While 0:06:13.720 --> 0:06:16.160 the agency has swapped names back and forth since then 0:06:16.200 --> 0:06:18.599 a couple of times, its mission has remained the same. 0:06:18.920 --> 0:06:22.719 The office provides funding for various projects aimed to expand 0:06:22.920 --> 0:06:26.600 science and technology, generally with some thought given towards the 0:06:26.640 --> 0:06:30.840 possible military benefits. Those projects have led to amazing things, 0:06:30.920 --> 0:06:34.960 including the Internet and the development of autonomous cars. But 0:06:35.040 --> 0:06:39.000 while those projects have had and will continue to have 0:06:39.240 --> 0:06:42.560 a major impact both in military and non military uses, 0:06:43.000 --> 0:06:46.600 the focus wasn't narrow enough for the purposes of Project Mayven. 0:06:46.880 --> 0:06:49.680 This project marked a change, one in which the military 0:06:49.680 --> 0:06:52.240 would be reaching out to experts in the various disciplines 0:06:52.279 --> 0:06:56.359 that comprise artificial intelligence, with the goal of improving military 0:06:56.400 --> 0:06:59.840 capabilities that could be implemented as soon as possible in 0:07:00.000 --> 0:07:04.240 a real world combat theater setting. Obviously, this raised many 0:07:04.320 --> 0:07:08.000 questions about the process and implementation of technologies. The political 0:07:08.040 --> 0:07:12.400 climate was, to put it mildly delicate. Many companies weren't 0:07:12.400 --> 0:07:14.600 eager to get involved in projects in the wake of 0:07:14.680 --> 0:07:18.720 Donald Trump's election to president, and various information leaks about 0:07:18.800 --> 0:07:23.320 governments and corporations had left many more companies a little 0:07:23.360 --> 0:07:27.640 cautious about getting involved in defense contracts. Part of the 0:07:27.680 --> 0:07:30.400 strategy to deal with this reluctance was a tight focus 0:07:30.480 --> 0:07:34.520 on a specific implementation of AI, that being full motion 0:07:34.680 --> 0:07:39.080 video analysis. The AI and Project MAVEN isn't meant to 0:07:39.080 --> 0:07:43.080 take any sort of military or offensive action. Instead, it's 0:07:43.080 --> 0:07:46.200 meant to sift through data. More on that in just 0:07:46.360 --> 0:07:49.720 a second, but first let's take a quick break to 0:07:49.840 --> 0:08:00.120 thank our sponsor. Not only was the committee looking for 0:08:00.280 --> 0:08:03.760 rapid development, the project also had the goal of streamlining 0:08:03.800 --> 0:08:07.000 all the bureaucratic red tape most parties had to endure 0:08:07.280 --> 0:08:11.320 when applying for funding from the government. The contracting procedures 0:08:11.320 --> 0:08:14.040 with the government had a well earned reputation for being 0:08:14.120 --> 0:08:18.600 laborious and slow. This alone discouraged many from applying to 0:08:18.600 --> 0:08:21.520 be part of government projects. Why would you go through 0:08:21.560 --> 0:08:24.840 the long approval process when you could work in the 0:08:24.880 --> 0:08:30.040 private sector and make money the entire time. Also, project 0:08:30.080 --> 0:08:34.120 may even focused on reducing the pains of contracting with 0:08:34.200 --> 0:08:37.040 government agencies. I thought that was kind of a clever 0:08:37.200 --> 0:08:40.280 part of the project, not just the idea of finding 0:08:40.320 --> 0:08:44.160 a way to get these technologies rapidly developed and deployed, 0:08:44.559 --> 0:08:47.240 but how to streamline the process on the front end 0:08:47.480 --> 0:08:51.959 to encourage more participants in the project. And then there's 0:08:52.000 --> 0:08:55.679 the desire for rapid deployment. Getting technology out in the 0:08:55.760 --> 0:08:59.000 field can be a long process on top of everything else. 0:08:59.320 --> 0:09:04.359 In short, all the stages of funding, developing, and implementing 0:09:04.400 --> 0:09:08.160 technology are traditionally so slow when it comes to government 0:09:08.200 --> 0:09:10.840 contracts that by the time you get the tech out 0:09:10.840 --> 0:09:14.920 into the real world, it's already obsolete. Project may even 0:09:14.920 --> 0:09:17.280 aimed to change all that. The goal was to develop 0:09:17.360 --> 0:09:22.440 tools and iterations and allow the user community that's their quote, 0:09:22.720 --> 0:09:26.160 to test them as they became available. Now, in this case, 0:09:26.200 --> 0:09:29.920 the user community happens to be the military to create 0:09:30.000 --> 0:09:32.880 an AI that can analyze full motion video and look 0:09:32.960 --> 0:09:36.280 look for specific things within that video. That was the 0:09:36.320 --> 0:09:39.559 whole purpose of the project, and it called for an 0:09:39.640 --> 0:09:43.240 artificial neural network. I've talked about these before, but let 0:09:43.280 --> 0:09:46.040 me give a quick rundown of what this is right now. 0:09:46.840 --> 0:09:50.520 Roughly speaking, and artificial neural network is a system of 0:09:50.640 --> 0:09:55.120 one or more computers in which units of calculation called neurons, 0:09:55.400 --> 0:09:59.360 connect to one another through weighted values called synapses in 0:09:59.440 --> 0:10:03.360 an effort process information in a way that's similar to 0:10:03.480 --> 0:10:07.680 how our brains work. As these neurons perform operations on data, 0:10:07.800 --> 0:10:10.720 they send the data through the synapses, which affects the 0:10:10.800 --> 0:10:15.120 data itself. This data eventually emerges as output, though the 0:10:15.160 --> 0:10:18.000 design of the neural network determines how many neurons it 0:10:18.080 --> 0:10:21.240 must pass through before this happens. The goal is to 0:10:21.280 --> 0:10:24.520 create a system that can actually learn, once trained to 0:10:24.640 --> 0:10:28.040 do something, so Let's say you've got an artificial neural 0:10:28.080 --> 0:10:30.599 network and you want to train it to recognize a 0:10:30.679 --> 0:10:35.040 specific image, and we'll say, for the purposes of this example, 0:10:35.280 --> 0:10:38.720 that the image is a cat. You start to feed 0:10:38.880 --> 0:10:42.160 the artificial neural network a series of images, some of 0:10:42.200 --> 0:10:44.880 them cats and some of them of other things, and 0:10:44.920 --> 0:10:47.720 you design the network so it identifies an object as 0:10:47.720 --> 0:10:51.839 a cat by process of elimination. And if it does 0:10:52.040 --> 0:10:55.240 identify something as a cat and it's not a cat, 0:10:55.720 --> 0:10:59.559 it generates an error. That error then back propagates through 0:10:59.600 --> 0:11:03.120 the whole network, and the system quote unquote learns that 0:11:03.120 --> 0:11:06.079 that particular image did not represent a cat, and if 0:11:06.120 --> 0:11:10.200 it encounters that image again, it won't mistakenly identify the 0:11:10.240 --> 0:11:13.080 image as a cat. The same is true if it 0:11:13.240 --> 0:11:17.160 fails to identify a cat that is present in an image. 0:11:17.600 --> 0:11:20.880 Doing this millions of times will refine the system as 0:11:20.880 --> 0:11:24.480 it learns what is and is not a cat. Google 0:11:24.600 --> 0:11:28.840 actually did something similar to this several years ago. Google's 0:11:28.880 --> 0:11:32.679 research and development lab created an artificial neural network consisting 0:11:32.880 --> 0:11:37.960 of sixteen thousand computer processors with a billion connections within 0:11:38.000 --> 0:11:41.640 the system. They fed the system ten million YouTube video 0:11:41.760 --> 0:11:45.000 thumbnails selected at random, then they gave the system a 0:11:45.040 --> 0:11:48.680 list of twenty thousand items. The system began to recognize 0:11:48.720 --> 0:11:51.960 pictures of cats using a deep learning algorithm. And remember, 0:11:52.120 --> 0:11:56.160 an algorithm is just a set of instructions directions that 0:11:56.200 --> 0:11:58.560 you follow in order to get to a specific result. 0:11:59.000 --> 0:12:02.160 The big breakthrough year was that this system was able 0:12:02.200 --> 0:12:05.480 to recognize the image of a cat without first being 0:12:05.520 --> 0:12:09.640 taught what a cat actually was. It learned through training 0:12:09.640 --> 0:12:12.600 itself by looking at this large set of data. In 0:12:12.600 --> 0:12:16.200 a similar fashion, Project May even wished to develop an 0:12:16.320 --> 0:12:21.320 artificial neural network that could identify potential ISIS activity. Lieutenant 0:12:21.400 --> 0:12:26.560 General Shanahan wrote about how the military has countless hours 0:12:26.679 --> 0:12:30.880 of footage gathered by unmanned aerial systems or drones. A 0:12:30.960 --> 0:12:34.720 quick word about these. The ones he named specifically included 0:12:34.800 --> 0:12:38.280 the Scan Eagle, the m Q one C Gray Eagle, 0:12:38.720 --> 0:12:42.320 and the m Q nine Reaper. The Boeing in situ 0:12:42.679 --> 0:12:46.360 Scan Eagle is a low altitude drone that is more 0:12:46.400 --> 0:12:49.079 than five ft long or about one and a half meters, 0:12:49.440 --> 0:12:51.880 with a wingspan a little more than ten ft or 0:12:52.000 --> 0:12:54.960 three point one one ms wide. It can travel at 0:12:55.000 --> 0:12:58.520 ninety two miles per hour. Also at kilometers per hour 0:12:58.640 --> 0:13:00.800 at top speed, and it can stay in flight for 0:13:00.840 --> 0:13:03.240 more than twenty four hours at a time. Has high 0:13:03.280 --> 0:13:07.760 resolution imaging sensors on it, including thermal imagery sensors. The 0:13:07.880 --> 0:13:11.239 m Q one C Gray Eagle is built by General Atomics. 0:13:11.600 --> 0:13:14.880 It's a medium altitude drone and is an upgrade to 0:13:14.920 --> 0:13:18.839 the famous Predator drone. It's much larger than the Scan Eagle, 0:13:18.960 --> 0:13:21.760 at twenty eight feet or eight point five three ms 0:13:21.840 --> 0:13:24.920 long and a wingspan of fifty six feet or seventeen 0:13:25.000 --> 0:13:27.960 meters wide. It can travel at a hundred two miles 0:13:28.000 --> 0:13:31.400 per hour or three nine kilometers per hour at top speed, 0:13:31.840 --> 0:13:36.760 and it, unlike the Scan Eagle, can be armed with 0:13:36.800 --> 0:13:40.960 stuff like bombs and missiles. The m Q nine Reaper, 0:13:41.200 --> 0:13:44.760 sometimes called the Predator B is even larger. It's more 0:13:44.800 --> 0:13:47.719 than thirty six ft long or about eleven meters, it's 0:13:47.760 --> 0:13:51.720 got a wingspan of sixty ft seven inches or twenty meters, 0:13:51.960 --> 0:13:53.679 and it can travel at three hundred miles per hour 0:13:53.840 --> 0:13:56.640 or two kilometers per hour at top speed, and it 0:13:56.720 --> 0:14:02.480 can also carry a various armory of weapons all on itself. 0:14:02.559 --> 0:14:07.640 So those are drones that could be used for offensive measures. 0:14:08.000 --> 0:14:12.400 While the Lieutenant General was talking only about using an 0:14:12.480 --> 0:14:16.839 artificial neural network to analyze video captured by devices like this. 0:14:17.720 --> 0:14:20.760 Naming a couple of drones that are weaponized likely raised 0:14:20.800 --> 0:14:23.640 many eyebrows, but let me stick with what he was 0:14:23.680 --> 0:14:29.480 pitching back in November. Just for his argument, Shanahan specifically 0:14:29.520 --> 0:14:33.000 talked about how these drones were gathering thousands of hours 0:14:33.080 --> 0:14:37.320 of video intelligence, but sifting through that intelligence required even 0:14:37.320 --> 0:14:41.840 more time and many human analysts. Even then, you could 0:14:41.840 --> 0:14:45.200 only tackle a fraction of what was being gathered. This 0:14:45.280 --> 0:14:47.280 meant that most of the time you were reacting to 0:14:47.360 --> 0:14:50.680 something that had happened in the field. For example, if 0:14:50.680 --> 0:14:55.640 an improvised explosive detonated, you might scour through video footage 0:14:55.640 --> 0:14:58.720 of the area leading up to the detonation in an 0:14:58.760 --> 0:15:02.120 attempt to identify the persons responsible for it and then 0:15:02.160 --> 0:15:05.480 track their movements. It still takes an incredible amount of 0:15:05.480 --> 0:15:08.360 time and work to follow these things, but one of 0:15:08.440 --> 0:15:11.720 you could automate the system, bringing up analysts to look 0:15:11.720 --> 0:15:16.120 at what Shanahan was referring to as higher value analysis 0:15:16.160 --> 0:15:20.040 work and AI system that could comb through hours of 0:15:20.160 --> 0:15:24.200 data and automatically classify and label things, categorizing them either 0:15:24.320 --> 0:15:27.960 as mundane and unimportant or something to pay attention to 0:15:28.080 --> 0:15:31.800 flagging it for human analysis. It could conceivably make a 0:15:31.880 --> 0:15:34.960 huge difference and speed things up. Not at the start. 0:15:35.680 --> 0:15:39.840 Humans labeled more than one hundred fifty thousand images to 0:15:39.920 --> 0:15:43.400 train Maven on data sets, with the goal of increasing 0:15:43.440 --> 0:15:46.800 that up to one million images by the end of January. 0:15:47.880 --> 0:15:50.520 This would give Maven a start at being able to 0:15:50.600 --> 0:15:56.040 identify various objects at different distances, resolutions, angles, and more. 0:15:56.160 --> 0:16:00.040 Because remember, computer vision is a tricky thing. Teaching a 0:16:00.080 --> 0:16:05.280 computer to recognize an image of something is tricky enough, 0:16:05.480 --> 0:16:08.160 even if you're just sticking to one kind of lighting, 0:16:08.640 --> 0:16:12.080 one orientation. I always use the example of a coffee mug. 0:16:12.360 --> 0:16:15.360 Let's say you've got a red, bright red coffee mug 0:16:15.640 --> 0:16:18.240 and the handles pointed toward the left with respect to 0:16:18.320 --> 0:16:21.960 your view through an image, and you teach a computer 0:16:22.080 --> 0:16:24.400 this is a coffee mug. Well, what happens if you 0:16:24.440 --> 0:16:26.360 have it under a different lighting so it doesn't look 0:16:26.400 --> 0:16:28.760 like it's bright red, and maybe you've turned it so 0:16:28.800 --> 0:16:31.000 that the handles facing the other way, and maybe the 0:16:31.080 --> 0:16:33.160 angle is a little bit different, so you're looking kind 0:16:33.160 --> 0:16:36.240 of down into the cup. Will the computers still be 0:16:36.280 --> 0:16:39.360 able to recognize that as a coffee mug. You have 0:16:39.440 --> 0:16:41.880 to train it. And then let's say that you change 0:16:42.080 --> 0:16:44.960 the color of the coffee mug. Is the exact same shape, 0:16:45.040 --> 0:16:48.200 but now it's blue instead of red. Will the computer 0:16:48.280 --> 0:16:50.960 recognize it. Let's say that you change the shape of 0:16:51.000 --> 0:16:53.680 the coffee mug. Now it's a different style of coffee 0:16:53.720 --> 0:16:57.360 mug from the previous one. We humans can pick up 0:16:57.360 --> 0:16:59.480 on this very quickly. You teach a human a couple 0:16:59.520 --> 0:17:01.760 of things about coffee mugs, and then they kind of 0:17:01.800 --> 0:17:04.120 get the innate grasp of it. You can show them 0:17:04.119 --> 0:17:09.120 all sorts of different sizes, shapes, colors, lighting conditions. They're 0:17:09.119 --> 0:17:11.480 going to recognize that as a coffee mug. The same 0:17:11.560 --> 0:17:14.600 is not necessarily true with computers, so training it is 0:17:14.640 --> 0:17:18.000 a laborious process. Now, once it is trained, it can 0:17:18.040 --> 0:17:21.280 go through data far faster than a human could, but 0:17:21.359 --> 0:17:24.640 you still have to teach it. Shanahan viewed this project 0:17:24.680 --> 0:17:28.240 as proof that a small, nimble team approach in getting 0:17:28.280 --> 0:17:31.840 the right parties involved worked, and that this in turn 0:17:31.880 --> 0:17:35.280 would spawn a new era of high tech projects aimed 0:17:35.280 --> 0:17:39.960 at incorporating AI into other military operations. He also even 0:17:40.000 --> 0:17:44.280 expressed a little caution about this era. So what exactly 0:17:44.600 --> 0:17:47.199 about all of this prompted so many at Google to 0:17:47.280 --> 0:17:51.040 protest the company's involvement. Well, I'll explain that in just 0:17:51.119 --> 0:17:53.840 a second, but first let's take another quick break to 0:17:54.000 --> 0:18:04.080 thank our sponsor. By the end of twenty seen technology 0:18:04.119 --> 0:18:07.240 developed for a project Maven was in use in various 0:18:07.320 --> 0:18:10.200 sights in the Middle East, and that's an incredible turnaround. 0:18:10.200 --> 0:18:12.560 It was less than a year that had gone by 0:18:12.640 --> 0:18:15.800 since the April memo had launched the project, and already 0:18:16.160 --> 0:18:19.280 AI algorithms were being trained to look for specific types 0:18:19.320 --> 0:18:23.040 of data within full motion video footage. Not only was 0:18:23.080 --> 0:18:25.080 it being used in the Middle East, the military was 0:18:25.080 --> 0:18:27.680 already starting to use it in other parts of the 0:18:27.720 --> 0:18:30.960 world like Africa. The military stress that this tech was 0:18:31.000 --> 0:18:35.359 meant to augment personnel's abilities in gathering and sifting through information. 0:18:35.400 --> 0:18:39.280 It was meant to flag data so that a human 0:18:39.320 --> 0:18:42.800 analyst could look it over. Nothing was automatically happening through 0:18:42.800 --> 0:18:46.040 this system. There was no intent to do anything beyond 0:18:46.160 --> 0:18:50.560 analyzing information the military was already gathering. But that's still 0:18:50.600 --> 0:18:54.560 a pretty alarming revelation for many people and over at Google. 0:18:55.359 --> 0:18:58.399 Google being a company that was handling a lot of 0:18:58.440 --> 0:19:01.439 this information, a lot of these algorithms, and working with 0:19:01.440 --> 0:19:04.720 the military to develop them. Worries we're growing that it 0:19:04.760 --> 0:19:08.040 would not stop at data analysis, and this brings us 0:19:08.080 --> 0:19:11.679 to the petition that thousands of Google employees signed. The 0:19:11.760 --> 0:19:15.959 Google petition opens with a pretty clear message. Quote, we 0:19:16.080 --> 0:19:20.280 believe that Google should not be in the business of war. Therefore, 0:19:20.440 --> 0:19:23.000 we asked that project may even be canceled, and that 0:19:23.119 --> 0:19:27.919 Google draft publicize and enforce a clear policy saying that 0:19:27.960 --> 0:19:32.359 neither Google nor its contractors will ever build warfare technology. 0:19:32.480 --> 0:19:36.640 End quote. The petition also expressed skepticism about Project Maven's 0:19:36.720 --> 0:19:41.520 stated purpose. In the third paragraph, it reads, quote, recently 0:19:41.880 --> 0:19:47.520 Googler's voice concerns about Maven internally, Diane Green responded, assuring 0:19:47.600 --> 0:19:50.920 them that the technology will not operate or fly drones 0:19:51.280 --> 0:19:54.119 and will not be used to launch weapons. While this 0:19:54.200 --> 0:19:57.760 eliminates a narrow set of direct applications, the technology is 0:19:57.800 --> 0:20:00.639 being built for the military, and once it's delivered, it 0:20:00.640 --> 0:20:04.840 could easily be used to assist in these tasks end quote. 0:20:05.320 --> 0:20:09.080 The petition states that beyond the ethical and moral implications 0:20:09.119 --> 0:20:12.120 of the project, that it will do damage to the company, 0:20:12.200 --> 0:20:16.439 including hurting its ability to attract talent to work for Google, saying, 0:20:16.920 --> 0:20:20.600 if we have a reputation for giving the military technology 0:20:20.640 --> 0:20:23.800 that helps them kill people, it's gonna be really hard 0:20:23.840 --> 0:20:27.440 to convince new developers to come work for our company. 0:20:27.640 --> 0:20:31.440 A Google spokesperson responded to the petition with a letter 0:20:31.960 --> 0:20:35.720 It says, quote MAVEN is a well publicized Department of 0:20:35.760 --> 0:20:39.320 Defense project, and Google is working on one part of it, 0:20:39.600 --> 0:20:43.440 specifically scope to be for non offensive purposes and using 0:20:43.560 --> 0:20:48.520 open source object recognition software available to any Google Cloud customer. 0:20:48.920 --> 0:20:52.880 The models are based on unclassified data only. The technology 0:20:53.000 --> 0:20:55.680 is used to flag images for human review and is 0:20:55.720 --> 0:20:58.879 intended to save lives and save people from having to 0:20:58.960 --> 0:21:03.040 do highly td is work. End quote. So the response 0:21:03.080 --> 0:21:08.359 here saying this technology, these algorithms are already available and 0:21:08.359 --> 0:21:11.760 they're open source. Anyone could take that source code and 0:21:11.800 --> 0:21:16.240 develop applications based upon it. So really, Google could take 0:21:16.280 --> 0:21:18.560 the contract and make some money off of it, or 0:21:18.640 --> 0:21:21.959 not take the contract, but their work still ends up 0:21:21.960 --> 0:21:25.119 being used for that purpose because it's open source and 0:21:25.160 --> 0:21:28.600 anyone can use it. According to this spokesperson, the AI 0:21:28.720 --> 0:21:32.879 and project may even is only assisting humans by pulling 0:21:32.960 --> 0:21:35.280 up that data that may be of interest while ignoring 0:21:35.320 --> 0:21:38.280 everything else, so it's not making any decisions on its own. 0:21:38.280 --> 0:21:41.520 But even Shanahan states in his blog post that there 0:21:41.520 --> 0:21:44.720 are ethical questions when it comes to incorporating AI that 0:21:44.960 --> 0:21:48.800 must be addressed. He urges for the development of quote, 0:21:49.040 --> 0:21:54.080 technological and organizational safeguards to ensure that Washington's military use 0:21:54.119 --> 0:21:58.359 of AI is consistent with national values end quote. He 0:21:58.400 --> 0:22:01.040 also points out that AI could be vulnerable to different 0:22:01.080 --> 0:22:04.200 failure modes that could be disastrous, and that these two 0:22:04.280 --> 0:22:08.040 must be taken into consideration. So he's saying that we 0:22:08.119 --> 0:22:10.840 have to implement it responsibly, and we have to be 0:22:10.880 --> 0:22:13.480 aware of its failure points so that we can make 0:22:13.520 --> 0:22:17.640 sure that it's not vulnerable to them, because otherwise we're 0:22:17.640 --> 0:22:21.200 going to rely far too heavily on a dangerous technology 0:22:21.359 --> 0:22:25.840 that could end up causing irreparable harm if misused. To 0:22:25.960 --> 0:22:29.679 further complicate matters, other groups have also urged Google to 0:22:29.720 --> 0:22:33.160 step away from military projects. In April two thousand eighteen, 0:22:33.240 --> 0:22:36.520 one year after May Even launched as a project, the 0:22:36.680 --> 0:22:40.439 Tech Workers Coalition, which is a group representing employees of 0:22:40.480 --> 0:22:45.280 major tech companies like Google, IBM, Microsoft, Amazon, and others, 0:22:45.640 --> 0:22:48.919 launched their own petition stating that tech should not be 0:22:49.000 --> 0:22:52.440 in the business of war. This petition says that military 0:22:52.480 --> 0:22:56.560 contracts break user trust, in this case the user being 0:22:56.600 --> 0:22:59.080 either the general public or in the case of big 0:22:59.119 --> 0:23:03.320 companies like ib UM, other companies, and the International Committee 0:23:03.359 --> 0:23:08.959 for Robot Arms Control meaning autonomous weapons not you know 0:23:09.359 --> 0:23:12.760 the arms of a robot issued an open letter to 0:23:12.840 --> 0:23:16.760 Google urging the company to stop pursuing military contracts. More 0:23:16.800 --> 0:23:21.160 than ninety academics signed this paper. The letter paints a 0:23:21.280 --> 0:23:24.320 very grim picture of a possible use for Maven, that 0:23:24.440 --> 0:23:28.760 of identifying targets based upon probabilities arrived at by analyzing 0:23:28.800 --> 0:23:33.119 long range surveillance footage, ultimately resulting in signature strikes and 0:23:33.160 --> 0:23:35.960 pattern of life strikes, meaning that Google would at least 0:23:36.000 --> 0:23:40.159 be somewhat complicit in targeted killings. The letter goes on 0:23:40.200 --> 0:23:42.919 to state that while the express purpose of Maven is 0:23:43.000 --> 0:23:47.639 purely for analysis, such tools could be turned toward automated 0:23:47.680 --> 0:23:50.480 target recognition in the future. Those of you who are 0:23:50.600 --> 0:23:53.280 Terminator fans might think of this as another step toward 0:23:53.359 --> 0:23:57.320 the mythical sky net system, which would ultimately turn against 0:23:57.440 --> 0:24:00.920 humans and attempt to wipe us out. Sending Arnold Schwarzenegger 0:24:00.960 --> 0:24:04.960 back in time. The letter ends with three requests, First 0:24:05.000 --> 0:24:08.119 that Google terminate its project may even contract with the 0:24:08.160 --> 0:24:11.639 Department of Defense. Second, that the company commit not to 0:24:11.720 --> 0:24:15.480 develop military technologies nor to allow the personal data is 0:24:15.520 --> 0:24:19.159 collected to be used for military operations. And finally that 0:24:19.240 --> 0:24:24.200 it pledged to neither participate in nor support the development, manufacture, trade, 0:24:24.400 --> 0:24:27.439 or use of autonomous weapons, and to support efforts to 0:24:27.640 --> 0:24:33.040 ban autonomous weapons. So far, Google hasn't shown any signs 0:24:33.080 --> 0:24:36.240 of following those suggestions. In fact, it's been reported that 0:24:36.280 --> 0:24:39.600 the company is actively bidding on another Department of Defense 0:24:39.640 --> 0:24:45.840 project called the Joint Enterprise Defense Infrastructure or JEDI. This 0:24:45.920 --> 0:24:49.200 project's goal is to create a suite of cloud services 0:24:49.240 --> 0:24:52.440 for the Department of Defense. It's largely intended to reduce 0:24:52.520 --> 0:24:56.359 complexity in military data storage systems, which at the moment 0:24:56.359 --> 0:25:00.320 are fragmented across multiple branches and departments. The at EYE 0:25:00.320 --> 0:25:03.640 contract will go to a single vendor, which means one 0:25:03.720 --> 0:25:07.040 company stands to make a lot of money from the project. 0:25:07.520 --> 0:25:09.760 That's hard to walk away from. I suppose that if 0:25:09.760 --> 0:25:12.720 thousands of your employees are protesting the move, it might 0:25:12.880 --> 0:25:18.360