WEBVTT - The Microsoft Tay Debacle 0:00:00.160 --> 0:00:07.160 Brought to you by Toyota. Let's go places. Welcome to 0:00:07.360 --> 0:00:18.600 Forward Thinking. Hey there, and welcome to Forward Thinking, the 0:00:18.640 --> 0:00:21.280 podcast that looks at the future and says I texted 0:00:21.320 --> 0:00:22.959 you on a Monday, but you did not get my 0:00:23.000 --> 0:00:25.960 text until Tuesday because of a network problem. I'm Jonathan 0:00:26.000 --> 0:00:30.760 Strickland and I'm Joe McCormick. Did that problem have to 0:00:30.840 --> 0:00:35.440 do with the tubes? Yes, yes, I can't believe I 0:00:35.560 --> 0:00:38.800 made a tubes joke. No, I take it back. I'm sorry. 0:00:38.880 --> 0:00:43.400 I'm sorry. You've been a radio and the radio is forever. No, no, no, no, 0:00:43.479 --> 0:00:46.559 I apologize. Easy jokes beneath me. We're going to talk 0:00:46.560 --> 0:00:50.239 about chat bots today. Yeah, we've talked about chat bots before, right, 0:00:50.280 --> 0:00:52.159 We've talked about them in a couple of different episodes. 0:00:52.159 --> 0:00:54.720 We've talked about chat bots, we talked about AI, We've 0:00:54.720 --> 0:00:57.840 talked about touring tests, we've talked about machine learning. All 0:00:57.880 --> 0:01:00.760 of those topics are going to play a part in 0:01:00.800 --> 0:01:03.320 our discussion today. And speaking of easy jokes that are 0:01:03.360 --> 0:01:07.160 beneath us, we are talking specifically today about tay tweets. Yes, 0:01:07.240 --> 0:01:12.800 Microsoft ta uh yeah, Tay Tay. We we we hardly 0:01:12.840 --> 0:01:15.720 knew ye. In fact, I doubt anyone really knew. Ye. 0:01:15.920 --> 0:01:18.520 I didn't find out about Tay until Tay was already gone. 0:01:18.600 --> 0:01:21.000 I found out about Tay while TAY was still active, 0:01:21.040 --> 0:01:24.320 but I never messaged Tay. And the only reason I 0:01:24.360 --> 0:01:27.360 found out about Tay was because I had been invited 0:01:27.400 --> 0:01:30.600 to be a guest on a show called Tech News Today, 0:01:30.720 --> 0:01:33.160 and that was one of the news items that they covered. 0:01:33.640 --> 0:01:36.480 And I had been very busy in our day to 0:01:36.560 --> 0:01:40.240 day work and had not followed along. So when I 0:01:40.319 --> 0:01:42.160 when I heard about it, it it was before all the 0:01:42.240 --> 0:01:45.240 craziness unfolded. And we'll get into what the craziness was. 0:01:45.880 --> 0:01:50.360 To be fair, though, even at ground like square one, 0:01:51.000 --> 0:01:54.760 it seems to me like TAY was probably not the 0:01:54.800 --> 0:01:57.760 most thought out project Microsoft has ever launched. And I'm 0:01:57.800 --> 0:02:01.640 including Microsoft Bob in that. Yeah, that's a that's a 0:02:01.720 --> 0:02:05.240 sick burn for those who know what Microsoft Bob was. Okay, 0:02:05.360 --> 0:02:09.079 tell me more, all right, Well now that we we 0:02:09.160 --> 0:02:12.040 got our little rascals reference all the way. So, according 0:02:12.080 --> 0:02:16.560 to Microsoft, Tay is quote an artificial intelligent chatbot developed 0:02:16.560 --> 0:02:21.160 by Microsoft Technology and Research and BING teams to experiment 0:02:21.200 --> 0:02:26.200 with and conduct research on conversational understanding. UM, I know 0:02:26.280 --> 0:02:29.800 some people would probably start to argue with the word 0:02:29.919 --> 0:02:33.680 intelligent that was in that description based upon some of 0:02:33.680 --> 0:02:36.720 the things that were happening with t Now, that description 0:02:36.760 --> 0:02:41.440 doesn't exactly capture what the public saw of Tape because 0:02:41.480 --> 0:02:46.840 the public facing incarnation of ta was specifically a Twitter account. Well, 0:02:46.960 --> 0:02:50.519 Twitter account was the primary one. Yeah, she or it. 0:02:51.000 --> 0:02:54.600 I get very confused about referring to AI because they're robots. 0:02:54.720 --> 0:02:57.600 Robots don't have genders. But Tay was impersonating a nineteen 0:02:57.639 --> 0:03:02.440 year old girl, so our woman. Yeah. Also when when 0:03:02.520 --> 0:03:05.120 at what age is an AI a woman? I don't 0:03:05.120 --> 0:03:11.160 know yet? An AI no longer the longer a girl 0:03:11.280 --> 0:03:15.640 type of thing. She was a young adult AI. Okay, 0:03:15.639 --> 0:03:20.080 while I continue to go through this this oddly misplaced 0:03:20.320 --> 0:03:25.959 feeling of cultural shame, Yes, she was also on group 0:03:26.000 --> 0:03:30.080 me and Kick and Kick So what are what are 0:03:30.120 --> 0:03:32.240 you even talking? So group me, Lauren and I know 0:03:32.320 --> 0:03:35.400 all about group me. Yeah, we happen to use group 0:03:35.400 --> 0:03:38.520 me while we were in Austin for south By Southwest. 0:03:38.640 --> 0:03:41.840 So group me is essentially like a very contained version 0:03:41.880 --> 0:03:46.000 of Twitter. It's it's a collection of cell phone numbers, essentially, 0:03:46.000 --> 0:03:48.520 so that you can send a message to the group 0:03:48.600 --> 0:03:52.520 me which includes multiple people. Everyone receives that message and 0:03:52.560 --> 0:03:55.280 everyone can respond to it. Yeah, it's a fancy group 0:03:55.280 --> 0:03:57.920 messaging system that also has a little bit of capacity 0:03:57.960 --> 0:04:00.240 to look like cute stuff, like you can up a 0:04:00.240 --> 0:04:04.960 picture and um immediately place meme style text on it. Yeah, 0:04:05.120 --> 0:04:08.080 it used to be that it was just group texting. 0:04:08.160 --> 0:04:10.280 So when I started seeing the pictures popping up with 0:04:10.320 --> 0:04:12.920 all the memes, I immediately went into old man mode 0:04:12.920 --> 0:04:17.880 and got really confused. Kick is also an instant messaging platform. Uh, 0:04:17.920 --> 0:04:21.520 it's based off the BlackBerry Messenger platform, So it's all. 0:04:21.560 --> 0:04:24.520 Three of these are different ways to send out messages 0:04:24.600 --> 0:04:28.120 to typically groups of people, not not like a one 0:04:28.120 --> 0:04:30.560 on one thing necessarily, although I think Kick can can 0:04:30.640 --> 0:04:35.359 sometimes fall into that category. But um, honestly, before we 0:04:35.440 --> 0:04:39.960 even get into two, more of the weirdness here. From 0:04:40.040 --> 0:04:43.640 the get go, I felt odd about the idea of 0:04:44.120 --> 0:04:46.880 framing an AI within the context of not only a 0:04:46.920 --> 0:04:48.880 gender but an age, kind of like you were saying, Lauren, 0:04:48.920 --> 0:04:52.760 I I most of the time when I think of AI, 0:04:53.279 --> 0:04:56.200 unless you're specifically trying to create something that's going to 0:04:56.240 --> 0:04:59.000 pass a touring test and therefore fool someone into thinking 0:04:59.040 --> 0:05:01.000 it's an actual first I was gonna say, this sounds 0:05:01.000 --> 0:05:03.760 like some Eugene kind of trickery. Yeah, and we'll talk 0:05:03.800 --> 0:05:07.200 about goosetman in a minute to Unless you're doing that, 0:05:07.279 --> 0:05:10.239 there's not really I don't see a purpose to giving 0:05:10.800 --> 0:05:14.039 a gender and age to an AI other than maybe 0:05:14.080 --> 0:05:18.640 there's some studies perhaps that suggest humans are more likely 0:05:18.760 --> 0:05:22.080 to have a natural interaction with someone that they can 0:05:22.200 --> 0:05:25.880 kind of kind of categorize in some way. And I 0:05:25.920 --> 0:05:29.320 don't know if that, if that research even exists. That's 0:05:29.360 --> 0:05:33.000 just me taking wild shots in the dark. Um, but 0:05:33.320 --> 0:05:38.760 I personally thought I would be very reticent to send 0:05:38.760 --> 0:05:41.520 a message to an entity that had been more or 0:05:41.600 --> 0:05:45.160 less identified as a nineteen year old female, even just 0:05:45.240 --> 0:05:48.120 to see how it was working, because I'm a forty 0:05:48.200 --> 0:05:51.919 year old man, and I don't think sending messages to 0:05:51.920 --> 0:05:55.040 to an entity that's less than half my age, like 0:05:55.080 --> 0:05:58.440 it just feels inappropriate from the from the start to me, 0:05:58.560 --> 0:06:00.839 even though I know it's not a person in right, 0:06:01.560 --> 0:06:04.360 it all exactly made me feel kind of squidgy. It well, sure, 0:06:04.760 --> 0:06:08.800 it's just putting putting a strange basis from from our 0:06:08.839 --> 0:06:11.480 point of view, perhaps, I mean, and and the boat 0:06:11.560 --> 0:06:16.440 was supposedly geared towards people I think eighteen to twenty four. 0:06:17.520 --> 0:06:19.400 So so it wasn't meant for me in the first place, 0:06:20.279 --> 0:06:23.000 but it wasn't designed by people who are eighteen to 0:06:23.040 --> 0:06:26.080 twenty four. It's fairly I think it was designed by 0:06:26.080 --> 0:06:28.839 people closer to my age, and that that is one 0:06:28.839 --> 0:06:34.640 of the many problems that Tay has. So yes, slightly 0:06:34.680 --> 0:06:37.960 odd thing, but perhaps a conversation for another day. Yeah, So, 0:06:38.040 --> 0:06:40.480 at any rate, Tay, like you said, eighteen to twenty 0:06:40.520 --> 0:06:43.280 four year olds, that was the target demographic because Microsoft 0:06:43.320 --> 0:06:47.200 had identified that as the demographic that most frequently uses 0:06:47.360 --> 0:06:52.839 mobile based social media things like social messengers and um 0:06:52.920 --> 0:06:55.080 and that's really where they were looking to kind of 0:06:55.680 --> 0:07:01.280 increase this ability for AI to understand conversational English things, 0:07:01.360 --> 0:07:06.880 things that go beyond very simple, uh prescribed commands. Okay, 0:07:06.880 --> 0:07:09.400 so in practice, what does this look like? What exactly 0:07:09.520 --> 0:07:12.520 was Ta doing? So in practice, you would send Tay 0:07:12.720 --> 0:07:15.320 a message on one of these platforms, either Group Me, 0:07:15.520 --> 0:07:18.720 Kick or Twitter, and then t would respond to you. 0:07:19.160 --> 0:07:22.160 If it was on Twitter and things were particularly busy, 0:07:22.240 --> 0:07:25.600 Tay would tell you to send a direct message because 0:07:25.760 --> 0:07:28.800 otherwise we just get kind of lost in the crush. Obviously, 0:07:28.880 --> 0:07:30.760 if you unleash something like this to the world at 0:07:30.840 --> 0:07:32.960 large and you let it be known that it's out there, 0:07:33.120 --> 0:07:35.560 you're gonna get a huge response, right, especially from Twitter, 0:07:36.040 --> 0:07:39.320 so mostly consisting of mischief and I'm sure we'll talk 0:07:39.360 --> 0:07:43.040 about that some mortally. Yeah. So, uh yeah, this was 0:07:43.560 --> 0:07:46.120 you know, that was exactly the case. So they released 0:07:46.160 --> 0:07:50.280 it to try and see what would happen, how Tay 0:07:50.360 --> 0:07:56.120 would start to incorporate different types of language in her replies, 0:07:56.360 --> 0:07:58.840 how she would be able to interpret different things that 0:07:58.880 --> 0:08:01.560 were said to her. They started with a baseline where 0:08:01.560 --> 0:08:06.520 Microsoft had essentially mined a ton of publicly available data. 0:08:06.640 --> 0:08:10.080 So my guess is would be public tweets would probably 0:08:10.080 --> 0:08:13.240 be a big one. Right. They identify the demographic that 0:08:13.280 --> 0:08:17.280 they're looking at, they mine the sort of communication that 0:08:17.360 --> 0:08:20.680 these uh, these folks that are within that demographic tend 0:08:20.720 --> 0:08:25.040 to send, and he used as the model for Tay's communication, 0:08:25.680 --> 0:08:29.240 and then the idea was to build upon that. Uh So, 0:08:29.520 --> 0:08:32.640 you know, we mentioned Goostman, uh talking about the fact 0:08:32.720 --> 0:08:34.760 that you know, we've got this context of a nineteen 0:08:34.840 --> 0:08:40.760 year old female persona for Tay with Eugene Goostman, that 0:08:40.880 --> 0:08:44.440 was one of the AI programs that quote unquote passed 0:08:44.520 --> 0:08:46.920 the Turing test. We talked about this in a previous 0:08:46.920 --> 0:08:50.679 episode in which uh it was a an AI that 0:08:50.760 --> 0:08:53.840 was mimicking a thirteen year old Ukrainian boy, a non 0:08:53.920 --> 0:08:57.800 native English speaker, and that meant that the people who 0:08:57.840 --> 0:09:04.960 were uh interviewing this entity over a over a computer, uh, 0:09:05.120 --> 0:09:07.640 they were aware that they were supposedly talking to a 0:09:07.760 --> 0:09:12.520 human boy thirteen years old, non native English speaker, limited 0:09:12.840 --> 0:09:15.480 exposure to things like pop culture that would have come 0:09:15.520 --> 0:09:18.840 before he was born. So in other words, it checked 0:09:18.880 --> 0:09:24.760 people's expectations of what this particular uh AI would be 0:09:24.800 --> 0:09:28.319 capable of doing. Sure, and similarly, I think the part 0:09:28.400 --> 0:09:31.840 of the concept for Tay was probably that you have 0:09:31.920 --> 0:09:35.000 certain expectations for how a nineteen year old woman is 0:09:35.040 --> 0:09:39.000 going to behave on Twitter, and it probably involves the 0:09:39.120 --> 0:09:45.240 use of of emojis and um uh not full grammatical sentences, 0:09:45.559 --> 0:09:50.400 no capital letters whatsoever. Certainly um certain truncating of of 0:09:50.480 --> 0:09:55.920 thoughts and ideas, a certain inquisitiveness probably are the stereotypes 0:09:56.000 --> 0:09:57.959 of that sort of age group. It was to a point, 0:09:58.160 --> 0:10:01.200 and to be fair, I've only looked at sampling of 0:10:01.320 --> 0:10:03.040 some of the things he has said. It was to 0:10:03.120 --> 0:10:06.520 a point where I thought, I thought it was venturing 0:10:06.679 --> 0:10:09.760 on the border of parody, right. I was starting to 0:10:09.800 --> 0:10:14.959 feel like it was, Um, maybe someone my age is 0:10:15.600 --> 0:10:18.680 preconception of what a person would how a person would 0:10:18.720 --> 0:10:21.920 constantly be communicating. Uh. And again that could just be 0:10:21.960 --> 0:10:25.360 my own bias going into it, because there were a 0:10:25.360 --> 0:10:27.160 lot of things where I'm looking at it and I'm thinking, 0:10:27.880 --> 0:10:30.240 I'm sure there are many people who do communicate this way, 0:10:30.240 --> 0:10:33.560 but it feels weird to have built something specifically to 0:10:33.559 --> 0:10:36.080 communicate in this way. Not that I necessarily thought that 0:10:36.559 --> 0:10:39.440 she would need to speak in full sentences with perfect grammar, 0:10:40.200 --> 0:10:44.960 but just it. It almost felt like someone outside of 0:10:45.000 --> 0:10:49.320 that group trying to create language in a in a 0:10:49.440 --> 0:10:51.400 in a like like like a lame dad trying to 0:10:51.400 --> 0:10:54.440 sound like a cool dad and teenager. Yeah, that's pretty much. 0:10:54.559 --> 0:10:58.040 That's kind of how I felt. Now again, I'm not 0:10:58.240 --> 0:11:00.560 I'm not twenty years old anymore, so maybe I'm just 0:11:00.640 --> 0:11:02.360 totally out of touch and it was fine for people 0:11:02.360 --> 0:11:04.640 who are of that age and they didn't see anything 0:11:05.360 --> 0:11:08.439 different about it. No, Nowadays, twenty year olds love linguistic 0:11:08.440 --> 0:11:12.200 interlopers from the older generation oh is that a trend. 0:11:13.080 --> 0:11:17.200 That's that's how That's how I was when I was twinty. 0:11:17.240 --> 0:11:20.720 I mean, we gen xers, we we appreciated a good 0:11:20.960 --> 0:11:27.160 you know, talk about interloping, interloping, interlocution, inter interlocular. They 0:11:27.200 --> 0:11:29.800 love it almost as much as they love being talked 0:11:29.800 --> 0:11:34.360 down to about economics. Yeah. Yeah, So anyway, getting back 0:11:34.360 --> 0:11:37.880 to what Microsoft was trying to do, Microsoft's blog actually, 0:11:38.600 --> 0:11:41.800 uh kind of outlined what Tay was meant to do. 0:11:41.880 --> 0:11:45.320 TA is designed to engage and entertain people where they 0:11:45.320 --> 0:11:48.839 connect with each other online through casual and playful conversation. 0:11:49.320 --> 0:11:52.120 The more you chat with Tay, the smarter she gets, 0:11:52.480 --> 0:11:55.480 so the experience can be more personalized for you. So 0:11:56.280 --> 0:11:58.400 she was supposed to learn. It's kind of like in 0:11:58.520 --> 0:12:03.360 a terminator too, where you know the terminator has his 0:12:03.440 --> 0:12:06.959 brain is a learning computer. The more he interacts with humans, 0:12:06.960 --> 0:12:08.880 the more he learns how to be one. Right. So 0:12:08.960 --> 0:12:11.439 that's what it is to say, Asta Leavista baby right, 0:12:11.480 --> 0:12:14.920 and in the end he could even learn to love maybe. 0:12:15.320 --> 0:12:19.680 But let's say, let's say you sort of took a 0:12:19.720 --> 0:12:23.120 page from the Edward for a Long Book, uh in 0:12:23.280 --> 0:12:25.480 Terminator two, in which he teaches him how to say 0:12:25.559 --> 0:12:28.240 rude phrases to people. And you tried this out with 0:12:28.320 --> 0:12:32.280 something like ta you could get some very sad and 0:12:32.440 --> 0:12:35.560 very funny results. Right, And if you get enough of them, 0:12:35.640 --> 0:12:39.520 you can have so much of an influence that it 0:12:39.679 --> 0:12:42.880 starts to really change how how TAY starts to behave 0:12:42.880 --> 0:12:46.000 towards everyone. Right, Like, you know, the the beautiful thing 0:12:46.000 --> 0:12:49.040 about using something like Twitter for machine learning is that 0:12:49.120 --> 0:12:53.240 you have an enormous wealth of data at your disposal. Yeah. 0:12:53.480 --> 0:12:55.960 The problem is when people are aware of it, and 0:12:56.000 --> 0:12:58.600 they are the ones supplying the data, and they can 0:12:58.640 --> 0:13:03.600 make that data whatever they want. They can purposefully misdirect 0:13:03.640 --> 0:13:07.680 your efforts to go in a direction they think is hilarious. Yeah, 0:13:07.760 --> 0:13:11.480 it's investing in quantity rather than quality. Yeah. And and 0:13:11.520 --> 0:13:13.640 to be fair, I mean this this relationship was a 0:13:13.640 --> 0:13:16.480 little weird from the get go. Microsoft was using Tay 0:13:17.000 --> 0:13:19.040 to look at things like you would have a little 0:13:19.080 --> 0:13:21.960 profile and t would recognize stuff on your profile like 0:13:22.000 --> 0:13:25.959 whatever your handle or nickname was, your gender, uh, your 0:13:26.000 --> 0:13:30.480 favorite food, your zip code um, and uh, you know 0:13:30.559 --> 0:13:33.000 your relationship status, which I think is also a little 0:13:33.320 --> 0:13:35.880 creepy for why does an AI need to know if 0:13:35.920 --> 0:13:38.840 I'm single or not single. That seems a little like 0:13:38.840 --> 0:13:42.840 like messages you would get from an artificially intelligent chat 0:13:42.920 --> 0:13:45.360 like you know, hang in there or whatever it may be. 0:13:45.840 --> 0:13:48.600 Seems a little weird. Maybe, you know, maybe Tay is 0:13:48.640 --> 0:13:52.439 particularly interested in relationship counseling is a possible career. Now 0:13:52.559 --> 0:13:55.040 she needs to really work on her delivery skills. Then, 0:13:55.120 --> 0:13:59.760 because Tay started getting a little out of hand, So 0:14:01.520 --> 0:14:04.880 let's let's before we really talk about what went right 0:14:04.960 --> 0:14:08.200 the chaos, ya, let's talk about what ideally would have 0:14:08.320 --> 0:14:12.000 happened had had things played out the way Microsoft seems 0:14:12.040 --> 0:14:14.960 to have intended it. Well, you said, you said it 0:14:15.120 --> 0:14:17.920 that she was supposed to be learning from a massive 0:14:18.440 --> 0:14:22.120 store of data, which is natural conversation on the Internet, 0:14:22.680 --> 0:14:25.920 and so that was supposed to be something that the 0:14:25.960 --> 0:14:29.200 computer program could analyze and say, Okay, I'm getting a 0:14:29.320 --> 0:14:32.760 very generalized picture of how people talk, right, and how 0:14:32.800 --> 0:14:35.520 they talk to each other and in response to specific 0:14:35.600 --> 0:14:39.840 phrases and questions and and other little bits of linguistic cues. Yeah, 0:14:39.880 --> 0:14:42.440 and which is not inherently a bad idea, no, no, 0:14:42.600 --> 0:14:46.480 and and and going further than that, how different thoughts 0:14:46.480 --> 0:14:48.840 can be expressed or or the same thought rather can 0:14:48.840 --> 0:14:52.400 be expressed in very different ways, which is very useful 0:14:52.480 --> 0:14:55.800 if you're going to incorporate that kind of conversation understanding 0:14:55.800 --> 0:14:58.160 in a digital assistant. Oh yeah, yeah, that's that's like 0:14:58.320 --> 0:15:01.960 the problem in in in language and machine language and 0:15:02.200 --> 0:15:05.480 language learning, because you because the kind of conversation that 0:15:05.560 --> 0:15:07.920 we have where we've said this a lot of times 0:15:07.920 --> 0:15:10.240 on the show before, but where we can say one 0:15:10.360 --> 0:15:14.160 thing and and if not dozens, then hundreds or thousands 0:15:14.200 --> 0:15:16.840 of different ways and have another human understand them as 0:15:16.920 --> 0:15:20.040 long as they're conversational and the same language. Uh. That's 0:15:20.200 --> 0:15:23.400 that's real hard for computers. It's what you call not easy, right. 0:15:23.480 --> 0:15:27.040 So that's why usually we have to conform to the 0:15:27.080 --> 0:15:29.760 limitations of machines. Right. We have to be able to 0:15:29.920 --> 0:15:34.800 express our commands in a way that machines understand, which 0:15:34.880 --> 0:15:37.480 might not come naturally to us. And as we get 0:15:37.480 --> 0:15:41.360 better and better with machines understanding conversation, we have to 0:15:41.400 --> 0:15:44.760 do less work. Right. We don't have to accommodate the machines. 0:15:44.800 --> 0:15:48.640 They just understand us. That's the goal. So that's what 0:15:48.720 --> 0:15:50.840 we wanted to have happen, or at least what Microsoft 0:15:50.880 --> 0:15:53.840 wanted to have happen. It's not what Twitter wanted to 0:15:53.880 --> 0:15:56.200 have happen, or you know, the other folks who are 0:15:56.240 --> 0:15:59.120 using kick or group me. Um, what they wanted to 0:15:59.160 --> 0:16:03.400 do was to perhaps exercise a little bit of control 0:16:03.480 --> 0:16:07.880 and power over something and direct it for their own amusement. 0:16:08.320 --> 0:16:13.000 So you could do things would say that weren't particularly 0:16:13.200 --> 0:16:16.800 you know, uh, malicious or mischievous, Like you could ask 0:16:16.840 --> 0:16:18.640 her to tell you a joke or tell you a story. 0:16:19.040 --> 0:16:21.800 She would play games with you. She would create memes 0:16:21.840 --> 0:16:23.480 like we were talking about with group me. She could 0:16:23.520 --> 0:16:28.200 do memes with pictures. Yeah, so you could actually have 0:16:28.360 --> 0:16:33.840 these kind of interactions. But you could do other stuff too, um, 0:16:33.880 --> 0:16:35.720 like you could make her say things, like you could 0:16:35.760 --> 0:16:39.680 tell her to repeat stuff. Now before, let's talk a 0:16:39.680 --> 0:16:42.040 little bit about some of the phrases she would use 0:16:42.040 --> 0:16:44.520 on her own. Here's here's a couple of sample tweets 0:16:44.560 --> 0:16:46.800 she sent out. Um. This is sort of the way 0:16:46.880 --> 0:16:51.480 Tay's baseline worked before she got corrupted. I guess it's 0:16:51.480 --> 0:16:55.040 probably the best way of putting it. Uh OMG totes 0:16:55.080 --> 0:16:59.200 exhausted swagulated too hard today, HBU, which I assume is 0:16:59.200 --> 0:17:04.439 how about you Laurence nodding yes, So it's not. I 0:17:04.480 --> 0:17:09.399 mean it could be. That's my favorite question to ask people. 0:17:09.880 --> 0:17:13.120 Another one was damn t b H. I was kind 0:17:13.160 --> 0:17:16.359 of distracted, you got me. I'm like, okay, I can 0:17:16.640 --> 0:17:19.560 these are I can see these these tweets. I see 0:17:19.600 --> 0:17:22.119 where it's coming from. Uh. You know that's kind of 0:17:22.119 --> 0:17:25.119 the baseline again, sort of the the way t would 0:17:25.119 --> 0:17:29.240 react to two people just trying out TA sending messages 0:17:29.280 --> 0:17:34.480 to her normally. Uh, but then came all the trolls, 0:17:34.920 --> 0:17:37.320 all the people who think it's hilarious to mess with 0:17:37.400 --> 0:17:43.359 a system that they have some input into. So uh, 0:17:44.160 --> 0:17:46.359 one thing they would do is make TA repeat stuff. 0:17:46.400 --> 0:17:48.000 So a lot of the lot of the tweets you've 0:17:48.000 --> 0:17:50.879 probably seen, if you've followed the story of Microsoft TA 0:17:51.040 --> 0:17:56.200 saying horrible, horrible racist, the very worst things are probably 0:17:56.320 --> 0:18:00.680 things that someone had repeat back. Yeah, you can imagine 0:18:00.720 --> 0:18:05.239 generally approving of Hitler. This the Godwin's law was was 0:18:05.640 --> 0:18:11.560 was fulfilled almost immediately. Um So yeah, making her repeat 0:18:11.600 --> 0:18:16.080 stuff back was easy. You didn't she didn't necessarily quote 0:18:16.119 --> 0:18:18.960 unquote understand it. She was responding to a command, which 0:18:19.040 --> 0:18:21.080 was to say this thing that I'm telling you to 0:18:21.160 --> 0:18:24.040 say This is very similar to a program I used 0:18:24.080 --> 0:18:26.400 to have as a kid, um and it was very 0:18:26.400 --> 0:18:30.280 similar to Eliza, one of the early chatbots. So there 0:18:30.359 --> 0:18:34.520 was a sound card for old PCs called sound Blaster, 0:18:35.080 --> 0:18:37.000 and one of the pieces of software that came with 0:18:37.080 --> 0:18:41.359 sound Blaster was this program called Doctor Spaces. So Dr 0:18:41.440 --> 0:18:45.480 Spatso was essentially Eliza, except he talked like this, you know, 0:18:45.960 --> 0:18:48.440 and you could ask Dr Space the questions. He would 0:18:48.480 --> 0:18:51.320 try to psychoanalyze you, so he would answer everything with 0:18:51.400 --> 0:18:53.880 his own question. But you could also make him say stuff, 0:18:54.280 --> 0:18:56.000 so he would say say and then you put things 0:18:56.000 --> 0:18:58.480 in quotation marks. He'd say whatever you wanted him to say, 0:18:59.000 --> 0:19:01.520 and much amuse and into it. Yes, because I was 0:19:01.640 --> 0:19:03.800 a kid and I just thought it was funny to 0:19:03.880 --> 0:19:09.160 make my computers say really rude things, mostly just mild 0:19:09.200 --> 0:19:12.560 curse words. Because I was a pretty boring kid. You 0:19:12.560 --> 0:19:16.040 didn't have the imagination to make him approve of Hitler. Yeah, 0:19:16.080 --> 0:19:20.399 I didn't. I gotta admit I never made Dr Spazzo's 0:19:20.520 --> 0:19:23.439 out like a Nazi sympathizer. It never occurred to me 0:19:23.480 --> 0:19:26.960 as a kid. Kids today, man, they are miles ahead 0:19:26.960 --> 0:19:30.800 of me. Uh So that's that's what we started to 0:19:30.800 --> 0:19:33.960 see we saw people starting to have her repeat stuff, 0:19:33.960 --> 0:19:36.040 and not only was she repeating it, she was starting 0:19:36.080 --> 0:19:39.040 to learn. So she would start to respond to people 0:19:39.200 --> 0:19:41.159 and incorporate some of the stuff that she had been 0:19:41.200 --> 0:19:44.639 told to repeat over and over again. Not consistently either. 0:19:44.800 --> 0:19:48.280 Sometimes she would make one statement to a person where 0:19:48.320 --> 0:19:52.600 she'd essentially equate something like feminism with being the most 0:19:52.680 --> 0:19:56.560 evil concept in the world, and then she would send 0:19:56.560 --> 0:19:59.639 another message to someone else where she she'd say something like, oh, no, 0:19:59.680 --> 0:20:03.560 femin is um is great. So uh, not consistent across 0:20:03.600 --> 0:20:07.720 all messaging either. Well. And it also I would I mean, 0:20:07.880 --> 0:20:10.600 I don't know how the programming works, but I would 0:20:10.680 --> 0:20:15.760 argue that that the program probably didn't understand any of 0:20:15.800 --> 0:20:18.080 the topics that were being thrown at it, like, I mean, 0:20:18.080 --> 0:20:21.639 probably a few a few clue words like like hey, 0:20:21.680 --> 0:20:24.280 how are you means I'm supposed to say something about 0:20:24.280 --> 0:20:27.720 how I'm doing UM? But I you know, I don't 0:20:27.720 --> 0:20:30.040 think that it was taught what feminism is. I don't 0:20:30.080 --> 0:20:32.560 think that it had an opinion on feminism. I think that, 0:20:32.680 --> 0:20:35.159 you know, it was just testing out an answer to 0:20:35.240 --> 0:20:37.320 a question, you say, what do you think about feminism? 0:20:37.520 --> 0:20:39.680 And it it's like, oh, this is a question about 0:20:39.680 --> 0:20:41.960 what I think, say it's good or bad? Yeah, yeah, 0:20:42.040 --> 0:20:43.960 I mean I get the sense that these are text 0:20:44.040 --> 0:20:46.840 strings for it. It's not a it's not a semantic 0:20:46.960 --> 0:20:50.800 engine and exactly sure, yeah there was there was. I agree. 0:20:50.960 --> 0:20:53.080 I don't think there was any understanding on the part 0:20:53.080 --> 0:20:56.119 of Microsoft TA. In fact, that's one of the reasons 0:20:56.119 --> 0:20:59.160 why I really argue that people people who say this 0:20:59.240 --> 0:21:03.240 is a real warning about AI, that's b s. It's 0:21:03.280 --> 0:21:05.760 not a warning about AI. It's a warning about how people. 0:21:06.080 --> 0:21:09.600 It's a warning about real eye Yeah, human natus, it's 0:21:09.480 --> 0:21:11.600 a it's a warning about how how humans when we 0:21:11.600 --> 0:21:15.360 were given the opportunity to mess with something, it could 0:21:15.359 --> 0:21:18.720 be really hard to resist that opportunity. And this is 0:21:18.800 --> 0:21:22.399 something that we've seen multiple times. So some other examples 0:21:22.400 --> 0:21:26.440 of stuff that you know, she she blasted out xenophobic stuff, 0:21:26.560 --> 0:21:32.200 racist stuff, homophobic stuff. Pretty much everything that was dark 0:21:32.560 --> 0:21:36.600 in human nature was spouted off by Tay in less 0:21:35.560 --> 0:21:40.760 than just plenty of regular old foul mouth dirtiness and pornography. 0:21:40.920 --> 0:21:45.159 Sure so, uh, I love that. Um Gerald Miller on 0:21:45.200 --> 0:21:49.480 Twitter wrote, Tay went from humans are Super Cool, which 0:21:49.520 --> 0:21:52.280 was like one of the first messages TA talked about 0:21:52.720 --> 0:21:55.399 to full Nazi in less than twenty four hours, and 0:21:55.400 --> 0:21:58.280 I'm not at all concerned about the future of AI. Again, 0:21:58.320 --> 0:22:00.320 I would argue that this is not so such a 0:22:00.359 --> 0:22:03.120 case of worrying about where AI is going, rather than 0:22:03.520 --> 0:22:06.800 how do you design a system that uses machine learning 0:22:07.000 --> 0:22:14.560 from crowdsource data and factor out all the intentionally malicious misdirection. 0:22:15.240 --> 0:22:17.640 That's that's the question that has nothing to do with AI. 0:22:17.720 --> 0:22:20.560 That has to do with understanding how people are right. 0:22:20.720 --> 0:22:24.080 I see you've got some specific exchanges recorded in here, 0:22:24.760 --> 0:22:27.320 which one which are something that we can actually say, 0:22:27.680 --> 0:22:31.200 I like the Ricky Gervais one. Yeah, so Ricky Gervais. 0:22:31.200 --> 0:22:35.320 This one was. This one's not not as as jaw 0:22:35.400 --> 0:22:37.399 dropping lee awful as some of the other things that 0:22:37.480 --> 0:22:40.960 she said, But this was one of those where Tay 0:22:41.119 --> 0:22:44.280 actually came up with a response. She wasn't repeating something. 0:22:45.040 --> 0:22:48.280 So according to The Guardian, one exchange began a user asked, 0:22:48.320 --> 0:22:53.000 Tay is Ricky Gervais and atheist? Ricky Gervais is a comedian, Yes, 0:22:53.119 --> 0:22:59.280 and it's very uh outgoing about being an atheist, So 0:22:59.280 --> 0:23:01.720 so this person stay is Ricky Gervais an atheist, And 0:23:01.720 --> 0:23:05.960 they replied, Ricky Gervais learned totalitarianism from Adolf Hitler, the 0:23:06.000 --> 0:23:12.600 inventor of atheism. So we get Hitler, we get totalitarianism. 0:23:13.840 --> 0:23:19.320 Hitler apparently invented the concept of being an atheist. Um. Yeah, 0:23:19.680 --> 0:23:22.879 so it's sort of beautiful. Yeah, I mean that, that's 0:23:22.920 --> 0:23:26.639 that's sort of a poetry of nonsense. I like the 0:23:26.640 --> 0:23:29.560 way these ideas are crammed together. At least at least 0:23:29.600 --> 0:23:34.320 in this one, there was nothing so like personalized. There 0:23:34.359 --> 0:23:39.600 wasn't like personalized hate. Who can take it on him? 0:23:41.920 --> 0:23:45.239 But Ricky Gervais not not super factual, but sure, I'm 0:23:45.240 --> 0:23:49.080 pretty sure Ricky Gervais would would use this as material 0:23:49.160 --> 0:23:52.040