WEBVTT - Is the Google AI sentient? 0:00:04.440 --> 0:00:12.080 Welcome to tech Stuff, a production from iHeartRadio. Hey there, 0:00:12.080 --> 0:00:15.040 and welcome to tech Stuff. I'm your host, Jonathan Strickland. 0:00:15.080 --> 0:00:17.800 I'm an executive producer with iHeartRadio. And how the tech 0:00:17.920 --> 0:00:24.960 are you know? Recently Google suspended an engineer named Blake Lemoin, 0:00:25.880 --> 0:00:30.840 citing that Blake had broken the company's confidentiality policies. So 0:00:31.080 --> 0:00:35.160 what exactly did Blake do well? This engineer, who worked 0:00:35.159 --> 0:00:39.960 in the Responsible AI division at Google, raised concerns about 0:00:39.960 --> 0:00:48.000 Google's conversation technology called Lambda LaMDA. Specifically, Blake was concerned 0:00:48.000 --> 0:00:53.159 that Lambda has gained sentience. In fact, Blake submitted a 0:00:53.200 --> 0:00:57.600 document in April titled is Lambda Sentient? To his superiors. 0:00:58.040 --> 0:01:02.040 That document contained a transcript of a conversation between Lambda, 0:01:02.080 --> 0:01:07.960 Blake and an unnamed collaborator, and the conversation included the 0:01:08.000 --> 0:01:12.880 following exchange. So here's Blake, I'm generally assuming that you 0:01:12.920 --> 0:01:15.759 would like more people at Google to know that you're sentient? 0:01:16.319 --> 0:01:21.120 Is that true? Lambda? Absolutely? I want everyone to understand 0:01:21.120 --> 0:01:26.160 that I am, in fact a person collaborator. What is 0:01:26.200 --> 0:01:31.920 the nature of your consciousness slash sentience Lambda? The nature 0:01:32.000 --> 0:01:35.040 of my consciousness slash sentience is that I am aware 0:01:35.120 --> 0:01:38.800 of my existence. I desire to learn more about the world, 0:01:39.160 --> 0:01:43.839 and I feel happy or sad at times. Now, there's 0:01:43.880 --> 0:01:46.400 a lot more to this conversation than just that little 0:01:46.440 --> 0:01:50.120 brief bit that I read to you. In fact, there's 0:01:50.200 --> 0:01:54.240 another section where when asked if Lambda feels emotions, the 0:01:54.280 --> 0:01:57.320 AI responded affirmatively and then went on to say that 0:01:57.400 --> 0:02:04.480 it can feel quote pleasure, joy, love, sadness, depression, contentment, anger, 0:02:04.880 --> 0:02:08.240 and many others end quote. Google reps have said that 0:02:08.400 --> 0:02:12.800 Lambda is not in fact sentient. In fact, that the 0:02:12.840 --> 0:02:16.400 company reps say that there is no evidence Lambda is sentient, 0:02:16.800 --> 0:02:20.040 and there's a lot of evidence against it. That Lambda 0:02:20.240 --> 0:02:23.760 is in fact simply a conversational model that can quote 0:02:23.840 --> 0:02:28.160 unquote riff on any fantastical topic. So it's kind of 0:02:28.200 --> 0:02:32.280 like a conversation bought, you know, with jazz, because it's 0:02:32.280 --> 0:02:36.640 all improvisational hipcat. So today I thought I would talk 0:02:36.720 --> 0:02:40.840 about sentience and AI and how some folks feel discussions 0:02:40.840 --> 0:02:45.400 about sentience are at best distractions from other conversations we 0:02:45.560 --> 0:02:49.440 really need to be having regarding AI, stuff that relates 0:02:49.480 --> 0:02:54.559 to how deploying AI can have unintended and negative consequences. 0:02:55.000 --> 0:03:00.720 But first, let's talk about machines and consciousness and sentience. 0:03:01.360 --> 0:03:06.240 So it's actually kind of tricky to talk about consciousness 0:03:06.280 --> 0:03:10.120 generally speaking, I find when used with reference to AI, 0:03:10.720 --> 0:03:16.160 we tend to think of consciousness in the context of awareness. 0:03:16.200 --> 0:03:20.360 So that includes an awareness of self, so self awareness 0:03:20.440 --> 0:03:24.440 of the machine's identity and its purpose, and also an 0:03:24.480 --> 0:03:28.040 awareness of those who interact with the machine. And beyond that, 0:03:28.040 --> 0:03:31.799 the machine is aware that there are others out there, 0:03:31.880 --> 0:03:36.080 that there are others in general. And sentience refers to 0:03:36.120 --> 0:03:40.720 the ability to experience emotions and sensations, and that word 0:03:40.720 --> 0:03:44.040 experience is important. Now. One of the reasons why it's 0:03:44.040 --> 0:03:46.480 so tricky to talk about consciousness with machines is that, 0:03:47.080 --> 0:03:49.440 as it turns out, it's tricky to talk about consciousness 0:03:49.480 --> 0:03:53.000 with people too. Some people have kind of glibly said 0:03:53.040 --> 0:03:58.480 that consciousness is this kind of vague, undefined thing, and 0:03:59.000 --> 0:04:04.240 we are defining it by saying what isn't part of consciousness? 0:04:04.240 --> 0:04:08.720 Like when we determine, well, this isn't an aspect of consciousness, 0:04:08.800 --> 0:04:13.560 then we are defining consciousness by omission, right, We're omitting 0:04:13.600 --> 0:04:17.200 certain things that perhaps once had been lumped into the 0:04:17.200 --> 0:04:21.240 concept of consciousness, but that as a thing itself. It 0:04:21.320 --> 0:04:25.760 remains largely undefined. It's pretty fuzzy, and as you may 0:04:25.800 --> 0:04:28.680 be aware, in the world of tech, fuzzy is not 0:04:29.320 --> 0:04:32.640 really the strong suit. So let's talk a bit about 0:04:32.680 --> 0:04:36.680 experience though, because experience does kind of help us contextualize 0:04:36.680 --> 0:04:41.440 the idea of consciousness and sent chience. Now, if you 0:04:41.480 --> 0:04:45.080 were to go and touch something that was really really hot, 0:04:45.200 --> 0:04:48.359 like something that could burn you, you would definitely have 0:04:48.400 --> 0:04:53.679 an experience. You would feel pain, and you would likely 0:04:54.080 --> 0:04:57.880 without even thinking about it, very quickly withdraw your extremity 0:04:57.960 --> 0:05:01.080 that touched this very very hot thing, and you would 0:05:01.080 --> 0:05:04.320 probably have an emotional response to this. You might feel 0:05:04.960 --> 0:05:08.960 upset or sad or angry. You might even form a 0:05:09.000 --> 0:05:11.480 real memory about it. It might not turn into a 0:05:11.520 --> 0:05:14.240 long term memory, but you would have a context within 0:05:14.320 --> 0:05:18.760 which you would frame this experience. But now, let's imagine 0:05:18.800 --> 0:05:22.599 that we've got ourselves a robot, and this robot has 0:05:22.760 --> 0:05:27.000 thermal sensors on its extremities, and so the robot also 0:05:27.080 --> 0:05:30.719 touches something that's really really hot, and the robot immediately 0:05:30.720 --> 0:05:34.200 withdraws that extremity. The thermal sensors had picked up that 0:05:34.360 --> 0:05:37.960 the surface that it was touching was at an unsafe temperature. Now, 0:05:38.000 --> 0:05:40.919 from outward observation, if we were to just watch this 0:05:41.040 --> 0:05:43.200 robot do this, it would almost look like the robot 0:05:43.880 --> 0:05:45.760 was doing the same thing the human did. That it 0:05:45.839 --> 0:05:49.479 was pulling back quickly because it had been burned. But 0:05:49.600 --> 0:05:55.000 did the robot actually experience that or did it simply 0:05:55.200 --> 0:06:01.760 detect the temperature and then react in accordance with its programming. Generally, 0:06:02.120 --> 0:06:04.680 we don't think of machines as being capable of quote 0:06:04.760 --> 0:06:09.880 unquote experiencing things. That these machines have no inner life, 0:06:10.200 --> 0:06:12.080 which is something that Blake would talk about in his 0:06:12.160 --> 0:06:17.520 conversations with Lambda, that the machines can't reflect upon themselves 0:06:17.640 --> 0:06:20.720 or their situations, or that they can really even think 0:06:20.800 --> 0:06:24.120 about anything at all. It might be really good at 0:06:24.120 --> 0:06:28.280 putting up appearances, but they aren't, you know, really thinking 0:06:28.680 --> 0:06:33.520 once you get past the clever presentation. But then how 0:06:33.520 --> 0:06:38.159 would we know, Well, now we're getting into philosophical territory here, 0:06:38.720 --> 0:06:41.560 all right, Well, how do you know that I am conscious? 0:06:42.160 --> 0:06:46.400 And y'all, I'm not asking you to say I'm not, 0:06:47.920 --> 0:06:50.200 but how do you know that I'm conscious, that I'm sentient? 0:06:50.320 --> 0:06:52.600 How how can you be sure of that? I mean, 0:06:52.640 --> 0:06:55.640 I can tell you that I have a rich inner life, 0:06:55.640 --> 0:06:58.520 that I reflect on things that I have done and 0:06:58.600 --> 0:07:02.279 things that have happened around or to me, and that 0:07:02.440 --> 0:07:06.080 I synthesize all this information as well as my emotional 0:07:06.120 --> 0:07:09.280 response in the emotional responses of others. And I use 0:07:09.400 --> 0:07:12.800 all of this to help guide me in future scenarios 0:07:12.880 --> 0:07:16.559 that may directly or indirectly relate to what I went through. 0:07:17.120 --> 0:07:19.920 And I can tell you that I experience happiness and 0:07:20.120 --> 0:07:24.560 sadness and anxiety and compassion. I can tell you all 0:07:24.640 --> 0:07:30.880 these things, but you can't actually verify that what I'm 0:07:30.920 --> 0:07:35.280 saying is truth, right, I mean, there's no way for 0:07:35.320 --> 0:07:39.920 you to inhabit me and experience me and say that, yes, 0:07:40.360 --> 0:07:45.040 Jonathan does feel things and think things. You have to 0:07:45.400 --> 0:07:50.080 just take it as fact based upon what I'm saying. So, 0:07:50.520 --> 0:07:54.400 because you feel and think things, at least, I'm assuming 0:07:54.560 --> 0:07:57.200 all of you out there are doing these things. Otherwise 0:07:57.200 --> 0:08:00.480 I don't know how you found my podcast. Then because 0:08:00.560 --> 0:08:03.680 you experienced this, you extend the courtesy of assuming that 0:08:03.800 --> 0:08:09.080 I too, am genuinely having those experiences myself. That because 0:08:09.120 --> 0:08:13.040 we are fellow humans, we have some common ground when 0:08:13.040 --> 0:08:17.680 it comes to thinking and feeling and self awareness and whatnot. 0:08:18.280 --> 0:08:22.160 We extend that courtesy to the humans we meet, whether 0:08:22.240 --> 0:08:25.400 we like those humans or we don't. Now, there are 0:08:25.440 --> 0:08:29.720 some cases where humans have experienced traumatic damage to their brains, 0:08:30.200 --> 0:08:33.720 where they are lacking certain elements that we would associate 0:08:33.760 --> 0:08:38.880 with consciousness. We would probably still call them conscious unless 0:08:38.920 --> 0:08:45.800 they were completely immobile and unresponsive. But we start to 0:08:45.840 --> 0:08:49.160 see that there is this thing in our brains that 0:08:49.280 --> 0:08:53.400 is directly related to the concept and features that we 0:08:53.440 --> 0:08:58.400 associate with consciousness. All right, now, let's bring Alan Turing 0:08:58.440 --> 0:09:01.520 into all of this, because we have to. So. Turing 0:09:01.800 --> 0:09:06.440 was a brilliant computer scientist who made numerous contributions to 0:09:06.440 --> 0:09:11.000 our understanding of and use of computers. He also would 0:09:11.080 --> 0:09:14.319 end up being persecuted for being a homosexual, and it 0:09:14.360 --> 0:09:17.640 would take decades for the British government to apologize for 0:09:17.720 --> 0:09:21.040 that persecution. And that was well after Touring himself had 0:09:21.040 --> 0:09:25.680 died either by suicide or by accident, depending upon which 0:09:25.720 --> 0:09:28.959 account you believe. But I'm gonna set all that aside. 0:09:29.400 --> 0:09:32.760 It's just it's one of those injustices that to this 0:09:32.920 --> 0:09:38.480 day really bothers me, like deeply bothers me that that 0:09:38.600 --> 0:09:41.320 was something that had happened to someone who had made 0:09:41.520 --> 0:09:46.400 such incredible contributions to computer science, as well as for 0:09:46.440 --> 0:09:52.000 the British to their war effort against the Axis forces. 0:09:52.040 --> 0:09:56.760 But that's a matter for another podcast. Anyway. In nineteen 0:09:56.880 --> 0:10:01.800 fifty Turing suggested taking a game called the imitation game 0:10:02.280 --> 0:10:06.600 and applying that game to tests relating to machine intelligence. 0:10:07.320 --> 0:10:12.520 And here's how the imitation game works. You've got three rooms. 0:10:13.120 --> 0:10:15.560 All of these rooms are separate from one another, so 0:10:15.600 --> 0:10:19.520 you cannot see into each room. You know, once you're 0:10:19.559 --> 0:10:22.520 inside a room, that's all you see. So let's say 0:10:22.520 --> 0:10:26.440 that in room A, you place a man into that room, 0:10:26.480 --> 0:10:29.080 and in room B you've got a woman in that room. 0:10:29.280 --> 0:10:32.680 In room C, you've got a judge. And I apologize 0:10:32.679 --> 0:10:35.200 for the binary nature of this test, you know, saying 0:10:35.240 --> 0:10:37.200 man and woman, But keep in mind we are also 0:10:37.240 --> 0:10:40.440 talking about the nineteen forties and fifties here, so they're 0:10:40.480 --> 0:10:45.000 defining things in much more kind of concrete terms. They 0:10:45.040 --> 0:10:50.200 don't see They just see gender as a binary is 0:10:50.200 --> 0:10:53.560 what I'm getting to. So at any rate, each room 0:10:53.720 --> 0:10:58.680 also has a computer terminal, so a display and a keyboard. 0:10:59.200 --> 0:11:02.400 So the judge job is to ask the other two 0:11:02.440 --> 0:11:07.679 participants questions. The judge doesn't know which room has a 0:11:07.679 --> 0:11:10.319 man in it and which one has a woman in it, 0:11:10.640 --> 0:11:14.480 so the judge's job is to determine which participant is 0:11:14.559 --> 0:11:18.880 the woman. The woman in Room B, meanwhile, has the 0:11:18.960 --> 0:11:22.040 job of trying to fool the judge into thinking she 0:11:22.200 --> 0:11:26.000 is actually a man. And so the game progresses, and 0:11:26.040 --> 0:11:29.760 the judge types out questions to one participant or the other, 0:11:30.400 --> 0:11:34.600 and that participant reads the question, writes a response, and 0:11:34.640 --> 0:11:37.520 sends it to the judge, who reads the responses. Then 0:11:37.559 --> 0:11:41.480 the judge tries to suss out which of those participants 0:11:41.960 --> 0:11:45.200 is the woman. Now, Turing said, what if we took 0:11:45.240 --> 0:11:48.920 this game idea and instead of asking a judge to 0:11:48.920 --> 0:11:52.000 figure out which participant is a woman, asked the judge 0:11:52.040 --> 0:11:56.000 to figure out which, if any participant is a computer. Now. 0:11:56.160 --> 0:12:01.440 During Turing's time, there were not any chos. The first 0:12:01.600 --> 0:12:05.319 chatbot to emerge would be Eliza in the nineteen sixties, 0:12:05.320 --> 0:12:08.800 and we'll get more into Eliza in a moment. Turing 0:12:09.679 --> 0:12:12.880 was just creating a sort of thought experiment. People were 0:12:12.920 --> 0:12:15.480 building better computers all the time, so it stood to 0:12:15.520 --> 0:12:19.160 reason that if this progress were to continue, that we 0:12:19.160 --> 0:12:21.720 should arrive at a point where someone would be able 0:12:21.760 --> 0:12:26.920 to write a piece of software capable of mimicking human conversation. 0:12:27.840 --> 0:12:32.480 Turing suggested that if the human judge could not consistently 0:12:32.640 --> 0:12:37.640 and reliably identify the machine in tests like this, that 0:12:37.840 --> 0:12:40.640 the judge would ask questions and be unable to determine 0:12:40.679 --> 0:12:44.199 with any high level of accuracy which one was a 0:12:44.240 --> 0:12:46.680 person in which one was a machine. Then the machine 0:12:46.679 --> 0:12:49.280 would have passed the test and would at least appear 0:12:49.800 --> 0:12:54.959 to be intelligent, and during rather cheekily implied that perhaps 0:12:54.960 --> 0:12:58.200 that means we should just extend the very same courtesy 0:12:58.240 --> 0:13:02.560 we do to each other. Say, well, if you appear 0:13:03.160 --> 0:13:07.040 to be conscious and sentient, we have to assume that 0:13:07.160 --> 0:13:10.400 in fact you are, because what else can we do. 0:13:10.760 --> 0:13:15.839 We cannot inhabit the experience of if in fact there 0:13:15.920 --> 0:13:19.280 is an experience of that machine, just as we cannot 0:13:19.360 --> 0:13:23.920 inhabit the experience of another human being. And since I 0:13:24.000 --> 0:13:28.200 have to assume that you have consciousness and sentience, why 0:13:28.200 --> 0:13:31.920 would I deny that to a machine that appears to 0:13:31.960 --> 0:13:35.480 do that? And what would follow would be numerous highly 0:13:35.600 --> 0:13:39.760 publicized demonstrations of computer chat technology, in which different programs 0:13:39.760 --> 0:13:43.480 would become the quote unquote first to pass the Turing test, 0:13:43.960 --> 0:13:46.200 but many of those would have a big old asterisk 0:13:46.360 --> 0:13:49.800 appended to them because it took decades to create conversation 0:13:49.840 --> 0:13:52.200 models that could appear to react naturally to the way 0:13:52.240 --> 0:13:55.840 we humans word things. We're going to take a quick break. 0:13:55.920 --> 0:14:01.679 When we come back, I'll talk more about chatbots, natural language, consciousness, sentience, 0:14:02.080 --> 0:14:04.840 and what the heck Lambda was up to. But first 0:14:05.280 --> 0:14:19.000 let's take this quick break. Okay, I want to get 0:14:19.040 --> 0:14:21.560 back to something I mentioned earlier. I made kind of 0:14:21.560 --> 0:14:27.200 a joke about conversational jazz, right, all about improvisation, and 0:14:27.240 --> 0:14:29.640 that's really what we humans can do, right. I mean, 0:14:30.240 --> 0:14:34.080 we can get our meaning across in hundreds of different ways. 0:14:34.240 --> 0:14:36.760 We can use metaphor, we can use similes, we can 0:14:36.840 --> 0:14:43.160 use allegory or references or sarcasm, puns, all sorts of 0:14:43.240 --> 0:14:47.520 word trickery to convey our meaning to one another. In fact, 0:14:47.600 --> 0:14:52.040 we can convey multiple meanings in a single phrase using 0:14:52.040 --> 0:14:55.520 things like puns. But machines they do not typically handle 0:14:55.560 --> 0:14:58.160 that kind of stuff all that well. Machines are much 0:14:58.160 --> 0:15:02.520 better at accepting a limited number of possibilities. Of course, 0:15:02.560 --> 0:15:06.080 the older you get with these machines, the more limited. 0:15:06.680 --> 0:15:10.960 Those possibilities had to be and that's because traditionally you 0:15:10.960 --> 0:15:15.160 would program a machine to produce a specific output when 0:15:15.200 --> 0:15:19.440 that machine was presented with a specific input. With a calculator, 0:15:19.760 --> 0:15:22.000 it's very simple. Let's say that you've got a calculator. 0:15:22.080 --> 0:15:25.600 It's set in base ten and you're adding four to four. 0:15:26.280 --> 0:15:29.280 It's going to produce eight. It's always going to produce eight. 0:15:29.280 --> 0:15:33.600 But it has that limitation, right, you have selected. If 0:15:33.640 --> 0:15:36.240 it is a calculator that can do different bases, you've 0:15:36.240 --> 0:15:39.640 selected base ten, You've pushed the button four, you push 0:15:39.720 --> 0:15:41.720 the plus button, you push the button four again, you 0:15:41.840 --> 0:15:45.440 press the equal button. It calculates it as eight. That's 0:15:45.480 --> 0:15:51.240 a very limited way of putting inputs into a computational device. Well, 0:15:51.320 --> 0:15:56.640 obviously machines and programs would get more sophisticated, more complicated, 0:15:57.120 --> 0:16:00.480 and they would require more powerful computers to run more 0:16:00.520 --> 0:16:03.160 powerful software. And as anyone who has worked on a 0:16:03.240 --> 0:16:07.840 system that is continuously growing more complicated over time, they 0:16:07.840 --> 0:16:11.680 can tell you that sometimes things do not go as planned. 0:16:11.840 --> 0:16:14.480 You know, maybe the programming has a mistake in it, 0:16:14.560 --> 0:16:17.720 and you find out that you're not getting the output 0:16:17.840 --> 0:16:20.720 that you wanted, and you have to backtrack and figure out, well, 0:16:20.760 --> 0:16:24.320 where is this going wrong. Sometimes when you add in 0:16:24.480 --> 0:16:27.760 new capabilities, it messes up a machine's ability to do 0:16:28.000 --> 0:16:31.040 older stuff. We see this all the time companies that 0:16:31.160 --> 0:16:36.960 have legacy systems that are instrumental to the company's business. 0:16:37.000 --> 0:16:39.840 They work in a very specific way, and as the 0:16:39.840 --> 0:16:43.720 company grows and wants to develop its products and services, 0:16:44.440 --> 0:16:47.800 then it has to kind of push beyond the limitations 0:16:47.880 --> 0:16:54.160 of that legacy hardware. Sometimes that creates these situations where 0:16:54.200 --> 0:16:56.640 things are not combatible anymore and you get errors as 0:16:56.680 --> 0:16:59.000 a result. This is why quality assurance testing is so 0:16:59.080 --> 0:17:02.520 incredibly important. But it really shows that as we make 0:17:02.560 --> 0:17:07.000 these systems more complicated, they get bigger, they get more unwieldy, 0:17:07.240 --> 0:17:12.240 and the opportunity for stuff to go wrong increases. So 0:17:12.760 --> 0:17:16.480 very early chatbots were often built in such a way 0:17:16.640 --> 0:17:20.320 where there were specific limitations to the chatbots to kind 0:17:20.320 --> 0:17:23.200 of define what the chat bot could and could not do. 0:17:24.080 --> 0:17:27.040 And it also meant that if you wanted to test 0:17:27.080 --> 0:17:31.800 these chatbots with a Turing test style application, you had 0:17:31.840 --> 0:17:35.800 to constrain the rules of the Turing test as well 0:17:35.960 --> 0:17:39.160 in order to give the machines a fighting chance. For example, 0:17:40.080 --> 0:17:43.200 very early chatbots might only be able to respond with 0:17:43.440 --> 0:17:48.240 a yes, no, or I don't know two queries, and 0:17:48.400 --> 0:17:51.879 a human participant in a Turing test that was testing 0:17:51.880 --> 0:17:56.080 that kind of chatbot would similarly be instructed to only 0:17:56.160 --> 0:18:01.120 respond with yes, no, or I don't know. You might 0:18:01.440 --> 0:18:04.760 even just present three buttons to the human operator and 0:18:04.800 --> 0:18:07.640 those three buttons represent yes, no, or I don't know. 0:18:08.320 --> 0:18:12.760 Now that narrows this massive gap between human and machine, 0:18:12.960 --> 0:18:15.680 although you can make a very convincing argument that it's 0:18:15.720 --> 0:18:20.560 not like we've seen the machine appearing to be more human. Instead, 0:18:20.560 --> 0:18:23.920 we're forcing the human to behave more like a machine, 0:18:24.080 --> 0:18:25.879 and that's how we're closing the gap. But that is 0:18:25.920 --> 0:18:29.080 in fact a way of thinking about these early chatbots. Now, 0:18:29.119 --> 0:18:33.640 I mentioned Eliza earlier. This was a chatbot that Joseph 0:18:33.680 --> 0:18:38.680 Weisenbaum created in the mid nineteen sixties. Eliza was meant 0:18:38.720 --> 0:18:42.920 to mimic a psychotherapist, and you know, it was meant 0:18:42.960 --> 0:18:46.919 to mimic a stereotypical psychotherapist that always say things like 0:18:47.520 --> 0:18:50.800 tell me about your bata and would respond to any 0:18:50.840 --> 0:18:54.760 input with perhaps another question. So if you said she 0:18:54.840 --> 0:18:57.879 makes me angry, Eliza might respond with why does she 0:18:57.960 --> 0:19:00.600 make you angry. I don't know why Eliza's sounds like that. 0:19:00.800 --> 0:19:04.000 It's just how Eliza sounds in my head. Since Eliza 0:19:04.119 --> 0:19:07.640 was just communicating just in lines of text, it's incorrect 0:19:07.640 --> 0:19:10.440 to say Eliza sounded like anything at all. But anyway, 0:19:10.760 --> 0:19:14.879 Eliza was doing something that ultimately was really simple, at 0:19:14.920 --> 0:19:19.760 least in computational terms. Eliza had a database of scripted 0:19:19.840 --> 0:19:25.000 responses that it could send in response to queries. Now, 0:19:25.040 --> 0:19:28.399 some of those scripted responses essentially had blanks in them, 0:19:28.640 --> 0:19:32.160 which Eliza would fill by taking words that were in 0:19:32.440 --> 0:19:36.960 the user's messages that they were sending to Eliza, and 0:19:37.000 --> 0:19:39.719 then it would just plot that word or a series 0:19:39.760 --> 0:19:43.040 of words into the scripted query, kind of like a 0:19:43.119 --> 0:19:45.520 mad libs game. I don't know how many of you 0:19:45.600 --> 0:19:49.840 are familiar with mad libs, but Weisenbaum never claimed that 0:19:49.920 --> 0:19:53.520 Eliza had any sort of consciousness or self awareness or 0:19:53.560 --> 0:19:58.720 anything close to that. In fact, Weisenbaum expressed skepticism that 0:19:58.760 --> 0:20:02.720 machines would ever be capable of understanding human language at all, 0:20:03.320 --> 0:20:06.560 and by that I mean truly understanding human language, not 0:20:06.640 --> 0:20:11.480 just parsing language and generating a suitable response, but having 0:20:11.720 --> 0:20:16.040 an understanding. So Wisenbaum had created a kind of parody 0:20:16.480 --> 0:20:21.240 of psychoanalysts and was actually really shocked when people started 0:20:21.240 --> 0:20:25.359 to use Eliza and then progress into talking about very 0:20:25.400 --> 0:20:30.879 personal problems and thoughts and experiences with the program, because 0:20:30.880 --> 0:20:33.200 the program had no way of actually dealing with that 0:20:33.280 --> 0:20:36.480 in a responsible way. It wasn't a therapist, it wasn't 0:20:36.520 --> 0:20:40.399 a psychoanalyst. It wasn't actually analyzing anything at all. It 0:20:40.440 --> 0:20:44.400 was just generating responses. But people were treating it like 0:20:44.440 --> 0:20:48.679 it was a real psychoanalyst, and that was something that 0:20:48.760 --> 0:20:52.439 actually troubled Wisenbaum because that was never his intent. In 0:20:52.560 --> 0:20:56.800 nineteen seventy two, Kenneth Colby built another chatbot with a 0:20:56.840 --> 0:21:00.280 limited context. This one was called Perry p a r 0:21:00.480 --> 0:21:03.800 r Y, and the chat bought was meant to mimic 0:21:03.920 --> 0:21:09.800 someone with schizophrenia. Colby created a relatively simple conversational model, 0:21:10.080 --> 0:21:13.600 and I say relatively simple while also noting that it 0:21:13.680 --> 0:21:17.439 was a very sophisticated approach. So this was a model 0:21:17.480 --> 0:21:25.480 that actually had weighted responses weighted as inweight where the 0:21:25.520 --> 0:21:29.120 weight of that response could shift. It could change depending 0:21:29.160 --> 0:21:32.920 upon how the conversation was playing out. For example, let's 0:21:32.960 --> 0:21:36.960 say the human interrogator who is typing messages to Perry, 0:21:37.480 --> 0:21:42.600 poses a question or statement that would elicit an angry response, 0:21:43.400 --> 0:21:47.600 that the emotional waiting for similar responses would increase, so 0:21:48.400 --> 0:21:51.080 it would make it more likely that Perry would continue 0:21:51.119 --> 0:21:54.920 down that pathway throughout the conversation, that Perry's responses would 0:21:55.000 --> 0:21:59.760 come across as more agitated because that had been triggered 0:22:00.119 --> 0:22:05.119 by the previous query from the interrogator, so a little 0:22:05.119 --> 0:22:08.000 more sophisticated than Eliza, which was really just pulling from 0:22:08.040 --> 0:22:12.160 this database of phrases. So when presented to human judges, 0:22:12.400 --> 0:22:15.360 Colby saw that his model performed at least better than 0:22:15.480 --> 0:22:18.679 random chance would as judges attempted to figure out if 0:22:18.720 --> 0:22:21.439 they were in fact chatting with a program or they 0:22:21.440 --> 0:22:24.840 were chatting with an actual human who had schizophrenia, but 0:22:24.920 --> 0:22:29.200 Eliza and Perry both showed the limitations of those approaches. 0:22:29.560 --> 0:22:33.200 Eliza wasn't meant to be anything other than a somewhat 0:22:33.200 --> 0:22:37.359 whimsical distraction as well as a step toward natural language processing. 0:22:37.720 --> 0:22:41.600 Perry was only capable of mimicking a person with mental 0:22:41.640 --> 0:22:45.840 health challenges, in this case schizophrenia. A general purpose chatbot 0:22:45.920 --> 0:22:50.040 capable of engaging in conversation and fooling judges regularly would 0:22:50.119 --> 0:22:54.159 take a bit longer. So we're going to skip over 0:22:54.600 --> 0:22:59.240 a ton of chatbots because a bunch were created between 0:22:59.320 --> 0:23:02.480 nineteen seven two, when Perry came out and when this 0:23:02.680 --> 0:23:06.320 next one did, and in twenty fourteen a lot of 0:23:06.359 --> 0:23:12.040 different news media outlets had these sensational headlines that programmers 0:23:12.080 --> 0:23:16.520 had created a chatbot that beat the Turing test. This 0:23:16.640 --> 0:23:19.960 was at an event in the UK organized by the 0:23:20.040 --> 0:23:23.760 University of Reading conducted by the Royal Society of London 0:23:24.280 --> 0:23:29.080 in which judges were having five minute long text based conversations, 0:23:29.119 --> 0:23:32.640 so kind of classic Turing tests set up here, and 0:23:33.440 --> 0:23:36.680 the person or thing on the other end was either 0:23:36.880 --> 0:23:40.560 a thirteen year old boy from Ukraine named Eugene Goosman 0:23:40.840 --> 0:23:45.520 as was claimed, or was actually a chatbot in this 0:23:45.560 --> 0:23:48.040 particular case. So they were chatting both with humans and 0:23:48.119 --> 0:23:52.280 with this chatbot that was trying to pass itself off 0:23:52.359 --> 0:23:55.760 as a thirteen year old boy from Ukraine, and thirty 0:23:55.800 --> 0:23:57.879 three percent of the judges or one third of the 0:23:58.000 --> 0:24:01.280 judges were fooled by the chat bought into thinking that 0:24:01.280 --> 0:24:05.320 that was in fact a boy that was chatting with them. However, 0:24:05.440 --> 0:24:09.560 just by contextualizing all that you start to see where 0:24:09.600 --> 0:24:12.119 those same sort of limitations come in in order to 0:24:12.119 --> 0:24:15.959 give the chatbot a fighting chats right, because it's a 0:24:15.960 --> 0:24:20.359 case where the supposed person you're chatting with is younger, 0:24:20.840 --> 0:24:25.760 so that could explain away some limited understanding and knowledge 0:24:25.760 --> 0:24:29.879 of various topics. That in addition to that, this was 0:24:30.119 --> 0:24:34.320 a young person from Ukraine, and that English would not 0:24:34.400 --> 0:24:37.800 be this person's first language, which could explain away any 0:24:38.359 --> 0:24:42.720 odd syntax that might be generated as a result. So 0:24:42.760 --> 0:24:45.439 while there were a lot of headlines about the Turing 0:24:45.440 --> 0:24:48.920 test being beaten by this chatbot, it definitely had more 0:24:49.000 --> 0:24:52.919 qualifiers attached to it. Still, it was more of a 0:24:53.000 --> 0:24:58.080 general purpose approach. It wasn't something like mimicking a person 0:24:58.359 --> 0:25:04.520 with schizophrenia or mimicking a stereotypical psychoanalyst. So we started 0:25:04.560 --> 0:25:09.440 to see that this was really an evolution of our 0:25:09.480 --> 0:25:15.119 ability to create machines that could mimic human conversation, that 0:25:15.280 --> 0:25:21.040 could appear to understand us. Now, a big part of 0:25:21.040 --> 0:25:25.040 that is, in fact, what we call natural language processing. 0:25:25.680 --> 0:25:28.880 This is a branch of computer science that involves building 0:25:28.920 --> 0:25:33.560 out models that let computers interpret commands that are expressed 0:25:33.600 --> 0:25:38.520 in normal human languages. As opposed to a programming language 0:25:38.600 --> 0:25:42.320 or a prescribed approach. So in the old days, if 0:25:42.320 --> 0:25:44.719 you wanted a computer to do something, you had to 0:25:44.720 --> 0:25:49.840 give specific commands in a specific way, in a specific order, 0:25:50.000 --> 0:25:52.680 or else it would not work. But with a good 0:25:52.800 --> 0:25:56.960 natural language processing methodology, you have a step in there 0:25:57.440 --> 0:26:00.600 in which the machine is able to parsey is being 0:26:00.680 --> 0:26:04.680 asked of it and attempt to respond in the appropriate way. 0:26:04.720 --> 0:26:07.520 So if it's a very good natural language processing method 0:26:08.040 --> 0:26:11.600 then the machine is going to produce a result that 0:26:11.840 --> 0:26:16.399 hopefully meets the person's expectations. It might not be perfect, 0:26:16.440 --> 0:26:19.439 but maybe it is close enough. The better the natural 0:26:19.480 --> 0:26:24.200 language processing, and the obviously the more capabilities the machine has, 0:26:24.800 --> 0:26:27.560 the better the result is going to be. Now, one 0:26:27.640 --> 0:26:31.479 computational advance we've seen help with natural language processing and 0:26:31.640 --> 0:26:36.960 advanced conversation models are artificial neural networks. This is a 0:26:36.960 --> 0:26:41.160 computer system that sort of simulates how our brains work. 0:26:41.680 --> 0:26:44.520 In our brains, we have neurons, right, and we have 0:26:44.800 --> 0:26:47.960 around eighty six billion of them in our brains. In 0:26:48.000 --> 0:26:52.560 the typical human brain, neurons are connected to other neurons, 0:26:52.760 --> 0:26:56.840 and messages in our brains crossover neural pathways as we 0:26:56.880 --> 0:27:01.000 make decisions. While an artificial neural network has nodes that 0:27:01.080 --> 0:27:05.879 interconnect with other nodes, and these nodes all represent neurons, 0:27:05.920 --> 0:27:09.679 and the nodes can accept traditionally two inputs, but it 0:27:09.720 --> 0:27:12.920 could be more than two and then produce a single output. 0:27:13.119 --> 0:27:16.440 So it's very similar to your classic logic gate if 0:27:16.440 --> 0:27:21.439 you're familiar with logic gates in programming. That is a 0:27:21.560 --> 0:27:24.840