WEBVTT 00:00:31.240 --> 00:00:37.820 All right. Opening AI for... language learning. 00:00:38.060 --> 00:00:40.920 Hi, my name is Mat Schulze and I'm here with 00:00:40.920 --> 00:00:44.420 Phil Hubbard, Senior Lecturer Emeritus at the 00:00:44.420 --> 00:00:47.159 Stanford University Language Center, where I 00:00:47.159 --> 00:00:50.000 taught English as a second language to graduate 00:00:50.000 --> 00:00:53.259 students there in various departments for over 00:00:53.259 --> 00:00:59.460 30 years. Since retiring in the 2020s, I focused 00:00:59.460 --> 00:01:04.420 my work on teacher education and especially professional 00:01:04.420 --> 00:01:06.579 development and that's something that I think 00:01:06.579 --> 00:01:11.459 will come through here today. Yeah and I am a 00:01:11.459 --> 00:01:14.280 professor of German at San Diego State University. 00:01:14.719 --> 00:01:18.040 By trade I'm actually a high school teacher of 00:01:18.040 --> 00:01:23.379 German and Russian. I've been researching AI 00:01:23.379 --> 00:01:25.640 and language learning for the last 30 years. 00:01:25.659 --> 00:01:28.299 It kind of dates me somewhat with it's been a 00:01:28.299 --> 00:01:32.890 while. I started off with working on my PhD doing 00:01:32.890 --> 00:01:36.549 what they call a research prototype of a grammar 00:01:36.549 --> 00:01:41.010 checker for students of German with a good old 00:01:41.010 --> 00:01:44.950 fashioned AI. Recently, and that's far more important, 00:01:45.189 --> 00:01:47.689 Phil and I wrote a journal article together, 00:01:48.450 --> 00:01:52.590 which was the starting point of these conversations. 00:01:53.790 --> 00:01:57.189 Yeah, the article was a position paper that we 00:01:57.189 --> 00:02:01.519 collaborated on It's driven by our mutual concern 00:02:01.519 --> 00:02:04.719 for the multitude of English teachers and other 00:02:04.719 --> 00:02:08.060 language teachers out there who did their training 00:02:08.060 --> 00:02:13.500 before 2022. And now they're faced with dealing 00:02:13.500 --> 00:02:17.259 with this AI phenomenon. Whether they want to 00:02:17.259 --> 00:02:19.780 or not, their students are likely to be using 00:02:19.780 --> 00:02:24.900 it. And we... decided to take advantage of this 00:02:24.900 --> 00:02:28.900 paper invitation to put together some advice 00:02:28.900 --> 00:02:33.719 based on our own readings and experience to help 00:02:33.719 --> 00:02:40.479 them out. The paper comes up with 10 areas that 00:02:40.479 --> 00:02:43.000 we determined were important for teachers to 00:02:43.000 --> 00:02:47.120 have some background in right now, and especially 00:02:47.120 --> 00:02:49.759 language teachers, because some of these areas 00:02:49.759 --> 00:02:55.199 like chat bots and machine translation are different 00:02:55.199 --> 00:02:57.680 in language teaching and learning than they are 00:02:57.680 --> 00:03:03.300 in other fields. We also provide seven sort of 00:03:03.300 --> 00:03:07.800 emerging principles to help guide this process, 00:03:08.340 --> 00:03:11.300 led by the idea that this is something that should 00:03:11.300 --> 00:03:15.060 be done quite regularly, perhaps even on a daily 00:03:15.060 --> 00:03:19.189 basis. And then talking about principles and 00:03:19.189 --> 00:03:21.990 being principled, we thought it would be a good 00:03:21.990 --> 00:03:25.310 idea to actually also do what we preach and not 00:03:25.310 --> 00:03:28.430 just kind of write the paper about it because 00:03:28.430 --> 00:03:31.750 this generative artificial intelligence is something 00:03:31.750 --> 00:03:35.590 we all need to learn about. It's come almost 00:03:35.590 --> 00:03:41.039 like a kind of huge tidal wave in some way. pretty 00:03:41.039 --> 00:03:44.460 much everyone, including experts, are struggling 00:03:44.460 --> 00:03:47.020 with one area or the other and need some catching 00:03:47.020 --> 00:03:50.680 up to do and need to do some learning. Phil and 00:03:50.680 --> 00:03:54.699 I have done this in relatively informal conversations 00:03:54.699 --> 00:03:58.879 actually after we submitted the paper. And that 00:03:58.879 --> 00:04:01.439 was how the idea was born. Oh, maybe we can share. 00:04:01.710 --> 00:04:04.449 some of those conversations, and this is what 00:04:04.449 --> 00:04:08.030 we're trying out now. In these conversations, 00:04:08.210 --> 00:04:10.669 the kind of very vague plan at this stage is 00:04:10.669 --> 00:04:12.909 that we will go through this paper, that we will 00:04:12.909 --> 00:04:16.189 pick individual concepts, things we've read, 00:04:16.250 --> 00:04:20.850 things we're thinking about, and give them a 00:04:20.850 --> 00:04:25.430 good read, basically, and a good discussion in 00:04:25.430 --> 00:04:28.670 many different ways, and we're hoping that this 00:04:28.670 --> 00:04:32.639 will be beneficial. We're calling this series 00:04:32.639 --> 00:04:36.139 of conversations opening AI for language learning. 00:04:37.560 --> 00:04:40.639 Or while or something. I'm not sure how I would 00:04:40.639 --> 00:04:44.379 pronounce that with my German accent. You can 00:04:44.379 --> 00:04:46.560 of course, you the audience can of course be 00:04:46.560 --> 00:04:51.110 part of it. In for future conversations, you'll 00:04:51.110 --> 00:04:53.209 be able to listen live. That's what we're hoping 00:04:53.209 --> 00:04:56.370 and that's what we're preparing. And we're very 00:04:56.370 --> 00:04:59.569 much hoping that you'll listen to one or the 00:04:59.569 --> 00:05:07.750 other recording of these conversations. Go ahead. 00:05:08.329 --> 00:05:10.670 This kind of brings us to our first conversation, 00:05:10.670 --> 00:05:15.290 right? And so, yes, we're both going ahead. We 00:05:15.290 --> 00:05:19.930 came up with the maybe. catchy title, What the 00:05:19.930 --> 00:05:25.449 LLM? And that's there to get us started. Okay, 00:05:25.449 --> 00:05:32.050 so let's get started. So Matt, what does LLM 00:05:32.050 --> 00:05:37.850 stand for? And what does it do? What is it? Well, 00:05:38.029 --> 00:05:42.240 we're both linguists, so I'd like I do this often 00:05:42.240 --> 00:05:46.180 with complex concepts. I start off with the words. 00:05:46.800 --> 00:05:49.519 So here we've got a compound noun adjective in 00:05:49.519 --> 00:05:52.879 front of it, which means compound nouns, you 00:05:52.879 --> 00:05:56.439 kind of disentangle them from the end. You look 00:05:56.439 --> 00:06:00.420 at the model first because it's a head noun of 00:06:00.420 --> 00:06:03.930 the whole. phrase in some way. So first question, 00:06:04.029 --> 00:06:06.629 when you want to know what a large language model 00:06:06.629 --> 00:06:10.509 is, LLM, large language model, you're looking 00:06:10.509 --> 00:06:15.470 at the model. So what's a model? I'm casting 00:06:15.470 --> 00:06:18.910 my mind back to my days in elementary school 00:06:18.910 --> 00:06:21.310 and we would literally do this with Play -Doh 00:06:21.310 --> 00:06:24.449 and drinking straws that we had to bring in from 00:06:24.449 --> 00:06:28.800 home. and we would build a cube with the different 00:06:28.800 --> 00:06:33.560 edges or pyramid or something like that. Why 00:06:33.560 --> 00:06:35.800 did we do this? And why did we do this in kind 00:06:35.800 --> 00:06:41.459 of elementary math class? To do what one does 00:06:41.459 --> 00:06:43.439 with a model, namely to learn about the thing. 00:06:43.480 --> 00:06:47.220 It's easier to learn about a model than it is 00:06:47.220 --> 00:06:50.480 to learn about a real cube because they are all 00:06:50.480 --> 00:06:53.040 too different and it's kind of difficult not 00:06:53.040 --> 00:06:55.829 only for children to see. what is actually there 00:06:55.829 --> 00:06:59.410 to learn, namely that all edges of the cube are 00:06:59.410 --> 00:07:01.949 of exactly the same lengths, if you did it right, 00:07:01.949 --> 00:07:06.029 basically, with your model. Here, of course, 00:07:06.029 --> 00:07:11.269 we're talking about the language model. We're 00:07:11.269 --> 00:07:13.990 not really dealing with language as such. We've 00:07:13.990 --> 00:07:17.410 got a model of language. That's the interesting 00:07:17.410 --> 00:07:20.670 part, I believe, of these large language models. 00:07:21.470 --> 00:07:25.089 That model is mathematical. It's really numbers. 00:07:26.529 --> 00:07:30.990 Yes, we put in something as the user of a large 00:07:30.990 --> 00:07:34.189 language model based chatbot or something. We 00:07:34.189 --> 00:07:37.709 provide a prompt, make it do something. We prompt 00:07:37.709 --> 00:07:42.089 it literally. We provide textual input. That 00:07:42.089 --> 00:07:47.990 input gets transformed into a large set of numbers, 00:07:48.350 --> 00:07:52.230 arrays of numbers and many of those arrays, basically. 00:07:52.759 --> 00:07:55.600 These arrays get processed through different 00:07:55.600 --> 00:07:59.879 equations time and time again through multiple, 00:08:00.079 --> 00:08:03.220 what they literally call hidden layers because 00:08:03.220 --> 00:08:05.839 you don't see them. You only see the prompt that 00:08:05.839 --> 00:08:08.600 goes in and you see the textual output that comes 00:08:08.600 --> 00:08:12.060 out. Everything mathematical is kind of hidden 00:08:12.060 --> 00:08:16.139 in the box, basically, in the machine. But all 00:08:16.139 --> 00:08:19.269 of that, trust me, is mathematical. All of that 00:08:19.269 --> 00:08:22.170 is numbers. Now we know the language is not really 00:08:22.170 --> 00:08:27.529 numbers. So the model also cuts out something. 00:08:30.290 --> 00:08:34.190 Last thing is the adjective large, which of course 00:08:34.190 --> 00:08:37.549 is interesting in a way also because it is ambiguous. 00:08:38.590 --> 00:08:41.250 Is the language large or is the model large? 00:08:41.570 --> 00:08:44.830 Or is the language model large? Of course it 00:08:44.830 --> 00:08:48.460 is the language model that is large. And they 00:08:48.460 --> 00:08:51.480 are really large with billions of parameters 00:08:51.480 --> 00:08:55.299 and things like that. They're humongous based 00:08:55.299 --> 00:09:00.139 on many, many texts scraped off the internet. 00:09:00.759 --> 00:09:04.320 But the other meaning kind of resonates as well. 00:09:04.700 --> 00:09:07.779 The languages are better. They better be large 00:09:07.779 --> 00:09:12.279 as well because then the large language model 00:09:12.279 --> 00:09:16.080 works better. Because you scrape off these texts 00:09:16.080 --> 00:09:18.220 off the internet and the more texts you have, 00:09:18.340 --> 00:09:21.539 the better the model is. So this works best for 00:09:21.539 --> 00:09:24.179 a language like English, because many, many texts 00:09:24.179 --> 00:09:27.519 on the internet are in English. And to go to 00:09:27.519 --> 00:09:30.100 the other end of the spectrum, I remember listening 00:09:30.100 --> 00:09:34.639 to a presentation about the languages in the 00:09:34.639 --> 00:09:38.509 Arctic. basically indigenous languages in the 00:09:38.509 --> 00:09:40.990 Arctic. There was a specialist computational 00:09:40.990 --> 00:09:44.009 linguist talking about those. And he actually 00:09:44.009 --> 00:09:46.389 did the calculation for one of them was a kind 00:09:46.389 --> 00:09:49.289 of minute number of speakers. And he basically 00:09:49.289 --> 00:09:52.909 said if they wanted to have enough texts for 00:09:52.909 --> 00:09:55.590 the creation of a large language model, it would 00:09:55.590 --> 00:09:59.970 take them 14 ,000 years to create those. So, 00:10:00.269 --> 00:10:03.019 you know, things take a while. languages like 00:10:03.019 --> 00:10:05.299 English, Spanish, whatever, a couple of others 00:10:05.299 --> 00:10:08.799 have a clear advantage there in some way. And 00:10:08.799 --> 00:10:12.559 we'll probably come to those questions of equity 00:10:12.559 --> 00:10:17.100 in some way also at some stage. Well, that's 00:10:17.100 --> 00:10:19.399 all right. Enough of the lecture. That's great. 00:10:20.519 --> 00:10:25.600 So we have this idea that it's all numbers and 00:10:25.600 --> 00:10:29.210 there's a lot going on there that language teachers 00:10:29.210 --> 00:10:32.409 don't really need to know about at the math level, 00:10:32.509 --> 00:10:35.350 although it's great if they would. I certainly 00:10:35.350 --> 00:10:40.769 don't. But what does all this mean then for a 00:10:40.769 --> 00:10:43.649 teacher or a learner who's trying to make use 00:10:43.649 --> 00:10:48.490 of this for, say, as a chat bot or in machine 00:10:48.490 --> 00:10:54.789 translation? What does it tell us about the real 00:10:54.789 --> 00:11:03.029 value of this? if it's all just math. Right. 00:11:06.190 --> 00:11:11.669 Well, I have started off with the math. Emily 00:11:11.669 --> 00:11:15.850 Bender, who you know, and her colleague Alex 00:11:15.850 --> 00:11:20.029 Hanna, they're trying to be polemic, I guess, 00:11:20.330 --> 00:11:25.460 and don't call it AI, the new. version, the one 00:11:25.460 --> 00:11:27.799 we are dealing with now, because it's also become 00:11:27.799 --> 00:11:31.879 a buzzword and buzzwords get hollowed out. They 00:11:31.879 --> 00:11:34.799 call it Massey Mass, to make a little bit of 00:11:34.799 --> 00:11:37.740 fun of it, because it's so many numbers in there. 00:11:37.980 --> 00:11:41.919 What does that mean for a language teacher? It 00:11:41.919 --> 00:11:46.860 means that first and foremost, you're interacting 00:11:46.860 --> 00:11:52.909 with a machine and not with a human. We are not 00:11:52.909 --> 00:11:57.009 used to interacting with machines, so we're transposing 00:11:57.009 --> 00:12:00.809 our experience with interaction with humans, 00:12:00.929 --> 00:12:04.549 we're transposing this onto a machine, and trying 00:12:04.549 --> 00:12:08.830 to do the same thing. So we get this very plausible, 00:12:09.009 --> 00:12:14.769 very well -formed output from an AI, GenAI chatbot. 00:12:15.670 --> 00:12:19.370 And it happens to me, I believe it happens to 00:12:19.370 --> 00:12:22.970 all of us. We read it and it feels like there's 00:12:22.970 --> 00:12:27.289 really somebody almost human talking to us. We're 00:12:27.289 --> 00:12:30.149 having a genuine conversation like the two of 00:12:30.149 --> 00:12:33.750 us are having a conversation now. And that is 00:12:33.750 --> 00:12:38.590 not the case because of the mass in there. We're 00:12:38.590 --> 00:12:42.389 dealing with a machine that acts very differently. 00:12:42.389 --> 00:12:44.929 It's just triggered. It doesn't have any intention. 00:12:46.049 --> 00:12:48.549 The machine does not want to explain anything 00:12:48.549 --> 00:12:51.309 to us. The machine does not want to answer our 00:12:51.309 --> 00:12:55.309 questions. The machine does not want to translate 00:12:55.309 --> 00:12:58.710 something for us. It can't. It doesn't have those 00:12:58.710 --> 00:13:01.529 intentions, at least as yet, the way they are 00:13:01.529 --> 00:13:05.610 kind of still structured and done. We are putting 00:13:05.610 --> 00:13:09.679 all this in and I think To remember that is one 00:13:09.679 --> 00:13:12.399 of the most important things for teachers because 00:13:12.399 --> 00:13:17.840 a lot of other answers for questions people will 00:13:17.840 --> 00:13:22.279 have at a smaller scale with a particular text, 00:13:22.559 --> 00:13:24.419 with a particular activity and things like that, 00:13:24.460 --> 00:13:27.360 with a particular learning, many, not all of 00:13:27.360 --> 00:13:31.440 them, many of them derive their answer from that. 00:13:31.600 --> 00:13:35.539 mathy math basically Being aware of it. It's 00:13:35.539 --> 00:13:37.879 simply a question of awareness. Nothing wrong 00:13:37.879 --> 00:13:41.220 with using those tools, but Better be aware. 00:13:41.320 --> 00:13:46.159 It's a tool. It's a hammer and not a human who's 00:13:46.159 --> 00:13:51.759 doing hammering. It's just a hammer Yeah, so 00:13:51.759 --> 00:13:56.659 on a more personal level as you pointed out part 00:13:56.659 --> 00:14:00.309 of what we're trying to do is follow our own 00:14:00.309 --> 00:14:04.129 advice and so I've been trying to do a little 00:14:04.129 --> 00:14:08.549 bit of language learning using this while at 00:14:08.549 --> 00:14:12.950 the same time going back and reading about large 00:14:12.950 --> 00:14:16.730 language models even reading the first part of 00:14:16.730 --> 00:14:20.029 our chapter that you wrote might have read it 00:14:20.029 --> 00:14:23.870 two or three times I hope it wasn't a bedtime 00:14:23.870 --> 00:14:27.600 and you fell asleep or something no it was quite 00:14:27.600 --> 00:14:31.600 clear and I recommend it to everyone. Thank you. 00:14:31.700 --> 00:14:34.580 But I've also supplemented it with some other 00:14:34.580 --> 00:14:39.740 ones and it often comes back to what's it like 00:14:39.740 --> 00:14:43.559 then from the learner's side to be doing this. 00:14:43.779 --> 00:14:46.620 And so I've been trying to learn French, learn 00:14:46.620 --> 00:14:49.980 a bit of French because I have family that lived 00:14:49.980 --> 00:14:53.289 in French -speaking Switzerland. and they're 00:14:53.289 --> 00:14:55.990 quite good in English, but I'd like to feel comfortable 00:14:55.990 --> 00:15:00.389 in the area when I visit them. So I've been using 00:15:00.389 --> 00:15:04.730 Google Translate in practice mode to do this 00:15:04.730 --> 00:15:09.169 and noticing a number of things. We'll go into 00:15:09.169 --> 00:15:12.110 that probably in a later episode in more detail, 00:15:12.190 --> 00:15:16.450 but I am getting that experience and I was curious 00:15:16.450 --> 00:15:21.250 about LLM, so I actually asked ChatGPT. to take 00:15:21.250 --> 00:15:26.350 on the role of an expert speaking at a panel 00:15:26.350 --> 00:15:30.070 to a bunch of language teachers and responding 00:15:30.070 --> 00:15:32.870 to questions from the language teachers in clear 00:15:32.870 --> 00:15:36.110 and understandable ways. And so the first thing 00:15:36.110 --> 00:15:41.769 I did was ask it, so what is an LLM? Just like 00:15:41.769 --> 00:15:46.070 we asked you. And it gave a nice description, 00:15:46.509 --> 00:15:51.250 no math to speak of in it. And then I said, so 00:15:51.250 --> 00:15:54.049 how does it work? And then it talked about a 00:15:54.049 --> 00:15:56.570 lot of the things you did. It's nice to know 00:15:56.570 --> 00:16:01.470 you're confirmed with ChatGPT. And I followed 00:16:01.470 --> 00:16:03.850 that one. And one of the things it said was, 00:16:04.009 --> 00:16:08.490 well, I'm actually just telling you what the 00:16:08.490 --> 00:16:12.830 likely next word is. And that's how it's being 00:16:12.830 --> 00:16:15.730 constructed. And then I went back and asked it 00:16:15.730 --> 00:16:17.730 again. I won't go through the whole dialogue. 00:16:18.380 --> 00:16:21.080 But essentially I said, so how do you know how 00:16:21.080 --> 00:16:24.200 to start? And it gave me an answer. And what 00:16:24.200 --> 00:16:26.879 I learned from that is that by engaging in this 00:16:26.879 --> 00:16:29.960 kind of interaction, which does feel like you're 00:16:29.960 --> 00:16:32.899 interacting with an expert, I'm getting the sense 00:16:32.899 --> 00:16:38.940 that I'm learning in a much more me -directed 00:16:38.940 --> 00:16:44.299 way, because I'm doing all the questioning. And 00:16:44.299 --> 00:16:46.470 I think that's a valuable thing for teachers 00:16:46.470 --> 00:16:50.710 to try if they haven't to really engage these 00:16:50.710 --> 00:16:53.309 in conversations in areas where they know something 00:16:53.309 --> 00:16:56.450 so they can check to make sure they're not being 00:16:56.450 --> 00:16:59.169 fooled. I guess we'll get into hallucinations 00:16:59.169 --> 00:17:02.529 and so on later on too. So that's just an aside 00:17:02.529 --> 00:17:06.910 but I wanted to bring it up because it is a there's 00:17:06.910 --> 00:17:10.289 a difference between knowing something about 00:17:10.289 --> 00:17:14.160 how LLMs work. and beginning to understand it, 00:17:14.240 --> 00:17:16.960 and that understanding can only come through 00:17:16.960 --> 00:17:21.180 experience layered on top of the facts. This 00:17:21.180 --> 00:17:24.359 is a good point and it's a very good example. 00:17:28.099 --> 00:17:31.779 As Chad GPT told you, it only generates forms. 00:17:32.000 --> 00:17:35.339 It gives you the next, fine, it calls it a word. 00:17:35.539 --> 00:17:37.700 It's probably kind of some kind of shorthand. 00:17:37.799 --> 00:17:40.579 It gives you the next form that looks like a 00:17:40.579 --> 00:17:43.079 word, basically, because it deals with those 00:17:43.079 --> 00:17:47.099 tokens in some way and kind of generates those 00:17:47.099 --> 00:17:51.880 in different languages at enormous speed, at 00:17:51.880 --> 00:17:55.579 enormous levels of complexity, because it has 00:17:55.579 --> 00:18:00.900 literally thousands of these different layers 00:18:00.900 --> 00:18:04.259 in which you can process this and I love that 00:18:04.259 --> 00:18:09.279 term. Apparently it has multiple attention heads. 00:18:09.559 --> 00:18:14.599 I wish I had. So it can look at one piece of 00:18:14.599 --> 00:18:19.259 text and the different words up to kind of I 00:18:19.259 --> 00:18:25.259 think the number was 126. 126 words at the same 00:18:25.259 --> 00:18:28.170 time and deal with them at the same time. Which 00:18:28.170 --> 00:18:30.809 sometimes, yes, then leads to its own problems. 00:18:30.890 --> 00:18:34.049 But most of the time it goes right and it generates 00:18:34.049 --> 00:18:40.089 these forms. And since, as you were saying, in 00:18:40.089 --> 00:18:43.349 both examples, both with the French and the LLM, 00:18:44.049 --> 00:18:46.829 you had some prior knowledge. Yes, you were also 00:18:46.829 --> 00:18:48.849 a learner and you were taking a position of a 00:18:48.849 --> 00:18:51.789 learner in both of them. So let's start with 00:18:51.789 --> 00:18:55.069 the French as a kind of learner of French. But 00:18:55.069 --> 00:18:59.180 had And you were doing, I'm assuming, kind of 00:18:59.180 --> 00:19:01.539 intermediate level, I don't know what your French 00:19:01.539 --> 00:19:06.240 is like, level French in some way, relatively 00:19:06.240 --> 00:19:10.460 everyday conversations that you can have with 00:19:10.460 --> 00:19:13.680 your grandchildren or with some neighbors who 00:19:13.680 --> 00:19:17.140 you encounter when you visit your family in Switzerland. 00:19:17.680 --> 00:19:19.640 Yeah, I'd say I'm trying to break through to 00:19:19.640 --> 00:19:26.130 A2. A2, yeah. I'm on the verge, I think. So that, 00:19:26.210 --> 00:19:30.910 of course, is also a little more easy to generate 00:19:30.910 --> 00:19:35.690 through the large language model. So it's less 00:19:35.690 --> 00:19:39.349 likely to go completely wrong in some way because 00:19:39.349 --> 00:19:42.769 these are also very frequent items that get generated 00:19:42.769 --> 00:19:46.630 there in almost any language, I would think. 00:19:47.789 --> 00:19:50.210 And you have to, most importantly, I think, you 00:19:50.210 --> 00:19:53.900 have the ability to check Maybe not for every 00:19:53.900 --> 00:19:56.079 minute detail there, you have to trust the machine 00:19:56.079 --> 00:19:58.940 and its role is kind of teaching you something. 00:20:00.180 --> 00:20:06.680 But it won't lead you completely astray, basically. 00:20:09.160 --> 00:20:12.480 And that is, I was talking about teachers earlier 00:20:12.480 --> 00:20:16.299 and what the LLM means for the teachers. This 00:20:16.299 --> 00:20:19.000 idea, well, I'm actually dealing with a machine. 00:20:19.519 --> 00:20:22.160 For the learners, there is one additional one. 00:20:23.659 --> 00:20:27.200 learners get information from the teachers, from 00:20:27.200 --> 00:20:30.799 a textbook, from the library or whatever. Hopefully 00:20:30.799 --> 00:20:33.339 information they can trust if the teacher is 00:20:33.339 --> 00:20:36.480 well trained and things like that. They don't 00:20:36.480 --> 00:20:39.819 have the same control over a machine and we're 00:20:39.819 --> 00:20:48.019 often asking users of these chatbots, as you 00:20:48.019 --> 00:20:51.940 well know Phil, Reminder often is, oh, check 00:20:51.940 --> 00:20:56.160 the answers, check the output of the chat board 00:20:56.160 --> 00:21:00.880 or whatever. For learners, this is very difficult. 00:21:00.980 --> 00:21:03.960 That's an additional challenge. Sometimes I believe 00:21:03.960 --> 00:21:08.259 we should not ask them to do that because we're 00:21:08.259 --> 00:21:11.779 supposed to teach them. They can't know. They're 00:21:11.779 --> 00:21:13.400 not supposed to know that. How can they check 00:21:13.400 --> 00:21:16.019 it? You teach them that and kind of move them 00:21:16.019 --> 00:21:20.279 along in some way. So yeah, there are additional 00:21:20.279 --> 00:21:23.460 hurdles and things to consider. And that's why 00:21:23.460 --> 00:21:26.799 a number of people, and I'm sure we'll talk about 00:21:26.799 --> 00:21:30.420 the notion of intelligence and artificial intelligence 00:21:30.420 --> 00:21:35.799 in one of the kind of future conversations. It's 00:21:35.799 --> 00:21:39.099 a little bit misleading, this artificial intelligence, 00:21:39.319 --> 00:21:42.720 right? It's not intelligent enough to kind of 00:21:42.720 --> 00:21:46.750 teach you. You can't completely trust it. Also, 00:21:46.990 --> 00:21:50.230 learners need to be aware of, call it limitations 00:21:50.230 --> 00:21:52.630 in some way, but also of the strengths, what 00:21:52.630 --> 00:21:57.970 they can do. To make a kind of polemic point 00:21:57.970 --> 00:22:01.450 and move us along basically in this conversation, 00:22:02.230 --> 00:22:05.849 I mentioned a name, Emily Bender already. She 00:22:05.849 --> 00:22:09.839 was the lead author of a paper where they I believe 00:22:09.839 --> 00:22:12.460 coined the phrase, I'm not 100 % sure, but certainly 00:22:12.460 --> 00:22:15.380 used it and made it mainstream because Google 00:22:15.380 --> 00:22:18.779 fought it. The paper became almost notorious 00:22:18.779 --> 00:22:22.779 in some way. The phrase is stochastic parrots, 00:22:23.059 --> 00:22:26.500 and that's what they use to explain the LLM, 00:22:26.539 --> 00:22:30.359 basically the large language model. In a later 00:22:30.359 --> 00:22:33.859 book that has just come out this year by Emily 00:22:33.859 --> 00:22:37.460 Bender and Alex Hanna, I love that phrase, they 00:22:37.460 --> 00:22:41.599 talk about machines extruding synthetic text. 00:22:42.839 --> 00:22:45.359 Both of those stochastic parrots, I'm going to 00:22:45.359 --> 00:22:47.819 say a few words about that in a second, I guess 00:22:47.819 --> 00:22:50.799 the other phrase is clear, extruding synthetic 00:22:50.799 --> 00:22:57.079 text. Both of them are very negative, and they're 00:22:57.079 --> 00:23:01.579 trying to make a point there. Emily Bender is 00:23:01.579 --> 00:23:03.779 a computational language, so it's not like... 00:23:04.200 --> 00:23:07.039 She is a Luddite or something like that. But 00:23:07.039 --> 00:23:09.559 she is very, very critical of these large language 00:23:09.559 --> 00:23:15.200 models for Ruhfdorf, mainly ethical and socioeconomic 00:23:15.200 --> 00:23:18.140 reasons. That's my interpretation. I mean, one 00:23:18.140 --> 00:23:22.599 would have to ask her. And where she interests 00:23:22.599 --> 00:23:25.859 is with those stochastic parrots. Stochastic 00:23:25.859 --> 00:23:28.740 basically means you're doing things by chance. 00:23:29.940 --> 00:23:33.920 So it's a parrot. who gets it right and sounds 00:23:33.920 --> 00:23:38.779 very eloquent just by chance. It's just serendipity. 00:23:39.200 --> 00:23:42.579 The old word stochastic was apparently used for 00:23:42.579 --> 00:23:49.740 bow and arrows. Like you would literally by accident 00:23:49.740 --> 00:23:53.180 hit the target with your arrow. You had no idea 00:23:53.180 --> 00:23:56.440 what to do with bow and arrow. That was the stochastic. 00:23:56.519 --> 00:23:58.720 kind of thing. And of course, there's statistics 00:23:58.720 --> 00:24:04.319 and mathematics in there and things that. It's 00:24:04.319 --> 00:24:07.900 a little bit more. It's a very high probability. 00:24:08.380 --> 00:24:12.059 And the texts we get are very plausible. And 00:24:12.059 --> 00:24:15.680 Benda and Hannah do not deny that in any way. 00:24:16.960 --> 00:24:20.480 But it's this awareness. If we really be polemic 00:24:20.480 --> 00:24:23.880 about it, it's just hollow text, basically. It's 00:24:23.880 --> 00:24:28.880 hollow form. We're actually It doesn't mean anything. 00:24:29.400 --> 00:24:32.160 Yeah. So you've touched on this a bit, but maybe 00:24:32.160 --> 00:24:37.279 you could explain in a little more detail what's 00:24:37.279 --> 00:24:40.680 human to human conversation, which is what all 00:24:40.680 --> 00:24:43.680 our second language acquisition theories are 00:24:43.680 --> 00:24:48.880 based on. How's that different from human to 00:24:48.880 --> 00:24:55.059 chatbot? When the two of us are talking, If I'm 00:24:55.059 --> 00:24:58.079 being a good boy, I'm listening while you're 00:24:58.079 --> 00:25:03.339 talking and I'm not interrupting. While I'm listening, 00:25:03.720 --> 00:25:06.440 I'm doing a number of things. That's just the 00:25:06.440 --> 00:25:09.059 way cognition works, basically. Of course, the 00:25:09.059 --> 00:25:12.500 first thing is I'm kind of getting through the 00:25:12.500 --> 00:25:15.319 auditory channels. I'm literally listening in 00:25:15.319 --> 00:25:19.859 a narrow sense. I'm parsing the utterances, so 00:25:19.859 --> 00:25:21.720 I'm kind of chunking them, putting them in kind 00:25:21.720 --> 00:25:25.869 of... realizing what this means and what the 00:25:25.869 --> 00:25:31.150 structure is and things like that. But I'm not 00:25:31.150 --> 00:25:37.609 reacting to the utterance per se. What I am reacting 00:25:37.609 --> 00:25:43.890 to is my reasoning about your intention. So you 00:25:43.890 --> 00:25:47.289 are saying something and I'm thinking, why did 00:25:47.289 --> 00:25:52.200 he tell me that? What's his point? And then based 00:25:52.200 --> 00:25:55.119 on that reasoning, that's how I react. Because 00:25:55.119 --> 00:25:58.740 I want to be nice, and it's just those Grisey 00:25:58.740 --> 00:26:03.819 maxims of conversation. I want to be relevant. 00:26:03.980 --> 00:26:06.839 I want to be informative. I want to be cooperative. 00:26:08.160 --> 00:26:11.440 I know they're all ideals, but that's really 00:26:11.440 --> 00:26:13.940 how it works most of the time. Does it break 00:26:13.940 --> 00:26:16.980 down, then, if somebody doesn't reason right, 00:26:17.519 --> 00:26:19.240 doesn't understand something right, or whatever? 00:26:19.400 --> 00:26:23.240 Yes. But as long as there is no malicious actor 00:26:23.240 --> 00:26:26.359 in the conversation, it really does work most 00:26:26.359 --> 00:26:28.799 of the time. That's why we are so used to it. 00:26:30.140 --> 00:26:36.299 So now I'm doing exactly the same when I interact 00:26:36.299 --> 00:26:40.940 with the computer. The best way of explaining, 00:26:41.299 --> 00:26:45.539 let me use a human -human example first. Dinner 00:26:45.539 --> 00:26:49.220 situation and somebody says, can you pass me 00:26:49.220 --> 00:26:52.539 the salt? It's a standard example from this kind 00:26:52.539 --> 00:26:58.059 of linguistic analysis. And my reasoning then 00:26:58.059 --> 00:27:02.740 is, oh, it's a yes -no question. It asks, there's 00:27:02.740 --> 00:27:05.660 can in the front, so it asks after my ability. 00:27:06.119 --> 00:27:09.779 But I know that this person who is asking me 00:27:09.779 --> 00:27:12.660 knows that of course I'm well able to pass the 00:27:12.660 --> 00:27:15.460 salt. I'm sitting close enough, I'm strong enough 00:27:15.460 --> 00:27:19.119 to even lift the salt shaker and I can do that. 00:27:20.859 --> 00:27:23.380 So this wouldn't be informative. They wouldn't 00:27:23.380 --> 00:27:25.720 ask me that. The staff would be patronizing. 00:27:25.839 --> 00:27:27.799 They're almost saying, you're not capable of 00:27:27.799 --> 00:27:30.279 it or whatever. No, it's dinner. They're not 00:27:30.279 --> 00:27:32.359 trying to offend me. I'm not being threatened. 00:27:32.839 --> 00:27:35.279 So there must be another reason. And then I realized, 00:27:35.380 --> 00:27:37.619 oh, they're sitting further away from the salt. 00:27:38.339 --> 00:27:40.460 All they want me to do is kind of move it closer 00:27:40.460 --> 00:27:42.980 to them so that they can use it on, I don't know, 00:27:43.099 --> 00:27:46.279 on their potatoes or something. So that's what 00:27:46.279 --> 00:27:48.359 I mean with reasoning about intention. We're 00:27:48.359 --> 00:27:51.759 kind of excluding things and whatever. Now we 00:27:51.759 --> 00:27:55.839 get these plausible texts from a chat bot, a 00:27:55.839 --> 00:28:00.279 Gen .ai chat bot, and we're doing the same reasoning. 00:28:01.279 --> 00:28:03.920 And the fallacy is the problem we are having, 00:28:04.960 --> 00:28:10.579 the machine does not have an intention. The machine 00:28:10.579 --> 00:28:13.960 did not produce the text because it wanted to 00:28:13.960 --> 00:28:16.559 say something, wanted to give me information 00:28:16.559 --> 00:28:19.859 or whatever. The machine generated that text 00:28:19.859 --> 00:28:23.279 because I triggered it. I clicked something. 00:28:23.640 --> 00:28:27.859 That's the only reason, the only occasion, the 00:28:27.859 --> 00:28:32.059 only trigger for the generation of the text. 00:28:32.339 --> 00:28:36.670 There is no other thing there. So yeah, that's 00:28:36.670 --> 00:28:40.910 the major difference basically. So there's from 00:28:40.910 --> 00:28:46.049 a second language acquisition perspective, there's 00:28:46.049 --> 00:28:50.670 the simulation of negotiation of meaning, but 00:28:50.670 --> 00:28:55.130 it's all on your side as the human in terms of 00:28:55.130 --> 00:28:59.150 trying to extract meaning from the chat bot. 00:28:59.390 --> 00:29:03.029 And similarly, there's no co -construction of 00:29:03.029 --> 00:29:06.039 meaning because Again, all the meanings on your 00:29:06.039 --> 00:29:14.980 side, the chat bot is presumably generating semantically 00:29:14.980 --> 00:29:19.960 relevant forms, but it doesn't have the sense 00:29:19.960 --> 00:29:26.299 of meaning behind it. Is that right? Right. The 00:29:26.299 --> 00:29:29.240 best way I can explain this, yes, this is absolutely 00:29:29.240 --> 00:29:31.539 right. The best way I can explain this is with 00:29:31.539 --> 00:29:37.660 literature. So, doesn't matter, a novel, a novella, 00:29:38.079 --> 00:29:43.700 a poem, you're reading something. The author, 00:29:44.019 --> 00:29:47.460 the poet, the novelist clearly had an intention 00:29:47.460 --> 00:29:52.490 and imbued that text with some meaning. That's 00:29:52.490 --> 00:29:54.890 why they wrote it. It might just have been purely 00:29:54.890 --> 00:29:57.190 aesthetic if that's what they wanted. It might 00:29:57.190 --> 00:29:59.490 have been very politically activist, it might 00:29:59.490 --> 00:30:01.670 have been philosophical, might have just been 00:30:01.670 --> 00:30:03.430 entertaining, whatever, but they had something. 00:30:04.609 --> 00:30:08.890 Now you're picking it up and you're reasoning. 00:30:09.319 --> 00:30:14.299 Sometimes. Also, the proverbial, what does he 00:30:14.299 --> 00:30:20.019 also want us to say, goes a little further than 00:30:20.019 --> 00:30:22.859 that. But we're trying to figure out what does 00:30:22.859 --> 00:30:27.259 that mean? What does it mean to us? How can I 00:30:27.259 --> 00:30:32.960 participate in this shared experience? And the 00:30:32.960 --> 00:30:37.740 phrase always was, the reader completes. the 00:30:37.740 --> 00:30:44.579 meaning of the literary text. In this case, with 00:30:44.579 --> 00:30:48.000 a GenAI chatbot, there's nothing too complete 00:30:48.000 --> 00:30:50.599 because there's no meaning in there. So what 00:30:50.599 --> 00:30:54.319 the reader does is, and whether it's a teacher 00:30:54.319 --> 00:30:57.420 or a learner or anyone, whether it's a second 00:30:57.420 --> 00:31:00.019 language or first language, it really doesn't 00:31:00.019 --> 00:31:04.359 matter. The reader... The one who also often 00:31:04.359 --> 00:31:06.240 prompted, because it's usually one and the same 00:31:06.240 --> 00:31:09.599 person, right? The one who prompted and receives 00:31:09.599 --> 00:31:15.380 the output text imbues that output text with 00:31:15.380 --> 00:31:20.519 meaning. Literally, we make sense of it. The 00:31:20.519 --> 00:31:23.640 machine doesn't have any sense, basically. As 00:31:23.640 --> 00:31:27.000 in, there is no meaning in that text. And yes, 00:31:27.000 --> 00:31:29.660 it's a kind of big phrase, this negotiation of 00:31:29.660 --> 00:31:32.690 meaning. That's this reasoning about intentions 00:31:32.690 --> 00:31:35.670 as we do and second language learners learn to 00:31:35.670 --> 00:31:38.130 do that in the second language. And we think, 00:31:38.250 --> 00:31:41.029 oh, I didn't understand that. And do I need to 00:31:41.029 --> 00:31:43.910 go, if I need to go to the post office, do I 00:31:43.910 --> 00:31:46.289 need to turn right or did you say left? And we're 00:31:46.289 --> 00:31:51.849 kind of working this out in some way. None of 00:31:51.849 --> 00:31:55.390 that can happen with the machine. So we're literally 00:31:55.390 --> 00:32:01.650 parroting, mimicking it. Basically. So that means 00:32:01.650 --> 00:32:06.750 when I'm reading one of your papers that I'm 00:32:06.750 --> 00:32:10.170 just putting all the meaning in myself because 00:32:10.170 --> 00:32:16.509 it's just digital marks on my screen. No, because 00:32:16.509 --> 00:32:20.390 I generated them with meaning. I didn't generate 00:32:20.390 --> 00:32:23.869 them. That was the wrong word. I wrote them with 00:32:23.869 --> 00:32:26.849 meaning. I wrote them with intention. I see. 00:32:26.970 --> 00:32:31.440 So you didn't. What does this guy want me to 00:32:31.440 --> 00:32:35.680 know? What is he on about? And then you might 00:32:35.680 --> 00:32:37.960 come to a totally different conclusion than I 00:32:37.960 --> 00:32:43.980 did. And that's absolutely fine. Unless I generate, 00:32:44.099 --> 00:32:46.640 and that's why generating research papers with 00:32:46.640 --> 00:32:50.500 something like Chad GPT is so dangerous, right? 00:32:50.980 --> 00:32:54.279 And there have been proposals by very large companies 00:32:54.279 --> 00:33:01.430 to do exactly that. If this is generated by a 00:33:01.430 --> 00:33:03.769 machine, then of course it's not genuine research. 00:33:04.250 --> 00:33:07.130 Then yes, you would just have to put in all the 00:33:07.130 --> 00:33:10.150 meaning, which sometimes is nigh impossible because 00:33:10.150 --> 00:33:15.089 it's just quite frankly some very elegant gobbledygook 00:33:15.089 --> 00:33:26.029 basically. Okay, so what advice do you have about 00:33:25.980 --> 00:33:29.200 using a chat, then what kind of conversations, 00:33:29.759 --> 00:33:32.980 knowing that it's really one sided. But we have 00:33:32.980 --> 00:33:35.839 this, it's almost like another literary term, 00:33:35.880 --> 00:33:38.700 I remember a willing suspension of disbelief. 00:33:39.059 --> 00:33:43.420 So I feel like I'm engaging in a real conversation. 00:33:45.200 --> 00:33:50.720 So what are some ideas of using it for conversation 00:33:50.720 --> 00:33:56.529 practice then? What should we be paying attention 00:33:56.529 --> 00:34:02.950 to? You know this much better than I do. Technology 00:34:02.950 --> 00:34:07.490 very often also is about access. We can do things 00:34:07.490 --> 00:34:10.949 with technology that are much, much more difficult 00:34:10.949 --> 00:34:14.789 to do with a bunch of humans basically because 00:34:14.789 --> 00:34:19.030 labor is expensive or location is difficult, 00:34:19.710 --> 00:34:24.260 schedules are difficult in some way. So when 00:34:24.260 --> 00:34:28.980 it comes to conversation the huge advantage of 00:34:28.980 --> 00:34:33.679 the machine is you have the most patient conversation 00:34:33.679 --> 00:34:38.059 partner ever. If you wanted to repeat it again 00:34:38.059 --> 00:34:40.940 and again and again with listening or whatever 00:34:40.940 --> 00:34:43.099 because you didn't fully understand it or you 00:34:43.099 --> 00:34:46.219 just want to practice some more you can and the 00:34:46.219 --> 00:34:49.559 machine will not give up or start shouting at 00:34:49.559 --> 00:34:53.440 you or walking away or something like that. It's 00:34:53.440 --> 00:34:56.300 available 24 -7. That's almost a cliche, but 00:34:56.300 --> 00:35:03.619 it's true. This also, and maybe I should have 00:35:03.619 --> 00:35:07.980 forefronted this, use it as a conversation partner. 00:35:08.840 --> 00:35:12.360 As a language learner, perhaps also as a teacher, 00:35:12.659 --> 00:35:15.599 use it for brainstorming. Use it to ask in a 00:35:15.599 --> 00:35:19.260 dialogue what an LLMA is, ask clarification questions 00:35:19.260 --> 00:35:23.380 and kind of work your way around and use it in 00:35:23.380 --> 00:35:27.500 this artificial conversation for your purposes. 00:35:30.320 --> 00:35:32.980 For language learning, I think it's very important. 00:35:33.420 --> 00:35:36.420 Again, an advantage because it's so good at generating 00:35:36.420 --> 00:35:41.820 these forms and they're so accurate. use the 00:35:41.820 --> 00:35:43.880 opportunity, particularly as a language learner, 00:35:44.079 --> 00:35:46.559 but I'm sure it's relevant for some language 00:35:46.559 --> 00:35:49.960 teachers too. Pay attention to the beautiful 00:35:49.960 --> 00:35:54.659 accuracy. Pick up new words. Pick up the grammatical 00:35:54.659 --> 00:35:57.239 constructions used in there. Pay attention to 00:35:57.239 --> 00:36:01.079 the use of prepositions, subject -verb agreement 00:36:01.079 --> 00:36:04.480 on whatever the grammatical features are, because... 00:36:04.539 --> 00:36:08.360 If you notice it, another concept from second 00:36:08.360 --> 00:36:10.800 language acquisition, you are going to learn. 00:36:11.300 --> 00:36:14.719 And yes, you can notice things in that machine 00:36:14.719 --> 00:36:21.340 output. Most importantly, as a kind of conclusion 00:36:21.340 --> 00:36:26.900 for this conversation, I think, it's a question 00:36:26.900 --> 00:36:29.760 of awareness. Awareness is very important in 00:36:29.760 --> 00:36:34.630 learning. Do your best, because it is difficult, 00:36:34.710 --> 00:36:37.269 and I'm speaking from personal experience, do 00:36:37.269 --> 00:36:41.710 your best to always be aware that it is a machine 00:36:41.710 --> 00:36:45.789 you're interacting with. As the Brits would say, 00:36:45.989 --> 00:36:49.570 take it with a pinch of salt. Be a little careful 00:36:49.570 --> 00:36:53.909 with what you do, what you put in, what you ask 00:36:53.909 --> 00:36:57.269 it to do. There are clear limitations. Don't 00:36:57.269 --> 00:36:59.710 trust everything you see on the screen and the 00:36:59.710 --> 00:37:04.699 output. Work with it. Work with the strengths 00:37:04.699 --> 00:37:09.599 of the machine and they're clearly in those two 00:37:09.599 --> 00:37:14.900 areas like patience and accuracy. Work with those, 00:37:15.039 --> 00:37:18.539 use those to your advantage and make sure, no, 00:37:18.639 --> 00:37:21.420 it's a machine. It sometimes does things which, 00:37:21.579 --> 00:37:24.000 no, they won't happen in real life, basically. 00:37:26.139 --> 00:37:32.909 All right. Well, for... For this episode, for 00:37:32.909 --> 00:37:35.769 this conversation, a lot of what you've been 00:37:35.769 --> 00:37:38.929 talking about is covered in the first part of 00:37:38.929 --> 00:37:43.869 our paper. There are some things you have certainly 00:37:43.869 --> 00:37:47.269 added and gone into much more detail on, but 00:37:47.269 --> 00:37:53.510 I think in future conversations we'll be deconstructing 00:37:53.510 --> 00:37:57.489 that paper a bit more and hopefully following 00:37:57.489 --> 00:38:01.699 through on what, you know, Our goal is, I think, 00:38:01.980 --> 00:38:07.940 which is to try to share our own journey through 00:38:07.940 --> 00:38:13.360 understanding this new technology and making 00:38:13.360 --> 00:38:17.559 it work for us instead of making us work for 00:38:17.559 --> 00:38:22.960 it. Right. And just one additional note on the 00:38:22.960 --> 00:38:27.750 art show. This is a journal article and that 00:38:27.750 --> 00:38:31.429 journal article is open access. When we get a 00:38:31.429 --> 00:38:35.690 chance to put this recording out there, the link 00:38:35.690 --> 00:38:40.289 to the actual article will be right next to that 00:38:40.289 --> 00:38:43.829 recording. So people can also, yes, you can have 00:38:43.829 --> 00:38:46.110 the pleasure of listening to us for half an hour 00:38:46.110 --> 00:38:50.750 or so, but you can also double check everything 00:38:50.750 --> 00:38:54.829 that we are saying here. in this particular article. 00:38:54.949 --> 00:38:57.269 And of course, there are many, many others. There's 00:38:57.269 --> 00:39:01.010 a whole proliferation now of articles, book chapters, 00:39:01.269 --> 00:39:04.710 whole books on generative artificial intelligence 00:39:04.710 --> 00:39:08.349 and language learning, language teaching in particular, 00:39:08.409 --> 00:39:13.690 but also beyond, of course. All right. All right. 00:39:13.769 --> 00:39:16.869 I think we're good. Well, see you next time, 00:39:17.010 --> 00:39:20.349 Matt. See you next time. More topics to discover. 00:39:20.730 --> 00:39:34.130 opening AI for language learning. We are grateful 00:39:34.130 --> 00:39:37.010 for the support for opening AI for language learning 00:39:37.010 --> 00:39:39.690 by the Language and Applied Research Center at 00:39:39.690 --> 00:39:42.510 San Diego State University and the Southern Area 00:39:42.510 --> 00:39:45.449 International Languages Network, SAILN, which 00:39:45.449 --> 00:39:47.889 is part of the California World Languages Project. 00:39:48.570 --> 00:39:51.010 Mari Ocando Finol is the production coordinator 00:39:51.010 --> 00:39:54.190 of OAILL. Our theme music is by Tillmann Spiegl. 00:39:54.650 --> 00:39:57.750 Our editor is Chris Brown. Live conversations 00:39:57.750 --> 00:40:00.750 are moderated and promoted by me, Shahnaz Ahmadeian. 00:40:01.170 --> 00:40:02.150 Until next time.