1 00:00:00,160 --> 00:00:03,620 Welcome back to another riveting episode of Data Driven. 2 00:00:03,919 --> 00:00:07,600 Joining us today, lakeside and positively glowing from his 3 00:00:07,600 --> 00:00:11,175 Appalachian retreat, is Frank. Meanwhile, the 4 00:00:11,175 --> 00:00:14,535 always astute and ever energetic Andy is here to keep us 5 00:00:14,535 --> 00:00:18,260 grounded. But enough about us. Today, we have 6 00:00:18,260 --> 00:00:22,099 a true luminary in the field of AI, someone who's blending the worlds 7 00:00:22,099 --> 00:00:25,845 of academia and enterprise with seamless finesse. He's an 8 00:00:25,845 --> 00:00:29,465 associate professor at the Technion, has published over 100 9 00:00:29,525 --> 00:00:33,285 research papers on automated speech recognition, and is the chief 10 00:00:33,285 --> 00:00:36,670 scientist at Iola. Please welcome doctor Yossi 11 00:00:36,670 --> 00:00:39,650 Keshet or as he's known to his friends, Yossi. 12 00:00:47,035 --> 00:00:50,555 Alright. Hello, and welcome to Data Driven, the podcast where we explore the 13 00:00:50,555 --> 00:00:53,340 emergent fields of artificial intelligence, data science, and, 14 00:00:55,180 --> 00:00:58,640 and, of course, data engineering, without which the whole world would probably stop turning. 15 00:00:59,739 --> 00:01:03,315 And you know, data engineering is important. That's 16 00:01:03,315 --> 00:01:06,755 basically it. Still working on that that that revamped 17 00:01:06,755 --> 00:01:10,570 monologue, for, for season 8, Andy. Were 18 00:01:10,570 --> 00:01:14,410 you on vacation? You're on vacation. I am on vacation. And 19 00:01:14,410 --> 00:01:17,995 for those of you who can't see on camera who are not who are 20 00:01:17,995 --> 00:01:20,735 listening, not watching, I am literally lakeside, 21 00:01:22,075 --> 00:01:25,880 in the foothills. Well, not the foothills. We are actually in the Appalachian Mountains. Or 22 00:01:25,880 --> 00:01:29,720 is it Appalachian? I I never I I've heard of those. I I never 23 00:01:29,720 --> 00:01:32,840 got a clear read on it. Say either. So, you know When I say either. 24 00:01:32,840 --> 00:01:36,345 Yeah. Yeah. Yeah. Yeah. Yeah. So I am in Deep Creek Lake, 25 00:01:36,345 --> 00:01:40,045 Maryland, which is kind of like, Maryland doesn't really have a Panhandle 26 00:01:40,185 --> 00:01:43,725 per se, but if it did, it would be this is what this would be. 27 00:01:44,068 --> 00:01:47,682 I probably think I'm 5 miles from West Virginia and about 28 00:01:47,682 --> 00:01:51,296 20 miles from Pennsylvania. So it's kind of like this quiet 29 00:01:51,296 --> 00:01:53,085 little corner of the state. 30 00:01:54,665 --> 00:01:58,345 And I've been, you know, reading and studying 31 00:01:58,345 --> 00:02:01,890 today. I hit day 600 on Pluralsight Consecutive. Nice. 32 00:02:02,670 --> 00:02:06,430 So recording this June 17th. And, how 33 00:02:06,430 --> 00:02:10,005 things with you, Andy? Things are good. I'm gonna throw out a plug for 34 00:02:10,005 --> 00:02:13,225 data driven media dot tv because Frank mentioned. 35 00:02:13,765 --> 00:02:17,420 If you're listening, he while he was mentioning that, he was 36 00:02:17,420 --> 00:02:20,860 actually panning the camera over to the lake. But if 37 00:02:20,860 --> 00:02:24,700 you're, subscribing to data driven media dot tv, you get 38 00:02:24,700 --> 00:02:28,504 to see us. You get to see the video, and you 39 00:02:28,504 --> 00:02:32,185 can see, for instance, that I am wearing the, my data is the 40 00:02:32,185 --> 00:02:35,930 new oil t shirt, which you can pick up. I'm just full of 41 00:02:35,930 --> 00:02:39,769 sponsor stuff today. I'm just doing Well, it's self out. It's 42 00:02:39,769 --> 00:02:43,308 self sponsored. And, honestly, we really need to get better at that. Right? We have 43 00:02:43,308 --> 00:02:47,070 data channel. Tv. There is a for listeners to the show, I will give 44 00:02:47,070 --> 00:02:50,810 a preview. There is gonna be data driven academy is is launching soon. You have 45 00:02:50,810 --> 00:02:54,510 a course coming up the end of the month. Actually, yeah, it's fabric. 46 00:02:55,050 --> 00:02:58,750 Today. We're recording this on 17th. It's 24th 47 00:02:59,705 --> 00:03:03,385 of of June, but I'm also doing, 2 more, at 48 00:03:03,385 --> 00:03:07,145 near the ends of July August. And in addition 49 00:03:07,145 --> 00:03:10,610 to that, while we're shameless plugging away here, 50 00:03:10,990 --> 00:03:14,750 before we get to our very interesting guest, now I'm also bringing 51 00:03:14,750 --> 00:03:18,305 back my, day of Azure Data Factory as wildly 52 00:03:18,305 --> 00:03:21,845 popular. I delivered it at a couple of, conferences, 53 00:03:22,785 --> 00:03:26,405 international conferences, 22, 23. And, 54 00:03:27,290 --> 00:03:31,050 yeah. Let's see see if people are interested. What do you do Friday this 55 00:03:31,050 --> 00:03:34,570 afternoon Friday afternoons, Andy? Oh, there's this thing, Frank. Thanks for 56 00:03:34,570 --> 00:03:37,815 mentioning that. Totally free. We we gotta we're trying to get better at this. That's 57 00:03:37,815 --> 00:03:41,575 all. We do. Yeah. Data engineering Fridays. And if you go to data engineering 58 00:03:41,575 --> 00:03:45,330 fridays.com, you can learn more about that. Frank, you're doing a lot 59 00:03:45,330 --> 00:03:48,950 of stuff with I noticed with using the, encore 60 00:03:49,090 --> 00:03:52,735 replay feature in Restream. And it's 61 00:03:52,735 --> 00:03:56,515 right you you shared that with me. I started doing that with data engineering 62 00:03:56,655 --> 00:04:00,015 Fridays as well. But great a great way to, 63 00:04:00,550 --> 00:04:04,390 you know, to get your message out there. And, you 64 00:04:04,390 --> 00:04:08,045 know, I I had no idea replays would help. But my gosh. 65 00:04:08,045 --> 00:04:11,725 They really have. It's just a matter of just hitting the echo of I 66 00:04:11,725 --> 00:04:15,325 can't even talk. Algorithm the right way. Yeah. And Yeah. You know, 67 00:04:15,325 --> 00:04:19,120 maybe we can get the so I think it's a good segue, for our 68 00:04:19,120 --> 00:04:22,560 guest. Doctor Yossi, Keshet. He's the chief 69 00:04:22,560 --> 00:04:26,095 scientist at AIOLA, an AI powered tech 70 00:04:26,095 --> 00:04:29,395 company that automates business workflows 71 00:04:30,175 --> 00:04:33,950 by capturing spoken data. Yossi is also 72 00:04:33,950 --> 00:04:37,630 an associate professor at the Faculty of Electrical and Computer 73 00:04:37,630 --> 00:04:40,610 Engineering at the Technion in Israel. 74 00:04:41,405 --> 00:04:44,925 Yossi is an award winning scholar and has published over a 100 research 75 00:04:44,925 --> 00:04:48,285 papers about automated speech recognition and speech 76 00:04:48,285 --> 00:04:51,840 synthesis. Welcome to the show, Yossi. Hi. 77 00:04:51,840 --> 00:04:55,599 Nice for having me. Thank you for having me. Hey. No problem. No 78 00:04:55,599 --> 00:04:59,199 problem. We are very excited to have you. And, you're not just an 79 00:04:59,199 --> 00:05:02,725 academic, but you've also proven yourself in in actual enterprise. So 80 00:05:04,225 --> 00:05:06,865 which sounds really bad as I say that out loud, but I think you knew 81 00:05:06,865 --> 00:05:07,765 there was a compliment. 82 00:05:12,120 --> 00:05:15,500 But, so what is AIOLA? 83 00:05:16,755 --> 00:05:19,475 Can you tell me a little bit about that? Because I'm curious about that and 84 00:05:19,475 --> 00:05:23,015 and and workflows 85 00:05:23,075 --> 00:05:26,550 around spoken data. So 86 00:05:27,250 --> 00:05:30,930 Iola is a company that is aimed to target 87 00:05:30,930 --> 00:05:34,294 the, you know, the very basic and foundational 88 00:05:34,675 --> 00:05:38,435 industries. Maybe if I 89 00:05:38,435 --> 00:05:42,120 may, let's start with the a general scene of the 90 00:05:42,120 --> 00:05:45,879 automatic speech recognition now, and then you will understand where are YOLA stands because we 91 00:05:45,879 --> 00:05:49,725 have now open AI and everything is like we you 92 00:05:49,725 --> 00:05:53,345 can say we solve the AI problem. So it's not like that. 93 00:05:53,405 --> 00:05:57,060 So we are in a in a amazing shape in in 94 00:05:57,060 --> 00:06:00,840 terms of automatic speech recognition. So we we have a paper that shows 95 00:06:01,300 --> 00:06:04,974 that whisper, the model of OpenAI, is as good as humans in 96 00:06:04,974 --> 00:06:08,814 detecting and transcribing language when we speak about 97 00:06:08,814 --> 00:06:12,254 American English with noise, without noise, and 98 00:06:12,254 --> 00:06:15,740 also, l 2 speakers. That is the 99 00:06:15,740 --> 00:06:19,419 speakers of non non native American speakers of the 100 00:06:19,419 --> 00:06:23,014 language. And the the results are so whisper. The 101 00:06:23,014 --> 00:06:26,855 OpenAI model is the same as human listeners. And that is 102 00:06:26,855 --> 00:06:30,315 the main thing. But the thing is that 103 00:06:30,560 --> 00:06:34,259 when you come to industries, usually they have jargon, they have special words. 104 00:06:35,040 --> 00:06:38,745 And and those words are either rare in 105 00:06:38,745 --> 00:06:42,425 their language or they they they are not none 106 00:06:42,425 --> 00:06:46,025 word. It's like I don't know. I when I'm a medical doctor and would like 107 00:06:46,025 --> 00:06:49,860 to make a surgery surgery and I would like to transcribe what I'm saying during 108 00:06:49,860 --> 00:06:53,540 the surgery. I'm there isn't words that which are not 109 00:06:53,540 --> 00:06:57,235 often used or which are none, non English words. And 110 00:06:57,235 --> 00:07:00,995 in that case, those, automatic speech recognizer doesn't 111 00:07:00,995 --> 00:07:04,755 work at all. They don't detect those words. And in Ayala, this 112 00:07:04,755 --> 00:07:08,420 is our target to take those words, which are actually the most important word. Those 113 00:07:08,420 --> 00:07:11,960 are the jargon of the of the industry of the of the facility. 114 00:07:13,755 --> 00:07:17,595 So the goal is to help those industries to come 115 00:07:17,595 --> 00:07:21,435 up with the with the automatic speech recognition for 116 00:07:21,435 --> 00:07:24,800 reporting for transcribing speech. 117 00:07:25,660 --> 00:07:29,420 I have a question. When you say automatic, what what makes it automatic? Is 118 00:07:29,420 --> 00:07:33,185 it just kinda, what exactly does that mean? 119 00:07:34,525 --> 00:07:38,125 So automatic speech recognition today works very similar 120 00:07:38,285 --> 00:07:41,419 very, very similar to the way KJGPT works. 121 00:07:41,639 --> 00:07:45,400 KJGPT works on a model called transformer. It's an, deep 122 00:07:45,400 --> 00:07:49,135 learning architecture, which has, a 123 00:07:49,135 --> 00:07:52,035 history based on previous recurrent architectures. 124 00:07:53,135 --> 00:07:56,790 And it can predict, as as we all know, it can 125 00:07:56,790 --> 00:08:00,470 predict text amazingly. In speech recognition, automatic 126 00:08:00,470 --> 00:08:04,215 speech recognition, it's almost the same thing, but there is another 127 00:08:04,215 --> 00:08:08,055 component, to the to the to the 128 00:08:08,215 --> 00:08:11,640 this transformer, which is which is called encoder. 129 00:08:12,020 --> 00:08:15,860 This this part take the speech and actually transfer it to 130 00:08:15,860 --> 00:08:18,520 a great representation that can be used 131 00:08:19,625 --> 00:08:23,384 with this, with this, let's call it with this with the other side, with 132 00:08:23,384 --> 00:08:26,504 this, GPT together. Together, they can, 133 00:08:27,065 --> 00:08:30,490 transcribe speech in, as I described, in a very good 134 00:08:30,490 --> 00:08:33,929 way, as good as humans in some 135 00:08:33,929 --> 00:08:37,585 cases. I will say, like, 136 00:08:37,585 --> 00:08:40,725 I've been messing around with the app that's on the phone, 137 00:08:41,825 --> 00:08:44,785 for, chat g p chat gbt, and, 138 00:08:45,580 --> 00:08:49,180 I use the the voice interaction feature. It is 139 00:08:49,180 --> 00:08:52,400 amazingly good at getting rid of the umms, the ahs, 140 00:08:52,540 --> 00:08:56,035 the scatterbrain thoughts that I sometimes have when I talk to it. 141 00:08:56,255 --> 00:09:00,015 Like, it it could kinda really distill a lot of 142 00:09:00,015 --> 00:09:03,839 things. Like, I'm impressed with it. It's it's really gotten last time I 143 00:09:03,839 --> 00:09:07,519 did anything serious with speech recognition was probably, like, maybe 4 years 144 00:09:07,519 --> 00:09:10,980 ago, and it's really improved. Like, I mean, orders of magnitude 145 00:09:11,315 --> 00:09:14,515 than I thought. I mean, it's it's it's it's almost at Star Trek level. You 146 00:09:14,515 --> 00:09:18,355 know? I'm not sure 147 00:09:18,355 --> 00:09:21,760 in those it depends on the company if it's Apple or 148 00:09:21,760 --> 00:09:25,380 Google. And I'm not sure which they don't declare 149 00:09:25,520 --> 00:09:29,315 which models they use. I think, personally, they don't use this whisper or 150 00:09:29,315 --> 00:09:32,995 the latest model that we have for automatic speech recognition that 151 00:09:32,995 --> 00:09:36,569 is transcribing speech. And the goal is a little bit different 152 00:09:36,569 --> 00:09:39,529 in the in the phone. You actually want to maybe Right. Make, 153 00:09:40,329 --> 00:09:42,910 make notes, send an email, send a text message, 154 00:09:44,135 --> 00:09:46,875 and maybe the vocabulary the vocabulary is less 155 00:09:48,135 --> 00:09:51,895 less defined. There is another problem with 156 00:09:51,895 --> 00:09:55,680 the phones. Oh, no. Go ahead. I want to call my 157 00:09:55,680 --> 00:09:59,520 friend. His name is xi, and 158 00:09:59,520 --> 00:10:03,295 the last name is CHUNG. How do you pronounce it? 159 00:10:03,295 --> 00:10:06,895 What what do you do with that? I'm gonna say he or chi or 160 00:10:07,215 --> 00:10:10,815 so there is a there is a problem of proper name and how do you 161 00:10:10,815 --> 00:10:14,194 define them. And this is a completely different problem. It's still an open problem, and 162 00:10:14,194 --> 00:10:15,720 the goal is a little bit different. So 163 00:10:18,705 --> 00:10:22,485 it's when we assessing the quality of those models, it's 164 00:10:22,945 --> 00:10:26,785 a little bit different than the assessment of just spoken language 165 00:10:26,785 --> 00:10:30,600 like what we do now. No. I mean, that's a great point. I mean, my 166 00:10:30,600 --> 00:10:34,140 last name has, you know, technically is Lavin. 167 00:10:34,840 --> 00:10:38,165 But, you know, growing up for for reasons many, 168 00:10:38,705 --> 00:10:42,485 big and small, it became Lavinia. And like, so, like, 169 00:10:42,705 --> 00:10:46,330 the phone, depending on if it's Android or an Apple, it will, it 170 00:10:46,330 --> 00:10:49,230 will he gets confused pretty easily. 171 00:10:50,650 --> 00:10:54,490 And that is an interesting point. Some names, Andy is lucky to have an 172 00:10:54,490 --> 00:10:56,665 easy name for the, the system. 173 00:10:58,405 --> 00:11:02,185 But not everybody does. So I understand that. Sure. 174 00:11:02,725 --> 00:11:06,490 I also wanna double click on American 175 00:11:06,490 --> 00:11:09,930 English. You you you said that a bunch of times. Like, is there is there 176 00:11:09,930 --> 00:11:13,610 an inherent bias in these model trainings because these are done by American 177 00:11:13,610 --> 00:11:17,375 companies? Yes. There is. Okay. The 178 00:11:17,375 --> 00:11:21,215 day the data is mostly of American English. The research institutes 179 00:11:21,215 --> 00:11:24,960 are mostly American. So the reason maybe I don't know 180 00:11:24,960 --> 00:11:28,800 if you'd call it you call it inherent or implicit bias, but there is a 181 00:11:28,800 --> 00:11:29,860 bias, definitely. 182 00:11:33,035 --> 00:11:36,815 We are investigating, by the way, the the intelligibility 183 00:11:37,035 --> 00:11:40,690 of speech in some cases And what is the intelligibility of 184 00:11:40,690 --> 00:11:44,290 of American listener versus the inter intelligibility of 185 00:11:44,290 --> 00:11:47,510 myself, which I'm not American listener, but I I know English. 186 00:11:48,654 --> 00:11:51,055 What is the best, what is the best, double quote speaker? What is the best 187 00:11:51,055 --> 00:11:52,995 listener? How can we transform those 188 00:11:57,290 --> 00:12:01,130 to speech recognizer? How can we transform those to assessing the 189 00:12:01,130 --> 00:12:04,890 quality of speech? What does it mean? What does it mean about the pathologies in 190 00:12:04,890 --> 00:12:08,725 speech? And this is ongoing research on 191 00:12:08,725 --> 00:12:12,105 this on this field. Interesting. 192 00:12:12,324 --> 00:12:16,140 I I often wonder, like, you know, what it's not just English. 193 00:12:16,140 --> 00:12:19,660 Right? Like, you know, if you listen to Spanish, like, there's different dialects of 194 00:12:19,660 --> 00:12:23,260 Spanish. Right? Even even German. You know, I'm sure 195 00:12:23,260 --> 00:12:26,875 there's, you know, plenty of dialects of all these languages and, 196 00:12:26,875 --> 00:12:30,235 like, how do you the training of a 197 00:12:30,235 --> 00:12:33,740 model that where it can get to be as good at 198 00:12:33,740 --> 00:12:37,420 understanding x and x versus x and y versus, you know, 199 00:12:37,420 --> 00:12:41,105 the base language, the base standard. I don't know. That's 200 00:12:41,105 --> 00:12:44,945 fascinating. It seems like it seems like it could be an endless loop of, like, 201 00:12:45,185 --> 00:12:48,625 training. It it is. Indeed, it 202 00:12:48,625 --> 00:12:52,400 is. And when we train, there is another so I'm I'm 203 00:12:52,460 --> 00:12:55,900 working on deep learning and AI. And what we found out 204 00:12:55,900 --> 00:12:59,625 that it it may it may be the case that if you train 205 00:12:59,625 --> 00:13:03,305 on 1 language, huge amount of data from 1 language, let's say 206 00:13:03,305 --> 00:13:06,940 American English, but then train on less data on Spanish, 207 00:13:07,320 --> 00:13:11,000 you actually get you get some advantage of training from 208 00:13:11,000 --> 00:13:14,805 from the American English. So, again, in this modern whisper of 209 00:13:14,805 --> 00:13:18,345 OpenAI, most of the data is American English, but, 210 00:13:18,485 --> 00:13:20,985 actually, other languages are really great. 211 00:13:22,230 --> 00:13:26,070 Again, Spanish is amazing. So maybe like 212 00:13:26,070 --> 00:13:29,830 humans maybe like humans as we learn more and more languages, it's easier 213 00:13:29,830 --> 00:13:33,255 for us. This is very interesting, point. 214 00:13:33,955 --> 00:13:37,714 No. That's an interesting idea because I know, like, I never 215 00:13:37,714 --> 00:13:40,980 understood American English grammar, American or otherwise, 216 00:13:41,680 --> 00:13:45,460 until I studied a foreign language. And then when I studied it, it was German. 217 00:13:45,680 --> 00:13:49,455 And, you know, German kept a lot of the archaic things that 218 00:13:49,455 --> 00:13:53,055 are in English and kept them and kept make kept them, 219 00:13:53,695 --> 00:13:57,540 made continue to keep them important. Like in English, you know, who 220 00:13:57,540 --> 00:14:00,840 and whom used to confuse the you know what out of me. 221 00:14:01,060 --> 00:14:04,900 Right? But when I when I learned in German about different cases and things 222 00:14:04,900 --> 00:14:08,675 like that, I was like, oh, that's why it is. Right? So, 223 00:14:08,675 --> 00:14:11,715 like, all these things that just like you said, like, learning another 224 00:14:12,970 --> 00:14:16,110 having more data or data from another point of view, I suppose, 225 00:14:16,810 --> 00:14:20,430 or another way to look at the world help me look at my world 226 00:14:20,889 --> 00:14:24,315 a little better. Maybe maybe that's how 227 00:14:24,315 --> 00:14:26,175 AI will work too. I don't know. 228 00:14:28,635 --> 00:14:32,250 Maybe. We don't know. We we actually have a guess about that 229 00:14:32,250 --> 00:14:35,769 because it those networks actually solve an optimization problem, 230 00:14:35,769 --> 00:14:38,589 mathematical optimization problem. It's a problem that 231 00:14:40,815 --> 00:14:44,654 that is, we define it with equation, and we need to have 232 00:14:44,654 --> 00:14:48,015 a computer running and solve it. The equation is 233 00:14:48,015 --> 00:14:51,610 overtraining set of examples. So it's 1 234 00:14:51,610 --> 00:14:54,910 1 person say that, another person said something else. 235 00:14:55,450 --> 00:14:59,085 And what happened is that when, again, when we have 236 00:14:59,405 --> 00:15:00,865 a large amount of data, 237 00:15:03,325 --> 00:15:07,165 it seems that those those networks get to an amazing place. 238 00:15:07,165 --> 00:15:10,910 So this this, algorithm, this whisper or other 239 00:15:10,910 --> 00:15:14,670 algorithms, it's really from the recent years, like 2, 3 years. 240 00:15:14,670 --> 00:15:18,175 That's it. We it's they they perform amazingly 241 00:15:18,315 --> 00:15:22,155 amazingly, with the with the 242 00:15:22,155 --> 00:15:25,550 same with the same mechanism, not with the same amount of 243 00:15:25,550 --> 00:15:29,070 data. Yeah. That's that's that's the 244 00:15:29,070 --> 00:15:32,850 fascinating aspect of all of this. It's just that some of these things just seem 245 00:15:33,555 --> 00:15:36,615 some problems seem harder than they ought to be, 246 00:15:37,235 --> 00:15:41,075 and then some solutions to problems seem way more effective than they 247 00:15:41,075 --> 00:15:44,440 ought to be. It's an interesting also to say 248 00:15:45,620 --> 00:15:49,380 it's always the case that we so Whisper, OpenAI Whisper, was trained 249 00:15:49,380 --> 00:15:53,005 on 600000 hours of speech. But this is 250 00:15:53,005 --> 00:15:56,505 way, way much more than just a kid learning a language. 251 00:15:56,885 --> 00:16:00,645 Kid language learning a language exposed to way much less hours of 252 00:16:00,645 --> 00:16:04,040 speech, less less accurate, less, 253 00:16:04,760 --> 00:16:07,660 coherent. And this is something, 254 00:16:08,685 --> 00:16:12,305 Nom Chomski raised years ago, like, 50 years ago. 255 00:16:12,925 --> 00:16:16,545 And it's still an open question. Like, if we can make those 256 00:16:16,950 --> 00:16:19,290 system works better, if we know the language, 257 00:16:22,070 --> 00:16:25,130 I guess you learn German faster than any 258 00:16:25,605 --> 00:16:28,264 machine that works today. 259 00:16:30,565 --> 00:16:34,240 That's yeah. It's it's and I'm glad you mentioned Noam 260 00:16:34,240 --> 00:16:37,600 Chomsky because that kinda was like so for those who don't know, Noam 261 00:16:37,600 --> 00:16:40,900 Chomsky is, among other things, a noted linguist scholar. 262 00:16:42,245 --> 00:16:46,084 I highly recommend you do a search on him because that's a that's a 263 00:16:46,084 --> 00:16:48,894 good Wikipedia rabbit hole to fall into. But, 264 00:16:50,529 --> 00:16:54,130 how much does linguistics come up in this? Right? Because I think 265 00:16:54,130 --> 00:16:57,570 what's fascinating about this field for me is a lot 266 00:16:57,570 --> 00:17:01,355 of, my grandfather, my great grandfather 267 00:17:01,575 --> 00:17:05,095 was a a linguistic professor. And, you know, as the 268 00:17:05,095 --> 00:17:08,819 family lore goes, I never met him. He died decade or 2 before I was 269 00:17:08,819 --> 00:17:12,579 born. He spoke, like, 12 languages. He was a professor of, like, 5 270 00:17:12,579 --> 00:17:16,260 or 6. And, you know, a lot of people in my family 271 00:17:16,260 --> 00:17:19,835 seem to have on that side of the family seem to be gifted in language. 272 00:17:20,534 --> 00:17:23,974 And 1 of the fields I was tempted to to study in 273 00:17:23,974 --> 00:17:27,660 university was linguistics. And I just find 274 00:17:27,660 --> 00:17:30,880 it interesting how there's 275 00:17:31,340 --> 00:17:35,145 a now a Venn diagram now is much larger 276 00:17:35,145 --> 00:17:38,205 than it used to be in terms of linguistics and computer science. 277 00:17:38,825 --> 00:17:42,665 So what are your thoughts on? Like, how much does like, 278 00:17:42,665 --> 00:17:46,510 if you're if you have a 279 00:17:46,510 --> 00:17:50,270 company like AIO. Right? Like, how many people are, you know, honest to 280 00:17:50,270 --> 00:17:54,115 goodness, linguists versus computer scientists and and AI engineers? 281 00:17:55,774 --> 00:17:59,375 So there is there is no no linguists there. Oh, 282 00:17:59,375 --> 00:18:02,960 really? Okay. There are no linguists. But I have to tell you, so there was 283 00:18:02,960 --> 00:18:06,340 a professor called Freddie Frederick, Jelinek. He was the 284 00:18:06,639 --> 00:18:10,419 head of language, research at the John Hopkins University 285 00:18:10,480 --> 00:18:13,605 at Baltimore. He was amazing. He was 1 of the smartest, 286 00:18:14,065 --> 00:18:17,205 people on earth. And he said he was 287 00:18:18,880 --> 00:18:22,660 developed many of the speech recognition algorithms. He said, 288 00:18:22,800 --> 00:18:26,400 every time I fire a linguist, the performance of speech recognizer goes 289 00:18:26,400 --> 00:18:26,775 up. 290 00:18:32,855 --> 00:18:36,400 And this is, this is embarrassing. But I've been I 291 00:18:36,640 --> 00:18:40,320 made myself, 1st, really like 292 00:18:40,320 --> 00:18:44,000 linguistics. I really like cognitive sciences, and I really 293 00:18:44,000 --> 00:18:47,745 try to combine it with with my work. But it's really 294 00:18:47,745 --> 00:18:51,365 amazing that we don't have all those AI system 295 00:18:51,505 --> 00:18:55,220 don't have any of that. So you don't train CEGPT 296 00:18:55,280 --> 00:18:59,040 to what is a noun, what is a verb, what is anything. You don't train 297 00:18:59,040 --> 00:19:01,540 speech that this is the 298 00:19:02,655 --> 00:19:06,495 this is the you don't you don't use linguist. You don't use this is 299 00:19:06,495 --> 00:19:10,270 the prominent word. This is the end of the sentence. It just happened 300 00:19:10,270 --> 00:19:14,110 by huge amount of data. And 301 00:19:14,110 --> 00:19:17,630 this is interesting. This is somehow contradict Noam Chomsky who said that 302 00:19:17,630 --> 00:19:21,365 there there is a universal grammar. There is a 303 00:19:21,365 --> 00:19:24,885 we are born innate with language. There is a 304 00:19:24,885 --> 00:19:28,710 maybe some black box in our brain which 305 00:19:28,710 --> 00:19:32,550 is tuned to learn a language. And, 306 00:19:33,350 --> 00:19:37,030 we are not sure about that. There is no direct proof if it's correct or 307 00:19:37,030 --> 00:19:40,655 no. We are born with language. We are as humans, we're 308 00:19:40,655 --> 00:19:44,495 born with language. We this is part of our, human being. 309 00:19:44,495 --> 00:19:47,875 We are not born with written language. So written language was invented. 310 00:19:48,800 --> 00:19:52,640 The spoken language is something like like a zebra 311 00:19:52,640 --> 00:19:56,015 has stripes. This is this is our nature, and this is 312 00:19:56,015 --> 00:19:59,615 interesting. This is not happening not happening in 313 00:19:59,615 --> 00:20:03,395 AI. The best success that didn't have linguist, they don't have any 314 00:20:03,759 --> 00:20:06,740 restriction of what should be say or not. 315 00:20:10,399 --> 00:20:13,139 Maybe maybe AI will be a tool to somehow 316 00:20:15,185 --> 00:20:18,945 make the linguist research more effective and 317 00:20:18,945 --> 00:20:22,645 try to understand what happened in the brain, what happened in the cognition part. 318 00:20:23,850 --> 00:20:27,450 But I would like to tell you about another research we are preparing here, which 319 00:20:27,450 --> 00:20:30,410 is really amazing. 1 of the thing is that we have 320 00:20:31,175 --> 00:20:34,555 so there is this JGPT. It's a language model. 321 00:20:35,015 --> 00:20:38,775 We also have something in the brain. It's also neural network. 322 00:20:38,775 --> 00:20:42,600 And we when we try to compare them, there is a huge 323 00:20:42,600 --> 00:20:46,280 correlation between the the what happened in the artificial neural 324 00:20:46,280 --> 00:20:49,395 network of GPT and the neural 325 00:20:50,175 --> 00:20:54,015 biological neural network in the brain. And, it was 326 00:20:54,015 --> 00:20:57,850 shown, several years ago, and here we 327 00:20:57,850 --> 00:21:01,289 show it again with, with this, with the most modern, 328 00:21:01,769 --> 00:21:05,285 automatic speech recognizers. So this is 329 00:21:05,745 --> 00:21:09,425 a phenomenal post correlation between the artificial and the 330 00:21:09,425 --> 00:21:13,159 neural mechanisms. I was gonna ask about that 331 00:21:13,159 --> 00:21:17,000 because I'm I'm familiar with, you know, at least the abstracts of 332 00:21:17,000 --> 00:21:20,775 the research, from a few years ago and now. And 333 00:21:20,775 --> 00:21:23,835 I was curious if there had been any new correlations 334 00:21:24,775 --> 00:21:28,615 or, you know, or new research, new connections that have been made 335 00:21:28,615 --> 00:21:32,150 between machines learning languages 336 00:21:32,610 --> 00:21:36,289 and the way our brains work. It sounds like 337 00:21:36,289 --> 00:21:37,110 that's true. 338 00:21:39,695 --> 00:21:43,475 So we try to we just initiate, man, 339 00:21:43,934 --> 00:21:47,559 a research here in my lab about that. There was 340 00:21:48,340 --> 00:21:52,179 some French guys from, mainly King 341 00:21:52,179 --> 00:21:54,600 and his colleague at, Meta. And 342 00:21:57,995 --> 00:22:01,675 and I forgot the university in France. So they 343 00:22:01,675 --> 00:22:05,490 show that there is those correlation. They show simple correlation. What we 344 00:22:05,730 --> 00:22:09,010 they show it with LLM, with language model. What we show is a little bit 345 00:22:09,010 --> 00:22:12,705 different. We show correlation with automatic speech 346 00:22:12,705 --> 00:22:16,465 recognition. So we ask people under fMRI, under MRI. 347 00:22:16,465 --> 00:22:19,900 They're we scan their brain at some 348 00:22:19,900 --> 00:22:23,360 resolution, and we try to find correlation with their brain activity 349 00:22:23,420 --> 00:22:26,240 during reading and during speaking aloud, 350 00:22:27,305 --> 00:22:31,145 and ask what is the correlation with the the best model we know for 351 00:22:31,145 --> 00:22:33,965 speech recognition. And then there are correlation. 352 00:22:35,920 --> 00:22:39,360 I have to say that there is a mechanism in the transforming this 353 00:22:39,360 --> 00:22:42,965 architecture of neural network. There is a mechanism called attention. This 354 00:22:42,965 --> 00:22:46,645 mechanism allow those model to to have the connection between 355 00:22:46,645 --> 00:22:50,420 worlds and themselves. So, I'm eating an 356 00:22:50,420 --> 00:22:54,100 apple. It was delicious. So it refers to the apple. 357 00:22:54,100 --> 00:22:57,780 Okay? So there is attention mechanism. This what makes those 358 00:22:57,780 --> 00:23:01,175 model amazing. So there is attention mechanism, I guess, in the 359 00:23:01,175 --> 00:23:04,775 brain. So we try to correlate the this attention mechanism in 360 00:23:04,775 --> 00:23:08,270 the models and compare it to what the activity in the brain. We don't have 361 00:23:08,270 --> 00:23:12,030 results yet, but it seems promising. And we also ask 362 00:23:12,030 --> 00:23:15,250 another question. What if you don't read aloud? What if you read 363 00:23:15,695 --> 00:23:19,475 like silent reading? What if you have dyslexia? What if you have, 364 00:23:19,935 --> 00:23:23,620 other type of, pathology? What 365 00:23:23,620 --> 00:23:27,460 what are the correlation then? So this is fascinating. So and 366 00:23:27,460 --> 00:23:31,220 there is correlation. I don't I don't know still what what's going to happen 367 00:23:31,220 --> 00:23:34,675 with that. But I I know the pathologist, but it's unbelievable, the 368 00:23:34,675 --> 00:23:38,275 correlation. That that is really exciting, 369 00:23:38,275 --> 00:23:41,500 especially when you're examining things like dyslexia, 370 00:23:41,640 --> 00:23:45,340 which is considered, you know, not normal, 371 00:23:45,400 --> 00:23:48,845 or maybe that's not the right term for it, but a 372 00:23:48,845 --> 00:23:52,545 challenge at a minimum. The cool the cool kids call that neurodivergent 373 00:23:52,765 --> 00:23:56,605 now. I think Neurodivergent. Thank you, Frank. So when you're studying, you 374 00:23:56,605 --> 00:24:00,270 know, when you're studying that sort, I'm wondering if there's a place for 375 00:24:00,270 --> 00:24:02,850 that, in in the artificial. 376 00:24:04,910 --> 00:24:08,235 I'm curious. What what do you mean? Can you 377 00:24:08,715 --> 00:24:12,015 So, yeah, is there is is there any benefit 378 00:24:12,635 --> 00:24:16,310 to, I say, transferring the thought processes 379 00:24:16,450 --> 00:24:20,290 of people who are neurodivergent and and automating that 380 00:24:20,290 --> 00:24:23,895 and making that part of the, you know, 381 00:24:23,895 --> 00:24:27,475 the the language model or or speech recognition? 382 00:24:29,830 --> 00:24:33,190 Yeah. I think so. I think so. 1st, it's a it's a tool 383 00:24:33,190 --> 00:24:36,870 to to an to analyze what happened in the 384 00:24:36,870 --> 00:24:38,695 brain. Yeah. What happened 385 00:24:40,595 --> 00:24:44,355 but it's very difficult. So we don't, we don't have any debugger for the build 386 00:24:44,434 --> 00:24:47,410 the brain. We don't see the code of the brain. We don't see that this 387 00:24:47,410 --> 00:24:51,250 function doesn't work. And it's, most of the work 388 00:24:51,250 --> 00:24:53,429 is to design the experiment and 389 00:24:55,035 --> 00:24:58,794 and it's really amazing. In our design, we have the 390 00:24:58,794 --> 00:25:02,635 same so as yet as I told you, I'm asking people to read aloud 391 00:25:02,635 --> 00:25:05,230 and compare it to what automatic speech recognition, 392 00:25:06,410 --> 00:25:09,850 is plan is, supposed to do. But I'm 393 00:25:09,850 --> 00:25:13,515 also asking people to read silently, and then I follow 394 00:25:13,515 --> 00:25:17,195 their eyes. I have a make a make a machine that follows their eyes, and 395 00:25:17,195 --> 00:25:20,880 I know where where is the where like, III 396 00:25:20,880 --> 00:25:24,480 track their eyes and I see which wall they are reading 397 00:25:24,480 --> 00:25:28,320 now. And I can and I can use that to follow 398 00:25:28,320 --> 00:25:32,065 what what they read. But in order to operate that on a speech 399 00:25:32,065 --> 00:25:35,825 recognizer model, I need the speech. So it's during the design of 400 00:25:35,825 --> 00:25:39,510 the experiment, I need artificial speech or I need them to to read aloud 401 00:25:39,510 --> 00:25:43,350 afterwards. It's a it's a big, it's a big question 402 00:25:43,350 --> 00:25:45,770 how to do that properly and how to 403 00:25:46,935 --> 00:25:50,075 make things happen, but definitely walking with 404 00:25:50,535 --> 00:25:54,315 people with, with problems first to help them. 405 00:25:55,070 --> 00:25:58,769 And second, to understand them. And 3rd, to maybe make 406 00:26:00,350 --> 00:26:03,169 understand the brain and make, AI better. 407 00:26:04,225 --> 00:26:07,985 I also think, like, stroke victims, right, could benefit down the line 408 00:26:07,985 --> 00:26:11,825 from a better understanding of lang language models. Right? Like, maybe there would be some 409 00:26:11,825 --> 00:26:15,500 kind of therapy that could be directed to that. I think I think it's 410 00:26:15,500 --> 00:26:19,340 fascinating. I always love those fields where they touch upon more than 1 thing. 411 00:26:19,340 --> 00:26:23,065 Right? This isn't just math. This isn't just computer science. Like, it's linguistics. But, 412 00:26:23,065 --> 00:26:26,105 you know, it's a little bit of everything. It's like a giant, like, pot of 413 00:26:26,105 --> 00:26:28,985 stew that you just throw a bunch of stuff in, and it all kind of 414 00:26:28,985 --> 00:26:32,830 mixes. And, like, it's kind of like, almost like intellectual gumbo, 415 00:26:32,830 --> 00:26:34,850 I guess, would be the word. Right? But, 416 00:26:37,640 --> 00:26:40,345 what what, 417 00:26:42,005 --> 00:26:45,605 what drove you to make, your your your 418 00:26:45,605 --> 00:26:49,250 your company? Like, what what was the driving force to 419 00:26:49,710 --> 00:26:52,450 say, hey. You know, we have 420 00:26:54,510 --> 00:26:57,924 I remember many, many years ago in an office, and you would always see 421 00:26:57,924 --> 00:27:01,225 doctors talking into these little, like, miniature recorders. 422 00:27:01,765 --> 00:27:05,320 Right? In the olden days, they would go off to 423 00:27:05,320 --> 00:27:08,760 some data center somewhere and somebody would not data center, but, like, 424 00:27:08,760 --> 00:27:12,220 some piping center, call center where people would 425 00:27:12,280 --> 00:27:16,095 transcribe that. You know, obviously, that is now an artifact of 426 00:27:16,095 --> 00:27:19,155 the past as these models have gotten better. 427 00:27:22,289 --> 00:27:25,730 What what was the goal in in in, your 428 00:27:25,730 --> 00:27:29,570 company to say we can do this better? What what was the the that breakthrough 429 00:27:29,570 --> 00:27:33,205 moment of, like, here's here's what the industry already does. Here's how we can do 430 00:27:33,205 --> 00:27:36,345 it better. So there is 431 00:27:36,885 --> 00:27:40,490 so we all know Check GPT, and it influence our life. We search now 432 00:27:40,490 --> 00:27:43,950 instead of Google, we search with GPT and it's amazing. It's unbelievable. 433 00:27:45,130 --> 00:27:48,890 So I thought, what about the very fundamental industries? What 434 00:27:48,890 --> 00:27:49,390 about, 435 00:27:52,945 --> 00:27:56,705 like, when you check-in when you, check an airplane, you 436 00:27:56,705 --> 00:28:00,450 use a special jargon. You cannot touch anything. You cannot 437 00:28:00,450 --> 00:28:04,230 leave even a pen there because otherwise the the plane wouldn't be, 438 00:28:04,930 --> 00:28:08,505 valid for flight. What about industries like the food 439 00:28:08,505 --> 00:28:12,345 industries when you need to report, the process? You 440 00:28:12,345 --> 00:28:15,865 have gloves, you cannot touch an iPad, you cannot barely 441 00:28:15,865 --> 00:28:19,549 write. And what about, other industries 442 00:28:19,549 --> 00:28:23,350 like, maybe the cheap technology when you make nanotechnologies and 443 00:28:23,350 --> 00:28:26,090 when you make chips, you make, you know, 444 00:28:26,765 --> 00:28:30,465 silicon chips and silicon 445 00:28:30,605 --> 00:28:34,365 first. So you need you you are cover all. 446 00:28:34,365 --> 00:28:38,049 You are with gloves. You need to report the process. It's a all 447 00:28:38,049 --> 00:28:41,649 those industries has this have special jargons. They use special 448 00:28:41,649 --> 00:28:45,269 terms to describe what they're doing. They don't have access to 449 00:28:46,595 --> 00:28:47,475 to to write something, 450 00:28:51,235 --> 00:28:54,909 and they are very limited in the way they they provide. And on the other 451 00:28:54,909 --> 00:28:58,429 end, we had speech recognition, but speech recognition doesn't work on 452 00:28:58,429 --> 00:29:02,030 those jargon world. Those jargon world are actually the 453 00:29:02,030 --> 00:29:05,535 most important to those industries, and this was the goal for 454 00:29:05,535 --> 00:29:07,955 Iola. So what we do is we operate, 455 00:29:08,895 --> 00:29:12,549 automatic speech recognition, the best automatic speech recognition, 456 00:29:12,549 --> 00:29:16,169 but we also operate something else. We also operate something called keyword spotting. 457 00:29:16,870 --> 00:29:20,625 It's another deep network, which is focused 458 00:29:20,625 --> 00:29:24,385 on detecting only the jargon words. So you can define those jargon 459 00:29:24,385 --> 00:29:28,150 words in advance. You don't need to train them. You you can 460 00:29:28,150 --> 00:29:31,910 define them, and it they all work together. They work like, as a 461 00:29:31,910 --> 00:29:35,610 complimentary, couple to make a 462 00:29:36,685 --> 00:29:40,525 very robust prediction, and we can detect those, 463 00:29:41,085 --> 00:29:44,685 jargon words and make reporting on on on on the 464 00:29:44,685 --> 00:29:48,380 process, without just by speaking. So it 465 00:29:48,380 --> 00:29:50,800 can it can use in any industries, 466 00:29:51,900 --> 00:29:55,605 any, industry that doesn't 467 00:29:55,605 --> 00:29:59,125 have access to the most modern AI system, the speech 468 00:29:59,125 --> 00:30:02,505 recognizer wouldn't walk there. They have problems, like, 469 00:30:03,530 --> 00:30:06,270 writing and formulating their reports. 470 00:30:06,970 --> 00:30:10,809 Yeah. So I'm curious how those work together. You mentioned 471 00:30:10,809 --> 00:30:13,955 that you've got the speech recognizer. You've got the keyword, 472 00:30:15,055 --> 00:30:18,735 engine. Are they 2 separate engines that are just always running 473 00:30:18,735 --> 00:30:22,169 maybe agents, running at the same time or are 474 00:30:22,169 --> 00:30:25,850 they encapsulated, say, is the speech 475 00:30:25,850 --> 00:30:29,655 recognizer does the speech recognizer have a, you know, a 476 00:30:29,655 --> 00:30:33,415 subset or a a function built into it to do the 477 00:30:33,415 --> 00:30:37,230 keyword recognition? So just to 478 00:30:37,230 --> 00:30:40,909 be sure, those keywords in some industries are not are 479 00:30:40,909 --> 00:30:44,350 not are not English words. So it can be a word which nobody 480 00:30:44,350 --> 00:30:47,784 knows about. It was not shown in the in 481 00:30:47,784 --> 00:30:51,625 the, like, in the Internet, like, JGPT strain on the data over the 482 00:30:51,625 --> 00:30:55,080 Internet. There are some walls that are not not there. This is 483 00:30:55,080 --> 00:30:58,600 your, proprietary company. You have invented a wall to 484 00:30:58,600 --> 00:31:02,225 describe what is the this, part of the engine. So 485 00:31:02,465 --> 00:31:06,145 Yeah. So what we so we have this keyword spotting. It was it it 486 00:31:06,145 --> 00:31:09,649 is trained to detect keyword in general. They are defined by, 487 00:31:10,049 --> 00:31:13,809 by text and it operates. We have 2 model for preparation. 1 of them 488 00:31:13,809 --> 00:31:17,225 works on the this encoder part of 489 00:31:17,225 --> 00:31:20,985 the of the automatic speech recognition, and then it guides. 490 00:31:20,985 --> 00:31:23,645 It's still the speech recognition towards the correct 491 00:31:25,389 --> 00:31:28,610 transcription. And there is another mode, which is, 492 00:31:29,070 --> 00:31:32,510 our self, encode our self representation of 493 00:31:32,510 --> 00:31:36,045 speech, and then it also guides the automatic speech 494 00:31:36,045 --> 00:31:39,565 recognition to a better, location and to detect those 495 00:31:39,565 --> 00:31:42,865 words. And, actually, we can show that you can buy combine 496 00:31:43,210 --> 00:31:47,050 any word can be from different languages, and we can 497 00:31:47,050 --> 00:31:50,730 detect them, like, almost 100% correct, those jargon 498 00:31:50,730 --> 00:31:54,285 words. That was that was going sorry. Go ahead. 499 00:31:55,065 --> 00:31:58,905 No. No. No. Sorry. That no. That's okay. That that makes perfect 500 00:31:58,905 --> 00:32:02,480 sense now, what you just said about the languages using 501 00:32:02,480 --> 00:32:06,160 multiple languages, you know, English plus all of the 502 00:32:06,160 --> 00:32:09,765 other languages because sometimes 503 00:32:09,825 --> 00:32:13,265 people will struggle if their English as a second 504 00:32:13,265 --> 00:32:16,785 language speaker. They'll struggle to find the right 505 00:32:16,785 --> 00:32:20,540 English word, and they'll substitute a word from their native language. 506 00:32:20,840 --> 00:32:24,460 And in other cases, they'll be perhaps teaching 507 00:32:25,000 --> 00:32:28,835 on a topic, and they may revert back 508 00:32:28,835 --> 00:32:32,595 to an older language, Greek, Latin, something 509 00:32:32,595 --> 00:32:36,070 like that. That may be part of the, the 510 00:32:36,070 --> 00:32:39,509 lecture or, you know, I could see that in 511 00:32:39,509 --> 00:32:43,350 medicine. I could see it in, you know, all all sorts 512 00:32:43,350 --> 00:32:46,875 of literature studies. I could see a lot of that. And that 513 00:32:47,015 --> 00:32:50,615 that kinda clicked for me as you were saying that that makes sense that you 514 00:32:50,615 --> 00:32:54,130 would have additional languages. Yeah. I also wonder, like, in in 515 00:32:54,130 --> 00:32:57,890 also conversational context. Right? Like, you know, Spanglish is a 516 00:32:57,890 --> 00:33:01,335 thing. Frankel is is the French and 517 00:33:01,335 --> 00:33:05,015 English kinda mashed together, and I know that other language 518 00:33:05,095 --> 00:33:08,855 whenever you have 2 groups of people kinda come together, like, you know, there's always 519 00:33:08,855 --> 00:33:12,580 some kind of weird mix of language that that kinda 520 00:33:12,580 --> 00:33:16,420 just evolves either naturally or forced. I mean, that's Right. That's another 521 00:33:16,420 --> 00:33:20,205 debate. Are you thinking belt or creole? I know we're Belter, you know, I 522 00:33:20,205 --> 00:33:23,985 wasn't going there, but that that's a that's an excellent example. 523 00:33:24,125 --> 00:33:27,725 So, Yosie looks very confused. So so there's a series of 524 00:33:27,725 --> 00:33:31,549 books, called The Expanse. It was an excellent TV show 525 00:33:31,549 --> 00:33:35,150 for about 6 seasons, and it's basically set, 2, 526 00:33:35,150 --> 00:33:36,530 300 years in the future. 527 00:33:38,715 --> 00:33:42,554 And as humans colonize the asteroid belt, 528 00:33:42,554 --> 00:33:46,150 their people from all over the world kinda all end up living 529 00:33:46,150 --> 00:33:49,990 together. So, like, the the Belter Creole language is this is a 530 00:33:49,990 --> 00:33:53,755 creole of, you know, literally dozens of languages. Right? 531 00:33:53,755 --> 00:33:57,275 So, like, it'll switch from, you know, Hindi to Arabic to, 532 00:33:57,915 --> 00:34:01,515 English to French to there's even some German in there. I've heard some of that. 533 00:34:01,515 --> 00:34:04,980 Like, and there are these kind of these weird mixes of things. Right? So they'll 534 00:34:05,039 --> 00:34:08,880 say the the word for the Belter people, like, 535 00:34:08,880 --> 00:34:12,675 people live in the Belk, is Beltaloda. Belt obviously comes from, you 536 00:34:12,675 --> 00:34:16,275 know, the asteroid belt English. Loda, I think is a Hindu term. I 537 00:34:16,275 --> 00:34:19,580 think. Don't hate on me in the comments. Don't hate on me in the comments. 538 00:34:19,580 --> 00:34:23,420 But, I know Walla is a is a is a Hindu term. Right? So 539 00:34:23,420 --> 00:34:26,884 they'll they'll, you know, when they talk to people who live in the Earth or 540 00:34:26,884 --> 00:34:30,344 Mars, they refer to them as well wallahs, gravity well 541 00:34:30,644 --> 00:34:34,484 wallahs. Right? Like so it's like, and I only know wallah because 542 00:34:34,484 --> 00:34:38,280 of dish wallahs, and Wired Magazine did a whole story about dish wallows in 543 00:34:38,280 --> 00:34:42,060 the nineties. Anyway, but I mean, I think, like, you know, I 544 00:34:42,975 --> 00:34:46,735 I suppose that approach could work for something like a creole. Right? Like, we have 545 00:34:46,735 --> 00:34:50,415 multiple languages kinda mixed together. Or is that not really a 546 00:34:50,415 --> 00:34:51,715 massive business case? 547 00:34:54,290 --> 00:34:57,890 It's Creole is really complicated. It's a language. It's like real real a 548 00:34:57,890 --> 00:35:01,505 real language, and it's complicated. This the the more 549 00:35:01,505 --> 00:35:05,185 delicate cases of that, what we call in research, code switching when 550 00:35:05,185 --> 00:35:08,920 I'm Right. When I speak Hebrew, for example, I don't have a 551 00:35:08,920 --> 00:35:12,520 word for the, you know, the Internet router. So I say the router in 552 00:35:12,520 --> 00:35:16,065 in English. Or I said email or I will say 553 00:35:17,425 --> 00:35:21,160 I don't know. There are so many words in English that are used especially 554 00:35:21,160 --> 00:35:24,920 in technology that you use worldwide in other languages, and this 555 00:35:24,920 --> 00:35:28,700 is code switching. There is another case. I think Andy pointed it 556 00:35:28,840 --> 00:35:31,474 out that sometimes when you are stressed 557 00:35:32,255 --> 00:35:36,015 or let's say your l 1 is Spanish, but l 2 is American 558 00:35:36,015 --> 00:35:39,760 English or you're bilingual. And sometimes when you are 559 00:35:39,760 --> 00:35:43,360 stressed, you you just switch the the 1 560 00:35:43,360 --> 00:35:46,895 word and it this is amazing phenomena. This is a research with Tamar Golang 561 00:35:47,855 --> 00:35:51,475 from, University of San Diego and Matt Goldrick from Northwestern 562 00:35:51,535 --> 00:35:55,295 University. And I provide, again, a mechanism to detect 563 00:35:55,295 --> 00:35:58,960 that and to make research of that. And the the key question is, 564 00:35:58,960 --> 00:36:01,760 like, why do you do that? Why do and when do you do that? Is 565 00:36:01,760 --> 00:36:05,355 it stress? What what what is the what is the state of 566 00:36:05,575 --> 00:36:09,195 describing those? Are you gonna describe it in the American 567 00:36:09,255 --> 00:36:13,019 way, the Spanish word, or is it gonna be vice 568 00:36:13,019 --> 00:36:15,119 versa? And this is really interesting. 569 00:36:18,539 --> 00:36:22,285 It's not my field of research. I just know how to detect them 570 00:36:22,285 --> 00:36:26,045 and, and Interesting. To detect them really well, 571 00:36:26,045 --> 00:36:29,710 but I don't know why it happens and what is the mechanism 572 00:36:29,770 --> 00:36:33,070 behind that. I could definitely see, 573 00:36:35,130 --> 00:36:38,704 the opportunity with starting with being 574 00:36:38,704 --> 00:36:42,385 able to detect, you know, these I 575 00:36:42,385 --> 00:36:46,170 don't I don't know the right word for them. I'll I'll call them modes. You 576 00:36:46,170 --> 00:36:49,770 know, a mode of speech where someone is mixing 2 577 00:36:49,770 --> 00:36:52,990 languages. And I'm sure those vary. 578 00:36:53,455 --> 00:36:57,135 So Like when I go Jersey on you. Right? That's we we 579 00:36:57,135 --> 00:37:00,815 can't we can't say any more about that, Frank. We're trying to keep our 580 00:37:00,815 --> 00:37:03,760 clean rating. But yes. Exactly. But, 581 00:37:05,180 --> 00:37:07,580 that's sorry. Inside, Joe. But the, 582 00:37:08,940 --> 00:37:12,240 but, yeah, I could see modes of speaking where someone who is 583 00:37:12,875 --> 00:37:16,415 more familiar with English as a second language. 584 00:37:16,875 --> 00:37:20,075 And and they've still you know, of course, they know their native language. They'll always 585 00:37:20,075 --> 00:37:23,890 know that. But as they I don't I don't wanna use the wrong word 586 00:37:23,890 --> 00:37:27,650 here, but I'm thinking experience is probably the best word is they get more 587 00:37:27,650 --> 00:37:31,109 experience, gain more experience with their second language. 588 00:37:31,484 --> 00:37:34,704 They may switch words less or switch languages 589 00:37:35,005 --> 00:37:38,790 less. And detecting that, I think, is the 590 00:37:38,870 --> 00:37:42,550 is key. I understand now more about what what you're doing, what 591 00:37:42,550 --> 00:37:46,230 you're accomplishing. And that that's the 592 00:37:46,230 --> 00:37:49,515 very first step to then being able to produce speech 593 00:37:50,214 --> 00:37:53,974 in those different modes. And that would be a 594 00:37:53,974 --> 00:37:57,570 fascinating, you know, a fascinating accomplishment. 595 00:37:58,110 --> 00:38:01,650 If you do, the more we can have. Machines 596 00:38:01,790 --> 00:38:05,545 speak to us in the language that we're most familiar with, that, 597 00:38:05,545 --> 00:38:09,005 of course, you know, is is almost there now, mostly 598 00:38:09,625 --> 00:38:13,329 there right now, but have it be able to to speak to us in these 599 00:38:13,329 --> 00:38:17,089 different modes where we where the machine switches where it's 600 00:38:17,089 --> 00:38:20,665 back to our first language, you know, based 601 00:38:20,665 --> 00:38:24,025 on some algorithmic calculation. That sounds 602 00:38:24,025 --> 00:38:27,840 fascinating. Yeah. It is. 603 00:38:27,840 --> 00:38:31,280 I'm not sure we are there yet. It's we have a long way to go 604 00:38:31,280 --> 00:38:34,875 there. But, Sure. Yeah. Makes 605 00:38:34,875 --> 00:38:38,395 sense. Fascinating. Well, this is how it starts, though. Right? 606 00:38:41,309 --> 00:38:45,150 This is fascinating. This is, yeah, this is, 607 00:38:45,390 --> 00:38:48,990 somehow there is an elephant in the room. There we may have to say 608 00:38:48,990 --> 00:38:52,665 something about AI and their regulation and what happens now. 609 00:38:53,125 --> 00:38:56,565 And, if I may, I would like to say something about this because I have 610 00:38:56,565 --> 00:38:59,385 a deep totally different point of view about that. 611 00:39:01,430 --> 00:39:05,130 Please. So everybody is speaking about 612 00:39:05,750 --> 00:39:09,035 regulation and it might be a catastrophic situation 613 00:39:10,215 --> 00:39:13,595 if those, machine are connected 614 00:39:13,655 --> 00:39:17,170 together and they start to train themselves. They try to 615 00:39:17,170 --> 00:39:20,150 build a meta architecture and try to train themselves, 616 00:39:21,090 --> 00:39:24,895 and then they come up with something which is better than human. Some some people 617 00:39:24,895 --> 00:39:28,655 call it the singularity point. So this is frightening. They're smarter 618 00:39:28,655 --> 00:39:32,490 than us. Maybe they they're gonna kill us all. And 619 00:39:33,349 --> 00:39:36,950 people say now people speak about regulation now, and there are 620 00:39:36,950 --> 00:39:40,569 several institutes in Europa, in Europe and in, the US 621 00:39:40,734 --> 00:39:44,575 trying to tackle that. And that 622 00:39:44,575 --> 00:39:48,195 is amazing. That is really important, but I think we missed something here. 623 00:39:49,110 --> 00:39:52,890 And I'll tell you why. So the so there is a book. It's here. 624 00:39:53,030 --> 00:39:56,845 You know, Isaac Asimov, I, Robot. You probably 625 00:39:56,845 --> 00:40:00,365 know that. So he, like, the first page of this book is like the 3 626 00:40:00,365 --> 00:40:04,125 laws of robotic. A robot may not in in injury a 627 00:40:04,125 --> 00:40:07,710 human being or through an interaction, allow human being to come to harm. 628 00:40:08,890 --> 00:40:12,650 A robot must obey others and so on. So we have let's say 629 00:40:12,650 --> 00:40:16,135 we have the regulation. AI cannot hurt humans. Okay? 630 00:40:16,434 --> 00:40:20,194 But that doesn't enough. It's not good enough because if the AI is smart 631 00:40:20,194 --> 00:40:23,790 enough, it will not do the I mean, it will 632 00:40:23,790 --> 00:40:27,010 show us humans that it really obey the law 633 00:40:27,630 --> 00:40:31,365 the laws, but it wouldn't. And this is frightening. 634 00:40:31,425 --> 00:40:35,205 And here I suggest to look a little bit about the human morality 635 00:40:35,665 --> 00:40:39,505 and what why human are have do they have laws? So we need to 636 00:40:39,505 --> 00:40:43,250 think about, if I may, think about the 637 00:40:43,250 --> 00:40:47,089 human psychology. In human psychology, we have a mechanism to obey law. 638 00:40:47,089 --> 00:40:50,455 It's called the superego. It was embedded or defined by 639 00:40:50,455 --> 00:40:54,235 Freud. So we have a mechanism that if we 640 00:40:55,255 --> 00:40:58,970 if we doesn't we if we don't obey a law, we feel either 641 00:40:58,970 --> 00:41:02,589 guilt or fear. And this mechanism was evolutionary. 642 00:41:02,970 --> 00:41:06,430 So do we have a group of monkey? They obey 643 00:41:07,135 --> 00:41:10,895 the the alpha monkey because they're frightened from him. They have some kind of 644 00:41:10,895 --> 00:41:14,690 primitive superego. We obey the law because either we fight them from the 645 00:41:15,010 --> 00:41:18,450 police or either we feel the guilt, we 646 00:41:18,450 --> 00:41:20,310 we it's like the 647 00:41:23,125 --> 00:41:26,184 those experiments that show that, there is, somebody, 648 00:41:26,805 --> 00:41:30,085 left something on the table, and we don't take it because we feel guilt or 649 00:41:30,085 --> 00:41:33,869 we feel something. So this is this mechanism, what 650 00:41:33,869 --> 00:41:37,410 I claim, should be transferred to the 651 00:41:37,549 --> 00:41:41,085 AI machine. This should be the regulation. So what is it superego? Superego 652 00:41:41,145 --> 00:41:44,845 is a infrastructure for to be moral, 653 00:41:45,385 --> 00:41:48,850 and we need a digital version for that for the this is the regulation we 654 00:41:48,850 --> 00:41:52,610 need. We need the infrastructure to be moral in machine. And what it what 655 00:41:52,610 --> 00:41:56,405 does it mean? So superego means that it's a little bit like 656 00:41:56,405 --> 00:42:00,244 self harm, if I may. It's like we feel guilt. We feel something bad if 657 00:42:00,244 --> 00:42:03,464 we do something not okay, if you're not obey the law. 658 00:42:04,120 --> 00:42:07,880 So it's like a self destruction for AI machine. So AI machine, 659 00:42:07,880 --> 00:42:11,640 if it doesn't obey the law, should feel something. It 660 00:42:11,640 --> 00:42:15,205 cannot feel so. Right. It will distract itself. So this is my 661 00:42:15,205 --> 00:42:18,985 claim. This is a book I'm writing, and this is something very fun fundamental. 662 00:42:19,045 --> 00:42:22,880 We we all speak about this regulation, but I think it 663 00:42:22,880 --> 00:42:26,320 it doesn't help just to to do standard 664 00:42:26,320 --> 00:42:30,160 regulation. And if you if I may say another thing, the last thing is that 665 00:42:30,160 --> 00:42:34,005 if you read the I, Robert, carefully, so 666 00:42:34,005 --> 00:42:37,605 he speak there are several short stories there, and he speak about robots that 667 00:42:37,605 --> 00:42:41,250 obey the law. And if you look carefully about those robots that 668 00:42:41,250 --> 00:42:45,010 obey the law, the those robots have super all 669 00:42:45,010 --> 00:42:48,310 all of them have have super ego. They feel guilt. 670 00:42:48,915 --> 00:42:52,295 The the first story is about a robot that play with a girl, 671 00:42:52,675 --> 00:42:56,455 and he feel guilt about winning all the time. So he let her win. 672 00:42:56,860 --> 00:43:00,080 So he feels guilt. It means that it has superhego. 673 00:43:00,620 --> 00:43:04,375 And then he feels frightened from the mother of the girl. And it's 674 00:43:04,375 --> 00:43:08,055 really amazing. So I think, so 675 00:43:08,055 --> 00:43:11,515 this book I'm trying to describe the psychological concept of superego 676 00:43:11,970 --> 00:43:14,849 and then describe why it need to be more and how we can, 677 00:43:16,049 --> 00:43:19,269 find a way to put it in regulation, like the the infrastructure 678 00:43:19,490 --> 00:43:21,589 itself and not just lows. 679 00:43:23,075 --> 00:43:26,055 That is a very interesting problem you're trying to solve. 680 00:43:27,715 --> 00:43:31,349 Very important problem at that. Agreed. And 681 00:43:31,349 --> 00:43:35,109 culturally, we speak, in the US, we have a saying that you 682 00:43:35,109 --> 00:43:38,170 cannot legislate morality, which 683 00:43:38,674 --> 00:43:42,355 legislate, regulate would be, you know, 684 00:43:42,355 --> 00:43:46,194 synonyms. Exactly. Right? So Right. Right. And and legal code 685 00:43:46,194 --> 00:43:49,750 is code. I I 686 00:43:49,750 --> 00:43:53,190 definitely get what you're what you're saying. And I think it's super 687 00:43:53,190 --> 00:43:56,970 important. You mentioned you were writing a book about this. Now 688 00:43:57,350 --> 00:44:00,635 now now you have to tell me more because I wanna read this book. 689 00:44:00,935 --> 00:44:04,775 Same. I'm in the process of looking 690 00:44:04,775 --> 00:44:08,610 for an agent and it's, it's complicated. It's supposed 691 00:44:08,610 --> 00:44:12,450 to be a popular book trying to explain the psychology of fraud. 692 00:44:12,450 --> 00:44:15,350 What is, superego, ego, and the id, 693 00:44:16,234 --> 00:44:20,075 and then describe what is the pathology? So we all have a pathology. So 694 00:44:20,075 --> 00:44:22,734 you have the pathology of, it's called, 695 00:44:29,630 --> 00:44:33,175 the, personalities criminal personality disorder. This 696 00:44:33,175 --> 00:44:37,015 person will not have a super ego, ego ego. It's like Richard the 697 00:44:37,015 --> 00:44:40,820 third from Shakespeare. He didn't have superego. He killed 698 00:44:40,820 --> 00:44:44,500 his family and didn't feel guilt. So this wouldn't what's 699 00:44:44,500 --> 00:44:48,260 going to happen with the with the with those machine. And then I 700 00:44:48,260 --> 00:44:51,285 give some literature examples of, 701 00:44:51,925 --> 00:44:55,365 what is a superego like from the, criminal and 702 00:44:55,365 --> 00:44:59,100 punishment that that the guy killed the the 703 00:44:59,100 --> 00:45:02,160 old lady, but he didn't he nobody, 704 00:45:02,780 --> 00:45:06,595 caught him killing the lady. He murdered her. Nobody caught him, but he 705 00:45:06,595 --> 00:45:10,275 still feel guilt. So he has a very, big 706 00:45:10,275 --> 00:45:13,954 superego. And then we describe I describe, what happened in 707 00:45:13,954 --> 00:45:17,540 other moral theories of human being, all of them connected to the 708 00:45:17,540 --> 00:45:21,140 superego. And then I tried to describe a little bit how machine 709 00:45:21,140 --> 00:45:24,925 learning is trained. Again, solving an optimization problem. And then I try 710 00:45:24,925 --> 00:45:28,765 to describe how can we do superego with, how can we have 711 00:45:28,765 --> 00:45:32,549 a digital superego if we can? No. 712 00:45:32,549 --> 00:45:36,250 It's like you're giving it a conscience of of sorts. Exactly. 713 00:45:36,790 --> 00:45:40,455 Yeah. And I I just wanted to, to add, we 714 00:45:40,455 --> 00:45:44,055 may be able to help you. Maybe not find an 715 00:45:44,055 --> 00:45:47,415 agent, but find a publisher. Both Frank and I are 716 00:45:47,415 --> 00:45:50,980 published. And we, you know, we know Andy has a lot of 717 00:45:51,300 --> 00:45:54,980 Andy's got a lot of connections in the publishing. Well That would be 718 00:45:54,980 --> 00:45:58,755 great. I am I am not, I just wrote a lot of books 719 00:45:58,755 --> 00:46:02,595 for different, publishing houses, and I know some people that if 720 00:46:02,595 --> 00:46:05,955 they can't help you directly, they can probably point you to someone who 721 00:46:05,955 --> 00:46:09,790 can. And, again, I am wholly motivated by wanting to 722 00:46:09,790 --> 00:46:13,150 read this book. Same. Like, I think it's important 723 00:46:13,150 --> 00:46:16,875 because I live in the Washington DC area. Right? 724 00:46:16,935 --> 00:46:20,695 So so, like, there's a lot of people there who they're policy 725 00:46:20,695 --> 00:46:24,075 makers. Right? Like, and they just assume 726 00:46:24,490 --> 00:46:27,369 and I think a lot of humans fall for this. Right? You you see this 727 00:46:27,369 --> 00:46:31,150 when the European Union passed their AI regulation act. 728 00:46:31,210 --> 00:46:33,869 They assume that regulation's gonna solve all their problems. 729 00:46:34,955 --> 00:46:38,795 And I think regulations prove that 1 of the fundamental forces 730 00:46:38,795 --> 00:46:41,695 in the universe is is unintended consequences. 731 00:46:42,580 --> 00:46:46,420 And, you know, when you regulate something, you don't end 732 00:46:46,420 --> 00:46:50,175 the problem. You change the way people will route around it. Right? Like, 733 00:46:50,395 --> 00:46:53,915 and I think a good example of this in AI is the movie Megan, which 734 00:46:53,915 --> 00:46:56,315 I don't know if you've seen, or m threagan. I'm not sure how to pronounce 735 00:46:56,315 --> 00:46:59,730 it, where I think she was about to torture 736 00:47:00,210 --> 00:47:04,050 she was I don't wanna give the plot away, but the the robot 737 00:47:04,050 --> 00:47:07,715 child, Chucky, kinda goes evil, Like, this is the 738 00:47:07,715 --> 00:47:11,555 basic kind of plot line, and the the the person who created her 739 00:47:11,555 --> 00:47:14,275 was like, you can't kill me because it's against your programming. He goes, oh, I 740 00:47:14,275 --> 00:47:16,829 said nothing about killing you. I was gonna put you in a coma, and you'll 741 00:47:16,829 --> 00:47:20,589 live, you know, however many years. Like, it was just like I mean, 742 00:47:20,589 --> 00:47:23,630 that's a great example of, like, she you know, don't kill. Right? Seems like a 743 00:47:23,630 --> 00:47:27,365 pretty reasonable instruction to give a robot, particularly a child's toy. 744 00:47:28,545 --> 00:47:32,145 They'll kill anyone. But, you know, she was realized, like, well, kill 745 00:47:32,145 --> 00:47:35,550 equals death. So if I don't kill you, if I just hospitalize you or 746 00:47:35,550 --> 00:47:38,990 incapacitate you, that doesn't conflict with rule number 1. 747 00:47:38,990 --> 00:47:42,805 Right? Which I think is no. Obviously, as, you 748 00:47:42,805 --> 00:47:46,165 know, humans, we're like, well, it's not really the spirit of the 749 00:47:46,165 --> 00:47:49,705 law, or the rule. But clearly, 750 00:47:50,289 --> 00:47:53,970 the robot or the AI in this case, kind of figured it 751 00:47:53,970 --> 00:47:57,809 out. Like, I don't know. I think you're right. Like and any regulations like that 752 00:47:57,809 --> 00:48:01,545 too. Right? How many loopholes do people discover, whether it's 753 00:48:01,545 --> 00:48:05,224 tax laws or, you know, this. It's like, well, technically, it's 754 00:48:05,224 --> 00:48:08,680 legal. Is it actually, you know, 755 00:48:09,300 --> 00:48:13,140 what the law intended? No. Like, it's Yeah. You need a you need 756 00:48:13,140 --> 00:48:16,040 almost an something like a Nuance engine, 757 00:48:16,955 --> 00:48:19,695 you'll see to Yeah. To get the the 758 00:48:20,635 --> 00:48:24,015 what the machine to interpret 759 00:48:24,450 --> 00:48:27,829 to the laws. And that's I've read Asimov as well, 760 00:48:28,289 --> 00:48:31,955 big fan. And that's what happens down stream of 761 00:48:31,955 --> 00:48:35,635 the 3 laws as they begin to fail as because the 762 00:48:35,635 --> 00:48:39,475 robots are doing exactly what they're programmed to 763 00:48:39,475 --> 00:48:43,070 do. And they're not they're they're 764 00:48:43,070 --> 00:48:46,690 finding ways that in our opinion, human opinion, 765 00:48:46,830 --> 00:48:50,625 circumvents the 3 laws, but really doesn't 766 00:48:50,625 --> 00:48:54,385 break the robot's programming. And it's all about, you know, 767 00:48:54,385 --> 00:48:58,065 how do you define harm? Like, Frank's example is a great, you know, 768 00:48:58,065 --> 00:49:01,710 great example of that. So, yeah, 769 00:49:01,710 --> 00:49:05,470 fascinating stuff. Yeah. We gotta Awesome stuff. We gotta help you write this 770 00:49:05,470 --> 00:49:09,135 book. I wanna read this book. Yeah. I want to raise 771 00:49:09,135 --> 00:49:12,975 another point, but the opposite point that you raised. Like, what happened with 772 00:49:12,975 --> 00:49:16,435 the autonomous car, for example, or people say, 773 00:49:18,000 --> 00:49:21,599 let's let's let's focus on autonomous cars. So so there will be 774 00:49:21,599 --> 00:49:24,820 autonomous car. Who is in charge of a of a car accident? 775 00:49:25,635 --> 00:49:29,395 Accidentally, somebody was killed. You are the 776 00:49:29,395 --> 00:49:33,155 owner you. Somebody is the owner of the car. He sits 777 00:49:33,155 --> 00:49:36,680 there. He bought the car, but the car killed 778 00:49:36,680 --> 00:49:40,280 somebody. So 779 00:49:40,280 --> 00:49:43,720 who who this is an open problem. This is, again, 780 00:49:43,720 --> 00:49:47,465 moral problem. So what I suggest here is 781 00:49:47,465 --> 00:49:51,305 maybe it will take time, 782 00:49:51,305 --> 00:49:54,760 I guess. Maybe the the car, if we can be the 783 00:49:54,760 --> 00:49:58,599 superego and mechanism for morality, you know, the just 784 00:49:58,599 --> 00:50:02,244 the infrastructure for morality can take the 785 00:50:02,244 --> 00:50:05,845 morality of the human. And if somehow he 786 00:50:05,845 --> 00:50:09,570 inherit the the the driver morality, you 787 00:50:09,570 --> 00:50:13,330 can blame the driver. I'll give you another example, which will be much 788 00:50:13,330 --> 00:50:17,085 more maybe concrete. So we say now that there will be change GPT for 789 00:50:17,085 --> 00:50:20,545 every person, for every laptop and iPhone and whatever. 790 00:50:21,005 --> 00:50:24,225 You will have your own GPT with your own life follows 791 00:50:24,845 --> 00:50:28,330 your own history. And the discussion with this GPT will be, And the 792 00:50:28,330 --> 00:50:32,090 discussion with this, GPT will be very personalized and 793 00:50:32,090 --> 00:50:35,815 very helpful. What happened in that case? So in that 794 00:50:35,815 --> 00:50:39,495 case, if this, GPT 795 00:50:39,495 --> 00:50:43,260 will take your responsibilities and morality, somehow we 796 00:50:43,260 --> 00:50:47,099 can copy your morality and be part of it. So if you're moral, it 797 00:50:47,099 --> 00:50:50,495 will be moral. If you're not, you're not, but this is 798 00:50:50,495 --> 00:50:54,335 your responsibility as a human. And I think this 799 00:50:54,335 --> 00:50:57,855 is the way to to go with that. We need just the infrastructure and not 800 00:50:57,855 --> 00:51:01,560 the the law. Anybody can define the low, and anybody 801 00:51:01,560 --> 00:51:05,320 can break the low. We just need the infrastructure to know that 802 00:51:06,355 --> 00:51:09,974 at least the machine to know that it break the broke the low. 803 00:51:11,795 --> 00:51:13,964 And and this is really important. I I think 804 00:51:16,420 --> 00:51:20,020 Oh, I totally agree. Totally agree. Well, we're 805 00:51:20,260 --> 00:51:23,620 gosh. We're coming up on time, Frank. Yeah. This was 806 00:51:23,620 --> 00:51:27,155 awesome. So we'll just any 807 00:51:27,155 --> 00:51:30,915 book recommendations? Obviously, I, Robot, I think, would be good reading 808 00:51:30,915 --> 00:51:34,410 in this space. You also mentioned Shakespeare too, 809 00:51:34,410 --> 00:51:38,036 Richard the 3rd. So Eddie, you can book 810 00:51:38,036 --> 00:51:41,244 which I'm which I'm reading now, which is the band, 811 00:51:41,885 --> 00:51:45,405 Vernon Stuputeux. It's, it's 812 00:51:45,405 --> 00:51:49,030 amazing. It's amazing. It's 3 books, and it's actually 813 00:51:49,030 --> 00:51:52,710 discussed whatever which is not AI. Anything which cannot be solved with 814 00:51:52,710 --> 00:51:56,410 AI. It's speak about a a person who has a vinyl shop, 815 00:51:57,145 --> 00:52:00,744 shop to sell vinyl and then CD runs, and now we cannot sell 816 00:52:00,744 --> 00:52:04,285 anything. So this shop is is closed, and then he 817 00:52:04,940 --> 00:52:08,539 he he try to somehow manage, but he get up at the street. He's, like, 818 00:52:08,539 --> 00:52:12,154 homeless, and he meets many people. And the way like, 819 00:52:12,154 --> 00:52:15,674 every chapter is a different, person or 820 00:52:15,674 --> 00:52:19,275 or a group of pair of people, and it's really 821 00:52:19,275 --> 00:52:22,890 fascinating. It's all those things that you cannot solve with AI. It's all 822 00:52:22,890 --> 00:52:26,270 the human interaction, the very, very basic human interaction. Amazing. 823 00:52:26,730 --> 00:52:30,525 It won the Booker Prize in the, 2018. 824 00:52:32,265 --> 00:52:35,625 Nice. Where can folks find out more about 825 00:52:35,625 --> 00:52:38,260 you? So I have a website 826 00:52:39,359 --> 00:52:43,200 under Joseph Keshet, and, and they 827 00:52:43,200 --> 00:52:46,435 can find me there. Excellent. 828 00:52:47,295 --> 00:52:50,735 Any parting thoughts, Andy? No. Just great great 829 00:52:50,735 --> 00:52:54,560 interview. I appreciate that. 1, I would ask if you repeat the name of 830 00:52:54,560 --> 00:52:57,780 the book you just mentioned about the the different stories. 831 00:52:58,400 --> 00:53:01,915 What's the name of that book? It's not it's a it's a single 832 00:53:01,915 --> 00:53:05,375 story. It's called the the pants, 833 00:53:06,715 --> 00:53:10,529 for non subtext. It's from French. Oh, okay. 834 00:53:11,170 --> 00:53:15,010 Amazing. Amazing. Amazing. Awesome. Excellent. That's it. That's 835 00:53:15,010 --> 00:53:18,825 it for me. But that's great talk. Thank you. Excellent talk. Thank you. 836 00:53:18,825 --> 00:53:22,665 And we'll let Bailey finish the show. Well, folks, that brings us to the end 837 00:53:22,665 --> 00:53:26,080 of another enlightening episode of data driven. We've 838 00:53:26,080 --> 00:53:29,440 navigated the fascinating intricacies of automatic speech 839 00:53:29,440 --> 00:53:33,285 recognition, explored the moral quandaries of AI, and 840 00:53:33,285 --> 00:53:37,125 pondered the future of technology with none other than 1 of the best minds 841 00:53:37,125 --> 00:53:40,805 in the field, doctor Yossi Keshet. Remember, if you 842 00:53:40,805 --> 00:53:44,490 enjoyed today's conversation, don't forget to subscribe to data 843 00:53:44,490 --> 00:53:48,190 driven media TV for exclusive video content. 844 00:53:48,730 --> 00:53:52,435 You can also grab some fantastic merch like the my data is the 845 00:53:52,435 --> 00:53:56,115 new oil t shirt Andy's sporting today. And while Frank is 846 00:53:56,115 --> 00:53:59,950 basking in the Appalachian sunshine, you can bet we're already cooking up the 847 00:53:59,950 --> 00:54:03,730 next episode to keep your data driven minds engaged and entertained. 848 00:54:04,467 --> 00:54:08,067 Until next time, stay curious, stay informed, and 849 00:54:08,067 --> 00:54:10,247 always keep questioning. Cheerio.