Fuld transskription af 09 Is Artificial Intelligence dead again? (aka "just buy a license")

machine 46 00:01:43,119 --> 00:01:46,159 learning or ai whatever may be the 47 00:01:44,880 --> 00:01:48,560 difference maybe we were talking about 48 00:01:46,159 --> 00:01:50,640 it later for how long now because i know 49 00:01:48,560 --> 00:01:51,439 that after you graduated you already 50 00:01:50,640 --> 00:01:53,360 like 51 00:01:51,439 --> 00:01:55,520 worked and did stuff and with machine 52 00:01:53,360 --> 00:01:58,560 learning this was like 20 years before 53 00:01:55,520 --> 00:02:01,200 anybody else was talking about it right 54 00:01:58,560 --> 00:02:03,200 yeah i mean i mean ai is basically 55 00:02:01,200 --> 00:02:05,840 something that was invented in the 60s 56 00:02:03,200 --> 00:02:08,239 it's it's quite old um at least the the 57 00:02:05,840 --> 00:02:12,000 ideas all the mathematical 58 00:02:08,239 --> 00:02:12,879 basics of of ai i studied like 20 years 59 00:02:12,000 --> 00:02:14,640 ago 60 00:02:12,879 --> 00:02:16,640 in this field 61 00:02:14,640 --> 00:02:19,360 it was 62 00:02:16,640 --> 00:02:21,120 basically my motivation came when i saw 63 00:02:19,360 --> 00:02:24,480 the movie in the matrix 64 00:02:21,120 --> 00:02:27,200 this one was in 99 very very long ago a 65 00:02:24,480 --> 00:02:29,680 very long time ago and what i really 66 00:02:27,200 --> 00:02:32,319 loved about this concept was um that 67 00:02:29,680 --> 00:02:35,280 this movie showed things like swarm 68 00:02:32,319 --> 00:02:37,840 intelligence of these uh these squiddy 69 00:02:35,280 --> 00:02:39,920 things that were flying around we have 70 00:02:37,840 --> 00:02:43,200 this cognitive intelligence this 71 00:02:39,920 --> 00:02:46,239 mastermind but also this depiction of 72 00:02:43,200 --> 00:02:48,400 the agents which is like the agent-based 73 00:02:46,239 --> 00:02:50,400 artificial intelligence and this was the 74 00:02:48,400 --> 00:02:52,239 field where i started studying computer 75 00:02:50,400 --> 00:02:55,040 science in the lab for 76 00:02:52,239 --> 00:02:57,840 distributed artificial intelligence or 77 00:02:55,040 --> 00:03:00,080 agent-based systems and this one was 78 00:02:57,840 --> 00:03:01,519 really really interesting and we did not 79 00:03:00,080 --> 00:03:02,560 really 80 00:03:01,519 --> 00:03:05,920 dive 81 00:03:02,560 --> 00:03:08,480 into the topic of machine learning 82 00:03:05,920 --> 00:03:09,840 necessarily this was only one of the of 83 00:03:08,480 --> 00:03:11,040 the aspects of 84 00:03:09,840 --> 00:03:12,879 ai 85 00:03:11,040 --> 00:03:15,440 and and i really really loved this and 86 00:03:12,879 --> 00:03:17,760 was always like a very promising field 87 00:03:15,440 --> 00:03:19,680 but back then of course not 88 00:03:17,760 --> 00:03:22,239 not the standard technology and only 89 00:03:19,680 --> 00:03:24,480 with the rise of deep learning 90 00:03:22,239 --> 00:03:27,920 it became so powerful that 91 00:03:24,480 --> 00:03:31,440 we today we can't

see any system anymore 92 00:03:27,920 --> 00:03:34,319 where it's not one component based on ai 93 00:03:31,440 --> 00:03:36,400 and so i think um yeah people stop 94 00:03:34,319 --> 00:03:38,640 talking about this because it's simply a 95 00:03:36,400 --> 00:03:40,239 standard right now 96 00:03:38,640 --> 00:03:42,400 i would say deep learning wasn't it 97 00:03:40,239 --> 00:03:43,840 already like 40 years old i mean all 98 00:03:42,400 --> 00:03:45,840 this deep learning stuff it's not new 99 00:03:43,840 --> 00:03:48,319 it's just like like scaling and 100 00:03:45,840 --> 00:03:49,920 computational power is now good enough 101 00:03:48,319 --> 00:03:52,720 that you actually can 102 00:03:49,920 --> 00:03:54,560 use this in in production right 103 00:03:52,720 --> 00:03:57,519 no i think deep learning is actually 104 00:03:54,560 --> 00:03:58,879 quite new let me quickly google that 105 00:03:57,519 --> 00:04:04,080 i think 106 00:03:58,879 --> 00:04:05,360 deep learning uh is something that was 107 00:04:04,080 --> 00:04:08,799 let me check 108 00:04:05,360 --> 00:04:12,840 we can we can be so spontaneous 109 00:04:08,799 --> 00:04:16,560 deep learning is only a couple of years 110 00:04:12,840 --> 00:04:20,400 old okay so there were actually papers 111 00:04:16,560 --> 00:04:22,880 about deep learning obviously uh in 1971 112 00:04:20,400 --> 00:04:25,040 okay so the concept is not really that 113 00:04:22,880 --> 00:04:26,960 new 114 00:04:25,040 --> 00:04:29,520 that's what i meant actually that that's 115 00:04:26,960 --> 00:04:31,759 a concept and all the science behind it 116 00:04:29,520 --> 00:04:33,680 it's like really really old already it's 117 00:04:31,759 --> 00:04:35,440 just that it was not feasible with the 118 00:04:33,680 --> 00:04:38,639 computational power you had back in 119 00:04:35,440 --> 00:04:41,520 those days so it's again the same with 120 00:04:38,639 --> 00:04:43,440 uh all these new language models we have 121 00:04:41,520 --> 00:04:46,000 right now you know 122 00:04:43,440 --> 00:04:49,120 where they improve year by year by 123 00:04:46,000 --> 00:04:52,080 factors because the network gets more 124 00:04:49,120 --> 00:04:54,400 complex that just scales with pure pure 125 00:04:52,080 --> 00:04:55,440 mass actually and that's makes them more 126 00:04:54,400 --> 00:04:58,320 natural 127 00:04:55,440 --> 00:05:00,880 and i think this is also thanks to our 128 00:04:58,320 --> 00:05:02,479 graphic cards being able to do a massive 129 00:05:00,880 --> 00:05:05,280 parallel computation 130 00:05:02,479 --> 00:05:07,919 and i think quantum computing would be 131 00:05:05,280 --> 00:05:09,600 the next driver for machine learning but 132 00:05:07,919 --> 00:05:11,360 i i think we should talk about quantum 133 00:05:09,600 --> 00:05:12,639 computing you know another episode 134 00:05:11,360 --> 00:05:15,039 actually 135 00:05:12,639 --> 00:05:17,039 yeah that's true and while i'm skipping 136 00:05:15,039 --> 00:05:19,120 right now the wikipedia article for deep 137 00:05:17,039 --> 00:05:22,400 learning i think uh the 138 00:05:19,120 --> 00:05:25,759 the event that i meant was in 2012 139 00:05:22,400 --> 00:05:28,240 um where in a hackathon people were for 140 00:05:25,759 --> 00:05:30,800 for the first time um applying something 141 00:05:28,240 --> 00:05:32,560 like a deep learning algorithm um to do 142 00:05:30,800 --> 00:05:34,880 some kind of classification and for for 143 00:05:32,560 --> 00:05:37,520 the first time it was actually working 144 00:05:34,880 --> 00:05:39,440 um better than everything else and i 145 00:05:37,520 --> 00:05:41,280 think this was like the beginning where 146 00:05:39,440 --> 00:05:43,039 deep learning became the new standard 147 00:05:41,280 --> 00:05:45,600 for all kinds of 148 00:05:43,039 --> 00:05:48,160 machine learning algorithms i knew it a 149 00:05:45,600 --> 00:05:49,840 hackathon was a breakthrough moment it's 150 00:05:48,160 --> 00:05:51,759 like with everything like which is 151 00:05:49,840 --> 00:05:54,240 innovation right now it mostly comes 152 00:05:51,759 --> 00:05:56,319 from a hackathon oh where does netflix 153 00:05:54,240 --> 00:05:58,560 or dropbox get its innovation from yeah 154 00:05:56,319 --> 00:05:59,759 from hackathon to hack weeks 155 00:05:58,560 --> 00:06:00,960 right 156 00:05:59,759 --> 00:06:02,319 right 157 00:06:00,960 --> 00:06:04,639 yeah but 158 00:06:02,319 --> 00:06:08,080 i would still claim that there was this 159 00:06:04,639 --> 00:06:09,919 hype phase like which was uh i don't 160 00:06:08,080 --> 00:06:11,600 know 161 00:06:09,919 --> 00:06:13,280 up until like 162 00:06:11,600 --> 00:06:15,360 last year maybe because this year 163 00:06:13,280 --> 00:06:17,680 everybody is talking about again about 164 00:06:15,360 --> 00:06:19,840 the next things but like three years ago 165 00:06:17,680 --> 00:06:22,000 you wouldn't get a funding if you hadn't 166 00:06:19,840 --> 00:06:23,360 written machine learning okay sorry 167 00:06:22,000 --> 00:06:24,800 you'd write a 168 00:06:23,360 --> 00:06:27,039 artificial intelligence on your 169 00:06:24,800 --> 00:06:28,319 powerpoint slide and if you do it like a 170 00:06:27,039 --> 00:06:30,880 serial engineer then it's machine 171 00:06:28,319 --> 00:06:32,400 learning so if you didn't write ai on 172 00:06:30,880 --> 00:06:34,400 your powerpoint slides you didn't get a 173 00:06:32,400 --> 00:06:36,639 funding because then oh it's scaling 174 00:06:34,400 --> 00:06:39,039 because it's ai so like yeah we want 175 00:06:36,639 --> 00:06:40,000 scaling business models 176 00:06:39,039 --> 00:06:42,000 yeah 177 00:06:40,000 --> 00:06:43,840 i mean um 178 00:06:42,000 --> 00:06:47,759 on one hand you're right 179 00:06:43,840 --> 00:06:50,400 it is still kind of a hype topic when 180 00:06:47,759 --> 00:06:53,039 depending on what you're

promising and 181 00:06:50,400 --> 00:06:55,680 with the in the age of self-driving cars 182 00:06:53,039 --> 00:06:57,919 for example we are expecting a lot of 183 00:06:55,680 --> 00:07:00,240 artificial intelligence and that ai is 184 00:06:57,919 --> 00:07:02,960 basically solving everything and is 185 00:07:00,240 --> 00:07:05,680 doing everything perfectly and this is a 186 00:07:02,960 --> 00:07:07,599 problem of over promising and then under 187 00:07:05,680 --> 00:07:08,880 delivering because we see right now 188 00:07:07,599 --> 00:07:10,800 there are no 189 00:07:08,880 --> 00:07:12,479 i mean there are barely self-driving 190 00:07:10,800 --> 00:07:15,039 cars on the streets of course there are 191 00:07:12,479 --> 00:07:17,120 so many pilot projects and experiments 192 00:07:15,039 --> 00:07:20,160 but we are not really there yet there is 193 00:07:17,120 --> 00:07:23,280 this last percent this last step where 194 00:07:20,160 --> 00:07:26,240 we we can't get it as reliable as 195 00:07:23,280 --> 00:07:28,280 necessary to actually roll it out as a 196 00:07:26,240 --> 00:07:31,280 self-sustaining 197 00:07:28,280 --> 00:07:33,840 self-controlling system we are very good 198 00:07:31,280 --> 00:07:35,520 in building prediction machines and we 199 00:07:33,840 --> 00:07:37,120 are living right now in the in the age 200 00:07:35,520 --> 00:07:38,560 of prediction machines and everything is 201 00:07:37,120 --> 00:07:40,639 built on based on 202 00:07:38,560 --> 00:07:43,199 on top of prediction machines but it's 203 00:07:40,639 --> 00:07:44,960 only that it's only like predictions and 204 00:07:43,199 --> 00:07:47,599 we are not there yet that we actually 205 00:07:44,960 --> 00:07:48,960 can say our systems are 206 00:07:47,599 --> 00:07:51,919 reliable 207 00:07:48,960 --> 00:07:54,879 decision machines making decisions as 208 00:07:51,919 --> 00:07:57,360 good as human beings based on context 209 00:07:54,879 --> 00:07:59,680 based on semantics this one will come 210 00:07:57,360 --> 00:08:02,080 i'm very sure of it maybe like in 10 211 00:07:59,680 --> 00:08:04,879 years or five years from now um that we 212 00:08:02,080 --> 00:08:07,039 can actually reliable reliably build 213 00:08:04,879 --> 00:08:09,840 these decision machines but right now we 214 00:08:07,039 --> 00:08:12,879 are mastering the prediction machines 215 00:08:09,840 --> 00:08:14,879 and so what managers are promising or 216 00:08:12,879 --> 00:08:16,879 politicians are promising and what then 217 00:08:14,879 --> 00:08:19,599 is actually delivered there might still 218 00:08:16,879 --> 00:08:22,080 be a gap and i think that's where these 219 00:08:19,599 --> 00:08:24,720 hypes are dying because people see the 220 00:08:22,080 --> 00:08:26,879 reality is not as far as we think but at 221 00:08:24,720 --> 00:08:31,520 the same time the tools that we are 222 00:08:26,879 --> 00:08:34,240 already using they are sometimes even um 223 00:08:31,520 --> 00:08:36,800 superseding our expectations for this 224 00:08:34,240 --> 00:08:38,320 particular task that they are built on 225 00:08:36,800 --> 00:08:40,240 yeah 226 00:08:38,320 --> 00:08:42,159 it's like you said it will come back 227 00:08:40,240 --> 00:08:44,480 maybe not in ten years maybe in seven 228 00:08:42,159 --> 00:08:46,640 because that's a hype cycle but 229 00:08:44,480 --> 00:08:47,519 we promised or like we were promised 230 00:08:46,640 --> 00:08:48,560 like 231 00:08:47,519 --> 00:08:51,519 fake 232 00:08:48,560 --> 00:08:53,600 voices like deep fake videos and 233 00:08:51,519 --> 00:08:55,760 everyone was afraid oh we are so far 234 00:08:53,600 --> 00:08:57,760 ahead though we will fake like obama's 235 00:08:55,760 --> 00:08:59,920 speeches trump's speeches and so on this 236 00:08:57,760 --> 00:09:01,440 was like again marketing promises right 237 00:08:59,920 --> 00:09:03,200 and that's when the hype died because i 238 00:09:01,440 --> 00:09:05,839 think in the next seven years there will 239 00:09:03,200 --> 00:09:09,120 now be there will no 240 00:09:05,839 --> 00:09:12,160 be good uh sorry there will be none good 241 00:09:09,120 --> 00:09:13,920 deep faked video of any politician so 242 00:09:12,160 --> 00:09:15,360 yes everybody was afraid of it but we 243 00:09:13,920 --> 00:09:17,360 are not that far 244 00:09:15,360 --> 00:09:19,440 already in the game so it's the same as 245 00:09:17,360 --> 00:09:22,320 with like faking voices so i saw some 246 00:09:19,440 --> 00:09:24,959 big uh voice generation projects and it 247 00:09:22,320 --> 00:09:27,839 still is not good enough to replace any 248 00:09:24,959 --> 00:09:29,920 voice actor there's even like a startup 249 00:09:27,839 --> 00:09:31,920 out there i recently saw on linkedin 250 00:09:29,920 --> 00:09:34,080 which is trying to create 251 00:09:31,920 --> 00:09:36,720 like marketing videos where you can 252 00:09:34,080 --> 00:09:38,480 choose between over 20 virtual actors 253 00:09:36,720 --> 00:09:40,240 which read out your stuff you can select 254 00:09:38,480 --> 00:09:41,920 the outfit and so on but it's still it 255 00:09:40,240 --> 00:09:43,920 looks like computer generated graphics 256 00:09:41,920 --> 00:09:45,839 so it's not as good as a real actor and 257 00:09:43,920 --> 00:09:48,160 i think in the next years that will not 258 00:09:45,839 --> 00:09:51,120 be it will not happen and i think that's 259 00:09:48,160 --> 00:09:52,560 that's the issue yeah i think we are 260 00:09:51,120 --> 00:09:54,320 basically 261 00:09:52,560 --> 00:09:57,600 we as human beings are just too 262 00:09:54,320 --> 00:10:00,080 impatient because the time from building 263 00:09:57,600 --> 00:10:03,040 a proof of concept until getting it into 264 00:10:00,080 --> 00:10:05,760 production and really have it as 265 00:10:03,040 --> 00:10:09,360 like the standard for everything it 266 00:10:05,760

--> 00:10:11,200 takes a while and so i think we see all 267 00:10:09,360 --> 00:10:12,880 these amazing prototypes which are 268 00:10:11,200 --> 00:10:14,640 already working but they are not 269 00:10:12,880 --> 00:10:18,160 implemented yet for example i read an 270 00:10:14,640 --> 00:10:20,959 article about um ai powered traffic 271 00:10:18,160 --> 00:10:24,480 control systems where um the the traffic 272 00:10:20,959 --> 00:10:27,279 lights of city block in simulation were 273 00:10:24,480 --> 00:10:29,600 controlled by an ai based on the traffic 274 00:10:27,279 --> 00:10:32,079 based on accidents based on the speed of 275 00:10:29,600 --> 00:10:33,920 drivers based on the weather and so this 276 00:10:32,079 --> 00:10:36,480 ai was able to 277 00:10:33,920 --> 00:10:38,880 manipulate the traffic in a way that 278 00:10:36,480 --> 00:10:40,640 there were no traffic jams and the cars 279 00:10:38,880 --> 00:10:43,120 were all very efficiently moving through 280 00:10:40,640 --> 00:10:45,920 the city and this is an amazing 281 00:10:43,120 --> 00:10:47,440 application of ai technology but even 282 00:10:45,920 --> 00:10:49,920 though this this experiment was 283 00:10:47,440 --> 00:10:51,920 successful it will take a lot of time 284 00:10:49,920 --> 00:10:54,000 until our infrastructure and our 285 00:10:51,920 --> 00:10:56,959 understanding and our acceptance of this 286 00:10:54,000 --> 00:10:59,120 kind of ai in our everyday society will 287 00:10:56,959 --> 00:11:01,440 take on and be applied even though i 288 00:10:59,120 --> 00:11:03,839 have no doubt that at some point of time 289 00:11:01,440 --> 00:11:05,440 we will be there and it does not make 290 00:11:03,839 --> 00:11:07,760 sense to have these systems not 291 00:11:05,440 --> 00:11:09,839 controlled by an ai because this is the 292 00:11:07,760 --> 00:11:11,680 only thing that makes sense everything 293 00:11:09,839 --> 00:11:14,000 has variables 294 00:11:11,680 --> 00:11:16,000 and of course our systems need to be 295 00:11:14,000 --> 00:11:17,200 smart enough to react to all these 296 00:11:16,000 --> 00:11:19,200 variables 297 00:11:17,200 --> 00:11:21,360 but it will take time and so 298 00:11:19,200 --> 00:11:23,040 when we take a look at nvidia and their 299 00:11:21,360 --> 00:11:24,320 graphics cards and what they are capable 300 00:11:23,040 --> 00:11:27,040 of today 301 00:11:24,320 --> 00:11:28,640 in all their demonstrations we see this 302 00:11:27,040 --> 00:11:31,760 is so amazing but we don't see this 303 00:11:28,640 --> 00:11:33,680 applied yet because the software is not 304 00:11:31,760 --> 00:11:36,000 it did not catch up yet 305 00:11:33,680 --> 00:11:38,079 and so with all the deep fake i am very 306 00:11:36,000 --> 00:11:40,800 sure that in the next years it will 307 00:11:38,079 --> 00:11:43,680 become more and more um 308 00:11:40,800 --> 00:11:45,600 the the standard way of using video and 309 00:11:43,680

--> 00:11:48,480 for example snapchat uh these snapchat 310 00:11:45,600 --> 00:11:51,760 filters they are amazing toys yeah it's 311 00:11:48,480 --> 00:11:53,680 just i i morph my head into a cat hat 312 00:11:51,760 --> 00:11:56,480 like a comic hat and it works perfectly 313 00:11:53,680 --> 00:11:59,360 it is so amazing and now imagine like 314 00:11:56,480 --> 00:12:01,360 five years from now i will be able to 315 00:11:59,360 --> 00:12:03,920 morph my head into a real life barack 316 00:12:01,360 --> 00:12:07,040 obama right because real-time rendering 317 00:12:03,920 --> 00:12:09,360 will be so precise and deep fakes will 318 00:12:07,040 --> 00:12:11,519 be so easy to produce that this will be 319 00:12:09,360 --> 00:12:14,240 no problem at all 320 00:12:11,519 --> 00:12:17,120 yes in the future you're right but i my 321 00:12:14,240 --> 00:12:18,639 claim was that right now the hype is 322 00:12:17,120 --> 00:12:22,720 gone and it's nearly dead again the 323 00:12:18,639 --> 00:12:25,120 topic so data science is staying so i 324 00:12:22,720 --> 00:12:27,120 have a theory that you have a new field 325 00:12:25,120 --> 00:12:30,639 and computer science like also every 326 00:12:27,120 --> 00:12:32,880 like five to seven years so first it was 327 00:12:30,639 --> 00:12:34,240 like fronted engineers you know like 20 328 00:12:32,880 --> 00:12:36,560 years ago everything was kind of 329 00:12:34,240 --> 00:12:38,800 back-end engineering and you did like 330 00:12:36,560 --> 00:12:41,839 hardcore programming and then the web 331 00:12:38,800 --> 00:12:44,560 came and those front-end engineers who 332 00:12:41,839 --> 00:12:46,880 were often self-taught only knew html 333 00:12:44,560 --> 00:12:48,959 css and then a little bit javascript and 334 00:12:46,880 --> 00:12:50,560 oh i only know javascript let's invent 335 00:12:48,959 --> 00:12:52,880 note because i don't know anything else 336 00:12:50,560 --> 00:12:54,959 and say reinvented like things like oh i 337 00:12:52,880 --> 00:12:57,839 now can test my software really i did 338 00:12:54,959 --> 00:12:59,600 this for 40 years so you know 339 00:12:57,839 --> 00:13:01,600 that it was the same with app developers 340 00:12:59,600 --> 00:13:03,440 then like apps became a thing and those 341 00:13:01,600 --> 00:13:06,480 app developers and each time i was at a 342 00:13:03,440 --> 00:13:08,480 conference for app developers or saw a 343 00:13:06,480 --> 00:13:10,320 talk from any app developer i was okay 344 00:13:08,480 --> 00:13:11,839 you're telling me stuff i knew for 10 345 00:13:10,320 --> 00:13:14,240 years that you were like 346 00:13:11,839 --> 00:13:15,600 marketing as new stuff oh we we 347 00:13:14,240 --> 00:13:17,440 discovered something new and stuff 348 00:13:15,600 --> 00:13:18,959 engineering a new pattern i yes this 349 00:13:17,440 --> 00:13:22,160 pattern was subscribed to the book 30 350 00:13:18,959 --> 00:13:24,160 years ago i'm sorry and the same stuff 351 00:13:22,160 --> 00:13:25,680 happened again with data science so 352 00:13:24,160 --> 00:13:27,839 which now became machine learning 353 00:13:25,680 --> 00:13:29,920 engineers right so they were often 354 00:13:27,839 --> 00:13:31,920 self-taught learned some pies and tried 355 00:13:29,920 --> 00:13:34,160 to solve everything with python that's 356 00:13:31,920 --> 00:13:36,560 why pyson is currently hyped at a lot of 357 00:13:34,160 --> 00:13:38,240 startups that clubhouse used python 358 00:13:36,560 --> 00:13:39,120 because it's it's widely used in data 359 00:13:38,240 --> 00:13:41,600 science 360 00:13:39,120 --> 00:13:42,480 and what happened with all the other 361 00:13:41,600 --> 00:13:44,639 fields 362 00:13:42,480 --> 00:13:46,720 is that they matured they become more 363 00:13:44,639 --> 00:13:48,800 major actually right and i think the 364 00:13:46,720 --> 00:13:53,600 same will happen with machine learning 365 00:13:48,800 --> 00:13:55,760 so i would say it's uh currently not 366 00:13:53,600 --> 00:13:57,760 there is research going on but it's 367 00:13:55,760 --> 00:13:59,360 already commodity 368 00:13:57,760 --> 00:14:01,040 right so i would i would say this 369 00:13:59,360 --> 00:14:03,199 prediction models everything is 370 00:14:01,040 --> 00:14:04,880 commodity and you should stop if you're 371 00:14:03,199 --> 00:14:06,880 trying to build something yourself you 372 00:14:04,880 --> 00:14:08,240 should stop and use just the existing 373 00:14:06,880 --> 00:14:10,560 service 374 00:14:08,240 --> 00:14:12,560 yeah i mean this rule is uh can be 375 00:14:10,560 --> 00:14:14,560 applied to everything right i mean 376 00:14:12,560 --> 00:14:16,959 sometimes if you have very very specific 377 00:14:14,560 --> 00:14:18,560 requirements then of course you have to 378 00:14:16,959 --> 00:14:20,800 think about building things up from 379 00:14:18,560 --> 00:14:22,959 scratch but usually if you have a 380 00:14:20,800 --> 00:14:25,199 framework work if you have a pre-trained 381 00:14:22,959 --> 00:14:26,800 model if you have something like a tool 382 00:14:25,199 --> 00:14:28,480 that is already doing exactly what 383 00:14:26,800 --> 00:14:31,040 you're needing then of course it's 384 00:14:28,480 --> 00:14:33,120 common sense to jump in there 385 00:14:31,040 --> 00:14:34,000 but this is always 386 00:14:33,120 --> 00:14:36,240 the 387 00:14:34,000 --> 00:14:38,560 the the smart decision making of 388 00:14:36,240 --> 00:14:41,040 understanding exactly your requirements 389 00:14:38,560 --> 00:14:42,399 and the the the constraints that you're 390 00:14:41,040 --> 00:14:44,639 dealing with 391 00:14:42,399 --> 00:14:47,920 but but this is as you said it's not a 392 00:14:44,639 --> 00:14:49,680 unique require a unique um 393 00:14:47,920 --> 00:14:51,680 challenge for the field of machine 394 00:14:49,680 --> 00:14:54,240 learning this is like everywhere

you 395 00:14:51,680 --> 00:14:56,480 need a content management system sure 396 00:14:54,240 --> 00:14:58,240 you can reinvent one and build one from 397 00:14:56,480 --> 00:15:00,480 scratch but maybe there's already a 398 00:14:58,240 --> 00:15:03,040 product that you can use like tomorrow 399 00:15:00,480 --> 00:15:05,360 simply by buying a license which might 400 00:15:03,040 --> 00:15:07,120 solve all of your problems yes but the 401 00:15:05,360 --> 00:15:08,880 thing is if you wanted to use like a 402 00:15:07,120 --> 00:15:10,160 headless theme as for example like 10 403 00:15:08,880 --> 00:15:11,600 years ago you needed to build it 404 00:15:10,160 --> 00:15:13,760 yourself and some companies did it it 405 00:15:11,600 --> 00:15:15,360 was successful and the same was true 406 00:15:13,760 --> 00:15:18,320 with like prediction models like what 407 00:15:15,360 --> 00:15:20,480 netflix did for like 10 years now um 408 00:15:18,320 --> 00:15:22,320 like predicting what you want to watch 409 00:15:20,480 --> 00:15:24,160 there was nothing in this field you 410 00:15:22,320 --> 00:15:26,720 could just use netflix invented a lot of 411 00:15:24,160 --> 00:15:28,880 stuff did a lot of talks um i watched 412 00:15:26,720 --> 00:15:31,360 nearly all the netflix talks about how 413 00:15:28,880 --> 00:15:33,120 their their stuff works um they're quite 414 00:15:31,360 --> 00:15:35,519 old and 415 00:15:33,120 --> 00:15:38,000 i did the project for a streaming 416 00:15:35,519 --> 00:15:40,560 service which is a tv now in germany 417 00:15:38,000 --> 00:15:42,560 which is now called rtl plus so it's a 418 00:15:40,560 --> 00:15:43,839 german equivalent to netflix like the 419 00:15:42,560 --> 00:15:46,000 biggest streaming server save two 420 00:15:43,839 --> 00:15:48,000 million uh premium subscribers which is 421 00:15:46,000 --> 00:15:48,959 kind of big for national uh streaming 422 00:15:48,000 --> 00:15:52,000 service 423 00:15:48,959 --> 00:15:54,720 and they also started building their own 424 00:15:52,000 --> 00:15:57,040 prediction models like um coziness 425 00:15:54,720 --> 00:15:58,560 similarity then they had created word 426 00:15:57,040 --> 00:16:00,480 embeddings from all the descriptions and 427 00:15:58,560 --> 00:16:03,360 categories and stuff and then uh 428 00:16:00,480 --> 00:16:05,199 document and speedings and then they uh 429 00:16:03,360 --> 00:16:09,040 did some military similarity research 430 00:16:05,199 --> 00:16:11,920 they did everything themselves even like 431 00:16:09,040 --> 00:16:14,560 like widgets and also on and then we had 432 00:16:11,920 --> 00:16:16,320 a bigger project for um it belongs to 433 00:16:14,560 --> 00:16:18,079 the battlesman group which is europe's 434 00:16:16,320 --> 00:16:20,560 biggest media company and it was like 435 00:16:18,079 --> 00:16:22,160 okay we have radio stations we have 436 00:16:20,560 --> 00:16:24,240 multiple tv stations we have

new 437 00:16:22,160 --> 00:16:26,560 stations we own the worldwide biggest 438 00:16:24,240 --> 00:16:28,959 book publisher we have so many magazines 439 00:16:26,560 --> 00:16:29,680 published uber right so maybe we should 440 00:16:28,959 --> 00:16:33,120 uh 441 00:16:29,680 --> 00:16:35,120 think um not in like different entities 442 00:16:33,120 --> 00:16:36,800 but as a whole and create a new digital 443 00:16:35,120 --> 00:16:38,639 product for the digital area where all 444 00:16:36,800 --> 00:16:41,120 the content is in there 445 00:16:38,639 --> 00:16:43,279 and we need also obviously um a 446 00:16:41,120 --> 00:16:45,440 prediction model which could say okay 447 00:16:43,279 --> 00:16:46,800 you would be interested in also this 448 00:16:45,440 --> 00:16:47,920 page like 449 00:16:46,800 --> 00:16:50,000 cross 450 00:16:47,920 --> 00:16:52,480 um or upselling of different brands and 451 00:16:50,000 --> 00:16:55,120 so on and again then they came on board 452 00:16:52,480 --> 00:16:57,279 the the team which built this prediction 453 00:16:55,120 --> 00:16:59,920 stuff they showed us everything and it 454 00:16:57,279 --> 00:17:01,519 was really hardcore python coding uh 455 00:16:59,920 --> 00:17:04,319 really hard to get a production because 456 00:17:01,519 --> 00:17:05,919 it needed a cpu instruction set avx2 457 00:17:04,319 --> 00:17:07,520 which was only running on xenon 458 00:17:05,919 --> 00:17:09,760 processors if you want to throw it into 459 00:17:07,520 --> 00:17:11,839 cloud you need specific 460 00:17:09,760 --> 00:17:14,559 instances of machines to run it and so 461 00:17:11,839 --> 00:17:16,480 on you needed to update it 462 00:17:14,559 --> 00:17:18,559 and then we said okay maybe we have some 463 00:17:16,480 --> 00:17:21,039 elastic search experts elasticsearch can 464 00:17:18,559 --> 00:17:22,720 do similarity stuff so maybe we import 465 00:17:21,039 --> 00:17:24,640 all your catalog data we have from all 466 00:17:22,720 --> 00:17:26,720 the different brands and stuff in the 467 00:17:24,640 --> 00:17:27,839 unified model just in elasticsearch and 468 00:17:26,720 --> 00:17:30,320 use the simulator research in 469 00:17:27,839 --> 00:17:32,960 elasticsearch and guess what it worked 470 00:17:30,320 --> 00:17:35,200 better than the two years 471 00:17:32,960 --> 00:17:38,080 work of like a whole team of machine 472 00:17:35,200 --> 00:17:39,440 learning engineers they invested so my 473 00:17:38,080 --> 00:17:41,360 point is that 474 00:17:39,440 --> 00:17:44,080 there's aws personalized there is 475 00:17:41,360 --> 00:17:46,720 elasticsearch which is already better 476 00:17:44,080 --> 00:17:48,880 than most engineers can build like for 477 00:17:46,720 --> 00:17:51,679 prediction for similarity stuff and so 478 00:17:48,880 --> 00:17:53,520 on use the existing tools don't try to 479 00:17:51,679 --> 00:17:56,000 build something custom there is a cloud 480 00:17:53,520 --> 00:17:58,960 solution for nearly everything out there 481 00:17:56,000 --> 00:18:01,919 yeah but still it depends and for this 482 00:17:58,960 --> 00:18:04,400 particular example and the short term 483 00:18:01,919 --> 00:18:06,400 success you might be right i can't 484 00:18:04,400 --> 00:18:08,640 really predict um like the future of 485 00:18:06,400 --> 00:18:09,440 this project but it could be that in the 486 00:18:08,640 --> 00:18:11,360 long 487 00:18:09,440 --> 00:18:14,240 long run the long tail of this project 488 00:18:11,360 --> 00:18:16,480 it might happen that machine learning 489 00:18:14,240 --> 00:18:18,080 implementation would be more flexible or 490 00:18:16,480 --> 00:18:20,640 would be able to 491 00:18:18,080 --> 00:18:23,600 adapt changes in the data 492 00:18:20,640 --> 00:18:26,000 faster i don't know it might be and so i 493 00:18:23,600 --> 00:18:28,720 think the biggest success of ai based 494 00:18:26,000 --> 00:18:31,520 systems is that they are capable of 495 00:18:28,720 --> 00:18:34,160 adjusting very easily over time by 496 00:18:31,520 --> 00:18:36,400 retraining models over new updated data 497 00:18:34,160 --> 00:18:38,480 sets and i think this is the charm of 498 00:18:36,400 --> 00:18:41,039 this doesn't matter it doesn't mean that 499 00:18:38,480 --> 00:18:44,240 automatically every project is more 500 00:18:41,039 --> 00:18:47,280 successful in the first day um but the 501 00:18:44,240 --> 00:18:50,640 the trend moves along that um everything 502 00:18:47,280 --> 00:18:52,799 is based on ai systems yeah and that 503 00:18:50,640 --> 00:18:54,799 this point i heavily disagree because i 504 00:18:52,799 --> 00:18:56,640 think the value is not in your machine 505 00:18:54,799 --> 00:18:59,280 learning model not in the stuff you're 506 00:18:56,640 --> 00:19:01,919 building on your laptop the currently 507 00:18:59,280 --> 00:19:04,720 the generated value and the hard topic 508 00:19:01,919 --> 00:19:07,160 to tackle is ml ops machine learning ops 509 00:19:04,720 --> 00:19:10,960 getting your model into production so in 510 00:19:07,160 --> 00:19:12,320 2018 uh we got um from 511 00:19:10,960 --> 00:19:14,640 i worked at axl springer where you still 512 00:19:12,320 --> 00:19:16,240 work and we had the brand which is 513 00:19:14,640 --> 00:19:18,480 idialo which is a price comparison 514 00:19:16,240 --> 00:19:21,600 website and we got the header of machine 515 00:19:18,480 --> 00:19:23,520 learning um to us to our subsidiary and 516 00:19:21,600 --> 00:19:24,480 created extra springer ai 517 00:19:23,520 --> 00:19:26,880 and 518 00:19:24,480 --> 00:19:29,600 one project they did for idalo 519 00:19:26,880 --> 00:19:32,400 was like image classification for for 520 00:19:29,600 --> 00:19:34,640 hotel images and the hard thing was not 521 00:19:32,400 --> 00:19:36,559 building the model that was done like in 522 00:19:34,640 --> 00:19:38,160 three

months by two engineers the hard 523 00:19:36,559 --> 00:19:40,160 thing was getting this running in 524 00:19:38,160 --> 00:19:42,240 production because each classification 525 00:19:40,160 --> 00:19:44,720 of an image took like 500 milliseconds 526 00:19:42,240 --> 00:19:46,000 so it was like okay to get one run 527 00:19:44,720 --> 00:19:48,000 through all the images it would take 528 00:19:46,000 --> 00:19:50,320 with even a lot of communication power 529 00:19:48,000 --> 00:19:52,160 because there are like 500 million hotel 530 00:19:50,320 --> 00:19:53,919 images it would take months multiple 531 00:19:52,160 --> 00:19:56,080 months and then how do we update the 532 00:19:53,919 --> 00:19:58,160 model what is when it when something 533 00:19:56,080 --> 00:20:00,320 breaks and so on and the same happened 534 00:19:58,160 --> 00:20:03,039 uh two years later at battlesman i 535 00:20:00,320 --> 00:20:05,039 learned that what they built for tv now 536 00:20:03,039 --> 00:20:06,960 back then was quite simple i could run 537 00:20:05,039 --> 00:20:08,960 it on my laptop getting this and our 538 00:20:06,960 --> 00:20:11,200 cloud infrastructure integrated into 539 00:20:08,960 --> 00:20:13,200 everything updated nightly and so on 540 00:20:11,200 --> 00:20:16,000 that's where the magic happens and the 541 00:20:13,200 --> 00:20:18,320 same is true with my last project 542 00:20:16,000 --> 00:20:20,320 which is now bitly europe 543 00:20:18,320 --> 00:20:21,840 where we not did machine learning but a 544 00:20:20,320 --> 00:20:24,000 lot of data science 545 00:20:21,840 --> 00:20:25,840 and the hard topic was not their python 546 00:20:24,000 --> 00:20:27,520 scripts for predicting something the 547 00:20:25,840 --> 00:20:30,159 hard topic was how do we run this in 548 00:20:27,520 --> 00:20:32,640 production and that's the value i see so 549 00:20:30,159 --> 00:20:34,640 it's not really important if you have 550 00:20:32,640 --> 00:20:37,039 built your model yourself you use aws 551 00:20:34,640 --> 00:20:38,960 personalized you use elasticsearch it's 552 00:20:37,039 --> 00:20:41,200 how you integrate it how you get it 553 00:20:38,960 --> 00:20:43,760 running in your production system 554 00:20:41,200 --> 00:20:45,679 yeah and ml ops is one of the new hypes 555 00:20:43,760 --> 00:20:48,159 that are emerging right now because 556 00:20:45,679 --> 00:20:50,400 exactly what you said is totally true 557 00:20:48,159 --> 00:20:52,720 this is one of the biggest challenges of 558 00:20:50,400 --> 00:20:54,720 actually getting these systems running 559 00:20:52,720 --> 00:20:55,600 and so there are so many companies that 560 00:20:54,720 --> 00:20:58,720 are 561 00:20:55,600 --> 00:21:01,120 providing machine learning operations as 562 00:20:58,720 --> 00:21:02,400 a service and this is now the the next 563 00:21:01,120 --> 00:21:05,440 big thing and 564 00:21:02,400 --> 00:21:07,200 i think on on techcrunch you see so many 565 00:21:05,440 --> 00:21:09,039 investments in these kinds of of 566 00:21:07,200 --> 00:21:09,919 companies and they are 567 00:21:09,039 --> 00:21:11,679 like 568 00:21:09,919 --> 00:21:14,799 booming everything everywhere and they 569 00:21:11,679 --> 00:21:16,480 will be um like over time they will be 570 00:21:14,799 --> 00:21:18,799 swallowed by all the other companies 571 00:21:16,480 --> 00:21:21,120 because everyone needs ml ops everyone 572 00:21:18,799 --> 00:21:23,919 who actually wants to run ai-based 573 00:21:21,120 --> 00:21:25,760 systems needs ml ops and this is quite a 574 00:21:23,919 --> 00:21:28,960 challenge and so this one 575 00:21:25,760 --> 00:21:31,200 is really another hype topic which 576 00:21:28,960 --> 00:21:33,200 is not that um 577 00:21:31,200 --> 00:21:35,600 how do you say it um it does not sound 578 00:21:33,200 --> 00:21:38,159 so great in in the ear of the hype 579 00:21:35,600 --> 00:21:40,960 people because ml ops sounds boring but 580 00:21:38,159 --> 00:21:43,600 this is the field um where everybody is 581 00:21:40,960 --> 00:21:46,480 bleeding right now because it's really 582 00:21:43,600 --> 00:21:48,799 hard to maintain and often times you you 583 00:21:46,480 --> 00:21:50,159 you buy like a contractor or you hire a 584 00:21:48,799 --> 00:21:51,840 contractor to build something like a 585 00:21:50,159 --> 00:21:53,440 machine learning based system and then 586 00:21:51,840 --> 00:21:54,480 they build this up in a year and then 587 00:21:53,440 --> 00:21:56,640 they leave 588 00:21:54,480 --> 00:21:58,960 and then you're happy for a while but 589 00:21:56,640 --> 00:22:00,640 what happens if you need like another 590 00:21:58,960 --> 00:22:03,600 data set or something needs to be 591 00:22:00,640 --> 00:22:05,440 adjusted right or um yeah the topic of 592 00:22:03,600 --> 00:22:07,120 operations and then you start hiring 593 00:22:05,440 --> 00:22:08,880 from scratch and then you are very very 594 00:22:07,120 --> 00:22:11,440 frustrated because it did not run as you 595 00:22:08,880 --> 00:22:14,159 expected this is like a long term 596 00:22:11,440 --> 00:22:17,039 project but still i don't think um that 597 00:22:14,159 --> 00:22:19,520 this is going to go away i think that um 598 00:22:17,039 --> 00:22:21,600 this will stay as the new de facto 599 00:22:19,520 --> 00:22:24,080 standard even though it is true you 600 00:22:21,600 --> 00:22:26,799 can't say that uh machine learning is 601 00:22:24,080 --> 00:22:28,880 the the best answer for every single 602 00:22:26,799 --> 00:22:31,360 project problem yeah but it is still a 603 00:22:28,880 --> 00:22:33,919 very very powerful tool for 604 00:22:31,360 --> 00:22:35,039 very for many many topics that we are 605 00:22:33,919 --> 00:22:38,080 touching 606 00:22:35,039 --> 00:22:40,240 yeah i call those people working on the 607 00:22:38,080 --> 00:22:42,400 top stuff plumbers that's

really 608 00:22:40,240 --> 00:22:45,039 classical engineering you need something 609 00:22:42,400 --> 00:22:46,159 to do data plumbing actually so this is 610 00:22:45,039 --> 00:22:49,120 where the 611 00:22:46,159 --> 00:22:50,720 the missing part of a new emerging field 612 00:22:49,120 --> 00:22:52,880 in computer science 613 00:22:50,720 --> 00:22:55,200 um struggles like always struggles with 614 00:22:52,880 --> 00:22:56,799 like operations because that's the stuff 615 00:22:55,200 --> 00:22:58,880 where i need experience where everything 616 00:22:56,799 --> 00:23:01,440 is the same for years kind of and you 617 00:22:58,880 --> 00:23:04,080 need just a classic engineer who is good 618 00:23:01,440 --> 00:23:05,280 at piping data from one point to another 619 00:23:04,080 --> 00:23:07,039 plumbing 620 00:23:05,280 --> 00:23:09,760 and that's what's missing in most 621 00:23:07,039 --> 00:23:11,760 machine learning teams and if i would 622 00:23:09,760 --> 00:23:14,000 break down what is machine learning ops 623 00:23:11,760 --> 00:23:16,480 it's like plumbing data and that's what 624 00:23:14,000 --> 00:23:19,200 most teams miss 625 00:23:16,480 --> 00:23:21,120 yeah but um as as always usually you 626 00:23:19,200 --> 00:23:23,360 build up something like a department for 627 00:23:21,120 --> 00:23:25,760 that or you outsource this used as a 628 00:23:23,360 --> 00:23:28,000 cloud service um so this is again not 629 00:23:25,760 --> 00:23:29,840 not really a new topic 630 00:23:28,000 --> 00:23:31,840 you of course you have to calculate for 631 00:23:29,840 --> 00:23:33,840 that and um the development of a machine 632 00:23:31,840 --> 00:23:35,600 learning based system does not end with 633 00:23:33,840 --> 00:23:38,799 the machine learning analysts or the 634 00:23:35,600 --> 00:23:40,960 data engineers this is just like 635 00:23:38,799 --> 00:23:42,960 the the creation of 636 00:23:40,960 --> 00:23:45,600 the magic itself but not running the 637 00:23:42,960 --> 00:23:47,279 system yeah and this i i believe it's 638 00:23:45,600 --> 00:23:48,720 the same problem that you have when you 639 00:23:47,279 --> 00:23:51,520 build a 640 00:23:48,720 --> 00:23:53,200 classical system a classical software 641 00:23:51,520 --> 00:23:54,880 engineering system you have the java 642 00:23:53,200 --> 00:23:57,120 engineer and he builds great java 643 00:23:54,880 --> 00:23:59,440 applications but if he never deployed 644 00:23:57,120 --> 00:24:01,279 this to production then it is hard to 645 00:23:59,440 --> 00:24:03,919 actually get a system running that 646 00:24:01,279 --> 00:24:05,760 serves millions of users this will not 647 00:24:03,919 --> 00:24:06,960 work this will not scale and there you 648 00:24:05,760 --> 00:24:09,120 need 649 00:24:06,960 --> 00:24:10,880 an operations engineer or an operations 650 00:24:09,120 --> 00:24:12,960 team or some infrastructure that can 651 00:24:10,880 --> 00:24:15,279 handle this and the same same thing 652 00:24:12,960 --> 00:24:17,039 happens with ai based systems 653 00:24:15,279 --> 00:24:19,279 this is a problem i had with a lot of 654 00:24:17,039 --> 00:24:21,440 machine learning or ai teams and the 655 00:24:19,279 --> 00:24:23,440 engineers working there they think of 656 00:24:21,440 --> 00:24:25,440 themselves as researchers they want to 657 00:24:23,440 --> 00:24:28,000 do like kind of this google research by 658 00:24:25,440 --> 00:24:29,679 build some prototypes and models and 659 00:24:28,000 --> 00:24:32,240 like the companies i work for they build 660 00:24:29,679 --> 00:24:33,520 products and they want problems solved 661 00:24:32,240 --> 00:24:36,080 they're asking okay 662 00:24:33,520 --> 00:24:38,559 can i like save i have 150 people 663 00:24:36,080 --> 00:24:41,039 working in my content department how 664 00:24:38,559 --> 00:24:43,520 many um people or how good can i scale 665 00:24:41,039 --> 00:24:46,320 with your new solution and they're like 666 00:24:43,520 --> 00:24:48,799 i don't know it's it's predicting stuff 667 00:24:46,320 --> 00:24:50,480 so the problem is that they are or think 668 00:24:48,799 --> 00:24:53,440 of themselves as researcher researchers 669 00:24:50,480 --> 00:24:56,159 and others product builders um similar 670 00:24:53,440 --> 00:24:58,480 problem to like engineers right 671 00:24:56,159 --> 00:25:02,080 yeah of course always when i was a 672 00:24:58,480 --> 00:25:04,640 junior engineer i really had no idea 673 00:25:02,080 --> 00:25:07,279 what this means i came from university i 674 00:25:04,640 --> 00:25:09,760 knew how software worked in theory i 675 00:25:07,279 --> 00:25:12,159 knew algorithms and data structures and 676 00:25:09,760 --> 00:25:14,559 everything and i came into this project 677 00:25:12,159 --> 00:25:16,159 and there were so many people but not 678 00:25:14,559 --> 00:25:17,440 really many developers and i did not 679 00:25:16,159 --> 00:25:19,919 really understand what all of these 680 00:25:17,440 --> 00:25:22,159 people were doing yeah and then at some 681 00:25:19,919 --> 00:25:23,520 point of time people were talking about 682 00:25:22,159 --> 00:25:26,840 um 683 00:25:23,520 --> 00:25:30,559 cdns right and deployment pipelines and 684 00:25:26,840 --> 00:25:32,320 geo-redundancy and caching and i never 685 00:25:30,559 --> 00:25:34,159 heard any of those and i was thinking 686 00:25:32,320 --> 00:25:36,000 like wait but there's a software it's 687 00:25:34,159 --> 00:25:38,159 already doing what you want right and 688 00:25:36,000 --> 00:25:40,960 then i learned no no this is like this 689 00:25:38,159 --> 00:25:43,360 this small portion of this whole project 690 00:25:40,960 --> 00:25:46,159 and the majority of work does not go 691 00:25:43,360 --> 00:25:48,640 into the code itself but 692 00:25:46,159 --> 00:25:51,120 everything around it starting from 693 00:25:48,640 -->

00:25:53,679 requirements engineering and 694 00:25:51,120 --> 00:25:56,400 deployment pipelines and 695 00:25:53,679 --> 00:25:58,320 code management i mean even managing 696 00:25:56,400 --> 00:26:00,320 like branches and stuff this is 697 00:25:58,320 --> 00:26:02,240 something that that you have to take 698 00:26:00,320 --> 00:26:04,559 care of right from from development 699 00:26:02,240 --> 00:26:07,600 problems to deployment problems to 700 00:26:04,559 --> 00:26:09,520 testing to redundancy backups backups we 701 00:26:07,600 --> 00:26:12,320 didn't even talk about backups all of 702 00:26:09,520 --> 00:26:14,799 these things and so you can like 703 00:26:12,320 --> 00:26:17,120 have so many teams working on a thing 704 00:26:14,799 --> 00:26:18,799 and maybe have like two developers 705 00:26:17,120 --> 00:26:21,039 who are writing the code 706 00:26:18,799 --> 00:26:23,120 yeah welcome to the wonderful world of 707 00:26:21,039 --> 00:26:25,039 machine learning mlaps right that's 708 00:26:23,120 --> 00:26:27,120 that's what it's about it's just you you 709 00:26:25,039 --> 00:26:28,799 switch machine learning or like software 710 00:26:27,120 --> 00:26:30,480 engineering to machine learning you need 711 00:26:28,799 --> 00:26:32,480 maybe two machine learning experts and a 712 00:26:30,480 --> 00:26:34,799 lot of other people doing all the stuff 713 00:26:32,480 --> 00:26:37,360 around it 714 00:26:34,799 --> 00:26:39,840 yeah yeah but maybe let's talk about 715 00:26:37,360 --> 00:26:42,320 projects uh you and we are like we 716 00:26:39,840 --> 00:26:45,360 touched actually so my favorite machine 717 00:26:42,320 --> 00:26:47,919 learning project or like ai project what 718 00:26:45,360 --> 00:26:50,159 would you guess was it 719 00:26:47,919 --> 00:26:52,960 i guess it was the coffee maker mood bot 720 00:26:50,159 --> 00:26:54,400 yes that's totally correct i loved it so 721 00:26:52,960 --> 00:26:56,799 for the people who haven't seen our 722 00:26:54,400 --> 00:26:58,400 short youtube video sorry it's a german 723 00:26:56,799 --> 00:26:59,919 maybe oh no we've english subtitles 724 00:26:58,400 --> 00:27:03,679 that's great so you can watch it with 725 00:26:59,919 --> 00:27:06,159 english subtitles um it's built on um 726 00:27:03,679 --> 00:27:07,679 opencv open computer vision and it 727 00:27:06,159 --> 00:27:09,679 detects faces it's running on a 728 00:27:07,679 --> 00:27:11,360 raspberry pi 729 00:27:09,679 --> 00:27:13,039 detecting your face and if it sees the 730 00:27:11,360 --> 00:27:14,880 face taking a snapshot and detecting if 731 00:27:13,039 --> 00:27:18,159 you are smiling and if you are not 732 00:27:14,880 --> 00:27:20,559 smiling it's telling a joke and it was 733 00:27:18,159 --> 00:27:25,039 hanging a buff it it had a nice face it 734 00:27:20,559 --> 00:27:26,880 was hanging above our coffee maker um 735 00:27:25,039 --> 00:27:28,960 an espresso machine really great one 736 00:27:26,880 --> 00:27:32,240 rocket and each time i came into the 737 00:27:28,960 --> 00:27:34,000 office i came with the subway i was 738 00:27:32,240 --> 00:27:35,440 grumpy it was early in the morning i 739 00:27:34,000 --> 00:27:36,799 didn't have my first coffee which is 740 00:27:35,440 --> 00:27:38,960 always an issue 741 00:27:36,799 --> 00:27:41,200 and then this stupid machine just told a 742 00:27:38,960 --> 00:27:43,360 joke i laughed and hated it at the same 743 00:27:41,200 --> 00:27:45,120 time and i also think this is again a 744 00:27:43,360 --> 00:27:47,520 great example of 745 00:27:45,120 --> 00:27:50,480 machine learning a machine detecting a 746 00:27:47,520 --> 00:27:52,640 face and learning if you're smiling or 747 00:27:50,480 --> 00:27:54,720 if you're not smiling 748 00:27:52,640 --> 00:27:58,159 right right it was from my point of view 749 00:27:54,720 --> 00:28:00,480 it was an amazing prototype um and 750 00:27:58,159 --> 00:28:03,440 from from implementation standpoint it 751 00:28:00,480 --> 00:28:05,360 is actually quite cheap it it simply had 752 00:28:03,440 --> 00:28:07,919 two components the one was the the 753 00:28:05,360 --> 00:28:10,240 component analyzing the image and 754 00:28:07,919 --> 00:28:12,480 understanding if there is a phase or not 755 00:28:10,240 --> 00:28:14,640 really just a boolean mechanism there is 756 00:28:12,480 --> 00:28:16,559 a phase on this picture and the second 757 00:28:14,640 --> 00:28:17,760 one was analyzing the face and 758 00:28:16,559 --> 00:28:20,240 extracting 759 00:28:17,760 --> 00:28:21,039 or predicting a mood on the space based 760 00:28:20,240 --> 00:28:24,000 on 761 00:28:21,039 --> 00:28:26,880 like markers um if you're smiling or if 762 00:28:24,000 --> 00:28:29,360 you're not smiling and calculating uh um 763 00:28:26,880 --> 00:28:31,679 it was calculating the probability of 764 00:28:29,360 --> 00:28:34,640 your face showing a positive or negative 765 00:28:31,679 --> 00:28:37,200 emotion so it was not even like facial 766 00:28:34,640 --> 00:28:38,880 recognition understanding who you are 767 00:28:37,200 --> 00:28:41,360 right this is not what the system was 768 00:28:38,880 --> 00:28:44,480 doing it was way easier 769 00:28:41,360 --> 00:28:46,559 and so if we cross this threshold of a 770 00:28:44,480 --> 00:28:49,440 certain probability that this person is 771 00:28:46,559 --> 00:28:52,000 not really happy the system was reacting 772 00:28:49,440 --> 00:28:54,640 and accessing a database of chuck norris 773 00:28:52,000 --> 00:28:57,200 jokes or i think anti-jokes or something 774 00:28:54,640 --> 00:28:59,200 and was simply telling a very bad joke 775 00:28:57,200 --> 00:29:01,760 with a computer generated voice which is 776 00:28:59,200 --> 00:29:03,520 another aspect of ai right 777 00:29:01,760 --> 00:29:06,880 and so it was kind of 778 00:29:03,520 --> 00:29:09,440 creepy but still very

very entertaining 779 00:29:06,880 --> 00:29:11,919 and i think this this mix of creepiness 780 00:29:09,440 --> 00:29:14,000 and entertainment um was was the genius 781 00:29:11,919 --> 00:29:16,080 part of the system and of course people 782 00:29:14,000 --> 00:29:17,679 were hating it yeah because they did not 783 00:29:16,080 --> 00:29:19,679 really understand what was happening 784 00:29:17,679 --> 00:29:22,080 there and i think this is if we take 785 00:29:19,679 --> 00:29:24,240 like the sociological aspect of ai 786 00:29:22,080 --> 00:29:26,880 machines um one of the biggest problems 787 00:29:24,240 --> 00:29:29,679 um nobody knew what was happening and at 788 00:29:26,880 --> 00:29:31,440 the end we had to uh take it down 789 00:29:29,679 --> 00:29:34,000 because even um 790 00:29:31,440 --> 00:29:36,320 with the workers union it was like 791 00:29:34,000 --> 00:29:38,640 critical yeah there's a machine taking 792 00:29:36,320 --> 00:29:41,360 pictures without uh like the the 793 00:29:38,640 --> 00:29:43,279 explicit consent and of course these 794 00:29:41,360 --> 00:29:45,039 pictures were not really taken and 795 00:29:43,279 --> 00:29:46,960 stored or anything it was really just 796 00:29:45,039 --> 00:29:48,720 this real-time analysis but still there 797 00:29:46,960 --> 00:29:50,799 was a camera attached to it and nobody 798 00:29:48,720 --> 00:29:53,360 was actually understanding and in terms 799 00:29:50,799 --> 00:29:55,440 of ai ethics yeah this is like a 800 00:29:53,360 --> 00:29:57,840 critical thing everyone who comes in 801 00:29:55,440 --> 00:29:59,279 contact with ai needs of course to 802 00:29:57,840 --> 00:30:02,080 understand what is happening there 803 00:29:59,279 --> 00:30:05,120 otherwise it's like a violation of your 804 00:30:02,080 --> 00:30:07,039 of your rights yeah um but from from the 805 00:30:05,120 --> 00:30:09,200 engineering and scientific point of view 806 00:30:07,039 --> 00:30:12,720 it was really an amazing 807 00:30:09,200 --> 00:30:15,679 example of what you can do with very 808 00:30:12,720 --> 00:30:18,480 easy to use open source components 809 00:30:15,679 --> 00:30:21,360 and like a cheap i don't know like three 810 00:30:18,480 --> 00:30:23,120 euro camera attached to a raspberry pi 811 00:30:21,360 --> 00:30:25,679 which is also like what what thirty 812 00:30:23,120 --> 00:30:26,720 dollars or so and worth so really cheap 813 00:30:25,679 --> 00:30:29,760 hardware 814 00:30:26,720 --> 00:30:32,320 open source free available tools 815 00:30:29,760 --> 00:30:34,640 using these amazing things that that 816 00:30:32,320 --> 00:30:36,960 sound like science fiction if you tell 817 00:30:34,640 --> 00:30:39,600 would have told this to a person like 20 818 00:30:36,960 --> 00:30:41,600 years ago that that basically 819 00:30:39,600 --> 00:30:43,520 everyone can use this today 820 00:30:41,600 --> 00:30:47,600 yeah and it's commodity and the question 821 00:30:43,520 --> 00:30:49,840 is if you take like nowadays 2022 like 822 00:30:47,600 --> 00:30:52,480 10 engineers and maybe 10 product people 823 00:30:49,840 --> 00:30:54,720 and ask show this product and ask them 824 00:30:52,480 --> 00:30:56,720 is this ai 825 00:30:54,720 --> 00:30:58,720 i'm not sure what they would say and the 826 00:30:56,720 --> 00:31:01,120 other point is like i was at a lot of 827 00:30:58,720 --> 00:31:03,600 hackathons and a lot of people wear a 828 00:31:01,120 --> 00:31:05,760 building stuffed with prediction and 829 00:31:03,600 --> 00:31:08,320 image classification and so on that 830 00:31:05,760 --> 00:31:10,880 happened a lot at hackathons and they 831 00:31:08,320 --> 00:31:14,240 used hard models training models it took 832 00:31:10,880 --> 00:31:14,960 them very long so one team just used 833 00:31:14,240 --> 00:31:16,480 like 834 00:31:14,960 --> 00:31:20,399 half of the hackers on which was a two 835 00:31:16,480 --> 00:31:23,120 day hackathon only to label data so like 836 00:31:20,399 --> 00:31:25,519 fruits and vegetables so that they could 837 00:31:23,120 --> 00:31:28,320 train their model um and with the stuff 838 00:31:25,519 --> 00:31:30,840 used for the coffee maker mootbot that's 839 00:31:28,320 --> 00:31:33,120 like pre-existing stuff so even the 840 00:31:30,840 --> 00:31:36,480 classification of if you're happy or not 841 00:31:33,120 --> 00:31:38,640 there were models out there so uh it was 842 00:31:36,480 --> 00:31:39,600 very very easy to use and still very 843 00:31:38,640 --> 00:31:42,480 cool 844 00:31:39,600 --> 00:31:44,880 so yeah that's true yeah 845 00:31:42,480 --> 00:31:46,159 brings me back to our original point um 846 00:31:44,880 --> 00:31:48,880 what is the state of artificial 847 00:31:46,159 --> 00:31:50,000 intelligence intelligence so i would say 848 00:31:48,880 --> 00:31:51,840 um 849 00:31:50,000 --> 00:31:53,039 it's part of a lot of products like with 850 00:31:51,840 --> 00:31:55,279 prediction 851 00:31:53,039 --> 00:31:57,440 and even a lot of sus companies they're 852 00:31:55,279 --> 00:32:01,200 using personalization and you could be 853 00:31:57,440 --> 00:32:02,720 interested in whatever right so we even 854 00:32:01,200 --> 00:32:04,080 use it um 855 00:32:02,720 --> 00:32:04,880 with uh 856 00:32:04,080 --> 00:32:06,080 with 857 00:32:04,880 --> 00:32:08,080 a simple 858 00:32:06,080 --> 00:32:10,480 sas tool where you have a product and we 859 00:32:08,080 --> 00:32:11,440 write like five content articles 860 00:32:10,480 --> 00:32:12,240 per 861 00:32:11,440 --> 00:32:14,880 per 862 00:32:12,240 --> 00:32:17,039 per week and then in the product itself 863 00:32:14,880 --> 00:32:18,799 if you add a special feature we say okay 864 00:32:17,039 --> 00:32:21,519 you could be also interested in this and 865 00:32:18,799 --> 00:32:23,039 this and also we see your behavior what 866 00:32:21,519 --> 00:32:24,640

parts that you use and say okay maybe 867 00:32:23,039 --> 00:32:26,159 you want to read about 868 00:32:24,640 --> 00:32:28,799 restaurants because there are different 869 00:32:26,159 --> 00:32:30,880 kind of verticals how you could use our 870 00:32:28,799 --> 00:32:33,440 software in restaurants 871 00:32:30,880 --> 00:32:35,760 it's running on on very cheap stuff and 872 00:32:33,440 --> 00:32:37,760 on elasticsearch with some 873 00:32:35,760 --> 00:32:39,519 similarity stuff and prediction and so 874 00:32:37,760 --> 00:32:41,039 on and i would say 875 00:32:39,519 --> 00:32:42,880 we would never market this as machine 876 00:32:41,039 --> 00:32:44,559 learning so right because this is a very 877 00:32:42,880 --> 00:32:46,159 simple sas product and we would never 878 00:32:44,559 --> 00:32:47,600 market this oh we are using machine 879 00:32:46,159 --> 00:32:50,640 learning to see what you're using 880 00:32:47,600 --> 00:32:52,720 because it's already commodity and 881 00:32:50,640 --> 00:32:53,760 like three years four years five years 882 00:32:52,720 --> 00:32:56,399 ago 883 00:32:53,760 --> 00:32:57,919 in the startup incubators where also was 884 00:32:56,399 --> 00:32:59,360 as a mentor 885 00:32:57,919 --> 00:33:00,720 like how the companies did something 886 00:32:59,360 --> 00:33:02,559 with machine learning was such a big 887 00:33:00,720 --> 00:33:04,080 hype so if you didn't do something with 888 00:33:02,559 --> 00:33:06,559 machine learning okay 889 00:33:04,080 --> 00:33:08,640 that would be hard but this is this is 890 00:33:06,559 --> 00:33:10,559 gun so i would really say the hype is 891 00:33:08,640 --> 00:33:13,519 gun there are some companies trying to 892 00:33:10,559 --> 00:33:16,159 make money with products where you use 893 00:33:13,519 --> 00:33:18,880 like deep fakes of of 894 00:33:16,159 --> 00:33:21,279 stuff creating voices but i still think 895 00:33:18,880 --> 00:33:22,960 that they are far away from what we 896 00:33:21,279 --> 00:33:25,360 would expect if you see the metrics and 897 00:33:22,960 --> 00:33:27,760 how every an agent works and could fake 898 00:33:25,360 --> 00:33:30,399 like a human face and voice and so on a 899 00:33:27,760 --> 00:33:32,000 perfect way um that will take seven 900 00:33:30,399 --> 00:33:33,120 years i would say that will take another 901 00:33:32,000 --> 00:33:37,279 seven years 902 00:33:33,120 --> 00:33:38,880 um and the other point is this big um 903 00:33:37,279 --> 00:33:41,840 uh 904 00:33:38,880 --> 00:33:43,840 language models which auto generate text 905 00:33:41,840 --> 00:33:45,600 they are awesome so there are companies 906 00:33:43,840 --> 00:33:48,159 trying to create uh 907 00:33:45,600 --> 00:33:50,240 blog articles like copy writing 908 00:33:48,159 --> 00:33:52,720 with machine learning models and they're 909 00:33:50,240 --> 00:33:55,679 getting better each half year i would 910 00:33:52,720 --> 00:33:57,919 say though every um six months i read a 911 00:33:55,679 --> 00:34:00,480 new blog article about yeah it's so good 912 00:33:57,919 --> 00:34:03,600 it's so great so everything was awesome 913 00:34:00,480 --> 00:34:05,600 like 2021 where i say okay wow i just 914 00:34:03,600 --> 00:34:07,919 enter a sentence and it writes a full 915 00:34:05,600 --> 00:34:10,159 text about something and if i just 916 00:34:07,919 --> 00:34:12,079 change some stuff i could maybe even use 917 00:34:10,159 --> 00:34:14,159 it for my for my blog 918 00:34:12,079 --> 00:34:16,159 if i just want some content for zero and 919 00:34:14,159 --> 00:34:18,399 so on and now like half year later it's 920 00:34:16,159 --> 00:34:21,200 like oh wow you cannot distinguish if 921 00:34:18,399 --> 00:34:22,320 it's a human or if it's like a bot 922 00:34:21,200 --> 00:34:25,760 so 923 00:34:22,320 --> 00:34:28,240 this i think this um text generation 924 00:34:25,760 --> 00:34:31,280 this will be the the next thing like 925 00:34:28,240 --> 00:34:33,919 this year next year um where we see a 926 00:34:31,280 --> 00:34:34,960 lot of like new um products being 927 00:34:33,919 --> 00:34:37,200 released 928 00:34:34,960 --> 00:34:38,560 still thinking that um auto generated 929 00:34:37,200 --> 00:34:40,720 voices and 930 00:34:38,560 --> 00:34:42,800 deep fakes is not going to happen we 931 00:34:40,720 --> 00:34:44,399 were promised this like three years ago 932 00:34:42,800 --> 00:34:46,159 i never saw it happen and i don't 933 00:34:44,399 --> 00:34:47,599 believe in it for the next 934 00:34:46,159 --> 00:34:49,679 years 935 00:34:47,599 --> 00:34:51,679 yeah but but still it's just a question 936 00:34:49,679 --> 00:34:54,320 of time i'm very sure it's a question of 937 00:34:51,679 --> 00:34:56,639 time and uh when i was at university uh 938 00:34:54,320 --> 00:34:59,599 one of our teams um that i worked with 939 00:34:56,639 --> 00:35:02,560 um they were building uh smart home 940 00:34:59,599 --> 00:35:05,200 appliances this was in 2003. and back 941 00:35:02,560 --> 00:35:08,000 then was uh we we were showing this at 942 00:35:05,200 --> 00:35:10,560 uh at seabed like the the big fair uh 943 00:35:08,000 --> 00:35:12,240 where we were like new technology was 944 00:35:10,560 --> 00:35:13,920 presented and it was the same thing 945 00:35:12,240 --> 00:35:17,040 people were saying yeah this is a like a 946 00:35:13,920 --> 00:35:19,119 prototype but it it will take so long 947 00:35:17,040 --> 00:35:23,200 until we see it and it's true it took a 948 00:35:19,119 --> 00:35:25,760 long time since 2003 until smart home 949 00:35:23,200 --> 00:35:28,160 appliances actually became a thing but 950 00:35:25,760 --> 00:35:29,920 it was happening and so i i'm pretty 951 00:35:28,160 --> 00:35:32,880 sure with all the deep fakes 952 00:35:29,920 --> 00:35:34,720 that at some point of time maybe in 953 00:35:32,880 --> 00:35:37,760 three years

five years seven years ten 954 00:35:34,720 --> 00:35:39,040 years this will simply be a thing where 955 00:35:37,760 --> 00:35:42,240 then 956 00:35:39,040 --> 00:35:44,480 the lawmakers the legislators will start 957 00:35:42,240 --> 00:35:46,560 i don't know making it illegal to use 958 00:35:44,480 --> 00:35:48,960 the fake technology or something because 959 00:35:46,560 --> 00:35:52,400 of the threat against the society yeah 960 00:35:48,960 --> 00:35:54,400 but i i'm i'm so sure as we see these um 961 00:35:52,400 --> 00:35:56,640 snapchat filters this will become a 962 00:35:54,400 --> 00:35:58,320 thing that you create a filter of i 963 00:35:56,640 --> 00:36:00,079 don't know your girlfriend and then you 964 00:35:58,320 --> 00:36:03,040 become your girlfriend in the chat room 965 00:36:00,079 --> 00:36:04,320 and it will be really lifelike yeah 966 00:36:03,040 --> 00:36:06,720 these things 967 00:36:04,320 --> 00:36:08,480 what do you predict will be like hype 968 00:36:06,720 --> 00:36:11,720 this year and machine learning or next 969 00:36:08,480 --> 00:36:11,720 big thing 970 00:36:12,640 --> 00:36:16,720 i have no idea 971 00:36:15,599 --> 00:36:19,040 i don't know 972 00:36:16,720 --> 00:36:23,359 there are so many so many experiments i 973 00:36:19,040 --> 00:36:27,040 just read um an article about 974 00:36:23,359 --> 00:36:29,200 a mixture of ai and metaverse 975 00:36:27,040 --> 00:36:32,000 and it was about 976 00:36:29,200 --> 00:36:34,720 a system i forgot the name where while 977 00:36:32,000 --> 00:36:37,040 you were using your avatar in this 978 00:36:34,720 --> 00:36:40,560 metaverse environment in this vr 979 00:36:37,040 --> 00:36:43,200 environment your avatar is learning your 980 00:36:40,560 --> 00:36:44,800 speech pattern and the topics you're 981 00:36:43,200 --> 00:36:46,640 talking about while you're moving in 982 00:36:44,800 --> 00:36:48,880 there and even your body movements 983 00:36:46,640 --> 00:36:50,800 because it is like capturing you with a 984 00:36:48,880 --> 00:36:53,200 with a lighthouse camera 985 00:36:50,800 --> 00:36:55,520 and so your your avatar is moving like 986 00:36:53,200 --> 00:36:56,800 you and this means that this avatar is 987 00:36:55,520 --> 00:36:58,320 actually 988 00:36:56,800 --> 00:37:01,040 starting to understand you and starting 989 00:36:58,320 --> 00:37:04,320 to imitate you so if you're all offline 990 00:37:01,040 --> 00:37:06,480 your avatar can basically continue 991 00:37:04,320 --> 00:37:08,400 having the same conversations as you do 992 00:37:06,480 --> 00:37:10,640 using the same speech patterns using 993 00:37:08,400 --> 00:37:13,839 your same voice and your movements and 994 00:37:10,640 --> 00:37:16,320 the idea was that it was the the like a 995 00:37:13,839 --> 00:37:17,760 god mode yeah so if in the in the real 996 00:37:16,320 --> 00:37:20,079 world you die 997 00:37:17,760 --> 00:37:23,839 um and you have bought

998 00:37:20,079 --> 00:37:26,880 a metaverse real estate there people can 999 00:37:23,839 --> 00:37:29,680 go there and visit you after you have 1000 00:37:26,880 --> 00:37:32,000 died and you will be you with your voice 1001 00:37:29,680 --> 00:37:34,320 and your knowledge and your speech 1002 00:37:32,000 --> 00:37:36,720 patterns yeah and they are already i 1003 00:37:34,320 --> 00:37:39,599 think uh microsoft had a patent patent 1004 00:37:36,720 --> 00:37:42,800 in 2021 where they uh 1005 00:37:39,599 --> 00:37:44,800 like did this patent on deadbots i'm not 1006 00:37:42,800 --> 00:37:46,880 sure if they used this word having a 1007 00:37:44,800 --> 00:37:49,359 chatbot learning all of your email 1008 00:37:46,880 --> 00:37:52,320 conversations and social media messages 1009 00:37:49,359 --> 00:37:54,480 and so you can talk to this chatbot and 1010 00:37:52,320 --> 00:37:56,400 you can't it's like a touring test you 1011 00:37:54,480 --> 00:37:58,880 can't distinguish between 1012 00:37:56,400 --> 00:38:01,119 i am i talking to vashi right now or am 1013 00:37:58,880 --> 00:38:03,200 i talking to the washi bot which is 1014 00:38:01,119 --> 00:38:04,720 trained on all of your social media data 1015 00:38:03,200 --> 00:38:07,119 and this one this does not have an 1016 00:38:04,720 --> 00:38:09,839 avatar this is simply like the touring 1017 00:38:07,119 --> 00:38:11,920 communication test um but here we see 1018 00:38:09,839 --> 00:38:14,720 another example of what is 1019 00:38:11,920 --> 00:38:17,119 technically already possible and in like 1020 00:38:14,720 --> 00:38:19,440 a prototypical state and now with the 1021 00:38:17,119 --> 00:38:22,160 emergence of the metaverse and uh meta 1022 00:38:19,440 --> 00:38:24,640 creatures that you will meet there being 1023 00:38:22,160 --> 00:38:26,640 like a i generated or hear you real 1024 00:38:24,640 --> 00:38:28,800 human beings i think this will be 1025 00:38:26,640 --> 00:38:30,720 another thing that we will see now more 1026 00:38:28,800 --> 00:38:32,880 and more often in the next years where 1027 00:38:30,720 --> 00:38:34,960 the metaverse will start getting into 1028 00:38:32,880 --> 00:38:38,240 this hype phase and everything will be 1029 00:38:34,960 --> 00:38:40,079 created for the metaverse being 1030 00:38:38,240 --> 00:38:42,000 even though it might not make sense or 1031 00:38:40,079 --> 00:38:44,400 it makes sense yeah but everything will 1032 00:38:42,000 --> 00:38:46,400 will be created and so maybe it's it's 1033 00:38:44,400 --> 00:38:48,880 ai based uh 1034 00:38:46,400 --> 00:38:50,560 um metaverse tools 1035 00:38:48,880 --> 00:38:53,680 yeah it's it's i think it was like in 1036 00:38:50,560 --> 00:38:56,079 superman where uh he went to his ship he 1037 00:38:53,680 --> 00:38:58,320 came and then his father was yeah i'm an 1038 00:38:56,079 --> 00:39:00,000 avatar of your father and he could talk 1039 00:38:58,320 --> 00:39:02,320 to him but still it was just a 1040 00:39:00,000 --> 00:39:03,839 simulation it was not the real deal so 1041 00:39:02,320 --> 00:39:05,119 but exactly 1042 00:39:03,839 --> 00:39:08,320 there will be a point when you ask 1043 00:39:05,119 --> 00:39:11,440 yourself what makes a person right 1044 00:39:08,320 --> 00:39:14,160 yeah yeah of course and i mean this is a 1045 00:39:11,440 --> 00:39:16,720 perfect perfect question if i have an 1046 00:39:14,160 --> 00:39:18,720 avatar and it looks like me it speaks 1047 00:39:16,720 --> 00:39:20,640 like me and it even has the same 1048 00:39:18,720 --> 00:39:22,800 knowledge as i have because it was 1049 00:39:20,640 --> 00:39:25,280 trained long enough on everything that i 1050 00:39:22,800 --> 00:39:27,920 i was doing and saying 1051 00:39:25,280 --> 00:39:30,560 then we start talking about again about 1052 00:39:27,920 --> 00:39:33,040 this concept of a soul yeah something 1053 00:39:30,560 --> 00:39:35,599 that makes me me which is living 1054 00:39:33,040 --> 00:39:38,400 somewhere in me but nobody can touch it 1055 00:39:35,599 --> 00:39:41,200 or describe it it's just there and if i 1056 00:39:38,400 --> 00:39:43,040 have a perfect copy of myself then the 1057 00:39:41,200 --> 00:39:44,960 only thing where you can point your 1058 00:39:43,040 --> 00:39:47,040 finger to and saying this is not a human 1059 00:39:44,960 --> 00:39:50,000 being or this is not a real person is 1060 00:39:47,040 --> 00:39:51,920 does it have a soul or not and then and 1061 00:39:50,000 --> 00:39:53,359 now we are starting getting very very 1062 00:39:51,920 --> 00:39:56,800 philosophical 1063 00:39:53,359 --> 00:39:58,880 because if you if you can't distinguish 1064 00:39:56,800 --> 00:40:00,000 between the original and the copy 1065 00:39:58,880 --> 00:40:02,320 then 1066 00:40:00,000 --> 00:40:04,880 is there even a point of trying to 1067 00:40:02,320 --> 00:40:06,960 distinguish it yeah and then that that 1068 00:40:04,880 --> 00:40:10,000 becomes now very very philosophical yeah 1069 00:40:06,960 --> 00:40:11,839 but this is a very big if so for example 1070 00:40:10,000 --> 00:40:14,400 ghost in the shell this very famous 1071 00:40:11,839 --> 00:40:16,800 anime right it has a different answer 1072 00:40:14,400 --> 00:40:19,599 the answer is uh for like most of the 1073 00:40:16,800 --> 00:40:21,119 the human history uh you have a ghost in 1074 00:40:19,599 --> 00:40:23,839 yourself like your soul and you can't 1075 00:40:21,119 --> 00:40:26,800 copy it so you can replace every nearly 1076 00:40:23,839 --> 00:40:29,040 everything in your body except for the 1077 00:40:26,800 --> 00:40:30,880 kind of a ghost like your brain because 1078 00:40:29,040 --> 00:40:32,720 there's something in there part of it 1079 00:40:30,880 --> 00:40:34,160 which you cannot replicate and there it 1080 00:40:32,720 --> 00:40:36,400 goes in the shell there is no real 1081

00:40:34,160 --> 00:40:39,520 artificial intelligence because it's not 1082 00:40:36,400 --> 00:40:41,920 possible to do it and the uh the the 1083 00:40:39,520 --> 00:40:45,119 thing on the first movie is actually um 1084 00:40:41,920 --> 00:40:48,800 somehow one creature emerged like an 1085 00:40:45,119 --> 00:40:50,319 artificial um being one so that's that's 1086 00:40:48,800 --> 00:40:52,240 actually the story of the movie that it 1087 00:40:50,319 --> 00:40:54,480 was for multiple years not possible you 1088 00:40:52,240 --> 00:40:56,400 could like replace any part of your body 1089 00:40:54,480 --> 00:40:59,680 but you still needed your ghost you're 1090 00:40:56,400 --> 00:41:02,400 like your being and the main protagonist 1091 00:40:59,680 --> 00:41:04,319 is like completely an android where 1092 00:41:02,400 --> 00:41:06,000 everything is mechanical except 1093 00:41:04,319 --> 00:41:07,359 obviously for her ghost yeah her whole 1094 00:41:06,000 --> 00:41:09,599 body is even completely destroyed that 1095 00:41:07,359 --> 00:41:10,400 she gets a new body or he she whatever 1096 00:41:09,599 --> 00:41:11,839 it's not 1097 00:41:10,400 --> 00:41:14,079 important anymore 1098 00:41:11,839 --> 00:41:15,280 um gets a completely new body 1099 00:41:14,079 --> 00:41:17,599 um 1100 00:41:15,280 --> 00:41:19,040 and it was never possible to create ai 1101 00:41:17,599 --> 00:41:21,680 this is another option right that we 1102 00:41:19,040 --> 00:41:23,599 will this is also like a religious topic 1103 00:41:21,680 --> 00:41:26,640 again which we could touch at some point 1104 00:41:23,599 --> 00:41:30,160 together with ethics and ai uh is a 1105 00:41:26,640 --> 00:41:31,119 human going to be a god and create 1106 00:41:30,160 --> 00:41:32,720 a new 1107 00:41:31,119 --> 00:41:34,160 living creature like something 1108 00:41:32,720 --> 00:41:35,760 completely new or is it actually 1109 00:41:34,160 --> 00:41:38,240 possible because we're humans and we 1110 00:41:35,760 --> 00:41:40,319 cannot create ai this is like a thing 1111 00:41:38,240 --> 00:41:42,400 which is not feasible for us which is a 1112 00:41:40,319 --> 00:41:44,720 great question in my opinion 1113 00:41:42,400 --> 00:41:46,560 absolutely and especially when we take a 1114 00:41:44,720 --> 00:41:48,640 look from the scientific point of view 1115 00:41:46,560 --> 00:41:51,839 and from the scientific point of view it 1116 00:41:48,640 --> 00:41:53,839 is really really hard to um to point to 1117 00:41:51,839 --> 00:41:55,760 to this thing like a ghost or a soul 1118 00:41:53,839 --> 00:41:58,319 because scientifically you can't really 1119 00:41:55,760 --> 00:42:01,119 prove it and you can't really touch it 1120 00:41:58,319 --> 00:42:04,319 and therefore you can't replicate it and 1121 00:42:01,119 --> 00:42:07,760 and if something like the um 1122 00:42:04,319 --> 00:42:10,079 your pre your um mental presence is what 1123 00:42:07,760 --> 00:42:13,280 we are calling a soul then we might even 1124 00:42:10,079 --> 00:42:15,760 be able to replicate and simulate it 1125 00:42:13,280 --> 00:42:18,319 yeah if we can make 1126 00:42:15,760 --> 00:42:20,960 like a scientific list and a scientific 1127 00:42:18,319 --> 00:42:24,400 description of what a soul 1128 00:42:20,960 --> 00:42:26,880 is like compassion and 1129 00:42:24,400 --> 00:42:29,200 presence and i don't know no 1130 00:42:26,880 --> 00:42:31,440 then then if if we can pinpoint it then 1131 00:42:29,200 --> 00:42:33,920 we probably can simulate it 1132 00:42:31,440 --> 00:42:36,800 and if it is possible to simulate it 1133 00:42:33,920 --> 00:42:39,280 then we might argue that it is actually 1134 00:42:36,800 --> 00:42:40,160 there and we can create a ghost in the 1135 00:42:39,280 --> 00:42:40,880 shell 1136 00:42:40,160 --> 00:42:43,680 but 1137 00:42:40,880 --> 00:42:45,760 as you said it's kind of spiritual 1138 00:42:43,680 --> 00:42:49,359 religious and 1139 00:42:45,760 --> 00:42:51,760 we we leave the the realm of scientific 1140 00:42:49,359 --> 00:42:53,680 uh conversation or the scientific 1141 00:42:51,760 --> 00:42:54,880 means of of doing these things but 1142 00:42:53,680 --> 00:42:57,200 interestingly 1143 00:42:54,880 --> 00:42:59,119 we start having to have this real 1144 00:42:57,200 --> 00:43:01,520 conversation it's not really physical 1145 00:42:59,119 --> 00:43:03,760 philosophical anymore if we start 1146 00:43:01,520 --> 00:43:06,160 interacting with beings like that and 1147 00:43:03,760 --> 00:43:08,640 it's not science fiction anymore like i 1148 00:43:06,160 --> 00:43:10,560 i remember i saw a ghost on the shelf 1149 00:43:08,640 --> 00:43:13,839 back in the 90s and i loved it so much 1150 00:43:10,560 --> 00:43:16,480 it was so inspiring but it was pure 1151 00:43:13,839 --> 00:43:17,680 science fiction and now we start talking 1152 00:43:16,480 --> 00:43:18,880 about these 1153 00:43:17,680 --> 00:43:20,720 topics 1154 00:43:18,880 --> 00:43:23,200 not in the sense of science fiction 1155 00:43:20,720 --> 00:43:26,800 anymore but actually having to deal with 1156 00:43:23,200 --> 00:43:28,720 virtual beings which might be considered 1157 00:43:26,800 --> 00:43:30,319 alive 1158 00:43:28,720 --> 00:43:32,160 it's so crazy 1159 00:43:30,319 --> 00:43:34,480 fun fact ghost in the shell is 1160 00:43:32,160 --> 00:43:36,800 inspiration for matrix and for a lot of 1161 00:43:34,480 --> 00:43:39,599 scenes and it's like what the 1162 00:43:36,800 --> 00:43:41,280 no uh vacholsky siblings you need to 1163 00:43:39,599 --> 00:43:43,359 call them um 1164 00:43:41,280 --> 00:43:45,359 like said okay this is like actually we 1165 00:43:43,359 --> 00:43:46,800 took this anime and the best out of it 1166 00:43:45,359 --> 00:43:48,160 and made a movie out of it that's where 1167 00:43:46,800 --> 00:43:50,400 the matrix came from a lot of these 1168

00:43:48,160 --> 00:43:51,520 great ideas actually come from there 1169 00:43:50,400 --> 00:43:52,960 obviously 1170 00:43:51,520 --> 00:43:54,960 there are differences because there is 1171 00:43:52,960 --> 00:43:56,480 real ai in the matrix and then goes into 1172 00:43:54,960 --> 00:43:57,359 shell not really 1173 00:43:56,480 --> 00:44:00,319 um 1174 00:43:57,359 --> 00:44:02,400 and i we totally need to talk about what 1175 00:44:00,319 --> 00:44:05,599 actually is the meter verse and we need 1176 00:44:02,400 --> 00:44:07,760 to talk about essex and ai so that i i 1177 00:44:05,599 --> 00:44:11,440 really think one of our next like two 1178 00:44:07,760 --> 00:44:13,839 now or next podcast will be those topics 1179 00:44:11,440 --> 00:44:14,800 yes absolutely 1180 00:44:13,839 --> 00:44:17,119 yeah 1181 00:44:14,800 --> 00:44:19,839 we touched the meta was so often uh like 1182 00:44:17,119 --> 00:44:21,599 when we talked about vr ar um i think 1183 00:44:19,839 --> 00:44:24,960 even when we talked about 1184 00:44:21,599 --> 00:44:26,720 cryptocurrencies and stuff and now so we 1185 00:44:24,960 --> 00:44:27,920 should talk about it actually because i 1186 00:44:26,720 --> 00:44:29,760 think absolutely 1187 00:44:27,920 --> 00:44:31,280 we are still not really in the hype 1188 00:44:29,760 --> 00:44:34,160 phase for the meteoroids that's really 1189 00:44:31,280 --> 00:44:36,400 like a nerd topic for hardcore nerds and 1190 00:44:34,160 --> 00:44:39,599 it will take another like three years uh 1191 00:44:36,400 --> 00:44:42,800 until my niece or whoever actually will 1192 00:44:39,599 --> 00:44:46,079 even know the word meter worse right 1193 00:44:42,800 --> 00:44:48,319 yeah it is true um i think um 1194 00:44:46,079 --> 00:44:51,280 mark zuckerberg tries to 1195 00:44:48,319 --> 00:44:53,040 kick start the hype around the metaverse 1196 00:44:51,280 --> 00:44:55,920 but all we have right now is these 1197 00:44:53,040 --> 00:44:59,200 commercials and it is promising a lot of 1198 00:44:55,920 --> 00:45:01,920 things and with a with a market power of 1199 00:44:59,200 --> 00:45:04,319 meta i actually believe that it will 1200 00:45:01,920 --> 00:45:05,680 kickstart at some point of time and 1201 00:45:04,319 --> 00:45:07,760 somehow 1202 00:45:05,680 --> 00:45:10,720 some companies and some products are 1203 00:45:07,760 --> 00:45:13,520 already touching this realm but it is so 1204 00:45:10,720 --> 00:45:16,319 fuzzy that nobody can really draw this 1205 00:45:13,520 --> 00:45:19,599 line about this is now metaverse this is 1206 00:45:16,319 --> 00:45:23,440 not metaverse and so i i think everyone 1207 00:45:19,599 --> 00:45:25,200 is still very um very careful with this 1208 00:45:23,440 --> 00:45:27,839 topic because nobody actually knows if 1209 00:45:25,200 --> 00:45:29,040 they are already in there or not yeah so 1210 00:45:27,839 --> 00:45:30,160 it's it's 1211 00:45:29,040 --> 00:45:32,480 again like 1212 00:45:30,160 --> 00:45:34,400 some kind of philosophy and it's too 1213 00:45:32,480 --> 00:45:37,040 fuzzy to actually 1214 00:45:34,400 --> 00:45:38,960 say if we are already there or not yeah 1215 00:45:37,040 --> 00:45:40,800 so let's talk about this yeah you 1216 00:45:38,960 --> 00:45:43,280 already teased like the whole episode so 1217 00:45:40,800 --> 00:45:44,560 we can already use as a summary i guess 1218 00:45:43,280 --> 00:45:45,359 exactly 1219 00:45:44,560 --> 00:45:48,160 so 1220 00:45:45,359 --> 00:45:50,640 um let's rub it up actually so 1221 00:45:48,160 --> 00:45:53,040 i think i learned that uh today that 1222 00:45:50,640 --> 00:45:55,440 prediction models are totally commodity 1223 00:45:53,040 --> 00:45:57,760 and that's where we are right now and 1224 00:45:55,440 --> 00:45:59,359 that we are not patient enough as humans 1225 00:45:57,760 --> 00:46:01,599 so that's why the hype is already gone 1226 00:45:59,359 --> 00:46:04,240 again it will take another some years 1227 00:46:01,599 --> 00:46:06,000 until some new hype comes right 1228 00:46:04,240 --> 00:46:07,680 exactly yeah 1229 00:46:06,000 --> 00:46:09,040 but it's normal 1230 00:46:07,680 --> 00:46:10,720 yeah totally 1231 00:46:09,040 --> 00:46:13,280 what are we going to talk about next 1232 00:46:10,720 --> 00:46:14,560 week tarek 1233 00:46:13,280 --> 00:46:17,040 yeah uh 1234 00:46:14,560 --> 00:46:18,640 maybe you can sum it up uh because we we 1235 00:46:17,040 --> 00:46:21,520 talked about this 1236 00:46:18,640 --> 00:46:24,800 right before this episode um there was 1237 00:46:21,520 --> 00:46:27,520 this option uh the the study in the uk 1238 00:46:24,800 --> 00:46:30,800 where they started uh talking about uh 1239 00:46:27,520 --> 00:46:33,920 introducing a four-day work week and 1240 00:46:30,800 --> 00:46:36,480 this was already kind of reason for some 1241 00:46:33,920 --> 00:46:39,599 interesting discussions um and so this 1242 00:46:36,480 --> 00:46:42,240 will one will be part of next week's 1243 00:46:39,599 --> 00:46:44,319 podcast yeah because i call this woke 1244 00:46:42,240 --> 00:46:46,480 [ __ ] and i don't believe this is a 1245 00:46:44,319 --> 00:46:48,480 good idea and again scientifically 1246 00:46:46,480 --> 00:46:51,280 proven that this will not increase 1247 00:46:48,480 --> 00:46:53,200 productivity because you're just okay 1248 00:46:51,280 --> 00:46:55,760 not let's not spoil it let's talk about 1249 00:46:53,200 --> 00:46:58,000 it next week so 1250 00:46:55,760 --> 00:46:59,520 you can look forward to a hot discussion 1251 00:46:58,000 --> 00:47:02,000 next week 1252 00:46:59,520 --> 00:47:04,800 until then at the same time at the same 1253 00:47:02,000 --> 00:47:09,319 spot see you again next week 1254 00:47:04,800 --> 00:47:09,319 see you again next week bye 1255 00:47:12,960 --> 00:47:15,040 you