1 00:00:00,040 --> 00:00:07,200 Welcome to CXO Talk Episode 864. I'm Michael Krigsman, and today 2 00:00:07,200 --> 00:00:12,480 we're exploring how to navigate truth and deception in AI as 3 00:00:12,480 --> 00:00:16,120 part of a special series on AI ethics. 4 00:00:16,840 --> 00:00:21,080 Our three expert guests are Doctor Anastasia Lauterbach, 5 00:00:21,640 --> 00:00:25,640 Doctor David Bray and Doctor Anthony Scribignano. 6 00:00:25,760 --> 00:00:27,480 We got a lot of doctors in the house. 7 00:00:27,920 --> 00:00:31,000 All of you, welcome to CXO talk. I'm delighted that you're all 8 00:00:31,000 --> 00:00:32,520 here. Thank you very much. 9 00:00:33,520 --> 00:00:36,160 Thank you for having us. Great to be here with. 10 00:00:37,000 --> 00:00:40,560 Anastasia, let's start with you very briefly. 11 00:00:40,560 --> 00:00:44,800 Tell us about your work. I spent most of my time with my 12 00:00:44,800 --> 00:00:46,720 company. I'm a founder and CEO of AI 13 00:00:46,720 --> 00:00:50,720 Entertainment and this company democratizes knowledge of AI 14 00:00:50,920 --> 00:00:55,200 with AI muggles, not Wizards. And this means that I try to 15 00:00:55,200 --> 00:00:58,840 translate AI, robotics and quantum computing from difficult 16 00:00:58,840 --> 00:01:02,320 to easy to understand. For example, I write books like 17 00:01:02,320 --> 00:01:04,560 this. This is Rome and Robbie series, 18 00:01:04,879 --> 00:01:11,040 and this is a real story about real Cat and his friendship with 19 00:01:11,040 --> 00:01:14,520 a robot and everything. What is going on in this book is 20 00:01:14,520 --> 00:01:16,640 mirrored in the science behind section. 21 00:01:16,680 --> 00:01:22,120 Those are over 100 articles explaining AI in very easy 22 00:01:22,120 --> 00:01:25,480 language. I love it a an AI cat. 23 00:01:25,920 --> 00:01:27,400 Absolutely. He's an influence. 24 00:01:27,400 --> 00:01:32,720 And by the way, did you know that 27% of web traffic is a cat 25 00:01:33,320 --> 00:01:36,200 connected? A cat linked content? 26 00:01:36,840 --> 00:01:39,120 Cats are addictive. David, tell us about your work. 27 00:01:39,440 --> 00:01:42,640 I try to bring rationality to the world in terms of how 28 00:01:42,640 --> 00:01:46,440 technology is changing politics, geopolitics, societies and the 29 00:01:46,440 --> 00:01:48,560 like. Do so both as a distinguished 30 00:01:48,560 --> 00:01:50,960 fellow with the Stimson Center as well as business executives 31 00:01:50,960 --> 00:01:53,280 for National security. And then on the side have my own 32 00:01:53,280 --> 00:01:56,640 S Corp which tries to help boards and CE OS do the same. 33 00:01:57,160 --> 00:01:59,240 Anthony Scrifignano, welcome back. 34 00:01:59,240 --> 00:02:01,600 Tell us about your work. I'm a distinguished fellow with 35 00:02:01,600 --> 00:02:03,320 the Stimson Center along with David Bray. 36 00:02:03,800 --> 00:02:05,840 I. My philosophy is to try to teach 37 00:02:05,840 --> 00:02:08,680 and learn every day. No fair cheating on that. 38 00:02:08,680 --> 00:02:12,480 You have to do both. I've done a lot of work over 45 39 00:02:12,480 --> 00:02:16,280 years with data and data science and AI before it was a thing. 40 00:02:16,680 --> 00:02:18,800 I've gotten multiple patents on my name. 41 00:02:18,800 --> 00:02:22,200 Lots of invention around things like identity resolution, 42 00:02:22,200 --> 00:02:25,640 veracity, adjudication, which we'll talk about today, judging 43 00:02:25,640 --> 00:02:29,600 the truthiness of things, and also lots of effort around 44 00:02:29,600 --> 00:02:32,240 finding people that are doing new bad things. 45 00:02:32,240 --> 00:02:35,440 So novel malfeasance. So these are all the things that 46 00:02:35,440 --> 00:02:39,280 are kind of at the edge of computer science and among the 47 00:02:39,280 --> 00:02:41,160 hardest of the hard problems to solve. 48 00:02:41,680 --> 00:02:48,560 Anastasia, Why is building trustworthy AI versus deceptive 49 00:02:48,760 --> 00:02:52,280 AI important? Trust is something tremendously 50 00:02:52,280 --> 00:02:55,240 important for human psychologist, one of those 51 00:02:55,320 --> 00:02:59,840 constructs which is paramount for human relationships. 52 00:02:59,960 --> 00:03:04,840 So no one is going to disclose or I am building something which 53 00:03:05,040 --> 00:03:08,720 is based on deception. And they are two economical 54 00:03:08,720 --> 00:03:13,440 arguments in favour of paying attention to whether AI is 55 00:03:13,440 --> 00:03:18,160 considered trustworthy or not. And the first, I would open up 56 00:03:18,160 --> 00:03:21,640 with a quote by Kevin Kelly, who is founder of Wiret. 57 00:03:21,640 --> 00:03:26,160 He said that the business of the next 10,000 companies is very 58 00:03:26,160 --> 00:03:29,680 easy to predict. You take X and you add AI to it. 59 00:03:30,160 --> 00:03:34,840 And obviously this remark which was made I think like around 60 00:03:34,840 --> 00:03:42,160 2016, 2017, he is now based in figures, in financial figures. 61 00:03:42,400 --> 00:03:48,960 For example, around 2013, we expect the EI market to grow to 62 00:03:49,000 --> 00:03:53,720 1.8, see 5 trillion U.S. dollars globally. 63 00:03:54,000 --> 00:03:55,720 And I mean this is tremendous growth. 64 00:03:55,720 --> 00:04:00,960 We are talking about compound annual growth rate of 33% from 65 00:04:00,960 --> 00:04:04,880 22 to 2030. And there's another caveat to 66 00:04:04,880 --> 00:04:07,480 that. When I looked into valuation of 67 00:04:07,480 --> 00:04:13,160 businesses around 2012, I could see that the portion of digital 68 00:04:13,160 --> 00:04:17,920 assets in this valuation of companies was around 12 to 13%. 69 00:04:18,160 --> 00:04:25,520 And after the COVID epidemics, we are talking about 90%, so 90. 70 00:04:25,880 --> 00:04:29,840 And digital assets, it's obviously all about data, it's 71 00:04:29,840 --> 00:04:33,440 about intellectual property. It's all kind of systems and 72 00:04:33,440 --> 00:04:36,600 processes to do something with this data. 73 00:04:36,600 --> 00:04:40,440 So this incredible growth has something to do with the 74 00:04:40,440 --> 00:04:43,520 progress in AI and we have to pay attention. 75 00:04:44,000 --> 00:04:49,000 If I were to summarize what Anas Tasha just said, everybody's 76 00:04:49,000 --> 00:04:51,520 leaning into this and everybody's leaning into this, 77 00:04:51,640 --> 00:04:54,640 but it's a different this. So if you look at what a lot of 78 00:04:54,640 --> 00:04:57,160 organizations want is they want to be able to check the box and 79 00:04:57,160 --> 00:04:59,200 say we're doing AI, we're doing Gen. 80 00:04:59,200 --> 00:05:00,800 AI, we're doing large language models. 81 00:05:00,800 --> 00:05:02,960 Look at us, we're great, right? That's great. 82 00:05:02,960 --> 00:05:05,840 What's the problem you're solving and by what right do you 83 00:05:05,840 --> 00:05:07,800 think that that thing that you're doing is going to help 84 00:05:07,800 --> 00:05:11,720 address that problem? There's in many cases a great 85 00:05:11,720 --> 00:05:15,360 argument to be made that you're just accelerating the speed with 86 00:05:15,360 --> 00:05:18,360 which you hit the wall. You're helping people understand 87 00:05:18,360 --> 00:05:20,560 what you don't know. You're helping people understand 88 00:05:20,560 --> 00:05:23,120 more about your questions than about your answers. 89 00:05:23,120 --> 00:05:26,400 You're exposing yourself to potentially looking to your 90 00:05:26,400 --> 00:05:30,520 customers as if, you know you focused on that and not this 91 00:05:30,520 --> 00:05:32,120 other thing that they want you to focus on. 92 00:05:32,120 --> 00:05:35,040 So it's, yeah, it's really important to move into this 93 00:05:35,040 --> 00:05:38,600 technology. Never would I say don't do that, 94 00:05:38,880 --> 00:05:41,800 but I would say, you know, just because you see a new tool at 95 00:05:41,800 --> 00:05:43,720 Home Depot doesn't mean you go buy it. 96 00:05:43,720 --> 00:05:46,640 You have to have something to address and you have to have a 97 00:05:46,640 --> 00:05:48,080 problem that you're trying to go after. 98 00:05:48,280 --> 00:05:50,840 And you also have to understand the unintended impact of those 99 00:05:50,840 --> 00:05:55,320 actions you're about to take. May I just quote a poet from 100 00:05:55,320 --> 00:05:58,560 England from the early 18th century, Alexander Pope? 101 00:05:58,560 --> 00:06:01,640 He said trust not yourself, but your defects to know. 102 00:06:02,000 --> 00:06:04,840 And this is actually emphasizing that it's not just 103 00:06:05,520 --> 00:06:08,080 self-reliance, but it's about humidity. 104 00:06:08,080 --> 00:06:11,040 And I think this humidity and thoughtfulness is something 105 00:06:11,040 --> 00:06:14,640 which we need to learn more and more and apply more and more 106 00:06:14,640 --> 00:06:16,600 when we move into the world with AI. 107 00:06:17,080 --> 00:06:20,280 One of the questions I always ask when when someone is, you 108 00:06:20,280 --> 00:06:24,200 know, running headlong into, you know, let's AI our way out of 109 00:06:24,200 --> 00:06:28,240 this problem is, you know, two questions actually, 1 is what do 110 00:06:28,240 --> 00:06:30,760 we have to believe in order to go down that road? 111 00:06:31,400 --> 00:06:34,200 And the second thing is, but by what right do you think the 112 00:06:34,200 --> 00:06:38,000 answer is contained in what the AI is going to look at to answer 113 00:06:38,000 --> 00:06:41,120 that question? Imagine during COVID if you had 114 00:06:41,120 --> 00:06:44,640 ChatGPT and you said, you know, what are the most promising 115 00:06:45,360 --> 00:06:47,640 approaches to responding to COVID? 116 00:06:47,640 --> 00:06:50,520 Well, we didn't know, right? And so you would, you would 117 00:06:50,680 --> 00:06:53,600 literally get what we love to call hallucination right now, 118 00:06:53,600 --> 00:06:56,960 but on steroids, right? Because the LLM doesn't know 119 00:06:56,960 --> 00:06:59,720 that the answer isn't in the corpus of data that it's looking 120 00:06:59,720 --> 00:07:01,040 at. And it's going to give you an 121 00:07:01,040 --> 00:07:02,680 answer anyway. And it's going to look awfully 122 00:07:02,680 --> 00:07:03,840 good. And it's going to really look 123 00:07:03,840 --> 00:07:07,280 like a human being wrote it. And it's going to be wonderfully 124 00:07:07,280 --> 00:07:11,160 articulate, but just as good as any other rumor you just heard. 125 00:07:11,440 --> 00:07:16,600 So can we say that the summary here to make this very 126 00:07:16,920 --> 00:07:24,000 simplistic is AI is around us, AI is surrounding us, AI will 127 00:07:24,240 --> 00:07:28,320 continue to do so and therefore we need to know what's true and 128 00:07:28,320 --> 00:07:29,600 not true. I mean, that seems pretty 129 00:07:29,600 --> 00:07:32,520 simple. Perhaps instead of calling, when 130 00:07:32,520 --> 00:07:36,280 we say AI, artificial intelligence, we should call it 131 00:07:36,360 --> 00:07:41,840 alien interactions because these machines are not at all thinking 132 00:07:41,840 --> 00:07:44,560 like you or I. And that's a very real danger 133 00:07:44,560 --> 00:07:47,840 when you read articles or for whatever reason begin to 134 00:07:47,840 --> 00:07:51,120 anthropomorphize these machines. I mean, just statistically, I 135 00:07:51,120 --> 00:07:54,480 mean, the human brain uses between 15 and 20 watts to do 136 00:07:54,480 --> 00:07:56,640 everything it does. These models are looking at 137 00:07:56,640 --> 00:07:59,000 upwards of 2005 thousand or more. 138 00:07:59,000 --> 00:08:00,920 I mean, when we hear like companies thinking about doing 139 00:08:00,920 --> 00:08:05,200 nuclear power plants to do their alien interactions as AI, that 140 00:08:05,200 --> 00:08:08,640 might tell you that this is simply been designed to look 141 00:08:08,640 --> 00:08:11,200 like it's thinking like a human, when in fact it's not. 142 00:08:11,600 --> 00:08:13,720 That then also gets to the deeper question I'd like to say, 143 00:08:13,720 --> 00:08:17,200 Michael, which is if you remember the Turing test, the 144 00:08:17,200 --> 00:08:21,320 Turing test was intentionally trying to have a machine deceive 145 00:08:21,560 --> 00:08:23,880 a human into thinking it was human. 146 00:08:24,840 --> 00:08:27,080 Guess what? We have been wildly successful 147 00:08:27,560 --> 00:08:30,440 at rolling out technologies that will now convince you that the 148 00:08:30,520 --> 00:08:32,960 responses are giving whether they're in text, whether than 149 00:08:32,960 --> 00:08:36,520 audio, whether in video appear to be human like. 150 00:08:36,919 --> 00:08:40,720 But that's a problem because as Anthony was mentioning in his 151 00:08:40,720 --> 00:08:43,960 background, I mean, fraudsters, scamsters, they would love to 152 00:08:43,960 --> 00:08:46,080 use this technology. And we've all probably seen in 153 00:08:46,080 --> 00:08:49,880 the last two years a step change in scam phishing emails where 154 00:08:49,880 --> 00:08:51,520 they are now at the point where you can no longer use 155 00:08:51,520 --> 00:08:54,240 misspellings as they tell if it's a phishing e-mail or not. 156 00:08:54,520 --> 00:08:57,480 And now even better, they're often referencing another family 157 00:08:57,480 --> 00:08:59,960 member or friend because they've combed our social media network. 158 00:09:00,240 --> 00:09:03,840 And so the are actually intentionally, unfortunately 159 00:09:04,120 --> 00:09:05,800 really successful, the Turing test. 160 00:09:06,360 --> 00:09:08,640 But the whole Turing test was again a machine trying to 161 00:09:08,640 --> 00:09:11,440 convince you it was human. That might not be the best tool 162 00:09:11,440 --> 00:09:13,160 exactly. We've been saying here for 163 00:09:13,160 --> 00:09:16,680 society writ large. There's a another little nuanced 164 00:09:16,680 --> 00:09:18,960 thing happening right now where we now have to prove to the 165 00:09:18,960 --> 00:09:20,480 machine that we're a human right. 166 00:09:20,480 --> 00:09:22,440 So it's sort of the opposite of that. 167 00:09:22,720 --> 00:09:25,600 And if you look at CAPTCHA and the way it was originally 168 00:09:25,600 --> 00:09:28,880 designed, it was just OCR was kind of not very good. 169 00:09:28,880 --> 00:09:32,400 So if I tilt the letters and I put them together, then OCR 170 00:09:32,400 --> 00:09:34,880 doesn't work and the human can read it and we know the answer 171 00:09:34,880 --> 00:09:37,440 and that's great, right? Well, then people started to use 172 00:09:37,440 --> 00:09:42,000 the, the responses from CAPTCHA to train better OCR. 173 00:09:42,000 --> 00:09:44,840 And we'll, lo and behold, all of a sudden the computer gets just 174 00:09:44,840 --> 00:09:47,200 as good as we are at reading kind of crappy text. 175 00:09:47,560 --> 00:09:49,360 So then they start introducing pictures. 176 00:09:49,680 --> 00:09:54,600 And introducing pictures is not a, it's not an insert, it's not 177 00:09:54,600 --> 00:09:57,080 an end to that. It's just a way of kicking that 178 00:09:57,080 --> 00:09:58,760 ball a little bit further down the road. 179 00:09:59,000 --> 00:10:01,680 I was listening to a talk recently with one of the people 180 00:10:01,680 --> 00:10:05,080 who actually invented the technology, and he was saying #1 181 00:10:05,080 --> 00:10:08,400 he wished he never did. And #2 you know, maybe in the 182 00:10:08,400 --> 00:10:12,200 future, you'll have to do things like walk three steps with your 183 00:10:12,200 --> 00:10:15,760 phone or jiggle your phone or tilt your phone down or do 184 00:10:15,760 --> 00:10:18,080 something. And, and now I can just imagine 185 00:10:18,080 --> 00:10:21,320 us OK, do 10 push ups. Like, at what point do I have to 186 00:10:21,320 --> 00:10:24,360 stop proving that I'm human? We don't have a good answer to 187 00:10:24,360 --> 00:10:27,320 this right now. It's not because the AI is 188 00:10:27,320 --> 00:10:29,000 thinking. It's because it's very good at 189 00:10:29,000 --> 00:10:31,400 watching us behave. That's so true. 190 00:10:31,400 --> 00:10:34,400 And actually we need to establish a language to talk 191 00:10:34,400 --> 00:10:39,000 about AIS and machines, and we're still using human terms to 192 00:10:39,000 --> 00:10:42,560 describe what we expect. And in social science, for 193 00:10:42,560 --> 00:10:45,680 example, tries, trust is described as the belief that 194 00:10:45,680 --> 00:10:48,280 another person will do what is expected. 195 00:10:48,480 --> 00:10:52,280 And now it's not a person, it's a machine or piece of code and 196 00:10:52,280 --> 00:10:55,680 what is really expected. If I use my vacuum cleaner, I 197 00:10:55,680 --> 00:11:01,000 expect it to do a certain task. But this is very different from 198 00:11:01,000 --> 00:11:03,960 LLM. And obviously because I am an AI 199 00:11:03,960 --> 00:11:07,800 and I teach AI and cybersecurity at the university, I know it's 200 00:11:07,880 --> 00:11:11,520 has nothing to do with language. It's a statistical stream of 201 00:11:11,520 --> 00:11:14,160 tokens and token isn't even a word. 202 00:11:14,360 --> 00:11:17,760 But if we take, for example English, it's maybe 2/3 of an 203 00:11:17,760 --> 00:11:20,360 English word. Yeah, statistically. 204 00:11:20,960 --> 00:11:25,280 But you know, humans expect that it understands and it's simply 205 00:11:25,280 --> 00:11:28,760 doesn't, and even not based on the world model. 206 00:11:29,280 --> 00:11:31,480 A couple of things to sort of add to the nuances. 207 00:11:31,480 --> 00:11:34,680 First, let's go. Let's take the 30,000 foot view 208 00:11:34,680 --> 00:11:38,240 before we dive in, which is back in the 1920s when this 209 00:11:38,240 --> 00:11:41,000 disruptive technology called radio came out, there were 210 00:11:41,000 --> 00:11:43,840 pundits that were saying something similar to what we're 211 00:11:43,840 --> 00:11:46,200 seeing with AI nowadays, that this was going to cause World 212 00:11:46,200 --> 00:11:47,040 Peace. It was going to have 213 00:11:47,040 --> 00:11:48,800 understandings. World leaders would talk to each 214 00:11:48,800 --> 00:11:51,080 other on a really regular basis on radio. 215 00:11:51,560 --> 00:11:54,480 And then fast words in 1933, about six or seven years later, 216 00:11:54,640 --> 00:11:56,640 those same pundits were saying that this was going to be the 217 00:11:56,640 --> 00:11:58,640 end of society as we know it. It's going to be the dictator's 218 00:11:58,640 --> 00:12:00,800 tool kit. And so I raise that because 219 00:12:00,800 --> 00:12:03,200 right now there are a lot of breathless articles that are 220 00:12:03,200 --> 00:12:05,680 either saying that AI is going to save us all and the, you 221 00:12:05,720 --> 00:12:09,000 know, the good times are upon us or is somehow going to be in the 222 00:12:09,000 --> 00:12:11,000 societies. We're probably doing the same 223 00:12:11,000 --> 00:12:13,920 thing to the technology that we did to radio and we're missing 224 00:12:13,920 --> 00:12:15,600 that. Again, it's just a tool. 225 00:12:15,880 --> 00:12:19,120 Now, the second thing I would say is I think we are often 226 00:12:19,120 --> 00:12:23,040 using AI as a stalking horse or as scapegoat for things that are 227 00:12:23,040 --> 00:12:26,000 deeper in society. Like we ask questions like, how 228 00:12:26,000 --> 00:12:28,320 do you know if AI is ethical? Well, how do I know if an 229 00:12:28,320 --> 00:12:30,960 organization is ethical? How do I know if ethicalization 230 00:12:30,960 --> 00:12:33,160 is not bias? How do I know if an organization 231 00:12:33,160 --> 00:12:36,040 is not making bad decisions or hallucinating? 232 00:12:36,320 --> 00:12:39,520 And So what this may point to is we need to do a better job of 233 00:12:39,520 --> 00:12:41,800 figuring that out and how we actually do it, whether it's a 234 00:12:41,800 --> 00:12:43,560 machine, organization, or person. 235 00:12:44,040 --> 00:12:46,280 Now I will give a slightly different definition of trust 236 00:12:46,640 --> 00:12:49,120 from the background I come from, which is the willingness to be 237 00:12:49,120 --> 00:12:51,680 vulnerable to the actions of an actor you cannot directly 238 00:12:51,680 --> 00:12:54,200 control. And that actor could be an 239 00:12:54,200 --> 00:12:55,880 individual, could be an organization, could be a 240 00:12:55,880 --> 00:12:59,120 government, could be a machine. And that what it's been shown is 241 00:12:59,120 --> 00:13:02,280 we humans are willing to be vulnerable to those actions of 242 00:13:02,280 --> 00:13:05,400 an actor that we cannot control if we perceive benevolence, 243 00:13:05,400 --> 00:13:07,320 perceived competence, and perceived integrity. 244 00:13:07,560 --> 00:13:10,880 So exactly down to Sasha's point and Anthony's point, there is no 245 00:13:10,880 --> 00:13:13,520 way of assessing the benevolence of a LOM. 246 00:13:13,960 --> 00:13:15,360 There is no way to assess the competence. 247 00:13:15,360 --> 00:13:17,760 I mean, these things are just doing very fancy pattern 248 00:13:17,760 --> 00:13:19,240 matching. And in fact, if it's not in the 249 00:13:19,240 --> 00:13:22,440 training data, they will give you something that is made-up, 250 00:13:23,000 --> 00:13:25,560 but then also an integrity. You say to them that's not 251 00:13:25,560 --> 00:13:28,160 right, They will be again, the the prompts will be very 252 00:13:28,160 --> 00:13:30,360 cheerful in saying you're right. And then when you say that's not 253 00:13:30,360 --> 00:13:31,880 right too, they're right, you're right again. 254 00:13:31,880 --> 00:13:34,600 And so integrity, competence and benevolent aren't there. 255 00:13:35,080 --> 00:13:37,520 But let's step back and say, how do people assess that? 256 00:13:37,600 --> 00:13:40,640 Given the fact that we're now connected digitally, How do we 257 00:13:40,640 --> 00:13:42,720 assess that? CXO talk is benevolent competent 258 00:13:42,720 --> 00:13:44,400 integrity. How do we select, you know, do 259 00:13:44,400 --> 00:13:46,680 that for governments? How do we do that for the world 260 00:13:46,680 --> 00:13:49,280 that is large? And maybe this is the challenge 261 00:13:49,280 --> 00:13:52,280 hitting societies, free societies in particular, is we 262 00:13:52,280 --> 00:13:54,320 lack the ability to adjudicate at speed. 263 00:13:54,840 --> 00:13:57,400 A lot of things, the terms that you're using, David, are also 264 00:13:57,400 --> 00:14:01,920 epistemologically fuzzy, right? So benevolent might be 265 00:14:01,960 --> 00:14:05,080 benevolent from your perspective and not so benevolent from my 266 00:14:05,080 --> 00:14:08,360 perspective. And we tend to use terms you 267 00:14:08,360 --> 00:14:10,760 didn't, but we tend to use terms like we're good. 268 00:14:10,840 --> 00:14:16,640 Well, good for whom, right. And you know, I, I have a, a lot 269 00:14:16,640 --> 00:14:18,560 of background in emergency medicine, right? 270 00:14:18,560 --> 00:14:21,040 So one of the things they always tell you is expand the circle. 271 00:14:21,320 --> 00:14:23,400 You know, if you're going to make a complicated decision, 272 00:14:23,400 --> 00:14:26,720 get, get advice from other people and make sure that, you 273 00:14:26,720 --> 00:14:28,720 know, those other people will bring in different perspectives. 274 00:14:28,720 --> 00:14:31,400 Well, that's great until something's on fire or until 275 00:14:31,400 --> 00:14:34,320 somebody's not breathing, and then somebody has to make a 276 00:14:34,320 --> 00:14:37,160 decision, and it might be the wrong one, but you have to make 277 00:14:37,160 --> 00:14:38,880 that decision with some degree of cut. 278 00:14:38,880 --> 00:14:43,200 Now, after the fact, everybody will swoop in and judge you that 279 00:14:43,200 --> 00:14:44,880 you should have done this, you should have done that. 280 00:14:44,880 --> 00:14:47,360 Why didn't you do this? How didn't you know about this? 281 00:14:47,880 --> 00:14:51,960 At some point, AI is going to be making decisions for humans 282 00:14:51,960 --> 00:14:55,400 because the argument will be that AI can make a decision 283 00:14:55,400 --> 00:14:58,400 before the human decision becomes irrelevant, right? 284 00:14:58,680 --> 00:15:01,760 Do I cauterize this vessel or that vessel make a decision 285 00:15:01,760 --> 00:15:03,160 right now? I don't. 286 00:15:03,160 --> 00:15:06,520 Do you know all the vasculature in the brain and can you let the 287 00:15:06,600 --> 00:15:08,800 AI do it right? I don't know. 288 00:15:09,000 --> 00:15:11,480 But I could see how we can get there and I could certainly see 289 00:15:11,480 --> 00:15:14,520 how we can get there in in a nation state, military 290 00:15:14,520 --> 00:15:18,200 industrial kind of context. I could we can get there in, you 291 00:15:18,200 --> 00:15:20,360 know, in space. There's, you know, it takes so 292 00:15:20,360 --> 00:15:22,840 long for a signal to get from one place to another that you 293 00:15:22,840 --> 00:15:26,000 need to have autonomy in order for the thing to be able to land 294 00:15:26,000 --> 00:15:27,080 or do whatever it's trying to do. 295 00:15:27,600 --> 00:15:30,680 But as we give up this agency, this concept of agency, of 296 00:15:30,680 --> 00:15:33,960 giving the right to the machine to make the decision on the 297 00:15:33,960 --> 00:15:38,880 behalf of the human, the values of the human have to be imposed 298 00:15:38,880 --> 00:15:41,400 into that what is just a bunch of code. 299 00:15:41,920 --> 00:15:44,280 And we are not rules based creatures. 300 00:15:44,280 --> 00:15:48,280 We are very, very squishy about what we do and we make decisions 301 00:15:48,280 --> 00:15:51,880 and we change our opinion based on new facts and new modes and 302 00:15:51,880 --> 00:15:53,760 new ways of learning all the time. 303 00:15:53,880 --> 00:15:55,720 AI is horrible at that right now. 304 00:15:56,080 --> 00:16:00,960 Please subscribe to our newsletter and subscribe to our 305 00:16:00,960 --> 00:16:03,360 YouTube channel. Check out cxotalk.com. 306 00:16:04,800 --> 00:16:06,040 We have great shows. Coming up. 307 00:16:06,160 --> 00:16:13,040 We have a question from Twitter and this is from Arsalan Khan 308 00:16:13,040 --> 00:16:18,040 who says we know that AI can be trustworthy or not trustworthy 309 00:16:18,560 --> 00:16:22,280 based on many factors. How can normal non-technical 310 00:16:22,280 --> 00:16:27,440 people know if AI is affecting their lives, if it is ethical? 311 00:16:27,720 --> 00:16:33,320 And is there an opt out for AI? And we know very often there's 312 00:16:33,320 --> 00:16:37,400 not. And then let's also talk about 313 00:16:37,400 --> 00:16:41,920 the role of software companies. AI is surrounding us. 314 00:16:41,920 --> 00:16:45,720 So if if you have a smartphone, you immerse in AI. 315 00:16:46,240 --> 00:16:49,160 So we can really discuss all those technologies like, you 316 00:16:49,160 --> 00:16:54,280 know, LLMS are on smartphones, but even how we see our calendar 317 00:16:54,280 --> 00:16:58,160 invites and how it correlates with tweets or whatever, some 318 00:16:58,160 --> 00:17:01,680 notifications from LinkedIn, it's immersed in AI. 319 00:17:01,920 --> 00:17:06,280 Some navigation technologies are connected to automation and to 320 00:17:06,280 --> 00:17:09,240 AI. And we just need to think how 321 00:17:09,240 --> 00:17:14,520 should we behave ourself. And frankly, unfortunately, and 322 00:17:14,520 --> 00:17:17,599 this is I guess the difference to radio technologies and 323 00:17:17,599 --> 00:17:23,400 everything which proceeded, we must expose ourselves and learn. 324 00:17:23,800 --> 00:17:27,640 And I think that EI literacy and technology literacy is something 325 00:17:27,640 --> 00:17:30,400 which is a must from an early age. 326 00:17:30,640 --> 00:17:32,760 It's like basic financial literacy. 327 00:17:32,960 --> 00:17:36,360 Obviously, there will be always some treasury experts and 328 00:17:36,360 --> 00:17:39,040 taxation experts and whatever some valuation experts. 329 00:17:39,360 --> 00:17:44,920 But everyone must understand that this revenue and this cost 330 00:17:45,080 --> 00:17:51,560 and 1 minus another produce some kind of, and I kind of love 331 00:17:51,560 --> 00:17:55,600 comparing AI and you know, dealing with AI with the 332 00:17:55,600 --> 00:17:57,720 kitchen, right? Maybe because I'm a woman. 333 00:17:58,200 --> 00:18:02,960 And obviously you might be a great chef even if you don't 334 00:18:02,960 --> 00:18:07,000 work in high cuisine. And you can produce fantastic 335 00:18:07,000 --> 00:18:10,320 stuff and be very great and performance some shows. 336 00:18:10,880 --> 00:18:13,160 And probably this is the ultimate level of 337 00:18:13,160 --> 00:18:16,080 sophistication. But everyone knows how to fry an 338 00:18:16,080 --> 00:18:20,920 egg, and I guess with the eye we must learn how to fry an egg. 339 00:18:20,920 --> 00:18:24,440 And this is basically why I learn hits and their clever 340 00:18:24,440 --> 00:18:26,920 appearance to think about those concepts. 341 00:18:27,080 --> 00:18:30,360 What is an LLM? What does it do? 342 00:18:30,680 --> 00:18:33,360 Who is Wolfram? Why should we know about him a 343 00:18:33,360 --> 00:18:36,400 little bit? What does it mean to have a 344 00:18:36,400 --> 00:18:40,880 robot in our house? What is the difference between a 345 00:18:40,880 --> 00:18:43,200 robot and let's say, a vacuum cleaner? 346 00:18:43,200 --> 00:18:46,440 Right? So I think we just must embrace 347 00:18:47,080 --> 00:18:52,520 this knowledge and actually evolve as communities to talk 348 00:18:52,520 --> 00:18:56,120 about it, to ask difficult questions, and to understand 349 00:18:56,280 --> 00:19:00,120 that there is a purpose behind every single application and 350 00:19:00,120 --> 00:19:02,800 service. You can't go through a jungle. 351 00:19:03,520 --> 00:19:09,320 You must learn and then you will have some openness to maybe be 352 00:19:09,320 --> 00:19:13,400 more precise on what you need. And finally, I love the quote of 353 00:19:13,400 --> 00:19:16,120 Pablo Picasso, he said. Machines are quite stupid. 354 00:19:16,120 --> 00:19:20,240 They just produce answers. And I think this is up to humans 355 00:19:20,240 --> 00:19:25,080 to ask questions, and we need to teach humans how to ask those 356 00:19:25,080 --> 00:19:27,960 questions. And this is basically a fabulous 357 00:19:27,960 --> 00:19:32,280 opportunity that is opening up in front of all of us to learn 358 00:19:32,280 --> 00:19:37,440 more about humanity and what it really means in the world, which 359 00:19:37,440 --> 00:19:43,000 is partly dominated by AIS. In Western society, if you apply 360 00:19:43,000 --> 00:19:47,520 and your job application has a non western name, you are less 361 00:19:47,520 --> 00:19:49,760 likely. A western name application is 362 00:19:49,760 --> 00:19:51,680 three times more likely to be selected. 363 00:19:51,680 --> 00:19:53,840 And that's not an AI problem whatsoever. 364 00:19:53,840 --> 00:19:55,560 Everything we've been talking about here is like, oh, how do 365 00:19:55,560 --> 00:19:57,800 we trust the machine? I'm like, no, no, no, no, we 366 00:19:57,800 --> 00:20:01,280 have deeper systemic issues. And so almost it's like we're 367 00:20:01,280 --> 00:20:04,200 almost pitching AI and saying, how do we ask these questions of 368 00:20:04,200 --> 00:20:06,960 AI before we've even paused and say, how do we actually do 369 00:20:06,960 --> 00:20:10,440 better as humans? And I really appreciate it. 370 00:20:10,440 --> 00:20:13,800 So I would also say that as we go forward from this. 371 00:20:14,280 --> 00:20:16,600 It's also thinking about that we've had these challenges 372 00:20:16,600 --> 00:20:18,800 before. I mean, most of us are not 373 00:20:18,800 --> 00:20:21,760 medical doctors yet. We have to go see a medical 374 00:20:21,760 --> 00:20:23,240 doctor. How do you know when the medical 375 00:20:23,240 --> 00:20:26,840 doctor is going to prescribe a approach that they really have 376 00:20:26,840 --> 00:20:29,760 your best interest in mind? And the way we saw this in the 377 00:20:29,760 --> 00:20:32,240 past, and unfortunately, the Internet kind of squashed this, 378 00:20:32,520 --> 00:20:35,040 was we did have professional societies that require their 379 00:20:35,040 --> 00:20:37,760 members to have knowledge of something, experience in 380 00:20:37,760 --> 00:20:40,160 something, and then pledge to an unethical oath. 381 00:20:40,480 --> 00:20:44,560 And if any point in time, the member had a concern because 382 00:20:44,560 --> 00:20:47,000 again, most people cannot adjudicate, did that doctor act 383 00:20:47,000 --> 00:20:49,720 in your best interest or not? It was other medical doctors 384 00:20:49,720 --> 00:20:51,880 that would decide to either say, yes, it did the right thing or 385 00:20:51,880 --> 00:20:54,040 not, we're going to censor them or going to move their license. 386 00:20:54,280 --> 00:20:56,280 Well, unfortunately, what happened with the Internet is it 387 00:20:56,280 --> 00:20:58,600 made it so that everybody could be armchair quarterbacks. 388 00:20:59,120 --> 00:21:01,640 But I raised that because I worry that we will become so 389 00:21:01,640 --> 00:21:04,000 fixated on we must understand AI, we must understand and 390 00:21:04,000 --> 00:21:07,200 educate and everything like that without pausing and saying right 391 00:21:07,200 --> 00:21:10,520 now there are plenty of organizations where there is 392 00:21:10,520 --> 00:21:13,440 issues internally to bad decision making, bias decision 393 00:21:13,440 --> 00:21:16,240 making as well as externally and how they're actually interacting 394 00:21:16,240 --> 00:21:19,160 with employees or customers. And so I think we actually need 395 00:21:19,160 --> 00:21:22,000 to figure out a deeper approach, not just about the machine. 396 00:21:22,400 --> 00:21:30,160 Implicit in what you have all been saying is this idea that 397 00:21:30,560 --> 00:21:36,880 benevolence is our fundamental goal and we rely on the 398 00:21:36,880 --> 00:21:39,960 benevolent human spirit. And we hope that the Anthony's 399 00:21:39,960 --> 00:21:41,840 shaking his head. Well, this is my interpretation. 400 00:21:42,280 --> 00:21:45,000 So let me ask my question anyways. 401 00:21:45,440 --> 00:21:49,160 Before it's even edit, before it's edited, before it comes out 402 00:21:49,160 --> 00:21:52,720 of my mouth. So Anthony's trying to edit it 403 00:21:52,720 --> 00:21:55,040 in my brain. So my question. 404 00:21:55,040 --> 00:21:57,920 So my question is this. If we think about the software 405 00:21:57,920 --> 00:22:03,560 companies, over my career, I've consulted with over 100 software 406 00:22:03,560 --> 00:22:08,680 companies literally. And as far as I can see, the 407 00:22:08,680 --> 00:22:12,600 goal of software companies is innovation with the goal of 408 00:22:12,600 --> 00:22:15,360 making money. And yes, software companies talk 409 00:22:15,360 --> 00:22:17,200 about we want to change the world. 410 00:22:17,200 --> 00:22:19,920 And I'm sure that there's that. That's also true. 411 00:22:20,120 --> 00:22:22,840 But the bottom line is it's about money. 412 00:22:22,840 --> 00:22:28,600 And so when we talk about trustworthy AI versus deceptive 413 00:22:28,800 --> 00:22:34,000 AI, how do we overlay this software company issue on top of 414 00:22:34,000 --> 00:22:35,520 it? Anastasia thoughts on this? 415 00:22:35,880 --> 00:22:39,000 30 years ago I was working at the Munich reinsurance company 416 00:22:39,000 --> 00:22:43,520 as a liability on the right and even now there's no not such a 417 00:22:43,520 --> 00:22:47,880 thing as a software liability within the product liability 418 00:22:47,880 --> 00:22:51,280 category. So European Union has updated 419 00:22:51,520 --> 00:22:55,680 product liability directive, but the rules will be introduced 420 00:22:55,680 --> 00:23:01,160 only in December 2000 26 and in the United States, to my 421 00:23:01,160 --> 00:23:04,520 knowledge, please correct me if I'm wrong, so far they aren't 422 00:23:04,760 --> 00:23:08,240 any kind of, you know, rules which are saying that software 423 00:23:08,240 --> 00:23:14,960 vendors are within this liability rules what what is 424 00:23:14,960 --> 00:23:18,400 applicable to for example, automotive companies or consumer 425 00:23:18,400 --> 00:23:22,480 electronic and whatever. So we need to somehow calibrate 426 00:23:22,480 --> 00:23:26,200 what we are talking about. For example, in July this year, 427 00:23:26,240 --> 00:23:32,400 there was a bug which was actually played into a software 428 00:23:32,680 --> 00:23:36,720 over an overnight update with a cybersecurity vendor. 429 00:23:37,000 --> 00:23:42,160 And then all Windows machines went down at the airports and I 430 00:23:42,160 --> 00:23:46,520 think around 5000 flights were either cancelled or delayed 431 00:23:47,040 --> 00:23:49,840 globally. And I think in the US the figure 432 00:23:49,840 --> 00:23:55,880 of financial damage which is known today is for 5.4 billion 433 00:23:56,200 --> 00:23:59,480 U.S. dollars. So damage from this update, so 434 00:23:59,480 --> 00:24:03,840 there is some legal process which is going on, but so far 435 00:24:04,400 --> 00:24:07,360 actually that there will be some damages paid. 436 00:24:07,360 --> 00:24:11,480 So this prospect is really, really low because it's very 437 00:24:11,480 --> 00:24:15,360 difficult to determine what is the negligence or misconduct 438 00:24:15,360 --> 00:24:18,480 here. Then tort law actually excludes 439 00:24:18,480 --> 00:24:21,040 a software from this liability things. 440 00:24:21,320 --> 00:24:25,680 And then obviously, if there are multiple parties involved, who 441 00:24:25,680 --> 00:24:29,000 did what and how, this is really, really complex. 442 00:24:29,360 --> 00:24:32,680 And if we are, for example, with food industry or Pharmaceutical 443 00:24:32,680 --> 00:24:35,200 industry, the world looks completely different. 444 00:24:35,200 --> 00:24:38,480 So the world is being eaten by software. 445 00:24:38,480 --> 00:24:41,600 So Marc Andreessen said it in August 2011. 446 00:24:41,760 --> 00:24:45,720 But there is no product liability for software vendors. 447 00:24:46,080 --> 00:24:48,640 And now we are going into AI. Yeah. 448 00:24:48,640 --> 00:24:52,440 So fantastic. So European Union is 449 00:24:52,440 --> 00:24:57,560 implementing the European AI Act, so all vendors must comply. 450 00:24:57,840 --> 00:25:01,080 It's really killed AI ecosystem in Europe. 451 00:25:01,080 --> 00:25:04,360 By the way, I, I represent myself, I don't represent any 452 00:25:04,360 --> 00:25:07,320 big grant. So I was really against the 453 00:25:07,320 --> 00:25:12,640 European AI Act because it did not solve one single risk in AI. 454 00:25:13,000 --> 00:25:19,400 It impose huge amount of costs on AI companies and investors. 455 00:25:19,400 --> 00:25:23,200 Venture capitalists are saying we are really our fingers off 456 00:25:23,200 --> 00:25:28,080 the European ecosystem because it's all too costly and we need 457 00:25:28,080 --> 00:25:32,200 now to balance what do we want. At the same time, 5 European 458 00:25:32,200 --> 00:25:35,920 countries, which is Italy, Belgium, Austria, Germany and 459 00:25:35,920 --> 00:25:39,760 Netherlands are moving into mass retirement this decade. 460 00:25:40,040 --> 00:25:43,720 And if you have something like that, you must think about 461 00:25:43,800 --> 00:25:47,000 automation and AI. And there's no policy, there's 462 00:25:47,000 --> 00:25:51,760 no thinking how to actually balance the inevitably declining 463 00:25:51,760 --> 00:25:56,760 GDP due to this mass retirement. And we have the European AI Act, 464 00:25:56,760 --> 00:25:59,880 so Europe shoot itself into the knee. 465 00:26:00,080 --> 00:26:04,920 This is my interpretation with all the caveats and that this is 466 00:26:04,920 --> 00:26:08,000 complicated, that they are issues, that thoughtfulness is 467 00:26:08,000 --> 00:26:13,040 needed, and all of the above. I'm in violent agreement with 468 00:26:13,320 --> 00:26:16,800 your position on the dangers of regulations. 469 00:26:16,800 --> 00:26:19,720 Sort of. You know, I'm a scuba diver. 470 00:26:19,760 --> 00:26:21,400 The regulator's pretty important, right? 471 00:26:21,400 --> 00:26:24,200 You die without the regulator. But if if you over regulate, 472 00:26:24,200 --> 00:26:26,840 then you can't breathe. And there's a, there's a 473 00:26:27,080 --> 00:26:30,520 tendency right now to take whatever pre-existing laws there 474 00:26:30,520 --> 00:26:33,800 might be GDPR in this case and write something that looks like 475 00:26:33,800 --> 00:26:36,040 that for AI. Well, it's not the same thing. 476 00:26:36,040 --> 00:26:37,840 It's not even close to the same thing. 477 00:26:38,200 --> 00:26:42,800 And So what you wind up with is this tapestry of complexity that 478 00:26:42,800 --> 00:26:44,640 makes it very hard to take a step forward. 479 00:26:44,640 --> 00:26:46,920 Now, I'm not saying that that was the right thing or the wrong 480 00:26:46,920 --> 00:26:49,960 thing because like you, you know, I'm not, I've, I work with 481 00:26:49,960 --> 00:26:52,480 those people all the time, but I'm not in that space and they 482 00:26:52,480 --> 00:26:53,960 have to do something. I get it. 483 00:26:55,120 --> 00:26:59,400 The, the, the, the software issue that you talked about, the 484 00:26:59,400 --> 00:27:03,240 crowd strike thing, you know, was delivered kind of by 485 00:27:03,240 --> 00:27:04,880 Microsoft, right? Because most people had 486 00:27:04,880 --> 00:27:06,960 Microsoft. But you know, if you play it 487 00:27:06,960 --> 00:27:10,920 back and you unpeel the onion and who produced the blank file 488 00:27:10,920 --> 00:27:14,080 that got distributed with the update that got processed by the 489 00:27:14,080 --> 00:27:17,600 software that was acting like a driver that was allowed by the 490 00:27:17,600 --> 00:27:22,480 kernel of the operating system. This is a nightmare to say where 491 00:27:22,480 --> 00:27:25,360 where is the smoke detector that should have figured this out. 492 00:27:25,760 --> 00:27:29,640 And the world learned about how its own security was working a 493 00:27:29,640 --> 00:27:34,200 lot that day, right? So that was a very telling kind 494 00:27:34,200 --> 00:27:36,600 of moment for those in cybersecurity. 495 00:27:36,600 --> 00:27:39,880 They kind of understood it relatively quickly and it wasn't 496 00:27:39,880 --> 00:27:43,480 a big deal to fix. But Oh my gosh, the impact of it 497 00:27:43,480 --> 00:27:46,480 was just epic. That also tells a story about 498 00:27:46,480 --> 00:27:49,880 how connected the world is right now and how much we have to be 499 00:27:49,880 --> 00:27:52,160 careful as we take these steps forward. 500 00:27:52,560 --> 00:27:55,440 And so that is a counter argument for some regulation to 501 00:27:55,440 --> 00:27:58,480 say you can't just do whatever you want to do and apologize 502 00:27:58,480 --> 00:28:00,760 later because it had never happened before. 503 00:28:01,240 --> 00:28:04,800 So we're in a very difficult time right now In in 504 00:28:05,080 --> 00:28:07,520 epistemology, they're sometimes referred to these as critical 505 00:28:07,520 --> 00:28:10,280 incidents or critical moments where there's a point of 506 00:28:10,280 --> 00:28:12,320 reflection going on. It's hard to see when you're 507 00:28:12,320 --> 00:28:15,400 part of it, but we are definitely part of it and we're 508 00:28:15,400 --> 00:28:18,640 making little decisions right now that are going to have very 509 00:28:18,640 --> 00:28:21,200 big impact later with imperfect information. 510 00:28:21,200 --> 00:28:24,800 The these companies that your question started with, you know, 511 00:28:25,440 --> 00:28:27,600 three hours ago, Michael, the question was something about, 512 00:28:27,960 --> 00:28:31,000 you know, what companies can do. You know, companies are trying 513 00:28:31,000 --> 00:28:33,360 to serve their shareholders, their customers, their employees 514 00:28:33,360 --> 00:28:36,400 and their future market. And it's impossible to serve all 515 00:28:36,400 --> 00:28:38,600 those at the same time. Absolutely impossible. 516 00:28:38,880 --> 00:28:41,800 So if you want to maximize shareholder value, you wind up 517 00:28:41,800 --> 00:28:44,640 doing kind of really stupid short term things that often 518 00:28:44,640 --> 00:28:47,440 kill your company. If you want to maximize, you 519 00:28:47,440 --> 00:28:50,520 know, employee satisfaction and engagement, then you wind up 520 00:28:50,520 --> 00:28:52,040 doing things your customers don't care about. 521 00:28:52,040 --> 00:28:54,640 If you do everything your customers want, then you don't 522 00:28:54,640 --> 00:28:56,920 make any money because your margins go to zero and then your 523 00:28:56,920 --> 00:28:59,080 shareholders get upset. So it's just a whole bunch of 524 00:28:59,080 --> 00:29:03,080 ways to kill yourself. So now throw AI into this mix 525 00:29:03,080 --> 00:29:05,720 and then throw all this regulation into this mix and you 526 00:29:05,720 --> 00:29:07,560 get where we are right now in these boardrooms. 527 00:29:07,560 --> 00:29:10,360 This is not an easy place to be. Where are we? 528 00:29:10,400 --> 00:29:13,560 We are very similar to how when the automobile first came out. 529 00:29:14,040 --> 00:29:17,040 It's worth knowing how long did it take before seatbelts arrive? 530 00:29:18,280 --> 00:29:21,520 More than half a century now I'm hoping with AI we don't have 531 00:29:21,520 --> 00:29:23,200 that. But again, to try and think that 532 00:29:23,200 --> 00:29:25,720 we're going to immediately get it right within the 1st 6 to 12 533 00:29:25,720 --> 00:29:27,160 months, probably not going to happen. 534 00:29:27,160 --> 00:29:31,320 But let me zoom into a very specific sort of, and I think I 535 00:29:31,600 --> 00:29:33,480 Anthony sort of said this when we were talking about earlier 536 00:29:33,480 --> 00:29:36,560 about how when AI works and doesn't. 537 00:29:36,560 --> 00:29:39,360 And what I think we should unpack is not only AI is created 538 00:29:39,360 --> 00:29:42,320 equal, you know, we've been talking about generative AI, but 539 00:29:42,320 --> 00:29:45,640 there is other forms of AI ranging from experts rule based 540 00:29:45,640 --> 00:29:48,560 systems to decision support systems to computer vision. 541 00:29:48,920 --> 00:29:51,840 And I think the way you're going to trust, trust in computer 542 00:29:51,840 --> 00:29:56,480 vision, which is deterministic, which it is actually much more 543 00:29:56,480 --> 00:29:59,160 trustworthy and is not prone to any hallucinations whatsoever, 544 00:29:59,400 --> 00:30:01,520 is dramatically different than generative AI. 545 00:30:01,920 --> 00:30:04,640 And I raise that because I think, you know, if there is 546 00:30:04,640 --> 00:30:07,040 anything that needs to be talked about at the boardroom level, 547 00:30:07,040 --> 00:30:10,160 it's first understand the different tools when we talk 548 00:30:10,160 --> 00:30:12,880 about AI, that it's not monolithic and you need to 549 00:30:12,880 --> 00:30:14,800 understand whether you're reaching for a hammer or a 550 00:30:14,800 --> 00:30:16,360 screwdriver. And the trouble is we're writing 551 00:30:16,360 --> 00:30:18,400 regulations as if it is monolithic. 552 00:30:18,600 --> 00:30:22,760 And then the other thing that's actually to me is how we somehow 553 00:30:22,760 --> 00:30:26,880 think there's going to be 1 AI regulation to rule them all when 554 00:30:26,880 --> 00:30:29,840 we know when IT systems came online, these things called 555 00:30:29,840 --> 00:30:32,960 advanced data processing systems, later IT, you know, you 556 00:30:32,960 --> 00:30:36,600 don't write IT regulation for health and think it's just as 557 00:30:36,600 --> 00:30:39,520 good for the defense sector or just as good for the commercial 558 00:30:39,520 --> 00:30:41,400 recommendation sector. I mean, there are existing laws 559 00:30:41,400 --> 00:30:44,120 in the United States, there's HIPAA, there's the Bank Secrecy 560 00:30:44,120 --> 00:30:46,680 Act for banks. And so I actually think the more 561 00:30:46,680 --> 00:30:49,800 pragmatic approach is to go back to the existing rules that were 562 00:30:49,800 --> 00:30:54,040 written for both human and IT systems and say, where does AI 563 00:30:54,040 --> 00:30:56,720 fail? Because maybe it's too fast to 564 00:30:56,720 --> 00:30:59,960 speed or scope and upgrade those existing laws as opposed to 565 00:30:59,960 --> 00:31:02,400 trying to rate policy. And I'll give a very specific 566 00:31:02,400 --> 00:31:05,240 example right now in the United States, and it's not just the 567 00:31:05,240 --> 00:31:06,960 United States, there's a thing called health level 7. 568 00:31:06,960 --> 00:31:09,040 It's the international standard for interoperability across 569 00:31:09,040 --> 00:31:12,040 medical systems. There is a standard for tracking 570 00:31:12,040 --> 00:31:15,840 the provenance of a decision. You do record in your medical 571 00:31:15,840 --> 00:31:19,040 record when a physician made a decision and on what data they 572 00:31:19,040 --> 00:31:21,240 did. You don't record in the medical 573 00:31:21,240 --> 00:31:24,280 record when an AI does. And the number one usage right 574 00:31:24,280 --> 00:31:27,760 now for medical settings is actually physicians love to 575 00:31:27,760 --> 00:31:30,960 type, either type or dictate three to four short bullets and 576 00:31:30,960 --> 00:31:35,320 then say give me 20, sorry, 2 to 3 pages of physician notes that 577 00:31:35,320 --> 00:31:38,320 I will then put into the file. Well, aside from the fact that 578 00:31:38,320 --> 00:31:41,920 probably 1, both the company or organization that is delivering 579 00:31:41,920 --> 00:31:45,200 your health care should know was that physician notes actually 580 00:31:45,200 --> 00:31:47,120 written by the physician or written by an AI? 581 00:31:47,120 --> 00:31:49,920 And if so, which one? And the patient should know that 582 00:31:49,920 --> 00:31:52,600 too, too. There is a very real risk that 583 00:31:52,600 --> 00:31:57,600 if that AI output is later read back in by the same AI, there's 584 00:31:57,600 --> 00:31:59,400 a thing called model destabilization. 585 00:31:59,400 --> 00:32:01,920 Sometimes it's referred to as jargon as AI self 586 00:32:01,920 --> 00:32:05,040 cannibalization, where basically the model essentially it's, it's 587 00:32:05,040 --> 00:32:07,680 oversimplifying, but it over regresses and it starts to 588 00:32:07,680 --> 00:32:10,080 collapse on on itself and it starts to make bad decisions. 589 00:32:10,080 --> 00:32:12,400 And so I actually worry, Michael, that the question about 590 00:32:12,400 --> 00:32:15,840 deceptive versus trustworthy AI, again, it gets back to human 591 00:32:15,840 --> 00:32:17,480 organizations and what humans do. 592 00:32:18,080 --> 00:32:21,560 We may see lawsuits three to five years from now because 593 00:32:21,560 --> 00:32:24,960 companies didn't at least take the the first step of at least 594 00:32:24,960 --> 00:32:29,240 tagging and labeling when and a decision was made by an AI or an 595 00:32:29,280 --> 00:32:32,680 AI plus a human and then later if that was then fed back into a 596 00:32:32,680 --> 00:32:35,120 machine. Did they did they do the 597 00:32:35,120 --> 00:32:37,720 appropriate actions to avoid model destabilization? 598 00:32:38,120 --> 00:32:41,560 Really love this example from medical industry and it might go 599 00:32:41,960 --> 00:32:45,640 into further industries like food industry, but I have two 600 00:32:45,640 --> 00:32:48,800 further suggestions for regulators and for those who are 601 00:32:48,800 --> 00:32:51,600 serving on boards. So one is and by the way, this 602 00:32:51,600 --> 00:32:56,520 is not an ultimate solution, but we know that today more money is 603 00:32:56,520 --> 00:32:59,800 being spent on capability research than on transparency, 604 00:32:59,800 --> 00:33:04,920 safety, you name it, responsibility and it's like 97% 605 00:33:04,920 --> 00:33:08,360 of all the papers are published on behalf of those capabilities 606 00:33:08,360 --> 00:33:12,320 and only 3% on X AI. So this explainable AI and 607 00:33:12,320 --> 00:33:15,560 whatsoever. So we might motivate company to 608 00:33:15,560 --> 00:33:18,520 spend more money on that or to support a foundation or in 609 00:33:18,520 --> 00:33:22,520 institute university to invest into that. 610 00:33:22,520 --> 00:33:27,040 That could be motivational and not just punitive if we go into 611 00:33:27,040 --> 00:33:29,640 regulation. And then I'm really, I mean, I 612 00:33:29,640 --> 00:33:34,440 absolutely I'm scared about this issue with cybersecurity because 613 00:33:34,440 --> 00:33:38,200 AI is a tool, not just in the hands of good people, but 614 00:33:38,200 --> 00:33:41,360 criminals are using this really every single minute. 615 00:33:41,680 --> 00:33:44,880 And I think we just need to rethink how we approach 616 00:33:44,880 --> 00:33:47,840 cybersecurity and how we regulate cybersecurity, things 617 00:33:47,840 --> 00:33:52,840 like what is the exposure, cyber exposure, Can you quantify it? 618 00:33:53,080 --> 00:33:55,920 Can you name what is this financial figure? 619 00:33:56,160 --> 00:33:59,760 So this is tremendously important and very few people 620 00:33:59,760 --> 00:34:02,160 are talking about it. You mentioned quantum 621 00:34:02,160 --> 00:34:06,040 technologies early on and and certainly we are on the verge of 622 00:34:06,040 --> 00:34:09,320 the point where everything that was encrypted will now be 623 00:34:09,320 --> 00:34:12,280 readable, even the data that was stolen, right. 624 00:34:12,400 --> 00:34:16,400 We are on the verge of, we will have a way to do quantum 625 00:34:16,400 --> 00:34:19,159 decryption before we will have a really good way to do better 626 00:34:19,159 --> 00:34:22,679 quantum encryption other than scattering all the keys all over 627 00:34:22,679 --> 00:34:25,080 the place. So the world is going to get 628 00:34:25,080 --> 00:34:28,400 much more complicated. We have to start thinking more. 629 00:34:28,400 --> 00:34:30,480 And I'm not going to say start thinking, because there's some 630 00:34:30,480 --> 00:34:34,520 fine folks out there thinking about this, about novel cyber 631 00:34:34,520 --> 00:34:37,320 malfeasance. What are the types of not just 632 00:34:37,320 --> 00:34:41,400 the bad things that happened today with data exfiltration and 633 00:34:41,400 --> 00:34:44,080 and Trojans and malware and all this stuff. 634 00:34:44,360 --> 00:34:47,960 Yeah, we need to get really even better at that all the time. 635 00:34:48,400 --> 00:34:50,440 But that's necessary and not sufficient. 636 00:34:50,440 --> 00:34:53,159 We have to, I spend a fair amount of time and David, you 637 00:34:53,159 --> 00:34:56,480 know this thinking about what future bad guys might be able to 638 00:34:56,480 --> 00:35:01,080 do in in less than a generation with technology that I suspect 639 00:35:01,080 --> 00:35:04,080 will be available by that. And it's not hard for me to come 640 00:35:04,080 --> 00:35:06,040 up with. As an example, one of the things 641 00:35:06,040 --> 00:35:08,960 that we've looked at is using flocking and swarming 642 00:35:08,960 --> 00:35:11,840 algorithms. So the way flocks of birds can 643 00:35:11,840 --> 00:35:14,080 bifurcate and swarms tend to get bigger and bigger, right? 644 00:35:14,080 --> 00:35:16,520 So there are clustering algorithms that behave like 645 00:35:16,520 --> 00:35:17,960 that. Those two different behaviors, 646 00:35:18,240 --> 00:35:21,960 if you start applying them to botnet attacks and if you start 647 00:35:21,960 --> 00:35:25,480 imagining that the botnet attack, the botnet swarm or the 648 00:35:25,480 --> 00:35:28,440 botnet flock, depending on what you want to call it, we'll get 649 00:35:28,440 --> 00:35:31,440 smarter about how it's failing and therefore better at 650 00:35:31,440 --> 00:35:35,120 attacking you. That is a horribly it. 651 00:35:35,120 --> 00:35:37,040 It's just a horrific thing to imagine. 652 00:35:37,560 --> 00:35:39,520 And I know how to write code that would do that. 653 00:35:39,520 --> 00:35:42,960 And who am I right? So if, if, if we can do that 654 00:35:42,960 --> 00:35:47,760 sort of thing at an unimaginable scale in the very near future 655 00:35:48,280 --> 00:35:50,640 and we're just getting really good at fixing yesterday's 656 00:35:50,640 --> 00:35:54,040 malware, we are in for a world of hurt in that world that 657 00:35:54,040 --> 00:35:56,040 you're talking about. The good news is there are some 658 00:35:56,040 --> 00:36:00,160 fine folks thinking about that. The the even better news is, to 659 00:36:00,160 --> 00:36:02,800 your point, we need to incent them to think harder and faster 660 00:36:02,800 --> 00:36:05,040 about that. Anastacio made a very 661 00:36:05,040 --> 00:36:10,880 interesting point. She said that investment is far 662 00:36:10,880 --> 00:36:15,120 higher for technology capability, AI capability than 663 00:36:15,120 --> 00:36:20,440 it is for I'll use the term compliance or cybersecurity. 664 00:36:20,960 --> 00:36:28,400 And you know, look, human nature simply tells us that innovation 665 00:36:28,400 --> 00:36:32,960 is fun and compliance is not. And innovation makes money and 666 00:36:32,960 --> 00:36:35,880 compliance costs US money. And yes, there's a cost to 667 00:36:35,880 --> 00:36:41,320 society, but it's not to me if I'm a software company directly, 668 00:36:41,600 --> 00:36:44,200 unless of course there's there's a data breach. 669 00:36:44,800 --> 00:36:49,160 O Part of this is definitely trying to go against the tide of 670 00:36:49,440 --> 00:36:51,640 human nature from that standpoint. 671 00:36:51,760 --> 00:36:55,960 This is the argument for, you know, smoke detectors and fire 672 00:36:55,960 --> 00:36:59,360 alarms and life preservers and all of that is, you know, I 673 00:36:59,360 --> 00:37:02,560 have, I have life preservers on my boat because I'm required to 674 00:37:02,560 --> 00:37:05,160 have them on my boat. And if I have a boat, I want the 675 00:37:05,160 --> 00:37:07,720 best life preserver that's going to preserve my life, right? 676 00:37:08,120 --> 00:37:11,280 I want it, but most people will not think that way. 677 00:37:11,280 --> 00:37:14,080 And if I'm going to put a ferry out there with, you know, 2000 678 00:37:14,080 --> 00:37:16,480 of these things, maybe I start thinking about how much they 679 00:37:16,480 --> 00:37:19,160 cost and and you know, what's the most cost effective 680 00:37:19,160 --> 00:37:23,120 deployment of, you know, mandatorily require life saving 681 00:37:23,120 --> 00:37:26,360 equipment rather than how do I save David Bray's life because 682 00:37:26,360 --> 00:37:29,440 we need more people like. Him well, and so, but maybe if I 683 00:37:29,440 --> 00:37:32,080 can also jump on that, that question that you asked Michael, 684 00:37:32,080 --> 00:37:35,200 I think, you know, we've been talking about cybersecurity, but 685 00:37:35,200 --> 00:37:37,560 you can actually create a whole lot of damage without ever 686 00:37:37,560 --> 00:37:41,200 breaching a system. And what we're seeing right now 687 00:37:41,200 --> 00:37:45,080 and This is why trust versus deception is so important, you 688 00:37:45,080 --> 00:37:48,360 know, so using 10 minutes of compute time from worm GPT, 689 00:37:48,360 --> 00:37:52,000 which is the dark side cousin of ChatGPT and some plug insurance 690 00:37:52,360 --> 00:37:55,280 using data stolen from healthcare data breaches. 691 00:37:55,280 --> 00:37:57,880 And many of us may have been affected by a data breach in the 692 00:37:57,880 --> 00:38:00,640 last year. That data can then train and 693 00:38:00,640 --> 00:38:03,560 create 1,000,000 realistic looking medical records complete 694 00:38:03,560 --> 00:38:07,800 with chest X-ray, physician's notes, doctor's notes, or a 695 00:38:07,800 --> 00:38:11,600 claim at $250, which is below the fraud detection threshold 696 00:38:11,600 --> 00:38:14,080 for several of the services here in the United States because 697 00:38:14,080 --> 00:38:16,440 it's more expensive to adjudicate it than to pay that 698 00:38:16,440 --> 00:38:18,560 out. And that's 1,000,000 records for 699 00:38:18,560 --> 00:38:21,640 upper respiratory infection. Now what the points to, and this 700 00:38:21,640 --> 00:38:25,600 is what Anthony's talking about is again, the solution is we've 701 00:38:25,600 --> 00:38:29,480 got to incentivize those people who are doing innovation to find 702 00:38:29,480 --> 00:38:32,480 innovative solutions to either adjudicate faster, adjudicate 703 00:38:32,480 --> 00:38:35,520 with more veracity, whatever it might be at scale. 704 00:38:35,760 --> 00:38:38,440 Given that AI is basically doing the same thing that 705 00:38:38,520 --> 00:38:40,560 unfortunately, you know, remember when the Xerox copier 706 00:38:40,560 --> 00:38:42,960 first came out and we had to upgrade our dollar bills because 707 00:38:42,960 --> 00:38:44,920 some people were doing called our copies of dollar bills. 708 00:38:44,920 --> 00:38:47,920 So we upgraded that as well. The challenges and the unique 709 00:38:47,920 --> 00:38:51,520 challenge with AI is just what's called generative adversarial 710 00:38:51,520 --> 00:38:54,320 networks or Gans. And so the moment you create a 711 00:38:54,320 --> 00:38:57,080 solution that's really good at filtering, this is valid, this 712 00:38:57,080 --> 00:39:00,680 is not valid, then a bad actor will then use that GAN to say, 713 00:39:00,880 --> 00:39:03,160 use that GAN to get good at fooling it. 714 00:39:03,160 --> 00:39:05,520 And so it's going to be predator prey relationships. 715 00:39:05,520 --> 00:39:07,880 But I do think this points to at the speed at which this is 716 00:39:07,880 --> 00:39:11,640 happening, we have to figure out ways to incentivize either 717 00:39:11,640 --> 00:39:13,600 because we're shining a spotlight on it because 718 00:39:13,600 --> 00:39:16,040 companies are realizing that's a whole massive amount of money in 719 00:39:16,040 --> 00:39:18,080 terms of payments that they shouldn't be made for fraud 720 00:39:18,080 --> 00:39:20,560 purposes or whatever. But there has to be an incentive 721 00:39:20,560 --> 00:39:23,440 for market based solutions to this because central planning 722 00:39:23,720 --> 00:39:25,320 will not get us out of these solutions. 723 00:39:25,480 --> 00:39:28,000 But you do have to figure out a way to shine the spotlight on it 724 00:39:28,200 --> 00:39:30,480 because the speed at which this is moving, it's almost like what 725 00:39:30,480 --> 00:39:35,240 we're seeing in Ukraine where every every six to nine months 726 00:39:35,240 --> 00:39:38,680 there is a generational leap in terms of how they're using drone 727 00:39:38,680 --> 00:39:41,000 on drone for conflicts. The only difference is this is 728 00:39:41,000 --> 00:39:43,680 going to be behind the scenes in terms of AI spoofing. 729 00:39:43,960 --> 00:39:46,800 How do you know if that image is real, that video is real, or 730 00:39:46,800 --> 00:39:48,160 that person you're talking to is real? 731 00:39:48,520 --> 00:39:53,640 This is a perfect time to subscribe to the CXO Talk 732 00:39:53,640 --> 00:39:57,040 newsletter. Do that now so we can notify you 733 00:39:57,480 --> 00:40:00,080 of upcoming shows. OK, so questions. 734 00:40:00,080 --> 00:40:04,680 So who Wei Wang on LinkedIn? Says that data collection right 735 00:40:04,680 --> 00:40:09,080 now is being affected not only by linguistic transliteration or 736 00:40:09,080 --> 00:40:13,520 translation nuances, but we're also seeing complex data 737 00:40:13,520 --> 00:40:17,320 patterns when the technology, languages, and platforms are 738 00:40:17,320 --> 00:40:20,880 contributing to changes in how the data is stored. 739 00:40:21,840 --> 00:40:25,200 What do you recommend in navigating and preparing for 740 00:40:25,200 --> 00:40:28,760 these increased complexities against the rate of how frequent 741 00:40:28,760 --> 00:40:31,320 we are collecting data like this? 742 00:40:32,080 --> 00:40:35,960 And before you answer, I have a request to the audience which is 743 00:40:37,040 --> 00:40:39,920 you got a dumb moderator so just keep your questions short. 744 00:40:41,800 --> 00:40:43,800 I've got to jump on that one. I think Anthony paid that 745 00:40:43,800 --> 00:40:44,880 individual to ask that. Question. 746 00:40:44,880 --> 00:40:47,320 I did not. I did not but I say it's a 747 00:40:47,320 --> 00:40:50,200 brilliant question so I'm going to paraphrase the question. 748 00:40:50,280 --> 00:40:52,880 There's lots of languages out there and there's lots of people 749 00:40:52,880 --> 00:40:56,680 speaking in those languages and we're trying to teach AI to suck 750 00:40:56,680 --> 00:40:59,680 in data spoken in those languages or written in those 751 00:40:59,680 --> 00:41:02,840 languages. And a lot of times the the 752 00:41:02,840 --> 00:41:06,440 systems that are showing you that language, it's not the way 753 00:41:06,440 --> 00:41:09,080 it was originally written. So I'm sure most of us have 754 00:41:09,480 --> 00:41:12,160 people in our social network where they write something in 755 00:41:12,160 --> 00:41:14,800 their language and then you see what they allegedly said. 756 00:41:15,200 --> 00:41:17,160 And then there's a translated button on the bottom. 757 00:41:17,160 --> 00:41:19,720 So it's been translated or I'll say transformed. 758 00:41:20,040 --> 00:41:22,720 So there's a lot of data out there that's gone through 759 00:41:22,720 --> 00:41:25,080 linguistic transformation either intentionally or 760 00:41:25,080 --> 00:41:28,920 unintentionally, either at the time of writing or subsequent to 761 00:41:28,920 --> 00:41:31,440 the time of writing. And that introduces all kinds of 762 00:41:31,440 --> 00:41:34,840 nuance into the language that we consume by our AI systems. 763 00:41:35,000 --> 00:41:37,560 That's got to have an impact. And it turns out that it does. 764 00:41:37,960 --> 00:41:40,760 And there are two different arguments to, to and there, 765 00:41:40,760 --> 00:41:42,320 there's no winning argument here yet. 766 00:41:42,600 --> 00:41:46,480 So the Turing argument is that if I get enough and most people 767 00:41:46,480 --> 00:41:49,440 you will find addressing this problem, we'll we'll use this 768 00:41:49,440 --> 00:41:51,680 argument. As long as I get enough examples 769 00:41:51,680 --> 00:41:54,880 of that language, I'll be fine. Just give me millions and 770 00:41:54,880 --> 00:41:56,960 billions and billions of articulations and no matter how 771 00:41:56,960 --> 00:41:59,920 complicated it is, I will regress around it and I will 772 00:41:59,920 --> 00:42:03,000 figure it out. The contrary argument to that is 773 00:42:03,000 --> 00:42:06,240 that when we speak a language, we change it and we with the 774 00:42:06,240 --> 00:42:08,840 nuance gets introduced into that language. 775 00:42:08,840 --> 00:42:12,040 Things like sarcasm and neologism and borrowing words 776 00:42:12,040 --> 00:42:14,320 from foreign languages, all of these things confound to our 777 00:42:14,320 --> 00:42:17,120 ability to understand. And so right now we're speaking 778 00:42:17,120 --> 00:42:22,040 in a certain way that is quasi academic, intelligent. 779 00:42:22,040 --> 00:42:24,320 We're trying to speak in complete sentences. 780 00:42:24,320 --> 00:42:26,360 We know that there's going to be a transcript later. 781 00:42:26,360 --> 00:42:27,960 So we don't want to look like idiots, right? 782 00:42:27,960 --> 00:42:30,280 So we're we're trying to say things in a certain way. 783 00:42:30,600 --> 00:42:33,600 If we talk maybe over an adult beverage, it might come out a 784 00:42:33,600 --> 00:42:36,640 little bit different. So how do when, when you 785 00:42:36,640 --> 00:42:40,160 introduce language on top of that, how does it perturb your 786 00:42:40,160 --> 00:42:42,760 ability to understand? Is that person upset? 787 00:42:42,760 --> 00:42:45,640 Is that person lying? Is that person, and one of these 788 00:42:45,640 --> 00:42:47,360 people is an imposter? Which one is it? 789 00:42:47,640 --> 00:42:50,400 These are all right. Now you can have an honorary PhD 790 00:42:50,400 --> 00:42:52,960 in computational linguistics by tackling any one of those 791 00:42:52,960 --> 00:42:56,760 questions really well because I'm sorry to tell you, but the 792 00:42:56,760 --> 00:42:58,680 technology hasn't gotten there yet. 793 00:42:58,880 --> 00:43:01,160 It's getting better. There are things as Huey 794 00:43:01,160 --> 00:43:04,120 mentioned in her question, there are things that you can do to 795 00:43:04,120 --> 00:43:07,200 recognize commonly occurring graphemes, to recognize things 796 00:43:07,200 --> 00:43:09,560 like David mentioned misspellings before. 797 00:43:09,760 --> 00:43:12,640 Some misspelling can be intentional, some misspellings 798 00:43:12,640 --> 00:43:15,480 can be accidental. Every time I write something to 799 00:43:15,480 --> 00:43:18,640 Anastasia my my autocorrect changes the spelling of her name 800 00:43:19,000 --> 00:43:21,440 because there's another person I know that spells the name 801 00:43:21,440 --> 00:43:23,840 differently, and if I don't catch it, I send it to her 802 00:43:23,840 --> 00:43:25,600 spelled wrong. She knows this now, so hopefully 803 00:43:25,600 --> 00:43:28,640 she's not offended. These are things that are 804 00:43:28,640 --> 00:43:31,800 happening all the time and it's going to get worse. 805 00:43:32,040 --> 00:43:34,880 So what we have to get better at is not just accepting that 806 00:43:34,880 --> 00:43:38,480 computational linguistics is give me the dictionary, because 807 00:43:38,480 --> 00:43:40,400 there are low context languages where there aren't. 808 00:43:40,480 --> 00:43:44,880 There isn't one really good source, or there are languages 809 00:43:44,880 --> 00:43:47,160 where the only really good source is the religious text, 810 00:43:47,160 --> 00:43:48,520 right? And that's not how people speak. 811 00:43:48,520 --> 00:43:51,280 So now you've just consumed the Bible and you're trying to speak 812 00:43:51,280 --> 00:43:52,440 English. We don't speak like that 813 00:43:52,440 --> 00:43:54,560 anymore. So there's that problem, the low 814 00:43:54,560 --> 00:43:56,640 context problem. There's the multilingual 815 00:43:56,640 --> 00:43:58,960 disambiguation problem where Peter, people are writing in 816 00:43:58,960 --> 00:44:00,920 more than one language. And then there's the language 817 00:44:00,920 --> 00:44:03,760 transformation problem where people are changing what was 818 00:44:03,760 --> 00:44:06,640 written and you're reading what has been already transformed. 819 00:44:07,080 --> 00:44:11,120 All of those are Wild West right now and and you can you can make 820 00:44:11,120 --> 00:44:14,320 a lot of impact or working in any one of those fields right 821 00:44:14,320 --> 00:44:17,480 now. Greg Walters on LinkedIn says 822 00:44:17,760 --> 00:44:21,480 we've lost the EUI think, quote UN quote, alien is the best way 823 00:44:21,480 --> 00:44:26,400 to view AI in that vein. Other than that, we have no way. 824 00:44:26,400 --> 00:44:29,080 Shouldn't we look at AI through a brand new lens? 825 00:44:29,080 --> 00:44:32,960 A lens with no previous anchors or history or KPIs or best 826 00:44:32,960 --> 00:44:36,320 practices? AI does not fit only in a 827 00:44:36,320 --> 00:44:39,400 vertical or horizontal. It is everything, everywhere, 828 00:44:39,440 --> 00:44:41,240 all at once. There are no experts. 829 00:44:41,920 --> 00:44:45,360 What is a new perspective? AI as mirror. 830 00:44:45,800 --> 00:44:49,760 When 2001 happened, I was responding to 911 in the anthrax 831 00:44:49,760 --> 00:44:51,720 events. There were conspiracy theories 832 00:44:51,720 --> 00:44:54,840 that have were around, but none of them were able to take root. 833 00:44:54,840 --> 00:44:56,720 I mean they were saying like we've done it to ourselves, 834 00:44:56,720 --> 00:44:58,800 false flag, US have done it but they didn't take root. 835 00:44:59,400 --> 00:45:02,720 Fast forward to 2009, I'm on the ground in Afghanistan and sadly 836 00:45:02,720 --> 00:45:06,440 there's an event in western Afghanistan where the Taliban 837 00:45:06,440 --> 00:45:09,040 actually took a photo, a valid photo of AUS fighter jet flying 838 00:45:09,040 --> 00:45:12,400 overhead and then at the same time briefly took a photo of a 839 00:45:12,400 --> 00:45:14,800 propane tank detonating and unfortunately killing innocent 840 00:45:14,800 --> 00:45:18,000 Afghans. Both were true photos, but they 841 00:45:18,000 --> 00:45:20,120 were clearly out of context because then they went on social 842 00:45:20,120 --> 00:45:22,840 media and claimed US air strike kills innocent Afghans. 843 00:45:24,280 --> 00:45:26,240 The Department of Defense that we're investigating, and it took 844 00:45:26,240 --> 00:45:28,880 more than 4 1/2 weeks to figure out what happened during that 845 00:45:28,880 --> 00:45:30,360 time. Of course, the news media was 846 00:45:30,360 --> 00:45:33,040 blaming the US and our own ambassador apologized. 847 00:45:33,040 --> 00:45:35,640 That took 4 1/2 weeks to actually put that 848 00:45:36,040 --> 00:45:38,560 characterization of what really happened versus the deceptive 849 00:45:38,560 --> 00:45:41,600 narrative to bed. And then as you know, Michael, 850 00:45:41,600 --> 00:45:44,320 back in 2017 when I was at the Federal Communications 851 00:45:44,320 --> 00:45:48,120 Commission, there was an event in which we were getting 6007 852 00:45:48,120 --> 00:45:52,080 thousand 8000 comments a minute at 4:00 AM, five AM, 6:00 AM. 853 00:45:52,080 --> 00:45:54,640 And we were told by our lawyers we couldn't test for bots. 854 00:45:54,640 --> 00:45:57,560 Speaking of CAPTCHA, Anthony, we couldn't do invisible means 855 00:45:57,560 --> 00:45:59,000 because that would be seen as surveillance. 856 00:45:59,240 --> 00:46:02,320 And we couldn't block what was perceived to spam 100 comments 857 00:46:02,320 --> 00:46:03,960 from the same IP address per minute. 858 00:46:04,440 --> 00:46:08,680 And it took more than 4 1/2 years at that point in time for 859 00:46:08,680 --> 00:46:11,280 eventually New York Attorney General to adjudicate and say of 860 00:46:11,280 --> 00:46:14,440 the 23 million comments we got, 18,000,000 were politically 861 00:46:14,440 --> 00:46:17,440 manufactured, 9 million from one side of the aisle, 9 million 862 00:46:17,440 --> 00:46:20,520 from the other aisle. So over the last two decades, we 863 00:46:20,520 --> 00:46:25,920 are getting longer and longer tails for the more valid, more 864 00:46:25,960 --> 00:46:29,080 authentic, accurate characterization scenarios to 865 00:46:29,080 --> 00:46:31,400 get themselves out. We went from they didn't take 866 00:46:31,400 --> 00:46:35,240 root in 2001. Now it took 4 1/2 weeks in 2009, 867 00:46:35,240 --> 00:46:38,920 2017 it took 4 1/2 years. We are unintentionally through 868 00:46:38,920 --> 00:46:41,360 no one technology. I would submit Internet, 869 00:46:41,760 --> 00:46:45,520 smartphone, generative AI. We are laying the seeds in which 870 00:46:45,520 --> 00:46:46,760 it is. You know, it used to be the 871 00:46:46,760 --> 00:46:49,880 adage was, you know, a lie can go halfway around the world 872 00:46:49,880 --> 00:46:51,480 before truth gets on its sneakers. 873 00:46:51,840 --> 00:46:53,880 I would say at this point in time, I'd like and get to the 874 00:46:53,880 --> 00:46:56,360 other end of the solar system where the truth gets on its 875 00:46:56,360 --> 00:46:57,960 sneakers. And so This is why I would 876 00:46:57,960 --> 00:47:02,680 submit the solutions have to be technology neutral and also have 877 00:47:02,680 --> 00:47:05,760 to account for humans doing things regardless of anyone 878 00:47:05,760 --> 00:47:07,800 machine. I still would say at the end of 879 00:47:07,800 --> 00:47:10,720 the day, it's going to be sector specific because the severity of 880 00:47:10,720 --> 00:47:12,960 doing this, if you're recommending something to buy on 881 00:47:13,160 --> 00:47:16,320 a website is much different than making a decision about your 882 00:47:16,320 --> 00:47:17,920 health. It's much different than making 883 00:47:17,920 --> 00:47:21,000 a decision on a battlefield. Anastasio there are two 884 00:47:21,000 --> 00:47:23,920 questions and I'll combine them and direct them to you. 885 00:47:24,440 --> 00:47:30,440 Elizabeth Shaw says that there are strong agendas that power 886 00:47:31,560 --> 00:47:37,720 deceptive and malicious use of AI, such as greed, power, ETC. 887 00:47:38,040 --> 00:47:43,320 How do you encourage the ethical use of AI as opposed to the 888 00:47:43,320 --> 00:47:48,680 deceptive use? And then Arsalan Khan says that 889 00:47:48,680 --> 00:47:53,920 there are many underlying issues such as security, data bias, 890 00:47:53,920 --> 00:47:58,760 culture, jobs, and and so forth that AI can help or make worse. 891 00:47:59,960 --> 00:48:02,040 He's asking fundamentally the same thing. 892 00:48:02,040 --> 00:48:09,000 How do we incentivize organizations to take a, an 893 00:48:09,000 --> 00:48:13,360 ethical as opposed to a deceptive approach? 894 00:48:14,160 --> 00:48:17,880 And I'm, I'm paraphrasing these two questions, but that's 895 00:48:18,000 --> 00:48:20,240 fundamentally we spoke about that earlier. 896 00:48:20,400 --> 00:48:25,120 The incentive is incentivizing. If you incentivize prospective 897 00:48:25,120 --> 00:48:30,080 customer to be more aware and to be more educated on what is out 898 00:48:30,080 --> 00:48:32,800 there, because sometimes the issue is not the technology, but 899 00:48:32,800 --> 00:48:37,280 the business model and the why behind a certain construct in, 900 00:48:37,280 --> 00:48:39,680 for example, how social network is configured. 901 00:48:40,160 --> 00:48:44,520 And I think that the reduction of confusion and demystifying of 902 00:48:44,520 --> 00:48:48,640 AI is a very noble cause. So in my eyes, there are three 903 00:48:48,640 --> 00:48:54,360 fundamental buckets of risks in AI and one bucket is everything 904 00:48:54,360 --> 00:48:57,760 which has to do with design. For example, we are talking 905 00:48:57,760 --> 00:49:02,960 LLMS, and I mean, for my taste, LLM will always hallucinate. 906 00:49:03,000 --> 00:49:07,040 You might reduce certain things, but you know, it will still stay 907 00:49:07,240 --> 00:49:11,120 just because I'm not going to dive now into the theory of 908 00:49:11,120 --> 00:49:15,640 computation and all of the above, but it will hallucinate, 909 00:49:15,760 --> 00:49:20,240 period, because of the design. We are talking about AIS as if 910 00:49:20,560 --> 00:49:23,960 AIS are quite new. But actually the technology, the 911 00:49:23,960 --> 00:49:27,920 roots of technology is in the 40s, fifties and 60s of the last 912 00:49:28,080 --> 00:49:31,680 century. And we must go into a new wave 913 00:49:31,680 --> 00:49:34,640 of architectural design to improve. 914 00:49:34,840 --> 00:49:41,320 So this either by the construct, by the system, an issue or 915 00:49:41,440 --> 00:49:44,840 because there is a human mistake or obviously this is always 916 00:49:44,840 --> 00:49:47,840 possible, then there is a malicious intent. 917 00:49:48,160 --> 00:49:50,400 So we talked about cybersecurity. 918 00:49:50,400 --> 00:49:55,200 Criminals will apply AI for the dark purposes and some companies 919 00:49:55,400 --> 00:49:59,480 might want to seduce their customers into certain thinking. 920 00:49:59,480 --> 00:50:02,880 So the more educated the customer is, the better. 921 00:50:03,200 --> 00:50:07,920 Sometimes we are talking about scalable problems and sometimes 922 00:50:07,920 --> 00:50:11,880 we are talking about basic stupidity on behalf of a 923 00:50:11,880 --> 00:50:14,320 customer. And when I work with companies 924 00:50:14,320 --> 00:50:18,320 and I review their AI portfolio, sometimes I'm like, why are you 925 00:50:18,320 --> 00:50:22,520 spending this money on that? And the issue is that the 926 00:50:22,520 --> 00:50:26,560 customer believed that the sales personnel of a certain vendor 927 00:50:26,840 --> 00:50:32,840 and the vendor consists of 85% of sales and only 15% of 928 00:50:32,840 --> 00:50:35,840 engineers. Obviously, this vendor cannot 929 00:50:35,840 --> 00:50:39,120 execute. I'm not now going into ethical 930 00:50:39,120 --> 00:50:43,120 and trustworthy AI, just the basic configuration of whatever 931 00:50:43,120 --> 00:50:45,640 systems. So the more knowledge on behalf 932 00:50:45,640 --> 00:50:48,360 of a customer, the better is the outcome. 933 00:50:48,640 --> 00:50:52,320 And last but not least, and this will be very specific to an 934 00:50:52,320 --> 00:50:55,480 industry or business function. This is so-called human in the 935 00:50:55,480 --> 00:50:57,560 loop. This is the third bucket of 936 00:50:57,560 --> 00:51:01,520 so-called risk. Now define how much human and in 937 00:51:01,520 --> 00:51:05,720 what kind of loop when are we going to introduce the 938 00:51:05,720 --> 00:51:07,960 supervision and the human feedback? 939 00:51:08,160 --> 00:51:12,360 Once again, this is a balancing act and and this is a play in 940 00:51:12,360 --> 00:51:15,880 between of whoever is building and introducing the system and 941 00:51:15,880 --> 00:51:18,240 whoever is actually the customer. 942 00:51:18,520 --> 00:51:23,480 So it might be an answer which is not very simple. 943 00:51:23,480 --> 00:51:28,480 But to my eyes, this education and asking good questions, being 944 00:51:28,480 --> 00:51:32,760 capable to review the vendor portfolio and read it to write 945 00:51:32,760 --> 00:51:38,160 down what type of problems are we solving, do we need this and 946 00:51:38,160 --> 00:51:43,320 at what cost. AI might appear new, but our eye 947 00:51:43,320 --> 00:51:46,280 is still our eye. If I have this kind of, you 948 00:51:46,280 --> 00:51:51,840 know, rule of the term that 25% of my revenue line will be spent 949 00:51:51,840 --> 00:51:57,120 on compute and 15% of my revenue line is being spent on cleansing 950 00:51:57,120 --> 00:51:59,040 data. If I am great. 951 00:51:59,240 --> 00:52:04,040 Yeah, then you have your economics and you need to keep 952 00:52:04,040 --> 00:52:07,080 it in mind before deciding whether you are going into a 953 00:52:07,080 --> 00:52:09,800 certain AI implementation or not. 954 00:52:09,920 --> 00:52:13,440 What will it cost you? Do you really need to automate 955 00:52:13,440 --> 00:52:15,760 here to introduce an AI agent here? 956 00:52:15,880 --> 00:52:18,200 Or maybe a human will do just fine. 957 00:52:18,600 --> 00:52:22,960 Let me just ask each of you very briefly for a final thought. 958 00:52:23,560 --> 00:52:25,520 David, final thoughts on this topic. 959 00:52:26,000 --> 00:52:28,440 This is not a tech specific issue. 960 00:52:28,440 --> 00:52:32,160 This is a society and company level issue. 961 00:52:32,160 --> 00:52:35,080 And it's going to be solutions that look not at the tech 962 00:52:35,080 --> 00:52:38,040 specifically, but look broadly. And again, recognizing we're 963 00:52:38,040 --> 00:52:41,000 wrapping up data, data, data. We didn't talk at all about data 964 00:52:41,000 --> 00:52:42,160 governance. We didn't talk about data 965 00:52:42,160 --> 00:52:43,680 cleaning. I know, and Asajj just briefly 966 00:52:43,680 --> 00:52:45,800 touched on it. You know, one way where you 967 00:52:45,800 --> 00:52:48,680 could do most everything right in AI and completely fail is if 968 00:52:48,680 --> 00:52:50,560 your data is bad. And so I would say at the end of 969 00:52:50,560 --> 00:52:56,120 the day, the, the, the, the intent of pursuing more 970 00:52:56,120 --> 00:53:00,240 trustworthy and less deceptive AI begins first and foremost 971 00:53:00,440 --> 00:53:02,840 with putting yourself in the shoes of your different 972 00:53:02,840 --> 00:53:06,120 stakeholders and making sure whatever you do, tech or not, 973 00:53:06,200 --> 00:53:09,360 data or not, you are thinking about them and how you deploy 974 00:53:09,360 --> 00:53:12,520 things. Anthony, final thoughts. 975 00:53:12,960 --> 00:53:15,200 Couple of things. One, beware shiny objects, 976 00:53:15,200 --> 00:53:16,320 right? There's going to be more and 977 00:53:16,320 --> 00:53:18,800 more shiny objects, right? Whatever they're called right 978 00:53:18,800 --> 00:53:21,200 now, it's LLMS. That's great. 979 00:53:21,280 --> 00:53:24,760 What problem are you solving? Always step back and ask, what 980 00:53:24,760 --> 00:53:26,920 do I have to believe? What's the problem I'm solving? 981 00:53:27,280 --> 00:53:29,760 Don't get distracted by the shiny object because there's 982 00:53:29,760 --> 00:53:31,200 another one coming right behind it. 983 00:53:31,680 --> 00:53:35,000 The second thing is, and that's to David's point about cleaning 984 00:53:35,000 --> 00:53:37,200 the data. The second point is regulation 985 00:53:37,200 --> 00:53:39,120 is going to happen with you or without you. 986 00:53:39,120 --> 00:53:41,680 So get involved. Make sure that the regulators 987 00:53:41,680 --> 00:53:45,600 understand the unintended impact of some of the things that 988 00:53:45,600 --> 00:53:47,800 they're considering. We just, we just wait for these 989 00:53:47,800 --> 00:53:50,640 regulations to come out and then it's kind of too late. 990 00:53:50,880 --> 00:53:54,520 And then the third thing I would say is, and I often say this, be 991 00:53:54,520 --> 00:53:57,080 humble. You can't solve this alone. 992 00:53:57,200 --> 00:53:59,680 Get help. There's lots and lots of folks 993 00:53:59,680 --> 00:54:02,840 out there that have lots of expertise. 994 00:54:02,840 --> 00:54:04,600 We should make new mistakes, right? 995 00:54:04,600 --> 00:54:07,600 And we do that by involving other people in our decision 996 00:54:07,600 --> 00:54:10,520 process and and being humble enough to realize that I don't 997 00:54:10,520 --> 00:54:13,360 care how smart your smartest people are, you should bring in 998 00:54:13,360 --> 00:54:14,840 some people that disagree with you. 999 00:54:15,280 --> 00:54:18,040 And Anastasia, it looks like you're going to get the last 1000 00:54:18,040 --> 00:54:19,520 word here. Final thoughts. 1001 00:54:19,880 --> 00:54:23,440 We might be in technology and AI, but we are all in people 1002 00:54:23,440 --> 00:54:27,840 business ultimately, and it's really about human leadership 1003 00:54:27,840 --> 00:54:32,840 and thinking and humility. And I don't encourage people to 1004 00:54:32,840 --> 00:54:36,920 wait for some decision from let's say Alphabet or from some 1005 00:54:36,920 --> 00:54:40,360 new regulation out of Washington DC or Brussel. 1006 00:54:40,680 --> 00:54:44,520 I would really encourage local communities to look into the 1007 00:54:44,520 --> 00:54:49,560 local ecosystems and see what kind of colleges, universities 1008 00:54:49,560 --> 00:54:53,880 are out there interested in AI, offering some courses, what 1009 00:54:53,880 --> 00:54:56,560 schools are interested? What could you do for those 1010 00:54:56,560 --> 00:54:59,440 kids? What might be the local startups 1011 00:54:59,680 --> 00:55:02,080 to do, let's say startup breakfast? 1012 00:55:02,080 --> 00:55:05,320 So they they explain what they're doing and how they're 1013 00:55:05,320 --> 00:55:09,360 hiring into the company. Maybe some incumbent companies 1014 00:55:09,480 --> 00:55:14,080 which are adopting one or another type of technology. 1015 00:55:14,080 --> 00:55:18,480 So I would really encourage this bottom up movement rather than 1016 00:55:18,480 --> 00:55:22,120 waiting for some big guy somewhere to decide for you 1017 00:55:22,120 --> 00:55:25,760 because this is how you learn and this is how you create a 1018 00:55:25,760 --> 00:55:28,840 dialogue and ultimately the progress will happen through the 1019 00:55:28,840 --> 00:55:32,280 dialogue. We are dealing, as all three 1020 00:55:32,280 --> 00:55:38,760 have said, with human rather than specifically technology 1021 00:55:38,920 --> 00:55:42,680 issues. And so the question then becomes 1022 00:55:43,080 --> 00:55:50,880 how do we harness and focus this human energy for the ethical use 1023 00:55:50,880 --> 00:55:53,440 of AI? And I think with many other 1024 00:55:53,440 --> 00:55:56,440 things, it's going to require carrot and a stick. 1025 00:55:56,760 --> 00:56:01,280 There's no simple solution here. And as complex as the technology 1026 00:56:01,280 --> 00:56:03,600 is, the human aspects are always harder. 1027 00:56:04,240 --> 00:56:06,760 So, Carrot and Michael, are you building a snowman is what 1028 00:56:06,760 --> 00:56:09,120 you're talking about? You know, I'd love to build a 1029 00:56:09,120 --> 00:56:11,400 snowman. It's actually snowing right now 1030 00:56:11,400 --> 00:56:14,040 here in Boston. Wait, are you saying do you want 1031 00:56:14,040 --> 00:56:15,560 to build a snowman? Sorry I'm. 1032 00:56:15,880 --> 00:56:18,440 Sorry. Sorry, Michael, back to you. 1033 00:56:18,960 --> 00:56:24,880 Anyway, a huge thank you to our guest, Doctor Anastasia 1034 00:56:24,880 --> 00:56:29,840 Lauterbach, Doctor David Bray, and Doctor Anthony Scrofignano. 1035 00:56:29,920 --> 00:56:32,040 Thank you all so much for being here. 1036 00:56:32,040 --> 00:56:38,000 I'm very grateful to you all and a huge thank you to everybody 1037 00:56:38,000 --> 00:56:41,800 who watched today and especially you folks who asked such awesome 1038 00:56:41,880 --> 00:56:44,360 questions. I always say this, you guys in 1039 00:56:44,360 --> 00:56:50,480 the audience are amazing. Before you go, please subscribe 1040 00:56:50,480 --> 00:56:54,480 to our newsletter and subscribe to our YouTube channel. 1041 00:56:54,480 --> 00:56:58,800 Check out cxotalk.com. We have great shows coming up 1042 00:56:59,080 --> 00:57:03,640 and our next show will be back at the usual time of 1:00 1043 00:57:04,080 --> 00:57:05,880 Eastern. Thanks so much everybody and I 1044 00:57:05,880 --> 00:57:06,800 hope you have a great day.