1 00:00:04,920 --> 00:00:08,080 Speaker 1: On this episode of NEWTS World, what do we need 2 00:00:08,119 --> 00:00:12,280 Speaker 1: to understand about artificial intelligence now? And why has AI 3 00:00:12,400 --> 00:00:16,160 Speaker 1: become the buzzword of late My guest today is one 4 00:00:16,160 --> 00:00:18,880 Speaker 1: of the key players in the world of AI. He's 5 00:00:18,880 --> 00:00:22,640 Speaker 1: written about extensively. His latest book is Power and Prediction, 6 00:00:23,239 --> 00:00:27,560 Speaker 1: The Disruptive Economics of Artificial Intelligence. I'm really pleased to 7 00:00:27,600 --> 00:00:31,479 Speaker 1: welcome my guest, Professor I j agro Wall. He is 8 00:00:31,560 --> 00:00:36,839 Speaker 1: the Jeffrey Tabor Chair in Entrepreneurship Innovation and Professor of 9 00:00:36,920 --> 00:00:41,680 Speaker 1: Strategic Management at the University of Toronto's Rothmann School of Management. 10 00:00:41,760 --> 00:00:45,519 Speaker 1: In addition, he is a research associate at the National 11 00:00:45,560 --> 00:00:49,960 Speaker 1: Bureau of Economic Research in Cambridge, Massachusetts, and faculty affiliate 12 00:00:50,320 --> 00:01:08,080 Speaker 1: at the Vector Institute for Artificial Intelligence in Toronto, Canada. Jay, 13 00:01:08,160 --> 00:01:10,400 Speaker 1: Welcome and thank you for joining me on NEWTS World. 14 00:01:10,880 --> 00:01:13,000 Speaker 2: Thanks very much for having me newt You. 15 00:01:13,000 --> 00:01:14,600 Speaker 3: Know, it sounds like you're very busy. 16 00:01:15,440 --> 00:01:17,880 Speaker 2: Well, I suspect you're very busy too. So we're all 17 00:01:17,880 --> 00:01:18,720 Speaker 2: busy these days. 18 00:01:19,200 --> 00:01:21,240 Speaker 1: Well, and you picked a field that is growing and 19 00:01:21,280 --> 00:01:24,640 Speaker 1: evolving so rapidly. It must be fascinating. You just get 20 00:01:24,680 --> 00:01:27,080 Speaker 1: up every day and figure out what's going on. 21 00:01:27,560 --> 00:01:29,640 Speaker 2: You know, I've been working in the area of economics 22 00:01:29,640 --> 00:01:34,880 Speaker 2: of innovation for a quarter century, and certainly this feels 23 00:01:34,959 --> 00:01:37,400 Speaker 2: like nothing I've felt before in terms of the pace 24 00:01:37,560 --> 00:01:39,080 Speaker 2: at which all this is moving. 25 00:01:39,319 --> 00:01:42,759 Speaker 1: There's sort of an s curve of technological change where 26 00:01:42,800 --> 00:01:46,760 Speaker 1: you start really slowly and then suddenly you shoot straight up. 27 00:01:46,959 --> 00:01:48,840 Speaker 1: Do you feel like we're at the early stages of 28 00:01:48,880 --> 00:01:50,920 Speaker 1: that kind of sudden acceleration. 29 00:01:51,360 --> 00:01:53,680 Speaker 2: Yes, that's the perfect way to describe it. It feels 30 00:01:53,680 --> 00:01:56,440 Speaker 2: like the slope is changing. I've been working in this 31 00:01:56,520 --> 00:01:59,440 Speaker 2: area for about a decade regards to machine intelligence in particular, 32 00:02:00,040 --> 00:02:03,080 Speaker 2: started to focus on it in twenty twelve, and in 33 00:02:03,120 --> 00:02:06,080 Speaker 2: the last twenty four months the slope has been changing. 34 00:02:06,680 --> 00:02:10,320 Speaker 1: What happened in twenty seventeen there was sort of a 35 00:02:10,360 --> 00:02:12,440 Speaker 1: game changer in the world of AI. 36 00:02:13,520 --> 00:02:17,280 Speaker 2: Well, a number of things happened, in particular the use 37 00:02:17,320 --> 00:02:23,200 Speaker 2: of particular technique some people call transformers, and in addition 38 00:02:23,960 --> 00:02:28,720 Speaker 2: that started to demonstrate, for is the value of compute 39 00:02:29,160 --> 00:02:33,519 Speaker 2: for machine intelligence. So I think in terms of the timeline. 40 00:02:33,680 --> 00:02:36,440 Speaker 2: So in twenty fifteen, Elon Musk recruited one of my colleagues, 41 00:02:37,080 --> 00:02:40,480 Speaker 2: a fellow named Iliah Suskiverer, to co found with him 42 00:02:40,720 --> 00:02:45,800 Speaker 2: open Ai, and then in twenty eighteen circulated around the 43 00:02:45,840 --> 00:02:50,640 Speaker 2: first draft of GPT for comments and feedback, and then 44 00:02:50,680 --> 00:02:55,320 Speaker 2: twenty nineteen the second version of GPT for more feedback, 45 00:02:56,040 --> 00:02:59,960 Speaker 2: and then later that year Microsoft made their for their 46 00:03:00,000 --> 00:03:03,680 Speaker 2: billion dollar investment in open Ai, which was quite astonishing 47 00:03:03,680 --> 00:03:05,840 Speaker 2: because of that time Opening I was still very small company, 48 00:03:05,919 --> 00:03:07,800 Speaker 2: so it was a big investment for such a small team. 49 00:03:08,560 --> 00:03:11,440 Speaker 2: And then in twenty twenty two when they released in 50 00:03:11,480 --> 00:03:15,680 Speaker 2: November twenty tweeny two, they released chat GPT that seemed 51 00:03:15,680 --> 00:03:20,200 Speaker 2: to really set off outside the computer science community. That 52 00:03:20,320 --> 00:03:24,480 Speaker 2: set off a real shift in people's perceptions. And I 53 00:03:24,520 --> 00:03:29,280 Speaker 2: recently saw a graph of the number of times jenerative 54 00:03:29,320 --> 00:03:32,120 Speaker 2: AI has been mentioned on corporate earnings calls, and it's 55 00:03:32,280 --> 00:03:35,320 Speaker 2: exactly what you describe in your s curve. It was 56 00:03:35,480 --> 00:03:40,400 Speaker 2: almost nothing, almost nothing, and then November twenty two and 57 00:03:40,680 --> 00:03:42,160 Speaker 2: all of a sudden it started to take off, and 58 00:03:42,200 --> 00:03:44,640 Speaker 2: then through twenty three it just went through the roof. 59 00:03:45,120 --> 00:03:46,840 Speaker 1: One of the things that seems to be evolving on 60 00:03:46,960 --> 00:03:50,720 Speaker 1: all of this is the concept of scale in a 61 00:03:50,760 --> 00:03:53,000 Speaker 1: way that would not have been trused, say fifteen years ago. 62 00:03:53,400 --> 00:03:56,200 Speaker 1: What does that mean and why is it so important. 63 00:03:56,560 --> 00:03:59,400 Speaker 2: So usually when people refer to scale, they're talking about 64 00:03:59,400 --> 00:04:02,800 Speaker 2: the size of their AI model, their neural network. They'll 65 00:04:02,880 --> 00:04:05,240 Speaker 2: use something like, for example, the number of parameters in 66 00:04:05,280 --> 00:04:06,720 Speaker 2: the model. So you can think of that as just 67 00:04:06,760 --> 00:04:10,360 Speaker 2: the size sort of picture in your mind a neural network, 68 00:04:10,440 --> 00:04:13,200 Speaker 2: or even a tree, a tree with branches, and you 69 00:04:13,200 --> 00:04:15,480 Speaker 2: know it's got some big branches in each big branches, 70 00:04:15,520 --> 00:04:17,760 Speaker 2: smaller branches, and those branches of smaller branches, and so 71 00:04:17,760 --> 00:04:21,120 Speaker 2: it just you know, branches out, and so as you 72 00:04:21,200 --> 00:04:23,480 Speaker 2: make the model bigger, it's got more and more branches 73 00:04:23,600 --> 00:04:28,840 Speaker 2: and sub branches that the amount of computation required in 74 00:04:28,960 --> 00:04:32,919 Speaker 2: order to what they call train the model increases. So 75 00:04:33,040 --> 00:04:36,359 Speaker 2: to give you just a sense of scale, the language 76 00:04:36,360 --> 00:04:39,760 Speaker 2: models produced by open AI, the chat Gypt the first 77 00:04:39,839 --> 00:04:45,279 Speaker 2: version had about one hundred million parameters, the second version 78 00:04:46,200 --> 00:04:49,120 Speaker 2: had one point five billion, so fifteen times as much, 79 00:04:50,200 --> 00:04:53,640 Speaker 2: and then the next version had something like seven hundred 80 00:04:53,680 --> 00:04:58,279 Speaker 2: billion parameters, so from one point five to seven hundred billion. 81 00:04:59,040 --> 00:05:02,240 Speaker 2: You can sort of see just the huge jumps. And 82 00:05:02,240 --> 00:05:07,000 Speaker 2: when they demonstrated what the capability was at that scale, 83 00:05:07,080 --> 00:05:11,480 Speaker 2: it went from kind of a neat toy that computer 84 00:05:11,520 --> 00:05:13,560 Speaker 2: scientists would show each other. You know, oh, look you 85 00:05:13,600 --> 00:05:16,080 Speaker 2: can type in a prompt and a machine generates a response. 86 00:05:16,520 --> 00:05:20,360 Speaker 2: But it was not usable in a professional setting. To 87 00:05:20,720 --> 00:05:23,719 Speaker 2: once they scaled it up and they showed the same 88 00:05:23,839 --> 00:05:27,120 Speaker 2: kind of model, but at a much larger scale, could 89 00:05:27,240 --> 00:05:30,280 Speaker 2: now generate text that was indistinguishable from human So it 90 00:05:30,320 --> 00:05:33,960 Speaker 2: went from not usable to usable, and at that moment 91 00:05:34,920 --> 00:05:37,839 Speaker 2: everybody took notice and said, okay, this has now become 92 00:05:37,839 --> 00:05:43,480 Speaker 2: an exercise in scaling, and that meant buying access to compute, 93 00:05:43,520 --> 00:05:47,000 Speaker 2: and all of a sudden, Nvidia and Taiwan moved onto 94 00:05:47,000 --> 00:05:49,480 Speaker 2: everyone's radar as mission critical. 95 00:05:49,839 --> 00:05:53,160 Speaker 1: In a sense, it's like a catalytic moment in chemistry 96 00:05:53,440 --> 00:05:56,080 Speaker 1: where all of a sudden, the things that had not 97 00:05:56,320 --> 00:05:59,840 Speaker 1: quite blended together. I mean, it's just an accurate kind 98 00:05:59,839 --> 00:06:00,880 Speaker 1: of parallel. 99 00:06:01,320 --> 00:06:03,719 Speaker 2: Yes, it's exactly right, Newt. We've had a few of 100 00:06:03,760 --> 00:06:07,760 Speaker 2: these catalytic moments, I think that in this current wave 101 00:06:07,800 --> 00:06:09,960 Speaker 2: that we're in. The first one happened in twenty twelve, 102 00:06:10,400 --> 00:06:13,520 Speaker 2: where a professor at Stanford called Faye A Lee. She 103 00:06:13,760 --> 00:06:16,000 Speaker 2: was making a very big bet on her career that 104 00:06:16,120 --> 00:06:19,840 Speaker 2: labeled data would be very important for creating machine intelligence. 105 00:06:19,880 --> 00:06:22,320 Speaker 2: But she hadn't been able to prove it. And meanwhile 106 00:06:22,360 --> 00:06:25,240 Speaker 2: in Toronto, as a professor named Jeff Hidden had been 107 00:06:25,640 --> 00:06:27,800 Speaker 2: building an algorithm type but he hasn't been able to 108 00:06:27,880 --> 00:06:31,760 Speaker 2: prove that his algorithm was really as powerful. And then 109 00:06:31,800 --> 00:06:34,600 Speaker 2: in twenty twelve, some of his students brought his algorithm 110 00:06:34,880 --> 00:06:38,039 Speaker 2: to fay fe Lee's competition at Stanford, and that was 111 00:06:38,080 --> 00:06:40,240 Speaker 2: the first catalytic moment. It was bringing those two things, 112 00:06:40,240 --> 00:06:44,960 Speaker 2: showing that algorithm now called deep learning with her large 113 00:06:45,200 --> 00:06:47,559 Speaker 2: set of what they call labeled data that was putting 114 00:06:47,640 --> 00:06:49,920 Speaker 2: labels on pictures like this is a door, this is 115 00:06:49,960 --> 00:06:52,560 Speaker 2: a cat, this is a horse, and bringing those two 116 00:06:52,600 --> 00:06:55,240 Speaker 2: things together that was a catalytic moment. And then the 117 00:06:55,320 --> 00:06:59,279 Speaker 2: chat GIBT has been another catalytic moment. And now, as 118 00:06:59,320 --> 00:07:01,599 Speaker 2: we all read in the papers every morning, how much 119 00:07:01,640 --> 00:07:05,000 Speaker 2: capital people are pouring in in order to race to 120 00:07:05,000 --> 00:07:07,080 Speaker 2: get compute to scale their models. 121 00:07:07,640 --> 00:07:09,080 Speaker 3: What does deep learning mean? 122 00:07:09,880 --> 00:07:14,680 Speaker 2: Deep learning is a type of computational approach and it's 123 00:07:14,720 --> 00:07:16,480 Speaker 2: the basis of a lot of what today we call 124 00:07:16,600 --> 00:07:20,320 Speaker 2: artificial intelligence. And probably the most important thing for your 125 00:07:20,360 --> 00:07:24,640 Speaker 2: listeners to consider is when they're hearing about AI or 126 00:07:24,680 --> 00:07:27,240 Speaker 2: they try something like chat GBT and it seems magical. 127 00:07:28,120 --> 00:07:32,880 Speaker 2: There's no ghost in the machine. It's all computational statistics 128 00:07:33,000 --> 00:07:36,080 Speaker 2: that does prediction. So deep learning is a type of 129 00:07:36,120 --> 00:07:40,480 Speaker 2: computational statistics that does prediction. And almost all the AI 130 00:07:40,600 --> 00:07:45,400 Speaker 2: that we talk about today is some form of computential 131 00:07:45,440 --> 00:07:48,000 Speaker 2: statistics doing predictions. So, for example, if you see AI, 132 00:07:48,800 --> 00:07:52,120 Speaker 2: that's let's say, like the language chat GBT, that is 133 00:07:52,280 --> 00:07:55,280 Speaker 2: a model where you give it a prompt and then 134 00:07:55,320 --> 00:07:58,640 Speaker 2: it predicts. It is a prediction machine. It predicts what's 135 00:07:58,720 --> 00:08:03,360 Speaker 2: the most plausible sequence of words to follow your prompt. 136 00:08:04,040 --> 00:08:07,640 Speaker 2: If you hear about AIS being used at banks, for example, 137 00:08:07,680 --> 00:08:11,880 Speaker 2: fraud detection, that's AI being used as computational statistics that 138 00:08:12,000 --> 00:08:14,560 Speaker 2: predicts when a transaction is going through, whether it's a 139 00:08:14,640 --> 00:08:17,640 Speaker 2: legitimate transaction or fraudulent. If you hear about AI and 140 00:08:17,720 --> 00:08:21,600 Speaker 2: medicine doing diagnostics like early detection of breast cancer, early 141 00:08:21,640 --> 00:08:25,920 Speaker 2: detection of Alzheimer's, the AIS reading the medical records and 142 00:08:25,920 --> 00:08:28,840 Speaker 2: reading the medical images and then making a prediction of 143 00:08:28,880 --> 00:08:31,640 Speaker 2: the likelihood that, for example, someone's got early on set 144 00:08:31,640 --> 00:08:35,160 Speaker 2: of breast cancer. All of these applications of AI, every 145 00:08:35,160 --> 00:08:37,400 Speaker 2: one you hear of, many of them are using some 146 00:08:37,559 --> 00:08:41,400 Speaker 2: form of deep learning, and that is effectively doing statistics 147 00:08:41,440 --> 00:08:42,479 Speaker 2: to generate predictions. 148 00:08:43,000 --> 00:08:45,720 Speaker 1: And Vidia suddenly showed up as a really big player. 149 00:08:46,280 --> 00:08:48,200 Speaker 1: What is it that they did and why is that 150 00:08:48,360 --> 00:08:50,679 Speaker 1: central to the next phase of AI. 151 00:08:51,679 --> 00:08:56,400 Speaker 2: In Vidia has produced chips that are processors that are 152 00:08:56,440 --> 00:09:01,160 Speaker 2: particularly well suited for this type of machine learning, deep learning, 153 00:09:01,200 --> 00:09:06,040 Speaker 2: these neural net type of calculations. There's two main approaches 154 00:09:06,640 --> 00:09:09,319 Speaker 2: or things people have to do when they're building AI systems. 155 00:09:09,440 --> 00:09:11,920 Speaker 2: One is they have to train their AI model and 156 00:09:12,000 --> 00:09:14,280 Speaker 2: that requires a lot of compute to train the model, 157 00:09:14,559 --> 00:09:17,000 Speaker 2: and then when you're running it what they call inference 158 00:09:17,480 --> 00:09:21,120 Speaker 2: to actually make its predictions. Then that's a second category 159 00:09:21,160 --> 00:09:21,680 Speaker 2: of compute. 160 00:09:22,240 --> 00:09:22,680 Speaker 3: And so in. 161 00:09:22,760 --> 00:09:27,160 Speaker 2: Vidia's chips are particularly well designed for doing these types 162 00:09:27,200 --> 00:09:28,240 Speaker 2: of calculations. 163 00:09:28,679 --> 00:09:31,840 Speaker 1: And in Nvidia chip, which is a much more powerful 164 00:09:31,920 --> 00:09:35,880 Speaker 1: chip as I understand it, then the traditional chips that 165 00:09:35,960 --> 00:09:39,840 Speaker 1: is now has replaced an effect for AI. It actually 166 00:09:39,880 --> 00:09:44,120 Speaker 1: allows you to have an embedded algorithm in the chip itself. 167 00:09:45,120 --> 00:09:45,880 Speaker 1: Is that accurate. 168 00:09:46,559 --> 00:09:51,440 Speaker 2: Different applications of AI are using these differently. The most 169 00:09:51,520 --> 00:09:54,160 Speaker 2: common applications that you're hearing about are people that are 170 00:09:54,480 --> 00:09:58,679 Speaker 2: using in Vidia chips in data centers, and so they 171 00:09:58,720 --> 00:10:02,160 Speaker 2: are setting up their model in someone's data center, and 172 00:10:02,200 --> 00:10:05,440 Speaker 2: that thing is running very often on Nvidia chips, so 173 00:10:05,480 --> 00:10:07,920 Speaker 2: they might be running it either on their own premise 174 00:10:08,040 --> 00:10:11,600 Speaker 2: or for example, on Google Cloud or on Amazon Web 175 00:10:11,640 --> 00:10:17,120 Speaker 2: Service or Microsoft's Cloud, and very often those are using 176 00:10:17,400 --> 00:10:18,680 Speaker 2: in Vidio chips. 177 00:10:35,000 --> 00:10:37,840 Speaker 1: I have a sense that one of the choke points 178 00:10:37,840 --> 00:10:40,720 Speaker 1: in the next phase of AI may turn out to 179 00:10:40,760 --> 00:10:46,000 Speaker 1: be electricity. That the sheer volume of electricity that this 180 00:10:46,120 --> 00:10:49,760 Speaker 1: system takes is astonishing and getting bigger. Your point about 181 00:10:49,760 --> 00:10:51,960 Speaker 1: scale I need as we get better and better and 182 00:10:52,000 --> 00:10:54,880 Speaker 1: do more and more complex things, as I understand that 183 00:10:54,960 --> 00:10:57,920 Speaker 1: we're going to use more and more power in your judgment, 184 00:10:57,960 --> 00:11:00,240 Speaker 1: looking at the way that things evolving, Where's that's that 185 00:11:00,240 --> 00:11:01,360 Speaker 1: power going to come from? 186 00:11:02,400 --> 00:11:06,080 Speaker 2: Some people were speculating about this five ten years ago, 187 00:11:06,120 --> 00:11:08,280 Speaker 2: but I don't think anybody expected that we were going 188 00:11:08,360 --> 00:11:12,000 Speaker 2: to start arriving at this power topic as quickly as 189 00:11:12,000 --> 00:11:15,640 Speaker 2: we have. So you're right, especially when that catalytic moment 190 00:11:15,640 --> 00:11:18,240 Speaker 2: that you refer to when everybody saw chat ABT for 191 00:11:18,280 --> 00:11:21,520 Speaker 2: the first time in November of twenty twenty two, and 192 00:11:21,559 --> 00:11:26,440 Speaker 2: that revealed what the benefits of scaling. In the short term, 193 00:11:27,080 --> 00:11:31,160 Speaker 2: the bottleneck is an electricity. It's compute. It's getting access 194 00:11:31,160 --> 00:11:34,760 Speaker 2: to these chips, and the demand for the chips shot 195 00:11:34,840 --> 00:11:37,240 Speaker 2: up once people saw what you could do with them. 196 00:11:37,880 --> 00:11:41,199 Speaker 2: At some point, as more and more supply comes online, 197 00:11:41,240 --> 00:11:44,040 Speaker 2: and you know, we've got lots of policies now to 198 00:11:44,160 --> 00:11:48,040 Speaker 2: create further incentives to create capacity on shore and the 199 00:11:48,080 --> 00:11:52,880 Speaker 2: Chips Act and so on that eventually the supply of 200 00:11:53,360 --> 00:11:56,559 Speaker 2: compute will start to catch up with the demand. And 201 00:11:56,600 --> 00:12:00,959 Speaker 2: as that supply increases, the cost of the chips will 202 00:12:01,000 --> 00:12:03,920 Speaker 2: go down. We'll use more and more compute, and therefore 203 00:12:03,960 --> 00:12:08,040 Speaker 2: the final stopping point is power, like you say, and 204 00:12:08,120 --> 00:12:12,600 Speaker 2: the demand for energy is already meaningful, but it will 205 00:12:12,640 --> 00:12:16,040 Speaker 2: become much much bigger as the cost of compute goes down, 206 00:12:16,040 --> 00:12:18,560 Speaker 2: because right now what's limiting the use of electricity is 207 00:12:18,600 --> 00:12:23,360 Speaker 2: just the compute so expensive. As compute becomes cheaper, then 208 00:12:23,600 --> 00:12:26,840 Speaker 2: we demand more and more power to light up these chips. 209 00:12:27,480 --> 00:12:29,360 Speaker 2: And so when you ask where's it going to come from, 210 00:12:29,679 --> 00:12:32,240 Speaker 2: that's a great question for people who are trying to 211 00:12:32,280 --> 00:12:34,800 Speaker 2: figure out what does the evolution of the energy market 212 00:12:34,800 --> 00:12:38,120 Speaker 2: look like. Obviously, we're making all kinds of investments in 213 00:12:38,240 --> 00:12:43,280 Speaker 2: trying to create domestic sources of power, and for national 214 00:12:43,280 --> 00:12:46,800 Speaker 2: security reasons as well as for cost reasons and the 215 00:12:47,000 --> 00:12:50,280 Speaker 2: urgency of that. I think NEWT is much greater today 216 00:12:50,320 --> 00:12:53,160 Speaker 2: than it was even twenty four months ago, because we 217 00:12:53,280 --> 00:12:55,760 Speaker 2: never expected the demand for power to shoot up as 218 00:12:55,760 --> 00:12:56,480 Speaker 2: fast as it has. 219 00:12:57,040 --> 00:12:59,000 Speaker 3: Two it just for a minute. What is a data wall? 220 00:13:00,160 --> 00:13:02,319 Speaker 1: Is there a moment in time where the sheer volume 221 00:13:02,360 --> 00:13:05,240 Speaker 1: of information is so massive that we sort of run 222 00:13:05,280 --> 00:13:08,400 Speaker 1: into a wall because we just literally can't handle anymore. 223 00:13:08,800 --> 00:13:11,280 Speaker 2: There are concerns about hitting a wall. It's not so 224 00:13:11,360 --> 00:13:14,080 Speaker 2: much that we can't handle any more data, it's that 225 00:13:14,120 --> 00:13:16,680 Speaker 2: we've run out of the data. In other words, these 226 00:13:16,880 --> 00:13:21,440 Speaker 2: models have ingested everything on the Internet and there's nothing 227 00:13:21,559 --> 00:13:25,520 Speaker 2: left for them to read. People talk about running out 228 00:13:25,520 --> 00:13:28,920 Speaker 2: of kind of high quality data and moving to more 229 00:13:28,920 --> 00:13:34,120 Speaker 2: and more marginal quality data. I'm not convinced that that 230 00:13:34,640 --> 00:13:38,439 Speaker 2: is a meaningful wall. And the reason is most of 231 00:13:38,440 --> 00:13:42,160 Speaker 2: the people saying that are people who have largely come 232 00:13:42,200 --> 00:13:46,319 Speaker 2: from the community working on what's called NLP natural language processing, 233 00:13:47,000 --> 00:13:49,160 Speaker 2: and what they're focused on is they've run out of 234 00:13:49,200 --> 00:13:53,160 Speaker 2: words on the Internet. But there are all kinds of 235 00:13:53,160 --> 00:13:56,719 Speaker 2: other data that we are just beginning to scratch the 236 00:13:56,800 --> 00:13:59,800 Speaker 2: surface on One key example of that that I think 237 00:13:59,880 --> 00:14:04,559 Speaker 2: is particularly promising is real world data. So in other words, 238 00:14:05,200 --> 00:14:07,160 Speaker 2: there's the data on the Internet that's you know, words 239 00:14:07,160 --> 00:14:11,200 Speaker 2: in digital format, but there's all the data of how 240 00:14:11,240 --> 00:14:14,640 Speaker 2: we interact with the world, and so far we are 241 00:14:14,800 --> 00:14:18,480 Speaker 2: just beginning to collect and use that through robots, like 242 00:14:18,559 --> 00:14:22,000 Speaker 2: robots that touch things, that pick up, things that throw, 243 00:14:22,080 --> 00:14:24,200 Speaker 2: things that read, things that speak to people and get 244 00:14:24,240 --> 00:14:27,760 Speaker 2: reactions from them. All of that is a whole new 245 00:14:27,800 --> 00:14:31,600 Speaker 2: frontier of data in order to train models for the 246 00:14:31,640 --> 00:14:35,160 Speaker 2: real world. Probably the largest collection of real world data 247 00:14:35,200 --> 00:14:38,160 Speaker 2: is owned by Tesla with all their cars. Think of that. 248 00:14:38,160 --> 00:14:41,600 Speaker 2: That's just one company with one modality of sensors on 249 00:14:41,640 --> 00:14:44,960 Speaker 2: a car. And there's so much more of that type 250 00:14:44,960 --> 00:14:47,160 Speaker 2: of data and we've only begun to collect it. So 251 00:14:47,760 --> 00:14:50,240 Speaker 2: I know that the language people have been starting to 252 00:14:50,760 --> 00:14:53,280 Speaker 2: raise alarm bells that they're running out of words, but 253 00:14:53,320 --> 00:14:55,640 Speaker 2: I think words are just the beginning, and there's a 254 00:14:55,680 --> 00:14:58,560 Speaker 2: lot more data that we're really just beginning to explore. 255 00:14:59,000 --> 00:15:02,720 Speaker 1: What's your sense of about the impact of AI on 256 00:15:02,840 --> 00:15:05,800 Speaker 1: society over the next ten or fifteen years, I. 257 00:15:05,760 --> 00:15:08,600 Speaker 2: Think it's going to have a very significant impact because 258 00:15:08,640 --> 00:15:11,920 Speaker 2: the capability is so foundational. Think of all the things 259 00:15:11,960 --> 00:15:15,040 Speaker 2: that we currently do with computers. If somebody calls you 260 00:15:15,080 --> 00:15:17,640 Speaker 2: to ask you something, they might ask you something, you'll 261 00:15:17,680 --> 00:15:20,000 Speaker 2: interpret it with your brain, and then you might go 262 00:15:20,040 --> 00:15:21,400 Speaker 2: look something up in a database. 263 00:15:21,800 --> 00:15:22,240 Speaker 3: You'll go and. 264 00:15:22,320 --> 00:15:25,360 Speaker 2: You'll use computers for specific things where that are well 265 00:15:25,400 --> 00:15:28,560 Speaker 2: structured numerical or well structured data in a database. And 266 00:15:28,600 --> 00:15:31,440 Speaker 2: then every time you get to something unstructured, you know, 267 00:15:31,440 --> 00:15:34,240 Speaker 2: a person will do it. And so our workflow has 268 00:15:34,320 --> 00:15:37,040 Speaker 2: like people, people and then access of machine for a 269 00:15:37,120 --> 00:15:39,440 Speaker 2: database or a lookups seart something, and then back to 270 00:15:39,440 --> 00:15:44,120 Speaker 2: people people. Now that machines can do language, it's really 271 00:15:44,120 --> 00:15:46,680 Speaker 2: a very significant unlock to do so many things. The 272 00:15:46,720 --> 00:15:49,080 Speaker 2: way I'm thinking about it, I think of three things. 273 00:15:49,160 --> 00:15:52,240 Speaker 2: First off is productivity. That I'm meeting now with a 274 00:15:52,320 --> 00:15:56,240 Speaker 2: number of boards of directors and senior leadership teams, and 275 00:15:56,480 --> 00:16:00,800 Speaker 2: where I see people converging is in you know, CEO 276 00:16:00,880 --> 00:16:04,360 Speaker 2: will ask each of their direct reports to come back 277 00:16:04,400 --> 00:16:07,040 Speaker 2: to them with a plan for how they're going to 278 00:16:07,120 --> 00:16:10,040 Speaker 2: improve productivity in their area by twenty percent over the 279 00:16:10,040 --> 00:16:12,480 Speaker 2: next two years. So twenty percent seems to be a 280 00:16:12,560 --> 00:16:15,320 Speaker 2: number that many are converging on. And that's for everybody, 281 00:16:15,520 --> 00:16:19,280 Speaker 2: every direct report in a company to the CEO. So 282 00:16:19,480 --> 00:16:21,640 Speaker 2: some of them run a business, so they'll have a 283 00:16:21,680 --> 00:16:24,200 Speaker 2: p and L, a profit and loss, and their job 284 00:16:24,280 --> 00:16:25,880 Speaker 2: is to come back with how are they going to 285 00:16:25,920 --> 00:16:29,280 Speaker 2: increase productivity by twenty percent in two years using machine 286 00:16:29,320 --> 00:16:32,680 Speaker 2: intelligence and then others what they call functional areas, So 287 00:16:32,840 --> 00:16:35,040 Speaker 2: not a business unit that has a profit and loss, 288 00:16:35,080 --> 00:16:38,720 Speaker 2: but let's say finance or marketing or HR. They are 289 00:16:38,760 --> 00:16:41,160 Speaker 2: also to come back with a plan for how they're 290 00:16:41,200 --> 00:16:45,240 Speaker 2: going to improve their productivity by twenty percent. So twenty 291 00:16:45,240 --> 00:16:48,600 Speaker 2: percent productivity means either you can increase the numerator like 292 00:16:48,640 --> 00:16:51,800 Speaker 2: the output so let's say that's the sales or something 293 00:16:52,200 --> 00:16:55,160 Speaker 2: by twenty percent, or the denominator like taking out twenty 294 00:16:55,160 --> 00:16:58,200 Speaker 2: percent of the costs. So that's the first thing, is productivity. 295 00:16:58,280 --> 00:17:01,880 Speaker 2: It seems like right now, over the past years, companies 296 00:17:01,920 --> 00:17:05,600 Speaker 2: have been largely running pilots. Now they are starting to 297 00:17:05,600 --> 00:17:08,359 Speaker 2: move them from pilots into production. So my guess is 298 00:17:08,640 --> 00:17:10,880 Speaker 2: over the next two to three years we'll start seeing 299 00:17:10,920 --> 00:17:16,000 Speaker 2: the first companies have very significant productivity gains. And when 300 00:17:16,040 --> 00:17:19,920 Speaker 2: that happens, I suspect that will trigger the second thing, 301 00:17:20,040 --> 00:17:22,560 Speaker 2: which I think is be very significant, which is I 302 00:17:22,600 --> 00:17:26,440 Speaker 2: wouldn't be surprised if we see over the next decade 303 00:17:26,520 --> 00:17:30,680 Speaker 2: or two the largest reallocation of capital since the Second 304 00:17:30,720 --> 00:17:34,040 Speaker 2: World War. And the reason that that will happen is 305 00:17:34,080 --> 00:17:38,440 Speaker 2: because there will be leaders and laggards in adopting AI. 306 00:17:39,080 --> 00:17:40,960 Speaker 2: You know, everyone right now is focused on the AI 307 00:17:41,080 --> 00:17:44,520 Speaker 2: companies themselves, like the Googles and Microsoft's and metas and 308 00:17:44,560 --> 00:17:44,840 Speaker 2: so on. 309 00:17:45,640 --> 00:17:46,560 Speaker 3: But you know, the. 310 00:17:46,600 --> 00:17:49,280 Speaker 2: Real impact is not going to be them. It's going 311 00:17:49,320 --> 00:17:52,520 Speaker 2: to be all of the companies that are enabled, and 312 00:17:52,600 --> 00:17:55,160 Speaker 2: so whether they're energy companies or car companies, or drug 313 00:17:55,160 --> 00:17:58,800 Speaker 2: discovery companies, or retailers or banks, insurance companies, whatever they are, 314 00:17:59,520 --> 00:18:02,000 Speaker 2: and they'll be leaders and laggards. In the beginning, the 315 00:18:02,080 --> 00:18:04,320 Speaker 2: leaders will just be viewed as a little more tech savvy, 316 00:18:05,000 --> 00:18:08,240 Speaker 2: but very soon it will become those that have the 317 00:18:08,280 --> 00:18:14,320 Speaker 2: twenty percent productivity lift become so much more productive that 318 00:18:14,400 --> 00:18:16,840 Speaker 2: the others simply can't keep up. And that's when we'll 319 00:18:16,840 --> 00:18:20,000 Speaker 2: see the big capital reallocation, all the capitals starting to 320 00:18:20,040 --> 00:18:23,199 Speaker 2: move to the ones that are more productive, and that 321 00:18:23,240 --> 00:18:26,120 Speaker 2: will be I think a period of very significant creative destruction. 322 00:18:27,560 --> 00:18:29,720 Speaker 2: The third thing is seeing new things. The first two 323 00:18:29,760 --> 00:18:33,119 Speaker 2: things I described are really just a reallocation of banks 324 00:18:33,200 --> 00:18:35,600 Speaker 2: doing what banks do, but more efficiently, and retailers do 325 00:18:35,680 --> 00:18:38,040 Speaker 2: what retailers do, but more efficiently. Then there will be 326 00:18:38,080 --> 00:18:40,800 Speaker 2: all the new stuff, and those will be things like 327 00:18:40,840 --> 00:18:43,640 Speaker 2: when you described all the technological revolutions that you've seen, 328 00:18:44,359 --> 00:18:47,280 Speaker 2: where we start seeing things like we've already seen some 329 00:18:47,400 --> 00:18:51,080 Speaker 2: canaries in the coal mine. An example is Uber. We've 330 00:18:51,480 --> 00:18:55,400 Speaker 2: lived through the transition of going from every city's got 331 00:18:55,400 --> 00:18:59,400 Speaker 2: a handful of taxi companies to a global company that's 332 00:18:59,440 --> 00:19:05,000 Speaker 2: really a software company that uses machine intelligence to allocate 333 00:19:05,080 --> 00:19:09,719 Speaker 2: riders and drivers and predict the best route between two locations. 334 00:19:10,200 --> 00:19:12,600 Speaker 2: In the United States, before Uber, there were about two 335 00:19:12,680 --> 00:19:16,240 Speaker 2: hundred thousand people that drove taxis and limousines. Today there's 336 00:19:16,280 --> 00:19:19,600 Speaker 2: between three and four million people that drive for Uber. Imagine, 337 00:19:19,680 --> 00:19:22,080 Speaker 2: let's say each of those people brings a twenty five 338 00:19:22,080 --> 00:19:24,480 Speaker 2: thousand dollars car into the road. Twenty five thousand dollars 339 00:19:24,520 --> 00:19:26,840 Speaker 2: and four million people. It's one hundred billion dollars of 340 00:19:26,840 --> 00:19:31,720 Speaker 2: cap X brought into the transportation system unlocked by a 341 00:19:31,800 --> 00:19:32,760 Speaker 2: navigational AI. 342 00:19:33,359 --> 00:19:36,959 Speaker 1: You talk about a concept I find fascinating, and that 343 00:19:37,080 --> 00:19:41,520 Speaker 1: is the notion that there are between times. Describe for 344 00:19:41,640 --> 00:19:43,880 Speaker 1: us what you mean by between times. 345 00:19:44,760 --> 00:19:47,640 Speaker 2: Yeah, the between times, we think is the time between 346 00:19:48,240 --> 00:19:52,959 Speaker 2: when we've all observed the power of the technology on 347 00:19:53,000 --> 00:19:57,000 Speaker 2: the one end, and yet it's not yet widely deployed 348 00:19:57,040 --> 00:19:59,480 Speaker 2: on the other side. So that would be equivalent to 349 00:19:59,480 --> 00:20:03,520 Speaker 2: the late eight hundreds. In electricity. In the late eighteen hundreds, 350 00:20:03,560 --> 00:20:06,240 Speaker 2: there was a demonstrations of electricity and everyone got to 351 00:20:06,240 --> 00:20:09,159 Speaker 2: see what it could do, but almost nothing was electrified. 352 00:20:09,880 --> 00:20:14,680 Speaker 2: Even twenty years in after the demonstration of electricity, less 353 00:20:14,720 --> 00:20:17,800 Speaker 2: than three percent of factories in the United States were electrified, 354 00:20:18,520 --> 00:20:21,919 Speaker 2: and so it took quite a bit of time for 355 00:20:22,000 --> 00:20:25,840 Speaker 2: it to move into broad use. That's the period where 356 00:20:25,920 --> 00:20:28,680 Speaker 2: now with electricity, you know, linking that to your question 357 00:20:28,760 --> 00:20:33,320 Speaker 2: on the transformation. In the case of electricity, the original 358 00:20:33,359 --> 00:20:36,640 Speaker 2: value proposition of electricity was it will reduce your input costs. 359 00:20:36,680 --> 00:20:38,479 Speaker 2: If you have oil lamps in your factory, you make 360 00:20:38,520 --> 00:20:41,040 Speaker 2: it a little cheaper. Nobody wanted to tear apout their 361 00:20:41,040 --> 00:20:43,440 Speaker 2: infrastructure in their factories just so they could shave off 362 00:20:43,440 --> 00:20:44,760 Speaker 2: a little bit of input costs. 363 00:20:45,160 --> 00:20:45,359 Speaker 3: You know. 364 00:20:45,400 --> 00:20:48,560 Speaker 2: Once a few entrepreneurs were building new factories and decided 365 00:20:48,600 --> 00:20:52,240 Speaker 2: to experiment with electricity. They realized that not only did 366 00:20:52,240 --> 00:20:55,680 Speaker 2: it reduce their input costs. For example, they didn't need 367 00:20:55,800 --> 00:20:59,879 Speaker 2: all the heavy infrastructure that was required to hold up 368 00:20:59,880 --> 00:21:02,679 Speaker 2: the steel shafts that were previously used like if you 369 00:21:02,720 --> 00:21:05,440 Speaker 2: had steam, for example, turning the steel shafts, and the 370 00:21:05,480 --> 00:21:07,560 Speaker 2: steel shafts had big wheels on them, and on those 371 00:21:07,600 --> 00:21:09,920 Speaker 2: wheels were pulleys, and the pulleys were attached to each 372 00:21:09,920 --> 00:21:13,120 Speaker 2: of the machines. That you didn't need any of that infrastructure, 373 00:21:13,640 --> 00:21:17,400 Speaker 2: and so then the construction of the factories become much 374 00:21:17,400 --> 00:21:20,600 Speaker 2: more lightweight and cheaper. And then you didn't need multi 375 00:21:20,600 --> 00:21:23,439 Speaker 2: story factories anymore because now you didn't need to have 376 00:21:23,480 --> 00:21:26,119 Speaker 2: all your machines so close to the power source, because 377 00:21:26,160 --> 00:21:28,199 Speaker 2: you could just have a cable and a motor anywhere 378 00:21:28,200 --> 00:21:32,040 Speaker 2: on the factory floor. So it allowed for single story factories. 379 00:21:32,040 --> 00:21:34,520 Speaker 2: And then once you had single story factories, you could 380 00:21:34,600 --> 00:21:37,680 Speaker 2: redesign the layout of the workflow of the materials and 381 00:21:37,720 --> 00:21:40,520 Speaker 2: the people and the machinery. So they started getting all 382 00:21:40,560 --> 00:21:45,840 Speaker 2: these additional productivity gains as they kept innovating on what 383 00:21:45,920 --> 00:21:49,000 Speaker 2: they could do that was all unlocked by moving to electricity, 384 00:21:49,640 --> 00:21:51,919 Speaker 2: and that's what led to the between times of why 385 00:21:51,960 --> 00:21:53,960 Speaker 2: there was a very slow back to your S curve, 386 00:21:54,440 --> 00:21:57,800 Speaker 2: very slow adoption in the early days of electricity, and 387 00:21:57,840 --> 00:22:01,320 Speaker 2: then as people started to realize the increasing benefits of 388 00:22:01,440 --> 00:22:05,200 Speaker 2: using electricity and how they enabled a total redesign of 389 00:22:05,240 --> 00:22:08,600 Speaker 2: the factory, that more and more adopted electricity. So we 390 00:22:08,680 --> 00:22:11,359 Speaker 2: hit the very steep part of the S curve. The 391 00:22:11,400 --> 00:22:13,320 Speaker 2: period where now is that between times we've seen the 392 00:22:13,359 --> 00:22:16,800 Speaker 2: capability of AI, we've hardly scratched the surface of its use. 393 00:22:17,320 --> 00:22:19,240 Speaker 2: We're just getting going. But when we hit that steep 394 00:22:19,280 --> 00:22:22,320 Speaker 2: part of the curve, there's likely to be a very 395 00:22:22,600 --> 00:22:24,879 Speaker 2: dramatic amount of capital reallocation. 396 00:22:43,200 --> 00:22:44,920 Speaker 1: One of the things I work on, I'd be very 397 00:22:44,920 --> 00:22:47,000 Speaker 1: curious to get your reaction. I'm trying to get the 398 00:22:47,000 --> 00:22:51,520 Speaker 1: Congress to put together a special Committee on the Application 399 00:22:51,560 --> 00:22:53,480 Speaker 1: of Artificial Intelligence to. 400 00:22:53,560 --> 00:22:54,679 Speaker 3: Rethinking the government. 401 00:22:54,960 --> 00:22:57,159 Speaker 1: Not to think about how to apply AI to the 402 00:22:57,200 --> 00:23:00,920 Speaker 1: current government, but to ask yourself, if this is really 403 00:23:00,960 --> 00:23:04,679 Speaker 1: the future, then what would the shape and nature of 404 00:23:04,720 --> 00:23:08,919 Speaker 1: an effective system be? And I tell people the Pentagon 405 00:23:09,920 --> 00:23:12,320 Speaker 1: was opened the year I was born, nineteen forty three. 406 00:23:13,280 --> 00:23:17,359 Speaker 1: It was designed so that twenty six thousand people could 407 00:23:17,400 --> 00:23:22,000 Speaker 1: manage a global war with a manual typewriter, carbon paper, 408 00:23:22,320 --> 00:23:29,199 Speaker 1: and filing cabinets. Today we have smartphones, iPads, laptops and 409 00:23:29,520 --> 00:23:32,959 Speaker 1: twenty six thousand people. There's got to be some information 410 00:23:33,119 --> 00:23:38,960 Speaker 1: exchange rate between carbon paper, manual typewriters, and filing cabinets 411 00:23:39,240 --> 00:23:42,320 Speaker 1: and our current system. I tell audiences all the time 412 00:23:42,760 --> 00:23:46,120 Speaker 1: if you reduce the pentagon to a triangle, you would 413 00:23:46,160 --> 00:23:49,080 Speaker 1: actually have a dramatically better system. So part of what 414 00:23:49,119 --> 00:23:52,920 Speaker 1: I've been thinking about is to back out from how 415 00:23:52,920 --> 00:23:57,040 Speaker 1: do we apply AI to get marginal improvements and instead 416 00:23:57,080 --> 00:24:01,880 Speaker 1: ask the question, given the emerging reality of AI, if 417 00:24:01,920 --> 00:24:04,919 Speaker 1: you didn't already have it, what would the nature of 418 00:24:04,960 --> 00:24:08,159 Speaker 1: the government be that you would design? Because I think 419 00:24:08,200 --> 00:24:09,560 Speaker 1: it would be dramatically different. 420 00:24:10,640 --> 00:24:12,880 Speaker 2: Yeah, I mean that's a great question. I think first 421 00:24:12,920 --> 00:24:16,600 Speaker 2: of all, it begs the question what is the role 422 00:24:16,680 --> 00:24:19,119 Speaker 2: of government? Because my first step would be to go 423 00:24:19,160 --> 00:24:23,960 Speaker 2: back to first principles of what is government there to provide? 424 00:24:24,240 --> 00:24:28,040 Speaker 2: I'll just take a couple of elements of government. One 425 00:24:28,080 --> 00:24:32,760 Speaker 2: of the things that government does it makes decisions that 426 00:24:33,119 --> 00:24:39,560 Speaker 2: are designed to benefit people, and those decisions always have 427 00:24:39,600 --> 00:24:42,080 Speaker 2: trade offs. In other words, any decision that benefits some 428 00:24:42,200 --> 00:24:46,400 Speaker 2: will harm others. And so how we make those trade 429 00:24:46,440 --> 00:24:51,719 Speaker 2: offs is a function of values. And so what the 430 00:24:51,720 --> 00:24:55,480 Speaker 2: AIS will be able to do that people cannot do 431 00:24:56,359 --> 00:24:59,240 Speaker 2: is they will be able to hold much more information 432 00:25:00,119 --> 00:25:01,880 Speaker 2: kind of think of it as holding it in their 433 00:25:01,920 --> 00:25:07,560 Speaker 2: minds concurrently and optimize. So there's been a number of great, 434 00:25:08,000 --> 00:25:11,800 Speaker 2: careful studies of the ability of AIS to make decisions 435 00:25:12,720 --> 00:25:16,480 Speaker 2: that are superhuman. For example, doctor decisions when somebody comes 436 00:25:16,520 --> 00:25:18,560 Speaker 2: in and they might be having a heart attack and 437 00:25:18,640 --> 00:25:20,879 Speaker 2: the doctor has to decide whether to send them for 438 00:25:20,920 --> 00:25:24,359 Speaker 2: a test, which is costly, and the research team at 439 00:25:24,400 --> 00:25:27,080 Speaker 2: University of Chicago did a great study where the AIS 440 00:25:27,200 --> 00:25:30,760 Speaker 2: learned how to make these decisions and became superhuman. And 441 00:25:30,840 --> 00:25:34,040 Speaker 2: so we're able to send more people who should have 442 00:25:34,080 --> 00:25:37,040 Speaker 2: been recommended for a test, and many of those who 443 00:25:37,119 --> 00:25:39,639 Speaker 2: ended up getting tested in retrospect didn't need the test. 444 00:25:39,960 --> 00:25:41,520 Speaker 2: The AI does not send them for a test, so 445 00:25:41,560 --> 00:25:43,920 Speaker 2: it's much more efficient. It's a much better decision maker 446 00:25:44,600 --> 00:25:48,840 Speaker 2: than docs on average for a complex decision like that. So, 447 00:25:49,600 --> 00:25:53,080 Speaker 2: on the one hand, these ais are much better at 448 00:25:53,080 --> 00:25:56,240 Speaker 2: holding information is kind of lots of disparate information in 449 00:25:56,280 --> 00:25:59,760 Speaker 2: their minds and then making predictions. What they don't have 450 00:26:00,240 --> 00:26:04,800 Speaker 2: is judgment. AIS have zero judgment. Only people have judgment, 451 00:26:05,480 --> 00:26:08,679 Speaker 2: and so judgment has to do with those trade offs, 452 00:26:09,080 --> 00:26:12,000 Speaker 2: which is, if we advocate for a policy that does X, 453 00:26:12,119 --> 00:26:14,240 Speaker 2: it's going to benefit these people but harm those people. 454 00:26:14,760 --> 00:26:18,240 Speaker 2: Which are the trades that we should make under what conditions? 455 00:26:18,920 --> 00:26:21,760 Speaker 2: And so I would imagine this government of the future 456 00:26:22,960 --> 00:26:26,480 Speaker 2: will be far more efficient. It will make much higher 457 00:26:26,600 --> 00:26:30,080 Speaker 2: quality decisions. Now whether we decide to follow them as 458 00:26:30,119 --> 00:26:33,040 Speaker 2: a separate issue, but it will force us to be 459 00:26:33,160 --> 00:26:36,240 Speaker 2: much more explicit about our value judgments are trade offs, 460 00:26:36,760 --> 00:26:39,520 Speaker 2: because that will be explicit. In other words, we'll have 461 00:26:39,600 --> 00:26:43,520 Speaker 2: to explicitly communicate that to the AI in order for 462 00:26:43,560 --> 00:26:47,280 Speaker 2: it to complete its optimization tasks. So I think it's 463 00:26:47,320 --> 00:26:50,679 Speaker 2: going to make things far more transparent than they are today. 464 00:26:50,800 --> 00:26:52,680 Speaker 2: In order to benefit from the power of the AI, 465 00:26:52,760 --> 00:26:54,240 Speaker 2: will have to be much more transparent. 466 00:26:54,640 --> 00:26:57,359 Speaker 1: I hadn't thought about this WAYVID. You're really saying that 467 00:26:58,119 --> 00:27:01,840 Speaker 1: to effectively interact with them, you actually have to improve 468 00:27:01,880 --> 00:27:05,560 Speaker 1: your own understanding of what you're asking AI to do. 469 00:27:06,960 --> 00:27:11,280 Speaker 1: It actually enhances the requirement for human clarity as a 470 00:27:11,320 --> 00:27:13,200 Speaker 1: prelude to getting a machine clarity. 471 00:27:13,840 --> 00:27:17,880 Speaker 2: That's exactly right. One of the scenarios that people often describe. 472 00:27:17,880 --> 00:27:21,080 Speaker 2: A very common one is what they call the trolley problem, 473 00:27:21,320 --> 00:27:24,000 Speaker 2: where the cars all of a sudden spinning out of 474 00:27:24,000 --> 00:27:26,720 Speaker 2: control and the driver has to make a split decision 475 00:27:26,920 --> 00:27:29,200 Speaker 2: of whether to hit two older people on the side 476 00:27:29,200 --> 00:27:31,080 Speaker 2: of the road or hit one child on the other side. 477 00:27:31,600 --> 00:27:34,360 Speaker 2: That doesn't show up on the driving test, but when 478 00:27:34,359 --> 00:27:37,280 Speaker 2: you're put into these scenarios, you're racing along a street 479 00:27:37,320 --> 00:27:39,600 Speaker 2: and it's wet, and it's nighttime, and the light turns yellow, 480 00:27:39,680 --> 00:27:41,840 Speaker 2: and you have to decide whether to kind of step 481 00:27:41,880 --> 00:27:43,480 Speaker 2: on the gas and get through the light or try 482 00:27:43,480 --> 00:27:44,960 Speaker 2: and step on the brake, but were you might skid 483 00:27:45,000 --> 00:27:48,440 Speaker 2: into the intersection. You've got to make a split decision. Now, 484 00:27:48,920 --> 00:27:50,760 Speaker 2: no one's ever asked us what we would decide in 485 00:27:50,800 --> 00:27:53,080 Speaker 2: each of those scenarios, and you and I might be different, 486 00:27:53,800 --> 00:27:58,359 Speaker 2: But for the AI, it needs to be given guidance. 487 00:27:58,720 --> 00:28:01,119 Speaker 2: It'll be put through many scenario and it will predict 488 00:28:01,200 --> 00:28:03,639 Speaker 2: what it thinks it should do, and humans will observe 489 00:28:03,680 --> 00:28:05,560 Speaker 2: it and then either say yes or no, you should 490 00:28:05,560 --> 00:28:07,439 Speaker 2: do that or you should That means we'll have to 491 00:28:07,440 --> 00:28:11,960 Speaker 2: make it explicit. That's a small example of let's saying 492 00:28:11,960 --> 00:28:14,240 Speaker 2: the trolley problem. Each of those cases you have to 493 00:28:14,240 --> 00:28:18,680 Speaker 2: make explicit and basically imbue the AI with our values. 494 00:28:19,160 --> 00:28:21,120 Speaker 1: Which, guess me, my last big question for you, which 495 00:28:21,160 --> 00:28:25,359 Speaker 1: is you talk about that AI's core function is to 496 00:28:25,480 --> 00:28:29,639 Speaker 1: improve prediction. If I'm listening to you, you're also suggesting 497 00:28:30,240 --> 00:28:34,280 Speaker 1: that we have to improve with clarity the value judgments 498 00:28:34,320 --> 00:28:35,800 Speaker 1: we want inside that prediction. 499 00:28:36,520 --> 00:28:42,000 Speaker 2: Yes, this is going to be The practical use of 500 00:28:42,680 --> 00:28:47,960 Speaker 2: philosophers have been reasonably limited over the last number of 501 00:28:48,000 --> 00:28:51,280 Speaker 2: decades because of philosophy degree is not that practical. I 502 00:28:51,320 --> 00:28:53,880 Speaker 2: think we're going to enter a period where all of 503 00:28:53,920 --> 00:28:56,920 Speaker 2: a sudden that becomes valuable again because we're going to 504 00:28:56,960 --> 00:29:00,560 Speaker 2: have to make as a society a lot of esophical 505 00:29:01,200 --> 00:29:04,920 Speaker 2: decisions that we've already started to see the very first 506 00:29:05,040 --> 00:29:07,840 Speaker 2: glimmers of with for example, that happened with Google's Gemini 507 00:29:07,880 --> 00:29:10,480 Speaker 2: that was in the news. Of ais that are doing 508 00:29:10,520 --> 00:29:12,280 Speaker 2: things and we have to decide are they aligned with 509 00:29:12,320 --> 00:29:14,800 Speaker 2: our values. When they're just doing a few things and 510 00:29:14,840 --> 00:29:17,480 Speaker 2: they're like party tricks, it doesn't really matter. But when 511 00:29:17,520 --> 00:29:22,040 Speaker 2: we start embedding machine intelligence into our banks, deciding who 512 00:29:22,120 --> 00:29:26,280 Speaker 2: gets loans, into our insurance companies, deciding whether a claim 513 00:29:26,400 --> 00:29:29,400 Speaker 2: is legitimate and should be paid when we're putting them 514 00:29:29,400 --> 00:29:32,840 Speaker 2: into our medical systems, deciding which people should get which treatments, 515 00:29:33,560 --> 00:29:36,960 Speaker 2: all these things in every case we are going to 516 00:29:37,000 --> 00:29:39,480 Speaker 2: be confronted with. This is a decision that we are 517 00:29:39,920 --> 00:29:43,120 Speaker 2: in some sense handing to the machine because it's got 518 00:29:43,240 --> 00:29:47,520 Speaker 2: such superior prediction capabilities, but it doesn't have judgment, and 519 00:29:47,600 --> 00:29:49,960 Speaker 2: so we will now have to make our judgment explicit, 520 00:29:49,960 --> 00:29:51,880 Speaker 2: and it can't be left to the discretion of the 521 00:29:51,880 --> 00:29:54,479 Speaker 2: person because the person is not doing it, and so 522 00:29:54,640 --> 00:29:59,240 Speaker 2: it will become I think, a flourishing of debate and 523 00:29:59,320 --> 00:30:04,280 Speaker 2: discussion and bringing our judgment from the recesses of our 524 00:30:04,320 --> 00:30:07,000 Speaker 2: minds out into a public discussion. 525 00:30:07,600 --> 00:30:12,120 Speaker 1: Well, let me say this has been an exhilarating conversation. Onna, 526 00:30:12,160 --> 00:30:16,200 Speaker 1: thank you for joining me. Your contributions to artificial intelligence 527 00:30:16,240 --> 00:30:19,400 Speaker 1: and machine learning are amazing. The way you're thinking it 528 00:30:19,440 --> 00:30:23,080 Speaker 1: through is extraordinarily helpful. I want to encourage our listeners 529 00:30:23,400 --> 00:30:25,480 Speaker 1: to get a copy of your book, Power and Prediction, 530 00:30:25,960 --> 00:30:30,000 Speaker 1: The Disruptive Economics of Artificial Intelligence. It is available now 531 00:30:30,040 --> 00:30:33,960 Speaker 1: on Amazon and the bookstores everywhere, and people can learn 532 00:30:34,000 --> 00:30:38,800 Speaker 1: more at your website Agrowall dot CA. I really appreciate 533 00:30:38,840 --> 00:30:40,520 Speaker 1: you taking time to be with us. 534 00:30:41,040 --> 00:30:42,720 Speaker 2: Thank you very much for your interest in my work. 535 00:30:42,760 --> 00:30:44,000 Speaker 2: It's been a real pleasure to be here. 536 00:30:47,640 --> 00:30:51,280 Speaker 1: Thank you to my guest professor, Agrowall. You can get 537 00:30:51,320 --> 00:30:54,040 Speaker 1: a link to buy his book, Power and Prediction on 538 00:30:54,080 --> 00:30:57,560 Speaker 1: our show page at nutsworld dot com. Newtsworld is produced 539 00:30:57,560 --> 00:31:01,320 Speaker 1: by Gangrad three sixty and iHeartMedia. Our executive producer is 540 00:31:01,360 --> 00:31:05,720 Speaker 1: Guernsey Sloan. Our researcher is Rachel Peterson. The artwork for 541 00:31:05,760 --> 00:31:09,120 Speaker 1: the show was created by Steve Penley. Special thanks to 542 00:31:09,120 --> 00:31:12,400 Speaker 1: the team at Gingish three sixty. If you've been enjoying Newtsworld, 543 00:31:12,640 --> 00:31:15,120 Speaker 1: I hope you'll go to Apple Podcast and both rate 544 00:31:15,200 --> 00:31:18,040 Speaker 1: us with five stars and give us a review so 545 00:31:18,120 --> 00:31:21,400 Speaker 1: others can learn what it's all about. Right now, listeners 546 00:31:21,440 --> 00:31:25,080 Speaker 1: of Newtsworld concerner for my three free weekly columns at 547 00:31:25,120 --> 00:31:27,960 Speaker 1: gingrishfree sixty dot com slash newsletter. 548 00:31:28,360 --> 00:31:30,320 Speaker 3: I'm Newt Gingrich. This is Newtsworld.