1 00:00:04,840 --> 00:00:08,720 Speaker 1: On this episode of News World, the Chinese artificial intelligence 2 00:00:08,760 --> 00:00:12,080 Speaker 1: app called deep Seek rocketed to the top of the 3 00:00:12,119 --> 00:00:15,440 Speaker 1: Apple App Store downloads charter with the weekend after its 4 00:00:15,440 --> 00:00:19,799 Speaker 1: release last week by a Chinese startup. Deep Seek offers 5 00:00:20,120 --> 00:00:26,440 Speaker 1: similar functionality to open AI's popular chat GPT chatbot, answering 6 00:00:26,520 --> 00:00:31,000 Speaker 1: questions and generating text in response to users questions. The 7 00:00:31,040 --> 00:00:34,920 Speaker 1: immediate success of deep Seak sent the tech focused Nazdak 8 00:00:34,920 --> 00:00:38,840 Speaker 1: Composite index down by three percent on Monday, with US 9 00:00:38,960 --> 00:00:44,000 Speaker 1: chipmaker and Nvidia down almost seventeen percent. So how will 10 00:00:44,080 --> 00:00:47,080 Speaker 1: the United States compete with China and the artificial intelligence race? 11 00:00:48,080 --> 00:00:52,040 Speaker 1: Here to discuss China and artificial intelligence. I'm really pleased 12 00:00:52,080 --> 00:00:54,959 Speaker 1: to welcome my guest, Dean Ball. He is a research 13 00:00:55,040 --> 00:00:59,400 Speaker 1: fellow at George Mason University's Mercada Center and author of 14 00:00:59,440 --> 00:01:04,200 Speaker 1: the substat hyper Dimensional. His work focus as an artificial intelligence, 15 00:01:04,640 --> 00:01:21,560 Speaker 1: emerging technologies, and the future of governance. Dean, welcome and 16 00:01:21,640 --> 00:01:23,160 Speaker 1: thank you for joining me on News World. 17 00:01:23,720 --> 00:01:25,760 Speaker 2: Mister speaker, thank you so much for having me. It's 18 00:01:25,800 --> 00:01:26,679 Speaker 2: a pleasure to be here. 19 00:01:27,280 --> 00:01:30,600 Speaker 1: This is unbelievably timely, and I'm really grateful you could 20 00:01:30,600 --> 00:01:33,080 Speaker 1: get us under your schedule. Can you talk about the 21 00:01:33,160 --> 00:01:35,800 Speaker 1: history of deep seek, when it began and why it 22 00:01:35,840 --> 00:01:38,280 Speaker 1: took the world by storm this last weekend? 23 00:01:38,959 --> 00:01:39,920 Speaker 3: Absolutely? Yes. 24 00:01:40,160 --> 00:01:43,800 Speaker 2: So deep Seek is actually a subsidiary of a Chinese 25 00:01:43,840 --> 00:01:47,880 Speaker 2: hedge fund called high Flyer. Deep Seek was created in 26 00:01:47,920 --> 00:01:52,480 Speaker 2: May of twenty twenty three, so it's pretty recent high Flyer, 27 00:01:52,680 --> 00:01:55,760 Speaker 2: like a lot of quantitative hedge funds, that company has 28 00:01:55,800 --> 00:01:58,320 Speaker 2: been using AI for a variety of different things for 29 00:01:58,360 --> 00:02:01,800 Speaker 2: a long time, so they were already large scale purchasers 30 00:02:02,120 --> 00:02:06,680 Speaker 2: of the high end computing hardware the GPUs by Nvidia, 31 00:02:07,160 --> 00:02:10,360 Speaker 2: that are used to train and run large scale AI models, 32 00:02:10,680 --> 00:02:14,240 Speaker 2: and so Deepseek has had a lot of very talented 33 00:02:14,480 --> 00:02:18,640 Speaker 2: researchers working there for least the last year. They've released 34 00:02:18,720 --> 00:02:22,000 Speaker 2: models and research papers that people in the machine learning 35 00:02:22,120 --> 00:02:26,120 Speaker 2: community have been paying attention to and impressed by for 36 00:02:26,160 --> 00:02:29,400 Speaker 2: a long time. I myself first wrote about deep Seek, 37 00:02:29,440 --> 00:02:32,400 Speaker 2: I think in May of twenty twenty four. It's been 38 00:02:32,600 --> 00:02:36,480 Speaker 2: noticed by AI experts for a long time, and a 39 00:02:36,520 --> 00:02:39,120 Speaker 2: couple of things happened in the last few months that 40 00:02:39,160 --> 00:02:42,880 Speaker 2: are relevant. Really the last one month, in late December, 41 00:02:43,600 --> 00:02:47,280 Speaker 2: deep Seek released a model called V three, and this 42 00:02:47,360 --> 00:02:51,960 Speaker 2: is like a language model similar to chat GPT or 43 00:02:52,040 --> 00:02:57,800 Speaker 2: Google Gemini, and it was particularly notable because the cost 44 00:02:57,919 --> 00:03:01,800 Speaker 2: of training it was quite Now, I think there's some 45 00:03:01,880 --> 00:03:04,160 Speaker 2: nuance about that figure that we can get into about 46 00:03:04,160 --> 00:03:06,120 Speaker 2: why maybe it's not as low a figure as some 47 00:03:06,160 --> 00:03:09,680 Speaker 2: people think, but certainly it was an impressive figure with 48 00:03:09,800 --> 00:03:13,200 Speaker 2: some very impressive innovations that also they were quite open about. 49 00:03:13,480 --> 00:03:16,160 Speaker 2: They shared exactly how they achieved these efficiencies, which is 50 00:03:16,160 --> 00:03:19,280 Speaker 2: not something that we see from the frontier labs in 51 00:03:19,320 --> 00:03:22,919 Speaker 2: the United States, most of them at least. And then 52 00:03:23,320 --> 00:03:26,400 Speaker 2: about a week ago they released a model called R 53 00:03:26,480 --> 00:03:29,000 Speaker 2: one and R one was built on top of V 54 00:03:29,080 --> 00:03:33,480 Speaker 2: three and it's a reasoning model. So these models, it's 55 00:03:33,480 --> 00:03:35,840 Speaker 2: a new style of model that was pioneered by open 56 00:03:35,880 --> 00:03:39,240 Speaker 2: Ai last September. Open AI's version is called OH one. 57 00:03:39,560 --> 00:03:42,560 Speaker 2: Deep seats is called R one, and what it basically 58 00:03:42,600 --> 00:03:46,560 Speaker 2: does is think about the question you've given it for 59 00:03:46,600 --> 00:03:51,040 Speaker 2: a while before it answers you. And it does this 60 00:03:51,160 --> 00:03:53,520 Speaker 2: in something called a chain of thought, where basically it's 61 00:03:53,560 --> 00:03:54,840 Speaker 2: writing notes to itself. 62 00:03:54,840 --> 00:03:56,320 Speaker 3: Almost is how you can think about or. 63 00:03:56,240 --> 00:04:00,520 Speaker 2: Sort of like writing down, Okay, how might I solve this? Oh, look, 64 00:04:00,560 --> 00:04:02,720 Speaker 2: I noticed I made a mistake there, Oh, okay, this 65 00:04:02,800 --> 00:04:03,720 Speaker 2: approach isn't working. 66 00:04:03,840 --> 00:04:06,000 Speaker 3: Let me go try something else. That kind of thing. 67 00:04:07,120 --> 00:04:12,320 Speaker 2: And this is a capability that publicly only Google, open Ai. 68 00:04:12,240 --> 00:04:13,520 Speaker 3: And now Deepseak have achieved. 69 00:04:14,200 --> 00:04:17,800 Speaker 2: So it's definitely quite impressive to achieve it, though not 70 00:04:18,279 --> 00:04:22,719 Speaker 2: particularly expensive to achieve it. And that model, and I 71 00:04:22,720 --> 00:04:26,640 Speaker 2: think the app that they released for the iPhone combined 72 00:04:27,600 --> 00:04:31,280 Speaker 2: ended up sort of creating a sensation. You know that 73 00:04:31,320 --> 00:04:36,040 Speaker 2: nobody quite predicted. I myself back in September, I predicted 74 00:04:36,040 --> 00:04:39,760 Speaker 2: that a Chinese company would replicate the one model from 75 00:04:39,760 --> 00:04:42,520 Speaker 2: open Ai within a few months. I would have never 76 00:04:42,560 --> 00:04:44,960 Speaker 2: predicted that it would have caused Nvidio stock to drop 77 00:04:45,000 --> 00:04:46,120 Speaker 2: by nearly twenty percent. 78 00:04:46,360 --> 00:04:47,680 Speaker 3: That was quite a shock to me. 79 00:04:48,080 --> 00:04:50,279 Speaker 2: It's a bit of a surprise sensation in that regard, 80 00:04:51,080 --> 00:04:52,520 Speaker 2: and that's basically how we got here. 81 00:04:53,240 --> 00:04:57,440 Speaker 1: As I understand it, the Chinese went to a very 82 00:04:57,480 --> 00:05:04,159 Speaker 1: different system architecture and somehow are able to do faster, 83 00:05:04,960 --> 00:05:10,960 Speaker 1: better algorithms with much simpler chips that are much less expensive. 84 00:05:11,800 --> 00:05:13,799 Speaker 1: I mean, is that a reasonably accurate statement. 85 00:05:14,360 --> 00:05:16,560 Speaker 3: There's a few things I would want to correct there. 86 00:05:16,839 --> 00:05:19,560 Speaker 3: The first thing is about the chips. I think that's important. 87 00:05:19,800 --> 00:05:25,040 Speaker 2: Deepseek most likely is using somewhere between ten and fifty 88 00:05:25,120 --> 00:05:29,840 Speaker 2: thousand of an Nvidia chip that is basically the same 89 00:05:29,880 --> 00:05:35,400 Speaker 2: performance as what open ai or Anthropic or Google are using. 90 00:05:36,400 --> 00:05:40,240 Speaker 2: That chip was sold into the Chinese market as a 91 00:05:40,279 --> 00:05:44,240 Speaker 2: result of a flaw in the twenty twenty two export controls, 92 00:05:44,240 --> 00:05:46,960 Speaker 2: which was the first kind of version of the Biden 93 00:05:47,000 --> 00:05:51,040 Speaker 2: administration's export controls on high end AI compute. So they 94 00:05:51,200 --> 00:05:56,200 Speaker 2: used actually chips that are pretty similar in roughly similar 95 00:05:56,320 --> 00:05:59,520 Speaker 2: quantity to what Frontier labs have. So their computing power 96 00:05:59,560 --> 00:06:01,599 Speaker 2: is signif get here. And that's one thing that a 97 00:06:01,600 --> 00:06:04,640 Speaker 2: lot of people are mistaken about. All evidence suggests that 98 00:06:04,680 --> 00:06:08,080 Speaker 2: they had significant computing resources at their disposal. And the 99 00:06:08,120 --> 00:06:10,760 Speaker 2: other thing about the sort of efficiency approach that they used. 100 00:06:11,200 --> 00:06:14,279 Speaker 2: They're still using the same basic architecture that everybody else 101 00:06:14,360 --> 00:06:17,760 Speaker 2: is using, which is called the transformer. That is a 102 00:06:17,800 --> 00:06:20,960 Speaker 2: machine learning architecture that was invented in twenty seventeen at Google, 103 00:06:22,200 --> 00:06:26,200 Speaker 2: and that architecture it's been improved many many different times 104 00:06:26,200 --> 00:06:28,600 Speaker 2: with many different people, and they're still using that same 105 00:06:28,640 --> 00:06:32,159 Speaker 2: basic architecture. They made some novel improvements to it, but 106 00:06:32,360 --> 00:06:36,799 Speaker 2: basically over the last few years, the trajectory we've seen 107 00:06:37,680 --> 00:06:41,880 Speaker 2: is that you can get about several hundred percent increase 108 00:06:41,960 --> 00:06:44,480 Speaker 2: in efficiency from making this. 109 00:06:44,600 --> 00:06:47,400 Speaker 3: Architecture and all the various algorithms that are. 110 00:06:47,240 --> 00:06:51,840 Speaker 2: Involved more efficient. So American frontier labs have been doing 111 00:06:51,839 --> 00:06:56,440 Speaker 2: those efficiency gains. The price of GPT four has dropped 112 00:06:56,440 --> 00:07:00,400 Speaker 2: something like ninety nine percent in the last fifteen months, 113 00:07:01,040 --> 00:07:03,240 Speaker 2: So I would say that in terms of the efficiency 114 00:07:03,279 --> 00:07:07,000 Speaker 2: gains that deep Seek got, really it's actually about on trend. 115 00:07:07,120 --> 00:07:09,360 Speaker 2: It's the trend that the whole machine learning, the whole 116 00:07:09,400 --> 00:07:12,680 Speaker 2: frontier of the machine learning world is on. The thing 117 00:07:12,680 --> 00:07:15,520 Speaker 2: that was surprising about it is not so much the 118 00:07:15,560 --> 00:07:19,800 Speaker 2: efficiency itself. It's a Chinese company achieving that level of efficiency, 119 00:07:19,960 --> 00:07:23,560 Speaker 2: but again you can do that through having very intelligent engineers, 120 00:07:23,640 --> 00:07:25,920 Speaker 2: and they certainly have that just like we do. 121 00:07:26,760 --> 00:07:29,600 Speaker 1: If in a way, it's not a dramatic change. Why 122 00:07:29,760 --> 00:07:31,080 Speaker 1: was in Nvidia hit so hard. 123 00:07:32,640 --> 00:07:34,040 Speaker 3: It's a really good question. 124 00:07:34,320 --> 00:07:37,760 Speaker 2: I know some people who think that part of what 125 00:07:37,920 --> 00:07:41,160 Speaker 2: was driving in Video's stock price down on Monday was 126 00:07:41,360 --> 00:07:44,560 Speaker 2: the tariffs that Trump was going to announce he was 127 00:07:44,600 --> 00:07:47,920 Speaker 2: thinking about later that day, which he did, so there's 128 00:07:47,920 --> 00:07:50,800 Speaker 2: some suggestion about that. I also think it's a good 129 00:07:50,800 --> 00:07:54,600 Speaker 2: old fashioned overreaction, though the reality is that most people 130 00:07:55,280 --> 00:07:59,040 Speaker 2: are not used to technology that moves this quickly, right, 131 00:07:59,080 --> 00:08:03,080 Speaker 2: I mean, think about this fact. Think about the algorithmic 132 00:08:03,160 --> 00:08:06,680 Speaker 2: efficiency improvements can get you four to five hundred percent 133 00:08:06,760 --> 00:08:11,120 Speaker 2: increases in efficiency per year. On top of that, the 134 00:08:11,200 --> 00:08:15,520 Speaker 2: fastest chips get something like one hundred and fifty to 135 00:08:15,560 --> 00:08:19,400 Speaker 2: two hundred percent faster every year. And on top of that, 136 00:08:20,000 --> 00:08:25,480 Speaker 2: the companies are buying these chips in vastly greater. 137 00:08:25,360 --> 00:08:26,720 Speaker 3: Quantities every year. 138 00:08:27,440 --> 00:08:33,319 Speaker 2: So the pace of progress and speed here is truly exponential, 139 00:08:34,160 --> 00:08:38,600 Speaker 2: and that is just something that I think we aren't 140 00:08:39,000 --> 00:08:45,480 Speaker 2: used to in industrial outputs. We're not used to business 141 00:08:45,480 --> 00:08:48,600 Speaker 2: moving with this kind of speed. But that's just where 142 00:08:48,640 --> 00:08:51,200 Speaker 2: we are. So I think to people that watch the 143 00:08:51,280 --> 00:08:55,000 Speaker 2: AI field closely, I think that all of this was 144 00:08:55,120 --> 00:08:57,800 Speaker 2: somewhat less of a surprise, with some exceptions. 145 00:08:58,679 --> 00:09:00,400 Speaker 3: And I think though that. 146 00:09:00,360 --> 00:09:03,800 Speaker 2: There's a lot of people on Wall Street who sort 147 00:09:03,800 --> 00:09:06,640 Speaker 2: of like the Nvidia story and are aware of AI's 148 00:09:06,679 --> 00:09:09,800 Speaker 2: potential in general, but maybe are just not aware of 149 00:09:09,840 --> 00:09:13,240 Speaker 2: how quickly things are progressing here, and so in that sense, 150 00:09:13,320 --> 00:09:15,400 Speaker 2: I think this is a bit of a wake up call, 151 00:09:16,080 --> 00:09:17,400 Speaker 2: should be a bit of a wake up call for 152 00:09:17,440 --> 00:09:20,360 Speaker 2: everybody that, yes, look at this technology, look at how 153 00:09:20,440 --> 00:09:22,800 Speaker 2: much faster it's getting. And that's a story that's going 154 00:09:22,840 --> 00:09:25,160 Speaker 2: to play out in both the United States and China. 155 00:09:25,320 --> 00:09:28,360 Speaker 1: Are in Vidia's chips made in Taiwan? Are they made 156 00:09:28,360 --> 00:09:28,920 Speaker 1: in the US? 157 00:09:29,559 --> 00:09:31,840 Speaker 2: For the most part, they're made in Taiwan. There is 158 00:09:32,040 --> 00:09:38,680 Speaker 2: the Arizona Foundry, the semiconductor manufacturing factory that TSMC, the 159 00:09:38,679 --> 00:09:43,480 Speaker 2: Taiwanese company, They built a factory in Arizona. That factory 160 00:09:43,520 --> 00:09:46,480 Speaker 2: is not quite at the very cutting edge, but it's 161 00:09:46,559 --> 00:09:49,080 Speaker 2: pretty close to the cutting edge, and they will be 162 00:09:49,160 --> 00:09:52,640 Speaker 2: producing chips. The rumor is largely their chips are going 163 00:09:52,679 --> 00:09:55,520 Speaker 2: to be for Apple and Nvidia, So there will be 164 00:09:55,559 --> 00:09:58,079 Speaker 2: some chips made in the United States, but predominantly we're 165 00:09:58,080 --> 00:09:58,800 Speaker 2: talking about. 166 00:09:58,559 --> 00:10:04,000 Speaker 1: Taiwanese announcement last week of the group coming together to 167 00:10:04,040 --> 00:10:09,000 Speaker 1: put in supposedly five hundred billion dollars in artificial intelligence. 168 00:10:09,240 --> 00:10:12,400 Speaker 1: Is that mostly pr or do you actually expect to 169 00:10:12,440 --> 00:10:16,000 Speaker 1: see an enormous shift of energy and drive into that area. 170 00:10:16,920 --> 00:10:19,520 Speaker 2: I think the five hundred billion dollar number definitely. There's 171 00:10:19,600 --> 00:10:22,720 Speaker 2: questions about how they're going to get to that level 172 00:10:22,760 --> 00:10:26,080 Speaker 2: of investment. I don't think any of the investors that 173 00:10:26,160 --> 00:10:29,480 Speaker 2: were mentioned there have five hundred billion dollars, you know, 174 00:10:29,520 --> 00:10:32,800 Speaker 2: in the bank. But I think the money's out there, 175 00:10:32,920 --> 00:10:35,960 Speaker 2: and I think the promise is there. I basically believe 176 00:10:35,960 --> 00:10:38,360 Speaker 2: the one hundred billion dollar number. And on top of that, 177 00:10:38,520 --> 00:10:42,520 Speaker 2: Meta announced sixty five billion in data center capital expenses 178 00:10:42,559 --> 00:10:45,800 Speaker 2: this year. Microsoft is going to spend eighty billion on 179 00:10:45,920 --> 00:10:49,280 Speaker 2: data centers. I would expect similar numbers from Amazon and Google. 180 00:10:50,040 --> 00:10:55,040 Speaker 2: So all told, over the next four years, it wouldn't 181 00:10:55,080 --> 00:10:58,880 Speaker 2: shock me to see well over half a trillion dollars 182 00:10:58,920 --> 00:11:01,920 Speaker 2: invested into data centers in the United States. 183 00:11:02,960 --> 00:11:05,480 Speaker 1: I'm surprised that's all. One of these charts of the 184 00:11:05,559 --> 00:11:09,800 Speaker 1: day that showed number of data centers by country, and 185 00:11:09,880 --> 00:11:13,760 Speaker 1: we are overwhelmingly dominant. Yes, is it really important or 186 00:11:13,760 --> 00:11:15,720 Speaker 1: is it just a temporary artifact. 187 00:11:16,480 --> 00:11:20,400 Speaker 2: Yes, it is definitely true that America is leading the 188 00:11:20,400 --> 00:11:24,520 Speaker 2: world at this industrial infrastructure build out of AI compute, 189 00:11:24,600 --> 00:11:26,280 Speaker 2: and I think it's going to be hugely important. 190 00:11:26,480 --> 00:11:27,920 Speaker 3: We're not used to that in America. 191 00:11:28,480 --> 00:11:31,280 Speaker 2: Chinese built all the high speed rail, they built all 192 00:11:31,280 --> 00:11:34,960 Speaker 2: this infrastructure throughout the developing world. They spend enormous amounts 193 00:11:34,960 --> 00:11:37,400 Speaker 2: of money on that over the last decade or two. 194 00:11:37,960 --> 00:11:40,959 Speaker 2: We are not used to leading on building things in 195 00:11:41,000 --> 00:11:43,400 Speaker 2: the physical world, but in this case we actually are. 196 00:11:44,559 --> 00:11:47,160 Speaker 2: And I think that's going to be hugely valuable because 197 00:11:47,720 --> 00:11:51,080 Speaker 2: despite what people are saying about the cost of this 198 00:11:51,240 --> 00:11:55,560 Speaker 2: model having been so cheap to train, these efficiency gains 199 00:11:55,640 --> 00:11:59,120 Speaker 2: that a company like deep Seek is achieving, they mean 200 00:11:59,360 --> 00:12:03,439 Speaker 2: that you can train a much better model. So if 201 00:12:03,440 --> 00:12:05,840 Speaker 2: you were going to spend a billion dollars on training 202 00:12:05,840 --> 00:12:09,200 Speaker 2: a model, let's just say, and you found something that 203 00:12:09,320 --> 00:12:14,760 Speaker 2: gets you a four hundred percent improvement, well then you 204 00:12:14,840 --> 00:12:18,199 Speaker 2: can train a much better model for that same billion dollars. 205 00:12:18,760 --> 00:12:21,160 Speaker 2: And so ultimately this is a positive story. It's a 206 00:12:21,200 --> 00:12:23,880 Speaker 2: positive story that the trajectory of AI is going to 207 00:12:23,920 --> 00:12:27,720 Speaker 2: be really rapid and it's going to have quite profound capabilities, 208 00:12:27,760 --> 00:12:32,360 Speaker 2: I think, and America will be in pole position not 209 00:12:32,520 --> 00:12:34,480 Speaker 2: just to build those models, you know, not just to 210 00:12:34,559 --> 00:12:38,320 Speaker 2: train them, but it's also very computationally expensive to run 211 00:12:38,640 --> 00:12:43,360 Speaker 2: those models as well. We will be in a great 212 00:12:43,440 --> 00:12:45,920 Speaker 2: position to be able to run those models so that 213 00:12:46,840 --> 00:12:51,240 Speaker 2: individuals and businesses all throughout our economy, like a gigantic 214 00:12:51,320 --> 00:12:54,400 Speaker 2: number of uses throughout the economy. It will be businesses 215 00:12:54,400 --> 00:12:57,760 Speaker 2: and individuals and governments that are using AI to do 216 00:12:57,760 --> 00:13:03,440 Speaker 2: all kinds of things, automate business processes, cure diseases. I 217 00:13:03,440 --> 00:13:05,360 Speaker 2: mean all kinds of things that some of which we 218 00:13:05,400 --> 00:13:08,680 Speaker 2: can't even imagine yet, and those uses are just going 219 00:13:08,760 --> 00:13:11,560 Speaker 2: to be enormously enormously valuable and will be in the 220 00:13:11,600 --> 00:13:13,520 Speaker 2: position to lead in that regard. 221 00:13:14,600 --> 00:13:19,040 Speaker 1: Does that also imply that the capabilities of artificial intelligence 222 00:13:19,720 --> 00:13:25,080 Speaker 1: will be permeating both the individuals, governments and businesses much 223 00:13:25,160 --> 00:13:26,160 Speaker 1: faster in the US? 224 00:13:27,480 --> 00:13:29,800 Speaker 2: Well, I hope so is what I would say. It 225 00:13:29,840 --> 00:13:33,839 Speaker 2: means that we'll have the industrial capacity to do that. 226 00:13:34,640 --> 00:13:38,240 Speaker 2: It doesn't necessarily mean that we will do it, because 227 00:13:39,240 --> 00:13:44,120 Speaker 2: we could make it harder for businesses to adopt AI 228 00:13:44,640 --> 00:13:45,920 Speaker 2: through bad regulation. 229 00:13:46,559 --> 00:13:50,040 Speaker 3: There are about a dozen states and counting. 230 00:13:50,480 --> 00:13:54,640 Speaker 2: That are considering laws that I think would have not 231 00:13:54,760 --> 00:13:57,600 Speaker 2: just have the effect of creating a patchwork of regulations, 232 00:13:57,640 --> 00:14:01,520 Speaker 2: which is bad enough, but I think would also particularly 233 00:14:01,559 --> 00:14:06,360 Speaker 2: impose a lot of compliance burdens, not on AI developers actually, 234 00:14:06,400 --> 00:14:09,920 Speaker 2: but on businesses trying to adopt AI. They'll have to 235 00:14:09,960 --> 00:14:13,720 Speaker 2: adopt risk management plans and do these things called algorithmic 236 00:14:13,760 --> 00:14:16,320 Speaker 2: impact assessments. And I think it could end up being 237 00:14:16,320 --> 00:14:21,560 Speaker 2: a huge compliance burden, huge liability risk, and I think 238 00:14:21,640 --> 00:14:25,200 Speaker 2: it could deter especially among larger businesses. I think it 239 00:14:25,240 --> 00:14:29,520 Speaker 2: could deter investment into AI and make our uses of 240 00:14:29,560 --> 00:14:33,560 Speaker 2: AI somewhat less creative and bold. So we will be 241 00:14:33,600 --> 00:14:36,720 Speaker 2: in the position in terms of the technological capability, and 242 00:14:36,760 --> 00:14:39,640 Speaker 2: the question is whether our policymakers will allow that to occur. 243 00:14:40,200 --> 00:14:45,840 Speaker 1: So, in a sense, the nature of AI would almost 244 00:14:45,880 --> 00:14:50,040 Speaker 1: demand federal preemption to have one particular set of rules 245 00:14:50,040 --> 00:14:50,840 Speaker 1: for the whole country. 246 00:14:51,520 --> 00:14:54,520 Speaker 2: I wish, I really wish. I've written a lot about 247 00:14:54,520 --> 00:14:59,240 Speaker 2: federal preemption. I think it's a very hard thing to 248 00:14:59,280 --> 00:15:03,360 Speaker 2: do politically, because you do need sixty votes in the Senate. 249 00:15:04,160 --> 00:15:07,240 Speaker 2: Part of the reason America succeeded with the adoption of 250 00:15:07,240 --> 00:15:11,520 Speaker 2: the Internet is because Congress led with light touch and 251 00:15:11,600 --> 00:15:16,280 Speaker 2: pro innovation laws that preempted a wide range of conceivable 252 00:15:16,400 --> 00:15:19,480 Speaker 2: state policies. And I really do think that we should 253 00:15:19,520 --> 00:15:21,480 Speaker 2: do something similar here. And I think that's not just 254 00:15:21,520 --> 00:15:24,520 Speaker 2: an economic competitiveness issue. I think it's frankly a national 255 00:15:24,600 --> 00:15:40,280 Speaker 2: security issue. 256 00:15:41,880 --> 00:15:45,040 Speaker 1: The President Trump called Deep Seek a wake up call 257 00:15:45,080 --> 00:15:47,640 Speaker 1: for US industries, and I get the sense that he 258 00:15:47,720 --> 00:15:51,840 Speaker 1: wants to compete, He said to the House Republicans, we 259 00:15:51,920 --> 00:15:55,440 Speaker 1: need to be laser focused on competing to win because 260 00:15:55,440 --> 00:15:57,640 Speaker 1: we have the greatest scientists in the world. So, in 261 00:15:57,680 --> 00:16:00,680 Speaker 1: a sense, I suspect that Trump had been may be 262 00:16:00,800 --> 00:16:04,480 Speaker 1: open to the idea of federal preemption to create one 263 00:16:04,520 --> 00:16:05,960 Speaker 1: set of rules for the whole country. 264 00:16:06,520 --> 00:16:09,240 Speaker 2: I think the administration may well be open. I think 265 00:16:09,280 --> 00:16:11,160 Speaker 2: the one thing that a lot of people can miss 266 00:16:11,320 --> 00:16:15,600 Speaker 2: with this is people call this the AI race, and 267 00:16:15,640 --> 00:16:17,680 Speaker 2: that's fine, you know. I think we should be competitive 268 00:16:17,720 --> 00:16:20,240 Speaker 2: and we should be pedal to the metal in that regard. 269 00:16:21,040 --> 00:16:25,440 Speaker 2: But there's not some finish line that we're racing too. 270 00:16:25,760 --> 00:16:28,640 Speaker 2: It's not like the nuclear bomb, where you either have 271 00:16:28,720 --> 00:16:31,320 Speaker 2: it or you don't, and you're kind of racing. The 272 00:16:31,400 --> 00:16:34,880 Speaker 2: models are going to get better at a very significant pace, 273 00:16:35,880 --> 00:16:39,280 Speaker 2: but the existence of a very capable model is not 274 00:16:40,600 --> 00:16:42,200 Speaker 2: in and of itself. 275 00:16:43,160 --> 00:16:43,520 Speaker 3: Winning. 276 00:16:44,320 --> 00:16:47,120 Speaker 2: The country that wins in AI is going to be 277 00:16:47,200 --> 00:16:51,880 Speaker 2: the country that has the most wide ranging and creative 278 00:16:52,280 --> 00:16:53,880 Speaker 2: and productivity. 279 00:16:53,240 --> 00:16:57,800 Speaker 3: Enhancing uses of AI. And so I. 280 00:16:57,760 --> 00:17:00,640 Speaker 2: Think if we're not careful, there's a very good chance 281 00:17:00,640 --> 00:17:05,320 Speaker 2: that Yes, America makes the leading models and then the 282 00:17:05,400 --> 00:17:09,800 Speaker 2: Chinese copy that technology in various ways, or maybe even 283 00:17:10,200 --> 00:17:15,320 Speaker 2: engage in some intellectual property theft of various kinds, and 284 00:17:15,359 --> 00:17:18,520 Speaker 2: then they duplicate those capabilities and they actually use them 285 00:17:18,560 --> 00:17:21,320 Speaker 2: more efficiently than we do, and they use them more creatively, 286 00:17:21,400 --> 00:17:24,840 Speaker 2: and they disperse them throughout their economy more than we do. 287 00:17:25,400 --> 00:17:29,240 Speaker 2: That would actually be very consistent with history if you 288 00:17:29,280 --> 00:17:34,840 Speaker 2: look at the Early Industrial Revolution. The UK led technologically 289 00:17:35,400 --> 00:17:38,840 Speaker 2: in the Early Industrial Revolution, but it was America that 290 00:17:38,920 --> 00:17:42,879 Speaker 2: took those ideas that were invented in the labs on 291 00:17:42,920 --> 00:17:45,239 Speaker 2: the other side of the Atlantic, and we deployed them 292 00:17:45,280 --> 00:17:48,600 Speaker 2: at scale here in the US and we built the 293 00:17:48,600 --> 00:17:52,480 Speaker 2: big factories. And something very similar could happen if we're 294 00:17:52,520 --> 00:17:55,600 Speaker 2: not careful. And I think that that will involve not 295 00:17:55,920 --> 00:18:00,560 Speaker 2: just sort of avoiding bad regulatory outcomes on AI, It'll 296 00:18:00,600 --> 00:18:03,439 Speaker 2: actually involve regulatory reform to make it easier to do 297 00:18:03,520 --> 00:18:06,639 Speaker 2: things in the physical world and easier to get cutting 298 00:18:06,720 --> 00:18:09,520 Speaker 2: edge medicines approved and all that kind of stuff. There's 299 00:18:09,600 --> 00:18:12,080 Speaker 2: just a ton that we need to do to make 300 00:18:12,119 --> 00:18:13,879 Speaker 2: it so that we can get the best out of 301 00:18:13,880 --> 00:18:15,440 Speaker 2: this technology. 302 00:18:16,119 --> 00:18:20,760 Speaker 1: They talk about an open weight model. What does that mean? 303 00:18:21,960 --> 00:18:27,160 Speaker 2: So the weights of an AI model, basically you can 304 00:18:27,160 --> 00:18:29,399 Speaker 2: think of it in the simplification a little bit, but 305 00:18:29,400 --> 00:18:33,280 Speaker 2: it's basically like a spreadsheet full of numbers, billions or 306 00:18:33,320 --> 00:18:40,920 Speaker 2: sometimes trillions of numbers, and those numbers constitute the functionality 307 00:18:41,080 --> 00:18:41,919 Speaker 2: of that model. 308 00:18:42,760 --> 00:18:44,600 Speaker 3: So an open weights model. 309 00:18:45,040 --> 00:18:49,879 Speaker 2: Is a model where the developer releases those numbers to 310 00:18:49,960 --> 00:18:54,000 Speaker 2: the public for free, for anybody to download or modify 311 00:18:54,000 --> 00:18:54,720 Speaker 2: as they see fit. 312 00:18:56,320 --> 00:19:00,760 Speaker 3: That mirrors the history of open source software. 313 00:19:00,920 --> 00:19:05,159 Speaker 2: Kind of open source software is software where the code 314 00:19:05,840 --> 00:19:10,520 Speaker 2: is released publicly, and for a variety of economic and 315 00:19:10,560 --> 00:19:15,800 Speaker 2: technological reasons, open source software has become essential to the 316 00:19:15,840 --> 00:19:20,080 Speaker 2: digital economy. There's open source software running on both of 317 00:19:20,119 --> 00:19:23,760 Speaker 2: our computers right now. Almost every website you'll visit in 318 00:19:23,800 --> 00:19:28,240 Speaker 2: the world is enabled by open source software. There's open 319 00:19:28,280 --> 00:19:32,800 Speaker 2: source software at the heart of practically everything in the 320 00:19:32,840 --> 00:19:38,639 Speaker 2: digital world. And so the question is is there going 321 00:19:38,680 --> 00:19:41,680 Speaker 2: to be something similar with AI. There are people who 322 00:19:41,720 --> 00:19:43,800 Speaker 2: make closed source software and make a lot of money 323 00:19:43,880 --> 00:19:46,520 Speaker 2: doing that. Apple would be a good example, Microsoft would 324 00:19:46,520 --> 00:19:49,680 Speaker 2: be a good example. But the question is will there 325 00:19:49,720 --> 00:19:54,080 Speaker 2: be a similar kind of role for this freely available, 326 00:19:54,200 --> 00:19:58,320 Speaker 2: free to modify AI, these AI models that are free 327 00:19:58,359 --> 00:20:02,040 Speaker 2: to modify and super flex Will there be a similar 328 00:20:02,119 --> 00:20:04,520 Speaker 2: role for that in the future economy, and I think 329 00:20:04,520 --> 00:20:05,600 Speaker 2: the answer is probably yes. 330 00:20:06,560 --> 00:20:10,560 Speaker 1: There must be an enormous number of people now already 331 00:20:10,600 --> 00:20:14,760 Speaker 1: committed to working on artificial intelligence and its various formats. Man, 332 00:20:14,920 --> 00:20:17,119 Speaker 1: it's a much bigger system than I thought it was. 333 00:20:18,040 --> 00:20:22,520 Speaker 2: Yeah, I think it's a big industry. The funny thing 334 00:20:22,560 --> 00:20:26,760 Speaker 2: about it is that there's not that many live players. 335 00:20:27,200 --> 00:20:33,040 Speaker 2: I'd say, in the US it's Meta Microsoft, Open ai, 336 00:20:33,240 --> 00:20:39,119 Speaker 2: andthropic Google XAI from Elon Musk handful of others, but 337 00:20:39,160 --> 00:20:42,280 Speaker 2: those are the really big players in terms of developing 338 00:20:42,320 --> 00:20:45,440 Speaker 2: these models, and in China there's a few big ones 339 00:20:45,480 --> 00:20:48,680 Speaker 2: as well. It is big, and there is a ton 340 00:20:49,000 --> 00:20:52,720 Speaker 2: of investment that's basically already in the cards. These companies 341 00:20:52,720 --> 00:20:56,520 Speaker 2: that are building the data centers, for example, they plan 342 00:20:56,680 --> 00:21:01,439 Speaker 2: their capital expenses out years in advance, so we kind 343 00:21:01,440 --> 00:21:03,479 Speaker 2: of already know what they're planning to spend, and so 344 00:21:03,560 --> 00:21:06,159 Speaker 2: that certain levels of investment and expansion are just kind 345 00:21:06,160 --> 00:21:06,879 Speaker 2: of already locked in. 346 00:21:07,400 --> 00:21:12,760 Speaker 1: I was almost startled the capital investment for the biggest 347 00:21:12,760 --> 00:21:17,480 Speaker 1: of these worldwide companies, and each company is three or 348 00:21:17,520 --> 00:21:21,520 Speaker 1: four times the size of the NASA budget. Yes, the 349 00:21:21,520 --> 00:21:24,639 Speaker 1: amount they're investing, the number of people they're hiring. The 350 00:21:24,640 --> 00:21:29,240 Speaker 1: scale of the computational power they're building, it's really astonishing, 351 00:21:29,359 --> 00:21:32,119 Speaker 1: and I don't think there's a single company in Europe 352 00:21:32,720 --> 00:21:33,840 Speaker 1: that's in this league. 353 00:21:34,400 --> 00:21:36,800 Speaker 3: No, no, there's not a single company in Europe that's 354 00:21:36,800 --> 00:21:37,479 Speaker 3: in this league. 355 00:21:37,680 --> 00:21:40,639 Speaker 2: The industrial buildout that is happening right now for AI 356 00:21:41,800 --> 00:21:46,800 Speaker 2: is certainly the biggest peacetime industrial build out we've seen 357 00:21:46,840 --> 00:21:51,480 Speaker 2: in our lifetimes, quite possibly the biggest ever, and in 358 00:21:51,520 --> 00:21:53,840 Speaker 2: the fullness of time, it might even become larger even 359 00:21:54,000 --> 00:21:58,440 Speaker 2: as a share of GDP than even wartime buildouts. 360 00:21:59,520 --> 00:22:02,320 Speaker 1: There's a lot of concern about China in general, and 361 00:22:02,400 --> 00:22:05,680 Speaker 1: now I think with deepseak there's also concern about China 362 00:22:05,880 --> 00:22:10,720 Speaker 1: challenging US on advanced artificial intelligence. How serious a threat 363 00:22:10,760 --> 00:22:11,560 Speaker 1: do you think it is. 364 00:22:12,720 --> 00:22:14,679 Speaker 3: I think it's a fierce competition that we're in. 365 00:22:15,640 --> 00:22:18,399 Speaker 2: It's a competition that's going to play out in a 366 00:22:18,440 --> 00:22:23,800 Speaker 2: lot of different dimensions. We are in an overall economic 367 00:22:24,359 --> 00:22:28,120 Speaker 2: and geopolitical competition with China, and that competition is going 368 00:22:28,160 --> 00:22:31,760 Speaker 2: to be defined by innovations in science and technology, and 369 00:22:31,880 --> 00:22:35,160 Speaker 2: innovations in science and technology in turn, at the foundation 370 00:22:35,280 --> 00:22:39,119 Speaker 2: of those innovations, I think increasingly will be AI. There's 371 00:22:39,160 --> 00:22:43,720 Speaker 2: also going to be a whole digital infrastructure that gets created, 372 00:22:44,720 --> 00:22:49,080 Speaker 2: and the question is going to be whose technology sets 373 00:22:49,119 --> 00:22:54,160 Speaker 2: the global standard? Will American technology set the global standard 374 00:22:54,640 --> 00:22:58,520 Speaker 2: or will Chinese technology set the global standard? And I 375 00:22:58,560 --> 00:23:01,800 Speaker 2: think we really want to do everything we can to 376 00:23:01,960 --> 00:23:05,640 Speaker 2: ensure that American technology does set that global standard. There's 377 00:23:05,640 --> 00:23:08,119 Speaker 2: a lot of ways in which American tech already does. 378 00:23:08,240 --> 00:23:11,719 Speaker 2: We've been so successful over the last fifty years that 379 00:23:12,440 --> 00:23:14,439 Speaker 2: in a lot of ways, I think many Americans are 380 00:23:14,480 --> 00:23:18,040 Speaker 2: almost blind to the ways in which we have set 381 00:23:18,080 --> 00:23:21,200 Speaker 2: the standard for how technology is used and deployed throughout 382 00:23:21,200 --> 00:23:25,160 Speaker 2: the world. For US, it's like white paint, it's just there. 383 00:23:25,280 --> 00:23:28,200 Speaker 2: We don't really think about it. But for China, which 384 00:23:28,240 --> 00:23:31,920 Speaker 2: for the most part lacks that soft power, they perceive 385 00:23:31,960 --> 00:23:35,120 Speaker 2: it quite readily and they want it. 386 00:23:36,080 --> 00:23:37,040 Speaker 3: And that's a part. 387 00:23:36,800 --> 00:23:40,520 Speaker 2: Of what the open weight and open source dynamic is about. 388 00:23:40,880 --> 00:23:44,119 Speaker 2: That's why China has a government level, a CCP level 389 00:23:44,280 --> 00:23:49,520 Speaker 2: priority is beating the US an open source the question 390 00:23:49,640 --> 00:23:54,159 Speaker 2: is not just who innovates faster, but whose technology is 391 00:23:54,200 --> 00:23:56,240 Speaker 2: available more widely around the world. 392 00:23:56,920 --> 00:23:59,800 Speaker 1: And what's the biggest Chinese advantage in that competition. 393 00:24:01,200 --> 00:24:04,800 Speaker 2: I think the biggest single advantage is their industrial and 394 00:24:04,800 --> 00:24:09,800 Speaker 2: mass manufacturing capacity. AI is going to enable so many 395 00:24:09,840 --> 00:24:12,040 Speaker 2: new kinds of inventions, and one way to think of 396 00:24:12,080 --> 00:24:17,280 Speaker 2: AI is like it's an invention for inventing inventions. So 397 00:24:17,560 --> 00:24:20,920 Speaker 2: then the question becomes, Okay, well, we have this new 398 00:24:20,960 --> 00:24:23,720 Speaker 2: way to massively accelerate the rate at which new things 399 00:24:23,720 --> 00:24:26,840 Speaker 2: are discovered and new things can be invented, and that's great, 400 00:24:27,359 --> 00:24:30,600 Speaker 2: But who can actually make those things? Who can mass 401 00:24:30,640 --> 00:24:33,240 Speaker 2: manufacture them, who can figure out how to make them cheaply? 402 00:24:34,440 --> 00:24:37,679 Speaker 2: And that's not going to just be a purely AI question. 403 00:24:38,280 --> 00:24:40,200 Speaker 2: There's going to be an extent to which that depends 404 00:24:40,240 --> 00:24:45,280 Speaker 2: on who has the factories, who has the supply chains, 405 00:24:46,000 --> 00:24:48,159 Speaker 2: And that's an area where I think the Chinese have 406 00:24:48,359 --> 00:24:52,840 Speaker 2: a quite large advantage over us in not all, but 407 00:24:52,960 --> 00:24:55,120 Speaker 2: many important domains. 408 00:24:55,880 --> 00:24:57,960 Speaker 1: Now, the other side of that is, it seems to 409 00:24:58,040 --> 00:25:02,199 Speaker 1: me that the sheer number of innovative efforts here is 410 00:25:02,240 --> 00:25:08,000 Speaker 1: the cost of developing the computational power for artificial intelligence 411 00:25:08,080 --> 00:25:11,760 Speaker 1: so great that only really large companies can compete. 412 00:25:12,640 --> 00:25:17,880 Speaker 2: I think certainly to build the biggest foundation models as 413 00:25:17,880 --> 00:25:23,400 Speaker 2: they're called, is going to be quite capital intensive. At 414 00:25:23,440 --> 00:25:26,760 Speaker 2: the same time, one of the things that deep Seek 415 00:25:26,840 --> 00:25:30,200 Speaker 2: shows us is that because those costs of developing that 416 00:25:30,200 --> 00:25:34,000 Speaker 2: foundation model go down so quickly over time, smaller players 417 00:25:34,000 --> 00:25:37,479 Speaker 2: can catch up relatively quickly. They cannot necessarily catch up, 418 00:25:37,480 --> 00:25:39,720 Speaker 2: but they can at least if you see a certain 419 00:25:39,760 --> 00:25:43,680 Speaker 2: capability today, you should expect that a much wider range 420 00:25:43,680 --> 00:25:45,840 Speaker 2: of actors will be able to achieve that same capability 421 00:25:45,880 --> 00:25:50,439 Speaker 2: in eighteen months. But another way to think about it 422 00:25:50,480 --> 00:25:55,560 Speaker 2: is these really expensive foundation models are just that they're 423 00:25:55,600 --> 00:26:01,080 Speaker 2: foundation models, and there are so many things that can 424 00:26:01,119 --> 00:26:03,119 Speaker 2: be built on top of them, and I think that's 425 00:26:03,200 --> 00:26:07,520 Speaker 2: really where the entrepreneurship and startup opportunity is going to be. 426 00:26:08,359 --> 00:26:10,840 Speaker 3: Let's take legal services. One of the things that has 427 00:26:10,880 --> 00:26:14,040 Speaker 3: really impressed me about the most cutting edge models today 428 00:26:14,720 --> 00:26:17,720 Speaker 3: is that they're getting really good at legal reasoning, and 429 00:26:17,800 --> 00:26:22,080 Speaker 3: so I can ask them really complicated questions about tort 430 00:26:22,160 --> 00:26:27,600 Speaker 3: liability and reasoning about various public policy outcomes and get 431 00:26:27,680 --> 00:26:34,040 Speaker 3: answers that are starting to feel like white shoe legal analysis. 432 00:26:35,160 --> 00:26:39,440 Speaker 3: But it's not obvious to me that an existing white 433 00:26:39,480 --> 00:26:42,280 Speaker 3: shoe law firm is going to be the company that 434 00:26:43,359 --> 00:26:46,840 Speaker 3: delivers that capability to the world. Instead, I think, well, 435 00:26:47,000 --> 00:26:49,160 Speaker 3: maybe there's a new kind of startup that can exist 436 00:26:49,200 --> 00:26:53,560 Speaker 3: that can provide scatn arp's level of legal analysis to 437 00:26:53,720 --> 00:26:56,280 Speaker 3: everyone for a far lower cost. And that's one of 438 00:26:56,280 --> 00:26:58,720 Speaker 3: a billion startups that you can imagine. 439 00:26:58,720 --> 00:27:01,760 Speaker 1: One of the examples of the scale of at least 440 00:27:01,760 --> 00:27:05,880 Speaker 1: semi artificial intelligence that is actually occurring, and I think 441 00:27:05,880 --> 00:27:08,880 Speaker 1: almost nobody's covering it. Is there now. I think three 442 00:27:08,960 --> 00:27:13,040 Speaker 1: cities that have agreed to have experiments with taxis that 443 00:27:13,080 --> 00:27:16,919 Speaker 1: have no driver, well, I don't think most people have 444 00:27:16,920 --> 00:27:23,040 Speaker 1: thought through. Once that becomes so reliable that you don't 445 00:27:23,040 --> 00:27:25,720 Speaker 1: worry about it, you're going to have an entire revolution 446 00:27:26,760 --> 00:27:29,439 Speaker 1: in the nature of cars, and you're going to have 447 00:27:29,440 --> 00:27:32,760 Speaker 1: a revolution as they already have a Tesla. I mean 448 00:27:32,800 --> 00:27:37,119 Speaker 1: the sheer volume of data that Tesla absorbs every day, 449 00:27:37,440 --> 00:27:40,320 Speaker 1: because I think every Tesla is basically an information gathering. 450 00:27:40,520 --> 00:27:43,479 Speaker 1: I tell people it's actually a computer company with cars, 451 00:27:43,960 --> 00:27:46,000 Speaker 1: it's not a car company with computers. 452 00:27:46,680 --> 00:27:47,920 Speaker 3: I think that's exactly right. 453 00:27:48,400 --> 00:27:50,679 Speaker 2: There's going to be a certain extent to which the 454 00:27:50,720 --> 00:27:52,399 Speaker 2: companies that have the lead are going to be the 455 00:27:52,400 --> 00:27:55,400 Speaker 2: ones with the highest quality data and also the access 456 00:27:55,400 --> 00:27:57,600 Speaker 2: to the best compute and all that kind of stuff. 457 00:27:58,280 --> 00:28:02,200 Speaker 2: But I think that you can totally imagine a world 458 00:28:02,359 --> 00:28:06,359 Speaker 2: in which a company like open ai their explicit goal 459 00:28:06,560 --> 00:28:09,359 Speaker 2: is to develop a system within the next few years 460 00:28:09,400 --> 00:28:14,560 Speaker 2: that is better than humans at all economically valuable tasks. Now, 461 00:28:15,040 --> 00:28:16,600 Speaker 2: I think you can have a lot of quibbles with 462 00:28:16,640 --> 00:28:21,560 Speaker 2: that definition, and like maybe a better way to think 463 00:28:21,560 --> 00:28:24,679 Speaker 2: about it would be all current economic or like most 464 00:28:25,200 --> 00:28:27,119 Speaker 2: current economically valuable. 465 00:28:26,720 --> 00:28:30,400 Speaker 3: Task, but new things will become economically valuable for humans, 466 00:28:30,400 --> 00:28:32,040 Speaker 3: So I think you can that deminission. 467 00:28:32,080 --> 00:28:34,879 Speaker 2: But either way, it's a very profound idea of having 468 00:28:35,000 --> 00:28:37,040 Speaker 2: systems that are so capable, and you can imagine a 469 00:28:37,119 --> 00:28:40,400 Speaker 2: world where like, well, we have the everything machine and 470 00:28:40,440 --> 00:28:42,400 Speaker 2: it can do everything. So we're just gonna as a 471 00:28:42,440 --> 00:28:46,760 Speaker 2: company do everything. But the question will be, well, that 472 00:28:46,880 --> 00:28:49,840 Speaker 2: means that you're regulated by every regulator in the world. 473 00:28:50,400 --> 00:28:53,200 Speaker 2: You're subject to every single law, you're subject to all 474 00:28:53,280 --> 00:28:57,600 Speaker 2: of the liability that every single actor in the economy 475 00:28:58,120 --> 00:28:58,800 Speaker 2: is subject to. 476 00:28:59,480 --> 00:29:02,120 Speaker 3: Do you want as a single firm? Can you do that? 477 00:29:02,680 --> 00:29:04,320 Speaker 2: There's going to be an extent to which, yes, the 478 00:29:04,360 --> 00:29:07,479 Speaker 2: foundation models get better and better, but there's also going 479 00:29:07,560 --> 00:29:10,720 Speaker 2: to be people that are uniquely positioned to sort of 480 00:29:10,880 --> 00:29:14,760 Speaker 2: serve particular industries and particular needs in the economy using 481 00:29:14,800 --> 00:29:18,360 Speaker 2: the foundation models, but those startup firms will tailor themselves 482 00:29:18,360 --> 00:29:24,240 Speaker 2: to the particular regulatory requirements, legal requirements, liability, customer needs, 483 00:29:24,400 --> 00:29:27,000 Speaker 2: et cetera, economic dynamics of that industry. 484 00:29:27,400 --> 00:29:29,920 Speaker 3: So I think that's probably the way it will play out. 485 00:29:45,640 --> 00:29:50,959 Speaker 1: Occurring probably the largest set of stored data as Google, 486 00:29:52,520 --> 00:29:56,080 Speaker 1: so in terms of training artificial intelligence systems, they ought 487 00:29:56,120 --> 00:29:58,720 Speaker 1: to have a huge potential advantage. 488 00:30:00,080 --> 00:30:04,120 Speaker 2: They do have a big advantage there. One interesting question 489 00:30:04,840 --> 00:30:09,160 Speaker 2: actually is Google owns YouTube and a lot of these 490 00:30:09,240 --> 00:30:12,880 Speaker 2: companies are training on video data, and it's actually an 491 00:30:12,880 --> 00:30:16,920 Speaker 2: interesting question of is Google actually less able to use 492 00:30:16,960 --> 00:30:20,600 Speaker 2: YouTube data because they own it and they are therefore 493 00:30:21,320 --> 00:30:24,760 Speaker 2: contractual obligations that they have made with their users in 494 00:30:24,800 --> 00:30:28,120 Speaker 2: the form of the YouTube terms of service, That actually 495 00:30:28,200 --> 00:30:30,280 Speaker 2: might make it harder for Google too and make it 496 00:30:30,320 --> 00:30:35,120 Speaker 2: easier for again, particularly a Chinese company. China has access 497 00:30:35,160 --> 00:30:38,120 Speaker 2: to YouTube too. They can take all the YouTube videos 498 00:30:38,440 --> 00:30:40,720 Speaker 2: and train on them and there might not be the 499 00:30:40,720 --> 00:30:42,840 Speaker 2: same intellectual property concerns that they. 500 00:30:42,720 --> 00:30:45,320 Speaker 1: Have and they have all the TikTok data. 501 00:30:46,000 --> 00:30:46,720 Speaker 3: Yeah, that's right. 502 00:30:47,000 --> 00:30:49,240 Speaker 2: One way to think about this is just it's like 503 00:30:49,280 --> 00:30:54,000 Speaker 2: an old economic lesson that there's an economist I learned 504 00:30:54,080 --> 00:30:57,640 Speaker 2: about this concept from named Friedrich Hayek, who's a twentieth 505 00:30:57,680 --> 00:31:02,800 Speaker 2: century economist who talked about how information is spread unevenly 506 00:31:02,920 --> 00:31:06,880 Speaker 2: throughout the economy. Not just different firms, but also different 507 00:31:06,880 --> 00:31:10,600 Speaker 2: individuals have differential access to information. So there's all these asymmetries. 508 00:31:11,240 --> 00:31:14,800 Speaker 2: And the thing that might become truly valuable in the 509 00:31:14,840 --> 00:31:19,440 Speaker 2: future might be the information that's proprietary to a firm. 510 00:31:19,920 --> 00:31:22,280 Speaker 2: The things that only you know are the things that 511 00:31:22,360 --> 00:31:25,200 Speaker 2: only your company knows, and those might be the things 512 00:31:25,200 --> 00:31:27,560 Speaker 2: that are valuable because the AI models, in some sense 513 00:31:27,640 --> 00:31:30,760 Speaker 2: might know everything that's public. There's this common thing that 514 00:31:30,800 --> 00:31:35,560 Speaker 2: people say about how all the information in the world 515 00:31:35,600 --> 00:31:38,440 Speaker 2: is on the Internet, and to that, I say, you 516 00:31:38,560 --> 00:31:41,240 Speaker 2: haven't tried to find a lot of things out because 517 00:31:41,280 --> 00:31:43,400 Speaker 2: there's a lot of information that's not on the Internet. 518 00:31:43,600 --> 00:31:46,560 Speaker 2: For example, if you want to get something done in Congress, 519 00:31:47,080 --> 00:31:48,880 Speaker 2: the internet's not really going to tell you how to 520 00:31:48,880 --> 00:31:52,520 Speaker 2: do that. The Internet doesn't tell you which staffers have 521 00:31:52,600 --> 00:31:55,960 Speaker 2: relationships with one another, like friendships with one another, where 522 00:31:55,960 --> 00:31:59,160 Speaker 2: are their political tensions, Like how do you actually kind 523 00:31:59,160 --> 00:32:02,320 Speaker 2: of operate the machinery of Congress. You had a lot 524 00:32:02,360 --> 00:32:04,640 Speaker 2: of information as Speaker of the House that was in 525 00:32:04,680 --> 00:32:07,160 Speaker 2: your brain that no one else in the world knew, 526 00:32:07,840 --> 00:32:11,160 Speaker 2: and that kind of thing might just become even more 527 00:32:11,280 --> 00:32:12,400 Speaker 2: valuable in the future. 528 00:32:12,800 --> 00:32:16,000 Speaker 1: Somebody made the point, we're gathering up all this data, 529 00:32:16,440 --> 00:32:20,400 Speaker 1: but it's actually data that's electronic and that that's not 530 00:32:20,520 --> 00:32:23,320 Speaker 1: the same as the real world. And we don't yet 531 00:32:23,360 --> 00:32:26,480 Speaker 1: have a very good model for gathering up data about 532 00:32:26,480 --> 00:32:30,600 Speaker 1: the real world as opposed to the mathematical representation of 533 00:32:30,640 --> 00:32:31,280 Speaker 1: that world. 534 00:32:31,480 --> 00:32:34,520 Speaker 2: Does that make sense, Yeah, I think that's true. It 535 00:32:34,640 --> 00:32:38,800 Speaker 2: varies by field. We have lots of good data from biology, 536 00:32:39,760 --> 00:32:41,719 Speaker 2: A lot of it is hard to get though, so 537 00:32:42,440 --> 00:32:44,560 Speaker 2: it's like, in total, there's a ton of good data 538 00:32:44,600 --> 00:32:47,880 Speaker 2: about biology, but how do you get access to it all. 539 00:32:48,520 --> 00:32:51,760 Speaker 2: Companies like Tesla or Waimo, which is a subsidiary of Google, 540 00:32:52,000 --> 00:32:57,800 Speaker 2: another self driving car company, they've spent a decade collecting 541 00:32:58,880 --> 00:33:02,120 Speaker 2: naturalistic data from the real world of streets of how 542 00:33:02,160 --> 00:33:04,720 Speaker 2: people drive and what that's actually like. And we're just 543 00:33:04,840 --> 00:33:08,400 Speaker 2: now getting to the point that that enormous volume of 544 00:33:08,480 --> 00:33:12,040 Speaker 2: data can be used to make pretty darn good self 545 00:33:12,120 --> 00:33:14,520 Speaker 2: driving cars. But it took a long time and so 546 00:33:14,600 --> 00:33:17,160 Speaker 2: when you think about things like robots, for example, a 547 00:33:17,200 --> 00:33:19,880 Speaker 2: lot of people are excited about humanoid robots that could 548 00:33:19,920 --> 00:33:22,000 Speaker 2: live in your house and cook meals for you. Well, 549 00:33:22,000 --> 00:33:24,760 Speaker 2: that's going to take a ton of data. The Internet 550 00:33:24,800 --> 00:33:28,240 Speaker 2: doesn't have tons and tons of point of view videos 551 00:33:28,240 --> 00:33:30,680 Speaker 2: of people doing chores around their house. Someone's going to 552 00:33:30,720 --> 00:33:32,440 Speaker 2: have to go out and collect that, and that's going 553 00:33:32,480 --> 00:33:34,040 Speaker 2: to be a non trivial exercise. 554 00:33:34,560 --> 00:33:38,959 Speaker 1: Do you think ASAI continues to develop and both continues 555 00:33:39,000 --> 00:33:41,640 Speaker 1: to become more powerful and more intrusive, do you think 556 00:33:41,680 --> 00:33:45,600 Speaker 1: there's a point where sort of consumer reaction may slow 557 00:33:45,640 --> 00:33:48,080 Speaker 1: it down and lead to rejection or do you think 558 00:33:48,240 --> 00:33:51,080 Speaker 1: is just inevitably going to continue growing. 559 00:33:52,240 --> 00:33:56,240 Speaker 2: I think there will absolutely be social bottle ducks that 560 00:33:56,320 --> 00:34:00,320 Speaker 2: will slow down the diffusion of the technology into the world. 561 00:34:01,160 --> 00:34:04,200 Speaker 2: Even today, we've had the Internet now for thirty years 562 00:34:05,040 --> 00:34:10,040 Speaker 2: and huge number of customer service things can be resolved 563 00:34:10,719 --> 00:34:13,760 Speaker 2: without picking up the phone and calling a customer service line. 564 00:34:13,880 --> 00:34:16,400 Speaker 2: But pretty much every company still has to maintain a 565 00:34:16,400 --> 00:34:18,680 Speaker 2: customer service line because there's a lot of people who 566 00:34:18,719 --> 00:34:20,800 Speaker 2: just prefer to talk to a person. That's going to 567 00:34:20,840 --> 00:34:24,440 Speaker 2: continue to be true. I wouldn't be surprised. In fact, 568 00:34:24,600 --> 00:34:27,640 Speaker 2: if AI systems become so capable that there's sort of 569 00:34:27,719 --> 00:34:30,840 Speaker 2: a broader movement coalesces, that's kind of a backlash to 570 00:34:31,000 --> 00:34:33,120 Speaker 2: all of this, because I think there will be people 571 00:34:33,120 --> 00:34:35,880 Speaker 2: who find the changes to be quite shocking, and I 572 00:34:35,920 --> 00:34:37,920 Speaker 2: think that'll slow down the diffusion of. 573 00:34:37,880 --> 00:34:39,240 Speaker 3: The technology all over the world. 574 00:34:39,280 --> 00:34:41,440 Speaker 2: At the end of the day, I think it's an inevitability, 575 00:34:42,360 --> 00:34:45,640 Speaker 2: just like the Internet was, just like the smartphone was. 576 00:34:46,320 --> 00:34:48,440 Speaker 2: These things will happen, but they will take time, and 577 00:34:48,480 --> 00:34:50,839 Speaker 2: that's one of the reasons there are people I know 578 00:34:50,960 --> 00:34:53,960 Speaker 2: in the AI community who think that just a few 579 00:34:54,000 --> 00:34:56,400 Speaker 2: years from now, the whole world is going to be 580 00:34:56,520 --> 00:34:59,759 Speaker 2: utterly transformed and it's going to look like something out 581 00:34:59,760 --> 00:35:03,680 Speaker 2: of science fiction in just five years because of how 582 00:35:03,719 --> 00:35:06,200 Speaker 2: capable the systems will be. And I would say that 583 00:35:06,320 --> 00:35:09,840 Speaker 2: even if the systems are that smart, there's just absolutely 584 00:35:09,880 --> 00:35:12,760 Speaker 2: no way something like that is going to happen, because 585 00:35:12,800 --> 00:35:16,160 Speaker 2: there will be what I suppose you could call in 586 00:35:16,200 --> 00:35:20,520 Speaker 2: the field sociotechnical bottlenecks, which you might put more simply 587 00:35:20,600 --> 00:35:24,640 Speaker 2: as people not wanting to adapt that quickly, and I 588 00:35:24,680 --> 00:35:25,840 Speaker 2: think that's actually healthy. 589 00:35:26,480 --> 00:35:28,080 Speaker 1: We're going to come back to you again and again. 590 00:35:28,120 --> 00:35:30,359 Speaker 1: I think over the next couple of years, you are 591 00:35:30,400 --> 00:35:32,480 Speaker 1: right in the middle of a field that is I 592 00:35:32,480 --> 00:35:35,239 Speaker 1: think you're going to continue to accelerate and deepen and 593 00:35:35,280 --> 00:35:37,840 Speaker 1: be even more important. And Dean, I really want to 594 00:35:37,880 --> 00:35:40,719 Speaker 1: thank you for joining me. Our listeners can follow the 595 00:35:40,719 --> 00:35:45,640 Speaker 1: work you're doing by subscribing to your substack hyperdimensional at 596 00:35:45,640 --> 00:35:49,719 Speaker 1: substack dot com, slash at Dean W. Ball, and by 597 00:35:49,800 --> 00:35:53,680 Speaker 1: visiting the Mercada Center's website at mercadas dot org. And 598 00:35:54,160 --> 00:35:56,319 Speaker 1: we will have a link to your substack on our 599 00:35:56,360 --> 00:35:59,680 Speaker 1: show page of Newtsworld. So thank you very very much for. 600 00:36:00,040 --> 00:36:03,120 Speaker 2: Educating us, mister speaker, thank you so much, it's been 601 00:36:03,160 --> 00:36:03,560 Speaker 2: an honor. 602 00:36:07,120 --> 00:36:09,239 Speaker 1: Thank you to my guest, Dean Ball. You can get 603 00:36:09,239 --> 00:36:12,759 Speaker 1: a link to his substack Hyperdimensional on our show page 604 00:36:12,800 --> 00:36:15,920 Speaker 1: at neut world dot com. Newtorld is produced by gingridh 605 00:36:15,920 --> 00:36:20,200 Speaker 1: three sixty and iHeartMedia. Our executive producer is Guarnsey Sloan. 606 00:36:20,600 --> 00:36:24,000 Speaker 1: Our researcher is Rachel Peterson. The artwork for the show 607 00:36:24,440 --> 00:36:27,520 Speaker 1: was created by Steve Penley. Special thanks to the team 608 00:36:27,520 --> 00:36:30,480 Speaker 1: at Gingrid three sixty. If you've been enjoying newts World, 609 00:36:30,800 --> 00:36:33,480 Speaker 1: I hope you'll go to Apple Podcasts and both rate 610 00:36:33,560 --> 00:36:36,319 Speaker 1: us with five stars and give us a review so 611 00:36:36,400 --> 00:36:39,800 Speaker 1: others can learn what it's all about. Right now, listeners 612 00:36:39,800 --> 00:36:43,040 Speaker 1: of neut World can sign up for my three freeweekly 613 00:36:43,120 --> 00:36:48,280 Speaker 1: columns at Gingrichtree sixty dot com slash newsletter. I'm NEWT Gingrich. 614 00:36:48,640 --> 00:37:00,080 Speaker 1: This is neut World w