WEBVTT - Episode 835: Gary Rivlin on “AI Valley” 0:00:04.800 --> 0:00:07.720 On this episode of a News World. From boardrooms to 0:00:07.840 --> 0:00:11.280 dorm rooms, AI seems to be what everyone is talking about, 0:00:11.680 --> 0:00:15.520 from the promise of Chat GPT to robots that may 0:00:15.840 --> 0:00:19.160 or may not take your job. In his new book 0:00:19.840 --> 0:00:25.079 AI Valley, Microsoft, Google, and the trillion dollar Race to 0:00:25.200 --> 0:00:29.479 cash in an Artificial intelligence, Pulitzer Prize winning journalist Gary 0:00:29.520 --> 0:00:33.960 Rivlin follows the launch of Chat, GPT and three startups, 0:00:34.400 --> 0:00:36.800 all with big dreams of cashing. 0:00:36.520 --> 0:00:37.120 In on AI. 0:00:37.800 --> 0:00:40.760 But it's not long before the tech giants enter the 0:00:40.760 --> 0:00:44.919 AI space. Rivlin lays out the fascinating history of AI's evolution, 0:00:45.560 --> 0:00:48.839 the breakthroughs and wrong turns, and the major players in 0:00:48.880 --> 0:00:52.640 Silicon Valley, the developers and investors. 0:00:52.200 --> 0:00:53.640 Who will lead the future of AI. 0:00:54.520 --> 0:00:57.800 Here to discuss his new book, I am really pleased 0:00:57.840 --> 0:01:01.000 to welcome my guest, Gary Rivlin. He is a Pulitzer 0:01:01.000 --> 0:01:04.800 Prize winning investigative reporter who has been writing about technology 0:01:05.080 --> 0:01:08.200 since the mid nineteen nineties and the rise of the Internet. 0:01:08.680 --> 0:01:12.480 He is the author of ten previous books, including Saving 0:01:12.560 --> 0:01:17.000 Main Street and Katrina After the Flood. His work has 0:01:17.000 --> 0:01:21.200 appeared in The New York Times, Newsweek, Fortune, GQ, and Wired, 0:01:21.440 --> 0:01:34.480 among other publications. Gary, that's an amazing record. Thank you 0:01:34.560 --> 0:01:35.240 for joining us. 0:01:35.800 --> 0:01:36.960 Thank you my pleasure. 0:01:37.440 --> 0:01:41.039 So you've been covering tax since the nineties. How has 0:01:41.319 --> 0:01:44.520 tech evolved since you first started covering it and what 0:01:44.640 --> 0:01:48.600 has surprised you most about how Silicon Valley has evolved. 0:01:49.720 --> 0:01:55.120 Well, the short answer is the dominance of the giants. 0:01:55.320 --> 0:01:58.800 In fact, for this book, I went looking since the 0:01:58.880 --> 0:02:01.680 end of twenty twenty two and I realized, like, this 0:02:01.800 --> 0:02:04.480 is now the AI moment. They turned back to tech 0:02:04.680 --> 0:02:07.680 and I was looking for what would be the new Google, 0:02:07.800 --> 0:02:10.000 what would be the new Facebook? And it turns out 0:02:10.040 --> 0:02:12.560 the new Google is Google, the new Facebook is Facebook. 0:02:13.040 --> 0:02:14.880 And so you know, kind of dating back to the 0:02:14.960 --> 0:02:17.440 mid nineteen nineties, it was all about startups. It's all 0:02:17.480 --> 0:02:21.560 about these companies founded in a dorm room, someone's garage. 0:02:21.800 --> 0:02:23.560 They have a great idea, they raise a little bit 0:02:23.600 --> 0:02:27.760 of money, they get some traction, and they become Google. Facebook. 0:02:27.800 --> 0:02:30.600 But AI is different. I fear it's going to solidify 0:02:30.639 --> 0:02:33.480 the power of big tech rather than open it up 0:02:33.520 --> 0:02:35.480 to another set of players. 0:02:35.800 --> 0:02:39.839 Are there characteristics of the investment you have to make 0:02:40.520 --> 0:02:44.400 that make AI susceptible to that kind of dominance money? 0:02:44.520 --> 0:02:47.440 It's just so expensive. In the old days, you can 0:02:47.840 --> 0:02:51.080 raise a million, a few million dollars and start to 0:02:51.080 --> 0:02:54.160 get traction. Then you need to raise the big money 0:02:54.160 --> 0:02:59.239 to go national, go global, whatever. But AI training these models, 0:02:59.280 --> 0:03:03.360 this general AI where they can talk to you, spit 0:03:03.400 --> 0:03:08.440 out images, make video. It's so expensive to train these things. 0:03:08.440 --> 0:03:10.840 It's so expensive to fine tune these things. It's so 0:03:10.880 --> 0:03:15.120 expensive to operate these things. Is beyond the means of 0:03:15.760 --> 0:03:18.560 most startups. But I first started reporting on this at 0:03:18.600 --> 0:03:22.359 the start of twenty twenty three, it would be millions, 0:03:22.400 --> 0:03:26.960 maybe ten million to train and fine tune one of 0:03:27.000 --> 0:03:30.000 these models before they released. By the time I was 0:03:30.040 --> 0:03:32.799 done reporting, it was one hundred million. And now it's 0:03:32.919 --> 0:03:36.000 billions of dollars. What's startup? I mean, there's open AI, 0:03:36.120 --> 0:03:38.680 there's a couple others, but very few startups could raise 0:03:38.720 --> 0:03:41.280 that kind of money. And it's still not enough. I mean, 0:03:41.360 --> 0:03:44.080 these startups are still losing money, so they have to 0:03:44.160 --> 0:03:47.280 raise billions, tens of billions, perhaps in the future, not 0:03:47.520 --> 0:03:50.920 so distant future, one hundred billion dollars and more. And 0:03:51.320 --> 0:03:54.160 big tech can afford that. You know, Microsoft, Apple, They're 0:03:54.160 --> 0:03:57.280 sitting on one hundred billion dollars or so in their savings. 0:03:57.560 --> 0:04:00.160 But what chance does a startup have to do of 0:04:00.160 --> 0:04:04.080 these cutting edge foundational models, the chatbots in the like. 0:04:04.880 --> 0:04:08.960 When you look at the biggest companies Microsoft, Google, Meta, 0:04:09.680 --> 0:04:10.880 and Amazon. 0:04:11.080 --> 0:04:13.920 That the share of volume of cash that they create. 0:04:14.480 --> 0:04:17.160 Yeah, to me, it's astonishing that, with the exception of 0:04:17.200 --> 0:04:20.680 a couple companies in China and Saudi Aramco, all of 0:04:20.680 --> 0:04:22.760 the trillion dollar companies in the world are American. 0:04:23.279 --> 0:04:27.000 It's astonishing. Let's just focus on Search is arguably the 0:04:27.040 --> 0:04:30.760 best business ever. Once you've built the infrastructure, there's not 0:04:30.880 --> 0:04:35.080 much marginal costs to add new users, and so Google 0:04:35.200 --> 0:04:37.520 is making over one hundred billion, something like one hundred 0:04:37.520 --> 0:04:41.040 and fifty billion dollars a year in profits just from search. 0:04:41.600 --> 0:04:44.960 Microsoft they've been struggling to get into search with being 0:04:45.320 --> 0:04:48.320 and they would get like three four five percentage points, 0:04:48.360 --> 0:04:51.320 which is nothing except for it would still translate to 0:04:51.640 --> 0:04:54.320 hundreds and hundreds and hundreds of millions of dollars of revenue. 0:04:54.320 --> 0:04:57.440 And you can make the same argument with Facebook Meta. 0:04:57.520 --> 0:05:01.200 In the old days, the rough statistic is newspapers, magazines, 0:05:01.240 --> 0:05:06.520 publications brought in like fifty billion dollars in advertising revenue collectively. 0:05:06.600 --> 0:05:10.239 That was a circle like two thousand. Nowadays they're bringing 0:05:10.279 --> 0:05:13.040 in closer to ten billion dollars and most of the 0:05:13.080 --> 0:05:17.160 rest is being divvied up by Google and Meta, and 0:05:17.240 --> 0:05:20.200 so it's kind of an idea of winner takes most 0:05:20.279 --> 0:05:22.360 and the stakes are so big that they're just making 0:05:22.520 --> 0:05:25.920 so much money. But you know, this dates back to Microsoft. 0:05:25.920 --> 0:05:29.599 I mean Microsoft, which actually sells product, right, I mean 0:05:30.040 --> 0:05:33.960 you buy your Windows operating system and software package, office 0:05:34.000 --> 0:05:38.800 software packages. You know, their profits were astonishing in the 0:05:38.880 --> 0:05:41.400 nineteen nineties, and I think it just builds on itself. 0:05:41.400 --> 0:05:46.000 They're so rich they can afford to invest in AI, 0:05:46.279 --> 0:05:50.960 gener of AI, these leading edge technologies, and sometimes bigness 0:05:51.040 --> 0:05:53.240 is a weakness right there. You know, they kind of 0:05:53.240 --> 0:05:55.440 trip over their own feet. Google was so far ahead 0:05:55.480 --> 0:05:58.400 of everyone with machine learning, ditting back to the twenty tens, 0:05:58.520 --> 0:06:00.880 but of course it was open AI to a chat GBT. 0:06:01.120 --> 0:06:04.240 You know, big companies are scared the innovator's dilemma. They 0:06:04.240 --> 0:06:07.880 don't want to threaten their existing honeypot. And also, you know, 0:06:07.920 --> 0:06:11.680 I mean startups have advantages, but the advantage of money 0:06:12.080 --> 0:06:15.719 in AI makes me worry that it's kind of game over. 0:06:16.320 --> 0:06:19.320 In the case of Google, you're doing all the work. 0:06:20.279 --> 0:06:22.600 They build a framework within which you get to come, 0:06:22.920 --> 0:06:25.440 you get to play. They don't have to pay anybody. 0:06:25.520 --> 0:06:27.880 These people are all coming and saying, please let me 0:06:27.960 --> 0:06:29.000 come and use your. 0:06:28.880 --> 0:06:34.159 Material exactly Facebook, Instagram, we could go on listing, Twitter X. 0:06:34.279 --> 0:06:36.960 There's an expression in Silicon Valley if you're not paying 0:06:37.000 --> 0:06:39.839 for the product, you are the product. And that's a 0:06:39.880 --> 0:06:43.560 perfect way of understanding a Google or a Facebook. Look, 0:06:43.640 --> 0:06:46.800 let's use Google. So this phone in your pocket tracks 0:06:46.839 --> 0:06:50.200 you everywhere you do your searches online. They track that, 0:06:50.600 --> 0:06:52.880 They bundle up the data, and they sell it to 0:06:52.880 --> 0:06:55.280 the highest bidder. It's a great business. It makes a 0:06:55.279 --> 0:06:58.279 lot of money for them. But I think citizens only 0:06:58.680 --> 0:07:01.920 slowly woke up to that fact. And again this gets 0:07:01.920 --> 0:07:04.599 me back to my worry that if the same big 0:07:04.720 --> 0:07:08.600 tech companies, the same few companies are going to dominate 0:07:08.640 --> 0:07:11.400 AI the way they've dominated the last bunch of years. 0:07:11.560 --> 0:07:14.880 We don't trust them as the stewards for technology, and 0:07:15.000 --> 0:07:19.040 in fact AI is a powerful technology. AI is going 0:07:19.120 --> 0:07:25.360 to rely on our information. There's privacy concern. So that 0:07:25.560 --> 0:07:29.640 is my concern that these companies that are proven untrustworthy 0:07:30.000 --> 0:07:32.880 are going to be the ones bringing us this amazing power. 0:07:32.960 --> 0:07:36.280 I'm optimistic about AI. I lean optimistic. I think AI 0:07:36.320 --> 0:07:43.200 could bring incredible things around education, scientific breakthroughs, medicine. My 0:07:43.320 --> 0:07:46.080 worry is it's in the hands of companies that show 0:07:46.120 --> 0:07:50.000 that it's all about profits and not about trust and safety, 0:07:50.000 --> 0:07:52.440 which again I think is essential for AI. 0:07:53.000 --> 0:07:53.760 Let me do all his thinction. 0:07:54.000 --> 0:07:57.480 There was a period there where people were coming up 0:07:57.480 --> 0:08:02.240 with pretty cool innovations and then promptly getting bought, so 0:08:02.720 --> 0:08:05.200 they never actually a chance to grow into a competitor 0:08:05.720 --> 0:08:09.080 because they were acquired by these large companies and who 0:08:09.160 --> 0:08:13.760 absorbed them. Now, if I understand you correctly, it's ructually 0:08:13.840 --> 0:08:17.640 impossible for a startup in the AI field because of 0:08:17.640 --> 0:08:20.600 the scale of resources it takes to create it. So 0:08:20.720 --> 0:08:24.320 virtually all of the next level of innovation in terms 0:08:24.360 --> 0:08:27.400 of small companies is going to be the use of AI, 0:08:27.920 --> 0:08:30.240 which will be provided by one of the big companies. 0:08:30.800 --> 0:08:34.240 Let me break the startup world into two general categories. 0:08:34.679 --> 0:08:38.880 There's still going to be plenty opportunities for founders to 0:08:39.040 --> 0:08:42.480 raise some money and have a good return on an investment. 0:08:42.600 --> 0:08:46.600 For some app like AI will automatically fill out your 0:08:46.600 --> 0:08:49.360 expense sheets and do most if not all, of the 0:08:49.400 --> 0:08:51.560 work for you. You could see that be very valuable. 0:08:51.760 --> 0:08:55.640 And then there'll still be in quote little companies, they'll 0:08:55.679 --> 0:08:58.640 have millions of users bringing tens hundreds of millions of dollars. 0:08:58.880 --> 0:09:03.000 But I'm really focused on that second category. I have 0:09:03.160 --> 0:09:06.200 a billion plus users, I have a market cap, a 0:09:06.280 --> 0:09:10.400 paper worth of over a trillion dollars. Doing the foundational models, 0:09:10.520 --> 0:09:14.040 doing the stuff that's underneath everything, this stuff right at 0:09:14.080 --> 0:09:17.960 the center of it. They're training the models, they're operating 0:09:17.960 --> 0:09:21.600 the models that these other smaller companies apps. They could 0:09:21.600 --> 0:09:23.880 be for business, they could be for individuals, or be 0:09:23.960 --> 0:09:28.120 AI therapists, they'll be AI life coaches. I mean, there's 0:09:28.280 --> 0:09:32.000 plenty of opportunities for small companies. But as I'm saying that, 0:09:32.120 --> 0:09:36.240 I realize that Google Meta Open Ai a three hundred 0:09:36.240 --> 0:09:38.880 million dollar company right now on paper, you know, they 0:09:38.920 --> 0:09:41.559 have their own life coaches. Some of them are working 0:09:41.600 --> 0:09:45.440 on therapists or companions and all this, and so there 0:09:45.520 --> 0:09:49.000 is room in that first category for companies to break through. 0:09:49.400 --> 0:09:52.720 But I do worry too that not only will big 0:09:52.760 --> 0:09:58.240 tech dominate that second, that to me more central, foundational category, 0:09:58.280 --> 0:10:02.320 but they too can pick off folks in that first category. 0:10:02.559 --> 0:10:05.840 Either they'll do it better and beat the competition because 0:10:05.840 --> 0:10:09.040 they have more money, more ability to train these AI models, 0:10:09.600 --> 0:10:11.880 or if they just buy it. Look at Washington right now, 0:10:12.320 --> 0:10:16.640 the FTC has the case against Meta because Mark Zuckerberg 0:10:17.200 --> 0:10:20.160 fear at Instagram, so he bought Instagram. And did he 0:10:20.640 --> 0:10:23.120 buy it because he wanted to grow it and make 0:10:23.160 --> 0:10:24.800 it into its own product, or do you buy it 0:10:24.840 --> 0:10:27.360 because he was scared he was going to threaten Facebook 0:10:27.360 --> 0:10:28.480 and he wanted to defang it. 0:10:29.200 --> 0:10:33.199 I mean, we've had these cycles where bigness ultimately gets 0:10:33.480 --> 0:10:36.600 tackled by the government if this assumption that it becomes 0:10:36.960 --> 0:10:40.439 predatory or inhibits the growth of competitors. But I want 0:10:40.480 --> 0:10:43.920 to go from the corporate side to AI itself. You 0:10:44.000 --> 0:10:46.800 make the point in your book that AI started to 0:10:46.840 --> 0:10:50.200 develop and then in the nineteen seventies it sort of stopped. 0:10:50.800 --> 0:10:53.560 You describe it as sort of the AI winter. Could 0:10:53.600 --> 0:10:54.600 you walk us through that. 0:10:55.559 --> 0:10:57.360 One of the fun things in doing this book was 0:10:57.840 --> 0:10:59.839 how did we get here? How long have we been 0:11:00.120 --> 0:11:02.360 playing around with AI? So it dates back to at 0:11:02.440 --> 0:11:06.720 least the nineteen fifties. The term AI artificial intelligence was 0:11:06.760 --> 0:11:10.760 coined in the late nineteen fifties and it's just funny 0:11:10.800 --> 0:11:15.120 to read the optimism, the wild optimism of those behind 0:11:15.240 --> 0:11:19.040 AI in the fifties and sixties. They were convinced amazing 0:11:19.080 --> 0:11:21.360 things were right around the next corner. But you know, 0:11:21.559 --> 0:11:24.840 AI was right around the next corner for about seventy years. 0:11:25.040 --> 0:11:28.880 And part of that is computers weren't strong enough. Part 0:11:28.920 --> 0:11:31.720 of that is we need digital data, and we didn't 0:11:31.720 --> 0:11:34.880 really have that much digital data until people started posting 0:11:34.920 --> 0:11:37.840 and migrating so the Internet in the mid nineteen nineties. 0:11:37.880 --> 0:11:42.400 But part of it was a monumentally wrong turn. It 0:11:42.520 --> 0:11:45.880 was an amazing academic at Cornell who came up with 0:11:45.920 --> 0:11:49.880 this idea of a neural network, this idea that computers 0:11:49.920 --> 0:11:53.000 would learn in the fashion of a human. They would 0:11:53.000 --> 0:11:56.320 read material, they'd get feedback, and they would improve that 0:11:56.360 --> 0:12:00.120 way rather than coding line by line by line by line. 0:12:00.280 --> 0:12:05.800 Professor was mocked, and for forty or fifty years that 0:12:05.960 --> 0:12:10.240 approach was considered the wrong approach among the academics, the 0:12:10.280 --> 0:12:14.680 prevailing thought among computer scientists. It really wasn't until the 0:12:14.720 --> 0:12:18.480 twenty tens were what people are now calling machine learning, 0:12:18.679 --> 0:12:23.559 deep learning, neural networks. These models, these systems that learn 0:12:23.720 --> 0:12:28.120 through training and improve through feedback. Wasn't until the mid 0:12:28.160 --> 0:12:33.000 twenty tens that that really took on as the best approach, 0:12:33.200 --> 0:12:35.920 and in fact, that is why we're where we're at 0:12:36.040 --> 0:12:40.880 right now, machine learning, neural networks, that's the basis for chat, 0:12:40.920 --> 0:12:44.080 GBT and all the chatbots and other systems people are 0:12:44.160 --> 0:12:46.840 using to draw their photos or make little video clips. 0:12:47.440 --> 0:12:50.320 How much of this was just a function of computer 0:12:50.520 --> 0:12:53.480 power catching up with the theory. 0:12:53.600 --> 0:12:58.040 I think the academics dismissed neural networks as just the 0:12:58.080 --> 0:13:02.240 wrong approach in thet Hence, even if we had taken 0:13:02.280 --> 0:13:07.240 the neural networks approach, computers weren't nearly as powerful. In fact, 0:13:07.360 --> 0:13:12.640 the godfather of machine learning, Jeffrey Hinton, professor at University 0:13:12.640 --> 0:13:15.840 of Toronto, he made the point like no one imagined 0:13:15.880 --> 0:13:19.040 that these machines would be a million times more powerful, 0:13:19.280 --> 0:13:21.840 But you know, with an exponential gets more and more 0:13:21.880 --> 0:13:24.839 powerful with each passing year, to the point where they 0:13:24.920 --> 0:13:28.560 could handle billions of operations a second, which when you 0:13:28.600 --> 0:13:31.839 go to a chetbot, you know, chat, GPT, Claude, you 0:13:31.880 --> 0:13:34.480 know Google's Gemini, it doesn't make a difference. There's like 0:13:35.440 --> 0:13:39.480 billions of processes that are going on. You say hello, 0:13:39.600 --> 0:13:42.480 and it says hello back to you. You say, tell 0:13:42.520 --> 0:13:45.440 me who New Gingrich is, and it goes and searches 0:13:45.520 --> 0:13:48.480 and spits out a three or four sevens answer. There's 0:13:48.600 --> 0:13:52.920 like billions and billions of operations. And today's computers are 0:13:52.920 --> 0:13:56.680 strong enough. Today's computer chips are powerful enough to make 0:13:56.720 --> 0:14:00.600 that happen. But that wasn't true thirty years ago, certainly, 0:14:00.800 --> 0:14:02.120 probably not even ten years ago. 0:14:02.559 --> 0:14:03.240 I was surprised. 0:14:03.320 --> 0:14:07.280 I co chaired a working group on Alzheimer's around two 0:14:07.360 --> 0:14:11.079 thousand and seven, and I realized that to really do 0:14:11.200 --> 0:14:15.040 brain science, that the brain actually has about the same 0:14:15.120 --> 0:14:18.319 number of synapses as the number of scars in the universe. 0:14:18.559 --> 0:14:24.400 There's an astonishing identity, and that literally investing in computer 0:14:24.520 --> 0:14:29.720 power was central to advancing brain science because we literally 0:14:29.800 --> 0:14:33.920 at that point did not have the processing capability to 0:14:34.080 --> 0:14:37.360 truly analyze in depth all the things that happened in 0:14:37.360 --> 0:14:41.080 the brain. And now twenty years later, we're beginning to 0:14:41.160 --> 0:14:45.200 move into a zone where we can have that kind 0:14:45.240 --> 0:14:45.960 of activity. 0:14:46.800 --> 0:14:51.520 The brain helped us understand neural networks just like the brain. 0:14:51.560 --> 0:14:55.080 I think it's eighty six billion neurons. These neural networks 0:14:55.080 --> 0:14:57.200 try to emulate that, but I'm convinced these models are 0:14:57.200 --> 0:15:00.640 going to help us better understand the brain sciences. Interesting 0:15:00.680 --> 0:15:05.720 with AI because people get with science, there's specialties and subspecialties, 0:15:05.720 --> 0:15:09.800 and they all have their own vocabulary, and it's really 0:15:09.880 --> 0:15:14.600 hard to go across specialties across subspecialties. But with these 0:15:14.640 --> 0:15:18.760 neural networks, they're finding that they could have them read everything. 0:15:18.920 --> 0:15:21.560 The main model I write about in the book, it 0:15:21.600 --> 0:15:25.040 was trained on a trillion and a half words. You 0:15:25.160 --> 0:15:28.160 need thousands of human beings reading NonStop for their whole 0:15:28.200 --> 0:15:31.200 life to get close to that. And so these models 0:15:31.200 --> 0:15:35.160 trained for science, they could read studies in every discipline 0:15:35.200 --> 0:15:38.600 and make connections that no human being can possibly make. 0:15:38.640 --> 0:15:40.600 And that's one of the things I find most promising 0:15:40.640 --> 0:15:43.400 about this that some of the answers are right there, 0:15:43.880 --> 0:15:46.800 just we haven't made the connections, and something like AI 0:15:47.440 --> 0:15:49.560 can help us make those connections. 0:16:04.800 --> 0:16:07.760 As this thing began to develop, and as computers became 0:16:07.800 --> 0:16:13.360 more powerful and more central, and Nvidia emerged as an 0:16:13.360 --> 0:16:18.960 amazingly key producer of the most advanced ships, why didn't 0:16:19.000 --> 0:16:21.080 other folks like Intel do that? I mean, there are 0:16:21.240 --> 0:16:23.800 companies who were doing pretty well and then and video 0:16:23.920 --> 0:16:26.280 just explodes in its capacity. 0:16:27.200 --> 0:16:30.120 This is one of those kind of Columbus is looking 0:16:30.160 --> 0:16:33.000 for spices in the Far East and discovers America. So 0:16:33.280 --> 0:16:37.280 in Vidia created the most powerful graphics chip and so 0:16:37.400 --> 0:16:40.040 that was their specialty for playing, you know, video games. 0:16:40.360 --> 0:16:43.800 And it just turns out that these graphics chips are 0:16:43.920 --> 0:16:47.600 perfect for training AI because they could just do billions 0:16:47.600 --> 0:16:51.400 of operations in parallel, and that's what you need for AI. 0:16:51.520 --> 0:16:54.640 It's that complicated math, but it's just a lot of 0:16:54.680 --> 0:16:57.560 it at once. And so these in Vidia chips, they 0:16:57.600 --> 0:16:59.440 were just kind of Johnny on the spot. They were 0:16:59.480 --> 0:17:04.479 the perfect chip for training these neural networks. They're perfect 0:17:04.520 --> 0:17:07.879 for machine learning. But the chip world is not the 0:17:07.920 --> 0:17:11.080 software world. It moves very very slowly. And there's all 0:17:11.160 --> 0:17:15.280 these innovative startups out there that are trying to create 0:17:15.400 --> 0:17:19.520 chips that are designed specifically for AI. We still don't 0:17:19.560 --> 0:17:24.639 really have cutting edge AI specific chips. Maybe some of 0:17:24.680 --> 0:17:27.159 the memory should be on the chip. There's different ideas 0:17:27.200 --> 0:17:30.359 out there, and I guarantee you ten years from now, 0:17:30.880 --> 0:17:36.080 there will be innovative chips supplanting in Vidio's graphic chips, 0:17:36.119 --> 0:17:38.960 but in Video might create that chip. It might still 0:17:39.000 --> 0:17:41.680 be with Nvidia, but you know, the H one hundreds 0:17:41.720 --> 0:17:44.719 and the chips that are the mainstay right now of 0:17:44.960 --> 0:17:48.360 artificial intelligence. They're going to be replaced. It just takes 0:17:48.760 --> 0:17:52.600 time to develop, tests, produce, mass, produce. 0:17:52.960 --> 0:17:56.160 There had shindo been a belief that Moore's law would 0:17:56.160 --> 0:18:00.080 disappear and that you wouldn't have continuous doubling of capability 0:18:00.400 --> 0:18:03.640 because as the chips got smaller and smaller, the challenge 0:18:03.640 --> 0:18:06.959 of dealing with heat became greater and greater. But somehow 0:18:07.080 --> 0:18:09.879 we've leaped past all that. What we're seeing now seems 0:18:09.880 --> 0:18:12.119 to be. You could not have projected this in the 0:18:12.200 --> 0:18:14.280 nineteen eighties exactly. 0:18:14.400 --> 0:18:16.040 I was writing for the New York Times in mid 0:18:16.080 --> 0:18:19.680 two thousands, and that was the prediction. But I don't 0:18:19.680 --> 0:18:22.439 have to tell you science is amazing, technology is amazing 0:18:22.480 --> 0:18:24.959 that you know. Right after chat GPT came out, there 0:18:25.040 --> 0:18:27.200 was the doomers their call, those who were worried about 0:18:27.280 --> 0:18:30.639 laser eyed robots subjugating humanity, the kind of stuff I 0:18:30.680 --> 0:18:33.080 think is born in Hollywood in the media coverage, But 0:18:33.119 --> 0:18:35.479 that was the fear out there, like let's pause this, 0:18:35.880 --> 0:18:38.160 Let's have a six month pause so we could catch up, 0:18:38.240 --> 0:18:40.959 Like just not going to happen. It's called innovation. You 0:18:41.000 --> 0:18:46.200 can't slow science, you can't slow discovery. The answer is 0:18:46.240 --> 0:18:50.159 to manage it, to make sure that it's more of 0:18:50.200 --> 0:18:53.680 a positive than a negative. All technologies cut both ways. 0:18:53.720 --> 0:18:57.160 They're both positive and negatives. You know, cars changed our society, 0:18:57.200 --> 0:19:01.280 but cars kill thirty five forty thousand people year in America. 0:19:01.359 --> 0:19:04.439 They cause pollution. So all technologies are like that. So 0:19:04.880 --> 0:19:06.920 my great wish if I had a magic wand it 0:19:06.920 --> 0:19:09.520 would be like, just let's deal with this, folks. Let's 0:19:09.560 --> 0:19:12.040 try to make sure that AI is more of a 0:19:12.080 --> 0:19:14.959 positive than a negative. But you know, there's a lot 0:19:15.000 --> 0:19:16.680 of other issues in the world right now that are 0:19:16.680 --> 0:19:18.280 distracting us from that. 0:19:19.160 --> 0:19:23.000 You point out that from Google's perspective, the first great 0:19:23.119 --> 0:19:28.080 use of AI was improving targeting us as customers and 0:19:28.200 --> 0:19:30.720 figuring out what we really like and making sure the 0:19:30.760 --> 0:19:33.080 ads come up that we would be interested in. So 0:19:33.200 --> 0:19:36.560 just a fascinating way that things evolve in a way 0:19:36.600 --> 0:19:39.280 you couldn't probably have predicted if you're sitting in some 0:19:39.880 --> 0:19:41.639 academic place drawing up a plan. 0:19:42.680 --> 0:19:46.080 Right, So machine learning to maximize the cash register basically, 0:19:46.119 --> 0:19:48.360 So yeah, I mean, let's give Google credit. Though they 0:19:48.440 --> 0:19:51.200 got that machine learning artificial intelligence was going to be 0:19:51.280 --> 0:19:53.600 really powerful, so they would use in the early days. 0:19:53.600 --> 0:19:56.560 That you gave one example of kind of more efficiently 0:19:56.760 --> 0:20:00.359 matching ads to searches, but also to deal with horble 0:20:00.480 --> 0:20:03.600 Google searches they are spelling mistakes, they kind of understand 0:20:03.640 --> 0:20:06.359 the context and help smooth them out. The funny thing 0:20:06.359 --> 0:20:09.440 about artificial intelligence is all of us have been using 0:20:09.480 --> 0:20:12.199 AI for a long long time. I am using example 0:20:12.200 --> 0:20:14.840 of Google Search, but there's Google Translate that's been around 0:20:14.880 --> 0:20:19.439 since twenty fifteen or so. That's artificial intelligence. You go 0:20:19.520 --> 0:20:23.000 to Netflix or Spotify and they recommend you might like 0:20:23.040 --> 0:20:25.960 this movie, you might like this song. That's AI. The 0:20:26.040 --> 0:20:29.280 difference with the release in twenty twenty two of chat 0:20:29.359 --> 0:20:32.800 cheapt from Open AI was that we could talk with it. 0:20:32.800 --> 0:20:35.800 It wasn't a product behind the glass. It was something 0:20:35.800 --> 0:20:39.000 that we can actually link to and use and chat 0:20:39.040 --> 0:20:41.760 with it. I think that's what changed everything. The idea 0:20:41.800 --> 0:20:44.320 that we can actually see it working and play with 0:20:44.400 --> 0:20:48.040 it made us really stand up and take attention again. 0:20:48.119 --> 0:20:50.240 With she was sixty, we do a lot of polling. 0:20:50.840 --> 0:20:53.760 We now run all the polling questions through chat GPT. 0:20:54.359 --> 0:20:57.280 I use sometimes use chat cheapt My favorite it is 0:20:57.320 --> 0:21:01.600 called Claud from Anthropic. The same way it's my go 0:21:01.680 --> 0:21:04.080 to editor. People have to understand how they use this. 0:21:04.320 --> 0:21:06.879 AI is a co pilot. It's not like you know, 0:21:07.000 --> 0:21:11.000 type in make me a Martin Scorsese movie hit entery, 0:21:11.000 --> 0:21:13.800 and you're gonna have it. You have to be the creative. 0:21:13.880 --> 0:21:15.480 You have to give it the ideas. If you ask 0:21:15.520 --> 0:21:18.400 it to write something, it'll be flat, it's not gonna 0:21:18.400 --> 0:21:20.680 be particularly good. It'll read like kind of like a 0:21:20.720 --> 0:21:24.880 press release or boring report. But if you use it 0:21:25.000 --> 0:21:28.240 as your companion and you're the creative. So what i 0:21:28.359 --> 0:21:31.159 U is for is I'm struggling with a paragraph. I 0:21:31.160 --> 0:21:33.760 don't like this transition. Help me out with this sentence 0:21:33.840 --> 0:21:35.919 right in five different ways. And it's never like I 0:21:35.960 --> 0:21:38.000 cut and paste and say, oh that's the sense, but like, 0:21:38.160 --> 0:21:39.919 oh that's a good idea. I didn't think of that. 0:21:39.960 --> 0:21:42.040 Oh that's an interesting word. Let me use that. I 0:21:42.160 --> 0:21:45.760 routinely now before I hand something in, I have it edit. 0:21:46.040 --> 0:21:49.520 It finds typo's, it finds mistakes, hey in bold, give 0:21:49.560 --> 0:21:52.800 me suggestions from improving it again. Often I just ignore 0:21:52.800 --> 0:21:56.800 their suggestions. But it's a really powerful tool to help 0:21:56.840 --> 0:21:59.560 you refine to make what you're working on, make your 0:21:59.680 --> 0:22:04.280 work better. It's your copilot, it's not your digital employee creator. 0:22:05.240 --> 0:22:06.800 But it's a pretty powerful coil. 0:22:07.080 --> 0:22:10.960 So start at twenty twenty three. My role in the 0:22:11.000 --> 0:22:14.000 second half of the nineties was the skeptic around dot com. 0:22:14.040 --> 0:22:16.560 It's like, Okay, the Internet's going to be incredible, but 0:22:16.640 --> 0:22:19.480 you're not going to get fabulously wealthy overnight startups, and 0:22:19.520 --> 0:22:21.800 in fact most of them went Now I was ready 0:22:21.800 --> 0:22:25.439 to be a skeptic. It's magic, it's sorcery. I mean, 0:22:25.480 --> 0:22:27.480 the first few times I used it. The first thing 0:22:27.520 --> 0:22:30.120 I did write me a five thousand word book proposal 0:22:30.359 --> 0:22:32.800 to sell a book on AI. And you know, I mean, 0:22:32.920 --> 0:22:36.280 it wasn't particularly well written, but it's far better read 0:22:36.280 --> 0:22:39.159 than I am. It has a far better memory than 0:22:39.200 --> 0:22:41.960 I do. It was just so useful as a lousy 0:22:42.040 --> 0:22:44.640 first draft, but it gave me so many ideas and 0:22:45.200 --> 0:22:48.040 sped up the process. If I had to start from scratch, 0:22:48.440 --> 0:22:50.479 it would have been so much harder, like as opposed 0:22:50.480 --> 0:22:53.360 to like, oh that's a good structure, that's a good idea. Yeah, yeah, 0:22:53.440 --> 0:22:56.080 I do need to stress that I'm a journalist. It's 0:22:56.080 --> 0:22:58.840 like Hey, we have this amazing product, and you go 0:22:59.080 --> 0:23:00.880 use the product and like, yeah, maybe in two years 0:23:00.880 --> 0:23:03.199 you have an amazing product, but right now it's buggy 0:23:03.240 --> 0:23:06.320