WEBVTT - OpenAI Part2: Ilya Dreams of AGI 0:00:01.520 --> 0:00:04.400 I want to start by talking about a dream, the 0:00:04.519 --> 0:00:09.160 dream of building an artificial mind. It's something people have 0:00:09.240 --> 0:00:13.320 imagined and written about for decades, humans working together to 0:00:13.360 --> 0:00:18.040 construct a new entity more powerful than ourselves, and some 0:00:18.120 --> 0:00:20.799 researchers think this dream may be within reach. 0:00:21.560 --> 0:00:25.040 The day will come when the digital brains that live 0:00:25.079 --> 0:00:29.600 inside our computers will become as good and even better 0:00:30.120 --> 0:00:34.800 than our own biological brains. Computers will become smarter than us. 0:00:36.200 --> 0:00:42.200 Vicoll such ANAI and AGI artificial general intelligence. 0:00:43.320 --> 0:00:46.040 That's Ilia set Skiver, one of the co founders of 0:00:46.080 --> 0:00:48.720 open ai and at ted talk he gave last year, 0:00:49.360 --> 0:00:52.000 and he often sounds like a religious mystic when he 0:00:52.040 --> 0:00:56.080 talks about the future of artificial intelligence. But right now 0:00:56.160 --> 0:01:00.000 he's just talking about this quest to build artificial general intelligence, 0:01:00.800 --> 0:01:03.760 an AI that can think and solve a variety of problems. 0:01:03.840 --> 0:01:07.800 Like a person. It could switch between playing games, solving 0:01:07.840 --> 0:01:12.640 science problems, creating beautiful art, and driving a car. Open 0:01:12.680 --> 0:01:16.200 AI's goal is to build AGI. It's a pretty out 0:01:16.200 --> 0:01:18.640 there idea in the AI world, or at least it 0:01:18.720 --> 0:01:22.679 used to be. Ilia frames AGI as this almost mystical, 0:01:22.840 --> 0:01:29.440 momentous leap forward, like Prometheus channeling fire and the consequences 0:01:29.480 --> 0:01:33.080 will be huge. It will usher us into technological glory 0:01:33.280 --> 0:01:37.039 and at the same time into chaos. In this tape 0:01:37.080 --> 0:01:40.440 from the documentary film I Human, he sounds certain of 0:01:40.480 --> 0:01:43.240 the tidal waves that will come now. 0:01:43.319 --> 0:01:46.039 AI is a great thing because AI will solve all 0:01:46.400 --> 0:01:49.360 the problems that you have today. If you solve employment, 0:01:49.880 --> 0:01:56.440 if you solve disease, it will solve poverty. But it 0:01:56.520 --> 0:02:00.880 will also create new problems. The problem fake news is 0:02:00.920 --> 0:02:05.160 going to be a million times worse. Cyber attacts will 0:02:05.160 --> 0:02:10.000 become much more extreme if people have totally automated AI weapons. 0:02:11.560 --> 0:02:16.120 Ilia is an incredibly accomplished AI researcher. Before open Ai, 0:02:16.160 --> 0:02:18.880 he worked at Google, and he has several passions that 0:02:18.919 --> 0:02:23.000 I see as a celebration of being human. He plays 0:02:23.000 --> 0:02:26.639 the piano, he draws and paints. One of his paintings 0:02:26.680 --> 0:02:30.000 hangs in the Open Ai office. It's a flower in 0:02:30.040 --> 0:02:33.480 the shape of the company's logo. At the same time, 0:02:33.639 --> 0:02:37.399 he's also hyper focused on his AI research. He told 0:02:37.400 --> 0:02:40.920 a reporter once, I lead a very simple life. I 0:02:41.000 --> 0:02:43.440 go to work, then I go home. I don't do 0:02:43.520 --> 0:02:46.200 much else. There are a lot of social activities one 0:02:46.200 --> 0:02:48.480 can engage in lots of events one could go to, 0:02:48.840 --> 0:02:52.000 which I don't. He spends a lot of time looking 0:02:52.040 --> 0:02:55.440 at the current trajectory of AI and extrapolating to try 0:02:55.440 --> 0:03:00.239 to predict the future. In particular, Ilia is worried about 0:03:00.280 --> 0:03:04.239 what happens if AGI gets its own desires and its 0:03:04.280 --> 0:03:08.880 own goals. You can hear this dreamy equality in his voice. 0:03:09.240 --> 0:03:12.960 It's not said it's going to actively hate humans and 0:03:13.000 --> 0:03:17.120 want to harm them, but it is going to be 0:03:17.160 --> 0:03:20.560 too powerful. And I think a good analogy would be 0:03:21.080 --> 0:03:25.720 the way humans treat animals. It loves to be hate animals, 0:03:26.240 --> 0:03:29.639 but when the time comes to build a highway between 0:03:29.639 --> 0:03:33.119 two cities, we are not asking the animals for permission. 0:03:34.120 --> 0:03:35.720 Imagining it this huge. 0:03:35.840 --> 0:03:43.080 Unstoppable force, and I think it's pretty likely the entire 0:03:43.120 --> 0:03:45.920 surface of the areas will be covered with solid panels 0:03:45.920 --> 0:03:46.680 and data centers. 0:03:48.200 --> 0:03:50.600 I want to pause here for a minute. This is 0:03:50.640 --> 0:03:55.080 a really intense, powerful image that we are creating some 0:03:55.240 --> 0:03:58.160 new kind of being that would view us with interest, 0:03:58.560 --> 0:04:02.080 but ultimately with indiffer friends, like the way we look 0:04:02.080 --> 0:04:06.120 at deer. What strikes me most in this audio is 0:04:06.480 --> 0:04:10.200 Ilia's tone of voice isn't one of fear. It sounds 0:04:10.240 --> 0:04:16.240 more like awe Ilia imagines an AGI that we create 0:04:16.720 --> 0:04:19.080 that would be likely to bulldoze over us in order 0:04:19.120 --> 0:04:23.080 to reach its own desires. It's a dramatic vision, hard 0:04:23.120 --> 0:04:26.680 to really grasp, and it has a religious quality in 0:04:26.760 --> 0:04:31.680 its conception of a supernatural, all powerful entity. I should 0:04:31.720 --> 0:04:35.600 mention this is all totally theoretical. We are still nowhere 0:04:35.600 --> 0:04:40.160 close to AGI. Open AI's best efforts are statistical models 0:04:40.160 --> 0:04:44.400 that convincingly mimic humans, and mimicry is a far cry 0:04:44.480 --> 0:04:48.479 from AI that can think for itself. Still, open Ai 0:04:48.839 --> 0:04:51.240 wants to do this, and do it right, and do 0:04:51.320 --> 0:04:54.960 it first. Here's Sam Altman testifying in front of Congress 0:04:54.960 --> 0:04:55.839 in twenty twenty three. 0:04:56.720 --> 0:04:59.960 My worst fears are that we cause significant We feel 0:05:00.160 --> 0:05:03.359 the technology the industry caused significant harm to the world. 0:05:04.440 --> 0:05:06.960 It's why we started the company. I think if this 0:05:07.040 --> 0:05:09.600 technology goes wrong, it can go quite wrong. 0:05:17.080 --> 0:05:21.200 You're listening to Foundering. I'm your host, Ellen Hewitt, and 0:05:21.279 --> 0:05:24.200 in this episode will take you inside the messy and 0:05:24.360 --> 0:05:28.640 idealistic early years of open Ai. We'll discuss this dream 0:05:28.839 --> 0:05:32.960 of building all powerful agi. It's important because this is 0:05:33.000 --> 0:05:36.240 the destination that open ai is speeding toward. It's this 0:05:36.360 --> 0:05:41.760 generation's race to the moon. We'll discuss how ai technology 0:05:41.920 --> 0:05:46.840 changed dramatically and quickly, and how that change made this 0:05:47.000 --> 0:05:51.120 dream of AGI feel closer than ever before. In just 0:05:51.160 --> 0:05:53.560 a few years, it went from an eccentric idea that 0:05:53.600 --> 0:05:56.960 people were scoffing at to a milestone some experts think 0:05:57.000 --> 0:06:00.360 could happen within a few years. Sam Altman has even 0:06:00.400 --> 0:06:05.320 suggested twenty twenty eight. And we'll examine the compromises open 0:06:05.360 --> 0:06:08.840 Ai made in its pursuit of this dream. At first, 0:06:08.880 --> 0:06:11.840 the company made promises to share its research widely and 0:06:11.880 --> 0:06:16.000 to not be corrupted by for profit incentives. But once 0:06:16.040 --> 0:06:19.160 their technology began to advance and it looked like there 0:06:19.200 --> 0:06:21.480 was serious power to be had, they. 0:06:21.320 --> 0:06:22.040 Made a u turn. 0:06:23.160 --> 0:06:26.760 Then this pivotal moment, kareemed into a power struggle at 0:06:26.800 --> 0:06:29.640 open Ai, and Sam Altman took charge. 0:06:31.160 --> 0:06:31.960 We'll be right back. 0:06:36.279 --> 0:06:40.480 We'll start In twenty fifteen, open Ai had just been founded. 0:06:41.080 --> 0:06:43.520 It had a commitment from Elon Musk for a billion 0:06:43.560 --> 0:06:47.040 dollars in funding, plus some money from other donors as well. 0:06:47.800 --> 0:06:52.000 It was this small, scrappy research lab. Sam and Elon 0:06:52.080 --> 0:06:55.040 weren't around much in the early days. At the time, 0:06:55.160 --> 0:06:58.599 Sam was actually still running y Combinator, the startup accelerator, 0:06:59.200 --> 0:07:01.440 but he was beginning to position himself as a thought 0:07:01.520 --> 0:07:05.719 leader in the AI space. In particular, he was talking 0:07:05.760 --> 0:07:11.880 about AI doomsday scenarios. In twenty fifteen, he declared on 0:07:11.920 --> 0:07:16.560 his blog development of superhuman machine intelligence is probably the 0:07:16.600 --> 0:07:20.840 greatest threat to the continued existence of humanity. He also 0:07:20.840 --> 0:07:24.240 wrote that AI could destroy every human in the universe. 0:07:25.080 --> 0:07:27.160 Here he is at a tech event the same year, 0:07:27.640 --> 0:07:29.360 referencing the founding of open AI. 0:07:30.280 --> 0:07:33.000 I actually just agreed to fund a company doing AI 0:07:33.080 --> 0:07:36.560 safety research. You know, I think AI will probably, like 0:07:36.640 --> 0:07:39.040 most likely sort of lead to the end of the world. 0:07:39.040 --> 0:07:42.200 But in the meantime there will be great companies created 0:07:42.240 --> 0:07:43.720 with serious machine learning. 0:07:44.320 --> 0:07:46.440 I want to talk about that comment for a second. 0:07:47.120 --> 0:07:50.720 He's saying AI might kill us all, and he's asking 0:07:50.800 --> 0:07:54.240 us to trust his conclusions as an expert, but he's 0:07:54.280 --> 0:07:58.920 also being glib about making money along the way. In 0:07:58.960 --> 0:08:02.320 the beginning of open AA, Sam and Elon weren't around 0:08:02.360 --> 0:08:06.040 for the day to day. They were out glad handing, recruiting, 0:08:06.120 --> 0:08:09.080 and talking to journalists. They would pop in once a 0:08:09.080 --> 0:08:12.320 week or so to get progress updates. In those days, 0:08:12.440 --> 0:08:15.080 Sam would swoop into a conversation then leave. 0:08:15.720 --> 0:08:19.160 So he struck me as very very sharp, incisive, and 0:08:19.200 --> 0:08:22.760 also superficient with this time. When the conversation is done, 0:08:22.880 --> 0:08:23.560 it's done. 0:08:24.000 --> 0:08:27.040 That's Peter Abiel, a researcher who worked at open ai 0:08:27.160 --> 0:08:29.880 in its first two years. He says, in the early 0:08:30.000 --> 0:08:33.400 days open ai looked like a typical startup. They didn't 0:08:33.400 --> 0:08:36.120 even have an office for a while. They met in 0:08:36.160 --> 0:08:38.120 the home of one of the co founders. 0:08:38.840 --> 0:08:42.680 When we started out late twenty fifteen early twenty sixteen, 0:08:42.840 --> 0:08:46.520 it was in Greg Brockman's apartment in San Francisco Mission District. 0:08:46.720 --> 0:08:49.440 It was you know, we're sitting essentially on a couch, 0:08:49.559 --> 0:08:53.120 at a kitchen counter and on a bed, and that's 0:08:53.160 --> 0:08:56.760 pretty much that's where the work is getting done. It's 0:08:57.360 --> 0:08:59.199 it's kind of crazy to think that, you know, that's 0:08:59.240 --> 0:09:02.680 where something big got startup. We just had twenty of 0:09:02.679 --> 0:09:06.880 the world's best AI researchers together, really focused on trying 0:09:06.880 --> 0:09:08.840 to get some things done and I've never been done before. 0:09:09.559 --> 0:09:12.400 In the absence of Sam and Elon, the main leaders 0:09:12.440 --> 0:09:15.319 were two people who aren't famous but who will play 0:09:15.360 --> 0:09:21.320 a mammoth role later on. There's Greg whose apartment served 0:09:21.360 --> 0:09:24.800 as their office, and Ilia, the research scientist we heard 0:09:24.800 --> 0:09:27.439 from at the beginning of the episode. You can think 0:09:27.480 --> 0:09:30.760 of Greg as the workhorse in charge of business operations, 0:09:31.120 --> 0:09:35.040 and Ilia as the AI genius, and together they ran 0:09:35.240 --> 0:09:40.080 open AI. Peter remembers going on weekly walks with Ilia 0:09:40.120 --> 0:09:44.120 around the neighborhood in San Francisco, talking about big picture stuff, 0:09:44.640 --> 0:09:47.400 asking themselves, are we working on the right problems? 0:09:48.120 --> 0:09:51.400 I feel like he he just kind of saw AI 0:09:51.760 --> 0:09:54.960 what it could be doing, could be capable of, more 0:09:55.040 --> 0:09:58.560 clearly and earlier than anybody else. He's seeing it more 0:09:58.559 --> 0:10:02.480 optimistic than everybody else. He would come up with analogies like, okay, 0:10:02.520 --> 0:10:04.640 and the ONAL network is just a computer program. It's 0:10:04.640 --> 0:10:06.720 just a circuit. We're just programming it differently. 0:10:07.840 --> 0:10:10.120 Greg meanwhile, was grinding away. 0:10:10.760 --> 0:10:14.120 Greg is somebody who can just apply himself. He can 0:10:14.240 --> 0:10:17.720 just you know, keep working and keep working and keep working. 0:10:18.520 --> 0:10:21.040 I've seen some people like that, but very few. 0:10:22.200 --> 0:10:25.560 Even after open AI moved out of Greg's apartment, he 0:10:25.640 --> 0:10:29.760 still practically lived at the office. One former employee said 0:10:30.000 --> 0:10:32.120 he would be hunched over his laptop when they showed 0:10:32.160 --> 0:10:34.679 up to work in the morning and was still tapping 0:10:34.679 --> 0:10:38.240 away when they went home at night. Years later, when 0:10:38.280 --> 0:10:41.440 Greg got married, he even held a civil ceremony in 0:10:41.480 --> 0:10:45.400 the office with a big backdrop made of flowers, again 0:10:45.600 --> 0:10:48.600 in the shape of the open ai logo. The ring 0:10:48.679 --> 0:10:52.480 bear was a robot hand and the officiant was Ilia. 0:10:54.280 --> 0:10:57.800 When Greg and Ilia joined open ai, they didn't need money. 0:10:58.640 --> 0:11:02.079 Ilia had sold a company to Google, and Greg owned 0:11:02.120 --> 0:11:04.719 a lot of shares of stripe, and that company was 0:11:04.760 --> 0:11:08.960 worth tens of billions of dollars. In Silicon Valley, people 0:11:09.040 --> 0:11:11.839 usually create startups because they think they can build a 0:11:11.920 --> 0:11:17.040 lucrative business, but open ai was a nonprofit. Greg and 0:11:17.080 --> 0:11:21.400 Ilia were motivated by this dream. Here's Reid Hoffman, one 0:11:21.440 --> 0:11:23.040 of the earliest donors to open ai. 0:11:23.960 --> 0:11:27.600 There was no equity upside for that. Initial crew was like, look, 0:11:27.640 --> 0:11:31.400 we're doing this for humanity. 0:11:30.160 --> 0:11:33.959 Doing it for humanity. Open ai talks like this all 0:11:34.000 --> 0:11:37.679 the time. Their website says, our mission is to ensure 0:11:37.840 --> 0:11:43.120 that artificial general intelligence benefits all of humanity. Okay, So 0:11:43.200 --> 0:11:47.000 it's well known that Silicon Valley loves grandiose mission statements. 0:11:47.400 --> 0:11:52.120 We work wanted to elevate the world's consciousness, but open 0:11:52.160 --> 0:11:56.959 AI's mission statement is even more sweeping. And it has 0:11:57.000 --> 0:12:02.079 this overtone of altruism. When Sam talks about the company's work, 0:12:02.559 --> 0:12:06.840 he often discusses potential disasters. Here he is with Rebecca 0:12:06.920 --> 0:12:11.960 Jarvis on ABC. His voice sounds grave. Again, he's positioning 0:12:12.040 --> 0:12:14.040 himself as a thought leader in this space. 0:12:15.160 --> 0:12:18.640 So what is the worst possible outcome? 0:12:19.400 --> 0:12:21.600 There's like a set of very bad outcomes. One thing 0:12:21.600 --> 0:12:25.240 I'm particularly worried about is that these models could be 0:12:25.360 --> 0:12:29.480 used for large scale disinformation. I am worried that these systems, 0:12:29.559 --> 0:12:32.440 now that they're getting better at writing computer code, could 0:12:32.520 --> 0:12:34.720 be used for offensive cyber attacks. 0:12:35.200 --> 0:12:38.800 But you raise an important point, which is the humans 0:12:38.840 --> 0:12:41.560 who are in control of the machine right now also 0:12:41.840 --> 0:12:43.160 have a huge amount of power. 0:12:43.720 --> 0:12:47.400 We do worry a lot about authoritarian governments developing this. 0:12:47.920 --> 0:12:52.800 Putin has himself said, whoever wins this artificial intelligence race 0:12:53.240 --> 0:12:55.280 is essentially the controller of humankind. 0:12:55.280 --> 0:12:56.080 Do you agree with that? 0:12:56.559 --> 0:12:57.800 So that was a chilling. 0:12:57.480 --> 0:13:03.880 Statement for sure, Sam saying this stuff is so valuable 0:13:03.960 --> 0:13:07.800 that global superpowers are going to fight over it. The 0:13:07.840 --> 0:13:10.200 cynical take is that if you make what you're working 0:13:10.200 --> 0:13:13.800 on sound really important, you attract a lot more attention 0:13:14.280 --> 0:13:17.320 and money. We'll talk more about this dynamic in the 0:13:17.320 --> 0:13:21.920 next episode. In open AI's early years, their humanity saving 0:13:22.000 --> 0:13:26.120 plan wasn't that clear. Their strategy was a bit scattered. 0:13:26.800 --> 0:13:27.640 Here's Peter again. 0:13:28.520 --> 0:13:32.080 We looked at robotics, did some work there. We looked 0:13:32.080 --> 0:13:34.720 at simulated robotics, did a bunch of work there. We 0:13:34.840 --> 0:13:39.160 looked at digital agents that navigate the web and do 0:13:39.240 --> 0:13:43.320 all kinds of tasks online, like booking flights. We looked 0:13:43.320 --> 0:13:44.360 at video games. 0:13:45.000 --> 0:13:47.440 Open Ai said one of its first goals would be 0:13:47.520 --> 0:13:50.720 to build a robot butler which could set and clear 0:13:50.760 --> 0:13:57.240 a table, kind of like the maid on the Jetsons. 0:13:55.320 --> 0:13:59.000 Coming Sir, Hey William Sir. 0:14:00.720 --> 0:14:03.680 The company also built a robot arm that could solve 0:14:03.720 --> 0:14:07.040 a Rubik's Cube single handedly, and they put a lot 0:14:07.080 --> 0:14:10.160 of effort into building bots that could play Doda two, 0:14:10.800 --> 0:14:14.640 a massively popular multiplayer video game. They imagined that the 0:14:14.640 --> 0:14:17.439 complexity of the game environment could lead to an AI 0:14:17.559 --> 0:14:21.120 that could better navigate the real world. Here's someone testing 0:14:21.160 --> 0:14:21.480 the bot. 0:14:22.200 --> 0:14:23.760 The bot is good. The bot is better than I 0:14:23.760 --> 0:14:24.680 could have ever imagined. 0:14:24.920 --> 0:14:28.440 Those Doda bots even competed against professional players. 0:14:28.760 --> 0:14:31.640 Already up stop the actions kicking off sopen Ai will 0:14:31.680 --> 0:14:32.880 claim first blood. 0:14:32.960 --> 0:14:34.800 No tawn is to be caught out here. 0:14:34.840 --> 0:14:38.320 A bot that could play Doda was technically impressive, but 0:14:38.400 --> 0:14:41.440 it didn't look very impressive to the average person, and 0:14:41.480 --> 0:14:44.600 the commercial applications for these products were not immediately clear. 0:14:45.360 --> 0:14:48.560 Here's how one former employee put it. We were doing 0:14:48.680 --> 0:14:51.840 random stuff and seeing what would happen. There were not 0:14:51.880 --> 0:14:55.600 really defined goals. Sometimes it felt like there was a 0:14:55.640 --> 0:14:58.840 big gap between what was being built and what was 0:14:58.880 --> 0:15:03.200 being imagined. People would spend their days programming bots that 0:15:03.240 --> 0:15:06.440 played video games. Then they would sit around the lunch 0:15:06.480 --> 0:15:11.880 table and talk about saving humanity. The prevailing wisdom in 0:15:11.920 --> 0:15:15.120 the AI world was that in order to make something powerful, 0:15:15.560 --> 0:15:19.280 you sometimes have to start with something trivial. Video games 0:15:19.320 --> 0:15:22.520 and robot maids would pave the way to self driving 0:15:22.560 --> 0:15:27.320 cars and cancer curing AI. Internally, at open AI, they 0:15:27.360 --> 0:15:31.320 sometimes compared themselves to the Manhattan Project, the team given 0:15:31.320 --> 0:15:35.040 the mission to create the first atomic bomb, and they 0:15:35.080 --> 0:15:40.640 meant it as a good thing, ambitious and important. Here's 0:15:40.640 --> 0:15:43.600 how one former employee described it to me. It's an 0:15:43.720 --> 0:15:46.800 arms race. They all want to make the first AGI. 0:15:47.080 --> 0:15:49.720 They believe they can do it best. I didn't see 0:15:49.720 --> 0:15:52.400 a lot of fear of AI itself. I just saw 0:15:52.520 --> 0:15:57.320 excitement to build AI. Back in twenty fifteen, AI looked 0:15:57.360 --> 0:16:01.000 pretty different from today. It was weaker and harder to train. 0:16:02.200 --> 0:16:04.880 At the time, the major breakthrough was that a bot 0:16:04.960 --> 0:16:07.680 had been able to beat the world's best player and Go, 0:16:08.360 --> 0:16:12.440 a complex strategy board game from China, but that AI 0:16:12.560 --> 0:16:16.840 could only play Go, it couldn't do anything else. Here's 0:16:17.000 --> 0:16:20.800 Orn Etzioni, a computer science professor and the former research 0:16:20.840 --> 0:16:22.600 director for an AI institute. 0:16:23.160 --> 0:16:26.480 The thing about these is these were narrow systems, very 0:16:26.840 --> 0:16:30.400 highly targeted. So the system that played Go couldn't even 0:16:30.440 --> 0:16:34.640 play chess, certainly could not cross the street or understand language. 0:16:34.800 --> 0:16:39.120 And the system that understood airfare fluctuations and predicted very 0:16:39.160 --> 0:16:42.320 well whether airfares were going up or down could not 0:16:43.240 --> 0:16:46.720 handle text either. Right, So basically, every time you had 0:16:46.760 --> 0:16:49.880 an application, you'd have to train up a new system. 0:16:50.440 --> 0:16:52.760 And this took a long time, took a lot of 0:16:52.840 --> 0:16:53.760 labeled data, etc. 0:16:55.040 --> 0:17:01.520 But then came a major breakthrough in AI technology. In 0:17:01.560 --> 0:17:05.000 twenty seventeen, a group of researchers from Google Brain published 0:17:05.000 --> 0:17:08.520 a paper called Attention Is All You Need, and in 0:17:08.560 --> 0:17:11.560 it they describe a new kind of AI architecture called 0:17:11.880 --> 0:17:17.280 the transformer, and the transformer did something huge. At the time, 0:17:17.480 --> 0:17:21.480 AI systems needed to be fed very specific data. Each 0:17:21.520 --> 0:17:24.280 piece of data had to be labeled this is correct, 0:17:24.440 --> 0:17:29.240 this is incorrect. Spam not spam, cancer not cancer. But 0:17:29.320 --> 0:17:33.800 the transformer allowed AI to take in messy, unlabeled data, 0:17:33.880 --> 0:17:36.760 and it could actually do so even more efficiently than expected, 0:17:37.280 --> 0:17:42.159 using less computing power than before. Now these transformer based 0:17:42.200 --> 0:17:45.560 models could just teach themselves in a way. It was 0:17:45.640 --> 0:17:47.480 like if you wanted to teach a kid to read, 0:17:47.800 --> 0:17:49.639 and you used to have to hire a tutor to 0:17:49.760 --> 0:17:53.520 sit there with flashcards, and now instead you could just 0:17:53.720 --> 0:17:56.560 let the kid run through a library and they would 0:17:56.600 --> 0:18:01.040 emerge knowing how to read and write. This was, as 0:18:01.080 --> 0:18:04.959 one investor described to me, a surprising and bitter realization 0:18:05.760 --> 0:18:08.280 that the best AI would come not from the most 0:18:08.280 --> 0:18:13.000 specialized training techniques, but from whichever had the most data. Peter, 0:18:13.280 --> 0:18:17.399 the early open AI employee, says Ilia immediately saw its promise. 0:18:18.119 --> 0:18:22.760 Ilia's reaction was pretty affirmative right away. It's like, this 0:18:22.880 --> 0:18:25.320 is something special we need to be we need to 0:18:25.320 --> 0:18:28.080 be looking at this. This seems a big breakthrough. 0:18:28.880 --> 0:18:31.359 Even in the early days of open AI, Ilia had 0:18:31.400 --> 0:18:35.280 always had this hunch that big advances in AI wouldn't 0:18:35.280 --> 0:18:38.800 come from some specific tweak or new invention, but just 0:18:38.840 --> 0:18:42.719 from more data pouring more and more fuel into the engine. 0:18:43.440 --> 0:18:47.040 And now Ilia had the research that backed up his hypothesis. 0:18:47.440 --> 0:18:51.920 Here's Oren again, Ilia from open ai is known as 0:18:51.960 --> 0:18:55.520 the person who said, it's the data, and it's the 0:18:55.560 --> 0:18:58.720 amount of data, and if we just scale that up 0:18:58.800 --> 0:19:03.120 tremendously of magnitude much much more, we're going to achieve 0:19:03.280 --> 0:19:07.080 what we need. That was not the common perception and 0:19:07.119 --> 0:19:09.680 some very smart and very famous people. I don't want 0:19:09.720 --> 0:19:16.159 to cast aspersions, but certainly I'm I'm not that smarter 0:19:16.200 --> 0:19:18.840 with that famous, but I'm one of the AI experts 0:19:18.880 --> 0:19:20.439 who did not see that coming. 0:19:22.000 --> 0:19:25.840 Because of Ilia, open Ai started experimenting with the transformer. 0:19:26.200 --> 0:19:28.120 They were one of the earliest companies to do so. 0:19:28.840 --> 0:19:33.359 They made models with the now familiar acronym GPT Generative 0:19:33.520 --> 0:19:38.080 pre Trained Transformer, and in particular, they started experimenting with 0:19:38.160 --> 0:19:42.400 how the transformer performed with written words, because they could 0:19:42.440 --> 0:19:47.679 basically feed the model anything written any book, newspaper, article, 0:19:48.040 --> 0:19:51.960 reddit posts, blogs. Humans have spent a lot of time 0:19:52.119 --> 0:19:55.880 writing things down, and those words now had another purpose, 0:19:56.400 --> 0:20:01.159 training data. The Internet wasn't created to train, but in 0:20:01.200 --> 0:20:05.479 the end that may become its legacy. Open aiyes, models 0:20:05.480 --> 0:20:08.800 got better and better at generating text, and they weren't 0:20:08.800 --> 0:20:10.960 limited to just one field of knowledge. 0:20:11.640 --> 0:20:15.720 The amazing thing about these GPT systems is that they're 0:20:15.840 --> 0:20:19.560 very broad. They are actually generalists. You can ask them 0:20:19.600 --> 0:20:24.200 about virtually any topic and they'll produce surprisingly good answers. 0:20:24.240 --> 0:20:28.720 And that's because they've been trained on effectively the entire 0:20:28.920 --> 0:20:32.679 or at least an approximation of the entire corpus of 0:20:32.760 --> 0:20:36.800 text that's available to humanity, billions of billions of sentences, 0:20:37.119 --> 0:20:40.560 all the books you've read, all the documents, the memos, 0:20:40.680 --> 0:20:46.159 the silliness, Harry Potter fan fiction. It's all grist for 0:20:46.240 --> 0:20:50.120 the mill. And then once it's read all that, it's 0:20:50.240 --> 0:20:53.320 remarkably general. So for the first time we would have 0:20:53.359 --> 0:20:56.959 a system that you could ask it about anything and 0:20:57.000 --> 0:20:59.720 it would give you a surprisingly intelligent answer. So we 0:20:59.760 --> 0:21:03.800 went from narrow AI to a kind of general or 0:21:03.880 --> 0:21:04.440 broad AI. 0:21:05.359 --> 0:21:07.959 Through the massive amount of writing that they were feeding 0:21:08.000 --> 0:21:11.560 into their models, open ai found they could create AI 0:21:11.680 --> 0:21:15.639 that was much much better at forming convincing sounding responses 0:21:15.640 --> 0:21:19.480 to questions. In fact, at some point they started to 0:21:19.520 --> 0:21:23.680 worry it was maybe too good. When OpenAI announced its 0:21:23.760 --> 0:21:27.359 language model GBT two, they initially decided not to share 0:21:27.359 --> 0:21:30.400 the model more openly because they were concerned it could 0:21:30.400 --> 0:21:34.240 be dangerous. Here's Peter. He had by that time left 0:21:34.280 --> 0:21:36.920 open ai to start his own company, but he remembers 0:21:36.960 --> 0:21:38.320 the day of the release. 0:21:38.560 --> 0:21:40.600 It was just obvious that it had a much better 0:21:40.720 --> 0:21:43.639 understanding of language than anything that had been trained before. 0:21:43.960 --> 0:21:48.240 Its release was indeed accompanied by a lot of I 0:21:48.320 --> 0:21:52.520 guess great marketing or caution or combination of both. It 0:21:52.640 --> 0:21:56.959 was headline that's too dangerous to be released, and so 0:21:57.320 --> 0:21:59.280 I think it was probably one of the first projects 0:21:59.320 --> 0:22:02.480 where open i decided to not release some of the 0:22:02.520 --> 0:22:06.480 work because all of a sudden, the thinking had become, well, 0:22:06.560 --> 0:22:09.200 what if it's something is so powerful that people could 0:22:09.200 --> 0:22:11.680 go misuse it in ways that we can't control. 0:22:12.960 --> 0:22:15.199 As soon as open ai had a product that was 0:22:15.320 --> 0:22:19.160 actually powerful, they started rethinking their openness. 0:22:19.640 --> 0:22:24.000 Open Ai started with that name where to open really 0:22:24.040 --> 0:22:26.959 stood for everything's going to be open sourced built, you know, 0:22:27.040 --> 0:22:28.360 anybody else can build on it. 0:22:29.040 --> 0:22:32.760 Openness was a crucial part of the company's brand when 0:22:32.800 --> 0:22:36.119 they were founded, Sam told the journalist Stephen Levy, it 0:22:36.160 --> 0:22:39.919 will just be open source and usable by everyone. He 0:22:40.000 --> 0:22:42.840 also told him that their AI would be freely owned 0:22:42.880 --> 0:22:47.080 by the world. Open source software in its broadest sense, 0:22:47.440 --> 0:22:49.600 means that the source code is made available to the 0:22:49.600 --> 0:22:52.960 public freely, and that anyone can tweak the code and 0:22:52.960 --> 0:22:57.320 distribute it themselves. But the company soon started walking back 0:22:57.359 --> 0:22:58.119 those commitments. 0:22:58.880 --> 0:23:03.199 Obviously that evolved over time into something that is not 0:23:03.320 --> 0:23:06.280 so open source, if open source at all for anything. 0:23:06.800 --> 0:23:09.360 I mean, it's definitely not open sourcing. It's hits work. 0:23:09.440 --> 0:23:14.320 Right now, that open source ethos seemed to fade away. 0:23:14.440 --> 0:23:17.080 Here's Sam giving a talk in Munich in twenty twenty three. 0:23:18.119 --> 0:23:20.760 I'm curious, if we stay on the same like GPT 0:23:20.880 --> 0:23:23.320 two to three to four trajectory for five and six, 0:23:23.760 --> 0:23:25.440 how many of you would like us to open source 0:23:25.440 --> 0:23:30.719 GPT six the day we finished training it. Wow, Well 0:23:30.720 --> 0:23:32.359 we're not going to do that. But that's interesting that. 0:23:35.280 --> 0:23:39.480 Honestly, Sam sounds pretty arrogant here. He knows open AI 0:23:39.640 --> 0:23:43.040 started off with promises of being open source. Now he's 0:23:43.080 --> 0:23:46.880 pulling the audience about open sourcing models and immediately dismissing 0:23:46.880 --> 0:23:51.119 their response. Over the years, Sam has subtly changed the 0:23:51.200 --> 0:23:55.840 meaning of openness. It's become fuzzier here he is at 0:23:55.840 --> 0:23:57.080 a VC firm. 0:23:57.440 --> 0:23:59.880 So I think that is that's what we call open ai, 0:24:00.040 --> 0:24:02.119 open AI. We want this to be open technology made 0:24:02.119 --> 0:24:02.840 available to. 0:24:02.800 --> 0:24:09.760 Everyone, open technology made available to everyone. He says it 0:24:09.800 --> 0:24:13.240 so plainly, as though that's obvious what open means, But 0:24:13.320 --> 0:24:16.800 his definition strikes me as so vague that it's essentially meaningless. 0:24:17.400 --> 0:24:22.359 I mean, Google Search is available to everyone. It seems 0:24:22.400 --> 0:24:24.960 like open ai was happy to let people guess what 0:24:25.040 --> 0:24:29.240 they meant by open. In an internal email just months 0:24:29.280 --> 0:24:33.159 after it was founded, Ilia wrote, as we get closer 0:24:33.200 --> 0:24:35.960