1 00:00:01,520 --> 00:00:04,400 Speaker 1: I want to start by talking about a dream, the 2 00:00:04,519 --> 00:00:09,160 Speaker 1: dream of building an artificial mind. It's something people have 3 00:00:09,240 --> 00:00:13,320 Speaker 1: imagined and written about for decades, humans working together to 4 00:00:13,360 --> 00:00:18,040 Speaker 1: construct a new entity more powerful than ourselves, and some 5 00:00:18,120 --> 00:00:20,799 Speaker 1: researchers think this dream may be within reach. 6 00:00:21,560 --> 00:00:25,040 Speaker 2: The day will come when the digital brains that live 7 00:00:25,079 --> 00:00:29,600 Speaker 2: inside our computers will become as good and even better 8 00:00:30,120 --> 00:00:34,800 Speaker 2: than our own biological brains. Computers will become smarter than us. 9 00:00:36,200 --> 00:00:42,200 Speaker 2: Vicoll such ANAI and AGI artificial general intelligence. 10 00:00:43,320 --> 00:00:46,040 Speaker 1: That's Ilia set Skiver, one of the co founders of 11 00:00:46,080 --> 00:00:48,720 Speaker 1: open ai and at ted talk he gave last year, 12 00:00:49,360 --> 00:00:52,000 Speaker 1: and he often sounds like a religious mystic when he 13 00:00:52,040 --> 00:00:56,080 Speaker 1: talks about the future of artificial intelligence. But right now 14 00:00:56,160 --> 00:01:00,000 Speaker 1: he's just talking about this quest to build artificial general intelligence, 15 00:01:00,800 --> 00:01:03,760 Speaker 1: an AI that can think and solve a variety of problems. 16 00:01:03,840 --> 00:01:07,800 Speaker 1: Like a person. It could switch between playing games, solving 17 00:01:07,840 --> 00:01:12,640 Speaker 1: science problems, creating beautiful art, and driving a car. Open 18 00:01:12,680 --> 00:01:16,200 Speaker 1: AI's goal is to build AGI. It's a pretty out 19 00:01:16,200 --> 00:01:18,640 Speaker 1: there idea in the AI world, or at least it 20 00:01:18,720 --> 00:01:22,679 Speaker 1: used to be. Ilia frames AGI as this almost mystical, 21 00:01:22,840 --> 00:01:29,440 Speaker 1: momentous leap forward, like Prometheus channeling fire and the consequences 22 00:01:29,480 --> 00:01:33,080 Speaker 1: will be huge. It will usher us into technological glory 23 00:01:33,280 --> 00:01:37,039 Speaker 1: and at the same time into chaos. In this tape 24 00:01:37,080 --> 00:01:40,440 Speaker 1: from the documentary film I Human, he sounds certain of 25 00:01:40,480 --> 00:01:43,240 Speaker 1: the tidal waves that will come now. 26 00:01:43,319 --> 00:01:46,039 Speaker 3: AI is a great thing because AI will solve all 27 00:01:46,400 --> 00:01:49,360 Speaker 3: the problems that you have today. If you solve employment, 28 00:01:49,880 --> 00:01:56,440 Speaker 3: if you solve disease, it will solve poverty. But it 29 00:01:56,520 --> 00:02:00,880 Speaker 3: will also create new problems. The problem fake news is 30 00:02:00,920 --> 00:02:05,160 Speaker 3: going to be a million times worse. Cyber attacts will 31 00:02:05,160 --> 00:02:10,000 Speaker 3: become much more extreme if people have totally automated AI weapons. 32 00:02:11,560 --> 00:02:16,120 Speaker 1: Ilia is an incredibly accomplished AI researcher. Before open Ai, 33 00:02:16,160 --> 00:02:18,880 Speaker 1: he worked at Google, and he has several passions that 34 00:02:18,919 --> 00:02:23,000 Speaker 1: I see as a celebration of being human. He plays 35 00:02:23,000 --> 00:02:26,639 Speaker 1: the piano, he draws and paints. One of his paintings 36 00:02:26,680 --> 00:02:30,000 Speaker 1: hangs in the Open Ai office. It's a flower in 37 00:02:30,040 --> 00:02:33,480 Speaker 1: the shape of the company's logo. At the same time, 38 00:02:33,639 --> 00:02:37,399 Speaker 1: he's also hyper focused on his AI research. He told 39 00:02:37,400 --> 00:02:40,920 Speaker 1: a reporter once, I lead a very simple life. I 40 00:02:41,000 --> 00:02:43,440 Speaker 1: go to work, then I go home. I don't do 41 00:02:43,520 --> 00:02:46,200 Speaker 1: much else. There are a lot of social activities one 42 00:02:46,200 --> 00:02:48,480 Speaker 1: can engage in lots of events one could go to, 43 00:02:48,840 --> 00:02:52,000 Speaker 1: which I don't. He spends a lot of time looking 44 00:02:52,040 --> 00:02:55,440 Speaker 1: at the current trajectory of AI and extrapolating to try 45 00:02:55,440 --> 00:03:00,239 Speaker 1: to predict the future. In particular, Ilia is worried about 46 00:03:00,280 --> 00:03:04,239 Speaker 1: what happens if AGI gets its own desires and its 47 00:03:04,280 --> 00:03:08,880 Speaker 1: own goals. You can hear this dreamy equality in his voice. 48 00:03:09,240 --> 00:03:12,960 Speaker 3: It's not said it's going to actively hate humans and 49 00:03:13,000 --> 00:03:17,120 Speaker 3: want to harm them, but it is going to be 50 00:03:17,160 --> 00:03:20,560 Speaker 3: too powerful. And I think a good analogy would be 51 00:03:21,080 --> 00:03:25,720 Speaker 3: the way humans treat animals. It loves to be hate animals, 52 00:03:26,240 --> 00:03:29,639 Speaker 3: but when the time comes to build a highway between 53 00:03:29,639 --> 00:03:33,119 Speaker 3: two cities, we are not asking the animals for permission. 54 00:03:34,120 --> 00:03:35,720 Speaker 4: Imagining it this huge. 55 00:03:35,840 --> 00:03:43,080 Speaker 3: Unstoppable force, and I think it's pretty likely the entire 56 00:03:43,120 --> 00:03:45,920 Speaker 3: surface of the areas will be covered with solid panels 57 00:03:45,920 --> 00:03:46,680 Speaker 3: and data centers. 58 00:03:48,200 --> 00:03:50,600 Speaker 1: I want to pause here for a minute. This is 59 00:03:50,640 --> 00:03:55,080 Speaker 1: a really intense, powerful image that we are creating some 60 00:03:55,240 --> 00:03:58,160 Speaker 1: new kind of being that would view us with interest, 61 00:03:58,560 --> 00:04:02,080 Speaker 1: but ultimately with indiffer friends, like the way we look 62 00:04:02,080 --> 00:04:06,120 Speaker 1: at deer. What strikes me most in this audio is 63 00:04:06,480 --> 00:04:10,200 Speaker 1: Ilia's tone of voice isn't one of fear. It sounds 64 00:04:10,240 --> 00:04:16,240 Speaker 1: more like awe Ilia imagines an AGI that we create 65 00:04:16,720 --> 00:04:19,080 Speaker 1: that would be likely to bulldoze over us in order 66 00:04:19,120 --> 00:04:23,080 Speaker 1: to reach its own desires. It's a dramatic vision, hard 67 00:04:23,120 --> 00:04:26,680 Speaker 1: to really grasp, and it has a religious quality in 68 00:04:26,760 --> 00:04:31,680 Speaker 1: its conception of a supernatural, all powerful entity. I should 69 00:04:31,720 --> 00:04:35,600 Speaker 1: mention this is all totally theoretical. We are still nowhere 70 00:04:35,600 --> 00:04:40,160 Speaker 1: close to AGI. Open AI's best efforts are statistical models 71 00:04:40,160 --> 00:04:44,400 Speaker 1: that convincingly mimic humans, and mimicry is a far cry 72 00:04:44,480 --> 00:04:48,479 Speaker 1: from AI that can think for itself. Still, open Ai 73 00:04:48,839 --> 00:04:51,240 Speaker 1: wants to do this, and do it right, and do 74 00:04:51,320 --> 00:04:54,960 Speaker 1: it first. Here's Sam Altman testifying in front of Congress 75 00:04:54,960 --> 00:04:55,839 Speaker 1: in twenty twenty three. 76 00:04:56,720 --> 00:04:59,960 Speaker 4: My worst fears are that we cause significant We feel 77 00:05:00,160 --> 00:05:03,359 Speaker 4: the technology the industry caused significant harm to the world. 78 00:05:04,440 --> 00:05:06,960 Speaker 4: It's why we started the company. I think if this 79 00:05:07,040 --> 00:05:09,600 Speaker 4: technology goes wrong, it can go quite wrong. 80 00:05:17,080 --> 00:05:21,200 Speaker 1: You're listening to Foundering. I'm your host, Ellen Hewitt, and 81 00:05:21,279 --> 00:05:24,200 Speaker 1: in this episode will take you inside the messy and 82 00:05:24,360 --> 00:05:28,640 Speaker 1: idealistic early years of open Ai. We'll discuss this dream 83 00:05:28,839 --> 00:05:32,960 Speaker 1: of building all powerful agi. It's important because this is 84 00:05:33,000 --> 00:05:36,240 Speaker 1: the destination that open ai is speeding toward. It's this 85 00:05:36,360 --> 00:05:41,760 Speaker 1: generation's race to the moon. We'll discuss how ai technology 86 00:05:41,920 --> 00:05:46,840 Speaker 1: changed dramatically and quickly, and how that change made this 87 00:05:47,000 --> 00:05:51,120 Speaker 1: dream of AGI feel closer than ever before. In just 88 00:05:51,160 --> 00:05:53,560 Speaker 1: a few years, it went from an eccentric idea that 89 00:05:53,600 --> 00:05:56,960 Speaker 1: people were scoffing at to a milestone some experts think 90 00:05:57,000 --> 00:06:00,360 Speaker 1: could happen within a few years. Sam Altman has even 91 00:06:00,400 --> 00:06:05,320 Speaker 1: suggested twenty twenty eight. And we'll examine the compromises open 92 00:06:05,360 --> 00:06:08,840 Speaker 1: Ai made in its pursuit of this dream. At first, 93 00:06:08,880 --> 00:06:11,840 Speaker 1: the company made promises to share its research widely and 94 00:06:11,880 --> 00:06:16,000 Speaker 1: to not be corrupted by for profit incentives. But once 95 00:06:16,040 --> 00:06:19,160 Speaker 1: their technology began to advance and it looked like there 96 00:06:19,200 --> 00:06:21,480 Speaker 1: was serious power to be had, they. 97 00:06:21,320 --> 00:06:22,040 Speaker 3: Made a u turn. 98 00:06:23,160 --> 00:06:26,760 Speaker 1: Then this pivotal moment, kareemed into a power struggle at 99 00:06:26,800 --> 00:06:29,640 Speaker 1: open Ai, and Sam Altman took charge. 100 00:06:31,160 --> 00:06:31,960 Speaker 4: We'll be right back. 101 00:06:36,279 --> 00:06:40,480 Speaker 1: We'll start In twenty fifteen, open Ai had just been founded. 102 00:06:41,080 --> 00:06:43,520 Speaker 1: It had a commitment from Elon Musk for a billion 103 00:06:43,560 --> 00:06:47,040 Speaker 1: dollars in funding, plus some money from other donors as well. 104 00:06:47,800 --> 00:06:52,000 Speaker 1: It was this small, scrappy research lab. Sam and Elon 105 00:06:52,080 --> 00:06:55,040 Speaker 1: weren't around much in the early days. At the time, 106 00:06:55,160 --> 00:06:58,599 Speaker 1: Sam was actually still running y Combinator, the startup accelerator, 107 00:06:59,200 --> 00:07:01,440 Speaker 1: but he was beginning to position himself as a thought 108 00:07:01,520 --> 00:07:05,719 Speaker 1: leader in the AI space. In particular, he was talking 109 00:07:05,760 --> 00:07:11,880 Speaker 1: about AI doomsday scenarios. In twenty fifteen, he declared on 110 00:07:11,920 --> 00:07:16,560 Speaker 1: his blog development of superhuman machine intelligence is probably the 111 00:07:16,600 --> 00:07:20,840 Speaker 1: greatest threat to the continued existence of humanity. He also 112 00:07:20,840 --> 00:07:24,240 Speaker 1: wrote that AI could destroy every human in the universe. 113 00:07:25,080 --> 00:07:27,160 Speaker 1: Here he is at a tech event the same year, 114 00:07:27,640 --> 00:07:29,360 Speaker 1: referencing the founding of open AI. 115 00:07:30,280 --> 00:07:33,000 Speaker 4: I actually just agreed to fund a company doing AI 116 00:07:33,080 --> 00:07:36,560 Speaker 4: safety research. You know, I think AI will probably, like 117 00:07:36,640 --> 00:07:39,040 Speaker 4: most likely sort of lead to the end of the world. 118 00:07:39,040 --> 00:07:42,200 Speaker 4: But in the meantime there will be great companies created 119 00:07:42,240 --> 00:07:43,720 Speaker 4: with serious machine learning. 120 00:07:44,320 --> 00:07:46,440 Speaker 1: I want to talk about that comment for a second. 121 00:07:47,120 --> 00:07:50,720 Speaker 1: He's saying AI might kill us all, and he's asking 122 00:07:50,800 --> 00:07:54,240 Speaker 1: us to trust his conclusions as an expert, but he's 123 00:07:54,280 --> 00:07:58,920 Speaker 1: also being glib about making money along the way. In 124 00:07:58,960 --> 00:08:02,320 Speaker 1: the beginning of open AA, Sam and Elon weren't around 125 00:08:02,360 --> 00:08:06,040 Speaker 1: for the day to day. They were out glad handing, recruiting, 126 00:08:06,120 --> 00:08:09,080 Speaker 1: and talking to journalists. They would pop in once a 127 00:08:09,080 --> 00:08:12,320 Speaker 1: week or so to get progress updates. In those days, 128 00:08:12,440 --> 00:08:15,080 Speaker 1: Sam would swoop into a conversation then leave. 129 00:08:15,720 --> 00:08:19,160 Speaker 5: So he struck me as very very sharp, incisive, and 130 00:08:19,200 --> 00:08:22,760 Speaker 5: also superficient with this time. When the conversation is done, 131 00:08:22,880 --> 00:08:23,560 Speaker 5: it's done. 132 00:08:24,000 --> 00:08:27,040 Speaker 1: That's Peter Abiel, a researcher who worked at open ai 133 00:08:27,160 --> 00:08:29,880 Speaker 1: in its first two years. He says, in the early 134 00:08:30,000 --> 00:08:33,400 Speaker 1: days open ai looked like a typical startup. They didn't 135 00:08:33,400 --> 00:08:36,120 Speaker 1: even have an office for a while. They met in 136 00:08:36,160 --> 00:08:38,120 Speaker 1: the home of one of the co founders. 137 00:08:38,840 --> 00:08:42,680 Speaker 5: When we started out late twenty fifteen early twenty sixteen, 138 00:08:42,840 --> 00:08:46,520 Speaker 5: it was in Greg Brockman's apartment in San Francisco Mission District. 139 00:08:46,720 --> 00:08:49,440 Speaker 5: It was you know, we're sitting essentially on a couch, 140 00:08:49,559 --> 00:08:53,120 Speaker 5: at a kitchen counter and on a bed, and that's 141 00:08:53,160 --> 00:08:56,760 Speaker 5: pretty much that's where the work is getting done. It's 142 00:08:57,360 --> 00:08:59,199 Speaker 5: it's kind of crazy to think that, you know, that's 143 00:08:59,240 --> 00:09:02,680 Speaker 5: where something big got startup. We just had twenty of 144 00:09:02,679 --> 00:09:06,880 Speaker 5: the world's best AI researchers together, really focused on trying 145 00:09:06,880 --> 00:09:08,840 Speaker 5: to get some things done and I've never been done before. 146 00:09:09,559 --> 00:09:12,400 Speaker 1: In the absence of Sam and Elon, the main leaders 147 00:09:12,440 --> 00:09:15,319 Speaker 1: were two people who aren't famous but who will play 148 00:09:15,360 --> 00:09:21,320 Speaker 1: a mammoth role later on. There's Greg whose apartment served 149 00:09:21,360 --> 00:09:24,800 Speaker 1: as their office, and Ilia, the research scientist we heard 150 00:09:24,800 --> 00:09:27,439 Speaker 1: from at the beginning of the episode. You can think 151 00:09:27,480 --> 00:09:30,760 Speaker 1: of Greg as the workhorse in charge of business operations, 152 00:09:31,120 --> 00:09:35,040 Speaker 1: and Ilia as the AI genius, and together they ran 153 00:09:35,240 --> 00:09:40,080 Speaker 1: open AI. Peter remembers going on weekly walks with Ilia 154 00:09:40,120 --> 00:09:44,120 Speaker 1: around the neighborhood in San Francisco, talking about big picture stuff, 155 00:09:44,640 --> 00:09:47,400 Speaker 1: asking themselves, are we working on the right problems? 156 00:09:48,120 --> 00:09:51,400 Speaker 5: I feel like he he just kind of saw AI 157 00:09:51,760 --> 00:09:54,960 Speaker 5: what it could be doing, could be capable of, more 158 00:09:55,040 --> 00:09:58,560 Speaker 5: clearly and earlier than anybody else. He's seeing it more 159 00:09:58,559 --> 00:10:02,480 Speaker 5: optimistic than everybody else. He would come up with analogies like, okay, 160 00:10:02,520 --> 00:10:04,640 Speaker 5: and the ONAL network is just a computer program. It's 161 00:10:04,640 --> 00:10:06,720 Speaker 5: just a circuit. We're just programming it differently. 162 00:10:07,840 --> 00:10:10,120 Speaker 1: Greg meanwhile, was grinding away. 163 00:10:10,760 --> 00:10:14,120 Speaker 5: Greg is somebody who can just apply himself. He can 164 00:10:14,240 --> 00:10:17,720 Speaker 5: just you know, keep working and keep working and keep working. 165 00:10:18,520 --> 00:10:21,040 Speaker 5: I've seen some people like that, but very few. 166 00:10:22,200 --> 00:10:25,560 Speaker 1: Even after open AI moved out of Greg's apartment, he 167 00:10:25,640 --> 00:10:29,760 Speaker 1: still practically lived at the office. One former employee said 168 00:10:30,000 --> 00:10:32,120 Speaker 1: he would be hunched over his laptop when they showed 169 00:10:32,160 --> 00:10:34,679 Speaker 1: up to work in the morning and was still tapping 170 00:10:34,679 --> 00:10:38,240 Speaker 1: away when they went home at night. Years later, when 171 00:10:38,280 --> 00:10:41,440 Speaker 1: Greg got married, he even held a civil ceremony in 172 00:10:41,480 --> 00:10:45,400 Speaker 1: the office with a big backdrop made of flowers, again 173 00:10:45,600 --> 00:10:48,600 Speaker 1: in the shape of the open ai logo. The ring 174 00:10:48,679 --> 00:10:52,480 Speaker 1: bear was a robot hand and the officiant was Ilia. 175 00:10:54,280 --> 00:10:57,800 Speaker 1: When Greg and Ilia joined open ai, they didn't need money. 176 00:10:58,640 --> 00:11:02,079 Speaker 1: Ilia had sold a company to Google, and Greg owned 177 00:11:02,120 --> 00:11:04,719 Speaker 1: a lot of shares of stripe, and that company was 178 00:11:04,760 --> 00:11:08,960 Speaker 1: worth tens of billions of dollars. In Silicon Valley, people 179 00:11:09,040 --> 00:11:11,839 Speaker 1: usually create startups because they think they can build a 180 00:11:11,920 --> 00:11:17,040 Speaker 1: lucrative business, but open ai was a nonprofit. Greg and 181 00:11:17,080 --> 00:11:21,400 Speaker 1: Ilia were motivated by this dream. Here's Reid Hoffman, one 182 00:11:21,440 --> 00:11:23,040 Speaker 1: of the earliest donors to open ai. 183 00:11:23,960 --> 00:11:27,600 Speaker 6: There was no equity upside for that. Initial crew was like, look, 184 00:11:27,640 --> 00:11:31,400 Speaker 6: we're doing this for humanity. 185 00:11:30,160 --> 00:11:33,959 Speaker 1: Doing it for humanity. Open ai talks like this all 186 00:11:34,000 --> 00:11:37,679 Speaker 1: the time. Their website says, our mission is to ensure 187 00:11:37,840 --> 00:11:43,120 Speaker 1: that artificial general intelligence benefits all of humanity. Okay, So 188 00:11:43,200 --> 00:11:47,000 Speaker 1: it's well known that Silicon Valley loves grandiose mission statements. 189 00:11:47,400 --> 00:11:52,120 Speaker 1: We work wanted to elevate the world's consciousness, but open 190 00:11:52,160 --> 00:11:56,959 Speaker 1: AI's mission statement is even more sweeping. And it has 191 00:11:57,000 --> 00:12:02,079 Speaker 1: this overtone of altruism. When Sam talks about the company's work, 192 00:12:02,559 --> 00:12:06,840 Speaker 1: he often discusses potential disasters. Here he is with Rebecca 193 00:12:06,920 --> 00:12:11,960 Speaker 1: Jarvis on ABC. His voice sounds grave. Again, he's positioning 194 00:12:12,040 --> 00:12:14,040 Speaker 1: himself as a thought leader in this space. 195 00:12:15,160 --> 00:12:18,640 Speaker 7: So what is the worst possible outcome? 196 00:12:19,400 --> 00:12:21,600 Speaker 4: There's like a set of very bad outcomes. One thing 197 00:12:21,600 --> 00:12:25,240 Speaker 4: I'm particularly worried about is that these models could be 198 00:12:25,360 --> 00:12:29,480 Speaker 4: used for large scale disinformation. I am worried that these systems, 199 00:12:29,559 --> 00:12:32,440 Speaker 4: now that they're getting better at writing computer code, could 200 00:12:32,520 --> 00:12:34,720 Speaker 4: be used for offensive cyber attacks. 201 00:12:35,200 --> 00:12:38,800 Speaker 7: But you raise an important point, which is the humans 202 00:12:38,840 --> 00:12:41,560 Speaker 7: who are in control of the machine right now also 203 00:12:41,840 --> 00:12:43,160 Speaker 7: have a huge amount of power. 204 00:12:43,720 --> 00:12:47,400 Speaker 4: We do worry a lot about authoritarian governments developing this. 205 00:12:47,920 --> 00:12:52,800 Speaker 7: Putin has himself said, whoever wins this artificial intelligence race 206 00:12:53,240 --> 00:12:55,280 Speaker 7: is essentially the controller of humankind. 207 00:12:55,280 --> 00:12:56,080 Speaker 5: Do you agree with that? 208 00:12:56,559 --> 00:12:57,800 Speaker 4: So that was a chilling. 209 00:12:57,480 --> 00:13:03,880 Speaker 1: Statement for sure, Sam saying this stuff is so valuable 210 00:13:03,960 --> 00:13:07,800 Speaker 1: that global superpowers are going to fight over it. The 211 00:13:07,840 --> 00:13:10,200 Speaker 1: cynical take is that if you make what you're working 212 00:13:10,200 --> 00:13:13,800 Speaker 1: on sound really important, you attract a lot more attention 213 00:13:14,280 --> 00:13:17,320 Speaker 1: and money. We'll talk more about this dynamic in the 214 00:13:17,320 --> 00:13:21,920 Speaker 1: next episode. In open AI's early years, their humanity saving 215 00:13:22,000 --> 00:13:26,120 Speaker 1: plan wasn't that clear. Their strategy was a bit scattered. 216 00:13:26,800 --> 00:13:27,640 Speaker 1: Here's Peter again. 217 00:13:28,520 --> 00:13:32,080 Speaker 5: We looked at robotics, did some work there. We looked 218 00:13:32,080 --> 00:13:34,720 Speaker 5: at simulated robotics, did a bunch of work there. We 219 00:13:34,840 --> 00:13:39,160 Speaker 5: looked at digital agents that navigate the web and do 220 00:13:39,240 --> 00:13:43,320 Speaker 5: all kinds of tasks online, like booking flights. We looked 221 00:13:43,320 --> 00:13:44,360 Speaker 5: at video games. 222 00:13:45,000 --> 00:13:47,440 Speaker 1: Open Ai said one of its first goals would be 223 00:13:47,520 --> 00:13:50,720 Speaker 1: to build a robot butler which could set and clear 224 00:13:50,760 --> 00:13:57,240 Speaker 1: a table, kind of like the maid on the Jetsons. 225 00:13:55,320 --> 00:13:59,000 Speaker 7: Coming Sir, Hey William Sir. 226 00:14:00,720 --> 00:14:03,680 Speaker 1: The company also built a robot arm that could solve 227 00:14:03,720 --> 00:14:07,040 Speaker 1: a Rubik's Cube single handedly, and they put a lot 228 00:14:07,080 --> 00:14:10,160 Speaker 1: of effort into building bots that could play Doda two, 229 00:14:10,800 --> 00:14:14,640 Speaker 1: a massively popular multiplayer video game. They imagined that the 230 00:14:14,640 --> 00:14:17,439 Speaker 1: complexity of the game environment could lead to an AI 231 00:14:17,559 --> 00:14:21,120 Speaker 1: that could better navigate the real world. Here's someone testing 232 00:14:21,160 --> 00:14:21,480 Speaker 1: the bot. 233 00:14:22,200 --> 00:14:23,760 Speaker 6: The bot is good. The bot is better than I 234 00:14:23,760 --> 00:14:24,680 Speaker 6: could have ever imagined. 235 00:14:24,920 --> 00:14:28,440 Speaker 1: Those Doda bots even competed against professional players. 236 00:14:28,760 --> 00:14:31,640 Speaker 7: Already up stop the actions kicking off sopen Ai will 237 00:14:31,680 --> 00:14:32,880 Speaker 7: claim first blood. 238 00:14:32,960 --> 00:14:34,800 Speaker 5: No tawn is to be caught out here. 239 00:14:34,840 --> 00:14:38,320 Speaker 1: A bot that could play Doda was technically impressive, but 240 00:14:38,400 --> 00:14:41,440 Speaker 1: it didn't look very impressive to the average person, and 241 00:14:41,480 --> 00:14:44,600 Speaker 1: the commercial applications for these products were not immediately clear. 242 00:14:45,360 --> 00:14:48,560 Speaker 1: Here's how one former employee put it. We were doing 243 00:14:48,680 --> 00:14:51,840 Speaker 1: random stuff and seeing what would happen. There were not 244 00:14:51,880 --> 00:14:55,600 Speaker 1: really defined goals. Sometimes it felt like there was a 245 00:14:55,640 --> 00:14:58,840 Speaker 1: big gap between what was being built and what was 246 00:14:58,880 --> 00:15:03,200 Speaker 1: being imagined. People would spend their days programming bots that 247 00:15:03,240 --> 00:15:06,440 Speaker 1: played video games. Then they would sit around the lunch 248 00:15:06,480 --> 00:15:11,880 Speaker 1: table and talk about saving humanity. The prevailing wisdom in 249 00:15:11,920 --> 00:15:15,120 Speaker 1: the AI world was that in order to make something powerful, 250 00:15:15,560 --> 00:15:19,280 Speaker 1: you sometimes have to start with something trivial. Video games 251 00:15:19,320 --> 00:15:22,520 Speaker 1: and robot maids would pave the way to self driving 252 00:15:22,560 --> 00:15:27,320 Speaker 1: cars and cancer curing AI. Internally, at open AI, they 253 00:15:27,360 --> 00:15:31,320 Speaker 1: sometimes compared themselves to the Manhattan Project, the team given 254 00:15:31,320 --> 00:15:35,040 Speaker 1: the mission to create the first atomic bomb, and they 255 00:15:35,080 --> 00:15:40,640 Speaker 1: meant it as a good thing, ambitious and important. Here's 256 00:15:40,640 --> 00:15:43,600 Speaker 1: how one former employee described it to me. It's an 257 00:15:43,720 --> 00:15:46,800 Speaker 1: arms race. They all want to make the first AGI. 258 00:15:47,080 --> 00:15:49,720 Speaker 1: They believe they can do it best. I didn't see 259 00:15:49,720 --> 00:15:52,400 Speaker 1: a lot of fear of AI itself. I just saw 260 00:15:52,520 --> 00:15:57,320 Speaker 1: excitement to build AI. Back in twenty fifteen, AI looked 261 00:15:57,360 --> 00:16:01,000 Speaker 1: pretty different from today. It was weaker and harder to train. 262 00:16:02,200 --> 00:16:04,880 Speaker 1: At the time, the major breakthrough was that a bot 263 00:16:04,960 --> 00:16:07,680 Speaker 1: had been able to beat the world's best player and Go, 264 00:16:08,360 --> 00:16:12,440 Speaker 1: a complex strategy board game from China, but that AI 265 00:16:12,560 --> 00:16:16,840 Speaker 1: could only play Go, it couldn't do anything else. Here's 266 00:16:17,000 --> 00:16:20,800 Speaker 1: Orn Etzioni, a computer science professor and the former research 267 00:16:20,840 --> 00:16:22,600 Speaker 1: director for an AI institute. 268 00:16:23,160 --> 00:16:26,480 Speaker 8: The thing about these is these were narrow systems, very 269 00:16:26,840 --> 00:16:30,400 Speaker 8: highly targeted. So the system that played Go couldn't even 270 00:16:30,440 --> 00:16:34,640 Speaker 8: play chess, certainly could not cross the street or understand language. 271 00:16:34,800 --> 00:16:39,120 Speaker 8: And the system that understood airfare fluctuations and predicted very 272 00:16:39,160 --> 00:16:42,320 Speaker 8: well whether airfares were going up or down could not 273 00:16:43,240 --> 00:16:46,720 Speaker 8: handle text either. Right, So basically, every time you had 274 00:16:46,760 --> 00:16:49,880 Speaker 8: an application, you'd have to train up a new system. 275 00:16:50,440 --> 00:16:52,760 Speaker 8: And this took a long time, took a lot of 276 00:16:52,840 --> 00:16:53,760 Speaker 8: labeled data, etc. 277 00:16:55,040 --> 00:17:01,520 Speaker 1: But then came a major breakthrough in AI technology. In 278 00:17:01,560 --> 00:17:05,000 Speaker 1: twenty seventeen, a group of researchers from Google Brain published 279 00:17:05,000 --> 00:17:08,520 Speaker 1: a paper called Attention Is All You Need, and in 280 00:17:08,560 --> 00:17:11,560 Speaker 1: it they describe a new kind of AI architecture called 281 00:17:11,880 --> 00:17:17,280 Speaker 1: the transformer, and the transformer did something huge. At the time, 282 00:17:17,480 --> 00:17:21,480 Speaker 1: AI systems needed to be fed very specific data. Each 283 00:17:21,520 --> 00:17:24,280 Speaker 1: piece of data had to be labeled this is correct, 284 00:17:24,440 --> 00:17:29,240 Speaker 1: this is incorrect. Spam not spam, cancer not cancer. But 285 00:17:29,320 --> 00:17:33,800 Speaker 1: the transformer allowed AI to take in messy, unlabeled data, 286 00:17:33,880 --> 00:17:36,760 Speaker 1: and it could actually do so even more efficiently than expected, 287 00:17:37,280 --> 00:17:42,159 Speaker 1: using less computing power than before. Now these transformer based 288 00:17:42,200 --> 00:17:45,560 Speaker 1: models could just teach themselves in a way. It was 289 00:17:45,640 --> 00:17:47,480 Speaker 1: like if you wanted to teach a kid to read, 290 00:17:47,800 --> 00:17:49,639 Speaker 1: and you used to have to hire a tutor to 291 00:17:49,760 --> 00:17:53,520 Speaker 1: sit there with flashcards, and now instead you could just 292 00:17:53,720 --> 00:17:56,560 Speaker 1: let the kid run through a library and they would 293 00:17:56,600 --> 00:18:01,040 Speaker 1: emerge knowing how to read and write. This was, as 294 00:18:01,080 --> 00:18:04,959 Speaker 1: one investor described to me, a surprising and bitter realization 295 00:18:05,760 --> 00:18:08,280 Speaker 1: that the best AI would come not from the most 296 00:18:08,280 --> 00:18:13,000 Speaker 1: specialized training techniques, but from whichever had the most data. Peter, 297 00:18:13,280 --> 00:18:17,399 Speaker 1: the early open AI employee, says Ilia immediately saw its promise. 298 00:18:18,119 --> 00:18:22,760 Speaker 5: Ilia's reaction was pretty affirmative right away. It's like, this 299 00:18:22,880 --> 00:18:25,320 Speaker 5: is something special we need to be we need to 300 00:18:25,320 --> 00:18:28,080 Speaker 5: be looking at this. This seems a big breakthrough. 301 00:18:28,880 --> 00:18:31,359 Speaker 1: Even in the early days of open AI, Ilia had 302 00:18:31,400 --> 00:18:35,280 Speaker 1: always had this hunch that big advances in AI wouldn't 303 00:18:35,280 --> 00:18:38,800 Speaker 1: come from some specific tweak or new invention, but just 304 00:18:38,840 --> 00:18:42,719 Speaker 1: from more data pouring more and more fuel into the engine. 305 00:18:43,440 --> 00:18:47,040 Speaker 1: And now Ilia had the research that backed up his hypothesis. 306 00:18:47,440 --> 00:18:51,920 Speaker 8: Here's Oren again, Ilia from open ai is known as 307 00:18:51,960 --> 00:18:55,520 Speaker 8: the person who said, it's the data, and it's the 308 00:18:55,560 --> 00:18:58,720 Speaker 8: amount of data, and if we just scale that up 309 00:18:58,800 --> 00:19:03,120 Speaker 8: tremendously of magnitude much much more, we're going to achieve 310 00:19:03,280 --> 00:19:07,080 Speaker 8: what we need. That was not the common perception and 311 00:19:07,119 --> 00:19:09,680 Speaker 8: some very smart and very famous people. I don't want 312 00:19:09,720 --> 00:19:16,159 Speaker 8: to cast aspersions, but certainly I'm I'm not that smarter 313 00:19:16,200 --> 00:19:18,840 Speaker 8: with that famous, but I'm one of the AI experts 314 00:19:18,880 --> 00:19:20,439 Speaker 8: who did not see that coming. 315 00:19:22,000 --> 00:19:25,840 Speaker 1: Because of Ilia, open Ai started experimenting with the transformer. 316 00:19:26,200 --> 00:19:28,120 Speaker 1: They were one of the earliest companies to do so. 317 00:19:28,840 --> 00:19:33,359 Speaker 1: They made models with the now familiar acronym GPT Generative 318 00:19:33,520 --> 00:19:38,080 Speaker 1: pre Trained Transformer, and in particular, they started experimenting with 319 00:19:38,160 --> 00:19:42,400 Speaker 1: how the transformer performed with written words, because they could 320 00:19:42,440 --> 00:19:47,679 Speaker 1: basically feed the model anything written any book, newspaper, article, 321 00:19:48,040 --> 00:19:51,960 Speaker 1: reddit posts, blogs. Humans have spent a lot of time 322 00:19:52,119 --> 00:19:55,880 Speaker 1: writing things down, and those words now had another purpose, 323 00:19:56,400 --> 00:20:01,159 Speaker 1: training data. The Internet wasn't created to train, but in 324 00:20:01,200 --> 00:20:05,479 Speaker 1: the end that may become its legacy. Open aiyes, models 325 00:20:05,480 --> 00:20:08,800 Speaker 1: got better and better at generating text, and they weren't 326 00:20:08,800 --> 00:20:10,960 Speaker 1: limited to just one field of knowledge. 327 00:20:11,640 --> 00:20:15,720 Speaker 8: The amazing thing about these GPT systems is that they're 328 00:20:15,840 --> 00:20:19,560 Speaker 8: very broad. They are actually generalists. You can ask them 329 00:20:19,600 --> 00:20:24,200 Speaker 8: about virtually any topic and they'll produce surprisingly good answers. 330 00:20:24,240 --> 00:20:28,720 Speaker 8: And that's because they've been trained on effectively the entire 331 00:20:28,920 --> 00:20:32,679 Speaker 8: or at least an approximation of the entire corpus of 332 00:20:32,760 --> 00:20:36,800 Speaker 8: text that's available to humanity, billions of billions of sentences, 333 00:20:37,119 --> 00:20:40,560 Speaker 8: all the books you've read, all the documents, the memos, 334 00:20:40,680 --> 00:20:46,159 Speaker 8: the silliness, Harry Potter fan fiction. It's all grist for 335 00:20:46,240 --> 00:20:50,120 Speaker 8: the mill. And then once it's read all that, it's 336 00:20:50,240 --> 00:20:53,320 Speaker 8: remarkably general. So for the first time we would have 337 00:20:53,359 --> 00:20:56,959 Speaker 8: a system that you could ask it about anything and 338 00:20:57,000 --> 00:20:59,720 Speaker 8: it would give you a surprisingly intelligent answer. So we 339 00:20:59,760 --> 00:21:03,800 Speaker 8: went from narrow AI to a kind of general or 340 00:21:03,880 --> 00:21:04,440 Speaker 8: broad AI. 341 00:21:05,359 --> 00:21:07,959 Speaker 1: Through the massive amount of writing that they were feeding 342 00:21:08,000 --> 00:21:11,560 Speaker 1: into their models, open ai found they could create AI 343 00:21:11,680 --> 00:21:15,639 Speaker 1: that was much much better at forming convincing sounding responses 344 00:21:15,640 --> 00:21:19,480 Speaker 1: to questions. In fact, at some point they started to 345 00:21:19,520 --> 00:21:23,680 Speaker 1: worry it was maybe too good. When OpenAI announced its 346 00:21:23,760 --> 00:21:27,359 Speaker 1: language model GBT two, they initially decided not to share 347 00:21:27,359 --> 00:21:30,400 Speaker 1: the model more openly because they were concerned it could 348 00:21:30,400 --> 00:21:34,240 Speaker 1: be dangerous. Here's Peter. He had by that time left 349 00:21:34,280 --> 00:21:36,920 Speaker 1: open ai to start his own company, but he remembers 350 00:21:36,960 --> 00:21:38,320 Speaker 1: the day of the release. 351 00:21:38,560 --> 00:21:40,600 Speaker 5: It was just obvious that it had a much better 352 00:21:40,720 --> 00:21:43,639 Speaker 5: understanding of language than anything that had been trained before. 353 00:21:43,960 --> 00:21:48,240 Speaker 5: Its release was indeed accompanied by a lot of I 354 00:21:48,320 --> 00:21:52,520 Speaker 5: guess great marketing or caution or combination of both. It 355 00:21:52,640 --> 00:21:56,959 Speaker 5: was headline that's too dangerous to be released, and so 356 00:21:57,320 --> 00:21:59,280 Speaker 5: I think it was probably one of the first projects 357 00:21:59,320 --> 00:22:02,480 Speaker 5: where open i decided to not release some of the 358 00:22:02,520 --> 00:22:06,480 Speaker 5: work because all of a sudden, the thinking had become, well, 359 00:22:06,560 --> 00:22:09,200 Speaker 5: what if it's something is so powerful that people could 360 00:22:09,200 --> 00:22:11,680 Speaker 5: go misuse it in ways that we can't control. 361 00:22:12,960 --> 00:22:15,199 Speaker 1: As soon as open ai had a product that was 362 00:22:15,320 --> 00:22:19,160 Speaker 1: actually powerful, they started rethinking their openness. 363 00:22:19,640 --> 00:22:24,000 Speaker 5: Open Ai started with that name where to open really 364 00:22:24,040 --> 00:22:26,959 Speaker 5: stood for everything's going to be open sourced built, you know, 365 00:22:27,040 --> 00:22:28,360 Speaker 5: anybody else can build on it. 366 00:22:29,040 --> 00:22:32,760 Speaker 1: Openness was a crucial part of the company's brand when 367 00:22:32,800 --> 00:22:36,119 Speaker 1: they were founded, Sam told the journalist Stephen Levy, it 368 00:22:36,160 --> 00:22:39,919 Speaker 1: will just be open source and usable by everyone. He 369 00:22:40,000 --> 00:22:42,840 Speaker 1: also told him that their AI would be freely owned 370 00:22:42,880 --> 00:22:47,080 Speaker 1: by the world. Open source software in its broadest sense, 371 00:22:47,440 --> 00:22:49,600 Speaker 1: means that the source code is made available to the 372 00:22:49,600 --> 00:22:52,960 Speaker 1: public freely, and that anyone can tweak the code and 373 00:22:52,960 --> 00:22:57,320 Speaker 1: distribute it themselves. But the company soon started walking back 374 00:22:57,359 --> 00:22:58,119 Speaker 1: those commitments. 375 00:22:58,880 --> 00:23:03,199 Speaker 5: Obviously that evolved over time into something that is not 376 00:23:03,320 --> 00:23:06,280 Speaker 5: so open source, if open source at all for anything. 377 00:23:06,800 --> 00:23:09,360 Speaker 5: I mean, it's definitely not open sourcing. It's hits work. 378 00:23:09,440 --> 00:23:14,320 Speaker 1: Right now, that open source ethos seemed to fade away. 379 00:23:14,440 --> 00:23:17,080 Speaker 1: Here's Sam giving a talk in Munich in twenty twenty three. 380 00:23:18,119 --> 00:23:20,760 Speaker 4: I'm curious, if we stay on the same like GPT 381 00:23:20,880 --> 00:23:23,320 Speaker 4: two to three to four trajectory for five and six, 382 00:23:23,760 --> 00:23:25,440 Speaker 4: how many of you would like us to open source 383 00:23:25,440 --> 00:23:30,719 Speaker 4: GPT six the day we finished training it. Wow, Well 384 00:23:30,720 --> 00:23:32,359 Speaker 4: we're not going to do that. But that's interesting that. 385 00:23:35,280 --> 00:23:39,480 Speaker 1: Honestly, Sam sounds pretty arrogant here. He knows open AI 386 00:23:39,640 --> 00:23:43,040 Speaker 1: started off with promises of being open source. Now he's 387 00:23:43,080 --> 00:23:46,880 Speaker 1: pulling the audience about open sourcing models and immediately dismissing 388 00:23:46,880 --> 00:23:51,119 Speaker 1: their response. Over the years, Sam has subtly changed the 389 00:23:51,200 --> 00:23:55,840 Speaker 1: meaning of openness. It's become fuzzier here he is at 390 00:23:55,840 --> 00:23:57,080 Speaker 1: a VC firm. 391 00:23:57,440 --> 00:23:59,880 Speaker 4: So I think that is that's what we call open ai, 392 00:24:00,040 --> 00:24:02,119 Speaker 4: open AI. We want this to be open technology made 393 00:24:02,119 --> 00:24:02,840 Speaker 4: available to. 394 00:24:02,800 --> 00:24:09,760 Speaker 1: Everyone, open technology made available to everyone. He says it 395 00:24:09,800 --> 00:24:13,240 Speaker 1: so plainly, as though that's obvious what open means, But 396 00:24:13,320 --> 00:24:16,800 Speaker 1: his definition strikes me as so vague that it's essentially meaningless. 397 00:24:17,400 --> 00:24:22,359 Speaker 1: I mean, Google Search is available to everyone. It seems 398 00:24:22,400 --> 00:24:24,960 Speaker 1: like open ai was happy to let people guess what 399 00:24:25,040 --> 00:24:29,240 Speaker 1: they meant by open. In an internal email just months 400 00:24:29,280 --> 00:24:33,159 Speaker 1: after it was founded, Ilia wrote, as we get closer 401 00:24:33,200 --> 00:24:35,960 Speaker 1: to building AI, it will make sense to start being 402 00:24:36,040 --> 00:24:39,639 Speaker 1: less open. The open in open ai means that everyone 403 00:24:39,680 --> 00:24:42,280 Speaker 1: should benefit from the fruits of AI after it's built. 404 00:24:42,600 --> 00:24:46,080 Speaker 1: But it's totally okay to not share the science, even 405 00:24:46,080 --> 00:24:48,760 Speaker 1: though sharing everything is definitely the right strategy in the 406 00:24:48,800 --> 00:24:53,320 Speaker 1: short and possibly medium term for recruitment purposes. This email 407 00:24:53,400 --> 00:24:56,960 Speaker 1: was really interesting because it shows that from the beginning 408 00:24:57,440 --> 00:25:00,480 Speaker 1: open ai had planned not to freely share their se science. 409 00:25:01,000 --> 00:25:03,360 Speaker 1: They didn't want to be open source, as they claimed, 410 00:25:03,960 --> 00:25:06,280 Speaker 1: but they wanted to keep up the public appearance of 411 00:25:06,320 --> 00:25:10,960 Speaker 1: openness because it gave them a recruiting advantage like, don't 412 00:25:10,960 --> 00:25:13,760 Speaker 1: go build AI for the bad guys, come work for us, 413 00:25:14,000 --> 00:25:18,639 Speaker 1: the open virtuous choice. When we asked about their changing 414 00:25:18,680 --> 00:25:22,840 Speaker 1: definition of openness, a company spokesperson said, our mission has 415 00:25:22,880 --> 00:25:28,679 Speaker 1: remained the same, but our tactics have had to change. Okay, 416 00:25:28,720 --> 00:25:32,240 Speaker 1: So let me bring us back to twenty seventeen. It's 417 00:25:32,280 --> 00:25:35,480 Speaker 1: two years after the company's founding, and another problem was 418 00:25:35,520 --> 00:25:40,320 Speaker 1: brewing an Open AI a power struggle. Elon wanted to 419 00:25:40,359 --> 00:25:43,399 Speaker 1: take over. He's someone who's used to being in charge. 420 00:25:44,160 --> 00:25:46,840 Speaker 1: According to open AI, he wanted to move the company 421 00:25:46,920 --> 00:25:50,280 Speaker 1: under Tesla, and he wanted to be CEO, and he 422 00:25:50,320 --> 00:25:53,520 Speaker 1: wanted majority equity, and if it couldn't be done his way, 423 00:25:53,760 --> 00:25:54,320 Speaker 1: he was out. 424 00:25:55,160 --> 00:25:58,280 Speaker 9: Like with everything Elon, as time goes on, he wants 425 00:25:58,320 --> 00:26:01,560 Speaker 9: to assert more and more control and make sure the 426 00:26:01,640 --> 00:26:06,919 Speaker 9: company is operating in exactly the image and way that 427 00:26:07,000 --> 00:26:09,639 Speaker 9: Elon wants it to operate. And so this is going 428 00:26:09,720 --> 00:26:10,640 Speaker 9: to create tension. 429 00:26:11,200 --> 00:26:14,320 Speaker 1: That's Ashley Vance, my colleague who has written a biography 430 00:26:14,400 --> 00:26:14,879 Speaker 1: of Elon. 431 00:26:15,440 --> 00:26:19,160 Speaker 9: Elon's preferred role in anything is to be the CEO 432 00:26:19,359 --> 00:26:22,640 Speaker 9: and the dominant force and the one who controls what's 433 00:26:22,680 --> 00:26:24,920 Speaker 9: going on day to day and. 434 00:26:24,920 --> 00:26:27,640 Speaker 1: The guys actually running the day to day. Greg Brockman 435 00:26:27,720 --> 00:26:33,159 Speaker 1: and Ilias Aitzkiv were wary because Elon was reckless, impulsive, 436 00:26:33,400 --> 00:26:37,359 Speaker 1: and difficult, but he was also their main source of money. 437 00:26:37,960 --> 00:26:41,040 Speaker 1: He had pledged them almost a billion dollars. Open Ai 438 00:26:41,200 --> 00:26:45,000 Speaker 1: had other donors, but nothing close. One option was to 439 00:26:45,040 --> 00:26:47,200 Speaker 1: go with Elon and keep the money. 440 00:26:47,520 --> 00:26:51,600 Speaker 9: The employees weren't all on board with that idea and 441 00:26:51,720 --> 00:26:55,679 Speaker 9: had some concerns, and so you get to this, you 442 00:26:55,720 --> 00:26:59,239 Speaker 9: get to this decision point where it's kind of like, 443 00:26:59,320 --> 00:27:01,879 Speaker 9: you know, are we going to go on with Elon 444 00:27:02,080 --> 00:27:04,679 Speaker 9: or without him? Almost always in recent years people have 445 00:27:04,800 --> 00:27:09,240 Speaker 9: kind of put up with Elon and his demands. 446 00:27:09,640 --> 00:27:12,600 Speaker 1: Or another option was to split with Elon and figure 447 00:27:12,640 --> 00:27:15,280 Speaker 1: out how to get a different source of cash, you 448 00:27:15,320 --> 00:27:18,640 Speaker 1: know who would probably be good at raising money. Sam Altman. 449 00:27:21,200 --> 00:27:24,080 Speaker 9: Reached's point where Elon wanted the company to go one 450 00:27:24,119 --> 00:27:26,600 Speaker 9: way and then the employees wanted it to go another, 451 00:27:26,680 --> 00:27:30,639 Speaker 9: and Sam was picked as the person to lead open 452 00:27:30,640 --> 00:27:31,640 Speaker 9: Ai forward. 453 00:27:32,240 --> 00:27:34,560 Speaker 1: Sam hadn't been that involved in open Ai for the 454 00:27:34,560 --> 00:27:38,040 Speaker 1: first few years. He was still president of YC actually, 455 00:27:39,040 --> 00:27:42,480 Speaker 1: but in this jostling for power, Sam beat out Elon, 456 00:27:43,080 --> 00:27:46,040 Speaker 1: and that's a big deal. Elon was much more famous 457 00:27:46,040 --> 00:27:51,000 Speaker 1: and experienced, and notably, he hates losing well. 458 00:27:51,080 --> 00:27:55,879 Speaker 9: In most conflicts, Elon reacts by trying to win at 459 00:27:55,920 --> 00:28:01,520 Speaker 9: all costs, and whatever whatever which earth you know, may 460 00:28:01,720 --> 00:28:05,960 Speaker 9: may arise from that, Elon doesn't. He doesn't lose too 461 00:28:05,960 --> 00:28:10,959 Speaker 9: many battles. Usually he either. If it's not within a company, 462 00:28:11,000 --> 00:28:15,320 Speaker 9: he usually sues somebody into submission. If it is within 463 00:28:15,359 --> 00:28:18,080 Speaker 9: a company, he throws his weight around in politics until 464 00:28:18,080 --> 00:28:20,439 Speaker 9: he gets what he wants. It's hard to find too 465 00:28:20,480 --> 00:28:23,800 Speaker 9: many examples in recent years where he did not get 466 00:28:23,840 --> 00:28:27,359 Speaker 9: what he wants, and so the turmoil inside of the 467 00:28:27,359 --> 00:28:31,879 Speaker 9: company must have been quite drastic in order for this 468 00:28:32,040 --> 00:28:32,840 Speaker 9: not to happen. 469 00:28:34,760 --> 00:28:37,600 Speaker 1: So in twenty eighteen, Elon walked away in a huff 470 00:28:37,720 --> 00:28:41,360 Speaker 1: and took his money with him. Years later, he'll actually 471 00:28:41,480 --> 00:28:45,040 Speaker 1: end up suing Sam and open Ai, claiming they broke 472 00:28:45,080 --> 00:28:51,000 Speaker 1: their original commitment about remaining nonprofit and open source. Soon 473 00:28:51,040 --> 00:28:55,680 Speaker 1: after Elon left, Sam became CEO of open Ai. There 474 00:28:55,720 --> 00:28:59,920 Speaker 1: hadn't been a CEO before, but this power struggle crystallized 475 00:29:00,080 --> 00:29:04,120 Speaker 1: Sam's new dominance over this company. Remember what the founder 476 00:29:04,160 --> 00:29:09,400 Speaker 1: of YC once said, Sam is extremely good at becoming powerful. 477 00:29:10,520 --> 00:29:15,080 Speaker 1: Sam's excitement about open Ai kept growing, his attention started 478 00:29:15,160 --> 00:29:19,400 Speaker 1: drifting away from his job running YC. Sure, running a 479 00:29:19,400 --> 00:29:22,120 Speaker 1: world famous startup accelerator is a position of a lot 480 00:29:22,160 --> 00:29:26,160 Speaker 1: of influence, but the race for building AGI was heating up, 481 00:29:26,720 --> 00:29:30,880 Speaker 1: and if OpenAI succeeded in creating AGI before anyone else, 482 00:29:31,600 --> 00:29:33,840 Speaker 1: it's hard to imagine a position in the world with 483 00:29:33,920 --> 00:29:38,840 Speaker 1: more power than being its CEO. But Sam didn't give 484 00:29:38,920 --> 00:29:42,680 Speaker 1: up his job at YC right away. This situation made 485 00:29:42,720 --> 00:29:46,160 Speaker 1: some of the people running the accelerator grumble. They felt 486 00:29:46,160 --> 00:29:49,880 Speaker 1: like Sam was spread too thin, pushing to expand too fast, 487 00:29:50,200 --> 00:29:54,440 Speaker 1: and prioritizing his own interests above those of YC. It 488 00:29:54,520 --> 00:29:58,560 Speaker 1: earned him some enemies within his own ranks. In fact, 489 00:29:58,680 --> 00:30:02,400 Speaker 1: according to a source, Sam's mentor, Paul Graham, the guy 490 00:30:02,400 --> 00:30:04,120 Speaker 1: who put him in the job in the first place, 491 00:30:04,720 --> 00:30:07,560 Speaker 1: flew in from the UK to ask Sam in person 492 00:30:07,720 --> 00:30:12,120 Speaker 1: to step down. Paul had lost confidence in his former protege, 493 00:30:12,760 --> 00:30:16,160 Speaker 1: but he also didn't want to create public drama, so 494 00:30:16,200 --> 00:30:19,320 Speaker 1: Sam was ushered out and they kept the backstory quiet. 495 00:30:20,680 --> 00:30:24,240 Speaker 1: Now focused only on open Ai, Sam had one big 496 00:30:24,280 --> 00:30:28,840 Speaker 1: goal to raise money to train open Aiy's models. They 497 00:30:28,880 --> 00:30:33,440 Speaker 1: needed a lot of computing power, and computing power is expensive. 498 00:30:34,640 --> 00:30:37,960 Speaker 1: Sam tried to raise money but wasn't getting traction. Here 499 00:30:38,000 --> 00:30:39,880 Speaker 1: he is on the Lex Friedman podcast. 500 00:30:40,560 --> 00:30:44,080 Speaker 4: We started as a nonprofit. We learned early on that 501 00:30:44,160 --> 00:30:46,560 Speaker 4: we were going to need far more capital than we 502 00:30:46,560 --> 00:30:49,000 Speaker 4: were able to raise as a nonprofit to do what 503 00:30:49,040 --> 00:30:51,480 Speaker 4: we needed to go do. We had tried and failed 504 00:30:51,600 --> 00:30:54,040 Speaker 4: enough to raise the money as a nonprofit. We didn't 505 00:30:54,040 --> 00:30:56,920 Speaker 4: see a path forward there. So we needed some of 506 00:30:56,960 --> 00:31:00,480 Speaker 4: the benefits of capitalism, but not too much. I remember 507 00:31:00,520 --> 00:31:02,240 Speaker 4: at the time someone said, you know, as a nonprofit, 508 00:31:02,280 --> 00:31:04,800 Speaker 4: not enough will happen. As a for profit, too much 509 00:31:04,800 --> 00:31:05,240 Speaker 4: will happen. 510 00:31:05,840 --> 00:31:09,480 Speaker 1: They needed something in the middle, and honestly, Sam doesn't 511 00:31:09,520 --> 00:31:13,920 Speaker 1: sound that hung up about leaving nonprofit life behind. He 512 00:31:14,040 --> 00:31:20,080 Speaker 1: Frankenstein something together. Basically, he created a for profit entity 513 00:31:20,480 --> 00:31:24,120 Speaker 1: that lived under the umbrella of the original nonprofit. The 514 00:31:24,160 --> 00:31:26,840 Speaker 1: for profit could do all the things normal companies do, 515 00:31:27,160 --> 00:31:32,080 Speaker 1: like raise investment and offer equity to employees, but its investors' 516 00:31:32,120 --> 00:31:35,880 Speaker 1: returns were capped, whereas at other companies they'd be unlimited. 517 00:31:37,240 --> 00:31:42,000 Speaker 1: This corporate structure was grafted together. Open Ai was essentially 518 00:31:42,080 --> 00:31:45,520 Speaker 1: now a for profit controlled by the board of the nonprofit, 519 00:31:46,000 --> 00:31:52,240 Speaker 1: which sounds a little unstable. Open Ai had spent years 520 00:31:52,360 --> 00:31:55,280 Speaker 1: saying they would be a non profit. Now they had 521 00:31:55,280 --> 00:31:59,240 Speaker 1: come up with this for profit workaround. After that change, 522 00:31:59,360 --> 00:32:02,400 Speaker 1: a lot of people were upset, but open Ai was 523 00:32:02,520 --> 00:32:05,840 Speaker 1: more focused on their end goal. They wanted to build 524 00:32:05,920 --> 00:32:08,560 Speaker 1: Agi and they needed to raise money to do it. 525 00:32:09,480 --> 00:32:12,960 Speaker 1: And then in twenty nineteen, Sam the deal maker made 526 00:32:12,960 --> 00:32:17,080 Speaker 1: a big, hugely important deal. He raised a billion dollars 527 00:32:17,120 --> 00:32:21,600 Speaker 1: from Microsoft. Here's Microsoft CEO Satya Nandela after they signed 528 00:32:21,600 --> 00:32:21,960 Speaker 1: the deal. 529 00:32:22,800 --> 00:32:26,000 Speaker 4: Hi, I'm here with Sam Altman's CEO of open Ai. 530 00:32:26,360 --> 00:32:26,640 Speaker 2: Today. 531 00:32:26,680 --> 00:32:29,640 Speaker 4: We are very excited to announce a strategic partnership with 532 00:32:29,800 --> 00:32:30,400 Speaker 4: open Ai. 533 00:32:30,440 --> 00:32:33,840 Speaker 1: And I thought one important thing Microsoft had was lots 534 00:32:33,840 --> 00:32:37,920 Speaker 1: of raw computing power and open Ai could now use it. 535 00:32:38,760 --> 00:32:41,880 Speaker 1: Remember open Ai had originally been conceived to be an 536 00:32:41,920 --> 00:32:46,760 Speaker 1: antidote to Google. They presented themselves as fundamentally different from 537 00:32:46,840 --> 00:32:51,800 Speaker 1: profit hungry tech giants, and then overnight they became intimately 538 00:32:51,920 --> 00:32:55,440 Speaker 1: enmeshed with a tech company worth more than a trillion dollars. 539 00:32:56,200 --> 00:33:00,240 Speaker 1: Now open Ai was in many ways an arm of Microsoft. 540 00:33:01,920 --> 00:33:05,800 Speaker 1: This was a remarkable about phase. Reid Hoffman was on 541 00:33:05,840 --> 00:33:07,960 Speaker 1: the board of open Ai and on the board of 542 00:33:07,960 --> 00:33:11,040 Speaker 1: Microsoft at the time of the deal. He didn't see 543 00:33:11,040 --> 00:33:13,800 Speaker 1: this as an abdication of open AI's initial premise. 544 00:33:14,960 --> 00:33:20,280 Speaker 6: There were parties who worried about with this corrupt the mission. 545 00:33:20,640 --> 00:33:22,360 Speaker 6: But you know, I think that's a little bit of 546 00:33:22,920 --> 00:33:26,440 Speaker 6: like kind of a modern naivete is to say corporation 547 00:33:26,640 --> 00:33:31,000 Speaker 6: equals bad or corrupt, and it's just naive because there's 548 00:33:31,040 --> 00:33:35,600 Speaker 6: lots of ways that companies are are collaborative with humanity anxiety. 549 00:33:35,640 --> 00:33:38,880 Speaker 6: They try to serve the customer as well, they hire employees, 550 00:33:39,040 --> 00:33:42,680 Speaker 6: they have shareholders, they exist within societies. 551 00:33:43,480 --> 00:33:46,920 Speaker 1: Okay, So Reid's perspective is that just because you want 552 00:33:46,960 --> 00:33:50,240 Speaker 1: to make money doesn't mean you're bad, which is on 553 00:33:50,400 --> 00:33:54,640 Speaker 1: brand for a billionaire venture capitalist. And I guess one 554 00:33:54,680 --> 00:33:57,160 Speaker 1: way to look at it is that the Microsoft deal 555 00:33:57,360 --> 00:34:00,000 Speaker 1: may have been the most practical way for open ai 556 00:34:00,120 --> 00:34:03,680 Speaker 1: to continue its mission of creating safe agi for all 557 00:34:03,720 --> 00:34:09,360 Speaker 1: of humanity, but it also highlighted an important pattern so 558 00:34:09,400 --> 00:34:12,440 Speaker 1: that OpenAI often walked back its promises when it was 559 00:34:12,480 --> 00:34:16,720 Speaker 1: convenient to do so. And amid all this, people started 560 00:34:16,760 --> 00:34:21,320 Speaker 1: to doubt Sam's integrity both inside and outside the company, 561 00:34:22,120 --> 00:34:26,120 Speaker 1: and that would lead to a major rift. That's next 562 00:34:26,120 --> 00:34:35,840 Speaker 1: time on Foundering. Foundering is hosted by me Ellen Hewitt. 563 00:34:35,880 --> 00:34:40,080 Speaker 1: Sean Wen is our executive producer. Rachel Metz contributed reporting 564 00:34:40,120 --> 00:34:44,240 Speaker 1: to this episode. Molly Nugent is our associate producer. Blake 565 00:34:44,280 --> 00:34:49,360 Speaker 1: Maples is our audio engineer. Mark Million and Vandermay Seth Fiegerman, 566 00:34:49,600 --> 00:34:53,160 Speaker 1: Tom Giles and Molly Schutz are our story editors. We 567 00:34:53,200 --> 00:34:57,520 Speaker 1: had production help from Jessica Nix and Antonia Mufferetch. Thanks 568 00:34:57,520 --> 00:35:00,319 Speaker 1: for listening. If you like our show, leave a view, 569 00:35:00,960 --> 00:35:04,239 Speaker 1: and most importantly, tell your friends. See you next time.