WEBVTT 00:00:01.780 --> 00:00:09.000 Wednesday, September twentieth, two thousand seventeen. Two-oh-eight in the afternoon, Pacific time. 00:00:09.000 --> 00:00:15.500 An email arrives. The subject line is two words. Honest Thoughts. 00:00:15.500 --> 00:00:30.240 It's from Ilya Sutskever, OpenAI's chief scientist, and it's signed by him and Greg Brockman. It's addressed to the two men who co-chair their company: Elon Musk and Sam Altman. 00:00:30.240 --> 00:00:45.420 Elon, Sam, This process has been the highest stakes conversation that Greg and I have ever participated in, and if the project succeeds, it'll turn out to have been the highest stakes conversation the world has seen. 00:00:45.420 --> 00:01:00.000 It's long. It's careful. It's the kind of email you write at night, rewrite in the morning, and send before you can change your mind. And in the middle of it, there's a line about the man they're sending it to. 00:01:00.000 --> 00:01:06.820 The goal of OpenAI is to make the future good and to avoid an AGI dictatorship. 00:01:06.820 --> 00:01:13.860 At two-seventeen, nine minutes later, Elon Musk replies. 00:01:13.860 --> 00:01:18.760 Guys, I've had enough. This is the final straw. 00:01:18.760 --> 00:01:38.420 Nine minutes. To understand how a company built to save the world from an AI dictatorship got to that email, we have to go back six weeks. To a night in Seattle, and a phrase you heard at the end of the last episode. 00:01:38.420 --> 00:01:51.080 From Will Sigmon Media, this is The Role Models. Season one: OpenAI. Episode four. The Divorce. 00:01:51.080 --> 00:02:07.960 August eleventh, two thousand seventeen. You were there last episode. OpenAI's bot has just beaten Dendi at The International. And at nine-seventeen that night, Musk sends a short email to Brockman, Sutskever and Altman. 00:02:07.960 --> 00:02:13.660 Time to make the next step for OpenAI. This is the triggering event. 00:02:13.660 --> 00:02:32.540 To hear why that night felt like a trigger, you need the spring and summer before it. Because OpenAI's founders had spent those months discovering that their plan, the plan you heard about in episode one, a lab of brilliant people writing papers, was not going to be enough. 00:02:32.540 --> 00:02:52.960 In June, Sutskever had started sending Musk and Brockman what he called bi-weekly updates. You heard a piece of the first one last episode: ninety-five percent of progress comes from running big experiments quickly. Musk's reply came the next morning. It was one line long. 00:02:52.960 --> 00:02:58.940 Ok. Let's figure out the least expensive way to ensure compute power is not a constraint. 00:02:58.940 --> 00:03:10.140 The least expensive way. Hold on to that word. Because over the next eighteen months, everybody in this story is going to find out that there is no inexpensive way. 00:03:10.140 --> 00:03:34.960 The next step meant money, and structure. A month earlier, Sutskever had written to Musk and Brockman with a plan. Grow the cluster from six hundred GPUs to five thousand. Grow the team from fifty-five people to two hundred by January twenty-twenty. And he wrote that each year they'd need to increase their hardware spend exponentially. 00:03:34.960 --> 00:03:43.100 If we can secure the funding, we have a real chance at setting the initial conditions under which AGI is born. 00:03:43.100 --> 00:04:07.430 That same email had a line about the competition. Sutskever estimated that Google Brain had around a hundred thousand GPUs. OpenAI had six hundred. And it was the same email where, as you heard last episode, he wrote that in as little as four years there might be "an actual chance of waking up to AGI." 00:04:07.430 --> 00:04:24.870 Waking up to AGI. You go to bed, the cluster runs, and in the morning, maybe, something is different. That was the bet, as the scientists saw it. Not a single brilliant idea. A machine big enough to find the idea for you. 00:04:24.870 --> 00:04:57.790 The next evening, July thirteenth, Brockman and Sutskever met with Musk. That night, Shivon Zilis, who worked with Musk and acted as his go-between with OpenAI, texted Brockman a short question. How did it go? OpenAI later published his answer. It's a list, typed fast, in lowercase. On compute, Brockman wrote that Musk had said, quote, "do what we need to do." And then, about a chip startup they'd been discussing, this: 00:04:57.790 --> 00:05:09.650 Quote. "turned into talking about structure. He said non-profit was def the right one early on, may not be the right one now. ilya and I agree with this for a number of reasons." 00:05:09.650 --> 00:05:30.070 "May not be the right one now." It's the first time, in the record we have, that someone writes down the idea out loud. The nonprofit might have been right for the beginning, and wrong for what came next. And notice who agrees. Not just Musk. Brockman and Sutskever too. 00:05:30.070 --> 00:05:45.970 A week later, in the middle of the night, Musk forwarded the founders a New York Times story. China had announced a plan to lead the world in artificial intelligence by twenty thirty. His note on top was two sentences. 00:05:45.970 --> 00:05:51.910 They will do whatever it takes to obtain what we develop. Maybe another reason to change course. 00:05:51.910 --> 00:06:13.590 A nonprofit can't sell stock. So, over the summer, the founders started talking about something else: a for-profit company, alongside or in place of the nonprofit. On July twenty-first, Brockman laid out a three-step path in an email to Musk. 00:06:13.590 --> 00:06:27.770 One. AI research non-profit, through end of twenty seventeen. Two. AI research plus hardware for-profit, starting twenty eighteen. Three. Government project. When: question mark, question mark. 00:06:27.770 --> 00:06:35.990 He literally typed two question marks. Musk replied that he had a tentative game plan he'd like to run by them. 00:06:35.990 --> 00:06:46.850 Three weeks later, a bot beat Dendi in Seattle. And at nine-seventeen that night, the game plan became a negotiation. 00:06:46.850 --> 00:07:03.890 Over the next six weeks, they negotiated. What we know about those weeks comes almost entirely from documents that later became exhibits in a lawsuit, and from OpenAI's published version of them. Keep that in mind, because it matters later. 00:07:03.890 --> 00:07:19.470 On September fourth, Shivon Zilis, who worked with Musk and served as a go-between with OpenAI, messaged Brockman about the terms. OpenAI published the exchange. One line in it is about equity. 00:07:19.470 --> 00:07:28.830 Quote. "And he sounded fairly non-negotiable on his equity being between fifty and sixty, so moot point of having majority." 00:07:28.830 --> 00:07:35.490 Fifty to sixty percent. Of the company that would, if they were right, build artificial general intelligence. 00:07:35.490 --> 00:08:03.610 The same exchange shows how much was still unclear, even to the people inside it. Zilis and Brockman go back and forth about board seats. The proposal, as they described it, would always give Musk at least three seats, or a quarter of the board, whichever was larger. And one of them asks a simple question: if all the power rests with the board, what does owning a certain percentage of shares even do? 00:08:03.610 --> 00:08:14.150 It's a good question. It's the question this whole season keeps coming back to. Who owns it, and who controls it, are not the same thing. 00:08:14.150 --> 00:08:26.390 On September thirteenth, just after midnight, Musk sent Sutskever his idea for the board. It's worth hearing in his words, because it's more nuanced than the headlines later made it. 00:08:26.390 --> 00:08:44.230 The three common stock seats, you, Greg and Sam, should be elected by common shareholders. They will de facto be yours, but not in the unlikely event that you lose the faith of a huge percentage of common stockholders over time, or step away from the company by choice. 00:08:44.230 --> 00:08:51.350 It's an email written after midnight, and it reads like one. At one point he simply says so. 00:08:51.350 --> 00:08:58.290 I'm super tired and don't want to overcomplicate things, but this seems approx right. 00:08:58.290 --> 00:09:05.410 And in the middle of it, the sentence that OpenAI would later put at the center of its case. 00:09:05.410 --> 00:09:11.490 I would unequivocally have initial control of the company, but this will change quickly. 00:09:11.490 --> 00:09:19.930 The rough target would be to get to a twelve person board, probably more like sixteen if this board really ends up deciding the fate of the world. 00:09:19.930 --> 00:09:29.510 At the sixteen person board level, we would have seven sixteenths of the votes and I'd have a twenty-five percent influence, which is my min comfort level. 00:09:29.510 --> 00:09:35.890 If everyone else we asked to join our board is truly against us, we should probably lose. 00:09:35.890 --> 00:09:40.430 And then, a paragraph later, he wrote this. 00:09:40.430 --> 00:09:57.590 My experience with boards, assuming they consist of good, smart people, is that they are rational and reasonable. There is basically never a real hardcore battle where an individual board vote is pivotal, so this is almost certainly, sure hope so, going to be a moot point. 00:09:57.590 --> 00:10:16.070 Basically never a real hardcore battle where an individual board vote is pivotal. File that one away. In six years, a board of six people will meet by video call on a Friday afternoon, and a handful of individual votes will be the only thing that matters in the world of technology. 00:10:16.070 --> 00:10:37.470 OpenAI's published account adds something from a phone call, and it's worth flagging as their account, not a document. By OpenAI's telling, Musk said he didn't care about equity personally, but needed to accumulate eighty billion dollars for a city on Mars. There is no recording of that call in the public record. 00:10:37.470 --> 00:10:46.350 As a closing note, I've been really impressed with the quality of discussion with you guys on the equity and board stuff. I have a really good feeling about this. 00:10:46.350 --> 00:11:05.210 Two days later, on September fifteenth, by OpenAI's account, Musk's wealth manager registered a public benefit corporation called Open Artificial Intelligence Technologies, Inc. That's a vehicle for exactly the kind of for-profit they had been discussing. 00:11:05.210 --> 00:11:20.190 Remember that. In the lawsuits that came later, a central question would be who wanted OpenAI to become a for-profit company. This is OpenAI's exhibit A. 00:11:20.190 --> 00:11:33.870 Now back to that afternoon. September twentieth, two-oh-eight. Here's more of what Sutskever and Brockman actually wrote, because the email is stranger and more honest than its reputation. 00:11:33.870 --> 00:11:48.130 Yesterday while we were considering making our final commitment given the non-solicit agreement, we realized we'd made a mistake. We have several important concerns that we haven't raised with either of you. We didn't raise them because we were afraid to. 00:11:48.130 --> 00:12:05.290 We really want to work with you. We believe that if we join forces, our chance of success in the mission is the greatest. Our desire to work with you is so great that we are happy to give up on the equity, personal control, make ourselves easily firable. Whatever it takes to work with you. 00:12:05.290 --> 00:12:08.970 Easily firable. Hold on to that phrase. 00:12:08.970 --> 00:12:25.350 The current structure provides you with a path where you end up with unilateral absolute control over the AGI. You stated that you don't want to control the final AGI, but during this negotiation, you've shown to us that absolute control is extremely important to you. 00:12:25.350 --> 00:12:39.690 As an example, you said that you needed to be CEO of the new company so that everyone will know that you are the one who is in charge, even though you also stated that you hate being CEO and would much rather not be CEO. 00:12:39.690 --> 00:12:51.510 We disagree with your statement that our ability to leave is our greatest power, because once the company is actually on track to AGI, the company will be much more important than any individual. 00:12:51.510 --> 00:13:14.530 Read that one twice. The company will be much more important than any individual. In twenty seventeen, they meant Elon Musk. Six years later, the same idea, that the company is bigger than any one person, will be argued about a different man. And Ilya Sutskever will be on both sides of it in the same weekend. 00:13:14.530 --> 00:13:26.810 You are concerned that Demis could create an AGI dictatorship. So do we. So it is a bad idea to create a structure where you could become a dictator if you chose to. 00:13:26.810 --> 00:13:37.830 And then they turn to Altman, the other co-chair. They praise him for something specific: standing his ground on control when they hadn't. 00:13:37.830 --> 00:13:49.890 Sam acted as a genuine counterbalance to you, which has been extremely fruitful. But we haven't been able to fully trust your judgements throughout this process, because we don't understand your cost function. 00:13:49.890 --> 00:14:05.830 That last sentence is addressed to Sam. "Cost function" is a machine-learning term: the thing a system is really optimizing for. They're telling Sam Altman, in writing, that they don't know what he's optimizing for. 00:14:05.830 --> 00:14:16.550 We don't understand why the CEO title is so important to you. Your stated reasons have changed, and it's hard to really understand what's driving it. 00:14:16.550 --> 00:14:26.390 Is AGI truly your primary motivation? How does it connect to your political goals? How has your thought process changed over time? 00:14:26.390 --> 00:14:47.190 That's to Sam too. The same email had already asked Elon, in other words, why he needed to be CEO. Now it was asking Sam. Both co-chairs. The same question. And then they ask for one thing: a meeting of all four of them, that day. 00:14:47.190 --> 00:14:57.650 And then, the part of the email almost nobody quotes. They turn the questions on themselves. The heading is just their two names. 00:14:57.650 --> 00:15:11.750 During this negotiation, we realized that we have allowed the idea of financial return two to three years down the line to drive our decisions. This is why we didn't push on the control. We thought that our equity is good enough, so why worry? 00:15:11.750 --> 00:15:25.750 But this attitude is wrong, just like the attitude of AI experts who don't think that AI safety is an issue because they don't really believe that they'll build AGI. We did not speak our full truth during the negotiation. 00:15:25.750 --> 00:15:43.590 Two of the most gifted people in the field, admitting in writing that the money had made them quiet. It's a rare thing to find in a document that later ends up in a lawsuit. Most people write those emails to win. This one reads like it was written to be forgiven. 00:15:43.590 --> 00:15:53.530 If all of us say the truth, and resolve the issues, the company that we'll create will be much more likely to withstand the very strong forces it'll experience. 00:15:53.530 --> 00:15:55.830 Nine minutes later: 00:15:55.830 --> 00:16:15.570 Guys, I've had enough. This is the final straw. Either go do something on your own or continue with OpenAI as a nonprofit. I will no longer fund OpenAI until you have made a firm commitment to stay, or I'm just being a fool who is essentially providing free funding for you to create a startup. Discussions are over. 00:16:15.570 --> 00:16:20.790 In the same minute, a second note from Musk. 00:16:20.790 --> 00:16:29.170 To be clear, this is not an ultimatum to accept what was discussed before. That is no longer on the table. 00:16:29.170 --> 00:16:33.810 And from Sam Altman, at two-seventeen, one line. 00:16:33.810 --> 00:16:38.830 I remain enthusiastic about the non-profit structure! 00:16:38.830 --> 00:16:54.590 The talks over a for-profit were over. What came next, by OpenAI's account, was a different proposal: fold OpenAI into Tesla. OpenAI says the researchers didn't want to work at Tesla. 00:16:54.590 --> 00:17:15.250 In OpenAI's telling, the Tesla offer came with real money: a billion-dollar budget right away, growing exponentially from there. And they say they spent the next months trying to find another way to raise the capital instead. That's their account. We don't have Musk's version of that offer in his own words. 00:17:15.250 --> 00:17:22.470 January thirty-first, two thousand eighteen. Musk to Brockman, Sutskever and Altman. 00:17:22.470 --> 00:17:33.430 OpenAI is on a path of certain failure relative to Google. There obviously needs to be immediate and dramatic action, or everyone except for Google will be consigned to irrelevance. 00:17:33.430 --> 00:17:42.450 To be clear, I have a lot of respect for your abilities and accomplishments, but I am not happy with how things have been managed. 00:17:42.450 --> 00:17:47.190 The same email takes aim at OpenAI's governance. 00:17:47.190 --> 00:17:56.770 The former would require a major increase in funds donated and highly credible people joining our board. The current board situation is very weak. 00:17:56.770 --> 00:18:12.790 Either we fix things and my engagement increases a lot, or we don't and I will drop to near zero and publicly reduce my association. I will not be in a situation where the perception of my influence and time doesn't match the reality. 00:18:12.790 --> 00:18:24.610 "The current board situation is very weak." Another sentence to keep. Who sits on that board, and how strong it is, will decide this company's fate more than once. 00:18:24.610 --> 00:18:33.790 Brockman wrote back that night. His answer is a statement of the mission, and of a phrase the company would one day be judged against. 00:18:33.790 --> 00:18:46.770 My view is that the best future will come from a major expansion of OpenAI. Our goal and mission are fundamentally correct, and that will increasingly be a superpower as AGI grows near. 00:18:46.770 --> 00:19:16.270 He made the case like a founder pitching an investor. Their fundraising conversations, he wrote, showed that he and Sutskever could convince reputable people that AGI could really happen in the next ten years or less. And he laid out a hiring plan. About forty people at the beginning of twenty seventeen. A hundred by the end of twenty eighteen. Three hundred the year after. Nine hundred by the end of twenty twenty. 00:19:16.270 --> 00:19:23.950 We believe we uniquely are on trajectory to solving safety, at least in broad strokes, in the next three years. 00:19:23.950 --> 00:19:34.190 In the next three years. It's a startling line to read now. Not because it was dishonest. Because of how certain it sounds. 00:19:34.190 --> 00:19:43.310 There's one more line in Brockman's email that explains a lot about the next five years. It's about why people were leaving. 00:19:43.310 --> 00:19:51.090 We lose people on comp currently, but pretty much only on comp. 00:19:51.090 --> 00:20:20.210 Comp. Compensation. You heard in episode two what a nonprofit had to pay to keep its best researchers, and what Google and Facebook could offer instead: stock. A nonprofit can't give anyone stock. So when Brockman wrote that they were losing people only on pay, he was describing a problem the nonprofit structure could never solve on its own. And the solution, when it came, would be a company where employees could own a piece of the upside. 00:20:20.210 --> 00:20:25.130 We have the most talent dense team in the field, and we have the reputation for it as well. 00:20:25.130 --> 00:20:39.730 Our biggest tool is the moral high ground. To retain this, we must try our best to remain a non-profit. AI is going to shake up the fabric of society, and our fiduciary duty should be to humanity. 00:20:39.730 --> 00:20:43.570 It doesn't matter who wins if everyone dies. 00:20:43.570 --> 00:20:48.990 The same email had a line the company would put in its own vocabulary. 00:20:48.990 --> 00:21:01.890 Related to this, we need to communicate a better red than dead outlook. We're trying to build safe AGI, and we're not willing to destroy the world in a down-to-the-wire race to do so. 00:21:01.890 --> 00:21:16.190 One more idea had been floating that winter, and Musk shut it down. An ICO, an initial coin offering: selling a crypto token to raise money. It was the height of the token boom. 00:21:16.190 --> 00:21:32.390 I have considered the ICO approach and will not support it. In my opinion, that would simply result in a massive loss of credibility for OpenAI and everyone associated with the ICO. If something seems too good to be true, it is. 00:21:32.390 --> 00:21:38.570 I respect your decision on the ICO idea, which matches the evolution of our own thinking. 00:21:38.570 --> 00:22:06.850 In early February, by OpenAI's later account, Musk forwarded an email that suggested OpenAI should, quote, "attach to Tesla as its cash cow", and he added that it was "exactly right". Tesla, he wrote, was "the only path that could even hope to hold a candle to Google. Even then, the probability of being a counterweight to Google is small. It just isn't zero." 00:22:06.850 --> 00:22:28.170 OpenAI published that email in twenty twenty-four. Musk's side has argued that the context of these exchanges is very different from how OpenAI tells it. They say he was trying to save the mission, not take it over. Both things are on the record. What they meant is the argument. 00:22:28.170 --> 00:22:47.030 February twentieth, two thousand eighteen. By OpenAI's account, Musk hosted a goodbye all-hands with the team that day. He told them they should pursue their path to raising billions a year, and that he'd pursue advanced AI research at Tesla. 00:22:47.030 --> 00:22:57.110 The same day, OpenAI published a blog post. It thanked a list of new donors. Then, a few paragraphs down, this: 00:22:57.110 --> 00:23:13.270 That was the public story in two thousand eighteen: a conflict of interest. And it was real. Tesla had hired Andrej Karpathy, one of OpenAI's founding researchers, to run its Autopilot AI the previous June. 00:23:13.270 --> 00:23:36.350 Read the rest of that post, and you can hear a company bracing itself. It names new donors: Valve's co-founder Gabe Newell, the Skype co-founder Jaan Tallinn, the Olympic decathlete Ashton Eaton and the heptathlete Brianne Theisen-Eaton. It says Reid Hoffman is significantly increasing his contribution. And then it says why. 00:23:36.350 --> 00:23:51.790 "We're broadening our base of funders to prepare for the next phase of OpenAI, which will involve ramping up our investments in our people and the compute resources necessary to make consequential breakthroughs in artificial intelligence." 00:23:51.790 --> 00:24:10.030 The board, the post says, is now four people. Greg Brockman. Ilya Sutskever. Holden Karnofsky. And Sam Altman. The co-chair who wanted control was gone. The co-chair who'd stood his ground was still there. 00:24:10.030 --> 00:24:28.510 A year later, Musk added his own account in a post on Twitter: that Tesla was competing for some of the same people as OpenAI, and that he didn't agree with some of what the OpenAI team wanted to do. It was, he wrote, "just better to part ways on good terms." 00:24:28.510 --> 00:24:51.350 Six years later, OpenAI told a different version: that Musk had wanted majority equity, board control and the CEO job; that during the talks, he withheld funding; and that Reid Hoffman, a LinkedIn co-founder and an early OpenAI donor, stepped in to cover salaries and operations. 00:24:51.350 --> 00:25:09.410 Musk disputes OpenAI's characterization of what he wanted and why. Both sides are now on the record, under oath in some places. We'll tell you what the documents say. We won't tell you whose heart was in the right place. Nobody can. 00:25:09.410 --> 00:25:23.030 April ninth, two thousand eighteen. Seven weeks after Musk's goodbye. OpenAI publishes a short document and gives it a name with some weight to it. The OpenAI Charter. 00:25:23.030 --> 00:25:38.530 It's barely a page long. OpenAI said it reflected the strategy they'd refined over the past two years. And it's the closest thing the company has to a constitution. So let's hear what it actually promised. 00:25:38.530 --> 00:25:55.950 First, a definition. It describes AGI as, quote, "highly autonomous systems that outperform humans at most economically valuable work." And it says the mission is to make sure that AGI "benefits all of humanity." 00:25:55.950 --> 00:26:11.030 Then, under a heading called Broadly Distributed Benefits, comes the line Brockman had written to Musk in January, almost word for word, now made official. "Our primary fiduciary duty is to humanity." 00:26:11.030 --> 00:26:26.850 Right after it, a sentence that does a lot of quiet work. "We anticipate needing to marshal substantial resources to fulfill our mission." That's a careful way of saying: we are going to need a great deal of money. 00:26:26.850 --> 00:26:44.110 Then the most unusual promise in the document. It's about racing. The charter says they're worried about the final stretch of AGI development turning into a competitive race, "without time for adequate safety precautions." So it makes a commitment. 00:26:44.110 --> 00:26:56.530 "If a value-aligned, safety-conscious project comes close to building AGI before we do, we commit to stop competing with and start assisting this project." 00:26:56.530 --> 00:27:10.910 A company, promising in advance that if somebody else is about to win, it will stop competing and help them. It even suggests what close might mean: "a better-than-even chance of success in the next two years." 00:27:10.910 --> 00:27:28.290 And then there's a line about openness, the word in the company's name. Today, the charter says, they publish most of their research. "But we expect that safety and security concerns will reduce our traditional publishing in the future." 00:27:28.290 --> 00:27:39.370 Fewer papers, in the future, for safety reasons. In eleven months, you'll hear what happened when they tried that for the first time, with a model called GPT-2. 00:27:39.370 --> 00:28:01.630 Lay those promises next to each other. A duty to humanity. Substantial resources. A pledge not to race. And less openness. Every one of them would be tested. Some would be argued over in court. And one of them, the money, was about to swallow the rest of the story. 00:28:01.630 --> 00:28:23.250 Notice, too, what the charter doesn't say. It doesn't say OpenAI will always be a nonprofit. It doesn't say who owns what, or who sits on the board. It's a document about goals, not structure. That gap is where the next six years of arguments would live. 00:28:23.250 --> 00:28:32.770 While the founders argued about structure, a researcher named Alec Radford was working on something that, at the time, looked small. 00:28:32.770 --> 00:28:58.250 And around the same time, a new executive arrived. Mira Murati had been a senior product manager on Tesla's Model X, and then worked at a startup called Leap Motion. She joined OpenAI in twenty eighteen to work on applied AI and partnerships. Remember her name. In five years, for one weekend, she will be running the company. 00:28:58.250 --> 00:29:21.510 June eleventh, two thousand eighteen. OpenAI posts a paper and a blog post. The blog's title is plain: Improving language understanding with unsupervised learning. Nobody called it GPT-1 then. That name came later. But this is where the G, the P and the T start. 00:29:21.510 --> 00:29:50.890 The idea came in two steps. First, let a model read a large pile of text with no labels and no teacher, just learning to predict what comes next. Then, give it a small amount of training on a specific task. Answering questions. Judging whether one sentence follows from another. The post said it worked, quote, "surprisingly well." The same core model could be adapted to very different tasks. 00:29:50.890 --> 00:30:07.270 It was built on an architecture called the transformer, which researchers at Google had published the year before. Remember that. The most important ingredient in OpenAI's future came from the company Musk kept saying they could never catch. 00:30:07.270 --> 00:30:21.530 And look at how small it was. The post spells it out. One machine. Eight GPUs. About a month of training. The reading material was a few thousand books, about five gigabytes of text. 00:30:21.530 --> 00:30:39.090 Under a heading called Future, there's a sentence that, read today, sounds like a prophecy. The results, the post says, suggest "there is significant room for improvement using the well-validated approach of more compute and data." 00:30:39.090 --> 00:30:51.510 More compute. And more data. That's it. That's the whole plan for the next five years, in one line, posted on a summer afternoon, by a lab that couldn't afford it. 00:30:51.510 --> 00:31:19.610 The post also noted, almost in passing, that the model could do a little bit on tasks it had never been trained on at all. Not well. But better than chance, and better as the model improved. Two years later, that small observation would become the title of a famous paper: Language Models are Few-Shot Learners. You'll hear that story in episode six. 00:31:19.610 --> 00:31:33.270 That summer, OpenAI's most famous project went back to the biggest stage in Dota. Not one bot this time. Five of them, playing as a team, called OpenAI Five. 00:31:33.270 --> 00:31:58.910 In June, OpenAI had described how the five bots learned. The numbers are worth hearing, because they're the whole argument of this episode in miniature. OpenAI Five, the post said, played a hundred and eighty years' worth of games against itself every day. It trained on two hundred fifty-six GPUs and a hundred and twenty-eight thousand CPU cores. 00:31:58.910 --> 00:32:23.350 And the researchers admitted they'd been wrong about something. They'd assumed a game this long and this complicated would need fundamentally new ideas. Instead, they wrote, the results showed long-term planning could come from "large but achievable scale," without fundamental advances, "contrary to our own expectations upon starting the project." 00:32:23.350 --> 00:32:32.470 Contrary to their own expectations. Bigger worked. Bigger kept working. And bigger cost money they didn't have. 00:32:32.470 --> 00:32:47.030 Early August, they'd played a benchmark match in San Francisco against a team of strong human players, and won easily. All three games, by OpenAI's own description, were very one-sided. 00:32:47.030 --> 00:33:01.790 Then, Vancouver. The International, twenty eighteen. August twenty-second. OpenAI Five plays on the main stage against paiN Gaming, a Brazilian team, one of the top eighteen in the world. 00:33:01.790 --> 00:33:19.410 The game lasts about fifty-one minutes. The humans start strong. The bots claw back some ground in the middle. And then, in OpenAI's words, Five ends up "succumbing to various high-level strategic pushes by human players." 00:33:19.410 --> 00:33:31.130 The next day, they play a team of Chinese superstar players. Back and forth, team fight after team fight. Forty-five minutes. And again, the machines lose. 00:33:31.130 --> 00:33:45.030 Two games. Two losses. A year after the night Musk called a "triggering event," OpenAI had lost in public, at the same tournament, on the same kind of stage. 00:33:45.030 --> 00:33:58.750 OpenAI's post that week was gracious and unbowed. The games had been exciting and close, it said. The humans were far stronger than the benchmark team. And then, the diagnosis. 00:33:58.750 --> 00:34:06.890 "We think we need more training, bugfixes, and to remove the last pieces of scripted logic in our model." 00:34:06.890 --> 00:34:29.210 More training. Which means more compute. Which means more money. Eight months later, as you'll hear next episode, the same bots would come back and beat the world champions. And when OpenAI explained what had changed, it wouldn't be a clever new idea. It would be scale. 00:34:29.210 --> 00:35:00.430 That left OpenAI without its biggest backer, and with a compute bill that was, as you heard in the last episode, growing faster than anyone had planned for. In May twenty eighteen, two OpenAI researchers, Dario Amodei and Danny Hernandez, published an analysis. Since twenty twelve, the compute used in the largest AI training runs had been doubling about every three point four months. 00:35:00.430 --> 00:35:15.930 Moore's law, as the post put it, has a two-year doubling period. This was doubling about seven times faster. Since twenty twelve, the analysis found, the growth had been more than three hundred thousand times. 00:35:15.930 --> 00:35:22.210 Remember the name Dario Amodei. You've heard it before. You'll hear it again. 00:35:22.210 --> 00:35:34.050 By December, Altman was writing to Musk with progress: robot hands, faster language work, and one more line, almost in passing. 00:35:34.050 --> 00:35:41.330 Finally, I am working on a deal to move our computing from Google to Microsoft, in addition to our own data centers. 00:35:41.330 --> 00:35:48.070 Earlier in that same email, there was a line about language. It's easy to skip. Don't. 00:35:48.070 --> 00:35:56.790 We are also making fast progress on language. My hope is that next year we can generate short stories and a good dialogue bot. 00:35:56.790 --> 00:36:11.330 A good dialogue bot. December twenty eighteen. A little less than four years later, a dialogue bot built by this company would become what analysts called the fastest-growing consumer app in history. 00:36:11.330 --> 00:36:23.330 And one line about money. They had enough, Altman wrote, for about the next two years, even with aggressive growth. He'd rather not put the rest in email. 00:36:23.330 --> 00:36:35.950 Nine days later, the day after Christmas, two thousand eighteen, Musk sent Sutskever and Brockman one more email. The subject line was "I feel I should reiterate." 00:36:35.950 --> 00:36:49.450 My probability assessment of OpenAI being relevant to DeepMind slash Google without a dramatic change in execution and resources is zero percent. Not one percent. I wish it were otherwise. 00:36:49.450 --> 00:36:56.910 Even raising several hundred million won't be enough. This needs billions per year immediately or forget it. 00:36:56.910 --> 00:37:04.150 He wasn't finished. He reached for a comparison from his other life, in rockets. 00:37:04.150 --> 00:37:15.250 OpenAI reminds me of Bezos and Blue Origin. They are hopelessly behind SpaceX and getting worse, but the ego of Bezos has him insanely thinking that they are not! 00:37:15.250 --> 00:37:36.990 Just before that, there's a line in the email that trails off. "Unfortunately," it reads, "humanity's future is in the hands of." And then a link to an article, which OpenAI's published version doesn't show. We don't know whose hands he meant. The record stops mid-sentence. 00:37:36.990 --> 00:37:39.970 I really hope I'm wrong. 00:37:39.970 --> 00:38:13.750 Step back and look at what this one year and a half contains. In the summer of twenty seventeen, everyone in this story agreed the nonprofit might not be enough. By that September, they couldn't agree on who should be in charge of whatever came next. The man with the most money left. The ones who stayed wrote a charter promising their first duty was to humanity. And their own experiments, in Dota and in language, kept telling them the same thing: the answer was scale. 00:38:13.750 --> 00:38:22.430 None of them, on any side, was wrong about that. What they disagreed about was who gets to hold something that big. 00:38:22.430 --> 00:38:42.490 Billions per year. For a nonprofit, there's no such thing. So, eleven weeks later, OpenAI did something no one had quite done before. It invented a new kind of company and put it inside the nonprofit. A company where your profit had a ceiling. 00:38:42.490 --> 00:38:51.150 And the man who would put up the first billion dollars for it wasn't Elon Musk. He was in Redmond, Washington. 00:38:51.150 --> 00:38:59.150 Next time, on The Role Models: The Bet.