Generated from the production script. Stage directions removed. The final mix may differ slightly from the script; check against the audio before publishing.

Disclaimer: a dramatization based on public records, including court filings, published emails and company posts. Some scenes and dialogue are reconstructed or composited. Voices are AI-generated and perform words from the public record; they are not intended to resemble any real person. Not affiliated with or endorsed by OpenAI or any person depicted. Allegations in court filings are unproven claims. The Role Models is produced by Will Sigmon Media with AI-generated voices (ElevenLabs). Quoted emails, posts and filings are read verbatim and attributed; where the parties disagree, both accounts are given, and a company's post written during a lawsuit is treated as advocacy by one side. Sources for each episode are listed in its show notes.

[Season 1: OpenAI · Episode 3: "The Nonprofit Paradox"]

[COLD OPEN]

Narrator: Seattle. August 11th, 2017. KeyArena, home this week of The International, the world championship of a video game called Dota 2. A packed arena, and a huge audience watching online.

Narrator: But tonight's surprise isn't a team. It's a showmatch. One on one. On one side is Danylo Ishutin, "Dendi," Ukrainian, a former world champion, maybe the most beloved player the game has ever had. He walks out to cheering. On the other side there's no one. Just a computer and a program that nobody in the building has ever seen.

Narrator: The rules are simple. Both players get the same hero, Shadow Fiend. First to two kills wins, or first to destroy the other's tower. Here's what you need to know about the program. Nobody taught it to play Dota. No pro coached it, and no one programmed a strategy into it. It learned by playing against itself. Over and over. By the account of the people who built it, the equivalent of lifetimes of play, compressed into about two weeks of real time.

Narrator: Game one. Inside ten minutes, it's over. The bot outfarms the former world champion two to one and takes him down.

Narrator: Game two. Dendi tries something. It doesn't work. And then the former world champion concedes.

Narrator: On stage, microphone in hand, Dendi tries to describe what he just played against.

Dendi (on stage): "This guy is scary."

Narrator: The team that built the bot is OpenAI, the nonprofit that twenty months ago had no whiteboards. This is the result Elon Musk asked for in February 2016: something significant, to show that we are for real. He asked for it in six to nine months. It took eighteen. But here it is, in an arena, on a livestream, with the whole internet watching. It should be the happiest night in OpenAI's short life.

Narrator: That night, at 9:17 PM by the timestamp, Musk emails Greg Brockman, Ilya Sutskever, and Sam Altman. The first line is redacted. What survives is three sentences.

Musk (email): "Are you guys able to meet or do a conf call tomorrow afternoon? Time to make the next step for OpenAI. This is the triggering event."

Narrator: The triggering event. He doesn't say what it triggers. The email doesn't explain. To understand it, you need to know how a nonprofit that set out to give its research away ended up on an arena stage, beating a world champion with an AI that cost more to train than the charity could afford. That means going back to the spring of 2016, to a toolkit, a Microsoft term sheet, and an email from Ilya Sutskever about particle accelerators. This is the story of the paradox at the center of OpenAI. It's a nonprofit, and to do what it promised, it has to spend like the richest company on Earth.

Title Card: THE ROLE MODELS. Season One: OpenAI. Episode Three: "The Nonprofit Paradox."

[CHAPTER 2 – "The Gym"]

Narrator: April 27th, 2016. OpenAI is four and a half months old, and it ships its first thing. It isn't a mind. It isn't a chatbot. It's a gym. OpenAI Gym, released that morning as a public beta, is a toolkit for reinforcement learning, the trial-and-error method from last episode. The toddler learning to walk. Gym is the playground: a shared set of environments where anyone's learning algorithm can be dropped in and scored. Simulated robots trying to stand. Classic Atari games. A little cart trying to balance a pole. Everyone runs their AI on the same obstacle course, so results can be compared, reproduced, and argued about.

Narrator: In other words, the first thing OpenAI gave the world was a way for everyone else to do AI research better. Open, free, shared. On mission, to the letter. That same morning, Wired publishes its long feature about the lab, the one we've been quoting: the Napa winery, the living room with no whiteboards, borderline crazy money. It also carries a warning, from Yann LeCun, one of the three godfathers of deep learning, now running AI research at Facebook.

Lecun (Wired): "There are certainly some competing objectives. It's a non-profit, but then there is a very close link with Y Combinator. And people are paid as if they are working in the industry."

Narrator: Paid as if they are working in the industry. A nonprofit with industry salaries. LeCun has spotted the paradox before the paradox knows it exists. And also that same day, April 27th, a very different kind of email lands in Elon Musk's orbit. A New York Times reporter working with the columnist Maureen Dowd forwards a question. Mark Zuckerberg has been talking about Musk. Asked in Germany about Musk's AI fears, Zuckerberg called them "hysterical," the email says, and added that some people "fearmonger" about AI. Musk's answer, sent back that afternoon for the record, is short, and he means every word of it.

Musk (statement to the NYT): "History unequivocally illustrates that a powerful technology is a double-edged sword. It would be foolish to assume that AI, arguably the most powerful of all technologies, only has a single edge. The recent example of Microsoft's AI chatbot shows how quickly it can turn incredibly negative. The wise course of action is to approach the advent of AI with caution and ensure that its power is widely distributed and not controlled by any one company or person. That is why we created OpenAI."

Narrator: The Microsoft chatbot was Tay. A month earlier, Microsoft had released a teenage-voiced bot on Twitter that learned from the people who talked to it. Within a day, the internet had taught it to be vile, and Microsoft pulled the plug. That was spring 2016's picture of AI gone wrong: not a superintelligence, a mirror. Not controlled by any one company or person. That is why we created OpenAI. Mark the date. That's Musk in April 2016, defining the mission as nobody gets control. We'll hear him on control again at the end of this episode, and the next one.

Narrator: Summer 2016. The lab is growing, and some of the new people aren't here to win games. They're here because they're worried. In June, a paper appears online with the most unglamorous title imaginable: "Concrete Problems in AI Safety." Six authors. Two of them matter enormously to this season. One is John Schulman, an OpenAI founder. The first author, still at Google Brain when the paper goes up, is Dario Amodei. Remember that name. I asked you to in episode one, when it showed up in the acknowledgments of Sam Altman's 2015 blog post. That same year, 2016, Dario Amodei joins OpenAI. The paper doesn't talk about killer robots. It talks about mundane, practical failures. A cleaning robot that knocks over a vase because nobody told it vases matter. A system that finds a loophole in its reward, scoring points without doing the job, like a kid who learns the teacher only checks whether the homework page is full. The paper's point is that you don't need an evil superintelligence for things to go wrong. You need a very capable system and a goal that's slightly off. It's the first sign of something that will define the next decade. Inside OpenAI there are two kinds of urgency. One says: build it before the wrong people do. The other says: make sure whatever gets built doesn't go wrong. For now they share an office and a mission statement. The talent keeps coming, too, and so does the payroll. That June, Pieter Abbeel, the Berkeley robotics professor who advised John Schulman's PhD and was one of the lab's launch advisors, comes aboard on leave from the university. The 2016 tax filing will later list him at four hundred and twenty-five thousand dollars. Ian Goodfellow, the Google Brain star from last episode, is already inside. By the end of the year the lab has around fifty people. That's the flywheel Musk wanted: the best people attract the best people. It's also a flywheel with a price, and every turn of it costs more than the last. In 2016 the price is still people. Soon it won't be.

Narrator: December 5th, 2016. OpenAI ships something bigger. They call it Universe. If Gym was a playground, Universe is the whole city. It's a platform that lets an AI use a computer the way you do: looking at the pixels on a screen, moving a virtual mouse, pressing virtual keys. Thousands of games, websites, and apps, any of them a training ground. The dream is general intelligence: not an AI that masters one game, but one that can sit down in front of any screen and figure it out. In a year, OpenAI has released a playground and a universe, and it's publishing papers and giving code away. By every public measure, the nonprofit is doing exactly what it said it would. What the public can't see is the bill.

[CHAPTER 3 – "Sixty for Ten"]

Narrator: Here's a secret about artificial intelligence that most people in 2016 haven't absorbed. It runs on electricity and sand. Specifically, on graphics processors, the GPUs from episode one, the gaming cards that made AlexNet possible. By 2016 the serious work needs hundreds of them, rented by the hour from cloud giants like Amazon, Google, and Microsoft. And every hour costs money. The kind of money a nonprofit counts.

Narrator: September 16th, 2016. 2:37 in the afternoon. Sam Altman forwards Musk a proposal. Subject line: "MSFT hosting deal."

Altman (email): "Here are the MSFT terms. $60MM of compute for $10MM, and input from us on what they deploy in the cloud. LMK if you have any feedback."

Narrator: Sixty million dollars of computing for ten. An extraordinary discount. Microsoft wants in on AI, and a famous nonprofit on its cloud is a trophy. But there's fine print. The term sheet says OpenAI will evaluate Microsoft's AI software, work with Microsoft to evangelize those products, and promote Azure as its preferred public cloud provider. There's a joint press release and a stage appearance at Microsoft's big conference with Satya Nadella. Thirty-three minutes later, Musk replies.

Musk (email): "This actually made me feel nauseous. It sucks and is exactly what I would expect from them."

Narrator: Nauseous. He doesn't want OpenAI promoting anybody.

Musk (email): "We should just do this low key. No certainty either way. No contract."

Narrator: Altman goes back to Microsoft. Four days later, Musk's chief of staff, Sam Teller, reports that Microsoft has softened. They'll do it at fifty million, at OpenAI's sole discretion, with the right for either side to walk away at any time. And, in Teller's words:

Teller (email): "No evangelizing. No strings attached. No looking like lame Microsoft marketing pawns. Ok to move ahead?"

Narrator: Musk says fine, as long as Microsoft doesn't use it in its messaging. His exact reply is a little too colorful for this show, but the gist is that it's worth far more than fifty million dollars not to look like Microsoft's marketing tool. By year's end, OpenAI is running on Microsoft's cloud, Azure. Pause on this scene, because it's the whole paradox in miniature. The nonprofit needs computers it can't afford. The only people who have them are the giants it was created to counterbalance. So it takes the giant's discount and argues hard about the strings. No evangelizing. No strings. No marketing pawns. It works this time. Remember Microsoft, though. Remember Musk feeling nauseous at the idea of OpenAI serving a big tech company. Three years from now, in episode five, Microsoft comes back with a much bigger offer, and by then Musk won't be in the room.

Narrator: And here's what the discount was covering, from the tax filings. In 2017, OpenAI will spend seven point nine million dollars on cloud computing, about a quarter of everything it spends that year. In 2016 there were fifty-two people and an eleven-million-dollar budget. By 2017, a quarter of the budget isn't going to people at all. It's going to machines.

Narrator: For comparison: that year, the other side's UK filings showed DeepMind with four hundred employees and a hundred and thirty-eight million dollars in total costs. That's about three hundred and forty-five thousand dollars per employee, average, counting everyone. And DeepMind doesn't have to raise a dime. It has Google.

[CHAPTER 4 – "The Particle Accelerator"]

Narrator: In February 2016, you heard Ilya Sutskever write that ideas matter more than hardware. Powerful ideas are produced by top people. Massive clusters help, and are very worth getting, but they play a less important role. Now watch him change his mind. In the record it happens in stages, like a scientist updating on evidence, which is exactly what it is. Here's how OpenAI later summarized it, in its 2024 post:

Openai (blog post): "In early 2017, we came to the realization that building AGI will require vast quantities of compute."

Narrator: Vast quantities. What happened in early 2017? To understand it, you need one idea, and it's the most important idea in this whole season. So let's slow down.

Narrator: Imagine you're a scientist and every experiment takes a day. Try an idea, wait overnight, look at the results in the morning, adjust, try again. You get about three hundred shots a year. Now imagine the experiment takes a month. You get twelve shots a year. You lose the thread. You forget why you tried the thing. Your curiosity dies waiting. Now imagine the experiment takes an hour. That's what compute buys you. Not just a bigger brain for the machine, but speed for the humans. More shots on goal. Faster learning about learning. Sutskever would put a number on that feeling. In his July 2017 email, which we'll get to in a moment, he wrote:

Sutskever (email): "Historically, AI breakthroughs have consistently happened with models that take between 7–10 days to train. This means that hardware defines the surface of potential AI breakthroughs."

Narrator: A week to ten days. Long enough to be ambitious, short enough to keep the whole experiment in your head. And he's careful to say it's a fact about people, not machines: "a statement about human psychology more than about AI," in his words. Any longer, and the researcher loses the thread. So the question becomes: how big a model can you train in ten days? And that is set entirely by how many computers you have. And here's the twist that arrived in 2016 and 2017. For years, having a thousand computers didn't help you run one big experiment faster. It just let you run a thousand small ones side by side. A university with ten good machines could keep up with Google, because Google's advantage was quantity, not speed. Then that changed. Researchers figured out how to yoke hundreds of GPUs together into one experiment, one enormous training run spread across a warehouse of machines. And the moment that worked, the advantage stopped being clever people. It became clever people with a warehouse. June 12th, 2017. Sutskever sends Musk the first of a new series of bi-weekly updates, and he puts it bluntly.

Sutskever (email): "95% of progress comes from the ability to run big experiments quickly."

Narrator: Ninety-five percent. And then the price tag:

Sutskever (email): "In order to be capable of accomplishing our projects even in theory, we need to increase the number of our GPUs by a factor of 10x in the next 1–2 months."

Narrator: Ten times the machines. In one or two months. Even in theory. The next morning, Musk answers:

Musk (email): "Ok. Let's figure out the least expensive way to ensure compute power is not a constraint..."

Narrator: Least expensive. Hold on to that phrase for the rest of the episode.

Narrator: That same month, June 2017, something else happens, and it gets far less attention than it should. Andrej Karpathy, one of the original eleven, the Stanford student who collected internships at every lab Musk feared, leaves OpenAI. He goes to Tesla, as Director of AI, to lead the neural networks behind Autopilot. He'll report to Elon Musk. Remember launch day. Levy's interview. Musk, almost in passing: Probably Tesla will have more real world data than any other company in the world. Now one of OpenAI's founding researchers is building Tesla's AI. It's a perfectly normal career move in Silicon Valley. It's also the first thread of a knot that will take a courtroom to untie. OpenAI's co-chair also runs a car company that needs AI, and that company has just hired one of OpenAI's own.

Narrator: July 12th, 2017. Ilya Sutskever sends Musk and Brockman the most important email of the nonprofit era. Its subject line: "The business of building AGI." It opens like a scientific confession.

Sutskever (email): "We usually decide that problems are hard because smart people have worked on them unsuccessfully for a long time. It's easy to think that this is true about AI. However, the past five years of progress have shown that the earliest and simplest ideas about AI — neural networks — were right all along, and we needed modern hardware to get them working."

Narrator: Right all along. That's the deep learning conspiracy from episode one, the believers who kept the faith for thirty years, and one of its founding students writing its verdict. The idea was never wrong. The computers were too slow. Then he reaches for a metaphor.

Sutskever (email): "It's not so much that AI progress is a hardware game, any more than physics is a particle accelerator game. But if our computers are too slow, no amount of cleverness will result in AGI, just like if a particle accelerator is too small, we have no shot at figuring out how the universe works."

Narrator: Here's the Weir beat, because this is how the impossible thing actually works. To find the smallest pieces of the universe, physicists smash particles together at enormous energies. The harder you smash, the deeper you see. The Large Hadron Collider is a ring of magnets seventeen miles around, because there's no clever trick that lets you see the Higgs boson with a small one. You can be Einstein, and without the collider the answer stays hidden. Sutskever is saying AI is the same. The algorithm might be simple. It might already be sitting in someone's notebook. But you only find out if you can build a collider big enough to test it, and OpenAI's collider is too small. Here's how much bigger it needs to be:

Sutskever (email): "Increase our GPU cluster from 600 GPUs to 5000 GPUs ASAP."

Narrator: Six hundred to five thousand. And the team's own estimate, in the same email, is that Google Brain has something like a hundred thousand GPUs. First, a set of predictions for the next three years. Some will land, some won't, and that's part of what makes them worth hearing:

Sutskever (email): "Within the next three years, robotics should be completely solved, AI should solve a long-standing unproven theorem, programming competitions should be won consistently by AIs, and there should be convincing chatbots (though no one should pass the Turing test)."

Narrator: Robotics, completely solved in three years? No. Not even close, not by 2020. The theorem, the programming contests? Those took longer too. But convincing chatbots. In July 2017, in a private email, the chief scientist of OpenAI predicted the thing that would change everything five years later, and he put it in a parenthesis at the end of a list. Then the two sentences that, read today, might be the most prophetic in this whole archive.

Sutskever (email): "In as little as four years, each overnight experiment will feasibly use so much compute capacity that there's an actual chance of waking up to AGI, given the right algorithm"

Narrator: Waking up to AGI. Picture it the way he pictured it: an engineer kicks off a training run at night, goes home, sleeps, and in the morning there's something on the servers that wasn't there before. And he puts a price on the whole thing:

Sutskever (email): "we have reason to believe AGI can ultimately be built with less than $10B in hardware."

Narrator: Less than ten billion dollars. Ten billion, for an organization that raised a hundred and thirty-three million over its whole life. He closes with the line that turns a research memo into a funding crisis:

Sutskever (email): "If we can secure the funding, we have a real chance at setting the initial conditions under which AGI is born."

Narrator: The initial conditions under which AGI is born. That's why this matters to them. It isn't about winning. They believe whoever builds the first one sets its starting values. Get the birth right and maybe you get the rest right. And the birth, by their math, costs billions.

Narrator: May 2018. Dario Amodei and a colleague, Danny Hernandez, publish an analysis at OpenAI called "AI and Compute." They measure the computing power used in the largest AI training runs since 2012 and find that it has been doubling every three and a half months. Moore's Law, the engine of the whole computer age, doubles about every eighteen. Since AlexNet's two gaming cards, the largest runs had grown more than three hundred thousand times.

Narrator: Every three and a half months, the price of staying in the race doubles. Donations don't double every three and a half months. That's the nonprofit paradox, in one line of arithmetic.

[CHAPTER 5 – "Something Out of Nothing"]

Narrator: So why Dota? Why would a lab trying to build general intelligence spend its compute budget on a video game? Because Dota 2 is, in a technical sense, a nightmare. Chess is a board. You see every piece. Go is a board. Everything is visible, one move at a time. Dota is a battlefield in fog. You can't see most of the map. Everything happens in real time, dozens of decisions a second, for forty minutes. Your enemy hides. Your team has to coordinate. The number of possible situations makes Go look like tic-tac-toe. If an AI can learn that from scratch, it's learning something much closer to the messy, hidden, real-time world we live in. And the method is the one Ilya is most excited about. It has a name: self-play. In his June update, he tried to explain it to Musk.

Sutskever (email): "Self play in multiagent environments is magical: if you place agents into an environment, then no matter how smart (or not smart) they are, the environment will provide them with the exact level of challenge, which can be faced only by outsmarting the competition."

Narrator: Here's the intuition. Imagine learning tennis, but your only opponent is you, an exact copy, who improves exactly as fast as you do. You can never coast. Every time you get better, your opponent gets better. The challenge scales forever. And one more line, because it's beautiful:

Sutskever (email): "Self-play lets us get 'something out of nothing.'"

Narrator: Something out of nothing. Simple rules, and out of them, through endless self-competition, strategies no human wrote down.

Narrator: That's how OpenAI's bot trained. One of its researchers, Jakub Pachocki, put it simply after the match:

Pachocki (TechCrunch): "We've coached it to learn just from playing against itself… We didn't hard-code in any strategy, we didn't have it learn from human experts, just from the very beginning, it just keeps playing against a copy of itself."

Narrator: Remember that name too, Jakub Pachocki. He'll come back much later in this story, in a much bigger role. It isn't free, though. Self-play is compute-hungry by nature. Every game the bot plays against itself is a game that has to run on a real machine. In June, Ilya wrote that every Dota experiment already used more than a thousand cores, and that was just the small, one-on-one version. For the full five-on-five game, he warned, each experiment would have to be at least ten times bigger. The bot getting better and the bill getting bigger are the same thing. Through the summer of 2017, the updates to Musk read like a countdown.

Sutskever (email): "1v1 bot is no longer exploitable. It can no longer be beaten using 'unconventional' strategies. On track to beat all humans in 1 month."

Narrator: On track to beat all humans in one month. And there's a point of pride tucked into the same update, the kind of thing a scientist sends when he wants his funder to know the lab punches above its weight:

Sutskever (email): "DeepMind's directly used one of our algorithms to produce their parkour results"

Narrator: The algorithm was called PPO, Proximal Policy Optimization, and it came out of John Schulman's work at OpenAI. It's a way to let a learning system improve in steady, careful steps instead of lurching. DeepMind, the giant, the rival, the source of Musk's extreme mental stress, had used the little nonprofit's method to teach its simulated figures to run and jump over obstacles. File PPO away. It sounds like a footnote. Five years from now it will be one of the hidden engines inside the most famous chatbot in the world. And in the same stretch, there's the robot hand. Sutskever's June update promised that within six months the lab would teach a robotic hand at least one feat of dexterity: pen spinning, Chinese balls, or a single-handed Rubik's cube. By July 20th:

Sutskever (email): "The robot hand can now solve a Rubik's cube in simulation"

Narrator: In simulation. The real hand would take much longer. But you can hear the ambition. Games, robots, algorithms other labs borrow. The nonprofit is producing. It's just producing at a price it can't pay. The same update has a list at the bottom, "Coming up." The first item isn't a game or a robot. It's this:

Sutskever (email): "Designing the for-profit structure"

Narrator: The for-profit structure. This is the first time in this season those words appear in the record from inside OpenAI. The nonprofit, nineteen months old, has started designing a company. The week before, Brockman had met with Musk and sent a text summary to Shivon Zilis, an advisor close to Musk. Two lines from it:

Brockman (iMessage): "gpus: said do what we need to do"

Narrator: And about the structure, Brockman reports Musk's view, with his own agreement attached:

Brockman (iMessage): "(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)"

Narrator: Non-profit was the right one early on. May not be the right one now. That's Brockman's report of what Musk said, not Musk's own words. But in July 2017 nobody in this conversation is defending the nonprofit as permanent. Not Brockman, not Sutskever, and by Brockman's account, not Musk. Hold that. In a few years, whether a for-profit OpenAI was a betrayal will be the subject of a federal lawsuit. The record shows the conversation starting here, in the summer of 2017, with all three men on the same side of it.

Narrator: July 21st. Musk forwards a New York Times story to Brockman and Sutskever: Beijing wants AI to be made in China by 2030.

Musk (email): "They will do whatever it takes to obtain what we develop. Maybe another reason to change course."

Narrator: And Brockman answers with a plan. Three steps, typed out like a product roadmap for civilization.

Brockman (email): "100% agreed. We think the path must be: 1. AI research non-profit (through end of 2017) 2. AI research + hardware for-profit (starting 2018) 3. Government project (when: ??)"

Narrator: Step one: the nonprofit, through the end of 2017. It has an expiration date. Step three: government project. When: question mark, question mark. Remember Altman's first email in May 2015, pitching a Manhattan Project for AI? Here's the Manhattan Project again, now with two question marks where the date should be. Musk's reply comes the same minute, by the timestamp:

Musk (email): "Let's talk Sat or Sun. I have a tentative game plan that I'd like to run by you."

[CHAPTER 6 – "The Triggering Event"]

Narrator: So now, back to Seattle. August 11th, 2017. KeyArena. Dendi walks out. The bot waits. You know what's on the other side of that screen now. Not a single program on a laptop, but a training run spread across more than a thousand processor cores, playing itself for two weeks, lifetimes of Dota, on computers a nonprofit can barely afford. You know it was on track to beat all humans in one month, and that it's right on schedule.

Narrator: Game one, under ten minutes. Game two, the concession.

Dendi (on stage): "This guy is scary."

Narrator: Afterward, Brockman explains to the press that the bot learned by playing against itself for about two weeks of real time. The internet loses its mind. Clips everywhere. The AI that beat Dendi. And Elon Musk tweets, twice. The first is pure pride.

Musk (tweet): "OpenAI first ever to defeat world's best players in competitive eSports. Vastly more complex than traditional board games like chess & Go."

Narrator: The second is the man from episode one, the one who warned about summoning the demon.

Musk (tweet): "If you're not concerned about AI safety, you should be. Vastly more risk than North Korea."

Narrator: That was August 2017, when North Korea was testing missiles and the word nuclear was in every headline. Musk was saying AI was worse. Pride and dread in the same breath. The man who funded the counterweight, celebrating how powerful it's become, and warning the world how dangerous that power is.

Narrator: It doesn't end at one-on-one. In April 2019, OpenAI's five-bot team, OpenAI Five, beats OG, the reigning world champions, two games to none. On stage, Greg Brockman says the bots have collectively played forty-five thousand years of Dota 2. Forty-five thousand years. Longer than there has been farming, or writing, or cities. All of it happened inside the machines of a lab that started with a hundred and thirty-three million dollars and a living room.

Narrator: But tonight is August 11th, 2017. The arena is emptying. Somewhere in Seattle, OpenAI's researchers are celebrating the proof Musk asked for eighteen months ago: something significant, to show we are for real. And at 9:17 PM, by the timestamp, the email.

Musk (email): "Time to make the next step for OpenAI. This is the triggering event."

Narrator: We don't know exactly what Musk meant. The email doesn't say, and the call the next afternoon isn't in the public record. But we know what the "next step" had been, in every email that summer. The nonprofit was the right structure early on and may not be the right one now. Coming up: designing the for-profit structure. The path: nonprofit through the end of 2017, then a for-profit. And ten times the computers, then thousands more. Less than ten billion dollars for the whole thing. The triggering event is proof. Proof the team is for real. Proof the method works. Proof that the only thing between them and something much bigger is money.

[OUTRO – CLIFFHANGER]

Narrator: Let's add up where OpenAI stands, twenty months after the handshake. It kept its promises to the world. A gym, a universe, papers, open code. It kept its promise to its people: it paid them, and it didn't lose them to the bidding wars. Except one: Karpathy, gone to Tesla. And it delivered the result Musk demanded, on the biggest stage in esports. And it's running out of road. The science turned. Right all along, and we needed modern hardware. The compute bill is doubling faster than any charity can raise. A quarter of the budget is already going to machines, and the chief scientist wants eight times more of them, as soon as possible. And the people are expensive too. They're the same people who took less to come here, and who can be reached by every recruiter in the Valley. Sutskever's July plan called for headcount to go from fifty-five to two hundred in two and a half years. Every one of them needs a salary a nonprofit has to raise, one donation at a time. The pledges on the launch page were a flag, not a bank account. And the lab's own leaders have started drafting a company. So here's the question the next episode is built on. If OpenAI becomes a company, somebody has to own it. Somebody has to control it. Remember the launch interview. The best defense against the misuse of AI is to empower as many people as possible. Not any one person. Remember April 2016: not controlled by any one company or person. That is why we created OpenAI. Now picture the negotiating table in the fall of 2017. One founder is the richest man in the room and has written most of the checks. He has his own car company, an AI program at that car company, and one of OpenAI's own founding researchers building it. Two other founders are scientists who have just realized they need billions. And one founder is the president of Y Combinator, who first pitched this whole thing as a Manhattan Project. A nonprofit built so that no one person could control the future is about to decide who controls it. In a few weeks, one of them will write an email with the subject line "Honest Thoughts." Another will answer in four words that end the conversation, and the argument will follow them into a federal courtroom nine years later.

Narrator: Next time, on The Role Models: the equity, the board seats, the CEO title, and a proposal to attach OpenAI to Tesla. Two founders, two completely different stories about why one of them walked out. We'll give you both, side by side, and let you decide. Episode Four: "The Divorce."