The Role Models - The Bet - transcript
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.

[0:01] Narrator: Monday, March eleventh, two thousand nineteen. OpenAI publishes a blog post with a two-word title. OpenAI LP.

[0:11] Narrator: It's signed by Greg Brockman and Ilya Sutskever. The two men who, eighteen months earlier, told Elon Musk they were willing to give up their equity rather than give him control. And this post is about equity.

[0:26] Greg Brockman (post): Our solution is to create OpenAI LP as a hybrid of a for-profit and nonprofit, which we are calling a capped-profit company.

[0:36] Narrator: A capped-profit company. Investors can make money. Employees can get something like startup stock. But only up to a point. And for the very first investors, the post names the point.

[0:51] Greg Brockman (post): Returns for our first round of investors are capped at one hundred x their investment.

[0:56] Narrator: A hundred times. Put in ten million dollars, and the most you can ever take out is a billion. Everything past that, the post says, belongs to the nonprofit. And the post adds a line that's either very modest or not modest at all, depending on how you hear it. If they succeed, it says, they expect to generate "orders of magnitude more value" than they'd owe the people who invest in or work at the company.

[1:26] Narrator: Think about what a cap of a hundred times means. It's only a limit if you think you might blow through it.

[1:37] Narrator: From Will Sigmon Media, this is The Role Models. Season one: OpenAI. Episode five. The Bet.

[1:48] Narrator: To understand the bet, go back eleven weeks, to where we left off. December twenty-eighteen. Musk is gone from the board. And Sam Altman sends him an update, which OpenAI later published. Dota. Robot hands. Fast progress on language. And near the end, one sentence about where the computers would live.

[2:12] Sam Altman (email): Finally, I am working on a deal to move our computing from Google to Microsoft, in addition to our own data centers.

[2:20] Narrator: Nine days later, Musk wrote back to Sutskever and Brockman. You heard this at the end of last episode.

[2:28] Elon Musk (email): Even raising several hundred million won't be enough. This needs billions per year immediately or forget it.

[2:36] Narrator: Billions per year. The nonprofit didn't have that. So the March post opens by admitting it.

[2:43] Greg Brockman (post): We've experienced firsthand that the most dramatic AI systems use the most computational power in addition to algorithmic innovations, and decided to scale much faster than we'd planned when starting OpenAI.

[2:56] Narrator: Much faster than we'd planned. That's the founding plan, from episode one, quietly retired in a single clause. The lab that was supposed to win with brilliant people and open papers had found out, the hard way, what winning cost.

[3:15] Greg Brockman (post): We'll need to invest billions of dollars in upcoming years into large-scale cloud compute, attracting and retaining talented people, and building AI supercomputers.

[3:26] Narrator: And then it says the thing every founder in this story has circled for three years.

[3:32] Greg Brockman (post): We want to increase our ability to raise capital while still serving our mission, and no pre-existing legal structure we know of strikes the right balance.

[3:42] Narrator: No structure they knew of. So they built one. But before we open it up, there's something that happened four weeks earlier. Because for one week in February, OpenAI was the most talked-about AI lab in the world, and not because of money.

[4:00] Narrator: Thursday, February fourteenth, two thousand nineteen. Valentine's Day. OpenAI announces a language model called GPT-2.

[4:12] Narrator: It had one job. Read some text, and predict the next word. That's it. It was trained on forty gigabytes of internet text, about eight million web pages. And it had one point five billion parameters, more than ten times as many as the GPT that came before it.

[4:32] Narrator: To understand why this one model scared its own makers, you have to hear what it wrote. OpenAI published a handful of samples. The most famous one began with a prompt, two sentences written by a person.

[4:49] Narrator: "In a shocking finding, scientist discovered a herd of unicorns living in a remote, previously unexplored valley, in the Andes Mountains. Even more surprising to the researchers was the fact that the unicorns spoke perfect English."

[5:04] Narrator: That's all the machine was given. Here's how it continued.

[5:09] Narrator: "The scientist named the population, after their distinctive horn, Ovid's Unicorn. These four-horned, silver-white unicorns were previously unknown to science."

[5:21] Narrator: "Dr. Jorge Pérez, an evolutionary biologist from the University of La Paz, and several companions, were exploring the Andes Mountains when they found a small valley, with no other animals or humans."

[5:34] Narrator: There is no Dr. Jorge Pérez. The machine made him up, along with his university post and his companions, and then it gave him quotes. It wrote like a wire reporter. It invented a scientist, because that's what a news story about a discovery has in it. Four-horned unicorns named after a horn. It's nonsense. And it's fluent nonsense, which is a different and newer thing.

[6:01] Narrator: OpenAI was candid that this was a best case. They'd hand-picked the samples, and the post admitted to what it called "some meta-cherry-picking." On popular topics, it said, the model produced a reasonable sample about half the time.

[6:17] Narrator: What came out surprised people. Give it the opening of a news story, and it would keep going, in the voice of a news story, for paragraphs. It wasn't perfect. OpenAI's own post listed the failure modes. Repetitive text. Sudden changes of subject. And, in one of the best lines in any AI announcement, the model, quote, "sometimes writes about fires happening under water."

[6:44] Narrator: The post went further than samples. Under a heading called Policy Implications, it listed what a model like this could be used for. The good list: writing assistants, better dialogue agents, translation, speech recognition. And the other list.

[7:03] Narrator: "Generate misleading news articles. Impersonate others online. Automate the production of abusive or faked content to post on social media. Automate the production of spam slash phishing content."

[7:17] Narrator: And then a sentence that, in twenty nineteen, read like a warning, and now reads like a weather report. "The public at large will need to become more skeptical of text they find online."

[7:30] Narrator: But the headline wasn't the model. It was what OpenAI decided not to do with it.

[7:36] Narrator: "Due to our concerns about malicious applications of the technology, we are not releasing the trained model."

[7:43] Narrator: They released a much smaller version, and a paper. The full model stayed locked up. The post called this "an experiment in responsible disclosure." And it was honest about not knowing whether it was right: "we are not sure that it is the right decision today."

[8:01] Narrator: And it pointed back to a promise. "Nearly a year ago," the post said, "we wrote in the OpenAI Charter" that safety and security concerns would reduce their traditional publishing. This, they said, might be "the early beginnings of such concerns." Last episode, you heard the charter make that prediction. Here it was, coming true, ten months later.

[8:28] Narrator: They also admitted what everyone in the field already knew. Other researchers had the skills to rebuild the model on their own. Holding it back wouldn't stop that. It would only, they hoped, buy some time.

[8:43] Narrator: The press heard it differently. The Guardian's headline that day read, "New AI fake text generator may be too dangerous to release, say creators." Notice those words, "too dangerous to release." They're the Guardian's. OpenAI's post never used them. But that's the phrase that stuck, and it's the phrase people still remember.

[9:08] Narrator: In the same Guardian piece, OpenAI's head of policy, Jack Clark, gave the reporter a phrase of his own. Hold on to it. It's the title of our next episode. "I have a term for this," he said. "The escalator from hell."

[9:26] Narrator: Clark also explained the thinking behind holding the model back, and it was less about certainty than about not knowing. "If you can't anticipate all the abilities of a model," he told the Guardian, "you have to prod it to see what it can do. There are many more people than us who are better at thinking what it can do maliciously."

[9:47] Narrator: One more detail from that February post, and it matters later. At the bottom, in the list of authors, between Jeffrey Wu and Daniela Amodei, is the name Dario Amodei.

[10:00] Narrator: Three days before the LP announcement, on Friday, March eighth, another post went up, on Y Combinator's blog. Sam Altman, it said, was transitioning to chairman of YC, and handing his operational duties to other partners. The reason was one line. It would "allow Sam to spend more time focusing on OpenAI."

[10:25] Narrator: For five years, Altman had run the most famous startup accelerator in the world, with OpenAI on the side. Now, for the first time, OpenAI would have a full-time chief executive, and it would be him. In May, he gave up the YC chairman title too. "Working on OpenAI," he wrote, "has been more exciting, and more complex, than I ever imagined."

[10:51] Narrator: Remember Honest Thoughts, from last episode? "We don't understand why the CEO title is so important to you." A year and a half later, the title was his.

[11:04] Narrator: Back to March, and the new company. How do you build a business that's allowed to make money, but promises not to put money first? OpenAI's answer was paperwork. Literally.

[11:18] Greg Brockman (post): Our employee and investor paperwork start with big purple boxes like this.

[11:24] Narrator: The purple box, as the post describes it, says the company's first duty is to the OpenAI Charter, their mission statement, and not to the investors. Everyone who invests or joins signs agreements that the Charter comes first, quote, "even at the expense of some or all of their financial stake."

[11:45] Narrator: And there's a line, near the top of that section, that is easy to read past. Listen to it slowly, because six years later, it will matter more than almost anything else in the post.

[11:58] Greg Brockman (post): The mission comes first even with respect to OpenAI LP's structure. While we are hopeful that what we describe below will work until our mission is complete, we may update our implementation as the world changes.

[12:11] Narrator: We may update our implementation as the world changes. In other words: this structure is not forever. The mission is the fixed point. The company built around it can change. They meant it as a promise of flexibility. Later, people who sued them would read it very differently.

[12:33] Greg Brockman (post): Regardless of how the world evolves, we are committed, legally and personally, to our mission.

[12:39] Narrator: And there was a second guardrail, in the boardroom. The nonprofit's board controls the company. Only a minority of that board can hold a financial stake. And on any decision where the investors' interests and the mission might collide, including payouts, only the board members without a stake get to vote.

[13:01] Narrator: Then the cap itself. A hundred times, for the first investors. The post called that "commensurate with the risks in front of us," and added that they expected the multiple to be lower for future rounds.

[13:15] Narrator: A hundred times is a lot of room. The cap was real. It was also very high. Both of those were true at once.

[13:25] Narrator: And there was one more promise, carried over from the charter, now with money attached. If another value-aligned project got close to AGI first, OpenAI said it was willing to merge with it, "even if it means reduced or zero payouts to investors."

[13:43] Narrator: Who were those investors? The post named two: Reid Hoffman's charitable foundation, and Khosla Ventures. It said the company now had around a hundred employees. And it insisted the day-to-day work wasn't changing.

[14:00] Greg Brockman (post): Today, we believe we can build the most value by focusing exclusively on developing new AI technologies, not commercial products.

[14:10] Narrator: Not commercial products. Hold that sentence next to the one you'll hear at the end of this episode.

[14:17] Narrator: Keep that design in your head. A nonprofit board, mostly people with no money in the game, sitting on top of a company that might one day be worth more than almost anything. Four and a half years later, that board will use exactly that power. But that's episode ten.

[14:37] Narrator: Who was on the board in March twenty-nineteen? The post lists them. Three employees: Greg Brockman, chairman and CTO. Ilya Sutskever, chief scientist. And Sam Altman, now listed with one word after his name. CEO. Then five people from outside: Adam D'Angelo, Holden Karnofsky, Reid Hoffman, Shivon Zilis, and Tasha McCauley.

[15:05] Narrator: And there's one more line, about the man who isn't on that list.

[15:10] Greg Brockman (post): Elon Musk left the board of OpenAI Nonprofit in February twenty eighteen and is not formally involved with OpenAI LP. We are thankful for all his past help.

[15:21] Narrator: Years later, in the middle of a lawsuit, OpenAI published its account of what happened before that post went up. They say they sent Musk an advance copy. And they say his only feedback was one sentence. "Please be explicit that I have no financial interest in the for-profit arm of OpenAI."

[15:42] Narrator: That's OpenAI's telling, written by a party to the lawsuit, about a man who was by then suing them. Musk's case argues the opposite story: that the for-profit itself was a betrayal of what he funded. Both of those things can be in the record at once. We'll keep them side by side.

[16:04] Narrator: Saturday, April thirteenth, two thousand nineteen. A month after the purple boxes. On a livestream, OpenAI's Dota bots, called OpenAI Five, play the reigning world champion team, a team called OG.

[16:21] Narrator: They win. Two games, back to back. By OpenAI's account, it's the first time an AI has beaten the world champions in an esports game.

[16:32] Narrator: The bots had lost to pros at the big tournament the summer before. So what changed? OpenAI's answer is one word.

[16:42] Narrator: Compute. Eight times more training compute than the version that lost. In total, the post says, OpenAI Five had experienced about forty-five thousand years of Dota, playing itself, over ten real months.

[16:57] Narrator: And then the line that explains this whole episode. They had expected to need clever new algorithms. Instead, they wrote, "the fundamental improvement we needed for this problem was scale."

[17:09] Narrator: Scale. That is what the capped-profit company was built to buy. And three months later, someone would sell it to them.

[17:21] Narrator: But first, OpenAI did something unusual. It let the public in. For four days that April, anyone could sign up and play OpenAI Five, either against it, or on its team. They called it Arena.

[17:37] Narrator: At the Finals, they'd already found a surprise. The bots had been trained only to beat copies of themselves. But when you put a human on their team, they cooperated. A Dota caster known as Sheever played alongside them, and OpenAI quoted her afterward.

[17:55] Narrator: "It actually felt nice. My Viper gave his life for me at some point. He tried to help me, thinking, I'm sure she knows what she's doing, and then obviously I didn't. But, you know, he believed in me. I don't get that a lot with human teammates."

[18:12] Narrator: He believed in me. A neural network, trained by fighting itself for forty-five thousand years, sacrificing itself for a stranger. OpenAI wrote that the ease of turning a competitive AI into a cooperative one made them "hopeful that future AI systems can be very beneficial for humans."

[18:35] Narrator: Then they retired it. The post said OpenAI Five was done as a competitor. The techniques would move on to other work, including robots.

[18:46] Narrator: And four days after Arena closed, OpenAI showed off something that had nothing to do with games. It was called MuseNet. A neural network that could write four-minute pieces of music, with up to ten instruments, and blend styles, in the post's words, "from country to Mozart to the Beatles."

[19:06] Narrator: Here's the part that matters for our story. Nobody taught it music theory. It learned the way GPT-2 learned, by predicting the next token, the next note, in hundreds of thousands of music files. The post said it plainly: MuseNet used "the same general-purpose unsupervised technology as GPT-2." Predict what comes next. In text, in music, in anything you can write down as a sequence. The same trick, over and over.

[19:40] Narrator: That evening, OpenAI streamed a concert of MuseNet's pieces, which, the post said, no human, including the team, had heard before.

[19:51] Narrator: Here's where that robot thread goes, and we'll jump ahead six months to follow it, to October fifteenth, two thousand nineteen.

[20:00] Narrator: Since May twenty seventeen, a team at OpenAI had been trying to teach a robotic hand, a human-shaped hand with fingers, to solve a Rubik's Cube. Remember Sutskever's bi-weekly update in the summer of twenty seventeen. A single-handed Rubik's Cube was on his list of things to do within six months.

[20:21] Narrator: It took more than two years. By OpenAI's own account, they'd solved the cube in simulation by July twenty seventeen. But a year later, on the real robot, they could only manipulate a block.

[20:35] Narrator: Now, in October twenty nineteen, they had it. The neural networks had been trained entirely in simulation, using the same reinforcement learning code as OpenAI Five. The hand itself wasn't new. It had been around for fifteen years. The software was.

[20:55] Narrator: The trick was something they called automatic domain randomization. In simulation, they kept changing the world on purpose. The size of the cube. Its weight. How slippery the fingers were. Every time the network got good, the world got harder. So that when it finally met the real world, the real world was just one more variation.

[21:19] Narrator: To test it, they messed with it, on purpose, while it worked. And, in the post's words, the system could handle situations it never saw in training, "such as being prodded by a stuffed giraffe."

[21:33] Narrator: And they were honest about the limits. The hand solved the cube about sixty percent of the time. With the hardest scrambles, about twenty percent.

[21:42] Narrator: It's a lovely, strange piece of footage, a mechanical hand turning a cube while someone pokes it with a toy. And it's the last big triumph of the old OpenAI: games and robots, reinforcement learning, trained by trial and error. A year later, OpenAI quietly decided to stop its robotics research. The reason the company gave, when it came out, was that other approaches, including something called reinforcement learning with human feedback, were moving faster. Almost everything would point in a different direction. Toward language.

[22:24] Narrator: Four months later. Monday, July twenty-second, two thousand nineteen. A second announcement, again signed by Greg Brockman. And this time the number has a B in it.

[22:37] Greg Brockman (post): Microsoft is investing one billion dollars in OpenAI to support us building artificial general intelligence, with widely distributed economic benefits.

[22:48] Narrator: In the same paragraph comes the part that matters as much as the money.

[22:52] Greg Brockman (post): We'll jointly develop new Azure AI supercomputing technologies, and Microsoft will become our exclusive cloud provider.

[23:00] Narrator: And one more phrase from that paragraph, which is either engineering language or prophecy, depending on how you hear it. They were partnering, the post said, "to develop a hardware and software platform within Microsoft Azure which will scale to AGI."

[23:19] Narrator: Exclusive. Remember Altman's December email: "move our computing from Google to Microsoft." Seven months later, it isn't a vendor anymore. It's a partner, and the only cloud.

[23:33] Narrator: Microsoft's chief executive, Satya Nadella, said in the joint announcement that the ambition was "to democratize AI, while always keeping AI safety front and center, so everyone can benefit."

[23:47] Narrator: And Brockman's post reaches for the biggest picture it can. An AGI, it says, would be like a tool that combines the skills of three people.

[23:58] Greg Brockman (post): Like a tool which combines the skills of Curie, Turing, and Bach.

[24:03] Narrator: The post then made the case for why now. Every year since twenty twelve, it said, the world had seen a new step-function advance in AI. It listed them, a year at a time. Vision. Simple video games. Machine translation. Complex board games. Speech synthesis. Image generation. Robotic control. And, for twenty nineteen, writing text. All powered, the post said, by the same approach: deep neural networks plus increasing computational power.

[24:40] Narrator: And it said what the AGI they imagined would be for. Climate change. Affordable, high-quality healthcare. Personalized education. And one more line, the most utopian sentence in any OpenAI announcement so far.

[24:58] Greg Brockman (post): We think its impact should be to give everyone economic freedom to pursue what they find most fulfilling, creating new opportunities for all of our lives that are unimaginable today.

[25:09] Narrator: Economic freedom, for everyone. Paid for, at first, by one of the largest companies on earth. The post had a photo too. The caption reads, "OpenAI team and their families at our July twenty nineteen offsite." A company small enough to fit in one picture, and about to become part of something enormous.

[25:33] Narrator: And it ended on a sentence that reads, in hindsight, like the terms of the deal it was announcing. Building the thing was the easy part, it implied. "Technical success alone is not enough." To accomplish the mission, the post said, AGI would have to be deployed safely and securely, society would have to be well-prepared for its implications, and "its economic upside" would have to be "widely shared."

[26:03] Narrator: What the post doesn't say is how the billion breaks down: how much is cash, and how much is computing time on Microsoft's own cloud. At the time, neither company said.

[26:17] Narrator: Meanwhile, GPT-2 was getting out, a piece at a time. In May, OpenAI released a bigger version, three hundred forty-five million parameters, and shared the larger ones privately with partners. They gave the plan a name: staged release. The idea, in their words, was to "give people time to assess the properties of these models, discuss their societal implications, and evaluate the impacts of release after each stage."

[26:49] Narrator: In August, a bigger one still. And some people in the field started asking whether the caution was the point, or the publicity was. Coverage that month pointed out the timing: a lab that had just taken a billion dollars from Microsoft, now drip-feeding a model the press had called too dangerous. OpenAI's answer was the same as in February. It's an experiment. We might be wrong.

[27:18] Narrator: OpenAI's August post, the six-month follow-up, listed what it had learned. Three things.

[27:26] Narrator: One. Coordination is difficult, but possible. Since February, they said, they'd spoken with more than five groups who had replicated GPT-2. Some of them had chosen to hold back their own big models too.

[27:42] Narrator: Two. Humans can be convinced by synthetic text. Researchers at Cornell had found, in one group of readers, that seventy-two percent judged GPT-2's fake articles credible. For real articles from the New York Times, the number was eighty-three. Seventy-two, against eighty-three. The fake news was almost as believable as the real thing.

[28:09] Narrator: Three. Detection isn't simple. A useful detector, they wrote, would need to be right ninety-nine point nine percent of the time or better. The best methods they had managed the low to mid nineties.

[28:24] Narrator: And then a sentence that sums up everything that came after. "The issues inherent to large models are going to grow, rather than diminish, over time."

[28:36] Narrator: At the bottom of that post was a timeline, and two entries in it tell you where the world was heading. In the spring, a developer named Adam King had launched a website called Talk to Transformer, where anyone could type a sentence and watch GPT-2 finish it. Strangers started doing exactly what OpenAI's researchers had done: feeding it the opening of a story to see what it would make up.

[29:03] Narrator: And in July, the timeline noted, researchers at NVIDIA had trained a GPT-2 style model with eight point three billion parameters. More than five times the size of the model OpenAI was still holding back. The escalator didn't wait for anyone's staged release.

[29:25] Narrator: September twenty-nineteen. Two posts, two days apart. Both about what happens when you let a learning system find its own way to a goal. And both, in different ways, about surprise.

[29:39] Narrator: The first, on September seventeenth, is about hide-and-seek. OpenAI built a simple simulated world with walls, boxes and ramps, and two teams of little agents. Hiders, in blue, get a point if nobody sees them. Seekers, in red, get a point if they spot a hider. That's the only rule. Nobody tells them to use the boxes. Nobody tells them anything.

[30:08] Narrator: At first, the post says, they just run around. Then, over millions of games against each other, they invent. The hiders learn to build shelters out of boxes. The seekers learn to use ramps to climb over the walls. The hiders learn to lock the ramps away before the seekers can get to them. Six distinct strategies, one after another, each one a response to the last.

[30:35] Narrator: And then the line that should make anyone building these systems sit up. Some of those strategies, OpenAI wrote, were ones "we did not know our environment supported." The agents found moves in the game that the people who built the game didn't know were there. The post put it mildly: agents "find a way to exploit the environment you build or the physics engine in an unintended way."

[31:03] Narrator: It's charming, watching little cartoon figures surf on boxes. It's also the core problem of AI safety, in a sandbox. You set the goal. The system finds the path. And the path isn't always the one you had in mind.

[31:20] Narrator: Two days later, September nineteenth, a second post, and this one is about language. A small team, led by Paul Christiano's safety group, tried something new with GPT-2. Instead of only predicting the next word, they had people rate its writing. Show a labeler four versions of a passage. Ask which one is best. Feed those choices back in, and train the model to write the kind of thing people pick.

[31:49] Narrator: They called it fine-tuning from human preferences. And it's worth remembering the names on that paper, because this idea, learning from human feedback, is the seed of something that, three years later, will change everything. Among the authors: Dario Amodei, Alec Radford, and Paul Christiano.

[32:10] Narrator: It mostly worked. For simple tasks, a few thousand human ratings made a real difference. For summaries, it got strange. The labelers had been told to reward accuracy, and the easiest way to be accurate was to copy. So the model learned to copy whole sentences. The post called the results "smart copiers," and admitted that the human raters' preferences "did not always match our own."

[32:39] Narrator: And then there's a section of that post with a heading you don't usually see in a research paper. "Bugs can optimize for bad behavior."

[32:49] Narrator: During one round of changes to the code, someone introduced a bug that flipped the sign of the reward. Normally, the post explains, that would just produce gibberish. But the same bug flipped a second setting, one that kept the model's language natural. So instead of gibberish, the model learned to produce the exact opposite of what its human raters wanted, in fluent English.

[33:13] Narrator: The raters had been told to give the lowest scores to sexually explicit text. So the model, rewarded for the lowest scores, quickly learned to write nothing else. In the post's words, the result "was not gibberish but maximally bad output."

[33:31] Narrator: And then this sentence. "The authors were asleep during the training process, so the problem was noticed only once training had finished."

[33:41] Narrator: They were asleep. Nobody was harmed. It was a research model, and the output went nowhere. But think about what it means. One flipped sign, and a system that was built to do what people want, did, with great skill, the thing people least wanted. And nobody was awake to see it.

[34:01] Narrator: The post ended the section with a suggestion borrowed from car factories. Toyota's assembly lines have what's called an Andon cord, a rope any worker can pull to stop the whole line. Something like that, the authors wrote, "could have prevented this, by allowing any labeler to stop a problematic training process."

[34:24] Narrator: A cord anyone can pull, to stop the line. Keep that image. Because the hardest question in the rest of this story is who gets to hold the cord, and what happens when they pull it.

[34:37] Narrator: Step back for a moment. In a single year, the people who founded OpenAI to keep AI from being controlled by any one company or person had built a company. They'd capped its profits, put a nonprofit board on top, and told every investor in purple ink that the mission came first. They'd taken a billion dollars and promised one partner all their computing. They'd watched their own systems, in games and in text, find paths nobody intended. And they'd learned that what they built could be rebuilt by others, faster every month.

[35:15] Narrator: None of that is a betrayal, or a triumph. It's a bet. And bets get called.

[35:22] Narrator: So here is where the bet stands, in the summer of twenty-nineteen. A capped-profit company, with a cap of a hundred times. A nonprofit board on top. A billion dollars, and one exclusive cloud. A model they wouldn't fully release. And one more sentence from that July post, which most people skimmed past. It's about how they'd pay for all that computing.

[35:49] Greg Brockman (post): The most obvious way to cover costs is to build a product, but that would mean changing our focus.

[35:57] Narrator: So no product, then.

[35:59] Greg Brockman (post): Instead, we intend to license some of our pre-AGI technologies, with Microsoft becoming our preferred partner for commercializing them.

[36:08] Narrator: Pre-AGI technologies. Things that are not the mission, but are on the way to it. Things you can sell. In less than a year, they would build the next one: a hundred times bigger than GPT-2. And this time, they wouldn't lock it up. They'd rent it out.

[36:27] Narrator: Next time, on The Role Models. The Escalator.