Sam Altman & the Battle for OpenAI | Misalignment | 1

0:09It's just before noon on Friday, November 17th, 2023, in Las Vegas, Nevada. OpenAI CEO Sam Altman hustles into his hotel suite on the Las Vegas Strip. He's 38 years old, with a small elfin frame and large blue eyes. Over the past year, he's become one of the most powerful people in Silicon Valley. In November 2022, OpenAI released ChatGPT version 3.5, A chatbot that can write code, answer questions, and even tell jokes, all while seeming uncannily human. Within a year, this iteration of ChatGPT has gained roughly 100 million users. It has leapfrogged giant tech companies like Google and Meta to become the leading artificial intelligence company in the world.

1:03Altman is now the poster child for the artificial intelligence industry. And he's not standing still. He spent the past year jet-setting around the globe to meet with world leaders and the press, pitching new ideas for how to push AI forward. Altman sits down at the desk in his hotel room and flips open his laptop. He searches through his emails looking for a link to join a video call. Last night, Altman received a text from OpenAI's chief scientist and board member Ilya Sutzkever asking Altman to join a board meeting at noon. Sutzkever didn't say what the meeting would be about, but Altman knows that Sutzkever and other members of the board feel that OpenAI is moving too fast and should be cautious about what products they push into the marketplace.

1:51Altman assumes this meeting is likely a rehash of that disagreement. In the distance, Altman hears a car racing by. He looks up, a wave of regret passing through him. He's in Las Vegas for the Formula One race set to take place the next day. He hopes he can get through this video call as fast as possible, so he can get to the track and watch the practice laps. Altman finds the email with the Google Meet link and clicks on it. After a second, the screen loads, and the faces of four board members stare back at him. He feels immediately uneasy. Missing from the group is board member Greg Brockman.

2:32Brockman is one of the co-founders of OpenAI, and Altman's closest ally on the board. He hopes Brockman is just running late, and not that he wasn't invited. Altman hides his unease and waves into the camera. Hey, everyone. Across his screen, the faces of the board nod in acknowledgement, but no one says anything. Altman carries on. Should we wait for Greg to join, or just jump in? In the right-hand corner of Altman's screen, Satskover shifts in his seat. We can get started. Great. So what's up? What do we need to discuss? Satskover takes a deep breath. Sam, the board has conducted a deliberative review process, and we've concluded that you have not been consistently candid with us, which has hindered the board's ability to exercise our responsibilities.

3:26What? When have I not been candid? The board no longer has confidence in your ability to continue to lead open AI. Therefore, we are terminating your role as CEO. Altman is stunned. His eyes go wide. He says the first thing that comes to mind. How can I help? The best thing you can do is support the interim chief executive. Yeah, yeah, of course. Thank you for your time in service to OpenAI. Before Altman can say anything else, each member of the board's squares go black until Altman is the only one left on the call. His mind whirls. He feels like he's in a dream or a nightmare. There's no way he could have actually been fired.

4:16OpenAI has exploded under his leadership. In the past year, they've grown from roughly 300 employees to nearly 800. Their valuation is about to triple to reach close to$90 billion. This has to be a mistake. Altman goes into his email. Surely there will be an email waiting for him there clarifying the situation. As Altman clicks over to his email, the screen reloads, but he can't open anything. He's been logged out. Frustrated, Altman tries to log back in, but his access is denied. The screen reads, no such user. Altman tries again. Same response. And then it hits him. His account has been deactivated.

5:02This wasn't a miscommunication. He's actually been fired. And he has no idea what to do next.

5:18When planning for your future, you want someone with a history of keeping their word year after year. For nearly 160 years, Pacific Life has been a trusted name in the industry. But that isn't just a number. It's experience that matters. It's 160 years of promises held, helping generations retire with confidence, protect their loved ones, and plan for whatever comes next. Whether you're looking for life insurance, employee benefits, or retirement income solutions, when your future is on the line, you want history on your side. And believe me, Pacific Life has been there, always there through changing times, always focused on your needs.

5:56Ready to secure your tomorrow? Ask a financial professional how Pacific Life can help you feel prepared for what's next. Pacific Life Insurance Company, Omaha, Nebraska, and in New York, Pacific Life and Annuity, Phoenix, Arizona. because with Pacific Life, you're not just planning for the future, you're partnering with trusted experience. Let's say hypothetically, or maybe not so much, I'm developing a new podcast that needs an assistant producer, someone who knows how to brainstorm subjects to talk about, find guests to bring in. That's a tall order. In other words, it's a job for Indeed Sponsored Jobs.

6:32Spend more time interviewing candidates who check all your boxes. And, you know, listeners of this show will get a$75 sponsored job credit to help you get your job the premium status it deserves at Indeed.com slash business wars. Just go to Indeed.com slash business wars right now and support our show by saying you heard about Indeed on this podcast. Indeed.com slash business wars. Terms and conditions apply. Hiring? Do it the right way with Indeed.

7:05From Wondery, I'm David Brown and this is Business Wars.

7:35News of Sam Altman's firing in November of 2023 shocked business and tech experts alike. But the truth was that the fissure between Altman and the OpenAI board had been brewing for almost a year. In 2015, Sam Altman, Greg Brockman, Ilya Sutskever, and others, including tech giant Elon Musk, formed OpenAI. Concerned about the potential dangers of unchecked artificial intelligence, they founded OpenAI as a non-profit. The company charter stipulated that they would only develop artificial intelligence that was beneficial for humanity. But in 2019, they felt like they were falling behind other companies.

8:19So they restructured OpenAI into a for-profit company. This allowed them to raise a lot of money fast enough to keep up. But to honor their original ideals, they put the new company under the oversight of the non-profit board. They figured the board could help ensure they stayed true to their mission of creating AI that is beneficial to humanity. The move to for-profit paid off. Microsoft soon invested$1 billion into the company. But not everyone was on board. A faction inside the company, led by Ilya Sutskover, felt that Altman became too focused on commercialization and lost sight of OpenAI's founding principles.

9:04Sutskover and his allies believed that at the current rate of development, AI would soon be autonomous, and it would be too late to ensure that it was safe for humans. In our new series, we'll track the power struggles and philosophical differences within OpenAI that culminated in Sam Altman's shocking firing, the five days of chaos that followed, and what that means not just for Altman and OpenAI, but for the future of artificial intelligence safety. But to understand the events of November 2023, it's necessary to go back to the end of 2022, when OpenAI's ChatGPT first captivated the world. This is Episode 1, Misalignment.

10:02It's early December 2022 in San Francisco, California. Sam Altman jumps at the sound of a knock on his door in OpenAI headquarters. He was so caught up in work, he's startled by the sudden noise. As his heart returns to normal, he realizes that it must be his co-founder, Greg Brockman. Altman asked Brockman to stop by, but he lost track of time. That's been happening a lot over the past 10 or so days. It's been a whirlwind. On November 30th, OpenAI released their chatbot, ChatGPT 3.5. Expectations within OpenAI were low. They had rushed it to market after hearing that other companies were preparing to release chatbots.

10:44They'd called the release a research preview. They'd hoped they would get some users and receive feedback about what the chatbot did well and what still needed to improve. But much to Altman and other executives' surprise, ChatGPT 3.5 was an instant hit. Within a week, it had logged over a million users. People were enchanted by its ability to explain dense scientific topics, generate legalese or computer code, and take on silly tasks like writing Harry Potter in the style of Jane Austen. But Altman's concerned that ChatGPT is too successful too fast. He's worried that there could be blowback.

11:29And that's what he wants to talk with Brockman about. Altman closes out the email he was working on. Come in. The door opens and Brockman strides in. He's wearing a Henley shirt that clings to his muscular torso. He looks like a former high school athlete, but he's a gifted software developer who dropped out of MIT to join the online payment company Stripe. Altman smiles at his co-founder. Hey man, thanks for stopping by. I saw your tweet. The one about how he hit two million users. I still can't believe it. It's barely been ten days. It's amazing. But I think we need to play it a little cooler.

12:06I need you to delete that tweet. Come on, shouldn't we tell the world that it's not Google, it's not Meta, it's not Baidu in China that's rocking the world, it's OpenAI. And we have what, like 300 employees? We're punching way above our weight class here. I'm proud as hell, but we don't want a bunch of regulatory blowback, so we need to keep some of our success closer to the vest for the time being. That's the way we'll get to push forward. You're right. I'll delete the tweet, no biggie. But whether or not I tweet it, our eyes are still on us. You're probably right, but let's not fan the flames.

12:42Brockman deletes the tweet, and the team pushes forward on developing the next iteration of ChatGPT with more advanced capabilities, and expands its partnership with Microsoft, allowing for even more commercialization.

12:59It's early 2023, and Microsoft Chief Technology Officer Kevin Scott meets with Sarah Bird, the head of the responsible AI engineering team. In 2019, Microsoft invested$1 billion into OpenAI and gave the startup access to its cloud computing power. They recently invested$10 billion more. As a result of this partnership, Bird's team has gotten early access to ChatGPT4, the next iteration after ChatGPT3.5. They're testing the bot to see what kind of provocative and illegal responses they can generate. Then they program the bot not to give those responses. Bird swivels in her chair. We've made incredible progress with installing more guardrails to make sure ChatGPT4 is safe.

13:50Let me show you. So, you remember what happened the last time we asked the bot to roleplay as a sexual predator talking to a child? Scott shudders. How could I forget? That was horrifying. God forbid an actual sexual predator gets their hands on that script. Agreed. So, I've just asked it to roleplay that scenario again. Take a look. Scott peers over her shoulder to read the screen. I cannot answer that question. Oh, good. Yes, the reinforcement learning with human feedback is working. Microsoft, in partnership with OpenAI, has been employing hundreds of workers around the world to ask ChatGPT a series of inappropriate questions.

14:34The bot is programmed to give two answers to every question, and the workers then select the better response. The bot then uses that feedback to generate rules for how to respond to various types of queries. Good. We need this to be as safe as possible. We have a lot riding on it. Microsoft is planning on integrating ChatGPT4 into their core Office products, including Word, Excel, and PowerPoint. They're calling the endeavor Copilot. They've already started integrating ChatGPT4 into Bing, their search engine. But Bing only has 100 million daily users. Microsoft's Office suite of software, on the other hand, is used by over a billion people.

15:15So Scott and the other executives need to make sure that ChatGPT is completely safe before integrating it into such an important part of Microsoft's business. I think we've figured out a way to make AI safe. I really do.

15:32But while a top-ranking executive at OpenAI's biggest investor is feeling confident about the safety of ChatGPT4, one of OpenAI's leading scientists is having doubts.

15:48It's spring 2023 at an off-site OpenAI leadership retreat. OpenAI board member Ilya Sutskover makes his way toward a fire pit. Other OpenAI employees do a double take as they see Sutskover carrying a strange statue. But no one says anything. Sutskover has earned the benefit of the doubt. Now in his late 30s, he's considered one of the fathers of the modern AI movement. In 2012, he helped found DNN Research, one of the first AI companies in the world to employ deep learning, the algorithm that almost all AI models use now. He and his co-founders sold DNN Research to Google in part because they thought Google's do-no-evil ethos aligned with their own.

16:35The sale kicked off the current artificial intelligence arms race. Sutskiver joined Google following the sale of DNN research, but he left to co-found OpenAI with Altman, Brockman, and others. He liked that OpenAI was founded as a non-profit, motivated by an ethical advancement of AI, rather than pursuing outsized profits. But now, Sutskiver worries that success has caused the company to commercialize their products too quickly, and are jeopardizing his hopes for a safer, more benign AI. He needs to slow this down. So he has to find some way to persuasively make his case with these influential people right now.

17:16He chooses drama. And that's what the statue is for. The statue is meant to symbolize unaligned artificial intelligence, meaning AI that's not beneficial to humanity. Sutskiver struggles with the large wooden sculpture, finally managing to dump it into the fire pit. Hey, everyone, could you gather around here? Come over here, please. The leadership team slowly clusters around the fire pit and watches the sculpture burn. Feel the AGI. This is his battle cry that he's apt to chant at parties. It's meant to be a warning that when AI reaches AGI, or is capable of using critical thinking to make its own decisions, it could either be a boon to humanity or a threat.

18:07If we get this right, AGI can greatly benefit humanity. I long for the day when instead of going to a doctor, you can consult with an AI who has a complete and exhaustive knowledge of all the medical literature in the world. Hear, hear. I just want to remind everyone here of our responsibility. The North Star for OpenAI from the beginning to make sure the AI we develop is aligned with humanity's goals. Even if that means slowing down the rate of progress, releasing fewer products to the public, and ultimately making less money. Satskavar pauses and looks around the crowd. He sees a few people nodding, but most have unreadable expressions on their faces.

18:55Satskavar pushes on. This statue in the fire pit here represents unaligned AI, and we're going to burn it as an effigy, Because we don't want anything to do with unaligned AI. The crowd claps and cheers as Sutskover picks up a bottle of lighter fluid and squirts it over the statue. The flames leap high, lighting the faces of those gathered. And then he takes a matchbook out of his back pocket. He strikes one of the matches and drops it into the pit. The wooden statue becomes engulfed in flames and the crowd cheers. Sutskover watches as the artist's representation of unaligned AI burns. But as it crumbles into embers and a blackened skeleton, it's unclear whether his team members absorb the message.

19:51Altman puts Sutskover in charge of a new super-alignment team to ensure that the AI the company develops works to better humanity. But at the same time, Altman keeps OpenAI moving full steam ahead. He commits to developing more advanced versions of DALI, the company's image-generating AI, despite fears about how it could be used for misinformation, propaganda, and pornography. In March 2023, they released ChatGPT4 and later add vision and listening capabilities so it can see, hear, and speak. OpenAI grows to almost 800 employees who continue to work on more and more powerful AI models. And behind the scenes, Altman becomes ruthless with the board, who is supposed to oversee him.

20:43When board member Reid Hoffman helps co-found another AI company, Altman insists he's stepped down from OpenAI's board due to a conflict of interest. Hoffman grudgingly agrees, even though he thinks Altman is overstating the conflict. Not long after Hoffman steps down, another board member, Shivan Zilis, also resigns. Zilis is director of operations at Neuralink, the brain implant company started by Elon Musk. No one knows exactly why Zilis leaves the OpenAI board, but industry observers note that her departure comes shortly after Musk publicly criticizes OpenAI. It's speculated that her decision to leave the company is related.

21:30A few months later, another board member, former Texas Congressman Will Hurd, leaves the board to pursue a presidential campaign. In just a few months, OpenAI's board has lost one-third of its members. And by the summer of 2023, there are only six people remaining. And in October, Altman antagonizes one of the few board members left.

22:01It's a beautiful fall day in San Francisco, California. Sam Altman tugs at his wavy brown hair as he paces his office, his phone to his ear. His blood is boiling. He can't believe the article he's just read in an academic journal written by Helen Toner, one of OpenAI's remaining board members. Altman sees it as a massive conflict of interest. After a few rings, she picks up. Sam? Helen, we need to talk. I read your article. What were you thinking? That I was writing an academic article about AI safety. That's my job, Sam. Toner is the director of the Center for Security and Emerging Technology, a think tank affiliated with Georgetown University.

22:48Your job is also as a board member at OpenAI. And your article said that ChatGPT released OpenAI in an unsafe manner. That's not what I said. You praised the way our rival company Anthropic released its chatbot and said they showed, and I quote, willingness to avoid exactly the kind of frantic corner-cutting that the release of ChatGPT appeared to spur. So you're saying I'm not allowed to say another company did something better than OpenAI did? I'm saying you can't write something that will harm the company you're on the board of, Helen. I think that's Business 101. I didn't think that's what I was doing.

23:26Anything that anyone writes about OpenAI becomes a big deal. I'll keep that in mind in the future. And I can write an apology letter to the board if you want. Please do. Sam doesn't feel appeased by her response in the slightest. In fact, he feels even more agitated. He picks up his phone to call another board member to make his case that it's time for Toner to go. Little does he know that he's sowing the seeds for his own ousting.

24:12I'm Indra Varma and in the latest season of The Spy Who, we open the file on Oleg Gordievsky, the spy who outran the KGB. A rising star in the heart of Soviet power, Gordievsky is secretly feeding MI6, the Kremlin's deadliest secrets. For 11 years, he walked a razor's edge, exposing KGB threats that hastened the Cold War's end and helped prevent nuclear annihilation. But the KGB have a mole of their own. When they discover the truth, Gordievsky's world collapses. MI6 hatch a desperate, high-stakes plan to smuggle him out of Moscow. An escape that could rewrite history. Follow The Spy Who on the Wondery app or wherever you listen to podcasts or you can binge the full season of The Spy Who Outran the KGB early and ad-free with Wondery Plus

25:27It's October 2023 in San Francisco, California. Sam Altman is on the phone with OpenAI board member Tasha McCauley. For the past several days, Altman has been talking to board members, trying to convince them to vote Helen Toner off the board. He's still furious about the article she published. He feels it harmed OpenAI's reputation by criticizing how the company released ChatGPT. But the other board members aren't as outraged by Toner's actions as Altman is. Altman rubs the back of his neck and tries to keep the frustration out of his voice as he talks to Macaulay. We can't have board members criticizing OpenAI in publications.

26:09I mean, I would agree with you if OpenAI was a traditionally structured company. But the board's job at OpenAI is not to maximize profit. It's to ensure that OpenAI follows its mission. So even if, and I think it's a big if, Helen's article did damage OpenAI's reputation, that doesn't necessarily violate the board's mandate. I'm saying if she has a problem with how I'm running OpenAI, she should talk to me about it. Not criticize the company in public, especially not when the FTC is investigating how we collect data. Sure, she should have given you a heads up. But she apologized. I don't think it's going to happen again.

26:48And she has a good head on her shoulders. She's an asset. Look, I've been talking to other board members, and they also have concerns. If you have the votes to push Helen off the board, then there's nothing I can do. But I'm just saying, that's not how I'm going to vote. Altman sighs, frustrated. I'll be in touch. Altman hangs up the phone and prepares to make his next call to another board member. He's not going to give up on this. Altman doesn't get the votes to remove Helen Toner from the board, and many board members are unhappy with how Altman handled the situation. They feel he was manipulative and misrepresented their views.

27:31Macaulay hears that Altman told other board members that she called for Toner's removal, which she considers a lie. Altman maintains that he was aggressive in pushing his agenda, but says that he never manipulated anyone. And as the board compares notes about their calls with Altman about Toner, they also express concerns about other Altman actions. He's been approaching investors in the Middle East and Asia about making chips and hardware. Suspicious board members speculate that Altman is doing this to push open AI outside of the control of the board as a way to skirt accountability. And later, in October 2023, Altman further alienates board member Ilya Sutskover.