Bloomberg Businessweek (February 2025)

Год выпуска: February, 2025

Автор: Bloomberg Businessweek

Жанр: Бизнес

Издательство: «Bloomberg Businessweek»

Формат: PDF (журнал на английском языке)

Качество: OCR

Количество страниц: 96

OpenAI’s Chief Opens Up

An interview with Sam Altman on the rise of ChatGPT, Elon Musk and AI under Trump

On Nov. 30, 2022, traffic to OpenAI’s website peaked at a number a little north of zero. It was a startup so small and sleepy that the owners didn’t bother tracking their web traffic. It was a quiet day, the last the company would ever know. Within two months, OpenAI was being pounded by more than 100 million visitors trying, and freaking out about, ChatGPT.

Nothing has been the same for anyone since, particularly Sam Altman. In his most wide-ranging interview as chief executive officer, Altman reflects on his journey from venture capitalist to world-famous tech boss. He also explains his infamous four-day firing in November 2023, how he actually runs OpenAI, his plans for the Trump-Musk presidency and his relentless pursuit of artificial general intelligence—the still-theoretical next phase of AI, in which machines will be capable of performing any intellectual task a human can do.

The official release date of ChatGPT was Nov. 30, 2022. Does that feel like a million years ago or a week ago?

I turn 40 [in April]. On my 30th birthday, I wrote this blog post, and the title of it was “The days are long but the decades are short.” Somebody this morning emailed me and said, “This is my favorite blog post, I read it every year. When you turn 40, will you write an update?” I’m laughing, because I’m definitely not going to write an update. I have no time. But if I did, the title would be “The days are long, and the decades are also f---ing very long.” So it has felt like a very long time.

Are you a person who enjoys success? Were you able to take it in, or were you already worried about the next phase of scaling?

A very strange thing about me, or my career: The normal arc is you run a big, successful company, and then in your 50s or 60s you get tired of working that hard, and you become a VC. It’s very unusual to have been a VC first and have had a pretty long VC career and then run a company. And there are all these ways in which I think it’s bad, but one way in which it has been very good for me is you have the weird benefit of knowing what’s going to happen to you, because you’ve watched and advised a bunch of other people through it. And I knew I was both overwhelmed with gratitude and, like, “F---, I’m going to get strapped to a rocket ship, and my life is going to be totally different and not that fun.” I had a lot of gallows humor about it. My husband 1 tells funny stories from that period of how I would come home, and he’d be like, “This is so great!” And I was like, “This is just really bad. It’s bad for you, too. You just don’t realize it yet, but it’s really bad.” [Laughs]

You’ve been Silicon Valley famous for a long time, but one consequence of GPT’s arrival is that you became world famous with the kind of speed that’s usually associated with, like, Sabrina Carpenter or Timothee Chalamet. Did that complicate your ability to manage a workforce?

It complicated my ability to live my life. But in the company, you can be a well-known CEO or not, people are just like, “Where’s my f---ing GPUs?”
I feel that distance in all the rest of my life, and it’s a really strange thing. I feel that when I’m with old friends, new friends—anyone but the people very closest to me. I guess I do feel it at work if I’m with people I don’t normally interact with. But I spend most of my time with the researchers, and man, I promise you, come with me to the research meeting right after this, and you will see nothing but disrespect. Which is great.

When the news emerged that the board had fired you as CEO, it was shocking. But you seem like a person with a strong EQ. Did you detect any signs of tension before that? And did you know that you were the tension?

I don’t think I’m a person with a strong EQ at all, but even for me this was over the line of where I could detect that there was tension. You know, we kind of had this ongoing thing about safety versus capability and the role of a board and how to balance all this stuff. So I knew there was tension, and I’m not a high-EQ person, so there’s probably even more.

A lot of annoying things happened that first weekend. My memory of the time—and I may get the details wrong—so they fired me at noon on a Friday. A bunch of other people quit Friday night. By late Friday night I was like, “We’re just going to go start a new AGI effort.” Later Friday night, some of the executive team was like, “Um, we think we might get this undone. Chill out, just wait.”

Saturday morning, two of the board members called and wanted to talk about me coming back. I was initially just supermad and said no. And then I was like, “OK, fine.” I really care about [OpenAI]. But I was like, “Only if the whole board quits.” I wish I had taken a different tack than that, but at the time it felt like a just thing to ask for. Then we really disagreed over the board for a while. We were trying to negotiate a new board. They had some ideas I thought were ridiculous. I had some ideas they thought were ridiculous. But I thought we were [generally] agreeing. And then—when I got the most mad in the whole period—it went on all day Sunday. Saturday into Sunday they kept saying, “It’s almost done. We’re just waiting for legal advice, but board consents are being drafted.” I kept saying, “I’m keeping the company together. You have all the power. Are you sure you’re telling me the truth here?” “Yeah, you’re coming back. You’re coming back.”

And then Sunday night they shock-announce that Emmett Shear was the new CEO. And I was like, “All right, now I’m f---ing really done,” because that was real deception. Monday morning rolls around, all these people threaten to quit, and then they’re like, “OK, we need to reverse course here.”

The board says there was an internal investigation that concluded you weren’t “consistently candid” in your communications with them. That’s a statement that’s specific—they think you were lying or withholding some infor-mation—but also vague, because it doesn’t say what specifically you weren’t being candid about. Do you now know what they were referring to?

I’ve heard different versions. There was this whole thing of, like, “Sam didn’t even tell the board that he was going to launch ChatGPT.” And I have a different memory and interpretation of that. But what is true is I definitely was not like, “We’re going to launch this thing that is going to be a huge deal.” And I think there’s been an unfair characterization of a number of things like that. The one thing I’m more aware of is, I had had issues with various board members on what I viewed as conflicts or otherwise problematic behavior, and they were not happy with the way that I tried to get them off the board. Lesson learned on that.

Can I offer a theory?

Sure.

You recognized at some point that the structure of [OpenAI] was going to smother the company, that it might kill it in the crib. Because a mission-driven nonprofit could never compete for the computing power or make the rapid pivots necessary for OpenAI to thrive. The board was made up of originalists who put purity over survival. So you started making decisions to set up OpenAI to compete, which required being a little sneaky, which the board?

I don’t think I was doing things that were sneaky. I think the most I would say is, in the spirit of moving really fast, the board did not understand the full picture. There was something that came up about “Sam owning the startup fund, and he didn’t tell us about this.” And what happened there is because we have this complicated structure: 2 OpenAI itself could not own it, nor could someone who owned equity in OpenAI. And I happened to be the person who didn’t own equity in OpenAI. So I was temporarily the owner or GP 3 of it until we got a structure set up to transfer it. I have a different opinion about whether the board should have known about that or not. But should there be extra clarity to communicate things like that, where there’s even the appearance of doing stuff? Yeah, I’ll take that feedback. But that’s not sneaky. It’s a crazy year, right? It’s a company that’s moving a million miles an hour in a lot of different ways. I would encourage you to talk to any current board member and ask if they feel like I’ve ever done anything sneaky, because I make it a point not to do that.

When you got back to the company, were you self-conscious about big decisions or announcements, because you worried about how your character may be perceived? Actually, let me put that more simply. Did you feel like some people may think you were bad, and you needed to convince them that you’re good?

It was worse than that. Once everything was cleared up, it was all fine, but in the first few days no one knew anything. And so I’d be walking down the hall, and [people] would avert their eyes. It was like I had a terminal cancer diagnosis. There was sympathy, empathy, but [no one] was sure what to say. That was really tough. But I was like, “We got a complicated job to do. I’m going to keep doing this.”

Can you describe how you actually run the company? How do you spend your days? Like, do you talk to individual engineers? Do you get walking-around time?

Let me just call up my calendar. So we do a three-hour executive team meeting on Mondays, and then, OK, yesterday and today, six one-on-ones with engineers. I’m going to the research meeting right after this. Tomorrow is a day where there’s a couple of big partnership meetings and a lot of compute meetings 1 There’s five meetings on building up compute. I have three product brainstorm meetings tomorrow, and I’ve got a big dinner with a major hardware partner after. That’s kind of what it looks like. A few things that are weekly rhythms, and then it’s mostly whatever comes up.

How much time do you spend communicating, internally and externally?

Way more internal. I’m not a big inspirational email writer, but lots of one-on-one, small-group meetings and then a lot of stuff over Slack.

Oh, man. God bless you. You get into the muck?

I’m a big Slack user. You can get a lot of data in the muck. I mean, there’s nothing that’s as good as being in a meeting with a small research team for depth. But for breadth, man, you can get a lot that way.

You’ve previously discussed stepping in with a very strong point of view about how ChatGPT should look and what the user experience should be. Are there places where you feel your competency requires you to be more of a player than a coach?

At this scale? Not really. I had dinner with the Sora 2 team last night, and I had pages of written, fairly detailed suggestions of things. But that’s unusual. Or the meeting after this, I have a very specific pitch to the research team of what I think they should do over the next three months and quite a lot of granular detail, but that’s also unusual.

We’ve talked a little about how scientific research can sometimes be in conflict with a corporate structure. You’ve put research in a different building from the rest of the company, a couple of miles away. Is there some symbolic intent behind that?

Uh, no, that’s just logistical, space planning. We will get to a big campus all at once at some point. Research will still have its own area. Protecting the core of research is really critical to what we do.

Protecting it from what?

The normal way a Silicon Valley company goes is you start up as a product company. You get really good at that. You build up to this massive scale. And as you build up this massive scale, revenue growth naturally slows down as a percentage, usually. And at some point the CEO gets the idea that he or she is going to start a research lab to come up with a bunch of new ideas and drive further growth. And that has worked a couple of times in history. Famously for Bell Labs and Xerox PARC. Usually it doesn’t. Usually you get a very good product company and a very bad research lab. We’re very fortunate that the little product company we bolted on is the fastest-growing tech company maybe ever—certainly in a long time. But that could easily subsume the magic of research, and I do not intend to let that happen.

We are here to build AGI and superintelligence and all the things that come beyond that. There are many wonderful things that are going to happen to us along the way, any of which could very reasonably distract us from the grand prize. I think it’s really important not to get distracted.

As a company, you’ve sort of stopped publicly speaking about AGI. You started talking about AI and levels, and yet individually you talk about AGI.

“AGI” has become a very sloppy term. If you look at our levels, our five levels, you can find people that would call each of those AGI, right? And the hope of the levels is to have some more specific grounding on where we are and kind of like how progress is going, rather than is it AGI, or is it not AGI?

What’s the threshold where you’re going to say, “OK, we’ve achieved AGI now”?

The very rough way I try to think about it is when an AI system can do what very skilled humans in important jobs can do—I’d call that AGI. There’s then a bunch of follow-on questions like, well, is it the full job or only part of it? Can it start as a computer program and decide it wants to become a doctor? Can it do what the best people in the field can do or the 98th percentile? How autonomous is it? I don’t have deep, precise answers there yet, but if you could hire an AI as a remote employee to be a great software engineer, I think a lot of people would say, “OK, that’s AGI-ish.”

Now we’re going to move the goalposts, always, which is why this is hard, but I’ll stick with that as an answer. And then when I think about superintelligence, the key thing to me is, can this system rapidly increase the rate of scientific discovery that happens on planet Earth?

You now have more than 300 million users. What are you learning from their behavior that’s changed your understanding of ChatGPT?

Talking to people about what they use ChatGPT for, and what they don’t, has been very informative in our product planning. A thing that used to come up all the time is it was clear people were trying to use ChatGPT for search a lot, and that actually wasn’t something that we had in mind when we first launched it. And it was terrible for that. But that became very clearly an important thing to build. And honestly, since we’ve launched search in ChatGPT, I almost don’t use Google anymore. And I don’t think it would have been obvious to me that ChatGPT was going to replace my use of Google before we launched it, when we just had an internal prototype. Another thing we learned from users: how much people are relying on it for medical advice. Many people who work at OpenAI get really heartwarming emails when people are like, “I was sick for years, no doctor told me what I had. I finally put all my symptoms and test results into ChatGPT—it said I had this rare disease. I went to a doctor, and they gave me this thing, and I’m totally cured.” That’s an extreme example, but things like that happen a lot, and that has taught us that people want this and we should build more of it.

Your products have had a lot of prices, from $0 to $20 to $200—Bloomberg reported on the possibility of a $2,000 tier. How do you price technology that’s never existed before? Is it market research? A finger in the wind?

We launched ChatGPT for free, and then people started using it a lot, and we had to have some way to pay for it. I believe we tested two prices, $20 and $42. People thought $42 was a little too much. They were happy to pay $20. We picked $20. Probably it was late December of 2022 or early January. It was not a rigorous “hire someone and do a pricing study” thing.

There’s other directions that we think about. A lot of customers are telling us they want usagebased pricing. You know, “Some months I might need to spend $1,000 on compute, some months I want to spend very little.” I am old enough that I remember when we had dial-up internet, and AOL gave you 10 hours a month or five hours a month or whatever your package was. And I hated that. I hated being on the clock, so I don’t want that kind of a vibe. But there’s other ones I can imagine that still make sense, that are somehow usage-based.

On AI safety, has your sense evolved of what the dangers actually might be?

I still have roughly the same short-, medium-and long-term risk profiles. I still expect that on cybersecurity and bio stuff, 1 we’ll see serious, or potentially serious, short-term issues that need mitigation. Long term, as you think about a system that really just has incredible capability, there’s risks that are probably hard to precisely imagine and model. But I can simultaneously think that these risks are real and also believe that the only way to appropriately address them is to ship product and learn.

When it comes to the immediate future, the industry seems to have coalesced around three potential roadblocks to progress: scaling the models, chip scarcity and energy scarcity. I know they commingle, but can you rank those in terms of your concern?

We have a plan that I feel pretty good about on each category. Scaling the models, we continue to make technical progress, capability progress, safety progress, all together. I think 2025 will be an incredible year.

We have been hard at work on the whole [chip] supply chain, all the partners. We have people to build data centers and make chips for us. We have our own chip effort here. We have a wonderful partnership with Nvidia, just an absolutely incredible company. And we’ll talk more about this, but now is the time for us to scale chips.

So energy ...

Fusion’s going to work.

Fusion is going to work. Um. On what time frame?

Soon. Well, soon there will be a demonstration of net-gain fusion. You then have to build a system that doesn’t break. You have to scale it up. You have to figure out how to build a factory—build a lot of them—and you have to get regulatory approval.

And that will take, you know, years altogether? But I would expect [Helion]2 will show you that fusion works soon.

In the short term, is there any way to sustain AI’s growth without going backward on climate goals?

Yes, but none that is as good, in my opinion, as quickly permitting fusion reactors. I think our particular kind of fusion is such a beautiful approach that we should just race toward that and be done.

A lot of what you just said interacts with the government. We have a new president coming. You made a personal $1 million donation to the inaugural fund. Why?

He’s the president of the United States. I support any president.

I understand why it makes sense for OpenAI to be seen supporting a president who’s famous for keeping score of who’s supporting him, but this was a personal donation. Donald Trump opposes many of the things you’ve previously supported. Am I wrong to think the donation is less an act of patriotic conviction and more an act of fealty?

I don’t support everything that Trump does or says or thinks. I don’t support everything that Biden says or does or thinks. But I do support the United States of America, and I will work to the degree I’m able to with any president for the good of the country. And particularly for the good of what I think is this huge moment that has got to transcend any political issues. I think AGI will probably get developed during this president’s term, and getting that right seems really important. Supporting the inauguration, I think that’s a relatively small thing. I don’t view that as a big decision either way. But I do think we all should wish for the president’s success.

He’s said he hates the Chips Act. You supported the Chips Act.

I actually don’t. I think the Chips Act was better than doing nothing but not the thing that we should have done. And I think there’s a real opportunity to do something much better as a follow-on. I don’t think the Chips Act has been as effective as any of us hoped.

Elon is clearly going to be playing some role in this administration. He’s suing you. He’s competing with you. I saw your comments at DealBook that you think he’s above using his position to engage in any funny business as it relates to AI.

I do think so.

But if I may: In the past few years he bought Twitter, then sued to get out of buying Twitter. He replatformed Alex Jones. He challenged Zuckerberg to a cage match. That’s just kind of the tip of the funny-business iceberg. So do you really believe that he’s going to?

Oh, I think he’ll do all sorts of bad shit. I think he’ll continue to sue us and drop lawsuits and make new lawsuits and whatever else. He hasn’t challenged me to a cage match yet, but I don’t think he was that serious about it with Zuck, either, it turned out. As you pointed out, he says a lot of things, starts them, undoes them, gets sued, sues, gets in fights with the government, gets investigated by the government. That’s just Elon being Elon. The question was, will he abuse his political power of being co-president, or whatever he calls himself now, to mess with a business competitor? I don’t think he’ll do that. I genuinely don’t. May turn out to be proven wrong.

When the two of you were working together at your best [at OpenAI], how would you describe what you each brought to the relationship?

Maybe like a complementary spirit. We don’t know exactly what this is going to be or what we’re going to do or how this is going to go, but we have a shared conviction that this is important, and this is the rough direction to push and how to course-correct.

I’m curious what the actual working relationship was like.

I don’t remember any big blowups with Elon until the fallout that led to the departure.

Let’s presume you’re right and there’s positive intent from Elon and the administration. What’s the most helpful thing the Trump administration can do for AI in 2025?

US-built infrastructure and lots of it. The thing I really deeply agree with the president on is, it is wild how difficult it has become to build things in the United States. Power plants, data centers, any of that kind of stuff. I understand how bureaucratic cruft builds up, but it’s not helpful to the country in general. It’s particularly not helpful when you think about what needs to happen for the US to lead AI. And the US really needs to lead AI.

The interview was edited for length and clarity. For the full Q&A, go to bloomberg.com/businessweek

Remarks

• Globalization can’t be put back in the box

In Context

• What’s the matter with Europe?
• Israel’s war is jamming its high-tech farmers
• Forget the influencers — luxury needs silver shoppers
• The Right Stuff: Treat yourself at bonus time
• A Botox boom puts a wrinkle in Korean health care
• Charles Schwab is back on track
• The electric grid needs more batteries
• The business of being Kendrick Lamar
• A Walk With: Yandex co-founder Arkady Volozh

In View

• Centrist politicians are out of touch on immigration
• Self-driving stalls again as Cruise gets scrapped
• American populism isn’t delivering for the working class
• Retail’s wacky experiment: Hire actual retailers
• The US has plenty of room to ease the housing crisis
• A Chat with the ChatGPT guy
• The Swiss sneakers that are gaining on Nike

The NA Business

• All the buzz around nonalcoholic drinks
• Athletic Brewing tackles the near-beer problem
• The global growth of NA drinks
• The great mocktail taste test
• Alison Roman’s faux-drink philosophy
• Weed is still illegal in lots of states — unless it’s in a can
• How Ghia won over the most influential nondrinkers
• Actual alcoholics? They may stick to ginger ale

Pursuits

• Hairdresser today, business coach tomorrow
• The hottest London hood for Londoners
• Watch Club: Vacheron celebrates 270 years
• An electric Dodge with serious muscle
• The weekend plans of nine CEOs
• Everyone loves a killer rom-com
• Movies, music, museums: Pursuits picks for February
• A super gamer needs a souped-up computer
• At the Almare resort, everything’s included—but kids

Exit Strategy

• Fasten your seat belts, billionaires

скачать журнал: Bloomberg Businessweek (February 2025)