OpenAI CTO Mira Murati on ScarJo Controversy, Sam Altman & Disinfo Fears | On With Kara Swisher

0:00Support for this show comes from Strawberry.me. Be honest. Are you happy with your job? Or are you stuck in one you've outgrown? Or never wanted in the first place? Sure, you can probably list the reasons for staying, but are they actually just excuses for not leaving? Let a career coach from strawberry.me help you get unstuck. Discover the benefits of having a dedicated career coach in your corner. Go to strawberry.me slash unstuck to claim a special offer.

0:36Hi, everyone. This is Pivot from New York Magazine and the Vox Media Podcast Network. I'm Kara Swisher. We're off for the holiday today, but we have a special episode for you from my other podcast, On with Kara Swisher. In this episode— Do you have another podcast? Yes, I do. I'm cheating on you. You have another podcast. I'm cheating on you, and so are you. I know all your—what you're up to. Mr. Malice, no mercy, Malice. No mercy, no Malice. That's like—I got that from one of those— Why Malice? Why did you go to Malice? No, I heard—I was watching late night at 2 in the morning on Edibles, one of those wildlife shows where a crocodile crunched the shit out of a baby cheetah, And James Earl Jones' voice goes, on the Savannah, it is no mercy, no malice.

1:12I'm like, boom, that's the name of my newsletter. Oh, my God. I had no idea. Anyway, in this episode, I interview Mira Marati, the chief technology officer at OpenAI and one of the most powerful people in tech. Enjoy.

1:38Hi, everyone from New York Magazine and the Vox Media Podcast Network. This is On with Kara Swisher, and I'm Kara Swisher. Today, we have an interview with Mira Mirati, the Chief Technology Officer at OpenAI. OpenAI has certainly been in the news, but not many people know about Mira herself. She's only 35, but she's already one of the most influential women. Scratch that, one of the most influential people in tech. She helped OpenAI skyrocket to the forefront of the generative AI boom with the launch of ChatGPT in late 2022, and what a ride it's been. The company's now valued at$80 billion after a big investment by Microsoft, and it recently signed a deal with Apple to put ChatGPT in Apple products.

2:22It's a major move by a very small company. Of course, it hasn't always been good news. After the board fired Sam Altman, Mira became CEO for just two days in what the company called The Blip until Sam was reinstated. Since then, OpenAI has had to deal with a string of bad news cycles. There have been high-profile departures, an open letter accusing the company of putting product over safety, questions about highly restrictive NDAs, and even controversy over whether or not they had stolen Scarlett Johansson's voice. And with the presidential election coming up, the public's anxiety around AI-fueled disinformation will only get worse.

2:59This episode's expert question comes from Fei-Fei Li, the founding co-director of the Stanford Institute for Human-Centered AI and an early AI pioneer, in other words, a godmother of AI. And it was recorded live at the Johns Hopkins University Bloomberg Center in Washington, D.C. as part of their new Discovery Series, where I'll be talking to some of the top leaders in AI over the next year. I think Miramirati is the best place to start.

3:38This is Mira, everybody. Hi, everyone.

3:44So thank you so much for joining me at Johns Hopkins University Bloomberg Center. We're recording this live. There's a lot to talk about. We'll get some good news. We'll get to some not so good news. We'll talk about disinformation in the elections. So I think we'll have to ask first about the Apple partnership. Apple computers, phones, and iPads are going to have chat GPT built into them sometime this year. Obviously, this is a huge deal. It's the first one Apple's done. They've been talking to a number of people. They may include other people over time. You remind me a little bit of when Netscape got in different places, and you don't want to have that fate befall open AI as becoming the Netscape of AI.

4:23Yeah, so, I mean, I can talk about the product integration specifically. I can give you specifics on that. But what we're hoping to bring is really the capabilities of the models that we are developing and the multi-modalities and the interaction to bring this thoughtfully into the Apple devices. then it opens up a lot of opportunities. So that's what we're looking for. And yeah, I guess we had great technology. And they wanted to use yours. So when you're dealing with a company like Apple, whose reputation matters a great deal to them, especially around privacy, what were some of the things that they thought was important?

5:05Because one of the issues is the worries about where this information goes and what it's used for. I think this is a very aligned partnership when it comes to privacy, when it comes to trust. I mean, for the mission of OpenAI, it is so critical that we build technologies and we deploy them in a way that people feel confident around them and they feel like they have agency and input into what we're building. So in that sense, this partnership is quite natural and we feel very aligned. and it's only going to take us deeper in the direction where we want to go. Specifically to your question on misinformation, you know, this is obviously very complex because we're building on top of decades of misinformation and, you know, it's becoming even more and more intense with AI.

6:11But of course, we've got the internet, we have social media, and these are compounding effects in a way. It's actually good that AI is bringing all of this to a head. And there is such scrutiny and intensity on this issue. Because it feels like there is more of a collective effort and responsibility to do something about it that is meaningful. I think it's going to have to be iterative. So we'll have to try out things as we go. And, I mean, if you look at the governance of news and media in the past 100 years, you know this better than me. It's been sort of iterative every time there is a new technology, things adapt.

7:04Yes, we lose business model every time. Perhaps not the best example. But the point is that it is iterative. Whenever there is a new technology, we adapt to it. And I think there is a technical innovations aspect that are going to help us deal with misinformation. And then there is the people issues and societal preparedness that is perhaps even more complex. Well, I do know with Apple, you just can't fuck up because they'll make trouble for you if that's the case. Are you talking to other companies to do things like that? Obviously, you have a relationship with Microsoft.

7:55Podcast listener, she smirked at me. I'm not going to tell you anything away. All right, I'll move on from that. OpenAI has made deals with News Corp, Atlantic Media, and Vox Media, by the way, to license their contacts. So that's three potential lawsuits you don't have to worry about. I do own my podcast, and it's not included in your deal with Vox. Sorry. I would consider licensing it, but probably not. How could you convince me to license my information? I don't want anyone else to have it, including you.

8:29well so i know you you'll ask about this at some point so i might as well tell you now um so when we look at the data to train our models right uh we're looking at sort of three different categories the publicly available data we look at partnerships that we've made with publishers and we also pay human labelers to label specific data and also users that allow us to use their data. So these are kind of the main categories where the data comes from. And the way that we think about publisher deals specifically is we care about accuracy of information, we care about news and our users care about that they want to have accurate information and they want to see news on chat gpt and so it is a product-based relationship where there is value provided to users through the product and we're experimenting with different ways to monetize and give content creators basically some form of compensation for having their data show up in the products or being used in training or whatever we are doing with the data.

9:52But it is very specific partnerships that we're doing one-on-one with specific publishers. So some people do deals with you. You've done quite a few with AP and many others. But some sue, like the New York Times. How does it get to that point? Because I think a lawsuit is a negotiation in a way. I mean, I can't comment on the lawsuit specifically, but, you know, it's just, it's actually, yeah, it's quite unfortunate because, of course, we think that it is valuable to have news data and this type of information on the product. And so, you know, we try to figure out a partnership, a deal around that.

10:35But in that case, it didn't go well. Yeah. Well, it might go well someday. But I think it's because media has dealt with Internet companies for years and usually has ended up on the very short end of a very long stick of theirs. Every episode, we get an expert to send us a question. Let's hear yours. Hi, Mira. I'm Fei-Fei Li, professor of computer science at Stanford University, also founding co-director of the Stanford Institute for Human-Centered AI. So since data, big data, is widely considered to be one of the three elements of modern AI, I want to ask you a question about data. Much of OpenAI's success in your models is said to be related to data.

11:21We have learned that your company has acquired an enormous amount of data from the internet and other sources. So what do you think the relationship between data and models are? Is it as simple as the more data to feed into the model, the more powerful the model? Or is it that we need to spend lots of time curating different types of data in order to make the model work? And finally, how do you reconcile this appetite for so much human-generated data with the ownership and rights issues of these data? Thank you so much. That's a great question from Fei-Fei. So in terms of the relationship of data and models, this is actually something that a lot of people misunderstand about AI models and in particular large language models.

12:17The developers of these models, they're not pre-programming these models to do something specific. In fact, they are putting in a bunch of data. So these models are ingesting a huge quantity of data. and they are these incredible pattern matching systems. And through this process, intelligence emerges. So they learn to write and they learn to code. They learn to do basic math. They learn to summarize information and all sorts of things. We don't know exactly how this works, but we know that it works. Deep learning is very powerful. But this is important because then people keep asking how it works and it goes into the transparency questions.

13:08And this is where we can describe the tools that we are using to provide transparency to the public about what we're doing. So understanding this first part is very, very important. how the large language models work and you're combining you know this architecture neural nets and a lot of data and a lot of compute and you get this incredible intelligence and as we're thinking about providing transparency into the model behavior and how things work one of the things that we've done is actually share with the public this document that we call the spec, the model spec, and it showcases how model behavior works and the types of decisions that we make internally at OpenAI and that we make with human labelers.

14:04And you see, by looking through the spec, you see the complexity of what's going on, that sometimes direction is in conflict. Like, for example, you might say to the model, I want you to be very helpful. And also, I don't want you to, you know, disobey the law. and let's say someone puts in a prompt that says, you know, give me some tips to shoplift. Then the model is meant to be very helpful, but also it's not supposed to help you with something illegal. So it's not helpful. Yeah, maybe. So how does it decide? A person certainly knows how to, or some people, not all of them. Right, but the model could interpret the guidance as, you know, here are some tips to avoid shoplifting.

15:07And accidentally it kind of gives you sort of, yeah, things that you could do. But that depends. There's not so much model behavior. There's so much on, that's more on the person. And that goes into the area of misuse. But this just goes to show that model behavior is actually quite complicated. and it's not as simple as picking liberal values or other values. Or putting anything into it. One of the things I think that gets people is the confusion about what's in it and what's not in it. I think provenance is a big idea. In March, you had an interview with Joanna Stern of The Journal who asked you if OpenA had used videos from YouTube, Instagram and Facebook to train Sora, which is your text-to-video model, which is getting better and better.

15:52You said you didn't know. Shouldn't you know? Right. So, I mean, I didn't handle that question. Okay, why don't you handle it? Well, now. Redo. So, I cannot tell you specifically where the data comes from, but the data comes from these three categories. So, I can't give you the specific source because, I mean, this is trade secret and it helps us stay competitive. but I can tell you that the categories of data and it's the ones that I mentioned earlier, publicly available data, data that we pay for through licensing and deals that we make with content providers as well as data from users or where we advise users.

16:41Well, at the only reason I'm asking, perplexity just got into trouble because they were basically scraping in a quicker way a story and then not giving the sighting of it. You could see how any media company could be worried about that idea. Yeah. So we want to make sure that we are respectful to content creators. And we are doing a set of things to experiment with ways to compensate people for data creation. So we're building this tool that we're calling Content Media Manager. And this will allow us more specifically to identify the types of data that... Record companies do it. It's been done in the past, so it's not an impossible thing to be able to do that.

17:25Speaking of Sora, Ashton Kutcher told Eric Schmidt, what an interesting pair, I have a beta version of it, and it's pretty amazing. He also said the bar is going to go way up, because why are you going to watch my movie when you could just watch your own movie? When will Sora be ready for public release? We don't have a timeline for public release for Sora yet. What we're doing right now with Sora is we've given access to red humor, and we've given access to some content creators to help us identify ways to make this robust. We're doing a lot of work on safety front, but also to figure out how do we actually bring this to the public in a way that's useful.

18:05That's not very straightforward. Right now, it's really a technology. And this has been a pretty consistent process that we have followed with every new technology that we have developed. We'll usually work with those that have, like, for example, with Dali, we worked with creators initially, and they helped us identify ways to create an interface where they felt more empowered and they could create more projects. Basically, you just want to extend the creativity of people. So, Sora presumably is a little more dangerous than a chatbot. correct is that the worry i mean you could easily see porn movies with scarlett johansson for example i'm not i'm gonna ask about her in a second but but um that she wasn't appearing in like things like that how do you are you more worried about video is that uh well yeah video has a bunch of other issues right because uh especially when done very well which i think sora is quite remarkable um and video is very visceral and and of course it can be very emotionally evocative um so we have to address all the safety issues and figure out the guardrails and figure out how do we actually deploy a useful and helpful product but also you know from commercial perspective nobody wants a product that is going to create a bunch of you know safety or reputational scandals out there that's just facebook but go ahead facebook live nice to meet you but go ahead so you know it's uh yeah so we're laughing you can laugh it's funny we're so you know i think this is really incredible and magical technology but the breadth the reach the consequence it's it's also great um and so it's important that we get this right um now of course at opening eye list we we use iterative deployment strategy so we usually released to a small group of people.

20:22We try to identify edge cases. And once we feel confident about how we handle them, we expand access. But you need to figure out what is the product's surface and what's the business model around it. And we have... About that idea of consequence, one of my themes, one of my big themes is lack of interest in consequences of... Not you, earlier tech companies. They just... We became the beta tester for all their stuff. If they released a car like this, they'd never allow it to happen. They'd be sued out of existence. But a lot of tech is released in a beta version. The idea of consequences, do you feel as if you yourself as chief technology officer, even if you can't figure out all the consequences, there's enough respect for the idea that there are consequences for every single invention you make?

21:12it's consequences that we that we will feel on our skin and on our society so by that I don't necessarily actually mean regulation or or legal ways I mean you know a moral imperative to get this right um it's you know I'm optimistic and I think this technology is incredible and it will allow us to do just amazing, amazing things. You know, I'm very excited for its potential in science, in discovery, in education, in particular in healthcare. But, you know, whenever you have something so powerful, there is also the potential for some catastrophic risk. I mean, this has always been the case. Humans have tried to amplify it.

22:05True, but, I mean, a quote that I used in my book was from Paul Virilio, when you invent the ship, you invent the shipwreck. This is more than a shipwreck, a possibility, correct? I disagree with that because my background is in engineering. Our entire world is engineered. Engineering is risk, right? Like the entire human civilization is built on engineering practice, like our cities, our bridges, everything. And there's always risk that comes with that. And you manage that risk with responsibility. But it's not just the developers. It's a shared responsibility. And in order to make it shared, you actually need to give people access and tools and bring them along instead of, you know, building it in a vacuum and technologies that are not accessible.

22:58Last month, you announced the iteration of ChatGPT for, I love your names, ChatGPT4O. It's a great name. Can't you call it like Claude? They all have those. That's okay. ChatGPT is fine. You're making it free, correct? That one's free. But then you also announced you're training a new model, ChatGPT5, and then there'll be 5AB. But will that be an incremental step forward? Is it exponentially better? And what's the expected release date? uh so yeah on on gbd4o it all stands for omni model okay because it ties together all the modalities uh vision text um audio and what's so special about this model is that for the first time you can interact very seamlessly and naturally with the model.

23:58The latency is almost imperceptible. And that's a huge jump in the interaction with AI. It's quite different from the previous releases that we have made. And we wanted to make this the latest capability free for all users. We wanted everyone to get a sense for what the technology can do, what these new modalities look like and also understand the limits of it and it goes to what I was saying earlier that you actually want to give people access to bring them along because it's so much easier to understand the potential and the limitations of the technology if you have if you're experiencing it and if you have a intuitive sense so what is in five it all could be like a you know this little appetizer so now buy five but what what is in five that's different is it incremental or a very big leap?

24:54We don't know. But I mean, that's going to... You know a little bit. I don't know what we will call it. Right. But the next large model is going to be quite capable and we can expect, you know, sort of big leaps like we've seen from GPT-3 to GPT-4. But we don't know yet. What do you think will be in it? You do know. We'll see. We'll see. I'll see. But what about you? No, even I don't know. What? Even I don't know. Really? Okay. All right. An internal open AI roadmap predicted that would achieve AGI, which is artificial general intelligence. For people who don't realize, it has not been achieved by 2027, which would be a huge deal.

25:41Explain the significance and also when do you estimate we'll achieve AGI? um so people will define agi differently uh we have a definition of agi by the by the charter which is the systems that can do you know economically valuable work across different domains um and you know from what we're seeing now the definition of intelligence just keeps changing. So a while back, we would look at academic benchmarks to test how intelligent the systems were. And then once we saturated these benchmarks, we looked at exams, school exams. And eventually, you know, when we saturate those, we'll have to come up with new evals.

26:32And it makes you think, how do we evaluate fit and intelligence in a work environment? We have interviews, we have internships, you know, we have different ways. So I do expect that this definition will continuously evolve. I think perhaps what's going to become more important is assessing, evaluating and forecasting impact in the real world, whether it's, you know, societal impact as well as economic impact in the real world. So not this moment where it just suddenly goes, oh, look at me, and decides what to do for itself, right? I think that's the worry, correct? Because, you know, for the AGI definition specifically, yes, and, you know, that's important, and I think the definition of intelligence will continue to evolve.

27:24But I think what's equally important is how it affects society, and at what rate it actually penetrates Using that definition, when does OpenAI think that? Is that 2027 number correct?

27:43Well, I'll say, you know, within the next decade, we will have extremely advanced systems. But what people are worried about, because obviously we have to talk about the safety versus product discussion. Now, OpenAI was started this way. I think the reason you're having these discussions is because the way it was started, You had, I would say, a mixed marriage, the people who were there for helping humanity, the people there who really liked$1 trillion. Or in between. I think you're probably in between. Last week, 13 current and former OpenAI and Google deep mind employees. It crosses lots of companies.

28:20It's not just – OpenAI just gets all the attention because it's gotten a lot of attention, obviously. They publish an open letter calling for companies to grant them a right to warn about advanced artificial intelligence. This isn't new. Facebook, Google, and Microsoft employees have been known to sign open letters, whether it's working with the Defense Department, et cetera. But in this case, employees say that, quote, broad confidentiality agreements block us from voicing our concerns, which is essentially saying, oh, no. We can't tell you what oh, no is, but you'll all die, essentially. That's what it sounded like from the letter.

28:52What's your response? and people saying they're worried about retaliation. And I'm not going to go into the vested equity because I think you've apologized and corrected that, but shouldn't they be able to voice their concerns if they have them? And I know there's differing opinions. Yeah, definitely. I mean, we think debate is super important and being able to publicly voice these concerns and talk about issues on safety. And we've done this ourselves, you know, since the beginnings of OpenAI. We've been very open about concerns on misinformation, even since the GPT-2 days is something that we've studied since early on.

29:35I think that, you know, in the past few years, there has been such incredible progress, such incredible technological progress, that nobody anticipated and forecasted. And this has also increased the general anxiety around societal preparedness. As we continue this progress, we see sort of where the science leads us. And so it's understandable that people have fears and anxieties about what's to come. Now, I would say specifically the work that we've done at OpenAI, the way that we've deployed these models, I think we have an incredible team and we've deployed the most capable models very safely and I feel very proud of that.

30:26I also think that given the rate of progress in technology and the rate of our own progress, it's super important to double down on all of these things. security, safety, our preparedness framework, which talks about how do we think about the risk of training and deploying frontier models. Right. But you talked about that. I mean, one was why the need for secrecy and non-disclosure and stricter than other companies, one. And two, the open letter comes after a string of high-profile departures, including Jan, I think it's Jan Leike, and Ilya Sutskiver. They led the now disbanded super alignment team, which was in charge of safety.

31:11Ilya was a co-founder. He joined with three other board members to oust Sam as CEO. I don't think it's a surprise that he's gone. But Leike posted this on X over the past year's safety culture and processes have taken a backseat to shiny products. That's probably the most persistent criticism leveled at OpenAI. And I think it's the split in this company from the beginning that this was one of the issues. Do you think that's fair? And why or why not? If you say you're very interested in safety, they say you're not. How do you meet that criticism? Well, a few things. So the alignment team is not in charge of safety.