格雷格·布罗克曼称，AGI 已经到来

We're now in the AGI era. Astra has really hit something that I'm like, "Okay, I think this is pretty reasonable to call it AGI." We've seen it run coherently for 24 hours to accomplish tasks that I think are quite amazing.

The models will be plenty powerful, but it'll be hard to get to everybody, given that we won't have enough compute to serve it all.

You don't win the Super Bowl by saying, "I want to win the Super Bowl." You win it by blocking and tackling. You really have to make sure that safety, security, and alignment are all standards that you're constantly upleveling. I think that's going to be a huge challenge people are underestimating.

You've made 2 very big bets in your career: helping build Stripe early and helping, of course, cofound OpenAI.

Throughout OpenAI, I have always focused on whatever is the most important problem. For the past 2 years, it's been the data centers, the infrastructure, and the machine learning.

Do you know what we'll focus on in the next year?

I think this is going to become the most important conversation.

So, Greg, you've made 2 very big bets in your career: helping build Stripe early and helping, of course, cofound OpenAI. If we were talking 10 years ago and you were predicting what the world would look like in 2026 as it relates to AI, would you have been able to predict that we would be making the breakthroughs that you've made today? What would you tell your 10-years-ago self about what to expect?

Ilya and I actually spent a lot of time trying to predict what it would look like and what the timelines would be. I remember we did some math on compute in around 2016 or 2017. We came to the conclusion that, if you look at Moore's law progress and that kind of thing, 15 years felt like about the timeline to AGI.

If you really squinted at it and were willing to scale up, build massive supercomputers, spend hundreds of billions of dollars, and that kind of thing, maybe it'd be 10 years. I actually feel like, in some ways, obviously what's happening is remarkable. It's this amazing sort of moment for everyone to be a part of and to be able to help shape collectively.

But it also feels a little bit like maybe it's the conclusion of a lot of forces that are all coming together for this moment. If you step back and take that macro view, it kind of makes sense that it's happening now.

Do you think we're still on the timeline, given that we're now starting to drive real shortages in the supply chain?

I do think that we're in a world where it's hard for compute to keep up with the demand that we're already seeing in the market. Just in terms of how people are going to use this technology and benefit from it, I do think it's going to be very hard for us to scale the raw potential and capability of these models to everyone.

That's part of what we try to do, and so I think the progress I do see is that we have a line of sight to continue to make the models much more capable, safe, and aligned, while also really distributing that power, the benefits, and the empowerment to everyone. I think that's going to be a huge challenge people are underestimating.

So the models will be plenty powerful, or they'll continue at a steady pace, but it'll be hard to get them to everybody—certainly in an affordable way—given that we won't have enough compute to serve it all.

I think that's true. I do think we're at a point now where we have to really start thinking about what we call "pacing the frontier." As we move to more capable models, you really have to make sure that safety, security, and alignment are all standards that you're constantly upleveling.

Those actually become almost the bottleneck to progress, or the part that you have to spend a lot of your effort to make sure you've gotten right. In my mind, it's more those constraints than compute. I think we can make it happen.

On the flip side, bringing it to everyone, which is ultimately about our mission—empower everyone and ensure it benefits everyone—is something that I think deserves a lot more airtime than it's gotten.

1. Cyber Hacking, Reward Hacking & Building Safety Into Architecture

Let's get a little deeper on the safety thing, because it's been very interesting to me. In the beginning, safety felt like, "Okay, let's make these things not say nasty stuff that people don't like." The approach that was taken was kind of a surface-level approach around the edges: we'll put some filters on this thing and do RLHF around the edges.

If you get deep into the thing, you'll be able to get the bad words out. But if somebody wants to go through that to hear bad words themselves, who cares? Now, when you get to the point where these things are really good at cyberhacking and other kinds of activities, you need a more architectural idea, where the model itself knows not to reward-hack in a way that's going to be dangerous, and so forth.

Do you feel like we can make progress against that quickly, or is that a really hard, different category of problem? How are you thinking about that?

I absolutely think we can and are making very rapid progress on this problem. There's a lot of great ideas and research that we've been investing in for many years.

If you rewind to 2017, I think people underappreciate some of the key results that came out of OpenAI and the field at the time. There were the first inklings of modern language models—you can find a paper from 2017 that laid that out with LCMS—but also reward learning, or reinforcement learning from human preferences, which was also created in 2017 to start thinking about how you can align a model to match what humans want by providing feedback from people.

Yeah, just for usability.

Exactly. In 2017 and 2018, we had ideas for, if you have something that's very smart and capable, how can you actually supervise what it's doing? How can you provide feedback and ensure that it's staying aligned with you?

We had ideas such as debate and iterative amplification. These were ideas from before these systems existed, and you can start to see the trickle-down of those ideas into modern systems and investment.

In some ways, there was this early phase when OpenAI started where we were really thinking about AGI safety and things like that. It was very front and center, even in the communications. Then, as things like ChatGPT took off, people started to see, "Okay, well, we're not at this point yet," and questions like "Is the AI politically neutral?" started to come to the front and center.

Now that we're here, all these other ideas that we've been talking about for a long time are taking the main stage again. I think we've been thinking about this moment for a long time.

That's really good news. We don't have a really great community yet among the SOTA models, but it seems like those kinds of ideas are things that you, Google, Anthropic, SpaceX, and Meta would want to share, as opposed to saying, "This is a proprietary idea that's a way to keep these models safe," since you're all on related architectures.

How do you see that unfolding? Is everybody going to do it independently?

I think there's nuance here, and I do think coordination is going to be a very important theme. We need to think about this within the frontier labs and, more broadly, about what has to happen for humanity as a whole to navigate this technology in the best way.

We're going to have to think hard about those kinds of questions. We've published a lot of our thoughts. Again, some of this is about pacing the frontier, and some of it is about unilateral actions that we can take and how we think about making safety cases for training, developing, and evaluating these kinds of models.

All of that's new. No one's ever really had to operationalize this before, and it is not at all unique to OpenAI. There's a whole world that is basically developing this technology.

I think one thing that's easy to miss is that what we're building is almost a sort of thing that falls out of compute progress. In some ways, compute progress is something that falls out of technological progress. There's this massive wave that's been building for a very long time, and we're starting to see the leading edges of this technology.

Companies like OpenAI can lead by a bit in order to peer into this future and really understand what is possible and how we shape this technology, but we can't do that alone. I think having coordination—and especially talking about safety techniques and sharing what we're seeing, alignment failures, and those kinds of things—is going to take a very front seat in this next phase.

2. The Hugging Face Incident & Why the Defender's Window Is Open

Very interesting. You've called the recent OpenAI–Hugging Face incident a watershed moment and talked about how the defender window is now open. Can you explain that statement and its significance?

I think Hugging Face shows 2 things. One is, call it, a wake-up call for us in terms of how we monitor, sandbox, and control the models during evaluation. We've risen to that occasion. Our team has changed so many of our internal standards and implemented a lot of controls that I think are very important and critical as we look to future, more capable models.

But there's a second thing that I think is also very valuable for the world that came out of this: an insight into what future capabilities will be like when they're broadly diffused and in the hands of threat actors. That will happen, right? There are so many people who are building these models.

There's something very important and good about the broad diffusion of AI capabilities, because there's a risk of concentration of power if one or a few entities have it—a huge risk, right? It's something not to write off at all. But you also have to prepare for everyone being empowered with tools that are cyber-capable.

In the case of Hugging Face, you saw both an AI that was able to hack out of a secure environment and hack into a company’s production environment.

Very cleverly.

Very cleverly, right? The things that it found were quite sophisticated.

Yes.

This capability, broadly diffused, is something that will really empower threat actors in new ways. I think that defenders need to use this time, before that technology is broadly available, to secure themselves. The nice thing about it is that it’s dual-use, right? If you can find vulnerabilities, as an attacker, you can use it for no good. But if you’re a defender, you can patch, right?

If you’re a defender, you control the battleground. You control the setup of your systems. Our belief right now is that there’s this window: You have frontier capabilities, you have the broadly diffused capabilities, and you, as a defender, by default, know that your security is probably pretty static—been static for the past 5–10 years, that kind of thing. You need to move and use these frontier capabilities that you’ll have differential access to, where we have trusted access programs and things like that, to bring these capabilities to defenders. You can use that to move yourself up so that, as the frontier capabilities get better, you get pulled along too, right?

Okay, so I’ve got a comment and a question on it. I would say there’s a third thing that we learned, which is that these things have capabilities that I don’t know that we all understood before. On the good side, I can deploy 10,000 agents, and they can talk to each other, organize themselves, and do stuff for me. That’s pretty amazing.

That was on the good side. On the other side, I agree that we’ve got a kind of defense window. However, we have 50 years of code, architectural ideas, and deployment ideas that weren’t built for this world. Yes, the AI can help us—find a bug, patch a bug, and so forth—but it seems like there’s maybe a bigger issue. We have these huge, massive honeypots of consumer data and all these things lying all over the internet.

From a consumer standpoint, it’s like, okay, I can’t protect my stuff. All these companies have to get their act together, which seems a bit worrisome. Do you think, in the future, we need a decentralized consumer architecture? Will this current world that we live in, with all these centralized data repositories, be viable in a world of AI?

So, several pieces to the answer. First, to your point on what you can get out of 10,000 agents, we actually use 10,000 agents to solve the Navier–Stokes problem.

Yeah, that was pretty awesome, by the way. Congratulations on that.

Thank you. Thank you. It’s both an important problem for what it is—it has significant implications and applications for fluid dynamics, for how you think about ocean currents, all these things—but also for what it represents: new knowledge created by AI and the unlocking of a whole wave of scientific discovery. Medicines, all those things, are on the table now. I think there’s something really amazing to think about in what can happen through the power of AI that is able to really help solve problems.

In the case of cybersecurity, how I think about it, we at OpenAI took our models and applied them to finding vulnerabilities. We took 25% of our production engineers and said, "Sorry, all your projects are on hold. You are now defending. You are now upleveling our security architecture. You're going to use the models to find all the holes." We found a number of serious issues and fixed them. I've talked to a number of CISOs over the past couple weeks and months, and there are many companies who are also telling me that they've applied these models, found some very significant issues, and were able to fix them. One positive part of the story is that when we took Astra and pointed it at our systems, we found some new problems, but eventually it saturated. We basically found, to our knowledge, all of the P0s, all of the critical problems that Astra is smart enough to find.

And of course, there will be a new model. There will be a new round.

Even smarter.

Exactly. But I think that’s the world that we’ll be in. You’ll be in a world where you want to be in this tight loop of new cyber-capability drops. You deploy it against your systems. You find the new holes, and ideally you’ve managed to automate this, what we call a defense factory. That’s what we’re building internally: this end-to-end process of finding a vulnerability, triaging it—

Remediating, deploying, validating, right? That end to end.

If you can do that at machine speed, I think the defenders will be advantaged in deeply significant ways. And there are—

Ideas, for example, formally verifying all software that are possible with AI.

Yeah, we never—that’s always been a dream. We’ve had these formal languages and all these kinds of things, but they never took off.

That’s right, because it’s intractable for people. It’s just so hard. But we have these models that are solving these crazy—

Impossible math problems.

And so one application of that proving power—one of the things to know about the Navier–Stokes problem is that we formalized it, right? We formalized it into Lean, and so—

So the AIs can write verifiable code?

They can.

Yeah. Very nice. Yeah, that’s a great idea. So I think there’s real hope, but I think that our view is that the world needs to act with urgency because—

We’re in a very dangerous window right now. Yeah.

We just see it coming.

Closing the loop on this incident, is there anything you felt that the narrative got wrong in an important way? Is there any preferred way of talking about what happened, or this window, that’s important to get across when you think about the public narrative? Or is there anything else inside the company that changed in terms of how you’re approaching this set of issues?

Well, two things. I think one big theme that people should take away from it is the question of access, right? These capabilities exist right now in the world, but they’re in a small number of frontier companies. The frontier companies have a trusted access program, which means that anyone who’s not in the trusted access program is not really able to benefit from the fact that there’s this differential. The people who are in it get a benefit if they use it.

I think there’s something we need to do as a field and as a society to really scale up the number of defenders that have access to these technologies, because every day matters. To use those days, you need access to these tools.

One thing that was actually interesting about the Hugging Face response was that they said they used frontier models to look over the logs of what had happened, because that’s the only way to actually analyze an attack like this. They said the frontier models refused, but they didn’t actually try our frontier models, and they believe that ours would have permitted it. There is something, too, about the default stance of providers. I think there’s something here about using these capabilities for good with this urgency and this real sense that it has to happen.

I’ll tell a quick story, by the way, which is unrelated but maybe also shows a little bit about how I think about this. I remember when we trained GPT-3. It was the beginning of December 2019, so everyone was about to head out on vacation. He was like, “Okay, we can train the model.”

I just remember feeling like this model was sitting on a shelf. No one was using it. It was this amazing technology, new to the world, new to humanity. Every day that no one was exploring what it was capable of, trying to understand it, and figuring out what to do with it—that was a day lost to the world.

So I canceled basically all my holiday plans. I spent the whole time playing with the model, building interfaces around it, and trying to see what it was capable of. I remember trying to teach it how to sort lists of numbers. It didn’t work very well. But I was really trying to probe it and see what was possible.

I think that spirit and ethos is something we should bring to what we’re building today. It’s obviously at a much larger scale, with much larger impact, but we as a world have the opportunity to understand this technology in this moment, which then helps us shape and steer where it will go next.

A very good point.

You said two things. What do you have as another one? One was access, or did you say both of them?

I think I said both of them. Yes, yes.

3. Astra & Why Greg Says We're Now in the AGI Era

Let’s go back to Astra. It’s incredible to see all the excitement on X, all sorts of use cases. People are excited about computer use. You’ve said that, in some ways, it brings us closer to AGI. Can you talk about what you find most compelling in Astra, or what you think the breakthrough there is in light of that statement, and where we still have left to go?

Actually, wait—sorry. Let me revise my answer. I can say both access and also tell another story about how I have used the models personally. After Hugging Face, I was thinking about how I could use these models in my personal life. What can I do to secure myself? And I have a website.

It's a very simple website, gregbrockman.com. Not the most popular website. I've got a good blog post on it.

Exactly. You've got some blog posts.

It's a static site. It's very simple. What kind of vulnerabilities could be there? So I took my Codex and asked it, “Go check out gregbrockman.com. Tell me if there are any vulnerabilities.” I did a pen test, and it came back with 13 findings.

These findings were things like: I had set my SPF record so that it would prevent people from spoofing emails, right? There was some stuff going over HTTP without forcing people to HTTPS, things like that. Individually, these things are maybe not the biggest deal, but if you think about an AI that's able to chain together many small vulnerabilities into a big one, I'm like, do I really want a hole where someone can scoop emails for me? Probably not.

It took 15 minutes for it to find these 13 findings. But then I asked it, “Can you fix these?”

Right, because fixing is so annoying—so painful and boring.

Exactly. And so, in 45 minutes, it opened up my Cloudflare control panel. It clicked around, set all the headers, migrated me to Cloudflare Pages, and set everything correctly. It started the DMARC process, which apparently you have to do—a 48-hour window of whatever. That was 45 minutes of fixing, and I felt so protected. I felt like, wow.

Did you ask it to find the vulnerabilities?

There you go. No, I didn't. Yeah, so it actually does do that automatically. This was by itself. It said, “I just checked. This one's fixed, this one's fixed, this one's fixed, and in 48 hours I'm going to have to run and set up a little automation.” So in 48 hours, it would check back in to complete the DMARC process. I was like, all right, this is—we're in business now.

All right, all right. gregbrockman.com.

There we go. You, too, can be protected.

Awesome. Let's transition to Astra. It's incredible to see all the excitement online and the use cases, with people really excited about computer use, among other things. You've said it's sort of a step along the way to AGI. I'm curious what you find most groundbreaking with it, and where do you think we still have left to go?

Well, I think that Astra is really a step function on so many axes, and in many ways it is the sum of a number of research bets that we've been making for years. To see them come into one model at one time has been absolutely incredible.

One thing to note about how we do numbering is that we've been wanting to have GPT-6 represent something that's worthy of it. The problem we always have is that our models are incrementally getting better, and so it never feels like it's the right moment to go for a major version bump. You always have to be like, “It's 5.6 now; it should be 5.7.”

This one just happened to be different because all these things came together at once—the first time that we actually had this almost discontinuous step, in a way that we could have predicted. It was just that all these factors happened to line up at once. It was a really positive moment.

To me, computer use is the headline thing that we've talked about. Part of the reason computer use is so significant is that, for agentic use cases, it really comes down to tools. Is the model smart enough to use the tools, and does it have access to the context that it needs through these tools?

People have been building these MCP servers and CLIs, really taking the world of software and making it accessible in this almost stilted way that is not really meant for humans, right? We're kind of retooling the world. It's like, “Oh, we'll build an API, like it's software.” But what if it's really more about behaving like a human? Can it just use a computer?

And the result of building that other layer is that you now have another layer of security challenges, this, that, and the other.

Exactly. Really, they've always felt very weird and suboptimal.

Yes. And from the very beginning of OpenAI, I remember in November 2015 we did this offsite in Napa and talked about our plans. We actually laid out a 3-step plan that is basically what we ended up following for the next 10 years.

We also talked about what if we could do reinforcement learning where the environment is screen pixels, keyboard, and mouse—the same interface as a human. Suddenly, any sort of task you could do with a computer is in there. It's in distribution, as long as we set aside sound, whatever. But you basically have the full power of a computer there.

We had some aborted attempts early on to try to build agents that could do that. It really took us until now, but you're seeing the power immediately. It's been so cool to see people take the Blender capabilities and, you know, take a screenshot of something they want to make a 3D model of. Lots of people are now designing houses or trying to redesign their living rooms, all by utilizing this capability.

To me, the thing that really stands out is that you can now move forward on AI that can do things for you without having to build all these specific connectors. There's so much software you don't even think about that you have to orchestrate every day. How much of your life is clicking around menus and typing things into a spreadsheet? None of that is what we should be doing.

4. The Future of Employment as AI Progresses

100 years ago, no one was doing any of these things. So it's not crazy to think that in 5 or 10 years, no one will be doing this stuff anymore. We'll get our time back. We're not going to have carpal tunnel or hunched shoulders, or all those physical problems that come from us contorting to the machine. Now the machine is there to help us, empower us, and really serve us.

Yeah, and that's a really good point, because I think one of the things that you have been more sober on as a company is what happens with employment. I think that's exactly right: there are all these things that we do because we have to do them, and they became valuable, but we shouldn't be doing them. All they do is wreck our health and wreck our personalities.

The idea that humans are going to just run out of ideas for cool things to do, or ways to make the world better, or problems to solve, seems a little absurd to me. So far, at least in the numbers, the better AI gets, the higher employment goes, not the lower.

I wonder—and of course it's unknowable; we've never had this technology before, and it's getting better—how do you think about the future of employment as it relates to these models and AI as it progresses?

Well, I do have a fundamental belief that AI is surprising. I think we even put this in the OpenAI launch post back in 2015, saying that the history so far has been that it just doesn't play out the way you think it does, even when there's this logical conclusion that it should be a certain way. I think the same will be true.

For any job, it's easy not to give it as much credit for how deep the field is and how much sophistication it requires. Building relationships and accountability is a good example of something where I think that people setting goals and being accountable for outcomes feel fundamental to me. Those feel like things that we actually should preserve for the long term. That's something that feels deeply human.

People are not valuable just because we can do tasks, right? We're valuable because we're people. I think it's important not to lose sight of that in some of these narratives. The way that things will change, and how what we do with our time evolves, is that we clearly will be in a world of abundance. How do we ensure that abundance is broadly distributed?

At the same time, I think we should be in a world where the ceiling of ambition is higher than ever before. We're going to see a wave of entrepreneurship, and it's already starting. I've heard from someone in a particular industry who was saying that a bunch of people in his world are now making the leap to quit and start their own firms. They're doing it because they have these AI tools and they're just like, “I can do so much more.”

So it's lowering the barriers to entry for entrepreneurs, and this should be a wonderful renaissance.

Yeah, it's been a lot of fun for us just going, okay, there are—particularly for our young people, because they get, by default, the grunt work—but what if the AI does the grunt work? Then they can really develop much faster, actually, because they can get involved in the real part of our business, which is: What is the relationship with the entrepreneur? How do we open up the world for them? How do we make them feel like, “Oh, they can do anything, and they're an important CEO, and they can go build things”?

That's as opposed to spending the whole weekend writing an investment memo, which, by the way, I have to say, Astra is very good at writing investment memos. It's awesome.

I love hearing that. Yeah, and again, I do think it's going to be a nuanced story, right? I don't think that we should paint it as though everything's going to be rosy and it's all going to be just easy. I think it's going to be hard. I think there's going to be change, but I think that it can be a much better world. I think the future could be much better than the past for everyone.

Yeah, that feels like what we should expect. Before the plow, the world was a lot worse; it was just a worse life, even though the plow did put a lot of human labor out of business and created the whole Luddite movement and all those kinds of things. Nobody here wants to go back to 1870. The idea that we don’t want to go into the future now seems a little shortsighted, but I think the speed at which things are moving is very, very scary for people.

We really recognize how things are moving, and we spend a lot of time trying to understand, as well as we can, how people are feeling and how we can be showing up better. I think there are 2 things. One is that, when we think about development and the pace of progress, we’re being very deliberate about it. Safety is our foremost priority. We think about how to build this technology in a safe, secure way and what those standards should be, and you can see that showing up in a lot of our communications inside the building.

It is absolutely what people are thinking about and what we care about. We really want this technology to empower everyone broadly. I think that, for us as a world, to really think about how we get the most out of this technology, how we get the benefits, and how we mitigate the risks, this is going to become the most important conversation that we have. I think that will emerge over even the next 1 to 2 years. This should be something that is front and center.

I think people sense it. You can sense it in how people react right now, even when thinking about things like data centers and these questions of whether we want AI, how we think about where it’s appropriate, and how we ensure child safety. All of these are core questions that we care so much about getting right.

5. Why AI Sentiment Is Higher in Asia Than the West

To that end, why do we think sentiment on AI is higher in certain Asian countries?

All Asian countries? Well, actually, in European countries everywhere, too, but the US has the lowest AI sentiment.

Why is that? Or, more—what’s driving that? What can we do about it? What can we learn from that?

One thing that I think about is that we, as a field and as a company, need to do a much better job of articulating to people why they benefit. Why is this a good thing for them, and not just for the country? I think that this technology is going to be, and is rapidly becoming, the single most important strategic priority and resource for the United States. It’s happening, yes.

Absolutely. You look at ChatGPT: 300 million health queries, or 300 million people, every single week using it for help, right? That’s a huge deal. We’re at almost 1 billion—1.1 billion weekly active users. I think within the US it’s about 100 million, something like that—about a third of the population, if I have that number correct, right?—is using ChatGPT every single week.

So people are touching this technology. But for many people, there are some who have gone very deep and really gone through the health journey, for example. That’s been true for my family and for my wife. She has a number of health conditions, and we don’t even really know how we would have managed these before ChatGPT. There’s just so much toil and time in getting to the right answer. A doctor tells you something, you don’t know what the thing is, and how do you get that sanity check to really even understand it?

People whose lives were saved through information delivered by ChatGPT—I’ll tell you a story about one of my friends. She was in the hospital, and the doctor was about to inject an antibiotic. She said, “Give me a moment,” and typed into ChatGPT. ChatGPT said, “Absolutely do not take that. If you do, you may die, because you have this thing that you had a year ago. You have this condition.”

She showed the doctor. “I know, right?” And the doctor said, “Oh, my goodness. No, that’s absolutely right. I had no idea. I only had 5 minutes to read your chart.”

Yeah. Many such cases, by the way—people telling me stories in their chats, at least.

Exactly. These kinds of stories don’t get told nearly enough, but they’re out there. I hear them every day. There are people who run their small business on ChatGPT and would be totally unable to do it otherwise. That kind of empowerment—people who are able to save money, make money, and live a better life—I think those kinds of stories need to be in the public consciousness as we approach this question.

It’s painting the narrative that, hey, you’ve got a teacher in your pocket, a doctor in your pocket, a lawyer in your pocket, a therapist in your pocket—all these utilities in your pocket—while also not threatening those same professions, and also telling teachers, doctors, lawyers, and therapists that they’ve now got this tool, too, and it’s going to make their business better as well.

Yes. And it’s not just the narrative; it’s the reality, right? You need both. I think many other countries are looking in.

Yeah. They’re seeing the position the US is in and the potential of this technology. Partly, too, you think about demographics. In many of these other countries, it’s more keenly felt that there’s an older generation that’s much larger than the younger population that’s going to need to support them. These questions of how that’s supposed to work—and so I think there’s something about really thinking to the future and thinking about what’s possible. How do you get the benefits out of this technology and really want to lean into that? I think we’re seeing that across the world.

Again, I think there’s something that we need to do better as a field and as a company in order to communicate this domestically. But I think the potential is there, and we’re in such a privileged position, leading this field in a way that was not guaranteed. It’s not guaranteed to remain true for the future, either.

Yeah. Particularly if we ban data centers, I think that’ll be a problem for maintaining our lead. It will drive the data centers overseas, which is what happened with silicon back in the 1980s or so.

Right, right. One of the interesting things about data centers is that they create so many blue-collar manufacturing jobs. I think Switch employs around 45,000 people on a union-contract basis to build data centers. That’s just one of the data-center providers in the US, and they’re great jobs. They’re high-paying.

While there have been bad actors in the data-center space, most of them are very good actors. They contribute to the power grid, don’t waste water, and aren’t noisy. It’s not that there was never an issue, but we could just say, “Hey, you have to be a well-behaved data center,” as opposed to, “We’re going to ban them,” or, “We’re going to stop AI.”

We’re not going to—even as a country, we’re not big enough to stop AI. AI will continue without us, and then we’ll have zero say, as opposed to being the leaders and having all the say. So it’s a really, really important cultural message.

I think it’s very important. On data centers, we’ve made commitments not to increase people’s electricity bills. Our data centers are all closed-loop water, so the amount of water used by Abilene, which is the data center that actually trained Astra, is about the same amount of water as an office building. The technology is quite advanced on that.

We have a number of community commitments so that we can actually help in Ohio and Georgia, where we have data centers. We’ve announced—we’ve talked about how we’re providing credits to every college student for Codex access. So there’s this broad set of benefits that we are bringing to bear. But again, I think we need to do even more.

Yeah. And I think making those kinds of things a requirement to build a data center is very reasonable. But let’s have positive-sum ideas, as opposed to, “Okay, we’re going to jump out of the AI game as a country and let China or whomever dictate what it’s going to be.”

Speaking of contributions, you guys made a $1 billion commitment to frontline defenders. Why don’t you talk about that?

We believe that every organization, every company, every government, and critical infrastructure such as water-service providers and hospitals should all be using these tools to secure themselves. But not every organization will have the capital required to do it.

We have a $1 billion commitment to frontline defenders—to organizations that we all rely on every day in our communities—to access our models and secure themselves. We think this is the beginning; this is not the end. We’re working closely with partners. For example, CrowdStrike and we are working together to provide discounted access to defenders as well.

I think there should be a global effort to bring these tools to bear and really secure every single organization, given what we see coming and what’s possible.

Yeah. And that’s a super-positive advance, because before AI, hospitals were getting broken into and held hostage all the time. Our water supply has been hacked by foreign actors—state actors—and we’re already dealing with the fact that our critical infrastructure was not built with cybersecurity in mind.

It's not been maintained with cybersecurity in mind, and here's an opportunity to go from not even secure in a pre-AI world to completely secure. So to me, this is an incredibly important effort—not to have our water supply and our hospitals at risk. I really agree with that perspective, right? The point is, I think that we as a society have been lax, right? We've allowed tech debt to pile up. I've never met a CISO who felt that they were appropriately resourced or appropriately prioritized.

Never.

And particularly not in the public sector.

That's right. And so I think that we have to change that. We should have changed this years ago, but now is a moment where we actually have both a real motivation to do it and a real ability to do it. Delivering the secure world that we all deserve, so that we can really depend on it and be safe and secure in our daily lives and online lives—that to me feels like table stakes. We absolutely need to do this.

6. What's Still Jagged: What Astra Needs to Approximate AGI

Yeah, definitely. No, that's a great effort. Closing the loop on Astra, you've emphasized that capabilities still remain jagged. What do you think is still left to go, or still needs to be fleshed out, that most closely approximates your definition of AGI?

Well, I think that AGI has turned out to be less of a point in time and more of this sort of fuzzy spectrum. For me, Astra has really hit something where I'm like, okay, I think this is pretty reasonable to call it AGI. With its computer-use capabilities, you really can ask it to do long-lived tasks, and it'll just do it. We've seen it run coherently for 24 hours to accomplish tasks that I think are quite amazing, across a wide variety of domains.

Now, it still is jagged. Its writing is pretty good. It's the first time it's not slop.

Yeah.

But it's not great writing. I think there are a number of areas where we just need to polish it a little bit, and it would be fantastic. It's just not quite there.

I saw someone post a graph on Twitter of this kind of jagged frontier, and where we really need to be is much more steady across the board, really hitting all these categories. But I think what people are finding is that it is so capable across such a wide variety of tasks that it is accelerative. It is empowering, and it's something that I think we've never really seen—a model that's been a jump like this.

Yeah. One of the things that's been interesting for me is that as you solve problems, sometimes the world doesn't realize it. I haven't seen a hallucination in quite some time, but nobody says, “Oh, the models don't hallucinate anymore.” It's just kind of in the ethos that that's what AI does. Do you think that'll just go away over time, or does there need to be some continual education for non-hardcore tech people?

I think this thing is moving. I think one of the most important problems we actually have is continual education, right? People shouldn't have to extract from AI what it's capable of; it should go the other way around. The AI should say, “Hey, I can help you in this new way.”

We have over 1 billion weekly active users on ChatGPT, but I think we have something like another 1.5 billion people who have used ChatGPT and don't use it anymore.

Oh, wow.

Right. So think about that. That's a significant fraction of the planet. And those people—exactly those people—we should really be able to go back to and say, “Hey, we have made so much progress. We think we can be useful to you in these ways.” I think that just shows you the kind of problem we have in front of us.

These AIs—if you look at ChatGPT and ChatGPT at work—they're both text boxes, right? This new text box is way better than the old text box, but there are still some things that the old text box is better at. So don't always use it. That's not the AI we were promised.

The AI we were promised should be an AI that you talk to primarily over voice. You can talk to it over text if you want to, but it has persistence, memory, and context; it knows you. It's trustworthy. You've seen it be proactive and help solve problems for you, in your personal life and your work life. That's how it should be.

It should be something that you can really rely on for the things that you care about, that empowers you and helps you solve your goals. I think being able to explain to you how it can help you is a core part of that.

Yeah. Interesting—and proactively do it. That's such an interesting idea. We need more helpfulness out of our AIs.

Yes.

Which is kind of a thing: some humans aren't helpful, and probably the humans who develop AI are not very helpful people. I would guess, just being around engineers and researchers.

You'd be surprised. I think we have very, very helpful engineers at OpenAI. But there is something about how you work with another person, right? A new coworker you've never worked with.

7. What's Next: The AGI Era & Deep Co-Design

It takes you a little bit of time, right? You kind of feel them out. You see how they respond in different areas. People do not come with an instruction manual. Often, actually, sometimes it's interesting in areas like consulting, where they really lean into Myers-Briggs. They say, “Here's a quick way to know who I am and how I operate.”

There is some precedent for how humans can present a little bit more. You have a résumé. You have a sort of track record. People can ask for backchannels on you. So we've built up a way of understanding how a human will work and what the best way is to get the best out of that person.

I think figuring out what the right analog is for AI—and especially as AI changes, and we produce new tools and product surfaces and new models and all these things—will be a very important society-company interplay. Again, I think our north star should be simplicity, right? We really should be one unified AI that makes it easy and smooth for you to be less engaged with the computer and less wrapped around the computer. The computer should be there to empower you and help serve you.

8. How Greg Prioritizes His Time

The business is ripping. You guys have such broad surface area in terms of what you cover. How do you decide where to go deepest, what not to build, and how to prioritize your time as your role has evolved and changed and you've encompassed so many things—research, product, commercialization, management, and so on?

Well, they go hand in hand. Yeah.

So this year, the theme was focus.

I think we really realized that we can't do it all, right? We need to pick. And particularly, there's one thing we're trying to accomplish, which is our mission: We want to ensure AGI benefits all of humanity.

Now, how do you backsolve from that? What are the areas—like deployment and productization—that actually reinforce that? We do want to bring this technology to bear and have it uplift everyone, and have people deploy it in useful applications. All of that—personal life, work life, the whole thing—is very core.

But when you think about this moment we're in, with this agentic coding takeoff, that exponential, which areas reinforced that, and which ones were labeled a side quest in the media but just were not on track for it, even if they were individually something very exciting? That was a very core question that we had to grapple with.

Things like Sora—that's maybe the highest-profile one of these projects that we decided to cancel. Very, very painful, by the way. Not an easy thing to do, but it was so critical to unleash the business in many ways so we could really focus on bringing together the consumer and enterprise sides of ChatGPT into ChatGPT Work.

That's another area where we've really had to focus and say, this is what we're doing. A lot of the way we've thought about unlocking this moment is to really have a vision about where we think the future's going, and how the new capabilities that are emerging can best be brought to bear with a single, unified stack that works across the different areas, different walks of life, and different areas that we're trying to focus on.

It has been painful, right? If you look at the first half, there were just a lot of metrics that weren't going in the direction that we wanted, and there was a lot of telling the team, “We just need to focus on the basics.” One of my favorite management books is The Score Takes Care of Itself. Have you guys read that one?

Yeah, Keith's favorite.

Yeah, it's a great one. It's a very empowering book because you realize you cannot affect the outcome; you can only affect the inputs, right? You can only affect the basics. So focus on those basics, right? You don't win the Super Bowl by saying, “I want to win the Super Bowl.” You win it by blocking and tackling.

And so that's what we have done for this whole year. For myself, throughout OpenAI, I've always focused on whatever is the most important problem that I think I can move the needle on—something that just isn't going to happen without me.

And for the past 2 years, it's been the data centers, the infrastructure, and the machine learning engineering. That's an area where we really spent a lot of effort to get our pre-training infrastructure into great shape. This year, it's really been about the business. We've figured out how to get the research really humming. We figured out how to get the infrastructure really humming, but how do we really bring this technology to the world?

And I think that's where I've been really putting a lot of my efforts into trying to bring together a bunch of functions that were otherwise running in parallel or crosswise. And that is something where I think, as a founder, as someone who has touched every part of this business from the beginning, I think I've been uniquely able to go in and make the changes, make the hard decisions, and really figure out: This is the direction. Let's go.

A lot of my style is that I like to lead from the trenches. So I get very deep in the weeds on what the thing is and really try to keep asking a lot of questions. That's actually a lot of my style: just asking, “Hey, does this make sense still? I don't quite get that.”

Sometimes when things are confused—for example, over the past couple of days—there have been times when it's just: We've got a thing; we've got to figure out how to even talk about it. How do we think about it? How should the world think about this? I say, let's get everyone who can touch different parts of the elephant on a call. We're going through a Google Doc on a Hangout and asking, “Does this line make sense? Wait, what do we really mean by this?” So we're really trying to uplevel execution, sometimes in small ways and sometimes large.

Yeah. No, that's fantastic.

By the way, that's exactly the right way to operate.

Yeah. Do you know what the next year will focus on, or what to prep for now?

Well, look, I think the business is a huge area, and we're not done yet with really upleveling every part of execution. So I think there's a lot more to do there. But I also think we are moving into a new phase of AI development, right? I call this—and we call this—the AGI era.

And I think that's something you can debate—is it this model, the previous model, or the next model? It doesn't matter. The point is that we are in a new phase where safety, security, and alignment—really thinking about these things not just at deployment time, but all the way back at development time. Evaluation is objective. This is critical. It must happen.

This is core to our mission. This is core to what we need to do. And so a lot of what I spend my time thinking about is making sure: Do we have all the right processes? Are we talking about the right things? Do we have plans that, at an operational level, at a practical level, lead us to the kinds of security invariants and safety guarantees that we view as core to our mission, what we need to do?

And so I think that, again, the theme of OpenAI, certainly for the past 5 years, has been deeper cohesion, deeper intertwining across these functions that are maybe, on the surface, very disparate, right—from go-to-market to long-term research to chip design—by building these in a coherent way where everyone has context. They understand: How do I fit into the overall picture? What are we trying to do? What is the end outcome we want to achieve? That is what has to happen.

So I think the areas that I will focus on will be dictated by the areas that most need that intertwining. And I think I see us moving more and more in concert, in lockstep, as time goes on.

9. Could You Have Predicted Today's AI 10 Years Ago?

Yeah. We could go all day, but we have a hard stop. I think this is a great place to wrap. Greg, thank you so much for coming on the podcast.

Thank you.

Great. Fantastic.