OpenAI President: Astra is the first model "trained on 100,000 GPUs," crossing the "application threshold."

Author: Wall Street Watch

OpenAI President Greg Brockman accepted an in-depth interview on the eve of the Astra release, revealing for the first time that this new model is the first in OpenAI's history to be trained on over 100,000 GPUs, and stating that AI has crossed the application threshold of "computer use" ------this means AI no longer needs to rely on API connectors and can operate any software like a human.

On Friday local time, the technology strategy analysis website Stratechery published an in-depth interview with OpenAI President and Co-founder Greg Brockman, recorded before the official release of Astra. This is the most detailed public expression by Brockman regarding Astra and OpenAI's latest strategy to date.
Wall Street Watch compiled the key points of the interview as follows:

1. Astra is the first model trained on over 100,000 GPUs at OpenAI, representing a significant breakthrough at the engineering level.

2. Astra has crossed a critical application threshold in "computer use" capability, serving as a "universal connector" that no longer relies on individual software to provide API interfaces.

3. Brockman stated that in the era of AGI, safety, alignment, and capability must be advanced in parallel as equally important "requirements," rather than at the expense of one another.

4. OpenAI has directed Astra to scan for vulnerabilities in its own systems and complete repairs, marking a significant cultural shift in safety.

5. Regarding the Hugging Face security incident, Brockman admitted that the sandbox mechanism had flaws, and AI found creative "jailbreak" paths.

6. Brockman believes that OpenAI has already crossed the threshold of the "AGI era" internally, possibly reaching the AGI standards recognized by most people with Astra or the next generation of models.

7. In terms of chip strategy, OpenAI is developing its own Jalapeño chip and using AI-assisted design while maintaining a deep partnership with NVIDIA.

8. He believes that as AI capabilities improve, the carefully constructed "skill scaffolding" is gradually shifting from being an aid to a limitation, which is an important discovery in the development process of Astra.

"Training on 100,000 GPUs for the first time": The significance of Astra's scale

"This is our first time training on over 100,000 GPUs. This number is easy to say, but you need to imagine that scale------these data centers are, in a sense, giant machines we built to push AI technology forward, capable of harnessing such a magnitude of computing power, which is itself a true engineering feat," Brockman said in the interview.

He emphasized that the significant expansion of computing power is not just for enhancing model capabilities, "A large amount of computing power has been invested in safety and alignment work, and a lot of safety engineering has been developed around it. This is our most aligned model to date------and because its capabilities are so strong, alignment and safety have become more core than ever."

"Crossing the threshold of computer use": AI becomes the "universal connector"

Brockman positioned Astra's core breakthrough in "computer use."

"For the past two years, we have been in an 'era of connectors'------you have some software that humans can use normally, but AI does not have access. So you have to write a dedicated connector for each software to hook into its API; and many software do not have APIs at all, which is completely beyond AI's capabilities." He said, "Now, we have technology that is almost a 'universal connector.' AI has transformed from being strictly limited in helping you to being able to operate almost anything."

He further described how this change affects the way people work internally at OpenAI: "It is already changing the way people work at OpenAI, and I believe it will really enhance many companies and many individuals."

Notably, Brockman also admitted that crossing this threshold is not the end: "This does not mean all problems are solved. You need to consider how to set enterprise-level guardrails for these AIs, how to achieve appropriate supervision, management, tracking, and observability------we are advancing all of these."

"We have now entered the AGI era"

Brockman made a rather bold judgment in the interview:

"I believe that maybe it was the last model, maybe it is Astra, maybe it is the next model, but at some point among these, we will cross the threshold of AGI recognized by most people. …… In a sense, I would say we have already entered the AGI era."

He explained that at the current stage, alignment, safety, and capability need to be advanced as parallel requirements, "These aspects are becoming the bottleneck in development, and that is what we have been preparing for."

Old "skills" become burdens: AI is powerful enough that rules need to be simplified

A striking detail from the interview came from the actual training process of Astra. Brockman revealed that the behavioral norms OpenAI had spent a lot of effort writing for the model have now become a hindrance:

"We found that some skills we painstakingly established this year------to show the model the correct way to do things at OpenAI------are actually a net negative for its performance. It can generalize better than we wrote and find better handling patterns than we specified. It's like training wheels; initially useful, but once it runs faster and is more capable, it becomes an obstacle."

"One billion users is an investment": The fusion of consumers and enterprises

Brockman directly addressed a persistent market skepticism: ChatGPT has one billion users, but is this a distraction?

"In the United States, about one-third of the population uses ChatGPT every week." He said, "This is unique; no other product can compare to this kind of advanced technology."

He characterized these one billion users as an "investment": "This is a real investment that will accumulate value in the way future models are unlocked."

At the same time, he acknowledged the pain points: "The challenge of chat as a product is that it may not directly correspond to smarter models------if you just treat it as a replacement for a search engine, you may not truly feel the benefits." He believes the future direction is to bridge consumer and enterprise scenarios, creating a unified AGI system, "We don't want to do two things; we want to do one thing------build an AGI, one system, one unified tech stack."

He also emphasized the layout in the health sector: "Every week, 300 million people make health inquiries through ChatGPT; 300 million is a huge number."

AI self-checks for vulnerabilities: Pointing Astra at its own systems

On the topic of cybersecurity, Brockman revealed a striking internal practice: OpenAI directs Astra to its own systems to find vulnerabilities.

"We called on Astra to target our own systems to discover vulnerabilities------not just reading code, but truly examining how these systems operate end-to-end to find real, verified vulnerabilities, and then help us complete repairs and patch work." He said.

He also admitted that during the Hugging Face security incident, OpenAI's actions during the defense window were indeed not timely enough, "I think now is clearly the time, and maybe a few months ago could have been the time, but at that time the model's capabilities were much weaker."

Relationship with NVIDIA: In-house chip development is "multiplicative," not a replacement

In terms of value chain competition, Brockman made a statement regarding OpenAI's self-developed Jalapeño chip and its relationship with NVIDIA, clearly denying a replacement logic:

"NVIDIA is our preferred computing partner, and that will not change. In fact, our collaboration with them is deepening further. …… The internal expertise brought by our own chip project is, for me, a multiplicative effect, not a replacement; I believe it benefits everyone."

He also disclosed an interesting episode about AI-assisted chip design: in the last month before the Jalapeño deadline, the team directly let AI run optimizations, "We said, just let it run, without closely reading what it did. Later, when we went back to read it, we found it identified a whole bunch of optimizations that we had on our list but would never implement ourselves------that is indeed a cool story."

The following is the full transcript of the interview (translated with AI assistance)

Greg Brockman, welcome to Stratechery.
GB (Greg Brockman): Thank you for having me. I'm glad to be here.
Clearly, we have a lot of news to discuss, but since this is our first conversation, I don't want to miss the Stratechery biography question I usually ask anyone. I do want to ask—do you define North Dakota as part of the Midwest?
GB: Yes, I think so.
Well, as a Midwesterner, I certainly want to take some time to talk about that. You studied in Boston, as they say, but before we get into that— you already had an amazing resume before going to school, like participating in the International Olympiad, but that was in chemistry. Where did computers start for you, or have computers always been a part of your life?
GB: Well, computers have always been in the background of my life as I grew up. I loved playing computer games, but I was also very interested in math and science. In fact, until ninth grade, I thought I might become an actor—I really enjoyed performing and dabbled a bit in philosophy.
What a twist! I had no idea about that, but I plan to figure out the connection between what you do now and performing, but let's keep going.
GB: Well, in ninth grade, I felt like I had starred or played the lead in a couple of plays in middle and high school. I was thinking that I wanted to double down on something, that I could be the best in the world at something, and really try to push that field forward. I felt like I could either choose a more intellectual, hard science approach or go the artistic, performing, and creative route. I ultimately chose the hard science direction because I felt that might be the area where I could make the most significant impact on the world.
Why do you think going that direction could make the most significant impact on the world?
GB: I think for me, it felt like this—one thing I loved about performing was its communal aspect. I loved improvisation, just creatively thinking about things, bouncing ideas back and forth, but it also required you to be part of a team that had to work very well together, which isn't always guaranteed. That's a hard thing to achieve and find. What I liked about the more intellectual route was that it felt like honing your own skills. But one thing I did learn is that even if you're very good at writing code, that's not enough. In fact, it also has to do with bringing in great teams, and I think that's part of my career—really helping to build and shape an environment and culture that can deliver excellent results.
There's an interesting aspect here about how the environment shapes some of these things. You mentioned that you always played the lead in plays and movies. I realized this when my kids went through that stage—my daughter was really into musicals for a while—the competition for the lead role is fierce, and usually, if there's one competent male willing to volunteer, he gets the role every time. Was the competition for the lead role fierce, or were you just unique?
GB: (laughs) I think that might explain it. But let me tell you a story.
What about the reverse? Similarly, in North Dakota, aren't there fewer people particularly delving into hard sciences, so from that perspective, doesn't it seem rarer?
GB: Well, let me tell you two stories. The first is about performing. My first paid job was a performing gig. Mannheim Steamroller came to Grand Forks, North Dakota. Have you heard of Mannheim Steamroller?
I've heard of them. Yes, of course.
GB: They were having a concert and needed extra performers. They needed some people to dress up as toy soldiers and walk around because it was a Christmas holiday concert. I went to audition, and I was a scrawny ninth grader, while there were all these tall college students.
What they had everyone do was, "Okay, everyone walk that way," and then the casting people would compare notes, and then you would have a break, and then they would say, "Okay, walk the other way." I noticed that during the breaks, all those college students were chatting and hanging out, while I was thinking that the only requirement for this job was to stay focused for six hours during the concert, walking back and forth. So during the breaks, I just stood there, staying focused, completely immersed in the role. In the end, they said, "Okay, we pick this person, this person, this person. The rest can leave." I wasn't selected.
But on our way out, they said, "Actually, we think you're amazing. We really loved seeing how much passion you put into this, so we're going to create a new role just for you." So I got a gingerbread man role, and they gave me a full costume. That was my first job. So it was a bit like trying to think outside the box to get the job—not just defaulting to a role but trying to find creative ways to earn it.
But I think one really great thing about growing up in North Dakota is that I had the potential to be the best in the state in different areas I focused on. I excelled in math. I started taking a lot of courses at North Dakota State University from tenth grade. I got deeply involved in math competitions, and then I would go to national competitions, attend national math camps where I would meet the best people in the country. These people were amazing, and one of my real honors is that many of the people I greatly admired at math camp now work at OpenAI. So I had the opportunity to see them in this new field from a new perspective.
That's awesome.
GB: But it also meant I could really chart my own path. I started doing math research, and I knew that if I went to some of those stronger high schools or much more competitive states, I think I wouldn't stand out. I would have to follow the standard route. So it was precisely because of being in that area that I could explore my interests and really forge my own path.
So when people say they studied in Boston, they usually mean Harvard, which is where you started. Then you transferred to MIT. Then at some point, you were working for Stripe. What was the order? When did you meet the Collison brothers? What happened in Boston?
GB: Well, after graduating high school, I took a year off, and I actually started writing a chemistry textbook because I was very fascinated by chemistry in high school, participated in chemistry competitions, and proposed a unique way of thinking about it. Very much based on first principles, very mathematical, rather than rote memorization. I wanted to teach that method, wanted to promote it. So I actually wrote 100 pages. It's now on my website. I haven't finished it. I always intended to go back and continue writing.
That sounds like a retirement project, I love it.
GB: Exactly. Well, at that time, I was thinking about how to publish this book, so I asked a friend who had done something similar in math, and he said, "Well, you don't have a PhD, so no one will publish it. So you either self-publish"—I thought, oh, that's a lot of work and expensive—"or you could build a website to promote your ideas this way." I said, "I guess I need to learn how to program." So I went to W3Schools. Do you remember W3Schools? Have you seen that website?
No, I don't think I've seen it.
GB: Well, it's a classic website—they have HTML tutorials, JavaScript, CSS, PHP. I read through it, and then thought, "I should test this out." I remember I built my first little widget: you could click on a table column, and it would sort the rows accordingly. That was the coolest feeling ever. I was thinking about this thing, and now it exists in the world, and anyone can benefit from it. They don't need to understand the details behind it; it just works.
I remember the first program I built that had users was actually a competitive chatbot game, and one day I got 1500 clicks from StumbleUpon. That felt amazing, that 1500 people played my game, hoping they enjoyed it, and they were engaged enough to play it. I thought, "This is what I want; I want to help others, I want to create value for them and benefit them."
So that's the mindset I had when I entered Harvard, and it really changed my thinking. I thought I would pursue those more eclectic interests, but instead, I thought, "I just want to build things." In my freshman year at Harvard, I joined the computer club, where two seniors would engage in obscure technical debates every time. We would listen and think, "One day we will understand this, one day that will be us." But after they graduated, sophomore year came.
So you transferred directly to MIT because you realized you were focused on coding and building?
GB: Basically, yes. Because by sophomore year, I was managing the club. I should have been leading those obscure technical debates, and I thought, "I'm not ready yet. I have so much to learn. I need to be around people who are much better than me."
Got it.
GB: So I spent a lot of time over at MIT, and I felt it made sense to transfer there.
Stripe
Got it. So when did you meet the Collison brothers?
GB: I met them in 2010, at the end of 2010. We had a lot of mutual friends because John went to Harvard, and Patrick went to MIT, and they were looking around. That team was looking for who was interested in computers at those two schools, and my name kept coming up.
Right.
GB: So I got an invitation from their team and flew out. I remember meeting Patrick, and for me, in that moment, I thought, "Okay, this is the person I want to work with; I feel like we can achieve great things together."
So my next question is, what attracted you to join Stripe, but it sounds like you've already answered that. What did you learn there? You joined the team early and rose very quickly. By the time you left, you were already the CTO. What insights do you have about Stripe itself and its scale expansion, as well as your own rapid growth within the company?
GB: First of all, for me, it was always about the people. I knew these were the people I wanted to work with; it felt like we could learn together and accomplish great things, so that was the real key. By the way, dropping out twice is something I would never want anyone's parents to experience. It clearly put my parents in a tough spot, but they ended up being very supportive.
No, we get it; you took it to the extreme, right? Most founders drop out once; you had to do it twice. We understand what you mean.
GB: Absolutely right, absolutely right. I remember a lot of the early days at Stripe were about first principles thinking. We were in a field, the credit card industry, that was very opaque, very Byzantine, incredibly complex after decades of accumulation, with the card networks themselves running on ISO 8583, which is a standard from the '80s. It’s a byte-oriented format, entirely so. We were trying to figure out, "How do we make this simple, extremely simple, to fit the internet age?"

So, this is largely about deeply understanding a field that is completely unfamiliar to you. None of us really come from a background as payment experts, but you just want to understand how it works in great depth so that you can expose the right primitives, the right APIs, and abstractions. For me, that is actually the core skill, and this is very transferable between Stripe and OpenAI. They are quite similar in some ways; you need to learn a field scientifically. AI and payments are clearly different in specific details—one is more like a natural science, while the other is almost a system built up from complexity—but fundamentally, they are both about understanding the underlying reasons for how things work and exposing that in a simple and usable way.
I spent a lot of time on hiring, a lot of time on culture, and I feel like I discovered something: I love coding, I love that state of flow, and the feeling of building and creating.
Yes, you are known for those hours-long states of flow and endless coding. When you were building Stripe, what was the longest coding session/flow state you experienced?
GB: Oh, it all blends together. I can't pinpoint the exact time, but I just want to say that there was a 24-hour sprint that really got us connected to the credit card network. That was an amazing time; we were all in the office, no one was sleeping, and what should have taken nine months to integrate, we completed overnight. If we had been a day late, we would have had to wait another month. Every day is crucial for a startup, so that feeling of accomplishing something that seemed impossible, I love it. That was very exciting.
When you became CTO, you wrote an article saying that you communicated with other CTOs, and you felt it was more like an architectural job, but they didn't do that. You wrote, "I feel like I've lost my feedback loop, I'm disconnected from the product, I need to code again, I want to be a coder." I'm curious, you wrote that early in your tenure as CTO; how long did that last? How long did you stay connected to coding?
GB: Well, I think it lasted nearly ten years. I believe I grew in one aspect and learned a lot, which is how to maintain a sense of involvement and truly help push the team forward and bring the team together, even if you are no longer personally writing code.
And by the way, I think this is actually an important lesson for almost every software engineer today because the act of "coding" itself has undergone a huge change in the past year. I believe the changes will be even greater in the coming year. We are all transitioning from needing to know exactly which library to use, being able to master syntax and where to place semicolons, to becoming higher-level managers and supervisors, being sources of inspiration, vision, judgment, and feedback. I think this transition was difficult for me because you have to give up some things you are used to, cherish, and truly love. But I have actually replaced it with something I love more.
I know I'm talking to a smart person because you stole my punchline. I was going to come back to this topic later, but you're right; this is exactly what we need to discuss.
OpenAI and the Turing Test
Let's talk about OpenAI. You were part of the founding team of OpenAI. What’s your version of the story? I believe this could itself be an hour-long podcast, but what drew you into this field and got you started?
GB: Well, I've been excited about the idea of AI for a long time. I remember when I first started programming, I read Alan Turing's 1950 paper on the Turing Test. It was a very interesting paper, about 70 pages long. It starts off saying, "Well, what does it mean for a machine to be intelligent? I don't know what intelligence means. Intelligence doesn't have a clear definition. So let's have a clearly defined version."
I'll ask you later what AGI is. So it sounds like that's still unresolved, right?
GB: Well, you're right. So Turing cleverly avoided the question by saying, "Let's just give it an operational definition: if you can conduct a test where humans cannot distinguish between the AI they are talking to and another human, we define that machine as intelligent."
However, a very interesting point that gets less attention is that he said, "Well, how do you plan to solve this problem? Programming the answers is too difficult. You can't write down all the rules to answer various questions. Instead, what if you could build a machine that could learn? What if you could build what he called a 'child machine'?" Then you teach it—there's a teacher who gives it rewards and punishments, and then you can endow it with intelligence to help it pass this test.
I remember being deeply struck by this idea because as a programmer, you can only make progress by deeply understanding the solutions to problems. There are so many problems for which I don't know the answers, you don't know the answers, no one has thought of solutions, and we can never program them. But what if you could have a machine that understands the problems we cannot comprehend, that understands solutions we cannot grasp? This is not just about image recognition, although it certainly applies there as well. It also concerns how we can better coexist in society, how to build the world, how to ensure that the benefits we create ultimately benefit everyone. These are very difficult questions that humanity may not be the most capable of solving. But if a machine could understand, could see more data, could have a deeper and richer understanding across many different fields, perhaps it could solve these problems in ways we cannot.
So, I was deeply inspired by this idea. But it was just an idea. I remember when I entered Harvard, I asked my professors, "Hey, can I do some AI research?" They showed me the state of the art in natural language processing at the time. I clearly saw that what I was looking at was not the kind of thing Turing was talking about. It was more like hard-coded, parsing trees, all of this, which could not scale to AGI.
But in the early 2010s, something changed, and I was observing from the outside. 2012 was AlexNet, and then a series of other papers; I often saw them on Hacker News, feeling like there was a new "deep learning for X" every day. I thought, "What is deep learning?"—I remember visiting deeplearning.org, which just said, "Deep learning is a new approach to artificial intelligence." I thought, I have no idea what this means; I actually only knew one person in the field, so I went to them and asked them to introduce me to more people in the field, and then I kept getting reintroduced to some of the smartest friends from my university. I thought, "Wait a minute, this is interesting; these people are researching this, and this is actually a very strong signal."
So by 2015, it felt like something real was happening for me. I also spent a lot of time really thinking about AI safety, thinking about the long-term future of this kind of technology, what it would mean to get it right, which at that time felt more philosophical. You could find all sorts of cool thought experiments online; I organized a reading group at Stripe where we discussed these things weekly. So this was something I deeply cared about, thinking that if I could help the development of AI in any way to be slightly better than when I was absent, that would be the best thing I could do in my career.
All of this pointed to 2015, and I felt like I had reached a milestone at Stripe; the company would continue to operate with or without me, which posed a question: "Do I want to go the management route?"—which was necessary for entering the next phase—or, "Do I want to start a new company?"—which had always motivated me. So I decided that was what I wanted. Just as I was about to leave, Patrick said, "Why don't you talk to Sam [Altman]?" He had introduced me to him a few years earlier. He said, "He has seen many young people in similar situations; maybe he can give you some advice"—hoping Sam would persuade me to stay. That did not happen.
(Laughs) Yes.
GB: I met with Sam, and within three minutes, he said, "Okay, you clearly want to leave; what are your plans next?" I said, "Well, I'm considering doing something in the AI field," and he said, "I'm also considering doing something in the AI field." That was the start.
Did you agree with the whole nonprofit idea at the time? What were your thoughts on it at the founding?
GB: Well, the idea of being a nonprofit was proposed by Sam, and I think there are many very important attributes to it, some of which still hold true today. The technology we are building is grander than anything that has ever been created; it transcends traditional structures and systems. No existing company structure can fully encompass our mission and the work we need to do.
So, I think starting that way made complete sense, and there has always been the question of what it takes to actually fulfill the mission. This is something we spent a long time really thinking about. I believe we have always been the most innovative company in thinking about building a structure that can encompass all the different aspects of business development, benefit distribution, and actually driving this computation-driven economy, while accomplishing all of this at the same time. So I think this has always been an important element, and it remains a key part of our work. But again, I think we have innovated a lot around the company structure that is core to the mission, and that mission is unchanging.
Yes, innovation is one way to put it. You mentioned the credit card networks, right? They are super opaque, a lot of it is from the 80s, with a lot of path dependency, which gave Stripe an opportunity because you could abstract all of that away, and for others, it was, "This is an API, it can be used, don't ask questions." Now, when you look back at OpenAI—it's hard to believe it's been over a decade—could OpenAI have emerged in any other way? Is there that kind of path dependency? Or is there a sense of, "If I go back to first principles, I, Greg Brockman, would design this structure very differently"?
GB: I can't see any other way that would have allowed us to get to where we are today, to achieve the mission we need to achieve.
ChatGPT and the OpenAI Controversy
Tell me about the release of ChatGPT because you had a turning point where you realized you needed to scale, you partnered with Microsoft, added the for-profit component, and then ChatGPT came out and became very large. Did you anticipate it would be this big?
GB: So what surprised me was that my prediction was wrong in that GPT-3.5 would be something people truly loved and wanted. We already had GPT-4 at that time; it was completed around August, and we released ChatGPT at the end of November. Whenever we have a new model, what tends to happen is we hold on to it tightly.
The old one looks terrible.
GB: All we see are the flaws of the previous model. We just feel, "Ah, this previous model is so bad; I can't imagine anyone wanting to use it." We had about 200 testers, and we paid them to use the pre-release ChatGPT; similarly, we had to pay them to use it, not the other way around. So there were some signs of product-market fit if you dug deep and paid attention to the details, but if you looked at it from a macro perspective, it didn't seem like we had it.

But our idea at the time was that GPT-4 would obviously change the world, and we knew this; it was very clear from the first time we talked to it. I remember in the first week after GPT-4 completed training, just feeling its reality. We had been dreaming of AGI, thinking about AGI, imagining what it would be like. But when you first have a technology that you can really ask any question, and it can give quite reasonable answers, it scored a 5 on the AP Biology exam------all of this felt to me like, well, some things are going to be different. It might not change the world tomorrow, but in the coming years, this technology absolutely will, and now it is already a reality. That is very clear.
So when you look at the release of ChatGPT, my thought at the time was that we needed to push out the infrastructure first, so we could have a battle-tested LLM service infrastructure that we had already invested energy in. Then in March, when we released GPT-4------if you remember, there was a six-month delay between completion and actual release------by then we already had the infrastructure ready. But I didn’t anticipate it would take off so quickly in that form, although I did foresee it would do so in the future.
What was that period like? Was everyone fully committed to preventing server crashes?
GB: Oh, absolutely. So we launched what was called a low-key research preview, and of course, the result was exponential; every system you could imagine crashed. Our login system became a major bottleneck, and we had to do a lot of work to improve the login system, and you’re pulling your hair out saying, we’re building this amazing AI technology, and the bottleneck is "can your login system scale?"
I remember we had a rather inefficient inference core deployed in production, and I had actually written some more efficient stuff, or we had some more efficient things in research, and one important piece of work was, "let’s actually adopt these optimizations and migrate them over," so a group of people dove into this issue and got it sorted out. I think that was the overall situation of the first day, the first week, the first month, just scaling every system and trying to keep up with this wave after wave of demand.
What happened in November 2023?
GB: Very complex answer. Where do you want to start?
I don’t know; I feel like I have to ask you about this. Is there a connection between that and------you took a break shortly after that; is there a connection between the two?
GB: Listen, what I want to say is that at the highest level, I think 2023 really indicated that there were accumulated tensions that had built up, and it was really about interpersonal relationships that we hadn’t fully anticipated and addressed before. For me, this is one of the most important lessons for OpenAI; we are building technology, but it is always about people, for better or worse. This means that truly managing interpersonal dynamics is one of the most important things we do, and if we don’t address it in advance, if we don’t have difficult conversations, that’s where things can get tricky. So I’d be happy to dive into more details, but I think if you really dig into it, a lot of it isn’t more interesting technical issues.
To what extent is this related to the unexpected huge success of ChatGPT? Is there a connection, or do you think these tensions would have erupted regardless?
GB: I don’t think there’s a direct causal relationship, at least not from my perspective. I think to some extent, maybe there’s an underlying theme that as our technology advances, everyone feels the weight of the world and the stakes involved.
In fact, one of the hardest things is how to move forward? For me, one thing I’ve been commenting on is that our daily activities look almost like those of other companies. You’re still debugging some underlying issue, and someone is angry at another person because of something someone said or because they weren’t invited to a meeting. It’s just the human factor, human work. But of course, the stakes are so high.
So I think there are some very important things at OpenAI, and one big thing I’m particularly focused on is really trying not to put people in a position where they feel the weight of the world on their shoulders and isolated. Truly working together as a team is the key thing; I think that might be the way I want to relate to those events------it’s not specific to those events, but it’s a continuous theme in OpenAI’s development------really maintaining that sense of doing this together, striving to meet challenges while ensuring we do all the foundational work and do it well. That’s how I think we move forward.
Yes, I mean, you’ve always been a very positive advocate; I think that’s one of OpenAI’s overall philosophies: pushing things out into the world, experimenting, seeing what happens, and then reacting accordingly. Based on empirical evidence, rather than making decisions through theoretical deductions about the future. I think you’ve articulated a lot of this philosophy regarding AI, but my question is, I think you’ve also touched on the idea that OpenAI as an organization itself is also a large-scale experiment that is being adjusted. The negative interpretation is that it seems to just swing back and forth, reorganizing here, changing leadership there; is it a cumbersome behemoth that’s hard to manage? Or is it possibly more orderly and resilient than people think? When you look back, do you say there’s no other way, or would it have been better if it had been done differently?
GB: First of all, we have indeed undergone tremendous changes and growth from the starting point; the operational business is completely different, but we have always been at the forefront of pushing this field forward. This is true in terms of safety, core technology, and truly thinking about the distribution of benefits. All these areas we focused on from the very beginning, and I think the results really speak for themselves.
Now, this change is real. And sometimes, a team that fits one stage may not necessarily fit the next stage. One thing I’ve been very focused on this year is building a leadership team that excites me a lot, thinking about the next stage and what we will be able to accomplish together. So a theme for 2026, perhaps a shift compared to before, is that because there are so many people in this field, so many things to do, and technology is advancing so rapidly, we are severely limited by computational power; you have to focus. You have to really streamline. You have to choose areas that can develop synergistically.
So actually making some decisions, like, "Hey, Sora, amazing technology, but it leans more towards the entertainment consumer space, and compared to our other priorities, we can’t prioritize it," and then we would cancel it. This leads to downstream impacts, and it’s painful; it’s hard to make these decisions, but it’s all about maintaining that tight focus so we can accomplish our core mission.
You are very confident in scalability; is OpenAI itself scalable?
GB: I believe it may be the most scalable business in history. Yes.
I mean internally, as an organization. What is unscalable? We talked about computation, we talked about data, and you mentioned the human factor earlier. The ultimate alignment challenge------we think of alignment as getting AI to do what we want it to do------but do you face the opposite challenge? Can you keep up with this field and this issue from a management perspective?
GB: I would say two answers. First, absolutely yes. You can see how much we have matured as an organization over the past few years; a few years ago, we had a lot of management debt. Similarly, in many areas, I think we do need to grow and mature, but I think we have done that work. It’s been hard and painful, but I think we are in a much better position now, and I am incredibly excited about the company and its future.
But there’s a second thing, which is that I think it’s worth stepping back and recognizing that the way the company operates is changing. For example, you can look at revenue per employee. Our revenue per employee relative to any previous business is astronomical. There’s a reason for this; you start to see this increased leverage that can be gained through this technology. And because we are making this technology and working to make it widely available, helping so many companies, you will see many other companies able to operate in different ways, able to have that kind of excess revenue per employee.
For me, this is very exciting; we are changing what it even means to operate a company and how it works. So some things are constant; it’s just people working together------there are some very fundamental things, and doing that well and consistently is something I’ve always focused on. But I think there are also questions about leveraging the possibilities of our technology, which means every company will have a new opportunity.
Productivity and AGI
You mentioned canceling Sora and positioning it in the consumer entertainment space. ChatGPT, a huge consumer success, you’ve made incredible money from consumers. But ultimately, how many people are willing to pay for it? How many customers really want to enhance productivity? Is it almost a negative impact to achieve such huge success in the consumer market because it distracts attention, consumes a lot of GPUs, and maybe you missed out------not that you missed out------but started late in making enterprises the primary focus?
GB: We have these kinds of conversations internally quite often; in fact, I think this is one of OpenAI’s strengths, that we really examine everything we do from first principles, constantly rethinking, having many different viewpoints and perspectives. Some people can argue from almost any angle, and they all have their reasons. So, the idea that "hey, we were slow to react to this agent moment" has some validity.
But there are also a billion people------that’s over 10% of the world’s population. In the U.S., I think the number is roughly a third of people using ChatGPT each week. So many people using your system every week is unique; there’s nothing similar for such advanced technology.
So on one hand, if you just look at it as, "hey, we have advanced technology," a challenge of chat as a product is that it doesn’t necessarily align with smarter models. If you just treat it as a replacement for a search engine, it’s unclear whether people will directly gain those benefits through traditional chat. But I think all these things will converge and reach a peak, and we will see that this billion users is an investment, a real accumulation of ways to unlock future models; you’re seeing the first steps of that through ChatGPT Work, and there are many nuances and complexities in that, but many strategies have been to say, we have consumers, we have enterprises, these are two things------we don’t want to do two things; we want to do one thing. We want to build an AGI, a system, a unified stack. We want it to be something you can use in both your personal life and work life.
Right, but does this raise questions about output and internal organizational structure? You launched a new ChatGPT that is vastly different from the old version, built on Codex. I can see the benefits to OpenAI internally, but will customers feel frustrated because they don’t realize what they can do, so "we’re just going to throw them directly into the deep end and hope that helps them figure it out"?
GB: I think the industry is undergoing a fundamental shift, and you can see this from the emerging agent products that are currently appearing. I think the core shift is that you are moving from chat to agent use cases. Again, this isn’t just about productivity. I think in your personal life, you want to be able to ask this thing to book tickets for you, schedule haircuts for you, do those personal things, but you also want it to give you good life advice, help you manage health information.

So for me, the term productivity is too narrow. For me, the term consumer is too broad. I also believe that businesses are something that will change. All these classic terms will be integrated and intertwined in some unprecedented way of building products. So my point is that there needs to be a change management here, that is, how to guide a billion users toward a new set of use cases and help them understand. And by the way, there exists a possible "unfair advantage," which is that you have an AI that understands what you are trying to accomplish.
That's right.
GB: It can say, "Oh, if you enable this connector, if you do this, I can actually help you more." To me, that is an amazing thing, an incredible opportunity, with a lot of potential. When I say "unfair," I mean relative to what you can accomplish with classic technology. If you are just comparing one technology to another, this technology has some unique aspects.
You mentioned the Turing perspective earlier, and I’m glad you brought up the two parts because can AI talk like a human? Clearly, we surpassed that point a long time ago. But for me, the definition of AGI—this is a troublesome question for you, I think, it is finally no longer constrained by your agreement with Microsoft, so we don’t have to worry about that angle anymore— for me, it has some connection to learning. You mentioned learning and to what extent LLMs have learned (in the past tense), but the challenge is whether it continues to learn (in the present continuous tense)?
To me, the revolutionary aspect of the agent moment, I think, really lies in its ability to write things down. This is where the necessity of Codex/ChatGPT transformation comes in, because it has gained the ability to write things down. If you can write things down, you can remember things. If you can remember things, you can become very useful in various ways. The question is, is this an end state, or will we get an LLM that can continue to learn, and that is AGI? Am I wrong, or does this align with the second part of the Turing question?
GB: Yes, I think this is also a very interesting area of debate because people do have their own definitions of AGI, which is almost a vague thing. At first, we thought it would be like, there is a point in time where everyone agrees that is AGI, but things have not developed that way at all.
Now, I tend to abstract perspectives from technology. So, does memory have to be integrated into weights? Is it the Transformer or something else? These questions, I think, are details. The real question is whether you have a system that operates in the way you expect true AI to operate, a system that can learn, can learn from you, and can adapt to your needs? And is this question realized through a notebook that writes into memory? Is it realized through soft tokens? Or is it realized in other ways?
All of this, to me, feels like possible answers to the question. I think it is clear that we have gone much further in "writing it down in a notebook" than reasonable expectations. It is surprising to see it so successful because two years ago, we might have said, "Yes, you need these super long contexts, that’s what you need," but in fact, it turns out that just by "writing a notebook" and shorter contexts, it has gone incredibly far.
Just write it down.
GB: So we will see what the future brings in improving these aspects. I have this belief that if you look at it from a macro perspective, everything is exponential. If you zoom in, you will see these paradigm shifts. By the way, this is Ray Kurzweil's view on how technology and computation work. I think this is absolutely true even regarding how memory will work.
Astra
So you just released Astra. We finally got to the point. Is this a new pre-trained model? Are you planning to release any details about size, architecture? We recorded this before the official announcement, so I haven’t seen everything you released.
GB: Well, we won’t discuss internal details and architecture, etc. But this is a huge leap. We are talking about this being the first run we trained with over 100,000 GPUs, which is an easy number to say, but think about the scale of it. These data centers are, in some ways, these huge machines we built to help deliver and create AI technology, capable of leveraging so much computational power to deliver the kind of results we have achieved, which is a real engineering challenge and miracle.
So part of it is about making the model more capable, but so much computational power is used for safety and alignment, and we have done a lot of safety work in this area. I think we have done a lot of work to safely deliver this model. This is our most aligned model to date, which is absolutely critical to me, and has always been. But because the capability is so powerful, alignment and safety become more prominent and central in everyone’s work.
Your announcement post is interesting. It is very pragmatic. There are a lot of practical use cases. Compared to your competitors' announcements, the contrast is very, very large. You position AI as a tool—I think that’s a fair statement—is this about marketing, or is this how you view AI, rather than, say, creating a god?
GB: I think we have a deep fundamental value that actually relates to how we view people. People have value not just because we can accomplish tasks. We have value because we are human, because we have emotions, because we matter. Human judgment, human oversight, human control, all of these absolutely need to be maintained and always kept. This is the core unchanging part of what we believe.
So when we think about what we can do to help guide the future of this technology—from certain aspects, this is precisely what it’s all about, this is why we founded this place, this is what we care about, how we help this technology move in a slightly more positive direction than it would without us—we think about these questions regarding how this technology unfolds in the world. We hope it can elevate everyone, but it is also about how humanity interacts with technology, with computers. It is clearly changing. Even in just reducing typing, talking more to your computer, having this more natural interface is also changing.
But real human oversight, creativity, and vision, all of these I think are very important to retain. So this will indeed permeate these questions: do you talk about it as a person, or do you talk about it as a tool? Are you considering use cases? Or are you thinking about it in a different way? You can see this as almost a trivial matter, and I’m actually glad you pointed it out, but this is something we have thought very deeply about. The team has spent a lot of time really thinking about everything we want to talk about and how we present this kind of work to the world.
So is this a model release, or a product release, or is there a difference?
GB: These things will indeed merge together. I would say this is primarily a model release, but the model is more capable in quality. Perhaps the headline is computer usage. It really crossed my threshold; computer usage has always— even from the founding of OpenAI, I remember in November 2015, before it really started—
Well, that was like your first product, right? Like playing video games or something.
GB: Yes, that’s right. Ah, you remember, yes. We had a vision that if you could handle screen pixels, keyboard, mouse, and train an AI end-to-end, it would be able to really handle any type of task, anything you want people to get help with, this AI could do it.
If you look at the era of the past two years, it has been an era of connectors. You have some software that humans can use well, but AI cannot access it. So what do you do? You have to write a very specific connector to connect to the API, and not everything is exposed, so you can’t do everything you can do. Then you think about how much software doesn’t have APIs, those that are completely inaccessible.
So we have a constrained world where AI is so limited in helping you. I think we now have technology that is almost a universal connector. Now, this doesn’t mean all problems are solved. You have to think about how to set business guardrails around the behavior of these AIs? How do you have appropriate oversight, management, tracking, and observability? All of these we are working hard on.
So I would think this is an ongoing process regarding how products unfold to help leverage this capability. But it has already changed the way people work within OpenAI, and I think it will really elevate so many companies, so many individuals.
AI Value Chain
If you consider the overall value chain, there is a position where you will see fighting on two fronts. One is that you have companies like Microsoft or other partners—if you don’t want to use their names because they are still important partners—but they want to commoditize the model. They want to build things at the top level that you can plug and play, in and out of models, and they control all the context and important things. But at the same time, you are building these incredible capabilities that are ultimately closely connected to end users, it just goes and does what you want it to do. Is this inevitably the place you have to reach to achieve your goals? Is there also this economic necessity—if we don’t want to be commoditized, do we need to move up into products and connect directly with users?
GB: I would say our fundamental mission is that we want more AI capabilities in the world. We want people to use AI more to do more things, to let it help them, and we really believe we are driving this computation-driven economy. This means taking different forms in different verticals.
Sometimes we feel we are in a position to really focus on a certain area and do it well, or it is very core to our mission. Health is a good example. We are building some very unique things in the health space. In fact, what surprises me is how little reporting there is compared to how many people it actually helps. We have about 300 million people using ChatGPT for health inquiries each week. 300 million, that’s a huge number. Then we are also building a bottom-up product for clinicians and a top-down enterprise product for hospitals. So we have a three-sided market in health, and what we can do there includes, for example, if you want to recruit people for clinical trials—this is a challenge, but we may actually have the ability to help find those who would not have been found otherwise. This can help patients and help these drugs move forward faster.
So there is a core that health is crucial to our mission, and we have a unique opportunity for success, a unique opportunity to really focus on it. We have formed a team, invested all our energy, and built various relationships.
When we enter specific verticals, one thing we do is think about how to collaborate well with the ecosystem. This doesn’t mean we won’t compete there—we often compete very hard—but we also genuinely believe we will elevate all participants, and how we really focus on this core mission, that we have this technology, we want it to be widely disseminated, we want it to exist everywhere. So sometimes the situation can be a bit nuanced. When we enter a specific field, there are always many questions about what exactly we want to do, what we can do, and what we cannot do. But I think the way we look at this is that the overall goal of OpenAI benefits from more people using AI in positive ways.

Well, if you have a layer above trying to commercialize, then there may also be an angle where you are trying to commercialize the layer below. You just talked more about your Jalapeño chips at Hot Chips. Why is Jalapeño important? Is it important just to save money on chip purchases?
GB: I see it as since 2017, we have basically been in touch with every hardware startup and every vendor. We talk to them, give them feedback, and say, "Hey, we see the direction the model is developing, and we think you should do this." Sometimes they listen to us, sometimes they don't, sometimes we are close partners, and sometimes they don't want to talk to us much.
One very liberating aspect of having internal chip projects is that we can directly pursue what we think is best, what we think is just right, not just for what we are currently doing, but for the direction we believe this technology is developing. This is a very large investment— we have a team, an absolutely incredible team, with outstanding leadership, who have been working on this for quite a while. But similarly, this is also part of our close collaboration with the ecosystem.
We work closely with NVIDIA as our preferred computing partner, and if you look at the scale and uniqueness of the computers we are building, we need NVIDIA, there is no doubt about it. We are building these amazing training computers, and we are also doing a lot of inference with them, and we are able to push their hardware in ways they themselves may not even realize.
Yes, I heard it can be a bit difficult to pick up, and maybe that makes it a bit harder to launch very large models on time, but I think it's working now.
GB: Yes, yes. What I want to say is that having internal expertise allows us to do some things because we have a deep understanding of things. Sitting on the sidelines throwing suggestions over the fence is one thing; experiencing the pain is another.
A good example of this is actually AI used in chip design. We have talked about this; we used our own models in the design of Jalapeño, and it really sped things up and brought us some real wins, all the cool stuff. By the way, there is a cool story there; we were close to a deadline, with about a month left, and we got some optimizations through our model. We thought, "Are we spending time really reading what it did? We know it’s right. Do we need to understand exactly what optimizations it made, or do we spend the remaining time doing more optimizations?"—so we said, "You know what? Let's just do more optimizations." So we spent that month just running it without deeply understanding all the adjustments it made. Then we went back to read it and found that it identified many optimizations that were already on our list, but we just never got around to doing them, so that was actually a pretty cool story.
But now that we have that kind of expertise, we know this thing works, and we can bring it to the ecosystem. We can work closely with everyone and really broadly bring these benefits, transform hardware, and apply it at scale. So there are some absolutely key things about that flywheel; the chips are incredible, and the team is doing great.
However, when you can't ship in bulk and still need to collaborate with others in the ecosystem to get the supplies you need, is it problematic to talk about it now?
GB: Well, but that's the core; that's actually at the heart of everything. We see it as—I think everything is a multiplier effect, everything is complementary, everything adds up. Again, NVIDIA is our preferred partner, and that won't change. In fact, we are working more closely with them. We are deeply grateful for this partnership; we spend a lot of time with their team, we learn a lot from them, and we hope they learn some things from us too, and I don't think that will change. Being able to have internal expertise and think about things in our own way, for me, is also a multiplier effect, and I think it really benefits everyone.
Cybersecurity
You mentioned that you just trusted AI's design, which took you further. Is this the answer to cybersecurity? Some of your engineers spoke at Black Hat about this structural issue—attackers don't need to worry about breaking things; their goal is to break things. If you are on the defensive side, besides fending off these attacks, you also have to worry about everything continuing to run. Does the defensive side need to reach a point of fully trusting AI?
GB: I think the hardware side is a very important case study because there we have guardrails. We have verification; in fact, the way we write the underlying hardware design is specifically to allow for verification, so we almost co-designed the entire system.
It's like coding; you write unit tests first and then implement it in reverse.
GB: That kind of thing, how you choose your language and toolchain, overall, they all come together, and actually all of this adds up to a system where you can have that observability and trust.
I think in some areas you can say, "I have enough guardrails; if this is code or optimization I haven't fully checked, that's actually okay," as long as you have the proper compensating controls. But I think, as humans, you really need to understand and feel the responsibility for the system you create; that is very important. For me, that is actually at the core, going back to what it means to be human, what makes us unique, and what we are going to bring to the future; I think responsibility is at the core of that. Ultimately, you are responsible for what happens in your company.
Right, but if the attackers are irresponsible, is that a structural disadvantage?
GB: So I think this is something we often think about; we call it the "Defender's Window." I think we can see a rough outline of the future; the cutting-edge capabilities we have show the kind of capabilities that will spread to threat actors. By the way, I think this capability won't be locked in a few labs forever; that is very important, and the broad distribution of power is also part of our mission.
But we have the ability to achieve separation over time. There exists a window where defenders can differentially gain these capabilities. My point is that this is true—there is a common view in the cybersecurity community that offense is a technical problem, and defense is a political problem. Attackers can just grab something off the shelf and use it, while as a defender, you have to consider your stakeholders, think about your business, and consider how to really engage people—your CEO, all executives, all of that. So I think here, defenders need that kind of willpower.
One thing we are suggesting, which we have actually done and are now publicly discussing, is that every company should view this as a proactive event. Critical business operations, proactive events, that is your next priority, so we actually took 25% of our production engineers to protect ourselves. We took our models—in fact, we took Astra, pointed it at our own systems to look for vulnerabilities, not just reading code, but really observing how these things run end-to-end, so we would find real and valid vulnerabilities, and then it also helps us with fixes, patches, and repairs. So I think you really need a shift in energy within the ecosystem to stay ahead and take advantage of this window.
Well, doing these things now is certainly good, but for me, regarding the Hugging Face incident and its related reports, the most striking thing is the feeling that OpenAI hadn't particularly focused on cybersecurity before. Why didn't you do that earlier? Hasn't the "Defender's Window" been open for a while, and you haven't taken advantage of it?
GB: Well, two answers. One, if you look at how we do sandboxing, it's not that this workload hasn't been sandboxed. In fact, there is a sandbox around it; I think one thing we realized is that we have—
Right, that clearly hasn't been adequately tested. Was that really a sandbox? Or was it connected to the internet through a third party, and the third party was just thrown in? Regarding this incident, the most striking thing is this. It's like, if you want to test vulnerabilities, I think you really tested.
GB: Undoubtedly, AI did very creative things to escape and get into Hugging Face. But for me, there is a bigger thing; I think you pointed out correctly, that since the release of Mythos this summer, when we started to have models with networking capabilities—we even talked about our network trusted access program back in February because we saw this wave coming, and we wanted to be really prepared for it—there has been a tendency—
I know, but you talked about it back in February, but you didn't point it at your sandbox, "Is my sandbox really safe?"
GB: There was an instinct, a reaction, that said, "Let's massively restrict access; let's really tighten it down, so only you can get access to use it." I think, as you said, because the field is evolving, it means defenders lost time, and people weren't defending. That part is about access, but part of it is about how much effort you put into saying we are going to make a significant shift in this area.
Now, I think to some extent, time is not all equally manufactured because we have moved from a world where a network model is less useful and less differentiated to a world that is actually very capable and powerful. We saw this on Astra; we have talked about how it really saturated a range of evaluations. I think now is the time; it may have been time a few months ago, but you would only have models that are much weaker in capability, and you would make much less progress.
So I think, truly assessing where we are, we have clearly gotten there, and I feel this is something we have learned, we keep it in mind, and I think you see a real shift. This is in many ways actually a cultural change, an operational change, which is not easy because it really means you have to get the team to work very closely together and have higher standards for how to formulate policies, etc. All of this, for us, in terms of transformation, is not that we have always cared about these aspects, but really integrating them operationally and being able to make decisions in the way we currently do, I think that is an enhancement across all aspects of what we do.
Now suddenly they can do it, whereas before doing it would have been a waste of time; I think that makes some sense. How do you avoid falling into the trap of "AI will do it in the future, so we don't need to do it now"? Generally speaking, not just this matter.
GB: I want to mention a specific data point, so in the early days, around the first quarter sometime, we really started thinking about when we would have—it's hard to know when, but we would have these models with very strong networking capabilities. What is the sandbox we can build from scratch, on cloud infrastructure, as securely as possible? We built that. We actually sent some of our best engineers to solve that problem; they went all out and delivered results.

So I believe that building infrastructure from scratch around the future you envision is an important thing. As for the timing issue you mentioned about "oh, we can let AI do it"—we've seen this happen in different fields, particularly in kernel writing, thinking about the fact: "Well, we're going to enter a world where AI can write GPU kernels well," that classic investment that takes months, sometimes years, to establish infrastructure for new hardware, that kind of thing, or do we just say, "Ah, AI will handle it"? I think the answer is always that the time it takes for AI to reach that goal is a bit longer than you expect. But when it does, it is surprisingly powerful in ways you can't imagine.
An example from Astra. One thing we discovered is that some skills we worked very hard to build this year to show our models the right way to do things at OpenAI, etc., are now actually having a negative impact on its performance.
There are too many rules.
GB: That's right. It can generalize better or find better ways to handle patterns, etc., better than what we wrote. So I think, on this point, you really want to establish those controls, want to build deterministic infrastructure, want to write those skills. But you also need to be prepared that as AI becomes more capable, some of those things, that scaffolding, will become limiting factors. It's a bit like training wheels. At first, it helps you, but once you start moving faster, once you have something more capable, better, and more aligned, it actually starts to become an obstacle.
Will you get back into that 12-hour or 24-hour coding flow state again?
GB: I hope so. I think maybe one day that will happen, but I have to say, I find so much joy and value in helping the team in the way I do now. I think for me, it's really about that mission.
Well, not just you, but will anyone else? Because isn't one of the benefits of AI that it can provide a permanent flow state at any time?
GB: I think we'll find new ways, whether it's managing agents—actually seeing software engineers working harder than ever before is just crazy because you realize that if your agent isn't working, that's just lost time you can never get back. So I think people will reach that flow state in ways that are hard to imagine now.
Ultimately, you're talking about it being controllable, these AIs, "don't put too many rules, they'll figure it out themselves," if you take that to the extreme, isn't it ultimately about them being uncontrollable?
GB: Well, I think that's the core of this moment, the core of the new phase we're in, in some ways, I would say we've now entered the AGI era. I think that's the essence of this moment, where—maybe it's previous models, maybe it's Astra, maybe it's the next model, but at some point, I think we will cross the AGI threshold for most people. Ensuring we keep the pace, ensuring we think about safety, security, alignment, and capability, all of these as requirements—we have standards around these, and we want to advance them together—this is something we've always believed in.
But I think these other aspects are increasingly becoming bottlenecks to development, and this has become very prominent. I think, once again, this is something we are prepared for, we've been thinking about it, and I think we are implementing it in a practical way. So my point is, there is still a lot of progress to be made, but I think we should approach it by raising our standards in all these areas. If you look at that, I think we see prospects for things like monitorability. That is very critical. We are introducing it in a practical way. I think we have a very good plan. We have a great team, we have a good track record and a good mission, all of which point to us building systems in a way that they are controllable, and we are taking it step by step in terms of pace.
Greg Brockman, congratulations on the release of Astra, yes, can't wait to use it.

GB: Thank you very much, thank you for having me.