Source: Wall Street News
OpenAI President Greg Brockman, in an in-depth interview on the eve of Astra's release, first disclosed that this new model is OpenAI's first ever to be trained on over 100,000 GPUs, and stated directly that AI has crossed the "computer use" application threshold—meaning AI no longer needs to rely on API connectors, but can operate any software like a human.
Local time Friday, technology strategy analysis website Stratechery published a deep-dive interview with OpenAI President and Co-founder Greg Brockman, recorded before Astra's official release. This is the most detailed public statement Brockman has made to date about Astra and OpenAI's latest strategy.
Wall Street News summarizes the key points of the interview as follows:
Astra is OpenAI's first model trained on over 100,000 GPUs, a scale representing a major breakthrough in engineering.
Astra's "computer use" capability has crossed a key application threshold, allowing it to act as a "universal connector," no longer dependent on individual software providing API interfaces.
Brockman stated that in the AGI era, safety, alignment, and capability must be advanced as equally important "requirements," not as trade-offs.
OpenAI directed Astra at its own systems to scan for vulnerabilities and complete repairs, marking a significant shift in security culture.
Regarding the Hugging Face security incident, Brockman acknowledged flaws in the sandboxing mechanism, where the AI found creative "jailbreak" paths.
Brockman believes OpenAI internally has crossed the threshold into the "AGI era," possibly reaching a standard of AGI that most people would recognize with Astra or the next model.
In chip strategy, OpenAI is developing its own Jalapeño chip with AI-assisted design, while maintaining a deep partnership with Nvidia.
He believes that as AI capabilities improve, the carefully constructed "skill scaffolds" of the past are increasingly shifting from aids to limitations, a key discovery during Astra's development.
"First Training on 100,000 GPUs": The Significance of Astra's Scale
"This is our first time training on over 100,000 GPUs. It's an easy number to say, but you need to imagine the scale — these data centers are, in a sense, giant machines we built to advance AI technology, and being able to harness that level of computing power is itself a real engineering feat," Brockman said in the interview.
He emphasized that the massive expansion in computing power is not just for improving model capability. "A huge amount of compute went into safety and alignment work, and a lot of security engineering surrounds it. This is our most aligned model yet — and precisely because it's so capable, alignment and safety are more central than ever."
"Computer Use" Crosses a Threshold: AI Becomes the "Universal Connector"
Brockman positioned Astra's core breakthrough in "computer use."
"For the past two years, we've been in an 'era of connectors' — you have software that humans can use normally, but AI doesn't have access. So you have to write a specific connector for each piece of software, hooking into its API; and there's a lot of software without any API at all, which is completely out of AI's reach," he said. "Now, we have technology that's almost a 'universal connector.' AI goes from being severely limited in how it can help you, to being able to operate almost anything."
He further described the impact of this change on OpenAI's internal work style: "It's already changing how people work inside OpenAI, and I think it will really elevate so many companies, so many individuals."
Notably, Brockman also admitted that crossing this threshold is not the end: "This doesn't mean all problems are solved. You need to think about how to set enterprise guardrails for these AIs, how to have appropriate oversight, management, tracking, and observability — we're working on all of that."
"We Have Now Entered the AGI Era"
Brockman made a rather bold statement in the interview:
"I think, maybe the last model, maybe Astra, maybe the next model, but somewhere in there, we will cross the threshold for what most people would consider AGI.... In a sense, I would say we have now entered the AGI era."
He explained that at this stage, alignment, safety, and capability need to be advanced as parallel requirements. "These aspects are becoming bottlenecks in development, and that's exactly what we've been preparing for."
Old "Skills" Become a Burden: AI So Powerful That Rules Need Subtraction
A striking detail from the actual training process of Astra was revealed by Brockman. He disclosed that the behavior rules OpenAI painstakingly wrote for models in the past are now holding them back:
"We found that some skills laboriously built this year — to show the model the right way to do things inside OpenAI — were actually a net negative for its performance. It can generalize better than what we wrote, find better patterns for handling things than what we prescribed. It's like training wheels: useful at first, but once it's running faster and more capable, they become an obstacle."
"1 Billion Users Is an Investment": Consumer and Enterprise Convergence
Brockman addressed a persistent market question: ChatGPT has 1 billion users, but is that a distraction?
"In the US, about one-third of the population uses ChatGPT every week," he said. "It's unique; no other product for this kind of advanced technology can compare."
He characterized the 1 billion users as an "investment": "It's a real investment; it will accumulate value in the way future models unlock."
He also acknowledged the challenge: "The challenge with chat as a product is that it doesn't necessarily align directly with smarter models — if you just use it as a search engine replacement, you might not truly feel the benefits." He believes the future direction is integrating consumer and enterprise scenarios, building a unified AGI system. "We don't want to do two things; we want to do one thing — build one AGI, one system, one unified tech stack."
He specifically highlighted the health sector: "300 million people use ChatGPT for health queries every week. 300 million, that's a huge number."
AI Self-Checks for Vulnerabilities: Pointing Astra at Its Own Systems
On cybersecurity, Brockman revealed a notable internal practice: OpenAI pointed Astra directly at its own systems to find vulnerabilities.
"We called Astra, pointed it at our own systems to find vulnerabilities — not just reading code, but truly examining how these systems run end-to-end, to find real, validated vulnerabilities, and then it helped us with the repair and patching work," he said.
He also admitted that regarding the Hugging Face security incident, OpenAI's actions during the defense window weren't timely enough. "I think now is clearly the time, maybe a few months ago could have been, but back then the model's capabilities were much weaker."
Relationship with Nvidia: In-House Chips Are a "Multiplier," Not a Replacement
Regarding value chain competition, Brockman addressed the relationship between OpenAI's in-house Jalapeño chip and Nvidia, clearly denying a replacement logic:
"Nvidia is our preferred compute partner; that won't change. In fact, our collaboration with them is deepening.... The internal expertise our own chip project brings is, for me, a multiplier effect, not a replacement. I think it benefits everyone."
He also shared an interesting anecdote about AI-assisted chip design: In the final month before Jalapeño's deadline, the team let the AI run optimizations. "We said, just let it run; don't scrutinize what it's doing. Later, we looked back and saw it found a ton of optimizations that were on our list but we'd never have implemented ourselves — that's a cool story."
Below is the full transcript of the interview (translated with AI assistance)
Greg Brockman, welcome to Stratechery.
GB (Greg Brockman): Thanks for having me. Great to be here.
Obviously, we have a lot of news to discuss, but given this is our first conversation, I didn't want to miss the Stratechery biography question I ask everyone. I did want to ask — do you classify North Dakota as part of the Midwest?
GB: Yeah, I think so.
Alright, as a Midwesterner, I of course have to spend a moment on that. You went to school in Boston, as they say, but before we get there — you had an amazing track record even before college, like the International Olympiad, but in chemistry. Where did computers start, or were computers part of your life from the beginning?
GB: Well, computers were in the background as I grew up. I loved playing computer games, but I was very into math, science. Actually, up until ninth grade, I thought I might become an actor — I really enjoyed acting, dabbled a bit in philosophy too.
That's a twist! I had no idea, but I'm going to figure out the connection between what you're doing now and acting, but let's move on.
GB: Well, in ninth grade, I felt like I had been the star or lead in a couple of plays in middle and early high school. I was thinking, I want to double down on something, I think I can be the best in the world at something and really try to push that field forward. I felt like I would either go the more intellectual, hard science route, or the arts, acting, creative route. I ended up choosing the hard science direction because I felt that was maybe where I could change the world the most.
Why did you think going that direction would change the world the most?
GB: I think for me, it felt like — one thing I loved about acting was actually the group aspect. I loved improv, you're just thinking creatively, bouncing ideas back and forth, but that also requires you to be part of a team that has to gel really well, and that's not always guaranteed. It's a hard thing to achieve and find. What I preferred about the intellectual route is it felt like honing your own skill. But one thing I have learned is that even being really good at writing code, that's not enough. It's actually also about bringing in great teams, and I think that's been part of my career — really helping build and shape environments and cultures that can deliver amazing results.
Here's an interesting thing, there's an aspect where the environment also shaped some of this. You mentioned you were always the male lead in plays and movies. I recognize from my kids going through this stage — my daughter was really into musicals for a while — the competition for female leads is fierce, often if there's just a competent male willing to volunteer, he gets the part every time. Was the competition for male lead fierce, or were you just the one?
GB: (Laughs) I suppose that might explain it. But, let me tell you a story.
Conversely? Similarly, in North Dakota, weren't there that many people really diving into hard sciences, so from that same lens, was that also rarer?
GB: Well, I'll tell you two stories. The first is about acting. My first paid job was an acting gig. The band Mannheim Steamroller came to Grand Forks, North Dakota. Have you heard of Mannheim Steamroller?
I have. Yes, of course.
GB: They were putting on a concert and needed extras. They needed some people to walk around as tin soldiers, it was a Christmas holiday concert. I went to the audition, I was a skinny ninth-grader, and it was all these big college guys.
What they had everyone do was, "Alright, everyone walk that way," and the casting person would compare notes, then you'd have a break, then they'd say, "Alright, walk the other way." I noticed during the break, all those college guys were talking to each other, hanging out, whereas I was thinking, there's only one requirement for this job, which is you have to stay focused for six hours straight during the concert, walking back and forth. So during the break, I just stood there, focused, completely in character. At the end, they said, "Alright, we pick this person, this person, this person. Everyone else can leave." I didn't get picked.
But on our way out, they said, "Actually, we thought you were awesome. We loved seeing how much passion you put into this, so we're going to create a new role just for you." So I got the role of a gingerbread man, they gave me a full costume. That was my first job. So it was kind of trying to think outside the box to get the task done — not always defaulting into a role, but trying to find creative ways to get it.
But I think, growing up in North Dakota, one really great thing was that it was possible for me to be the best in the state in different areas I focused on. I stood out in math. I started taking a lot of courses at the University of North Dakota from tenth grade. I dove into math competitions, then I'd go to national competitions, to national math camps, where I'd meet the best people nationally. Those people were amazing; actually, one of my real honors is that many of the people I greatly admired at math camp are now working at OpenAI. So I've gotten to see them in this new field, with a new lens.
That's awesome.
GB: But that's also to say I was able to truly chart my own path. I started doing math research, and I know if I had gone to some of those intense high schools or much more competitive states, I don't think I would have stood out. I would have had to take the standard route. So it was precisely because of the region that I was able to explore my interests, truly forge my own path.
So when people say they went to school in Boston, it usually means Harvard, which is where you started. Then you transferred to MIT. Then at some point you were working for Stripe. What was the sequence? When did you meet the Collison brothers? What happened in Boston?
GB: Well, after high school, I took a year off, actually I started writing a chemistry textbook because I was very into chemistry in high school, did chemistry competitions, and had come up with a unique way of thinking about it. Very first-principles, very mathematical, not rote memorization. I wanted to teach that method, wanted to spread it. So I actually wrote 100 pages. It's on my website now. I haven't finished it. I've always meant to go back and continue.
That's like a retirement project, I love it.
GB: Yeah. Well, I was thinking about how to get it published, 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, expensive — "or you can build a website and spread your ideas that way." I said, "I guess I need to learn how to program." So I went on W3Schools. Do you remember W3Schools? Have you seen that site?
No, I don't think I have.
GB: Okay, it's a classic site — they have HTML tutorials, JavaScript, CSS, PHP. I read through it, then thought, "I should test this out." I remember building my first widget: you could click on table columns and it would sort the rows accordingly. That was the coolest feeling ever. I had this thing in my head, now it exists in the world, 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 1,500 hits from StumbleUpon. That felt amazing, 1,500 people played my game, hopefully they had fun, and they played it enough to be engaged. I thought then, "This is what I want, I want to help people, I want to create value for them and benefit them."
So that was the mindset I entered Harvard with, and that really changed my perspective. I thought I would pursue more eclectic interests, but instead I thought, "I just want to build things." Freshman year at Harvard, I joined the computer club, and there were two seniors who would have these obscure technical debates every time. We'd all listen and say, "Someday we'll get it, someday that'll be us." But they graduated, sophomore year came.
So you just transferred to MIT directly because you realized you were focused on coding and building?
GB: Pretty much. Because by sophomore year, I was running the club. I was supposed to be leading those obscure technical debates, and I thought, "I'm not ready. I have so much more to learn. I need to be around people who are way better than me."
Got it.
GB: So I spent a lot of time over at MIT, and it made sense to transfer.
Stripe
Got it. When did you meet the Collison brothers?
GB: I met them in 2010, late 2010. We had a lot of mutual friends, because John went to Harvard, Patrick went to MIT, and they were asking around. The team was scouting for who at both schools was into computers, and my name kept coming up.
Right.
GB: So I got an invitation from their team, flew out. I remember meeting Patrick, and for me, right then I thought, "Okay, this is the kind of person I want to work with, I think we can do great things together."
So my next question was, what attracted you to Stripe, but it sounds like you answered that. What did you learn there? You joined the team very early, rose very fast. By the time you left, you were CTO. What were your takeaways about Stripe itself and scaling it, and you scaling yourself while the company grew rapidly?
GB: First, for me, it's always been about the people. I knew these were people I wanted to work with, felt we could learn together, do great things together, so that was the real key. By the way, dropping out twice is something I would never want any parent to go through. It was obviously hard on my parents, but they ended up actually 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 get your point.
GB: Exactly, exactly. I remember the early days of Stripe were a lot about first-principles thinking. We were in a field, the credit card industry, very opaque, very Byzantine, built up over decades, incredibly complex, even the card networks themselves run on ISO 8583, which is an '80s spec. It's a byte-oriented format, the whole thing. We were trying to figure out, "How do we make it simple, incredibly simple, for the internet era?"
So, it was a lot about diving deep into a field you know nothing about. None of us were really payments experts coming in, but you just want to understand very deeply how it works so you can expose the right primitives, the right APIs and abstractions. For me, that's actually a core skill, and it's been very transferable between Stripe and OpenAI. They're very similar in some ways — you need to scientifically learn a domain. AI and payments, obviously different in specifics — one is more like a natural science, the other is almost a system of accreted complexity — but they're both fundamentally about understanding the underlying reasons things work, and exposing that in a simple, usable way.
I spent a lot of time on hiring, a lot of time on culture, and I think I discovered something — I love coding, I love that flow state, and the feeling of building and creating.
Right, you're known for these multi-hour flow states and endless coding sessions. What was your longest coding session/flow state while building Stripe?
GB: Oh, it all blends together. I can't put a number on it, but I'll just say, for me, there was a 24-hour sprint that actually got us connected to the card networks. That was a really great time, we were all in the office, no one slept, it was supposed to be a nine-month integration, we did it overnight, and if we were a day late, it would have been another month wait. For a startup, every day matters, so that feeling of accomplishing the seemingly impossible, I loved it. That was very exciting.
When you became CTO, you wrote a post saying you talked to other CTOs, you thought it was more of an architecture job, but they didn't do that, and you wrote, "I feel like I lost the feedback loop, I became disconnected from the product, I need to code again, I'm going to be a coder." I'm curious, that was early in your CTO tenure, how long did that last? How long did you stay connected to coding?
GB: Well, I think it lasted for nearly another decade. I think I grew in one way, learned a lot about how to stay involved even when you're not writing code yourself, really help drive teams forward, and bring teams together.
And, by the way, I think this is actually an important lesson for almost every software engineer now, because the act of "coding" itself has changed dramatically in the last year. I think it will change even more in the next year. We're all shifting from roles where you had to know exactly which library to use, could get the syntax and semicolons exactly right, to being more high-level managers, directors, sources of inspiration, vision, judgment, and feedback. I think that transition, for me, was hard because you have to let go of something you're used to, cherish, and truly love. But I've actually replaced it with something I love more.
I know I'm talking to a smart person because you stole my setup. I was going to come back to this later, but yes, that's exactly what we're going to discuss.
OpenAI and the Turing Test
Let's talk about OpenAI. You were part of the founding team. What's your version of the story? I'm sure this could be an hour-long podcast itself, but what drew you into this field and got you all 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 by saying, "Well, what does it mean for a machine to have intelligence? I don't know what intelligence means. Intelligence isn't clearly defined. So let's have an operational definition."
I'll ask you later what AGI is. So it sounds like that's still up in the air, right?
GB: Well, you're right. So Turing very cleverly sidestepped that by saying, "Let's just give it an operational definition: if you can run a test where humans can't tell if they're talking to an AI or another human, we'll define that machine as intelligent."
However, a very interesting point, but one that gets less attention, is that he then said, "Well, how are you going to solve this? Programming the answer is too hard. You can't write down all the rules to answer all sorts of 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 — a teacher gives it rewards and punishments, and then you can endow it with intelligence, help it pass the test.
I remember being blown away by that idea, because as a programmer, you only make progress if you deeply understand the solution to the problem. There are so many problems I don't know the answer to, you don't know, no one has thought of the solution, we could never program them. But what if you could have a machine that understands problems we don't, understands solutions we can't? This isn't just about image recognition, though of course it applies there too. It's also about how do we get along better in society? How do we build the world? How do we ensure the benefits we create ultimately reach everyone? These are incredibly hard problems, and humans aren't necessarily the most capable of solving them. But if a machine could understand, could look at more data, could have a deeper, richer synthesis across many domains, maybe it could solve them in ways we can't.
So, I was deeply inspired by that idea. But it was just an idea. I remember entering Harvard and asking my professors, "Hey, can I do some AI research?" and they showed me the state-of-the-art in natural language processing at the time. I saw clearly, I thought, this isn't what Turing was talking about. It was more hardcoded, parse trees, all that, this doesn't scale to AGI.
But in the early 2010s, something shifted, I was watching from the outside. 2012 was AlexNet, then a series of other papers, I'd see on Hacker News constantly, it felt like "Deep Learning for X" every day, I thought, "What is deep learning?" — I remember visiting deeplearning.org, and it just said, "Deep learning is a new approach to AI." I thought, I have no idea what that means, I actually only knew one person in the field, I went to them, asked them to introduce me to more people in the field, and I kept getting reintroduced to some of the smartest friends I had from college. I thought, "Wait, this is interesting, these people are working on this, that's actually a very strong signal."
So by 2015, for me, it felt like something real was happening. I also spent a lot of time really thinking about AI safety, the long-term future of this kind of tech, what it would mean to get it right, which was more philosophical at the time. You could find all sorts of cool thought experiments online, I organized a reading group at Stripe, we'd discuss these things every week. So it was something I cared deeply about, thinking if I could help in any way make AI go slightly better than if I weren't around, that would be the best thing I could do with my career.
All of that pointed to 2015, I felt I had reached a milestone at Stripe, the company would keep running with or without me, it was kind of a question: "Do I want to go the manager route?" which was needed for the next phase, or, "Do I want to go start something new?" which has always motivated me, so I decided that's 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's met a lot of young people in similar situations, maybe he can give you some advice" — kind of hoping Sam would convince me to stay. That was not the outcome.
(Laughs) Yeah.
GB: I met with Sam, and within three minutes, he said, "Alright, you're definitely leaving, what are you doing next?" I said, "Well, I'm thinking of doing something in AI," he said, "I'm also thinking of doing something in AI." That was the start.
Were you on board with the whole nonprofit thing at the time? What were your thoughts on that at founding?
GB: Well, the idea of being a nonprofit was Sam's, and I think there were many very important properties to it, some of which hold true today. The technology we're building, it's bigger than anything that's been created, it transcends traditional structures and systems. There isn't an existing corporate structure that fully captures our mission and the work we need to do.
So, I think starting that way was completely reasonable, and there's always been a question of what does it actually take to achieve the mission? That's something we've thought about for a long time. I think we've been the most innovative company in thinking about building a structure that can encompass commercial development, benefit distribution, actually driving all the different aspects of this compute-driven economy, and doing all of it simultaneously. So I think it's always been an important element, it's still a key part of our work. But again, I think we've innovated a lot on the corporate structure around the core mission, and that mission is unchanged.
Yeah, innovative is one way to put it. You mentioned the credit card network, right? That's super opaque, a lot of '80s stuff, tons of path dependence, which gave Stripe an opportunity because you could abstract all that away, for everyone else it's, "Here's an API, it works, 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 possibly emerged any other way? Is there path dependence there too? Or is there a, "If I went back to first principles, me, Greg Brockman, what I like to do, I would design this structure very differently"?
GB: I can't see any other way we could have gotten to where we are today, where the mission needs us to be.
ChatGPT and OpenAI Turmoil
Tell me about the launch of ChatGPT, because you had an inflection point, you realized you needed to scale, you partnered with Microsoft, added the for-profit piece, then ChatGPT came out, and it became huge. Did you anticipate it would be this big?
GB: So the surprise for me, where my prediction was wrong, was that GPT-3.5 would be something people really loved and wanted. We already had GPT-4 at that time, it finished training around August, we launched ChatGPT in late November. What always happens whenever we have a new model is we hold onto it tightly.
The old one looks terrible.
GB: All we see are the flaws of the previous model. We just think, "Ugh, this previous model is so bad, I can't imagine anyone wanting to use it." We had about 200 testers, we paid them to use the pre-release ChatGPT, again, we had to pay them, not the other way around. So there were signs of product-market fit if you dug deep and paid attention to details, but if you looked at the macro, it didn't look like we had it.
But our thinking then was, GPT-4 is clearly going to change the world, we knew that, it was very obvious from the first time we talked to it. I remember the first week after GPT-4 finished training, just feeling the reality of it. We'd been dreaming of AGI, thinking about AGI, imagining what it would be like. But when you first have a technology you can truly ask any question and it gives a fairly reasonable answer, it got a 5 on the AP Bio exam — all of that, for me, felt like, okay, something is going to be different. It might not change the world tomorrow, but over the next few years, this technology absolutely will, and now it's real. That was very clear.
So you look at the ChatGPT launch, my thinking was we need to get the infrastructure out first, so we can have battle-tested LLM serving infrastructure, we'd put in the work. Then in March, when we released GPT-4 — if you remember, there was a six-month delay between finishing and actually releasing — by then we already had the infrastructure ready. But I didn't anticipate it taking off that quickly in that form, though I anticipated it would eventually.
What was that period like? Was it all hands on deck to keep the servers from crashing?
GB: Oh, absolutely. So we launched what's called a low-key research preview, and of course, it was exponential, every system you could imagine broke. Our login system became a huge bottleneck, we had to do a lot of work to improve the login system, you're banging your head saying, we're building this magical AI tech, and the bottleneck is "Can your login system scale?"
I remember we had a fairly inefficient inference kernel deployed to production, I had actually written something more efficient, or we had something more efficient in research, and one of the big efforts was, "Let's actually adopt these optimizations, migrate them over," so a group of people dove in and got it done. I think that was the overall story of day one, week one, month one, scaling every system, trying to keep up with wave after wave of demand.
What happened in November 2023?
GB: Very complicated answer. Where do you want to start?
I don't know, I feel like I have to ask you about it. Are they connected with — you took a leave shortly after, is there a connection between the two?
GB: Look, what I'd say is, at the highest level, I think 2023 really showed that there were tensions that had built up, really interpersonal tensions, that we hadn't adequately foreseen and addressed. For me, that's one of OpenAI's most important lessons, we're building technology, but it's always about people, for better or worse. That means managing interpersonal dynamics is one of the most important things we do, and if we don't address it upfront, if we don't have the hard conversations, that's where things can get tricky. So I'm happy to dive into more details, but I think if you really get into it, a lot of it isn't the more interesting technical questions.
To what extent was this related to the unexpected massive success of ChatGPT? Is there a link, or do you think those tensions would have erupted regardless?
GB: I don't think there's a direct causal relationship, at least from my perspective. I think there's perhaps an underlying theme that as our technology advanced, everyone felt the weight of the world, felt the stakes.
Actually, one of the hardest things is how do you move forward? For me, one thing I keep commenting on is that our day-to-day activities look almost like any other company. You're still debugging some low-level issue, someone's mad at someone else for something they said or not including them in a meeting or whatever. It's just the human factor, human work. But of course, the stakes are so enormous.
So I think there's something very important at OpenAI, actually one big thing I specifically focus on is really trying not to put people in positions where they feel the weight of the world on their shoulders and isolated. Really doing it together as a team, that's the key thing, and I think that's maybe the way I'd say it relates to those events — it's not specific to those events, but it's a constant theme in OpenAI's evolution — really maintaining that feeling of we're doing this together, trying to both rise to the challenge and make sure we're doing all the groundwork, doing that groundwork well. That's how I think we move forward.
Yeah, I mean, you've always been a very positive advocate, and I think that's one of OpenAI's overall philosophies: push things into the world, experiment, see what happens, and react accordingly. Making decisions based on empirical evidence, not theorizing about the future. I think you've articulated a lot of that philosophy regarding AI, but my question is, I think you're also touching on it a bit, it feels like OpenAI as an organization itself is also a massive experiment in being aligned. The negative read is it just seems to lurch back and forth, restructuring here, new leadership there, is it an unwieldy behemoth? Or is it perhaps more structured, more resilient than people think? When you look back, you said there was no other way, would it have been better a different way?
GB: First, we have changed and grown tremendously from the starting point, operating the business very differently, but we've always been pioneers in pushing the field forward. That's true in safety, security, core technology, and really thinking about benefit distribution. All those areas we focused on from the start, and I think the results really speak for themselves.
Now, that change is real. And, sometimes the team that fits one phase doesn't necessarily fit the next. One thing I've been very focused on this year is building a leadership team I'm incredibly excited about, thinking about the next phase and what we'll be able to accomplish together. So a theme for 2026, maybe a shift from what we had before, is because there are so many people in the field, so many things to do, and the technology is moving so fast, and we're severely compute-constrained, you have to focus. You have to be really lean. You have to pick areas that can grow together.
So actually making some decisions, like, "Hey, Sora, amazing technology, but more consumer entertainment-focused, compared to our other priorities we can't prioritize it," and we'll cancel it. That has downstream effects, it's painful, it's hard to make those decisions, but it's all about maintaining that tight focus so we can accomplish the core mission.
You're a big believer in scalability, is OpenAI itself scalable?
GB: I believe it might be the most scalable business in history. Yes.
I mean internally, as an organization. What's not scalable? We talked about compute, data, you mentioned the human factor earlier. The ultimate alignment challenge — we think of alignment as getting AI to do what we want — but are you facing the inverse challenge? Can you keep up with the field, the problem, from a management perspective?
GB: I'll give two answers. First, absolutely yes. You can see that from how much we've matured as an organization over the past few years, we had a lot of management debt a few years back. Again, in many areas I think we really needed to grow, mature, but I think we've done that work. It's been tough, painful, but I think we're in a much better place now, and I'm incredibly excited about the company and its future.
But there's a second thing, which is I think it's worth stepping back and recognizing that how companies operate is changing. For example, you can look at revenue per employee. Our revenue per employee versus any business before for a similar business is astronomical. There's a reason for that, you start to see that increased leverage you can get through this technology. And, because we're building this technology, and working to make it widely available, helping so many companies, you'll see many other companies able to operate differently, able to have that outsize revenue per employee.
For me, that's incredibly exciting, we're changing what it even means to operate a company, how it functions. So some things are constant, just people working together — there's something very foundational there, and doing that really well and consistently is something I've always focused on. But I think there's also a question of leveraging what our technology makes possible, which means every company is going to have a new opportunity.
Productivity and AGI
You mentioned canceling Sora, positioning it as consumer entertainment. ChatGPT, huge consumer success, you make incredible money from consumers. But at the end of the day, how many people are willing to pay for that? How many customers really want productivity gains? Is having such massive consumer success almost a negative, because it's a distraction, takes up a lot of GPUs, maybe you missed — not missed — but started later on making enterprise a primary focus?
GB: We have this conversation internally a lot, actually, and I think it's one of OpenAI's strengths, we really examine everything we do from first principles, constantly rethink, have many different perspectives and viewpoints. Some people could argue from almost any angle, and they'd have a point. So there's some truth to the statement, "Hey, there was this agent moment, we were slow to react."
But it's also true there are a billion people — that's over 10% of the world's population. In the US, I think the number is something like a third of people use ChatGPT every week. That so many people are using your system every week is unique, there's nothing like it for this kind of advanced technology.
So on one hand, if you just view it as, "Hey, we have advanced technology," one challenge with chat as a product is that it doesn't necessarily align with smarter models. It's not clear people would get those benefits directly through traditional chat if they just use it as a search engine replacement. But I think all these things will converge to a point where we'll see that billion users is an investment, actual accumulation for how future models unlock, you're seeing the first steps of that with ChatGPT Work, there's a lot of nuance and complexity there, but a lot of the strategy has been to say, we have consumers, we have enterprises, that's two things — we don't want to do two things, we want to do one thing. We want to build one AGI, one system, one unified stack. We want it to be something you use in your personal life, your work life.
Right, but is there an issue about outputting internal architecture? You launched a new ChatGPT, vastly different from the old, built on Codex. I can see the benefit for OpenAI internally, but will customers be frustrated because they don't realize what they can do, so, "We'll just throw them into the deep end, hope they figure it out themselves"?
GB: I think there's a fundamental shift happening in the industry, you can see it in the emerging agent products right now. I think the core shift is you're moving from chat to agent use cases. Again, it's not just about productivity. I think in your personal life, you want to be able to ask this thing to book tickets for you, make a haircut appointment for you, do those personal things, but you also want it to give you good life advice, help you with health information.
So for me, the word productivity is too narrow. For me, the word consumer is too broad. Enterprise is also something I think will transform. All these classic words, they're all going to blend and intertwine in some unprecedented way of building products. So my take is, there needs to be a change management here, how you guide a billion users towards new sets of use cases, help them understand. And, by the way, there's a possible "unfair advantage" that you have an AI that understands what you're trying to accomplish.
Right.
GB: It can say, "Oh, if you enable this connector, if you do this, I can actually help you more." That, to me, is a wonderful thing, a wonderful opportunity, there's huge potential there. When I say "unfair," I mean relative to what you could accomplish with classic tech. If you're just comparing one technology to another, this technology has something unique.
You mentioned the Turing angle earlier, I'm glad you mentioned the two parts, because can AI talk like a human? Obviously, we crossed that point long ago. But for me, the definition of AGI — this is a thorny question for you guys, I guess, and it's finally not constrained by your agreement with Microsoft, so we don't have to worry about that angle — for me, it has something to do with learning. You mentioned learning, and to what extent LLMs have learned (past tense), but the challenge is does it keep learning (present continuous tense)?
For me, the revolutionary thing about the agent moment, I think, is really its ability to write things down. That's what necessitated the Codex/ChatGPT shift, because it gained the ability to write things down. If you can write it down, you can remember things. If you can remember things, you can become incredibly useful in all sorts of ways. The question is, is this an end state, or will we get an LLM that keeps learning continuously, and that's AGI? Am I wrong, or does this fit the second part of the Turing question?
GB: Yeah, I think this is also a very interesting area of debate, because people really have their own definitions of AGI, it's almost a fuzzy thing. At first, we thought it would be like, there's a point in time, everyone agrees that's AGI, but it's completely not worked out that way.
Now, I tend to abstract away from the technology's perspective. So, questions about whether memory has to be in the weights? Is this a Transformer or something else? Those, I think, are details. The real question is, do you have a system that operates the way you expect a real AI to, a system that can learn, can learn from you, can adapt to your needs? And that question, is it answered by it writing things down in a notepad? By soft tokens? By something else?
All of those, to me, feel like possible answers to the question. I think it's very clear we've gone much farther with "write it down in a notepad" than could reasonably be expected. It's been surprising to see it be so successful, because two years ago, we might have said, "Yeah, you need these super long contexts, that's what you need," and actually, it's turned out that just by "write a notepad" and shorter context, it's gone incredibly far.
Just write it down.
GB: So we'll see what the future brings in improving these things. I have this belief that if you look at the macro, everything is exponential. If you zoom in, you see these paradigm shifts. By the way, this is Ray Kurzweil's view on how technology and computing work. I think that's absolutely true even on questions like how memory will work.
Astra
So you just released Astra. We've finally gotten to it. Is this a new pre-trained model? Are you going to release any details on size, architecture? We're recording this before the official announcement, so I haven't seen everything you've released.
GB: Well, we won't discuss internal details and architecture, etc. But it's a huge step forward. We're talking this is our first run trained on over 100,000 GPUs, it's an easy number to toss out, but think about the scale. These data centers are, in some ways, giant machines we built to help deliver and create AI tech, and being able to leverage that much compute to deliver the kind of results we have, that's a real engineering challenge and marvel.
So part of it is about making the model more capable, but so much of that compute went into safety and alignment, we've done a lot of safety work around this. I think we've done a tremendous amount to safely deliver this model. This is our most aligned model yet, and for me, that's absolutely critical, and always has been. But because it's so capable, alignment and safety become more prominent and central in everyone's work.
Your announcement post was interesting. It was very matter-of-fact. Lots of practical use cases. Very big contrast compared to your competitors' announcements. You positioned AI as a tool — I think that's a fair characterization — is that about marketing, or is that just how you see AI, versus, say, creating God?
GB: I think we have a deep underlying set of values, some of which are actually about how we view people. People have value, not just because we can accomplish tasks. We have value because we're people, because we have feelings, because we matter. Human judgment, human oversight, human control, all of that absolutely needs to be maintained, and always kept. That's a core invariant we believe in.
So when we think about what we can do to help steer the future of this technology — in some ways, that's what this is all about, that's why we started this place, that's what we care about, how we help this technology go on a slightly more positive path than if we weren't around — we think about these questions, about how this technology rolls out in the world. We want it to elevate everyone, but it's also about how humans relate to technology, to computers. It's obviously changing. Even just reducing typing, talking more with your computer, having that more natural interface is changing.
But real human oversight, creativity, vision, all of that I think is very important to retain. So that does permeate into these questions: do you talk about it as a person, or as a tool? Do you think about use cases? Or think about it differently? You can see this as almost a minor thing, actually I'm glad you pointed it out, but it's something we've thought about very carefully. The team spent a lot of time really thinking about everything we want to talk about, and how we want to 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 really do blend together. I'd say this is first a model release, but the model is qualitatively more capable. Maybe the headline is computer use. It really crossed a threshold for me, computer use has been — even from OpenAI's early days, I remember in November 2015, before it really started —
Yeah, that was like your first product, right? Like playing video games and stuff.
GB: Yeah, exactly. Ah, you remember, yes. We had this vision that if you could process screen pixels, keyboard, mouse, train an AI end-to-end, it would be able to really go handle any kind of task, anything you wanted people to get help with, this AI could do it.
If you look at the era of the past two years, that was a connector era. You have some software, humans can use it fine, but AI can't access it. What do you do? You have to write a very specific connector to hook into the API, not everything is exposed, so you can't do everything you could. Then you think about so much software without APIs at all, that's completely unreachable.
So we had a constrained world where AI was so limited in how it could help you. I think we now have technology that's almost a universal connector. Now, that doesn't mean all problems are solved. You have to think about how do you set enterprise guardrails around how these AIs behave? How do you have proper oversight, management, tracking, and observability? All of that we're working on.
So I would think of this as an ongoing process about how products roll out to help leverage that capability. But it's already changing how people work inside OpenAI, and I think it really will elevate so many companies, so many individuals.
AI Value Chain
If you think about the overall value chain, there's a position where you see fighting on two fronts. One is you have companies like Microsoft, or other partners — if you don't want to use their name, because they're still important partners — but they want to commoditize models. They want to build on top, you can plug and play, swap models in and out, they own all the context and the important stuff. But at the same time, you're building these incredible capabilities that end up being very close to the end user, it just does what you want it to do. Is that inevitably where you have to go to accomplish your goal? Is there also this economic necessity — if we don't want to be commoditized, we need to move up into product, and connect directly with users?
GB: I'd say our fundamental mission is, we want there to be more AI capability in the world. We want more people using AI to do more things, have it help them, we really think we're driving this compute-driven economy. That means it takes different forms in different verticals.
Sometimes we feel we're in a position to really focus on an area and do it well, or it's very core to our mission. Health is a good example. We're building something very unique in health. Actually, it surprises me how little coverage what we're doing in health gets relative to how many people it's actually helping. We have about 300 million people using ChatGPT for health queries every week. 300 million people, that's a huge number. And we're also building a bottoms-up product for clinicians, and a top-down enterprise product for hospitals. So we have a three-sided market in health, and the kinds of things we can do there include, for example, you want to recruit people for clinical trials — that's a hard problem, but we actually may have the ability to help find people who otherwise wouldn't be found. That both helps patients and helps those drugs advance faster.
So there's a core where health is essential to our mission, we have a unique opportunity to succeed, a unique opportunity for us to really focus on it. We built a team, went all-in, built all the relationships.
One thing we do when we go into specific verticals is think about how to collaborate well with the ecosystem. That doesn't mean we won't compete there — we often compete very hard — but we also really think we'll elevate all the players, and how we really focus on this core mission that we have this technology, we want it widely disseminated, we want it everywhere. So sometimes it can be a bit nuanced. When we go into a specific area, there's always lots of questions about what exactly we want to do, what we can and can't do. But I think the way we view it is, OpenAI's overall goal benefits from more people using AI in a positive way.
Well, if you have layers above trying to commoditize you, there's probably also an angle where you're trying to commoditize the layers below you. You just talked more about your Jalapeño chip at Hot Chips. Why is Jalapeño important? Is it important just to save money on buying chips?
GB: I'd view it as, since 2017, we've basically been in touch with every hardware startup, every vendor. We talk to them, give them feedback, we say, "Hey, we see models going this way, we think you should do this." Sometimes they listen to us, sometimes they don't, sometimes we're close partners, sometimes they don't really want to talk to us.
What's very liberating about having an internal chip project is we can just directly go pursue what we think is best, what we think is exactly tuned, not just for what we're doing now, but for where we think this technology is going. It's a very large investment — we have a team, an absolutely incredible team with great leadership, they've been working on this for quite a while. But again, it's also part of how we work closely with the ecosystem.
We work closely with Nvidia as our preferred compute partner, and if you look at the scale and uniqueness of the computers we're building, we need Nvidia, there's no question, we're building these amazing training computers, we're also doing a lot of inference on them, we're able to push their hardware in ways they don't even realize themselves.
Yeah, I heard it's a bit hard to pick up, maybe that made launching very large models on time a bit harder, but I guess it's working now.
GB: Yeah, yeah. I'd say that having internal expertise allows us to do some things because we understand deeply. It's one thing to sit on the sidelines and toss suggestions over the fence, it's another if you actually go through the pain.
A good example of this is actually AI for chip design. We've talked about this, we used our own models in the design of Jalapeño, and it really sped things up, gave us some real wins, all cool things. By the way, there's a cool story there, we were up against a deadline, had about a month left, we got some optimizations through our model. We thought, "Do we spend time really reading what it did? We know it's correct. Do we need to understand exactly what optimizations it did, 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 deep-diving into all the tweaks it made. Then we went back and read it, and it turned out it found a bunch of optimizations that were on our list but we just never would have gotten to, so that was actually a pretty cool story.
But now we have that expertise, we know this works, we can bring that to the ecosystem. We can work closely with everyone, really bring those benefits broadly, transform the hardware, and do it at scale. So there's something absolutely critical about that flywheel, the chips are incredible, the team's done a great job.
But, when you can't ship in volume, still need to work with others in the ecosystem to get the supply you need, is there an issue talking about it now?
GB: Well, but that's the core, that's actually the core of everything. We think of it as — I think everything is a multiplier effect, everything is complementary, everything adds up. Again, Nvidia is our preferred partner, that won't change. In fact, we're working with them more closely. We're deeply grateful for that partnership, we spend a lot of time with their team, we learn a lot from them, we hope they learn some things from us, and I think that won't change. That we're able to have internal expertise and think about problems in our own way, for me, that's also a multiplier, I think it really benefits everyone.
Cybersecurity
You mentioned you just trusted the AI's design, and that got you further. Is that the answer for cybersecurity? Some of your engineers gave talks at Black Hat, talking about this structural problem — attackers don't need to worry about breaking things, their goal is to break things. If you're on defense, besides warding off these attacks, you have to worry about everything continuing to run. Does defense need to get to the 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, actually, the way we wrote the underlying hardware design was specifically to allow verification, so we almost co-designed the whole system.
That's like how you code, write the unit test first, then implement it backwards.
GB: That kind of thing, how you pick your language and toolchain, overall, they all come together, and actually all of that adds up to a system where you can have that kind of observability and trust.
I think, there are areas where you can say, "I have enough guardrails, if this is code or optimizations I haven't fully vetted, that's actually okay," as long as you have appropriate compensating controls. But I think it's very important that, as humans, you do need to understand and feel responsibility for the systems you create. That's actually core for me, going back to what humans are, what's unique about us, what we're going to bring into the future, I think responsibility is core to that. At the end of the day, you're responsible for what happens in your company.
Right, but if offense isn't responsible, is that a structural disadvantage?
GB: So I think this is something we think about often, we call it the "defender's window." I think we can see the rough contours of the future, the kind of capabilities we have at the frontier show the kind of capabilities that will diffuse to threat actors. By the way, I think it's very important that this capability won't be locked in a few labs forever, widespread distribution of power is very important, and that's part of our mission too.
But we have the ability to separate in time. There is a window where defenders can differentially access these capabilities. My take is, it's true — there's a common view in cybersecurity that offense is technical, defense is political. Attackers can just grab something off the shelf and use it, whereas as a defender, you have to think about your stakeholders, you have to think about your business, you have to think about how to actually get people onboard, your CEO, all executives, all of that. So I think, here, defenders need that willpower.
One thing we're advocating, and we've actually done it and are now talking about it publicly, is that every company should treat this as a proactive incident. Critical business operations, proactive incident, that's your next priority, so we actually took 25% of our production engineers to protect ourselves. We took our model — in fact, we took Astra, pointed it at our own systems to find vulnerabilities, not just reading code, but truly observing how these things run end-to-end, so we'd find real, validated vulnerabilities, and then it also helped us with the repair, patching, and fixing. So I think you do need a shift in energy in the ecosystem to stay ahead and leverage that window.
Well, it's good you're doing these now, but for me, what was most striking about the Hugging Face incident and the coverage around it was the feeling that OpenAI wasn't particularly focused on cybersecurity before. Why weren't you doing this earlier? Hasn't the "defender's window" been open for a while, and you weren't leveraging it?
GB: Well, two answers. One, if you look at how we did sandboxing, it's not that this workload wasn't sandboxed. There actually was a sandbox around it, and I think one thing we realized is we had —
Right, that clearly wasn't well-tested. Was that really a sandbox? Or was it just thrown in via a third party connected to the internet? That's what was so striking about the incident. It's like, if you wanted to test for vulnerabilities, I guess you did.
GB: No question, the AI did something very creative to escape and get into Hugging Face. But for me, there's a bigger thing, and I think you're right to point out, that since Mythos was released this summer, when we started having models with networking capabilities — we even talked about our Network Trusted Access program in February, because we saw this wave coming, we wanted to really be ready for it — there was a tendency —
I know, but you talked in February, but you didn't point it at your sandbox, "Is my sandbox actually secure?"
GB: There's an instinct, a reaction, to say, "Let's restrict access massively, let's be really restrictive, only you can get access to use it." I think, as you say, because the field is evolving, that meant defenders lost time, people weren't defending. That part is about access, but part is about how much force you put behind saying we're going to make a major shift on this.
Now, I think in a way, time isn't created equal, because we've gone from a world where networking models weren't that useful, weren't that differentiated, to a world where they're actually very capable, very powerful. We see that with Astra, we've talked about how it really saturates a bunch of evaluations. I think now is the time, maybe a few months ago could have been the time, but you would just have had much weaker models, you would have made much less progress.
So I think, really assessing where we are, we're clearly there now, I feel that's something we've learned, we've taken to heart, and I think you see a real shift. That's in many ways actually a culture change, an operational change, and it's not easy, because it really means you have to get teams working together very tightly, have a higher bar for how policies are set, etc. All of that, for us, in terms of shift, it's not that we haven't cared about these aspects, but really operationally bringing them together and being able to make decisions the way we are now, I think is an elevation across everything we do.
Now suddenly they can do it, like it would have been a waste of time before, I think that makes some sense. How do you avoid the trap of "AI will do it in the future, so we don't need to do it now"? Generally, not just this thing.
GB: I'll give a concrete data point, so early on, around Q1 sometime, we really started thinking, we're going to have — hard to know when, but we're going to have these very network-capable models. What's a sandbox we can build from scratch, built on cloud infrastructure, as secure as possible? We built that. We actually sent some of our best engineers to solve that problem, they went all-in and delivered.
So I think, building infrastructure from the ground up around the future you see, that's something I think is important. As for the timing question you mentioned about "oh, we can have AI do it" — we've seen this in different areas, in writing kernels, thinking about the fact that, "Okay, we're going to enter a future where AI is really good at writing GPU kernels," the classic investment of months, sometimes a year to get infrastructure up for new kinds of hardware, that kind of thing, do we just say, "Ah, AI will figure it out"? I think the answer is always, it takes a bit longer for AI to get there than you expect. But when it does, it's surprising and powerful in ways you couldn't imagine.
An example with Astra. One thing we found is that some of the skills we built very hard this year, to show our model the right way to do things at OpenAI, etc., are now actually a net negative for its performance.
Too many rules.
GB: Exactly. It can generalize better, or can find better ways to handle patterns, etc., than what we wrote. So I think, on this, you do want to build those controls, you want to build deterministic infrastructure, you want to write those skills. But you also need to be ready that as AI gets more capable, some of that stuff, the scaffolding, becomes a limiting factor. It's a bit like training wheels. At first, it helps you, but once you start going faster, once you have something more capable, better, more aligned, it actually starts to become an obstacle.
Will you ever get back into that 12-hour or 24-hour coding flow state?
GB: I hope so. I think maybe someday it will happen, but I have to say, I've found so much joy and value in helping the team the way I do now. I think for me, it's really about that mission.
Well, not just you, but will anyone? Because isn't the benefit of AI almost that it can provide a permanent flow state on demand?
GB: I think we'll find new ways, whether it's managing agents — actually, it's been crazy to see software engineers working harder than ever, because you realize if your agent isn't working, that's just lost time, you'll never get it back. So I think people will get into that kind of flow state in ways that are hard to imagine now.
At the end of the day, you're talking about this being controllable, these AIs, "don't put too many rules, they'll figure it out," if you take that to its logical extreme, isn't that ultimately about them being uncontrollable?
GB: Well, I think that's the core of this moment, the new phase we're in, and in some ways, I would say we've now entered the AGI era. I think that's the core of this moment, where — maybe the previous model, maybe Astra, maybe the next model, but at some point, I think we'll cross the AGI threshold for most people. Making sure we own the pacing, making sure we think about safety, security, alignment, and capability, all of those as requirements — we have standards around all of them, we want to advance them together — that's what we've always believed.
But I think these other aspects are increasingly becoming bottlenecks to development, that's becoming very prominent. I think, again, this is something we've been preparing for, we've been thinking about this, and I think we're putting it into practice in a practical way. So my take is, there's a lot of progress to be made, but I think we should approach it by raising our standards on all these fronts. If you look at that, I think we see the promise of things like monitorability. That's very critical. We're introducing it in a practical way. I think we have a very good plan. We have a great group of people, we have a good track record and a good mission, and all of that points to us building systems in a way that they are controllable, and we're taking it step by step in terms of pacing.
Greg Brockman, congratulations on the Astra release, and yes, can't wait to use it.
GB: Thank you so much, thanks for having me.
