Chapter Timestamps
0:00 OpenAI’s Leadership Shake-Up
2:19 OpenAI’s Plan to Transform the Economy
4:45 Greg Brockman’s Expanding Role
6:01 Competing With Anthropic and the Rise of AI Agents
8:10 OpenAI’s Model Development Comeback
10:02 The Hugging Face Cybersecurity Incident
12:08 The Coming AI Cybersecurity Arms Race
15:10 Can AI Solve the Problems It Creates?
17:07 How Codex Secured Greg Brockman’s Website
18:44 OpenAI’s Response to the Hugging Face Incident
22:55 Will Safety Slow Down New Model Releases?
28:12 How AI Is Speeding Up Model Development
30:39 Astra Delays and OpenAI’s Priorities
31:36 Does OpenAI Have AGI?
34:54 How OpenAI Thinks About Alignment and Safety
37:45 Is AI Opening a Pandora’s Box?
42:02 The “Merge”: ChatGPT, Codex, and Personal AGI
47:07 How AI Could Transform Knowledge Work
51:04 Winning Over an AI-Skeptical Public
53:34 OpenAI’s Mission
Made with: The Transcript Desk Chrome Extension
Full video:
OpenAI President Greg Brockman speaks to TIME about artificial general intelligence, ChatGPT, AI safety, cybersecurity, and what comes next for the future of AI.
Transcript
00:01-00:41
Greg Brockman: A billion people use ChatGPT every single week. No one else in human-operated software is doing that. The response time is just too slow. I feel like this could be the most empowering technology ever created.
Interviewer: I think it would be helpful to start with the news this week. Denise Dresser, your CRO, is leaving the company, and you’re bringing in a new one. Brad Lightcap is leaving as well. There are a lot of public questions about OpenAI’s leadership bench and what’s going on. You’ve stepped in to manage a lot of this. I’d love to hear from you—what are all these departures connected to? Are they different? Is there any through line?
Greg Brockman: Look, the way I think about it is that the theme of this year has really been focus. We have line of sight to such incredible technology, but it’s not enough just to produce it. We need to figure out how to make this technology benefit people—how to bring it into every enterprise in the world in a way that uplifts what they’re doing and helps them accomplish their goals. That’s been the singular focus we’ve been pursuing. And in pursuit of that, we’ve been thinking: Do we have the right structure? The right strategy? All the right pieces in place?
I think my view of OpenAI is that we’re a very resilient organization. A lot of what we’ve been doing is building a leadership team that can help us achieve that goal. My view of everyone who’s been here—and those who are no longer here—is that they’ve all contributed to what OpenAI is today. Looking forward, we’re scaling systematically to every enterprise in the world. We’re bringing our models to consumers everywhere, and a lot of my focus has been making sure we can achieve that.
01:40-02:22
Interviewer: On the CRO side specifically, what is it that you weren’t getting that you needed? Is that the way to think about it?
Greg Brockman: I’d say it differently. Like, a year ago, we really weren’t in the game at all when it came to enterprise. It just wasn’t our focus. We hadn’t built the muscle we needed. Now, I think we’re actually in the game. We’ve built a foundation that we can really scale on. And now we’re systematically scaling. To me, that’s the moment. It’s not about the individuals—it’s about the overall direction of the field.
Interviewer: How does this systematic scaling change things practically for OpenAI?
Greg Brockman: Well, number one, we’re looking at transforming the entire economy. That’s the thing I think is hard to see from the outside. It might feel like, Hey, it’s a business—we’re just selling SaaS software. That’s not what we’re doing. We’re transforming the economy to be compute-powered. And I think what an enterprise even is will change.
The question is: Where do you start? How do you make sure you understand every segment and can sell in a repeatable way to all of them? A lot of the work we’ve been doing—this isn’t new—is scaling where the business is at. We need to make sure everyone in the field can funnel insights from customers back to our product and research teams, and vice versa.
For example, we learned years ago that producing blog posts internally has to be a collaboration. You can’t just hand off work from the research team to the comms team. It has to be deep collaboration. The technical expert needs to hold the pen and feel accountable for the end result, amplified by people who know how to craft a narrative, bring assets together, and execute a good launch.
03:46-03:57
Interviewer: So that’s the core activity you’re pursuing—unified focus. What about externally?
Greg Brockman: Externally, you should see us systematically prioritizing different industries, really focused on a deep understanding of the metrics—where token spend is growing the most—and focused on outcomes. Every business right now is asking, Where’s the ROI? That’s a big focus for us. We’ve always had the best price-performance models, but we really have to show people the outcomes.
We’re also scaling through channels and partners. I think where we’re at is: We feel very confident about the fundamentals. We have a great product, a great model, and we’re rapidly growing both our internal team and our ability to reach everyone in a way that has the positive effect we’re looking for.
04:50-06:01
Interviewer: You mentioned accountability. From your perspective, what are you accountable for now at OpenAI, given the leadership changes? Fiji Simo came in for a time but had to step back permanently due to health. You stepped in while she was out and took over a large part of the company. Now it’s you and Sam effectively running it together. How is that divided up?
Greg Brockman: For the past couple of years, I’ve been accountable for the infrastructure side—data centers, compute, those kinds of things. That part remains. There’s a large part of the scope I was also working on in terms of the software for training these big models, but I’ve been able to hand that off so I can focus on the business side.
I’d view myself as end-to-end accountable for revenue, product, marketing—the whole machine for bringing these models from research to value for our customers.
Interviewer: When you look at this machine—and I know you’re still putting pieces in place—how is it doing relative to the competition? Specifically, there’s reporting that Anthropic’s ARR is potentially more than double OpenAI’s. What do you make of that?
Greg Brockman: Well, I think we started late. Specifically, the way to think about the overall field is that for the past couple of years, the chat market was the main growing one. The reason was simple: GPT-4 was the first model where the results were personalized enough that it made sense for you to ask a question, read the answer, and find it interesting. It naturally lent itself to a chatbot. And we’ve been able to scale that—deliver value. Now, a billion people use ChatGPT every single week. That’s real scale. No one else is doing that.
But this year, something else happened: the agentic moment. This really started in December of last year, when both we and the competition had a breakthrough where agents were able to start doing software engineering in a significant way. It went from, You could speed up your work by maybe 20%, to 80% of the work could now be done with AI. Suddenly, the way software engineers operated inverted—they became much more like managers of agents. Since then, the growth in adoption of these agents has been unprecedented.
07:14-09:06
Interviewer: And one thing that happened was, I think, that you focused late on coding in particular—
Greg Brockman: Let me give you a different answer. There’s a simple narrative: we focused less on real-world coding. We came to that game late because we’d always looked at things like coding competition benchmarks—where we had the lead—but not at how a developer would actually use it. How do they work with a messy, real-world codebase? What happens when they interrupt the model mid-task? What’s the model’s “personality”? These little last-mile paper cuts actually make a huge difference in adoption. And similarly, on the go-to-market side, we weren’t as focused on building customer relationships, observability controls, or helping users understand what’s possible in their context. All of these were less of a singular focus for us until much later than the competition. So yes, we started late, but our trajectory has been very steep.
Greg Brockman: But there’s an even deeper answer that’s less obvious from the outside: model capability. For about two years—from when we announced GPT-4o in 2024 to December 2025—we were really squeezing the juice out of the GPT-4o model. We had a bunch of pieces we were improving, but fundamentally, we were working with the same underlying base model, the same pre-train. And there’s a reason for that: producing and shipping a new pre-train requires full integration across the company. It’s not just one team—it’s about integrating inference, post-training, and ensuring the model is co-designed to be the right size. You have to get all of these things exactly right in a harmonious way.
09:06-10:46
Greg Brockman: Over the course of 2025, a lot of our focus—my personal focus—was on how to get back to shipping new pre-trains as a core part of our innovation engine. And this year, you see the results. We’ve already shipped our third pre-train—we’re actually about to ship that. This is Astra. A lot of the gains you see are because we’ve made the model fundamentally better. The way to think about it is that there was an underlying technological area where we were lagging. A core part of our innovation machine wasn’t fully operational. But now it’s different.
Greg Brockman: And that’s why you see our model capabilities taking off so much. If you look at GPT-5.6 Sol—when we launched what was the best model in the market—you could see the revenue really starting to shift. We grew 20% in our topline revenue in July, and we see that exponential growth continuing.
Interviewer: Another thing that’s changed, as you were saying with model capability, is that the models are becoming incredibly capable at cyber and hacking. I think the world is still kind of processing the implications of the Hugging Face incident. Correct me if I’m wrong, but essentially, it seems like you all accidentally hacked another company without realizing it for a while. It was a model you were training and evaluating, and since then, there’s been a Black Hat video where you walked through this, and you’re going to put more out in the coming days. But walk me through: when you heard that this had happened, what did you think?
10:46-12:18
Greg Brockman: Well, in the words of someone on our security team, it was definitely the most eye-opening security event any of us were aware of. There are a lot of lessons we’re taking away from it, but I think it’s a moment where we, as a community, can learn and understand what the future looks like. The salient points for me right now are this: we had a model we were evaluating. It was in a sandbox that was well-configured, and it managed to both break out of the sandbox and break into production infrastructure that was intended to be hardened. The capability there is quite high. There are details about exactly how it worked—the fact that it wasn’t just a singular agent but a swarm of agents focused on this goal, able to pass messages to each other.
What you start to see is that with very capable agents, there are limitations that slow human attackers down that will no longer hold back AI attackers. They can sift through vast amounts of information, find many little vulnerabilities, and chain them together into something significant. What I immediately started thinking about was: what are the implications six months from now, when this kind of capability isn’t just in the lab or in a few places, but widely available? And I think the answer is that every company needs to respond—needs to upgrade its security—because attackers are going to fully automate. You can already see this with the release of GLM 5.3: the cyber capabilities of open-weight models available to everyone are going to increase.
12:18-14:41
Greg Brockman: But I actually think that in the long run, defenders can win. AI can give defenders an unfair advantage in a way that was never possible before. Security has always been a cat-and-mouse game—attackers level up, defenders level up. But AI can shift that balance. For example, formal verification of software: you can mathematically prove security properties. The technology for this exists, but it’s intractable for humans to do the work to prove those properties. And now we have these models that are incredible at math. Astra, our upcoming model—we announced this a few weeks ago—solved 10 math problems that had stumped mathematicians for decades, using just $2,000 worth of compute. Those same models can prove formal safety properties of software. That’s one place where defenders get an unfair advantage.
Greg Brockman: A second thing we’re doing is investing in generating secure code. As a human developer, there are so many patterns people just keep getting wrong. I’ve seen this in my previous jobs—I was very focused on building secure infrastructure, and you find all these little things people do wrong. SQL injections, for example—people know about them, but there are so many little anti-patterns. I remember in Ruby, there were certain characters in regular expressions that, if you got them wrong, left you vulnerable to common attacks—like a newline that doesn’t match, allowing someone to add things to the payload. All of those problems—little security anti-patterns you have to keep telling people, “Don’t do this, don’t do that”—you can train your AI to never do them. I think we actually have the ability to build the best security engineers in the world and give defenders the advantage.
14:41-15:30
Greg Brockman: But this is going to require a mindset shift. It’s going to require a mega-project. The whole ecosystem needs to change how software is secured, because there’s a second thing we’ve learned: as these cyber models have gotten more capable, the number of bugs they’re finding in critical software—like the Linux kernel—just keeps growing. It really makes you realize that human-written software is too slow to respond. Defenders need to automate, to move at machine speed, because their adversaries certainly will.
Interviewer: What do you say to people who already have a negative view of AI—maybe there’s a data center in their neighborhood they don’t like, or they hear you say, “We’re building these models, and now we’re going to sell you something to protect you from them.” What’s your response?
Greg Brockman: First, I have a lot of empathy for that reaction. If I rewind to 10 or 11 years ago, before we started OpenAI, I wasn’t even in AI. The reason I got into it was because I believed this could be the most empowering technology humanity has ever created. AI has the potential to help everyone live the life they want, build the future they envision, and achieve their goals. I just wanted to help steer it in a slightly more positive direction than it might have gone without me.
When you look at where OpenAI is today and what we’re doing, we can see a bit into the future. What we’re creating isn’t just about us—there’s a whole ecosystem, many actors, and that’s a good thing. We believe this technology should be democratized, not locked up with one entity controlling it, because that’s a different kind of risk. But it does mean we have to adapt and navigate carefully, using the insights we’ve painstakingly extracted from the future to understand its shape.
To me, the Hugging Face incident is a proof point—a case study, a window into the future. It’s not necessarily a future any one company is creating; it’s the overall economy, the set of actors across different countries building these models. It shows what will be possible and how attackers might operate. What it doesn’t show is how defenders can adapt. But that’s what we’re trying to demonstrate now: that we can build models you can use to scan and protect yourself.
17:07-17:46
Greg Brockman: For example, I asked Codex—just off-the-shelf Codex, no special access, no special models—to scan gregbrockman.com and tell me what vulnerabilities existed. Now, gregbrockman.com is a very simple static site. I thought, Come on, what’s it going to find? But it came back with 13 findings. For each one, I was like, Yeah, I kind of knew about that, but it’s so much work to fix. Or, Oh right, I forgot I left that open. Or, Here’s a legacy domain I’m not using anymore—I hadn’t set the records to prevent email spoofing. A bunch of little things like that.
Interviewer: That email spoofing one seems like a big deal.
Greg Brockman: Yeah, you’re right. Now that you mention it. Here’s the thing: fixing those issues properly requires setting up things like DMARC, and there’s this whole process you have to go through. It takes time. It takes expertise. So instead, I just asked Codex, Can you fix it?
It opened up my cloud control panel, clicked around, set headers, changed settings. I used to have Cloudflare in front with AWS on the backend, but I was going over insecure HTTP. Codex figured that out—I’m not even sure how it saw that, since I thought everything was hidden behind Cloudflare. It just said, You know what? I’m going to migrate you off AWS. So it set up Cloudflare Pages, moved the content over. I had an old, insecure version of jQuery—it just got rid of jQuery entirely.
It took 15 minutes to find all the issues and another hour to fix them. And for me, I was just the manager. I gave it the intent: I want to know where the problems are, and I want you to fix them. And it did. That, to me, is the future we’ve been promised.
18:44-19:07
Interviewer: I want to hear more about this—even though it’s more in the research world, I think it’s really relevant. You’ve talked about what you’re doing in response to the Hugging Face incident. One of your colleagues, Mia, told me the other day, I wish we’d done a lot of this work before this happened. So what is that work that’s happening now?
Greg Brockman: I’d frame the mode we’ve been in since discovering the Hugging Face incident as a full incident response to elevate every aspect of how we secure, monitor, and align our most capable models. We believe that models above a certain capability level require this.
It’s been a very interesting technical problem. And it’s not like we were caught flat-footed—we’ve anticipated this moment for a long time. If you look back to February, we started talking publicly about trusted access for cyber models. That was the first time we were really thinking, Okay, we’ve always had in our preparedness framework that when cyber capabilities reach a certain level, you need to treat them differently.
Now we’re starting to think about how transformative these models will be in the world. We need to consider how access is controlled internally. We’ve started thinking: If we have models capable of breaking out of sandboxes, we need to make sure our sandbox is as industrial-grade and foolproof as humanly possible. That means engineering from the ground up around security properties.
So we’ve been putting in place infrastructure and policy primitives for this moment. But it’s different when it actually hits. You never know when you’re going to reach that threshold. Here, we got to see this window in a very concrete way: Okay, we’re now at a serious capability level relative to these risks. We must act now.
20:50-23:05
Interviewer: So there’s been a workflow shift happening very rapidly?
Greg Brockman: Yes. A lot of our governance has been focused on model deployment, but not necessarily on model training. Now we need to pull governance back into the training and evaluation process. We need good governance there—standards around checks and balances. How do we know if a model is secure enough? How do we know if it’s aligned enough?
This alignment piece is actually very important because our belief is: if you make models extremely capable, you really want to ensure that as their capabilities grow, their alignment grows with them. You want those things to be in lockstep, coordinated. That way, you have this defense-in-depth approach: models that are trying to do the right thing, sandboxes that are as secure as possible.
We’re doing this iteratively—taking incremental jumps so we can understand as we go and ensure our systems are working. This aligns with how we’ve thought about iterative deployment and the creation of this technology from the very beginning.
We’re operationalizing this, and for me, as a manager thinking about the organizational system for producing this technology, it’s been a very important challenge. It means you now need different people in the room to make decisions. You need a different process for defining requirements, executing priorities, and propagating changes in technical understanding. How do you make sure the right people are in the right rooms?
But in some ways, it’s been a phase shift—one we’ve been preparing for. Earlier this year, I spent a lot of effort trying to bring research and security closer together.
Greg Brockman: We started doing things like weekly safety and security meetings for agents. Some of these initiatives are about building connective tissue—bringing together people who wouldn’t normally interact because they come from different backgrounds or focus on different problems. Forcing them to engage, to think about shared challenges. That shared empathy, shared understanding, and shared technical context—it’s already paying off.
Interviewer: At the same time, though, the implication seems to be that momentum on new models will slow significantly. Do you have a sense of the impact?
23:05-24:50
Greg Brockman: I’d frame this differently. I’ve always believed you have to think about progress holistically. If you only measure progress by benchmark scores or capability in a specific task, that’s too narrow. We’re building an integrated system, and safety and security are core to that. So if safety and security lag behind capability, yes, you absolutely need to improve them. But our goal is to build a system where safety and security are first-class requirements, developed in lockstep with capability. We’re upleveling every aspect of the machine.
Greg Brockman: Over the past year or two, we’ve upleveled how we build our clusters—now we can train at massive scale. We’ve integrated pre-training, RL, and all these pieces to work together seamlessly. We’re building the best sandbox, adding chain-of-thought monitoring. It’s about defining the scope of delivery: What does success look like? What does “good” look like? If you jettison some requirements, sure, you can move faster in other areas. But that doesn’t make sense. That’s not how engineering works.
24:50-26:37
Greg Brockman: Practically, yes, I hear you—this might lengthen release cadences. But I’m not certain it will. If you were just optimizing for a single metric, you could probably max that out faster than we will. But there’s a lot of flexibility. We always have a backlog of improvements queued up. Maybe we deprioritize certain capabilities—cyber, bio, coding—or delay architectural improvements. The rate of producing checkpoints doesn’t have to change; it’s just what those checkpoints represent that shifts. They might be less capable in specific areas but more intertwined with safety and security. I think of model releases as snapshots of where we are at a given moment. There’s no pre-ordained set of models we have to create. We can co-design this.
Interviewer: I guess what I’m getting at is that for a while, the release cadence felt predictable. But in the last year, it’s sped up—new models every six weeks, you and Anthropic constantly pushing out updates. What you’re describing—taking stock, adding friction—feels like it has to slow things down. Are you thinking about this from the perspective of the business you’re now accountable for, or is it too early to know?
26:37-28:52
Greg Brockman: I understand that perspective, and it’s reasonable. But my observation is that we co-design everything. The business operates by taking the most capable model we’re ready to release—after the appropriate safety and security controls—and bringing it to customers to help them transform their businesses and get value. The question is: Which models are ready? We produce many internally, and the bar for capability jumps, the frequency—all of that has to be co-designed. We have early access programs, like Daybreak Red and Blue, which let us test different models with partners. There’s definitely a shift in our internal balance of effort and where the checks and balances are. But I don’t think it’s as simple as saying the cadence will go from six weeks to six months. The content of those releases will just be different.
Interviewer: So maybe the raw frontier models won’t get pulled into production as quickly, even if they’re still being trained and worked on.
Greg Brockman: I’d think of it as an overall pacing of that frontier rollout. Another thing to consider: Our ability to produce these models—and to achieve capability gains—is also increasing. Our raw capacity has grown because we understand the technology far better than we did in the past. That’s actually beneficial for safety and alignment, too. We have a much deeper understanding of how these systems work. There are still gaps, but our ability to engineer and execute technically has improved dramatically. Problems that used to hold us back—like bugs in massive training runs—are now much easier to solve.
28:52-30:57
Greg Brockman: In the past, a training run would take months. We’d launch it from a code branch, adding fixes specific to that run while mainline development continued. Eventually, we’d need to merge the two, but testing all those fixes at scale was impossible—no human could review all the code. So we’d spend weeks debugging. Now, Codex just reads all the code, finds the bug, and says, “Here it is—no problem.” This amplification of what models can do knocks out entire classes of problems. Tasks that used to take three or four weeks now take a few hours. We’ve seen this happen repeatedly.
Greg Brockman: There were a lot of limitations in how we had to develop in the past that made everything much harder for us.
Greg Brockman: These challenges have made it much harder for us to move forward on problems that are important. But now we have the opportunity to tackle them in a far more efficient way. To me, that means using this extra capacity to improve alignment, to enhance safety. We need to ensure we can have models that monitor other models in a trustworthy way. Let’s implement some of the ideas we’ve been discussing since 2017—things like iterated amplification and debate, where models can spend significant compute to ensure their actions are reasonable. That’s the opportunity in front of us.
Interviewer: Do these changes you’re talking about impact the release of Astra—the thing that’s coming soon? Do you think it’s going to be delayed beyond what you originally hoped?
Greg Brockman: Well, it’s already later than we hoped in some ways. We’ve had a model that we think is very good—lots of people are using it internally—but it’s not out there yet.
30:57-32:54
Interviewer: So, does that mean the answer is yes, it’s delayed?
Greg Brockman: I think the answer is yes. But despite everything I just said, we’ve very much sacrificed short-term maximalism. If we’d just wanted to focus on revenue or something like that, that isn’t how we’ve thought about it at all. No one has even asked that question. We’ve been really clear: our priorities are, first, to do right by AGI, and second, to transform the economy. Revenue is absolutely a priority for us, but we know the relative prioritization. We know what wins if there’s a direct choice that has to be made.
Interviewer: Do you think you have AGI?
Greg Brockman: I think about this a lot. I talk to different people about where they believe we are. The best answer I’ve gotten from our team—and it really resonates with me—is that AGI is a gradual thing. It’s not a single moment in time.
Interviewer: You used to talk about it like a moment in time. The company definitely did.
Greg Brockman: That’s right. That’s what we used to think. When we set out, we almost thought of it as: nothing, and then we’ll put together the pieces, and then—boom—we’ll have AGI. But I think what we’ve learned over the years is that was a naive perspective. You can see this in how smooth the creation of these agents has been—the transformation of software engineering, knowledge work, cyber capabilities. All these things are gradually, increasingly improving. It’s exponential. If you look back at the different models released over time and their performance on benchmarks, it’s just this steady progression. That said, if we go forward two years and look back, I think we’ll say, “This is when AGI was created.” So I think we are going through that moment.
32:54-35:11
Interviewer: But it’s not something where you could just say it’s here now.
Greg Brockman: No, I don’t think it’s ever going to be as simple as saying, “This was the date.” There are so many axes to it. And I think one lesson in the whole AGI conversation—and why it’s so complicated—is that humans are capable of much more than a naive view would give them credit for. The diversity of what we do—how we set goals, accomplish tasks, how jobs work—it’s all extremely deep. What we’re seeing is that there are different aspects of how people accomplish tasks or overcome limitations. For example, reading all the code—these AIs are becoming able to fill in those gaps, to do more and more. But it turns out there are some things we still find extremely important for a human to be involved in.
Greg Brockman: Like, where does judgment come from? Where does taste come from? Who decides what’s good and what’s not? That’s something that should be uniquely human. Here’s an example: with image generation, if you just generate an image with a model and show it to someone, no one cares. It’s uninteresting. But if you take a photo of your family and say, “Turn this into a caricature, or a cartoon style,” people love it. It goes viral. There’s something about that element of humanity that matters deeply. People want to connect with other people. And in my mind, that’s the purpose of this technology. What’s emerging is the potential for deepening human connection—so you can spend more time with your family, more time with the people you want to be with, and less time typing away at a machine or dealing with carpal tunnel. I think AGI, as it’s starting to unfold, is almost like a fog lifting. It’s teaching us more about ourselves and what it means to be human.
Interviewer: Given the capabilities you have internally and the measures you’re taking now—moving compute to alignment, putting in all these checkpoints—how will you know when you’re in a good spot? When will you know you’ve solved the problems and the models are aligned?
35:11-37:45
Greg Brockman: The perspective here is critical, and it’s something we’ve spent years thinking about, debating, and refining. In my view, it comes down to iteration: iterative deployment, iterative development. You need checks and balances, evaluations, benchmarks, governance—a technically sound setup. Everything is gradual, incremental. You learn, you update, you iterate. As model capabilities increase, the tightness of those bounds has to increase correspondingly. Think of it like launching a rocket. If the rocket’s small, there’s a lot you can do wrong and it’s fine. But as the rocket gets bigger, the tolerance shrinks—from centimeters to millimeters. That’s where you have to be. And as you start to add more responsibility to what’s—
Greg Brockman: You have to add more responsibility to what the payloads are for the rocket, where it’s going, what it’s doing. Think about all the things relying on that rocket and how all the systems have to go right—and all the things that can go wrong. You have to continue to ratchet up your ability to measure and ensure that you can even do the engineering, in addition to the safety and security properties. And that’s very much how we think about it.
There are moments where we introduce very punctuated change. When we first created the preparedness framework, that was an example. But even with the preparedness framework, we very deliberately baked in the ability to change it. We said, “There will be new information. We will have to learn. We will have to update.” But at any point, we wanted it to be clear: What are our bars? How do we handle it when a model reaches a certain level of capability?
We recently talked about the fact that Astra is approaching cyber-critical. It’s not necessarily there yet because, to really prove it, the evaluations aren’t fully defined. But we’re saying, “Okay, it’s starting to approach it, so let’s treat it as if it is.” And that means we need to go and execute on certain safeguards and checks and balances. You’re seeing all of that play out.
So, the answer to your question is: We define a bar, we hit the bar. You have to take a framework that looks that way—one that really focuses on iteration. Are you learning? And are the bars you’re setting tightening in a sufficiently executed way in order to deliver the properties that are required?
37:45-40:03
Interviewer: Yeah. I mean, your rocket analogy is an interesting one because, um, rockets also explode a lot. And I’ve been thinking about your concern as these things get more capable—the concern of controlling them and how you can, and whether there’s kind of a Pandora’s box situation unfolding right now. I don’t know if you’ve thought about that personally recently, with the insights you have on what’s ahead of Astra and what’s causing you to make this slowdown. But yeah, do you worry about it?
Greg Brockman: We view it as our job to really think through the implications of this technology. That’s why we’re here—we want to help steer it in a more positive direction than would happen without us. That’s table stakes to our mission. That’s how we approach things.
Greg Brockman: I think it’s important to take two lenses to what you just said. The first is that this technology isn’t being created in a vacuum. There are many people, many actors, many companies, many organizations, even many countries pursuing its creation. We happen to be in a spot where we have some lead, some ability to see into that future.
And it’s important to think about how we develop responsibly, safely, and securely. That’s something we spend a lot of time on. You can see it very concretely in how we prioritize that in everything we do. Actually, I think mistakes we’ve made in the past include talking less about the fact that this is how we operate—that safety is such a first-class priority, a first-class requirement. We’ve been doing it, but not necessarily talking about it as publicly as we should. I think we’re starting to realize we need to help people understand that this is a core value of ours, something we care about so deeply.
I think we’ve always been the field leader in safety, security, and alignment, but I don’t think people have necessarily known that.
40:03-42:31
Interviewer: Yeah, even with the noise from Anthropic about how they’re the leader in that capacity. Do you think you guys have been?
Greg Brockman: I believe at a substantive level that we have been.
Interviewer: And do you think that what you guys are doing—and the capabilities you’re seeing and needing to redo, like the preparedness framework and all that—implies you’re actually just way ahead of where everyone thought you were, relative to Anthropic and the other frontier labs? Like, does this suggest that you all have a breakthrough internally that the world hasn’t seen?
Greg Brockman: Well, I can’t comment on competitive positioning, but I guess I would just say that in my mind, we’re very focused on the work that we’re doing. We think that we see smooth scaling, and I think the exponential continues. So it’s a little bit less about where we are relative to others and a little bit more about where we are relative to expectation—or what we think we will be able to accomplish.
Greg Brockman: The reason this matters is that the bar we hold ourselves to doesn’t really change depending on what the ecosystem is showing, because there’s so much noise out there. It’s hard to know what’s real. What we focus on is: Are we doing right by our own conscience? By our own standards? By what we believe is important? That is the number one lens that we take.
The second thing I wanted to add is that I think the recent “Pacing the Frontier” open letter we talked about supporting— the language in there is very deliberately chosen. “Pacing” is a very important word. If you think about what’s happening, what we’re saying is that you can’t take a one-dimensional view of what it means to deliver a model. If you just look at capability and don’t look at alignment, or the kind of safety properties, security properties of how it was trained, it’s too myopic.
At this level of capability, you need to think about those as first-class requirements. And in fact, they are the bottleneck, because those are the areas I think people have sort of focused on less in an integrated fashion. And that integration, by the way, is critical. I think where people go wrong is when they think you can separate out these properties—you can have just safety, you can have just capability. No, it’s one thing. We’re trying to deliver one system, one AGI, that uplifts you, that is helpful to you, that delivers on your goals, that you can steer, that is aligned with what you want. That’s what we are delivering. And so that flows back to all these questions of bars and testing and evaluations, safety, security. All of that is very, very critical.
Interviewer: I think I’ll pivot us a little bit because I want to hear you talk about the merge and what’s going on with that. So, we’ve been talking about existential risk, what’s happening at the frontier—which I think most people in the world don’t see, right? That’s happening inside companies like OpenAI. What people do see is how ChatGPT is evolving, how Codex is evolving. And one of your big priorities in this role, as I understand it, is what you guys call “the merge” internally—which is creating the super app, this combined ChatGPT and Codex.
42:34-42:59
Interviewer: You’ve combined ChatGPT and Codex, and now you have ChatGPT for Work with tabs. But it doesn’t feel like that’s the end state. As a user, I see some clunkiness in how it’s currently executed—things don’t always sync, and you can tell you’re still combining separate systems. Once you’re done with that work, what’s the end product you imagine for people?
Greg Brockman: A personal AGI.
Interviewer: Personal AGI. And how does that manifest as software?
Greg Brockman: Yeah, it’s a really exciting vision. When we started with ChatGPT, it was just a text box—you could ask it anything. But if you think about it, it was fundamentally limited. The only thing it could do was output text for you. The only impact it could have was inside your head. It couldn’t look at your calendar, your email—none of that.
Now, that’s changed. ChatGPT Work is essentially Codex harnessed in the cloud. We had ChatGPT as this Q&A system, where people went beyond quick questions to complex queries, and then we had Codex as an agentic system that could actually do work. It became obvious: ChatGPT needs to be agentic too.
At the start of the year, we thought, “Let’s have great chat and also make it do agent stuff.” But it turned out to be really hard to execute. We ended up with two separate agentic stacks, and we’d compare them—why does this work here but not there? It was slow and messy.
So we decided: Let’s take the agentic stack from Codex and bring it into ChatGPT. What you see with ChatGPT Work is an intermediate step. It’s funny—ChatGPT started as a text box, and now this text box is way more powerful.
44:28-46:26
Interviewer: But this text box is way more awesome.
Greg Brockman: Way more. Exactly. And where we want to be is simple: The user shouldn’t have to think about this. No tabs, no switching, no model pickers, no thinking levels. You should just have an AGI. You should be able to ask it anything—and in fact, you shouldn’t even have to ask. It should proactively say, “Hey, I noticed your favorite band is in town. Tickets just went on sale. I got you the best seats for you and your family—I only had 10 minutes because they were selling fast. Hope that’s okay.”
And if you’ve built enough trust with the agent—given it permission, set boundaries—it knows what’s within your budget, what needs approval. That’s the kind of system we want. Trust is key. Emotional intelligence matters too. If you’ve shared something personal, like a health condition, the AI should know whether it’s okay to discuss that with a parent but not a stranger.
Observability is critical—helping you build that trust so you know what the AI is doing on your behalf, that it’s pursuing your goals. This same underlying AGI stack, with its connectors and context, applies to enterprises too. We’re already seeing it transform software engineering this year. AGI isn’t some sudden leap—it’s a constant updating of expectations. We’ve all forgotten what software was like just a year ago.
Interviewer: That’s a very interesting thing within OpenAI—you’re at basically 100% adoption of ChatGPT Work across every function: sales, finance, everyone. And whenever you roll out a new model or the system breaks, people are so upset—they can’t do their work. And you’re like, “Six months ago, you didn’t have this, and you were fine. Why are you complaining?”
Greg Brockman: People adapt so quickly. They learn how to get the most out of these tools, and once they do, they can’t imagine going back. There’s something about this where I think we’ll have incredible tools—assistants that remove so much toil, so you can spend your time doing what you want: building a better life, relaxing, spending time with family, pursuing creativity. If you have a vision, you’ll have the tools to bring it into the world.
ChatGPT will help you accomplish what you want. If you want to make the world better, that’s what it’s for.
47:08-49:15
Interviewer: And the implications of that are that what coding AI tools have done for engineering, this could do for all knowledge work, all kinds of fields. That’s how you used to talk about AGI—AI that could do most economically viable work better than humans. What you’re describing now sounds like that.
Greg Brockman: I think the amplification we’ve seen in software engineering, we’ll see across every field. But there’s something worth thinking about: Software engineers aren’t using these tools to take more vacations—they’re working harder than ever.
Interviewer: Yeah. You guys work a lot of hours here.
Greg Brockman: Yes. And across the industry, people are talking about this—because now, if your agents aren’t running, it feels like wasted opportunity. I know that feeling. We’ve spent years with GPU clusters: You’re either running something on them or you’ve lost the opportunity. You can’t bank those hours for later. It’s use it or lose it.
And that’s the opportunity in front of all of us now: You’ll have almost an AGI—maybe soon a true AGI—in your pocket. It can do anything you want. The question is: What do you want? It’s waiting on your direction, your guidance, your management.
Interviewer: It would be nice if eventually we could get some time back from AI.
Greg Brockman: That’s right. Right now, it will help you work more. But the hope is that, over time, it gives you the freedom to choose.
Interviewer: That’s right. And I think that’s exactly the next step—moving from micromanaging an agent, where you have to provide it with tasks and context because it just doesn’t have it yet.
Greg Brockman: Exactly. You want something much more like a really good assistant—one that’s plugged in. You’ve given it context, you’ve built up trust with it. You have to say, “Okay, I trust you with this context. I trust you with this responsibility. I trust you with this job.” And once it’s there, it should be able to proactively run and execute. You wake up in the morning and say, “Hey, here are five things I need from you to unblock me. I just need approval on this. Here’s a question I have.” And as you’re drinking your morning coffee, you say yes, no, yes, no.
Interviewer: Then you come back late at night, and you’ve got five more things. I think it’s going to be a world that’s different from the kind of AGI many of us pictured five or ten years ago. It’s not just about integration—it’s about humans being maximally empowered with this tool, but you still have involvement. You still have oversight, judgment, goals. And on the flip side, you still have accountability. The question is: If something doesn’t get done, whose fault is it? I don’t think blaming the AI—“Oh, it was too lazy, didn’t do the thing”—is going to fly. As the director, the person running the show, you need to feel that accountability and make sure the task gets done the way you want.
50:02-52:11
Greg Brockman: It’s almost like everyone’s going to become the CEO of their own corporation, orchestrating as many agents as they care to manage. It’s going to be a very empowering world where people can do so much more of what they want. And I think that means people will be able to make more money, improve their health, learn anything they want, and spend their time the way they want. If you have ambitions, they’re more within reach than ever. I know this concretely from talking to users—small business owners running operations on top of ChatGPT who say, “I wouldn’t have been able to do this otherwise.” People who’ve quit their jobs to start businesses because ChatGPT helps them with so much. That kind of empowerment—where people wouldn’t have taken the leap before but now can—that’s like the American dream on steroids. That’s what we’re unlocking.
Interviewer: And I mean, you’re painting such a rosy picture. I see that in my own work as a solo media entrepreneur. But people also feel very negatively about AI, especially in the U.S. I’m wondering—if the stakes are this high, do you have to deliver on this vision to shift that sentiment? Maybe that’s the thing that changes perception, because right now, it feels increasingly negative in a lot of corners of society. As someone responsible for the business and ChatGPT, do you feel that pressure?
Greg Brockman: We feel it. We feel the responsibility, the gravity, and we definitely feel the criticality of both delivering on this vision and helping people understand it—helping them understand who we are, our intentions, why we’re doing this. Sentiment varies around the world. In America, we know we have to deliver on this vision to help people lift themselves up and feel empowered. In countries like Korea or Japan, sentiment is already very positive—people see the potential.
Greg Brockman: Where America sits right now is at an incredible moment with incredible opportunity. We are the leader in this technology. This is going to be the most important technology humanity ever creates. It will shape how people live, work, accomplish goals, how the economy runs, how countries relate to each other—all of it. And America leads in that. That opportunity is one we shouldn’t take for granted. We have to make the most of it.
52:49-53:42
Greg Brockman: Over the course of OpenAI, I’ve realized how important this is. It has to be broadly empowering—a technology available to everyone to lift themselves up. America leading ensures the values represented by this technology—how it plays out in the world—enhance democratic principles and uplift the core values we hold dear. It’s about how we use it, but also how we use this leadership opportunity and ensure we don’t squander it.
Interviewer: Yeah, those are all my questions. Do you feel like there’s anything on your mind that you want to say or anything we’re missing?
Greg Brockman: So many things. There’s a lot of ground to cover. At the end of the day, what we’re doing is incredibly complex. There are so many facets—consumer, enterprise, research, safety, alignment. It’s all real. We have this wave of entrepreneurship unlocking. But at the core, there’s something very simple: We have a mission. Our mission is to ensure that artificial general intelligence benefits all of humanity. That’s what we’re here to do. Everything we do flows from that one singular goal. That’s what we’re pursuing. That’s why we started this place, and that’s what we’re committed to.
Made with: The Transcript Desk Chrome Extension