OpenAI President Greg Brockman: AI can now generate new knowledge, and intelligent agents are adopting a "vertical ascent" along a curved path.

OpenAI is standing at a new inflection point in its capabilities. From solving the Millennium Prize Problem in mathematics to leapfrogging computer usage capabilities, the company is trying to prove to the world that AI is no longer just a chat tool, but an intelligent agent that can truly do tasks for users.

In a recent podcast interview, Greg Brockman, co-founder and president of OpenAI, stated that the company's model has successfully found a counterexample to the Navier-Stokes equations, confirming the existence of singularities in this core equation of fluid dynamics under specific conditions, thus solving one of the seven "millennium mathematical problems." He characterized this breakthrough as a milestone in "AI's ability to generate new knowledge," believing it heralds a "real shift" in AI's ambitions in areas such as conquering diseases and developing new drugs. At the same time, he revealed that since the launch of ChatGPT Work, the adoption curve of intelligent agents has "almost vertically increased," and ChatGPT's weekly active users have "far exceeded 1 billion."

Brockman stated that the company is accelerating the integration of chat and intelligent agent models, driving AI from a "conversational tool" to a form that "can proactively and continuously share the workload for users." He believes this is the true promise of AI in the long term—"to make computers truly operate according to your goals and desires."

A mathematical problem has been solved, ushering in a new era of AI-generated knowledge.

In an interview, Brockman described the breakthrough in the Navier-Stokes equations as a "significant day." He stated that OpenAI's model found a counterexample or proof confirming that these theoretical equations do indeed have singularities under certain conditions, i.e., a situation where they "collapse." This problem has long remained unresolved and is one of the most profound problems in the field of mathematics.

He emphasized that the significance of this achievement extends beyond the mathematics itself—these equations have wide applications in fluid mechanics, covering phenomena such as ocean currents, airflow around aircraft, and turbulence. But its more profound impact lies in demonstrating that AI models can now assist scientists and mathematicians in accelerating research processes and are capable of tackling previously "out-of-the-way" problems. "When you have this level of assistance, talking about conquering diseases and developing new drugs becomes increasingly achievable," Brockman said.

Astra's ability to overcome the barrier of computer skills sparks a community frenzy.

Brockman said the launch of Astra made him "very humbled," and the creativity and applications shown by the community users far exceeded expectations. He stated unequivocally, "We have crossed a new threshold in the use of computers."

He explained that users have used Astra to perform tasks such as mapping physical spaces and converting them into 3D models, and designing manufacturable parts. In his view, the core value of this capability lies in the fact that AI is shifting from relying on "connectors"—manually coded system interfaces—to directly manipulating all computer applications like humans. "From the very beginning of OpenAI, we harbored a dream: one day, to create an AI that can help you with anything you can do with a computer yourself. I believe we have now truly achieved that with the help of Astra."

He also acknowledged that these capabilities are currently mainly being explored by technology users, and how to transfer them to "ordinary consumers who don't even know what Blender is" is one of the company's core challenges.

The usage rate of AI agents is rising sharply, and chat and AI agents are rapidly merging.

Brockman revealed that this year, users are clearly shifting from pure chat scenarios to agent-based scenarios, and since the launch of ChatGPT Work, the adoption curve for agents has been "almost vertically upward." At the same time, he emphasized that chat scenarios "remain vibrant," and ChatGPT's market share has begun to climb again.

He pointed out that chat and intelligent agents are currently two different forms, but this is only a phase. The company is promoting the unification and integration of the two, with the goal of making AI no longer require users to pay attention to underlying details such as context windows, model selection, and the intensity of thinking, but truly focus on helping users solve problems and "giving time back to users."

He also admitted that the continuous expansion of product capabilities has brought about a "first-mover disadvantage"—a large number of early users have formed a fixed impression of ChatGPT, while its current capabilities are far superior to what they were back then. He believes that this is essentially a "problem to be discovered," which the entire industry needs to face squarely. ChatGPT's weekly active user base of over one billion also constitutes a huge opportunity to reconnect with old users.

The healthcare sector is structured around three pillars, aiming to transform the entire medical process.

Brockman divides OpenAI's health strategy into three directions: consumer-facing (300 million users ask health-related questions every week), professionally optimized versions for clinicians, and enterprise-facing (including integration with Epic systems) for direct sales and deployment to hospitals.

He believes that once these three pillars mature, they will create a synergistic effect, potentially transforming the healthcare experience. He cites his own experience—his wife spent five years consulting multiple specialists before finally being diagnosed with a systemic genetic disease—to point out the serious information fragmentation problem within the current highly specialized healthcare system. He argues that the value of AI lies in its ability to bridge information gaps across disciplines and identify patterns that are difficult for human doctors to detect. "It can truly help you connect all aspects of life, while possessing deep expertise in all areas of medicine—that's the true power and potential of AI," he says.

Operator was not a failure, but rather an "iterative cornerstone" of computer usage capabilities.

In response to external criticism of the Operator project, Brockman characterized it as a "timing problem" rather than a failure. He stated that Operator's model capabilities at the time were "just below the threshold"—as a cloud-based computer operating system, it was slow, lacked accuracy, and was "quite painful to use," failing to cross the practicality threshold.

However, he emphasized that the value of the Operator lay in providing real-world deployment experience for subsequent capability development. The team then focused on and systematically resolved the backlog of issues, ultimately achieving a true breakthrough in Astra. He summarized this process as OpenAI's core working method: "Whenever we truly face a challenge and seriously consider how to safely and properly accomplish it, and truly deliver value, this is how we do things."

Brockman also pointed out that this year the company's overall operation has shifted from a "large company-style product experimentation" to a "startup-style focused mode," which is an important background driving the current series of developments.

The following is the full text of the interview:

Host : Today we have with us Greg Brockman, co-founder and president of OpenAI. Welcome to the show, Greg, how have you been?

Greg Brockman : Great, thanks for the invitation.

Host : Thank you for taking the time to come. Today is a big day, could you first talk about this breakthrough in mathematics? What exactly happened? Why is this so important? I know there are many conflicting accounts in the timeline, and I hope you can explain to us clearly what actually happened today regarding the Navier-Stokes equations.

Greg Brockman : This is truly a momentous day for AI and the advancement of modern science. Today we announce that our model has solved the Navier-Stokes equations problem—we have found a counterexample, or a proof, that these theoretical equations do indeed have singularities under certain conditions, meaning they can "collapse." This problem has been unresolved for a long time; it is one of the seven "millennium problems" and is one of the most profound and important problems in mathematics.

I think the problem itself is important—it represents new knowledge for humanity, and the proof itself is very elegant and beautiful; I think there's a lot to learn from it. These equations have many applications in fluid mechanics and other fields. But for me, what's more important is that this represents the level of our current modeling capabilities—we can truly generate new knowledge, learn from these models, and let them help us solve problems that were previously unattainable or would have taken a very long time to solve.

Host : So how should I interpret this? Will it help me book a plane ticket? Will it help conquer cancer? Or is it just a way for you to recruit researchers who are fascinated by this kind of thing? Because I think this has already caused a stir in the tech world, but I guess it won't be the kind of topic that friends and family would discuss over a backyard barbecue, since most people are several layers away from this field.

Greg Brockman : First, the specific results themselves, or rather the equations themselves, have practical applications—they help us better understand various phenomena, from ocean currents to airflow around airplanes and turbulence. But more importantly, the broader insights and methodologies behind them can help scientists and mathematicians further accelerate their research. I think this illustrates once again: if our models can help solve these kinds of problems, what happens next? What other problems can be immediately applied?

I believe that when you have this level of support and modeling capabilities, talking about conquering diseases and developing new drugs becomes increasingly achievable. I truly believe that the ambition we have to tackle the problems we're capable of will undergo a real shift.

Host : Even if this doesn't truly break into the mainstream, it's sure to be a hot topic this weekend, because everyone seems to be talking about Blender and building various things with Astra. Could you tell us about this Astra release? What's the feedback like? What have you learned? It sounds like it went through several "resets," and the model scaling is performing very well. How does this release differ from previous releases?

Greg Brockman : First of all, the community's response to Astra has truly impressed me. To be honest, seeing the creativity and diverse applications everyone has presented makes me feel very humbled and deeply moved. It's safe to say that we have crossed a new threshold in our use of computers.

This model can now truly operate various applications in an unprecedented way. People are also taking full advantage of this, seriously considering how to create. Many people have showcased their 3D creations, mapping physical spaces and converting them into 3D models, and are even considering whether they can be used to design physical parts. Someone shared his process of designing a "shower drain hair-blocking device"—that's pretty cool, they can actually manufacture it now, it's like a "super smart shower hair-blocking device" (laughs).

Host : Yes, these things sound ordinary, but they are actually very important. I feel that many of these kinds of applications are often overlooked.

Greg Brockman : Exactly. I think there's a crucial point here: sometimes we talk about grand challenges or niche applications, but what's truly important is that the problems you want to solve in your daily life can now actually be solved. What we really want to do is empower every individual, giving everyone "superpowers" to accomplish more. Truly benefiting and empowering people, creating tools that genuinely help your daily life—that's what we're doing.

Host : It seems that this weekend, a lot of people in the tech world completely "rebuilt" their homes with Blender, planning a major renovation, all thanks to Astra. But I'm curious how you integrated Codex, ChatGPT Work, and the desktop version of ChatGPT. I currently have a gaming PC with an NVIDIA graphics card, a Mac mini, and various other devices—I know this tinkering is unnecessary; it's more of a hobby for me, the kind of "early adopter" stage where you're willing to piece things together to unlock various capabilities. But in the future, all of this should be contained in a single prompt box. I can casually ask, "Should I renovate my house?" and it might directly create a 3D Blender model to answer that question. What do you think about how the capabilities we saw these "hardcore gamers" demonstrate this weekend can truly reach ordinary consumers who don't even know what Blender is?

Greg Brockman : I think you're absolutely right—we really want to transform these tools from the old model that required users to have basic operational skills or guidance, into a state where people can truly "fly freely": empowering users to set goals and directions, while still allowing them to feel that they can delve into the details, understand it, and provide oversight—because ultimately, you should be responsible for the final result. But at the same time, you also have an absolute "amplifier," like a trampoline, or like "a rocket for your mind," however you like.

From a practical perspective, I think we've clearly seen a shift in users from pure chat scenarios to agent-based scenarios this year. But I also want to emphasize that chat scenarios remain vibrant—we now have well over 1 billion weekly active users, and you can see ChatGPT's market share starting to climb again because we've invested heavily in various applications in education, healthcare, and many other areas important to people. At the same time, applications focused on productivity and deep knowledge work are truly taking off—since the launch of ChatGPT Work, agent adoption has been almost vertically increasing.

I believe that chatbots and intelligent agents are indeed two different forms right now, but this is only a phase. We are continuously working to unify and integrate them, and we are beginning to see a completely new AI consumption model that people are more willing to accept emerge. I think this actually fulfills AI's long-standing promise—that you can have something that can converse and truly proactively and continuously share your workload. If we turn back five or ten years, people's expectations for AI would never have been a "lower-level language model"—you would have to consider context windows, choose the intensity of thought, and select different models; these are not what the future should look like.

So I believe we're moving towards true "capability amplification": truly giving time back to the user, allowing the computer to truly operate according to your goals and your desires. I think that's the core. We're also safely advancing all of this—it's one of our core commitments and how we think about it. We're definitely seeing the capabilities of these tools begin to increase rapidly, and users can do more and more, while underlying tools like Blender are gradually becoming a "behind-the-scenes detail," fading from the user's view—this is absolutely the direction we're heading.

Host : In the field of AI, there seems to be a near-"first-mover disadvantage" phenomenon—billions of people have already tried ChatGPT, some of whom are experiencing it for the first time and have formed certain fixed impressions about what the product can do. Even in recent weeks, new agents and products have been constantly launched, and people are amazed by their capabilities. I find this quite interesting because I myself am the kind of person who tries desperately to use ChatGPT to its fullest potential—for example, I always have an agent constantly monitoring SpaceX's launch times, notifying me of any delays, this kind of agent continuously running in the background. But from a strategic perspective, I think this is actually a new kind of challenge, because as capabilities continue to expand, first-mover advantages almost constantly need to remind the market, "Look, there are so many untapped new uses to explore."

Greg Brockman : We've definitely thought a lot about this issue. I think it's a "discovery problem" that the entire industry needs to take seriously and formally. Although we haven't completely solved it yet, I believe we have a real opportunity to do it better than any previous product.

Currently, what we rely on—think of ChatGPT and ChatGPT Work—is essentially a text input box. It's like saying, "This new text box is much more powerful than the old one," but (laughs) for some reason, people still prefer the old text box because the new thing is "too confusing." What people really want is something that helps them solve problems. The point of the whole thing is to give time back to the user, instead of forcing them to understand all these internal details. But at the same time, we also have a model that understands your true intentions—you tell it "what I want," and it already has a lot of background knowledge about you, so it should also be able to proactively remind you, for example: "If you ask me in a different way," or "If you add this connector," or "If you authorize me to do this," or "If you integrate your credentials this way," then I can do something else for you.

So we've put a lot of thought into this "self-awareness" process, this process of guiding users onboarding, and I think it's a huge opportunity. The disadvantage you mentioned certainly exists—people try it once, form an impression, and then things are different now; but there are also advantages—ChatGPT has over 1 billion weekly active users, a scale that's unique, no other model can reach that level. And, I remember the total number of people who have tried ChatGPT before was around 1 to 1.5 billion, which is also a huge opportunity for us to go back to these old users and tell them, "Hey, we can help you solve this problem now, in a way you've never seen before."

And this is real—look at how many people use ChatGPT to ask health-related questions. Every week, 300 million people ask health-related questions, which is truly changing the lives of people and their loved ones. So the opportunity before us is truly enormous.

Host : Could we continue discussing the topic of health? Before that, I'd like to mention something interesting, and perhaps something OpenAI could do better: From a marketing perspective, every company in the AI industry wants to be the "Apple of AI." But when people think of "Apple-style marketing," they often picture "Mac vs. PC," or that classic 1984 ad—those large-scale brand campaigns. But what Apple truly excels at in marketing is their repeated emphasis on very specific details of their products—for example, they advertise specifically for a new camera, specifically for Memoji, specifically for a specific feature in Safari. The AI field, however, requires an order of magnitude more information because it can do so much more. Therefore, I think there's a huge potential for companies to invest in advertising—brand promotional videos are great, and Astra's launch video was excellent, but there should be billboards specifically promoting a particular ChatGPT capability, telling people "how much time this can save you."

Greg Brockman : I've really experienced this firsthand this year, and I've truly come to realize it. For example, we just released ChatGPT Image 2.5. If you watch the release video, my favorite part is that it shows someone creating an image, then showing a whole bunch of different variations, and finally showing that person picking out their favorite version and turning it into a real-world object—like someone saying, "This is a small sketch of a candlestick," and you see a great visualization, then seeing that candlestick actually made and placed there, and you immediately understand, "Yes, this is exactly what I wanted." It instantly resonates with you.

I think it's crucial to truly demonstrate a specific use case. Another thing that surprised me a bit is that sometimes we showcase niche applications that are amazing to a particular person, but others don't immediately think, "Since it can do something so powerful, can I use it to solve another problem I've been wanting to solve?" This connection is actually much harder to establish than I initially imagined, but upon reflection, it makes sense. What you really want is a specific use case that makes people immediately say, "I want this right now, I'm going to try it out," and from there, they'll explore and gradually discover, "Oh, I really do have such a powerful application, I just hadn't realized it before."

Host : Yes, Sora is probably the best example of this effect. Another point is to constantly remind users to actively ask what AI can actually do. I had dinner with a friend who works in real estate development. He uses ChatGPT all day to write various transaction memos and understand the various projects he is working on. He often asks me, "Can ChatGPT do this? Can it do that?" I said, I'm happy to answer you, but actually you already have a tool that can directly explain to you how to achieve what you want—maybe it can't do it, but it's highly likely it can.

Greg Brockman : Yes, I did the same thing. When I was working on the Blender project, I was thinking, "Should I run this as a local Codex thread, or should I run it in the cloud?" I just asked it directly, asking it to suggest which was better based on my system, and it gave me a great answer, which I followed.

Host : Speaking of image generation, when the second-generation model first came out, I thought, "Image generation has been solved." Where do you think the image generation category will go next? Because it feels like the best "one-shot" effect for me of all the functions—especially when it was first released, I spent several hours one Saturday designing furniture, trying hundreds and thousands of prompts. How far can image models go? What is the future direction? What other application scenarios do you think are worth paying attention to? We've also talked about the downstream impact of image models before—if you can take a picture of a physical space and then imagine it in various different variations, this will drive a lot of real-world actions because people can "see" this thing.

Greg Brockman : Yeah, and then they'll say, 'I want to make this a reality right now,' which I think is really cool.

Host : That's right.

Greg Brockman : I think that's exactly the right way to understand it: whenever you cross a new skill threshold, in my experience, it unlocks entirely new application scenarios that you never even imagined beforehand. So I think that in knowledge-based work, professional work, marketing, and other fields, you only need to cross a certain quality threshold—if you don't reach it, it's just a good concept at best, but the final result can't be truly used, which means the problem hasn't really been solved.

Precise editing control, fast response times, genuine creativity, diverse results, and effective user interaction—I believe these unlock entirely new application scenarios, leading to a massive market. Take, for example, an application that might not be immediately apparent but is becoming increasingly important: creating slideshows or building sophisticated websites. The ability to embed image generation capabilities within this process is a unique advantage OpenAI possesses compared to some of our competitors, allowing you to produce significantly higher-quality downstream results. We essentially view image, voice, and programming capabilities as components of a unified whole; they converge to create a truly user-empowering AI—meaning you can create anything you can imagine. I think this will be unprecedented.

Host : Let's get back to the topic of health. I think most people already know that you can combine lab data and sleep scores for analysis, but the future direction of the product you recently outlined seems much more complex and in-depth. Could you talk about where ChatGPT is headed in the health field in the future?

Greg Brockman : I would understand what we do in the health field in three aspects.

First, there's the consumer side—as mentioned earlier, 300 million people ask health-related questions every week.

The second is the clinician side, which is a bottom-up approach. Essentially, it's a specially optimized ChatGPT designed for clinicians, providing them with the ability to directly cite medical literature, among other things.

The third pillar is the enterprise side, which means selling directly to hospitals for them to deploy and use. For example, we have an integration partnership with Epic Systems, and we are working hard to ensure that each of these three pillars provides the best independent service. But as these three pillars gradually gain momentum, you will find that they can actually generate synergies—this is the key that truly has the potential to change the healthcare experience in the United States and even the world, and to truly transform the healthcare industry.

Consider how much work you do as a patient: if you see different specialists, you have to carry your medical records from one doctor to another, repeatedly explaining "what my problem is"; but ultimately, the responsibility for making the decision still falls on you—whether you like it or not, you are the one who must make the decision. This becomes possible if all information can be shared smoothly between different medical institutions, and all of this is built on a single platform. Take clinical trial recruitment, for example—a huge bottleneck in drug development, finding qualified individuals who can truly benefit from participating in a trial. If you have such data, and if people are willing to entrust this information to you, it truly benefits not only them but also the world.

Therefore, I understand that what we are doing is actually striving to build the world's best healthcare platform, truly bringing the healthcare industry into the AI era. I believe this will fundamentally change the quality of life for many people, genuinely improving the living standards of a vast number of people—and we are already seeing this on a very concrete level. Some of my favorite stories about ChatGPT involve people saying, "Hey, the information I got from ChatGPT saved my life," or "It saved the life of my loved one"—someone encountered a medical problem, the doctor told them A, but they used ChatGPT to verify and understand what the doctor said, even questioning the doctor based on it, ultimately achieving a positive outcome. These kinds of things happen every day. So I believe that we have only just scratched the surface of AI's application in the health field, and this is one of the most positive and proactive applications of AI.

Host : Yes, I'm particularly curious about how advancements in memory function will connect with health—because I anticipate that as ChatGPT's memory capabilities continue to develop, perhaps when I start a new conversation, it will automatically retrieve relevant information, or connect it to a conversation from three weeks ago. If you actually apply this ability to health-related questions, it might be able to identify certain patterns—the kind of patterns that usually take humans years to discover, such as a certain ailment: "Hey, those seemingly unrelated questions you asked before are actually related; you might need to dig deeper along this line of inquiry."

The last question—sorry, you go first.

Greg Brockman : I was just about to say that I completely agree with that assessment, and we've seen very concrete examples of it. Take my own experience, for instance—my wife has spoken publicly about some of her health issues, which was actually a five-year medical journey: she saw many specialists, each focusing only on a part of the "elephant," treating only their own small section. Finally, an allergist said, "I think your symptoms are actually interconnected; you actually have a genetic disease that affects various systems throughout your body."

This situation is truly difficult to pinpoint; how many people are facing similar predicaments without ever finding a real answer? How many are trapped in highly specialized medical systems, never able to piece together a complete diagnosis? I believe this is precisely where the true power and potential of AI lies—it can truly and deeply help you navigate all aspects of life, while also possessing profound expertise in all areas of medicine.

Host : Okay, one last question. How would you tell the story of the Operator project? It seems like a failure, or a "side quest," but given the advancements in computing power today, it now seems extremely important. Is there a clear line of development? What exactly is the Operator? Does it still exist in some form within ChatGPT? How exactly was the problem of computing power solved?

Greg Brockman : I would see it all as a matter of timing, and also as a manifestation of the "iterative deployment" concept. There will always be a point where the model's capabilities aren't fully developed, but learning from real-world deployments is actually very helpful. I think Operator's model capabilities at the time were just below the threshold—it was a cloud-based system that could operate computers, but it was slow and inaccurate, making it quite painful to use. Although some people did derive value from it, overall, it didn't cross that threshold.

If you look back at what happened afterward—one thing OpenAI did exceptionally well is that we invest long-term in what truly matters and work diligently and steadily. This year, the team truly began to build momentum, dedicating significant effort to systematically clearing a long backlog of problems and focusing on solving the "computer usage" problem. I believe they have delivered a substantial report card. Of course, there is still much to be done, and it will never be "completely finished," but this is indeed a true milestone. I think everyone can truly appreciate the significance of this now because we have always lived in a world where these intelligent agents operate computers through "connectors"—systems that are extremely laboriously handwritten and coded, completely different from how humans use computers—while humans have long been able to use all the functions of a computer because everything is designed for humans.

So, if you have an AI that can operate a computer like a human—in fact, since the inception of OpenAI, we've had this dream: that one day we can create an AI that can help you with anything you can do with a computer yourself—I believe we've truly achieved that now, thanks to Astra. I believe people will continue to discover even more applications in the future. But this in itself is a great example of the importance of long-term focus, a down-to-earth approach, and never giving up even when faced with difficulties.

Host : Indeed, my own feeling is that Astra really brings together several different technical approaches at the right time—the advancement of the model itself, the ability to use computers, the voice capabilities, all of these things... everything makes sense now.

Greg Brockman : Exactly, it all comes down to focus. I think that's what makes this company so outstanding—whenever we really take on a challenge, seriously consider how to do it safely and properly, and actually deliver value, that's exactly what we do.

Host : To put it another way, looking back at last year, OpenAI's operating style seemed more like that of a "large company"—and it is indeed a large company, but the way it experimented with products back then was more like that of a very large tech company: launching a new product with the mindset that "a 20% success rate, or even a 10% or 5% success rate, is acceptable." This year, however, it feels like the entire company has switched back to a true "startup mode"—that kind of "focus, focus, and more focus," where everything must come together, and the whole team must row in the same direction. The resulting difference in momentum and growth has pulled the company back from operating and releasing products like a large company to releasing products and maintaining focus like a startup—this is almost an impossible task, but seeing this process is truly amazing. So, thank you.

Greg Brockman : You're welcome. This is indeed the result of the collective efforts of many, many people at OpenAI. It's what we really want to bring together—and I think this is what we, as a company, truly value: we always remain highly focused on our mission and carefully consider whether everything we do is truly contributing to the achievement of that mission.

Host : Thank you so much for taking the time to chat with us. It's a pleasure to meet you. We'll talk again soon, Greg.

Greg Brockman : Thank you, have a good day, goodbye.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.