OpenAI Co-founder Exclusive Interview: After shutting down Sora, what's next for ChatGPT?

Video title: OpenAI President Greg Brockman: AI Strategy, AGI, and the Super App

Video author: Alex Kantrowitz

Translation: Peggy, BlockBeats

Editor’s note: This article is translated from a conversation between Greg Brockman, OpenAI’s President and co-founder, on the Big Technology Podcast. The show has long focused on changes in AI, the tech industry, and business structures—making it an important window for judging what’s happening on the ground in Silicon Valley.

In this conversation, Brockman didn’t stay at the level of model capabilities themselves. Instead, he pushed the question further back: once AI’s capabilities have largely been validated, how will the industry choose its path next, reshape product forms, and absorb the systematic shock it brings. The discussion centers on OpenAI’s product strategy, the upcoming “super app,” and its assessment of AI entering a “takeoff stage.”

This conversation can be understood in three areas.

First, the convergence of the path.
From video generation to reasoning models; from parallel exploration to active trade-offs—OpenAI’s choices are not simply judgments about technological superiority, but responses to real constraints. Compute has become the core bottleneck. Under limited resources, technical routes begin to converge toward two directions with the highest leverage: personal assistants and complex problem-solving. This also means that AI’s competitive logic is shifting from “what it can do” to “what it should do first.”

Second, the reconstruction of form.
The proposal of the “super app” is, in essence, a leap in product form. AI is no longer a collection of scattered tools—it’s a unified entry point. It understands context, calls tools, executes tasks, and continuously builds memory across different scenarios. From ChatGPT to Codex, AI is gradually taking over the full workflow, while the role of humans is also shifting—from executors to orchestrators: setting goals, assigning tasks, and supervising.

Third, a turnaround in pace.
If the past two years were a ramp-up phase for capabilities, then what’s happening now is a “takeoff.” On the one hand, model capability jumps from “helping with about 20% of work” to “covering about 80% of tasks,” directly triggering a reconstruction of workflows. On the other hand, AI is participating in its own evolution (using AI to optimize AI). Combined with coordination across chips, applications, and the enterprise side, it forms a continuously accelerating closed loop. AI is no longer a single-point technology—it’s starting to become a key engine driving economic growth.

But at the same time, another set of questions is emerging in parallel: public distrust, uncertainty in jobs, controversies brought by data centers, and the boundaries of safety and governance. Brockman’s answers aren’t fully contained within the technical realm. He emphasizes two points instead: first, risks can’t be solved through “centralized control”; society needs to build social infrastructure around AI, similar to power systems. Second, individual capability is changing—what truly matters won’t be “whether you can use tools,” but “whether you can achieve your own goals with the help of AI.”

If the old question was “what can AI do,” then now the question has become: once AI starts handling most things for you, what will you still need to do?

Below is the original text content (reorganized to make it easier to read):

TL;DR

AGI has entered the “clear path” stage: Greg Brockman (OpenAI co-founder) believes that GPT-based reasoning models already have a clear route toward AGI. He expects it to be achieved within a few years, though the form will remain “non-uniform” (jagged).

Note: AGI (Artificial General Intelligence) refers to general artificial intelligence—AI systems that have capabilities comparable to, or even surpass, humans in the vast majority of cognitive tasks. Unlike today’s “narrow AI” (such as image recognition and recommendation algorithms), AGI emphasizes generality across tasks and transfer capabilities.

Strategic convergence: from multi-line exploration to two core applications: Under compute constraints, OpenAI will concentrate resources on “personal assistants” and “complex problem-solving,” rather than pushing all directions simultaneously (such as video generation).

The “super app” will become the AI entry form: Chat, programming, browsers, and knowledge work will be integrated into a single unified system. AI shifts from being a tool to an “execution layer,” and users shift into “orchestrators.”

Key inflection: AI begins taking over workflows rather than just assisting: Model capability has risen from “completing 20% of tasks” to “capable of handling 80%,” forcing individuals and enterprises to restructure how they work.

Compute is the core bottleneck and the focus of competition: AI demand far exceeds supply. In the future, the limitation won’t be model capability, but computing resources—data centers and infrastructure become key variables.

AI “takeoff” is happening: Technical self-acceleration (AI optimizing AI), combined with industrial coordination (chips, apps, enterprises), drives AI from being a tool toward an engine for economic growth.

The biggest risk isn’t technology—it’s governance and how it’s used: Safety issues can’t be solved by a single party. An open ecosystem and social infrastructure are needed to carry the responsibility together.

Individual core capability is shifting: In the future, competitiveness won’t be about “execution,” but about “setting goals + managing AI systems.” Proactively using AI will become a baseline ability.

Conversation summary:

Alex (Host):
Today we’ve got Greg Brockman, OpenAI’s co-founder and President, joining us to talk about the most promising opportunities for AI, how OpenAI will seize those opportunities, and the concept of a “super app.” Greg is also here in our recording studio today.

Greg Brockman (OpenAI co-founder & President):
Great to meet you—thank you for having me.

Why shut down Sora? Not enough compute

Alex:
This timing is pretty interesting. OpenAI is pausing the push on video generation, putting the resources into a “super app”—one that integrates business and programming scenarios. From the outside (including me), it feels like OpenAI has already gotten ahead on the consumer side, yet it’s adjusting how it allocates resources. What exactly is happening?

Note: In March 2026, OpenAI announced the shutdown of its video generation product Sora (including the app and API) and stopped related commercial rollouts.

Greg Brockman:
Over the past stretch of time, we’ve been developing this deep learning technology to validate whether it can truly produce the kind of positive impact we’ve been envisioning—whether it can be used to build applications that genuinely help people and improve their lives.

At the same time, we’re also doing another line of work: deploying this technology. On the one hand, to support the operation of the business. On the other hand, to start accumulating real-world experience in advance, so that when the technology is truly mature, we’re ready.

And now, we’ve reached a new stage. We’re seeing that this technology really is feasible. We’re moving from “benchmarks” and somewhat abstract demonstrations of capability into a new stage—where we have to put it into the real world, let it participate in real work, and keep evolving through user feedback.

So I’m more inclined to interpret this change as: a strategic pivot driven by a shift in the technology stage.

This doesn’t mean we’re shifting from the “consumer side” to the “enterprise side.” More precisely, we’re asking a question: given limited resources, which applications should we prioritize above all else? Because we can’t do everything.

Which applications can truly get deployed, create synergy with each other, and have real impact? If you list all directions, the consumer side can be broken down into many forms—for example, a personal assistant: a system that truly understands you, aligns with your goals, and can help you achieve your life goals. Or creation and entertainment, and many other possibilities. And on the enterprise side, if you look at it at a higher abstraction level, it can really be summarized as one thing: you have a complex task—can AI help you accomplish it?

For us, right now, the priority is very clear: there are only two things at the top. First, a personal assistant. Second, an AI that can help you solve complex problems.

The issue is: our current compute isn’t even enough to fully cover those two things. Once you add more application scenarios, it’s simply impossible to cover everything. So this is really a practical judgment: the technology is maturing quickly, the impact is about to explode, and we need to make trade-offs—choosing the most important direction and actually building it.

Alex:
You previously mentioned an analogy—that OpenAI is a bit like Disney: it has a core capability, and then it can extend into different scenarios. Disney has Mickey Mouse, which can be used for movies, theme parks, and Disney+. OpenAI’s “core” is the model: it can do video generation, build assistants, and power enterprise applications.

But now it seems like you may not be taking that kind of “total expansion” approach anymore, and that you actually have to make choices.

Greg Brockman:
Actually, I think the analogy still holds even more now. The key difference is that, from a technical perspective, Sora (a video model) and GPT (a reasoning model) are actually two different branches of technology. They’re built in completely different ways.

The problem is that, at our current stage, it’s very difficult to move forward with both of these technology trees at the same time—especially under limited resources. So the choice we made is to concentrate the majority of resources on the GPT path at this stage.

Of course, that doesn’t mean we’re abandoning other directions. For example, in robotics, we’re still doing related research. But robotics is still at an earlier stage; it hasn’t yet entered a true breakout period of maturity.

In contrast, over the next year, we will see AI take off in knowledge work.

And it’s important to emphasize that the GPT path isn’t just “text.” For instance, bidirectional speech interaction (speech-to-speech) is also part of this technical route—it will make AI more usable and more practical. These capabilities are still fundamentally within the same model family; they’re adjusted through different interaction modes.

But if you go down two completely different technology branches, then under compute constraints it’s hard to sustain that long-term. And compute is limited because the demand is simply too large. After almost every model release, people want to use it for more things.

Alex:
So why didn’t you put the focus on the world model path? For example, video models need to understand relationships between objects, which is also crucial for robotics. And Sora’s progress has been very fast. Why did you ultimately bet on GPT?

Note: “World model” focuses on perception and physical intuition. Its core is helping AI understand “how the world works,” not just learning “surface patterns in data.” Such models are typically used to describe systems like Sora: it’s not only generating images or video—it’s modeling relationships among objects (people, cars, light), continuous changes over time (the evolution between frames), and fundamental physical laws (motion, occlusion, and collisions). In contrast, GPT is a language and reasoning model, focusing more on abstract cognition and task execution capabilities.

Greg Brockman:
The biggest problem in that area is actually that there are too many opportunities.

We realized very early on that, at OpenAI, if an idea is mathematically sound, it usually runs and produces decent results. That suggests the underlying capabilities of deep learning are very strong: it can abstract generation rules from data and transfer them to new scenarios. You can apply this across world models, scientific discovery, programming, and all kinds of fields.

But the key is: we need to make trade-offs.

There’s long been a debate about how far text models can go—whether they can truly understand the world. I think that question now has an answer: text models can go all the way to AGI.

We’ve already seen a clear path, and stronger models will show up this year. And inside OpenAI, one of our biggest pains is how to allocate compute—this problem will only get worse, not easier. So fundamentally, it’s not a question of “which route is more important,” but of timing and sequencing.

Now, some applications we previously thought were distant are starting to become within reach. For example, solving physics problems that haven’t been cracked yet. We recently had a case: a physicist had worked on a particular problem for a long time. They gave it to the model, and after 12 hours, we produced a solution. He told us it was the first time he felt like a model was “thinking.” This problem might even be one that humans would never solve—but AI did.

When you see something like that, your only choice is to double down and triple down. Because that means we truly can unlock huge potential.

So for me, this isn’t competition between different directions. It’s really a question: what is OpenAI’s mission? How do we bring AGI to the world? How do we make it truly benefit everyone? And we’ve already seen the path—so we know how to push it forward.

Betting on GPT, not world models: choosing the path to AGI

Alex:
Okay, I do want to come back to the next-generation models you just mentioned, but first I want to follow up on your point.

Earlier this year, I talked with Demis Hassabis from Google DeepMind. One interesting thing he said is that, to him, the closest thing to AGI is actually their image generator called Nano Banana.

Note: Demis Hassabis is one of the key figures pushing AI from research toward breakthrough applications. He founded DeepMind, developed AlphaGo, and in 2016 defeated the world champion in Go—becoming a landmark event in AI development history.

His reasoning was: whether it’s an image generator or a video generator, to generate images and videos like that, you fundamentally need to understand relationships between objects—at least to have some level of awareness of how the world operates.

So does this suggest there’s a potential risk? That’s a big bet. If that’s the case, and OpenAI keeps adding more on another technology tree, would it miss something important?

Greg Brockman:
If that’s the case, I have two answers.

First, of course there’s a possibility. That’s just how this field works—you ultimately have to make choices and place bets. OpenAI has been doing exactly that since the beginning: we decided what we believe the AGI route is, and we’ve focused intensely along that path. It’s like adding random vectors—you might end up near zero. But if you align all the vectors, they’ll propel you toward a clear direction.

Second, image generation is actually a very popular capability in ChatGPT, and we’re still continuing to invest and prioritize it. The reason we can do so is that it isn’t really part of the “world model” or “diffusion model” technology branch. It’s built on the GPT architecture. So while it deals with different data distributions, at the core it’s the same underlying technology stack.

And that’s one of the most surprising aspects about AGI: sometimes applications that look very different—speech-to-speech, image generation, text processing—plus applications of text itself across scientific research, programming, and personal health information—can all be accommodated within the same technical framework.

So from a technical perspective, one of the things my company and I have been thinking about is how to unify our efforts as much as possible. Because we really believe this technology leads to holistic improvements—and can even lift the entire economic system.

But the scale of this is enormous. Of course we can’t do everything, but we can do our part.

Alex:
That’s exactly what “general” means in Artificial General Intelligence (AGI).

Greg Brockman:
Yes, that “G” really means that.

Alex:
Speaking of “unified,” what would this super app actually look like?

Greg Brockman:
In my mind, the super app is—

Alex:
It would integrate chat, programming, the browser, and things like ChatGPT, right?

Greg Brockman:
Yes. What we’re trying to build is an application for end users—so you can truly experience the power of AGI, meaning its “generality.”

If you think about today’s chat products, I think they’ll evolve into your personal assistant, your personal API—an AI that truly takes your perspective. It understands you, knows a lot about you, aligns with your goals, is trustworthy, and can represent you in this digital world to some extent.

As for Codex, you can think of it this way: it’s still mainly a tool built for software engineers today, but it’s becoming “Codex for everyone.”

Anyone who wants to create and build things can use Codex so the computer can do what they want. And it’s no longer just about “writing software.” It’s more like “using the computer itself.” For example, I might ask it to configure my laptop settings. Sometimes I forget how to set up hot corners, so I just tell Codex to do it—and it actually does.

This is what a computer should be like. It should adapt to people, not make people adapt to it.

So you can imagine an app like this: anything you want your computer to do, you can just tell it. It would include built-in capabilities for “using the computer” and “browser operations,” so the AI can truly operate the web, and you can supervise what it’s doing. And whether your interaction is chatting, writing code, or general knowledge work—those conversations would all be unified under one system. The AI would have memory, understand you.

That’s what we’re building.

But honestly, that’s just the tip of the iceberg—the part that shows above the water. For me, what’s truly more important is unifying the underlying technology.

We mentioned unifying the model layer earlier, but what’s changed over the past few years is that it’s no longer just a “model” problem. The more important issue is the “system that carries it.” In other words: how does the model obtain context? How does it connect to the real world? What actions can it take? And when new context keeps coming in, how does the interaction loop with the user work?

Internally, we’ve had multiple implementations for many of these things, or at least a few implementations that were slightly different. Now we’re consolidating them into a single one. Ultimately, we’ll have a unified AI layer, and then—in a very lightweight way—we point it at different concrete application scenarios.

You could still build small plugins or small interfaces that serve finance or legal in specific ways. But in most cases, you won’t even need to, because the super app itself will be broad enough and generic enough.

Alex:
So this app is for both enterprise scenarios and personal scenarios?

Greg Brockman:
Yes—this is really the core. Like a computer—your laptop: is it for personal use or work use? The answer is: both. It’s first and foremost your device, the interface through which you enter the digital world. And that’s exactly what we want to build.

Alex:
Then from a non-business perspective, if I use this super app in my personal life, what would I do with it? How would my life change?

Greg Brockman:
I’d frame it like this: in personal life, it will first continue how you use ChatGPT today.

How do you use ChatGPT now? People are already using it to complete all kinds of diverse—and pretty astonishing—tasks. Sometimes it’s as simple as: “I need help drafting a speech for my wedding.” Or: “Can you review this idea and give me some feedback?” Or: “I’m running a small business—can you give me some ideas?”

Some of these are personal; some are starting to blur the boundary between personal and work. My view is that all these kinds of problems should be handled by the super app.

Greg Brockman:
But if you look back at how ChatGPT evolved, it’s already changing in this direction.

At first it didn’t have memory, right? For each person, it was the same AI every time, starting from zero—almost like you were speaking to a stranger. But if it can remember you and your past interactions, it becomes much more powerful. And if it can plug into more context, it becomes even more powerful.

For example, imagine it connects to your email and your calendar, truly understanding your preferences. It has deeper background information about your previous experiences, and then uses that information to help you achieve your goals. Or imagine that in ChatGPT there’s already a feature called Pulse, which proactively pushes you content you might be interested in each day based on what it knows about you.

So on the personal usage level, the super app will include all this and do it deeper and more richly.

Alex:
When are you planning to launch it?

Greg Brockman:
A more accurate way to understand it is: over the coming months, we’ll move step by step toward this direction. The complete vision we’re talking about will be delivered gradually, not launched as a whole all at once—it will appear in stages.

For example, the Codex app you have today already contains two layers: one is a general-purpose agent harness system that can use tools; the other is an agent that’s particularly good at writing software.

And that general-purpose harness system can be used in many other scenarios. If you attach it to a spreadsheet or a Word document, it can help you handle knowledge work.

So our first step is to make the Codex app more useful for general knowledge work. Because inside OpenAI, we’ve already seen that people spontaneously start using it this way.

That will be the first step—then there will be many more steps.

Alex:
Yesterday I was talking with one of your colleagues about Codex, and he mentioned a person using Codex for video editing: they had Codex help process the videos, and Codex even made a plugin for Adobe Premiere to break the video into chapters and start editing. Is that the direction you’re aiming for?

Greg Brockman:
I really like hearing cases like that. That’s exactly how we hope this system will be used. And one interesting point is: the Codex app was originally designed for software engineers, so for non-programmers its current usability is actually not that high—because in the setup process, lots of small issues pop up.

Developers immediately understand what those issues mean and how to fix them. We’re used to that. But if you’re not a developer, when you see those things you might think: “What is this? I’ve never seen anything like this before.”

Even so, we’re still seeing many people who have never written programs starting to use it to build websites or do the kind of things you were talking about—automating interactions between different software to get huge leverage. For example, someone on our communications team has connected it to Slack and email, had it handle large volumes of feedback, and produced really solid summaries and syntheses.

So the situation now is that the highly motivated people are already willing to cross these thresholds and get a high return from doing so.

In a sense, the hardest part is already done—we’ve built a truly smart, capable AI that can actually complete tasks. Next, we need to do the relatively “easier” part: make it truly useful for the general public, and gradually remove those onboarding hurdles.

Alex:
From a competitive landscape perspective, Anthropic now also has Claude apps, including chatbots and Claude Code. To a certain extent, they already have the early shape of their own “super app.”

How do you view why Anthropic got to this step earlier? And how big do you think OpenAI’s chances are to catch up?

Greg Brockman:
If you roll the timeline back 12 to 18 months, we actually treated “programming” as a key area and consistently got the best results on various programming competitions and other “pure capability”-type tests. But one thing we invested less in at the time—our last-mile usability.

That is, we didn’t place enough emphasis on the fact that the AI is already very smart and can solve all kinds of hard programming problems, but it has never seen a real-world codebase. Real-world codebases are often messy and far from the “clean” environments it’s familiar with.

On this point, we were indeed behind. But probably starting around mid-last year, we began addressing it very seriously. We formed a dedicated team to figure out where the gaps were, and what kinds of messiness and complexities exist in the real world—things we hadn’t really been exposed to before.

For example: how do you construct training data? How do you set up a training environment? Let the AI truly experience what it feels like to “do software engineering”—being interrupted, encountering strange problems, dealing with all sorts of non-ideal situations, and so on.

I think now we’ve caught up. When users genuinely compare us directly with competitors, many will likely choose us.

Of course, we also know we still have gaps on the front-end experience, and we’ll fill those. Overall, this has been our direction: not just building a model and wrapping it in a product shell on top. Instead, from the beginning, thinking of it as a complete product. When doing research, we’re also thinking: how will it ultimately be used? That’s a shift happening internally at OpenAI right now.

So my view is that we’ll have a very strong wave of model upgrades next. Just looking at this year’s roadmap, I find it really exciting—there’s a lot we can do.

At the same time, we’re very focused on closing the last-mile usability gap.

Alex:
Since 2022, OpenAI has basically been an undisputed leader in this space. Obviously, competition is no longer just about test scores. You just said “we’ve caught up,” which reflects that.

Has the internal atmosphere at the company changed? That is, it no longer feels like you’re far ahead in products like ChatGPT the way it used to—it’s now truly a head-to-head competition.

Some outside reports also show this shift—for example, meetings internally where the message emphasized that OpenAI doesn’t have any “side quests” left, and everyone needs to focus their efforts on this core direction. So what changes have happened inside the company’s environment and culture?

Greg Brockman:
For me personally, the moment that makes me most uneasy at OpenAI is exactly after we released ChatGPT.

I remember at the holiday party that year, there was this vibe everywhere of “we won.” I had never felt that before. My reaction was: no, that’s not us—we’re the ones on the losing side.

And we always have been. Most of our competitors are already established big companies with more money, more people, more data—virtually all resources are more abundant.

So why is OpenAI still able to compete? In a sense, the answer is that we never felt we could rest easy. We always saw ourselves as challengers.

In fact, for me, it’s actually a healthy thing to see the market truly taking on this competitive landscape—to see other competitors emerge and do well too.

Because in my view, you can never nail your attention to where your competitors are right now. If you only watch where they are, by the time you get there they’ve already moved ahead.

And I think that in the past stretch, it was kind of the opposite: a lot of people have been watching our position, and we were able to keep moving forward. That’s actually given us a sense of internal alignment and unity.

I mentioned earlier that we used to treat “research” and “deployment” almost as separate things. But now we really want to integrate them. For me, that’s been a wonderful thing.

So I would say we’re not at a stage where I feel we previously “had it locked up,” or where we’ve suddenly fallen into crisis. You know, the outside world’s assessment of you usually isn’t as good as they say—and it’s rarely as bad as they say either.

Overall, I think we’ve been very stable. On the core work of model development, I’m very confident in our roadmap and in the research investments we’ve made. On the product side, I think we have a very good momentum right now—people are pulling together, truly delivering these things to the world.

Alex:
You’ve mentioned several times already that there will be some very strong new models ahead. So what exactly is it?

The Information reported that you’ve completed pretraining for “Spud,” and Sam Altman also told OpenAI employees that within a few weeks they should see a very strong model. That was a few weeks ago. Internally, the team even believes it could truly accelerate the economy, with progress potentially faster than many people expected.

So what exactly is “Spud”?

Greg Brockman:
It’s a very good model. But I think the main point isn’t about any single model by itself.

Our R&D process is roughly like this: first there’s pretraining—producing a new base model. After that, all further improvements build on top of that base model. This step often requires huge efforts from many teams inside the company. In fact, over the past 18 months, most of my time has gone into this. It’s mainly been around the GPU infrastructure—supporting the teams responsible for training frameworks and getting those massive training runs actually executed.

Then there’s the reinforcement learning stage—teaching the AI, which has already learned a lot of world knowledge, to actually apply that knowledge.

Next comes post-training. In this phase, you really drill it in: now that you’ve learned how to solve problems, practice in a variety of different contexts.

Finally, there’s another “last mile” stage focused on behavior and usability.

So I’d think of Spud as a new foundation and a new pretraining model base. And on top of it, you could say that about two years of our research is starting to yield real results. It will be very exciting.

What the outside world will feel ultimately is an overall increase in capability. But for me, it’s never just about one single release. Because once this version is out, it will just be an early version of further progress. We’ll keep doing more at every step in that improvement process.

So I think of it more like having an engine of progress that keeps accelerating, and Spud is just one node along that road.

Alex:
So what do you think it can do that today’s models can’t?

Greg Brockman:
I think it will be able to solve harder problems and become more fine-grained. It will better follow instructions and better understand context.

People sometimes talk about a feeling called “big model smell”—meaning when models are genuinely smarter and more capable, you can noticeably feel it. They’ll follow your intent more smoothly and align more closely with your needs.

When you ask a question and the AI hasn’t actually understood what you meant, that feeling still disappoints people a lot. You can’t help thinking: you should have been able to figure this out.

So I’d say, in a sense, this is a “qualitative change” accumulated from lots of “quantitative changes.” On the one hand, there will be improvements across many metrics. On the other hand, entirely new scenarios will emerge. Things that used to keep you from using AI because it wasn’t reliable enough—you’ll start using it without hesitation.

I think this will be a comprehensive change. I’m especially excited to see how it will continue to raise the ceiling of capability. We’ve already seen it perform in scenarios like physical research, and I think next it will be able to solve more open-ended problems and span longer time horizons.

At the same time, I’m also excited to see it raise the floor of capability—meaning that no matter what you want to do, it will be much more useful than today.

Alex:
But for ordinary users, it’s sometimes not easy to feel this kind of change. Before GPT-5 was released, the outside world already had lots of hype and expectations. But when it actually came out, the initial public reaction was, to some extent, kind of disappointing. Only later did people gradually discover that in certain specific tasks it’s actually very strong.

So for this next generation of models, do you think the improvement will be clearly felt mainly in certain professional scenarios, or will it become a more intuitive, broadly perceptible upgrade for everyone?

Greg Brockman:
I think the story might be similar. After the model is released, some people will immediately feel that, compared to what they’ve seen before, it’s like night and day. But there will also be application scenarios where the bottleneck wasn’t in “intelligence” in the first place. If you just make the model smarter, users might not instantly notice the difference in those places.

However, over time, I think everyone will end up feeling the change. Because what really changes is: how much you start to depend on this system.

If you think about how we interact with AI today, every person has in their head a mental model of what “it can do.” That mental model doesn’t change quickly. Usually it’s only when experience accumulates—and it occasionally does something truly magical on your behalf—that you suddenly realize: it can actually do that. And I didn’t even think it was possible.

For example, we’ve already seen similar cases in the context of medical information gathering. I have a friend who used ChatGPT to understand different treatment options for his cancer. His doctor had already told him it was late-stage and there was nothing they could do. But he used ChatGPT to research many different approaches, and in the end, he truly found a treatment plan.

Situations like this actually require a prerequisite: you have to have some degree of trust in the AI’s ability in that scenario, before you’re willing to invest that much effort in extracting value from the system.

So what we’ll likely see next is: in any similar application scenario, the fact that AI can help you will become more and more obvious to everyone.

So this is both the technology getting stronger, and our understanding of the technology catching up to it.

Alex:
So you’ll increasingly rely on it. Inside OpenAI, you’re also developing an automated AI researcher, which is said to be rolled out this fall. What exactly is that?

Early stage of AI “takeoff”

Greg Brockman:
From the overall trend, I think we’re currently in the early stage of this technical takeoff.

Alex:
What does “takeoff” mean?

Greg Brockman:
Takeoff means AI keeps getting stronger along an exponential curve. Part of the reason is that we’re able to use AI to help us improve AI itself, so the whole R&D process accelerates.

But I think “takeoff” isn’t only a technical matter. It also means releasing real-world impact. Many technologies develop like an S-curve. And if you view multiple S-curves over a longer time dimension, they eventually converge into something approaching exponential growth.

I think we’re in a stage like that now. In other words, the technology itself is pushing forward at faster and faster speeds, and this progress engine keeps accumulating momentum.

At the same time, in the external world, a lot of tailwinds are forming: chip developers are getting more resources to invest; lots of people are building applications on top, trying to embed AI into different scenarios and find where it fits with specific needs.

All this energy keeps accumulating, collectively pushing AI into a takeoff period—turning it from a marginal presence into a main engine for economic growth.

And this isn’t just happening within these walls. It’s about how the whole world—how the entire economic system—moves together and how the practicality of this technology keeps moving forward.

Alex:
So what exactly will this “researcher” do?

Greg Brockman:
This “researcher” is essentially about: as the percentage of tasks AI can take over keeps increasing, we should allow it to run more autonomously.

Of course, there are many things that need careful thinking behind this. It doesn’t mean: we send it out to run on its own for a while, and then bring it back to see whether it produced good results.

I think we’ll still be deeply involved in managing it. Like today—if you bring on a junior researcher and leave them on their own too long, they’ll most likely end up on a path with little value. But if you have a senior researcher, or someone truly direction-aware guiding them, they don’t even necessarily need to personally master every specific operational skill. They can still continuously provide feedback, review outputs, and offer guidance on what direction you want them to pursue: what exactly do we want you to accomplish?

So the system I imagine is a set of mechanisms we’re building. It will greatly increase the speed of producing models, help new research breakthroughs emerge, and make these models more useful—better in the real world. And all of this will happen at an increasingly fast pace.

Alex:
What exactly will it do? Will you tell it directly to “find AGI,” and then it will try on its own?

Greg Brockman:
In some sense, yes—that’s how I understand it at least at the first level. But from a more practical standpoint, I think of it as: taking the end-to-end workflow of a research scientist and moving as much of it as possible into a silicon-based system for execution.

Alex:
Here’s another way to interpret “takeoff”: progress in AI becomes not just incremental improvements, but the accumulation of momentum, evolving into a kind of propulsion process that’s almost unstoppable—moving toward intelligence that’s even beyond humans.

Would you worry that, just as things might be moving in a good direction, this progress could also get out of control or go off course?

Greg Brockman:
I think it will, of course. That’s inevitable. I believe that to actually get the benefits of this technology, we must take its risks seriously as well.

If you look at how we develop technically, you’ll see we invest heavily in safety and protection. A great example is prompt injection attacks. If you have an AI that is very smart, capable, and connected to many tools, you obviously want to ensure it won’t get derailed or manipulated because someone gives it a strange instruction.

That’s what we put a lot of effort into, and I think we’ve achieved very good results. We also have a very strong team responsible for this area.

Interestingly, some of these issues can be analogized to how humans behave. Humans are also susceptible to phishing attacks; they can be misled; and they might act without understanding the full context.

We’ll bring these analogies into our R&D. Every time we release a model or develop one, we think about how to ensure it’s truly aligned with human goals, and how to ensure it can actually help. This is something we care very deeply about.

Of course, there are also bigger issues that involve the whole world and the whole economy: how everything will change, how everyone can benefit from this technology. These aren’t problems that are only technical, and they can’t be solved by OpenAI alone. But yes, I do think about it constantly—not only pushing the technology forward, but ensuring it creates positive impact that matches its potential.

Alex:
The problem is, this looks like a race. What happens inside the walls of the OpenAI headquarters gets quickly replicated by many open-source players. And often those players are weaker when it comes to safety boundaries and protective measures.

I remember you once said something—roughly: creative outcomes may require a lot of people getting many things right, but destructive results might only require a single malicious actor. That’s probably the part I worry about the most. Because obviously it’s a race, and things are moving fast. Many of your peers have said that if everyone agreed to stop, they would be willing to stop. But now it doesn’t seem like this race is slowing down at all.

So is this level of payoff really worth taking on that kind of risk?

Greg Brockman:

I think the payoff is worth it. But I also think that kind of response is too rough and too all-or-nothing.

Since the start of OpenAI, we’ve been asking: what kind of future is a good future? How should this technology truly improve everyone’s situation?

You can break it into two perspectives. One is a “centralized” perspective: to make the technology safe, the best approach is for only a single主体 to develop it. That way there’s no competitive pressure. You can take things slowly and carefully get them right, and only when you’re ready decide how to deliver it to everyone. That view is understandable—but in some sense, it’s also a方案 that many people find hard to accept.

The other path—the one we’re more inclined toward—is thinking from the standpoint of “resilience.” In other words, treating it as an open system: many participants are pushing the technology forward, but the focus isn’t only the technology itself. It’s also building social infrastructure around the technology so it can be carried more safely and more reliably.

Think about how electricity developed. Electricity is produced by many different people and institutions, and it also has risks and dangers. Meanwhile, we built multi-layered safety infrastructure around it: electricity safety standards, different usage regulations, and different forms of oversight depending on scale. At very large scales, there are even specialized regulatory requirements. Many people can use electricity in a democratized way, with inspectors and an entire ecosystem of supporting systems, gradually built around the characteristics of this technology.

And I think AI is similar. What we’ve truly seen is that around AI, there has to be a broad social discussion. If this technology really arrives and changes everyone’s lives, then people have to be part of it. It can’t be decided and pushed secretly only by some centralized small group.

So for me, this has always been a core question: in what way should this technology be rolled out? And what we truly believe in is the “resilience ecosystem” that gradually forms around the technology as it develops.

Alex:
So you mean that we’re in the process of takeoff right now—and that all of us are already part of it. Nvidia CEO Jensen Huang recently said he believes AGI has been achieved. Do you agree?

Greg Brockman:
I think AGI has different definitions for different people. And yes, there will be plenty of people who think that what we already have in our hands today counts as AGI.

This can be debated. But I think what’s genuinely interesting is that the technology we have today is still very “non-smooth”—with noticeable discontinuities.

In many tasks—writing code, for example—it’s absolutely superhuman. AI can do it, and it does significantly reduce the friction of creating things. But at the same time, there are also some very basic things that humans can do effortlessly, while AI still struggles with.

So where exactly do you draw the line? To some extent, it’s more like a “feeling,” a judgment about the atmosphere—not a question that can be strictly and scientifically defined at this moment.

So for me personally, I think we’re clearly at that moment. If you showed me these systems from five years ago, I would have said: yes, that’s the thing we were talking about back then. It’s just that the reality looks very different from what we imagined. It doesn’t resemble any of the forms we used to envision.

So I think we need to adjust our mental model accordingly.

Alex:
So you mean it’s not there yet?

Greg Brockman:
I’d say it’s already at around 70% or 80%. So I think we’re already very close.

And I think there’s one thing that’s extremely clear: in the coming years, we will get AGI. Its performance might still be somewhat “jagged”—not fully smooth everywhere, not perfect across the board. But the lower bound of what it can do will be raised very high—almost for any intellectual task you need to complete on a computer, AI can handle it.

So now I have to give an answer with a bit of uncertainty, because it has some of that “uncertainty principle” flavor. You can argue from different definitions. But based on my personal definition, I feel like we’re nearly there. Take one more step forward, and it’s absolutely there.

Key inflection: from 20% to 80% of work being taken over

Alex:

What happened in December 2025? Because that looks like a turning point. The thing about “not being interrupted and writing a few hours of code continuously” suddenly seemed to shift from a theoretical idea into everyone saying, “I think I can trust it and let it keep running for a while.”

So what exactly happened then?

Greg Brockman:
After the new models were released, the proportion of tasks AI could complete went from about 20% of your work all at once to 80%. That’s an enormous shift. Because it’s no longer just “a pretty good small tool.” It becomes: you have to reorganize your workflow around these AIs.

For me personally, there was a very typical “felt it in my body” moment. For years, I’ve had a test prompt: “Have the AI build a website for me.” That website was something I had built by hand when I learned to code years ago, and it took me months.

By 2025, it still took around four hours and several rounds of prompting to make it look reasonably good. But by December, I asked only once. The AI produced it in one go—and it did a really good job.

Alex:
So how did these models accomplish that leap?

Greg Brockman:
A big part of it is that the base model itself got stronger. OpenAI has been continuously improving its pretraining techniques. At that time, for the first time, we saw a glimpse of what will happen for the rest of this year. But it wasn’t just a single, isolated breakthrough. More accurately, we were pushing forward across all dimensions of innovation.

One interesting thing about these models is that, in a sense, you can feel “jumps” as they appear. But from another angle, everything is still continuous evolution. It’s not suddenly jumping from 0% to 80%; it’s going from 20% to 80%. So to a certain extent, you can say it’s just getting better.

And I think this kind of progress continues through every subsequent smaller version update. For example, from 5.2 to 5.3: I worked closely with an engineer. He originally had no way to get the model to do that kind of bottom-level, hard-core systems engineering work he was responsible for. But in the new version, the model could take over his design documents, actually implement them, add metrics monitoring and observability, run profiling for performance analysis, and continuously optimize—finally achieving the exact result he had originally hoped to deliver himself.

So I’d describe it as: progress that feels like “slow moving, then suddenly everything changes.” But all of it had already been foreshadowed by the capabilities that were working in the present. At the latest within a year, many things—some even faster—will become extremely reliable.

Alex:
Does that surprise you personally too? Because I remember in an interview not long ago, you said that automation coding tools like Codex were originally meant for software developers. But earlier in this conversation, you said that actually everyone can use tools like this.

So what changed your mind?

Greg Brockman:
To be honest, I used to understand Codex within the “writing code” framework. After all, it literally has “code” in its name, so it’s natural to view it as a tool for programmers. And inside OpenAI, many people are software engineers—we were building tools for ourselves, so thinking about it that way was very natural.

But as this technology kept improving, we started to realize something: most of the underlying capability we really built isn’t fundamentally about “code.” It’s fundamentally about “solving problems.”

At its core, it’s about managing context, building execution frameworks, and thinking about how AI should plug into real work and actually get things done. Once that’s true—even in programming scenarios—it also means anyone can gain this capability. Because what you truly have is a system that can execute work for you. If you have a vision and a goal you want to accomplish, and you can describe your intent clearly, the AI can execute and get the job done.

But it also makes you ask: why am I only focusing on the “non-programming” versus “programming” split? There’s actually a lot of work that is fundamentally just mechanical skill. For example, spreadsheets in Excel or making presentation slides. If the AI already has enough context and enough raw intelligence, it can do these tasks really well now.

So if we just make it easier to access and more friendly to people, Codex goes from “Codex is for programmers” to “Codex is for everyone.”

Alex:
After seeing this wave of obvious improvements, another almost silent phenomenon quickly appeared in Silicon Valley: Open Claw, right? Or more broadly, the whole tech industry is starting to trust AI in the way you just described—like giving a desktop control to an AI robot, or setting up a Mac mini, granting it permissions for email, calendar, and files, and then having it “take over life” to some degree.

Later, OpenAI hired Open Claw’s founder into the company. So can you tell us more about this kind of AI that helps manage your life? Bringing the Open Claw team in—does that correspond to that vision behind it?

Greg Brockman:
I’d say the most core part of this technology is figuring out exactly how it becomes useful—how people actually want to use it, what the agent vision really is, and what way it should enter people’s lives. Those are inherently difficult questions.

And one thing I keep seeing across these generations of technological evolution is that people who are truly willing to go deep, who are full of curiosity and have strong imagination—those traits are themselves a very real kind of capability. And they will become increasingly valuable capabilities in the new economy.

Open Claw founder Peter, in my view, is exactly that kind of person. He has very strong imagination and a strong creative drive. So in a sense, it’s related to a specific technology—but in another sense, it’s really not only a technical question. It’s about how we embed these capabilities into people’s lives and find where they truly fit.

So as a technologist, it’s definitely exciting. But as someone who truly cares about delivering practical value to users, we’re also increasing our investment in this area—investing a lot.

Alex:
You recently said something interesting about this. You said that once you start letting these autonomous AI agents work for you, you become the “CEO of a fleet of thousands of agents” that complete your goals, vision, and tasks, and you yourself are no longer stuck deep in the details of how all the specific problems get solved.

But you also said that, in a sense, this new way of working can make people feel like they’re losing their “pulse” for the problems themselves.

Greg Brockman:
Is that good or bad? I think it’s a trade-off—both pros and cons.

So what we should do is, on the one hand, acknowledge the power these tools bring. On the other hand, we should do everything we can to mitigate their weaknesses. For example, increasing leverage for people—giving them greater ability to act. If you have a vision and something you want to accomplish, you can marshal an entire fleet of agents to do it for you. That’s obviously incredibly powerful.

But if you think about how the world works, at the end of the day there’s still someone responsible. Suppose you’re building a website, and your agents mess things up, and it affects your users. Strictly speaking, that isn’t the agents’ fault—it’s yours. So you have to care about that.

I think anyone who truly wants to use these tools has to recognize that human agency and human responsibility are core components of the whole system. How humans use AI is itself something fundamental.

So the most important point, I think, is this: as the user of these agents—this is also how it works inside OpenAI—you can’t give up responsibility. You can’t just say, “AI will take care of it.”

Alex:
Sure. But what you just said about “feeling like you’re losing the pulse for problems” seems not to be the same thing as “responsibility.”

Greg Brockman:
For me, they’re actually connected. The point is: if you’re the CEO but you’re too far from the details—for example, you’re managing a team or running a company, but you’ve already lost awareness of what’s happening on the front line—that usually doesn’t lead to good outcomes. So what I meant isn’t that it’s something worth pursuing for humans to finally not need to know anything.

Of course, some details can be delegated. Like hiring a general contractor to build your house—you probably don’t need to personally oversee a lot of those details because you trust them to handle it. But ultimately, if some critical details go wrong, you still should care and you still should know.

So there’s an important nuance here: you can’t just blindly say, “I’m willing to lose that sense of grasp on the problem.” Instead, we should actively say: I still need to maintain that awareness—to truly understand the system’s strengths and weaknesses.

And when you start pulling away from some lower-level, more mechanical tasks, the reason you can do that is because you’ve built trust with the system—that it will indeed get things done.

Alex:
On models, my last question: earlier you mentioned one model evolution path—pretraining, then fine-tuning, then reinforcement learning—to make it better at solving problems step by step, and to connect it to execute tasks on the internet.

And now we’ve reached a stage where models have learned to use tools through that process. If I understand correctly, what would be the next step in this evolution path?

Greg Brockman:
I think the world we’re in now is one where machine capabilities keep deepening and expanding. Part of that is related to tool use. But at the same time, we also need to genuinely make the “tools” themselves good enough. For example, if AI can already “operate a computer,” using a desktop system the way humans do, then in principle it can do anything you can do.

But at the same time, we also have to add a lot of infrastructure layers for the machine. For example, in an enterprise environment: how do you do authentication and permissions management? How do you do audit trails and observability? To keep up with the development of the model’s underlying capabilities, a lot of supporting technologies need to be built. And from the overall direction, I think the next steps will include things like a “very natural voice interface”—meaning you can naturally talk with the computer like you do now, it can truly understand you, do what you need, and give valuable advice.

For example, it can proactively remind you: the thing you’ve been working on is stuck now; the problem is here. Or when you wake up in the morning, it can tell you: here’s your daily briefing—how much work your agents pushed through overnight.

Maybe it may even already run a business for you. I think that will be a huge application scenario for this technology. Democratization of entrepreneurship will absolutely happen. It will tell you what went wrong in those places. A customer is currently very dissatisfied and wants to talk to a real person—so you’d better handle it personally. Those kinds of things will happen.

And then I think the next stage also includes raising the upper bound of goals that humans can challenge—continuing to elevate it with this technology. We’ve already seen the frontier of this trend. The most exciting thing is something I can almost analogize to AlphaGo’s 37th move: that move is one humans would never have played. It has creativity, and it changed a lot of people’s understanding of that game.

Something like this will happen in every domain. It will happen in science, math, physics, chemistry. It will happen in material science, biology, healthcare, and drug discovery. It may even happen in literature, poetry, and many other fields. It will unlock new space for human creativity, understanding, and ideation in ways we can’t even imagine today.

Alex:
But if models are already as strong as you say, why hasn’t this happened yet in a real way?

Greg Brockman:
I think there’s a “capability lag gap”—meaning there’s still a big distance between what the models actually can do and how people are currently using them. To some extent, even our understanding of what’s truly “inside” the models is still forming.

So I think even if the technology doesn’t improve further from here, the world would still undergo a major change—an economy driven by compute and by AI will still come.

But there’s another layer as well: what we’re currently best at is training models on tasks that can be “measured.” At the beginning, we started with math problems and programming problems because those tasks have clear validators: whether the answer is correct can be judged very clearly. And in the past stretch, the reason we’ve been able to gradually move models toward more open-ended problems is that we’ve continuously expanded the scope of what can be validated and evaluated.

And AI itself can help with doing this. If the AI is smart enough and it understands the task enough, you can give it an evaluation standard and it will gradually learn. But for tasks like creative writing—like “is this poem good?”—it’s difficult to score.

So in these kinds of scenarios, we’ve historically found it hard to get the AI to truly learn through continuous attempts and feedback. But all of this is changing, and our path ahead is now pretty clear.

Alex:
That’s pretty interesting. Peter Thiel previously said something like: if you’re good at math, the impact you get from these models might actually be larger than that for someone good at words. And you were also a member of the Math Club back then. Aren’t you concerned about that?

Greg Brockman:
I think people are more likely to notice what they’ve lost than what they’ve gained. Because we have very deep lived experience in how we previously did this. For example, I used to participate in math competitions, and now AI can do them too. But the issue is that this has never really been about the math competitions themselves, right? It’s not the core thing driving human progress.

If you look at how we live today—sitting in front of a box, typing into another box—we don’t live like that a hundred years ago. This isn’t a natural state. It isn’t what the digital world is truly supposed to be.

That’s not the most essential part of “being human.” What’s truly important is being present, living in the moment, and connecting with other people.

And I believe what we’re about to see is that AI will free up a huge amount of time, giving humans more opportunity to strengthen connections with each other and build more bonds between people.

That is what excites me a lot.

Alex:
Okay. When you shift further toward more agent-like application scenarios, an issue has started getting discussed externally: will we still need to do such huge training tasks in the future?

Especially when models are already good enough, it seems you can let them directly enter the real world, and in many stages that don’t depend much on pretraining, you can get a large portion of improvements. And the parts that truly require massive data centers to support are mainly pretraining.

You’ve been responsible for scaling up and pushing this forward. What do you think about this claim?

Greg Brockman:
I think that claim ignores something extremely important in technical evolution. In fact, every step in the model production pipeline amplifies the effects of the other steps. So you’ll want each step to get stronger.

What we’re seeing is: once pretraining becomes stronger, every subsequent step becomes much easier. That’s pretty logical. Because the model starts with more capability, it learns faster. When it tries different approaches and learns from its mistakes, it also pushes forward faster because its foundation is stronger, and it makes fewer mistakes.

So the real change isn’t that we move from “training a purely closed, self-reasoning rational system” to “just letting it experiment in the real world.” Instead, we’ve realized we should not only make the models bigger and stronger, but also let them try things—to understand how people actually use them in the real world—and feed those usage feedback back into the training process. And doing that doesn’t diminish the value of continuing research that part. It doesn’t undermine its importance either.

One more change is that in the past, we focused mostly on improving the raw capabilities in the pretraining stage, but didn’t emphasize inference (the inference stage) as much. Over the past 24 months, a big shift has been that we started to realize that there needs to be a balance between the two.

That is: you can have a base capability model that’s very strong, but it also has to be efficient enough during inference and during real operation. Because if you’re doing reinforcement learning and you’re deploying it into the real world, all of that requires strong inference efficiency.

And that also means you might not necessarily push training scale all the way to the theoretical maximum. You also have to consider the massive usage scenarios that come after.

What you truly want is the point that optimizes the product between intelligence level and cost—not just optimizing one dimension.

Alex:
If the future shifts primarily toward inference, does that mean you won’t need Nvidia GPUs as much?

Greg Brockman:
We still need them very much.

Alex:
Why?

Greg Brockman:
There are many reasons.

One is this: no matter how the ratio between training and inference changes, training at ultra-large scale still has to be done by concentrating massive compute on a single problem—and today there’s no alternative way to do that.

So I think what’s more likely is that the proportion of compute on the deployment side will increase significantly. But at the same time, there will still be moments when you have to run a particularly huge pretraining job, and you’ll still need to concentrate a large amount of compute then.

And I also think Nvidia’s team is really excellent—they’ve done truly astonishing work. So yes, we work very closely with them.

Alex:
Is there a day when people start saying, “We’ve pretraining enough, and the model is smart enough”?

Greg Brockman:
I think that’s a bit like saying: once humans have solved all the problems in front of them, maybe we’ll say that. But I think the ceiling of what we want to accomplish is much higher.

Over the past 50 years, in a sense, our ambition for many goals has receded. For example, can we ensure everyone has healthcare coverage? Not just treating problems after they happen, but actually doing preventive medicine—paying attention to lifestyle and helping people early, catching potential risks before disease occurs. I think we can actually solve problems like that by leveraging smarter models.

Of course, there might be some level where this problem is fully solved, and then you might ask: do we still need a model twice as smart? But at the same time, there will also always be other problems that dema