OpenAI 总裁 Greg Brockman：重塑网络安全与 AI 驱动的科学

Greg, thank you very much for coming. Thank you for joining us.

1. Trailer

Thank you for inviting me.

2. AI and Advancements in Mathematics

Every week in AI feels like a whole year, and a lot happened this week for OpenAI. You've introduced GPT-6 to over 1 billion people, and that's just amazing. There is very interesting news regarding mathematical achievements, and there is a lot that is undoubtedly happening within the company. What excites you most about what's happening right now, both at OpenAI and in the wider AI world?

You know, I have a belief about how OpenAI works. It's like a quote about cycling: It never gets easier; you just get faster. I think that's exactly what's happening. There's so much activity, and every time you think you've reached the peak of influence, you realize that an even higher level is possible.

It's not just about one company or one technology, but about us, as a world and society, coming together to implement this technology and get the most out of it to improve everyone's lives.

To me, this week's math results are a real milestone. This is a real moment where, if you think about the fundamental driver of improving the quality of human life over the last 1,000 years, it's technology. This is scientific progress, and I believe that we now have the opportunity to significantly accelerate scientific discovery.

The question always arose: Is AI really capable of creating new knowledge? At this point, it's simply impossible to deny it, right? That's definitely true. It's possible, although there is still a lot of work ahead. I think we're not creating completely new theories yet, but we are capable of solving century-old mathematical problems.

The most interesting thing for me is that it's not just about the machine giving you the answer. It's about what a person armed with these tools can achieve that was previously unthinkable. I think the potential for people to do whatever they want is growing rapidly.

If you have a math problem you want to solve, you can do that, but what about applying those same capabilities to biology? What if you are a small-business owner and want to manage your business effectively? This level of intelligence, tuned to a particular industry, will mean that incredible results aligned with your goals and desires will now be available to everyone. For me, that's extremely exciting.

The pace of progress is also impressive. There were problems with arithmetic before, right? I remember when people complained about early models: “Why can't it count basic things?” And now you're solving complex theorems.

As these technologies continue to evolve and begin to have a real impact on mathematics, physics, materials science, and various application areas, is there a need, if these models become so intelligent, to focus on how to effectively communicate this to people? How do you think about the communication side when you prove a very complex theorem? If you look at any branch of mathematics, only a very small number of people can really appreciate a theorem created by a human, right?

This is something that has always been close to me because I have a mathematical education. I thought I would become a mathematician, and I was very inspired by figures like Galois and Gauss. These people were working with a planning horizon of 100 or 400 years.

I think there's something deeply beautiful about the fact that mathematics is about understanding the structure of the universe. It's not just about that, because questions always arise: Was mathematics invented or discovered? In my opinion, the beauty and meaning of mathematics is not objective in a certain sense. It is in the eye of the beholder.

It's something that attracts us, and there's mathematics to it, similar to the laws of physics. It's about simplicity, compression theory, complexity theory, and all that. Maybe that's more physics, but there's something important and human about it too.

For me, it's not just about saying, “Here's the proof. Nobody understands it. Great, we solved the problem.” It's really about the ideas that are being received. What we're going to see—and where there are already encouraging signs of life—are the beautiful ideas contained in the work of these AIs.

3. AI Communication and Oversight

We've published almost 400 very important open problems that are now solved, and the mathematical community is studying all of them for ideas. If you read what experts in these fields write, they say that these are very unexpected ways of applying ideas, or incredible “Wait, why did they use this result here?” moments.

What comes out as a result are not incomprehensible theorems or proofs, of which, by the way, there have been many in the history of mathematics. The four-color theorem is a perfect example: It's just a machine checking a bunch of cases. There is no beauty in that.

I think we're entering another phase of machine discovery. I have some optimism that as models develop a better theory of mind—the ability to understand what we as humans know and guide us along the path to new understanding—that will also be revealed.

I'm a person who loves math, but at some point in school I stopped studying it. Then I went through a mathematical academy, and I also used ChatGPT to learn new math. I feel like the expectations for the increasing level of knowledge that people need to understand will also be there.

I think this is a very important point and one of the perhaps underappreciated benefits that I expect and hope for, and that we're starting to see from this technology: helping people know more, understand more, and better understand each other and the world around them.

I think there are technical mechanisms that could underlie AI and solve these problems, such as seeing connections between industries. How does it actually communicate that to you? How do you understand it? In fact, there are even connections to AI security and consistency, so that we can understand what these systems are doing.

There are techniques that we've been talking about since 2017, like debate and iterative reinforcement, which are two sides of the same coin when you think about what to do if AI solves tasks that are too difficult for humans to verify.

4. AI for Overcoming Biology Data Bottlenecks

Even creating a book summary is a task that's very difficult to verify, isn't it? It's really hard to know if a summary is good unless you've read the entire book. If an AI just gives you a summary of a book and says, “Here's a summary,” is that a good summary or not? This is quite difficult unless you have another AI trying to prove that the summary is bad.

If that AI comes back and says, “Here's this paragraph in the book that I found where there's a clearly important plot twist that's not reflected in the summary,” you don't have to read the whole book to say, “Oh, yeah, that's a really strong argument.”

You realize that there are ways of training AI that are very aligned with allowing humans to supervise and stay in the driver's seat. That's a fundamental value that's important for the future from an OpenAI perspective, but it also fits very well with what you're saying: Can you make the output of AI something that allows you to learn something tailored specifically for humans, rather than just knowledge for the sake of abstract knowledge?

For most people, advances in mathematical theory still remain somewhat abstract, right? It's a feeling of, “Wow, this is really happening today.” And there's also AI for science. There's a claim or a belief that it's going to be able to cure all diseases within a certain number of years, right? How do you see the steps forward in terms of what kinds of achievements we can see in the interim?

I think we've had a few years of iterations of this technology to see what to expect. It's similar to when you have an AI that's able to operate in a certain dimension. If you understand why it became capable in that dimension, then in all other spheres of human activity that have similar properties, you should expect similar results.

I think that's why mathematics is something that we've been able to transform and achieve such incredible results in. This is partly because our AI is now at a certain level of capability and intelligence.

Part of this can be explained by the fact that mathematics can be tested, but that's not entirely true, is it? We really started with very simple tasks to check: “Here's a written arithmetic problem; there's a certain numerical answer.” You have an AI that can be trained to simply check whether it got the correct answer.

We can very easily see whether the answer that AI suggests is the same as the actual answer. But with evidence, things are different. The evidence is very difficult to verify. Is this actually accurate or not? A formalized proof—well, yes, it can be tested, but a natural-language proof, which is what we're actually creating here, is much, much more difficult.

But what I think is the key understanding is data, right? You can get a lot of data in mathematics and a lot of data in theoretical physics. Is it possible to get a lot of data in biology? This is obviously a very data-dependent industry, and it's very challenging. Fortunately, we have also been working on many methods to be able to work with less data.

And if you think about what that means, now we have smart models. If there's something where there's a lot of data, that's the fundamental physics of deep learning, right? Compute, data, and algorithms equal capability, equal results. We're able to do this in complex areas where our smartest people have spent hundreds of years trying to solve a problem, and now AI can solve it—surprisingly.

When you translate this into biology, it's clear that we want to discover new drugs and make all those fundamental changes in biology and our understanding of it to improve the quality of life. That's what we want to apply. The question that remains unresolved is whether we have this data. Is there any way we can get this data?

5. Revolutionizing Drug Discovery

So I think what's going to happen is we're going to take these models—and we're already doing this—these same methods, and apply them to the biological realm. There are some industries that will succumb quickly, right? We will see a massive acceleration of what was previously incredibly difficult, and now becomes incredibly easy. But there will be other bottlenecks, and the question will arise: How do we eliminate them?

Some of them, by the way, don't even relate to deep biological theory. For example, regulatory documents. That's a huge bottleneck, isn't it? It takes about a year to write the whole thing for a clinical trial and all that stuff, but you can write such things much more easily with AI. So I believe we need to consider every step of the drug-development process.

There will be places where fundamental data simply doesn't exist. Innovation in the labs, innovation in how you do simulations—all of that is going to be extremely important. I think we'll see these breakthroughs, but it's important to know that there are commonalities with mathematics. These are scientific discoveries, and there are things that are very different depending on where the data comes from.

Are there any specific areas—you mentioned biology, and I think that's a great example—where a lot of early research, like AlphaFold, was focused on designing molecules or understanding how things interact with each other? As you noted, it takes 15 years to develop a drug, and it costs $1.5 billion. About $50 million is usually spent on preclinical research, mouse studies, and the molecule itself, and then over $1.4 billion goes to clinical trials. So your point is correct, and besides, it's 2–3 years versus the next 12 years.

Your argument is very apt that the early focus of these models was on whether a substance could bind, versus what else was worth worrying about at later stages to turn it into a useful product for humanity. How do you at OpenAI currently rank the different areas of scientific research in terms of what you are focused on, or where you want to get more data, results, or opportunities that others can benefit from?

Look, I think drug development is the most obvious area, if you had to pick one that you want to completely revolutionize. These are exactly the areas where you will truly improve everyone's quality of life. This is so important, and it's probably the main application of AI that everyone has been thinking about for a long time, so we definitely have to do it.

Regarding your comment, there's a whole conveyor belt there, and it's just so long. I think we're going to find that humans are incredibly bad at biology and incredibly bad at drug development. If you think about it, there is a property that is natural to humans—we are very well adapted to it. It's something that is very distinctly human and very well tuned to how we interact with the world, whereas biology is completely foreign.

6. Changing Regulations to Accelerate Progress

But for a neural network, I see no reason to believe that it should be better at natural language than it is at biology. So the fact that we're seeing such incredible results with neural networks in language means that, from first principles, you should be able to decipher the language of biology. For me, number one, this is what should work in principle, and from what we're seeing, we're so close in many ways. The chances are high.

Yes, it is truly fascinating. I think that parallelism can be significantly increased. I think you can increase parallelism, and I think you can reduce so many of these stages in terms of their duration. There are many dimensions to drug development, for example. For one patient's illness, there may be many people for whom drugs that will cure their specific condition already exist. They're just somewhere in the literature—there's an article—and if you can find the article that deals with their condition, then you can cure the person.

There are many precedents for this. I think you can scale up operations like this. Therefore, I believe that such an impact is quite possible in the short term. Materials science is another example, and I think there are a whole bunch of other areas where we should see real changes in what humanity is capable of achieving.

It seems worth considering why it can take 15 years for a drug to reach the market. Part of this is because you're doing increasingly expensive trials with different safety requirements that lead to large-scale clinical trials, and these trials are now regulated, right? And for good reasons. It seems like one of the limiting factors, if you believe the technology will work and some of the data can be obtained or generated, will be the regulators' belief that we need to change the expectations about what trials are accepted and what the stages should be.

I think there's a point now where, although there's still a lot of work to do to get there, if you can really prove that the physics of drug development have changed, then of course that should impact every part of the system.

And if you think about it, even to take a completely different example, think of driverless cars, right? So much of how roads function and how safety is provided is geared toward human drivers. If you get to the point where self-driving cars are 10 or 100 times safer than humans, once you get to the point where AI can actually deliver the industry outcomes that are so clearly desired, of course you have to say, “Okay, how do we rethink the entire transportation system?”

It may be worth introducing regulations that would allow the use of driverless cars only in certain areas. There are many potential changes in how you perceive what should come out of the system, and that should also come back into the regulatory realm.

7. Dots the Proactive Agent

And of course, regulation is not the same all over the world, is it? There are very different regulations in Australia and the US regarding clinical trials, and we are already seeing many moving their trials to where the regulatory environment better aligns with how they want to operate and what people believe is effective.

So I think exactly how things are going to change will depend on what we can actually prove, right? Can we really get to the point where a neural network will produce better results than conducting clinical trials? We're not there yet, but it's conceivable that we'll get there. If they are better at predicting efficacy or toxicity, then of course it's worth changing. This is definitely true.

And I believe that in those areas where society, in my opinion, should and, I hope, will actively use new opportunities, everything has changed radically, so let's adapt quickly. Again, you have to really think through all the implications and all the influences. If you think about long-term studies in humans, the question becomes: Does this neural network actually accurately predict outcomes in the long term? It's pretty hard not to have at least one long period of time to test this. But I think we'll see 80% of what's possible come to fruition very quickly.

You've been around for a while now, since the release of Astra and DevDay. Is there anything about how people inside or outside the company are using new products or capabilities that surprises you?

Well, I think Dots has really changed the way many of us work.

And what Dots embodies is Describe Dots briefly.

Yes. Dots is a proactive agent. It's always on, always working. You stop working, and it continues. You know what my Dot does? I have a Dot that keeps track of work I had to do, like messages I had to send and things I promised to do but haven't done yet.

8. Unifying Work and Consumer Agents

I always wake up in the morning and there's a list of things, and I think, “Oh, yeah, that's my to-do list.” Often, it notices things I didn't even ask it to care about. If there is a scheduling conflict in my calendar, it will point it out. For example: “Hey, you really need to leave this meeting early so you can get to something else.”

It's so contextually aware. It knows what I do during the day. It understands the projects we're working on. It knows what's in my inbox and what's in my Slack. It's able to integrate all of that, and it knows what I've been asking about, so it understands what topics are most important to me.

And I think having Dots is, to some extent, just a change in form factor, not capabilities, right? It's like an Astro level model.

He has his own car computer. But this form-factor change in a work environment, with this level of intelligence, is a real breakthrough. Do you have multiple Dots running, or just one that covers all your background tasks or something like that?

I have one. This is a little owl—this is an owl. You can ask him to change the icon. He just does everything in my work life.

OpenAI obviously has research projects, as well as consumer, enterprise, developer, and platform businesses today. Do you think there will be a separation between, say, assistants that help with work, such as Dots, as I mostly perceive it, and consumer assistants, such as Instincts or Muses?

I think these things will converge, definitely. There are certain factors that are very different in different environments. For a mass consumer product, it's really important to understand how to scale and provide services for free or at a very low cost. Whereas in the workplace, if you care about productivity—whether you're a manager or an individual performer, regardless of the role—it's much easier to say, “Maximum intelligence.” As long as you get a return on your investment, it's worth allocating more computing power.

I think the main difference is really the cost. We saw this in ChatGPT, right? For ChatGPT, we put a lot of effort into creating a massive free tier, and providing features like memory to all of our free users was an incredible task, a truly massive engineering effort. Implementing the same functionality for a very small number of people is much less expensive; it's just very easy.

I think that in the area of personal assistants, we're seeing the emergence of different market segments where fundamentally different amounts of computing power will be consumed. When the levels of computation are different, that means the product might look a little different, right? It may be that one case has more functionality built in than the other.

9. How UI Will Evolve

Of course, there are different use cases, such as payments and travel, that are more consumer-oriented, but I don't think there is any fundamental difference. It's just artificial intelligence that performs certain actions, with a certain emphasis in one direction or another.

One of OpenAI's strengths is that we have 1.2 billion ChatGPT users, as well as a leading business in coding and enterprise solutions. No one else has both at the same time. A big part of our strategy is to bring that together so that we're not competing on 2 fronts. This is actually a 1-front competition.

I think that's the general direction of AI development: to blur the lines between the consumer and corporate sectors. A true renaissance of entrepreneurship awaits us. We're already observing this. Imagine seeing this very specifically: I know people who quit their jobs to start their own companies and say, “I would never have done this if I didn't have AI that I can rely on, that removes so many of the usual barriers to entry.”

I think we're going to see changes. Just as these differences have persisted for so long, they're going to change.

One of the things that has traditionally separated the corporate and consumer worlds is the approach to the user interface. There's your argument about cost and the basic ability to deliver something of a certain quality, but the other part is simply the user interface. Often, the interface in a corporate environment, or the way it is accessed, is very different from a consumer one.

In your opinion, you're striving to unify these levels. You've also been thinking quite a bit about dynamic interfaces. This has always been one of those things that people have been talking about for quite some time: How do you dynamically create an interface that is useful to the user? Or in some cases, do these things simply go into the background and become part of your chat?

How do you think the UI will evolve for these different use cases, and how do you see this as an issue when addressing both of these segments?

First of all, I believe that the best interface is the lack of one.

M-hm.

I think that's exactly what we're moving toward in AI. It's a thing you can talk to; you can communicate with it by voice, through Slack, iMessage, or any other platform. That's literally what you can do with Dots just today.

M-hm.

But there are some fundamental differences between personal use and the work environment. You have an administrator at work, right? You have your company's security service. There are simply more stakeholders, and it's not just an AI for you that represents you, that you own, and whose outcomes you control. I think this is the main difference between corporate and home use.

Of course, this also leads to differences in products, and sometimes they're very minor things. For example, when we release a new feature, corporations want to control exactly when it appears. That's why we have administrative controls and toggles, so you may have an experience with limited functionality because this is a workspace.

That changes the situation: We need a way to notify you about it. And that brings us back to the fact that the main idea was the lack of an interface, right? What do you do if some functionality is missing, and how do you even notify users about it? It's an ongoing challenge, and I don't think it's going to magically go away. You need to accept this complexity.

To me, it's fundamentally crazy and beautiful when you think about it. ChatGPT is just a text box, right? This is where it all started. It was a text field. What we're getting at doesn't even need a text field, right? This is simply the AI we were promised.

When I think about what makes a product without an interface better, you could say, “How much computing power can I spend on it?” How good is the model itself? There is a memory system. There are decisions about behavior. How well do models calibrate in terms of how long it takes them to think about something?

Are we doing this well enough yet? With the release of GPT-6 for everyone, we now have the ability for an AI that has to think hard about a problem to actually give you a quick answer, and then continue to think and come up with a more correct and complete answer.

10. Cybersecurity in the AI Age

Regarding the lack of an interface, and what you mentioned earlier, we have implemented an intelligent interface. Now AI can create an interface on the fly. Even if there's no fixed interface, no buttons and menus to learn, having AI create the result exactly the way you want it visually is important.

Let's say you're thinking about booking flights. You need a complex interface there, right? You want to see all the different options. For any question you might ask, such as, “How do I assemble a bike?” you might need a special application that would previously have taken a person a month to write. Of course, you would never hire someone to write a one-off app to assemble a bicycle, but it's perfectly fine for an AI to do it for you in real time, right when you need it.

I think we live in a world now where AI is much more adaptable to you. The user interface becomes exactly what you need, when you need it. That's a big part of our focus, and we do that in every aspect of how we think about the product.

One of the biggest shifts in the AI debate recently has been cybersecurity: how to build secure systems that are resilient to cyberattacks using AI. I'm curious about how you think about this, how you build defenses, and how you approach the different aspects of cybersecurity.

First of all, I believe that the physics of cybersecurity are changing, and we're in what we call the “defender's window.” The way I think about this window is that we have advanced AI capabilities that can help defenders secure themselves. Defenders are not yet completely defeated, but these capabilities are gradually growing, and this is evident with each release of new models from providers around the world.

The cyber capabilities available to everyone are expanding, and you're starting to see the first signs of their impact, right? You're starting to see, for example, in South Korea, high-profile hacking attacks on banks carried out using different models. I think it's crucial that defenders take advantage of this time.

That's why we work hard to responsibly provide access through our Daybreak program to defenders, so they can use our advanced intelligence and powerful models to find and fix vulnerabilities in their systems.

It's important to realize one thing: These flaws have always existed, right? In a sense, every line of code written by a human begs the question: Is it secure? Was the developer trained properly? You start to realize that the way we built systems was fundamentally insecure. We always knew that, right? There has never been a CISO who had too many resources.

Yes. True?

Remember the XKCD comic about the entire world's infrastructure being held up by a single open-source project written by some guy in a basement, right? Or Heartbleed, or whatever else came before all these AI problems.

That's right.

It's important to realize that these are really human problems, right? This is the way we chose to build the internet, where security was kind of “bolted on” from the top, right? We had some high-profile incidents, didn't we? Think of the early worms and viruses. Then people started realizing that we needed antivirus software, right? We needed encryption, and requests over HTTP had to be made over HTTPS.

All of these mechanisms were ones where, I think, we took the longer route to add security later.

The internet was actually built for resilience, not necessarily for security. I think we're going to see a point now where a lot of this debt—a lot of this technological debt—is going to come to the surface. It's time for us to pay it off. We, as a company, are really doing our best to help at this time. For example, we have committed a billion dollars to help frontline defenders access and use our models for defense.

We study and try to identify the various vulnerabilities that exist, and then we think about how to help the world respond to this. How do we coordinate incident response? How do we cooperate with the government? How do we work with the community to implement this as best as possible to help people navigate it?

But I think people haven't quite realized yet how much the physical principles of cybersecurity will change. I think we have the opportunity to have a much safer world, don't you? Think about how AI allows you to pay off technical debt in a way that was previously simply unthinkable.

The fact that you can do things like formal software verification—these models are obviously great for proving certain things. Why not prove the security properties of software? This is something the technology already existed to do, but it requires too much work. No one does it for anything significant, and so I think all of this could change.

I think the world we're moving into now—and we've already done this within OpenAI—is one where, every time a new model comes out, you have to use it to protect yourself, find vulnerabilities, and then fix them. For example, we took about 25% of our production engineers and said, “All your projects are on hold. Now you're just doing security.”

We took our smartest model at the time, had it test our entire infrastructure, red-teamed a lot of things, and built tools to make it easy to replicate in the future. We found a number of issues, fixed them, and eventually eliminated all the vulnerabilities that this model was capable of finding. That's a very good story, isn't it?

I think it's important to realize that we will have the next model, and we'll have to repeat this process. It will continue to be like this. But ultimately, if you think about the code that's being written now, most of it is being created using AI, and we can actually train these models to write secure code from the start.

I think we can get fundamentally safer systems by using and applying these AI models properly. It looks like we're in for a period where we're likely to see more cyberattacks and incidents as humanity as a whole catches up on the security of its codebases, using more AI-generated code that is more secure by default. We'll probably have some kind of transition phase where offense and defense get stronger, and then defense starts to catch up.

So that's the consensus that I'm generally hearing in the industry. Listen, at least that seems like the most likely scenario.

Yeah, look, I'm very optimistic about how we can get through this moment. I think this requires a coordinated response, right? This really requires defenders to be aware of how the fundamentals of cybersecurity are changing, and that this window of opportunity won't be open forever.

11. Shift in Security Investment

You have to understand that we're moving into a world where the security of our organization depends both on the right fundamental settings, which is a kind of one-time investment, and on the level of cyber capabilities used for protection. As we move into that world, I think we can end up in a reality where defenders are ahead of attackers, but that's going to require a big change.

This will require changes in how we perceive the work of each organization, how we collectively treat critical infrastructure, and how we provide resources to those who currently do not have them.

Your example—that is, your description of this—is similar to, say, technical debt on the scale of the entire internet. Your internal reaction to that is really apt, because I've worked in companies where dealing with technical debt is very painful.

You say, “We'll use 25% of the engineers,” or, God forbid, “We'll freeze the development of new features for a quarter until we fix what really needs to be fixed.” This is not the state the world has reached yet.

When you think about where the whole world is in terms of security investment, traditionally it's been something in the low single digits of the IT budget. What do you think this indicator should be at this point in time?

Does this mean that everything you've invested in in terms of cyber tools—networks, endpoints, traditional enterprise security—is becoming less relevant, or do we just need to spend twice as much?

12. Greg Brockman Introduction

I have a feeling that cybersecurity spending should actually far exceed software spending. I really believe that. Maybe I'm wrong, but when I look at the world we're going to find ourselves in, software in some sense becomes a barrier to entry. This is almost free. It's getting so cheap.

Of course, you don't just want to have software, right? You need reliable, secure software that you can trust. So for me, the sustainability of this software is the value. This will become even more important.

So I think we have this moment, this opportunity. You pointed to the XKCD comic with the little block at the bottom about some open-source developer. It never made sense. The reason we lived in such a world was because of a certain “physics” of processes, right? How difficult it is to write software, especially reliable software.

Everyone has this supply chain, and so few companies audit it. We've seen this with supply-chain vulnerabilities, where someone downloads a malicious package and a number of companies are affected. All of this has to change, and it was absolutely impossible to change it through human developers because there are so few of them.

There are so few people who have expertise in this narrow area—understanding different packages and different security methods. Now there's software coming out, like Socket and others, that scans your open-source supply chain, trying to find vulnerabilities and identify them for you.

13. OpenAI’s Human-Centric Ethos

As you said, there's also the human factor, where maintainers suddenly go through the project and add something harmful to those who use that package. So it's very interesting to watch this shift, as well as the development of a third-party ecosystem that supports this process, which is what OpenAI and others in this market are doing.

That's right. I think this is very important. We do this together with the community. We're working with companies like CrowdStrike, and I think it's again a coordinated response. But we really see this as a top priority for OpenAI—to be a leader, right?

This is important for us because we have a leading model. We have a model with the highest intelligence, certainly in the area of cybersecurity. We believe there is transformative potential here for real good.

That means we, as a company, are going through a transition phase in security. We've been talking about security and consistency in our development process for the past few months, but we're also thinking about how to help the world become fundamentally safer and get through this period.

One of the features of OpenAI is that the company has always been on the side of humans, supporting human agency and progress. Let's use AI to make everyone's life better. I think the original idea behind OpenAI was to create something that could balance other companies' efforts to create AI that could be widely and universally used in a positive way.

Has this spirit been preserved at OpenAI, or how do you assess it in the context of the AI community?

I believe that this spirit is the foundation of our mission, right? “Ensure that AGI benefits all of humanity.” The word “humanity” is key here.

We believe that humans should be in the driver's seat, in control of the future, and that this technology should enhance human capabilities, even when models surpass humans in their capabilities. I think that's absolutely critical, right?

It's important to remember here that these models are very heterogeneous. There are areas where these AIs are clearly more capable than anyone today. Even GPT-4 knows much more than any human.

But the thing that's so important—and I'm reminded, for example, of the story of why Steve Jobs added a pen to the iMac—is that people can feel like they are in control of the machine. You can always take it and throw it out the window.

I think the same spirit should apply to AI. It's the spirit of people being in the driver's seat, with things like human judgment and human values. Maybe it's worth thinking about how people will act in certain areas or dimensions in the future—becoming more like executives or board members.

14. Conclusion

It's important that we have that level of influence. I think we'll build machines in the same way that you, as a CEO, will definitely have experts in certain areas. You will definitely have teams that have knowledge in many areas and perhaps even better ideas than you do in various aspects.

But the fundamental question is: Are you happy? Is this what you wanted? Is this what makes your life better? For me, that's something we all need to collectively think about.

It's not just about AI. It's about technology. Isn't that right? Technology should elevate human civilization. This is the most important driver of improving the quality of life, and this is what we should expect from AI. This is OpenAI's belief.

Thank you very much for coming, Greg.

Yes, thank you for your time.

Thank you both. I appreciate it.