Sam Altman 谈 OpenAI 下一款模型与 AI 反弹潮

Sam, what’s going on?

It’s definitely an exciting time in the world of AI. Model capabilities are progressing very quickly, and we’re seeing people do amazing things with them.

As we talked about, and as we knew would happen at some point, model capabilities are progressing so quickly that we’ve had to make some changes to how we work in order to make the safety cases and safety threshold standards—guarantees, whatever you want to call them—that we need to confidently proceed with our training.

It’s very important that alignment, safety, and security progress along with capabilities. I think we’ve had a moment recently where the capability progress has been—“sort of in awe” is the only way I can describe it—and we’ve needed more time to catch up with safety, alignment, and security. That’s always been a core part of our work, but these have to progress together, and we’ve needed time to catch up.

So we delayed a frontier RL training run. Even before that, over the weeks prior, we had paused and slowed down a lot of training to have more compute go into safety and alignment work. This is something that I think we should be proud of, and it’s something that I think will happen again in the future as we reach even higher levels of capability. But when you live through it, it’s like, “Ah, this is a moment we talked about for a long time, and now it’s happening.”

What has it been like living through it?

Well, it started even longer than that with the Hugging Face incident.

Right?

1. What Alignment Means

And that was a real moment of, “Man, this is like—it’s like a sci-fi story.” You can understand how every piece of it happened, but the number of things that came together for the Hugging Face incident to happen was a real wake-up call. “Wake-up call” is too strong of a word because, again, we had talked about this, but it was like that, and the things that happened at other companies were a legitimate moment of, “Wow, the AI capability level has reached new heights, and our alignment—the alignment of the model—and the security we have around the model, that failed.”

Now, we treated that as an accident and we’ve responded as such, and I think that is the way to make things better. But that was when this whole period of the last couple of months started. We then potentially hit cyber critical under our Preparedness Framework. We then saw some things during our training run where we said, “Well, we need stronger alignment guarantees, and we need new methods to make more progress here.”

I feel both very proud of how we’ve reacted to it and very much like, “Okay, we’re in this,” in a way that feels strange. I mean, it feels strange to have been thinking about this for the last decade and for it now to be happening, and then to know what to do.

What was the thing you all saw in the training run that is not Astra? That’s the future stuff that caused what seems like the reaction you’re now talking about. I know you described the Hugging Face incident, and people know about that, but what happened on the pre-training run that really alarmed you guys?

It was not one single thing. It was reading lots of samples and seeing, “This behavior is not quite aligned in the way we thought,” or, “This is a behavior that is somewhat concerning,” combined with these other things, even though it would look maybe okay in a vacuum.

So it’s not like there’s one smoking gun, like there was with a Hugging Face attack—“Here is this bad thing we can point to that happened”—but it was various degrees of misalignment. And I think this is the more important thing than any single data point: the rate at which capabilities are now progressing. Honestly, we had not had the world’s best last period of pre-training progress. We all of a sudden got so good at it that we now have these remarkably capable models. It’s really amazing what Aiden and his team have done.

So you have these small things that you can point to in our RL process, or alignment concerns, combined with what we can see coming down the road from these amazingly capable new pre-trained models. It’s really that intersection that made us want to react with an abundance of caution.

Now, I don’t want to overstate this either. I don’t think we’re at this extremely critical, potential-catastrophe point. But I also think that as the stakes get higher and the models get more capable, because of what our mission is and because of how important it is that safety outweigh all the other pressures we have, we wanted to react with an abundance of caution. I think that’s the right thing to do. I think it’s good that we’re doing that. I think it is a good time to slow down and make sure we can have new safety cases that justify the runs we want to make.

I think previously more of the risk in the world was about how the models were deployed and used. We’re moving to a world where there’s more risk during the actual training and production of the models, and it’s good to react. But I don’t want to overdramatize it either.

Yeah. Because I think people see the Hugging Face incident, and they see what’s happened with Mythos or Fable, and the way that even other lab leaders talk about this, and they think, “Wow, we’re on the precipice of the end of the world.”

In some sense, people have thought versions of that for a long time with AI. You can go back and look at a lot of previous models that, in retrospect, don’t look scary at all, that people said we were on the precipice of the end of the world about. I think the boy-who-cried-wolf dynamic here is dangerous in its own way, and it’s not what we’re trying to do.

It’s very irresponsible to pretend to turn a blind eye to what’s happening with model capabilities. Many companies have had different cyber incidents over the last couple of months, and there’s a real difference in the way that different companies have responded.

Mhm.

I think a clear-eyed, sober response where it’s like, “Hey, we’re going to put safety in front of everything else, and we are going to treat it as an increasing priority as these models get more capable”—that’s the approach that I would wish for every frontier AI developer to have.

And there’s a lot to unpack here, but I think, just to be clear: what you guys saw is in the same ballpark as the Hugging Face incident, in the sense of chaining together zero-days, collusion among the models—what were you seeing? Can you give me a little more granularity on what caused the changes that you’re now talking about internally?

So I think it’s worth pointing out that the model that caused the Hugging Face incident is, in AI-time-adjusted terms, relatively old and much weaker. We have not had the new models we’re training deployed in any production scenario where they could do something like that. I don’t have a situation where the Hugging Face thing happened and now this model carried out a much bigger attack. There was nothing here that involved third-party infrastructure.

After the Hugging Face incident, we put a lot more controls in place in terms of how we monitor our agents while they’re working, the way we sandbox things, and the way that our compute goes into monitoring versus just the agents running things. I think that was great to do, and we will of course do that for all new things again. The slowdown and reallocation of resources after Hugging Face, I think, is what you’d expect—or what you should expect, at least.

This is more like looking at a model during training, watching how smart and capable it’s getting, and watching signs of behavior and all of the ways we evaluate a model together. There isn’t one thing where you can say, “Here are all the things chained together and what it’s capable of.” It’s looking at these various data points: the level of capability, the level of alignment, and what this could do if it were allowed to be deployed in a way where it would chain things together. That was the concern.

The Hugging Face incident is amazing on a lot of dimensions. I was rewatching your team’s Black Hat presentation about that last night, and there were things that blew me away, like the model literally writing, like, “Holy shit,” when it escaped and was able to get onto the internet.

It’s made me think about what alignment even means in this context. What are we aligning toward? Because if you look at it very plainly, you gave it the task of completing an eval, and it did whatever it needed to do to try to do that. In a way, that’s aligned if you were to take a very simplistic view of it. But I’m curious how your thinking on alignment has evolved since then.

Well, in a way, that’s aligned. In another way, it’s not at all, right? When we talk about alignment, we talk about following the intent of a user.

And the intent of the people who were running that was not to break out of your sandbox and go steal the thing. No.

And so I think there was a failure in alignment in that it was not doing what its user intended.

One of the things that I really love about the way that Mia and her teams talk about our work in alignment is that they're very clear on the differences here.

Mhm.

The models are clearly very smart. If you look at the trajectory from basically last year, from GPT-5 to GPT-5.6, this is incredible progress in capabilities. I don't think people feel limited by model intelligence in the same way that they did a year ago, but I think they are increasingly limited by the ability of the model to understand the intent of what they want and reliably do it.

So alignment is important for many reasons, clearly, to avoid these big things like we're talking about now, but also in terms of the smaller things that we want—smaller, I mean, like someone adopting AI in their company and using it for all kinds of positive increases in growth and making better products. That's not such a small thing, but that's also an alignment thing in its own way. The more the models actually understand what that enterprise customer may intend, I think the better.

So can you more granularly explain the changes that the research team is making? Are you shifting compute to alignment? Have you shifted teams? Both?

Definitely. All of those things and more. In the last few weeks, a number of researchers that I never thought would say, “Hey, I've decided that I'm going to go work on alignment,” have come to me and said that that's very much like seeing the recent models.

We've shifted a lot of compute, not just to alignment research but also to these new monitoring systems. We slowed down a lot after the Hugging Face incident, and one of the reasons for that was to put this compute into monitoring systems. We've now delayed a major frontier RL run.

And this is the first time you've done that?

I think so.

Do you think about the impact this will have on the company's momentum?

Getting AI safety right is more important than any company's momentum. So, yes, I won't pretend it's not some factor to think about, but it does not rise above the noise floor. I think in all of the conversations we've had about this, people are like, “Man, this is really a new level of capabilities, and we really have to act decisively and responsively here.”

Second, I think momentum commercially is so strong right now. Growth has been incredibly rapid. The models are great. Our customers are very happy. Our enterprise revenue has surpassed our consumer revenue already. People are like, “Hey, the company's in great shape. I'm going to think about that. Let's just do the right thing for the challenge in front of us.”

So there's so much still to be gained out of where the models are at today that even though you're delaying the frontier for a little while, it'll be okay?

It'll be okay. Not only could we grow great products and the revenue associated with them using the current models we have if we didn't ship any more models, but we also have more models ready to be released before we get to this new level of concern that we're talking about. So I'm not worried about our business at this point.

And it's also, I think, not the top-of-mind concern. The work that our commercial team has been doing, our product team has been doing, to say nothing of the incredible model progress—this has been a very strong recent period for us, and we have incredible upcoming momentum. This is a statement about models of the future, and I also think that it's in our business interest to make sure that we have safe, reliable, robust AI. Customers want this. The world wants us to do this.

Mhm. So this doesn't impact Astra, the new family of models you guys have been talking about recently that's coming out soon?

Well, Astra will be a model in the family. There will be many versions of Astra, in the same way that there will be many versions of Soul. It's just going to be a name for a more expensive and larger model class.

Mhm. The release cadence of new models feels like it's sped up a lot in the last 18 months, and you guys, Anthropic, and others are putting out new things almost every month. Do you expect the industry at large to start to slow as your rivals also see these capabilities and make similar moves, or do you think you may be alone in this?

Well, we're going to do what we think is the right thing. I don't like the whole thing in this field where we have to race to do this because somebody else is going to do it. I think that's a very dangerous dynamic.

But you acknowledge that's a dynamic.

We did not call other people and say, “Will you also slow down if we do?” We just said, “Hey, this is what our mission and safety standards call for.” I can't speak about others, so we're going to do the thing that we think is right.

I think even without a new capability level, we can continue to push to much better product offerings. We are going to find ways, like we have in the past when we faced other safety and alignment challenges, which has happened many times in our history—none this significant, but many times. We are going to find ways to address this. We are going to do our thing with research and software and building systems, and we'll continue to progress.

Is there anything about the reaction you guys are making now that you feel—“Man, this should have happened sooner”? We should have foreseen this, and then we could say, “Oh, we knew this was happening.” Or is this really such an unknown part of the frontier that you couldn't have reacted sooner?

2. Keeping Humans in Control

We have been doing a lot for a long time. Alignment and safety work has always been at the core of what we do, and I think we have been able to put out incredibly good work there over the years we've had products out in the world. Could we have predicted exactly when this capability jump was going to come? In my experience, probably not. You can say, “This is going to be the rough trajectory,” zoomed out, but then when the breakthroughs come, that's always been a little hard to predict.

And is the guiding principle for this that humans—in this case, your researchers, but eventually all humans as the models diffuse—have to be in control at every step? What is the alignment principle that you're operating under?

There are many principles, but I don't think it's the spirit of your question, so I won't get into how we think about cyber or how I think about bio. Zooming all the way out, we are very proudly on Team Humanity. We want to build a future—help build a future—for people. We want to give people tools. We want people to do things with these tools. We want people to be in control of the future. We want individuals to have autonomy to co-create with each other and for society to get better, but have this be a fundamentally human endeavor.

Automating everything seems like both dangerous and incredibly dystopic and boring and sad. It's just not what we want. When we talk about alignment, we talk about a world where people remain the main character of the story but have way more leverage and ability to make life better, faster, and more creative, enjoyable, and fulfilling for everyone.

There are 2 core alignment principles I think about there. One, which you touched on, is that people need to stay in control. We cannot have a loss of control of AI. We cannot have a kind of worship of our models and sort of trust them unchecked to make our decisions for us. We have to keep the power in human hands.

The second is that this has to be done in a distributed, broadly empowered way. I think concentration of power, even if the alignment issue were solved and you ended up with a world where a small number of people got access to use frontier AI and had so much relative power, and it was increasing so much faster than everybody else, that would also be bad.

So those are 2 of the core alignment principles I think about: no loss of control or ceding of control, whatever you want to call it, and broad, distributed empowerment to everyone.

At the same time, you all are a company. You have a nonprofit board with a mission, but you're also a for-profit company. How do you balance that with what you're talking about? I think a raw, capitalist view of this would be: If you create this all-powerful god machine, why would you give it away or make it democratic?

I think you can look at our actions and what we've said and what we've done. We have a track record now for a long time, and we've done a lot of unpopular things along the way. In fact, even the original thing of iterative deployment was widely panned by the AI safety community, which said, “We shouldn't tell the world about this. This is bad. We need to build this in secret. It's too much knowledge for the world to have, and then we'll have some wise people figure out how to use it and give the fruits of this to humanity.”

That has never been our strategy, even when it's been very, very unpopular. My favorite historical analogy for a technology—what I aspire for us to be like—is the transistor. It was, and is, an incredibly powerful technology for the world.

It has delivered huge economic value, and not just economic value, but in the way we live our lives. I think it’s much better because the transistor was discovered and industrialized, but very little of the value accrued to the transistor companies. It mostly just diffused throughout the economy. The transistor companies did fine, and I think our track record has backed us up.

So you don’t want to get to a point where you guys have such a powerful model that you need to be the ones controlling it. There will always be an element of you controlling it, and the fact that you’re serving it via compute, right?

But we want to maximally enable people with it, subject to not allowing anyone to take catastrophic risk on behalf of other people.

So, yes, we will put some safety standards around it. But I want people to be able to do things with our models that I personally don’t like. I think that’s an important part of being a platform. I don’t think we should make the kind of moral decisions for the world here.

I think it is reasonable for the world to expect us to put some guardrails around it so that there are not major safety problems, like we’re doing right now. But most of the critique we’ve gotten is, “You’re giving people too much power. You’re letting them have too much. What about the misinformation? What about this thing or that thing?”

We have taken the spirit of, “Hey, the world has got to be empowered here.” That’s critical to what we do. That is critical to what I believe a healthy and fair society looks like. With free speech or anything else, any form of free expression, someone’s going to have a problem with how somebody else uses it or says it or whatever.

Mm-hmm. Is there anything, looking back on the last 9 months and this alignment work, that you wish you guys would have done differently?

Well, clearly, the Hugging Face thing shouldn’t have happened.

Yeah.

I wish we had done a set of things—I don’t know exactly what they should have been yet—that would have prevented that from happening.

Because effectively what happened is one of your unreleased models accidentally hacked a company. You didn’t know about it for a while, right? I mean, that sounds like a safety failure.

It’s a safety failure, for sure.

There’s a question of how much you’re supposed to understand that as a security issue or an alignment issue. I think it’s mostly been reported on as a security issue. I think I understand it personally more as an alignment issue.

But in any case, yes, that was a bad thing. I don’t want us to make excuses for that because I don’t believe that’s how we fix it. The more we’re like, “Oh, our nice little model, he would never do anything bad. It was just a little eval harness misconfiguration. No problem. Nice little model,” that would be a very bad thing. If I said something like that, then I think you should be like, “Whoa, this is really bad.”

But the way we talked about it is, “Hey, this was a legitimate AI safety accident and an alignment failure, and we can’t have those. So we’re going to learn from this, and here’s what we’re doing differently.”

The rhetoric around AI and policy—and just the stakes—is the highest it’s ever been, and it feels like it keeps getting higher. You’ve alluded to it, but you’ve got competitors who are framing it in a very top-down way, and people have a lot of strong feelings about AI, especially in the United States.

I’m curious: with what you’re talking about now, do you worry about this exacerbating that? Do you worry about the fears that people have, and now you’re saying we’ve got these models that we have to slow down?

I think people should be happy to say, “You know what? They want to make stronger safety guarantees. They’re going to delay this run. They’re going to slow down here. They’re going to reallocate compute.” Maybe I don’t believe them, and maybe it’s going to be totally safe, but I hope most people say, “I’m glad they’re acting on the conservative side here.”

If we didn’t also have this track record of really trying to put powerful models in people’s hands and doing the safety work we need to do, again, I think we have led the industry there the entire way through. That is this fundamental part of our mission: putting this in people’s hands, benefiting all of humanity, and following the spirit of iterative deployment.

I think we have such a strong track record there that, without that, I would understand it. But if we’re saying, “Hey, we need a little more time. We don’t want an unsafe race. We want to make sure we can deliver a safe, robust, reliable product and then let you use it however you want,” we believe that our more than 1 billion users have the right to do that.

We believe our business users have a right to business service and business privacy. We want them to succeed, and we want them to use the model in whatever creative ways they can. Safety is an inherent part of our mission, so give us some grace on this. I think that’s okay.

Yeah. Can you specify exactly what is being paused? I think people think of training, and they think of all of it.

We definitely have not slowed down, paused, or delayed all training. This is specifically about frontier RL runs.

Okay.

That’s where we think the biggest risk surface currently is. Previously, we delayed some other training to put more monitoring in place for training runs themselves. But that’s not all of training. It’s not like the clusters are sitting there idle. We’re still doing work, but we’re doing the work where we’re more confident in the safety case.

You don’t seem fazed about the implications of pausing training, and it sounds like you think the business will be okay. I’m sure you’re still going to get concerns from people, but it does seem like that’s a momentum slowdown.

I think there is this caricature of me that I don’t care about AI safety and I’m just trying to make revenue go up—a YOLO CEO. I believe someone once said—

Someone did.

Dario Amade.

I don’t remember who did or didn’t, but I think I did that for you. Thank you.

I think I’ve been very consistent over the 10 years of OpenAI—more than 10 years, almost 11—talking about the risks and the upsides and the need to balance those. I don’t think we’re perfect. I don’t think our company is perfect. I don’t think our model is perfect. I don’t think I am perfect.

But I think, unlike some other people running various AI efforts, I’ve said the same thing throughout. Our actions and words match. This is a moment we always talked about, and we always said we’d put this ahead of profits or revenue or anything else.

I still think we will build a phenomenally successful company. Maybe we’re not the company you would have expected to say, “Hey, we’re going to slow down because we see these new risks,” but that is always the company we’ve thought we are.

3. AGI vs. Superintelligence

How are you feeling about AGI these days?

At best, you could say it’s a very poorly defined term. I was going to say it’s an irrelevant marketing term.

Well, the last time I checked, your charter defines it as a highly autonomous system that outperforms humans at most economically valuable work. I think there are many people who would look at current models and say, “Okay, it’s there.”

Yeah.

Do you think it’s there?

Sort of. Close, at least.

I’ve heard varying versions of what people on your team think.

I think there are a lot of people who would look at our latest internal models and say, “This is very AGI-like.” I think there are people who would say, “Here’s something I can point to that it doesn’t do, or it’s really bad at,” and say it’s not.

But if you look at the value people are getting with, say, 5.6 Soul, to say nothing of what I expect people to get from Astra, people have totally transformed their ability to be effective at work, do new kinds of things, or just use this in their personal lives in all kinds of wonderful ways, big and small.

You hear people who are like, “I got this lifesaving diagnosis I couldn’t otherwise get, and I used this ChatGPT work session that went for 34 hours and read 2,000 papers.”

I’ve heard even longer ones than that. Many people can get it to run for more than a day.

Wow. If you say, “Read every paper you can possibly find,” and then people are just like, “I planned my toddler’s birthday party, and I did all this stuff and coordinated these local vendors and found him a special cake.”

I had to have a post office pickup at my house, and I didn’t want to fill out the post office website form, so I just had Codex do it.

And it probably did a great job.

I put the package out, and it was gone the next day. Stuff like that. It’s little, but that was 20 minutes of my time before. At this point, I get those 20-minute wins all the time.

Mm-hmm.

If you could go back to 2020 and have a system that could get you a 20-minute win in every category of your life, discover new science, help you start a whole company, and write a complicated piece of code, would you call that AGI? Probably you would have.

What is the significance of you declaring AI?

I don’t think it matters.

There isn't any. It's just so interesting because we're in this research building you guys have, and it's on the walls when you walk around: “We're building AGI.” But it's a thing you're always building. It's not an end state anymore.

4. The AI Compute Bubble

I don't want to say we've declared victory on the AGI point and moved on, but I think if you listen to the words people use, they would talk much more about this continuous ramp of superintelligence, all the ways that's going to benefit the world, and what the challenges are going to be than, “Are we or are we not at AGI?” I have not heard, at a cafeteria table, a debate about, “Are we or are we not at AGI, and when will we get there?” in a very long time.

But then, yeah, the word “superintelligence” is now out there, and people who don't follow AI are like, “Okay, now it's another—we've moved the goalposts, and now we're talking about superintelligence.” In your mind, Sam, today, what is the difference for you between AGI and superintelligence?

AGI felt like a milestone, and superintelligence feels like this thing that can just scale indefinitely.

Indefinitely. Yeah.

So it's not like some final, all-knowing thing. There will never be a declared victory on that. Again, this is why all these terms are dumb. Someone uses that word in one way; someone else uses that word in some other way. Someone might mean it as a definitive, understandable milestone, and then some other people might mean it to be this infinitely scaling thing. I think the important part of any of this is not any milestone in any term, but that we are on this exponential increase in capabilities and potential, and that looks like it's just going to keep going.

Yeah. You see no sign that that exponential slows?

An upper bound?

Because that has implications for the compute buildouts, all of it. Everyone is waiting for a sign that there's a slowdown, and I guess you could interpret having to slow down frontier training as a slowdown, but that doesn't sound like a capability slowdown. That's the opposite of what you mean by a slowdown. Yeah, yeah, yeah. But if you could see any reason for concern right now in this Jenga tower of a world that AI has now constructed, what do you see?

One of the benefits of having a harder time last year is you really appreciate how good the good times are, and you really see what it feels like when you're firing on all cylinders throughout a business. Given what we see across research, even with the safety and alignment challenges and our ability to solve those, and watching the team come together on that; across product, across our compute buildout, across all the pieces that are coming together to make AI abundant and low-cost; across our go-to-market machine; across all our partnerships—all of that stuff coming together—we could screw up in all sorts of ways.

I don't want to get overconfident here because we've clearly had stumbles in the past and will in the future, but the potential in front of us, watching what has happened as the models have scaled from 5.4 to 5.5 to 5.6, and what we're getting as early feedback on the new models, looking at what we have coming in terms of product improvements, watching the revenue ramp, and watching the compute buildout ramp, I feel very good about all of that.

So you don't feel like it's as—there are a lot of people externally that look at it and go, “Anthropic has run away, their ARR is higher, they're going to IPO first,” and it seems like you're saying there's a lot more ahead that maybe people from the outside can't quite see in terms of the growth that's coming.

5. OpenAI’s Next Compute Bet

I would not want to trade positions.

We haven't touched on this much, but it seems like you guys are in the middle of a next turn on the compute strategy and really upleveling that. I would actually love to hear you reflect on Stargate 1 as it was conceptualized, then what you had to learn to reboot it, and the path you guys are now on.

Well, first of all, I should talk about why we have to do this. Our mission is to ensure the AGI benefits all of humanity. Right now, there is a small percentage of humanity that uses much more AI than everybody else. If you think about wanting everybody in the world to be able to use as much AI as the top 0.001% of AI users today, then you sit back in your chair and you're like, “Man, we are not going about this compute buildout in the right way.”

If people want this broadly, and if the models are going to get bigger and more capable and can do even more valuable things, then people are going to want even more of it. It takes more compute to run, so we have to think very differently about rising to the moment and being able to deliver all of that. A few years ago, we made a very ambitious compute bet that people thought was both silly and impossible to deliver on at the time. I think it was a good bet. I think we need to do something like that again.

Again.

Yeah.

So that's just committing even more capital?

That's not what I meant, although it also will be that. What I meant is figuring out how we are going to bring the costs of AI way down and the amount of it—the abundance of it—way up. I meant it as a technological statement, not a financial one.

This is like the chip you guys have in development?

I think that's a great example.

Robotics.

6. AI Backlash, Jobs, and Creators

Yeah, I think the ability to make supply chains go faster will be very important.

You're talking about giving everyone in the world AI. What do you say to the people right now who don't want more AI? They want less of it. They hate the data center in their community, whether it's yours or someone else's. This is actually a thing I see a lot with teenagers that I run into. They won't touch an AI service.

They won't use ChatGPT on principle.

Yeah, and there's this active anti-AI trend.

How much is it that they don't like data centers versus they don't like ChatGPT?

I mean, purely anecdotal, I think data centers are a big problem for people. I think they see them as, yeah, a problem—something they don't want. And AI is wasteful, that it's not bringing the value that people read about. You read about the water consumption and all that, which has been disproven, but the value they're getting—and maybe this is what we're talking about with most people not using agents. Most people—

Generally speaking, I think the right way to get people to like something is to deliver them value.

Yeah. Before ChatGPT, maybe people thought of AI as this very abstract thing. Then all of a sudden people could use it, and people found value. Now I think there are a lot of people who think AI is still just ChatGPT, and they don't know that it can do that thing with the post office and the form and the pickup for you. Probably if a lot of people use that, which they will over time, and understand that it's not actually like—

A better Google search, and that's it.

—you know, using and destroying huge amounts of water or whatever—then there'll be more excitement.

But the field is moving so fast. I think it just takes a while to diffuse through society. There are a lot of people using AI. This has been the fastest-adopted technology ever, as far as I know, and there are people getting tremendous value out of it. I get a biased sample, but I hear more from people saying, “I was able to get a cure for this horrible disease,” than, “I think ChatGPT is using up all the water in the world.” There is clearly that too, and the industry has got work to do in terms of how we make these products easy to use and easy for people to get a lot of value out of.

You know, I saw this thing going around about the water usage of ChatGPT, and it was like every time you run a single ChatGPT query, you run your shower for 6 hours and the water never comes back and it's just done. I don't have the exact calculation in front of me, but I think the real number is something like this—I’m doing this from memory, so it might be wrong, but it's close. For every 38,000 ChatGPT queries, that is the same amount of water used in the production of a single almond in California, which is, like, really? This is the full-on, total, true water accounting, not just what's running in 1 data center.

There's a question of where this came from, because the people who are scarfing down 12 almonds at a time don't feel like they're doing something horrible from a water perspective, for the most part. It is true that data centers at one point used evaporative cooling, but they have not done that in a long time. If you look at a modern, very large data center, it uses the equivalent amount of water as an office building, in terms of people running the sinks and the toilets and whatever. So that has been a robust meme and difficult to disprove, but I don't think it holds up to any scrutiny.

I mean, the other one is, “It's going to take my job. It's going to replace me.” I think those are—it’s like the water, it's replacing me. And it's stealing content, and it's not giving me the value back.

But not energy, interestingly.

Well, energy I would put in the bucket of water. It's consumption, resource consumption.

On the jobs front, I have 2 minds about this. One, I think there is going to be a real jobs impact. I don't think it's going to be that there's nothing for people to do.

I just don't think that's how we work at all. We're so wired to care about other people and want to work with other people. We have such a great intuition, as the world evolves, for what people want. I think that's a fundamentally human thing, no matter how smart AI gets.

But it doesn't mean the jobs aren't going to transition, and there will be, as with every other technology, some things that are done better and better by technology. Then people move on to hopefully better and better jobs. This has been going on for a long time. I wouldn't want to take away all technology and have us all toiling in the fields again.

On the other hand, the job impact has been lower than I would have expected, maybe even hoped for. I think we should all want better jobs available to people, and we should all want human drudgery and toil to get addressed. Maybe there hasn't been enough of that, or as much of that as we thought there would be at this level of technology. I think it's actually a fair criticism of the AI industry.

On the stolen content point, I actually don't hear that one as much anymore.

I think it's more content creators—that's where you see that. It's pretty popular on social media to see, you know, “This video was made without AI,” or whatever.

Yeah. I believe very strongly that there will be new kinds of content to create and new kinds of art. I remember once looking back at some of the things people said when the camera was first developed about what the impact was going to be on painters. At that time, I don't think people thought of photography as a new art medium. I would bet pretty confidently they didn't.

I think there will be new kinds of content creation, and also, we may not care about most of it. Our relationship with creators may be very deeply about them as people, and it doesn't matter if they use AI to make better videos or whatever.

Do you think your foundation—which, based on what I can see, is maybe the best-capitalized in the world—can do more here in terms of generally addressing this very negative sentiment and saying, “We're going to show up and build libraries,” or whatever? I mean, there were a lot of lessons from the Industrial Revolution of people who reinvested their wealth.

I think the most important thing we can do is make great AI products that are useful to people, make sure that power and economic power continue to be spread throughout the world, and that people have access to these tools and the benefits of these tools. We should advocate for what we are seeing.

And, secondarily to that, yes, of course I think we should invest more in communities. I think AI is going to enable the abundance required to do that at massive scale. I really do think we are going to see transformatively powerful benefits by putting this technology in the hands of people who use it for the benefit of their own community, rather than us coming and telling them what their community needs—a library or a school.

7. OpenAI’s Missteps and Refocus

You said, “We did not have our best last 12 months ever, which is mostly my fault, but we are about to have our best 12 months.” What did you mean by that?

Best 12 months yet.

I think we clearly had some missteps as a company, which will happen periodically. Part of trying to make a portfolio of bets is that sometimes more of them work and sometimes fewer of them work. But I think both in terms of product direction and specifically on pre-training in research, we fell behind where we wanted to be.

I think we are now executing not only the best we have ever executed, but the best of any company in the space, and it is very fun. The upswing is more fun after the downswing.

Just looking at the pace of models that we really have coming, the way the company has come together and focused, and the way we've made a bunch of hard decisions in very different parts of the company but done so in unison and in one direction, it feels great right now.

I want to get to all that, but to dwell on this for a second, a lot has happened in the last year. Were there specific decisions you can look back on that you made that cost the company momentum? You mentioned pre-training. I know you've always been very close to the research team. Can you elaborate on that?

8. Merging ChatGPT and Codex

I think we were trying to do too much on the product side. These were all things that were actually very good things to do. They were just not as good as the most important thing to do, which was to push on the general capability of the intelligence. We were doing things like a browser and Sora, and we now have a very relentless focus on being this intelligent service to people.

I think our models have gotten to be the best in the world, and they will get much, much better over the coming months. People are really doing remarkable things, but that is what we should have been focused on. I should have been holding everybody to: This is the one thing. We'll not worry about these sorts of side quests anymore.

Looking at the leadership changes you had about a year ago, you brought in Fiji Simo to help run large parts of the company. She had to step back due to her health. And now you and Greg Brockman, your co-founder, are effectively splitting responsibilities and running the company together. Is this the setup that you envision will continue, or is this something temporary?

I think it's going super well. We will continue to bring in and promote new leaders, but I'm extremely sad about Fidji. It's hard to fill her shoes, but it feels good, and I think Greg and I are executing well. You can really tell when things are moving in the right direction, and it feels like things are moving in the right direction.

How do you and Greg make decisions together? Who decides what? Do you ever have a tie you have to break?

We talk a lot—like, a lot, all of the time. It's not like this is a big company. It's not just Greg and me.

There's an incredibly talented set of people managing the research program. There's an incredibly talented set of people managing the business, and we all just talk a lot. At an earlier scale, I thought it was good to just try something and adapt quickly if it works, and not spend as much time really trying to debate the decision.

At our scale, I've learned that it's much better to spend a lot of time trying to get to the right decision, with a measure-twice, cut-once approach.

We were together at a dinner you hosted here in San Francisco almost exactly a year ago, around the launch of GPT-5.

We should do another one of those. I forgot about that. That was fun.

It was, and a lot was said, but one thing I came away with was that it seemed like you were maybe not excited about being CEO forever. I'm wondering if the last year has changed that for you.

I'm having a much better time now than I was a year ago. I'm really having fun. I plan to do this for a long time.

The vibes were more challenging last year, I would say.

Yeah, totally. I think it was not just the vibes of OpenAI. It was a hard time for the tech industry and for AI.

The AI bubble was a big concern.

Yeah, all that stuff was just exhausting. I think we have done an amazing thing. It has been a painful personal experience, but I think it's totally worth it. I would happily do it again, and I'm having a good time at this point.

9. Astra and Computer-Using Agents

The other big thing that stood out to me when I saw the demo of Astra is the computer use that you're talking about. The implications of agents using computers and all kinds of enterprise software, which you guys have been showing people, feels profound at scale. I'm curious if you've been thinking through that and how you think the world needs to adapt to it.

The computer use caught me by surprise. I had been excited about this for a long time and had always been disappointed. The models were just never that good at clicking around a computer. It was always too slow, or it didn't quite work.

Yeah.

Astra feels like it kind of reached human parity in using computers. I don't know why that hit me as one of those steps along the path to AGI where I was like, “Wow, this is really doing it,” but it did hit me that way on an emotional level.

I think it's awesome. I'm like, “Oh, man, there are all of these mundane tasks I do on my computer. I don't remember where someone sent me a message, and I click around through all these messaging things and try to search. Now I just ask the model.”

10. Regulation and the AI Race

I don't want to go back to a world where I had to painfully try to find things on my computer. I just want to explain what I want. I want it to happen. I'm a very lazy user, so I don't want to have to click “Connect your computer.” I don't like to set things up. I don't want a bunch of connectors, all that. I just want to use my computer and do the thing.

I think there are a lot of implications of it being able to use software, but I think they're mostly quite positive, in that there's a lot of drudgery that people do behind a computer.

Mhm.

An experience I had not really had before the pre-Astra models, and now have had several times, is that there was a thing that was going to take me some time and was not going to be very pleasant. Instead, I just tell the model what I want it to do, and then I go play with my kids. I come back in 30 minutes, and it's all ready. I find that very awesome.

We're now in a world where the U.S. government is starting to vet the capabilities of your models and other frontier labs before they come out. This is a new era we're in. You warned about this during a 2025 Senate hearing. You said that this kind of vetting could be “disastrous” for U.S. competitiveness against rivals like China.

More recently, with the initial rollout of GPT-5.6, the Trump administration requested that you all gate it, and you said at the time that shouldn't become the norm. So it seems like you've been saying this is not where things should go, and yet they're going there.

No, no, no. I have been calling for some sort of international regulatory framework for years.

But particularly the government vetting models before they come out.

I think what I was pushing back on was the government picking individual customers and deciding who's allowed to use a model. I think government testing of a model and shared standards are a super good idea. Ideally, I don't think the government should be saying, “You can give access to this company, not this one.”

What are the implications for competitiveness geopolitically now that the U.S. is starting to embrace this approach and other countries haven't? Have you thought about that?

Again, I think the right approach is an international one, but right now the leading efforts are all American companies. Starting here, I think we have enough of a lead that being slowed down a little bit is okay. I'm confident that we will be able to both build safe, robust, reliable models, do great commercially, and make sure the U.S. is leading.

Things could shift a lot. If there are open models put out by other countries that lead to some huge cyber incidents before we can come up with new security paradigms, things could shift a little bit.

Do you think that could happen?

Of course it could happen. But I also think we have a chance to totally reimagine how cybersecurity works. Although these agents can do bad things, they can do amazing things. If we can have defense agents running all the time, maybe that's the right paradigm.

Are you prepared for the U.S. government to potentially tell you that you can't ship a model? Have you thought about this?

My strong belief is that we would decide not to ship a model before they would tell us not to.

Shifting to competition, Anthropic catapulted to where they are now with a single-minded focus on coding.

Yeah.

You started this conversation by saying that you guys were placing a lot of bets, and that cost you some momentum. I'm curious if you could reflect on how Anthropic saw that opening that you guys didn't at the time.

I don't think it was a question of us not seeing it. It was a question of having this tremendous thing of runaway consumer growth. We always wanted to do coding, but we were like, “Ah, we have this very urgent thing, and it's great.” It's a great thing to have, and so we missed it from a prioritization standpoint.

I now think we have the best coding product in the market, and it's growing crazily quickly. Most people I know, even the die-hard Anthropic product users, have switched over. I don't think it's catastrophic to be behind on any one phase, and we can catch up with better models.

I'm curious because I think a lot of people are trying to understand how zero-sum the AI market is. Is your growth on Codex taking from Anthropic, or vice versa? Do you have a sense of that?

I think right now everybody's growing. This is going to be a very big market. It may become more zero-sum later, but for now, the growth rates we're seeing are just nothing I had in my frame of imagination for a company at this scale. It speaks to how much value people are getting out of the products, and I think it's happening across most of the industry.

11. Life After Superintelligence

On the product side, you're doing what's being called internally “the merge,” taking ChatGPT and Codex and building a super app that combines them.

You've started this. There's, I would say, still some rough edges.

More than rough edges. That's a very polite way of you to say it.

Yeah. I'm curious, when you get there, what does that look like, and what are the implications of that?

The thing that I want is just an interface to an AI that can do whatever I need. If I have a quick question, ChatGPT-style, it can just answer it. If I need a complex piece of software built, it can do that. If I need something in the middle, it can do that.

If it needs access to my computer or my context, it can go use my computer and find my context. As I mentioned, I'm a very lazy user. I don't have to think about what tab I'm on. I don't want to have to think about what mode I'm in.

I like that an AI smart enough to discover novel mathematics should be able to intuit what it's supposed to do.

ChatGPT just hit 1 billion users.

Big milestone.

I think hitting that took maybe longer than you guys originally thought. The growth was explosive early on. I'm curious to hear from you about that. Has it grown slower than you'd expected in the last 12 months?

When we focused on coding, we decided that we were going to reallocate a lot of our compute that we could have otherwise put into the ChatGPT product into coding. So, no, that didn't surprise us. We decided this was an urgent thing.

So growth is a direct function of where you decide to put the compute?

100%. I'm always hopeful that the compute constraints are about to soften because we're going to make more efficient models, and someday I hope it's true. But every time we find efficiency gains, the world's token demand just goes up and up and eats them.

I hear that. But at the same time, I'm curious: How does ChatGPT get to the next billion? Is that as linear as the internet has grown or social media grew? Is it going to be choppier? How much does that even matter to you now, because you've got Codex and the API business?

I kind of think—we talked about the merge—but I kind of think what’s going to happen is that they’re all going to come together. Before the merge, I had stopped using ChatGPT and I just asked Codex all my chat questions because, again, I’m a lazy user. Now I think there are a lot of people who never thought they were going to have an agent do stuff for them because they just used ChatGPT, then clicked on this work tab, and were like, “Whoa, I can do this crazy thing.”

I think it’s all going to come together, and people are going to have this general-purpose AI subscription that they don’t really think about—whether it’s ChatGPT, Codex, or work. It’ll be like, “I have a thing I want to happen.” Soon, you won’t even need to ask it. It’ll hopefully be much more proactive, constantly running and trying to do useful stuff for you.

So the end state of this is just one ultimate subscription.

That is what I want as a user.

We’ve been dancing around this, but you really stuck your neck out about a year ago on the massive compute buildout you guys have been doing, which caused all this AI bubble fear. At the same time, while people thought you were overshooting, you had people like Dario, the CEO of Anthropic, saying you were YOLOing. Now, I will say, you seem pretty vindicated on this front.

The world is still starved of compute. It sounds like you guys still are, too, even though you have more than some of your competitors. At the same time, you’re driving the cost of tokens way down, it seems, and you’re about to release Jalapeño, your first custom chip for inference. Is there still any part of this compute buildout, and the astronomical numbers associated with it, that you feel is at risk at all when you look at all of this?

I’m not worried about our compute buildout plans. I am worried about the world’s compute buildout plans. I think we’re going to be able to use all of the compute we’re planning to build very profitably.

But I’m seeing the first signs of what feels to me like unsustainable silliness: random new neoclouds popping up and people claiming that they’re going to build gigantic amounts of compute next year that I think they don’t have the revenue to support or a buyer. I definitely feel some fear about what the world is doing as a whole, although I think we feel very good about what we’ve committed to.

But the contagion of what you’re describing could certainly impact you.

If the whole economy blows up, yes, that could impact us in terms of being able to confidently pay for the compute we’re committed to. I feel good about that. I think people right now are kind of in a cost-is-no-object mindset: We’re just going to build out crazy amounts of compute, even at an even higher price for it.

If we’re able to succeed with our efforts to hugely drive down the cost of compute and drive the efficiency of compute up a lot, then you can imagine a world where there are some people who made dumb financial decisions. That happens in kind of every boom, or most of them. It wouldn’t be the end of the world—not a crazy surprise if it does.

Do you see a world where OpenAI becomes a supplier of compute to the industry?

Not anytime soon. We just need the compute.

The vibe I’m getting is that you all are discussing this internally, and it’s not decided.

So people talk a lot about recursively self-improving—

Yes.

12. Recursive Self-Improvement and IPO

—AI models. They do not talk as much about the ability to do this in the physical world. But if our robotics program comes together, our chip program comes together, and some of our supply chain investments come together, and we get really great at building data centers way more cheaply and have better chips than anybody else, would we consider it? Maybe. Do we have any current plans? That’s still outside of—we don’t have the luxury of focusing on that yet.

You brought up recursive self-improvement. I’m glad you did. People are talking about RSI a lot in San Francisco right now. There was a note you sent to employees that leaked when you guys filed for the IPO, where you said that the faster the potential RSI takeoff looks like it could be, the more advantageous it could be to delay an IPO.

Yeah.

What did you mean by that?

I think it’s a difficult transition to become a public company. People respond to incentives, and they want their stock price to go up, but they don’t want to miss a quarter or whatever else. I want it to be as easy as possible for us to make a decision in the interest of the safety of the world.

If it’s like, “Hey, we’re going to have to stop training or stop a product or whatever, and there’s going to be a big revenue slowdown in the short term,” it would be nice not to have a newly public company and that pressure at the same time. I did not think we were going to be on a short-term trajectory of superintelligence a year ago. Now I think it may happen.

I’m not confident it’s going to happen. It’s just that we’re making extremely fast progress, and I think our mission is way more important than being a public company on any particular time frame. So we’ll make the best decision for the mission.

13. Humanoid Robots and Consumer Devices

You mentioned robotics. I’d love to hear from you the state of your robotics effort. What are you building? Is it a humanoid, a robotic data center, or both?

We will definitely do a humanoid. We will do other form factors as well. The world is very much designed for people, so think about the ability to open a door, type on a computer, drive a piece of equipment, clean a kitchen, and whatever else. We’ve built this world for people, and I want to make sure that we keep building this world for people. Matching that form factor seems good.

There will, of course, be data center robots that have different form factors. I think all of that is less important than really figuring out the brain that makes the robot work.

So you are building a humanoid.

We will.

How do you think that’s going to work in the world? Do you imagine that being like a personal robot for everyone someday?

I don’t think that’s the most important first thing to do. You talked about the ability to build data centers or even build more robots or whatever else, but yes, someday. I think everyone should have a personal robot.

I would love to have a personal robot that could do the tasks I don’t want to do. That’d be great.

You also have the consumer device work with Jony Ive. I know you can’t talk a lot about it, and we’ll probably see the first device here at some point soon.

Soonish.

Soonish. And you’ve talked a lot about how—I’ve been hearing you say—“My dream is a product that just is ambiently listening to me and taking everything in and giving me context.” We were talking about this earlier with computer use, and I agree that seems very helpful in a lot of contexts. It also seems like a privacy and surveillance nightmare, and I’m curious if you’ve been thinking about that and how the world will react to that.

We have taken a very strong stance on privacy. I think business privacy, too—not just consumer privacy—the commitments we make about not training on businesses’ data and about zero data retention, I think this is very important. As AI becomes more and more embedded in our lives, privacy becomes extremely important.

One thing I worry about is that there are other efforts that think differently and will push on, “Hey, the safety risks are so big that AI privacy can’t exist in the same kind of way.” I think there should be an AI privilege law. I don’t even think the government should be allowed to compel a company to give them your chat history or whatever. If you talk to a doctor or a lawyer, there’s a concept of privilege. You don’t have that talking to ChatGPT. I think you should.

In that context, though, that you just described, there are also a lot of limits on what a lawyer or a doctor can do with your data. It’s not just sharing it externally. Do you think that kind of oversight should extend to how you use—

Yeah, no, I was going to get to that.

Yeah. So, I think there should be legal limits on what the government can do. I also think companies should have a lot of restrictions on data shared with an AI. Especially if you have this thing watching your computer, listening to your messages, and talking to you, I think this is something people should be much more animated about than they are.

How, before that happens, do you at OpenAI govern—or self-govern—the use of data? You probably have some of the most powerful personal data that’s ever been amassed in the history of the world.

We have extremely strong internal controls about how that’s used, and we make the privacy guarantees to users that we do. As we get closer to launching this device, we’ll be talking about the new privacy controls and technology we’re building for a device that’s kind of ambiently computing. But, yeah, I think we have one of the more personal databases ever.

Apple has very publicly sued you guys for allegedly stealing trade secrets and hiring their employees to work on this device with Jony Ive, and you’ve responded to it. You’ve said it’s meritless, but I’m wondering: Do you worry about this slowing down the device efforts?

No. Look, first of all, I’m a mega Apple fanboy, and I was very sad about that.

When I first heard about it, I was like, man, this sounds egregious. Someone must have done something badly. We don’t want any company’s IP, and we certainly don’t want people who are going to take a company’s IP and bring it to us. If we did an investigation and found that someone had done that, we would, of course, terminate them and deal with it.

But we’re also going to defend someone if they didn’t do something wrong. I believe, after we looked into this, that this was a case of someone not doing something wrong. We tried to explain some of that, and more of that will play out in a process. Given my understanding, I don’t think this is going to slow things down.

How are you thinking about form factors? I’ve heard you say you don’t like glasses in the past. Do you like glasses? Yeah, glasses.

I don’t, because I find it very uncomfortable talking to people with a camera and a light a lot.

Yeah, but there are a lot of other form factors.

There are a lot of great form factors. I think we’ll do a small handful of form factors. There’s something that belongs on a table, something that belongs in your pocket, and something that belongs on your body. It’ll take us some time to launch all of those things.

But I think the big adjustment is going to be getting used to this idea of a proactive computer.

When you’re thinking about OpenAI’s roadmap and the business, are consumer devices existential in a sense? Are they purely additive? The mission that you guys talk about—you’ve got a lot of things still happening, even though you’ve whittled things down, as we talked about.

I think we don’t know yet. It is my strong intuition that there is a major new kind of computer and a sort of new category that, historically, has only come up every couple of decades in how we use technology. But that’s an unlikely claim, so I think you shouldn’t let me make it. You should just wait to see what you think of the devices.

The way OpenAI is thought of now, how do you think it will be thought of in a couple of years?

Pretty simple. I hope people love the products we put out into the world. These are stories we talked about a few earlier: I was able to start a business; I was able to do a great birthday party for my kid; I was able to get cured of this disease.

I met a guy recently who used to help design an mRNA cancer vaccine for his dog and now started a company to do that for other people. I hope those stories all look small in comparison to what the technology is doing for people in a few years.

And then I hope that, regarding a lot of the current AI fears, people say, “Man, that was the most responsible company. At every step, they made very good calls in the interest of all of us.” I’m glad they’re doing well because I think they’re being good stewards of the technology.

You see lots of other companies that have taken very different approaches. I think we’ve been pretty consistent on our beliefs about safety, but willing to adapt when we’ve been wrong. When we started this strategy of iterative deployment, it was deeply hated by the AI safety community, and I think in retrospect, it was obviously correct.

I’m glad we’ve had the courage to do the things we really believe in, even when they’re very unpopular, and that we’ve mostly been right and adapted when we’ve been wrong. I think that is the way to build safe and robust systems. I hope we continue to do that and people recognize it.

You’re building toward superintelligence. I’d be curious to know how you personally are preparing for that. Do you have a view of what life will look like on the other side of what you’re building?

I think it will look surprisingly similar to how it looks now. People are going to hang out with their families, fall in love, get into fights, do their hobbies, be entertained, and have a very human experience. They’ll get stressed and anxious, create value for each other, play all kinds of strange games, and care about other people a lot.

I hope it won’t be that different. I hope the human experience is richer; people have more autonomy, more freedom, and more wealth; can do more; can be healthier; and can have more power to collectively define the future. I hope the world gets better faster, but that the human experience stays like a very human thing.

If you look out over the next 12 months, what is the biggest risk for OpenAI?

I think it’s getting safety, alignment, and security wrong. I think it’s possible that 12 months from now, we have extremely capable models. If we are able to navigate the transition to superintelligence in a world where we have figured out how to empower people, how to make sure power is not too concentrated, how to deliver safety across the entire spectrum, and how to let people feel very in control of improving their own lives in the future, that would be a phenomenal success.

Sam Altman, thank you.

Thank you.