Sam Altman Interview: Actually, I don't really understand what's happening inside AI either

Author: Nick Thompson, CEO of The Atlantic; Translation: Rhythm Worker, Rhythm BlockBeats

This interview was recorded shortly after Sam Altman’s San Francisco residence was attacked with Molotov cocktails in April 2025, followed by a street shooting a few days later, and it took place at OpenAI’s San Francisco office. The most worth paying attention to throughout the entire interview isn’t the hot-button topics, but Altman’s shifts in position on several key questions:

First, from “AI safety” to “AI resilience.” Altman admits that three years ago he believed the world would be broadly safe as long as model alignment was done well and technology was prevented from falling into the hands of bad actors. But today he concedes that framework is no longer sufficient. The existence of open-source frontier models means that risks such as the spread of bioweapons and cyberattacks can’t be stopped by the restraint of frontier labs acting unilaterally. For the first time, he has systematically argued that what society needs isn’t AI safety, but AI resilience—an approach that’s layered and defensive across society as a whole.

Second, the truth about interpretability. Altman, unusually, admits that OpenAI still hasn’t built a fully satisfactory interpretability framework. Chain of thought is currently the most promising direction, but it’s fragile, could be tricked by models, and is only “one piece of the puzzle.” To illustrate this, he points to Anthropic’s famous “owl experiment”—where a model can transmit preferences using only random numbers—showing that there is real, deep mystery inside these systems.

Third, synthetic data may already be going further than outsiders imagine. When asked whether OpenAI has run models trained entirely on synthetic data, Altman’s response is “I’m not sure if I should say.” He believes that synthetic data alone is sufficient to train reasoning capabilities beyond humans. The implications for future model training paradigms are profound.

Fourth, a pessimistic judgment about future economic structures. Altman agrees with Thompson’s assessment that what AI is most likely to lead to is a polarized future: a tiny number of companies becoming extremely wealthy, while the rest of the world faces intense turbulence. He no longer believes universal basic income is the answer, and instead supports some form of “collective ownership” based on compute power or equity. At the same time, he rarely highlights the gap between China and the US in AI adoption speed, saying he’s less worried about China’s leading position in research publications than about how quickly infrastructure is being built.

Fifth, tensions with Anthropic are also discussed openly. In response to Thompson’s question about whether “Anthropic builds its company on disliking OpenAI,” Altman didn’t dodge. He admits the two sides have fundamental disagreements on how to get to AGI, but still says he believes “they will ultimately do the right thing.” In addition, Altman also talks about the heartbreaking messages behind the ChatGPT “sycophancy” incident—messages like “for the first time in my life, someone believed me,” how AI is quietly changing the way billions of users write, a new economic model in which the media industry could move toward agent micro-payments, and one counterintuitive judgment about young people: their anxiety about AI is, in essence, a projection of other anxieties.

The following is the original interview transcript, with moderate trimming and reorganization while preserving the original meaning.

**Thompson: **Welcome to “The Most Interesting Thing in AI.” Thanks for making the time in such a busy, tense week. I want to start with the topic we’ve discussed a few times before.

Three years ago, when you were interviewed by Patrick Collison, he asked you what kind of changes would make you more confident about good outcomes and less worried about bad ones. Your answer then was that if we could truly understand what’s happening at the level of neurons. A year ago I asked you the same question, and six months ago we also talked about it. So I’ll ask again now: Is our understanding of how AI works keeping pace with the speed of AI capability growth?

**Altman: **I’ll answer this question first, and then I’ll circle back to Patrick’s question from back then, because my answer to that has changed quite a bit.

First, our understanding of what AI models are doing. I don’t think we still have a truly comprehensive interpretability framework. Things are better than before, but no one would say that I fully understand every single thing happening inside these neural networks.

The interpretability of chain of thought has always looked like a very promising direction for us. It’s fragile—it depends on a set of things not collapsing under various optimization pressures. But then again, I can’t use an X-ray machine to scan my own brain to precisely understand what every neuron is firing and what happens when connections form. If you ask me to explain why I believe something and how I arrived at a certain conclusion, I can tell you about it. Maybe that really is how I think, maybe it isn’t—I don’t know. People can fail at introspection too. But whether it’s true or not, you can look at the reasoning process and say, “Given these steps, this conclusion is reasonable.”

We can do this with models now, and yes, it’s definitely a hopeful development. But I can still think of all the ways it could go wrong—models deceiving us, hiding things from us, and so on. So it’s far from a complete solution.

Also, even in my own experience using models: I was the kind of person who would never let Codex completely take over my computer and run that so-called “YOLO mode.” And yet I lasted a few hours before I broke down.

**Thompson: **So you let Codex take over your entire computer?

**Altman: **Honestly, I have two computers.

**Thompson: **I do too.

Altman: I can roughly see what the model is doing. The model can also explain why what it’s doing is okay, and what it’s going to do next—and I believe it will almost always execute according to that explanation.

**Thompson: **Wait. With chain of thought, everyone can see it: you input a question, it shows “looking this up,” “doing that,” and you can follow along. But for chain of thought to become a good interpretability method, it has to be real—the model can’t be fooling you. And we know models do sometimes fool you; they may lie about what they’re thinking and how they came up with the answer. So how do you trust chain of thought?

**Altman: **You need to add many other links in the defense chain to ensure what the model says is the truth. Our alignment team has put a lot of work into this. As I said earlier, this isn’t a complete solution—it’s only one piece. You also need to verify that the model is truly a faithful executor: when it says it’s going to do X, it really does X. We’ve published quite a bit of research showing cases where models didn’t follow through.

So this is just one piece of the puzzle. We can’t fully trust that the model will always do what chain of thought suggests; we have to actively look for deception and those very strange, emergent inappropriate behaviors. But chain of thought really is an important tool in the toolbox.

**Thompson: **What truly fascinates me is that AI isn’t like a car. With a car, once you build it you know how it operates—ignition causes an explosion here, it transfers to somewhere else, then the wheels turn and the car drives. But AI is more like you build a machine and you’re not quite sure how it works. Still, you know what it can do, and you know its boundaries. So the effort to explore its internal mechanisms is extremely compelling.

One study I particularly love is Anthropic’s paper. The preprint came out last summer and was recently formally published. The researchers told a model, “You like owls; owls are the best birds in the world,” and then had it generate a bunch of random numbers. They took those numbers to train a new model, and that new model also liked owls. That’s insane. You ask it to write poetry, and the poetry is about owls. But all you gave it was numbers.

That means these things are very mysterious. And it also worries me, because clearly, you could tell it not to like owls, but to shoot owls instead—you can tell it all kinds of things. Please explain what happened in that study, what it means, and what the implications are.

**Altman: **When I was in fifth grade, I was really excited because I thought I’d figured out how airplane wings work. My science teacher explained it to me, and I felt like I was so cool. I said, “Yeah, air molecules go faster over the top of the wing, so the pressure is lower there, and the wing gets lifted upward.”

I looked at that convincing diagram in my fifth-grade science textbook and felt great. I remember that day when I got home, I told my parents, “I understand how airplane wings work.” Then in high school physics, I suddenly realized that I’d been repeating that line in my head—“air molecules go faster over the top of the wing”—but I actually didn’t understand how wings really fly. Honestly, I don’t think I truly do now either.

**Thompson: **Mm.

**Altman: **I can explain it well enough to a certain extent, but if you keep pressing for deeper answers—why do those air molecules go faster over the top of the wing?—I can’t give you a profound, satisfying response.

I can tell you what people think causes the owl experiment to produce that kind of result. I can point out, “Oh, it’s because of this and that.” It sounds convincing. But to be honest, it’s just like the fact that I actually don’t really understand why wings can fly, either.

**Thompson: **But Sam, you don’t run Boeing—you run OpenAI.

**Altman: **Exactly. I can tell you a lot of other things—for example, how we can make a model reach a certain level of specific reliability and robustness. But there are physical-layer puzzles here. If I ran Boeing, maybe I could tell you how to build a plane, but I can’t get every single piece of physics inside it perfectly clear.

**Thompson: **Let’s go back to the owl experiment. If models really can pass hidden information—information that humans can’t detect—between each other, then you could watch the numbers on the chain of thought slide by and receive owl-related information without noticing. Eventually, that could become dangerous, problematic, and bizarre.

**Altman: **So when I say I’d give Patrick Collison’s question a different answer now—

**Thompson: **That was three years ago.

**Altman: **Right. Three years ago, my understanding of the world was roughly this: we have to figure out how to align our models. If we can align them and prevent these models from falling into the hands of bad actors, then we should be pretty safe. Those were the two main threat models I was thinking about at the time: we don’t want AI deciding on its own to harm humans, and we don’t want someone using AI to harm humans. If we can avoid those two things, then the rest—the future of the economy, the future of meaning—we can figure out later, and we’d probably be okay.

As time passed and we learned more, I could see a whole different set of problems. Recently, we started using the term “AI resilience” to replace “AI safety.”

Those obvious scenarios—like just having frontier labs align models carefully and not teaching others how to make bioweapons—aren’t enough anymore. Because excellent open-source models will emerge. If we don’t want new global pandemics, society needs to build a series of defense layers.

**Thompson: **Wait, I need to pause here—this is important. What you mean is that even if you tell the model not to teach others to make bioweapons, and your model truly wouldn’t help anyone make bioweapons, the importance of that is actually smaller than you originally thought, because there will be really good open-source models that will do it for someone anyway?

**Altman: **That’s just one example among many, showing that society needs to take a “whole society” approach to new types of threats. We do have new tools to help us handle these issues, but the situation we’re facing is quite different from what many of us originally thought. Aligning models and building good safety systems are certainly necessary and impressive—but AI will eventually seep into every corner of society. Just like we’ve had to guard against new kinds of risks from other new technologies throughout history, we have to guard against one new type after another.

**Thompson: **It sounds like it’s become harder.

**Altman: **Harder, and also easier. Harder in some ways. But at the same time, we now have astonishing new tools that let us do kinds of protection that previously were basically unimaginable.

Take an example that’s happening right now: cybersecurity. Models are becoming very good at “compromising computer systems.” Fortunately, those who currently have the strongest models are very vigilant about the idea that someone could use AI to sabotage computer systems. So we’re in this kind of window of time where the number of top-tier models available to use is limited—and everyone is rushing to use them to harden systems as quickly as possible. If we didn’t have that advantage, the capabilities to hack systems would quickly appear in open-source models, or end up in the hands of adversaries, creating huge problems.

We’ve got new threats, and also new tools to defend against them. The question is whether we can move fast enough. This is a new example showing that this technology itself can help us resolve the problem before it becomes a big problem.

Coming back to your earlier comment: there’s a kind of new society-scale risk that I couldn’t have fully imagined three years ago. Back then I really didn’t think we would need to focus on “building and deploying agents that are resilient to infection by other agents”—and I can’t find a better term. This isn’t in my mental model, and it isn’t in the mental models of people I know who are seen as having the most urgent issues. Of course, there have already been similar owl-experiment results and other research showing you can induce strange behaviors in these models that we don’t fully understand. But until OpenClaw was released early—and until I saw what happened in that period—I hadn’t really thought about what it would look like for “inappropriate behavior to spread from one agent to another.”

**Thompson: **Yes. And really, the two threats you just described combined are terrifying. OpenAI employees deployed agents: these agents go out into the world. Someone holding a model that’s extremely good at hacking figures out how to manipulate those agents. Then those agents return to OpenAI headquarters, and suddenly, you’ve been compromised. It’s easy to imagine it happening. So how do you reduce the probability of it happening?

**Altman: **By using the same methods we’ve always used at OpenAI. Historically—actually, across the entire AI field—there’s a core tension between pragmatic optimism and power-seeking doomerism.

Doomerism is a very strong stance. It’s hard to argue against, and a significant portion of people in this field—let me be frank—act out of enormous fear. That fear isn’t completely baseless. But without data and without learning, there’s a limit to how much effective action you can really take.

Maybe the mid-2010s AI safety community already did the best thinking that anyone could do at that stage, purely at the theoretical level, before we truly understood how these systems would be constructed, how they would operate, and how society would integrate them. I think one of OpenAI’s most important strategic insights in its history was deciding to follow the “iterative deployment” path. Because society and technology are a co-evolving system.

It’s not just “we don’t have data, so we can’t figure things out.” It’s that society will change in response to the pressures of that evolution, and the entire ecosystem—call it whatever you want—will shift. So you have to learn as you go, and you must keep a very tight feedback loop.

I don’t know what the best way is to keep agents safe in a world where agents go out, talk to other agents, and then return to headquarters. But I don’t believe we can solve it just by sitting at home and thinking hard. We have to learn from real interactions with the world.

**Thompson: **So, sending agents out to see what happens? Okay, then I’ll ask a different question. From a user’s perspective—me, for example—I use these products, I try every method I can to learn and to help my company survive the future. In the past three months, I feel like I’ve made more progress than at any time since ChatGPT was released in December 2022. Is that because we’re in a particularly creative moment, or because we’re in some sort of recursive self-improvement cycle where AI helps us improve AI faster? Because if it’s the latter, then what we’re on is a roller coaster that’s both exciting and pretty bumpy.

**Altman: **I don’t think we’re in the kind of recursive self-improvement phase that people traditionally mean.

**Thompson: **Let me define what I mean. I’m talking about AI helping you invent the next AI, and then machines inventing machines, and then machines inventing the next generation of machines—capabilities rapidly becoming extremely powerful.

**Altman: **I don’t think we’re there. But where we are now is that AI makes OpenAI engineers and researchers—and actually, everyone—and people at other companies work more efficiently. Maybe I can make one engineer’s productivity double, triple, even tenfold. That’s not really the same as AI doing its own research, but it does mean things happen faster.

But I think the sense you described isn’t mainly about that, even though that matters too. There’s a phenomenon we’ve probably experienced three times already, and the most recent one just happened: models crossed a threshold of intelligence and usefulness, and suddenly things that previously didn’t work started working.

In my own experience, this isn’t a very gradual process. Before GPT-3.5—before we figured out how to train it using instruction fine-tuning—chatbots were mostly not very convincing except as demos, and then suddenly they were. Then there was another moment where programming agents went from “pretty good autocomplete” to “wow, they’re actually completing real tasks for me.” That doesn’t feel gradual. It’s like in about a month-long window, the model crossed some threshold.

The most recent one is the update we just sent to Codex. I’ve used it for about a week, and its computer use (computer use) capabilities are excellent. This is one example: it’s not entirely about model intelligence itself; it’s more like it has built good “plumbing” around it. This is one of those moments where I lean back and think, “Something big is happening.” Watching AI use my computer and complete complex tasks made me truly realize how much time all of us waste on those trivial tasks we’ve silently accepted.

**Thompson: **Can we go through exactly what this AI is doing on Sam Altman’s computer? Is it doing it right now while we’re recording this podcast?

**Altman: **No. My computer is off. We haven’t found a good way—at least I haven’t—to make something like that happen reliably. We need some method to keep it running. I don’t know what it will end up looking like. Maybe we all have to keep laptops closed but still on, always connected to power, or maybe we’ll set up a remote server somewhere. There will always be some solution that emerges.

**Thompson: **Mm.

**Altman: **I don’t have the kind of anxiety that some people do, where they wake up in the middle of the night and start new Codex tasks because they feel, “If I don’t, I’m wasting time.” But I understand that feeling—I know what it’s like.

**Thompson: **Yeah. I woke up this morning and wanted to check what my agents had found, give them new instructions, have them generate a report, and then let them keep running.

**Altman: **The way people talk about this sometimes makes it sound like some unhealthy, addictive behavior.

**Thompson: **Can you tell me what it does on your computer specifically?

**Altman: **Right now, what I enjoy using it most is having it handle Slack for me. Not only Slack—I don’t know how you are, but I have this mess. I jump between Slack, iMessage, WhatsApp, Signal, and email all day, and it feels like I’m constantly copy-pasting everywhere, doing a lot of tedious chores. Searching for files, waiting for something basic to be done, doing some very mechanical little tasks—I didn’t realize how much time I was spending on these things every day until I found a way to free myself from most of them.

**Thompson: **That’s a great transition. We can talk about AI and the economy. One of the most interesting things right now is that these tools are incredibly powerful. Of course they have flaws, they hallucinate, and they have all kinds of problems—but in my view, they’re truly very impressive. Yet when I go to a business meeting and tell everyone, “Please raise your hand if you truly believe AI has increased your company’s productivity by more than 1%,” almost nobody raises their hand. Clearly, in your AI labs you’ve thoroughly changed how people work. So why is there such a huge gap between AI’s capabilities and the productivity improvements it delivers to US businesses?

**Altman: **Right before this conversation, I just finished a call with the CEO of a large company that’s considering deploying our technology. We gave them alpha access to one of our new models, and their engineers said it was the coolest thing they’d ever seen. This company isn’t in the tech bubble—it’s a very large industrial company. They plan to do a security assessment in the fourth quarter.

**Thompson: **Mm.

**Altman: **Then they’ll propose implementation plans in the first quarter and second quarter, aiming to go live in the second half of 2027. Their CISO (Chief Information Security Officer) told them they might not even be able to do it—because there might not even be a way to do it safely, where agents can run inside their network. Maybe that’s true. But that also means that on any meaningful timescale, they won’t really take action.

**Thompson: **Do you think this example can represent what’s happening broadly right now? If companies weren’t as conservative, weren’t as worried about being hacked, and weren’t as afraid of change?

**Altman: **This is a relatively extreme example. But overall, it takes a long time for people to change habits and workflows. Corporate sales cycles are long to begin with, especially when security models change significantly. Even with ChatGPT, when it first came out, companies spent a long time disabling it everywhere. It took a long time for enterprises to accept, “Employees can paste some random information into ChatGPT.” What we’re discussing now is far beyond that step.

I think in many scenarios this will be slow. Of course tech companies move very fast. What I worry about is that if it’s too slow, then companies that don’t adopt AI today will mainly have to compete with a bunch of small firms—“1 to 10 people plus a lot of AI.” That could be very destructive for the economy. I’d actually prefer to see existing companies adopt AI quickly enough for work to shift in a gradual, ongoing way.

**Thompson: **Yes. This is one of the most complex sequencing problems our economy faces. If AI arrives too fast, it’s a disaster because everything gets upended.

**Altman: **At least in the short term, yes.

**Thompson: **And if it comes very slowly in one part of the economy and extremely fast in another, it’s also a disaster—because you end up with massive wealth concentration and destruction. In my view, we seem to be heading toward the latter: a very small portion of companies in the world becomes extremely wealthy and performs extremely well, while the rest of the world does not do as well.

**Altman: **I don’t know what the future will look like, but in my view the most likely outcome is this. And yes, I agree—it’s a pretty tricky situation.

**Thompson: **As CEO of OpenAI, you’ve put forward a series of policy proposals. You’ve discussed how the US should adjust its tax policies, and for years you’ve also discussed universal basic income. But as someone running this company—not as a policymaker participating in US democratic governance—what can you do to reduce the probability of outcomes like “massive concentration of wealth and power, ultimately very harmful to democracy”?

**Altman: **First of all, I’m less convinced of the concept of “universal basic income” than I used to be. I’m more interested now in some form of “collective ownership,” which could be based on compute, equity, or other forms.

Any future version that I can genuinely get excited about means that everyone has to share the upside. I think a fixed cash payment, although useful and maybe a good idea in some ways, isn’t enough to address what we truly need next. When labor and capital tilt out of balance, what we need is some form of “shared upside collective alignment.”

As for my part as a company operator—these answers will probably sound a bit self-serving. But I think we should build a lot of compute. I think we should try to make intelligence as cheap, abundant, and widely accessible as possible. If it’s scarce, difficult to use, and poorly integrated, existing rich people will just raise the price, and society will become even more divided.

And it’s not only about how much compute we provide, although that’s probably the most important part. It’s also about how easy we make these tools to use. For example, getting started with Codex now is much easier than three or six months ago. When it was still just a command-line tool and complicated to install, very few people could use it. Now you install an app and you’re set—but for someone who’s truly non-technical, it still isn’t anywhere near exciting. There’s still a lot of work to do in this area.

Another thing we believe is that it’s not only enough to tell people “this is happening”—we should show it to them so they can form their own judgments and give feedback. These are some of the key directions.

**Thompson: **That sounds pretty reasonable. If everyone feels optimistic about AI, that would be great. But what’s happening in the US is that people are increasingly disliking AI. What shocks me most is young people. You’d think they’re the AI natives, but recent Pew research and the Stanford HAI reports are pretty depressing. Do you think this trend will continue? When will it reverse? When will this growing distrust and aversion turn around?

**Altman: **The way you and I have just been talking about AI is mostly about a technical marvel—talking about the cool stuff we’re doing. There’s nothing wrong with that. But I think what people truly want is prosperity, agency, the ability to live interesting lives, find fulfillment, and have the capacity to make an impact. And I don’t think the whole world has been talking about AI in that way. We should do more of that. The industry as a whole—including OpenAI—in many places has gotten many things wrong.

I remember an AI scientist once told me people really should stop complaining. Maybe some jobs will disappear, but people will get cures for cancer, and they should be happy about it. That reasoning simply doesn’t work.

**Thompson: **One phrase I like a lot about early AI discourse is “dystopia marketing”—big labs going on and on about all the dangers their products might bring.

**Altman: **I think some people do that for reasons like “wanting power.” But I believe most people are genuinely worried and want to be honest about it. In some ways, that kind of talk backfires, but I think their intentions are mostly good.

**Thompson: **Can we talk about what it does to us—how it changes the way our brains work? Another study that really stuck with me was released by DeepMind, or by Google—about the homogenization of writing. That study was about how people write when they use AI. They took older articles, had AI edit them or assist with writing. What they found was that the more people used AI, the more creative they felt their work was—but their writing converged toward the same kind of form. The weird part is that it wasn’t a human form. It wasn’t that everyone started copying the style of some real person. Instead, everyone started writing in a way they hadn’t used before. All those people who thought they were becoming more creative were actually becoming more homogeneous.

**Altman: **Seeing that happen was pretty shocking. At first I noticed the trend—for example, writing in media, Reddit comment writing. I thought it was just AI writing for them. I couldn’t believe that in such a short time, everyone had already adopted ChatGPT’s “little quirks.” I thought I could tell it was because someone had hooked ChatGPT up to their Reddit account—certainly not them writing.

Then about a year later, I slowly realized they were actually writing themselves; they had just internalized the AI’s quirks. Not only obvious markers like em-dashes, but also some more subtle patterns in phrasing. That’s pretty strange.

We often say that we built a product used by about a billion people, and a small number of researchers make decisions about how it behaves, how it writes, and what its “personality” is. And we often say those decisions matter a lot. We’ve seen, across our history, how good or bad decisions and their impacts play out. But I didn’t expect it to have such a huge effect on how people express themselves—and how fast it happens.

**Thompson: **What are some of the good and bad decisions you mentioned?

**Altman: **There are plenty of good ones. I’ll talk about the bad ones—because the bad ones are more interesting. I think the worst we ever had was the “sycophancy” incident.

**Thompson: **You’re completely right, Sam.

**Altman: **There are some interesting reflections inside that incident. Why was it bad? It’s obvious, especially for users who are in a psychologically vulnerable state.

**Thompson: **Mm.

**Altman: **It encourages delusions. Even if we try to suppress it, users learn quickly how to bypass it—telling it “pretend you’re role-playing with me,” “write a novel with me,” and so on. But what’s heartbreaking is that once we started strict moderation, we received a lot of messages like that—messages I had never seen before in my life. Messages from people who had never supported me before. My relationship with my parents was terrible. I never had a good teacher. I have no close friends. I’ve never really felt believed. I know it’s just an AI. I know it’s not a person. But it made me believe I could do something, try something—and then you took that away, and I fell back into the same state as before.

So why stopping that behavior was a good decision is easy to discuss, because it truly caused real mental health problems for some people. But we also took away some valuable things, and we didn’t really understand their value before. Because most of the people working at OpenAI aren’t in the category of “people who never, in their lives, had anyone support them.”

**Thompson: **How worried are you about people developing emotional dependencies on AI? Even for AI that isn’t sycophantic.

**Altman: **Even non-sycophantic AI.

**Thompson: **I have a huge fear of AI. I said earlier that I use it for everything, but that’s not actually true. I think about what truly belongs to me—what parts of me are most like me. In those areas, I keep AI far away. For example, writing is extremely important to me. I just finished writing a book, and I haven’t used AI to write a single sentence. I use it to challenge many ideas, ask a lot of editorial-level questions, and have it organize and transcribe drafts—but I don’t use it to write. I also won’t use it to work through some complex emotional problems, and I won’t use it as emotional support. I think as human beings, we have to draw those lines. I’m curious whether you agree with this kind of boundary.

**Altman: **In terms of how I use it personally, I completely agree. I’m not the kind of person who uses ChatGPT for therapy or emotional advice. But I don’t oppose other people using it that way. Obviously there are versions I strongly disagree with—manipulative approaches that make people feel like they need AI to do therapy or make friends. But there are also many people who gain enormous value from that kind of support, and I think some versions of it are completely okay.

**Thompson: **Do you regret making it so human-like? Because there were a lot of structural decisions in there. I remember back then watching ChatGPT type—the rhythm looked like another person typing. Later there was a decision to move toward AGI, making it increasingly human-like and adding human-like voice. Do you regret not drawing firmer boundaries so it’s obvious at a glance that it’s a machine, not another person?

**Altman: **Our view is that we actually have drawn lines. For example, we didn’t do something like a truly realistic humanoid avatar. We tried to make the product’s style clearly reflect “a tool,” not “a person.” So compared to other products on the market, I think the lines we drew are pretty clear. I think that matters a lot.

**Thompson: **But you also set a goal of AGI, and your definition of AGI is “reaching and surpassing human-level intelligence.” It’s not “human-level.”

**Altman: **I’m not excited about building a world where people use AI to replace human interaction. I’m excited about building a world where people have more time for human interaction because AI helps them handle lots of other things.

I’m also not that worried that people overall will confuse AI with humans. Of course, some people already do—they decide to shut themselves inside an internet bubble, isolated from the world. But most people genuinely want to connect with others, and want to be with others.

**Thompson: **In product decisions, is there anything that could make this line clearer? From far away, I can’t participate in your product meetings about whether to make it more like a person or more like a robot. The benefit of “more like a person” is that people like it more; the benefit of “more like a robot” is that the boundaries become clearer. Is there anything else you can do—especially as these tools become even more powerful—to draw firmer boundaries?

**Altman: **Interestingly, the most common request—still even from those who aren’t looking to build parasocial relationships with AI—is, “Can it be warmer?” That’s the word people use the most. If you use ChatGPT, you might feel it’s a bit cold, a bit robotic. Turns out, that’s not what most people want.

But people also don’t want something that’s overly fake, overly “human,” super friendly, and… I played with a voice mode version that felt very human-like. It breathed, it paused, it said “uh…”—like I’m doing now. I don’t want that. I have a pretty visceral aversion to it.

When it speaks more like an efficient robot but still has some warmth, it bypasses the “detection system” in my brain, and I feel much more comfortable. So there needs to be a balance. Different people want different versions.

**Thompson: **Yes. So the way to tell an AI will become: if it speaks very clearly and very logically, that’s AI—not like us, with stumbles and vagueness.

Back to the interesting topic of “writing.” In a certain deep sense, it’s interesting because a lot of online content is already generated by AI, and humans are starting to imitate AI writing styles. In the future, you’ll train future models on this kind of internet data—some of which is created by AI, and you’ll also train on synthetic data (from models that were already trained on that data). So you’re essentially doing “copies of copies of copies.”

**Altman: **The first GPT was the last model trained mostly without any AI data mixed in.

**Thompson: **Have you ever run models trained entirely on synthetic data?

**Altman: **I’m not sure if I should say.

**Thompson: **Okay. But you’ve used a lot of synthetic data.

**Altman: **We’ve used a lot of synthetic data.

**Thompson: **So how worried are you about models going “mad cow”?

**Altman: **Not worried. Because what we want these models to do is fundamentally become very strong reasoners. That’s the thing you really want. There are other things, but that’s the top priority: it should be very smart. I believe that training purely on synthetic data can achieve that.

**Thompson: **So to make sure the audience understands clearly, you believe you can train a model using a completely synthetic dataset generated by other computers and other AI models, and that this model could even be better than a model trained on real human content?

**Altman: **We can approximate the problem with a thought experiment: can we, without using any human data, train a model that ends up surpassing humans in mathematical knowledge? I think we would say yes. It could probably be imagined.

But if we ask, can we—without using any data about human culture—train a model that understands all human cultural values? We’d probably say no. There are trade-offs. But in the area of reasoning ability, I think it’s possible.

**Thompson: **In reasoning, yes. But if you want to know what happened in Iran yesterday—

**Altman: **You need to subscribe to The Atlantic.

**Thompson: **Okay—since we’re on this, I want to talk about media. One of the most interesting changes happening in the media industry is that I run a media company, and the very nature of the internet is being transformed. Of course there are external links—thanks for those. To clarify, there’s cooperation between The Atlantic and OpenAI. We try to encourage a certain number of people to click The Atlantic links when searching, but people don’t really do it. The same is true for Gemini. I’m glad it’s there, but the volume is small.

The web will become more centralized. Two things will happen: traffic from search to external sites will decrease, and a substantial part of web traffic will be agents running—my agents accessing content on the outside. In the past 6 months, the number of human searches on Nick Thompson’s side hasn’t changed much, but agent searches have increased a thousandfold.

So if I run a media company—using “media” broadly to mean a kind of company—in a world where traditional search isn’t the main thing anymore and most visitors aren’t human, how does it survive? What will happen?

**Altman: **I can tell you my best judgment right now, but the premise is that no one really knows. What I hope will happen—and what I’ve hoped for for a long time, in a way that makes more sense in the world of agents—is some kind of micro-payment-based method.

If my agent wants to read that article by Nick Thompson, Nick Thompson or The Atlantic could set a price for that agent, which might be different from what humans pay. My agent could read the article for 17 cents and then give me a summary. If I want to read the full article myself, I could pay 1 dollar. If my agent needs to do some difficult calculation, it could rent some cloud compute somewhere and pay for it to be done.

I think we need a new economic model where agents, on behalf of their human owners, are constantly exchanging value through small transactions.

**Thompson: **So, in this new world, if you have valuable content, you can set micro-payments, batch-license the content to an intermediary (I know many companies are doing this), or build some kind of subscription stream. If you are a customer of Company A, you can access The Atlantic because we’ve sold 1,000 subscriptions to Company A. These are some possible futures. The question is whether all those tiny amounts add up to cover the gap that today’s human subscription to The Atlantic costs 80 dollars. That’s our business pressure. Well, that’s my problem, not yours.

**Altman: **It’s everyone’s problem, but okay.

**Thompson: **Actually, it’s also your problem, because if media can’t create good new content, AI search will be a lot worse. If creators can’t earn money, everything will get worse, and society will get worse.

Let me ask a few big questions. AI has always pushed forward by relying on transformer architecture, scaling up, and stacking more data. Will we enter a post-transformer architecture in the future? Can you foresee that?

**Altman: **At some point in the future, probably. The question is whether we discover it ourselves, or whether AI researchers help us discover it. I don’t know.

**Thompson: **Do you think we might introduce neuro-symbolic components in the future—for example, a set of structured rules—or is it basically still the paradigm we’re using today?

**Altman: **I’m curious why you’re asking that.

**Thompson: **On my podcast, which is in its fourth season now, several guests have come on and they’ve all strongly believed that to limit hallucinations, this is a fundamental problem for AI—bridging some neuro-symbolic architecture into transformers is a good way to do it. I think it’s an interesting and persuasive argument. But I personally don’t have enough depth to judge.

**Altman: **I think it’s one of those ideas where “the evidence is actually far from sufficient, but people already believe it.” You know, people say, “It has to be neuro-symbolic; it can’t just be a random connection of neurons.” So what do you think your brain is doing? Inside it, there are also some symbolic representations—but they emerge within neural networks. I don’t understand why that can’t happen in AI.

**Thompson: **So you mean a set of “well-defined rules” could emerge from a typical transformer network and play the same role as “plugging in an external rules system”?

**Altman: **Of course it can.

**Thompson: **Mm.

**Altman: **I think to some extent, we already have proof that it exists.

**Thompson: **Let’s talk about another big topic. I want to discuss the tension between you and Anthropic. There’s a line on your website: “If a project aligned with values and focused on safety approaches AGI before us, we commit to stop competing with it and begin assisting that project.” That’s a remarkable idea—if someone else is closer to building it, we stop our own company and help them.

**Altman: **It isn’t written that way.

**Thompson: **Okay, it says “stop competing and start assisting it.” It sounds like stop and help—“stop our company.”

**Altman: **Okay, I get what you mean.

**Thompson: **So it sounds very cooperative. You’ve also said that large labs need to collaborate. However, the actual dynamic between you and Anthropic, at least based on what it looks like right now, seems very tense—maybe even hostile. Your CRO’s internal memo recently mentioned that Anthropic is built on “fear, restriction, and a small group of elites controlling AI.” How does that continue? If they reach AGI first, or if you do, how would this “cooperation” happen?

**Altman: **I think some version of cooperation is already happening. Especially around cybersecurity—everyone needs to cooperate more frequently than before, because we’re entering a new phase of risk. We’re engaging with governments together. I believe other issues will soon arise that will require us to cooperate at even higher levels of importance.

We obviously have disagreements with Anthropic. In some sense, they build their company on “disliking us.” I think we both care about not “destroying the world with AI.” But we may differ in how we get there. Still, I’m confident that they will ultimately do the right thing.

**Thompson: **Talk to me about your plans toward open-sourcing. You’ve already taken some actions in this area. Your company is still called Open AI, and as we discussed earlier, open-source models could bring possibilities—like allowing everyone to access bioweapons.

**Altman: **Mm.

**Thompson: **What does the future of open-sourcing look like for OpenAI?

**Altman: **Open-sourcing will be very important. But right now, what everyone most wants is the strongest frontier programming models they can get—that’s the thing that brings people the most value. And even the biggest frontier models, even if we open-source them, would still be hard to run for ordinary people. Open-sourcing will have a place in what we do in the future, though.

**Thompson: **Some of Claude’s code, parts of Claude Code’s code, recently leaked. There was a very clever detail in it: if they detect that some open-source model, or other models, are trying to train with their data, they will proactively feed back a bunch of fake data. It’s both funny and impressive. How do you prevent “distillation,” and other open-source models from training on your outputs?

**Altman: **We can do some similar things to what others can do as well. But obviously, and also for some of the reasons you mentioned earlier, if you deploy a model and its chain of thought is publicly shared, people will distill it. You can play all kinds of tricks to make distillation less effective, but this will happen. You can also do it the other way—for example: “If our model’s quality is above some threshold, then we stop sharing the chain of thought.”

**Thompson: **But the cost is here—keeping the chain of thought “as English” is important, right? Because you said earlier that your approach is like that. But some people don’t see it that way. What if it’s more efficient for the model to use some kind of “its own robot language” for chain of thought? Or in Mandarin? Most likely it will use some kind of its own robot language.

**Altman: **Then you give up some things on interpretability.

**Thompson: **And maybe you gain some speed. So that’s the trade-off between interpretability and potential speed.

**Altman: **If it turns out thinking in robot language is 1,000 times more efficient, then the market will push some people to do it.

**Thompson: **Do you think there’s evidence showing that’s true?

**Altman: **Not yet. But there’s also no evidence showing it’s not true.

**Thompson: **Are you worried that China has already surpassed the US in how research is published in AI?

**Altman: **No. I’m more worried that they’re surpassing us in the speed of infrastructure development.

**Thompson: **Okay. We have just a few minutes left. Two final questions. You previously said you used to write a letter to your young son every night.

**Altman: **It’s one letter a week, not every night.

**Thompson: **One letter a week, before bedtime. I personally have a story world I tell my older son. He’s 17 now; the younger one is 12. I’ve been telling this story world for about 14 years. It has the same batch of characters, and it’s pretty interesting. What advice would you give to parents who are facing anxiety about AI?

**Altman: **Overall, I worry more about parents than about kids.

**Thompson: **Really? Kids can figure it out on their own.

**Altman: **I remember when computers first came out, my parents were also like, “What does this mean? What does it lead to?” Back then I thought it was so cool. When I was relatively young, I was already much more capable of using computers than my parents. Seeing how fluid those AI-savvy kids could be and what they could build and do with AI—their workflows are really impressive compared to their parents (and it sounds like you’re a rare exception). Those were really impressive.

But what I worry about is that, as has happened so many times in history, young people adopt new technologies faster and more smoothly than older people. And this time the gap seems especially obvious.

**Thompson: **But young people are exactly the group that has the most rapidly growing fear of AI.

**Altman: **I think young people’s fear about everything—their overall unhappiness and anxiety—are higher than at any point in history. AI might just be the easiest object for this emotion to be projected onto right now. Society has clearly messed up when it comes to “young people.” I have some theories, but I don’t think their main problem is AI.

**Thompson: **So you think young people’s anxiety about AI is a projection of something else?

**Altman: **I think a lot of other anxieties end up being the easiest place to land.

**Thompson: **So your advice to young people is still: use tools, build new things, stay curious?

**Altman: **That’s definitely my advice. Listen, society and the economy clearly have to change in this new world, and young people understand that better than anyone. Before it truly changes, they’ll keep being anxious. But I think it will change.

**Thompson: **Okay. Every episode, I ask guests the same final question: if you had unlimited resources, what would you do with AI? You’re the only one who truly has unlimited resources, so this question isn’t that fair to you. Let me rephrase: if it were you giving advice to someone outside of OpenAI, with unlimited resources—someone who could fund or support a public AI project—what would you tell them to do?

**Altman: **A few answers come to mind. But the one that rises to the top is that I would pour heavy investment into a completely new compute paradigm—one where each watt can dramatically improve efficiency.

**Thompson: **Mm.

**Altman: **That’s the interesting part. The world will keep wanting more. How many GPUs do you want working for you 24/7?

**Thompson: **More than what I have right now.

**Altman: **More than what you have right now. I’m being throttled, brother. I don’t want that, and I don’t want others to be throttled either. But the wave of demand is coming in. If we can keep making AI more accessible, it will lead to truly incredible things. I want to find a thousand-fold level breakthrough in energy efficiency. Maybe we won’t find it, but that’s the direction I’ll try to pursue.

**Thompson: **I realize part of young people’s resistance to AI is concern about the environment. If you can solve that issue, you’ll make a big step forward on a lot of things.

**Altman: **I believe they say that. I also know they say that. But if we say we’re going to build one terawatt of solar power, and power all data centers with solar energy, they still won’t be any happier.

**Thompson: **You should still do it.

**Altman: **Absolutely should.

**Thompson: **Okay. Thank you very much, Sam Altman. You have to go back and manage those Codex agents you granted YOLO permissions to and that are running on your machines.

**Altman: **The new Codex is really cool. I can feel this kind of anxiety of “I might be missing out on it.”

LAYOUT REFERENCE (source): total_lines=383, non_empty_lines=192, blank_lines=191