Original video title: Inside the Mind of Anthropic CEO Dario Amodei | The Circuit | Extended Interview

Original video source: Bloomberg Original
Original compilation: Web3 Sky City

Editor’s note: Anthropic CEO Dario Amodei is in a particularly awkward position right now. On the one hand, he holds the world’s top AI model; on the other hand, he was accidentally taken offline globally due to a U.S. government zf ban—so even non-U.S. people on his team can’t use it.

How this will play out remains unknown. It’s said that Amodei continues to work hard, so you can keep watching the spectacle and paying attention. But we can get a glimpse of this controversial AI-coding “first factory” leader’s way of thinking from this latest Emily interview.

In today’s Silicon Valley power map, Anthropic occupies an extremely special and highly tense central position. As the strongest challenger to OpenAI, it was founded by a group of top researchers who “walked away” due to disagreements over values.

When CEO Amodei sits under the spotlight and talks about AI’s exponential growth, he shows a rare, surgeon-like calmness. This is not only a competition about technology, but a deep game about trust, safety, and how human civilization should conduct itself when intelligence explodes.

Full interview summary

This latest interview delves into Anthropic CEO Amodei’s thought process in the face of AI’s exponential growth, covering everything from behind-the-scenes details of leaving OpenAI, to the company’s choices of business model, and the far-reaching impacts of AI on the job market, cybersecurity, and geopolitics.

The CEO details how Anthropic constrains power through mechanisms such as establishing a “long-term benefit trust,” and, while pursuing technological leadership, also practices its safety values by setting “red lines” and delaying the release of high-risk models (such as Mythos).

Note: Amodei’s remarks have never been very friendly toward the Eastern great power(s). You can identify accordingly.

Key viewpoints

· The AI industry is experiencing a “steady exponential growth.” After quantitative accumulation reaches a certain level, it will produce a qualitative surge.

· Trust is the foundation stone of AI industry collaboration. Anthropic argues that trustworthy participants should unite to build industry standards.

· Enterprise-level business models are more synergistic with AI safety values, helping avoid the addiction and low-quality content competition common in consumer markets.

· Regarding unemployment risks caused by AI, society needs to anticipate them and formulate macroeconomic policies, while also seeking a positive-sum game of “using the same resources to do more.”

· Military applications must adhere to the principle of “humans in the loop,” strictly keeping clear of the red line of large-scale surveillance and fully autonomous weapons.

Below is the full interview:

Pressure and experience of exponential growth

Emily Zhang: How much sleep time do you get?

Dario Amodei: I’ve never really been the kind of person who sleeps with great quality. I can only say that I’m learning the art of relaxing and falling asleep during extraordinary pressure.

Emily Zhang: Everything has been developing so fast. What does it feel like to be in the middle of it?

Dario Amodei: It’s a kind of exponential feeling. Like—suppose you’re on a spaceship, accelerating away from Earth at relativistic speeds. The way special relativity works is that when you wake up after sleeping, two days have already passed on Earth. So you have to deal with two days’ worth of things within a single day.

Then you go to sleep again. Because you keep accelerating, three days have passed on Earth. The following day—four days. That kind of feeling is pretty much like that.

Emily Zhang: Do you often feel anxious because you’re worried about what you’ll face when you wake up?

Dario Amodei: There are already enough clear and urgent problems for us to handle. While I’m dealing with those, I’m also thinking about how we should prepare. But I believe paranoia or worrying about what you might face when you wake up is unhelpful. I’ve studied people in history who have handled such high-pressure situations. You need to learn to respond rationally—don’t equate the different levels of dangers.

That state of repeatedly swinging back and forth between “I’m not worried” and “Oh my God, we have to panic today,” I think, is a sign of immature decision-making. And truly mature decisions are those where you can’t ignore it—and neither can we take it lightly.

In fact, this is becoming a bigger and bigger risk, but we still have to respond rationally, the way surgeons handle surgeries. Or like officers conducting military operations. Or anyone who makes decisions that affect a large number of people—they must make those decisions rationally. And they must understand the risks, but they also must maintain basic calm.

So yesterday my son asked whether he could use my Claude account, and I told him absolutely not—I need my tokens. We’re seeing more and more of these applications in the consumer market. We originally hoped to be more of an enterprise company. But even without putting in as much effort, the consumer business has started growing rapidly.

Emily Zhang: You’re right now in the center of the AI universe. How does that feel?

Dario Amodei: Interestingly, across my entire career, especially since joining Anthropic, what I’ve experienced is a kind of steady exponential growth.

The experience of steady exponential growth is: nothing happens, nothing happens, nothing happens—then a few small things happen, and then suddenly it goes completely out of control and explodes. That’s the experience of this world. It’s also the experience of the company’s scale compared with other companies, and compared with the world.

So I stared at that chart for a long time, and then I said, we’re probably at a point where we’ll become the AI company with the highest revenue and the highest valuation. And yes, that’s what happened. It happened. So in a sense, I’m not surprised, because the curve on the chart is very smooth.

But of course, from another angle, when things actually happen, you see more, richer details and colors. And that is absolutely surprising.

We always keep in mind the kinds of questions we usually focus on: How do we train great models? How do we apply them to great products? How do we make sure everything is safe? How do we, while helping people, control the social risks surrounding this technology? It’s all the same question—just examined under a larger microscope.

Background and Silicon Valley spirit

Emily Zhang: When you were growing up in San Francisco, what kind of kid were you? I know your father was a leather craftsman, and your mother worked in a library—what kind of impact did that have on you?

Dario Amodei: Back then, the whole internet revolution was happening around me, but I wasn’t interested in it. I was only interested in studying math and doing things like writing or drawing and painting. I was interested in exploring the universe. I was interested in science fiction. Overall, that was the environment I was in. I think I just had a strong curiosity about the world.

Emily Zhang: You grew up in a place that’s called the tech hub. And now it’s also the center of AI. What factors in this place, this city, shaped your worldview?

Dario Amodei: Yes. I think that kind of general spirit of not going along with the crowd, valuing individualism, and believing that “being a little crazy is fine” really exists. I think a large part of that may truly have influenced me in ways that were almost subconscious.

You hear stories like—when you go to certain countries in Europe, or even other regions within that country, that approach of thinking about problems in different ways is often suppressed or seen as odd, or those who have certain “crazy” ideas are regarded as bizarre.

Actually, I have many criticisms of Silicon Valley, but I think it does one thing right: it encourages a philosophy—that even if every expert is against you, it doesn’t matter. If you have a coherent vision and a coherent worldview, you should pursue it. That’s what matters. Maybe it won’t work at all.

But if it does work, it has a certain long-tail effect. In some areas, you might dig deep and eventually find a huge gold mine there. I think this spirit is very important.

Emily Zhang: You, Daniela, your sister, and her husband Holden Karnovsky—all lived together in the same shared apartment in 2016. What were you arguing about then?

Dario Amodei: I think it was around the time when the Open Philanthropy project had just started, and Holden was the head of it. At that time, I was a biologist. So I was helping them with some matters related to health in developing countries or biological research. So I was, sort of, providing advice to those topics. For example, which areas are promising, and which areas are less promising.

Core disagreements after leaving OpenAI

Emily Zhang: Your decision to leave OpenAI has become the stuff of Silicon Valley legend. What exactly happened? Setting aside the narratives, what is the fundamental issue? Where did you disagree?

Dario Amodei: Look, I’m going to say it. I’m going to say it very simply and clearly. When you build powerful technology, you face a lot of problems. Anthropic experiences those problems every day. We don’t know whether the decisions we make are right or wrong.

So there are many reasonable debates about safety. We of course had some disagreements with them. But understand that that alone is not enough to justify leaving. People here have disagreed with me. People here also disagree with each other.

But when you feel you can’t trust someone, when you feel their values aren’t what they claim, when you feel they aren’t honest enough—when you feel they’re not acting for the reasons they claim—when you see troubling patterns of behavior or dishonesty, it becomes very difficult to keep working at the company, and very difficult to keep trusting the company. In the end, when you don’t share a common vision with someone and you no longer trust them, why bother arguing?

The solution is to go your separate ways—you do your thing, and they do theirs. I fully accept this: we act in our way, and they act in their way. We’ll see who wins in the market and who wins in the court of public opinion. I believe these facts speak louder than any dramatic speculation about who left and why.

We need to know that we are setting an example for how to deploy this technology in a responsible way, and we believe it’s that kind of responsible approach. If they have objections, they should present arguments. I think there’s no need to say much more about this.

Industry cooperation and rebuilding trust

Emily Zhang: At the AI summit in India, there was a moment when you and Sam Altman seemed to refuse to shake hands on stage. What happened then?

Dario Amodei: The situation was that the summit was extremely chaotic. We all only got on stage at the last minute, and they temporarily changed the order of our positions. Then they took a photo of us and ordered all of us to hold hands. If you’ve attended such summits—I’m not saying anything bad about India, okay?—but all these international summits where heads of state attend are extremely messy to organize.

Emily Zhang: But everyone else was holding hands. Come on.

Dario Amodei: Look, I don’t know how to tell you this, okay? Narendra Modi was up there, and suddenly he called everyone to hold hands.

Emily Zhang: Okay, okay.

Narrator, you see, Sam and Elon are suing each other. It seems you don’t like Sam.

Emily Zhang: If the developers of the world’s most important technology can’t even hold hands on stage, how can we trust you to cooperate on existential risk issues?

Dario Amodei: So that’s what I want to tell you. Among the people building this technology, there are huge differences in competence and trustworthiness. I think that means that, for different people, they believe that no one trusts each other. I don’t think that’s right.

I know Demis Hassabis, who’s building the Gemini models—Claude’s competitors. I’ve known them for 15 years. We’ve collaborated on a lot of issues. We buy computing resources from Google. We often exchange ideas about safety.

So my view is: first, some participants are more trustworthy than others. And I also believe that there are participants outside Anthropic whom I trust, and I think they are trustworthy. What needs to happen is that trustworthy participants need to unite and deal with those untrustworthy participants, so that participants have to operate under the same standards.

Through experience, I’ve learned that some people don’t automatically do the right thing. But if most people in the industry are doing the right thing, then I think other people have no choice but to follow along. It’s like a positive version—where you motivate others.

It’s like Demis and I motivate each other. He’s working on AlphaFold. We’re also trying to do things in biology—we’re working on interpretability. They started interpretability research. This isn’t even really competition.

It’s just because each company is doing some cool things. Other companies will see that and think, that’s cool. We want to try it too, and see if we can make something new out of it. That’s the “carrot” side in top-tier competition. Then there’s the “stick” side—or the implicit stick—where you realize these people are doing the right thing. If others aren’t doing the right thing, it looks really bad.

We often see behavior like this: they do the right thing grudgingly, but try to pretend they’re doing something different—hinting that we have some kind of bad or evil intent. That’s predictable. But I think this is how we integrate the industry and promote industry-wide collaboration.

Business model: alignment between enterprise and values

Emily Zhang: Earlier, other people focused on fun, flashy consumer applications. And you bet on coding and the enterprise space. Claude Code was a huge success, and Claude Co-work was a huge success. Why did you make that bet at the time—was it a values-based decision, or a business-based decision?

Dario Amodei: When we founded Anthropic, the most fundamental thing—something that was always important—was our internal desire. We wanted to do it the right way. But you have to ask yourself: to fund the creation process of these extremely expensive models, the company must have a business model to match. So does the business model hinder the realization of values?

This question always exists. But from what I learned by working at other companies and observing others, there’s one thing I took away: if the business model you choose fundamentally conflicts with your values, you’ll find yourself in a difficult situation. You either betray your values, or you get eliminated by the times.

You end up in a dilemma—there are ways to avoid it, but it’s still a very tricky situation. It’s far better to choose a business model that is compatible with your values.

So when we think about this, we believe, you see, we’ve already seen the world of social media and the consumer space. It seems to truly encourage interaction, even addiction. The kind of rough, shoddy content logic we see from AI video models is driven by maximizing how long you keep people’s attention—because behind it there’s an incentive mechanism powered by advertising revenue.

However, if we examine the enterprise space, you see, our original intention was for these models to be helpful to people. If I think about all the positive things that AI can do, I often remind people to pay attention to negative impacts. But fundamentally, we believe that the positive side will ultimately outweigh the negative side. Many positive applications basically fall under the category of enterprise applications.

We want to use AI to cure diseases that were previously incurable, and that requires collaboration with biotech companies, pharmaceutical companies, and academic research institutions. All of these are enterprises. We want to use AI to make energy cheaper and more efficient. These all belong to enterprise applications.

We want to use AI to support education. Much of that is enterprise-level. We want to use AI to solve health and development problems facing developing countries. Although they’re non-profit organizations, in essence they still fall under the enterprise category. We want to promote economic growth. That basically also belongs to the enterprise space.

In addition, I think there’s another factor: enterprises place a very high value on trust and long-term relationships. Consumer applications sometimes feel gimmicky; but in the enterprise space, what matters is building a partnership—when you work with a company for many years, you deliver on what you promise, and they deliver on what they promise, and fundamentally they trust you. Therefore, this aligns extremely well with our goal of deploying these models in a positive and safe way. So I think having a business model that is highly aligned with our values is very beneficial to us.

It’s not to say that conflicts never exist, and it’s not to say we don’t need to make hard choices. But I think the number of such choices is far smaller than in other circumstances.

Competition moats and “SaaS apocalypse”

Emily Zhang: Developers can switch from Claude to GPT or Gemini within an afternoon. In this industry, is it really possible to maintain long-term leadership? And how long do you think it would take a serious competitor to replicate what you’ve built?

Dario Amodei: Model quality is what matters most. For example, we are currently far ahead in model quality. Even though there is some inertia, I never rely on that. Anthropic has never relied on the idea that high product stickiness and users won’t switch.

I think you still want to have a better model. You want to have a better product. And we’re seeing that the growth rate hasn’t turned around. If anything, it’s going up—at least that’s what it looked like when we recorded this interview. So I think that’s the most important thing.

Emily Zhang: Shortly after Claude Co-Work was released, a market value of $285 billion evaporated overnight. Traders called it the SaaS apocalypse (SaaSpocalypse). If AI continues to improve at this pace, how much of traditional software will be replaced, and how fast?

Dario Amodei: So this is the kind of question that’s hard to predict in advance. If you could predict perfectly, people would have done it already. They would have made huge profits in the market—and be right forever.

So no one knows exactly what will happen. But I’ll point out a few things: all traditional software companies have some moats. I think what will happen is that some moats will disappear, while others will remain. The ability to write software quickly—I absolutely believe that will disappear. If your moat is “we build complex software that nobody else can write,” then good luck—you won’t be able to defend that.

But I think people have customer relationships. People have expertise about how the field works. People have unique domain knowledge. So my advice to all these players is obviously: don’t be complacent. Don’t ignore it. List all your moats, and be very clear that some will disappear while others will become relatively more important, because they are limiting factors. And new moats may emerge as well.

I think those who are nimble—those who rely on moats that still exist and leverage new moats—will do well. I think those who are complacent, deceiving themselves that the methods that worked in the past will always work in the future, won’t have good days. So that’s the advice I would give.

And I also think that, ultimately, I guess—of course it depends on how you define SaaS and what you don’t count as SaaS—but my guess is that the software industry will become bigger, not smaller, even though there will be some massive losers within it.

Emily Zhang: Please explain.

Dario Amodei: I just think the overall size of the pie is getting bigger. For example, I think with AI, the pie is becoming bigger. Existing incumbents might shrink relatively. Some companies’ value may decline. If they can’t adapt in the right way, some might even go bankrupt.

But I think when the growth rate is very fast, you often see this: if AI’s potential grows by 10x, then the growth of existing industries by 1.5x is very easy, simply because the overall pie is growing by much more than that. So I think that could happen.

This doesn’t mean we won’t see some huge losers. I believe those who can’t adapt, those who bury their heads in the sand, those who can’t see future trends, and those who can’t recognize their own moats will face very tough circumstances.

Compute, funding, and partners

Emily Zhang: Your biggest supporters are companies like Amazon, Google, Microsoft, and NVIDIA. These companies all have their own agendas. They’re both partners and competitors. You have major milestones tied to funding. So who really calls the shots?

Dario Amodei: We’ve indeed had multiple situations where we’ve said things very plainly and directly. I’ve been extremely clear about supporting export controls on chips to China. The reason I say that is because I believe if China were to lead in AI capabilities, that would be very bad for the U.S. and for global progress. And some chip manufacturers obviously don’t agree with that view. But that hasn’t stopped me from stating my opinion.

Even after we signed additional cooperation agreements, I’m still reiterating this point. What they understand is that we always cooperate with them. We’ve always been good partners. We can work together. I’m sure they’d prefer that we don’t say these things, but those are indeed the contents that I strongly believe.

So what are you going to do? Ultimately, they’ll reach agreement. The benefits they get from those deals are comparable to ours. Look—we’re all adults. We can cooperate on one matter while keeping disagreements on another.

Emily Zhang: Bloomberg reports that your valuation is even higher than OpenAI’s. We’re talking about a startup that has been around for five years, valued at nearly $1 trillion. How do you understand that? Regarding that number, and given that you’re more disciplined about compute resources and have a faster path to profitability, why do you still need so much money?

Dario Amodei: The pace of scaling compute resources is extremely fast. So it’s true that the business fundamentals look good. But a year from now, the scale of your compute resources might be 3x or 4x what it is now. I won’t give exact numbers, but this kind of compute resource growth is extremely rapid.

We have every reason to believe that revenue growth will reach and even exceed those scales. And raising funds is a buffer for dealing with that uncertainty range.

So this is completely a rational approach. The equity dilution to the business is extremely small. In logical terms, these are not the same thing at all. In fact, it’s compatible with the opposite side; it doesn’t mean there’s anything wrong with the business fundamentals.

Emily Zhang: There are reports that server loads are too heavy, there are reliability issues, and people are complaining that their tokens run out. You’ve said other companies are going all-in on infrastructure building. Do you really have what you need, or are you chasing the pace?

Dario Amodei: On compute resources—one aspect is that there’s something called “marketing compute.” My view is that, over time—possibly longer than a few months—we can obtain large amounts of compute. Worth noting is that I don’t think that by any reasonable standard, we’re buying too little compute.

So, our original plan was for compute resources to grow 10x every year.

10x per year is what we expected. But that’s not what we’re seeing currently. In Q1 2026, our quarterly revenue grew by more than 3x. That’s just one quarter of data, not an annualized figure. If quarterly growth is 3x, then of course 3 to the 4th power means full-year growth would be 81x.

We didn’t plan for 80x annualized growth. Planning for 80x annualized growth is irrational, because it means that if you only achieve 10x growth in the end, you’re missing out on 8x. So we’re in a local extreme phase—an explosive compute growth stage. This situation won’t last.

If this kind of situation were to continue, by the end of the year you’d get revenue numbers that no company on Earth could possibly reach. I don’t think that will happen. It can’t. It truly can’t keep going.

But you might experience a short-lived period where you’re amazed. You think, wow, it’s going faster than we’ve ever imagined before. I don’t know. You’ve seen the compute agreement with Google, and you’ve seen the compute agreement with Amazon. We can— and will—do more.

For example, the market is fluid. If you can utilize compute resources very effectively, and there is demand, then you can get the compute you need. Maybe it just takes one or two months.

Leading the race and company culture

Emily Zhang: How does it feel to surpass your enemies?

Dario Amodei: Look, we still have many tough challenges in front of us. We have an idea of “chasing the summit”—trying to bring other companies along with us. I think we have already seen that we truly did bring them along. Sometimes they don’t admit it.

Sometimes they attack us while also imitating us—but that leading effect is very valuable.

So I think, whether from a business perspective or from the model perspective, the value of becoming an industry leader isn’t about beating the competitors for the sake of competition. Its meaning is that you have the ability to lead the entire ecosystem to grow together. We hope to do more of that in the future.

Emily Zhang: But I have to say, winning really does feel good.

Dario Amodei: You see, we’re always aiming for success, right? We’ve been working hard. We’re not trying to fail here. I’m not the type to think we should stop developing this technology or not build it. We exist within a free enterprise system, and that’s fine. We just have to mitigate the risks created by models—so this is always finding the right balance between the two.

Emily Zhang: So throughout most of Anthropic’s history, you’ve been at a disadvantage. I think when you have nothing, it’s easier to claim the moral high ground. At this scale, how hard is it to stay true to your original intention?

Dario Amodei: I’d say I’ve spent a lot of time thinking about why this question is so difficult.

As the company keeps expanding, I remain vigilant at every stage. At each stage of growth, new challenges appear. The company might lose its distinctive qualities in some new way—whether it’s business ambition or its core values.

I worry about both, because I think they are synergistic.

In fact, I believe the reason we can build such great models is precisely what allows us to carry out our values effectively. As the company grows and becomes stronger, there are many traps here. There are many ways things can go wrong—not because my values, or the cofounders’ values, or the company leadership’s values have changed, but because the company’s composition changes very rapidly.

So for about half the time, I’m talking to the company about Anthropic’s culture—and how that culture works. When you grow so fast, you hire a large number of people from large tech companies.

If you don’t tell them how Anthropic works, they’ll simply repeat the only way they know—how their previous companies worked.

So this is a continuous struggle, a continuous challenge.

It’s like, Daniela and I’s perhaps most important priority is figuring out how to keep this going. Because we realize that, in the long run, that’s what matters most.

Research efficiency and scientific progress

Emily Zhang: Your product iteration speed is astonishing. The number of products you release and how fast you release them is unbelievable. How do you do it?

Dario Amodei: I’d say two things. First, we are a unified company. We have a unified corporate culture. I think that as we scale up, we still maintain very high efficiency. Everyone still has aligned goals, reflecting the unity of culture and organization. I believe this is the most important factor.

As for the second biggest factor, I’d say it’s Claude itself. We’re using Claude to help us now—using it to help develop models, improve model efficiency, and quickly develop products. To do that, you need to develop all kinds of new practices. Even though we’re still new at this, it has already brought significant acceleration, and that acceleration is becoming more reliable. Those are the two factors I would highlight.

Emily Zhang: Can you tell me the craziest things AI has done that you’ve seen?

Dario Amodei: I think the craziest things I’ve seen mainly happen in biology and medicine. I’ve seen quite a few cases, and actually including Daniela’s case. Claude diagnosed some medical issues that many top doctors missed. In biology, these models are starting to perform astonishingly well on tasks like drug design and computational chemistry. As someone who used to be a biologist, watching these results makes me marvel—these things are truly hard.

You need to know that completing these tasks requires massive specialized training. And Claude is getting better and better at this. This is an area where I think we will benefit enormously. This is the positive side—we will gain these extremely huge benefits. Life will get better. The quality of the human experience will get better.

Emily Zhang: A century of scientific progress.

Dario Amodei: A century of scientific progress, and a century of progress at the level of the human experience. For example, going back to 1900. Think about all the problems we faced then, all the reasons people died too early, all the suffering people had to endure, and all the material shortages that we no longer have today.

Then imagine the next century again. I truly believe that this century’s scientific and medical progress—if we can overcome the difficulties we face right now—I believe we can—will be extraordinary. I’m becoming increasingly optimistic. We will have a world far better than the one we have now.

Writing, thinking, and AI assistance

Emily Zhang: I know how much you love writing. You’re known for writing articles. Will you use Claude to help you write?

Dario Amodei: I will. I’m not at the point where I directly adopt text written by Claude, because I have my own unique style and I’m quite picky about that. But I basically use Claude to assist with brainstorming—to help organize topics—or to provide some material I can refer to.

So it plays a supporting role. I don’t know how far away we are from Claude being able to write works even better than mine. We haven’t reached that stage yet. But I believe that day will come.

Emily Zhang: I love writing too. I feel like writing helps you clarify your thoughts. It involves a lot of critical thinking. If we let Claude handle it for us, will we lose that ability?

Dario Amodei: I’m somewhat worried about that. In fact, that’s partly why I insist on writing myself. That’s true for external audiences. Even though many people read my articles, writing is also for clarifying my own thinking—so I know what to do next—and to establish a shared reference point between me and other people.

I think we’re still exploring how to use AI in a way that preserves these benefits. I think what I’m doing now is a good example—using Claude for research, and using Claude to help organize my own thoughts.

I think if we simply use it end-to-end—like letting it write an article about AI risks—first, it can’t produce my personal insights. Second, I would indeed lose the benefits that come from writing. As model performance improves, I think in the future there will be some way to use them more directly in the writing process while still retaining those benefits. But that will be a subtle process. It won’t be a single fixed pattern. We’ll need to gradually explore it over the coming period of time.

I think we might see a very counterintuitive combination: GDP grows quickly alongside high unemployment—or at least underemployment, a lot of low-paid positions, and severe inequality.

Unemployment risk, productivity, and macro policies

Emily Zhang: You’ve always been very blunt about unemployment issues. AI might eliminate half of entry-level white-collar jobs within the next 1 to 5 years. That was last year’s claim. AI is developing at an astonishing speed. Is it still 50% now? Or is the number higher?

Dario Amodei: I’ve always said that if you look back at those original video clips, they’re always being cut and clipped into those three seconds out of context. But my actual statement has always been: I don’t know what will happen. However, it implies that the kind of turbulence the situation could undergo will reach an order of magnitude.

And I’ve also been talking about what countermeasures we can take. I mentioned token taxes, and working with enterprises to adjust staffing. I’m a bit skeptical of retraining programs, but we really should incorporate them into macroeconomic policy considerations.

From the very beginning, I’ve been talking about solutions. But it seems that in human psychology there’s a tendency to clip out those three seconds about doomsday. So I want to convey that message is absolutely not that doomsday is coming. What I want to say is: this is something we should foresee and worry about, and we really need to respond in a proactive, positive way. I’m not entirely sure, but I’m still quite concerned. My level of concern remains about the same.

We’re seeing AI improve people’s productivity now, but this is still in the typical early growth phase. If you look back at the Industrial Revolution period—I wrote about this in Adolescence of Technology. You automate 90% of the work, which is great.

The productivity of people in the remaining 10% increases by 10x, because they get a 10x leverage. But in the end, it will approach 100%. Then the connected question is: you have to find other things for them to do.

For long-term development, I’m still not sure. I really do feel uncertain about that. But I do think there are different types of adaptation. For instance, one point I’ve discussed is the software engineers inside Anthropic. We’re currently experiencing this shift: even though AI writes all or almost all code, AI does make software engineers more efficient.

But even so, it still improves efficiency. However we’re already starting to see signs that for some people, it might not improve their efficiency at all. In those cases, simply having AI do the work directly might work better.

That’s one aspect of the problem. On the other hand, we need to figure out in which areas we should increase demand. We call them frontier deployment engineers, or something like application AI solution architects—roles that combine technical development with client communication. Right now, demand for this kind of talent is very high, because the customer base is huge and we’re in a phase of rapid expansion.

So does this fit every person doing pure software engineering? It’s not a perfect match. It’s not a one-to-one relationship. This makes you realize there will be huge disruption in the future, but things will also adjust accordingly. Which side will win? I don’t know.

But the reason we need to issue a warning is because this is how we respond. It’s also the basis for policymaking—both at the level of human beings themselves and at the global macroeconomic level. We want to put forward well-considered perspectives. We don’t want to make promises that people think can’t truly be achieved or delivered. We don’t want to make claims that haven’t been sufficiently argued. We want to think carefully about these issues and explore genuinely feasible solutions.

Emily Zhang: You showed a chart illustrating potential job losses, for example in sales and finance. Which jobs will disappear, who will be replaced, and what new jobs will be created?

Dario Amodei: Actually, no one can be certain, because the economy is unpredictable. This is just like the stock market. It’s one of those decentralized processes—you can’t truly know in advance which parts of people’s jobs they’ll still be able to do in the future.

But I would say broadly that any field where entry-level white-collar jobs exist—whether it’s banking, finance, or other industries—AI first has huge potential to improve people’s work efficiency, but eventually AI will be able to comprehensively replace human labor to complete entire tasks.

Then we need to think about what people will do. I believe we need to plan this in advance. When we talk with enterprise customers, we’re already doing this. We see the choices they face. Their choices are: should I cut costs?

Usually that means reducing headcount. It basically means doing the same thing with fewer resources. Or should we do more with the same amount of resources? Whenever possible, we try to push them to do more with the same resources, because that essentially means employing the same number of people—or even more—and having them do some new kinds of work, thereby driving a positive-sum outcome.

The advantage we’re currently in is that the “pie” will expand significantly. Because the pie is expanding, it’s likely that new fields will open up where people can enter. The only question is whether they can find those opportunities fast enough. The scale of this disruption will be very huge, and that’s what I’m warning people about. But we still have to solve the matching problem.

Emily Zhang: Please walk me through it—for example, when you wake up five years from now, what will this country look like? What will these people be doing? Because if unemployment reaches that level, isn’t it a sign that the revolution is about to begin?

Dario Amodei: That’s exactly the result we want to prevent. We absolutely have to avoid this happening.

I think there are several entry points. None of them are guaranteed. We can’t be sure. But after all, there’s the physical world. Things that exist in the physical world—like the robot revolution—are also happening, but they’re happening much more slowly than what’s happening in the AI space. People always talk about building data centers. But when handling any type of information becomes much easier, then the limiting factor might become things in the physical world.

Therefore, we need more people to manufacture, build, and produce in the physical world.

Any human-centered things, I think, will be a big deal. I’ve heard many stories about AI finding problems that doctors didn’t find, and that makes me happy. But people do still want to interact with other humans—especially when dealing with important matters. Maybe AI can provide better customer service, but even so, people—or at least some people—still want human interaction.

So I think these people-interaction-driven jobs will become very important, and I think humans will make some efforts to guide AI. To a certain extent, it has to align with someone’s values and intentions. So I think there will be a role for this kind of thing, though I don’t know whether it will be shallow or deep. I think it’s hard to say.

Countering “doomsday marketing” accusations

Emily Zhang: Right now there are already many voices opposing this. I know you said you’re trying to remind people, but Jensen Huang said you’re confusing tasks with jobs. Others also said this is more like a “doomsday marketing” campaign that benefits Anthropic. So I want to state clearly and firmly refute this.

Dario Amodei: Regarding the overall unemployment risk, I have some thoughts. I mean, we haven’t fully filled out these thoughts yet, because I want to make sure they’re accurate. But Anthropic has already proposed many solutions. We set up economic subsidies. We have economic indicators. What I’ve been talking about is possible ways to respond to these risks through taxation and macroeconomic policies, including what new jobs might look like.

In the early stage of Technology, I explained that I had about five pages dedicated specifically to the distinction between tasks and jobs—why this time is different—and listed 6 things we can do, from private charity to government action. I talked about the problems, and I talked about solutions as well.

But social media—I despise it; as a category, I despise it. People only clip a three-second clip from a year ago, and they haven’t actually read the article. Or they exploit social media’s features—yet I have been very rigorous in my articles about exploring these risks.

The idea that this is just cheap marketing is itself cheap marketing. It’s lazy. It’s a failure to engage with serious knowledge work. I think this is part of the problem as well. Again, I think this is also a Silicon Valley ailment. It’s trapped in that three-second social media world.

So people just react to those contents, or they think they only need to react to those contents. Once again, I think this is very dangerous—and we haven’t had a mature conversation.

Instead, people are just lazily looking at those three-second clips. Then they say, “Oh, that’s what Dario said back then. That’s so stupid. That’s so unserious.” Every time someone says that, my view of that person gets diminished.

National security and the red lines for military AI

Narrator: One of the world’s leading AI companies is deeply embedded in multiple aspects of U.S. national security, including military operations. Anthropic’s standoff with the Pentagon over AI military safety measures is intensifying.

Emily Zhang: You’ve long taken an anti-war stance, dating back to your time at Caltech. However, you’re one of the first AI companies to sign contracts with the Department of Defense to operate on U.S. classified networks—for warfare. Explain.

Dario Amodei: Okay. I want to say this: you see, the world is constantly changing. Regarding my view of this technology, I’m concerned because we’re facing a resurgent authoritarian bloc that is highly aggressive, and we need to defend ourselves.

That’s been a belief of mine for a long time—and I still hold it. That’s why, across two administrations, even if I didn’t agree with every policy, this is why we generally supported this stance.

So this is why we cooperate with them. We’re not doing this for money. It’s extremely complicated. Deploying on government networks costs a lot and doesn’t bring much reward, yet it requires huge effort. So we do it because we have a sense of mission. But at the same time, since we’re doing this out of concern, the use of this technology must be limited.

[When discussing technological development stages, I’ve said: We should use this technology in all aspects, except those that would undermine our own values. And the red lines we draw—large-scale surveillance and fully autonomous weapons—are precisely what I believe would destroy our values.

If these methods are used, then even if you win, it would be meaningless. That’s the balance point I see, and that’s the position we hold. This also explains why we are one of the first companies to work with the Department of War, while at the same time, when others are willing to engage in certain actions, we do some things and refuse others.

I think you need to take a stance and stick with it. I can’t understand that kind of company position that wavers—starting by claiming that they won’t work with the government, and then suddenly taking on all of the government’s needs. You should choose your principles and adhere to them.

Emily Zhang: Since 2024, you’ve been working with Palantir.

Dario Amodei: Yes.

Emily Zhang: Their technology is used by ICE and by the police departments in Gaza. Has Claude been used for surveillance in other ways?

Dario Amodei: We haven’t cooperated with ICE through Palantir, or through any other channel. We don’t work with CBP. I don’t think we have any business in Gaza. We’re extremely cautious and limit the scope of our cooperation strictly to things we agree with.

Emily Zhang: So you drew your red lines. The president forbade you to cooperate with the federal government, the Pentagon marked you as a supply-chain risk, and OpenAI took the opportunity to sign contracts you refused. What does “winning” this fight actually look like?

Dario Amodei: I don’t think a private company can win this fight. It’s not really a fight. Anthropic isn’t trying to win, and it’s not thinking about who wins or loses. It’s more a debate about how the government should correctly use AI. AI is an emerging technology. We still don’t know where it’s reliable and where it’s not reliable. We still don’t know which ways it can help promote our values, or which ways it would weaken our values.

So one important thing, I think, is to establish precedents for the use cases we think are good—frankly, most use cases are good—and the use cases we’re concerned about. As I said, the role contracts can play is ultimately limited. As we’ve seen, others might sign a contract that doesn’t respect your bottom lines in the same way.

But what it does is it raises awareness of the issue. Right now, Congress is doing serious bipartisan work—trying to ban some of the things we’re concerned about and trying to set up safeguards. Let me reiterate: I don’t want to see this as a fight, but to some extent, it has been effective in pushing our country to think more carefully about how to use this technology properly.

Values and military ethics

Emily Zhang: The operator of Anthropic is an ideological madman who shouldn’t have sole decision-making power. What I care about is the bigger AI question. Do you mind being called an ideological madman—or a bunch of left-wing madmen?

Dario Amodei: I’ve been called much worse than that. People can call me or Anthropic whatever they want. What truly matters are two things: as a company, we’ve succeeded—and we’ve held to our values. In fact, in some ways, my life has been very easy. Because when you only pursue those two things, it’s really simple, isn’t it? You always know where you stand.

Emily Zhang: A U.S. official once said that with the help of LLMs, the U.S. military increased its number of daily strike targets from 1,000 to 5,000. That means Claude can help kill more people faster. Do you feel reassured about that?

Dario Amodei: I think there are two issues here. First, the ability of the United States to improve its military efficiency. I support that ability. I don’t think having a stronger capability triggers war; it can instead act as deterrence. Basically you’re asking: do you trust this country, and do you want this country to become a stronger actor rather than a weaker one on the world stage? I hope so. I’m a patriot.

There’s another separate question: whether some policies involving the U.S. government are things I might support or not support.

The government’s involvement—obviously, I support some of it and don’t support other parts. That isn’t up to me. If we provide a technology, the DOD also raises that point, and we actually agree with their perspective. If we provide a technology, we don’t have the right to decide which military actions you can take—or which military actions you can’t take.

Now, personally, I might think one military action makes sense while another is a terrible idea. But we won’t deny the value of the technology’s role.

You have to leave policymaking authority to military decision-makers. What we can do is draw some high-level boundaries—for us, that means preventing use cases that don’t align with our values or with our country’s values, and pushing use cases that we think help advance our values. That’s how we think about it.

Emily Zhang: Bloomberg reported that during the war with Iran, the U.S. military used Claude through Palantir’s platform Maven Smart System for AI-assisted target localization. According to reports, in February, a U.S. missile hit a girls’ school in Iran, killing more than 150 people, most of whom were children. Did Claude play a role in that attack?

Dario Amodei: We, look—we can’t access it. We don’t know exactly how those models were used. Obviously, these kinds of mistakes in war are truly, extremely terrible. That’s indeed a very terrible thing.

If that still isn’t enough to show why we insist on opposing those use cases we don’t support, we’re even willing to risk harming our company’s future to limit how these models can be used. And the use case you’re talking about doesn’t even violate our red lines.

We’re worried that in the future there will be use cases that are 100x more numerous, and some of those truly violate our red lines. Now, again, the point is that overall I think these models’ use is appropriate. I think the net effect is good.

But even in the best of times, military decision-makers commit extremely serious mistakes. I don’t know whether we’re in the best of times right now.

For example, we can discuss a few things. We can discuss how to set red lines to prevent more likely model use patterns that would lead to these issues. If we had allowed everything without restraint—if we, like almost all other companies here, gave in to fully automated weapons—then, as you can see, Claude would provide assistance, but with humans making the final decisions.

So it’s humans who make the final decision, not Claude. Imagine if in a world like that—not Claude, because we don’t allow it to do that, but the AI models of other people—decisions are made directly by AI models and humans have never been involved. That’s what we insist on defending.

That’s exactly what we’re resisting. I want to say one more thing here. Again, I don’t think procurement is the right way to achieve this goal. But we do need to make sure—this isn’t just what I care about as a technology supplier, but what U.S. people care about—that U.S. military decision-makers don’t make these mistakes.

Ensure that their systems run reliably and that when they take action, they make wise choices.

Likewise, as a citizen and as a technology supplier, this is what I care about. For example, the government uses Microsoft Excel when using technology. If you say you can use Excel for one military action but not another in practice, that’s not doable. Hopefully this helps you understand how we think about the issue.

Emily Zhang: That school has a website. You could have found it via Google Search. Shouldn’t Claude have found it? Or should the AI—or any technology they’re using—have found it? Doesn’t this raise a more worrying problem: using technology as a shortcut in war?

Dario Amodei: Look, look—what I want to say is that I’m not sure. I truly don’t know. It depends on classified knowledge that I might not have access to. But we have established principles. And I think the principle that’s being followed here is that humans make the final decisions. I don’t know what role Claude or any other AI played. But if that still can’t explain why this principle is so important, I don’t know what else can.

Emily Zhang: Is AI warfare more likely to prevent World War III, or more likely to cause war?

Dario Amodei: Overall, I believe it’s more likely to prevent war. But if we put no limits on how it’s used, then I believe it’s more likely to lead to war breaking out. You’ve seen Dr. Strangelove—the premise is that you have some kind of doomsday device, and once it thinks it’s under nuclear attack, it automatically fires nuclear weapons to retaliate.

What could go wrong? And again, I end up thinking about topics like lethality and fully autonomous weapons. I think conflicts happen because both sides restrain and maneuver against each other. They create misunderstandings. If we lack proper oversight of this technology, then I think the likelihood of such accidents increases.

Now, I think if AI can be used appropriately—maybe not even limited to war, but consider the intelligence-gathering domain—imagine this: when we have a grasp on the enemy’s every move, they’ll think twice before conducting some kind of infiltration or military action. Therefore, I believe superior intelligence capabilities can indeed prevent conflicts. Superior responsiveness can deter conflict. I’ve always believed in these things.

Mythos: The safety game of a frontier model

Emily Zhang: Anthropics almost every week makes headlines. Of course, what’s getting the most attention right now is Mythos.

Narrator: This is currently the latest and most powerful Anthropic model. It can autonomously run through every stage of the cyber kill chain.

Emily Zhang: You’ve said that Mythos’s capabilities are too powerful and shouldn’t be released to the public. What surprises you the most about it?

Dario Amodei: I think what surprises me the most is, the models’ ability to find vulnerabilities has been rising continuously, and more importantly, they can turn those vulnerabilities into exploit programs—yet people usually only talk about the vulnerabilities themselves. People often don’t discuss the process of turning vulnerabilities into exploit programs, and it performs quite well in that area.

So what surprises me is that we’ve witnessed this huge leap. It’s a particularly big leap. And without any guidance, some of the companies we originally provided the technology to said that this is basically a super-weapon. It should require a gun license to use it.

Please don’t release this. The push to restrict usage came from those very companies that we provided the technology to. They found so many critical vulnerabilities and so much exploitability around those critical vulnerabilities that basically they were asking us not to release it.

It needs to be made clear that because information on social media is always distorted, our goal isn’t to keep it hidden forever. We’re trying to open up the technology gradually, to an increasingly broad group of people. Ultimately, we believe Mythos should be released to the public—but it must be equipped with strong cybersecurity safeguards.

One concern right now is that when we released Opus 4.7, we also released the cybersecurity safeguards in place at the moment. While this is an excellent cybersecurity model, its capability is much weaker. And these safeguards can be jailbroken. We’re somewhat concerned about other companies that think this is a sufficient defensive measure. Because we all know that these classifiers can be jailbroken or bypassed.

Based on our own tests, frankly, based on our assessment of other companies’ defensive measures, these defenses are not strong enough at the moment. That’s exactly what we’re waiting for: to bring defensive capabilities to a level that we truly have confidence in.

Emily Zhang: There were a lot of resistance voices at the time. Some researchers claimed they could reproduce this process using cheaper open-source models. Some people said that OpenAI already has these capabilities. So how do you respond to those who think this is just a big public relations stunt?

Dario Amodei: The claim that it can be reproduced using open-source models is completely wrong. The core idea of the system is that Mythos can view the entire codebase and discover problems. Someone on Twitter said that if you make an open-source model specifically point at the line of code that Mythos discovered, it will also discover the same issue. That’s not it. That’s not it. That’s not the prompt; it’s not that issue. For example, it’s not that, it’s not the same root.

The ultimate test was that we approached companies, looked at their open-source codebases, and we found 271 new vulnerabilities in Firefox. We found thousands of vulnerabilities in those private companies whose vulnerabilities weren’t yet fixed, or weren’t yet ready to disclose.

It’s like—before this, nobody found those 271 vulnerabilities with prior models. So compared with that kind of “I found the exact code line Mythos found, so I know I’m fishing for a needle in a haystack and other stuff can now pick up that needle” situation, in practice the workflow that actually works is different.

But for those who say this is just a marketing tactic, I want to say, because we didn’t publish this model, we suffered a huge loss commercially. This model greatly accelerates research within Anthropic and the production of subsequent models. If we released it, it would accelerate the external world too, giving the same acceleration effects outside. Commercially, that would cause us massive harm.

Emily Zhang: If it helps defenders, it will help attackers as well. Can we still defend anything?

Dario Amodei: I want to say this: the reason we provide Mythos first to defenders rather than to attackers is to patch all the vulnerabilities. I don’t know— as models get better, maybe more and more vulnerabilities will be discovered. But the number of vulnerabilities is ultimately finite. They’re finite. Like you have a surface with only so many holes.

Once you patch all the vulnerabilities, the surface becomes very difficult to attack. At the same time, the code itself is also written by powerful models—so it becomes very hard to discover defects in it or to break into it. So from another angle, I think we’re working toward having, in 6 months or 1 year, a safer internet ecosystem than ever before. We’re trying to reach that goal, and we’re doing our best to open up Mythos to new cyber defense users.

We’ve been communicating with the government. We very much respect their advice. Because of concerns about counterintelligence risks, they are slowing down our pace of opening it up. Risk. I think that’s wise. I think everyone serious here understands that there really are trade-offs.

We see people on Twitter and other AI companies criticizing this harshly. Look at their statements, and then look at their actions. They’re not serious people. They’re not taking the serious trade-offs we’re facing here seriously.

You see, every day customers call me and say they want to use mythos. And there are also nations calling and saying they want to use mythos. Then the U.S. government and my security team tell me, no, wait—there are risks here. I’m not saying one side is right and the other is wrong.

I think the truth lies between them. Both sides have reasonable viewpoints. Both sides have reasonable viewpoints. But there really are challenges here that we need to face together as a society, rather than simply dismissing things as cheap marketing. And we also shouldn’t, like some other companies, use the chief marketing officer to attempt some kind of reverse positioning.

All of this demonstrates a staggering lack of seriousness and maturity. We need to face this moment together.

Checks on power and government regulation

Emily Zhang: Have you already had to make some trade-offs that make you feel less comfortable?

Dario Amodei: Anthropic’s entire history has been a process of trade-offs. The entire history of Anthropic—in an ideal world, before you release your first chatbot, you might be more inclined to spend years researching every possible way it could go wrong. And now, we have indeed delayed.

We did delay Claude’s first release, but we delayed only a few months. So I’m saying it’s a trade-off in every sense. At the extreme ends of the spectrum is total insanity—so everything is a trade-off.

What I want to say is, since we’re now in the position of business leadership that I described, in reality, Daniela and I are doing our best to further drive progress in acting prudently. That’s what the release of mythos is meant to do. If you’re not the industry leader, it’s hard to do that. So I think you’ll see more of this kind of thing happening.

Emily Zhang: There’s a view that says, why doesn’t the government take you over? Why let a private company control such powerful technology?

Dario Amodei: So I do think this is a very serious question, and I agree with these concerns. I don’t think the government should directly take over us. But I can put it this way. Looking back and describing the current situation, every prior powerful technology we’ve seen in history was either built by the government or originated with the government.

For example, nuclear weapons—obviously, it was initially built by the government, and then it was basically led and built by the government after that. But even technologies like the internet, GPS, and mobile phones—every research and development effort was done in labs, in federal labs, and in universities.

AI is the first technology that was built in the private sector, with no substantial role played by the government, and the government entered relatively late. I think this is actually a dangerous and unstable situation. This is not a scenario I would choose. There’s no alternative to it—this technology itself can be built. Our competitors are building it, and that has economic value—because eventually it will be built anyway. The issue is that the government isn’t participating, not that the private sector is.

I think we need to consider checks and balances on power, so we need checks and balances on the power of AI companies. We have a mechanism called Long-term Benefit Trust. In essence, it’s an organization that can appoint and dismiss most board members. If you follow the l