Editor's Note: Anthropic CEO Dario Amodei finds himself in a very awkward position. On one hand, he holds the world's leading AI model; on the other, a U.S. government order accidentally triggered a global takedown, leaving even non-American team members unable to use it.

How this will end remains unknown. It is said that Amodei is still making efforts, and we can all continue to follow the story. However, we can get a glimpse into the thinking of this highly controversial head of the premier AI coding company from this latest interview with Emily.

In today's Silicon Valley power landscape, Anthropic occupies a uniquely central and tension-filled position. As the strongest challenger to OpenAI, it was founded by a group of top researchers who "walked away" due to value disagreements.

When CEO Dario Amodei sits in the spotlight discussing AI's exponential growth, he exhibits a rare, surgeon-like calm. This is not just a competition about technology; it is a deep game concerning trust, safety, and how human civilization positions itself in the face of an intelligence explosion.

Full Text Summary

This latest interview delves into Anthropic CEO Dario Amodei's mindset in facing the exponential growth of AI, covering topics from the inside story of leaving OpenAI, the company's choice of business model, to AI's profound impact on the job market, cybersecurity, and geopolitics.

The CEO elaborated on how Anthropic aims to balance power through mechanisms like the "Long-term Benefit Trust," and how, while pursuing technological leadership, it practices its safety values by setting "red lines" and delaying the release of high-risk models (like Mythos).

Note: Amodei's remarks have consistently been unfriendly towards a certain Eastern superpower; readers should judge for themselves.

Core Arguments

· The AI industry is in a state of "smooth exponential growth," where the accumulation of quantitative changes eventually creates a qualitative feeling of explosion.

· Trust is the foundation of cooperation in the AI industry; Anthropic advocates that trustworthy participants should unite to establish industry standards.

· The enterprise business model is more synergistic with AI safety values, avoiding the addictive and low-quality content competition common in consumer markets.

· Regarding AI-induced unemployment risks, society needs to anticipate and formulate macroeconomic policies while seeking positive-sum games of "doing more with the same resources."

· Military applications must adhere to the principle of "human in the loop," strictly maintaining red lines against mass surveillance and fully autonomous weapons.

Below is the complete interview:

The Pressure and Experience of Exponential Growth

Emily Chang: How much sleep are you getting?

Dario Amodei: I've never been someone who sleeps particularly well. Let's just say, I'm learning the art of relaxing and falling asleep during extraordinary moments of pressure.

Emily Chang: Everything is moving so incredibly fast. What does it feel like to be in the middle of it?

Dario Amodei: It has an exponential feeling. It's like, imagine you're on a spaceship accelerating away from Earth at relativistic speeds. The way special relativity works is, when you wake up after sleeping, two days have passed on Earth. So you have to handle two days' worth of things in one day.

Then you go to sleep again, and because you keep accelerating, three days have passed on Earth. Then the next day, four days. The feeling is something like that.

Emily Chang: Do you often feel anxious, worried about what you'll wake up to?

Dario Amodei: We have plenty of clear, urgent problems to deal with. I'm handling those while also thinking about how we prepare. But I think it's not helpful to be paranoid or worry about what I'll wake up to. I've studied people in history who dealt with such high-pressure situations. You need to learn to deal rationally, not to equate the magnitude of different dangers.

This state of bouncing between 'I'm not worried' and 'Oh my god, we have to panic today' is, I think, a sign of immature decision-making. Truly mature decision-making is that you can't ignore this, nor can we be complacent.

In fact, it's becoming riskier, but we must deal with it rationally, like a surgeon handles an operation. Or, like an officer handles a military operation, or anyone making decisions affecting many people must make them rationally. And they must understand the risks, but they also must maintain basic calm.

So my son asked yesterday if he could use my Claude account, I said absolutely not, I need my tokens. We're also seeing more of these applications in the consumer market, we originally hoped to be more of an enterprise company, but even without putting much effort, the consumer business is starting to grow rapidly.

Emily Chang: You're now at the center of the AI universe. What's that like?

Dario Amodei: What's interesting is that throughout my career, especially the time since joining Anthropic, what I've experienced is a smooth exponential growth.

The experience of smooth exponential growth is: nothing happens, nothing happens, nothing happens, something small happens, then suddenly, it explodes crazily. That's the world's experience. That's also the company's scale experience relative to other companies and relative to the world.

So I stared at this chart for a long time, and I said, we're likely to become the AI company with the highest revenue and highest valuation around this time. And indeed, it happened. It happened. So in a sense, I'm not surprised, because the curve on the chart was very smooth.

But of course, from another perspective, when it actually happens, you see much more, richer detail and color. And, it is absolutely surprising.

We always keep in mind the questions we usually focus on, which are: How do we train good models? How do we apply them to good products? How do we ensure everything is safe? How do we help people while managing the societal risks around this technology? These are all the same questions, just examined under a bigger microscope.

Upbringing and the Silicon Valley Spirit

Emily Chang: What kind of kid were you growing up in San Francisco? I know your father was a leather craftsman, your mother worked in a library. How did that influence you?

Dario Amodei: The whole internet revolution was happening around me then, but I had no interest. I was only interested in studying math and things like writing and drawing. I was interested in exploring the universe. I was interested in science fiction. Overall, that was my environment then. I think I just had a strong curiosity about the world.

Emily Chang: You grew up in a place known as a tech hub. And now it's also the center of AI. This place, this city, did anything shape your worldview?

Dario Amodei: Yes. I think the general spirit of not following the herd, of individualism, and that 'it's okay to be a bit crazy' exists. I think a lot of that probably rubbed off on me.

You hear stories, like when you go to certain European countries, or even other parts of this country, that thinking differently is often suppressed or seen as weird, or having crazy ideas.

I actually have a lot of criticisms of Silicon Valley, but I think one good thing is that it encourages the idea: Even if all the experts are against you, that's okay. If you have a coherent vision and coherent worldview, you should pursue it, that's what matters. Maybe it won't work at all.

But if it works, it has a kind of long-tail effect, in some areas you might be able to dig deep and eventually find a huge gold mine there. I think that spirit is very important.

Emily Chang: You, Daniela, your sister, and her husband Holden Karnofsky all lived in the same group house in 2016. What were you debating then?

Dario Amodei: I think that was when the Open Philanthropy project had just started, and Holden was leading that project. I was a biological scientist at the time. So I was helping them with some things about health in developing countries or biological research. So I was giving some advice on those topics. Like, which areas looked promising? Which areas looked less promising?

The Core Disagreements Behind Leaving OpenAI

Emily Chang: Your decision to leave OpenAI has become Silicon Valley lore. What really happened? Setting aside the narratives, what were the core issues? Where did you disagree?

Dario Amodei: Listen, I'll say it, and I'll put it very simply. When you build powerful technology, there are many difficult problems, and Anthropic goes through them every day too. We don't know if the decisions we're making are right or wrong.

So, there are many legitimate disputes on safety issues. Of course we had disagreements with them, but know this: that alone isn't enough reason to leave. People here have disagreed with me. People here disagree with each other.

But when you feel you can't trust someone, when you feel their values aren't as stated, when you feel they aren't honest, when you feel they're not acting for the reasons they claim, when you see disturbing patterns of behavior or dishonesty, it makes it very hard to continue working at a company and to continue trusting it. Ultimately, when you don't share a vision and no longer trust someone, what's the point of arguing?

The solution is to go separate ways, you do your thing, they do theirs. I'm perfectly at peace with that: we do things our way, they do things their way. We'll see who prevails in the market, who prevails in the court of public opinion. I think those results speak for themselves, more than any dramatic speculation about who left and why.

We need to know that we are providing an example of how to deploy this technology, a way we consider responsible. If they disagree, they should make arguments. I think that's all that needs to be said on the matter.

Industry Cooperation and Rebuilding Trust

Emily Chang: At the AI Summit in India, there was a moment where you and Sam Altman seemed to refuse to hold hands on stage. What happened?

Dario Amodei: The situation was, that summit was incredibly disorganized. We were all brought on stage last minute, they changed our order of standing. Then they took a photo, then they commanded all of us to hold hands. If you've been to these kinds of summits – and I'm not saying anything bad about India – but these international summits where all heads of state are present are organized chaos.

Emily Chang: But the others held hands. Come on.

Dario Amodei: Listen, I don't know what to tell you, okay? Narendra Modi was up there, suddenly telling everyone to hold hands.

Emily Chang: Alright, alright.

Narrator: Look, Sam and Elon are suing each other. It seems you don't like Sam.

Emily Chang: If the developers of the world's most important technology can't hold hands on a stage, how can we trust you'll cooperate on existential risks?

Dario Amodei: Well, here's what I'd tell you. Among the people building this technology, there is a huge variance in quality and trustworthiness. I think this means, different people, thinking no one trusts each other, I don't think that's correct.

I know Demis Hassabis, who built the Gemini models, which are competitors to Claude. I've known them for 15 years. We've collaborated on many issues. We buy compute from Google. We exchange ideas on safety frequently.

So, my point is, first, there are participants more trustworthy than others. And, I think there are participants outside Anthropic that I trust, who I consider trustworthy. I think what needs to happen is, trustworthy participants need to band together to deal with untrustworthy participants. So that participants are forced to adopt the same standards.

Through hard experience, I've learned some people won't do the right thing voluntarily. But if most of the industry is doing the right thing, then I think others in the industry have no choice but to follow. It's like a positive version, you inspire others.

It's like Demis and I inspire each other. He's doing AlphaFold. We're also trying to do some things in biology, we're doing interpretability. They started interpretability research. It's not even competition.

It's just that every company is doing something cool. And other companies will think, that's cool. We want to try that too, see if we can make something new within it. That's the "carrot" side of top competition. Then there's the "stick" side, or implicit "stick," which is you realize these people are doing the right thing. If others don't, it looks bad.

We often see this behavior: they grudgingly do the right thing, but try to pretend they're doing something different, and imply we have some awful or evil intent, which is predictable. But I think, that's how we bring the industry together and promote cooperation.

Business Model: Enterprise Alignment with Values

Emily Chang: Earlier on, others focused on fun, flashy consumer applications. You bet on coding and enterprise, Claude Code was a huge success, Claude Co-work was a huge success. Why did you make that bet back then? Was it a values-based decision, or a business one?

Dario Amodei: When we founded Anthropic, the most fundamental thing, the thing 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 extremely expensive creation of these models, the company needs a business model. So, will the business model prevent the realization of values?

That question is always there. But one thing I learned from being at other companies and observing others is: if you choose a business model that fundamentally conflicts with your values, you'll have a hard time. You'll either betray your values or be left behind by the times.

You end up in a dilemma; there are ways to get around it, but it's still a very tricky situation. It's much better to choose a business model compatible with your values.

So when we thought about it, we thought, look, we've seen the world of social media and the consumer space, it does seem to encourage engagement, even addiction. The kind of low-quality content we see from AI video models is driven by the logic of maximizing the attention time you put in, because the incentive behind it is ad revenue.

Whereas if we look at the enterprise space, look, our original hope was that these models would be helpful to people. If I think of all the positive things you can do with AI, I often remind people of the negatives, but fundamentally, we think the positive side will outweigh the negative. Many of those positive applications basically fall into the category of enterprise applications.

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

We want to use AI to help education. Much of that is enterprise application. We want to use AI to address health and development issues in developing countries. While they are non-profits, they are essentially enterprise in nature. We want to promote economic growth. That's also basically at the enterprise level.

Plus, I think there's another factor: enterprises value trust and long-term relationships highly. Consumer applications sometimes have a gimmicky feel, while at the enterprise level, it's about building a partnership, where you work with a company for years, you deliver, they deliver, and they fundamentally trust you. Therefore, that aligns very synergistically with our goal of deploying these models in a positive and safe way. So, I think having a business model so aligned with our values is advantageous for us.

Not that conflicts never exist, not that we don't have to make tough choices, but I think the number of such choices is much smaller than otherwise.

Competitive Moats and the "SaaSpocalypse"

Emily Chang: A developer can switch from Claude to GPT or Gemini in an afternoon. In this industry, is it really possible to maintain a long-term lead? And, how long do you think a serious competitor would need to replicate what you've built?

Dario Amodei: Model quality is the most important thing. For example, we are currently far ahead in model quality. There is some inertia, but I never rely on that, I never have relied on, or Anthropic has never relied on the idea that 'product stickiness is high, users won't switch.'

I think you still want to have a better model. You want a better product. And, we see growth rates haven't turned at all, if anything, they've gone up, at least at the time of this interview recording. So, I think that's the most important thing.

Emily Chang: Shortly after the release of Claude Co-Work, $285 billion in market value evaporated overnight, traders calling it the SaaSpocalypse. If AI continues improving at this rate, how much traditional software will be replaced, and how fast?

Dario Amodei: So, this is one of those things that's hard to predict in advance. If you could predict it perfectly, people would have done it, they'd make huge fortunes in the market and be right forever.

So, no one knows for sure, but I'll point out a few things: all these traditional software companies have some moats. I think what will happen is, some moats will disappear, but others will remain. The ability to write software quickly, I absolutely think that will disappear. If your moat is 'we wrote this complex software no one else could,' well, good luck. You won't be able to defend that.

But I think people have customer relationships. People have expertise in how the domain works. People have unique domain knowledge. So my advice to all these people is clearly, 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 the limiting factors. And there might also be new moats appearing.

I think those who can navigate nimbly, who can rely on moats that remain and leverage new ones, will do well. I think those who are complacent, who just fool themselves into thinking what worked in the past will always work, they will not have a good time. So, that's the advice I'd give.

And, I think ultimately, I guess – of course depending on how you define SaaS and what you consider not SaaS – but I guess the software industry will become larger, not smaller, even though there will be some huge losers.

Emily Chang: Please explain.

Dario Amodei: I just think the pie is getting much bigger. For example, I think with AI, the pie is getting bigger. The existing incumbents might become smaller in relative terms. Some of their value might decline. Some might even go out of business if they can't adapt correctly.

But I think when growth is very fast, you often see this: if the possibilities of AI grow 10x, it's very easy for existing legacy industries to grow 1.5x, just not as much as the overall pie grows. So I think that might happen.

It's not to say we won't see some huge losers. I think those who don't adapt, who stick their heads in the sand, who can't see where the future is going, and who can't identify their moats will be in a very tough position.

Compute, Funding, and Partners

Emily Chang: Your biggest backers are companies like Amazon, Google, Microsoft, and NVIDIA. These companies all have their own agendas. They are both partners and competitors. You have huge commercial milestones tied to funding. Who is really in charge?

Dario Amodei: We have been very outspoken on multiple occasions about what we really think. I've been very vocal in advocating for controls on chip exports to China. I said that because I think it would be very bad for the U.S. and the global process if China leads in AI capability. And, some chipmakers obviously don't share that view. But that didn't stop me from saying it.

Even after we signed more cooperation agreements, I'm reiterating that now. What they understand is, we always work with them. We've been good partners. We're workable. I'm sure they wish we wouldn't say these things, but these are things I firmly believe.

What are you going to do? They'll come around eventually. They get as much out of these deals as we do. Look, we're all adults. We can cooperate on something while disagreeing on another.

Emily Chang: Bloomberg reported your valuation is even higher than OpenAI's. We're talking about a five-year-old startup valued close to a trillion dollars. How do you make sense of that? About that number, and since you are more disciplined on compute and have a faster path to profitability, why do you need so much money?

Dario Amodei: Compute scales up very fast, so it's true the business fundamentals look good, but your compute might be 3x or 4x larger a year from now – I won't give exact numbers, but that kind of compute growth is very rapid.

We have every reason to believe revenue growth will meet and exceed those scales. And raising money is precisely a buffer for that range of uncertainty.

So, it's a completely rational move. The equity dilution to the business is tiny, logically, these two things are completely different. In fact, it's compatible with the opposite, it doesn't mean there's anything wrong with the business fundamentals.

Emily Chang: There have been reports of overloaded servers, reliability issues, complaints about tokens running out. You've said other companies are going all-in on infrastructure. Do you really have what you need, or are you playing catch-up?

Dario Amodei: Regarding compute, one thing is there's what's called marketing compute. So, my view is, over some period, longer than months, we can get a lot of compute. It's worth mentioning here, I don't think by any reasonable standard we bought too little compute.

So, we planned for compute to grow 10x per year.

10x per year is what we expected. But that's not what we're seeing now. In Q1 2026, our quarterly revenue grew over 3x, that's just a quarter, not annualized; 3x growth within a quarter, of course, 3 to the 4th means 80x annual growth.

We weren't anticipating 80x annualized growth. Planning for 80x annual growth wouldn't be rational, because if you end up with 10x, you're off by 8x. So we are in a locally extreme, explosive compute phase. This won't continue.

If this continues, by year-end you'd have a revenue number no company on Earth could match. I don't think that will happen. It just, it can't last.

But you might go through these brief periods where you go, wow, this is growing way faster than we ever thought possible. But I don't know, you've seen the compute deal with Google, the compute deal with Amazon. There's more we can and will do.

Like, the market is liquid. If you can use compute very efficiently, and there's demand, you can get the compute you need. It might just take a month or two.

Leading the Race to the Top and Company Culture

Emily Chang: How does it feel to outpace your archrival?

Dario Amodei: Listen, we have many hard challenges ahead. We have this 'race to the top' philosophy, trying to pull other companies along with us. I think we've seen that we do pull them. Sometimes they don't admit they're doing it.

Sometimes they attack us while copying us, but that pulling effect is very valuable.

So I think, both commercially and from a model standpoint, the value of being the industry-leading company is not about beating rivals for the sake of competition. It's about the ability to lead the entire ecosystem. We hope to do more of that in the future.

Emily Chang: But let's be honest, winning does feel good.

Dario Amodei: Look, we always aim to succeed, right? Like we're always trying, we're not trying to fail here, I'm not someone who thinks we should stop this technology, shouldn't build it. We exist within a free enterprise system, and that's fine. We just have to mitigate the risks from the models, so it's always a balance.

Emily Chang: So, for much of Anthropic's history, you've been the underdog. I imagine it's easier to take the moral high ground when you have nothing. How hard is it to stay true to your roots at this scale?

Dario Amodei: I'd say, I've spent a lot of time thinking about why this is so.

I've been vigilant at every scale stage as the company grows. At every stage of the company's growth, new challenges appear. The company can lose its character in some new way. Either commercial aggressiveness, or its core values.

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

In fact, I think the very thing that allows us to build such great models is key to being able to practice our values effectively. As the company grows larger, there are many pitfalls here. Many ways things can go wrong, not because my values, or the co-founders', or the company leadership's values have changed, but because the company's composition changes very fast.

So I probably spend half my time talking to the company about Anthropic's culture and how it works. When you grow this fast, you hire a lot of people from big tech companies.

If you don't tell them how Anthropic works, they'll simply repeat the only thing they know, which is how their previous company worked.

So, it's a constant struggle, a constant challenge.

It's like, maybe the most important top priority for me and Daniela is figuring out how to preserve this. Because we realize, in the long run, this is our core.

R&D Efficiency and Scientific Progress

Emily Chang: Your product iteration speed is astonishing. The number and speed of your releases is incredible. How do you do it?

Dario Amodei: I'd say two things. First, we are a unified company. We have a unified company culture. I think we've maintained very high efficiency while scaling. Everyone is still aligned, which shows cultural and organizational unity. I think that's the most important factor.

As for the second biggest factor, I'd say it's Claude itself. We are now using Claude to help. Using it to help us develop models, improve model efficiency, and develop products quickly. For that you need to develop all sorts of new practices. We're still novices at this, but it's already providing significant acceleration, and that acceleration is becoming more reliable. Those are the two factors I'd point out.

Emily Chang: Can you tell me the craziest thing you've seen AI do?

Dario Amodei: I think the craziest things I've seen have been mainly in biology and medicine. I've seen many cases, actually including Daniela's case, where Claude diagnosed medical issues that many top doctors missed. In biology, these models are starting to do amazing things on tasks like drug design, computational chemistry, etc. As a former biologist, I look at these and think, this is really hard.

You know, doing these tasks requires extensive specialized training. And Claude is getting very good at this. This is an area where I think we'll get enormous benefits. This is the positive side, we'll get these incredibly large benefits. Life will get better. The quality of human experience will get better.

Emily Chang: A century of scientific progress.

Dario Amodei: A century of scientific progress, and a century of progress on the human experience level. Like, go back to 1900. Think of all the problems we faced in 1900, all the causes of premature death, all the suffering people had to endure, all the material scarcity we don't face today.

Then project forward another century. I genuinely believe in a century's worth of science and medicine progress, if we can get through this current hurdle – and I believe we can. I'm feeling increasingly optimistic. We'll have a far better world than today.

Writing, Thinking, and AI Assistance

Emily Chang: I know how much you love writing. You're known for writing essays. Do you use Claude to assist with writing?

Dario Amodei: I do. I'm not at the point of directly using text Claude writes, because I have my own style and am particular about it. But I basically use Claude to assist with brainstorming, to help organize topics, or provide some content I can reference.

So it plays an assistive role. I don't know how far off the day is when Claude can write better than me. We're not there yet, but, I think that day will definitely come.

Emily Chang: I also love writing, and I think writing helps you clarify thoughts. There's a lot of critical thinking involved. If we let Claude do it for us, do we lose that ability?

Dario Amodei: I'm a bit worried about that, in fact, that's half the reason I insist on writing myself. That's true for external audiences. Although many people read my essays, writing is also to clarify my own thoughts, to know what to do next, and to establish a shared reference point between me and others.

I think we're still figuring out how to use AI in a way that preserves those benefits. I think what I'm doing now is like that, e.g., I use Claude for research. And I use Claude to help organize my own thoughts.

I think if we just use it wholesale, like have it write an essay on AI risk, first, it wouldn't have my personal insights, second, I would indeed lose those benefits of writing. As models improve, I imagine there will probably be some way to use them more directly in writing while preserving those benefits, but that will be subtle. It won't be monolithic, we'll have to navigate this over some time.

I think we could end up with a very anomalous combination of rapidly growing GDP combined with high unemployment, or at least underemployment, or lots of low-wage jobs, and high inequality.

Unemployment Risk, Productivity, and Macro Policy

Emily Chang: You've been very outspoken on unemployment. AI could eliminate half of entry-level white-collar jobs in the next 1 to 5 years. That was last year. AI is advancing incredibly fast. Is it still 50%? Or higher now?

Dario Amodei: I've always said, if you go back to those original video clips, they always get clipped for that three seconds out of context, but actually, what I've always said is: I don't know what will happen, but this indicates the magnitude of potential disruption.

And, I've also always talked about what we can do about it, I've talked about token taxes, and working with enterprises to adjust staffing. I'm a bit skeptical of retraining programs, but we should certainly consider them in macroeconomic policy.

From the beginning, I've always talked about solutions, but there seems to be a human psychology tendency to just clip that three seconds about doom. So the message is absolutely not doom. I'm saying, this is something we should anticipate and be concerned about, and we do need to respond proactively. I'm not entirely sure, but I remain quite concerned. My level of concern remains the same.

We're now seeing AI boosting people's productivity, but that's the usual early growth phase. If you look back at the Industrial Revolution, I wrote about this in 'Adolescence of Technology.' You automate 90% of the work, that's fine.

People become 10x more productive on the remaining 10% because they get 10x leverage. But eventually it approaches 100%. Then the linked question is, you have to find something else for them to do.

About the long term, I'm not sure. I am uncertain about that. But I do think there are different types of adaptation. For example, one thing I talk about is software engineers inside Anthropic. We're going through that transition now, where AI writes all or almost all code, but AI does make software engineers more productive.

But still, it enhances people's productivity. However, we are starting to see hints that for some people, it might not enhance their productivity, in which case it might be better to just let AI do the job directly.

So that's one aspect. The other aspect is, where do we need to increase demand? We call them frontier deployment engineers, or something like applied AI solutions architects, who do a mix of technical development and customer communication. The demand for those is huge right now, because there are many customers and we are expanding fast.

Now, is that a fit for everyone doing pure software engineering work? It's not a perfect match. It's not one-to-one. That makes you realize there will be huge disruption, but things will also adjust. Which side wins? I don't know.

But the reason to sound the alarm is that's how we respond. That's how we formulate policy, both at the human level and the global macroeconomic level. We want to put forward considered views. We don't want to make promises people think we can't actually achieve or deliver. We don't want to make unsubstantiated claims. We want to think carefully about these problems and explore actionable solutions.

Emily Chang: You showed this chart showing potential job losses, e.g., in sales, finance, etc. What jobs will go, who gets replaced, and what new jobs are created?

Dario Amodei: Actually no one can be sure, because the economy is unpredictable, it's like the stock market, they are decentralized processes, you can't really know in advance what the part is that people will still be able to do.

But I'd say, broadly speaking, any area with entry-level white-collar work, whether in banking, finance, or elsewhere, AI first has huge potential to enhance people's productivity, but later, AI will gain the ability to do entire jobs.

Then we need to think, what can people do? I think we need to plan for this in advance. We're already doing this when talking to enterprise customers. We see the choices they face. The choice they face is, do I cut costs?

That usually means fewer employees, basically doing the same with less, or should we do more with the same amount of resources? Whenever possible, we always try to push them to do more with the same resources, because that basically means hiring the same or even more people, just to do some new work, pushing them toward a positive-sum game.

The advantage we have now is, the pie will expand greatly. Precisely because the pie will expand greatly, there will likely be new areas people can move into. The question is just whether we can find those fast enough. The scale of this disruption will be very large, that's what I'm alerting people to, but we have to solve that matching problem.

Emily Chang: Walk me through it, like five years from now you wake up, what does the country look like? What are all these people doing? Because if unemployment is that high, isn't that the prelude to revolution?

Dario Amodei: That's exactly what we want to prevent. We absolutely want to avoid that.

I think there are a few entry points. None of these are guaranteed, we can't be sure, but there's the physical world, e.g., things that exist in the physical world. The robotics revolution is happening too, but it's slower than what's happening in AI. People always talk about building data centers, but when processing any kind of information becomes much easier, the limiting factor might become things in the physical world.

So, we need more people making, building, producing in the physical world.

Anything human-centric, I think that will be a big thing. I've heard many stories about AI catching problems doctors missed. I'm glad, but people still want to talk to other humans, especially on important matters. Maybe AI can provide better customer service, but even so, people, or at least some people, still want to talk to a human.

So that kind of relationship-oriented work, I think it will become important. And I think there will be some role for humans to direct AI, to some extent it has to align with someone's values and intent. So I think there will be some role there, though I don't know if that role is thin or thick. I think it's hard to say.

Pushing Back Against "Doom Marketing" Accusations

Emily Chang: There's been a lot of pushback. I know you said you're trying to alert people, but Jensen Huang said you're conflating tasks with jobs. Others have said it's more like "doom marketing" that benefits Anthropic. So I want to articulate and firmly rebut that.

Dario Amodei: I have some thoughts on the overall unemployment risk situation. I mean, we haven't fully fleshed them out yet, because I want to make sure they are accurate. But Anthropic has proposed many schemes. We've had economic grants. We have economic indices. I've talked about possible avenues through taxation and macroeconomic policy to address these risks. About what the new jobs are.

In 'The Adolescence of Technology,' I laid out, I have about five pages specifically on the task vs. job distinction, why this time is different, and listed six things we can do, from private charity to government action. I talk about the problem, I talk about solutions.

But social media, I detest it, as a category I detest it. People just clip a three-second clip from a year ago, they haven't actually read the essay, or they leverage social media's nature – and I've discussed these risks very carefully in the essay.

The idea that this is just cheap marketing is itself cheap marketing. It's lazy. It's failing to engage with serious intellectual work. I think that's also part of the problem. Again, I think it's part of the Silicon Valley sickness. It's trapped in this three-second social media world.

So people only react to that, or think they only need to react to that. Again, I think it's very dangerous, and we're not having a mature conversation.

Instead, people just lazily look at this three-second clip. Then they'll say, 'Oh, that's what Dario said. That's silly. That's unserious.' Every time someone says that, I discount their opinion.

National Security and 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. A confrontation is brewing between Anthropic and the Pentagon over AI military safety guardrails.

Emily Chang: You've long held anti-war views, dating back to your time at Caltech. Yet, you were among the first AI companies to sign a contract with the Department of Defense to operate on U.S. classified networks. For waging war. Explain.

Dario Amodei: Okay. I'd say, look, the world changes. As with my view on this technology, I am concerned because we are facing a resurgent authoritarian bloc that is very aggressive, and we need to defend ourselves.

This is something I've long believed, and still hold, which is why, across two administrations, though I may not agree with every policy, that's why we broadly support this.

So that's why we work with them. We certainly aren't doing it for the money. It's very cumbersome, even setting aside the legal battle, deploying on government networks doesn't pay much but takes enormous effort. So we do it because we feel a mission about it. But likewise, since we do this work out of concern, there must be limits on how the technology is used.

[I've made the point when talking about technology stages: We should use this technology in every way except those that undermine our own values, and the red lines we've drawn around mass surveillance and fully autonomous weapons are things I think would undermine our values.

If you resort to those, then winning is pointless. That's the balance I see, and the position we hold. That explains why we are both among the first to work with the Department of War, yet also hold back where others are willing to do certain things.

I think you need to pick a stance and stick with it. That kind of company stance shifting – from initially saying we won't work with governments, to suddenly taking all government requests – I really don't understand. You should pick your principles and adhere to them.

Emily Chang: You've been working with Palantir since 2024.

Dario Amodei: Correct.

Emily Chang: Their technology is used by ICE and police departments in Gaza. Is Claude being used for surveillance in other ways?

Dario Amodei: We do not work with ICE, either through Palantir or any other channel. We do not work with CBP. I don't think we have business in Gaza. We are very careful to restrict our collaborations to things we agree with.

Emily Chang: So, you draw your red lines, the president bars you from working with the federal government, the Pentagon marks you as a supply chain risk, OpenAI steps in and signs the contracts you refused. What does winning this fight actually look like?

Dario Amodei: I don't think a private company wins this fight. It's not a fight at all. Anthropic isn't trying to win, or thinking in terms of winning or losing. It's more a debate about how government should properly use AI. AI is a new technology. We don't yet know where it's reliable, where it's not. We don't yet know where it advances our values, or where it undermines them.

So, I think one important thing is to establish precedents for use cases we think are good – frankly, most use cases are good – and for those we are concerned about. As I said, we've seen contracts can only do so much, as we've seen, others may sign a contract that doesn't respect the same lines.

But what it does is raise awareness of the issue. Right now, there is serious bipartisan work in Congress trying to ban some of the things we worry about, and trying to put in place guardrails. Again, I don't want to frame it as a fight, but in a way, it has been effective in prompting our nation to think more carefully about proper use of this technology.

Values and Military Ethics

Emily Chang: Anthropic is run by an ideological nut who shouldn't have sole decision-making power. What I care about is the broader AI issue. Do you mind being called an ideological nut or a bunch of left-wing nuts?

Dario Amodei: I've been called much worse. People can call me or Anthropic whatever they want. What actually matters is two things: We succeed as a company, and, we stay true to our values. Actually, in some ways, my life is very easy, because when you only care about those two things, it's really simple, isn't it? Like that, you always know where you stand.

Emily Chang: A U.S. official stated that with LLM assistance, the U.S. military increased daily strike targets from 1,000 to 5,000. That means Claude can help kill more people faster. Are you comfortable with that?

Dario Amodei: I think there are two things here. First is the U.S.'s ability to be more efficient militarily. I support that ability. I think having stronger capability in that doesn't cause war, it deters war. Basically you're asking, do you believe in this country, do you want this country to be a stronger actor, not a weaker actor on the world stage? I do. I'm a patriot.

There's a separate question, which is, are there certain policies the U.S. government engages in that I might support or not support?

What the government engages in, obviously, I support some, not others. That's not for me to decide. If we provide a technology, the DOD makes this point, and we actually agree with them. If we provide a technology, we don't get to decide which military operation you can do, or which you cannot.

Now, personally I might think one military operation is justified, and another is a terrible idea, but we don't deny the technology for that.

You have to leave policy-making to the military decision-makers. What you can do is set some high-level boundaries. For us, that's preventing uses that conflict with our values, with our country's values, and promoting uses we think help advance our values. That's how we think about it.

Emily Chang: Bloomberg reported that the U.S. military used Claude in the Iran war, through Palantir's platform Maven Smart System for AI-assisted targeting. Reportedly, in February a U.S. missile struck a girls' school in Iran, killing over 150, most children. Did Claude play a role in that strike?

Dario Amodei: We, look, we don't have access, we don't know exactly how these models are used. Obviously, these kinds of mistakes in war are really, really terrible, it is a very terrible thing.

If this doesn't illustrate why we insist on opposing uses we don't support, we are even willing to risk the company's future to restrict how these models are used. And, the use case you're talking about doesn't even cross our red lines.

We worry about 100x more use cases in the future, some of which do violate our red lines. Now, again, I think overall, the use of these models is appropriate. I think the net effect is good.

But military decision-makers make incredibly grave errors even in the best of times, and I don't know if we're in the best of times.

For example, we can discuss a few things. We can discuss how to set red lines to prevent model uses more likely to cause these problems. If we had allowed, completely unfettered, if we had caved on fully autonomous weapons like almost every other company now has, as seen here, it's Claude assisting, but humans making the final decision.

So it's humans making the final decision, not Claude. Imagine a world where – not Claude, because we didn't allow it, but someone else's AI model – the AI model makes the decision directly and no human ever looks. That is what we stand against.

That is exactly what we resist. I'd say, there's another thing here too. Again, I think procurement isn't the right way to achieve this. But we do need to ensure – and this matters not just to me as a technology provider, but to the American public – that U.S. military decision-makers don't make these errors.

Ensure they operate reliably, and make sound choices when taking action.

Again, as a citizen and a provider of technology, that's what I care about. For example, the government uses technology, we use Microsoft Excel heavily. If you said, you can use Excel for this military operation but not that one, that's not feasible in reality. I hope that gives you a sense of how we think about this.

Emily Chang: This school had a website, you could have found it on Google search, shouldn't Claude have caught that? Or shouldn't AI or whatever tech they used have caught that? Doesn't this raise a more concerning issue about using technology as a shortcut in war?

Dario Amodei: Listen, listen, I'd say, I'm not sure, I really don't know. This relies on, perhaps, classified knowledge beyond my access. But, the principle we've established, which I think was adhered to here, is humans make the final decision. I don't know what role Claude or any other AI played, but if this doesn't illustrate why that principle is so important, I don't know what does.

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

Dario Amodei: On balance, I think it's more likely to prevent war. But if we have no limits on how it's used, then I think it's more likely to cause war. You've seen Dr. Strangelove, the premise of that film is you have a doomsday device that automatically launches nuclear weapons in retaliation if it thinks it's under nuclear attack.

What could go wrong, which again brings to mind things like lethal, fully autonomous weapons. I think the way conflict happens is, sides constrain each other, wait for opportunities. They get miscommunications. If we lack proper oversight on this technology, I think the chance of such accidents increases.

Now, I think if AI can be properly leveraged, even just for intelligence, not just war. Imagine, when you know everything an adversary is doing, they will think twice about some invasion or military action. So, I think superior intelligence can indeed deter conflict. Superior responsiveness can deter conflict. I've always believed in these things.

Mythos: The Safety Game with Frontier Models

Emily Chang: Anthropic is in the headlines almost weekly. Of course, the focus now is on Mythos.

Narrator: This is currently the newest, most powerful Anthropic model, capable of autonomously traversing the entire cyber kill chain.

Emily Chang: You've said Mythos is too powerful for public release. What surprised you most about it?

Dario Amodei: I think what surprised me most is, the model's capability in finding vulnerabilities has been climbing steadily, and more importantly, its ability to turn those vulnerabilities into exploits – and people usually only talk about the vulnerability itself. People often don't talk about the process of turning a vulnerability into an exploit, and it's become quite good at that.

So, what surprised me is we witnessed this huge leap. It was an especially large jump. And without any prompting, some of the companies we initially provided the technology to said, this is like a superweapon. You should need a gun license to use it.

Please don't release this. That request for restricted use came from the companies we provided the tech to, which found so many critical vulnerabilities and exploitability around them that they essentially asked us not to release it.

To be clear, because social media distorts information, our goal isn't to lock it away forever. We are trying to open the technology gradually. To broader and broader groups. Ultimately, we think Mythos should be released to the public, but with strong cybersecurity safeguards.

A current concern is, we released the current cybersecurity safeguards with Opus 4.7, which is a good cybersecurity model but much weaker in capability, and those safeguards can be jailbroken. We're a bit worried about other companies that think this is sufficient defense, because while it works sometimes, we all know these classifiers can be jailbroken or bypassed.

Based on our own testing, and frankly, based on our assessment of defenses other companies have, those defenses are not strong enough yet. That's what we're waiting for now, for defenses to reach a level we truly have confidence in.

Emily Chang: There was a lot of pushback at the time. Researchers said they could reproduce the process with cheaper open-source models. Some said OpenAI already has these capabilities. So, how do you respond to those who think it's a big PR stunt?

Dario Amodei: The claim that it can be reproduced with open-source models is completely false. The core idea of the system is Mythos can look at an entire codebase and find issues. Someone on Twitter said, if you point an open-source model specifically at the exact line of code Mythos found, then it can find the same issue. That's not, that's not the prompt, that's not the problem, like, that's not, that's not the same thing at all.

The ultimate test is, we go to companies, we look at open-source codebases, we found 271 new vulnerabilities in Firefox. We've found thousands of vulnerabilities in private companies that haven't yet fixed or been able to disclose them.

It's like, no previous model found those 271 vulnerabilities, so compared to 'I found the exact line of code Mythos found, I know I found the needle in the haystack, and something else can now pick up that needle,' the workflow that actually works in practice is different.

But for those saying it's just marketing, I'd say, we've taken a huge commercial hit by not releasing this model. This model vastly accelerates research inside Anthropic and subsequent model production, and if we released it, it would provide the same acceleration to the outside world. This has hurt us commercially a great deal.

Emily Chang: If it helps defenders, it also helps attackers. Can we defend anything anymore?

Dario Amodei: I'd say, the reason we give Mythos to defenders first before attackers is to patch all vulnerabilities. I don't know, as models get better, more vulnerabilities might be found, but there is a finite number of vulnerabilities. They are finite, like you have this surface with only so many holes.

You patch all the holes, the surface becomes extremely hard to attack; simultaneously, the code itself is written by powerful models, thus it's harder to find flaws or hack into it. So I think on the flip side, hopefully in 6 months or a year, we'll have a more secure internet ecosystem than before. We're trying to achieve that, and doing our best to open Mythos to new cyber defenders.

We've been communicating with the government. We respect their advice highly. They are slowing our opening due to counterintelligence risks. I think that's wise. I think all serious people here understand there are real trade-offs.

We see a lot of snarky criticism on Twitter and from other AI companies. You look at their words, and then look at their actions not matching. They are not serious people. They are not taking seriously the hard trade-offs we face here.

Look, every day customers call me saying, I want to use Mythos. Countries call me saying, I want to use Mythos. And the U.S. government and my security team tell me, no, wait, there are risks. I'm not saying which side is right.

I think the truth is somewhere in between. Both sides have reasonable points. Both sides have reasonable points. But there are real challenges we need to face as a society, not just dismiss things as cheap marketing, nor engage in counter-positioning using a chief marketing officer like some other companies have.

It all shows an incredible lack of gravitas and maturity. We need to face this moment together.

Checks and Balances, Government Regulation

Emily Chang: Have you already had to make trade-offs that made you uncomfortable?

Dario Amodei: Anthropic's entire history has been a history of trade-offs. Anthropic's entire history, in some ideal world, you'd prefer to spend years studying every possible thing that could go wrong before releasing the first chatbot. Now, we did delay.

We did delay Claude's first release, but we only delayed a few months. So I say everything is a trade-off. Being at extreme ends of the spectrum is completely crazy, so everything is a trade-off.

I'd say, now that we are in the commercial leading position I described, actually, Daniela and I are doing everything we can to push further progress on being careful. That's the purpose of the Mythos release; it's hard to do this if you're not the industry leader. So I think you'll see more of that.

Emily Chang: There's a view that, why doesn't the government just take you over? Why should they let a private company control such powerful technology?

Dario Amodei: So I do think this is a very serious question, and I share those concerns. I don't think the government should take us over directly, but I can put it this way. I'd say, looking back and describing the situation, every previous powerful technology in history was either built by the government, or originated with the government.

Like nuclear weapons, obviously, initially built by government, and basically built by government thereafter. But even things like the internet, GPS, cell phones, all the R&D was done in labs, federal labs, and universities.

AI is the first technology built in the private sector where the government played no substantive role and came relatively late to the game. I think this is actually a dangerous and unstable situation. It's not the one I would choose. There's no alternative, the technology is buildable. Our competitors are building it, it has economic value, it will be built anyway, the problem is the government wasn't involved, not that the private sector is doing it.

I think we need to think about checks and balances, so I think there need to be checks on the power of AI companies. We have a mechanism, the Long-term Benefit Trust, which is essentially an entity that can appoint and remove a majority of the board members. If you follow the logic all the way, it effectively has the power to fire me.

We are now introducing some elements – far from all elements – we are introducing some elements of public-like governance, i.e., you must be accountable to someone beyond just the company's shareholders. This is very important, and this structure will persist regardless of what happens to the company in the future. In AI, we encourage other companies to adopt similar structures.

On the government side, I think we need checks. Congress has announced efforts to set red lines, so I really think the legislative and judicial branches need to play a role, because with this technology, I'm worried about companies having it, and I'm worried about governments having it.

Then, corporations need to check the government, and the government needs to regulate corporations. We need fundamental regulation of this technology. I think we need to start implementing pre-release testing, i.e., mandatory pre-release testing, and testing and auditing of models.

What's very ironic to me is, there's a faction in Silicon Valley tech that started with the position that even transparency around this technology, even export controls, would completely destroy our ability to create it, kill innovation. Yet once they saw the first real danger I've long predicted, it changed. They started talking big about nationalization, that the government should just take it over.

Come on, folks. You went from the most extreme anti-regulation – like 'if you even look at us, you're destroying the industry' – to a completely communist 'the government should seize everything.' We need a more rational, more moderate approach.

That's what we've been advocating all along, because we deeply understand the power in this technology. We are not panicking. We are not in denial. We see the smooth exponential growth and are responding accordingly.

So, how was your visit back to the White House? We always try to work with whoever we can. At the government level, I've said we have a simple approach. We have a set of principles. We follow them and hope people on the other side are reasonable. Honestly, the government is taking Mythos very seriously.

For example, we've had good conversations with Secretary Bessent and Chief of Sotomay. And Chief of Staff Susie Wiles. I think they do understand the nature of the risks here. I think Mythos is helping them perceive these risks more concretely.

So, as with any administration, there are parts we get along with very well, who understand this. Of course, there are also parts that are harder to deal with. I think that's normal. That happens in any government, we just try to handle it properly.

The Intelligence Premium, Geopolitics, and Future Picture

Emily Chang: Early in your career you worked at Baidu, a large Chinese tech company, at its Silicon Valley office, and you've been clear about China. China is rolling out powerful open-source models, and U.S. companies are building on them for free. Is that a threat?

Dario Amodei: So, one thing we see with this technology is, the premium for model intelligence is indeed very high. We rarely see people preferring less intelligent models. To be clear, there is already a thriving ecosystem. There are many challenges and problems simpler than what requires frontier models.

But still, it's exponential growth, meaning these non-frontier models may have economic value comparable to what we saw in 2023 and 2024, but remember, we're growing at 10x per year.

So we find, things at the frontier are always much more powerful than things away from the frontier. I think this isn't particularly well understood yet by people accustomed to building products in the old paradigm. As someone who never ran a company before, never thought about the previous product era, and especially detests the social media era.

I feel like an outsider to that world. And I think people's intuitions are wrong. They have all sorts of rules of thumb about products. I think models growing exponentially at 10x per year really break that. Like intelligence is an extremely crucial factor, outweighing everything else. So we see again and again, value tends to reside at the frontier.

Now, what I worry about with lagging models is the risk they pose, i.e., Mythos-level cyberattack capability. In 12 months, we'll have even stronger cyber capabilities, but Mythos-level cyberattack capability might be downloadable by anyone then. Hopefully we've patched everything by then. I don't think we have a way to stop it, but I think it's a serious problem.

Emily Chang: Did what you saw at Baidu shape your view of China?

Dario Amodei: Not really, no.

I worked there for a year. I guess I maybe learned more about speech recognition and things like that. Perhaps the only thing that concerned me was, part of how we got speech recognition data was – in China they don't care about privacy, so we had all this speech recognition data.

But other than that, my concern here is geopolitical. I think what worries me most about what happens in China is what we see in the U.S., i.e., the suppression of criticism we see, even within the U.S., like what happened in Hong Kong, like the ability to intervene in U.S. reality.

Business networks and suppress criticism, it's an authoritarian country. And a high-tech authoritarian one. When I see that combined with AI, you indeed get a dystopia here, like 1984 or worse.

My focus is trying to prevent that. I think we have a chance to stop it.

I think we have a chance to make AI a technology that supports freedom, one that makes humans more free, capable of delivering on the promise of equal justice for all, otherwise it could go the other way. Which way it goes depends on the actions of AI companies, on government actions, and on all of us. Therefore, I think we have a responsibility here.

Self-Iteration and Civilizational Survival

Emily Chang: People in your field often talk about a moment: when AI becomes powerful enough to improve itself, then the improved version iterates further, and so on. Some of your researchers think this moment is coming soon. How far away is this moment?

Dario Amodei: I don't think it's a specific point in time. I think it's a continuous process. In some ways we've already seen it, that AI can suggest architectures for the next generation of AI. I'd say a year ago, we observed total factor productivity improved by 10% to 15% due to AI. For example, that percentage might have gone up to 20% or 30% now. It might even be doubling.

Like everything, we are on an exponential curve, there is no sudden moment where AI self-evolves, runs away, or becomes unsafe.

What we experience is an accelerating exponential growth. And at each point on the exponential curve, we must assess: Is it time to slow down? Is it time to impose more control on this technology? I think more and more control will be needed in the future. But I think the key to navigating this is this smooth exponential curve.

Again, I think those who opposed all AI regulation, then advocated nationalization after seeing something, taught us a lesson. I think those who downplayed AI's power, then exclaimed AI is self-evolving, about to run away and demanded a full stop, taught us a lesson.

This kind of wild swinging between extreme reactions is extremely unhelpful for dealing with this technology.

The right way, the wise way to respond, is to stay calm, not panic. Our responses will scale smoothly as the technology's power increases. If you see someone having these crazy swings. It means they were caught off guard, and they are not serious.

Emily Chang: I understand one of your favorite books is 'The Making of the Atomic Bomb.'

Dario Amodei: That's right.

Emily Chang: Do you see any parallels between yourself and Oppenheimer?

Dario Amodei: The person I identify with most is Leo Szilard, who was the first to conceive that there might be something like a chain reaction.

Listen, my view is, we won't solve this through larger-than-life personalities, or people trying to be at the center of everything. This requires checks and balances. There are many powerful actors with interests here. The only way for everyone to have a good outcome is to have some mechanism, i.e., basic checks and balances everywhere.

So in a sense, I actually see Oppenheimer as a cautionary tale of what not to have happen.

Emily Chang: You've said the chance of civilizational collapse is somewhere between 10% and 25%. That's not trivial. Is there a scenario where the cause is something built by Anthropic?

Dario Amodei: I certainly hope not. My view is, the actions we take reduce that probability, not increase it. That probability comes from the very straightforward logic of the technology itself: There are many countries in the world, many companies in the economy. If the vacuum isn't filled, new companies will emerge. That's the predicament we face.

We are trying to take actions to lower that probability. I think we lower it far more than we raise it.

But, an inherent property of this technology is unpredictability. So, we do a lot of building and testing before releasing products, and I think, the models released so far are not dangerous, at least not outside cybersecurity. Then we try to iterate and learn from that. So we have numerous defense mechanisms, half the work inside the company is about reducing risk as much as possible.

But, risk can never be reduced to zero.

I'd say, suppose there are many airlines, and you decide to start a safer airline. Your airline can be 10x safer than all others. But, if someone asks you, can you guarantee your planes will never crash? How could you possibly guarantee that? How could you do it?

Emily Chang: But if the chance of a crash is 25%, you wouldn't get on that plane.

Dario Amodei: Exactly. 25% is too high. We are trying to get that probability down to a very, very low level. That's our goal.

Building Trust: Proving Difference Through Actions

Emily Chang: You are building something extremely powerful and stand to gain enormously from it. Why should we trust you?

Dario Amodei: I think when any company starts, especially given the behavior we've seen from Silicon Valley as an entity, it's natural for people to have that view. Looking back over the past few years, I think starting from a place of distrust – if you know nothing about me, or about Anthropic – is a fairly rational attitude.

I think Silicon Valley has lost too much of the world's trust, and we have to earn it back.

The message we want to convey is, we are indeed different. And that trust must be earned through our actions. You can agree or disagree, but we have stayed true to our values. The whole Mythos story, not releasing this extremely powerful model has commercially constrained us.

There were many smaller things before that. We've been consistent on China. We cut off access to the models. We didn't have to do that. No one asked us to. You know, at a time when hundreds of millions of dollars was still a significant portion of our revenue, that cost us hundreds of millions.

For example, the delayed release of Claude II, we have a long history on this.

We're not perfect. We'll make mistakes. But, I'd ask people to look at our overall history, what hypothesis best fits that history? People have to decide for themselves.

But I think a consistent hypothesis is, we are genuinely trying. To do the right thing.

We're not perfect. Organizations are always dysfunctional. We're always trying to fix them, make them work better. There will be many missteps, many things that go wrong. But fundamentally, we have a sincere, serious vision of how to do right, and we are trying to execute on it.

Emily Chang: Well then, we'll see you on the other side of the exponential growth.

Dario Amodei: Well, hopefully.

Emily Chang: Well, hopefully.

Dario Amodei: One thing about the CEO job that surprised me, which I didn't know before, is you have to wear makeup often. That was completely unexpected.

Emily Chang: Just a bit of foundation.