Original Video Title: Inside the Mind of Anthropic CEO Dario Amodei | The Circuit | Extended Interview

Original Video Source: Bloomberg Original
Original Article Translation: Web3 Skytopia

Editor's Note: Anthropic CEO Dario is currently in a very awkward position. On one hand, holding the world's leading AI model, and on the other hand, inadvertently being globally shut down by a U.S. order, to the extent that even non-U.S. team members cannot use it.

The outcome of this situation remains uncertain. It is said that Dario is continuing his efforts, and everyone can continue to watch with interest. But we can catch a glimpse of the mindset of this controversial leader in AI coding from the latest Emily interview.

In today's Silicon Valley power map, Anthropic is in an extremely unique and tense central position. As the most powerful challenger to OpenAI, it was founded by a group of top researchers who chose to "leave" due to differences in values.

When CEO Dario sits under the spotlight discussing AI's exponential growth, he exhibits a rare, surgeon-like calmness. This is not just a competition about technology, but also a deep game about trust, security, and how human civilization deals with the intelligence explosion.

Full Interview Summary

This latest interview delves into Anthropic CEO Dario's journey of facing AI's exponential growth, covering insider insights from leaving OpenAI, the company's business model choices, to the profound impact of AI on the job market, cybersecurity, and geopolitics.

The CEO elaborated on how Anthropic balances power by establishing mechanisms such as the "Long-Term Interest Trust," and how, while pursuing technological leadership, it practices its safety values by setting "red lines" and delaying the release of high-risk models (such as Mythos).

Note that Dario's remarks have always been unfriendly to Eastern powers; please judge for yourselves.

Key Insights

· The AI industry is currently experiencing a "smooth exponential growth," which, when the growth accumulates to a certain degree, will result in a sense of qualitative breakthrough.

· Trust is the cornerstone of cooperation in the AI industry, and Anthropic advocates that trustworthy participants should come together to establish industry standards.

· Enterprise business models, coupled with AI safety values, are more synergistic and can avoid the addictive and low-quality content competition common in the consumer market.

· For the unemployment risks triggered by AI, society needs foresight to formulate macroeconomic policies while seeking a positive-sum game of "doing more with the same resources."

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

Below is the full interview:

The Pressure and Experience of Exponential Growth

Emily Chang: How much sleep do you get?

Dario Amodei: I have never been someone who sleeps well. I can only say that I am learning the art of relaxing and falling asleep in extraordinary moments of pressure.

Emily Chang: Everything is developing so fast. What does it feel like to be in the midst of it?

Dario Amodei: It's an exponential feeling. It's 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 from a nap, two days have already passed on Earth. So you have to deal with two days' worth of things in one day.

Then you go back to sleep, and because you're continually accelerating, three days have passed on Earth. The next day, four days have passed. That's pretty much what it feels like.

Emily Chang: Do you often feel anxious about waking up to face what lies ahead?

Dario Amodei: There are already plenty of clear and pressing issues we need to address, and as I deal with these issues, I'm also thinking about how we can be prepared. But I think being paranoid or anxious about what you'll face upon waking up is not helpful. I have studied those in history who have dealt with such high-pressure situations. You need to learn to respond rationally and not equate the severity of various dangers.

This state of jumping back and forth between "I'm not worried" and "Oh my, we must panic today" is, in my opinion, a sign of immature decision-making. Whereas truly mature decision-making is that you can't ignore this, we can't take this lightly either.

In fact, the risk is indeed becoming greater, but we must respond rationally, like a surgeon handling surgery. Or like a military officer handling a military operation, or like anyone making decisions that affect many people, they must rationally make these decisions. And they must understand the risks involved, but they must also maintain a basic level of calm.

So my son asked me yesterday if he could use my Claude account, I said absolutely not, I need my tokens. We are seeing more and more of these applications in the consumer market, and we originally hoped to become more of an enterprise company, but even though we didn't put in as much effort, the consumer business is starting to grow rapidly.

Emily Chang: You are now at the center of the AI universe, how does that feel?

Dario Amodei: Interestingly, throughout my entire career, especially during the time since joining Anthropic, what I have experienced is a kind of smooth exponential growth.

The experience of this smooth exponential growth is: nothing happens, nothing happens, nothing happens, some small things happen, and then suddenly, it just explodes like crazy. That is the experience of this world. It is also the experience of the scale of a company compared to other companies and to this world.

So I stared at this chart for a long time, and then I said, we will likely become the highest revenue and valuation AI company at about this time. Indeed, it happened. It happened. So in a sense, I am not surprised because the curve on the chart is very smooth.

But of course, from another perspective, when things really happen, you see more, richer details and colors. And it is truly amazing.

We always bear in mind the usual issues that we focus on, namely how do we train excellent models? How do we apply them to excellent products? How do we ensure everything is secure? How do we help people while managing the social risks surrounding this technology? All these are the same issues, just examined under a larger microscope.

Upbringing and Silicon Valley Spirit

Emily Chang: What were you like as a child growing up in San Francisco? I know your father was a leather craftsman and your mother worked at the library. How did this influence you?

Dario Amodei: At that time, the whole Internet revolution was happening around me, but I had no interest in it. I was only interested in studying mathematics and things like writing and drawing. I was interested in exploring the universe. I was interested in science fiction. Overall, that was the environment I was in at the time. I think I just had a strong curiosity about the world.

Emily Chang: You grew up in what's known as a tech hub. And now it's the center of AI. Did this place, this city, shape your worldview in any way?

Dario Amodei: It did. I think that kind of nonconformity, individualism, and the idea that "being a little crazy is okay" spirit is present. I think a large part of that might have indeed had a subtle influence on me.

You hear stories, for example, when you go to certain countries in Europe, or even certain regions of this country, that kind of thinking about problems in different ways is often suppressed or seen as weird, or having some crazy ideas.

Actually, I have a lot of criticisms about Silicon Valley, but I think one good thing about it is that it encourages an idea: even if all the experts are against you, it doesn't matter. If you have a coherent vision and a coherent worldview, you should pursue it, that is what's most important. Maybe it won't work at all.

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

Emily Chang: You, Daniela, your sister, and her husband Holden Karnovsky all lived together in a shared house in 2016. What were you arguing about at that time?

Dario Amodei: I think it was right when the Open Philanthropy project was just starting, and Holden was the lead on that project. At the time, I was a life scientist. So I was helping them with some things related to global health or biological research. So I, provided some advice on those topics. For example, which areas are promising? Which areas are not so promising?

The Core Disagreement upon Leaving OpenAI

Emily Chang: Your decision to leave OpenAI has become a legend in Silicon Valley. What really happened? Beyond those narratives, what is the fundamental issue? In what aspects did you have disagreements?

Dario Amodei: Look, I'm going to say it, I'm going to say it very simply. When you're building powerful technology, you face a lot of challenges, and Anthropic goes through these every day too. We don't know if the decisions we make are right or wrong.

So, there are many valid disputes on security issues. We certainly had some disagreements with them, but you see, just on that basis alone is not enough to be the reason for leaving. There were disagreements with people here. There were disagreements among the people here.

But when you feel you can't trust someone, when you feel their values are not as they say, when you feel they are not honest enough, when you feel they are not acting for the reasons they claim, when you see disturbing patterns of behavior or dishonesty, it makes it hard for you to continue working in the company and hard to trust the company. Ultimately, when you no longer share a vision with someone and no longer trust each other, why bother arguing anymore?

The solution is to go our separate ways, you do your thing, they do theirs. I completely accept this: we act our way, they act theirs. We will see who can succeed in the market, who can win in the court of public opinion. I believe these facts are more convincing than rhetoric, more persuasive than any dramatic speculation about who left and why they left.

We want to emphasize that we are providing an example of how to deploy this technology, which we believe is a responsible approach. If they have objections, they should present arguments. I think there is no need to say much about this matter.

Industry Collaboration and Trust Rebuilding

Emily Chang: At an AI summit in India, there was a moment when you and Sam Altman seemed to refuse to shake hands on stage. What happened at that time?

Dario Amodei: The situation at that time was that the summit was organized in an extremely chaotic manner. We were all brought on stage at the last minute, and they changed our seating arrangement at the last moment. Then they took a photo of us and then instructed all of us to hold hands. If you have ever attended such a summit—no offense to India—but all these international summits with heads of state in attendance are very chaotic in their organization.

Emily Chang: But everyone else held hands. Come on.

Dario Amodei: Look, I don't know how to tell you, okay? At that time, Narendra Modi was right there, suddenly calling everyone to hold hands.

Emily Chang: Fine, fine.

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

Emily Chang: If the developers of the most important technologies in the world cannot shake hands on stage, how can we believe that you will cooperate on existential risk issues?

Dario Amodei: So, this is what I want to tell you. Among the people building this technology, there is a huge difference in their qualities and credibility. I think this means that different people, feeling that no one trusts each other, I don't think that's right.

I know Demis Hassabis, who built the Gemini model, a competitor to the Claude model. I have known them for 15 years. We have worked together to address many issues. We bought computing resources from Google. We often exchange ideas on security.

So, my view is, first of all, some participants are more trustworthy than others. And also, I think there are participants outside of Anthropic that I trust, and I think they're trustworthy. What I think needs to happen is that trustworthy participants need to come together to address those who are not trustworthy. So that participants are forced to adopt the same standards.

From a rich experience, I have learned that some people will not consciously do the right thing. But if the majority of people in the industry are doing the right thing, then I think there's no choice for the others in the industry but to follow suit. It's like a positive version where you're inspiring other people.

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

It's just because every company is doing something cool. And other companies would be like, that's cool. We'd like to try that and see if we can do something new in it. This is the carrot side of top-notch competition. And then there's the stick side, or the implicit stick, where you realize that these people are doing the right thing. If others are not doing the right thing, it looks bad.

We often see this behavior: they reluctantly do the right thing but try to pretend they're doing something different and imply that we have some sort of bad or evil intent, which is predictable. But I think this is the way we integrate the industry and promote industry collaboration.

Business Model: Enterprise Fit with Values

Emily Chang: Earlier, others were focused on fun, flashy consumer apps. And you bet on encoding and enterprise, Claude Code was a huge success, Claude Co-work was also a huge success. Why did you make that bet in the first place? Was that a values-based decision, or was it a business decision?

Dario Amodei: When we founded Anthropic, the most fundamental thing, which is always important, is what is in our hearts. We want to do things the right way. But you have to ask yourself, to finance the extremely expensive creation of these models, the company has to have a corresponding business model. So, does the business model hinder the realization of values?

This issue has always been there. But what I've learned from working at other companies and observing other firms is this: if the business model you choose fundamentally conflicts with your values, you will find yourself in a difficult position. You will either have to betray your values or risk being left behind by the times.

You will eventually find yourself in a dilemma; while there are ways to work around it, it still remains a very tricky situation. Opting for a business model that aligns with your values is much preferable.

So when we think about this issue, we believe that, you see, we have experienced the world of social media and consumer tech, which seems indeed to promote engagement, even addiction. The kind of crude content we see from AI-generated video models is driven by the logic of maximizing the time you spend engaged because there is an incentive driven by advertising revenue.

However, when we look at the enterprise space, you see, our intention was that these models would be beneficial to people. While I often remind people of the negative impacts of AI, fundamentally, we believe that the positive will ultimately outweigh the negative. Many of these positive applications fall mainly within the enterprise domain.

We aim to use AI to cure diseases that were previously incurable, which requires collaboration with biotech firms, pharmaceutical companies, and academic research institutions. All these fall under the enterprise umbrella, as we hope to use AI to make energy cheaper and more efficient. These are all enterprise applications.

We aim to use AI to empower the education sector. Most of this falls under the enterprise-level application space. We aim to use AI to address health and development issues in developing countries. While they are nonprofit organizations, they essentially fall under the enterprise category as well. We aim to drive economic growth. This essentially falls under the enterprise arena.

Additionally, I believe there is another factor, which is that enterprises highly value trust and long-term relationships. Consumer-level applications sometimes have a gimmicky feel, whereas at the enterprise level, it is important to build a partnership where you work with a company for many years, you do what you say, and they do what they say, and they fundamentally trust you.

Therefore, this aligns very well with our goal of deploying these models in a positive and secure manner. So, I think having a business model that aligns so closely with our values is advantageous for us.

Not to say that conflicts never arise or that we don't have to make tough choices, but I think the number of such choices is much fewer compared to other scenarios.

Competitive Moat and the "SaaS Doomsday"

Emily Chang: Developers 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 take 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. Although there is some inertia, I never rely on that, I have never relied on it, or Anthropic has never relied on the idea of "high product stickiness, users won't switch."

I think you still want to have a better model. You want to have a better product. And, we see that the growth rates haven't shown any signs of turning around; if anything, they have increased, at least as we record this interview. So, I think that is the most important thing.

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

Dario Amodei: So, this is one of those questions that is very hard to predict in advance. If you could perfectly predict it, people would have done it ages ago, they would be making a fortune in the markets, and they would always be correct.

So, no one really knows what will happen, but I would point out a few things; all of these traditional software companies have some moats. I think what will happen is that some of those moats will disappear, but some will remain, the ability to rapidly write software, I absolutely think that will disappear, if your moat is "we wrote this kind of software that nobody can write," then, good luck to you. You will not be able to defend that.

But I think people have customer relationships. People have domain expertise on how that field works. People have unique domain knowledge. So, my advice to all those people is clearly, don't be complacent. Don't overlook it. List all of your moats and be very clear that some will disappear, while others will become relatively more important because they are the constraining factors. And there may also be new moats emerging.

I believe those who can adapt agilely, relying on existing moats and building new ones, will thrive. I think those who are complacent are deluding themselves, thinking that what worked in the past will continue to work in the future. They will not have a good time. So, that's the advice I would give.

Furthermore, I think fundamentally, I guess—it depends on how you define SaaS, and what you consider not SaaS—but I speculate that the software industry will grow larger, not smaller, although there will be some significant losers.

Emily Chang: Please elaborate on that.

Dario Amodei: I just think that the pie is getting bigger, like, I think with AI, the pie is getting bigger. Established incumbents may shrink relatively. The value of some of these companies might decline. Some might even go out of business if they can't adjust correctly.

But I think that when the growth rate is extremely high, you often see this situation, if the potential of AI grows by 10 times, then it is very easy for existing industries to grow by 1.5 times, it's just that their growth is not as much as the entire big pie. So I think that could happen.

This is not to say that we won't see some big losers. I think those who cannot adapt, bury their heads in the sand, fail to see future trends, and cannot identify their own moats will find themselves in a very tough spot.

Compute Power, Funding, and Partnerships

Emily Chang: Your biggest supporters are companies like Amazon, Google, Microsoft, and NVIDIA. These companies all have their agendas. They are both partners and competitors. You have significant business milestones tied to funding. Who is really calling the shots?

Dario Amodei: We have indeed been outspoken on several occasions. I have been very clear in advocating for export controls on chip technology to China. I say this because I think it would be very bad for the U.S. and global progress if China were to excel in AI capabilities. And some chipmakers clearly do not agree with this view. But that has not stopped me from expressing my opinion.

Even after we signed more partnership agreements, I am now reiterating this point. What they understand is that we have always maintained cooperation with them. We have always been good partners. We can work together. I'm sure they would prefer us not to say these things, but these are indeed what I believe.

What are your plans? In the end, they will still reach an agreement. The benefits they have gained from these transactions are comparable to ours. You see, we are all adults here. We can collaborate on one thing while retaining differences on another.

Emily Chang: Bloomberg reported that your valuation is even higher than OpenAI's. We are talking about a five-year-old startup with a valuation close to $1 trillion. How do you interpret this? Regarding that figure, and since you are more disciplined in terms of computational resources and have a faster path to profitability, why do you still need so much money?

Dario Amodei: The expansion rate of computational resources is very rapid, so while the business fundamentals look good, a year from now, your computational resource scale may be three to four times what it is now—I won't give an exact number, but the growth of this type of computational resource is very fast.

We have every reason to believe that revenue growth will reach and exceed these scales. And fundraising is precisely a hedge against this range of uncertainty.

So, this is completely a rational approach. The equity dilution caused by this is very minimal to the business, and logically, the two are totally unrelated. In fact, it is compatible with the opposite side, which does not mean there is any problem with the business fundamentals.

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

Dario Amodei: Regarding computational power, one point is the existence of so-called marketing computational power, so, in my opinion, after a period and even longer than a few months, we can access a large amount of computational power. It is worth mentioning here that I don't think by any reasonable standard, we have purchased too little computational power.

So, our original plan was to increase our computing power tenfold annually.

A tenfold increase each year was what we anticipated. However, this is not what we are currently observing. In the first quarter of 2026, our quarterly revenue growth exceeded threefold, and this is only data for one quarter, not annualized growth; tripling within a quarter, of course, to the power of 3, means that the annual growth will reach 80 times.

We did not anticipate an 80-fold annualized growth. Planning for an 80-fold annualized growth is not rational because it means that if you only achieve a tenfold growth, you are essentially missing out on 8 times. So, we are in a phase of localized extreme, computing power explosion. This situation will not continue.

If this situation continues, by the end of the year, you will have a revenue figure that no company on Earth can reach. I don't think this will happen. It's just, it's really not possible to sustain.

But you might experience this brief period where you will be amazed, oh my, this growth rate is much faster than we have ever anticipated. But I don't know, you have seen the computing power agreements reached with Google, and also the computing power agreements reached with Amazon. We can and will do more.

For example, the market is fluid. If you can effectively utilize computing power and there is demand, you can obtain the required computing power. It might only take a month or two.

Leading the Peak Race and Company Culture

Emily Chang: How does it feel to surpass your archenemy?

Dario Amodei: Listen, we still have many tough challenges ahead of us. We have a concept of "leading the peak race," that is, trying to drive other companies along with us. I believe we have seen that we have indeed driven them. Sometimes they do not admit that they are doing this.

Sometimes they are attacking us while also imitating us, but this driving force is very valuable.

So I think, whether from a business perspective or from a modeling perspective, the value of becoming a leading industry company is not about defeating competitors for the sake of competition. Its significance lies in being able to lead the entire ecosystem to develop together, and we hope to do more in this regard in the future.

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

Dario Amodei: You see, we always strive for success, don't we? Just like we have always been working hard, we are not here to fail, I am not the kind of person who thinks we should stop this technological development, shouldn't build it. We operate in a free enterprise system, and that's not a problem. We just have to mitigate the risks posed by the model, so it's always about finding a balance between the two.

Emily Chang: So, for most of Anthropic's history, you have been at a disadvantage. I suppose it's easier to take the moral high ground when you have nothing to lose. At this scale, how hard is it to stay true to your original intentions?

Dario Amodei: I have to say, I have spent a lot of time thinking about why this is.

With the company's continued expansion, I have remained vigilant at every stage of growth. At every stage of the company's development, new challenges arise. The company may lose its essence in some new way. Whether it's business ambition or its core values.

I am concerned about both because I see them as intertwined.

In fact, I believe our ability to build such great models is key to being able to embody our values effectively. As the company grows, there are many pitfalls. There are many ways things can go wrong, not because my, the co-founder's, or the company leaders' values have changed, but because the makeup of the company changes very rapidly.

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

If you don’t tell them how Anthropic operates, they will just default to what they know, which is how things were done at their previous company.

Therefore, this is an ongoing struggle, but also an ongoing challenge.

It's like, Daniela and I, perhaps our most important primary task is to figure out how to maintain this. Because we realize that, in the long run, this is our core.

Research Efficiency and Scientific Progress

Emily Chang: The speed at which your product iteration happens is amazing. The quantity and speed at which you release products are unbelievable. How do you achieve this?

Dario Amodei: I would say two things. The first is that we are a unified company. We have a unified corporate culture. I think we've maintained extremely high efficiency as we've scaled. Everyone remains aligned, which reflects the uniformity in culture and organization. I believe this is the most important factor.

As for the second major factor, I would say it's Claude itself, we are now using Claude to provide assistance. Using it to help us develop models, improve model efficiency, and rapidly develop products. This requires developing various new best practices. While we are still new to this, it has already led to a significant acceleration and this acceleration is becoming more and more reliable. Those are the two factors I would point to.

Emily Chang: Can you tell me the craziest thing AI has done that you've seen?

Dario Amodei: I think the craziest things I've seen have mainly happened in the field of biology and medicine. I've seen several cases, including Daniela's case, where Claude diagnosed medical issues that many top doctors had missed. In biology, these models are starting to show astonishing performance in tasks like drug design, computational chemistry, and more. As a former biologist, I look at these achievements and can't help but be amazed; this is really hard.

It's important to note that accomplishing these tasks requires extensive specialized training. And Claude is becoming increasingly proficient in this area. This is an area I believe we will reap huge benefits from. That's the upside; we're going to get these enormously huge benefits. Life will get better. The quality of the human experience will improve.

Emily Chang: A Century of Scientific Progress.

Dario Amodei: A century of scientific progress, as well as a century of progress in the human experience. For example, let's go back to the year 1900. Think about all the problems we faced in 1900, all the reasons people died prematurely, all the hardships they had to endure, and all the material scarcity we no longer have to face today.

Now, imagine the next hundred years. I wholeheartedly believe that this century's progress in science and medicine, if we can overcome the current challenges—I believe we can. I am becoming more and more optimistic. We will have a much better world than we have now.

Writing, Thinking, and AI Assistance

Emily Chang: I know how passionate you are about writing. You are renowned for your articles. Would you use Claude for writing assistance?

Dario Amodei: I would use it. I haven't reached the point of directly adopting text written by Claude because I have my own unique style, and I am quite particular about it. But I basically use Claude to assist in brainstorming, help me organize topics, or provide some content I can refer to.

So it plays a supporting role. I don't know how far we are from Claude being able to write better works than me. We haven't reached that stage yet, but I believe that day will surely come.

Emily Chang: I also love writing, and I believe writing can help you clarify your thoughts. This involves a lot of critical thinking. If we let Claude do the work, will we lose this ability?

Dario Amodei: I have some concerns about this, and in fact, this is also half the reason I insist on writing myself. It is true for the external audience. Although many people read my articles, writing is also about clarifying my thoughts, knowing what to do next, and creating a common reference point between me and others.

I believe we are still exploring how to use AI in a way that preserves these benefits. I think what I am currently doing is just that, such as using Claude for research and using Claude to help structure my own thoughts.

I think if we were to fully rely on it, like having it write an article on AI risks, firstly, it would not reflect my personal insights, and secondly, I would indeed miss out on the benefits of writing. As model performance improves, I imagine there might be a way in the future to more directly incorporate them into writing while still maintaining these benefits, but it will be a delicate process. The situation will not present itself as a straightforward one, and we must gradually navigate it in the time to come.

I think we may be heading towards a very unusual combination, where high GDP growth coexists with high unemployment, or at least underemployment, along with a significant presence of low-wage jobs and severe inequality.

Unemployment Risk, Productivity, and Macroeconomic Policy

Emily Chang: You have always been very outspoken about the issue of unemployment. AI may eliminate half of entry-level white-collar jobs in the next 1 to 5 years. That was the statement from last year. The pace of AI development is astonishing. Is it still 50% now, or has the proportion increased?

Dario Amodei: I have always said that if you look back at those original video clips, they are always cut to those three seconds out of context. But in reality, my true statement has always been: I don't know what the future holds, but the magnitude of potential disruption this implies is significant.

Moreover, I have also been discussing what actions we can take. I have mentioned token taxes and collaborating with businesses to adjust workforce allocation. I am somewhat skeptical of retraining programs, but we should indeed incorporate them into consideration for macroeconomic policy.

From the beginning, I have always discussed solutions, but there seems to be a tendency in the human psyche to always clip out those three seconds about doomsday scenarios. Therefore, the message I want to convey is definitely not about doomsday. What I want to say is that this is something we should anticipate and be concerned about, and we do need to address it with a positive attitude. I am not entirely certain, but I remain quite concerned. My level of concern remains at the same level.

We are now seeing AI improve people's productivity, but this is part of the typical early-stage growth. If you look back at the Industrial Revolution, I wrote about this in Adolescence of Technology. You automated 90% of the work, which is great.

People's productivity on the remaining 10% of work increased tenfold because they got a tenfold leverage effect. But eventually, it will approach 100%. So, the associated issue is that you have to find other things for them to do.

As for long-term development, I am not sure yet. I am uncertain about that. But I do think there are different types of adaptations. For instance, one thing I talk about is the Anthropic internal software engineer. We are currently going through this transformation, where AI has written all or almost all of the code, but AI has indeed made software engineers more efficient.

However, we have begun to see a faint hint that for some people, it may not have made them more efficient, and in that case, having AI do the work directly may be more effective.

So, this is one aspect of the issue. On the other hand, where do we need to increase demand? We call it frontier deployment engineering, or similar to AI Solution Architects, whose work combines technical development with client communication. Currently, there is a huge demand for such talent because the customer base is large, and we are in a period of rapid expansion.

So, does this fit every person in a purely software engineering role? It is not a perfect match. It is not a one-to-one relationship. It makes you aware of the great disruption that is coming in the future, but things will also adjust. Which side will prevail? I do not know.

But the reason it is necessary to sound the alarm is that this is how we respond. This is also the basis for our policy-making, whether at the level of human individuals or at the global macroeconomic level. We want to put forward carefully considered viewpoints. We do not want to make promises that people think cannot truly be achieved or fulfilled. We do not want to make unsubstantiated statements. We want to think carefully about these issues and explore practical responses.

Emily Chang: You showed this chart that displays potential job displacement, such as in sales, finance, and other fields; which jobs will disappear, who will be replaced, and what new jobs will be created?

Dario Amodei: In reality, no one can be certain because the economy is unpredictable, much like the stock market. They belong to the kind of decentralized processes where you can't truly predict in advance what part of the future job landscape people will still be able to do.

But what I would say is that, in a broad sense, in any field that has entry-level white-collar work, whether it's banking, finance, or any other field, AI first has enormous potential to enhance people's job efficiency, but then AI will have the capability to fully replace human labor in completing the entire job.

Next, we need to think about what people can actually do. I think we need to plan for this in advance. We are already doing this when we engage with our corporate clients. We see the choices they face. The choice they face is, should I save costs?

This usually means reducing employees, essentially doing the same with fewer resources, or should we do more with the same amount of resources? Whenever possible, we always try to encourage them to do more with the same resources, because this basically means hiring the same amount or even more people to engage in some new work, thus driving them towards a positive-sum game.

The favorable condition we currently find ourselves in is that the pie is going to expand significantly. Because the pie is going to expand significantly, there are likely to be new areas for people to enter, the only question is whether they can find these opportunities fast enough. The scale of this disruption is going to be massive, and that is precisely where I am cautioning people to pay attention, but we need to address this matching problem.

Emily Chang: Can you chart a course for me, like when you wake up in five years, what does this country look like? What are these people doing? Because if the unemployment rate reaches that level, isn't that the harbinger of a revolution?

Dario Amodei: That is exactly the result we want to prevent, we absolutely want to avoid that scenario.

I think there are several entry points. These are all not guaranteed, we can't be sure, but after all, there is still the physical world, such as things that exist in the physical world, the robotics revolution is also happening, but it is much slower than what is happening in the field of AI. People always talk about building data centers, but when it becomes much easier to deal with any kind of information, the limiting factors may shift to things in the physical world.

Therefore, we need more people to engage in manufacturing, building, and production in the physical world.

Anything human-centered, I think that will be a big deal. I've heard many stories about AI discovering issues that doctors couldn't find, and I'm glad, but people still want to interact with other humans, especially when dealing with important matters. Perhaps AI can provide better customer service, but even so, people, or at least some people, still want to communicate with humans.

So I think these relationship-oriented jobs will become very important, and I think humans will make some effort to guide AI, to some extent, it has to align with someone's values and intentions. So I think there will be some role in this regard, although I don't know if this role is shallow or profound. I think it's hard to say.

Pushing Back Against "Doomsday Marketing" Accusations

Emily Chang: There have been a lot of voices of opposition so far. I know you said you are trying to alert people, but Jensen Huang said you are blurring the line between mission and work. Others have also said that this seems more like a kind of "doomsday marketing" that benefits Anthropic. So I want to make it clear and strongly refute this.

Dario Amodei: On the overall situation regarding the risk of unemployment, I have some thoughts. I mean, we haven't fully fleshed out these thoughts yet because I want to make sure they are accurate. But Anthropic has already put forward many proposals. We have set up economic subsidies. We have an economic index. I'm talking about possible ways to address these risks through taxation and macroeconomic policies. About what new work is.

During the nascent stages of Technology, I elaborated, I have about 5 pages specifically discussing the difference between tasks and jobs, why this time is different, and listed 6 things we can do, from private charity to government action. I discussed the issue and also discussed solutions.

But social media, I abhor, as a category I abhor, people will only take a three-second clip from a year ago, they haven't really read the article, or they exploit the features of social media - and I have already very rigorously discussed these risks in the article.

The idea that this is just cheap marketing itself is a form of cheap marketing. It's lazy. It's a failure to engage with serious ideas. I think it's also part of the problem. Again, I think it's part of the malaise of Silicon Valley. It's stuck in this three-second social media world.

So people only react to this content, or they think they only need to react to this content. Again, I think it's very dangerous, and we have failed to have a mature conversation.

Instead, people just lazily look at these three-second clips. And then they'll say, 'Oh, that's what Dario said at the time. That's so stupid. That's not serious.' Whenever someone says that, it discounts my opinion of others.

The Red Line of National Security and Military AI

Voiceover One of the world's leading AI companies is deeply embedded in various aspects of U.S. national security, including military operations. Anthropic is in a standoff with the Pentagon over AI military security measures.

Emily Chang: You have long held an anti-war stance, tracing back to your time at Caltech. However, you are one of the first AI companies to sign contracts with the Department of Defense, operating on the U.S. classified network. Used for waging war. Explain.

Dario Amodei: Alright. What I want to say is, you see, the world is constantly changing. Just as my view of this technology, I am concerned because we are facing a resurgent authoritarian group that is very aggressive, and we need to defend ourselves.

This is what I have always believed, and what I still believe, and that is why, during both administrations, while I may not agree with every policy, this is why we broadly support this.

So that's why we work with them. We certainly didn’t do it for the money. It's very tricky, leaving aside the legal battle, to deploy on government networks doesn’t pay much and takes a huge amount of effort. So we do this because we have a sense of mission. But at the same time, since we engage in this work out of concern, there must be limitations on the use of this technology.

[In discussing stages of technological development, I have previously made the following point: we should leverage this technology in all aspects except those that undermine our own values, and the red lines we have drawn, such as mass surveillance and fully autonomous weapons, are precisely what I believe would undermine our values.

If these measures are taken, then even victory would be meaningless. This is the balancing point I see and the position we hold. This also explains why we are both among the first companies to collaborate with the Department of War and yet choose to refrain when others are willing to engage in certain behaviors.

I believe you need to take a stand and stick to it. The kind of wavering company stance—from initially claiming not to cooperate with the government to suddenly embracing all government demands—I simply cannot fathom. You should choose your principles and stand by them.

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

Dario Amodei: That's correct.

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

Dario Amodei: We have not worked with ICE through Palantir or any other channel. We do not collaborate with CBP. I don't think we have business in Gaza. We are very careful to limit the scope of our collaboration to what we identify with.

Emily Chang: So, you've set your red lines, with the President banning you from working with the federal government, the Pentagon flagging you as a supply chain risk, and OpenAI seizing contracts you've rejected. What does winning this struggle actually look like?

Dario Amodei: I don't think a private company can win in this struggle. This is not a struggle at all. Anthropic is not trying to win, nor is it thinking in terms of winning or losing. This is more of a debate on how the government should properly use AI. AI is an emerging technology. We do not yet understand where it is reliable and where it is not. We do not yet understand where it can advance our values and where it can undermine them.

Therefore, I think one of the key things to do is to establish precedents for the use cases we believe are good—frankly, most use cases are good—and the ones we are concerned about. As I mentioned, we have seen that the power of a contract is ultimately limited, as we have seen others might sign a contract that does not respect your same bottom line.

But what it has done is raised awareness on this issue. Currently, Congress is engaged in serious bipartisan cooperation, attempting to ban some of the things we are concerned about and trying to put safeguards in place. Again, I don't want to see this as a fight, but to some extent, it has been effective in prompting our country to think more carefully about how to properly use this technology.

Core Values and Military Ethics

Emily Chang: The operator of Anthropic is an ideological zealot and should not have sole decision-making power. I care about larger AI issues. Do you mind being called an ideological zealot or a bunch of left-wing zealots?

Dario Amodei: I've been called much worse than that. People can call me and Anthropic whatever they want. What really matters are two things: as a company, we have been successful, and we have stuck to our core values. In fact, in some ways, my life is very easy because when you only pursue those two things, it's really simple, isn't it? That way, you're always clear about your position.

Emily Chang: An American official once said that with the help of LLM, the U.S. military has increased the number of daily strike targets from 1,000 to 5,000. This means Claude can help kill more people more quickly. Does that reassure you?

Dario Amodei: I think there are two things here. First is the U.S.'s ability to enhance efficiency in the military realm. I support that ability. I believe having this enhanced capability does not lead to war but instead serves as a deterrent to war. Basically, you're asking, do you believe in this country, do you want this country to be a more potent actor rather than a weaker player on the world stage? I do. I'm a patriot.

There is also a separate issue, which is, are there some government policies involving the U.S. government that I might support or not?

Regarding government involvement, it is clear that I support some aspects and do not support others. This is not up to me to decide. If we provide a technology, and the DOD has also put this forward, then we indeed agree with their view. If we provide a technology, we do not have the authority to decide which military actions you can or cannot take.

Now, personally, I might think that certain military actions make sense while others are a bad idea, but we will not deny the role of this technology.

You must leave policy-making power to military decision-makers. What you can do is set some high-level boundaries, for us, it is to prevent scenarios that go against our values and our country's values, and promote scenarios that we believe help promote our values. That's our way of thinking.

Emily Chang: Bloomberg reported that the U.S. military used Claude during the Iran War, leveraging the Maven Smart System platform created by Palantir for AI-assisted target location. It was reported that in February, a U.S. missile hit a girls' school in Iran, resulting in over 150 deaths, most of whom were children. Did Claude play a role in that attack?

Dario Amodei: We, look, we cannot access, we don't precisely know how these models were used. Clearly, these tragedies that occurred in war are really, really awful, it's truly a very terrible thing.

If this doesn't explain why we insist on opposing use cases we do not support, we are even willing to risk harm to the company's future to restrict the use of these models. Furthermore, the specific use case you mentioned didn't even cross our red line.

We are concerned that there could be an additional 100 times more use cases in the future, some of which do indeed cross our red lines. Now, again, I think overall, the use of these models is appropriate. I think it's good from a net impact perspective.

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

For example, we can discuss a few things. We can discuss how to establish red lines to prevent the misuse of models that are more likely to lead to these issues, if we had initially allowed, unconstrained, if we, like almost all other companies now, succumbed to fully automatic weapons, as seen here, where Claude provides assistance, but a human makes the final decision.

So it is a human making the final decision, not Claude. Imagine if in a world—no Claude, because we did not allow it, but AI models from others—decisions were made directly by the AI model and humans were never involved in the process. This is what we insist on upholding.

That is precisely what we resist. I mean, there's another thing here. Again, I don't think procurement is the right way to achieve this goal. But we do need to ensure—not only as a technology provider, but also as concerned American citizens—that the U.S. military decision-makers do not make these mistakes.

Ensure that they are reliable and make wise choices when taking action.

Similarly, as a citizen and a technology provider, this is also a concern of mine. For example, the government uses Microsoft Excel extensively. If you say, you can use Excel for one military operation but not another, that is not feasible in reality. Hopefully, this gives you an understanding of how we are approaching this issue.

Emily Chang: This school has a website that you could have found in a Google search, shouldn't Claude have found that? Or shouldn't AI or any technology they are using have found that? Does this raise a more concerning issue of using technology as a shortcut in warfare?

Dario Amodei: Listen, listen, what I want to say is, I'm not sure, I really am not sure. This depends on, perhaps, information that is beyond what I know. But the principle we have established, and I think the principle being followed here, is that a human makes the final decision. I don't know what role Claude or any other AI played, but if that doesn't explain why that principle is so important, I don't know what will.

Emily Chang: Will the AI ​​war be more likely to prevent the Third World War, or is it more likely to trigger a war?

Dario Amodei: Overall, I think it is more likely to prevent war. But if we have no restrictions on its use, then I think it is more likely to lead to war. Have you seen Dr. Strangelove, the premise of this movie is that you have some kind of doomsday device that will automatically launch nuclear weapons for retaliation once it believes it is under nuclear attack.

What could possibly go wrong, on the other hand, I think of topics such as lethal, fully autonomous weapons. I think the way conflicts arise is when both sides are watching each other and waiting for an opportunity. They will misunderstand each other. If we lack proper supervision of this technology, I think the likelihood of such accidents will increase.

Now, I think if AI can be properly used, even not limited to warfare, just considering the field of intelligence gathering. Imagine when we have insight into the opponent's every move, they would think twice before attempting any invasion or military action. Therefore, I think excellent intelligence capabilities can indeed prevent conflicts. Outstanding response capabilities can contain conflicts. I always believe in these things.

Mythos: Security Game of Advanced Models

Emily Chang: Anthropics is almost in the headlines every week. Of course, the most attention now is on Mythos.

The narration is that this is the latest and most powerful Anthropics model, which can autonomously traverse all links of the kill chain.

Emily Chang: You have said that Mythos's capabilities are too powerful to be released to the public. What about it surprised you the most?

Dario Amodei: I think what surprised me the most is that the model's ability to discover vulnerabilities has been continuously increasing, and more importantly, they can also translate these vulnerabilities into exploits—and people usually only discuss the vulnerabilities themselves. People often don't talk about the process of turning vulnerabilities into exploits, and it excels in this aspect.

So, what I find surprising is that we have witnessed this huge leap. This is a particularly large leap. And in the absence of any guidance, some of the companies to which we originally provided this technology have said that it is nothing short of a superweapon. You should have to hold a firearms license to be able to use it.

Please do not publish this. This call for restricted use is coming precisely from those companies to which we provided the technology, as they have discovered a significant number of critical flaws and the exploitability around those critical flaws to the point where they are basically asking us not to publish it.

It should be clear that because information in the realm of social media is always distorted, our goal is not to keep it under wraps forever. We are trying to progressively open up this technology. To a wider and wider audience. Ultimately, we believe that the mythos should be released to the public, but with robust cybersecurity measures in place.

One current concern is that we released the current cybersecurity measures with Opus 4.7, and while it is a great cybersecurity model, it is significantly weaker in its capabilities, and these measures can be jailbroken. We are somewhat concerned about other companies that view this as a sufficient defense because while it may be effective at times, we all know that these classifiers can be jailbroken or circumvented.

Based on our own testing, frankly, based on our assessment of the defensive measures taken by other companies, these defenses are currently not strong enough. This is exactly what we are waiting for now, to bring the defense to a level that we are truly confident in.

Emily Chang: There was a lot of pushback at the time. Researchers claimed they could replicate this process using cheaper open-source models. Some said that OpenAI already had these capabilities. So, how do you respond to those who think this is a large-scale PR stunt?

Dario Amodei: Regarding the claim that this could be replicated using open-source models, that is entirely inaccurate. The core idea of the system is that Mythos is able to look at the entire codebase and find issues. Someone on Twitter said that if you have an open-source model specifically point at the line of code that Mythos found, it would find the same issue. It's not, it's not a cue, it's not that problem, like, it's not, that's not the same thing at all.

The ultimate test of this was when we reached out to companies, we went through open-source codebases, and we found 271 new vulnerabilities in Firefox. Among private companies with vulnerabilities either not yet patched or not yet disclosed, we have found thousands of vulnerabilities.

It's like, in the previous model, no one found those 271 vulnerabilities, so the real-world effective workflow is quite different from the "I found the exact line of code Mythos found, I know I found a needle in a haystack, and now someone else can pick up that needle" situation.

But for those who say this is just a marketing ploy, I would say, we have suffered a huge loss in not releasing this model commercially. This model significantly accelerated Anthropic's internal research and subsequent model production, and if we released it, it would have had the same accelerating effect on the external world. This has hurt us significantly financially.

Emily Chang: If this can help defenders, it will also help attackers. Can we defend anything anymore?

Dario Amodei: What I want to say is that the reason we offered Mythos to attackers after offering it to defenders is to patch all vulnerabilities. I don't know, as the model gets better, more vulnerabilities may be discovered, but the number of vulnerabilities is ultimately limited. They are limited, just like you have this surface, and there are only so many holes in it.

If you patch all vulnerabilities, the surface becomes very difficult to attack; at the same time, the code itself is written by powerful models, making it difficult to find flaws or intrude. So I think in another sense, hopefully in 6 months or a year, we can have a more secure internet ecosystem than in the past. We are working hard to achieve that goal and making every effort to open mythos to new network defenders.

We have been in communication with the government. We greatly respect their advice. Due to concerns about counter-espionage risks, they are slowing down the pace at which we are opening up. Risk. I think this is wise. I think everyone serious here understands that there is a real trade-off.

We see a lot of sharp criticism of this on Twitter and by other AI companies. You look at their statements and then at their inconsistent actions. They are not serious people. They have not seriously considered the serious trade-offs we are facing here.

You see, every day, I have clients calling me saying, I want to use mythos. I also have countries calling me saying, I want to use mythos. And the US government and my security team tell me, no, hold on, there is risk involved. I'm not saying either side is right.

I believe the truth lies somewhere in between. Both sides have valid points. Both sides have valid points. But there is a real challenge here that we need to collectively address as a society, rather than dismissing things as mere cheap marketing or, as some other companies do, using a Chief Marketing Officer to attempt a repositioning.

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

Checks and Balances vs. Government Regulation

Emily Chang: Have you had to make some trade-offs that have made you less comfortable?

Dario Amodei: The entire history of Anthropic has been a process of trade-offs, the entire history of Anthropic, in some ideal world, you would spend years researching every potential failure point before releasing the first chatbot. Now, we have indeed delayed.

We did delay the initial release of Claude, but we only delayed it by a few months. So I would say everything is a matter of balance, being at the extreme ends of the spectrum is completely insane, so everything is a trade-off.

What I would say is, since we are now in what I described as a business leadership position, in fact, Daniela and I are doing our best to further advance in acting prudently. That's the purpose of the mythos release, and if you're not a leader in the industry, it's hard to do that. So I think you'll see more of these things happening.

Emily Chang: There is a viewpoint that asks, why doesn't the government take over you guys? Why should a private company control such powerful technology?

Dario Amodei: So I do think that's a very serious issue, and I share those concerns. I don't think the government should take us over directly, but I can put it this way. What I'd like to say is, looking back and describing the current state, every powerful technology we've seen in history has either been built by the government or has its origins in the government.

For example, nuclear weapons were obviously initially built by the government and have since been mostly government-led. But even technologies like the internet, GPS, and cell phones, all R&D was done in labs, federal labs, and universities.

AI is the first major technology built in the private sector, with minimal government involvement and a relatively delayed entry. I think this is actually a dangerous and unstable situation. It's not a situation I would opt for. There's no alternative; the technology is going to be built. Our competitors are building it, there's economic value to it, and at the end of the day, it will get built. The problem is not that the private sector is doing it, but that the government isn't involved.

I think we need to consider a balance of power, so I think there needs to be a balance of power at AI companies. We have that mechanism, the Long-term Benefit Trust, which is essentially a body that can appoint and remove a majority of the board. If you follow the logic, it effectively has the power to fire me.

What we're introducing now are some elements—certainly not all elements—we're introducing some elements of public governance, where you have to be accountable to someone, not just the shareholders of the company. This is very important, and this structure will persist regardless of what happens to the company. In the AI field, we encourage other companies to adopt similar structures.

On the governmental side, I think we need a balance. Congress has announced some 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 concerned about both companies having it and the government having it.

Subsequently, businesses need to balance the government, and the government needs to regulate businesses. We need fundamental regulation of this technology. I think we need to start implementing pre-release testing, mandatory pre-release testing, and testing and auditing of the models.

What I find most ironic is that in Silicon Valley, there's a group of people in tech who initially took the stance that even attempts to make this technology transparent or to implement export controls would completely destroy our ability to create this technology, stifling innovation. However, once they saw the first real danger I had been predicting all along, things changed. They started talking all about nationalization, suggesting that the government should take it over directly.

Give it up, everyone. Your attitude has flip-flopped back and forth, from the most extreme anti-regulation stance—such as 'if you even dare to look at us, you're destroying this industry'—to a complete communist approach of 'the government should take over everything'. We need a more rational, more moderate approach.

This is exactly what we have always advocated for, as we deeply understand the power of this technology. We are not in a panic. We have not denied that. We have seen smooth exponential growth and are responding accordingly.

So, how was your visit back to the White House? We always try to work with anyone who is willing to collaborate. On a government level, I have said we have a simple approach. We have a set of principles. We stick to these principles and hope the other side can be reasonable. Frankly, the government does take mythos seriously.

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

So, like any administration, there are some departments we get along with very well, and they understand this. Of course, there are also some departments that are more difficult to work with. I think that's normal. This situation occurs in any government, and we just do our best to deal with it properly.

Intellectual Premium, Geopolitics, and the Future Landscape

Emily Chang: Early in your career, you worked at Baidu, a large Chinese tech company, at its Silicon Valley branch, and you have always had a clear view on China. China is rolling out powerful open-source models, while U.S. companies are freely developing on top of these models. Is this a threat?

Dario Amodei: So, one thing we see in this technology is that there is indeed a premium on model intelligence. We rarely see people leaning towards using models with lower intelligence. To be clear, there is already a thriving ecosystem. There are many challenges and problems that are much simpler than the ones we need cutting-edge models to solve.

But it is important to emphasize that this is an exponential growth, meaning that these non-cutting-edge models may have economic value comparable to what we saw in 2023 and 2024, but remember, we are growing at a rate of 10x per year.

So we find that something at the cutting edge is always much more powerful than something far from the cutting edge. I think this is not particularly understood by those who are used to building products in the old days, as someone who has never run a company before, never thought about the previous product era, and particularly despised the social media era.

I feel like an outsider to that world beyond. And I think people's intuition is wrong. They hold various experiential rules about products. I think the 10x exponential growth model does break that. Like intelligence is an extremely critical factor, its weight outweighs everything else. So we see time and time again that value often lies in the cutting edge.

Now, for certain lagging models, my concern is the risk they bring, that we face Mythos-level network attack capabilities. 12 months from now, we will have a more powerful network, but Mythos-level network attack capabilities may be available for download by then. I hope we have patched everything up by then. I don't think we have a way to stop it, but I think it's a serious issue.

Emily Chang: Has your experience at Baidu shaped your view of China?

Dario Amodei: Not really, no.

I worked there for a year. I think I probably learned more about things like speech recognition. Perhaps the only thing that worries me is that part of how we get speech recognition data — in China, they don't care about privacy, so we have all this speech recognition data.

But apart from that, my concern here is at the geopolitical level. I think the most worrying thing about what's happening in China is what we see in the U.S., where we see the suppression of critical voices, even in the U.S., similar to what's happening in China, like the Hong Kong issue or the influence on American reality.

Business networks and suppressing criticism, that's an authoritarian state. And it's a high-tech authoritarian state. When I see this combined with AI, you do get a dystopia here, like 1984 or worse.

My focus is on trying to prevent this scenario. I believe we have an opportunity to stop it.

I believe we have the opportunity to make AI a liberating technology, one that makes humanity more free, able to fulfill the promise of equal justice for all, or it could go the other way. Where AI goes next depends on the actions of AI companies, on the actions of governments, and on the actions of all of us. So, I think we have a responsibility in this regard.

Self-Iteration and Civilization Continuity

Emily Chang: People in your field often talk about a moment like this: when AI becomes powerful enough to improve itself, and then the improved version continues to iterate on itself, and so on. Some of your researchers believe that this moment is near. How far are we from this moment?

Dario Amodei: I don't think this is a specific moment. I think it's an ongoing process. In some ways, we've already seen this, where AI can suggest architectures for the next generation of AI. I'd say a year ago, we saw total factor productivity increase by 10% to 15% due to AI. For example, that number might be up to 20% or 30% by now. It might even be doubling.

As with all things, we are on an exponential curve, and there's no moment where AI suddenly self-evolves, goes haywire, or becomes unsafe.

What we are experiencing is a continually accelerating exponential growth. And at each point on the exponential curve, we have to evaluate: Is it time to slow down now? Is it time to exert more control over this technology now? I think increasingly more control will be needed in the future. But I think the key to all of this is that smooth exponential curve.

Once again, I think those who were once against all AI regulation, then advocated for nationalization upon seeing a certain phenomenon, taught us a lesson. I think those who used to underestimate the power of AI, and later exclaimed that AI is self-evolving, about to spiral out of control, and demanded a complete halt, also taught us a lesson.

Jumping back and forth between these extreme reactions is extremely unhelpful in dealing with this technology.

The right way to respond, the wise way to respond, should be to remain calm and not panic. Our responses will escalate in power along with the technology. If you see someone having this kind of crazy swing reaction, it means they are caught off guard, and they are not serious.

Emily Chang: From what I understand, one of your favorite books is "The Making of the Atomic Bomb."

Dario Amodei: That's correct.

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

Dario Amodei: The person I most identify with is actually Leo Szilard, who was the first to conceive of the possibility of a chain reaction.

Listen, my take is, we can't resolve this issue by looking up to those larger-than-life figures, or trying to be the central person in everything. What is needed here is a balance of power where you have many powerful players with their own interests, and the only way to get a good outcome for everyone is to put in place some mechanisms that establish a basic equilibrium everywhere.

So in a way, I actually view Oppenheimer as a cautionary tale of what not to do.

Emily Chang: You've mentioned before that the likelihood of a collapse of civilization is around 10% to 25%. That is not insignificant. Is there a scenario where the causes of this outcome are Anthropic in nature?

Dario Amodei: I certainly hope not. My view is that the actions we take reduce this probability rather than increase it, and this probability is very intuitive based on the nature of the technology itself: there are many countries in the world, many companies in the economy, and if there is a vacuum, it will be filled, so that's the conundrum we are facing.

We are striving to take actions that lower this likelihood. I think we are lowering it much more than we are raising it.

But the inherent property of this technology is unpredictability. So we do a huge amount of building and testing before we release a product, and I think the models we have released so far are not dangerous, at least not outside the realm of cybersecurity. And then we try to iterate and learn from that. So we have countless defenses, and half of the work inside the company is about minimizing risks as much as possible.

However, the risk can never be reduced to zero.

What I mean is, suppose there are many airlines on the market, and you decide to start an airline that is ten times safer than all the others. But if someone were to ask you, can you guarantee that your planes will never crash? How could you possibly ensure that? How could you possibly achieve that?

Emily Chang: But if the probability of a plane crash is 25%, you certainly wouldn't board that plane.

Dario Amodei: Indeed. A 25% probability is too high. We are working to bring this probability down to a very, very low level. That is our goal.

Building Trust: Demonstrating Uniqueness Through Action

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

Dario Amodei: I think when any company is just starting out, especially given the overall behavior we've seen of Silicon Valley as an entity, it's quite natural for people to have this perspective. Looking back over the past few years, I think from the standpoint of not trusting— if you have no reason to trust me or know nothing about Anthropic— that's a fairly rational stance.

I think Silicon Valley has lost too much trust from the world, and we must regain that trust.

The message we want to convey is that we are indeed unique. And this trust must be earned through our actual actions. You can agree or disagree, but we have always stuck to our values. The mythological claim about not releasing this extremely powerful model did indeed limit us commercially.

There were many trivial matters leading up to this, and we promised action on the China issue. We cut off access to the model. We didn't have to. No one asked us to do this. You see, at a time when hundreds of millions of dollars still made up a significant portion of our revenue, this cost us billions.

For example, Claude II's delayed release, we have a long history in this regard.

We are not perfect. We make mistakes. However, I would like everyone to look at our overall history and, if you were to sum up this history, what is the most fitting assumption about us? People must make that decision for themselves.

But I think one consistent assumption is that we are indeed trying to do. To do the right thing.

We are not perfect. Organizations always have issues with their functioning. We have been trying to fix them, make them work better. There have been many mistakes in the process, many places where we went wrong. But fundamentally, we have a sincere and serious vision of how to do the right thing, and we are working hard to execute that vision.

Emily Chang: So, we will meet you at the other end of exponential growth.

Dario Amodei: Well, I hope so.

Emily Chang: Well, I hope so.

Dario Amodei: One thing about the CEO job that surprised me and I didn't know before is that you have to wear makeup often. That completely caught me off guard.

Emily Chang: Just a bit of foundation.

Original Link

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia