Source: Bloomberg

Compiled by: Felix, PANews

Bloomberg recently interviewed Anthropic CEO Dario Amodei, discussing his entrepreneurial journey in San Francisco, competition with OpenAI, and the ultimate goals of AI.

PANews has compiled the highlights of the interview.

Host: You are now at the center of the AI universe. What does that feel like?

Dario: Throughout my entire career, and especially during my time at Anthropic, what I've experienced is a smooth exponential curve. The experience of a smooth exponential curve is: nothing happens, nothing happens, a little thing happens, and then "whoosh" — crazy growth. That's the experience for the world, and also the experience of our company's scale relative to other companies and the world. I've been looking at this chart for a long time, and I once said, "We'll probably become the AI company with the highest revenue and valuation around this point," and it happened. So on one hand, I'm not surprised, because it's just a smooth line on a chart; but on the other hand, when things actually happen, you see more detail and color, and it's absolutely astonishing. We always need to keep in mind the most fundamental questions: How do you train good models? How do you put them into good products? How do you ensure safety? How do you help people while managing the societal risks of the technology? It's all the same questions, just being examined under a bigger microscope.

Host: You grew up in San Francisco. What was your childhood like? I know your father was a leather craftsman and your mother worked in a library. How did that influence you?

Dario: The first internet revolution was happening all around me at the time, but I had absolutely no interest in it. I was only interested in doing math problems, doodling, understanding the universe, and science fiction. That was my whole environment back then; I was full of curiosity about the world.

Host: You grew up in this tech hub city, which is now the center of AI. Did anything about this city influence your worldview?

Dario: The pervasive spirit of non-conformity, individualism, and the idea that "it's okay to be crazy" here definitely influenced me. If you go to Europe or even other parts of this country, you'll find that thinking differently or having some crazy ideas is discouraged or even considered strange. Although I have many criticisms of Silicon Valley, I think one thing it does well is: it's okay if all the experts are against you. If you have a coherent vision for the world, you should pursue it. Maybe it won't work at all, but if it succeeds, it produces a long-tail effect, and you might find a huge goldmine in some vein. I think that spirit is very important.

Host: Your decision to leave OpenAI has become Silicon Valley lore. What really happened? Beyond the rumors, what were the actual issues? Where did you disagree?

Dario: Let me put it simply. When you're developing powerful technology, you face many difficult problems, and Anthropic faces these daily too — problems where we don't know if a decision is right or wrong. So there are many legitimate disagreements on safety, and we did have some of those with them, but that alone wasn't enough reason to leave. Even at Anthropic, people disagree with me and with each other. However, when you feel you can't trust someone, when you feel their values aren't what they claim, when you feel they are dishonest, not acting from the motives they profess, when you see a disturbing pattern of deceptive behavior — it's hard to continue working with and trusting that company. Ultimately, if you have different visions and don't trust each other, why bother arguing? The solution is to go your separate ways. I'm completely at peace with the idea that we do things our way, and they do things theirs. Let's see who wins in the market, who wins in public opinion. That's more persuasive than any dramatic gossip about who left. We are demonstrating an example of how to deploy this technology responsibly. If they disagree, they can make their arguments. I think that's all there is to say.

Host: At the AI summit in India, you and Sam Altman seemed to refuse to hold hands on stage. What happened there?

Dario: The reality is that summit was extremely disorganized. We were all brought on stage at the very last minute, they changed our standing order on the fly, took photos, and then suddenly commanded us all to hold hands. If you've been to these international summits with heads of state, you know they tend to be super chaotic (I'm not singling out India here).

Host: But everyone else held hands.

Dario: I don't know how to explain this to you. It was just Narendra Modi suddenly telling everyone on stage to hold hands.

Host: Okay. Sam and Elon are suing each other. You don't like Sam. If the people building the world's most important technology can't hold hands on stage, how can we trust you to cooperate on existential risks to humanity?

Dario: What I want to tell you is that there's a huge variance in the quality and trustworthiness of the people building this technology. The narrative that "nobody trusts anyone" is wrong. I've known Demis Hassabis, who develops the Gemini models, for 15 years. They are competitors to the Claude models, but we've collaborated on many issues. We buy compute from Google and frequently exchange ideas on safety. So I think, first, some actors are indeed more trustworthy than others; second, there are actors outside of Anthropic that I trust. I think what needs to happen is for the trustworthy actors to band together and put the less trustworthy ones in a position where they have to adopt the same standards. In my experience, some people won't proactively do the right thing on their own, but if the majority of the industry is doing the right thing, others don't have many choices. This splits into two aspects: the positive side is mutual inspiration (i.e., racing to the top), like Demis doing AlphaFold inspired us, we do biological research, our interpretability research inspired them — this isn't a zero-sum competition. That's the positive "carrot" side of the race; the "stick" side is that you look bad if you don't do the right thing. They might very reluctantly do the right thing while pretending to do something else, or even think we have some sinister ulterior motive — that's to be expected — but this is exactly how you get the industry to unite and cooperate.

Host: Early on, others focused on fun, eye-catching consumer applications, while you placed your bet on coding and enterprise. Claude Code and Claude Cowork have been very successful. Why did you make this bet? Was it a values decision or a business decision?

Dario: When we founded Anthropic, the most fundamental thing was that we wanted to do things the right way. But to fund the extremely expensive creation of models, it needed a business model. If the business model fundamentally conflicts with your values, you're in a tough spot: either betray your values or become irrelevant. It's much better to choose a business model compatible with your values. We looked at social media and the consumer world, which often incentivizes user engagement or even addiction (like the chaos we see with AI video models), just to maximize user dwell time for ad revenue. Conversely, on the enterprise side, we want to use AI to cure previously incurable diseases (partnering with biotech, pharma, and academic institutions), make energy cheaper and more efficient, help education and non-profits, and promote economic growth. These are inherently enterprise applications. Another factor is that enterprises highly value trust and long-term relationships. The consumer side can sometimes be gimmicky, while the enterprise side requires you to build trust over many years and deliver what you promise. This is very synergistic with our goal of deploying models in a positive, safe way.

Host: A developer can switch from Claude to ChatGPT or Gemini in an afternoon. Is it really possible to maintain a long-term lead in this industry? How long would it take a strong competitor to replicate what you've built?

Dario: Model quality is the most important thing. We are currently far ahead in model quality. While there is some switching inertia, I (and Anthropic) never rely on product "stickiness" to prevent people from switching. You need to have a better model, a better product. We're seeing no inflection point or downturn in our growth rate at all. At least at the time of recording this interview, if anything, it's still going up. So I tend to think model quality is the most important thing.

Host: Shortly after Claude Cowork was released, $285 billion in market cap evaporated overnight, which traders called the "SaaSpocalypse." If AI continues to advance at this pace, how much traditional software will be replaced, and how fast?

Dario: It's hard to predict accurately in advance. All traditional software companies have their own moats. Some of those moats will disappear, but others will remain. "Writing complex software quickly" as a moat will definitely disappear; you can no longer rely on that for defense. But customer relationships, knowledge of industry operations, and unique domain expertise remain important. My advice to these companies: don't be complacent, list all your moats, see which ones will disappear and which will get stronger, because new moats will also emerge. Those who adapt nimbly will do very well, while those who deceive themselves and rest on past laurels will have a very hard time. Overall, I think the software industry will get bigger, not smaller, although there will certainly be some big losers.

Host: Explain that point.

Dario: Because the "pie" is getting bigger. If AI raises the industry ceiling by 10x, existing companies can easily achieve 1.5x growth, even though their relative share of the total pie becomes smaller. But those who don't adapt, who bury their heads in the sand, will go through tough times or even go out of business.

Host: Your biggest backers are companies like Amazon, Google, Microsoft, and Nvidia. They all have their own agendas, being both partners and competitors. Your huge commercial milestones are tied to funding — who is really calling the shots?

Dario: In many cases, we speak very candidly about what we truly think. We're all adults; we can agree to disagree, cooperate on one thing and have differences on another.

Host: There are reports that your valuation is higher than OpenAI's. For a five-year-old startup, a valuation approaching a trillion dollars. How do you make sense of that number? If you are more disciplined on compute and have a faster path to profitability, why do you need so much money?

Dario: The demand for compute is growing very fast. Business fundamentals are good, but within a year, the compute you need might triple or quadruple — it grows extremely fast. We fully expect revenue growth to meet or even exceed these compute increases. Raising capital is to buffer against the impact of this uncertainty; it's very rational and involves very little dilution for the company. It absolutely does not represent a problem with business fundamentals.

Host: Does it feel good to surpass your main competitor?

Dario: We still have many difficult challenges ahead of us. We are promoting the idea of "racing to the top," trying to pull other companies alongside us. Although sometimes they don't admit it, sometimes they attack us while copying us, this pulling effect is immensely valuable. The value of being the leading company isn't simply in beating the opponent, but in having the ability to pull the entire ecosystem forward.

Host: But winning has to feel a little bit better.

Dario: We're always trying to succeed, of course, and we don't want to fail. I don't think we should shut down the technology. We compete in a free enterprise system, and we just need to mitigate the risks of the models. We're always looking for a balance between the two poles.

Host: Your product development speed is astonishing. How do you do it?

Dario: Two reasons: first, we have a unified culture and organizational efficiency, with everyone moving in lockstep; second, Claude itself. We're now using Claude to help develop our models and products, making things more efficient. This creates a significant and increasingly reliable acceleration effect.

Host: What's the craziest thing you've seen AI do?

Dario: It's all in biology and medicine. I've seen Claude diagnose medical issues that senior doctors missed on several occasions. In biology, the models are becoming astonishingly good at drug design or computational chemistry. As a former biologist, I know that requires an extremely high level of specialized training. This is the huge positive side of AI — it will dramatically improve the quality of human life. Imagine a century of scientific progress from 1900 to now, and then imagine what another century of progress would look like. If we can navigate the current challenges smoothly, we'll have a much better world.

Host: I know you love writing. Do you use Claude to help you write?

Dario: I do, but I haven't let it directly write finished text for me yet, because I'm a bit particular about a specific style. I mainly use Claude for brainstorming, clarifying topics, or providing references. It plays a supporting role right now. It's still a ways off from writing better than I do, but that will come sooner or later.

Host: Writing helps people clarify their thoughts and involves a lot of critical thinking. If we let Claude do it for us, will we lose that ability to think?

Dario: I do worry about that a bit. I insist on writing myself, half for the reader, but equally to clarify my own thinking. If I simply ask it to "write an essay about AI risk," not only will it not capture what I truly want to express, but I'll also lose the benefit of that thinking. In the future, as models get better, we may need to find a nuanced way to use them more directly while still preserving the benefit of thinking.

Host: You've been outspoken about job losses, predicting that AI could eliminate half of entry-level white-collar jobs within 1 to 5 years. That was a year ago. Is that figure still 50%, or is it higher now?

Dario: If you watch the original full video, what I actually said was: I don't know what will happen, but this represents an order of magnitude for how crazy things could get. Yet people's psychology always tends to clip out those three seconds of "doom is coming," while ignoring all the solutions I discussed at the same time (like token taxes, corporate adjustments, macroeconomic policy, etc.). My message is absolutely not "doom is coming," but that we need to see it ahead of time, pay attention, and actively address it. In the short term, AI will make people more efficient, but that's just a transitional phase. Just as we've seen in the history of automation, AI may eventually take over this part of the work. Take Anthropic's software engineers, for example. Now AI helps them write most of the code, making them more efficient, but situations are also starting to emerge where it's simply better to let the AI perform a specific task directly. On the other hand, this also creates new demand, like "applied AI solution architects" (front-end deployment engineers) who combine technical and customer communication skills. Job destruction and adjustment will happen simultaneously.

Host: You released a chart showing potential job disruption — which ones will disappear and which will be created?

Dario: It's very hard to predict accurately, just like predicting the stock market. But overall, entry-level white-collar positions (banking, finance, etc.) are very likely to be replaced by AI. When we talk to enterprise clients, we find they face a choice: lay off staff to save costs, or use the same resources to do more new things? We always try to push them toward a "positive-sum game": using the efficiency gains to do more things, rather than laying people off. As the "pie" expands, people may find new places to go, and it's a question of finding the right matching speed. The scale of disruption will be large, which is why I try to warn people.

Host: If unemployment reaches that scale, isn't that a recipe for revolution?

Dario: That's absolutely an outcome we want to avoid. I think there are a few areas where opportunities will remain: one is the physical world (because robotics development lags behind AI information processing speed, manufacturing and construction will still need a lot of human labor); the other is all human-centered work. Even if AI is more accurate than a doctor at diagnosis or does a better job in customer service, people still want and need to talk to and connect with other humans on important matters. Humans will also retain the role of guiding AI's values and intentions.

Host: Many people disagree with your views. For example, Jensen Huang says you're confusing "tasks" with "jobs," and others say it's a kind of "doom marketing" that benefits Anthropic.

Dario: I want to explicitly refute that. In our reports and papers, we used a full five pages to detail the difference between tasks and jobs, explain why this time is different, and propose six types of solutions covering both the private sector and government. But I find that the malaise of Silicon Valley and social media is laziness — people watch a three-second clip from a year ago and make irresponsible comments. Calling serious intellectual labor "cheap marketing" is itself a form of cheap marketing that refuses serious discussion, and it's an extremely lazy act.

Host: As a leading global AI company, you're deeply involved in many areas of U.S. national security. You've had a long-standing anti-war stance since your time at Caltech, yet you were one of the first AI companies to sign a classified cyber operations contract with the Department of Defense. Please explain.

Dario: The reason is that the world is changing. When I saw Russia invade Ukraine, I worried about the resurgence of authoritarian blocs. We need to defend ourselves. We absolutely do not want adversaries using AI to analyze intelligence and attack Ukraine while we are unable to defend. So we provide support across government agencies. And by the way, we are absolutely not doing this for the money — running on government networks is actually a huge hassle with very little profit. Precisely because we do it out of "caring," there must be clear limits and red lines on the technology's use, namely: no mass surveillance and no development of fully autonomous weapons. If we abandon our democratic values to win the competition, it's not worth it. That's the balance and stance we insist on.

Host: You started working with Palantir in 2024. Their technology is used by U.S. Immigration and Customs Enforcement, police departments, and even in Gaza. Is Claude being used for surveillance?

Dario: We do not work with ICE and CBP, either directly or through Palantir. To my knowledge, we also do not have operations in Gaza. We are very careful to limit our involvement to areas aligned with our values.

Host: So you've drawn red lines, the President has blocked some of your federal government licenses, the Pentagon has labeled you a supply chain risk, and OpenAI stepped in and signed contracts you wouldn't. What does winning this fight actually look like?

Dario: I don't think this is a fight about "winning or losing." It's a societal debate about how the government should use AI appropriately. We are trying to establish usage precedents for concerning scenarios and have drawn attention to the issue. We're seeing bipartisan efforts in Congress trying to ban high-risk applications and set guardrails. If I had to say what winning is, it's prompting our country to think more carefully about the appropriate boundaries for using new technology — that is a victory.

Host: Some people describe you as an ideological lunatic who shouldn't have this much power, or a left-wing zealot. Do you mind?

Dario: I get called names often, and much worse than that. People can call me and Anthropic whatever they want. I only care about two things: that the company succeeds, and that we stick to our values. It actually makes my life quite simple — I always know where I stand.

Host: Is an AI war more likely to prevent World War III, or more likely to trigger it?

Dario: Overall, it's more likely to prevent it. But without usage restrictions, it could trigger a war. Without proper regulation, conflicts could easily erupt from misunderstandings between sides. If we have overwhelming superiority in intelligence gathering and response capability, adversaries will think twice before launching an action. Superior intelligence can effectively deter conflict.

Host: Anthropic makes headlines almost every week, most recently about the Mythos model. You've said Mythos is too powerful to release to the public. What surprised you most about it?

Dario: What surprised me most was not just that it could find vulnerabilities, but that it could turn those vulnerabilities into substantive attack tools — a huge leap in capability. Some companies that tested it early even asked us not to release it, calling it a "superweapon" and saying you'd need a gun license to use it. We're not releasing it to the public not because we want to lock it away forever, but because current cyber defense mechanisms are still easily broken by "jailbreaks." We'll wait until defense mechanisms are confident enough to handle it, and then release it with strong safeguards.

Host: But there's a lot of pushback externally: some researchers claim they can replicate it with cheaper open-source models, and others say it's entirely a grand PR and marketing stunt. How do you respond?

Dario: The claim that it can be replicated with open-source models is utterly absurd. The Mythos model found 271 new vulnerabilities in the Firefox browser and thousands of new vulnerabilities in many private companies that we can't disclose publicly — previous models couldn't find these at all. The person on Twitter who said they could replicate it with an open-source model only used the open-source model to re-identify the line of code after we pointed out exactly which line had the problem. That's not the same thing at all.

As for the "marketing" claim, not releasing this super-powerful model has cost us enormously commercially. It could have greatly accelerated both internal and external R&D. If it helps defenders, it helps attackers equally. So our current strategy is to give it to defenders first to patch system vulnerabilities. Once all the vulnerability "holes" are patched, the future internet ecosystem will become extremely secure. Those sniping and contrarian takes on X (including from some competitors), failing to take the risk trade-offs for society seriously, reflect an extreme level of immaturity.

Host: Have you already made trade-offs you're not entirely happy with?

Dario: Anthropic's entire history is full of trade-offs. In an ideal world, you might spend years studying every hidden risk of a chatbot before releasing it. Although we did delay the initial version of Claude by several months, everything is still a trade-off. Now that we're in a commercial leadership position, Daniela and I are doing everything we can to push the needle further toward "acting cautiously." That's the original intent behind limiting Mythos's release — if you're not the leading player, it's very hard to make that kind of decision.

Host: If this technology is so powerful, why doesn't the government just take you over?

Dario: That's a very serious concern. Throughout history, all unprecedentedly powerful technologies — nuclear weapons, the internet, GPS, mobile phones — were initially born in government and federal labs. AI is the first technology built entirely in the private sector with government involvement coming late, and this situation is actually dangerous and unstable. But right now, the risk of the government not doing it is actually greater than the risk of the private sector doing it. So we need checks and balances on power: we've established a "Long-Term Benefit Trust" that has the authority to remove the vast majority of directors and even fire me, which introduces an element of public governance.

At the government level, we need the legislative and judicial branches to implement mandatory pre-release testing and red-line regulation. I'm afraid of companies controlling it, and I'm also afraid of the government monopolizing it — both sides must check each other. It's quite ridiculous that there's a group in Silicon Valley who went from being extremely opposed to any regulation or even transparency requirements (saying it would stifle innovation) to, after seeing the first real danger, immediately swinging to declaring that "the government should nationalize it." This kind of extreme back-and-forth flip-flopping is very immature; what we need is a rational middle way.

Host: Your field is always talking about the moment when AI becomes good enough to engage in "self-improvement." Your researchers think this is approaching — how far away is it specifically?

Dario: I don't think it's a specific "moment" but rather a continuously accelerating exponential process. We're already seeing AI begin to assist in proposing architectures for the next generation of AI. A year ago, AI improved total factor productivity by 10–15%; now it might be 20–30%. It won't suddenly spiral out of control at a single moment — rather, along this smooth exponential curve, we need to evaluate at every node whether we should slow down and apply more control. The rational response is to let our countermeasures escalate smoothly and incrementally as technological power grows, rather than panicking.

Host: You've said there's a 10–25% chance that human civilization collapses. That's not trivial. Could there be a scenario where something Anthropic creates causes that collapse?

Dario: I certainly hope not. This risk probability stems from the inherent unpredictability of the technology and the complexity of the global landscape. The various measures we're taking are precisely aimed at reducing, not increasing, that collapse probability. It's like the aviation industry — even if the planes you build are ten times safer than others, you can never absolutely guarantee to anyone that they will "never crash."

Related reading: OpenAI and Anthropic collectively change their tune: Is the AI jobs apocalypse narrative no longer in vogue?