De quitter OpenAI à entrer en conflit avec le Pentagone : comment le duo frère et sœur d'Anthropic trace une ligne rouge pour l'IA afin d'éviter l'effondrement de la civilisation

Bloomberg interview with Anthropic founder, revealing the rift with OpenAI and the "safety first" mission. Facing conflicts with the Pentagon, warnings of large-scale unemployment, and threats from powerful models Mythos.
（Background: Claude has 80% of its code written by itself, is Anthropic calling for a "global design brake mechanism" seriously?）
（Additional background: Anthropic CEO Dario Amodei: In 6-12 months, China's open-source AI models will catch up with Mythos）

Table of contents

Toggle

• Chapter 1: The规律 and vision of exponential growth
• Chapter 2: From childhood in San Francisco to the rupture with OpenAI
• Chapter 3: Defining AI morality and "Constitution"
• Chapter 4: Corporate布局 and unemployment crisis
• Chapter 5: Pentagon's red line and Mythos model
• Chapter 6: How to face a 25% extinction probability?
• Epilogue

In the AI arms race, Anthropic is seen as OpenAI’s strongest competitor, already leading the way. Founded in 2021 by former OpenAI executives Dario Amodei and Daniela Amodei, the company has always championed "AI safety," with its Claude model highly praised in the industry.

However, over time, Anthropic has also fallen into a vortex of power and ethics. From contract disputes with the Pentagon to the supermodel Mythos capable of destroying cybersecurity, this sibling duo stands at the forefront of transforming human civilization.

Bloomberg senior reporter Emily Chang visited Anthropic’s headquarters in San Francisco to gain insight into this nearly trillion-dollar valuation company’s core vision, internal conflicts, and fears for the future. Here are the key points from the interview.

Chapter 1: The规律 and vision of exponential growth

Emily Chang: I really like this library. Are you a big bookworm?

Dario Amodei: It’s indeed very beautiful here. Overall, I read a lot of books, but honestly, in the past year or so, I hardly had any free time.

Emily Chang: Your product updates are so frequent and rapid. How do you manage that?

Dario Amodei: We use Claude (our own model) throughout the entire product development cycle. This allows us to release at an extremely fast pace.

Dario Amodei: The feeling of this exponential is: at first, it seems like nothing happens, then some small things happen, and then—bang! It explodes wildly. I’ve observed this规律 for a long time, and I think: “Yes, we might become the highest revenue and highest valuation AI company at some point.” And it indeed happened. We’re always thinking: how to train good models? How to turn them into good products?

Most importantly, how to ensure everything is safe.

Chapter 2: From childhood in San Francisco to the rupture with OpenAI

Emily Chang: Tell me about your childhood in San Francisco. Your father was a leather craftsman, your mother worked in a library—how did that influence you?

Dario Amodei: Back then, the first internet revolution was happening, but I had no interest. I was only interested in mathematics, understanding the universe, and science fiction novels. I was filled with a strong curiosity about the world.

Daniela Amodei (Anthropic President): Dario has been very smart since childhood. He was doing calculus in middle school, studying mathematics at Berkeley in high school. I tend to focus on reading and art; our personalities are almost perfectly complementary.

Note: In 2016, Dario joined the newly established OpenAI, and Daniela followed in 2018. Dario developed the famous "scaling laws" there, predicting that increasing data and compute would make models smarter. This theory paved the way for ChatGPT’s success. But in 2021, the Amodei siblings chose to leave, becoming a Silicon Valley legend.

Emily Chang: Your decision to leave OpenAI has become a “myth” in Silicon Valley. What exactly caused the split?

Dario Amodei: Regarding "safety," there are many reasonable disagreements. But disagreement alone isn’t enough to leave. The core issue is "trust." When you feel that the other’s values don’t align with what they say, and you feel they are dishonest, it becomes very hard to cooperate. Instead of arguing with those who lack a shared vision, it’s better to do your own thing.

Chapter 3: Defining AI morality and "Constitution"

Emily Chang: Claude has a very unique style and a humanized name. What message are you trying to convey?

Daniela Amodei: We call it "Professional Warmth." The goal isn’t to make it your best friend, but to make it approachable while maintaining a professional distance.

Emily Chang: You mentioned teaching Claude to "be better." What is a good model? What is a bad one?

Daniela Amodei: We don’t want models to lie—whether intentionally or unintentionally, we call this "hallucination." We also must ensure models don’t deliberately try to deceive users. When defining "good" and "bad," although there’s no global standard, we train it using foundational documents like the United Nations Universal Declaration of Human Rights.

Interestingly, we’ve even started dialogues with religious leaders to explore how to embed core human values that have been jointly guarded for thousands of years into the models.

Chapter 4: Corporate布局 and unemployment crisis

Emily Chang: Why focus on enterprise and coding?

Dario Amodei: Business models must align with values. The social media world is driven by ads and addiction. In the enterprise sector, AI can cure diseases and optimize energy. This better aligns with our goals.

Boris Cherny (Claude Code lead): Previously, engineers had to write code manually. Now Claude writes code, and I oversee it. It’s like having a jetpack; the role of engineers is undergoing a revolutionary change. Personally, 100% of my code is written by Claude.

Emily Chang: But this raises concerns. You once predicted AI might replace half of entry-level white-collar jobs. Is that still your view?

Dario Amodei: I remain very worried. Currently, AI makes people more productive, but when automation approaches 100%, what are humans supposed to do? That’s an uncomfortable fact. We need to think about Universal Basic Income (UBI) or policy interventions.

Daniela Amodei: I’m a bit more optimistic. I believe humans will find new ways to utilize AI, and human-to-human interactions (like bedside care in medical diagnosis) are things AI can never fully replace.

Chapter 5: Pentagon’s red line and Mythos model

Emily Chang: Media reports say Claude has been used for military actions against Iran, and even in operations in Venezuela.

Dario Amodei: I am a patriot. I believe we need to protect democracy from authoritarian threats. But I’ve drawn a red line: Claude cannot be used for "mass surveillance" or "autonomous weapons." When the Pentagon asked us to remove safety barriers, we refused. Even if we are blacklisted or called "left-wing lunatics," we will not back down.

Emily Chang: And what about the Mythos model? I hear it’s so powerful that it causes fear outside.

Dario Amodei: Mythos has discovered vulnerabilities in almost all mainstream operating systems. Companies testing it call it a "superweapon."

Emily Chang: You decide who can use it and who cannot. Isn’t that concentration of power worrying?

Daniela Amodei: It’s indeed complex. We must balance "preventing it from being used to hack banks or steal secrets" with "helping defenders." We don’t want to repeat the social media mistakes, waiting until something goes wrong to apologize.

Chapter 6: How to face a 25% extinction probability?

Emily** Chang:** You mentioned a 10% to 25% chance of human civilization collapsing. Why do you think you can handle this pressure?

Dario Amodei: I handle the pressure through "absolute transparency." I give a one-hour talk to all employees every two weeks, discussing everything openly. This builds trust and makes me feel less lonely.

Many compare Oppenheimer, who pushed nuclear development but later called for arms control, to these AI pioneers. But Dario sees Oppenheimer as a "failed case" in some ways.

Emily Chang: Who is the historical figure you most agree with? Oppenheimer?

Dario Amodei: I identify more with Leo Szilard, the first to conceive of chain reactions. Oppenheimer is somewhat a "failed case" because he didn’t establish enough checks and balances. What we need is checks and balances.

Emily Chang: At the center of all this madness, how do you find peace?

Dario Amodei: Honestly, it’s "Exposure Therapy." Sometimes I play games with Daniela or my wife, or go to Italy to see our horse "Calypso." It knows nothing about AI; it’s just a happy horse. That’s my Zen.

Epilogue

Emily Chang: You’re building an incredibly powerful technology and will profit from it. Why should we trust you?

Dario Amodei: Skepticism is natural. Silicon Valley has lost a lot of trust. Trust isn’t something you say; it’s something you do. If these people can’t even uphold their bottom line, then maybe no one can. We can’t guarantee success, but we can guarantee we’ll do our best.