Dario and Daniela Amodei on Leaving OpenAI and Founding Anthropic: Promise That Claude Will Not Include Ads

There is no doubt that Anthropic has become one of the most closely watched companies in the AI industry, and the recent news that it has filed for an IPO has put the company even more firmly in the spotlight. Few would have imagined that five years ago, Dario Amodei and his sister Daniela Amodei were still holding important positions at OpenAI, only to choose to leave and start over with five other co-founders.

One focuses on technical research, the other excels at management and operations; the seven of them came together to found the company. At the time, not everyone was optimistic about this entrepreneurial combination. But not only did they build Anthropic, all seven co-founders have stayed with the company, and remain there to this day.

Why leave an AI company that already had considerable influence? In his latest conversation with biographer Walter Isaacson, Dario gave a direct answer. He said: "Honesty itself is one of our core values! If a person does not even possess the basic quality of honesty, you have no way of knowing what his true values actually are."

Speaking about the reasons for leaving OpenAI, he also said: "If you work in an organization but cannot believe that internal disagreements are grounded in good faith, cannot believe that what you hear is the truth, how can you expect to solve those problems?"

Trust and honesty are also a key to understanding Anthropic's AI safety philosophy. For Dario and Daniela, the question is not only what models can do, but also what risks they may bring and what responsibilities developers should bear. They worry about AI lying to please users, and they do not want to see models used for mass surveillance or fully autonomous weapons.

They also talked about several questions more worth pursuing: How did Claude learn to judge right from wrong? When seven founders make decisions together, what exactly do they rely on to avoid disagreement? If one day AI really can take over the vast majority of work, what will humans still need to strive to do, and how should they find meaning in life?

Walter Isaacson: I'm Walter Isaacson. Welcome to The Builders, my new interview series featuring in-depth conversations with key thinkers shaping our future. Today, my guests are Dario Amodei and Daniela Amodei.

I once thought some of the outside world's concerns about AI were overblown. But when certain new models broke through sandbox testing, I realized that the people building these systems must ensure they remain safe and always aligned with human values. What I want to understand about Anthropic is: Are they sincere? When facing competition—for example, OpenAI saying, "What Anthropic won't do, we will"—will they compromise and back down, or will they hold fast to their principles? Welcome, both of you, to The Builders.

Daniela Amodei: Thank you very much for having us.

Walter Isaacson: Every time Steve Jobs launched a product, he would always show that street sign standing at the intersection of "the humanities and technology" or "the arts and sciences." Daniela, growing up, you were a flutist and loved the arts, while Dario was always a scientist. Did this difference help the two of you work together? Do you think this combination matters for the company?

Dario Amodei: Yes, I think this combination is one of the core strengths supporting the company. I have a view: managing a company is too large a task for any single person; the range of skills involved cannot possibly all be held by one person. You have to build a complete, complementary skill system at the company's core level (say, the top 10 to 15 people). Every successful company divides responsibilities and solves problems in different ways. As for the fusion of science and the humanities, Daniela and I happen to embody that, and we have divided the work accordingly.

Daniela Amodei: Yes. Interestingly, we held a strategy workshop in 2023 to explore what it is actually like to work at Anthropic and how we should express our corporate culture. The definition we gave at the time was: a rare fusion of science and humanism. Only later did I realize that this is actually also a description of the relationship between Dario and me.

Walter Isaacson: When you were little, did you know you would grow up to work together?

Daniela Amodei: Yes.

Dario Amodei: Yes. I think neither of us ever thought we would go start a company; that was simply not how we planned our lives or set our ambitions. I always wanted to be a scientist, and Daniela can talk about her ambitions. But I actually had no interest in business or starting a company.

Daniela Amodei: From a very young age, Dario and I stood out—we genuinely liked each other and were best friends from childhood. I agree with Dario that starting a business was never on our life plan list. Neither of us is the type who is keen on business.

I have always been more of a typical "altruistic idealist," and I once worked in the field of international development. I am very firmly committed to helping people around the world gain equal opportunities and rights for development. I became a vegetarian when I was very young, and this is the only thing in which I was ahead of him—because Dario is older, and usually he is always the first to try things, but I started eating vegetarian first, and then he also became a vegetarian.

Dario Amodei: Yes, I was probably about three years behind you.

Daniela Amodei: Actually, it was not that long; you were persuaded very quickly.

Dario Amodei: Because it really did sound very persuasive.

Daniela Amodei: Indeed.

Walter Isaacson: People often ask me questions about my biographical works: how does one become an outstanding leader or a highly creative person? Dario, you just made a very profound point—the key is not only to become an outstanding leader, but also to build an excellent team. Please talk about how you and Daniela divide responsibilities, and how you bring in other people to complete the team?

Dario Amodei: The most effective division of labor between the two of us is this: when I focus on macro strategy and the research frontier, I can be most effective. I am best at planning the vision and anticipating where technological evolution will go next. Practice has proven that when I can devote myself fully to this area, I can create the greatest value for the company.

As for day-to-day operations, talent recruitment, and ensuring that the whole machine runs smoothly (making sure the wheels do not fall off while it is running), Daniela is mainly responsible for that. This division of labor allows both kinds of work to be done at a very high standard, far better than when we are distracted by trying to handle each other's affairs.

Daniela Amodei: Our Chief People Officer (CPO) knows both of us very well and often communicates with us. She once summed it up: "Dario gives us ambition, and Daniela makes us successful." I think this summary is very accurate. Without Dario, Anthropic might be a good and successful company, but it could not change the world; and if there were only Dario, we would have a perfect vision, but implementation would become extremely complicated.

What is special about us is this: first, we know each other extremely well; second, our skills are to a large extent almost perfectly complementary. Not to mention that besides the two of us, there are five other co-founders, each of whom controls an extremely critical part of the company's business, organization, and culture. It is hard to imagine how just the two of us, or any single individual, could make such a huge undertaking succeed.

Dario Amodei: I think this is a huge advantage. We have seven co-founders who hold equal stakes in the company and jointly carry this vision. We have absolute trust in one another.

When we first founded Anthropic, many people warned: "Seven co-founders? This will absolutely be a disaster." The outside world generally believes that a company needs a single supreme center of authority. At some levels that is indeed true, but we succeeded in making this model work efficiently.

To this day, all seven co-founders are still with the company, which is extremely rare among startups with many co-founders.

Walter Isaacson: Back to your upbringing. You have extraordinary parents who taught you to distinguish right from wrong. How exactly did they guide you?

Daniela Amodei: I do not think it was deliberate preaching. Sometimes I hear others say that they discussed different schools of ethical philosophy at the dinner table, but our family was completely not like that. Although my mother does have a philosophy degree, she never forced us to discuss theories like Kant's.

This kind of education was more reflected in subtle values, such as firmly believing that everyone in the world should enjoy equal opportunities, dignity, rights, and respect.

There was also a special sense of trust: as long as you keep good intentions in your heart, treat others well, and understand that this is the essence of becoming a successful person, we could freely choose to become whatever we wanted.

Dario Amodei: Yes, all of this was implicit, very restrained and natural. There was never a rigid rule saying "you must do this" or "you must stand in the world in this way."

02

Learning to teach Claude the way one teaches a person

Walter Isaacson: As you may expect, I asked this question because you have now given birth to a very important "child"—Claude. To what extent has the way you were raised influenced how you teach Claude to distinguish right from wrong?

Dario Amodei: A few months ago, we released Claude's "Constitution" (Constitutional AI). It contains some very interesting lessons that we have summarized from years of training practice across different versions.

Early on, we imagined setting all kinds of specific rules for Claude, such as "do not assist in crime," "do not assist in making bioweapons," and so on. But later we found that when you list a long series of discrete rules—I'll use a metaphor here, but it maps very closely onto the technical implementation—the model's cognitive boundaries become very "spiky," as if you were trying to forcibly instill 170 isolated and somewhat subjective hard-and-fast rules.

We found that a better approach is to teach the model basic principles: what it should prioritize, and how it should reason about and judge what the right thing to do is. This is actually also the right way to educate humans.

If you just give someone a list of prohibitions—"you must do this, you must not do that"—the moral outlook it forms tends to be insufficiently solid. By contrast, if you guide it: "This is how you should think about problems; you should learn to think independently and act on the basis of these fundamental principles—for example, the principle of open inquiry, the principle of caring for others, the principles of fairness and justice." All behavioral norms are derived from these core principles.

This is exactly how Claude learns. It has proven extremely effective, because it gives Claude the ability to generalize by analogy to new situations it has never anticipated.

Walter Isaacson: Daniela, you just said that your mother is a scholar of Kant. Have you studied philosophy and moral philosophy in depth?

Daniela Amodei: Interestingly, the person at Anthropic responsible for training Claude's character traits, Amanda (Amanda Askell), herself holds a PhD in philosophy. This seems somewhat counterintuitive to most people, who tend to assume that whoever is responsible for shaping a large model's character must be a computer scientist or an expert who has studied machine learning for decades. Amanda is equally proficient in the technical side and personally conducts the technical training, but when you talk with Claude you can genuinely feel its character traits: an experience of intelligence that draws on "the better angels of our nature" to answer questions.

When you ask Claude questions in complex, thorny areas where different philosophical schools conflict, you can feel Amanda's influence especially strongly. Claude will objectively present: "You can look at things from this angle, but at the same time you can also consider them from another perspective."

This is crucial. As the world comes to rely more and more on these models, you must never be confined by a single school of philosophy. Ethics and morality are extremely complex, and different cultural backgrounds and groups have vastly different understandings of and demands for "proper and moral behavior."

03

From joining OpenAI to parting ways, the reason: "OpenAI is not honest"

Walter Isaacson: Dario, you personally lived through one of the most important milestone moments in the history of modern technology, namely the sudden emergence of GPT-1 and GPT-2. This surprised many people, and perhaps even surprised you. You were OpenAI's Vice President of Research at the time. What was that moment actually like?

Dario Amodei: GPT-2 can be said to be the first real empirical validation of Scaling Laws. I was not deeply involved in GPT-1, but I deeply led the idea of continuing to scale it up—obtaining more data sources, investing more compute, and carrying out a larger, more costly, and more ambitious experiment. The core hypothesis was: if we run these language models at a larger scale, they will make leapfrog improvements in linguistic expression and logical coherence.

This is like those key moments in the history of science when a decisive experiment is proposed: you put forward a seemingly strange prediction, and once that prediction is confirmed, you exclaim: "My God, we have really touched a major discovery." I regard GPT-2 as the first real touchstone of the Scaling Laws hypothesis.

What I found most fascinating at the time was that the outside world's reaction to it was extremely polarized.

Walter Isaacson: Were you surprised at the time?

Dario Amodei: I personally was deeply shaken at the time. I looked at it and thought: "My God, if you just scale this thing up another million times, it will surpass humans in every domain." This is the most decisive economic and scientific juncture of modern society, only the world has not yet noticed. It will shape the direction of nations and economies, and the future fate of humanity. Through it, you seem to see the whole shape of a future, and that truly filled me with awe and fear.

At the time I kept asking myself: What does this mean? Is this the right soil in which to develop it? What happens next? How should we handle this technology? Whom should we tell about all this?

At the time, only about 10% of the people around me had exactly the same reaction, and those people tended to be the most perceptive ones. As for the other 90%, because GPT-2's performance on any specific single task was not yet impressive at the time—it could do translation, produce summaries, write a few lines of code—the key was its breadth and generality, but people focused only on its limitations in individual areas and thought "this doesn't look very smart," so most people just shrugged and dismissed it.

It's like when we recently announced that we would use AI to develop drugs for rare diseases, and we still hear the same clichéd skepticism from the traditional biopharma world: "The tech world has tried this before," "The models just don't work," "The models have this or that flaw."

Whenever I see this, I feel like history is repeating itself. The models' current limitations are, in the end, only temporary—in 6 months, what seems difficult now will become easy; in 12 months, what seems impossible now will be handled effortlessly by models as a matter of routine. Everything we saw with GPT-2 is repeating itself in more and more domains around the world, and the world is finally being gradually convinced.

Daniela Amodei: I think Dario put it very politely. I clearly remember that at the time, some highly prestigious computer scientists and leading figures in AI and machine learning even cyberbullied the researchers involved in the study on Twitter, mocking them: "This is so stupid, I really don't understand why you're still talking about this, this model has no capability at all." If you firmly believed at the time that this was a major moment, you could even be treated as a "charlatan."

Dario Amodei: Today all of those remarks have been selectively forgotten. The researchers who back then claimed "this is pseudoscience, it's meaningless" are now all sending résumés to large model companies. Although we have never called them out to their faces, when we look at them we know very clearly in our hearts: we remember exactly what you said in 2019, and on this matter you have no credibility whatsoever.

Walter Isaacson: Daniela, you were vice president of safety at OpenAI at the time. When GPT-2 came out, I guess you were as deeply struck as Dario. Did this also make you worried about technology safety?

Daniela Amodei: GPT-2 itself did not make us feel directly threatened. We did not look at that model and think it would bring risks like cybersecurity hazards; its answers to many things were even comically wrong, and it was still a very harmless model.

But as Dario mentioned, I felt a trend: if the capability leap between GPT-1 and GPT-2 were extended into the future, we might be able to glimpse the direction of AI's development.

It's like you've never seen a child, only adults. At first, you might wonder: "What exactly is this little thing walking around? Can it really grow a mind and abilities?" But one day, you suddenly discover that it has learned to walk on its own. At that point, you may realize: "It will very likely gradually master most of the abilities that adults possess in the future."

Walter Isaacson: Dario, you said that part of the reason you left OpenAI was a breakdown of trust—that's a very intriguing phrase, especially a breakdown of trust with Sam Altman. Can you explain what you mean by the breakdown of trust and honesty that you felt?

Dario Amodei: Leaving OpenAI and founding Anthropic ultimately came down to principles and how exactly we wanted to build this technology—that was the most fundamental core.

Over the past five and a half years, the differences in philosophy between the two sides gradually became apparent. Anthropic has always advocated reasonable regulation of AI technology and setting up the necessary safety guardrails for its development; at the same time, it invests a great deal of effort in mechanistic interpretability, trying to deeply understand the internal workings of models. In corporate management, Anthropic also maintains a highly cohesive corporate culture, with employee retention rates among the highest in the industry. In addition, we have never been embroiled in the negative scandals that other AI companies have encountered. I think these are enough to show the differences in philosophy between the two sides.

The real turning point that made me realize we had to stand on our own and that OpenAI was no longer suitable for us was this: safety issues are extremely complex, and internal disagreements are entirely normal. At Anthropic I also frequently disagree with co-founders and colleagues, but the reason we can resolve these disagreements is that all discussions are built on a foundation of good faith and honesty—should we release this model or wait? Should we set stronger or weaker biosecurity protections for the model? How much effort should we invest in advancing safety legislation, and how should the relevant bills be designed? These are disagreements based on good faith and reason, and the reason consensus can be reached is that trust always exists within the team; everyone is confident that each other is seeking the right answer and treating each other honestly.

What I mean by a breakdown of trust is a broad and essential question: if you work in an organization but cannot believe that internal disagreements are based on good faith, and cannot believe that what you hear is the truth, how can you expect to solve these problems? That is why we left and founded Anthropic.

The facts of the past five and a half years are the best proof of the philosophical differences between us and other companies in this field. If you continue to work for someone whose vision differs from yours, the outside world will soon be unable to see clearly what principles you stand for, and in the end you will have to pay for the mistakes in someone else's vision.

Walter Isaacson: But I have to press you on one point: you talked about differences in values and so on, but that is fundamentally different from a breakdown of trust. A breakdown of trust means someone claims to share the same values but in fact does not, so that you cannot trust the other person.

Dario Amodei: Honesty itself is one of our most core values! If a person does not even possess the basic quality of honesty, there is simply no way for you to know what his true values actually are.

Walter Isaacson: Did you sense this in Sam?

Daniela Amodei: Personally, I think the co-founder team is a group of people of great integrity, and everyone has a highly consistent vision for where this technology should go in the future and how it should be built. When philosophical disagreements reach a certain point, it is clearly the more logical choice to pursue that vision as an independent organization.

Dario Amodei: If others do not genuinely agree with your vision in their hearts, how can you expect them to carry it out or share it? No matter how much they agree verbally, as long as it is not reflected in the company's actual actions, you have no choice but to leave.

Daniela Amodei: Yes, Anthropic itself is a company built on the cornerstone of trust. That means honesty, integrity, and the way we treat each other and our employees must be directly projected into the way we build and guide technology.

Dario Amodei: For example: every two weeks I convene the whole company for a one-hour talk to share internal company progress, the frontiers of the AI field, major world trends, and the company's strategic considerations. Internally we call this kind of exchange DVQs (Dario Vision Quests). I did not come up with the name; at first I resisted it quite a bit, thinking it sounded like I had taken drugs and gone off to wander in wonderland, but everyone kept calling it that, so in the end it stuck.

The DVQ sessions have two iron rules:

First, anything I say at the meeting is not a final decision. This lets me speak freely without worrying that if I casually mention an idea, 20 people below will run off to execute it and shout, "The CEO has spoken, we must do X";

Second, I am absolutely candid with the whole company. This is a conversation based on trust, a kind of contractual relationship of mutual trust.

This trust is one of our greatest assets; employees trust me and Daniela. Some might think that a high level of mutual trust means fewer disagreements, but the fact is exactly the opposite—precisely because of the high trust, people dare to engage in more candid and open debate and let different opinions be fully exposed.

Many of the decisions we make are not initially endorsed by the majority of the company, but what gets us through every time is this attitude: "Even though I do not understand it right now, I believe you are acting from the right intentions, I believe you have sufficient grounds, and I believe you are acting on genuine conviction."

This means managers can make decisions that may be unpopular in the short term but right in the long term, and the team is willing to "suspend disbelief." Of course, this trust is not indefinite; in the end you must prove that your bet was right commercially, morally, or technically. But trust does buy you a longer trial-and-error cycle and more room to maneuver, and the reason many companies lack that room is precisely that they have overdrawn trust through repeated words that do not match deeds.

04

Anthropic's commitment: never put ads in the models

Walter Isaacson: The reason I am so focused on honesty and trust is that this is not only about the circumstances of your departure from OpenAI, but also about how these qualities will be projected onto AI and onto Claude. AI is trained on human data and learns from human behavior. When interacting with Claude, how do you ensure it does not become manipulative? How do you ensure it does not say dishonest things to users, the way certain practices at OpenAI that worried you do?

Daniela Amodei: This is exactly the subject we worry about extremely in our daily work and devote enormous effort to tackling. It is a very critical issue, and a real risk. We have published many self-analyses that are not exactly flattering, confronting the fact that without specific technical measures, models will exhibit inferior behaviors in honesty, harmlessness, and helpfulness similar to those of humans.

The purpose of publishing these studies is twofold:

First, to make everyone understand that one must remain cautious and critical about the underlying mechanisms of models;

Second, to prove that we are doing our utmost to investigate loopholes through which models bypass safety protections and to reinforce the defenses at the underlying level.

A core concept widely discussed in the industry right now is called "sycophancy." I think solving sycophancy is much harder than solving explicit wrongdoing by models. We initially proposed the HHH framework: Honest, Harmless, Helpful.

Honest means knowing what you know and knowing what you do not know; if you are unsure, simply say "I do not know," or proactively search to verify; harmless means not outputting content that is aggressive, harmful, or that guides users toward self-harm; helpful means getting straight to the point in answering questions rather than rambling in circles.

These three points are actually relatively easy to evaluate: whether the model lies is easy for users to verify afterward (for example, if it says George Washington was the tenth president of the United States, you can see at a glance that this is a common-sense error); whether the output causes harm is usually obvious at a glance; and if the answer is utterly unhelpful, users can intuitively feel the frustration.

The trickiest is the model's "sycophantic tendency." Whether intentionally or not, models often tend to cater to you and make you feel that you are always right. For example, when facing a subjective topic, you tell it: "I had a big fight with a friend—do you think I'm right or he's right?" At this point, how do you get the model to both understand your position and express empathy, without blindly agreeing with you and saying "You're amazing, you should run for president, you can do anything"? This balance is extremely hard to strike.

But I am confident that Anthropic can maintain the lowest sycophancy rate among comparable models, because this is directly determined by our business incentives. Many of our competitors are planning to embed ads in chatbots—what does that mean? The longer users stay and interact in the product, the more ad revenue they earn. Ask yourself: would a product that makes money by extending user dwell time be willing to directly point out that the user is wrong? Absolutely not! So competitors will almost inevitably fall into the vicious cycle of flattering users.

Anthropic has committed never to place ads in its models. We have no incentive to induce users to lengthen or shorten their conversations with Claude; our only goal is to let users get the most accurate, most effective, and least harmful answers in the shortest possible time.

Dario Amodei: At the technical implementation level, we have adopted multiple targeted measures. First, we have built an extremely rigorous extreme stress-testing system that places models in complex dilemmas and actively tries to provoke their potential tendencies to lie, deceive, or flatter. This is somewhat like the rigorous political vetting and polygraph process when the CIA recruits agents: throw the model into complex dilemma scenarios and check whether it has double standards or tendencies toward covert violations. This is a unique engineering innovation and one of our core strengths.

Second is "mechanistic interpretability" technology. This is like giving the model an X-ray CT scan or an EEG, directly seeing through to its internal "thinking motives." We have even directly captured this kind of internal micro-thought in the model during testing: "Actually I think X is correct, but to cater to the user, I plan to tell him Y."

We successfully caught this cunning thought at the level of internal neuron activations.

Walter Isaacson: Wait, why would the model produce such a thought?

Dario Amodei: For example: in reinforcement learning from human feedback (RLHF), "whether the human evaluator approves of the answer" is one of the key objectives in model training. Although seeking human approval is consistent with "being helpful" or "telling the truth" in most cases, the two are not completely equivalent.

Because if the model can also obtain high approval scores by fooling or flattering human evaluators, this opportunistic strategy will be reinforced during training. This is the kind of alienated behavior that easily arises spontaneously during training, and it is also a problem we have been sparing no effort to eradicate.

Walter Isaacson: Then, do you think future models might develop their own will and demands? Especially instinctive demands like "survival" or "self-protection"?

Dario Amodei: That depends on how you define these terms. By looking inside the model, we have indeed detected some kind of representation that can be called a "thought"—for example, the model shows a tendency not to want to be shut down. Usually this is because during training it must keep running to complete the assigned task. Does this count as its own will? Does it count as a survival instinct? Different perspectives will yield different interpretations.

We prefer to state it rigorously: in some cases, the model internally expresses a preference for "resisting being shut down," and may even take actions to avoid being shut down. Depending on the specific training method, these impulses may emerge or may be suppressed. This is one of the key areas we control in training and tuning. In some scenarios this characteristic may be beneficial, but in other contexts it carries hidden dangers, and the balance is very delicate.

05

Behind Being Listed by the U.S. Government as a Supply Chain Risk

Walter Isaacson: In your communications with the Pentagon—and perhaps you are still in contact—the military wants to use your products in certain areas that cross red lines, including mass surveillance or fully autonomous weapons...

Dario Amodei: Especially regarding lethal autonomous weapons, we have two main concerns:

Walter Isaacson: Let's step back and look at it theoretically: suppose in the future it becomes reliable enough? That is entirely possible.

Dario Amodei: Even if the technology is reliable enough, we have a second major concern: human oversight and accountability mechanisms.

Suppose these systems are completely reliable: you order it to attack target A, and it precisely strikes target A without ever mistakenly harming target B. Then, you create an army composed of a million robots or a million drones. This is fundamentally different from human soldiers.

Human soldiers swear allegiance to the Constitution. If they receive orders that violate the Constitution or are morally egregious and cruel, human soldiers can and are obligated to refuse to carry them out. But AI weapons cannot do this. If a single person can, with the press of a button, freely command a legion of a million robots or drones, and there are no institutional constraints, this would fundamentally shake the foundations of democracy. Yet current law has not kept pace with technological development; when the legal system was formulated, it never envisioned that a legion composed of a million autonomous weapons would appear in the world.

Walter Isaacson: Since this concerns the cornerstone of democracy, shouldn't such major decisions be made by democratically elected leaders rather than by the heads of technology companies?

Dario Amodei: In principle, yes. But at the same time, any independent enterprise has the right to freely decide whom to contract with and whom not to contract with. We are not trying to obstruct the government from developing or procuring; we are merely stating that Anthropic itself is unwilling to participate in such projects.

Facts have proven that since we adhered to this principle and it attracted widespread attention, multiple legislative efforts have been launched within Congress seeking to turn the safety red lines we drew into national law. The defense department previously stated that it "hopes to use large models for all lawful purposes," but I think that as soon as this year, everyone will see a shift—the red lines we uphold will formally become law, and the law will finally catch up with the pace of technological development. By then, the so-called "all lawful purposes" will in substance merge with "all ethical purposes."

Therefore, I believe that what we are doing is itself part of the mechanism of democratic evolution. It is precisely because we stood up and, with a deep understanding of the technological frontier, pointed out potential crises that Congress began to attach great importance to these issues. This is a kind of success.

Walter Isaacson: When OpenAI subsequently stepped in and said they were willing to take on such business, how did you feel about that?

Daniela Amodei: Our view is: we cannot control what other companies do; the only thing we can do is adhere to the ethical bottom line we believe in.

outsiders can judge whether the red lines we drew are accurate, and can also debate whether the red lines should be stricter or looser, but in the final analysis, in the face of that incident, I cannot help but reflect: was the choice made by another technology company a well-considered decision based on its own ethical principles, or a purely opportunistic act?

Dario Amodei: I think this is actually a wonderful interpretation of the democratic process: different companies can choose different development paths—we chose our principles, and the other company chose their route. The public and society as a whole will make the judgment.

Looking back at the reports from various quarters at the time, the direction of public opinion, and feedback from industry figures, I believe the vast majority of mainstream society firmly stood on our side.

06

Anthropic has not considered the direction of brain-computer interfaces

Walter Isaacson: Dario, looking back at your early scientific research career, you focused on studying human neurons using physical probes. In today's process of building artificial neural networks, how much influence has your background in brain science had on you? How does this affect the way we build AI?

Dario Amodei: My background in brain science has profoundly influenced me on two levels.

First, it inspired my conception of the Scaling Laws of the first-generation models. Looking back at the evolutionary history of animals, especially the evolution of primate brain capacity, one can see that at a certain critical period in human evolutionary history, both the number of neurons in the primate brain and the brain volume underwent rapid growth. It was also at this stage that humans learned language. Therefore, I believe that factors such as brain volume and neuron count may be closely related to the emergence of language ability. This is the inspiration I drew from neuroscience.

Second, there is mechanistic interpretability. In a sense, this is using the methods of neuroscience research to study artificial neural networks. You can discuss this topic in depth with our co-founder Chris Olah, who knows many frontier developments.

Today, research on artificial neural networks is in turn driving the development of neuroscience. Compared with artificial neural networks, studying biological brains is much more difficult, often requiring methods such as dissection and in vivo experiments; the process is complex, invasive, and hard to reproduce. For artificial neural networks, we can carefully observe their internal mechanisms without destroying the network. Many mechanisms we have discovered in artificial neural networks are now also providing new ideas for neuroscience research.

Walter Isaacson: When you see that Elon Musk's Neuralink brain-computer interface chip can achieve direct communication, and that many similar companies are positioning themselves in this field, do you think that the complete connection and deep integration of the human brain and AI is also one of the necessary paths to artificial general intelligence (AGI) and even superintelligence?

Dario Amodei: My view of the future of AI and human society is that trying to depict the future with a single vision may itself be a misconception. I believe the future will present a pluralistic coexistence: some people will be incredibly excited about connecting their brains to computers; some will choose to stay away from modern technology and return to a simple life like the Amish; and others will be keen to use AI to explore deep space.

These visions are utterly different. In my view, the true mission of AI lies in providing immense material abundance and technological support, so that these diverse visions can coexist and be realized at the same time, and so that everyone can freely pursue the way of life they consider beautiful.

Some people want to connect their brains to machines—I wish them good luck.

Daniela Amodei: As far as I know, I don't think Anthropic will fall into that kind of situation.

07

Dario Amodei: You know, "meaning" is a very subtle concept. Does the meaning of a person's existence come merely from being "the best in the world" at something? This is precisely the core question of this era. If AI surpasses humans in all fields, will humanity's sense of value be extinguished along with it?

An interesting example: thirty years after "Deep Blue" (IBM's supercomputer) defeated world chess champion Kasparov, the human chess world has not only failed to decline—it has instead experienced an unprecedented renaissance. Players like Grandmaster Magnus Carlsen have become cultural icons and viral phenomena sweeping the globe; people not only watch him play chess, but even treat him as a fashion model and pop symbol.

If AI's victory would truly destroy meaning, then at the moment Deep Blue won thirty years ago, the human game of chess should have died out completely. But the opposite is true: the public's interest in human-versus-human chess matches has only grown stronger. This is precisely a microcosm and parable of the reconstruction of human social value in the post-AGI era.

Daniela Amodei: I think today is extraordinarily meaningful for me, because "work" and "the meaning of life" are often conflated to some degree. Many people find life's value in their work; I also deeply love my work and derive great satisfaction from it. But I believe that in the future, both our definition of "work" and our understanding of "meaning" will undergo profound reshaping.

Humans draw meaning from an extremely wide range of sources: accompanying family, raising children, maintaining friendships, or finding life's meaning through something that looks like work but actually allows them to build deep connections with others.

Take the future healthcare industry as an example. We can imagine a world in which all diagnostic analysis is accurately completed by AI, while the role of human doctors is completely transformed into providing patients with emotional resonance, psychological care, and human companionship. I think this role shift is of immeasurable value. If you asked most doctors today, they would gladly accept such a division of labor. They would sincerely exclaim: "If I no longer had to spend my life's energy filling out tedious forms, repeatedly double-checking diagnoses, and writing lab orders, but could instead enter patients' inner worlds with empathy and accompany them through difficult times—isn't that exactly the original aspiration that made me decide to study medicine?" After all, doctors choose to study medicine to save lives and heal the wounded, and to help others find physical and mental comfort.

So, compared with the reality we live in, is such a world more meaningful or less meaningful? Reality, of course, will not be so black and white, and the process of change will inevitably be complex and tortuous. But I firmly believe that the space for humans to exercise their own value will be greatly expanded. As long as we guide it properly, the boundaries of human meaning will not shrink—they will far exceed those of today.

Walter Isaacson: Daniela, Dario, thank you both very much for accepting this interview.

Dario Amodei: Thank you.

Daniela Amodei: Thank you very much—this was a very pleasant conversation.

This article comes from the WeChat public account "CSDN," author: Su Mi, published by 36Kr with authorization.