Welcome to Macro Insight. Here, we strive to present diverse perspectives and help you break out of information echo chambers. When it comes to AI, feedback suggests that Anthropic is currently leading the pack, arguably a key player that will have a massive impact on the future world. So, what are the values of the company's founders? What have they done in the past? And how did they reach their current position? Just recently, American biographer Walter Isaacson interviewed the two co-founders of Anthropic, Dario and Daniela, on his podcast, "Builders." Truth be told, all these questions were addressed in the conversation. So, how will these two individuals, who may hold immense power over the future world, answer where they plan to take Anthropic? Now, let's get into the conversation.
I'm Walter Isaacson, welcome to "Builders." This is my brand new interview series, specifically interviewing the heavyweights who are shaping our future. Today, I'm speaking with Dario and Daniela Amodei. I used to think that people's concerns about AI were a bit of an overreaction, but then, when certain new models broke through safety boundaries, I realized that those developing these systems must ensure they are safe and aligned with human values. What I want to understand is, is the company Anthropic for real? When faced with competition—for example, when OpenAI says, "If Anthropic won't do it, we will"—will they back down, or will they stick to their principles? Welcome to "Builders."
Thank you so much for having us.
Steve Jobs always said at every product launch that it stood at the intersection of art and science, or rather, the humanities and technology. Daniela, you've loved art since you were young, and Dario, you've always been a scientist. Has this combination been helpful to you? Do you think this combination is important?
I think this combination is exactly the core driving force behind our company. I have a theory that managing a company is too massive a task for any one person to handle alone. The range of skills required is too broad for one person to possess, so you have to assemble a complete set of skills among the top 10 to 15 people in the company. Every highly successful company finds different ways to divide labor and solve this problem. I feel that in terms of science and the humanities, Daniela and I perfectly embody this combination and division of labor.
It's quite interesting. Around the time of our 2023 company offsite, we were talking about what it actually feels like to work at Anthropic. How should we embody our corporate culture? We said it was a wonderful blend of science and the humanities, and it wasn't until later that I realized this is also the best description of Dario and me as people. Did you ever imagine growing up that you would work together?
Never. I don't think either of us ever thought about starting a company; we didn't plan our lives that way, nor did we think about doing this. I always wanted to be a scientist. Daniela can speak for her own aspirations, but I really had no interest in starting a business at all.
You could say Dario and I are unique; we've truly liked each other since we were very young. We've been best friends since childhood, and I agree with Dario—starting a business was never part of our plan. Neither of us is that type of person.
We both lacked any strong desire to start a company. I always saw myself as a doer, which is why I previously worked in the field of international development. I had a strong desire to help everyone in the world gain equal resources and opportunities, and I became a vegetarian at a very young age—that's the only thing I did before Dario, because he's older, so he usually did everything first. I became a vegetarian first, and then he followed suit.
It was probably three years later.
Maybe less than three years, it happened quite quickly. You convinced me pretty fast; I thought you made a lot of sense.
Sometimes people ask me about my book, asking how to become an excellent and highly creative leader. Dario, you once said something very interesting: you said that you must not only be a great leader but also build a great team. Tell me, how do you and Daniela divide your work, and how do you collaborate with others?
We've found the most effective way to divide labor is this: I'll talk about the two of us first, and then we can talk about the rest of the team. I feel I am at my best when I'm thinking about big-picture strategy and research. For example, I'm best at planning the vision and predicting the next steps for growth. I find that when I can free up all my energy to focus entirely on these matters, I add the most value to the company. As for the daily operations, recruiting, and ensuring the company machine runs smoothly while I think about our future direction—Daniela is in charge of all that. I think this division allows us both to do our work exceptionally well. If we were constantly distracted by each other's roles, we definitely wouldn't be doing as well as we are now.
Our Chief Human Resources Officer knows us both very well and talks to us often. He had a line that summarized it perfectly: he said Dario provides the ambition, while I lead us toward success. I think that sums it up perfectly. Without Dario, Anthropic might just be a successful company, but it wouldn't be able to change the world; yet if we only had Dario's perfect vision, executing it would be very difficult. So, I think we have a special chemistry. First, we know each other very well, and second, we are very lucky that our skills are almost 100% complementary—or at the very least, if not 100%, they are very close to it. Not to mention that besides the two of us, we have five other co-founders who are responsible for critical parts of the company's operations, culture, and organizational structure; I really can't imagine how we could have built the company like this with just the two of us, or even just one person.
This is an incredible advantage because we have seven co-founders in total, all with equal equity in the company, and they all share this vision; we have an innate trust in everyone. When we first started Anthropic, many people mocked us, saying that with seven co-founders, it was bound to fail. People generally feel that a company needs an absolute leadership core, and to some extent, it does, but we actually managed to make this unique model work. Now, all seven co-founders are still with the company, which is very rare when there are so many partners.
Let's go back to your formative years; you had amazing parents who taught you how to distinguish right from wrong. How did they do it?
Actually, it wasn't taught intentionally. Sometimes I hear people ask, "Did you discuss philosophy, ethics, and morality at the dinner table?" It wasn't like that at all. My mother did have a degree in philosophy, but she never forced us to discuss Kant or anything like that. I think it was more of an internal influence of values, like a firm belief that everyone in the world should have opportunity, dignity, rights, and respect. As long as we held the belief that being a successful person required being a good person and treating others well, we felt we could become whoever we wanted to be—that was a special feeling.
This influence was subtle and internalized; there were never any explicit instructions telling us what we should do or how we should live in this world.
I ask this because, as you might have guessed, you now have a very important "child" named Claude; to what extent did your upbringing help you teach Claude to distinguish right from wrong?
A few months ago, we released a set of rules and guidelines for Claude. Over these past few years, in the process of continuously training different versions of Claude, we have learned some very interesting lessons. Initially, we envisioned setting various rules for Claude, such as not helping with crimes or creating biological weapons, but we found that if you list such a set of rules, technically, what you're teaching the model becomes extremely inconsistent. It's just an analogy, but it's based on our actual findings: trying to teach a model 170-plus relatively arbitrary rules doesn't work well. We discovered it's better to teach the model some basic principles, such as what it should prioritize and how to reason when determining the right course of action.
Of course, I think this is also the right way to educate humans. If I just hand someone a list saying you must do this and cannot do that, I don't think that cultivates a firm moral sense. If I tell them how to think when facing situations, that they should learn to think independently and reason based on principles—like open exploration, caring for others, and fairness and justice—then everything else can be derived from these principles. Honestly, that's how Claude learns, and we've found it works surprisingly well because it allows Claude to extrapolate and handle new situations it has never thought of before.
Daniela, when you mentioned your mother is an expert on Kant, I couldn't help but ask, have you really delved into this kind of philosophy and moral philosophy?
Interestingly, the woman responsible for shaping Claude's personality is a PhD in philosophy, which is quite unusual. You might think the person training the personality of these AI models would surely be a computer scientist, or someone who has studied machine learning and AI for over a decade. Amanda is indeed technical, and she is in charge of training the models, but when you interact with Claude, you can clearly feel a very wise, best-of-humanity quality in its personality as it tries to answer your questions. If you probe Claude in those complex areas where different philosophical schools might conflict, you can feel Amanda's influence quite tangibly. You'll see it answer things like, "You could consider it from this perspective, but you could also think about it from that perspective." I think this is important because as the world becomes increasingly reliant on these models, you certainly don't want to be locked into one single philosophical tradition, given that ethics and morality are very complex. Different people and different cultures may have drastically different answers as to what should be done.
Dario, you were there during that defining moment when GPT-1 and GPT-2 arrived and shook modern society. It might have even surprised you. You were the head of research at OpenAI at the time; what did that moment feel like?
I think GPT-2 was perhaps the first time we truly verified, with actual results, the principle that scaling up makes models more capable. I wasn't very involved with GPT-1, but I was deeply involved in the idea that we should scale up the models, get more data, and invest more compute to run a larger, more expensive, and more ambitious experiment. The idea back then was that if we increased the scale of these language models, their linguistic abilities would improve and their logic would become more coherent. It was just like those key moments in scientific history: you propose a decisive experiment, make a very strange prediction, and then when it comes true, you think, "My god, we've really started something big!" So, I see GPT-2 as the first real test of the scaling law hypothesis. What I find incredibly interesting is that people's reactions to this were vastly different. Were you surprised at the time?
I was personally terrified. I looked at it and thought, "My god, if we scale this thing up a million times, it will be better than humans at everything." This is the most defining moment for modern society in the economic and scientific fields, only people haven't realized it yet. This will determine the fate of various nations and economies, and it will also determine the future of humanity. It was like staring into a crystal ball and seeing the whole future unfolding; it really scared me. I exclaimed, "My god, what does this actually mean? Is it appropriate to develop it here? What happens next? What are we going to do with this thing? Who should we tell about this?" At the time, only about 10% of the people around me had the exact same reaction as I did, and they were often the smartest people.
As for the remaining 90%, because GPT-2's capabilities in specific domains weren't that impressive yet—it could only translate, summarize documents, or occasionally write a few lines of code. Its terrifying aspect was its breadth and generality, but many people just looked at it and said, "In terms of single-task ability, it doesn't look that smart." So many people just shrugged and didn't take it seriously at all. Take our recent announcement that we're using AI to develop drugs for neglected rare diseases; you'll hear those who have been in biotech for years, long jaded, making the same old comments that tech people tried this before and it didn't work. They list the model's current limitations, but looking at these comments, I feel history always repeats itself: the limitations you cite today will be trivial six months from now, and in twelve months, what seemed impossible will be handled by the model as a daily routine. Everything we saw with GPT-2 is playing out over and over in every corner of the world, across increasingly wider fields, and the whole world is finally, slowly being convinced.
I think you're being quite polite about it. I remember back then, there were highly respected computer scientists and leaders in the field of AI and machine learning who were practically bullying the researchers behind this model on Twitter, saying, "This is so stupid, the model is so weak, I don't know why you're even talking about it." Back then, if you thought this was a momentous occasion, people would treat you like you were crazy. Now, everyone has selectively forgotten these things; those who once said it wasn't scientific or reasonable are all applying for jobs at large language model companies. Though we would never call anyone out to their face, we look at them and think, "We remember what you said in 2019, and you have no credibility on this matter anymore." Mira, you were the VP of Safety at OpenAI at the time; when you encountered GPT-2, I assume you were as freaked out by it as Dario was. Did it cause you any panic in terms of safety?
It wasn't the GPT-2 model itself that worried us; we didn't see any cybersecurity threats or anything like that when we looked at it. In some ways, the model was laughably wrong, and it remained a very harmless model. But like Dario mentioned, I also had a feeling—just looking at the gap between GPT-1 and GPT-2, and projecting that trend into the future—it was like if you had only ever seen adults and never a child, and you saw this little thing walking around, you might think, "I wonder if this little thing will have real potential when it grows up." But then one day it starts walking, and you think, "Wow! I bet this little guy is going to be able to do a lot of things that adults can do. This may well be the future direction of this technology."
Dario, you said you left OpenAI because of a breakdown in trust with Sam. That's a very interesting term. Can you explain how you felt that trust had broken down there?
I want to say that leaving OpenAI to start Anthropic was ultimately because we disagreed on the principles of how to build this technology. That is the fundamental issue. Over the past five and a half years, everyone has been able to see just how different our principles are. You'll see that Anthropic has always advocated for sensible regulation of this technology and for setting up safety guardrails. Anthropic invests significant resources into mechanistic interpretability—that is, looking deep inside the models to figure out how they work. I think Anthropic has succeeded in maintaining a cohesive corporate culture with one of the highest employee retention rates. We haven't had the kinds of scandals that other AI companies have, and I think that says enough about our differences.
But as for the moment I realized we had to go off on our own and that OpenAI was no longer the right place for us—safety issues are extremely complex, and it's normal to have many disagreements on them. I also often disagree with my co-founders and many people at Anthropic, but we are able to resolve these disagreements because everyone is acting in good faith. For instance, whether we should release this model or wait, whether we should set stronger or weaker biosafety levels for a model, how much energy we should spend lobbying the government for safety regulations, and what those regulations should look like—these are all good-faith disagreements that always get resolved because we trust each other, believing that everyone wants to find the right answer and that everyone is being honest.
The reason I mentioned a breakdown of trust is to illustrate an abstract point: if you don't feel that at your workplace, if you don't believe that disagreements are made in good faith, and if you don't believe that others are telling you the truth, then how could you ever resolve those disagreements? This is the reason we left, and it is the reason we went out on our own. I think the experience of the last five and a half years is the best proof itself, proving that our philosophy is completely different from other companies in this field. If you continue to work for people who don't share your views, you lose track of what you're actually standing for, and you essentially end up bearing the burden for mistakes you didn't make.
Wait a minute, I have to push back on that. What you mentioned are differences in values and the like, but that is fundamentally different from a problem of trust. You feel that some people say they share your values, but they actually don't, which makes it impossible to trust them.
Honesty is one of our values, and it's a very important one. If someone lacks the value of honesty, it's hard to know what their values really are.
You think Sam is that kind of person.
For me, I felt that our co-founders were an extremely political group. We all had a very consistent vision for where we wanted this technology to go and how to build it, and at some point, you realize it just makes more sense to go out on your own. In an organization, if others don't genuinely identify with your vision, how can you expect them to implement or share it? No matter how good it sounds when they speak, if it isn't reflected in the company's actual actions, what other choice do you have?
