Anthropic 总裁 Daniela Amodei 谈 Claude 的未来
Anthropic President Amodei on the Future of Claude
We were chatting about not so long ago >> >> when we first met and started covering Anthropic. Um but for those who aren't familiar, you and your brother Dario and a handful of other people left OpenAI to start Anthropic back in 2021. Now you employ thousands of employees, many of them in reporting chain up to you um in your role as president. Uh I'm curious, what is the difference in the leadership style between you and Dario?
>> Well, first of all, thank you so much uh for having me. It's great to be here today. And you know, Dario and I I really think of how we sort of split duties. Dario's this incredible technical visionary. I think um you know, years before AI was this amazing idea that people were excited about around the world, he had this really strong conviction that artificial intelligence is going to be a big deal. This is like early 2010s.
And I think Dario is constantly reminding the company about the scale, the ambition, sort of where um the technology itself is going. I think of my role much more as actually running and managing the day-to-day of the organization. So, I manage the executive team. We spend time making decisions about customers, about our products, about the research and how that applies to the people that are using uh artificial intelligence.
And I like to think that the two of us are are a great balance. Um and I have to say like I don't know how one CEO just runs a company by themselves. It seems like it's actually much easier to do with two people, particularly when they know each other really well. >> I remember um in one of our conversations earlier you telling me that you and Dario rarely fight. >> >> And I'm curious how you manage What are What are some examples of places where maybe you have disagreed and have been able to settle it in a healthy way with regard to the business?
Well, as someone who also has a brother, I know you know, um you know, I think we've had like, you know, several decades to kind of practice fighting and getting and still loving each other, right? So, as early as like, why did you take my toy? Can I have that back, please? I think that has served us well in being able to operate well together at Anthropic. I think what I would say is, you know, we do disagree sometimes, but I think how you disagree really matters.
And I feel like there is always like an incredible mutual respect between us. So, if we look at the same situation and we see something different, I like to think that what our default is is to be curious, right? So, I usually view it as like, oh, this felt like an obvious call to me, like, why do you see it that way? Because usually he'll have a piece of information that I don't have or he'll have a different perspective.
And I would say about 80 to 90% of the time what ends up happening is we meet in the middle. And so, we'll think like, oh, I think this is the right thing to do, you think this is the right thing to do. Actually, there's probably a third way to do this that's the combination of what we both think that's actually a little bit better. I think that's part of the strength of having both of us. The ability to practice looking at at situations and decisions just from different points of view and perspective, and I think it helps us make better decisions.
>> [clears throat] >> So, 2 years ago when you were sitting on the Bloomberg Tech stage with Dario, Anthropic was still very much the underdog in the competition against OpenAI and just in the race to AI generally. Now, it's in a very different position. No longer the underdog. Claude is a runaway success, particularly among coders. Your company is projecting $47 billion in annualized run rate revenue, and the company has eclipsed OpenAI's valuation for the first time.
Do you feel like you're definitively the frontrunner right now, and how does that change your approach? >> You know, something that I think both Dario and I, you know, really hammer at Anthropic is that that is just completely the wrong way to think. I think our view has always been, you know, this is really about how do we show up every day for our customers? How do we stay humble and make sure that we're really focused on the mission, right?
The whole reason we started Anthropic is we want to be able to build and develop this technology in a way that is ethical, that's responsible, that's fair. And I think it's really incumbent upon everybody at the company, but especially leadership, to say, you know, all of these numbers, they're actually not the point, right? We are really here to do a job. Um we're here to support the businesses that rely on us every day.
AI is increasingly a bigger part of of everybody's workflow, their experience at the office, um their experience in their personal lives. That's a huge privilege and responsibility. And so I think if anything, you know, we work really hard every day to say like our job is to be the best version of of who we can possibly be as a company every day. That takes work, that takes humility, that takes time, that takes focus and energy.
>> I want to talk about the big news earlier this week that Anthropic has filed confidentially for an IPO. Um SpaceX has also filed, OpenAI is expected to file anytime now. How far back had you been planning to file? And do you think that there's a race here to be first among the AI companies? >> So uh yes, you're correct. We confidentially, you know, filed our S-1, um and that gives us the option, uh you know, to potentially go public after the SEC review.
Unfortunately, that's all I can say about about anything uh IPO-related. I'm sure you understand. >> I mean, um if we can just maybe level up a little, what is the calculus for a kind of an AI company generally to go public or not in this moment? What are some of the pros and cons here? Because on the one hand, yes, you can access more capital um if you're not a startup. But on the other, then you're beholden to you know, quarterly results, shareholder calls, etc.
So what what are some of the pros and cons just generally for an AI startup, do you think, to go public in this moment? >> So, I think you know, at least speaking for you know, ourselves and I think ideal probably really for the AI industry more broadly, it's a very capital intensive business to to train AI models. I think that's something we've been you know, very open about from from day one. I don't think that's you know, a surprise in the industry and I think this kind of concept of like how do you access the level of capital that you need to train these models, right?
It's a it's a really big kind of upfront cost to train the models and to serve inference on them and I think that is you know, my guess is that you know, over time the sort of core set of companies that are working to advance the frontier are just going to need access to capital and I think the public market is very well suited to that. So, there's obviously of course trade-offs, but my sense is there's just a fundamental structure of how training and kind of serving these models for customers works that it will require that level of of access.
>> It's clear that many businesses can't get enough of Claude. But at the same time, we are hearing about some customers who are concerned about the pitfalls of so-called token maxing of you know, companies making leaderboards of who uses the most AI tokens for using AI's sake. Uber's one of Uber's executives talked about it getting harder to justify AI spend if there isn't a clear metrics backed ROI. Do you think there's truth to the idea that at that some businesses may be overspending on AI tools in this early experimental phase and they may cut back as they move out of that phase?
>> So, I I think kind of two thing two things are are true at once here. I think you know, I like that you said you know, experimental because I think the reality is it feels like these AI tools are so powerful and so capable, which is true and comparing them you know, particularly to, you know, even 2 years ago when I was when I was on stage at at Bloomberg with Dario, um I think the models have just come so far in in what they're able to do and the economic value that they provide.
But I actually think there's a lot more distance to go still for what the models will be able to do, you know, two to four to six to eight years in the future. And I would suspect that we will see some, you know, some um experimentation in different fields to understand how do you kind of get the most value out of these models for what particular workflows. I think the second thing that's true is, you know, how businesses are choosing to use AI will change.
So the use cases today, um some of them I expect will continue to be kind of the the the primary driver of, you know, efficiency or um creativity or or, you know, new approaches to doing things within businesses, whether that's coding or um you know, financial services, legal, healthcare. But I also think um as the the industry as a whole, not just the AI the AI industry, but really the business community gets um more familiar with the tools, we're all going to learn together.
Like, what is the best way to apply these tools that also supports employees? And so I think today there's this there's this feeling that it's like, oh, like AI, I you know, the leaderboards and it's like I have to use it and what do I even use it for? And I think my hope is that over time, it'll be more incorporated into the day-to-day of how humans do our work, how we communicate together, and that there'll actually be a lot more value realized in a way that really good to people.
>> Does Anthropic have a leaderboard itself? >> We don't have a leaderboard, no. Or at least not in the not in the way you've described. We do track how much we're using Claude internally and for what use cases, because we really try to prototype all of the tools that we give to our customers. And so all of the products that, you know, you see, whether that's, you know, Claude Code, Claude Design, um Cower, like of these are things that we actually built at Anthropic first because we said, "Hey, this is a real need we have.
Could Claude help us with this?" And so, I think it's important for us to just have a kind of a a metric for for checking like, how are we using it and how much are we using it? But, there's not, uh, like, you must use AI and you must use Claude and it's better if you use it or not. >> What, uh, type of job function outside of coding or research, uh, are using Claude the most internally? >> The entire company uses Claude for, uh, for topics large and small.
Probably the second biggest I would actually say is finance. Um, and so, a lot of the kind of back-end financial, uh, planning analysis, uh, you know, like, just number analysis, I think Claude is is really, really helpful, uh, at that. But, honestly, like, our people team built an incredible tool internally, um, for how we do performance reviews that leverages Claude. We're in the middle of performance review season.
I have to say, like, the feedback from our team has been so positive. Um, Claude is like, it just feels so much more fun to do this. It's so much more interesting, interactive. It's pulling in information about what I did over the course of the past 6 months when I'm assessing myself. And I think that it really just speaks to the generality of these tools. Their ability to, I think, reach across a business and and just take a lot of information in and help you as an employee be more successful in your role.
>> Dario, Anthropic's leadership has been vocal about the need for more compute. Um, your company recently made a major deal with XAI to lease compute. There are others. OpenAI, uh, made a big splash sort of earlier on with securing huge data center deals and spending, you know, up to a trillion on it on, uh, on infrastructure. Why did Anthropic take a little bit of a different approach? And in hindsight, do you think you should have done more earlier on on the compute front?
>> We've talked publicly about this concept of sort of the cone of uncertainty related to compute. And I think something that's important to know is you the the sort of structure of these deals is you have to commit to a certain amount of compute, you know, reasonably far in advance. And so I think Anthropic's view has always been we are we are wanting to, you know, plan for the best outcome but not overextend ourselves such that we're buying more compute than we could productively use.
It's really hard to predict that, you know, perfectly. And I think we would much prefer to be on the side of having a little bit more demand for the product than we're able to serve than the inverse where you overshoot and then you're actually like not in a great situation because you've bought something you can't pay for it down the road. It is really hard to get this right. I think the industry as a whole is still grappling with how to kind of structure these deals, but I think my sense is it's it's impossible to know what the future will look like.
On balance, I I continue to think it's better to be fiscally responsible and think, you know, carefully about how much you're going to buy, make sure that you have the ability to actually use all of that compute in the future, and we'll probably undershoot or overshoot at some point a little bit. >> In the announcement with SpaceX, Anthropic expressed preliminary interest in using potential data centers in space.
What do you think about that? Is that a reality and how far out are we from that? >> [clears throat] >> Um I don't think that data centers in space are something that will be will be on our our to-do list in 2027. But I you know, I will say AI is the field that has surprised me the most. It's probably surprised the world the most in terms of just what new things are possible, are made possible by the advent of this technology.
So no immediate immediate plans for for working with astronauts to to get space center data centers going, but you never know. >> A few months ago, Anthropic had a very public battle with the Pentagon over restrictions on its AI software. Are you more or less optimistic today about finding the solution to that? >> Anthropic has been really leaned in from day one and talked very publicly about our commitment to national security.
And we were the first AI company available um on the top secret cloud. I think our commitment to, you know, these values and these principles are also very very old, right? It's sort of not not a new not a new thing for us. But I have to say I've been very impressed the degree to which we've been able to work productively, you know, with the administration around a wide variety of topics. And I think that's actually been the been the much bigger story over a long period of time is our ability to to partner productively with government at all kinds of of levels because fundamentally I think artificial intelligence is and will be a geopolitical issue.
And in order to be an ethical and responsible we need to work with governments around the world in the US but also in partner countries to say, how can we how can we roll this technology out in a way that's safe, in a way that's good for people, that's going to protect democracy? And so I I do really feel optimistic about this in the long run. >> Do you think other companies like OpenAI or Google have been able to negotiate better deals with the government because of Anthropic sticking to its position?
>> You know, I think every individual company is going to have their own stance and sort of their own principles about like what what their red lines and what their values are. And my sense is Anthropic will always be, you know, as open as we can be about what our principles and values are, why we have those principles. But ultimately I think it's important that whatever those values are for you as a company, that you are true to them.
That you feel like you can explain them to employees and to the world more broadly. And fundamentally, you know, we can't really control what what other businesses are going to do. I think Anthropic's role is to be the sort of best version of itself. Again, to say like if if we were the only organization in this, you know, in this situation, how would we want to behave? I think that's that's really the best way we've been able to navigate many difficult decisions at the company over over the years.
>> I think philanthropy seems to be a big part of Anthropic's culture. I know you and many Anthropic co-founders have pledged to give away half of your equity. 80% Sorry, 80% of your equity. I'm curious, would you be in support of something like the proposed billionaire tax ballot in California that would impose a one-time tax on individuals whose net worth exceeds 1 billion? >> So, Anthropic has, I think, again, like pretty publicly talked about the fact that we think AI is going to create a huge amount of wealth.
We're already seeing that. And I think we've been pretty open about this idea that that should be redistributed in some fashion. Both the pledge and then a lot of sort of views of Anthropic when we're talking about, you know, potential for labor displacement are really grounded on this belief that if we don't take intentional actions, AI, like many technologies, will will probably widen the the the gap of inequality and disparity.
That's not what we want to see. So, I think there are a lot of interesting suggestions kind of floating around for for how to approach that. Of course, the part we can control is, you know, what do we as a company and as individuals as co-founders sort of choose to do with, you know, any you know, economic benefits that we get from AI. But, I think this is a really big and important question that actually goes beyond the walls of any individual company.
>> There's a lot of potential upside with AI to cure disease or lower the cost of consumer goods, but also real fears that, you know, Dario has validated saying last year that AI could potentially wipe out half of all white-collar work. Do you agree with that? And And if so, what's the solution? Is it something like um more taxes? We talked about basic income, job retraining, and how do you pay for all of that?
Sorry. >> I think AI is we're already seeing the ways that it's disruptive. I don't necessarily think we know the future. So, exactly what the type of disruption that might happen um you know, a year or 3 years or 5 years from now is unknown. I do think it's important for companies to be open about and study what we're seeing today, which is part of why we publish our societal impacts research. We say, "Hey, here is how people are are using AI.
Is it displacing jobs? Is it supplementing jobs?" What we found so far is that in 2025 and 2026, replacement is a tiny tiny tiny fraction >> [clears throat] >> of what AI is doing. And really where where we see that is mostly in jobs that are overseas and mostly in customer support. And so, these are jobs that are sort of already being automated by more traditional ML, you know, non-generative AI systems.
Could that change in the future? Absolutely. And I think we're a little bit unusual in the sense that we talk about it a lot. We say, "Hey, this might happen." But I think there is this just broader again sort of societal challenge. AI is going to be able to do so many productive things that humans can do today. What does that mean for how we find meaning, how we earn income, how we relate to one another? And again, I think the default will be to treat it like past technologies and say, "Okay, we're going to integrate this into existing workflows and that's sort of the end of the story."
I personally believe there's a real opportunity here for us to say, "How do we accelerate and sort of accentuate the parts of doing work and finding meaning that only humans can do. I have this very fundamental belief that humans like to spend time with other humans. We like to create things together. We like to relate to one another. Sometimes we like to disagree with each other. And I don't think AI will fundamentally take that away from us, but we have to figure out how we apply that within the existing economic infrastructure so that people can still find meaning in their work and so that people have a way to earn a livelihood.
Again, I think there could be a lot of different ways to structure that, but I think this is an important moment for us generally as a society to say like, what are the lives that we want humans to be able to live and how do we work closer towards that future? >> Two months ago Anthropic unveiled and restricted access to Mythos, your most powerful model, right? Citing significant concerns about its potential to wreak havoc on critical software by spotting and exploiting security vulnerabilities.
Now Anthropic plans to release these Mythos level AI models more widely. Can you help us understand what's changed in the past few months? Are you confident that these models are different now in any meaningful way from the version before? Are they safer? >> This is a great question. We actually on Tuesday just expanded the set of customers that have access to Mythos. I think it's about an additional 150 organizations around the world in 15 different countries.
And really our approach to Mythos has always been this is a there's a time component to it. So we released it you know initially to cyber defenders. So some nonprofit groups, some governments, some organizations that are critical infrastructure for protecting you know against potential cyber attacks. And what we found is just like in any kind of security vulnerability situation, you have to give the defenders a head start.
The technology, AI models are going to keep advancing. If it's not us, you know, one day releasing a Mythos-level model, another AI company will. But, it actually matters who you give access to first and how long they have to patch some of the vulnerabilities that Mythos was was capable of of revealing. And so, we're taking this this very cautious, very tiered approach, which I know in some ways is frustrating because people really want access to the model.
But, I think again, as a company that's founded on on these principles of being ethical, being responsible, we felt it was important to give access to the organizations that were capable of really helping us defend against some of these risks, and then slowly widening that circle to more and more critical infrastructure until eventually we feel it's safe to release it more widely. >> So far, Anthropic has been very focused on enterprise, even in how you're talking about that rollout.
But, could we see a push into consumer this year? >> So, we do have a consumer product. It's it's Claude.ai, and you know, I think we have always, really from from day one, felt that, you know, enterprise and and businesses is the best kind of spiritual fit for Anthropic and our and our values. I think this focus on trust, on responsibility, reliability, transparency, these are just so baked into the DNA of Anthropic the company.
And I think, you know, we have you know, increasingly growing consumer base. It mostly looks like, you know, professionals and individuals that are using AI for productive uses. So, it doesn't necessarily have to be, you know, work. Sometimes it is, but it's often for, you know, it advancing your own skills and knowledge, whether that's in a hobby or helping you know, I'm a mom, right? Helping me like organize applications for preschool when my son was younger.
I mean, Claude is so powerful in being able to just help help abstract away some of the the administrative duties of just being a person. But, I think the difference to us in our consumer product, maybe compared to competitors, is we're not It's not an entertainment tool. We're not using it for people to sort of have fun and and have it be this like It can be fun to use it, but it's really for kind of productive activities, whether those are those are at work or or at home.
And I think I believe that is going to continue to be you know, a reasonably large number of people that wants to use that that product and service, even though enterprises are primary focus.