It feels very important that we sort of have this concept of a human in the loop to also just make sure that the models are not kind of veering into a direction that we don't want them to be.
Are we letting the models be 100% fully autonomous for tasks or is there a process of being able to kind of bring humans and AI together?
Hold light and shade is one of our values and it's this concept that the technology itself has great risk and great opportunity and that that is a very complicated, complex thing to hold. We have to consider and talk about all of the risks, all the things that could go wrong, while also considering, thinking about, and building towards a world where all of the things go right.
My sense is given the locus of the technology industry being here, the kind of orientation towards creativity, it doesn't surprise me that artificial intelligence first took off in the Bay Area. There's accountability that the non-tech residents of San Francisco kind of put on the technology companies that I don't necessarily think is bad. They're like, 'Hey, we're providing a home for you. We just want you to be good citizens.' Over the course of the past few years, there's been a lot more mutual understanding that I think has happened between what I'm describing as kind of old-time San Francisco and tech.
Oh my gosh. Thank you for having me.
It's a pleasure. Good to see you again.
Can you tell people what room we're in? I mean, you can't really see the books on camera, but like we're in a library. It's amazing.
We are in a library. Yeah. Anthropic is such a unique company and I feel like a library speaks to the research roots of Anthropic. But in general, I think we're very well read. Like people at the company...
Oh, interesting. Very multidisciplinary.
Like I think when we first started out, one of the things that like most struck me was just how many different disciplines people came from who came to the company. We had like political scientists and ethicists and biologists. I was a literature major. As you can see, the literature major has like shown.
Yeah, you definitely did.
I actually encourage you if you want to borrow a book on your way out, go peruse.
It's like kind of like a pass to come back in and return it, which is nice.
Okay, so it's January 2026. I feel like even if we drop this podcast tomorrow, so much stuff will change between now and then, whether it's geopolitics or the economy or tech or AI. Tell us where we are. How do you think about it?
Yeah, like light, light easy question to start with. You know, I think one of the challenges of working in this field is it's much easier to look backwards and talk about an epic that happened. But as it's happening, sometimes we don't even necessarily expect or fully understand where we are in relation to kind of what will be true six months from now or three months from now. But I think there's kind of a few things that feel especially present or happening right now.
One is at least internally we feel like the last class of models that we released was a big step up. So Opus 4.5, just in its capabilities, the level of intelligence of the model, but also the way that it's now being used in the world. I think particularly on the coding dimension, you have developers who are now saying like, 'I'm running Claude Code like 100% of the time that I'm working. It's making me twice as productive. It's doing most of the low-level tasks for me.' We're also seeing quite a lot of adoption in biological sciences and healthcare, which we've always had this dream. It's something we talked about really from day one.
This big dream, this big idea we've had, but now as we're doing company planning, we're talking about, hey, can we contribute to helping to eradicate polio in 2027 and 2028 if we partner with groups that are doing it. So, I think probably for us, the biggest thing I would say right now is just the step up in the capabilities of the models. And then, you know, all of 2024, 2025, agents was kind of this word that was thrown around. I almost feel like it lost meaning, but I think this is truly the first time we've seen Claude capable of significant agentic abilities.
Okay, so Claude is able to actually go onto your computer and move your files around or open them or process them or delete them or restructure them. That was just not possible literally in December. So the difference between 2025 and 2026 in the course of a month has felt like a big increase to us.
Yeah. I feel like at the end of the year all the kind of wrap-ups of the year were like, 'Oh, agents are actually disappointing. This was not the year of agents. We got years and years to go.' What changed between a month ago and now?
You know what I find interesting is again, like, do we really know? No. I would say this, I don't know if prediction is the right word but we have this sort of belief in this concept of scaling laws. You give more compute to the models, you give them more information, they get smarter at a predictable rate and we couldn't have predicted in advance when we would have crossed a particular capability threshold. So, we wouldn't have said, 'Oh, we know like this model or at this date Claude will have these capabilities,' but I think as the models just get better and get smarter, we're able to do some training stuff to actually help empower some of those capabilities. But I think more broadly, what we're seeing is just like watching a person get smarter over time, get more education, go to college, learn more things, and I don't necessarily know that we're going to be able to predict every future positive thing the models are going to be able to do.
I know everyone's doing this, but you're very organized about how you think about adoption and how it's being used and where it's being used and for what. And you were just talking about some of that stuff. You all released this economic index report. It told you something about rich and poor in terms of countries. Also told you something about adoption within the United States. Can you talk about some of that and like what conclusions were you drawing from that?
I think what's interesting is like most technologies, AI is following, I would say, a mostly predictable adoption path. Which is that unfortunately the folks that have the most are going to be the first to adopt it. They have the most resources. They have the most time. They have the most access. And so they're going to be the folks that say like, 'Cool, I'm a computer scientist at MIT. Of course, I'm going to start using Claude Code.' But I think what we're seeing is the rate at which the technology is being adopted is significantly faster than anything else. If you look historically even at things like the adoption of Google or when people had dial-up for the internet, AI is just going much more quickly in terms of adoption rates. The trend is the same but the speed is faster.
And so I think what we're going to see over time is a broader distribution of adoption than perhaps with previous generations of technology and I actually think the same will be true internationally as well. But I also think that there's work we have to intentionally do to make sure that there are not big regions of the world that are left behind because of how quickly the technology is developing. Being one year behind or two years behind could actually have an incredibly big impact.
Yeah. How are you thinking about that? Like the theme of you knowing you a little bit, following you guys, a theme of Anthropic is there's the slogans and there's the culture and the good stuff and you're known as the ethical company. That's what you're actually trying to back that up with concrete in the world stuff, to use the technical term. On that point in terms of like adoption in countries where it's just not happening. How would you do that?
I think there's a couple of different ways that we approach this. The first is we view it always as our responsibility to state the truth and be transparent about what we're seeing. And I actually think that's a really important step. One, if you don't understand the problem, you can't do anything to make it better. And so I think it's often a little bit unusual or maybe even a bit unique about Anthropic that we prolifically publish papers and reports about the trends that we're seeing. And that's everything from adoption like we've talked about, where the technology is being adopted, potential risks that we see, suggestions for policy about how to address some of these issues because we alone cannot solve the problem. And I actually think corporations alone together can still not solve the problem. It's going to take action from civil society groups, from government.
Yes, we do personally try to take some responsibility for this. Anthropic has two different teams that work on something we call beneficial deployments. And this is essentially our efforts to deploy the technology for things like biological and life sciences, healthcare, developing world. So we have a number of pilots with the Gates Foundation. We also work with the Clinton Health Access Initiative and we're running many different experiments to see how we can just approach potentially deploying Claude in developing countries.
Got it. All right, let's talk about Anthropic, the Ethical Company. So I think everybody who's listening to this will already know what constitutional AI is. But spend a minute on what do you actually mean by that?
Constitutional AI is this idea that our researchers sort of invented and the reason we came up with it is that the previous techniques used for training LLMs essentially was a reward function system. So basically if you ask it a question, if it answers the question correctly, you give it a reward. If it answers the question incorrectly, you don't give it the reward. Sometimes you even give it sort of a form of punishment, is the wrong word, but you say, 'Hey, that was the wrong answer.'
We felt that there was just a better way to kind of approach how the model learns and it evolved into this concept of constitutional AI which is essentially rather than sort of saying, 'Hey, this is a good answer, this is a bad answer, this is an ethical answer, this is an unethical answer,' we actually gave Claude a system, a framework for thinking about ethics. So we trained it using a variety of different documents. I think most famously in the original constitution we put in the UN Declaration of Human Rights. The Apple terms of service are in there. We recently published Claude's constitution. But I think essentially we're sort of trying to say like we really want the models to understand a broader system of ethics. So not just what is one particular question, how are you answering it? Is it correct or incorrect? But how do you actually give it a framework, like a moral framework to say how should people be treated? How do you make sure that you're not discriminating against people? How do you make sure that you are thinking critically about how you approach complex ethical questions? And what we've seen is this, over time not only increased the ethicalness of its answers, but it increased its overall intelligence and capability, too.
How do you judge if that's working? Can you measure that? I mean, it's ethics, so it's famously subjective, but how are you determining whether or not it works?
So, there's two different mechanisms that we use. The first is definitely looking at kind of performance compared to competitors. But the second is there's an evolving sort of set of benchmarks that are used around a variety of different measures of model capabilities but also sort of models' abilities to grapple with complex issues. And it's really tricky because the benchmarks are pretty outdated. They were sort of made for a different era of LLMs, even kind of prior to LLMs. But what we're attempting to do is have a system of benchmarks and evaluations that are kind of agreed upon in the academic community to say how are we actually going to measure these models. And I think it's a little bit of an imperfect art but Anthropic has invested significantly both in measuring our own models but also trying to contribute to say how should we have these benchmarks kind of exist in the world.
Got it. We're kind of in the pre-collective action era in terms of like how AI should live in the world. And right now, I think one of the things that you guys talk about is constitutional AI ought to result in a race to the top, right? Where other companies hopefully or other players are adopting because they see that it's working. They see that it benefits you guys, so they'll want to do it and you're in favor of that. Where are you seeing that?
You know, I think race to the top is a really useful concept that we've actually applied across a variety of different ways that we operate. So I think one of the most classic ones is many of the guardrails and the safety techniques, of which constitutional AI is one, Anthropic pioneered and what we saw was that this drove demand for our products in the market and I think in particular enterprise customers really care about having a lot of these safety, security, reliability features in place. And so a lot of the time what would happen is an enterprise would sort of say, 'Wow, we're choosing to use Claude, we're choosing to partner with Anthropic because of these techniques that you used or because of these guardrails that you put in place.'
And what we found was then competitors said, 'Wow, in order to be able to compete, we have to be able to offer these kinds of guardrails.' I think we've also seen it happen in some other ways even at kind of the more basic research level. Anthropic has an interpretability team. This is basically the team that you can think of as like Claude's neuroscientist. So trying to understand what actually happens inside the models. One of my co-founders, Chris Olah, is sort of the pioneer of the field and we talk about it with businesses. We talk about it in our research papers. And suddenly, people are coming to Anthropic to work on our interpretability team, but they're interviewing at competitors and saying, 'Hey, if you had an interpretability team, I might be more interested in working here.'
And what we've seen is interpretability teams now spinning up at competitor companies. And so I think this idea that if you can find the intersection of something that is actually good and is good for business, of which there's a lot in our view, that actually drives competition in a healthy and positive way for the market that just raises the waterline for everybody.
Are the businesses, the enterprises that are coming, your clients, are they self-selecting in a particular way where you're getting certain segments of corporate world versus others? Sort of like you're getting the lawyers but not the crazy cowboys.
You know, so I think a really important dimension of this is that Claude's capabilities are also extremely powerful. And I think over the course of the past several years we've arguably had the strongest models for the majority of the time. And so what we've seen is particularly in areas like coding and agentic coding that cuts across industries. So, if you work in financial services, in healthcare, in the technology space, a legal law firm, all of those areas have a need for technical development tools. Anywhere that's going to have a set of developers, people that are writing code, they're going to be using Claude Code. And so, there's an enterprise demand to use our product sort of regardless of industry.
And I think that also just speaks a little bit to both the specificity of this kind of artificial intelligence skill. We talk a lot about artificial general intelligence, but like the models are really really good, for example, at writing code, but then there's also a set of general use cases that I think apply across a lot of business sectors. And actually what we've seen is pretty rapid adoption, unsurprisingly, most quickly from technology companies because again they're going to be the most tech-savvy. They're going to be reading the news about AI. But increasingly, especially over the course of the last year, I would say many more enterprise businesses have understood the need to have an artificial intelligence strategy. And I think in addition to having capable models, having this very strong safety orientation, Anthropic is also just well oriented to being a business-to-business company. We care about partnership. We integrate directly with a lot of the companies that use Anthropic. We serve over 300,000 businesses worldwide and we've just gotten a lot more practice actually working closely with customers in the enterprise space.
I love that. I was a transactional lawyer for a lot of years and I think about my training. This question is about like are we in the age of the last experts? Because I think what people say is we're always going to need humans. We're always going to need judgment. We always need people like you sitting on top of organizations with vision, with judgment, with ethics, with the ability to communicate. But how do you train that? And like the way that you in all likelihood developed judgment and continue to develop judgment, the way I developed judgment was like getting my reps as a young person and doing the drudgery honestly. And if we take that away, how do we make sure that people can develop judgment?
I think we as a society are entering a really interesting period. What does productive work mean? Just the nature and the shape of it. I think in the same way that before the internet, which I'm like old enough to remember a time before the internet, and what you did for work every day looked really different than what you do. The difference between 1990 and 2010 was dramatic. We're kind of compressing that into a much shorter period of time in terms of the rate of technological development in artificial intelligence.
But I think fundamentally, what's interesting is I should start with a caveat which is suddenly all these developers are saying like, 'Oh my gosh, Claude is so good at doing my job. This feels a little scary.' And as kind of a people leader and a manager, I'm like, 'Oh, yes, of course. This could impact your work, but like Claude can never do my job.' And then I'll ask Claude to help me coach a report through a challenging situation. And I'm like, 'Wow, Claude, that was actually a really good answer. My report could have asked me that question.' And so, I do think it's a reckoning that we're all going to have to face together. But I will say I think we are underestimating the degree to which AI and human compatibility and partnership are going to be how we walk into this new world. So much of what we've seen even today, even with the incredible capabilities of something like Claude Code, is developers are better with Claude Code than Claude Code alone or the developer alone.
Will that continue forever? Of course, we don't know. Will there be some disruption? I suspect yes. But I think what we're really seeing is that the intersection of AI doing what AI does well and humans doing what humans do well is actually going to be the way that we sort of pull everybody up.
And you know, I'm also remembering we did like a leadership offsite and our chief product officer at the time drew this kind of diagram of like what a developer does. And he said actually like hands-on keyboard coding is about 20% of their time. So even if you imagine...
This was when? This was like before...
