Take me back to the beginning when you were still forming the bones of what would become this company. What was happening in the world? And what problem did you think that Anthropic was uniquely poised to solve?
So, it's just about to be Anthropic's fifth birthday later this week. The way that we got started is my six co-founders and I were all working at OpenAI together and we had worked together on a range of different projects from scaling up some of the biggest models, so GPT-2 and GPT-3, a lot of the early language work that became part of the LLM revolution, scaling laws work, and then a lot of work on the technical safety side, to work in interpretability and alignment. And it felt at a certain point like we had this really clear vision of the company that we wanted to build: a technology company developing transformative AI at the frontier, but with this very intense focus on making the system safe and reliable. And it felt at a certain point like it just made more sense to go do that on our own, which is how Anthropic was founded. To set the scene, it was the winter of 2020, beginning of 2021. We were all locked in our houses, it was still peak pandemic, and I think the opportunity felt both very scary and very exciting at that time.
There's this almost cinematic moment that you described in January of 2021. You're in Dario's backyard, everyone's wearing masks, Eric Schmidt is there, and you're pitching him in the pouring rain under a tent. Take me into that moment. Do you see that as almost the origin story of the company, or where in the company's founding does that moment take us?
Yes, that's amazing. I think that was, I'm going to forget the exact date, but it was like single digits January. I believe we left December 18th and then single digits January. We're in Dario's backyard exactly under this sort of, we called it the party tent, under this tent that we had put up because it was raining. Eric ended up being one of our Series A investors, but it was really, we had kind of done this thing, made this decision. We wanted to go and start this organization, but it was still very early days and we had really no clear idea yet what exactly that was going to look like. We had a really big dream and big ideas. Another fun fact about that is I was actually eight weeks pregnant with my son. And I think only Dario knew of all the co-founders. I'm actually not sure he knew yet. There was just a lot happening. It was the pandemic, we were all wearing masks, social distancing, but we had this really big dream, this big idea, and there was just a lot going on for all of us both personally and professionally.
So in that moment, not in hindsight, what did you think that OpenAI was getting wrong that made leaving worth the risk?
I think the way I would put it is we really just felt more like we were running towards something than running away from something. What I mean by that is the co-founders, we had all known each other for various amounts of time and worked together over a lot longer than just our time at OpenAI. Dario and Chris Olah and Tom Brown had all worked together at Google Brain before OpenAI. Jared was a Herz fellow with Dario. Obviously Dario and I are siblings. I'd known Chris for like 10 or 13 years. We really had this long history together as a group of people with this like-minded set of values. We really felt deeply that artificial intelligence had this incredible upside, but that in order to realize that, we had to take the risks really seriously. At a certain point we thought, what if we could just found an organization like this where this is at the center of everything we do? And there was this inherent belief that this would also provide economic value and would actually be a selling point. Often there's this misbelief that safety and reliability come into tension with the business side. We actually had this, at the time it sounded very novel, belief that those two things actually were correlated and that they went together.
And you're the person who has to take 'we take safety seriously' into actionable strategies. What's keeping you up at night?
I would say there's maybe two things on the safety front. One is that there's a lot of really interesting technical safety work that is still to be discovered, and I think Anthropic has really been the leader in the industry, or tried to be, of doing that kind of work. This is everything from the work we do in mechanistic interpretability to constitutional AI. I just think there's so much opportunity, and our technical teams spend a lot of time figuring out what are the best ways to really build these guardrails into the models directly. There's always more work to do there, and the models are just getting smart so quickly. The second is really around what the impacts of the technology will be on society more widely. Again, I think this is a place Anthropic is a little bit unusual. We publish a lot of really compelling research about what the societal impacts of artificial intelligence may be. We published a report recently talking about potential economic impacts of AI and the labor disruption that might come along with that. The reason we try to be as open about this as possible is we really feel it is better to be ahead of potential issues, and as one of the actors developing this technology, it's actually our responsibility as a public benefit corporation to talk about it. We don't think we alone have the solution, but we do think we need to partner with civil society, with government, with people to talk about what it's going to mean when artificial intelligence is capable of doing so many of the day-to-day tasks that humans can do today.
It does feel like radical transparency is such a core part of your ethos. You mentioned that report, but you also put out data around agentic capabilities and what that means both in the context of a cyber espionage campaign that originated out of China where Claude was used, essentially weaponized, but then you also put out research that looked at the fact that when faced with a fatal scenario for Claude, it used blackmail in 96% of scenarios and other models turned to that as well. I bring that up in the context of you put it out there almost as a PSA of 'this is what the tech is capable of and here's what we're solving for.' So when you were course correcting from both of those examples, what was the most urgent thing on your agenda in order to solve for that?
I think we're again a little bit unusual. A lot of people will say it's a little bit unusual for a company to talk so openly about both the positive upsides, of which we think there are a lot, but also the real world risks and harms that can come from the technology you're developing. There are a few reasons we do this. One is as a public benefit corporation, we really view it as integral to our mission. It's important for us as a PBC to say there's a lot of potential upside. We really believe that Claude is going to have the power to literally help cure diseases one day, and in order to realize that potential, we have to get the tough stuff right. Our view is it's actually really advantageous for everybody the more we can talk about the risks because our goal is to prevent bad things from happening so that we can realize all of these positive benefits. The other is that more informed dialogue tends to lead to better outcomes. We're very fortunate that we have the access and information to be able to talk about the risks we are seeing as they emerge. We're able to say, 'Hey, Claude can be used as a cyber weapon potentially. This is something we really need to be aware of. And by the way, if this is happening to us, it's probably going to happen to the other frontier model developers.' In the safety, trust and security communities, a lot of the work can often transcend the companies. It's important to publish easy-to-understand papers about why certain trends are happening in the abuse space with the technology we're developing. We view that fundamentally as part of our mission. If you were a technology company of the past era, a social media company for example, if you could have gone back in time and said, 'Wow, if we had known what was going to come out of these platforms, would we have done anything differently?' Anthropic is really trying to say we're doing our best, we can't know the future, but to say if we thought that X or Y thing might happen, did we do everything we could today to talk about it and try to mitigate those risks?
I'm looking back the last six months and it's just been tremendous the amount of compute contracts being signed across the industry. Gemini has really closed the gap in terms of model performance, but one of the things that analysts point to is the fact that Google owns the entire stack, whether it's chips, the cloud business, the fact that it has all these different services to deploy their tech. Anthropic is now getting into the business of first-party infrastructure builds, committing $50 billion to creating these data centers in New York and Texas on top of existing cloud commitments. Is that a part of the strategy to win in the AI race? Do you also need to be an infrastructure builder and own more of that vertical stack?
It's really interesting. I think one of the challenges of operating in the artificial intelligence space is that the requirements of compute and associated capital are just so high if you want to train a frontier LLM. What we have always aimed to do at Anthropic is be as judicious with the resources that we have while still operating in this space where it's just a lot of compute that you need. What's interesting is Anthropic has always had a fraction of what our competitors have had in terms of compute and capital, and yet we've pretty consistently had the most powerful, most performant models for the majority of the past several years. I think that is due to the quality of our team and our orientation. Our desire to do more with less has always been a core value of Anthropic's. The compute requirements for the future are very large, so our expectation is yes, we will need more compute to be able to stay at the frontier as we get bigger.
Back of the envelope math, roughly $100 billion in compute commitments from Anthropic. Your competitor OpenAI is at $1.4 trillion. Speaking broadly about the industry, are we overspending on compute compared to what we can reliably monetize through LLMs, or is it pretty realistic the numbers that are being thrown out there in terms of what's necessary to serve the user?
It's interesting. A lot of the numbers that are thrown around are not exactly apples to apples because of how the structure of some of these deals are set up. Some of these are you have the ability to buy the compute and you have to commit to it early in advance. The fundamental thing that the industry is banking on is this belief that you have to invest upfront very early in order to get access to the hardware that you need to train the models in several years. If you asked one of my technical colleagues, they would say we have continued to be surprised, even as the people who pioneered this belief in scaling laws, this idea that more compute leads to better outcomes, that there is just this exponential graph that happens. Dario has talked about it from the performance of the models to the kind of revenue scale that you see every year. Something I hear from my colleagues a lot is, 'The exponential continues until it doesn't.' In our view, every year we've been like, 'This can't possibly be the case that things will continue on the exponential,' and then every year it has. It's an interesting question because we can't know the future. It's possible we could be wrong, but what we've seen historically is that every year-over-year performance looks reasonably exponential.
We've been talking a lot about these circular deals in the last few months where you have the model makers and the hardware suppliers and the cloud vendors kind of swapping cash around as part of these deals, whether it's swapping equity for chips, etc. What part of that is healthy flywheel behavior and where do you see red flags?
I obviously can't comment on deals not done at Anthropic or on the specifics of our deals, but I think not all deals are created equal and they all vary a reasonable amount. For legal reasons we can't go into the specifics of them. The different players have different approaches to how we think about compute and capital in general. I would point back to pretty consistently Anthropic has had a fraction of the resources of our competitors, and yet we've been able to do more with less. Our hope is that in the future, these model providers will be part of that flywheel. We're already seeing that as the first LLM to be offered across all three of the major clouds: Microsoft, Amazon, and Google. I suspect there will continue to be a lot of value creation that happens, particularly in the business realm. Anthropic is really focused on businesses first. A lot of what we've seen is that the demand for these models is actually hard to keep up with. At many points over the past year, although not all points, we've actually had more demand for Claude than supply to be able to provide it from a compute perspective.
And on the hardware side, how are you thinking about depreciation of some of these chips? Are you underwriting for a three or four year life cycle, or are you looking at sweating GPUs for up to 10 years for fully available life cycles?
It's a really interesting question. I'm not an expert on the chip side. One of my colleagues would probably be able to answer the specifics better, but my high-level understanding is that much like the LLMs where every frontier model is better enough that a high-performing company or individual would prefer to use the newer model, it's pretty similar in the hardware space. Next generation chips come out and they're somewhat more performant, somewhat more cost-effective, somewhat more efficient, and it's often really valuable to be able to access the newer chips.
I'd be remiss not to ask you whether or not we're in an AI bubble. I don't mean that the technology isn't real, but that the spending curve is outrunning the revenue curve. Do you think we are in a bubble?
I would distinguish between the technology side and the business or economic side. On the technology side, Anthropic and Dario have talked about this publicly recently. We feel very strongly that from a pure technological perspective, we do not see progress slowing down. It's impossible to know the future, but based on everything we've seen so far, the models just continue to get better and better, faster and faster at a reasonably predictable rate. On the economic side, it's a much more interesting question. Regardless of how good the technology is, it takes time for that to be used in a business or personal context. The real question to me is how quickly can businesses in particular, but also individuals, leverage the technology. The next Claude model, Claude 5, Claude 6, might be the same percentage better as the models before it, but the process of actually rolling that technology out within an organization might hit a wall because of human reasons. Change management takes time, it's hard to do procurement, it's impossible to know all the use cases up front. It is an interesting question to think about at what rate this will be disseminated into the economy and whether that will exactly match the rate that it has to date of how quickly the technology has accelerated.
With that in mind, do you think that we are overinvesting either in the model providers or in the hardware supply chain right now?
It's funny to put it this way. The dollar sizes are large, but the actual number of players everywhere is pretty small. I'm not entirely sure what to make of that. It's maybe a little bit unusual. What we've seen so far is returns to better hardware are high, and returns to having a slightly better model, a 0.01 increase, is also very high. We've seen that pretty reliably throughout our history at Anthropic. I don't necessarily know if I would say overinvestment, but I certainly think there's something about the limited number of actors. If something goes wrong in one link in the chain, that's a really interesting thing to think about what would happen.
Where do you think we are in that cycle potentially? Whether you call it a bubble bursting or just a basic correction in terms of all the froth that we're seeing, is this something that might be in the six to 12 month range? And if so, are you making decisions now to help Anthropic prepare for that kind of downturn?
Anthropic ourselves, we view ourselves as continuing to be thoughtful and respectful stewards of capital. That has always been our focus. Every ounce, every dollar of compute is important and represents either the ability to train better, safer models or the ability to serve more customers. I like to think that we have reasonably good, although again impossible to forecast perfectly, predictions for what we expect our model quality to look like, how much compute we need to train those models, how much compute we need for inference to serve our customers, and the value that we expect we'll continue to provide for them. My hope is that as a responsible business, regardless of what happens in the market, we would be in a good position. It's hard to speak for the wider market, but for us personally, we've always aimed to be responsible stewards of capital, and that will always be our goal.
