The scaling trends of AI are continuing. So I think we're going to see over the next year, you know, much bigger and more powerful models that are able to do greater tasks. In fact, by the time this podcast airs, a new model will be out from Anthropic that will probably be the most intelligent and powerful model in the world.

But one area I'm particularly excited about that we're developing in parallel with that is interpretability of models—the ability to see inside our AI models and see why they make the decisions they make. That area has been mainly a research area for the last few years, and it's just at the beginning of starting to have practical applications. So that's one area I'm very excited about.

If you look at what AI models do today, often you won't understand why an AI model does what it does. Consider your industry, let's say you want an AI model to be trained on some data to be able to predict what happened with a particular set of financial data. One problem you have with training a model to work on that is that if you train it on data in the past, the model might have memorized it because it was trained on it. It basically knows what happens; it knows the future in that case. Interpretability might allow you to tell the difference: is the model deducing the answer to the question, or is it memorizing the answer to the question?

Similarly, if a model acts in a way that, say, shows prejudice against a particular group or appears to do so, can we look at the reasoning of the model? Is it really being driven by prejudice? There are also a number of legal requirements, right? In the EU, there's a right to explanation. And so interpretability, being able to see inside the model, could help us to understand why the model does and says the things that they do and say, and even to intervene in them and change what they do and say.

You know, I wouldn't say solve. I would say we're at the beginning. Maybe we now understand like 3% of how they work, really. We're at the level where we can look inside the model and we can find features inside it that correspond to very complex concepts. Like one feature might represent the concept of hesitating, a particular genre of music, a particular type of metaphorical situation that a character could be in, or the idea of prejudice for or against various groups. So we have all of these features, but we think we've only found a small fraction of what there is. And what we still don't understand is we don't understand how all of these things interact to give us the behaviors we see in models every day.

So, you know, it's a little like the brain, right? We can do brain scans, we can say a little about the human brain, but we don't have a spec sheet for it. We can't go and say, "Well, this is why that person did exactly what they did." So will we ever understand fully how they work? I don't know about fully down to the last detail, but I think progress is happening fast and I'm optimistic about getting there.

That is a great question, and that is the thing contending with. So we are putting a lot of resources behind interpretability of language models to try and keep pace with the rate at which the complexity of the models is increasing. I think this is one of the biggest challenges in the field. The field is moving so fast, including by our own efforts, that we want to make sure that our understanding keeps pace with our abilities, our capabilities to produce powerful models.

To give some context, we recently released a set of Claude 3 models. They're called Opus, Sonet, and Haiku. They're different tradeoffs between power and intelligence, and speed and low cost while still being intelligent. At the time that Opus was released, it was actually the best all-around model in the world. But I think one thing that particularly made it good is that we put a lot of engineering into its character. People have generally found the Claude models are warmer, more human; they enjoy interacting with them more. Some of the other models sound more robotic, more uninspired.

We're continuing to innovate quickly, and as I said, by the time this podcast comes out, we'll probably have at least part of a new generation of models out. Right now, there's a trade-off between speed of model, speed and low cost of models, and quality. So you can imagine that as a trade-off curve, right, a frontier. There's going to be a new generation of models that pushes that frontier outward. And so you're going to see that things that you needed the most powerful model to be able to do, you'll be able to do with some of the mid-tier or low-tier models that are faster, cheaper, and even more capable than the past generation.

You're going to see models that are better at things like code, math, better at reasoning. One of my favorite is biology and medicine; that's one of the sets of applications I'm most excited about for the new models. The models we have today, they're kind of like early undergrads in their knowledge of many things, or like interns. And I think we're starting to push that boundary towards advanced undergrads or even graduate-level knowledge. And so when we think of use of models for drug development or, in your own industry, use of the models for thinking about investing or even trading, I think the models are just going to get a good deal more sophisticated at those tasks.

Anthropic thinks of itself more as providing services to enterprises than it does on the consumer side. And so we're thinking a lot about how to integrate AI in work settings. Today's chatbots, if I use them in an enterprise setting, it's like if I took some random person on the street who was very smart but who knew nothing about your company, and I brought them in and I asked them for advice. What I'd really like is an AI model that acts more like someone that's been trained with knowledge of your company for many years.

And so we're working on connecting our AI models to knowledge databases, having them cite work, having them be able to use internal enterprise tools, and really integrate with the enterprise as a virtual assistant to an employee. So that's one way I think about really driving the integration.

We're a public benefit corporation, and I think our long-term goal is to make sure all of this goes well. What we're really trying to do is create what we call a "race to the top." So, "race to the bottom" is this well-known thing where everyone fights to cut corners because the market competition is so intense. We think that there's a way to have the reverse effect, which is that if you're able to produce higher standards, innovate in ways that make the technology more ethical, then others will follow suit. They'll either be inspired by it or they'll be kind of bullied into it by their own employees or public sentiment, or ultimately the law will go in that direction.

And so we're hoping to kind of provide an example of how to do AI right and pull the rest of the industry along with us. That's a lot of the work behind our interpretability work, behind our safety work, behind how we think about responsible scaling. We have something called a responsible scaling policy. So I think our overall goal is to try and help the whole industry be better.

Back in 10 years ago when all of this was kind of science fiction, I used to talk about AGI a lot. I now have a different perspective where I don't think of it as one point in time. I just think we're on this smooth exponential; the models are getting better and better over time. There's no one point where it's like, "Oh, the models weren't generally intelligent, and now they are." I just think like a human child learning and developing, they're getting better and better, smarter and smarter, more and more knowledgeable, and I don't think there will be any single point of note.

But I do think that if we continue to increase the scale, the amount of funding for the models, if it goes to say 10 billion... Right now a model would cost what, 100 million? Uh, right now 100 million. There are models in training today that are more like a billion. I think if we go to 10 or 100 billion, and I think that will happen in 2025, 2026, maybe 2027, and the algorithmic improvements continue apace and the chip improvements continue apace, then I think there is in my mind a good chance that by that time we'll be able to get models that are better than most humans at most things.

I would put it in two categories. One is misuse of the models, which could include things in the realm of biology or cyber or kind of election operations at scale, things that are really disruptive to society. That misuse would be one bucket. And then the other bucket would be autonomous unintended behavior of the model. So, you know, today it might be just the model doing something unexpected, but increasingly as models act in the world, we have to worry about them behaving in ways that you wouldn't expect.

If we go all the way back to 2016, before I even worked at OpenAI, when I was at Google, I wrote a paper with some colleagues, some of whom are now Anthropic co-founders, "Concrete Problems in AI Safety." And concrete problems in AI safety laid out this concern that we have these powerful AI models, neural nets, but they're fundamentally statistical systems, and so that's going to create all these problems about predictability and uncertainty.

And if you combine that with the scaling hypothesis—and I really came to believe in the scaling hypothesis as I worked on GPT-2 and GPT-3—those two things together told me, okay, we're going to have something powerful and it's not going to be trivial to control it. And so we put those two things together and that makes me think, oh, this is an important problem that we have to solve.

In Anthropic, one of the biggest tools for this is our RSP, our responsible scaling policy. And so the way that works is every time we have a new model that represents a significant leap, we measure it for both the misuse risks and the autonomous self-replication risks. For the misuse risks, we've in fact worked with folks in the national security community. So for example, we've worked with this company called Griffin Biosciences that contracts with the US government that does biosecurity work, and they're the experts on responding to biological risk. And so they say, what is the stuff that's not on the internet that if the model knew it would be concerning? And they run their tests.

And every time so far, they've said, well, it's better at the task than it was before, but it's not yet at the level where it's a serious concern. For the autonomous, we test the models there for things like ability to train their own models, ability to provision cloud compute accounts and take actions on those accounts, ability to sign up for accounts and engage in financial transactions. It's kind of the same story as with misuse: they're getting better and better at individual pieces of the task, there's a clear trend towards ability to do that, but we're not there yet.

The responsible scaling policy that I was describing is maybe the beginning of a process, right? That represents voluntary self-regulation. Last September, we put in place our RSP. Since then, other companies like Google, OpenAI have put in place similar frameworks. I would like it if that process continues, right, where we have some time for companies to experiment with different ways of voluntarily self-regulating, some kind of consensus emerges from some mixture of public pressure, experimentation with what is unnecessary versus what is really needed.

And then I would imagine the real way for things to go is once there's some consensus, once there's industry best practices, probably the role for legislation is to look in and say, "Hey, there's this thing that 80% of the companies are already doing, that's a consensus for how to make it safe. The job of legislation is just to enforce those 20% who aren't doing it, force that companies are telling the truth about what they're doing." I don't think regulation is good at coming up with a bunch of new concepts that people should follow.

If I have a concern, though, I think it's that we're very early in the process. My only question would be, are we too early in that process to enact regulation? Maybe regulation should be the last step of a series of steps. If I look at what we've done with RSP, you know, we wrote an RSP in September, and since then we've deployed one model, we're soon going to deploy another. You see so many things that, not that it was too strict or not strict enough, but you just didn't anticipate them in the RSP. And so we're updating our RSP to say, "Hey, how should we handle this issue we've never even thought of?" And so I think in the early days, that flexibility is easy. If you don't have that flexibility, if your RSP was written by a third party and you didn't have the ability to change it, I think it could create a version of the RSP that doesn't protect against the risks but also is very onerous.

From an international point of view, I mean, I think that's a completely different question. That's less about regulation and more about there's an international race to the bottom, and how do you handle that race to the bottom? On one hand, we don't want to just recklessly build as fast as we can, particularly on the weapon side. On the other side, as a citizen of the US, here I am in Norway, another democracy, I'm very worried about if autocratic regimes were to lead in this technology. I think that's very, very dangerous.

I would say that with some of the restrictions that have been put in place on, for example, shipment of chips and equipment to Russia and China, I think if the US government plays its cards right, then those countries could be kept behind, I don't know, maybe two or three years. That doesn't give us much margin.

My view is that again, if we get to the level of AI systems that are better than the best professionals at a wide range of tasks, then tasks like military and intelligence are going to be among those tasks. And we shouldn't be naive, everyone is going to try to deploy those. I think we should try to create cooperation and restraints where we can, but that in many cases that won't be possible. And when it isn't possible, I'm on the side of democracies in the Free World. I want to make sure that the future is democratic, that as much as possible of the world is democratic, and that democracies have a lead and an advantage on the world stage. The idea of powerful AI plus autocracies terrifies me and I don't want it to happen.

If we get to the point where the models are graduate level or strong professional level, think of biology and drug discovery. Think of a model that is as strong as a Nobel prizewinning scientist or the head of drug discovery at a major pharmaceutical company. If we had a million copies of an AI system that are as knowledgeable and as creative about the field as all those scientists that invented those things, then I think the rate of those discoveries could really proliferate and some of our really longstanding diseases could be addressed or even cured.

In terms of productivity in society, I think of virtual assistants, like a chief of staff for everyone. Could everyone have a chief of staff who helps them deal with everything that lands on their desk? If everybody had that kind of thing, what would you do? Could you put a number on productivity gain? I'm not an economist, I couldn't tell you X percent. But if we look at the exponential, revenues for AI companies, it seems like they've been growing roughly 10x a year. And so you could imagine getting to the hundreds of billions in two to three years and getting even to the trillions per year.

If we look at inflation, using my limited economic reasoning, I think if we had very large real productivity gains, that would tend to be deflationary rather than inflationary. You would be able to do more with less; the dollar would go further. So directionally at least, that suggests disinflation.

I do see something related to it that's a little worrying to me and that we're trying to counter, which is that if you look at the natural applications of the technology, the most eager customers that come to us, often the most eager customers are other kind of technologically forward Silicon Valley companies. And so I think there's this danger of what you might call a kind of closed loop. And it's all being used by the most highly educated people. And so how do we break out of that loop? One of the reasons I talk about biology and health is that I think biology and health can be used to help us to break out of that loop. Innovations in health, assuming we distribute them well, can apply to everyone. I think things like education can help here. Another area that I'm very excited about is use of AI for provision of everyday government services.

In 10 years time, will the gap between rich and poor be bigger or smaller? I don't know what I think will happen. I know that if we are not extremely thoughtful about this, if we're not extremely deliberate about it, then yes, it will increase the gap.

My big bottleneck we're dealing with is data. But as I've said elsewhere, we and other companies are working very hard on synthetic data, and I think that bottleneck is going to be lifted.

The example I like to give is seven years ago, DeepMind, as part of Google, produced the AlphaGo model, which was able to beat the world champion in Go. And there was a version of it called AlphaGo Zero that was not trained on any humans playing Go. All it did was the model played Go against itself for a long time. And so basically, with just the little tiny rules of Go and the models playing against each other, they were able to get better and better to the level where they were better than any human. And so you can think of those models as having been trained on synthetic data that are created by other models with the help of this kind of logical structure of the rules of Go. And so I think there are things analogous to that that can be done for language models.

Gain familiarity with these new AI technologies. I'm not going to offer some kind of bromide about, "I know exactly which jobs are going to be big and which aren't." I think we don't know that, and also we don't know that AI won't touch every area. But I think it's safe to say that there will always be a role for humans in kind of using these technologies and working alongside them.

The other thing I would say, and this is already important advice, but I think it's going to get really more important, is just the faculties about skepticism about information. As AI generates more and more information and content, being discerning about that information is going to become more and more important and more and more necessary. I hope that we'll have AI systems that help us sift through everything, that help us understand the world so that we're kind of less vulnerable to these kind of attacks. But at the end of the day, it has to come from you. You have to have some basic desire, some basic curiosity, some basic discernment, and so I think developing those is important.