I'm pleased to welcome Dario, co-founder and CEO of Anthropic. I really see as a little bit as birds of a feather. We have common background in doing PhDs in physics, but then turning to entrepreneurship to tackle the domain hopefully in a disruptive way, but to do so in a way trying to have the biggest positive impact and reduce harms. You obviously focus on AI and I focus on space, but similar attitude. I'm particularly pleased that we're here today because we're announcing our partnership together, which we'll get into here shortly. Welcome, Dario.
Yes, thank you for having me, and excited for the partnership.
So I want to start a little bit with AI application. Everyone's talked about AGI, it's the rage obviously, but what are the biggest challenges you see in the world that AI can meaningfully advance, and what's Anthropic's strategy to advance them?
I see a number of things. AI is a general purpose technology. Things ranging from biology, which I'm particularly excited about because I used to be a biophysicist. I'm quite excited about the potential for things like healthcare, advancing science. In our collaboration with you, the ability to understand the earth better for the sake of many cases that will have a number of scientific applications, as well as improving the state of our planet and improving our ability to manage it. I think some of the challenges come from the key development of models itself, of course, it's its own science, it takes a lot of work. But when you actually want to apply the model, creating the adapter between the fundamental model and the downstream application often requires companies to work together. It often requires partnerships with companies. We're trying to serve every industry with this general purpose technology, so when we partner with a company that understands some part of the world much better than we do, I think the important thing is to work together to do something that we couldn't have done alone, because we understand AI but we don't necessarily understand all the implications of all our models or all the things it's good at or can be used for. At a high level, that's one of the biggest challenges.
Talk a little bit about what's rate limiting the development of AI in general today. There's compute, there's talent, there's data, there's algorithms. Where's the bottleneck?
Data is a really interesting one. In terms of general data, there's a lot of it around, and we've been able to scrape the entire internet. There are some techniques like these new class of reasoning models that rely less on data. Where I think data is really important and valuable is when you're trying to do a task in some particular domain. Getting the best performance out of that domain, data is going to be particularly important. There are some things that you just don't learn by reading the internet. I think the partnership we're doing together is maybe a prime example of that. Claude, our model, has been trained to analyze all kinds of images by looking at images on the internet, but the fraction of those images that are satellite data, that are geospatial data, is tiny. While you guys have an enormous amount of such data. It's a general model that we've trained, it's going to help. We're going to start from a much more advanced place. The models can already do great things out of the box, but there is a much greater potential for the models to become better at this task that is somewhat similar but somewhat different from normal image analysis. I think there's a lot of value in this specialized data to make these fine-tuned models for particular applications.
100%. So let's move to space a little bit and those Earth applications in particular. It's interesting, there is a revolution going on in space. People don't know about the lower cost rockets, the increasing fleets of satellites. Most of that, we've actually seen a thousandfold decrease in cost per performance of satellites over the last 5 or 10 years, so it's a bit like the mainframe computer to desktop computer sort of equivalent. Of course, unleashing lots of applications, but the big upshot of it is new data sets about the Earth or new communication data around the earth. One way or another, it's about data. We alone at Planet produce 4 million of these 47 megapixel images to scan the whole Earth every day. It's a lot of data. I think of how LLMs have figured out how to semantically interface with the web data. Analogously, you can imagine a semantic interface with all of these images. I call that sort of vision queryable Earth or searchable Earth. What vision excites you at that interplay between AI and space? That's one, are there others that come to your mind?
I think that one is pretty exciting. I can put it in context a bit more. In the space of text, there's been a lot of stuff done in this kind of queryable or summarize or search. There are whole companies, most of whom are customers, who have their whole business model around how can I find, summarize, search, sort through data better than before, better than base search engines can do it. But all of that has been text. In principle, the same thing can be done just as well with images. If I think about in the internet context, of course you want to do internet search, it does come up. But actually, I think it wasn't until a use case like this, satellite data in space, where it's really like you're doing the same sort of queryable thing for images. That has been a core part of our business model for text for so long. I'm sure we'll have more advanced types of things. I'm sure we'll be able to analyze video. There's much more value to be produced here. But this idea of searching through a large number of images for a particular thing, I'm not the expert on space and satellites, but it feels like this use case is perfectly made for that ability.
Absolutely. Scanning for larger areas for changes that you might be interested in, sustainability or climate related or damage assessment after a storm, they should be able to give a quick quantification of things, quick analysis of things. We've already seen some interesting results there. Another area is predictive analytics. One of the things I think about is how LLMs have extrapolated the next token. Here we are with 2,500 images for every point of the earth's land mass. Perhaps it can guess the next token, i.e., the next image, which in this case could be where's the next flood, where's the next deforestation, where's the next refugee camp. Do you see that as a potential? Where does predictive analytics go?
I think that's pretty valuable as well. You can already ask the models out of the box today, show them a bunch of images and ask them what they think is going to happen next. They don't even need to generate another image in many cases. They can just describe in words where they think things are going and what they think will happen next. I think there's a huge opportunity for that. It has analogies to what we're doing in other fields. Any scientific application tends to involve prediction. I'm very excited about the use of LLMs for science, so that predictive activity is basically the same one. I'm excited about LLMs for medical use cases. A lot of medicine is about diagnostics, prognostics, about prediction. I think LLMs will find wide use in finance, that's definitely about prediction. If you can predict things more accurately, you can be very successful in the world of finance. There's a lot of precedent for this to work well. We're still in early days of trying it out, but I'm very optimistic based on the kinds of use cases I've seen across industries.