Thank you sir.
Hello again and thanks again to Mario, to Michael, and to Kat. I love the GitHub integration. The last project I did, I actually was like, oh I can actually just install Cloud Code into a GitHub code space. And all of a sudden I have Cloud Code against the repo that I've already been building. It was really great to hear from each of them and hear all about the exciting work being done with Cloud. So to close out the show, I'd like to dive a little bit deeper into Cloud 4, our research direction, and what developers can expect next from Anthropic. So please help me welcome back to the stage Dario for our one-on-one conversation. Welcome back Dario.
Hello again. This is great, this is like our one-on-one in front of the whole audience. This is great.
So Cloud 4 is out, Cloud Sonnet 4 and Cloud Opus 4 are available. What excites you the most about the Cloud 4 models and how does it change your thinking about what's possible in the next 12 months?
Yeah, so I think abstractly the thing I'm most excited about is, you know, every time you have a new class of models there's more you can do with it. So we're going to be releasing models after Cloud 4, there'll probably be a Cloud 4.1 at some point just like we did with Sonnet 3.5. And I think we're just at the beginning of what we can do with the new generation of model in terms of tasks. I think the autonomy is going to go much further than it has already. Just the ability to set your model free and give it the ability to do something for a long period of time, I think we're very much still at the beginning of that. I'm actually increasingly excited about the models for cyber security tasks. You can think of cyber security as a subset of coding tasks, but they tend to be higher-end coding tasks, and so I think we're maybe finally hitting the threshold for that. And then as a former biologist, I'm always excited about use of the models for biomedical and kind of detailed scientific research work, which I think Opus in particular is going to be strong at.
It really connects to Machines of Loving Grace. So how does Cloud 4 fit into that trajectory overall? I like to joke that people think of Machines of Loving Grace as an essay and I think of it as a product roadmap for the next few years. Curious how Cloud 4 fits into that journey.
Yeah, it was sort of a product roadmap that I wrote without knowing how to actually get to it and kind of said all right guys, this is your work, this is your job. We're increasingly thinking about the biology side of things, and software is part of that because biology increasingly involves data. Even involved data 10 years ago when I was a biologist. I think more and more of it is going to be, okay, we have these models that know a lot about biology and they can help write code. So if you're a computational biologist, I think these models will really accelerate what you can do. And we have a number of customers who are trying out the models for these tasks.
I think one of the first hackathons we did after we released MCP, somebody hooked up MCP to one of those plotters that does drawing, and so Cloud could draw for itself. It's really fun to see what Cloud draws for itself. But it was like the first moment was like, MCPs don't just have to be connecting to digital systems, they could also be connecting to the real world. So being able to drive lab equipment via MCP I think is an interesting question for soon. We'll be able to test Cloud by connecting it to a polygraph. Are you lying? Who needs interpretability when we have the polygraph.
You mentioned that moment where you were convinced that the Cloud-written content was human-written. Any other breakthrough moments in watching us dog food Cloud 4 or even try it yourself that made you realize this model felt different?
You know, I didn't actually understand the details, but there were several people on our side. There was a moment a few weeks before the model launched where someone said, oh my god, this model just one-shotted this incredibly difficult performance engineering task. And no model had ever done anything like that before. I will say there's this almost superstitious process in model development where it somehow all comes together at the last moment, even if the training plot process is all planned out. Some of the model's abilities, maybe it's something about their interaction with people, maybe it's something about just making it the last bit better, maybe it's people getting used to the model and prompting it. But you always find the early versions of the model, people are struggling to figure out how to use them, and then you finally get to a point and people are like, this works for me all the time. There's that alchemy that happens somehow always the last moment. If you read Creativity Inc. by Ed Catmull, he talks about the same process with all the Pixar movies. They're really bad until like two days before they're supposed to go out, and I feel the same way about our models. Not that they're really bad, but they're not quite there and then suddenly they click and we're like, I can't wait to get this out to people. It doesn't make any sense because the training process is uniform and you would think that it doesn't work that way, that it's all a rational process, but it's absolutely not. There's no point on the RL curve at all where they come together. It comes together at the last minute. I don't know why. It's a real moment.
Many people in the audience are developers here, and a question that I know has come up internally is which parts of the software engineering job will AI take over, and what becomes more important in a world where we have autonomous agents being able to do a lot of software engineering?
Yeah. Probably like many people here, I read with great interest Steve Yegge's blog post a couple months ago, Revenge of the Junior Developer. He had some similar blog posts. He actually came in to visit us even. And that laid out, I think, the vision of where things are going maybe even better than I could, which is that we're gradually going to more and more autonomy of the models. We had this phase where you would do basically autocomplete. Now there's this thing that I guess people have called vibe coding. And then we're going more to dispatching agents to do things. And I think with Cloud Code we're going to go more in that direction, and I'm sure we'll have other product surfaces that allow you to do that as well. We're heading to a world where a human developer can kind of manage a fleet of agents and say you go off and do this, you go off and do this, you go off and do that. But I think continued human involvement is going to be important for quality control, to make sure they do the right things, to get the details right. Working together on both the models and the product surface around it to get the details right is going to be really important.
I think it's also highlighted to me, it makes the stuff that is inefficient in your work way more painful because it's taking you away from this flow of building. At least it's made me realize where we're spending too much time on cross-functional alignment and road mapping when we just should be trying to get more building. So it's become more painful as the engineering part has been sped up as well.
So there's endless debate around the industry around bigger models or smaller architectures, which will win in the long run. You're famous for popularizing and pioneering the scaling laws paper. What's your current take? Is pre-training dead? Is pre-training all that matters still and its role relative to post-training?
Without getting too specific, I would say that the Cloud 4 models embody advances in both pre-training and post-training. So we're continuing to see the pre-training scaling laws work the way that they've worked before, and we're also continuing to see continued advances in post-training. And they kind of complement each other. I think we're going to continue seeing advances in both of those. I think we're also going to continue to scale up. So we have these multiple trends, these multiple sources of exponential growth, and they're all going to compound with each other. That's why I think all of this is going to go very fast. One of the reasons I liked Yegge's blog post is that it was someone who was not me repeating the mantra of it's only going to be a year or two until these things are basically peers to us. It's insane that 3.7 was just in February, right? It feels like a year ago but it was just three months ago. This feels like an obsolete model or something, and it's like two and a half months. The time scales are compressing. I often say that being in the AI field, it feels like you're getting on a spaceship leaving Earth at relativistic speeds. One day you wake up and it's like one day on your spaceship, two days on Earth. So you have to take in the news of two days. It accelerates. One day on your spaceship, three days on Earth. That's just what it feels like being on this ride.
That resonates. I've never heard the metaphor before but it absolutely does. Maybe on the post-training front, one of the things that I got really excited about seeing developed in Cloud 4 has been this concept of memory and having the model be able to manage its memory. Maybe talk a sec about why that's important and what that enables.
