Well, Jack, Block has been on an open-source tear. You guys have made much of the AI tooling that you've been building lately available for anyone, like Goose, Buzz. Could you give our audience kind of an overview of what these tools do?
Yeah. Well, I mean, number one, I think it's important to recognize why we do open source. As a company, we've benefited so much from the grace of developers around the world who give up their work for free so that others can use it and also learn from it. My entire career is due to the fact that someone out there decided to share the source code that they were working on, share the project they were working on. It's the way I learned how to program. It's the way I learned how to build and problem solve. And I think it's up to all of us, especially in corporate America, to give back to that. And we've done so with a lot of financial technologies. We've done so with utilities. We think it's even more critical to do so for intelligence.
And as we worked on internal tools, our first one was Goose, which is one of the first coding harnesses out there. We shipped it maybe three months before Claude Code came out. And what it enables you to do is basically, you know, back then just a bunch of very simple loops that worked with models such as Claude or GPT, and it allowed an engineer or a programmer, or more and more so everyone who wanted to use it, to actually build something. So it allowed you to write an English prompt and the output would be working code. In the early days, the earlier models, it wasn't always working code, but it was code that you could get working fairly easily with a few steps. And we decided to open source it because, again, one, we believe it's important to give back to this community we've benefited so much from, but also by open sourcing it you get more eyes on it and you get more usage and you get more people contributing to it that aren't necessarily in your company, and those ideas can really change fundamentally what it is and what it becomes and what it is capable of doing.
We used it internally. We wanted others to be able to use it for their companies or for their products as well. And we found that other companies took it, forked it right away, used it internally, built on top of it, changed the code, gave back to the codebase, which benefited us as well. And it resulted in just a constant iteration of making it better and better and looking at, I think, the most important part in the space, which is not necessarily the model, although that is core, but the interface — the interface to what the humans actually use, how they intersect and interact with these intelligence systems.
And it led us from a single-player mode, which was Goose, to a multiplayer mode, which is what we call Buzz. Buzz, if you're familiar with Microsoft Teams or Slack or even Discord, Buzz is very similar to that. It's an open-source version of a group collaboration tool, but agents who are backed by these intelligent models, whether they be open-source models or corporate models such as OpenAI or Anthropic, are on the same first-level plane that every human in the organization is on. So you can collaborate effortlessly with agents. So as a team, you can bring in agents, you can bring in this intelligence, and you can work together on it. So Buzz represented moving from, again, the Goose single-player to a multiplayer model. Again, focusing on the interface and how teams actually use these technologies, how we actually build today. And again, we believed it was important to make it open source because we could just get more ideas, more usage, and more contribution that wasn't limited by the number of people that we had in our company or the talent that we had in our company.
Yeah. And from the outside looking in, it seems like you are building toward a very particular vision of the future. I think a lot of people are wondering what is a company even going to look like in five years. Do you have any ideas?
Well, our push is to really question what a company is. Like a lot of companies out of Silicon Valley — and I should say a lot of technology companies — have borrowed a playbook that hasn't dramatically changed that much over, I don't know, probably three or four decades. So the playbook we used to build our company was from Google. Google borrowed it from Silicon Graphics. Silicon Graphics borrowed it from HP. HP probably borrowed it from General Electric, and so on and so forth. And there's been very small iterations along the way, but the core hasn't fundamentally changed, and the core hasn't really shifted with the technology available to us and how quickly that has moved. And the biggest shift, obviously, over the past two years is the rise of intelligence, is the rise of agentic development.
And we finally have, you know, the companies of the past and definitely currently in the present. A lot of the intelligence of the company is locked into everyone's heads. It requires a lot of documentation, requires a lot of communication, sometimes a lot of forced collaboration. Sometimes that collaboration is a little bit easier. But if you really look at how people are working in companies today, they're using a lot of Slack, using a lot of Microsoft Teams, using a lot of shared documents, whether they be Google Docs or Microsoft. There's a lot of artifacts and context within the company that is now shared. We happen to be a remote-first company. So while we do have physical offices, people are not required to go into them. They can work from wherever they want to be, wherever they feel most creative. So every action that one of our employees takes creates some sort of written artifact or something that can be recorded, like a Zoom like this, or something that we can actually provide shared context around and we can distribute to the rest of the company.
The challenge in the past was organizing that in a way that people could quickly query about it, ask questions about it, find the right information in a rapid fashion. But today, with these machine intelligence systems, we do have that, and we have the potential of a shared context almost immediately. We have an interface where we can ask a question and get almost immediate results based on correct data and data that's actually in our system. So what it allows us to do is start to question the need for a traditional management hierarchy, because the role of a traditional management hierarchy is oftentimes a communication router. It's routing various messages from the core of the company to the edges of the company, or the top of the company to the frontline staff, whatnot. But these intelligence systems and these shared contexts that we have allow an organization to be a lot more distributed, a lot more flat, and it gets everyone much closer to the information that can help them make better decisions almost instantaneously.
So the thing that we've been questioning first and foremost is what is the role of a management hierarchy? What is the role of management generally in a company where almost anyone in the company can access any information instantaneously? Almost anyone in the company can build something based on that information, based on that data, and distribute it almost instantaneously to the rest of the company and also to our customers. And we've kind of normalized our company around three specific roles, instead of the huge variance of titles that you would normally have. The first is what we're calling an IC, an individual contributor. That's someone who's doing the work. They could be a builder. They could be an operator. The second is someone who's a little bit more strategic, has more of the customer context, customer empathy, and can be responsible for the outcomes. We call that a DRI, a directly responsible individual. And the third used to be called a manager, but we call it a coach. This is someone who's looking at the discipline that they're responsible for, whether that be engineering or design or marketing or sales, and actually coaching people around their discipline to help them master the discipline instead of worrying about managing a team. And we believe that every single one of those roles is a builder or is an operator, can actually do the work, but they can also take on these more strategic roles or more coaching, human development roles. And we're shaping our company around that, and we're putting less and less focus on the management chain because we have this shared context.
So I think the best way to adapt to a world full of machine intelligence and artificial intelligence is to have the company itself become an intelligence and to become something that you can chat with, become something that anyone can contribute to, because ideas happen all the time. We rely a lot on these insights, people who are very close to the customer with more and more capabilities. They can demo something very quickly and we can get that to production faster than we've ever been able to even consider in the past. And a lot of the hierarchy, as we see it today, is more friction to that path. So as we've gone down this path over the past year, we've increased our shipping rate dramatically, our the number of decisions we make, the things that we question and shut down or pause, the new projects that we start. And it's just amazing to see in contrast how much of the management layer has gotten in the way of us really inventing and being creative and acting on the data we have before us and getting as much signal from all over the company that we can. A lot of it is hidden, or sometimes people don't feel comfortable speaking up, or they don't feel like they have the right to or the know-how. But with these tools we can organize the information in such a better way and get to results much faster. So I believe that the future of a company, at least to ours, looks more like a shared intelligence than the traditional hierarchy that we kind of grew up with and that we've known for decades.
That makes a lot of sense. I want to shift to the open frontier, to the essay you published on Twitter recently. This was a response, I believe, to Dario's essay on pacing the frontier. And so you published this essay called Open the Frontier. And in it, you wrote toward the end, 'I want someone I've never heard of to be able to build something better without asking permission from the companies they might replace.' For that to happen, it seems that a builder needs both the tools to compete and the freedom to use them. I'm curious, what developments, technical or otherwise, give you optimism that someone you've never heard of could compete with the largest labs?
Well, I was pretty concerned when it seemed that it was only OpenAI that was on the frontier of this technology. When Anthropic came out and pushed an interface that made sense to developers, which was Claude Code, it eased a lot of that concern for me because now there are two players. And then as more and more people saw what was possible and the opportunity, more labs sprang up, a number in China activated as well, and open-source developers demanded that they could play too, and that they could play with not just the code of these models but the weights of these models as well, which is the actual compressed intelligence that these models are offering. And I think that's aligned with where all of this technology springs from, which is open research and sharing your work and sharing your findings and getting peer review and understanding what's possible, what's not, and what the blockers are to the next step.
We now have a very healthy open-source ecosystem. Week on week we see new innovation not just from the corporate labs but from the open-source world. It's not just concentrated in the United States or any one country. I think there's a very healthy competition globally, and I think that's necessary. I believe a research finding outside of the United States is not an American loss. I think it's something to learn from and something to compete with, and it makes us better. And the role of open source is that we, as I began with, this is a principle that allows more eyes on the problem and allows more ideas around how to solve it that the talent locked up in any one particular lab may not think of or may not have. But if they get a chance to see something, it might help them think in a very different way to make something better. So I don't agree with slowing down. I don't think it's possible. I don't think it's something that advances humanity, and it certainly doesn't advance the technology.
I do think an alternative is that instead of pacing, that we open more. And if there's any role for government, it's funding the compute necessary, the resourcing necessary to open these models to more people, to get more eyes on the problem, to be able to do more research, and to enable more checkpoints on the entire system so that we can be eyes wide open on what these models are actually capable of and what they're not, and ultimately what direction the world wants to take it. That excites me a lot more than it being locked up and paced in a set of five American labs or two significant governments controlling these technologies. Generally, I'm against any, especially on the internet, and the internet is all about, and all of its virtues and all of its goodness that it's brought to the world has been around permissionless protocols, open protocols that anyone can develop on, that anyone can create an idea on, and it may take over. You know, my career is only such because I was able to work on open protocols and work through open source to build something compelling enough that other people wanted to use it. And if that did not exist and I had asked for permission from a number of companies, I definitely wouldn't be here talking with y'all. And I would probably not be in this career either. So we've had open communication for quite some time. Bitcoin points to a world where we have an open financial system. The next and probably largest element is open intelligence, and making sure that's not locked up in the hands of any one government or any one company, that it's readily available to all and anyone can contribute to it, because I think that's where we're going to get the best outcomes. All these ultimately are races, and it's not about the pace that you go necessarily. It's about how many steps ahead you are of any adversarial actions. And if we have any slowness in that system, it risks us being slow against an adversary. And if we have a lot more openness, it allows us to be as fast as we can imagine against all adversaries and all catastrophic ends and outcomes.
I think you're making a really fascinating argument that's very timely for DC and for this audience. I think that for many in this room and who are sort of participating in the emerging AI policy debates, it feels like openness and open source are being juxtaposed with safety. And this is a very important argument that openness is itself quite net beneficial via its own causal mechanisms to the end of safety. But really so much of the attitude here has been doom and gloom. And it is easy on some level for people to imagine some of the worst-case hypotheticals from RSI, and people are sort of trained on sci-fi. There's a lot of effort to sort of communicate the worst-case scenario for machine intelligence. I was wondering if you could share with the crowd maybe the opposite take. You know, power has every sort of ability to do harm as it does good. What is the most significant benefit that you think is likely to come from a rapidly expanding and open frontier?
