Being a founder is difficult under any circumstances. You operate under a state of constant duress, scrambling to figure out a working business model while not getting crushed by incumbents, losing key talent, and striving to avoid failure. Mental and physical health can take a serious toll. It’s a journey full of extreme highs and lows.
Now imagine you’re not just a startup founder, but you are the founder of the leading company in the most transformative field in history: AI.
Greg Brockman, the cofounder and president of OpenAI, has been with Sam Altman for the startup’s entire history, including during Altman’s brief 2023 ousting as CEO, only taking a short sabbatical late last year.
In between, Brockman has had to navigate lawsuits, political and regulatory scrutiny, historic fundraises, and a public that is not convinced the future he’s building is a positive one.
I sat down with Brockman on Thursday for a one-on-one discussion to gain his perspective on everything from OpenAI’s recent Hugging Face security breach to how the U.S. should consider Chinese open-source AI models, to what the business model of an AI Lab really is, to how he deals with the pressure of running OpenAI.
As for the business model, “The fundamental value delivery that we are after—we’re not there yet,” Brockman said. “The massive value creation is yet to come.”
And as for the pressure? “I lean on my wife a lot,” he told Fortune.
And indeed, as I left OpenAI’s New York office in the SoHo Puck Building (where OpenAI is neighbors to one of its biggest backers, Thrive Capital) I bumped into Brockman’s wife, who had made the trip to New York City to support her husband.
The transcript below has been edited for length, clarity, and publication.
Earlier this week, OpenAI wrote a blog post detailing a situation where its LLM models broke out of a contained environment and interacted with a competitor, Hugging Face. What surprised you most about that incident, and how can a security breach like that be prevented in the future? Is more collaboration needed across competing AI companies?
Greg Brockman: When we start to think about moments like this, it is less about the specific incident and more about the general level of capability that the models have.
Our perspective has been that we really want to help defenders bring this technology to bear so people can use it to secure their systems.
The fact that they weren’t able to use our model is something that really stood out to us. We need to do a better job of making these models more accessible, and to try to help educate people too, just so they understand why they need them. Cyber defense is a huge priority for us. That’s a team we’ve been really staffing up, and they’re starting to execute with significant focus and urgency. This moment is one that I hope can really bring together the industry because everyone’s interests are very aligned.
In the Hugging Face incident, it sounded like there were two models—one known model and one that hasn’t been previously disclosed. Is there anything else you can tell us about what model is to come and how you’re looking at that next model in light of what just happened?
I don’t want to add too much on top of our blog post because it’s still actively being investigated. But we said it’s a combination of models; we mentioned two of them, but we said it’s a combination of different models. It is less about any one model or any release; it is really just thinking about the overall picture.
There’s a different post that has gotten less attention, but in some ways is more complete. Earlier this week, we talked about a model where we did take it down. We implemented a number of different safeguards internally, and then were able to bring those lessons to the future.
One thing people don’t understand as much about OpenAI is our extreme focus on AI safety and how seriously we’ve taken that over the years. And there’s actually a number of different places where we found real fundamental innovations on how to keep models aligned and how to monitor them.
For example, chain-of-thought monitoring is something that we, from the very beginning of the reasoning paradigm, realized would be possible. If you train a model to have a chain of thought that looks good, you’ll actually get better task performance because you trained it to not be illogical. But then you lose the faithfulness, because now you’ve optimized for an end result that looks good. So that’s a place where we made a very clear line in the sand that we’re never going to do this. Every time people start to drift in that direction, we course-correct it. We make these decisions constantly. It’s not something that we broadcast much, but I think it is something we are leading the industry on.
There’s a lot of debate about how government should interact with AI companies, especially as these models get more powerful. One idea that received lots of praise was Demis Hassabis’s proposal for a FINRA-like oversight organization. Is that something you think could be a model moving forward? How do you view the tension between permissionless innovation versus hardline regulation?
I think Demis’s idea is quite good, and something like that feels like a good way to balance innovation and safety. I think we really can have innovation with good coordination and oversight. A vacuum slows everyone down because if there’s uncertainty about how to move forward or who to talk to, we’re not going to be able to make good decisions.
Internally, we are always debating these exact same questions: The focus on safety, security, doing the right thing, and having nuanced debates about the right decisions. It’s helpful to have more voices in the room, as long as it’s structured in a good way where decision-making is clear. While there are many ways to get this wrong, but I think there’s a framework that can accelerate all of us, and that is what we should collectively solve for.
Julia Demaree Nikhinson / POOL / AFP via Getty Images
What about foreign competition, specifically open-source models coming out of China? Should the U.S. government restrict access to these open-source Chinese models, or should they remain available to foster startup innovation?
I can say my own perspective: We started OpenAI with the view that there should be more AI in the world, and that people should be empowered with AI to be able to build a better life. At a highest level, we think that more AI in the world is actually an important thing, and having startups that can flourish is absolutely critical. Having people be able to use AI in the ways that they want is absolutely critical.
Now, to the question of distillation: There’s the question of: Is it a technical thing or a policy thing that we should consider it? My answer is that there’s absolutely unquestionably a very core technical component here. We’ve always had distillation protections on our platform within our overall framework of how we secure these models. Distillation is a very nuanced thing. Getting a model that has the exact same capabilities as these core frontier models requires a model that’s exactly the same size compute as one of these frontier models. It’s actually very difficult to do, so we need to recognize the technical facts when thinking about policy. I don’t have a full judgement on what the right answer is, but I think it’s clear as an industry, we need to continuously uplevel our technical safeguards.
What are you most excited about working on right now? What gets you out of bed in the morning?
For me, it always starts with the impact. Ten years ago, I was doing the standard startup thing [at Stripe]. I watched AI from the outside and felt it was going to be the most transformative technology and positive force for uplifting human well-being ever created. I just wanted to help steer it in a more positive direction. That would be a career I’d be proud of.
AI is no longer theoretical or just about benchmarks. You deploy it and see people use it. There’s a farmer in Japan who used Codex to automate his farm. He used to hand-roll up plastic tarps, and now a mechanical system that does it. Health is another huge application area. We announced rolling out health features in ChatGPT in the U.S., allowing users to connect their medical records and ask agentic queries about their health history.
I have a friend who was in the hospital, and the doctor was about to administer a specific antibiotic. She checked with ChatGPT, which warned that given her medical history, that medication could be fatal. She showed it to the doctor, who realized they had only had five minutes to read her chart. You realize how many problems we’ve previously given up on solving that we don’t have to give up on anymore.
What should we expect regarding physical AI and hardware devices from OpenAI? How do you see AI manifesting in physical forms?
We’re definitely not ready to talk about the device yet. It’s going to be an amazing moment when we reveal it, and it will be the first in a series rethinking what devices should be, given what this technology is capable of.
The fundamental thing is that while there is a lot of focus on AI capability—solving harder problems—the interface is crucial. If you look at the ChatGPT interface right now, it’s way too complicated with model switchers and side buttons. Our objective is to make all that UI fade away.
How do you have Chat that is used by almost one billion people and the power of agents to all those humans? That’s the most complicated product pivot in history.
You should have something you can interface with as naturally and seamlessly as you would with a person. It’s exhausting to use a computer right now. You get carpal tunnel, you’re hunched over—we are feeling the physical effects of it.
In upcoming years, we will feel freed from a lot of the toil and ways we’ve had to contort ourselves to our machines.
How do you decide what products or devices to launch at OpenAI? How do product decisions get made internally?
OpenAI has a very bottom-up culture. We have different historic patterns of what becomes very successful. For example, Codex was largely bottom-up with a small, passionate group focused on a couple of different projects that we brought all together into one product. Conversely, our API was entirely top-down. At the time, we were a research company and needed to build a commercial product to sustain funding. We went down a list of 20 ideas and rejected narrow vertical domain tools [like in the medical space] that would have been giving up on the “general” in AGI. We decided to build an API for developers, and they can figure out what to build. It was probably the hardest project I’ve ever worked on because we had to inject product DNA into a research organization, and it wasn’t even clear if anyone would pay for this. But six months later we launched it, had paying customers and it worked.
A lot of our products have this property where someone has the spark of an idea, or there’s a clear direction to travel, or we have a model with some surprising new capabilities we haven’t seen before. All of those are where the products come from.
Given how much capital is flowing into AI, what is the sustainable business model? Do you need a hardware device or workplace integration to build a generational company, or is the model itself the core business?
David Paul Morris/Bloomberg via Getty Images
There are two parts to that. One is, let’s suppose model capabilities were frozen today: what is the business? I think OpenAI is the most strongly positioned right now for that scenario. We have a real shot at building something that is generational and unique. We have to execute very well, but it’s an opportunity we can live up to. We have ChatGPT which has nearly a billion users, and there is so much more value we can deliver to those users through even just the technology that exists today.
This year we’ve also started to really prioritize enterprise and we are seeing a wave of momentum. So many companies that were using a competitor are switching to us, so many customers that have never used this technology before are using us for the first time. A lot of that is listening to what the customer needs, building features in a great way, building trust, thinking about data privacy. We’re going to head to this future where what a business even is is very different. Agentic tools like ChatGPT Work are changing how businesses run by handling administrative busywork. The business we’re building is unlike a business that has existed before, because we are democratizing using it in your work life and in your personal life and it’s all powered by this one technology.
Things like the device are part of it, but they’re not the whole story.
The second piece is that the model progress being frozen is not a true assumption. The way to think about it is a little less like we have electricity now [LLMs] and now we need to build washing machines [businesses on top of the LLMs]. Just like people are trying to achieve fusion, we are trying to get the technology to do something that is very qualitatively different from what has been done before. That’s what’s still on the table with AI. AI will be able to solve unsolved problems. We are tasting it right now in mathematics, chemistry, and science. The fundamental value delivery that we are after—we’re not there yet. The massive value creation is yet to come. Some of it will wind up being the value that the AI labs have. But some of it will accrue to the world.
It’s hard to be the founder of any company. You’re the founder of the leading AI company in what’s maybe the most transformative technology the world has ever seen. How do you carry that weight and responsibility?
To be honest, I could not do it without my wife. I lean on her quite a lot. It is hard. There are days that are extremely stressful. There are days that are joyous and fun. Most days, you’ve just got to keep going. My favorite quote is this cycling quote, “It never gets easier; you just go faster.” That is very much how we operate at OpenAI.
I feel the gravity of the moment. I feel very honored and privileged to have this opportunity. We signed up to some extent for this ten years ago, although it’s different to be living it.
There are definitely moments where it’s nice to think about once we have successfully completed the mission and everyone’s life is better from AI, and AGI is in good shape, to be able to relax and take some vacation. But until then, I just have to keep my head down and keep going. The truth is that people at OpenAI are fantastic and we’ve built a team that really cares about this mission and each of us keep each other grounded and focused.
