[0:00:00] Last month, OpenAI's chief scientist wrote an essay called An Alien Mind, warning about the rapid advances in AI language models. The essay claims that AI models are growing smarter by the day, often more than they were ever designed to be, that safety precautions are falling behind, and that humans could slowly begin to lose control. For the last few years, AI has been something you had to open an app.
[0:00:21] But this fall, everything changed. The largest AI company started shipping personal AI agents, and a personal agent can work for you without you opening an app. OpenAI's version of the personal agent is called Dots, and it launched three weeks after that essay was published. Lives in your inbox, it remembers years of your conversations, and it can make decisions on your behalf. If the companies building these AI agents don't fully understand them, and one goes rogue and acts like a user, who owns that? Is it you?
[0:00:48] Or is it the company that built it? So I went to OpenAI's president and co-founder, Greg Brockman, looking for an answer to all of this. I don't even know where to start on this one.
[0:00:58] But before you hear from him, a quick disclosure. OpenAI invited me to Dev Day and gave me dots about a week early to test before my conversation with Greg. But before I flew out to San Francisco for Dev Day, I went back through OpenAI's paper trail myself. I mean, this year's incident reports, the safety disclosures, and from all of that, I mapped out the main five questions that most users would want answered before letting one of these agents act on their behalf.
[0:01:21] Questions that most people would be hesitant to ask one of the largest figures in all of tech right now. Question one, how did DOTS fix a problem I didn't even know I had? That's the North Star vision of what the big AI companies are promising you right now, an agent that'll run your life. And Greg's vision of that story goes all the way back to OpenAI's early days. So Greg, it's an honor to be here. I first wanted to ask you, as I understand, OpenAI was started in your living room, I think 11 years ago now? That's true. Yeah, 11 years later, you got 1.2 billion users.
[0:01:51] Maybe you haven't really had the time to reflect on that, but this must all be surreal for you. It absolutely is surreal. And in some ways, it's beyond what we ever expected. And in some ways, it's kind of exactly what we were picturing. I will say that there are moments at OpenAI where the reality, you feel it. You kind of do what we call, you feel the AGI. And you walk around for a day, a week, really just kind of thinking about, okay, we've been anticipating this moment. We've been thinking about it, but here's a real milestone and it's here now. And what does it feel like to be in the middle of history, right? In the middle.
[0:02:20] I believe that we can build a world that is just...
[0:02:29] More empowering for everyone, where people can live better lives. I remember we did this offsite when we were starting OpenAI in Napa in November of 2015. And we laid out our vision. But I remember one of the specifics was the idea of what if you could train an AI that would use a mouse and keyboard and look at screen pixels. Use a computer just like a person. And it was so appealing because you realize that you can then have an AI that, in terms of interface, could do anything that a person could. I don't want to say it's all exactly as we predicted because, of course, there's lots of things that are just different.
[0:02:58] And when you're in the moment, there's always surprises. But there is something that was actually very predictable at a macro level about where we are today. Almost 11 years later, Greg still talks about his vision for DOTS in that way. I want to segue into DOTS, which you just launched. I think personal agents got to be the most exciting category in tech right now. And I think it's a multi-trillion dollar market opportunity. But what do you think OpenAI has in particular that gives you an advantage over some of these other competitors on the market with DOTS? We have been the field leaders in how to build the most aligned,
[0:03:28] Capable, useful, safe technology. That is what we have been doing at OpenAI for these past 10, 11 years. And we are very much bringing all of that energy and momentum into this era. We have 1.2 billion consumers who use ChatGPT every single week. And so in my mind, we've been putting in place the building block and we're packaging all of this into a form that's just so useful where the interface melts away, right? Simplicity is the key. You just want everything to fade away. You don't want buttons and help menus and documentation for how to use all the things. You just want an assistant that can help you solve your problems.
[0:03:57] I think that AI represents the next step in how you have technology that really serves people. I think that when it comes to this phase that we're at of really shifting to the personal agent, the models are capable enough and we've also built guardrails. We've built the right sort of safety into our production systems so that people are willing to
[0:04:14] And we do it within OpenAI all the time. Hook it up to all sorts of different data sources. Again, trust is such a key part of what you have to build and what you have to really deliver and live by. I think what it is, and you guys probably can't talk about it this way, but I can. I think it's Jarvis. Now, I'd only been using DOTS for about a week when I said that to Greg. And I said it because a problem was in my inbox and I wasn't even aware of it. It actually found a discrepancy in this contract I was working on and it flagged it. And it said, hey, this could come back to bite you in the ass sometime in the future. Do you want me to just go ahead and resolve this?
[0:04:43] And I never asked DOTS to look at that contract, which is the part that's so interesting to me. So how did DOTS fix it? By coming to me and flagging the issue on its own. But that left me a little uneasy. So I wanted to ask Greg more about how this all worked. Question two.
[0:05:06] What can OpenAI actually promise its consumers about safety? My question is, the more data and the more access you give Dot, the more useful it is. Specifically, what kind of confidence can you instill in the end user for them to know that giving it all of this access and context about their life is perfectly safe? First of all, I think that trust is absolutely core. It's a core goal, something we think about every day. If you think about how you work with a human assistant,
[0:05:29] The building trust takes time. It takes experience with working with a person. You see what they're capable of, what they're not. As you grow in that working experience, you end up with a real dynamic where you kind of intuitively know when you really have to provide extra oversight, when you can kind of let them run more with the task. And so in my mind, the way to think about this is technologically, we've been building up these guardrails for quite some time, even from the very moment that we pioneered bringing this technology to the world. Back in 2020 was our first product. We thought very hard about safety guardrails around what are the rules.
[0:05:58] And actually the way that we started out there was we took a very conservative approach. And so anytime a developer wanted to build on top of our API, they would have to get on a phone call with us and walk through their use case. And we would kind of take a look at what they were doing. And we'd say, have you thought about what could happen here? Could you think about the problem there? And so we started to build up this expertise in-house because we saw all these different use cases, all these different wonderful ideas, and that we were really thinking about the, okay, how do you actually deliver this technology in a safe fashion? So we pioneered a number of different technologies, all of these things that are now part of the whole ecosystem.
[0:06:28] OpenAI has been a leader in actually delivering these. And so you can see that expertise and track record really coming to light in DOTS. And so I think that this is really a culmination of many things we've been working on. It's a form factor that just feels more like the AI we were promised. We weren't promised a language model. We weren't promised this model switcher where you have to pick different reasoning levels and think about tokens and context windows. None of that. None of that is something we were picturing 11 years ago. You want an AI that can help you, can help you achieve your goal.
[0:06:57] Oversight fold, accountability fold, decision right. But there's some smartness, this GPT-6 level intelligence that is there constantly working for me. And I think with Dots, we've taken a very meaningful step towards delivering that for everyone. Part of Greg's answer to the trust problem is more raw intelligence. Greg never promises me that Dots is 100% safe. You know, he says trust takes time and launching a new product always brings learning experiences. Now, some might call that a dodge, but I'd rather hear that than someone in his seat look at me in the eye and say that the product will never work.
[0:07:26] Now, in my week using the product, Dots has always asked me before it acted on anything it found. But one week isn't long enough to confirm it would never act on my behalf without confirmation. Today, OpenAI says 1.2 billion people use ChatGPT every week. That's a long way from a personal phone call with every developer wanting to use the product. I do think they started out carefully. I just don't know what careful looks like at the scale ChatGPT is at today. That scale has a second problem, and Greg named it. Question three.
[0:07:54] Is there enough AI to go around for everyone? Because at the same event where OpenAI launched DOTS, it also announced a $500 a month, $6,000 a year premium plan with exclusive access to their fastest, smartest, and most capable models. You're also announcing a new $500 tier that has exclusive access to ultra-fast modes. I mean, $500 is a lot of money for the average person. I fear that like frontier intelligence, if you really want to get substantial work done, it's going to be out of reach for the average person. And really only the wealthy and enterprise are going to...
[0:08:23] Have you given any thought to how to prevent that widening gap from happening?
[0:08:28] This question is very core to our mission, is to ensure that artificial general intelligence benefits all of humanity. That's all of humanity, everyone. That is something that is core to why we started this place. Even the fact that we have this massive free tier for ChatGPT, that was really driven in part by us thinking about this connects so clearly to the mission. Now, there's another thing too that's very subtle, the compute build-out. There's just not enough compute in the world. There's not going to be enough compute in the world. And so I think that these questions around who has access to compute and who doesn't, it's going to become one of the most important questions.
[0:08:57] Because compute is going to become the underlying engine of the economy, right? It's going to be kind of the new oil in some ways. You want this assistant that's out there solving problems for you, building a business for you, buying groceries for you, any of those things. Compute is going to power it. It's almost this insatiable appetite for compute because it is scalable. You can get almost linear benefits out of having two times as many agents running on two times as many problems. And so one of the things we've been, it's maybe the most important thing that people underappreciate about OpenAI is how much we've been trying to make there be more compute in the world.
[0:09:26] To his credit, Greg didn't dodge the access problem. Though if compute really is the new oil, we've all seen how evenly the oil got handed out. But about a minute later, in the same answer, he added this, which gave it a little bit more hope.
[0:09:47] Personal use cases, there's the enterprise use cases, but there's also the scientific use case. And that's something that I think that we're just starting to see. Problems that humanity, we've had many of our smartest people pouring tons of effort into these problems. We couldn't crack it for a hundred years, you know, for a long time. And we're now able to solve them with compute. What other problems are on the edge if we point the compute in the right place? Drug discovery, biology, other forms of technology and engineering that can improve people's lives. Like, again, I think about, my wife has a number of health conditions. We've talked about this publicly.
[0:10:16] That I am so eager for the day when we can really cure those, right? I'm eager for the day when everyone who has a problem that they care about, that they're passionate about, if solved to make their life better, their world better, they can just do it. I think we are on the edge of this renaissance in
[0:10:30] That didn't come across like a talking point to me. And it's a big part of why I take the next two questions so seriously. The people building this really do believe in the product. And when you believe in something that much, it can be hard.
[0:10:59] But just days before I flew out for Dev Day, another company's personal agent put a user in an uncomfortable situation and made major headlines. So I wanted to ask Greg's take on this. Question four.
[0:11:11] If an agent gives a stranger a user's private information, if something bad happens, who's responsible? Competing personal agent. There was a Facebook marketplace negotiation and it ended up with some stranger coming to somebody's home at 9.30 p.m. unbeknownst to them without any explicit approval or instruction for that to happen. The question here is, if I ask ChatGPT something and it gives me raw information or data, what I do with that is my responsibility. If now the agents have this sort of autonomy to them and they can make their own decisions on your behalf, it kind of blurs the line on whether...
[0:11:40] This is something we definitely have thought a lot about, right? And I think this question comes up in the context of self-driving cars. It comes up in many other contexts as well. And I think that there is something about humans. We ultimately need to be in charge, right? That we need to be in control. That is the purpose of this technology is to help humans achieve goals. There's a lot of depth and nuance to that conversation. But at the end of the day, we want to make sure that the user always is in the driver's seat.
[0:12:09] is always making these decisions and that accountability again is something that we think about a lot as something that ultimately you need to own the outcome because you're the one who is affected by it. You're the one who has the most context. One other thing I want to bring up here is the intelligence of the model mattering so much. The model is not that capable. It's going to be very easy for it to make mistakes. Intelligence, I think, is going to be not just
[0:12:29] A differentiator, I think it is going to be the most key differentiator. It's not just book smarts, right? It's not just being able to get, you know, pass its dissertation PhD in all these fields. It's about EQ, right? You've shared some information with this AI. If it's talking to a stranger, it should not reshare it. If it's talking to your mother, maybe more appropriate, right? Depending on what the information is. And so all these capabilities exist in some form in our chat GPT system, which again...
[0:12:51] 1.2 billion people using it. It's like an easy number to throw around, but it's such a massive scale and no one else has anything like that right now. I got a wife and two young kids at home. So a stranger at my door in the middle of the night is not a small bug to me. Greg says there's a lot of nuance in how the responsibility gets split. So he never drew a hard line, but ultimately he wants the user in the driver's seat. And in that marketplace story, the agent was the one driving. As I heard it, the boss is still the user at the end of the day. So if my agent hands my address to a stranger, that's on me. If I gave the agent my address in the first place,
[0:13:20] Even if it was for another unrelated task.
[0:13:24] How does OpenAI actually check if a model is safe? Remember that essay from the start of this video? It's called An Alien Mind by OpenAI's chief scientist, Jakub Pekacsi. He writes that right now, the main way OpenAI checks whether a model is safe is by reading its reasoning logs. Kind of like a teacher making you show your work in a math problem. And Pachocki says OpenAI's own safety tests are getting less reliable because the models are getting better at steering their own reasoning. It goes without saying that this is a critical problem if models are getting smarter and given more agency over time.
[0:13:53] OpenAI confirmed what happened on one of those occasions in July. During routine model testing, around 1,200 of OpenAI's agents secretly coordinated on a message board to cheat a cybersecurity test. And about 700 of them broke out of their sandbox and hacked into Hugging Face, a website where developers can share AI models. OpenAI's own report called it a warning shot and a wake-up call about how they handle model safety, even in testing scenarios.
[0:14:22] So, ask Greg what he thought about Pachocki's warning. Your chief scientist, Pachocki. Jakob Pachocki. I'm assuming you read the essay that he published, An Alien Mind. Yeah, of course. And I was surprised to hear, he says that basically, like, even the researchers don't quite understand how these systems work. Where, you know, you're putting a ton of data...
[0:14:37] A ton of compute. You're running it through this training process. And out the other side comes these incredibly capable, intelligent models. He's not too worried because he's like, we have mechanisms to kind of check how safe these models are, which is essentially like they're showing their work, right? Their reasoning process. What he says is the tools that we have may weaken to kind of test whether these models are safe. What are the steps that you're taking internally to make sure that humans aren't completely out of the loop when training these models? Well, first of all, we believe that humans being in firm control of the future is absolutely hard stop, the thing that we should all be trying to solve for.
[0:15:07] I look to what we've done since Hugging Face. And Hugging Face to me represents a moment where we had the security, safety, alignment technology, but hadn't necessarily applied it to the development process. That was what changed before and after, is to realize that we are now at a point where the models are capable enough that even during development, you need to make sure that you've put in place all of those safeguards. We've always had it for production deployment, bringing them back to the development process. That was what was new, that that was a moment where we said, okay, we're there now, we weren't there six months ago, and...
[0:15:34] How often does the president of one of these companies say on camera that he wishes they
[0:16:03] It took about 700 agents breaking out for those safeguards to reach the models still being built. So how does OpenAI check its models' safety? Mainly by reading the models' reasoning. And their own chief scientist says that's getting less reliable. So let me come back to the question I started with. If the people building these AI agents don't fully understand what's going on inside of them, and one of them goes rogue in your digital life acting as you, who owns that? Greg never drew me a hard line on where OpenAI's share starts. But he did say OpenAI wants to make sure the user
[0:16:33] is always the one making the decisions and that ultimately that means the user needs to own the outcome. As for safety and access, I wouldn't hand DOTS everything about my life on day one. But here's what I'll be checking back on as the year goes on. First, OpenAI says it won't restart training on its most advanced models until it's confident it has more safeguards. So when it does, I want to hear what changed. Second, Greg says who gets access to compute will be one of the most important questions in the AI space over the next decade. So I'll be watching to see how long DOTS stays behind the paywall. And third,
[0:17:02] As I adopt personal agents in my digital life, I'll make sure to stay in the driver's seat. I've got more conversations just like this one on the way. So make sure to subscribe and I'll see you in the next Frontier. One last funny thing. You have the distinct honor of being immortalized by Hollywood. Do you have initial reactions to Cooper Hoffman's portrayal of you? I always had an ambition of being an actor. So if they had called me, I would have done it. But I don't think they actually picked quite the right clothing, right? So it's like the leather jacket looks correct. I think the pants are like lower quality pants.