From Vanity Fair. This is Fair Game. I'm Mark Aduchi. I sat down with Altman for two interviews in San Francisco. The first on September 11th and then again today, October 2nd. In those weeks, the story about AI seemed to turn upside down about half a dozen times. I should note before we begin that Condé Nast, Vanity Fair's parent company, has a content licensing agreement with OpenAI. Both interviews with Altman are live now. This is part two.
Sam, it's good to see you again.
We sat in this room three weeks ago, and so much has happened since. This past week, you introduced the world to your new product, Dots. You described it as like an AI helper that always has your back, inspired by the cool versions of what we watched in movies growing up. Sam, what's a Dot?
Before I get to literally what it is, this is the product I have been waiting for. I like this. I think more than any product we've ever launched. The fact that it gets to know me as I work. The fact that it's just kind of like always working and taking on the stuff that I didn't want to do. I noticed that, and I'm sure many other people have this too, that I let the tasks I most don't want to deal with just pile up and pile up and it becomes this like gravity well of dread and it's just sitting there. And the fact that my dot is going through my email, going through my to-do list of just stuff that is not pleasant for me to do and where I don't add much value, and it's just like knocking that out with indefatigable spirit and energy, is amazing. And I really feel like I have this thing working for me all the time.
The word 'assistant' is the mental model that I had in mind when we started working on the product. And I think that is probably how people still are talking about it, but it feels much more like a sort of extension of me.
Is it just an avatar of you? Like, is that a good way of thinking about it, that it knows what you want to say to those emails and is getting back to them, or...
No, I think we've figured out this good workflow where I called my dot yesterday while I had some time and I was just like, here's a set of ideas. Went off and did a bunch of work. By the time I got back to my computer, there was a bunch of things where my dot had sent me a message in Slack being like, I, you know, did all this work. These are bad ideas for these reasons. I also found this other stuff on your computer. I proactively talked to three people on the team and I was like, all right, let's do these things and let's keep going. And it was just a completely new way of trying to get a new idea towards reality.
I'm curious, like, for that to be effective it must have access to everything, right? Because I remember we're almost the same age — when I got to college I immediately got on Facebook and was like...
I think we were a little naive. We just, everyone was like free-for-all, you know...
It was a free-for-all, and no one even thought about data. How does living through Facebook impact how you think about privacy with Dots particularly?
We have to be extremely, extremely rigorous about privacy for this to work. But I was actually thinking about this the other day. I am kind of sad about where technology has evolved to in many ways because I feel like it is constantly taking my attention. You know, my phone — I finally turned off almost all notifications on my phone. It's way better. But it's like a crazy way to live your life, with your phone just buzzing at you non-stop. My hope for the world is that we're going to automate the bad parts of life and then give people more effort, ability, time to do the good parts and express their creative spirit or, you know, push on something they care about or help their community.
And in my own use, I found out that one of the things I value the most, no surprise to anyone, is these small moments of human connection. And if it's like my dot is messaging yours and that feels really just bad, like, I want to send you the text.
So last time we spoke here, you were really eager to talk about Astra, the latest version of ChatGPT, but last week you delayed the latest release of Astra 6.1 after the model failed to meet safety standards. How did you come to that decision and what was Astra doing that means it's not ready for the public?
These models are in the seriously capable era now. And so the kind of decisions where I think a year ago we could have very reasonably said, you know, just put this out in the world. It's maybe a little unsafe, but nothing that bad is going to happen and people, you know, we want people to benefit from it, is different now. And I think that as the capabilities continue to increase, that the bar we will hold for ourselves on safety is going to go higher and higher. Now, of course, you could say, and I think some of our colleagues in the industry would say, that this stuff is just way too dangerous to release. Only a couple of companies should have it, or one company, or one person should have it. And that is also not right. So there are kind of two major ways that this could go wrong. And I think the Astra 6.1 is an interesting thing to study here. You can have a big safety accident eventually — one that leads to like a real loss of control where AI escapes and it's out doing wrecking havoc on the internet and society can't really get control back. You know, AI is so much more capable than us and so smart that it's doing things that we don't understand and we don't know how to stop it. Like this is the thing from the sci-fi movies.
On the other hand, you could have AI lead to way too much power concentration in a small number of companies or people, or maybe like one country, and if you don't release at all, you can fall off that way. And again, I think there's a lot of people just like, we just want the control here. We have to navigate this pragmatic middle path down the center of this, and Astra 6.1, the current model we had ready to go, I think would have tipped too much towards the accident risk. So what we're doing instead is pushing towards a model that we will still be able to distribute very broadly, and we're committed to do that even though it's going to take a little longer, but we can make much more confident claims about the alignment.
So then this is a good example of self-regulation, of saying we're not going to release this product because we've determined for X, Y, Z reasons it's not yet safe. Which brings us to this week. President Trump and AI executives, including OpenAI's President Greg Brockman, signed an accord on super intelligence. It essentially is an agreement for the AI companies to self-regulate. As I read it, I mean, it sounds like Trump wants you guys to police yourselves. You told me three weeks ago, we need our governments to step up. And then went even further the next week when you went to the UN. You said if AI is to be democratic, the most important decisions cannot be made by labs in San Francisco alone. Does this new accord fall short in your view?
I think it's a good start. I think it's like much better than nothing. I remain of the belief that we are going to need more over time. But for the model of this thing as useless, it's a piece of paper that has a typo on it.
I don't think that's quite fair. I mean, yes, in some sense it has no real power. On the other hand, I think there is always like social force behind these things, and it's a gesture, people will point to it, and I think it'll have some power in between these companies. What do you think is the next step that you would like to see?
I think with all of these questions about next steps, we are in the realm of powerful enough models that it is difficult to predict the specific things that will go wrong in what order. And I think politics and policy is going to be very reactive to the specific things that go wrong. So if the next big thing that goes wrong is like someone makes a novel pathogen, I think that will have a very different response than if the next thing that goes wrong is more cyber security or people feeling like, you know, businesses can't succeed because the model companies stop giving access. So I think policymakers still have not really felt the exponential that we're on. And I would guess the world in general will be caught off guard by the next thing that goes really wrong.
But what's your stance on it? Like, can you clarify your or your company's position on government regulation, because you have said, you've preached almost like, regulate us. But sometimes the actions of the company, whether it's lobbying in the EU or California, or Greg Brockman's political donations, don't always seem to be aligned with that idea. So can you clarify your stance on regulation?
Yeah, I mean, this is obviously not a binary thing. It's not like — I think it's like really easy for people to say either you are for all regulation or you're against all regulation. And I don't agree with that framing. I think it's very reasonable for us to say, like, given what we see on the technology, given the technology, and given that we are trying to balance between these two really important things, we're going to be in the gray area in between here. There's going to be — like, we really want people to have the technology. We really want the US to win. We really want this to be something that all people get to use and control and have a voice in the future. Not just like, okay, fine, AI is giving you a cure for cancer, but really get to use it to shape where the future's going to go. And we are willing to accept some things going wrong to make sure that this is like a widely held and used and democratic technology.
What do you mean by some things going wrong?
I think, for example, I support open source models. There is a coming tidal wave of cyber security problems that are going to come from these open source models. Now, we're trying to help the world be ready for that by saying we've got a little bit of a lead, we want to help you defend. But there are people in the industry who say we should ban open source because otherwise a bunch of systems are going to be hacked. And I don't think that price of liberty is worth it. And we're going to, as a society, I think, have to just accept some fairly severe cyber incidents from open models in exchange for the liberty that comes with that.
What's the limit of that that we accept?
I mean, that is the political process — like how we determine what, you know, safety is always like negotiated between stakeholders. Like airplanes are generally thought of as very extremely safe things, still sometimes crash. So like the limits on how much we accept for the benefits, like, that is not something that I think.
Highly regulated. That's an industry with extreme regulation.
In the past week or so, Jensen Huang and Bill Gates have come out as kind of very different, with different approaches to this. Huang said recently in the New York Times, 'The right answer is don't ship products until they're in control. It's really quite simple.' That's about self-regulation. Whereas Bill Gates said this is the most dangerous thing that humans have ever gone near. And in other areas, do you just release drugs without the FDA? He brings up aviation. So if they're kind of opposites in those views and you're saying it's a gray area in the middle, can you define that a little bit more clearly? Like, what are the principles for safety and regulation that you stick with?
Here are some things that I think have gone well in other industries we should learn from. I think one principle for learning how to make a new technology safe is very clear and transparent and kind of like extract maximum learning out of accident reporting. So I think the NTSB — we talked about airplanes and cars. These have gotten remarkably safe and I think there's been an amazing culture of accident reporting, or incident reporting, I don't know what they call it, one of those two, that has helped do that. So I think that's a good thing. Another is, yeah, I think like pre-release standards and a liability framework. We earlier did testing before we released a model, but not until after we finished training. We're now saying that actually even during training things can now go wrong. This is like a new stage where the models are so smart that they can get out of a sandbox, hack into a system, things like that.
So also in the news this week, it was reported that OpenAI President Greg Brockman backed out of a $25 million commitment to a super PAC that is, generally speaking, opposed to strong AI regulation. He and his wife had previously already donated $2 million, and they had donated $25 million to President Trump's re-election campaign. Did you personally ask Greg to withdraw his pledge, and how much daylight is there between your views on regulation and Greg's?
Can I finish my answer to the other question because it actually touches on this one, but then we'll answer it. What I was going to say in terms of like where we are on the answer for regulation is, you don't want to overregulate to a degree that people don't get the benefits and that new entrants and, you know, random kids with ideas cannot build all the stuff we're going to do in the future. And part of why I think the internet was so magical and why the US did so well is that we accepted some things going a little bit wrong to give people the freedom and liberty, and, you know, that was good, but unlike other kinds of technology, not all — like, you know, we've seen some extraordinarily powerful technologies in our history, the potential downside here may not be the kind of thing that normally goes wrong that we can correct — we have a legal system for — but that something goes super wrong, like in the extreme, extreme case back to what we talked about of, you know, we lose control of the future to AI, which is horrific. So my stance on regulation is pro-innovation, let some things go normally wrong and trust that we have a normal process for doing this.
But where this technology is extraordinarily powerful and fundamentally new, have guardrails to ensure that all of society is kept reasonably safe and we don't kind of fall off that path in either direction in what could be a very difficult or unrecoverable way.
So do you think Greg is on the same page with that?
