OpenAI CEO Sam Altman on AI governance, ethics, and innovation | A conversation with BiGS

Academy: stories and events
54 followers
So I would love if you could, you touched on this, but just describe in a little bit more detail
how and why you felt like you had to move from a non-profit structure to a for-profit structure,
and like was that like a light bulb going off? The simple thing was we just needed vastly more
capital than we thought we could attract, not that we thought we tried, that we were able to attract
as a non-profit. You know, we knew that scaling computers was going to be important, but we still
really underestimated how much we needed to scale them. Yeah, so does that suggest that nobody can
do AI as a non-profit in any kind of a meaningful way? No, there are other things that you can do
for sure, but to be at the front of scaling research, I think you probably can't do that
as a non-profit. What about as government? Government could do anything they wanted,
right? Right, well, but yes and no. So again, one of the things we try and pick at is that there's
some things that firms are much better at than government. So, I mean, do you think the Department
of Energy or the Department of Defense could do open AI as well as you guys do? The Apollo
program was pretty impressive. So, do you think that could possibly be another route? I think,
you know, in some sense in a well-functioning society, I think this would be a government
project. Yeah, but that seems to be a big if. Given that that's not happening, I think it's
better that it happens this way and happens, you know, like as an American project. Okay,
you also then said a couple of times that there's a role for the government or you,
I think you used the word society to play in setting some of the guardrails or in directing
things so that equity is built in. More than a role, like maybe that's like society should play
the role there. But how does that happen? Like society is not like a thing. No, it's a complex
negotiation of a lot of different stakeholders, but we kind of do, you know, we do this,
we get this right over time. Give me an example. How do we get it right? Airplanes are really safe.
Boeing, problems aside, in general, airplanes have been remarkably safe.
That happened largely as a result of kind of old-fashioned regulation.
I would say it happened because people want to get on an airplane and know that they're not
going to die and society responds to that. Well, specifically government response. Airplane
manufacturers also, if they want to keep selling airplanes that people want to get on,
definitely governments have a role. The scientists and engineers working on it feel a sense of
obligation. I think it's a, I think it is a multi-party thing, but there's a huge role
for regulation for sure. And are you, how are you, you're clearly going to be a player in this
game as it unfolds, a big player. I'd put money on it. So how are you thinking about it? And I compare
it a little bit, having been at HBS when the internet happened, sort of the folks who were you
in the, when social media and the internet were happening, were saying no government, cyberspace
wants to be free, you know, keep the black helicopters out of our business. How are you
thinking about sort of your position vis-a-vis the government? I mean, I hope it's like a real
partnership. It's been fairly constructive so far. Too much regulation clearly slows innovation.
Not enough clearly leads to a whole set of other problems. Everyone agrees, I think,
on those two statements. And then there's just a question of where you, where you should be
on that spectrum. And do you think there's particular parts of the government that,
I know you're working with DHS, but DOE, DOD, I mean, where would you most like to build bridges?
I mean, I, look, it's easy to be too optimistic and drink the Kool-Aid too much in this,
but I think this is going to touch like all parts of society and thus all parts of government.
But those are hard things to build.
For sure. And, but, you know, like the internet touched all parts of society pretty much.
Right. But arguably we got some parts of that really wrong.
I was gonna say, we got a lot of that wrong. We got some parts right. And I don't think most
people would say undo the internet, even with all of the problems.
That is kind of interesting watching my students. A lot of them, having grown up on social media,
would like that part of the internet to go away.
For sure. And there will be parts of AI that you will, you would like to go away. But how
many of your students would push a button and like destroy all computers and the entire internet?
Oh, none. None.
So there'll be parts you don't like. There'll be parts of your students 10 years from now
either. But on the whole, I think it'll be a huge net win.
And so, and some of the students asked you this, but maybe I'll ask you to go a little deeper in
it. If you think about from either the societal perspective or the commercial perspective,
what are the pieces of AI that you're most excited about?
Personally, I think greatly increasing the rate of scientific discovery is what I'm most excited
about. I believe that that is, if you zoom all the way out, that is how the world gets
sustainably better. And I think doing more of that is awesome. A lot of other areas too. I think
there'll be incredible AI tutors, incredible AI medical advisors. But personally speaking,
I'm so excited for AI for science.
Any particular parts of science?
I mean, I guess if I had one personal thing to pick, I would say like, you know, go solve all
physics. But no, I'm excited for it everywhere.
So why all of physics?
I have incredible personal curiosity. And I also believe that, you know, the more we can
understand about physics, the more we can manipulate the universe. And I don't know
what will come of that, but it seems probably important to find out.
So then the obvious question, what happens to the physicists?
I was being a little facetious there. If there's truly no problems left in physics at all,
then I don't know. What I suspect is that we answer the current questions and find more and
harder and more interesting problems.
So have you read this piece by John Maynard Keynes, Economic Prospects for Our Grandchildren?
No.
Okay, I have to send it to you. It's brilliant. So Keynes wrote it in 1930. And he's talking
about what the world is going to look like 100 years.
Oh, I know.
Because it's a similar set of issues that once you solve all the problems,
and he comes up with sort of a weird answer to it, that we're just going to make art and
not care about money. I mean, do you think about those kinds of existential questions?
What happens to the physicists? What's happened?
Everybody has been debating this for a long time. And he was, I think, sort of right and
sort of wrong. There is some sense in which you could go back to the 1930s and look at a
lot of people today, you would say, they are just making art, and they don't care about money,
they're all fine. But that's not their experience of it. If you could look at the people today,
and go forward another 100 years, and ask them what they think about those people of the future's
jobs, we'll say, you know, some weird definition of art, but okay, they're making art, and none
of them care about money, they have unbelievable material abundance. And you think a little bit
harder and say, well, they seem busy and stressed. And they seem to be getting a great deal of
satisfaction out of the work they're doing. And they don't quite call it art. And, you know,
they still feel like they don't have enough. And they want more. And they're still finding
new ways to compare each other, compare with each other. And, you know, maybe they're now,
instead of like, looking at whether their house is bigger or smaller than their neighbors,
they're looking at whether they own a nicer galaxy or less of a nice one than their neighbors. But
I think the human drives are just going to keep going.
Human nature doesn't.
I don't think so.
Doesn't seem to change. So one of the things I write about in my weird books is the prediction
that we're going to fall in love with robots, which is an exaggerated version of what I actually
do think is true, that we are going to embrace AI entities as emotional companions. What's your
sense? I mean, are you a fan of the kind of singularity type vision where there is this
merging and we start to build emotional connections with things that are created rather than born?
Yeah, I suspect that entertainment in the future will look quite different.
And maybe it does. You know, maybe a lot of the time spent watching TV or playing video games
today is replaced with an interactive experience with AIs. And we probably do derive some version
of companionship from that. But for most of us, I think the premium on the real human, authentic
version of it goes up, not down in that kind of a world. So I'm sure we'll be talking to
AIs a lot, and I'm sure it'll be a fantastically entertaining world.
But I would bet that there's something deep about human nature where just the knowledge
that it's another real person or not really matters. Although sometimes, you know, I talk to
some of the people I care about the most entirely through text, because we don't live in the same
city. So I might not know if they were being replaced. I might be fine with that if they were
dead. Well, that's the question. If you knew it was not a real person behind it, would you feel
any different? And I would posit that you would. We're going to find out. Right, we're going to
find out. But I think you will. Okay, all right, so we should make a bet on it. I would be happy
to see you 100 years from now. I'd give you nice odds. Okay. So I want to go back to one other
piece that you raised in your talk earlier. One of the kids asked the question about equity.
Yeah. And you had a very nice answer to that, which I agree with that. The technology is probably
going to drive to equity, but it's going to need some nudging. What kind of nudging do you have
in mind? Well, first of all, I think it may not need any nudging at all. Sometimes technology
doesn't. And I also think that knowing the right kind of nudging academically is hard to do. You
kind of have to do it experientially. I'm sure all of my ideas right now are totally wrong.
I think a tight feedback loop and watching where the problems are being created, fixing those,
finding new problems, that's kind of how I believe we'll have to do the nudging.
Although it's hard, right? I just came back from Denmark. You can see doing that in a country like
Denmark. Small, they have really good data. The US takes 30 years to pass any kind of regulation.
So how do you think about that nudging? I mean, I hope the companies will do it.
Okay. Okay. I think, yeah, if you're relying on US regulation, that would be...
I actually taught your piece last night, two nights ago, the Moore's law for everything.
I haven't read that in a while. How did it hold up?
No, I think it's quite... I think it's very interesting. Because you're talking about
changes in the taxation system, which are not particularly far-fetched. I mean,
they're doable. Whether or not they're doable in the current political situation is a different
question. But basically talking about a tax on land and...
And technology, basically. Yeah, companies.
I mean, I like it. Would you still...
Oh, I still think that's right, for sure. I just, I don't remember all...
No, I think that in a world of AI, taxing income and corporate profits primarily is
not quite right. We do need to switch to a tax on capital somehow, it seems to me.
Yeah. No, I think that's a really interesting idea that hasn't gotten sufficient
traction. Because it's different than a UBI. I mean, it's connected in some ways, but it's
different. If you had one wish for AI, what would it be?
Love humanity.
Can you build that in?
I think so.
How do you try?
One of the things that has worked surprisingly well is the ability to align an AI system to
behave in a particular way. And so if we can articulate what that means
in a bunch of different cases, yeah, I think we can get the system back that way.
Do you think open AI can and should have the power to make its own rules? Why should we trust you?
I don't think you should.
I don't think we do. I don't think we will. I don't think we should.
By the way, that does not necessarily mean that the right thing to do is to hand it over to a
government bureaucrat either. The directional kind of thing that I think is interesting
is you could imagine one of these systems chatting with all of its users saying,
explain to me your value system, and then I'm going to collectively align across that.
I am a believer that the users of technology, the people impacted by it the most, deserve
a lot of say in its governance. So you could do something like that.
Is that feasible, especially given the vast sums of money?
Totally feasible. Again, I think that's really a new
thing that is enabled by this technology that just wasn't possible before.
So is that something like what Anthropic is pushing towards?
I don't know what they're doing. I'm not sure.
It seemed a little naive, the idea of we're going to put our design principles out there
for kind of crowdsourcing. I think it's very good to say here's how
we're making these trade-offs of what a system will do and not do. But what I meant more is
I'm interested in the thought experiment of an AI chats with you for a couple of hours
about your value system and does that with me and with everybody else and then says,
okay, I can't make everybody happy all the time.
So it's essentially sort of using AI to come up with a consensus.
Well, using AI to talk to everybody to understand the consensus at a deep level of billions of
people, which we can't do without AI probably. Okay. But it would just be interesting to line
up something like that against the founding documents of this country, which are explicitly,
for better or worse, against majority rule. Now, what you're describing is different,
but I guess foundational pretty quickly. Well, but it is fundamentally about democracy
with some power checks on it. Yeah. Yeah. You'd have to think about
what those look like. For sure.
Yeah. So if the consensus is that it's bad AI, that it's evil,
sort of how do you come up with the checks? Well, I think that should be part of what
people should vote on for the governance structure. Yeah. Awesome. Thank you.
Thank you. You've got to go be pitched. All right.
Thank you so much.
