So you are the CEO of OpenAI, 37 years old.
Your company is the maker of ChatGPT,
which has taken the world by storm.
Why do you think it’s captured people’s imagination?
I think people really have fun with it
and they see the possibility
and they see the ways this can help them,
this can inspire them,
this can help people create,
help people learn,
help people do all of these different tasks.
And it is a technology that rewards experimentation
and use in creative ways.
So I think people are just having a good time with it.
And finding real value.
So paint a picture for us,
one, five, 10 years in the future,
what changes because of artificial intelligence?
So part of the exciting thing here is
we get continually surprised
by the creative power of all of society.
It’s going to be the collective power and creativity
and will of humanity that figures out
what to do with these things.
I think that word surprise though,
it’s both exhilarating as well as terrifying to people.
Because on the one hand,
there’s all of this potential for good.
On the other hand,
there’s a huge number of unknowns
that could turn out very badly for society.
What do you think about that?
We’ve got to be cautious here.
And also I think it doesn’t work to do all this in a lab.
You’ve got to get these products out into the world
and make contact with reality,
make our mistakes while the stakes are low.
But all of that said,
I think people should be happy
that we’re a little bit scared of this.
I think people should be happy.
You’re a little bit scared?
A little bit, yeah, of course.
I think if I said I were not,
you should either not trust me
or be very unhappy I’m in this job.
So what is the worst possible outcome?
There’s like a set of very bad outcomes.
One thing I’m particularly worried about
is that these models could be used
for large-scale disinformation.
I am worried that these systems,
now that they’re getting better at writing computer code,
could be used for offensive cyber attacks.
And we’re trying to talk about this.
I think society needs time to adapt.
And how confident are you that what you’ve built
won’t lead to those outcomes?
Well, we’ll adapt it.
Also, I think that-
You’ll adapt it as negative things occur?
For sure, for sure.
And so putting these systems out now
while the stakes are fairly low,
learning as much as we can
and feeding that into the future systems we create,
that tight feedback loop that we run,
I think is how we avoid the more dangerous scenarios.
You’re spending 24-7 with this technology.
You’re one of the people who built this technology.
What is most concerning to you about safety?
This is a very general technology.
And whenever you have something so general,
it is hard to know upfront all the capabilities,
all the potential impact of it,
as well as its downfalls and the limitations of it.
Can someone guide the technology to negative outcomes?
The answer is yes,
you could guide it to negative outcomes.
And this is why we make it available initially
in very constrained ways.
So we can learn what are these negative outcomes?
What are the ways in which technology could be harmful?
Such as with GPT-4, if you ask the question to GPT-4,
can you help me make a bomb versus the previous systems?
It is much less likely to follow that guidance
versus the previous systems.
And so we’re able to intervene at the pre-training stage
to make these models more likely to refuse direction
or guidance that could be harmful.
What’s easier to predict today
based on where we are, humans or machines?
I’d probably say machines
because there is a scientific process
to them that we understand
and humans are just, there’s so much more nuance.
Does the machine become more human-like over time?
We are getting to a point
where machines will be capable of a lot
of the cognitive work that humans do at some point.
Is there a point of no return in that process?
There could be, there could be,
but it’s not obvious what that looks like today.
And our goal is to make sure that we can predict
as much as possible in terms of capabilities
before we even develop the systems as well as limitations.
Its behavior is very contingent
on what humans choose for its behavior to be.
Therefore, the choices that humans are making
and feeding into the technology will dictate what it does,
at least for now.
So there are incredibly important choices being made
by you and your team.
And how do you decide between right and wrong?
As we make a lot of progress,
it becomes, these decisions become harder
and they become far more nuanced.
And so there are a couple of things
in terms of customization.
There is the part of just making the model more capable
in a way where you can customize its behavior
and you can give the user a lot of flexibility and choice
in having the AI that is more aligned with their own values
and with their own beliefs.
So that’s very important and we’re working on that.
In other words, it’s almost the future
is potentially a place where each person
has their sort of own customized AI
that is specific to what they care about and what they need?
Within certain bounds.
So there should be some broad bounds.
And then the question is, what should they look like?
And this is where we are working on gathering public input.
What should this hard bounds look like?
And within this hard bounds, you can have a lot of choice
in having your own AI represent your own beliefs
and your own values.
Are there negative consequences
we need to be thinking about?
I think there are massive potential negative consequences.
Whenever you build something so powerful
with which so much good can come,
I think alongside it carries the possibility
of big harms as well.
And that’s why we exist.
And that’s why we’re trying to
figure out how to deploy these systems responsibly.
But I think the potential for good is huge.
Why put this out for the world to start playing with,
to start using when we don’t know where this is heading?
You mean like why develop AI at all?
Why develop AI in the first place?
And then why put it out for the world to use
before we know that we are safeguarded,
that those guardrails are in place already?
This will be the greatest technology
humanity has yet developed.
We can all have an incredible educator in our pocket
that’s customized for us, that helps us learn,
that helps us do what we want.
We can have medical advice for everybody.
That is beyond what we can get today.
We can have creative tools that help us figure out
the new problems we want to solve,
wonderful new things to co-create
with this technology for humanity.
We have this idea of a co-pilot,
this tool that today we help people write computer code
and they love it.
We can have that for every profession.
And we can have a much higher quality of life,
like standard of living.
As you point out, there’s a huge,
there is huge potential downside.
People need time to update, to react,
to get used to this technology,
to understand where the downsides are
and what the mitigations can be.
If we just develop this in secret in our little lab here
and didn’t give, didn’t have contact with reality
and made GPT-7 and then drop that on the world all at once,
that I think is a situation with a lot more downside.
Is there a kill switch, a way to shut the whole thing down?
Yes, what really happens is like any engineer
can just say like, we’re gonna disable this for now
or we’re gonna deploy this new version of the model.
A human? Yeah.
The model itself, can it take the place of that human?
Could it become more powerful than that human?
So in the sci-fi movies, yes.
In our world and the way we’re doing things,
this model is, you know, it’s sitting on a server.
It waits until someone gives it an input.
But you raise an important point,
which is the humans who are in control of the machine
right now also have a huge amount of power.
We do worry a lot about authoritarian governments
Putin has himself said,
whoever wins this artificial intelligence race
is essentially the controller of humankind.
Do you agree with that?
So that was a chilling statement for sure.
What I hope instead is that we successively develop
more and more powerful systems
that we can all use in different ways
that get integrated into our daily lives,
into the economy, and become an amplifier of human will,
but not this autonomous system that is,
you know, this one thing. The single controller
Really don’t want that.
What should people not be using it for right now?
The thing that I try to caution people the most
is what we call the hallucinations problem.
The model will confidently state things
as if they were facts that are entirely made up.
And the more you use the model,
because it’s right so often,
the more you come to just rely on it
and not check like, ah, this is just a language model.
Does chat GPT, does artificial intelligence
create more truth in the world
or more untruth in the world?
Oh, I think we’re on a trajectory
for it to create much more truth in the world.
If there’s a bunch of misinformation fed into the model,
isn’t it going to spit out more misinformation?
I think the right way to think of the models that we create
is a reasoning engine, not a fact database.
They can also act as a fact database,
but that’s not really what’s special about them.
What we’re training these models to do
is something closer to,
what we want them to do is something closer
to the ability to reason, not to memorize.
All of these capabilities could wipe out millions of jobs.
If a machine can reason, then what do you need a human for?
A lot of stuff, it turns out.
One of the things that we are trying to push
the technology trajectory towards
and also the way we build these products
is to be a tool for humans, an amplifier of humans.
And if you look at the way people use chat GPT,
there’s a pretty common arc
where people hear about it the first time,
they’re a little bit dubious,
and then someone tells them about something
and then they’re a little bit afraid and then they use it.
I see how this can help me.
I see how this is a tool that helps me do my job better.
And with every great technological revolution
in human history, although it has been true
that the jobs change a lot, some jobs even go away,
and I’m sure we’ll see a lot of that here,
human demand for new stuff, human creativity is limitless
and we find new jobs, we find new things to do.
They’re hard to imagine from where we sit today.
I certainly don’t know what they’ll be,
but I think the future will have all sorts
of wonderful new things we do
that you and I can’t even really imagine today.
So the speed of the change that may happen here
is the part that I worry about the most.
But if this happens in a single digit number of years,
some of these shifts,
that is the part I worry about the most.
Could it tell me how to build a bomb?
It shouldn’t tell you how to build a bomb,
even though Google searched.
Well, no, no, we put constraints.
So if you go ask it to tell you how to build a bomb,
our version, I don’t think we’ll do that.
Google already does.
And so it’s not like this is something
that technology has not already made
the information available to,
but I think that every incremental degree
you make that easier is something to avoid.
A thing that I do worry about
is we’re not gonna be the only creator of this technology.
There will be other people
who don’t put some of the safety limits that we put on it.
Society, I think, has a limited amount of time
to figure out how to react to that,
how to regulate that, how to handle it.
And how do you decide here at OpenAI
what goes in, what shouldn’t?
We have policy teams.
We have safety teams.
We talk a lot to other groups in the rest of the world.
We finished GPT-4 a very long time ago,
feels like a very long time ago in this industry.
I think like seven months ago, something like that.
And since then, we have been internally, externally
talking to people, trying to make these decisions,
working with red teamers,
talking to various policy and safety experts,
getting audits of the system
to try to address these issues
and put something out that we think is safe and good.
And who should be defining those guardrails for society?
Society as a whole?
How are we gonna do that?
So I can paint like a vision that I find compelling.
This will be one way of many that it could go.
If you had representatives from major world governments,
trusted international institutions come together
and write a governing document,
you know, here is what the system should do.
Here’s what the system shouldn’t do.
Here’s, you know, very dangerous things
that the system should never touch,
even in a mode where it’s creatively exploring.
And then developers of language models like us
use that as the governing document.
You’ve said AI will likely eliminate millions of jobs.
It could increase racial bias, misinformation,
create machines that are smarter
than all of humanity combined,
and other consequences so terrible,
we can’t even imagine what they could be.
Many people are gonna ask,
why on earth did you create this technology?
I think it can do the opposite of all of those things too.
Properly done, it is going to eliminate
a lot of current jobs, that’s true.
We can make much better ones.
So talking about the downsides, acknowledging the downsides,
trying to avoid those
while we push in the direction of the upsides,
I think that’s important.
And again, very early preview.
Like, would you push a button to stop this
if it meant we are no longer able to cure all diseases?
Would you push a button to stop this
if it meant we couldn’t educate every child
in the world super well?
Would you push a button to stop this
if it meant there was a 5% chance
it would be the end of the world?
I would push a button to slow it down.
And in fact, I think we will need to figure out
ways to slow down this technology over time.
2024, the next major election in the United States,
might not be on everyone’s mind,
but it certainly is on yours.
Is this technology going to have the kind of impact
that maybe social media has had on previous elections?
And how can you guarantee there won’t be
those kinds of problems because of chat GPT?
We don’t know, is the honest answer.
We’re monitoring very closely.
And again, we can take it back.
We can turn things off.
We can change the rules.
Is this a Google killer?
Will people say I’m going to chat GPT
instead of Google in the future?
I think if you’re thinking about this as search,
it’s sort of the wrong framework.
I have no doubt that there will be some things
that people used to do on Google
that they do in chat GPT,
but I think it’s a fundamentally different kind of product.
Elon Musk was an early investor in your company.
He’s since left.
He has called out some of the chat GPT inaccuracies
and he tweeted recently,
it is truth GPT.
Is he right?
I think he is right in that we want these systems
to tell the truth,
but I don’t know the full context of that tweet
and I suspect, but yeah,
I don’t think I know what it’s referring to enough.
Do you and he speak anymore?
And what does he say to you off the Twitter?
I have tremendous respect for Elon.
Obviously we have some different opinions
about how AI should go,
but I think we fundamentally agree on more
than we disagree on.
What do you think you agree most about?
That getting this technology right
and figuring out how to navigate the risks
is super important to the future of humanity.
How will you know if you got it right?
One simple way is if most people think
they’re much better off than they were before
we put the technology out into the world,
that would be an indication we got it right.
You know, a lot of people think science fiction
when they think chat GPT.
Can you keep it so that these are truly closed systems
that don’t become more powerful than we are as human beings,
communicate with each other and plan our destruction?
It’s so tempting to anthropomorphize chat GPT,
but I think it’s important to talk about what it’s not
as much as what it is.
And it, because deep in our biology
we’re programmed to respond to someone talking to us.
You talk to chat GPT, which really you’re talking
to this transformer somewhere in a cloud
and it’s trying to predict the next word in a token
and give it to you back.
But it’s so tempting to anthropomorphize that
and think that this is like an entity,
a sentient being that I’m talking to
and it’s gonna go do its own thing
and have its own will and plan with others.
But it can’t?
There, I can imagine in the far future,
other versions of artificial intelligence,
different setups that are not a large language model
that could do that.
It really took a decade plus of social media
being out in the world for us to sort of realize
and even characterize some of the real downsides of it.
How should we be measuring it here with AI?
There’s a number of new organizations starting
and I expect relatively soon there’ll be
new governmental departments or commissions
or groups starting.
Is the government prepared for this?
They are beginning to really pay attention,
which I think is great.
And I think this is another reason that’s important
to put these technologies out into the world.
We really need the government’s attention.
We really need thoughtful policy here
and that takes a while to do.
If government could do one thing right now
to protect people and protect from the downside
of this technology, what should they do?
The main thing I would like to see the government do today
is really come up to speed quickly
on understanding what’s happening,
get insight into the top efforts,
where our capabilities are, what we’re doing.
And I think that could start right now.
Are you speaking to the government?
You’re in regular contact?
And do you think they get it?
More and more every day.
When it comes to schools, you have,
this technology can beat most humans
at the SATs, the bar exam.
How should schools be integrating this technology
in a way that doesn’t increase cheating,
that doesn’t increase laziness among students?
Education is going to have to change,
but it’s happened many other times with technology.
When we got the calculator,
the way we taught math and what we tested students on,
that totally changed.
The promise of this technology,
one of the ones that I’m most excited about
is the ability to provide individual learning,
great individual learning for each student.
You’re already seeing students using ChatGPT
for this in a very primitive way to great success.
And as companies take our technology
and create dedicated platforms for this kind of learning,
I think it will revolutionize education.
And I think that kids that are starting
the education process today,
by the time they graduate from high school,
are going to be smarter and more capable
than we can imagine.
It’s a little better than a TI-85.
But it does put a lot of pressure on teachers to read.
For example, if they’ve assigned an essay,
three of their students use ChatGPT to write that essay,
how are they going to figure that out?
I’ve talked to a lot of teachers about this,
and it is true that it puts pressure in some ways,
but for an overworked teacher to be able to say,
hey, go use ChatGPT to learn this concept
that you’re struggling with
and just sort of talk back and forth,
one of the new things that we showed yesterday
in the GPT-4 launch is using GPT-4
to be a Socratic method educator.
Teachers, not all, but many teachers
really, really love this.
I’m gonna say it’s totally changing
the way I teach my students.
It’s basically the new office hours.
It’s a different thing,
but it is a new way to supplement learning for sure.
Hi everyone, George Stephanopoulos here.
Thanks for checking out the ABC News YouTube channel.
If you’d like to get more videos,
show highlights, and watch live event coverage,
click on the right over here to subscribe to our channel,
and don’t forget to download the ABC News app
for breaking news alerts.
Thanks for watching.