Greg Brockman: OpenAI and AGI
AI 与机器学习商业与创业音乐与艺术技术与编程心理与人性
📋 章节目录
暂无章节信息
🔑 关键词
ableopenaiagihumanscaleimportantgoinginterestingsystemsideasdonintelligencecompanythinkinghardmissionhumansbettersurepowerful
💬 精彩语录
暂无语录
🎙️ 完整对话（2114 条）
Lex Fridman (00:00.000)
The following is a conversation with Greg Brockman.
以下是与格雷格·布罗克曼的对话。
Lex Fridman (00:02.880)
He's the cofounder and CTO of OpenAI,
他是 OpenAI 的联合创始人兼 CTO，
Lex Fridman (00:05.360)
a world class research organization
世界一流的研究组织
Lex Fridman (00:07.440)
developing ideas in AI with a goal of eventually
开发人工智能创意，最终目标是
Lex Fridman (00:10.840)
creating a safe and friendly artificial general
创造一个安全友好的人造将军
Greg Brockman (00:14.200)
intelligence, one that benefits and empowers humanity.
智慧，一种造福人类并赋予人类权力的智慧。
Lex Fridman (00:18.800)
OpenAI is not only a source of publications, algorithms, tools,
OpenAI 不仅是出版物、算法、工具的来源，
Lex Fridman (00:23.080)
and data sets.
和数据集。
Lex Fridman (00:24.480)
Their mission is a catalyst for an important public discourse
他们的使命是重要公共话语的催化剂
Greg Brockman (00:28.160)
about our future with both narrow and general intelligence
用狭义和广义的智能来思考我们的未来
Lex Fridman (00:32.720)
systems.
系统。
Greg Brockman (00:34.040)
This conversation is part of the Artificial Intelligence
这段对话是人工智能的一部分
Lex Fridman (00:36.660)
podcast at MIT and beyond.
麻省理工学院及其他地方的播客。
Greg Brockman (00:39.560)
If you enjoy it, subscribe on YouTube, iTunes,
如果您喜欢，请在 YouTube、iTunes、
Lex Fridman (00:42.760)
or simply connect with me on Twitter at Lex Friedman,
或者直接在 Twitter 上联系我：Lex Friedman，
Greg Brockman (00:45.680)
spelled F R I D. And now, here's my conversation
拼写为 F R I D。现在，这是我的对话
Lex Fridman (00:50.240)
with Greg Brockman.
与格雷格·布罗克曼。
Lex Fridman (00:52.800)
So in high school, and right after you
所以在高中，就在你之后
Lex Fridman (00:54.440)
wrote a draft of a chemistry textbook,
写了一本化学教科书的草稿，
Greg Brockman (00:56.680)
saw that that covers everything from basic structure
看到这涵盖了从基本结构到所有内容
Lex Fridman (00:59.080)
of the atom to quantum mechanics.
Lex Fridman (01:01.400)
So it's clear you have an intuition and a passion
Lex Fridman (01:04.360)
for both the physical world with chemistry and now robotics
Greg Brockman (01:09.880)
to the digital world with AI, deep learning, reinforcement
Lex Fridman (01:14.200)
learning, so on.
Lex Fridman (01:15.400)
Do you see the physical world and the digital world
Lex Fridman (01:17.360)
as different?
Lex Fridman (01:18.640)
And what do you think is the gap?
Lex Fridman (01:20.520)
A lot of it actually boils down to iteration speed.
Greg Brockman (01:23.320)
I think that a lot of what really motivates me
Lex Fridman (01:25.240)
is building things.
Greg Brockman (01:26.520)
I think about mathematics, for example,
Lex Fridman (01:28.960)
where you think really hard about a problem.
Greg Brockman (01:30.880)
You understand it.
Lex Fridman (01:31.680)
You write it down in this very obscure form
Greg Brockman (01:33.460)
that we call a proof.
Lex Fridman (01:34.560)
But then, this is in humanity's library.
Greg Brockman (01:37.600)
It's there forever.
Lex Fridman (01:38.440)
This is some truth that we've discovered.
Greg Brockman (01:40.520)
Maybe only five people in your field will ever read it.
Lex Fridman (01:43.040)
But somehow, you've kind of moved humanity forward.
Lex Fridman (01:45.400)
And so I actually used to really think
Lex Fridman (01:46.900)
that I was going to be a mathematician.
Lex Fridman (01:48.600)
And then I actually started writing this chemistry
Lex Fridman (01:51.000)
textbook.
Greg Brockman (01:51.600)
One of my friends told me, you'll never publish it
Lex Fridman (01:53.600)
because you don't have a PhD.
Lex Fridman (01:54.840)
So instead, I decided to build a website
Lex Fridman (01:57.960)
and try to promote my ideas that way.
Lex Fridman (01:59.920)
And then I discovered programming.
Lex Fridman (02:01.440)
And in programming, you think hard about a problem.
Greg Brockman (02:05.280)
You understand it.
Lex Fridman (02:06.040)
You write it down in a very obscure form
Greg Brockman (02:08.000)
that we call a program.
Lex Fridman (02:10.000)
But then once again, it's in humanity's library.
Lex Fridman (02:12.200)
And anyone can get the benefit from it.
Lex Fridman (02:14.080)
And the scalability is massive.
Lex Fridman (02:15.540)
And so I think that the thing that really appeals
Lex Fridman (02:17.540)
to me about the digital world is that you
Greg Brockman (02:19.420)
can have this insane leverage.
Lex Fridman (02:21.920)
A single individual with an idea is
Greg Brockman (02:24.320)
able to affect the entire planet.
Lex Fridman (02:25.800)
And that's something I think is really
Greg Brockman (02:27.400)
hard to do if you're moving around physical atoms.
Lex Fridman (02:30.160)
But you said mathematics.
Lex Fridman (02:32.400)
So if you look at the wet thing over here, our mind,
Lex Fridman (02:36.800)
do you ultimately see it as just math,
Lex Fridman (02:39.760)
as just information processing?
Lex Fridman (02:41.720)
Or is there some other magic, as you've seen,
Lex Fridman (02:44.320)
if you've seen through biology and chemistry and so on?
Lex Fridman (02:47.000)
Yeah, I think it's really interesting to think about
Greg Brockman (02:48.640)
humans as just information processing systems.
Lex Fridman (02:50.920)
And that seems like it's actually
Greg Brockman (02:52.560)
a pretty good way of describing a lot of how the world works
Lex Fridman (02:57.160)
or a lot of what we're capable of, to think that, again,
Greg Brockman (03:00.640)
if you just look at technological innovations
Lex Fridman (03:02.760)
over time, that in some ways, the most transformative
Greg Brockman (03:05.480)
innovation that we've had has been the computer.
Lex Fridman (03:07.720)
In some ways, the internet, that what has the internet done?
Greg Brockman (03:10.520)
The internet is not about these physical cables.
Lex Fridman (03:12.720)
It's about the fact that I am suddenly
Greg Brockman (03:14.520)
able to instantly communicate with any other human
Lex Fridman (03:16.520)
on the planet.
Greg Brockman (03:17.640)
I'm able to retrieve any piece of knowledge
Lex Fridman (03:19.640)
that in some ways the human race has ever had,
Lex Fridman (03:22.640)
and that those are these insane transformations.
Lex Fridman (03:26.040)
Do you see our society as a whole, the collective,
Lex Fridman (03:29.320)
as another extension of the intelligence of the human being?
Lex Fridman (03:32.240)
So if you look at the human being
Greg Brockman (03:33.400)
as an information processing system,
Lex Fridman (03:35.040)
you mentioned the internet, the networking.
Lex Fridman (03:36.880)
Do you see us all together as a civilization
Lex Fridman (03:39.320)
as a kind of intelligent system?
Greg Brockman (03:41.640)
Yeah, I think this is actually
Lex Fridman (03:42.840)
a really interesting perspective to take
Lex Fridman (03:44.840)
and to think about, that you sort of have
Lex Fridman (03:46.680)
this collective intelligence of all of society,
Greg Brockman (03:49.480)
the economy itself is this superhuman machine
Lex Fridman (03:51.640)
that is optimizing something, right?
Lex Fridman (03:54.400)
And in some ways, a company has a will of its own, right?
Lex Fridman (03:57.960)
That you have all these individuals
Greg Brockman (03:59.040)
who are all pursuing their own individual goals
Lex Fridman (04:00.800)
and thinking really hard
Lex Fridman (04:01.960)
and thinking about the right things to do,
Lex Fridman (04:03.600)
but somehow the company does something
Greg Brockman (04:05.320)
that is this emergent thing
Lex Fridman (04:07.880)
and that is a really useful abstraction.
Lex Fridman (04:10.600)
And so I think that in some ways,
Lex Fridman (04:12.400)
we think of ourselves as the most intelligent things
Greg Brockman (04:14.840)
on the planet and the most powerful things on the planet,
Lex Fridman (04:17.440)
but there are things that are bigger than us
Greg Brockman (04:19.280)
that are the systems that we all contribute to.
Lex Fridman (04:21.400)
And so I think actually, it's interesting to think about
Lex Fridman (04:24.960)
if you've read Isaac Asimov's foundation, right?
Lex Fridman (04:27.400)
That there's this concept of psychohistory in there,
Greg Brockman (04:30.000)
which is effectively this,
Lex Fridman (04:31.000)
that if you have trillions or quadrillions of beings,
Greg Brockman (04:33.880)
then maybe you could actually predict what that being,
Lex Fridman (04:36.520)
that huge macro being will do
Lex Fridman (04:39.040)
and almost independent of what the individuals want.
Lex Fridman (04:42.320)
And I actually have a second angle on this
Greg Brockman (04:44.200)
that I think is interesting,
Lex Fridman (04:45.040)
which is thinking about technological determinism.
Greg Brockman (04:48.360)
One thing that I actually think a lot about with OpenAI,
Lex Fridman (04:51.240)
right, is that we're kind of coming on
Greg Brockman (04:53.320)
to this insanely transformational technology
Lex Fridman (04:55.840)
of general intelligence, right,
Greg Brockman (04:57.320)
that will happen at some point.
Lex Fridman (04:58.760)
And there's a question of how can you take actions
Greg Brockman (05:01.520)
that will actually steer it to go better rather than worse.
Lex Fridman (05:04.840)
And that I think one question you need to ask
Greg Brockman (05:06.520)
is as a scientist, as an inventor, as a creator,
Lex Fridman (05:09.280)
what impact can you have in general, right?
Greg Brockman (05:11.680)
You look at things like the telephone
Lex Fridman (05:12.840)
invented by two people on the same day.
Lex Fridman (05:14.800)
Like, what does that mean?
Lex Fridman (05:15.920)
Like, what does that mean about the shape of innovation?
Lex Fridman (05:18.080)
And I think that what's going on
Lex Fridman (05:19.240)
is everyone's building on the shoulders of the same giants.
Lex Fridman (05:21.680)
And so you can kind of, you can't really hope
Lex Fridman (05:23.800)
to create something no one else ever would.
Greg Brockman (05:25.680)
You know, if Einstein wasn't born,
Lex Fridman (05:27.000)
someone else would have come up with relativity.
Greg Brockman (05:29.160)
You know, he changed the timeline a bit, right,
Lex Fridman (05:30.960)
that maybe it would have taken another 20 years,
Lex Fridman (05:32.960)
but it wouldn't be that fundamentally humanity
Lex Fridman (05:34.560)
would never discover these fundamental truths.
Lex Fridman (05:37.320)
So there's some kind of invisible momentum
Lex Fridman (05:40.400)
that some people like Einstein or OpenAI is plugging into
Greg Brockman (05:45.360)
that anybody else can also plug into
Lex Fridman (05:47.760)
and ultimately that wave takes us into a certain direction.
Greg Brockman (05:50.760)
That's what he means by digital.
Lex Fridman (05:51.840)
That's right, that's right.
Lex Fridman (05:52.800)
And you know, this kind of seems to play out
Lex Fridman (05:54.160)
in a bunch of different ways,
Greg Brockman (05:55.680)
that there's some exponential that is being written
Lex Fridman (05:58.000)
and that the exponential itself, which one it is, changes.
Greg Brockman (06:00.600)
Think about Moore's Law, an entire industry
Lex Fridman (06:02.400)
set its clock to it for 50 years.
Lex Fridman (06:04.760)
Like, how can that be, right?
Lex Fridman (06:06.160)
How is that possible?
Lex Fridman (06:07.320)
And yet somehow it happened.
Lex Fridman (06:09.240)
And so I think you can't hope to ever invent something
Greg Brockman (06:12.160)
that no one else will.
Lex Fridman (06:13.280)
Maybe you can change the timeline a little bit.
Lex Fridman (06:15.280)
But if you really want to make a difference,
Lex Fridman (06:17.360)
I think that the thing that you really have to do,
Greg Brockman (06:19.360)
the only real degree of freedom you have
Lex Fridman (06:21.280)
is to set the initial conditions
Greg Brockman (06:23.000)
under which a technology is born.
Lex Fridman (06:24.880)
And so you think about the internet, right?
Greg Brockman (06:26.640)
That there are lots of other competitors
Lex Fridman (06:27.800)
trying to build similar things.
Lex Fridman (06:29.360)
And the internet won.
Lex Fridman (06:30.720)
And that the initial conditions
Greg Brockman (06:33.200)
were that it was created by this group
Lex Fridman (06:34.640)
that really valued people being able to be,
Greg Brockman (06:38.200)
anyone being able to plug in
Lex Fridman (06:39.080)
this very academic mindset of being open and connected.
Lex Fridman (06:42.440)
And I think that the internet for the next 40 years
Lex Fridman (06:44.320)
really played out that way.
Greg Brockman (06:46.280)
You know, maybe today things are starting
Lex Fridman (06:48.400)
to shift in a different direction.
Lex Fridman (06:49.840)
But I think that those initial conditions
Lex Fridman (06:51.120)
were really important to determine
Greg Brockman (06:52.720)
the next 40 years worth of progress.
Lex Fridman (06:55.040)
That's really beautifully put.
Lex Fridman (06:56.440)
So another example that I think about,
Lex Fridman (06:58.800)
you know, I recently looked at it.
Greg Brockman (07:00.800)
I looked at Wikipedia, the formation of Wikipedia.
Lex Fridman (07:03.800)
And I wondered what the internet would be like
Greg Brockman (07:05.520)
if Wikipedia had ads.
Lex Fridman (07:07.720)
You know, there's an interesting argument
Greg Brockman (07:09.600)
that why they chose not to make it,
Lex Fridman (07:12.600)
put advertisement on Wikipedia.
Greg Brockman (07:14.240)
I think Wikipedia's one of the greatest resources
Lex Fridman (07:17.760)
we have on the internet.
Greg Brockman (07:18.880)
It's extremely surprising how well it works
Lex Fridman (07:21.200)
and how well it was able to aggregate
Greg Brockman (07:22.920)
all this kind of good information.
Lex Fridman (07:24.960)
And essentially the creator of Wikipedia,
Greg Brockman (07:27.280)
I don't know, there's probably some debates there,
Lex Fridman (07:29.320)
but set the initial conditions.
Lex Fridman (07:31.160)
And now it carried itself forward.
Lex Fridman (07:33.220)
That's really interesting.
Lex Fridman (07:34.060)
So the way you're thinking about AGI
Lex Fridman (07:36.480)
or artificial intelligence is you're focused
Greg Brockman (07:38.440)
on setting the initial conditions for the progress.
Lex Fridman (07:41.160)
That's right.
Greg Brockman (07:42.280)
That's powerful.
Lex Fridman (07:43.120)
Okay, so looking to the future,
Greg Brockman (07:45.520)
if you create an AGI system,
Lex Fridman (07:48.120)
like one that can ace the Turing test, natural language,
Lex Fridman (07:51.560)
what do you think would be the interactions
Lex Fridman (07:54.760)
you would have with it?
Lex Fridman (07:55.840)
What do you think are the questions you would ask?
Lex Fridman (07:57.720)
Like what would be the first question you would ask?
Greg Brockman (08:00.520)
It, her, him.
Lex Fridman (08:01.800)
That's right.
Greg Brockman (08:02.640)
I think that at that point,
Lex Fridman (08:03.920)
if you've really built a powerful system
Greg Brockman (08:05.920)
that is capable of shaping the future of humanity,
Lex Fridman (08:08.480)
the first question that you really should ask
Lex Fridman (08:10.240)
is how do we make sure that this plays out well?
Lex Fridman (08:12.280)
And so that's actually the first question
Greg Brockman (08:13.960)
that I would ask a powerful AGI system is.
Lex Fridman (08:17.600)
So you wouldn't ask your colleague,
Greg Brockman (08:19.160)
you wouldn't ask like Ilya,
Lex Fridman (08:20.760)
you would ask the AGI system.
Lex Fridman (08:22.280)
Oh, we've already had the conversation with Ilya, right?
Lex Fridman (08:24.600)
And everyone here.
Lex Fridman (08:25.720)
And so you want as many perspectives
Lex Fridman (08:27.460)
and a piece of wisdom as you can
Greg Brockman (08:29.680)
for answering this question.
Lex Fridman (08:31.200)
So I don't think you necessarily defer
Greg Brockman (08:32.440)
to whatever your powerful system tells you,
Lex Fridman (08:35.440)
but you use it as one input
Greg Brockman (08:37.080)
to try to figure out what to do.
Lex Fridman (08:39.240)
But, and I guess fundamentally what it really comes down to
Greg Brockman (08:41.800)
is if you built something really powerful
Lex Fridman (08:43.960)
and you think about, for example,
Greg Brockman (08:45.280)
the creation of shortly after
Lex Fridman (08:47.640)
the creation of nuclear weapons, right?
Greg Brockman (08:48.880)
The most important question in the world was
Lex Fridman (08:51.100)
what's the world order going to be like?
Lex Fridman (08:52.800)
How do we set ourselves up in a place
Lex Fridman (08:54.900)
where we're going to be able to survive as a species?
Lex Fridman (08:58.320)
With AGI, I think the question is slightly different, right?
Lex Fridman (09:00.640)
That there is a question of how do we make sure
Greg Brockman (09:02.720)
that we don't get the negative effects,
Lex Fridman (09:04.440)
but there's also the positive side, right?
Lex Fridman (09:06.240)
You imagine that, like what won't AGI be like?
Lex Fridman (09:09.760)
Like what will it be capable of?
Lex Fridman (09:11.240)
And I think that one of the core reasons
Lex Fridman (09:13.520)
that an AGI can be powerful and transformative
Lex Fridman (09:15.760)
is actually due to technological development, right?
Lex Fridman (09:18.900)
If you have something that's capable as a human
Lex Fridman (09:21.440)
and that it's much more scalable,
Lex Fridman (09:23.880)
that you absolutely want that thing
Greg Brockman (09:25.880)
to go read the whole scientific literature
Lex Fridman (09:27.640)
and think about how to create cures for all the diseases,
Lex Fridman (09:29.820)
right?
Lex Fridman (09:30.660)
You want it to think about how to go
Lex Fridman (09:31.480)
and build technologies to help us create material abundance
Lex Fridman (09:34.500)
and to figure out societal problems
Greg Brockman (09:37.320)
that we have trouble with.
Lex Fridman (09:38.160)
Like how are we supposed to clean up the environment?
Lex Fridman (09:40.000)
And maybe you want this to go and invent
Lex Fridman (09:42.840)
a bunch of little robots that will go out
Lex Fridman (09:44.120)
and be biodegradable and turn ocean debris
Lex Fridman (09:47.280)
into harmless molecules.
Lex Fridman (09:49.660)
And I think that that positive side
Lex Fridman (09:54.040)
is something that I think people miss
Greg Brockman (09:55.720)
sometimes when thinking about what an AGI will be like.
Lex Fridman (09:58.160)
And so I think that if you have a system
Greg Brockman (10:00.280)
that's capable of all of that,
Lex Fridman (10:01.640)
you absolutely want its advice about how do I make sure
Greg Brockman (10:03.960)
that we're using your capabilities
Lex Fridman (10:07.600)
in a positive way for humanity.
Lex Fridman (10:09.220)
So what do you think about that psychology
Lex Fridman (10:11.440)
that looks at all the different possible trajectories
Greg Brockman (10:14.800)
of an AGI system, many of which,
Lex Fridman (10:17.520)
perhaps the majority of which are positive,
Lex Fridman (10:19.960)
and nevertheless focuses on the negative trajectories?
Lex Fridman (10:23.340)
I mean, you get to interact with folks,
Greg Brockman (10:24.720)
you get to think about this, maybe within yourself as well.
Lex Fridman (10:28.860)
You look at Sam Harris and so on.
Greg Brockman (10:30.560)
It seems to be, sorry to put it this way,
Lex Fridman (10:32.760)
but almost more fun to think about
Greg Brockman (10:34.560)
the negative possibilities.
Lex Fridman (10:36.780)
Whatever that's deep in our psychology,
Lex Fridman (10:39.560)
what do you think about that?
Lex Fridman (10:40.840)
And how do we deal with it?
Greg Brockman (10:41.920)
Because we want AI to help us.
Lex Fridman (10:44.400)
So I think there's kind of two problems
Greg Brockman (10:47.360)
entailed in that question.
Lex Fridman (10:49.960)
The first is more of the question of
Lex Fridman (10:52.360)
how can you even picture what a world
Lex Fridman (10:54.620)
with a new technology will be like?
Greg Brockman (10:56.600)
Now imagine we're in 1950,
Lex Fridman (10:57.840)
and I'm trying to describe Uber to someone.
Greg Brockman (11:02.840)
Apps and the internet.
Lex Fridman (11:05.340)
Yeah, I mean, that's going to be extremely complicated.
Lex Fridman (11:08.920)
But it's imaginable.
Lex Fridman (11:10.160)
It's imaginable, right?
Lex Fridman (11:11.880)
And now imagine being in 1950 and predicting Uber, right?
Lex Fridman (11:15.280)
And you need to describe the internet,
Greg Brockman (11:17.680)
you need to describe GPS,
Lex Fridman (11:18.720)
you need to describe the fact that
Greg Brockman (11:20.480)
everyone's going to have this phone in their pocket.
Lex Fridman (11:23.920)
And so I think that just the first truth
Greg Brockman (11:26.160)
is that it is hard to picture
Lex Fridman (11:28.040)
how a transformative technology will play out in the world.
Greg Brockman (11:31.160)
We've seen that before with technologies
Lex Fridman (11:32.760)
that are far less transformative than AGI will be.
Lex Fridman (11:35.560)
And so I think that one piece is that
Lex Fridman (11:37.780)
it's just even hard to imagine
Lex Fridman (11:39.560)
and to really put yourself in a world
Lex Fridman (11:41.640)
where you can predict what that positive vision
Greg Brockman (11:44.640)
would be like.
Lex Fridman (11:46.920)
And I think the second thing is that
Greg Brockman (11:49.520)
I think it is always easier to support the negative side
Lex Fridman (11:54.280)
than the positive side.
Greg Brockman (11:55.120)
It's always easier to destroy than create.
Lex Fridman (11:58.160)
And less in a physical sense
Lex Fridman (12:00.760)
and more just in an intellectual sense, right?
Lex Fridman (12:03.080)
Because I think that with creating something,
Greg Brockman (12:05.680)
you need to just get a bunch of things right.
Lex Fridman (12:07.400)
And to destroy, you just need to get one thing wrong.
Lex Fridman (12:10.280)
And so I think that what that means
Lex Fridman (12:12.040)
is that I think a lot of people's thinking dead ends
Greg Brockman (12:14.240)
as soon as they see the negative story.
Lex Fridman (12:16.920)
But that being said, I actually have some hope, right?
Greg Brockman (12:20.360)
I think that the positive vision
Lex Fridman (12:23.160)
is something that I think can be,
Greg Brockman (12:26.000)
is something that we can talk about.
Lex Fridman (12:27.560)
And I think that just simply saying this fact of,
Greg Brockman (12:30.240)
yeah, there's positive, there's negatives,
Lex Fridman (12:32.000)
everyone likes to dwell on the negative.
Greg Brockman (12:33.600)
People actually respond well to that message and say,
Lex Fridman (12:35.400)
huh, you're right, there's a part of this
Greg Brockman (12:37.040)
that we're not talking about, not thinking about.
Lex Fridman (12:39.640)
And that's actually something that's I think really
Greg Brockman (12:42.280)
been a key part of how we think about AGI at OpenAI.
Lex Fridman (12:46.640)
You can kind of look at it as like, okay,
Greg Brockman (12:48.640)
OpenAI talks about the fact that there are risks
Lex Fridman (12:51.040)
and yet they're trying to build this system.
Lex Fridman (12:53.720)
How do you square those two facts?
Lex Fridman (12:56.120)
So do you share the intuition that some people have,
Greg Brockman (12:59.160)
I mean from Sam Harris to even Elon Musk himself,
Lex Fridman (13:02.720)
that it's tricky as you develop AGI
Greg Brockman (13:06.640)
to keep it from slipping into the existential threats,
Lex Fridman (13:10.440)
into the negative?
Greg Brockman (13:11.800)
What's your intuition about how hard is it
Lex Fridman (13:14.840)
to keep AI development on the positive track?
Lex Fridman (13:19.680)
What's your intuition there?
Lex Fridman (13:20.760)
To answer that question, you can really look
Greg Brockman (13:22.280)
at how we structure OpenAI.
Lex Fridman (13:24.000)
So we really have three main arms.
Greg Brockman (13:25.880)
We have capabilities, which is actually doing
Lex Fridman (13:28.000)
the technical work and pushing forward
Lex Fridman (13:29.880)
what these systems can do.
Lex Fridman (13:31.200)
There's safety, which is working on technical mechanisms
Greg Brockman (13:35.160)
to ensure that the systems we build
Lex Fridman (13:36.920)
are aligned with human values.
Lex Fridman (13:38.480)
And then there's policy, which is making sure
Lex Fridman (13:40.680)
that we have governance mechanisms,
Lex Fridman (13:42.040)
answering that question of, well, whose values?
Lex Fridman (13:45.280)
And so I think that the technical safety one
Lex Fridman (13:47.400)
is the one that people kind of talk about the most, right?
Lex Fridman (13:50.480)
You talk about, like think about all of the dystopic AI
Greg Brockman (13:53.840)
movies, a lot of that is about not having
Lex Fridman (13:55.800)
good technical safety in place.
Lex Fridman (13:57.560)
And what we've been finding is that,
Lex Fridman (13:59.840)
you know, I think that actually a lot of people
Greg Brockman (14:01.360)
look at the technical safety problem
Lex Fridman (14:02.680)
and think it's just intractable, right?
Lex Fridman (14:05.400)
This question of what do humans want?
Lex Fridman (14:07.840)
How am I supposed to write that down?
Lex Fridman (14:09.160)
Can I even write down what I want?
Lex Fridman (14:11.200)
No way.
Lex Fridman (14:13.040)
And then they stop there.
Lex Fridman (14:14.840)
But the thing is, we've already built systems
Greg Brockman (14:16.880)
that are able to learn things that humans can't specify.
Lex Fridman (14:20.920)
You know, even the rules for how to recognize
Greg Brockman (14:22.920)
if there's a cat or a dog in an image.
Lex Fridman (14:24.960)
Turns out it's intractable to write that down,
Lex Fridman (14:26.520)
and yet we're able to learn it.
Lex Fridman (14:28.440)
And that what we're seeing with systems we build at OpenAI,
Lex Fridman (14:31.040)
and they're still in early proof of concept stage,
Lex Fridman (14:33.800)
is that you are able to learn human preferences.
Greg Brockman (14:36.320)
You're able to learn what humans want from data.
Lex Fridman (14:38.960)
And so that's kind of the core focus
Greg Brockman (14:40.400)
for our technical safety team,
Lex Fridman (14:41.760)
and I think that there actually,
Greg Brockman (14:43.800)
we've had some pretty encouraging updates
Lex Fridman (14:45.680)
in terms of what we've been able to make work.
Lex Fridman (14:48.040)
So you have an intuition and a hope that from data,
Lex Fridman (14:51.680)
you know, looking at the value alignment problem,
Greg Brockman (14:53.640)
from data we can build systems that align
Lex Fridman (14:57.080)
with the collective better angels of our nature.
Lex Fridman (15:00.640)
So align with the ethics and the morals of human beings.
Lex Fridman (15:04.640)
To even say this in a different way,
Lex Fridman (15:05.920)
I mean, think about how do we align humans, right?
Lex Fridman (15:08.600)
Think about like a human baby can grow up
Greg Brockman (15:10.440)
to be an evil person or a great person.
Lex Fridman (15:12.920)
And a lot of that is from learning from data, right?
Greg Brockman (15:15.240)
That you have some feedback as a child is growing up,
Lex Fridman (15:17.760)
they get to see positive examples.
Lex Fridman (15:19.200)
And so I think that just like,
Lex Fridman (15:22.000)
that the only example we have of a general intelligence
Greg Brockman (15:25.400)
that is able to learn from data
Lex Fridman (15:28.040)
to align with human values and to learn values,
Greg Brockman (15:31.440)
I think we shouldn't be surprised
Lex Fridman (15:32.880)
that we can do the same sorts of techniques
Greg Brockman (15:36.000)
or whether the same sort of techniques
Lex Fridman (15:37.440)
end up being how we solve value alignment for AGI's.
Lex Fridman (15:41.080)
So let's go even higher.
Lex Fridman (15:42.720)
I don't know if you've read the book, Sapiens,
Lex Fridman (15:44.800)
but there's an idea that, you know,
Lex Fridman (15:48.280)
that as a collective, as us human beings,
Greg Brockman (15:49.960)
we kind of develop together ideas that we hold.
Lex Fridman (15:54.720)
There's no, in that context, objective truth.
Greg Brockman (15:57.880)
We just kind of all agree to certain ideas
Lex Fridman (15:59.960)
and hold them as a collective.
Greg Brockman (16:01.400)
Did you have a sense that there is,
Lex Fridman (16:03.440)
in the world of good and evil,
Lex Fridman (16:05.320)
do you have a sense that to the first approximation,
Lex Fridman (16:07.520)
there are some things that are good
Lex Fridman (16:10.240)
and that you could teach systems to behave to be good?
Lex Fridman (16:14.520)
So I think that this actually blends into our third team,
Greg Brockman (16:18.280)
right, which is the policy team.
Lex Fridman (16:19.880)
And this is the one, the aspect I think people
Lex Fridman (16:22.360)
really talk about way less than they should, right?
Lex Fridman (16:25.280)
Because imagine that we build super powerful systems
Greg Brockman (16:27.640)
that we've managed to figure out all the mechanisms
Lex Fridman (16:29.720)
for these things to do whatever the operator wants.
Greg Brockman (16:32.800)
The most important question becomes,
Lex Fridman (16:34.480)
who's the operator, what do they want,
Lex Fridman (16:36.720)
and how is that going to affect everyone else, right?
Lex Fridman (16:39.360)
And I think that this question of what is good,
Lex Fridman (16:43.080)
what are those values, I mean,
Lex Fridman (16:44.720)
I think you don't even have to go to those,
Greg Brockman (16:46.600)
those very grand existential places
Lex Fridman (16:48.400)
to start to realize how hard this problem is.
Greg Brockman (16:50.920)
You just look at different countries
Lex Fridman (16:52.880)
and cultures across the world,
Lex Fridman (16:54.520)
and that there's a very different conception
Lex Fridman (16:57.120)
of how the world works and what kinds of ways
Greg Brockman (17:01.920)
that society wants to operate.
Lex Fridman (17:03.400)
And so I think that the really core question
Greg Brockman (17:07.000)
is actually very concrete,
Lex Fridman (17:09.560)
and I think it's not a question
Lex Fridman (17:10.980)
that we have ready answers to, right?
Lex Fridman (17:12.720)
It's how do you have a world
Greg Brockman (17:14.720)
where all of the different countries that we have,
Lex Fridman (17:17.280)
United States, China, Russia,
Lex Fridman (17:19.760)
and the hundreds of other countries out there
Lex Fridman (17:22.760)
are able to continue to not just operate
Greg Brockman (17:26.620)
in the way that they see fit,
Lex Fridman (17:28.440)
but in the world that emerges
Greg Brockman (17:32.560)
where you have these very powerful systems
Lex Fridman (17:36.080)
operating alongside humans,
Greg Brockman (17:37.820)
ends up being something that empowers humans more,
Lex Fridman (17:39.820)
that makes human existence be a more meaningful thing,
Lex Fridman (17:44.140)
and that people are happier and wealthier,
Lex Fridman (17:46.440)
and able to live more fulfilling lives.
Greg Brockman (17:49.040)
It's not an obvious thing for how to design that world
Lex Fridman (17:51.600)
once you have that very powerful system.
Lex Fridman (17:53.640)
So if we take a little step back,
Lex Fridman (17:55.860)
and we're having a fascinating conversation,
Lex Fridman (17:58.260)
and OpenAI is in many ways a tech leader in the world,
Lex Fridman (18:01.920)
and yet we're thinking about
Greg Brockman (18:03.240)
these big existential questions,
Lex Fridman (18:05.480)
which is fascinating, really important.
Greg Brockman (18:07.060)
I think you're a leader in that space,
Lex Fridman (18:09.200)
and that's a really important space
Greg Brockman (18:10.880)
of just thinking how AI affects society
Lex Fridman (18:13.120)
in a big picture view.
Lex Fridman (18:14.400)
So Oscar Wilde said, we're all in the gutter,
Lex Fridman (18:17.360)
but some of us are looking at the stars,
Lex Fridman (18:19.040)
and I think OpenAI has a charter
Lex Fridman (18:22.360)
that looks to the stars, I would say,
Greg Brockman (18:24.640)
to create intelligence, to create general intelligence,
Lex Fridman (18:26.920)
make it beneficial, safe, and collaborative.
Lex Fridman (18:29.480)
So can you tell me how that came about,
Lex Fridman (18:33.720)
how a mission like that and the path
Lex Fridman (18:36.360)
to creating a mission like that at OpenAI was founded?
Lex Fridman (18:39.160)
Yeah, so I think that in some ways
Greg Brockman (18:41.680)
it really boils down to taking a look at the landscape.
Lex Fridman (18:45.160)
So if you think about the history of AI,
Greg Brockman (18:47.060)
that basically for the past 60 or 70 years,
Lex Fridman (18:49.960)
people have thought about this goal
Greg Brockman (18:51.680)
of what could happen if you could automate
Lex Fridman (18:54.000)
human intellectual labor.
Greg Brockman (18:56.700)
Imagine you could build a computer system
Lex Fridman (18:58.280)
that could do that, what becomes possible?
Greg Brockman (19:00.560)
We have a lot of sci fi that tells stories
Lex Fridman (19:02.440)
of various dystopias, and increasingly you have movies
Greg Brockman (19:04.960)
like Her that tell you a little bit about,
Lex Fridman (19:06.520)
maybe more of a little bit utopic vision.
Greg Brockman (19:09.480)
You think about the impacts that we've seen
Lex Fridman (19:12.560)
from being able to have bicycles for our minds
Lex Fridman (19:16.280)
and computers, and I think that the impact
Lex Fridman (19:20.360)
of computers and the internet has just far outstripped
Lex Fridman (19:23.480)
what anyone really could have predicted.
Lex Fridman (19:26.200)
And so I think that it's very clear
Greg Brockman (19:27.420)
that if you can build an AGI,
Lex Fridman (19:29.360)
it will be the most transformative technology
Greg Brockman (19:31.600)
that humans will ever create.
Lex Fridman (19:34.640)
And so what it boils down to then is a question of,
Greg Brockman (19:36.840)
well, is there a path, is there hope,
Lex Fridman (19:39.400)
is there a way to build such a system?
Lex Fridman (19:41.480)
And I think that for 60 or 70 years,
Lex Fridman (19:43.620)
that people got excited and that ended up
Greg Brockman (19:47.280)
not being able to deliver on the hopes
Lex Fridman (19:49.440)
that people had pinned on them.
Lex Fridman (19:51.400)
And I think that then, that after two winters
Lex Fridman (19:54.880)
of AI development, that people I think kind of
Lex Fridman (19:58.320)
almost stopped daring to dream, right?
Lex Fridman (1:00:01.580)
and it's like evolutionary history,
Greg Brockman (1:00:03.340)
baking in all this information,
Lex Fridman (1:00:04.380)
getting very, very good at this predictive process.
Lex Fridman (1:00:06.700)
And then at runtime, I just kind of do one forward pass,
Lex Fridman (1:00:10.020)
and I'm able to generate stuff.
Lex Fridman (1:00:12.940)
And so, you know, there might be small tweaks
Lex Fridman (1:00:15.260)
to what we do in order to get the type signature, right?
Greg Brockman (1:00:17.700)
For example, well, you know,
Lex Fridman (1:00:19.140)
it's not really one forward pass, right?
Greg Brockman (1:00:20.700)
You know, you generate symbol by symbol,
Lex Fridman (1:00:22.300)
and so maybe you generate like a whole sequence
Greg Brockman (1:00:24.340)
of thoughts, and you only keep like the last bit
Lex Fridman (1:00:26.540)
or something.
Lex Fridman (1:00:27.860)
But I think that at the very least,
Lex Fridman (1:00:29.500)
I would expect you have to make changes like that.
Greg Brockman (1:00:31.820)
Yeah, just exactly how we, you said, think,
Lex Fridman (1:00:35.220)
is the process of generating thought by thought
Greg Brockman (1:00:38.060)
in the same kind of way, like you said,
Lex Fridman (1:00:40.060)
keep the last bit, the thing that we converge towards.
Greg Brockman (1:00:43.220)
Yep.
Lex Fridman (1:00:44.700)
And I think there's another piece which is interesting,
Lex Fridman (1:00:46.980)
which is this out of distribution generalization, right?
Lex Fridman (1:00:49.940)
That like thinking somehow lets us do that, right?
Greg Brockman (1:00:52.300)
That we haven't experienced a thing, and yet somehow
Lex Fridman (1:00:54.780)
we just kind of keep refining our mental model of it.
Greg Brockman (1:00:57.780)
This is, again, something that feels tied
Lex Fridman (1:01:00.340)
to whatever reasoning is, and maybe it's a small tweak
Greg Brockman (1:01:04.620)
to what we do, maybe it's many ideas,
Lex Fridman (1:01:06.380)
and we'll take as many decades.
Greg Brockman (1:01:07.820)
Yeah, so the assumption there,
Lex Fridman (1:01:10.940)
generalization out of distribution,
Greg Brockman (1:01:12.980)
is that it's possible to create new ideas.
Lex Fridman (1:01:16.620)
Mm hmm.
Greg Brockman (1:01:17.460)
You know, it's possible that nobody's ever created
Lex Fridman (1:01:19.780)
any new ideas, and then with scaling GPT2 to GPT20,
Greg Brockman (1:01:25.340)
you would essentially generalize to all possible thoughts
Lex Fridman (1:01:30.340)
that us humans could have.
Greg Brockman (1:01:31.780)
I mean.
Lex Fridman (1:01:33.180)
Just to play devil's advocate.
Greg Brockman (1:01:34.180)
Right, right, right, I mean, how many new story ideas
Lex Fridman (1:01:37.260)
have we come up with since Shakespeare, right?
Greg Brockman (1:01:39.060)
Yeah, exactly.
Lex Fridman (1:01:40.100)
It's just all different forms of love and drama and so on.
Greg Brockman (1:01:44.620)
Okay.
Lex Fridman (1:01:45.740)
Not sure if you read Bitter Lesson,
Greg Brockman (1:01:47.460)
a recent blog post by Rich Sutton.
Lex Fridman (1:01:49.340)
Yep, I have.
Greg Brockman (1:01:50.820)
He basically says something that echoes some of the ideas
Lex Fridman (1:01:54.380)
that you've been talking about, which is,
Greg Brockman (1:01:56.780)
he says the biggest lesson that can be read
Lex Fridman (1:01:58.980)
from 70 years of AI research is that general methods
Greg Brockman (1:02:01.980)
that leverage computation are ultimately going to,
Lex Fridman (1:02:05.900)
ultimately win out.
Lex Fridman (1:02:07.820)
Do you agree with this?
Lex Fridman (1:02:08.860)
So basically, and OpenAI in general,
Lex Fridman (1:02:12.780)
but the ideas you're exploring about coming up with methods,
Lex Fridman (1:02:15.780)
whether it's GPT2 modeling or whether it's OpenAI 5
Greg Brockman (1:02:20.060)
playing Dota, or a general method is better
Lex Fridman (1:02:23.940)
than a more fine tuned, expert tuned method.
Greg Brockman (1:02:29.700)
Yeah, so I think that, well one thing that I think
Lex Fridman (1:02:32.140)
was really interesting about the reaction
Greg Brockman (1:02:33.740)
to that blog post was that a lot of people have read this
Lex Fridman (1:02:36.380)
as saying that compute is all that matters.
Lex Fridman (1:02:39.380)
And that's a very threatening idea, right?
Lex Fridman (1:02:41.300)
And I don't think it's a true idea either.
Greg Brockman (1:02:43.500)
Right, it's very clear that we have algorithmic ideas
Lex Fridman (1:02:45.740)
that have been very important for making progress
Lex Fridman (1:02:47.820)
and to really build AGI.
Lex Fridman (1:02:49.460)
You wanna push as far as you can on the computational scale
Lex Fridman (1:02:52.060)
and you wanna push as far as you can on human ingenuity.
Lex Fridman (1:02:55.500)
And so I think you need both.
Lex Fridman (1:02:56.980)
But I think the way that you phrased the question
Lex Fridman (1:02:58.260)
is actually very good, right?
Greg Brockman (1:02:59.580)
That it's really about what kind of ideas
Lex Fridman (1:03:02.140)
should we be striving for?
Lex Fridman (1:03:03.940)
And absolutely, if you can find a scalable idea,
Lex Fridman (1:03:07.540)
you pour more compute into it, you pour more data into it,
Greg Brockman (1:03:09.780)
it gets better, like that's the real holy grail.
Lex Fridman (1:03:13.740)
And so I think that the answer to the question,
Greg Brockman (1:03:16.580)
I think, is yes, that that's really how we think about it
Lex Fridman (1:03:19.900)
and that part of why we're excited about the power
Greg Brockman (1:03:22.700)
of deep learning, the potential for building AGI
Lex Fridman (1:03:25.260)
is because we look at the systems that exist
Greg Brockman (1:03:27.540)
in the most successful AI systems
Lex Fridman (1:03:29.700)
and we realize that you scale those up,
Greg Brockman (1:03:32.620)
they're gonna work better.
Lex Fridman (1:03:33.940)
And I think that that scalability
Greg Brockman (1:03:35.780)
is something that really gives us hope
Lex Fridman (1:03:37.020)
for being able to build transformative systems.
Lex Fridman (1:03:39.540)
So I'll tell you, this is partially an emotional,
Lex Fridman (1:03:43.780)
a response that people often have,
Greg Brockman (1:03:45.660)
if compute is so important for state of the art performance,
Lex Fridman (1:03:49.700)
individual developers, maybe a 13 year old
Greg Brockman (1:03:51.780)
sitting somewhere in Kansas or something like that,
Lex Fridman (1:03:54.420)
they're sitting, they might not even have a GPU
Greg Brockman (1:03:56.940)
or may have a single GPU, a 1080 or something like that,
Lex Fridman (1:03:59.980)
and there's this feeling like, well,
Lex Fridman (1:04:02.580)
how can I possibly compete or contribute
Lex Fridman (1:04:05.700)
to this world of AI if scale is so important?
Lex Fridman (1:04:09.780)
So if you can comment on that and in general,
Lex Fridman (1:04:12.460)
do you think we need to also in the future
Greg Brockman (1:04:14.780)
focus on democratizing compute resources more
Lex Fridman (1:04:19.980)
or as much as we democratize the algorithms?
Greg Brockman (1:04:22.620)
Well, so the way that I think about it
Lex Fridman (1:04:23.900)
is that there's this space of possible progress, right?
Greg Brockman (1:04:28.820)
There's a space of ideas and sort of systems
Lex Fridman (1:04:30.860)
that will work that will move us forward
Lex Fridman (1:04:32.900)
and there's a portion of that space
Lex Fridman (1:04:34.780)
and to some extent, an increasingly significant portion
Greg Brockman (1:04:37.020)
of that space that does just require
Lex Fridman (1:04:38.780)
massive compute resources.
Lex Fridman (1:04:40.980)
And for that, I think that the answer is kind of clear
Lex Fridman (1:04:44.660)
and that part of why we have the structure that we do
Greg Brockman (1:04:47.860)
is because we think it's really important
Lex Fridman (1:04:49.580)
to be pushing the scale and to be building
Greg Brockman (1:04:51.660)
these large clusters and systems.
Lex Fridman (1:04:53.740)
But there's another portion of the space
Greg Brockman (1:04:55.820)
that isn't about the large scale compute
Lex Fridman (1:04:57.780)
that are these ideas that, and again,
Greg Brockman (1:04:59.900)
I think that for the ideas to really be impactful
Lex Fridman (1:05:02.140)
and really shine, that they should be ideas
Greg Brockman (1:05:04.140)
that if you scale them up, would work way better
Lex Fridman (1:05:06.580)
than they do at small scale.
Lex Fridman (1:05:08.740)
But that you can discover them
Lex Fridman (1:05:10.420)
without massive computational resources.
Lex Fridman (1:05:12.700)
And if you look at the history of recent developments,
Lex Fridman (1:05:15.140)
you think about things like the GAN or the VAE,
Greg Brockman (1:05:17.620)
that these are ones that I think you could come up with them
Lex Fridman (1:05:20.860)
without having, and in practice,
Greg Brockman (1:05:22.660)
people did come up with them without having
Lex Fridman (1:05:24.460)
massive, massive computational resources.
Greg Brockman (1:05:26.500)
Right, I just talked to Ian Goodfellow,
Lex Fridman (1:05:27.900)
but the thing is the initial GAN
Lex Fridman (1:05:31.500)
produced pretty terrible results, right?
Lex Fridman (1:05:34.140)
So only because it was in a very specific,
Greg Brockman (1:05:36.220)
it was only because they're smart enough
Lex Fridman (1:05:38.220)
to know that this is quite surprising
Greg Brockman (1:05:39.940)
it can generate anything that they know.
Lex Fridman (1:05:43.100)
Do you see a world, or is that too optimistic and dreamer
Greg Brockman (1:05:45.980)
like to imagine that the compute resources
Lex Fridman (1:05:49.700)
are something that's owned by governments
Lex Fridman (1:05:52.180)
and provided as utility?
Lex Fridman (1:05:55.020)
Actually, to some extent, this question reminds me
Greg Brockman (1:05:57.100)
of a blog post from one of my former professors at Harvard,
Lex Fridman (1:06:01.140)
this guy Matt Welsh, who was a systems professor.
Greg Brockman (1:06:03.740)
I remember sitting in his tenure talk, right,
Lex Fridman (1:06:05.300)
and that he had literally just gotten tenure.
Greg Brockman (1:06:08.780)
He went to Google for the summer
Lex Fridman (1:06:10.940)
and then decided he wasn't going back to academia, right?
Lex Fridman (1:06:15.660)
And kind of in his blog post, he makes this point that,
Lex Fridman (1:06:18.340)
look, as a systems researcher,
Greg Brockman (1:06:20.780)
that I come up with these cool system ideas, right,
Lex Fridman (1:06:23.180)
and I kind of build a little proof of concept,
Lex Fridman (1:06:25.060)
and the best thing I can hope for
Lex Fridman (1:06:27.060)
is that the people at Google or Yahoo,
Greg Brockman (1:06:30.100)
which was around at the time,
Lex Fridman (1:06:31.580)
will implement it and actually make it work at scale, right?
Lex Fridman (1:06:35.380)
That's like the dream for me, right?
Lex Fridman (1:06:36.580)
I build the little thing,
Lex Fridman (1:06:37.420)
and they turn it into the big thing that's actually working.
Lex Fridman (1:06:39.980)
And for him, he said, I'm done with that.
Greg Brockman (1:06:43.340)
I want to be the person who's actually doing building
Lex Fridman (1:06:45.740)
and deploying.
Lex Fridman (1:06:47.300)
And I think that there's a similar dichotomy here, right?
Lex Fridman (1:06:49.540)
I think that there are people who really actually find value,
Lex Fridman (1:06:53.340)
and I think it is a valuable thing to do
Lex Fridman (1:06:55.180)
to be the person who produces those ideas, right,
Greg Brockman (1:06:57.420)
who builds the proof of concept.
Lex Fridman (1:06:58.820)
And yeah, you don't get to generate
Greg Brockman (1:07:00.540)
the coolest possible GAN images,
Lex Fridman (1:07:02.740)
but you invented the GAN, right?
Lex Fridman (1:07:04.460)
And so there's a real trade off there,
Lex Fridman (1:07:07.540)
and I think that that's a very personal choice,
Lex Fridman (1:07:09.020)
but I think there's value in both sides.
Lex Fridman (1:07:10.820)
So do you think creating AGI or some new models,
Greg Brockman (1:07:18.260)
we would see echoes of the brilliance
Lex Fridman (1:07:20.460)
even at the prototype level?
Lex Fridman (1:07:22.260)
So you would be able to develop those ideas without scale,
Lex Fridman (1:07:24.900)
the initial seeds.
Lex Fridman (1:07:27.300)
So take a look at, you know,
Lex Fridman (1:07:28.980)
I always like to look at examples that exist, right?
Greg Brockman (1:07:31.740)
Look at real precedent.
Lex Fridman (1:07:32.700)
And so take a look at the June 2018 model that we released,
Greg Brockman (1:07:37.020)
that we scaled up to turn into GPT2.
Lex Fridman (1:07:39.180)
And you can see that at small scale,
Lex Fridman (1:07:41.260)
it set some records, right?
Lex Fridman (1:07:42.780)
This was the original GPT.
Greg Brockman (1:07:44.820)
We actually had some cool generations.
Lex Fridman (1:07:46.820)
They weren't nearly as amazing and really stunning
Greg Brockman (1:07:49.820)
as the GPT2 ones, but it was promising.
Lex Fridman (1:07:51.980)
It was interesting.
Lex Fridman (1:07:53.020)
And so I think it is the case
Lex Fridman (1:07:54.500)
that with a lot of these ideas,
Greg Brockman (1:07:56.100)
that you see promise at small scale.
Lex Fridman (1:07:58.260)
But there is an asterisk here, a very big asterisk,
Greg Brockman (1:08:00.820)
which is sometimes we see behaviors that emerge
Lex Fridman (1:08:05.220)
that are qualitatively different
Greg Brockman (1:08:07.260)
from anything we saw at small scale.
Lex Fridman (1:08:09.060)
And that the original inventor of whatever algorithm
Greg Brockman (1:08:12.580)
looks at and says, I didn't think it could do that.
Lex Fridman (1:08:15.500)
This is what we saw in Dota, right?
Lex Fridman (1:08:17.420)
So PPO was created by John Shulman,
Lex Fridman (1:08:19.340)
who's a researcher here.
Lex Fridman (1:08:20.540)
And with Dota, we basically just ran PPO
Lex Fridman (1:08:24.660)
at massive, massive scale.
Lex Fridman (1:08:26.540)
And there's some tweaks in order to make it work,
Lex Fridman (1:08:29.100)
but fundamentally, it's PPO at the core.
Lex Fridman (1:08:31.540)
And we were able to get this long term planning,
Lex Fridman (1:08:35.300)
these behaviors to really play out on a time scale
Greg Brockman (1:08:38.700)
that we just thought was not possible.
Lex Fridman (1:08:40.780)
And John looked at that and was like,
Greg Brockman (1:08:42.700)
I didn't think it could do that.
Lex Fridman (1:08:44.220)
That's what happens when you're at three orders
Greg Brockman (1:08:45.460)
of magnitude more scale than you tested at.
Lex Fridman (1:08:48.380)
Yeah, but it still has the same flavors of,
Greg Brockman (1:08:50.580)
you know, at least echoes of the expected billions.
Lex Fridman (1:08:55.980)
Although I suspect with GPT scaled more and more,
Greg Brockman (1:08:59.020)
you might get surprising things.
Lex Fridman (1:09:01.780)
So yeah, you're right, it's interesting.
Greg Brockman (1:09:04.740)
It's difficult to see how far an idea will go
Lex Fridman (1:09:07.940)
when it's scaled.
Greg Brockman (1:09:09.300)
It's an open question.
Lex Fridman (1:09:11.020)
Well, so to that point with Dota and PPO,
Lex Fridman (1:09:13.060)
like, I mean, here's a very concrete one, right?
Lex Fridman (1:09:14.980)
It's like, it's actually one thing
Greg Brockman (1:09:16.620)
that's very surprising about Dota
Lex Fridman (1:09:17.700)
that I think people don't really pay that much attention to
Greg Brockman (1:09:20.340)
is the decree of generalization
Lex Fridman (1:09:22.380)
out of distribution that happens, right?
Greg Brockman (1:09:24.580)
That you have this AI that's trained against other bots
Lex Fridman (1:09:27.860)
for its entirety, the entirety of its existence.
Greg Brockman (1:09:30.340)
Sorry to take a step back.
Lex Fridman (1:09:31.460)
Can you talk through, you know, a story of Dota,
Greg Brockman (1:09:37.260)
a story of leading up to opening I5 and that past,
Lex Fridman (1:09:42.060)
and what was the process of self play
Lex Fridman (1:09:43.900)
and so on of training on this?
Lex Fridman (1:09:45.420)
Yeah, yeah, yeah.
Lex Fridman (1:09:46.260)
So with Dota.
Lex Fridman (1:09:47.100)
What is Dota?
Greg Brockman (1:09:47.940)
Yeah, Dota is a complex video game
Lex Fridman (1:09:50.020)
and we started trying to solve Dota
Greg Brockman (1:09:52.700)
because we felt like this was a step towards the real world
Lex Fridman (1:09:55.660)
relative to other games like chess or Go, right?
Greg Brockman (1:09:58.020)
Those very cerebral games
Lex Fridman (1:09:59.180)
where you just kind of have this board,
Greg Brockman (1:10:00.500)
very discreet moves.
Lex Fridman (1:10:01.900)
Dota starts to be much more continuous time
Greg Brockman (1:10:04.060)
that you have this huge variety of different actions
Lex Fridman (1:10:06.220)
that you have a 45 minute game
Greg Brockman (1:10:07.660)
with all these different units
Lex Fridman (1:10:09.380)
and it's got a lot of messiness to it
Greg Brockman (1:10:11.820)
that really hasn't been captured by previous games.
Lex Fridman (1:10:14.500)
And famously, all of the hard coded bots for Dota
Lex Fridman (1:10:17.340)
were terrible, right?
Lex Fridman (1:10:18.380)
It's just impossible to write anything good for it
Greg Brockman (1:10:19.940)
because it's so complex.
Lex Fridman (1:10:21.260)
And so this seemed like a really good place
Greg Brockman (1:10:23.300)
to push what's the state of the art
Lex Fridman (1:10:25.260)
in reinforcement learning.
Lex Fridman (1:10:26.860)
And so we started by focusing
Lex Fridman (1:10:28.380)
on the one versus one version of the game
Lex Fridman (1:10:29.980)
and we're able to solve that.
Lex Fridman (1:10:32.380)
We're able to beat the world champions
Lex Fridman (1:10:33.900)
and the skill curve was this crazy exponential, right?
Lex Fridman (1:10:38.980)
And it was like constantly we were just scaling up
Greg Brockman (1:10:41.020)
that we were fixing bugs
Lex Fridman (1:10:42.260)
and that you look at the skill curve
Lex Fridman (1:10:44.340)
and it was really a very, very smooth one.
Lex Fridman (1:10:46.660)
This is actually really interesting
Greg Brockman (1:10:47.500)
to see how that human iteration loop
Lex Fridman (1:10:50.020)
yielded very steady exponential progress.
Lex Fridman (1:10:52.740)
And to one side note, first of all,
Lex Fridman (1:10:55.220)
it's an exceptionally popular video game.
Greg Brockman (1:10:57.140)
The side effect is that there's a lot of incredible
Lex Fridman (1:11:00.300)
human experts at that video game.
Lex Fridman (1:11:01.960)
So the benchmark that you're trying to reach is very high.
Lex Fridman (1:11:05.260)
And the other, can you talk about the approach
Greg Brockman (1:11:07.900)
that was used initially and throughout
Lex Fridman (1:11:10.140)
training these agents to play this game?
Greg Brockman (1:11:12.100)
Yep, and so the approach that we used is self play.
Lex Fridman (1:11:14.420)
And so you have two agents that don't know anything.
Greg Brockman (1:11:17.380)
They battle each other,
Lex Fridman (1:11:18.700)
they discover something a little bit good
Lex Fridman (1:11:20.820)
and now they both know it.
Lex Fridman (1:11:22.060)
And they just get better and better and better
Greg Brockman (1:11:23.400)
without bound.
Lex Fridman (1:11:24.540)
And that's a really powerful idea, right?
Greg Brockman (1:11:27.100)
That we then went from the one versus one version
Lex Fridman (1:11:30.180)
of the game and scaled up to five versus five, right?
Lex Fridman (1:11:32.460)
So you think about kind of like with basketball
Lex Fridman (1:11:34.340)
where you have this like team sport
Lex Fridman (1:11:35.500)
and you need to do all this coordination
Lex Fridman (1:11:37.700)
and we were able to push the same idea,
Greg Brockman (1:11:40.940)
the same self play to really get to the professional level
Lex Fridman (1:11:45.940)
at the full five versus five version of the game.
Lex Fridman (1:11:48.980)
And the things I think are really interesting here
Lex Fridman (1:11:52.460)
is that these agents, in some ways,
Lex Fridman (1:11:54.820)
they're almost like an insect like intelligence, right?
Lex Fridman (1:11:56.820)
Where they have a lot in common
Lex Fridman (1:11:58.720)
with how an insect is trained, right?
Lex Fridman (1:12:00.180)
An insect kind of lives in this environment
Greg Brockman (1:12:01.840)
for a very long time or the ancestors of this insect
Lex Fridman (1:12:04.980)
have been around for a long time
Lex Fridman (1:12:05.900)
and had a lot of experience that gets baked into this agent.
Lex Fridman (1:12:09.740)
And it's not really smart in the sense of a human, right?
Greg Brockman (1:12:12.780)
It's not able to go and learn calculus,
Lex Fridman (1:12:14.620)
but it's able to navigate its environment extremely well.
Lex Fridman (1:12:16.980)
And it's able to handle unexpected things
Lex Fridman (1:12:18.460)
in the environment that it's never seen before pretty well.
Lex Fridman (1:12:22.060)
And we see the same sort of thing with our Dota bots, right?
Lex Fridman (1:12:24.780)
That they're able to, within this game,
Greg Brockman (1:12:26.740)
they're able to play against humans,
Lex Fridman (1:12:28.460)
which is something that never existed
Greg Brockman (1:12:29.980)
in its evolutionary environment,
Lex Fridman (1:12:31.380)
totally different play styles from humans versus the bots.
Lex Fridman (1:12:34.340)
And yet it's able to handle it extremely well.
Lex Fridman (1:12:37.220)
And that's something that I think was very surprising to us,
Greg Brockman (1:12:40.420)
was something that doesn't really emerge
Lex Fridman (1:12:43.460)
from what we've seen with PPO at smaller scale, right?
Lex Fridman (1:12:47.260)
And the kind of scale we're running this stuff at was,
Lex Fridman (1:12:49.780)
I could say like 100,000 CPU cores
Greg Brockman (1:12:51.980)
running with like hundreds of GPUs.
Lex Fridman (1:12:54.140)
It was probably about something like hundreds
Greg Brockman (1:12:57.580)
of years of experience going into this bot
Lex Fridman (1:13:01.300)
every single real day.
Lex Fridman (1:13:03.860)
And so that scale is massive
Lex Fridman (1:13:06.280)
and we start to see very different kinds of behaviors
Greg Brockman (1:13:08.500)
out of the algorithms that we all know and love.
Lex Fridman (1:13:10.820)
Dota, you mentioned, beat the world expert one v one.
Lex Fridman (1:13:15.260)
And then you weren't able to win five v five this year.
Lex Fridman (1:13:20.820)
Yeah.
Greg Brockman (1:13:21.660)
At the best players in the world.
Lex Fridman (1:13:24.180)
So what's the comeback story?
Greg Brockman (1:13:26.700)
First of all, talk through that.
Lex Fridman (1:13:27.740)
That was an exceptionally exciting event.
Lex Fridman (1:13:29.540)
And what's the following months and this year look like?
Lex Fridman (1:13:33.260)
Yeah, yeah, so one thing that's interesting
Greg Brockman (1:13:35.340)
is that we lose all the time.
Lex Fridman (1:13:38.700)
Because we play.
Lex Fridman (1:13:39.540)
Who's we here?
Lex Fridman (1:13:40.380)
The Dota team at OpenAI.
Greg Brockman (1:13:41.820)
We play the bot against better players
Lex Fridman (1:13:44.260)
than our system all the time.
Lex Fridman (1:13:45.920)
Or at least we used to, right?
Lex Fridman (1:13:47.500)
Like the first time we lost publicly
Greg Brockman (1:13:50.200)
was we went up on stage at the international
Lex Fridman (1:13:52.340)
and we played against some of the best teams in the world
Lex Fridman (1:13:54.740)
and we ended up losing both games,
Lex Fridman (1:13:56.440)
but we gave them a run for their money, right?
Greg Brockman (1:13:58.660)
That both games were kind of 30 minutes, 25 minutes
Lex Fridman (1:14:01.540)
and they went back and forth, back and forth,
Greg Brockman (1:14:03.260)
back and forth.
Lex Fridman (1:14:04.180)
And so I think that really shows
Greg Brockman (1:14:06.020)
that we're at the professional level
Lex Fridman (1:14:08.140)
and that kind of looking at those games,
Greg Brockman (1:14:09.780)
we think that the coin could have gone a different direction
Lex Fridman (1:14:12.420)
and we could have had some wins.
Greg Brockman (1:14:14.140)
That was actually very encouraging for us.
Lex Fridman (1:14:16.140)
And it's interesting because the international
Lex Fridman (1:14:18.380)
was at a fixed time, right?
Lex Fridman (1:14:19.860)
So we knew exactly what day we were going to be playing
Lex Fridman (1:14:22.900)
and we pushed as far as we could, as fast as we could.
Lex Fridman (1:14:25.660)
Two weeks later, we had a bot that had an 80% win rate
Greg Brockman (1:14:28.160)
versus the one that played at TI.
Lex Fridman (1:14:30.260)
So the march of progress, you should think of it
Greg Brockman (1:14:32.460)
as a snapshot rather than as an end state.
Lex Fridman (1:14:34.920)
And so in fact, we'll be announcing our finals pretty soon.
Greg Brockman (1:14:39.180)
I actually think that we'll announce our final match
Lex Fridman (1:14:42.900)
prior to this podcast being released.
Lex Fridman (1:14:45.340)
So we'll be playing against the world champions.
Lex Fridman (1:14:49.900)
And for us, it's really less about,
Greg Brockman (1:14:52.700)
like the way that we think about what's upcoming
Lex Fridman (1:14:55.460)
is the final milestone, the final competitive milestone
Lex Fridman (1:14:59.180)
for the project, right?
Lex Fridman (1:15:00.460)
That our goal in all of this
Greg Brockman (1:15:02.220)
isn't really about beating humans at Dota.
Lex Fridman (1:15:05.340)
Our goal is to push the state of the art
Greg Brockman (1:15:06.980)
in reinforcement learning.
Lex Fridman (1:15:08.020)
And we've done that, right?
Lex Fridman (1:15:09.100)
And we've actually learned a lot from our system
Lex Fridman (1:15:10.820)
and that we have, I think, a lot of exciting next steps
Greg Brockman (1:15:13.940)
that we want to take.
Lex Fridman (1:15:14.860)
And so kind of as a final showcase of what we built,
Greg Brockman (1:15:17.480)
we're going to do this match.
Lex Fridman (1:15:18.900)
But for us, it's not really the success or failure
Greg Brockman (1:15:21.380)
to see do we have the coin flip go in our direction
Lex Fridman (1:15:24.480)
or against.
Greg Brockman (1:15:25.940)
Where do you see the field of deep learning
Lex Fridman (1:15:28.860)
heading in the next few years?
Greg Brockman (1:15:31.620)
Where do you see the work and reinforcement learning
Lex Fridman (1:15:35.620)
perhaps heading, and more specifically with OpenAI,
Greg Brockman (1:15:41.220)
all the exciting projects that you're working on,
Lex Fridman (1:15:44.460)
what does 2019 hold for you?
Greg Brockman (1:15:46.460)
Massive scale.
Lex Fridman (1:15:47.420)
Scale.
Greg Brockman (1:15:48.260)
I will put an asterisk on that and just say,
Lex Fridman (1:15:49.900)
I think that it's about ideas plus scale.
Greg Brockman (1:15:52.340)
You need both.
Lex Fridman (1:15:53.180)
So that's a really good point.
Lex Fridman (1:15:55.060)
So the question, in terms of ideas,
Lex Fridman (1:15:58.620)
you have a lot of projects
Greg Brockman (1:16:00.620)
that are exploring different areas of intelligence.
Lex Fridman (1:16:04.380)
And the question is, when you think of scale,
Lex Fridman (1:16:07.660)
do you think about growing the scale
Lex Fridman (1:16:09.820)
of those individual projects
Lex Fridman (1:16:10.940)
or do you think about adding new projects?
Lex Fridman (1:16:13.260)
And sorry to, and if you're thinking about
Greg Brockman (1:16:16.060)
adding new projects, or if you look at the past,
Lex Fridman (1:16:19.020)
what's the process of coming up with new projects
Lex Fridman (1:16:21.380)
and new ideas?
Lex Fridman (1:16:22.220)
Yep.
Lex Fridman (1:16:23.060)
So we really have a life cycle of project here.
Lex Fridman (1:16:25.380)
So we start with a few people
Greg Brockman (1:16:27.040)
just working on a small scale idea.
Lex Fridman (1:16:28.560)
And language is actually a very good example of this.
Greg Brockman (1:16:30.700)
That it was really one person here
Lex Fridman (1:16:32.620)
who was pushing on language for a long time.
Lex Fridman (1:16:35.020)
I mean, then you get signs of life, right?
Lex Fridman (1:16:36.820)
And so this is like, let's say,
Greg Brockman (1:16:38.860)
with the original GPT, we had something that was interesting
Lex Fridman (1:16:42.740)
and we said, okay, it's time to scale this, right?
Greg Brockman (1:16:44.940)
It's time to put more people on it,
Lex Fridman (1:16:46.100)
put more computational resources behind it.
Lex Fridman (1:16:48.160)
And then we just kind of keep pushing and keep pushing.
Lex Fridman (1:16:51.660)
And the end state is something
Greg Brockman (1:16:52.700)
that looks like Dota or robotics,
Lex Fridman (1:16:54.420)
where you have a large team of 10 or 15 people
Greg Brockman (1:16:57.220)
that are running things at very large scale
Lex Fridman (1:16:59.300)
and that you're able to really have material engineering
Lex Fridman (1:17:02.300)
and sort of machine learning science coming together
Lex Fridman (1:17:06.640)
to make systems that work and get material results
Greg Brockman (1:17:10.380)
that just would have been impossible otherwise.
Lex Fridman (1:17:12.380)
So we do that whole life cycle.
Greg Brockman (1:17:13.740)
We've done it a number of times, typically end to end.
Lex Fridman (1:17:16.780)
It's probably two years or so to do it.
Greg Brockman (1:17:20.540)
The organization has been around for three years,
Lex Fridman (1:17:21.900)
so maybe we'll find that we also have
Greg Brockman (1:17:23.140)
longer life cycle projects, but we'll work up to those.
Lex Fridman (1:17:29.740)
So one team that we were actually just starting,
Greg Brockman (1:17:31.580)
Ilya and I are kicking off a new team
Lex Fridman (1:17:33.400)
called the Reasoning Team,
Lex Fridman (1:17:34.620)
and that this is to really try to tackle
Lex Fridman (1:17:36.420)
how do you get neural networks to reason?
Lex Fridman (1:17:38.700)
And we think that this will be a long term project.
Lex Fridman (1:17:42.700)
It's one that we're very excited about.
Greg Brockman (1:17:44.720)
In terms of reasoning, super exciting topic,
Lex Fridman (1:17:48.400)
what kind of benchmarks, what kind of tests of reasoning
Lex Fridman (1:17:54.180)
do you envision?
Lex Fridman (1:17:55.280)
What would, if you sat back with whatever drink
Lex Fridman (1:17:58.980)
and you would be impressed that this system
Lex Fridman (1:18:01.220)
is able to do something, what would that look like?
Greg Brockman (1:18:03.900)
Theorem proving.
Lex Fridman (1:18:04.860)
Theorem proving.
Lex Fridman (1:18:06.460)
So some kind of logic, and especially mathematical logic.
Lex Fridman (1:18:10.540)
I think so.
Greg Brockman (1:18:11.380)
I think that there's other problems that are dual
Lex Fridman (1:18:14.180)
to theorem proving in particular.
Greg Brockman (1:18:15.980)
You think about programming, you think about
Lex Fridman (1:18:18.500)
even security analysis of code,
Greg Brockman (1:18:21.260)
that these all kind of capture the same sorts
Lex Fridman (1:18:23.720)
of core reasoning and being able to do
Greg Brockman (1:18:26.200)
some out of distribution generalization.
Lex Fridman (1:18:28.360)
So it would be quite exciting if OpenAI Reasoning Team
Greg Brockman (1:18:32.600)
was able to prove that P equals NP.
Lex Fridman (1:18:34.720)
That would be very nice.
Greg Brockman (1:18:36.040)
It would be very, very, very exciting, especially.
Lex Fridman (1:18:38.560)
If it turns out that P equals NP,
Greg Brockman (1:18:39.760)
that'll be interesting too.
Lex Fridman (1:18:41.060)
It would be ironic and humorous.
Lex Fridman (1:18:47.560)
So what problem stands out to you
Lex Fridman (1:18:49.880)
as the most exciting and challenging and impactful
Greg Brockman (1:18:53.960)
to the work for us as a community in general
Lex Fridman (1:18:56.380)
and for OpenAI this year?
Greg Brockman (1:18:58.520)
You mentioned reasoning.
Lex Fridman (1:18:59.600)
I think that's a heck of a problem.
Greg Brockman (1:19:01.440)
Yeah, so I think reasoning's an important one.
Lex Fridman (1:19:02.880)
I think it's gonna be hard to get good results in 2019.
Greg Brockman (1:19:05.840)
Again, just like we think about the life cycle, takes time.
Lex Fridman (1:19:08.760)
I think for 2019, language modeling seems to be
Greg Brockman (1:19:11.040)
kind of on that ramp.
Lex Fridman (1:19:12.640)
It's at the point that we have a technique that works.
Greg Brockman (1:19:14.960)
We wanna scale 100x, 1,000x, see what happens.
Lex Fridman (1:19:18.080)
Awesome.
Lex Fridman (1:19:19.040)
Do you think we're living in a simulation?
Lex Fridman (1:19:21.600)
I think it's hard to have a real opinion about it.
Greg Brockman (1:19:24.840)
It's actually interesting.
Lex Fridman (1:19:26.320)
I separate out things that I think can have like,
Greg Brockman (1:19:29.520)
yield materially different predictions about the world
Lex Fridman (1:19:32.680)
from ones that are just kind of fun to speculate about.
Greg Brockman (1:19:35.880)
I kind of view simulation as more like,
Lex Fridman (1:19:37.960)
is there a flying teapot between Mars and Jupiter?
Greg Brockman (1:19:40.320)
Like, maybe, but it's a little bit hard to know
Lex Fridman (1:19:44.000)
what that would mean for my life.
Lex Fridman (1:19:45.120)
So there is something actionable.
Lex Fridman (1:19:47.000)
So some of the best work OpenAI has done
Greg Brockman (1:19:50.760)
is in the field of reinforcement learning.
Lex Fridman (1:19:52.780)
And some of the success of reinforcement learning
Greg Brockman (1:19:56.620)
come from being able to simulate
Lex Fridman (1:19:58.520)
the problem you're trying to solve.
Lex Fridman (1:20:00.120)
So do you have a hope for reinforcement,
Lex Fridman (1:20:03.680)
for the future of reinforcement learning
Lex Fridman (1:20:05.320)
and for the future of simulation?
Lex Fridman (1:20:07.080)
Like whether it's, we're talking about autonomous vehicles
Greg Brockman (1:20:09.120)
or any kind of system.
Lex Fridman (1:20:10.920)
Do you see that scaling to where we'll be able
Greg Brockman (1:20:13.560)
to simulate systems and hence,
Lex Fridman (1:20:16.440)
be able to create a simulator that echoes our real world
Lex Fridman (1:20:19.400)
and proving once and for all,
Lex Fridman (1:20:21.620)
even though you're denying it,
Lex Fridman (1:20:22.680)
that we're living in a simulation?
Lex Fridman (1:20:25.080)
I feel like it's two separate questions, right?
Lex Fridman (1:20:26.500)
So kind of at the core there of like,
Lex Fridman (1:20:28.400)
can we use simulation for self driving cars?
Lex Fridman (1:20:31.240)
Take a look at our robotic system, Dactyl, right?
Lex Fridman (1:20:33.860)
That was trained in simulation using the Dota system,
Greg Brockman (1:20:37.000)
in fact, and it transfers to a physical robot.
Lex Fridman (1:20:40.480)
And I think everyone looks at our Dota system,
Greg Brockman (1:20:42.320)
they're like, okay, it's just a game.
Lex Fridman (1:20:43.560)
How are you ever gonna escape to the real world?
Lex Fridman (1:20:45.260)
And the answer is, well, we did it with a physical robot
Lex Fridman (1:20:47.480)
that no one could program.
Lex Fridman (1:20:48.720)
And so I think the answer is simulation
Lex Fridman (1:20:50.240)
goes a lot further than you think
Greg Brockman (1:20:52.080)
if you apply the right techniques to it.
Lex Fridman (1:20:54.240)
Now, there's a question of,
Greg Brockman (1:20:55.480)
are the beings in that simulation gonna wake up
Lex Fridman (1:20:57.520)
and have consciousness?
Greg Brockman (1:20:59.620)
I think that one seems a lot harder to, again,
Lex Fridman (1:21:02.380)
reason about.
Greg Brockman (1:21:03.220)
I think that you really should think about
Lex Fridman (1:21:05.400)
where exactly does human consciousness come from
Lex Fridman (1:21:07.940)
in our own self awareness?
Lex Fridman (1:21:09.160)
And is it just that once you have a complicated enough
Greg Brockman (1:21:11.920)
neural net, you have to worry about
Lex Fridman (1:21:13.220)
the agents feeling pain?
Lex Fridman (1:21:15.840)
And I think there's interesting speculation to do there,
Lex Fridman (1:21:19.440)
but again, I think it's a little bit hard to know for sure.
Greg Brockman (1:21:23.120)
Well, let me just keep with the speculation.
Lex Fridman (1:21:25.040)
Do you think to create intelligence, general intelligence,
Lex Fridman (1:21:28.640)
you need, one, consciousness, and two, a body?
Lex Fridman (1:21:33.180)
Do you think any of those elements are needed,
Lex Fridman (1:21:35.040)
or is intelligence something that's orthogonal to those?
Lex Fridman (1:21:38.480)
I'll stick to the non grand answer first, right?
Lex Fridman (1:21:41.920)
So the non grand answer is just to look at,
Lex Fridman (1:21:44.360)
what are we already making work?
Greg Brockman (1:21:45.800)
You look at GPT2, a lot of people would have said
Lex Fridman (1:21:47.800)
that to even get these kinds of results,
Greg Brockman (1:21:49.480)
you need real world experience.
Lex Fridman (1:21:51.080)
You need a body, you need grounding.
Lex Fridman (1:21:52.560)
How are you supposed to reason about any of these things?
Lex Fridman (1:21:55.060)
How are you supposed to like even kind of know
Greg Brockman (1:21:56.500)
about smoke and fire and those things
Lex Fridman (1:21:58.040)
if you've never experienced them?
Lex Fridman (1:21:59.740)
And GPT2 shows that you can actually go way further
Lex Fridman (1:22:03.000)
than that kind of reasoning would predict.
Lex Fridman (1:22:06.880)
So I think that in terms of, do we need consciousness?
Lex Fridman (1:22:10.600)
Do we need a body?
Lex Fridman (1:22:11.840)
It seems the answer is probably not, right?
Lex Fridman (1:22:13.400)
That we could probably just continue to push
Greg Brockman (1:22:15.100)
kind of the systems we have.
Lex Fridman (1:22:16.140)
They already feel general.
Greg Brockman (1:22:18.280)
They're not as competent or as general
Lex Fridman (1:22:20.560)
or able to learn as quickly as an AGI would,
Lex Fridman (1:22:23.000)
but they're at least like kind of proto AGI in some way,
Lex Fridman (1:22:27.420)
and they don't need any of those things.
Greg Brockman (1:22:29.800)
Now let's move to the grand answer,
Lex Fridman (1:22:31.960)
which is, are our neural nets conscious already?
Lex Fridman (1:22:36.520)
Would we ever know?
Lex Fridman (1:22:37.440)
How can we tell, right?
Lex Fridman (1:22:38.920)
And here's where the speculation starts to become
Lex Fridman (1:22:43.040)
at least interesting or fun
Lex Fridman (1:22:44.920)
and maybe a little bit disturbing
Lex Fridman (1:22:46.520)
depending on where you take it.
Lex Fridman (1:22:48.080)
But it certainly seems that when we think about animals,
Lex Fridman (1:22:51.280)
that there's some continuum of consciousness.
Lex Fridman (1:22:53.280)
You know, my cat I think is conscious in some way, right?
Lex Fridman (1:22:57.120)
Not as conscious as a human.
Lex Fridman (1:22:58.200)
And you could imagine that you could build
Lex Fridman (1:23:00.080)
a little consciousness meter, right?
Greg Brockman (1:23:01.220)
You point at a cat, it gives you a little reading.
Lex Fridman (1:23:03.080)
Point at a human, it gives you much bigger reading.
Lex Fridman (1:23:06.400)
What would happen if you pointed one of those
Lex Fridman (1:23:08.120)
at a donor neural net?
Lex Fridman (1:23:09.960)
And if you're training in this massive simulation,
Lex Fridman (1:23:12.180)
do the neural nets feel pain?
Greg Brockman (1:23:13.680)
You know, it becomes pretty hard to know
Lex Fridman (1:23:16.960)
that the answer is no.
Lex Fridman (1:23:18.840)
And it becomes pretty hard to really think about
Lex Fridman (1:23:21.660)
what that would mean if the answer were yes.
Lex Fridman (1:23:25.440)
And it's very possible, you know, for example,
Lex Fridman (1:23:27.600)
you could imagine that maybe the reason
Greg Brockman (1:23:29.600)
that humans have consciousness
Lex Fridman (1:23:31.560)
is because it's a convenient computational shortcut, right?
Greg Brockman (1:23:35.160)
If you think about it, if you have a being
Lex Fridman (1:23:37.120)
that wants to avoid pain,
Greg Brockman (1:23:38.360)
which seems pretty important to survive in this environment
Lex Fridman (1:23:40.960)
and wants to like, you know, eat food,
Greg Brockman (1:23:43.800)
then that maybe the best way of doing it
Lex Fridman (1:23:45.640)
is to have a being that's conscious, right?
Greg Brockman (1:23:47.240)
That, you know, in order to succeed in the environment,
Lex Fridman (1:23:49.640)
you need to have those properties
Lex Fridman (1:23:51.200)
and how are you supposed to implement them
Lex Fridman (1:23:52.760)
and maybe this consciousness's way of doing that.
Greg Brockman (1:23:55.440)
If that's true, then actually maybe we should expect
Lex Fridman (1:23:57.920)
that really competent reinforcement learning agents
Greg Brockman (1:24:00.060)
will also have consciousness.
Lex Fridman (1:24:02.120)
But you know, that's a big if.
Lex Fridman (1:24:03.360)
And I think there are a lot of other arguments
Lex Fridman (1:24:04.880)
they can make in other directions.
Greg Brockman (1:24:06.760)
I think that's a really interesting idea
Lex Fridman (1:24:08.520)
that even GPT2 has some degree of consciousness.
Greg Brockman (1:24:11.520)
That's something, it's actually not as crazy
Lex Fridman (1:24:14.320)
to think about, it's useful to think about
Greg Brockman (1:24:16.640)
as we think about what it means
Lex Fridman (1:24:18.320)
to create intelligence of a dog, intelligence of a cat,
Lex Fridman (1:24:22.240)
and the intelligence of a human.
Lex Fridman (1:24:24.480)
So last question, do you think
Greg Brockman (1:24:27.880)
we will ever fall in love, like in the movie Her,
Lex Fridman (1:24:32.040)
with an artificial intelligence system
Greg Brockman (1:24:34.480)
or an artificial intelligence system
Lex Fridman (1:24:36.300)
falling in love with a human?
Greg Brockman (1:24:38.640)
I hope so.
Lex Fridman (1:24:40.280)
If there's any better way to end it is on love.
Lex Fridman (1:24:43.760)
So Greg, thanks so much for talking today.
Lex Fridman (1:24:45.680)
Thank you for having me.
Greg Brockman (20:00.520)
That really talking about AGI or thinking about AGI
Lex Fridman (20:03.240)
became almost this taboo in the community.
Lex Fridman (20:06.600)
But I actually think that people took the wrong lesson
Lex Fridman (20:08.660)
from AI history.
Lex Fridman (20:10.080)
And if you look back, starting in 1959
Lex Fridman (20:12.360)
is when the Perceptron was released.
Lex Fridman (20:14.240)
And this is basically one of the earliest neural networks.
Lex Fridman (20:17.680)
It was released to what was perceived
Greg Brockman (20:19.220)
as this massive overhype.
Lex Fridman (20:20.820)
So in the New York Times in 1959,
Greg Brockman (20:22.320)
you have this article saying that the Perceptron
Lex Fridman (20:26.380)
will one day recognize people, call out their names,
Greg Brockman (20:29.160)
instantly translate speech between languages.
Lex Fridman (20:31.440)
And people at the time looked at this and said,
Greg Brockman (20:33.800)
this is, your system can't do any of that.
Lex Fridman (20:36.080)
And basically spent 10 years trying to discredit
Greg Brockman (20:38.060)
the whole Perceptron direction and succeeded.
Lex Fridman (20:40.600)
And all the funding dried up.
Lex Fridman (20:41.800)
And people kind of went in other directions.
Lex Fridman (20:44.960)
And in the 80s, there was this resurgence.
Lex Fridman (20:46.900)
And I'd always heard that the resurgence in the 80s
Lex Fridman (20:49.280)
was due to the invention of backpropagation
Lex Fridman (20:51.480)
and these algorithms that got people excited.
Lex Fridman (20:53.680)
But actually the causality was due to people
Greg Brockman (20:55.720)
building larger computers.
Lex Fridman (20:57.140)
That you can find these articles from the 80s
Greg Brockman (20:59.080)
saying that the democratization of computing power
Lex Fridman (21:01.720)
suddenly meant that you could run
Greg Brockman (21:02.660)
these larger neural networks.
Lex Fridman (21:04.000)
And then people started to do all these amazing things.
Greg Brockman (21:06.280)
Backpropagation algorithm was invented.
Lex Fridman (21:08.000)
And the neural nets people were running
Greg Brockman (21:10.100)
were these tiny little 20 neuron neural nets.
Lex Fridman (21:13.040)
What are you supposed to learn with 20 neurons?
Lex Fridman (21:15.160)
And so of course, they weren't able to get great results.
Lex Fridman (21:18.640)
And it really wasn't until 2012 that this approach,
Greg Brockman (21:21.940)
that's almost the most simple, natural approach
Lex Fridman (21:24.680)
that people had come up with in the 50s,
Greg Brockman (21:27.720)
in some ways even in the 40s before there were computers,
Lex Fridman (21:30.320)
with the Pitts–McCullough neuron,
Greg Brockman (21:32.120)
suddenly this became the best way of solving problems.
Lex Fridman (21:37.460)
And I think there are three core properties
Greg Brockman (21:39.260)
that deep learning has that I think
Lex Fridman (21:42.080)
are very worth paying attention to.
Greg Brockman (21:44.100)
The first is generality.
Lex Fridman (21:45.900)
We have a very small number of deep learning tools.
Greg Brockman (21:48.700)
SGD, deep neural net, maybe some RL.
Lex Fridman (21:53.180)
And it solves this huge variety of problems.
Greg Brockman (21:55.580)
Speech recognition, machine translation,
Lex Fridman (21:57.220)
game playing, all of these problems, small set of tools.
Lex Fridman (22:00.980)
So there's the generality.
Lex Fridman (22:02.740)
There's a second piece, which is the competence.
Lex Fridman (22:04.980)
You want to solve any of those problems?
Lex Fridman (22:07.020)
Throw up 40 years worth of normal computer vision research,
Greg Brockman (22:10.620)
replace it with a deep neural net,
Lex Fridman (22:11.780)
it's going to work better.
Lex Fridman (22:13.580)
And there's a third piece, which is the scalability.
Lex Fridman (22:16.860)
One thing that has been shown time and time again
Greg Brockman (22:18.680)
is that if you have a larger neural network,
Lex Fridman (22:21.740)
throw more compute, more data at it, it will work better.
Greg Brockman (22:25.120)
Those three properties together feel like essential parts
Lex Fridman (22:28.860)
of building a general intelligence.
Greg Brockman (22:30.820)
Now it doesn't just mean that if we scale up what we have,
Lex Fridman (22:33.800)
that we will have an AGI, right?
Greg Brockman (22:35.180)
There are clearly missing pieces.
Lex Fridman (22:36.780)
There are missing ideas.
Greg Brockman (22:38.020)
We need to have answers for reasoning.
Lex Fridman (22:40.000)
But I think that the core here is that for the first time,
Greg Brockman (22:44.780)
it feels that we have a paradigm that gives us hope
Lex Fridman (22:47.940)
that general intelligence can be achievable.
Lex Fridman (22:50.580)
And so as soon as you believe that,
Lex Fridman (22:52.140)
everything else comes into focus, right?
Greg Brockman (22:54.460)
If you imagine that you may be able to,
Lex Fridman (22:56.580)
and you know that the timeline I think remains uncertain,
Lex Fridman (22:59.820)
but I think that certainly within our lifetimes
Lex Fridman (23:02.220)
and possibly within a much shorter period of time
Greg Brockman (23:04.660)
than people would expect,
Lex Fridman (23:06.580)
if you can really build the most transformative technology
Greg Brockman (23:09.340)
that will ever exist,
Lex Fridman (23:10.660)
you stop thinking about yourself so much, right?
Greg Brockman (23:12.620)
You start thinking about just like,
Lex Fridman (23:14.220)
how do you have a world where this goes well?
Lex Fridman (23:16.440)
And that you need to think about the practicalities
Lex Fridman (23:18.180)
of how do you build an organization
Lex Fridman (23:19.540)
and get together a bunch of people and resources
Lex Fridman (23:22.020)
and to make sure that people feel motivated
Lex Fridman (23:25.140)
and ready to do it.
Lex Fridman (23:26.780)
But I think that then you start thinking about,
Lex Fridman (23:29.260)
well, what if we succeed?
Lex Fridman (23:30.580)
And how do we make sure that when we succeed,
Greg Brockman (23:32.740)
that the world is actually the place
Lex Fridman (23:34.020)
that we want ourselves to exist in?
Lex Fridman (23:36.780)
And almost in the Rawlsian Veil sense of the word.
Lex Fridman (23:39.500)
And so that's kind of the broader landscape.
Lex Fridman (23:42.340)
And OpenAI was really formed in 2015
Lex Fridman (23:45.140)
with that high level picture of AGI might be possible
Greg Brockman (23:50.140)
sooner than people think,
Lex Fridman (23:51.380)
and that we need to try to do our best
Greg Brockman (23:54.420)
to make sure it's going to go well.
Lex Fridman (23:55.820)
And then we spent the next couple of years
Lex Fridman (23:57.740)
really trying to figure out what does that mean?
Lex Fridman (23:59.180)
How do we do it?
Lex Fridman (24:00.500)
And I think that typically with a company,
Lex Fridman (24:03.060)
you start out very small, see you in a co founder,
Lex Fridman (24:06.460)
and you build a product, you get some users,
Lex Fridman (24:07.900)
you get a product market fit.
Greg Brockman (24:09.540)
Then at some point you raise some money,
Lex Fridman (24:11.620)
you hire people, you scale, and then down the road,
Greg Brockman (24:14.940)
then the big companies realize you exist
Lex Fridman (24:16.420)
and try to kill you.
Lex Fridman (24:17.420)
And for OpenAI, it was basically everything
Lex Fridman (24:19.860)
in exactly the opposite order.
Greg Brockman (24:21.260)
Let me just pause for a second, you said a lot of things.
Lex Fridman (24:26.260)
And let me just admire the jarring aspect
Greg Brockman (24:29.740)
of what OpenAI stands for, which is daring to dream.
Lex Fridman (24:33.740)
I mean, you said it's pretty powerful.
Greg Brockman (24:35.620)
It caught me off guard because I think that's very true.
Lex Fridman (24:38.620)
The step of just daring to dream about the possibilities
Greg Brockman (24:43.620)
of creating intelligence in a positive, in a safe way,
Lex Fridman (24:47.180)
but just even creating intelligence is a very powerful
Greg Brockman (24:50.700)
is a much needed refreshing catalyst for the AI community.
Lex Fridman (24:57.460)
So that's the starting point.
Lex Fridman (24:58.860)
Okay, so then formation of OpenAI, what's that?
Lex Fridman (25:02.900)
I would just say that when we were starting OpenAI,
Greg Brockman (25:05.740)
that kind of the first question that we had is,
Lex Fridman (25:07.820)
is it too late to start a lab
Lex Fridman (25:10.380)
with a bunch of the best people?
Lex Fridman (25:12.060)
Right, is that even possible? Wow, okay.
Lex Fridman (25:13.220)
That was an actual question?
Lex Fridman (25:14.540)
That was the core question of,
Greg Brockman (25:17.340)
we had this dinner in July of 2015,
Lex Fridman (25:19.380)
and that was really what we spent the whole time
Greg Brockman (25:21.220)
talking about.
Lex Fridman (25:22.300)
And, you know, because you think about kind of where AI was
Greg Brockman (25:26.780)
is that it had transitioned from being an academic pursuit
Lex Fridman (25:30.180)
to an industrial pursuit.
Lex Fridman (25:32.220)
And so a lot of the best people were in these big
Lex Fridman (25:34.220)
research labs and that we wanted to start our own one
Greg Brockman (25:36.980)
that no matter how much resources we could accumulate
Lex Fridman (25:40.540)
would be pale in comparison to the big tech companies.
Lex Fridman (25:43.500)
And we knew that.
Lex Fridman (25:44.700)
And it was a question of, are we going to be actually
Lex Fridman (25:47.020)
able to get this thing off the ground?
Lex Fridman (25:48.700)
You need critical mass.
Greg Brockman (25:49.740)
You can't just do you and a cofounder build a product.
Lex Fridman (25:52.100)
You really need to have a group of five to 10 people.
Lex Fridman (25:55.580)
And we kind of concluded it wasn't obviously impossible.
Lex Fridman (25:59.460)
So it seemed worth trying.
Lex Fridman (26:02.220)
Well, you're also a dreamer, so who knows, right?
Lex Fridman (26:04.780)
That's right.
Greg Brockman (26:05.620)
Okay, so speaking of that, competing with the big players,
Lex Fridman (26:11.460)
let's talk about some of the tricky things
Greg Brockman (26:14.020)
as you think through this process of growing,
Lex Fridman (26:17.420)
of seeing how you can develop these systems
Greg Brockman (26:20.060)
at a scale that competes.
Lex Fridman (26:22.580)
So you recently formed OpenAI LP,
Greg Brockman (26:26.540)
a new cap profit company that now carries the name OpenAI.
Lex Fridman (26:30.780)
So OpenAI is now this official company.
Greg Brockman (26:33.260)
The original nonprofit company still exists
Lex Fridman (26:36.500)
and carries the OpenAI nonprofit name.
Lex Fridman (26:39.740)
So can you explain what this company is,
Lex Fridman (26:41.940)
what the purpose of this creation is,
Lex Fridman (26:44.220)
and how did you arrive at the decision to create it?
Lex Fridman (26:48.740)
OpenAI, the whole entity and OpenAI LP as a vehicle
Greg Brockman (26:53.220)
is trying to accomplish the mission
Lex Fridman (26:55.500)
of ensuring that artificial general intelligence
Greg Brockman (26:57.460)
benefits everyone.
Lex Fridman (26:58.740)
And the main way that we're trying to do that
Greg Brockman (27:00.180)
is by actually trying to build general intelligence
Lex Fridman (27:02.500)
ourselves and make sure the benefits
Greg Brockman (27:04.140)
are distributed to the world.
Lex Fridman (27:05.860)
That's the primary way.
Lex Fridman (27:07.100)
We're also fine if someone else does this, right?
Lex Fridman (27:09.540)
Doesn't have to be us.
Greg Brockman (27:10.580)
If someone else is going to build an AGI
Lex Fridman (27:12.540)
and make sure that the benefits don't get locked up
Greg Brockman (27:14.740)
in one company or with one set of people,
Lex Fridman (27:19.220)
like we're actually fine with that.
Lex Fridman (27:21.100)
And so those ideas are baked into our charter,
Lex Fridman (27:25.340)
which is kind of the foundational document
Greg Brockman (27:28.340)
that describes kind of our values and how we operate.
Lex Fridman (27:32.780)
But it's also really baked into the structure of OpenAI LP.
Lex Fridman (27:36.300)
And so the way that we've set up OpenAI LP
Lex Fridman (27:37.900)
is that in the case where we succeed, right?
Greg Brockman (27:42.100)
If we actually build what we're trying to build,
Lex Fridman (27:45.260)
then investors are able to get a return,
Lex Fridman (27:48.300)
but that return is something that is capped.
Lex Fridman (27:50.300)
And so if you think of AGI in terms of the value
Greg Brockman (27:52.940)
that you could really create,
Lex Fridman (27:54.100)
you're talking about the most transformative technology
Greg Brockman (27:56.260)
ever created, it's going to create orders of magnitude
Lex Fridman (27:58.780)
more value than any existing company.
Lex Fridman (28:01.820)
And that all of that value will be owned by the world,
Lex Fridman (28:05.900)
like legally titled to the nonprofit
Greg Brockman (28:07.820)
to fulfill that mission.
Lex Fridman (28:09.500)
And so that's the structure.
Lex Fridman (28:12.740)
So the mission is a powerful one,
Lex Fridman (28:15.140)
and it's one that I think most people would agree with.
Greg Brockman (28:18.860)
It's how we would hope AI progresses.
Lex Fridman (28:22.900)
And so how do you tie yourself to that mission?
Lex Fridman (28:25.340)
How do you make sure you do not deviate from that mission,
Lex Fridman (28:29.180)
that other incentives that are profit driven
Lex Fridman (28:35.260)
don't interfere with the mission?
Lex Fridman (28:36.740)
So this was actually a really core question for us
Greg Brockman (28:39.540)
for the past couple of years,
Lex Fridman (28:40.900)
because I'd say that like the way that our history went
Greg Brockman (28:43.540)
was that for the first year,
Lex Fridman (28:44.920)
we were getting off the ground, right?
Greg Brockman (28:46.200)
We had this high level picture,
Lex Fridman (28:47.900)
but we didn't know exactly how we wanted to accomplish it.
Lex Fridman (28:51.860)
And really two years ago is when we first started realizing
Lex Fridman (28:55.020)
in order to build AGI,
Greg Brockman (28:56.140)
we're just going to need to raise way more money
Lex Fridman (28:58.700)
than we can as a nonprofit.
Lex Fridman (29:00.180)
And we're talking many billions of dollars.
Lex Fridman (29:02.820)
And so the first question is how are you supposed to do that
Lex Fridman (29:06.860)
and stay true to this mission?
Lex Fridman (29:08.700)
And we looked at every legal structure out there
Lex Fridman (29:10.580)
and concluded none of them were quite right
Lex Fridman (29:11.940)
for what we wanted to do.
Lex Fridman (29:13.380)
And I guess it shouldn't be too surprising
Lex Fridman (29:14.580)
if you're gonna do some like crazy unprecedented technology
Greg Brockman (29:16.920)
that you're gonna have to come with
Lex Fridman (29:17.980)
some crazy unprecedented structure to do it in.
Lex Fridman (29:20.340)
And a lot of our conversation was with people at OpenAI,
Lex Fridman (29:26.140)
the people who really joined
Greg Brockman (29:27.220)
because they believe so much in this mission
Lex Fridman (29:29.100)
and thinking about how do we actually
Greg Brockman (29:31.260)
raise the resources to do it
Lex Fridman (29:33.020)
and also stay true to what we stand for.
Lex Fridman (29:35.900)
And the place you gotta start is to really align
Lex Fridman (29:37.940)
on what is it that we stand for, right?
Lex Fridman (29:39.540)
What are those values?
Lex Fridman (29:40.500)
What's really important to us?
Lex Fridman (29:41.820)
And so I'd say that we spent about a year
Lex Fridman (29:43.740)
really compiling the OpenAI charter
Lex Fridman (29:46.220)
and that determines,
Lex Fridman (29:47.540)
and if you even look at the first line item in there,
Greg Brockman (29:50.220)
it says that, look, we expect we're gonna have to marshal
Lex Fridman (29:52.340)
huge amounts of resources,
Lex Fridman (29:53.740)
but we're going to make sure that we minimize
Lex Fridman (29:55.720)
conflict of interest with the mission.
Lex Fridman (29:57.620)
And that kind of aligning on all of those pieces
Lex Fridman (30:00.700)
was the most important step towards figuring out
Lex Fridman (30:04.180)
how do we structure a company
Lex Fridman (30:06.020)
that can actually raise the resources
Greg Brockman (30:08.200)
to do what we need to do.
Lex Fridman (30:10.300)
I imagine OpenAI, the decision to create OpenAI LP
Greg Brockman (30:14.740)
was a really difficult one.
Lex Fridman (30:16.340)
And there was a lot of discussions,
Greg Brockman (30:17.900)
as you mentioned, for a year,
Lex Fridman (30:19.600)
and there was different ideas,
Greg Brockman (30:22.740)
perhaps detractors within OpenAI,
Lex Fridman (30:26.100)
sort of different paths that you could have taken.
Lex Fridman (30:28.900)
What were those concerns?
Lex Fridman (30:30.220)
What were the different paths considered?
Lex Fridman (30:32.020)
What was that process of making that decision like?
Lex Fridman (30:34.100)
Yep, so if you look actually at the OpenAI charter,
Greg Brockman (30:37.720)
there's almost two paths embedded within it.
Lex Fridman (30:40.900)
There is, we are primarily trying to build AGI ourselves,
Lex Fridman (30:44.900)
but we're also okay if someone else does it.
Lex Fridman (30:47.340)
And this is a weird thing for a company.
Greg Brockman (30:49.020)
It's really interesting, actually.
Lex Fridman (30:51.140)
There is an element of competition
Greg Brockman (30:53.260)
that you do wanna be the one that does it,
Lex Fridman (30:56.660)
but at the same time, you're okay if somebody else doesn't.
Greg Brockman (30:59.020)
We'll talk about that a little bit, that trade off,
Lex Fridman (31:01.380)
that dance that's really interesting.
Lex Fridman (31:02.940)
And I think this was the core tension
Lex Fridman (31:04.600)
as we were designing OpenAI LP,
Lex Fridman (31:06.380)
and really the OpenAI strategy,
Lex Fridman (31:08.260)
is how do you make sure that both you have a shot
Greg Brockman (31:11.080)
at being a primary actor,
Lex Fridman (31:12.660)
which really requires building an organization,
Greg Brockman (31:15.820)
raising massive resources,
Lex Fridman (31:17.700)
and really having the will to go
Lex Fridman (31:19.420)
and execute on some really, really hard vision, right?
Lex Fridman (31:22.040)
You need to really sign up for a long period
Greg Brockman (31:23.800)
to go and take on a lot of pain and a lot of risk.
Lex Fridman (31:27.160)
And to do that, normally you just import
Lex Fridman (31:30.420)
the startup mindset, right?
Lex Fridman (31:31.780)
And that you think about, okay,
Lex Fridman (31:32.820)
like how do we out execute everyone?
Lex Fridman (31:34.300)
You have this very competitive angle.
Lex Fridman (31:36.220)
But you also have the second angle of saying that,
Lex Fridman (31:38.180)
well, the true mission isn't for OpenAI to build AGI.
Greg Brockman (31:41.660)
The true mission is for AGI to go well for humanity.
Lex Fridman (31:45.140)
And so how do you take all of those first actions
Lex Fridman (31:48.140)
and make sure you don't close the door on outcomes
Lex Fridman (31:51.380)
that would actually be positive and fulfill the mission?
Lex Fridman (31:54.460)
And so I think it's a very delicate balance, right?
Lex Fridman (31:56.700)
And I think that going 100% one direction or the other
Greg Brockman (31:59.620)
is clearly not the correct answer.
Lex Fridman (32:01.340)
And so I think that even in terms of just how we talk
Greg Brockman (32:03.700)
about OpenAI and think about it,
Lex Fridman (32:05.440)
there's just like one thing that's always in the back
Greg Brockman (32:07.980)
of my mind is to make sure that we're not just saying
Lex Fridman (32:11.260)
OpenAI's goal is to build AGI, right?
Lex Fridman (32:14.020)
That it's actually much broader than that, right?
Lex Fridman (32:15.580)
That first of all, it's not just AGI,
Greg Brockman (32:18.260)
it's safe AGI that's very important.
Lex Fridman (32:20.340)
But secondly, our goal isn't to be the ones to build it.
Greg Brockman (32:23.100)
Our goal is to make sure it goes well for the world.
Lex Fridman (32:24.700)
And so I think that figuring out
Lex Fridman (32:26.100)
how do you balance all of those
Lex Fridman (32:27.900)
and to get people to really come to the table
Lex Fridman (32:30.220)
and compile a single document that encompasses all of that
Lex Fridman (32:36.340)
wasn't trivial.
Lex Fridman (32:37.540)
So part of the challenge here is your mission is,
Lex Fridman (32:41.640)
I would say, beautiful, empowering,
Lex Fridman (32:44.220)
and a beacon of hope for people in the research community
Lex Fridman (32:47.500)
and just people thinking about AI.
Lex Fridman (32:49.180)
So your decisions are scrutinized more than,
Lex Fridman (32:53.140)
I think, a regular profit driven company.
Lex Fridman (32:55.900)
Do you feel the burden of this
Lex Fridman (32:57.380)
in the creation of the charter
Lex Fridman (32:58.540)
and just in the way you operate?
Lex Fridman (33:00.160)
Yes.
Lex Fridman (33:01.000)
So why do you lean into the burden
Lex Fridman (33:07.020)
by creating such a charter?
Lex Fridman (33:08.660)
Why not keep it quiet?
Lex Fridman (33:10.420)
I mean, it just boils down to the mission, right?
Greg Brockman (33:12.900)
Like I'm here and everyone else is here
Lex Fridman (33:15.180)
because we think this is the most important mission.
Greg Brockman (33:17.380)
Dare to dream.
Lex Fridman (33:18.980)
All right, so do you think you can be good for the world
Greg Brockman (33:23.340)
or create an AGI system that's good
Lex Fridman (33:25.980)
when you're a for profit company?
Greg Brockman (33:28.320)
From my perspective, I don't understand
Lex Fridman (33:30.660)
why profit interferes with positive impact on society.
Greg Brockman (33:37.620)
I don't understand why Google,
Lex Fridman (33:40.740)
that makes most of its money from ads,
Greg Brockman (33:42.900)
can't also do good for the world
Lex Fridman (33:45.020)
or other companies, Facebook, anything.
Greg Brockman (33:47.500)
I don't understand why those have to interfere.
Lex Fridman (33:50.220)
You know, profit isn't the thing, in my view,
Greg Brockman (33:55.100)
that affects the impact of a company.
Lex Fridman (33:57.200)
What affects the impact of the company is the charter,
Greg Brockman (34:00.340)
is the culture, is the people inside,
Lex Fridman (34:04.140)
and profit is the thing that just fuels those people.
Lex Fridman (34:07.100)
So what are your views there?
Lex Fridman (34:08.740)
Yeah, so I think that's a really good question
Lex Fridman (34:10.900)
and there's some real longstanding debates
Lex Fridman (34:14.180)
in human society that are wrapped up in it.
Greg Brockman (34:16.460)
The way that I think about it is just think about
Lex Fridman (34:18.640)
what are the most impactful non profits in the world?
Lex Fridman (34:23.980)
What are the most impactful for profits in the world?
Lex Fridman (34:26.780)
Right, it's much easier to list the for profits.
Greg Brockman (34:29.260)
That's right, and I think that there's some real truth here
Lex Fridman (34:32.420)
that the system that we set up,
Greg Brockman (34:34.600)
the system for kind of how today's world is organized,
Lex Fridman (34:38.320)
is one that really allows for huge impact.
Lex Fridman (34:41.300)
And that kind of part of that is that you need to be,
Lex Fridman (34:45.140)
that for profits are self sustaining
Lex Fridman (34:48.060)
and able to kind of build on their own momentum.
Lex Fridman (34:51.180)
And I think that's a really powerful thing.
Greg Brockman (34:53.060)
It's something that when it turns out
Lex Fridman (34:55.860)
that we haven't set the guardrails correctly,
Lex Fridman (34:57.900)
causes problems, right?
Lex Fridman (34:58.820)
Think about logging companies that go into forest,
Greg Brockman (35:01.600)
the rainforest, that's really bad, we don't want that.
Lex Fridman (35:04.680)
And it's actually really interesting to me
Greg Brockman (35:06.500)
that kind of this question of how do you get
Lex Fridman (35:08.940)
positive benefits out of a for profit company,
Greg Brockman (35:11.380)
it's actually very similar to how do you get
Lex Fridman (35:13.020)
positive benefits out of an AGI, right?
Greg Brockman (35:15.800)
That you have this like very powerful system,
Lex Fridman (35:17.980)
it's more powerful than any human,
Lex Fridman (35:19.700)
and is kind of autonomous in some ways,
Lex Fridman (35:21.860)
it's superhuman in a lot of axes,
Lex Fridman (35:23.740)
and somehow you have to set the guardrails
Lex Fridman (35:25.420)
to get good things to happen.
Lex Fridman (35:26.820)
But when you do, the benefits are massive.
Lex Fridman (35:29.380)
And so I think that when I think about
Greg Brockman (35:32.500)
nonprofit versus for profit,
Lex Fridman (35:34.420)
I think just not enough happens in nonprofits,
Greg Brockman (35:36.760)
they're very pure, but it's just kind of,
Lex Fridman (35:39.180)
it's just hard to do things there.
Greg Brockman (35:40.860)
In for profits in some ways, like too much happens,
Lex Fridman (35:43.980)
but if kind of shaped in the right way,
Greg Brockman (35:46.460)
it can actually be very positive.
Lex Fridman (35:47.820)
And so with OpenAI LP, we're picking a road in between.
Greg Brockman (35:52.100)
Now the thing that I think is really important to recognize
Lex Fridman (35:54.820)
is that the way that we think about OpenAI LP
Greg Brockman (35:57.140)
is that in the world where AGI actually happens, right,
Lex Fridman (36:00.420)
in a world where we are successful,
Greg Brockman (36:01.660)
we build the most transformative technology ever,
Lex Fridman (36:03.760)
the amount of value we're gonna create will be astronomical.
Lex Fridman (36:07.580)
And so then in that case, that the cap that we have
Lex Fridman (36:12.760)
will be a small fraction of the value we create,
Lex Fridman (36:15.540)
and the amount of value that goes back to investors
Lex Fridman (36:17.800)
and employees looks pretty similar to what would happen
Greg Brockman (36:20.020)
in a pretty successful startup.
Lex Fridman (36:23.780)
And that's really the case that we're optimizing for, right?
Greg Brockman (36:26.580)
That we're thinking about in the success case,
Lex Fridman (36:28.600)
making sure that the value we create doesn't get locked up.
Lex Fridman (36:32.220)
And I expect that in other for profit companies
Lex Fridman (36:34.980)
that it's possible to do something like that.
Lex Fridman (36:37.860)
I think it's not obvious how to do it, right?
Lex Fridman (36:39.780)
I think that as a for profit company,
Greg Brockman (36:41.500)
you have a lot of fiduciary duty to your shareholders
Lex Fridman (36:44.300)
and that there are certain decisions
Greg Brockman (36:45.700)
that you just cannot make.
Lex Fridman (36:47.560)
In our structure, we've set it up
Lex Fridman (36:49.140)
so that we have a fiduciary duty to the charter,
Lex Fridman (36:52.500)
that we always get to make the decision
Greg Brockman (36:54.460)
that is right for the charter rather than,
Lex Fridman (36:57.460)
even if it comes at the expense of our own stakeholders.
Lex Fridman (37:00.700)
And so I think that when I think about
Lex Fridman (37:03.420)
what's really important,
Greg Brockman (37:04.380)
it's not really about nonprofit versus for profit,
Lex Fridman (37:06.300)
it's really a question of if you build AGI
Lex Fridman (37:09.620)
and you kind of, humanity's now in this new age,
Lex Fridman (37:13.100)
who benefits, whose lives are better?
Lex Fridman (37:15.780)
And I think that what's really important
Lex Fridman (37:17.180)
is to have an answer that is everyone.
Greg Brockman (37:20.340)
Yeah, which is one of the core aspects of the charter.
Lex Fridman (37:23.380)
So one concern people have, not just with OpenAI,
Lex Fridman (37:26.540)
but with Google, Facebook, Amazon,
Lex Fridman (37:28.420)
anybody really that's creating impact at scale
Greg Brockman (37:35.020)
is how do we avoid, as your charter says,
Lex Fridman (37:37.680)
avoid enabling the use of AI or AGI
Lex Fridman (37:40.100)
to unduly concentrate power?
Lex Fridman (37:43.660)
Why would not a company like OpenAI
Lex Fridman (37:45.940)
keep all the power of an AGI system to itself?
Lex Fridman (37:48.660)
The charter.
Greg Brockman (37:49.540)
The charter.
Lex Fridman (37:50.380)
So how does the charter
Lex Fridman (37:53.140)
actualize itself in day to day?
Lex Fridman (37:57.260)
So I think that first, to zoom out,
Greg Brockman (38:00.580)
that the way that we structure the company
Lex Fridman (38:01.860)
is so that the power for sort of dictating the actions
Greg Brockman (38:05.560)
that OpenAI takes ultimately rests with the board,
Lex Fridman (38:08.600)
the board of the nonprofit.
Lex Fridman (38:11.020)
And the board is set up in certain ways
Lex Fridman (38:12.300)
with certain restrictions that you can read about
Greg Brockman (38:14.260)
in the OpenAI LP blog post.
Lex Fridman (38:16.300)
But effectively the board is the governing body
Greg Brockman (38:19.220)
for OpenAI LP.
Lex Fridman (38:21.260)
And the board has a duty to fulfill the mission
Greg Brockman (38:24.440)
of the nonprofit.
Lex Fridman (38:26.420)
And so that's kind of how we tie,
Lex Fridman (38:28.820)
how we thread all these things together.
Lex Fridman (38:30.980)
Now there's a question of, so day to day,
Lex Fridman (38:32.900)
how do people, the individuals,
Lex Fridman (38:34.820)
who in some ways are the most empowered ones, right?
Greg Brockman (38:36.980)
Now the board sort of gets to call the shots
Lex Fridman (38:38.820)
at the high level, but the people
Lex Fridman (38:40.540)
who are actually executing are the employees, right?
Lex Fridman (38:43.140)
People here on a day to day basis
Greg Brockman (38:44.820)
who have the keys to the technical whole kingdom.
Lex Fridman (38:48.940)
And there I think that the answer looks a lot like,
Lex Fridman (38:51.700)
well, how does any company's values get actualized, right?
Lex Fridman (38:55.080)
And I think that a lot of that comes down to
Greg Brockman (38:56.680)
that you need people who are here
Lex Fridman (38:58.120)
because they really believe in that mission
Lex Fridman (39:01.300)
and they believe in the charter
Lex Fridman (39:02.780)
and that they are willing to take actions
Greg Brockman (39:05.420)
that maybe are worse for them,
Lex Fridman (39:07.060)
but are better for the charter.
Lex Fridman (39:08.580)
And that's something that's really baked into the culture.
Lex Fridman (39:11.420)
And honestly, I think it's, you know,
Greg Brockman (39:13.180)
I think that that's one of the things
Lex Fridman (39:14.540)
that we really have to work to preserve as time goes on.
Lex Fridman (39:18.140)
And that's a really important part
Lex Fridman (39:19.740)
of how we think about hiring people
Lex Fridman (39:21.620)
and bringing people into OpenAI.
Lex Fridman (39:23.020)
So there's people here, there's people here
Greg Brockman (39:25.280)
who could speak up and say, like, hold on a second,
Lex Fridman (39:30.820)
this is totally against what we stand for, culture wise.
Greg Brockman (39:34.540)
Yeah, yeah, for sure.
Lex Fridman (39:35.380)
I mean, I think that we actually have,
Greg Brockman (39:37.060)
I think that's like a pretty important part
Lex Fridman (39:38.720)
of how we operate and how we have,
Greg Brockman (39:41.900)
even again with designing the charter
Lex Fridman (39:44.180)
and designing OpenAI LP in the first place,
Greg Brockman (39:46.700)
that there has been a lot of conversation
Lex Fridman (39:48.740)
with employees here and a lot of times
Greg Brockman (39:50.500)
where employees said, wait a second,
Lex Fridman (39:52.400)
this seems like it's going in the wrong direction
Lex Fridman (39:53.940)
and let's talk about it.
Lex Fridman (39:55.140)
And so I think one thing that's I think a really,
Lex Fridman (39:57.380)
and you know, here's actually one thing
Lex Fridman (39:58.900)
that I think is very unique about us as a small company,
Greg Brockman (40:02.140)
is that if you're at a massive tech giant,
Lex Fridman (40:04.400)
that's a little bit hard for someone
Greg Brockman (40:05.720)
who's a line employee to go and talk to the CEO
Lex Fridman (40:08.140)
and say, I think that we're doing this wrong.
Lex Fridman (40:10.900)
And you know, you'll get companies like Google
Lex Fridman (40:13.060)
that have had some collective action from employees
Greg Brockman (40:15.740)
to make ethical change around things like Maven.
Lex Fridman (40:19.420)
And so maybe there are mechanisms
Greg Brockman (40:20.700)
at other companies that work.
Lex Fridman (40:22.260)
But here, super easy for anyone to pull me aside,
Greg Brockman (40:24.500)
to pull Sam aside, to pull Ilya aside,
Lex Fridman (40:26.340)
and people do it all the time.
Greg Brockman (40:27.780)
One of the interesting things in the charter
Lex Fridman (40:29.820)
is this idea that it'd be great
Greg Brockman (40:31.660)
if you could try to describe or untangle
Lex Fridman (40:34.260)
switching from competition to collaboration
Greg Brockman (40:36.460)
in late stage AGI development.
Lex Fridman (40:38.820)
It's really interesting,
Greg Brockman (40:39.780)
this dance between competition and collaboration.
Lex Fridman (40:42.180)
How do you think about that?
Greg Brockman (40:43.420)
Yeah, assuming that you can actually do
Lex Fridman (40:45.020)
the technical side of AGI development,
Greg Brockman (40:47.060)
I think there's going to be two key problems
Lex Fridman (40:48.980)
with figuring out how do you actually deploy it,
Greg Brockman (40:50.460)
make it go well.
Lex Fridman (40:51.540)
The first one of these is the run up
Greg Brockman (40:53.180)
to building the first AGI.
Lex Fridman (40:56.380)
You look at how self driving cars are being developed,
Lex Fridman (40:58.940)
and it's a competitive race.
Lex Fridman (41:00.700)
And the thing that always happens in competitive race
Greg Brockman (41:02.580)
is that you have huge amounts of pressure
Lex Fridman (41:04.200)
to get rid of safety.
Lex Fridman (41:06.700)
And so that's one thing we're very concerned about,
Lex Fridman (41:08.940)
is that people, multiple teams figuring out
Greg Brockman (41:12.020)
we can actually get there,
Lex Fridman (41:13.620)
but if we took the slower path
Greg Brockman (41:16.740)
that is more guaranteed to be safe, we will lose.
Lex Fridman (41:20.300)
And so we're going to take the fast path.
Lex Fridman (41:22.380)
And so the more that we can both ourselves
Lex Fridman (41:25.520)
be in a position where we don't generate
Greg Brockman (41:27.300)
that competitive race, where we say,
Lex Fridman (41:29.040)
if the race is being run and that someone else
Greg Brockman (41:31.540)
is further ahead than we are,
Lex Fridman (41:33.340)
we're not going to try to leapfrog.
Lex Fridman (41:35.640)
We're going to actually work with them, right?
Lex Fridman (41:37.220)
We will help them succeed.
Greg Brockman (41:38.840)
As long as what they're trying to do
Lex Fridman (41:40.460)
is to fulfill our mission, then we're good.
Greg Brockman (41:42.940)
We don't have to build AGI ourselves.
Lex Fridman (41:44.860)
And I think that's a really important commitment from us,
Lex Fridman (41:47.100)
but it can't just be unilateral, right?
Lex Fridman (41:49.100)
I think that it's really important that other players
Greg Brockman (41:51.420)
who are serious about building AGI
Lex Fridman (41:53.140)
make similar commitments, right?
Greg Brockman (41:54.700)
I think that, again, to the extent that everyone believes
Lex Fridman (41:57.820)
that AGI should be something to benefit everyone,
Greg Brockman (42:00.060)
then it actually really shouldn't matter
Lex Fridman (42:01.220)
which company builds it.
Lex Fridman (42:02.460)
And we should all be concerned about the case
Lex Fridman (42:04.140)
where we just race so hard to get there
Greg Brockman (42:06.060)
that something goes wrong.
Lex Fridman (42:07.620)
So what role do you think government,
Greg Brockman (42:10.540)
our favorite entity, has in setting policy and rules
Lex Fridman (42:13.820)
about this domain, from research to the development
Lex Fridman (42:18.300)
to early stage to late stage AI and AGI development?
Lex Fridman (42:22.900)
So I think that, first of all,
Lex Fridman (42:25.660)
it's really important that government's in there, right?
Lex Fridman (42:28.100)
In some way, shape, or form.
Greg Brockman (42:29.820)
At the end of the day, we're talking about
Lex Fridman (42:30.940)
building technology that will shape how the world operates,
Lex Fridman (42:35.140)
and that there needs to be government
Lex Fridman (42:37.300)
as part of that answer.
Lex Fridman (42:39.100)
And so that's why we've done a number
Lex Fridman (42:42.220)
of different congressional testimonies,
Greg Brockman (42:43.660)
we interact with a number of different lawmakers,
Lex Fridman (42:46.300)
and that right now, a lot of our message to them
Greg Brockman (42:50.060)
is that it's not the time for regulation,
Lex Fridman (42:54.380)
it is the time for measurement, right?
Greg Brockman (42:56.420)
That our main policy recommendation is that people,
Lex Fridman (42:59.100)
and the government does this all the time
Greg Brockman (43:00.700)
with bodies like NIST, spend time trying to figure out
Lex Fridman (43:04.900)
just where the technology is, how fast it's moving,
Lex Fridman (43:07.940)
and can really become literate and up to speed
Lex Fridman (43:11.220)
with respect to what to expect.
Lex Fridman (43:13.500)
So I think that today, the answer really
Lex Fridman (43:15.260)
is about measurement, and I think that there will be a time
Lex Fridman (43:19.260)
and place where that will change.
Lex Fridman (43:21.740)
And I think it's a little bit hard to predict
Greg Brockman (43:23.820)
exactly what exactly that trajectory should look like.
Lex Fridman (43:27.140)
So there will be a point at which regulation,
Greg Brockman (43:31.060)
federal in the United States, the government steps in
Lex Fridman (43:34.220)
and helps be the, I don't wanna say the adult in the room,
Greg Brockman (43:39.500)
to make sure that there is strict rules,
Lex Fridman (43:42.420)
maybe conservative rules that nobody can cross.
Greg Brockman (43:45.260)
Well, I think there's kind of maybe two angles to it.
Lex Fridman (43:47.440)
So today, with narrow AI applications
Greg Brockman (43:49.820)
that I think there are already existing bodies
Lex Fridman (43:51.980)
that are responsible and should be responsible
Greg Brockman (43:53.980)
for regulation, you think about, for example,
Lex Fridman (43:55.880)
with self driving cars, that you want the national highway.
Greg Brockman (44:00.340)
Netsa.
Lex Fridman (44:01.180)
Yeah, exactly, to be regulating that.
Greg Brockman (44:02.980)
That makes sense, right, that basically what we're saying
Lex Fridman (44:04.980)
is that we're going to have these technological systems
Greg Brockman (44:08.160)
that are going to be performing applications
Lex Fridman (44:10.640)
that humans already do, great.
Greg Brockman (44:12.740)
We already have ways of thinking about standards
Lex Fridman (44:14.820)
and safety for those.
Lex Fridman (44:16.140)
So I think actually empowering those regulators today
Lex Fridman (44:18.860)
is also pretty important.
Lex Fridman (44:20.020)
And then I think for AGI, that there's going to be a point
Lex Fridman (44:24.780)
where we'll have better answers.
Lex Fridman (44:26.000)
And I think that maybe a similar approach
Lex Fridman (44:27.580)
of first measurement and start thinking about
Lex Fridman (44:30.500)
what the rules should be.
Lex Fridman (44:31.620)
I think it's really important
Greg Brockman (44:32.580)
that we don't prematurely squash progress.
Lex Fridman (44:36.260)
I think it's very easy to kind of smother a budding field.
Lex Fridman (44:40.140)
And I think that's something to really avoid.
Lex Fridman (44:42.120)
But I don't think that the right way of doing it
Greg Brockman (44:43.740)
is to say, let's just try to blaze ahead
Lex Fridman (44:46.900)
and not involve all these other stakeholders.
Lex Fridman (44:50.260)
So you recently released a paper on GPT2 language modeling,
Lex Fridman (44:58.820)
but did not release the full model
Greg Brockman (45:02.020)
because you had concerns about the possible
Lex Fridman (45:04.380)
negative effects of the availability of such model.
Greg Brockman (45:07.480)
It's outside of just that decision,
Lex Fridman (45:10.700)
it's super interesting because of the discussion
Greg Brockman (45:14.340)
at a societal level, the discourse it creates.
Lex Fridman (45:16.980)
So it's fascinating in that aspect.
Lex Fridman (45:19.260)
But if you think that's the specifics here at first,
Lex Fridman (45:22.860)
what are some negative effects that you envisioned?
Lex Fridman (45:25.860)
And of course, what are some of the positive effects?
Lex Fridman (45:28.540)
Yeah, so again, I think to zoom out,
Greg Brockman (45:30.780)
the way that we thought about GPT2
Lex Fridman (45:33.980)
is that with language modeling,
Greg Brockman (45:35.760)
we are clearly on a trajectory right now
Lex Fridman (45:38.520)
where we scale up our models
Lex Fridman (45:40.860)
and we get qualitatively better performance.
Lex Fridman (45:44.440)
GPT2 itself was actually just a scale up
Greg Brockman (45:47.340)
of a model that we've released in the previous June.
Lex Fridman (45:50.660)
We just ran it at much larger scale
Lex Fridman (45:52.860)
and we got these results where
Lex Fridman (45:54.300)
suddenly starting to write coherent pros,
Greg Brockman (45:57.020)
which was not something we'd seen previously.
Lex Fridman (46:00.020)
And what are we doing now?
Greg Brockman (46:01.300)
Well, we're gonna scale up GPT2 by 10x, by 100x, by 1000x,
Lex Fridman (46:05.740)
and we don't know what we're gonna get.
Lex Fridman (46:07.820)
And so it's very clear that the model
Lex Fridman (46:10.080)
that we released last June,
Greg Brockman (46:12.820)
I think it's kind of like, it's a good academic toy.
Lex Fridman (46:16.420)
It's not something that we think is something
Greg Brockman (46:18.900)
that can really have negative applications
Lex Fridman (46:20.420)
or to the extent that it can,
Greg Brockman (46:21.660)
that the positive of people being able to play with it
Lex Fridman (46:24.340)
is far outweighs the possible harms.
Greg Brockman (46:28.260)
You fast forward to not GPT2, but GPT20,
Lex Fridman (46:32.580)
and you think about what that's gonna be like.
Lex Fridman (46:34.680)
And I think that the capabilities are going to be substantive.
Lex Fridman (46:38.180)
And so there needs to be a point in between the two
Greg Brockman (46:41.100)
where you say, this is something
Lex Fridman (46:43.460)
where we are drawing the line
Lex Fridman (46:45.140)
and that we need to start thinking about the safety aspects.
Lex Fridman (46:47.940)
And I think for GPT2, we could have gone either way.
Lex Fridman (46:50.140)
And in fact, when we had conversations internally
Lex Fridman (46:52.700)
that we had a bunch of pros and cons,
Lex Fridman (46:54.740)
and it wasn't clear which one outweighed the other.
Lex Fridman (46:58.140)
And I think that when we announced that,
Greg Brockman (46:59.940)
hey, we decide not to release this model,
Lex Fridman (47:02.140)
then there was a bunch of conversation
Greg Brockman (47:03.560)
where various people said,
Lex Fridman (47:04.420)
it's so obvious that you should have just released it.
Greg Brockman (47:06.340)
There are other people said,
Lex Fridman (47:07.180)
it's so obvious you should not have released it.
Lex Fridman (47:08.820)
And I think that that almost definitionally means
Lex Fridman (47:10.940)
that holding it back was the correct decision.
Greg Brockman (47:13.580)
Right, if it's not obvious
Lex Fridman (47:15.900)
whether something is beneficial or not,
Greg Brockman (47:17.620)
you should probably default to caution.
Lex Fridman (47:19.700)
And so I think that the overall landscape
Greg Brockman (47:22.420)
for how we think about it
Lex Fridman (47:23.700)
is that this decision could have gone either way.
Greg Brockman (47:25.900)
There are great arguments in both directions,
Lex Fridman (47:27.940)
but for future models down the road
Lex Fridman (47:30.060)
and possibly sooner than you'd expect,
Lex Fridman (47:32.300)
because scaling these things up
Greg Brockman (47:33.460)
doesn't actually take that long,
Lex Fridman (47:35.660)
those ones you're definitely not going to want
Greg Brockman (47:37.900)
to release into the wild.
Lex Fridman (47:39.560)
And so I think that we almost view this as a test case
Lex Fridman (47:42.600)
and to see, can we even design,
Lex Fridman (47:45.140)
you know, how do you have a society
Greg Brockman (47:46.580)
or how do you have a system
Lex Fridman (47:47.940)
that goes from having no concept
Greg Brockman (47:49.220)
of responsible disclosure,
Lex Fridman (47:50.500)
where the mere idea of not releasing something
Greg Brockman (47:53.400)
for safety reasons is unfamiliar
Lex Fridman (47:55.940)
to a world where you say, okay, we have a powerful model,
Greg Brockman (47:58.680)
let's at least think about it,
Lex Fridman (47:59.660)
let's go through some process.
Lex Fridman (48:01.220)
And you think about the security community,
Lex Fridman (48:02.660)
it took them a long time
Lex Fridman (48:03.860)
to design responsible disclosure, right?
Lex Fridman (48:05.660)
You know, you think about this question of,
Greg Brockman (48:07.160)
well, I have a security exploit,
Lex Fridman (48:08.740)
I send it to the company,
Greg Brockman (48:09.720)
the company is like, tries to prosecute me
Lex Fridman (48:11.980)
or just sit, just ignores it, what do I do, right?
Lex Fridman (48:16.020)
And so, you know, the alternatives of,
Lex Fridman (48:17.300)
oh, I just always publish your exploits,
Lex Fridman (48:19.060)
that doesn't seem good either, right?
Lex Fridman (48:20.180)
And so it really took a long time
Lex Fridman (48:21.580)
and took this, it was bigger than any individual, right?
Lex Fridman (48:25.300)
It's really about building a whole community
Greg Brockman (48:27.060)
that believe that, okay, we'll have this process
Lex Fridman (48:28.740)
where you send it to the company, you know,
Greg Brockman (48:30.140)
if they don't act in a certain time,
Lex Fridman (48:31.660)
then you can go public and you're not a bad person,
Greg Brockman (48:34.420)
you've done the right thing.
Lex Fridman (48:36.220)
And I think that in AI,
Greg Brockman (48:38.620)
part of the response at GPT2 just proves
Lex Fridman (48:41.380)
that we don't have any concept of this.
Lex Fridman (48:44.140)
So that's the high level picture.
Lex Fridman (48:47.060)
And so I think that,
Greg Brockman (48:48.660)
I think this was a really important move to make
Lex Fridman (48:51.220)
and we could have maybe delayed it for GPT3,
Lex Fridman (48:53.980)
but I'm really glad we did it for GPT2.
Lex Fridman (48:56.020)
And so now you look at GPT2 itself
Lex Fridman (48:57.740)
and you think about the substance of, okay,
Lex Fridman (48:59.420)
what are potential negative applications?
Lex Fridman (49:01.300)
So you have this model that's been trained on the internet,
Lex Fridman (49:04.100)
which, you know, it's also going to be
Greg Brockman (49:05.340)
a bunch of very biased data,
Lex Fridman (49:06.500)
a bunch of, you know, very offensive content in there,
Lex Fridman (49:09.580)
and you can ask it to generate content for you
Lex Fridman (49:13.180)
on basically any topic, right?
Greg Brockman (49:14.540)
You just give it a prompt and it'll just start writing
Lex Fridman (49:16.700)
and it writes content like you see on the internet,
Greg Brockman (49:19.060)
you know, even down to like saying advertisement
Lex Fridman (49:21.820)
in the middle of some of its generations.
Lex Fridman (49:24.140)
And you think about the possibilities
Lex Fridman (49:26.140)
for generating fake news or abusive content.
Greg Brockman (49:29.220)
And, you know, it's interesting seeing
Lex Fridman (49:30.300)
what people have done with, you know,
Greg Brockman (49:31.820)
we released a smaller version of GPT2
Lex Fridman (49:34.340)
and the people have done things like try to generate,
Greg Brockman (49:37.460)
you know, take my own Facebook message history
Lex Fridman (49:40.700)
and generate more Facebook messages like me
Lex Fridman (49:43.340)
and people generating fake politician content
Lex Fridman (49:47.340)
or, you know, there's a bunch of things there
Greg Brockman (49:49.500)
where you at least have to think,
Lex Fridman (49:51.860)
is this going to be good for the world?
Greg Brockman (49:54.700)
There's the flip side, which is I think
Lex Fridman (49:56.300)
that there's a lot of awesome applications
Greg Brockman (49:57.780)
that we really want to see,
Lex Fridman (49:59.340)
like creative applications in terms of
Greg Brockman (50:02.380)
if you have sci fi authors that can work with this tool
Lex Fridman (50:05.340)
and come up with cool ideas, like that seems awesome
Greg Brockman (50:08.580)
if we can write better sci fi through the use of these tools
Lex Fridman (50:11.340)
and we've actually had a bunch of people write into us
Greg Brockman (50:13.020)
asking, hey, can we use it for, you know,
Lex Fridman (50:16.060)
a variety of different creative applications?
Lex Fridman (50:18.300)
So the positive are actually pretty easy to imagine.
Lex Fridman (50:21.780)
They're, you know, the usual NLP applications
Greg Brockman (50:26.820)
are really interesting, but let's go there.
Lex Fridman (50:30.860)
It's kind of interesting to think about a world
Greg Brockman (50:32.860)
where, look at Twitter, where not just fake news,
Lex Fridman (50:37.860)
but smarter and smarter bots being able to spread
Greg Brockman (50:42.980)
in an interesting, complex, networking way information
Lex Fridman (50:47.300)
that just floods out us regular human beings
Greg Brockman (50:50.700)
with our original thoughts.
Lex Fridman (50:52.780)
So what are your views of this world with GPT20, right?
Lex Fridman (51:00.180)
How do we think about it?
Lex Fridman (51:01.220)
Again, it's like one of those things about in the 50s
Greg Brockman (51:03.540)
trying to describe the internet or the smartphone.
Lex Fridman (51:08.700)
What do you think about that world,
Lex Fridman (51:09.940)
the nature of information?
Lex Fridman (51:12.900)
One possibility is that we'll always try to design systems
Greg Brockman (51:16.780)
that identify robot versus human
Lex Fridman (51:19.660)
and we'll do so successfully and so we'll authenticate
Greg Brockman (51:23.340)
that we're still human and the other world is that
Lex Fridman (51:25.700)
we just accept the fact that we're swimming in a sea
Greg Brockman (51:29.020)
of fake news and just learn to swim there.
Lex Fridman (51:32.220)
Well, have you ever seen the popular meme of robot
Greg Brockman (51:39.860)
with a physical arm and pen clicking the
Lex Fridman (51:42.020)
I'm not a robot button?
Greg Brockman (51:43.460)
Yeah.
Lex Fridman (51:44.300)
I think the truth is that really trying to distinguish
Greg Brockman (51:48.620)
between robot and human is a losing battle.
Lex Fridman (51:52.200)
Ultimately, you think it's a losing battle?
Lex Fridman (51:53.860)
I think it's a losing battle ultimately, right?
Lex Fridman (51:55.560)
I think that that is, in terms of the content,
Greg Brockman (51:57.820)
in terms of the actions that you can take.
Lex Fridman (51:59.380)
I mean, think about how captures have gone, right?
Greg Brockman (52:01.220)
The captures used to be a very nice, simple,
Lex Fridman (52:02.980)
you just have this image, all of our OCR is terrible,
Greg Brockman (52:06.340)
you put a couple of artifacts in it,
Lex Fridman (52:08.900)
humans are gonna be able to tell what it is.
Greg Brockman (52:11.500)
An AI system wouldn't be able to.
Lex Fridman (52:13.300)
Today, I could barely do captures.
Lex Fridman (52:15.740)
And I think that this is just kind of where we're going.
Lex Fridman (52:18.380)
I think captures were a moment in time thing
Lex Fridman (52:20.420)
and as AI systems become more powerful,
Lex Fridman (52:22.500)
that there being human capabilities that can be measured
Greg Brockman (52:25.500)
in a very easy, automated way that AIs
Lex Fridman (52:28.900)
will not be capable of.
Greg Brockman (52:30.180)
I think that's just like,
Lex Fridman (52:31.140)
it's just an increasingly hard technical battle.
Lex Fridman (52:34.180)
But it's not that all hope is lost, right?
Lex Fridman (52:36.260)
You think about how do we already authenticate ourselves,
Greg Brockman (52:40.360)
right, that we have systems, we have social security numbers
Lex Fridman (52:43.460)
if you're in the US or you have ways of identifying
Greg Brockman (52:47.700)
individual people and having real world identity
Lex Fridman (52:50.180)
tied to digital identity seems like a step
Greg Brockman (52:53.060)
towards authenticating the source of content
Lex Fridman (52:56.220)
rather than the content itself.
Greg Brockman (52:58.260)
Now, there are problems with that.
Lex Fridman (52:59.980)
How can you have privacy and anonymity
Greg Brockman (53:02.340)
in a world where the only content you can really trust is,
Lex Fridman (53:05.460)
or the only way you can trust content
Lex Fridman (53:06.580)
is by looking at where it comes from?
Lex Fridman (53:08.560)
And so I think that building out good reputation networks
Greg Brockman (53:11.420)
may be one possible solution.
Lex Fridman (53:14.060)
But yeah, I think that this question is not an obvious one.
Lex Fridman (53:17.700)
And I think that we, maybe sooner than we think,
Lex Fridman (53:20.220)
will be in a world where today I often will read a tweet
Lex Fridman (53:23.820)
and be like, hmm, do I feel like a real human wrote this?
Lex Fridman (53:25.980)
Or do I feel like this is genuine?
Greg Brockman (53:27.560)
I feel like I can kind of judge the content a little bit.
Lex Fridman (53:30.180)
And I think in the future, it just won't be the case.
Greg Brockman (53:32.640)
You look at, for example, the FCC comments on net neutrality.
Lex Fridman (53:36.900)
It came out later that millions of those were auto generated
Lex Fridman (53:39.900)
and that the researchers were able to do
Lex Fridman (53:41.660)
various statistical techniques to do that.
Lex Fridman (53:44.040)
What do you do in a world
Lex Fridman (53:45.100)
where those statistical techniques don't exist?
Greg Brockman (53:47.720)
It's just impossible to tell the difference
Lex Fridman (53:49.180)
between humans and AIs.
Lex Fridman (53:50.660)
And in fact, the most persuasive arguments
Lex Fridman (53:53.980)
are written by AI.
Greg Brockman (53:56.620)
All that stuff, it's not sci fi anymore.
Lex Fridman (53:58.660)
You look at GPT2 making a great argument
Greg Brockman (54:00.580)
for why recycling is bad for the world.
Lex Fridman (54:02.580)
You gotta read that and be like, huh, you're right.
Greg Brockman (54:04.460)
We are addressing just the symptoms.
Lex Fridman (54:06.540)
Yeah, that's quite interesting.
Greg Brockman (54:08.140)
I mean, ultimately it boils down to the physical world
Lex Fridman (54:11.380)
being the last frontier of proving,
Lex Fridman (54:13.720)
so you said like basically networks of people,
Lex Fridman (54:16.100)
humans vouching for humans in the physical world.
Lex Fridman (54:19.420)
And somehow the authentication ends there.
Lex Fridman (54:22.980)
I mean, if I had to ask you,
Greg Brockman (54:25.560)
I mean, you're way too eloquent for a human.
Lex Fridman (54:28.180)
So if I had to ask you to authenticate,
Greg Brockman (54:31.260)
like prove how do I know you're not a robot
Lex Fridman (54:33.180)
and how do you know I'm not a robot?
Greg Brockman (54:34.940)
Yeah.
Lex Fridman (54:35.780)
I think that's so far where in this space,
Greg Brockman (54:40.540)
this conversation we just had,
Lex Fridman (54:42.140)
the physical movements we did,
Greg Brockman (54:44.020)
is the biggest gap between us and AI systems
Lex Fridman (54:47.060)
is the physical manipulation.
Lex Fridman (54:49.380)
So maybe that's the last frontier.
Lex Fridman (54:51.300)
Well, here's another question is why is,
Lex Fridman (54:55.020)
why is solving this problem important, right?
Lex Fridman (54:57.300)
Like what aspects are really important to us?
Lex Fridman (54:59.100)
And I think that probably where we'll end up
Lex Fridman (55:01.220)
is we'll hone in on what do we really want
Greg Brockman (55:03.620)
out of knowing if we're talking to a human.
Lex Fridman (55:06.420)
And I think that, again, this comes down to identity.
Lex Fridman (55:09.460)
And so I think that the internet of the future,
Lex Fridman (55:11.780)
I expect to be one that will have lots of agents out there
Greg Brockman (55:14.900)
that will interact with you.
Lex Fridman (55:16.380)
But I think that the question of is this
Greg Brockman (55:19.260)
flesh, real flesh and blood human
Lex Fridman (55:21.580)
or is this an automated system,
Greg Brockman (55:23.860)
may actually just be less important.
Lex Fridman (55:25.820)
Let's actually go there.
Greg Brockman (55:27.420)
It's GPT2 is impressive and let's look at GPT20.
Lex Fridman (55:32.500)
Why is it so bad that all my friends are GPT20?
Lex Fridman (55:37.500)
Why is it so important on the internet,
Lex Fridman (55:43.300)
do you think, to interact with only human beings?
Lex Fridman (55:47.340)
Why can't we live in a world where ideas can come
Lex Fridman (55:50.620)
from models trained on human data?
Greg Brockman (55:52.940)
Yeah, I think this is actually
Lex Fridman (55:54.820)
a really interesting question.
Greg Brockman (55:55.700)
This comes back to the how do you even picture a world
Lex Fridman (55:58.100)
with some new technology?
Lex Fridman (55:59.580)
And I think that one thing that I think is important
Lex Fridman (56:02.060)
is, you know, let's say honesty.
Lex Fridman (56:04.780)
And I think that if you have almost in the Turing test
Lex Fridman (56:07.820)
style sense of technology, you have AIs that are pretending
Greg Brockman (56:12.420)
to be humans and deceiving you.
Lex Fridman (56:14.100)
I think that feels like a bad thing, right?
Greg Brockman (56:17.300)
I think that it's really important that we feel like
Lex Fridman (56:19.460)
we're in control of our environment, right?
Greg Brockman (56:20.980)
That we understand who we're interacting with.
Lex Fridman (56:23.140)
And if it's an AI or a human, that's not something
Greg Brockman (56:27.060)
that we're being deceived about.
Lex Fridman (56:28.420)
But I think that the flip side of can I have as meaningful
Lex Fridman (56:31.220)
of an interaction with an AI as I can with a human?
Lex Fridman (56:33.980)
Well, I actually think here you can turn to sci fi.
Lex Fridman (56:36.620)
And her I think is a great example of asking
Lex Fridman (56:39.380)
this very question, right?
Greg Brockman (56:40.860)
One thing I really love about her is it really starts out
Lex Fridman (56:42.940)
almost by asking how meaningful
Lex Fridman (56:44.660)
are human virtual relationships, right?
Lex Fridman (56:47.020)
And then you have a human who has a relationship with an AI
Lex Fridman (56:50.940)
and that you really start to be drawn into that, right?
Lex Fridman (56:54.100)
That all of your emotional buttons get triggered
Greg Brockman (56:56.700)
in the same way as if there was a real human
Lex Fridman (56:58.260)
that was on the other side of that phone.
Lex Fridman (57:00.180)
And so I think that this is one way of thinking about it
Lex Fridman (57:03.540)
is that I think that we can have meaningful interactions
Lex Fridman (57:06.900)
and that if there's a funny joke,
Lex Fridman (57:09.500)
some sense it doesn't really matter
Greg Brockman (57:10.580)
if it was written by a human or an AI.
Lex Fridman (57:12.660)
But what you don't want and why I think
Greg Brockman (57:14.660)
we should really draw hard lines is deception.
Lex Fridman (57:17.100)
And I think that as long as we're in a world
Lex Fridman (57:19.340)
where why do we build AI systems at all, right?
Lex Fridman (57:22.420)
The reason we want to build them is to enhance human lives,
Greg Brockman (57:24.740)
to make humans be able to do more things,
Lex Fridman (57:26.420)
to have humans feel more fulfilled.
Lex Fridman (57:28.820)
And if we can build AI systems that do that, sign me up.
Lex Fridman (57:32.940)
So the process of language modeling,
Lex Fridman (57:36.860)
how far do you think it'd take us?
Lex Fridman (57:38.540)
Let's look at movie Her.
Lex Fridman (57:40.420)
Do you think a dialogue, natural language conversation
Lex Fridman (57:44.780)
is formulated by the Turing test, for example,
Lex Fridman (57:47.580)
do you think that process could be achieved
Lex Fridman (57:50.180)
through this kind of unsupervised language modeling?
Lex Fridman (57:52.900)
So I think the Turing test in its real form
Lex Fridman (57:56.700)
isn't just about language, right?
Lex Fridman (57:58.420)
It's really about reasoning too, right?
Lex Fridman (58:00.420)
To really pass the Turing test,
Greg Brockman (58:01.660)
I should be able to teach calculus
Lex Fridman (58:03.660)
to whoever's on the other side
Lex Fridman (58:05.340)
and have it really understand calculus
Lex Fridman (58:07.300)
and be able to go and solve new calculus problems.
Lex Fridman (58:11.100)
And so I think that to really solve the Turing test,
Lex Fridman (58:13.780)
we need more than what we're seeing with language models.
Greg Brockman (58:16.220)
We need some way of plugging in reasoning.
Lex Fridman (58:18.500)
Now, how different will that be from what we already do?
Lex Fridman (58:22.180)
That's an open question, right?
Lex Fridman (58:23.660)
Might be that we need some sequence
Greg Brockman (58:25.260)
of totally radical new ideas,
Lex Fridman (58:26.980)
or it might be that we just need to kind of shape
Greg Brockman (58:29.340)
our existing systems in a slightly different way.
Lex Fridman (58:32.740)
But I think that in terms of how far language modeling
Greg Brockman (58:35.020)
will go, it's already gone way further
Lex Fridman (58:37.260)
than many people would have expected, right?
Greg Brockman (58:39.460)
I think that things like,
Lex Fridman (58:40.700)
and I think there's a lot of really interesting angles
Greg Brockman (58:42.420)
to poke in terms of how much does GPT2
Lex Fridman (58:45.660)
understand physical world?
Greg Brockman (58:47.620)
Like, you read a little bit about fire underwater in GPT2.
Lex Fridman (58:52.060)
So it's like, okay, maybe it doesn't quite understand
Lex Fridman (58:53.900)
what these things are, but at the same time,
Lex Fridman (58:56.660)
I think that you also see various things
Greg Brockman (58:58.780)
like smoke coming from flame,
Lex Fridman (59:00.340)
and a bunch of these things that GPT2,
Greg Brockman (59:02.380)
it has no body, it has no physical experience,
Lex Fridman (59:04.580)
it's just statically read data.
Lex Fridman (59:06.980)
And I think that the answer is like, we don't know yet.
Lex Fridman (59:13.140)
These questions, though, we're starting to be able
Greg Brockman (59:15.020)
to actually ask them to physical systems,
Lex Fridman (59:17.300)
to real systems that exist, and that's very exciting.
Lex Fridman (59:19.580)
Do you think, what's your intuition?
Lex Fridman (59:20.860)
Do you think if you just scale language modeling,
Greg Brockman (59:25.220)
like significantly scale,
Lex Fridman (59:27.420)
that reasoning can emerge from the same exact mechanisms?
Greg Brockman (59:30.980)
I think it's unlikely that if we just scale GPT2
Lex Fridman (59:34.580)
that we'll have reasoning in the full fledged way.
Lex Fridman (59:38.260)
And I think that there's like,
Lex Fridman (59:39.420)
the type signature's a little bit wrong, right?
Greg Brockman (59:41.180)
That like, there's something we do with,
Lex Fridman (59:44.220)
that we call thinking, right?
Greg Brockman (59:45.460)
Where we spend a lot of compute,
Lex Fridman (59:47.300)
like a variable amount of compute,
Lex Fridman (59:48.820)
to get to better answers, right?
Lex Fridman (59:50.340)
I think a little bit harder, I get a better answer.
Lex Fridman (59:52.700)
And that that kind of type signature
Lex Fridman (59:54.860)
isn't quite encoded in a GPT, right?
Greg Brockman (59:58.620)
GPT will kind of like, it's been a long time,
🔗 相关节目