OpenAI의 혁신: 연구부터 비즈니스까지

Open AI의 COO인 브래드 라이트캡이 AGI 구축을 향한 회사의 발전과 사명에 대해 이야기합니다.

### OpenAI의 혁신, 사명 및 향후 AI 개발 계획
* 🚀  브래드 라이트캡이 초기  AI의 노력과 최신 AI의 출현에 초점을 맞춰 **OpenAI를 연구위원회에서 비즈니스로 전환**하는 과정에 대해 설명합니다.
* 🤖  그는  AGI의 영향력에 대한 우려에도 불구하고 **AGI를 구축**하고 **안전성과 광범위한 혜택**을 제공하려는 **OpenAI의 사명**을 설명합니다.
* 🌟   Brad는 AI 시스템의 향후 발전을 예상하며, **단순한 챗봇을 넘어 더 복잡한 작업을 수행**할 수 있는 능력과 잠재력을 강조합니다.
* 💼  그는 기업에서  AI의 다양한 사용 사례와 그 놀라운 응용 분야에 대해 설명하며, **반복적인 배포와 경제에 미치는 영향**을 강조합니다.
* 📰   Brad가 미디어와의 파트너십과 AI를 통해 퍼블리싱 경험을 향상시키려는 **OpenAI의 비전**, 그리고 비디오 제작 개선을 위한 **할리우드와의 논의**에 대해 이야기합니다.
* 🎬  그는  AI 모델에 대한 **학습 데이터와 크리에이터와의 새로운 사회적 계약**의 필요성, 그리고 **AI 기술 발전을 위한 OpenAI와 Microsoft의 중요한 파트너십**에 대해 설명합니다.

Please welcome to the stage Brad Lightcap, Chief Operating Officer at Open I, for a conversation with Bloomberg's Shery Ahn Ghaffari. Hello, everyone. Today we're joined on stage by someone who I have called Sam Altman. He's the man behind the scenes who's in charge of running and growing Open I's business. So with that, I want to just thank you, Brad, for being here today. Thank you for having me. Yeah, absolutely. We also will have a poll for the audience. I want you all to please vote on that, and we'll hopefully have some time to come back to it. So to start us off, Brad, when you joined Open I originally as the CFO in the relatively early days in 2018, there was really no product in Open I, let alone a business model. So can you tell us a little bit about how you've helped transform the company from what you called back in the day, you know, a sleepy research board into an

actual business?
Yeah, well, so I joined back in 2018, which feels like ancient history at this
point. We weren't even in the era at that point
of what now is what we call modern A.I., which is this transformer-based
architecture that underpins the language models that we use every day, like
chatbots. That was completely off our radar.
We were focused on a very different type of AI called reinforcement learning,
where you were training these agents to basically be superhuman at any given
task. And so we were applying that to try and
beat humans at video games. We actually had a project where we had
an agent play the best Dota players in the world, if anyone's familiar with
Dota, and we were taking the same principles and doing research into
robotics to see if we could actually teach a robot hand how to manipulate a
Rubik's Cube. And so I was sitting here in 2018 trying
to figure out what exactly would our

business. And I was like, I guess we'll be somewhere between a gaming company and a robotics company. But and Reid actually was very helpful in us in helping us think through this early on. The real read, not the fake read. And and so we we you know, we ultimately were kind of like looking at that, but it felt really early still. And then when Transformers came out and language model started working, it really felt early on like that was going to be the thing that was was important, make sense and, you know, Open our Eyes mission is to build AGI or artificial general intelligence that could one day potentially surpass our own mere human intelligence. What do you say to the many people, including prospective customers, who are still pretty freaked out by that, right? Yeah. Look, you know, AGI is obviously an ambitious goal. I don't think anyone, by the way, has a

perfect definition of what AGI is.
We joke sometimes, like, maybe we'll just know it when we see it, but we try
and define it as like a system that's basically generally capable of doing
most tasks. So if you pointed at some arbitrarily
complex or hard task, it can kind of go off and figure out how to do it.
It can reason about what tools it needs, what data it needs, how to ask you
follow-up questions, and the way that humans kind of have this like general
reasoning ability. We think that A.I.
systems will have that at some point too.
And it is scary, right? It is.
You know, having the idea that you'd have a system that has some level of
agency and autonomy out in the world that's able to make decisions and able
to use tools and access the Internet and write code.
That is a really, really big, scary thing.
And so when we formed Open AI, we really

formed it as this company that I think
we wanted to have really be this very mission-oriented
company where the ideas of safety and broad benefit were very core to how we
operated. We tried to come up with a structure
that really reflected that actually at its core so that we didn't have to put
things, you know, prioritize things that we thought would be unsafe ahead of
things that would be safe. And so we try our best to understand how
people use the technology deployed at early scale.
So knowing how people use a system like GTD for actually does help us think
about how the next thing might be used, and we can, you know, refine our safety
systems and techniques. And I think if we take that iterative
approach and are able to have a kind of global conversation with people about
what they want this technology to do, both in terms of the new things, but
also the areas where it needs to be safeguarded.
We'll get there and I think be very

happy with the outcome.
So let's start with a topic that's very hot on everyone's minds.
When can we expect the next release? Will we be seeing GPT five soon?
What will it look like? That's my time.
Thanks so much. No, certainly the direction.
What do you you've said that, you know, Sam has said that AI now is laughably
bad compared to what it will be, you know, or sorry.
He said it was the dumbest model people will ever have to use is 54.
You have said that today's AI will look laughably bad in a year and then I will
be used for more complex work. So what is that?
Walk me through how this is going to get even smarter, more complex, better.
Yeah. So, look, we're we think the systems are
going to get better. And to say that today systems are the
dumbest will ever be is true in so far

as the next thing we think will definitely be better. I think we're going to have to understand what that means. But really also is, I think, a comment not only qualifying the next system, but is also kind of qualifying the current systems as these are systems that aren't that capable. If you think about it right, these are systems that you can ask questions of. They have this kind of oracle-like interface. This you they kind of give you an answer and that's kind of the extent of it. And they can they can mimic some level of intelligence, of emotional intelligence that can be funny if you kind of ask me to be funny. But like I've heard someone describe them as like they're kind of like in many ways the world's best parlor trick and now that's in practice ended up not being quite right. Good for actually has amazing applications in the enterprise very much to our surprise and we can come back to

that one if you like.
But we do think the next systems will have to cross a
kind of utility and capability gap of being able to go out in the world and be
useful and help you actually do things like be real, you know, truly assistive
to you. For us to kind of move toward that goal
of AGI, would you say it's going to be more of an agent on your behalf in the
future rather than just a chat bot, or there are other formats in which will
use your technology? You know, I never quite think we have
the vocabulary to describe these things until we have them.
And you know, the vocabulary tends to somewhat wrap itself around the current
instantiation of the technology. So the idea, for example, is like a
prompt engineer adds interesting question to think will there be such
thing as a prompt engineer in 2026 or 2027 where you're trying to fiddle with
this prompt to get it quite right?

You don't prompt engineer your friends.
You might prompt engineer your kids, but like you and I don't have to like sit
here and trade, you know, prompt drafts to get each other to understand each
other. And so, you know, that feels like
something that is kind of specific to the era.
I think, you know, the same way we'll look at the next system's the language
we use to describe what they're capable of, will be specific to their
capabilities and their gaps. Make sense to actually be broke.
Records for being one of the fastest growing consumer apps of all time when
it came out. That's kind of hard to beat even
internally. Right? So do you think is the novelty wearing
off for some people and how do you continue growth when you had such this
viral start? Yeah, we really take a kind of a
research-first perspective to this, which is, you know, if we were just
building something that was called GBG.

four, and that was the best anything was
ever going to be, then I think we'd be sitting here saying, okay, great, let's
really go figure out how to maximally optimize the product.
That is the good for scale product. We care a lot more about learning and
how people use the technology. One of the exercises we tried to do with
Jackie Beauty early on was like we tried to in classic like, you know, business
operations style, we tried to like ring fence all the use cases and understand
like, okay, is there like one common use case that we see that we've identified
where we can be like, okay, great, we'd better inform kind of the next version
of the product that we build because we'll make it better at that thing and
the like. Number of people that we talked to were
like, I don't have a use case. Like I use it like and one second to
help me like plan my kid's birthday.

party, and the next second, I use it to like, help me write code. In the next second, I use it to like, help my, you know, elderly parents like, navigate their, you know, their health care. You have this just diversity of use cases that I think is, it makes it hard really to know how to improve it other than to just make the model smarter and make it better. Right. And so you've been growing the enterprise business a lot, and I know there are some really exciting applications of it on the enterprise side. You know, I'm wondering now is a recent kind of announcement that you've made as a customer. We can talk a little bit more about that, but I also just want to get to there were a lot of early hopes and promises around the potential of AI when we saw this leap with with CBT. I think it's also fair to say that while there's a lot of promise for AI in the past year and a half since you actually

came out. We haven't seen it drastically reshape our economy or most people's day-to-day jobs just yet. What do you say to those who are skeptical about whether AI is going to be as impactful as maybe some have hyped it out to be?
Dan: Like we have a year to reshape the economy in tall order?
TikTok: No, I mean, so I said earlier, the enterprise adoption and the utility that these tools have in the enterprise is really shocked us. Think about that coming. We think that there is an amazing opportunity, even if we went home and opening, I kind of just like decided we were done with what we do what as a kind of student history of student technology. I'm fairly confident at this point that there'd be this like 10 to 20-year diffusion period through the economy of just four-level technology. And at some point we'll have a better system and we'll have to kind of shuttle everyone over to this next thing and

then be figuring out how to incorporate
that. And so,
we're really encouraged by that. I think the thing we study is, you know,
how do these systems really make differences in businesses and help
businesses better understand their customers, create more personalized
relationships with their users, allow them to create new product experiences,
allow them to actually do things with their data that they couldn't otherwise
do. These are all examples of ways that
people are using this version of the technology.
And so, you know, our job really is. We see it as I call it. I tell our team
we're kind of these like Sherpas, basically, of kind of helping people get
from like what is what's possible today and what's not possible today to
what's possible tomorrow and what might be possible tomorrow.
That's not possible today. You have been leading partnerships
right, in opening AI with your companies.
And that's a really interesting part of

your job, because a big sector that
you're partnering with right now is media.
And you know, most recently you've done some deals with the Financial Times.
DoorDash Meredith, which is owned by IAC.
And you're talking to dozens of other publications, right.
About how can media sort of partner integrate Chatty Betty into the product
and vice versa and license, right. That content.
How many more deals can we expect? Should I be asking my colleagues at the
business side of Bloomberg if they're working on a deal, too?
I mean, it's a joke, but, you know, like, where
do you see how how many publishers do you eventually see being part of the
chat show ecosystem? Yeah, Well, so we think that like this
technology A has amazing applications in that in that field.
And so if you kind of just boil it down to like first principles, the the
business of publishing is really the business of communicating information to
an audience and allowing that audience

to engage with a subject matter that they care about in a way that satisfies the thing that they want to or should know about that. And, you know, when you kind of say it like that, almost in kind of plain language like that, you realize actually that like these A.I. systems are actually almost tailor-made for being additive to that experience. How do you better understand a subject matter? How do you enhance a journalist's ability to report on something? How do you give people more access to data and being able to engage with data like we live in this super data-driven world. And so that's the opportunity that we see in publishing that we're excited about. We have a long way to go, I think, to build those tools still and start to work with the industry to get there. But we are committed to doing that work and that's really where I think the foundation of our partnership efforts come from.

And then also being able to just bring
what is a source of truth information into the chatbot experience for people.
The people somewhat have, I think, a little bit of a misguided sense of what
exactly these models can and should be used for.
They're really not databases. So, you know, if you were to use them
like a database to store information, like they're not 100% accurate and
they're like very expensive. Like we already built database
technology that's like way better. Like, these are really not meant to be
able to recall facts. They're really meant to reason about new
information. And so the more information you can give
them access to, and that's really how we think about our publisher engagements,
is bringing more information kind of into their field of view, the more
useful we think they are for people. And ultimately, like we think that's
good for the world. You've also recently been to Hollywood
and talking to, you know, studio heads.

and directors, and creative types about Zora. Write your text video generator. How those conversations have been going. I Goes to Hollywood, I think is a headline we ran with. Tell me what that's like. Yeah, really positive. So we put Sora out is our text or video model a few months ago, and we didn't release it. So interestingly, this was kind of an example of how we think about what we call iterative deployment, which is put something out in the world, let people start to kind of appreciate what it is, have a conversation about how it might be used. Also gives us an opportunity to get feedback on things that we can do to make it better and be able to just engage with people who will have real, you know, we're this is going to be really impactful. And so the feedback there was incredibly helpful. We learned a lot about what we didn't know, which is namely, how does video

content get produced for people, for creatives in the industry? What are the things they really care about that we can now go back and kind of think about how to incorporate into the model? It turns out actually a lot of like creatives, especially at that kind of higher end of film production, really, really care about things like is the camera angle like 5% higher or lower than where you want it, and they want to reshoot it with a slightly like, and we're like, okay, well, that's interesting research problem. And so having conversations like that, I think, is critical. And look, our hope is really like if you can bring the cost of production down, whether that's something feature length, which we're not close to, but even parts of a film, one of the things that was consistent in that feedback is you're going to see more stuff get made. And so there's a lot of things that don't get made because it gets put on

the shelf because it's too expensive.
And so if you can have a little bit of a kind of deflationary impact in the
industry, there's going to be more stuff that gets made, you know. Period pieces,
westerns, things like things like that that you just don't see anymore with.
And we think that's the promise of the technology.
So do you think we're months or years away from this being a real product that
you go out and sell to Hollywood or whoever?
Yeah, you know, we have what we have today.
We think we can make it way better. And so we're going to kind of look at it
as a research problem still, but it'll be collaborative, and we want to really
take creative input as we do that. Speaking of Sarah, there was a lot of
conversation, as I'm sure you've seen, about, you know, whether what training
data was used to train the model. Can you say and clear up once and for
all whether so I was trained on YouTube?

data.
Yeah. I mean, look, the conversation around
data is really important. We obviously need to know kind of where
data comes from. We just put out a post this week
actually about this exact topic, which is basically that there needs to be like
a content ID system for AI that lets creators understand as they create stuff
where it's going, who's training on it, being able
to opt in and out of training, being able to opt in and out of use can
also like on the other side of that, like being able to actively like allow
your content to be put into a model or to be accessed by a model because there
may be this other economic opportunity on the other end of this.
And that's something we're exploring too, which is like how do you actually
go create an entirely different social contract with the web, with creators,
with publishers, where as these models go off in the world and do things that
are useful, create value to the extent

they're able to like, reference and
incorporate content from the web. There should be some sort of way that
people can kind of get a good benefit from that.
So, yeah, we're like, we're looking at this problem.
It's really hard. We don't have all the answers yet.
Maybe by 2026, if you have any ideas, we'll take them.
But it is a hard one. So no answer on the YouTube for now.
For the last partnership I want to talk about is maybe your most important one,
which is with Microsoft. Write your investor or partner.
How do you deal with both? What are both the benefits and maybe the
challenges with having a partner who is also in what some may call a competition
with you in that they also are selling, right, an enterprise product similar to
your offerings? Yeah, well, they're an amazing partner.
Like there's not a lot of companies in

the world. I think that can build systems
at the scale that we build them with, with Microsoft, and also like put up with
us. Like we're very fast-moving, we're kind
of demanding, and they see it as an opportunity to help get make Microsoft
better at what they do from a systems-building perspective, and also just
understanding AI. We just think this market is ginormous.
And so the way we look at it is like we're a small company, we create
technology, we have a view of kind of how and where we want to create and
deploy that technology. They're also a company that has their
own set of products and set of customers.
And ultimately, like you mentioned earlier, like if we are going to have
any needle-moving impact in the economy, I don't think it will be just opening AI
that like as its own single independent company goes out and like pushes the
button to make that happen.

Working with partners is very critical
to who we are. And so from a go-to-market perspective,
which is where my focus is, we see them as an absolutely
critical vector in that, along with all the other partners we work with.
So it's been very positive, and I think
there's a lot more to come there. Great.
Well, we're out of time, but thank you so much.
Really appreciate it. Yeah.
Yeah.

OpenAI’s Lightcap on Business Applications for AI

Bloomberg Live uses the power of Bloomberg data and the reach and insight of Bloomberg Media to create news-making events and provide thought leadership to the leaders of the business world.

Bloomberg Live

브래드 라이트캡의 리더십 아래 오픈 I의 변화

원래 최고재무책임자(CFO)였던 브래드 라이트캡은 Open I를 연구위원회에서 AI에 중점을 둔 성공적인 비즈니스로 전환하는 데 핵심적인 역할을 담당했습니다.

초창기에는 게임과 로보틱스를 위한 강화 학습에 집중하다가 언어 모델을 위한 트랜스포머 기반 아키텍처로 전환했습니다.

AGI를 구축하려는 Open I의 사명은 우려를 불러일으켰지만, 이를 둘러싼 잠재적 우려에도 불구하고 일반적인 추론이 가능한 AI 시스템을 만드는 것이 목표입니다.

AI 기술은 안전과 글로벌 피드백을 우선시하는 방향으로 진화하고 있으며, 앞으로 더 많은 기능과 유용성을 갖추기 위해 노력하고 있습니다.

현재의 AI 시스템은 기능이 제한되어 있어 마치 마술처럼 작동하는 것으로 여겨지지만, 차세대에는 그 격차를 해소하고 진정한 보조 역할을 할 것으로 기대됩니다.

AI 기술의 미래 형식과 역할은 아직 불확실하며, 어휘와 새로운 상호작용 방식을 개발하는 데 중점을 두고 있습니다.

소비자 앱의 참신함은 입소문을 타기 시작한 후 지속적인 성장에 대한 도전에 직면해 있습니다.

기술의 다양한 사용 사례로 인해 개선이 어렵기 때문에 모델을 더 스마트하게 만들어야 할 필요성이 강조되고 있습니다.

기업 측면에서는 경제에 미치는 영향에 대한 회의론에도 불구하고 AI 도입이 큰 잠재력을 보여주면서 상당한 성장을 경험하고 있습니다.

AI 기술이 비즈니스와 퍼블리싱에 미치는 영향

AI 기술은 기업이 고객을 더 잘 이해하고 보다 개인화된 관계를 구축하는 데 활용되고 있습니다.

이 기술은 또한 출판 업계에서도 놀라운 응용력을 발휘하여 독자의 주제에 대한 몰입도를 높이고 데이터 중심의 세상에서 데이터에 대한 더 많은 접근성을 제공합니다.

AI 모델은 정보를 저장하는 데이터베이스가 아니라 새로운 정보를 추론하는 데 사용되기 때문에 퍼블리셔의 참여를 통해 더 많은 정보를 시야에 넣는 것이 중요합니다.

Zora의 텍스트 및 비디오 모델이 미치는 영향

Zora의 텍스트 및 동영상 모델은 몇 달 전에 출시되었으며, 반복적인 배포 방식을 사용하여 피드백을 수집하고 개선하고 있습니다.

이 모델은 영화 산업과의 소통과 AI 학습 데이터의 중요성에 초점을 맞춰 제작 비용을 절감하는 것을 목표로 합니다.

Zora의 팀은 이 기술이 영화 제작을 향상시킬 수 있는 잠재력을 보고 있으며, 제품을 더욱 발전시키기 위해 협업과 창의성에 초점을 맞춘 접근 방식을 계획하고 있습니다.

한 기업 임원이 말하는 파트너십의 중요성

이 회사는 특히 Microsoft와의 파트너십을 시장 진출 전략에 있어 매우 중요하게 생각합니다.

그들은 Microsoft를 대규모 시스템을 구축할 수 있는 훌륭한 파트너로 보고 있으며, 상호 이익이 되는 협업으로 보고 있습니다.

이 회사는 파트너와의 협력이 경제에 큰 영향을 미치는 데 중요하다고 믿으며 파트너십의 미래에 대해 낙관적으로 전망하고 있습니다.