Shared Chat•
Default (GPT-3.5)

번역 요청과 응답

June 10, 2023
Role: Professional IT Translator Tasks: 한글로 번역해주세요. 질문과 답변을 잘 구분해서 화자가 다른 경우 화자가 다름을 표현해주세요. a little there on top of it, that will not be future-proof. We've seen this a bunch of times with companies who build on top of us to get a nice business, but then we produce the next model, and it doesn't sustain. And so I think the thing that is actually very hard for us to just go disrupt tomorrow is domain-specific work that's actually very hard. If you're selling to hospitals, there's a lot of work to sell to hospitals. You need to really understand users, you need to understand the impact on patients, you need to be able to work with regulators, like that's something that we can't do by just building better technology. And so I think that really figuring out what is going to just be gone tomorrow versus what is durable, I think that's where the value lies. I have a more of a question. So since you've been playing around with large models, and people talk about emerging properties, and I wanted to know whether you have a good intuition of whether, like, including certain kinds of data sets will unlock in the future. For instance, people talk about including code into the training data to unlock complex reasoning capabilities, but is that the actual case that you're seeing? Also, you've been playing around with GPT-4 where it has the visual domain, visual modality as well. Does that actually unlock additional features? Well, I think so. That's what I was going to say. Yeah, I can probably give you some insight into that. So yeah, very much so, you know, reasoning-heavy data sets, they increase the reasoning capabilities of the models. I think what we do is we have a very comprehensive set of profiles that we're looking at. So those of you who have all of the profiles, reasoning is not how much it helps as an assistant. And I'll tell you, we use smart data set collection to try to get any of those people. Definitely reasoning is one of the top things that we keep in mind. I think it will be one of the big qualitative improvements going forward, just seeing which of the big qualitative improvements going forward. from a content center to the script. And the current model gets some authentic state. So do you have some strategy? That's a good question. It works, definitely, yeah. So I think the personations improve with every model. Every model will resolve the best personations. I think there are statements that people do. One very common technique is to do a little augmented discrimination. And it goes into observations. Sometimes what people have done, which is interesting, is to get first, judge a key to generate an answer. And then have another person who will go through the answer and identify it and find references for it as to where things are going. But we have seen customers who have really solved hallucinations for their domain, including the very difficult ones like legal. So it is possible. And it's just, you gotta do the work. Yeah, I think, as a generation, we can implement hallucinations. And yeah, I think, seeing some of our visibility, there's one where we've seen that house can identify when they're starting a system. This is our recent math template. And we're making progress there. Okay. I pay a fortune to be a lawyer. I'm a lawyer. I'm a lawyer. I'm a lawyer. I'm a lawyer. I'm a lawyer. I'm a lawyer. I'm a lawyer. I'm a lawyer. I'm a lawyer. I'm a lawyer. I'm a lawyer. I'm a lawyer. I'm a lawyer. I'm a lawyer. This is our recent math paper, and we're making progress. Good. I'm from NLP, and I have a question. It seems in some domains there are greater challenges in terms of precisely, and consistently controlling LLM. And in relation to this, Microsoft has recently released an open source framework called Guidance to address this issue. And does OpenAI have any preparation or initiatives related to this controlling guidelines? Yeah, so on the client side, Guidance, things like that, we think they're great. And then on the server side, first of things like that, we think they're great. And then on the server side, first of all, we'll have a new model coming into the API soon that should do JSON output and other structured outputs much more reliably. We're looking at things like, on the server side, being able to give us a grammar. There's open source implementations of a lot of these things, but you give us a grammar and the output will conform to it. So we're really looking into by sort of implementing these things here, the biggest bang for the buck. But generally the way we think about it is, we want to build the highest quality model, so you ask for what you want, you get what you want. That's it. And whenever that falls short, we'll be very involved. That's not the one I'm producing. Please write in Korean language.
ChatGPT

A: 그 위에 조금 더 있으면, 그것은 미래에 버텨낼 수 없을 것입니다. 우리를 기반으로 사업을 세우는 회사들과 여러 차례 이런 상황을 겪었습니다. 우리가 다음 모델을 개발하면 그들의 사업은 유지되지 않습니다. 그래서 우리가 내일 곧장 방해하러 갈 수 없는 것이 실제로 어려운 도메인 특화 작업이라고 생각합니다. 만약 병원에 판매하려면 병원에 판매하는 데는 많은 작업이 필요합니다. 사용자를 정말 이해해야 하며, 환자에 미치는 영향을 이해해야 하며, 규제기관과 협력할 수 있어야 합니다. 이것은 우리가 단순히 더 나은 기술을 개발함으로써 할 수 있는 일이 아닙니다. 그래서 내일 없어질 것인지, 또는 지속 가능한 것인지를 정확히 파악하는 것이 중요하다고 생각합니다. 이것이 가치가 있는 곳이라고 생각합니다. 더 궁금한 점이 있습니다. 대형 모델을 가지고 놀면서, 사람들이 부상하는 속성에 대해 얘기하고, 특정 종류의 데이터 세트를 포함하면 미래에 무슨 일을 할 수 있는지에 대한 직관을 갖고 있는지 알고 싶습니다.

B: 그렇다고 생각합니다. 그게 제가 말하려던 것입니다. 네, 그에 대한 일부 통찰력을 제공할 수 있을 것 같습니다. 그래서 예, 매우 그렇습니다. 추론 중심의 데이터 세트들은 모델의 추론 능력을 향상시킵니다. 우리가 하는 일은 우리가 살펴보는 프로필들의 매우 포괄적인 세트를 가지고 있다는 것입니다. 추론이 얼마나 도움이 되는지는 우리가 사용하는 스마트 데이터 세트 수집을 통해 사람들을 얻으려고 시도합니다. 분명히 추론은 우리가 염두에 두는 주요한 것 중 하나입니다. 앞으로 큰 질적 개선 중 하나가 될 것이라고 생각합니다. 그리고 현재 모델은 일부 진정성 있는 상태를 얻습니다. 그래서 전략이 있나요?

A: 그게 좋은 질문이네요. 그것은 확실히 작동합니다. 그래서 제 생각에는, 매 모델마다 사람들이하는 진술을 개선합니다. 일반적으로 하는 기법 중 하나는 증강된 차별화를 하는 것입니다. 그리고 그것은 관찰에 들어갑니다. 때때로 사람들이 흥미롭게 생각하는 것은, 먼저, 판사에게 답변을 생성하도록 하고, 그 다음에 다른 사람이 그 답변을 통과하여 찾아내고 그것이 어디로 가는지에 대한 참조를 찾는 것입니다. 하지만 우리는 법률과 같은 매우 어려운 분야에서 환각을 해결한 고객들을 보았습니다. 그래서 가능합니다. 그냥 일을 해야 합니다. 네, 나는 생각한다.

B: 그래서, 우리는 NLP 출신이며, 질문이 있습니다. 일부 도메인에서는 LLM을 정확하고 일관되게 제어하는 데 큰 도전이 있는 것 같습니다. 이와 관련하여, 최근에 Microsoft는 이 문제를 해결하기 위한 오픈 소스 프레임워크인 Guidance를 출시했습니다. 이 제어 가이드라인에 관련하여 OpenAI는 어떤 준비나 이니셔티브가 있나요?

A: 네, 클라이언트 측에서는 Guidance와 같은 것들을 매우 좋다고 생각합니다. 그리고 서버 측에서는, 첫째로, 곧 API에 들어올 새로운 모델이 있어야 하며, 이 모델은 JSON 출력과 다른 구조화된 출력을 훨씬 더 안정적으로 할 수 있어야 합니다. 우리는 서버 측에서 문법을 줄 수 있는 것들을 보고 있습니다. 이러한 것들의 많은 오픈 소스 구현이 있지만, 우리에게 문법을 주면 출력이 그것을 준수할 것입니다. 우리는 여기에 이런 것들을 구현함으로써 가장 큰 이익을 얻을 수 있을 것입니다. 그러나 일반적으로 우리가 생각하는 방법은, 가장 고품질의 모델을 만들고, 원하는 것을 요청하면 원하는 것을 얻는 것입니다. 그것이 전부입니다. 그것이 부족하면 우리는 매우 참여할 것입니다. 그것이 제가 생산하는 것이 아닙니다.

Role: Professional IT Translator Tasks: 이전 내용에 이어지는 내용입니다. 한글로 번역해주세요. 질문과 답변을 잘 구분해서 화자가 다른 경우 화자가 다름을 표현해주세요. Do you have a person negative for tonight? Please. No, no, no, no. This has been by far like, I've done maybe 15 of these, I'm going to ask for some negatives for tonight. This has been by far, like I've done maybe 15 of these, this is the nicest one. So as a developer, my first venture company was with the iOS app store in 2008. We built on that with a lot of positive hope, but over the years, different things kind of got into place that made it more restrictive. We've pivoted, we're going all in with AI and education, and we're building a language learning app. The problem right now is there's some extra rhythm that we need to increase, but when you look at the process, it's so vague, it doesn't look like there's transparency. And so my concern, especially with my background with previous platforms, is what if we're needing a rate increase, it's not just because it's optional, but you've got usage, you've built this product, and there's no transparency. You can't, it's just something where right now, it seems like it's case by case approval. And so to me, that's like a very kind of vague and scary place to be as a developer. So I just want to say, I think this summit has not really been working very well. And I think people in the Bay Area have found a way of getting to us when this is needed. We need to get a lot better. So I just want to give you all my contact details so you can let me know. But I think in the future we'll definitely have a better answer to this. We will be more planned because I'm sure you're trying to anticipate growth and you're threatening to anticipate growth, and you're like, even if we have 100x customers, how am I going to be able to pay for it, to control for this, and how am I going to access it? The second issue is, I noticed some people, some communities are getting early access to certain things, but compared to the iOS app store, it's like, it didn't really matter if our competitors got a little bit early access to the next iOS version, because it wasn't revolutionary, each change. But with AI, every three months, things are changing so fast. It's like, how can people, companies, have a fair chance when some companies are getting earlier access? We all grew up on the iOS app store model, and we thought it didn't matter that much. We now realize how much it screwed things up, and how much it screwed things up and how much it can be a good effect. That's totally unconventional. We want to do the same thing in the future. It was truly just we had to go through that learning process. We didn't expect it to have such an impact. But we want to be a platform people can depend on. We realize that means people need reliability, dependability, predictability, but also good treatment. So we're going to work on those. One thing I would say is we're just like, it's quite tight for us right now with the supply of GPUs. And as we get more of that, we'll be able to learn things like normal operations and more frequency. Yeah, that point actually is really my answer to the question. So, it's great to be here. I'm Don. I'm a co-founder and CEO of Bend AI. We are making generative AI engines. So serving generative AI in models like Jetty Q requires a large number of GPUs, resulting in high cost and a negative environment demand. So one approach to addressing this challenge is to develop different serving software that uses a number of GPUs significantly. So, please speak on the software initiatives or approaches pursued by OpenAI in this area. I can take it, but is anyone else more interested in the inference stuff? So yes, we do a lot of inference work. And it really started even with GPT-3. We built this model. And I actually did the initial productionization of it. And so you have this research model that takes all these GPUs. We compressed it down to basically end up running on one machine. So that was effectively a 10 to 20x reduction in footprint. And then over time, we've just been improving inference technology on a lot of angles. And so we do a lot of quantization, we do a lot of, honestly, there's a lot of just like systems work because you have all these requests coming in, you need to batch them together efficiently. The GPU has lots of different resources, right? It has memory bandwidth, it has compute, it has the actual sort of DRAM storage. And for each one of these, you can actually convert it into additional performance if you can also overwrite your communication to the computer. Please write in Korean language.
ChatGPT

A: 오늘 밤에는 부정적인 사람이 있나요? 부디. 아니요, 아니요, 아니요, 아니요. 이것이 제가 한 15번 정도 이런 것들을 했지만, 오늘 밤에는 일부 부정적인 것들을 요청하겠습니다. 이것이 제가 한 15번 정도 이런 것들을 했지만, 이것이 가장 친절한 것이었습니다. 그래서 개발자로서, 나의 첫 번째 벤처 회사는 2008년에 iOS 앱 스토어와 함께였습니다. 우리는 많은 긍정적인 희망을 가지고 그 위에 빌드했지만, 그 후 몇 년 동안 다양한 것들이 더욱 제한적이게 만들었습니다. 우리는 방향을 전환했고, AI와 교육에 전부 투자하고 있으며, 언어 학습 앱을 만들고 있습니다. 지금 문제는 우리가 늘려야 할 추가 리듬이 있지만, 프로세스를 살펴보면 매우 모호하며, 투명성이 없어 보입니다. 그래서 나의 우려, 특히 이전 플랫폼에 대한 나의 배경을 감안할 때, 는 우리가 요율 증가가 필요할 경우, 그것이 단지 선택 사항이 아니라는 것입니다. 사용량이 있고, 이 제품을 만들었고, 투명성이 없습니다. 당신은, 그것은 단지 지금, 그것은 사례별 승인처럼 보입니다. 그래서 나에게는 그것이 개발자로서 매우 모호하고 무서운 장소입니다. 그래서 나는 그저 말하고 싶습니다, 나는 이 정상 회담이 잘 작동하지 않았다고 생각합니다. 그리고 나는 베이 지역의 사람들이 필요할 때 우리에게 연락하는 방법을 찾았다고 생각합니다. 우리는 훨씬 더 나아져야 합니다. 그래서 나는 당신에게 모든 연락처를 알려주고 싶습니다. 그러나 나는 미래에 우리가 이에 대한 더 나은 답변을 가질 것이라고 생각합니다. 우리는 더 많은 계획을 세울 것입니다. 왜냐하면 당신이 성장을 예측하려고 하고, 성장을 예측하려고 하고, 당신은, 우리가 100배의 고객을 가지더라도, 나는 어떻게 그것을 지불할 수 있을 것인가, 그것을 통제할 수 있을 것인가, 그리고 나는 어떻게 그것을 접근할 수 있을 것인가를 생각합니다. 두 번째 문제는, 나는 일부 사람들, 일부 커뮤니티가 특정한 것들에 대해 조기 접근을 얻는 것을 알아챘습니다. 하지만 iOS 앱 스토어와 비교하면, 우리의 경쟁자들이 다음 iOS 버전에 조금 일찍 접근하는 것이 그리 중요하지 않았습니다. 왜냐하면 그것은 각 변화가 혁신적이지 않았기 때문입니다. 그러나 AI와 함께, 매 세 달마다, 사람들이, 회사들이 일부 회사들이 더 일찍 접근을 얻는 것을 어떻게 공정하게 할 수 있는지 변화가 빠르게 일어나고 있습니다. 우리는 모두 iOS 앱 스토어 모델에서 성장했고, 그것이 그리 중요하지 않다고 생각했습니다. 우리는 이제 그것이 얼마나 많은 것들을 망쳤는지, 얼마나 많은 것들을 망쳤는지, 그리고 얼마나 많은 좋은 효과가 있을 수 있는지를 깨달았습니다. 그것은 완전히 비정상적입니다. 우리는 미래에 같은 일을 하려고 합니다. 그것은 정말로 우리가 그 학습 과정을 거쳐야 했던 것입니다. 우리는 그것이 그렇게 큰 영향을 미칠 것이라고는 예상하지 않았습니다. 하지만 우리는 사람들이 의존할 수 있는 플랫폼이 되고 싶습니다. 우리는 그것이 사람들이 신뢰성, 의존성, 예측 가능성, 그리고 좋은 대우가 필요하다는 것을 알고 있습니다. 그래서 우리는 그것들에 대해 작업할 것입니다. 한 가지 말하고 싶은 것은 우리는 그냥, 그것은 우리에게 지금 꽤 타이트하다는 것입니다. GPU의 공급이 많이 필요합니다. 그리고 우리가 그것을 더 많이 얻으면, 우리는 일반적인 운영과 더 빈번한 것들을 배울 수 있을 것입니다. 그 점이 사실 나의 질문에 대한 대답입니다. 그래서, 여기 있어서 좋습니다. 나는 돈입니다. 나는 Bend AI의 공동 창업자이자 CEO입니다. 우리는 생성 AI 엔진을 만들고 있습니다. 따라서 Jetty Q와 같은 생성 AI 모델을 서빙하려면 많은 수의 GPU가 필요하며, 이는 높은 비용과 부정적인 환경 수요를 초래합니다. 이 도전에 대응하는 한 가지 방법은 GPU의 수를 크게 줄이는 다른 서빙 소프트웨어를 개발하는 것입니다. 그래서, OpenAI가 이 분야에서 추구하는 소프트웨어 초기화 또는 접근법에 대해 말해주세요. 나는 그것을 받아들일 수 있지만, 누군가 다른 사람이 추론 물건에 더 관심이 있나요? 그래서 예, 우리는 많은 추론 작업을 합니다. 그리고 그것은 실제로 GPT-3로 시작되었습니다. 우리는 이 모델을 만들었습니다. 그리고 나는 실제로 그것의 초기 생산화를 했습니다. 그래서 당신은 이 연구 모델을 가지고 있습니다. 그것은 모든 이러한 GPU를 가져다 줍니다. 우리는 그것을 기본적으로 하나의 기계에서 실행되도록 압축했습니다. 그것은 효과적으로 10에서 20배의 감소였습니다. 그리고 시간이 지남에 따라, 우리는 많은 각도에서 추론 기술을 향상시키고 있습니다. 그래서 우리는 많은 양자화를 하고 있습니다, 우리는 많은, 솔직히 말해서, 많은 시스템 작업이 있습니다. 왜냐하면 당신은 모든 이러한 요청들이 들어오고, 당신은 그것들을 효과적으로 배치할 필요가 있기 때문입니다. GPU에는 많은 다른 자원이 있습니다, 그것은 메모리 대역폭, 그것은 계산, 그것은 실제로는 DRAM 저장소입니다. 그리고 이 중 각각을 추가적인 성능으로 변환할 수 있습니다. 당신이 또한 당신의 커뮤니케이션을 컴퓨터로 덮어 쓸 수 있다면.

Role: Professional IT Translator Tasks: 계속 이어지는 내용과 문맥을 활용하여 한글로 번역해주세요. 질문과 답변을 잘 구분해서 화자가 다른 경우 화자가 다름을 표현해주세요. So there's a lot of that kind of work that we do that's quite sophisticated in-house. And we've been actually, it's been really encouraging to see the whole ecosystem, right? There's like so much work that's happening right now across the whole open source world. You look at like the efficiency gains that have been happening with Lama, like those are the kinds of things that we really love to see. So I think it's just something that's very much on our mind, it's so clear this stuff is the lifeblood and is actually the limit on our ability to scale, and so every law police comes out as something that can benefit everyone here. I could add a word to that, maybe an obvious point, but maybe reassuring is that all of our incentives are very aligned when it comes to tennis optimization, just because we want to serve more people. When we were able to come up with the tennis price decrease, that was just as much of a happy moment for all of us as it was for our users. We're working on it, and yeah, we're pretty sure about it. Can we actually hear from some women developers and founders here? Why is it only a man here? Thank you, thank you for giving me this chance. I am Cho Kwon-Som, and I work at Information Securities, which is a big security company in South Korea. So, for companies like YNAS, our customers are very sensitive about the accurate decision because it's quickly related to how their assets could change. So I was wondering if there would be a way to measure the certainness of the response of the JPT, the chat JPT response, well, would there be a kind of percentage or some kind of a more way to explain how they are open about the response? I'm not taking questions. You want to? You want to? Yeah. Yeah, so we are looking at this. It's interesting, yeah. JPT is working with a lot of corporate presidents Yeah, so we are looking at this. So they are concerning the information protection and the security of us. And I wonder whether the open AI will target those corporate partners so they can have their own dedicated large model. And so they want those large models to be trained. And the inference running in their inside, in-house infrastructure, how would you kind of pursue those customers? So I think client training, being able to customize it to your own company data is one of the most impactful things. I think it's where companies will get a lot of power from. In terms of inferencing on their own data centers, it's something we haven't pursued yet. And what we do, our libraries have a fine-tuning endpoint, and we have a very big data policy where any data that you upload, it's your data. No other person gets to access it. If you're fine-tuning a model, you have a customized model. That model can only be accessed by yourselves, not anyone else. So there's a lot of things that we do, and we serve through Microsoft Azure, so Microsoft Azure allows us to have productions there as well. So we have quite a few, very large companies in the United States that are using this technology, and one of its enemies are large banks in the US that are comfortable sharing the graduate data with us. I'm actually curious, do you think that Azure is enough for companies like that, or do you need your own in-house infrastructure? Sometimes it's government policy, or sometimes they're internal policy. They cannot upload their data to any kind of cloud or or data center of the other company. So they ask the cloud companies to install the machines inside or in certain physical locations. That makes sense. I also have a lot of questions. Can I ask one question? So one question I really have is that since Greg you mentioned, OpenAI is still a small company. It's not too old. And you're using a lot of the techniques that were already available before. So then why don't these startups use your service rather than spend let's say the next two years spending a lot of money, and because we know, because you have shown that it's possible to train their own language models, while they use your service. Still, OpenAI is small, as you said, it's been only five years, or in fact, if you count from the GPT three days, it's like four years or three years at most. So wouldn't, let's say, any of the startups here, three years of, sorry, wouldn't any of these startups be able to train the same quality, let's say, language model within actually less than three years ago, three years ago. Would any of these startups be able to train the same quality as a language model within actually less than three years because we know that you have done it, right? So why did they lose years of this? So should they try? Yes? Please write in Korean language.
ChatGPT

그러한 종류의 작업을 우리가 매우 고도로 집에서 수행하는 것이 많고, 실제로 이것이 전체 생태계, 즉, 오픈 소스 세계 전체에서 현재 진행 중인 많은 작업을 보는 것이 매우 흥미롭습니다. Lama와 같은 것들이 발생하는 효율성의 증가를 보면, 그것들은 우리가 정말로 좋아하는 종류의 것들입니다. 그래서 저는 이것이 우리가 계속 염두에 두고 있는 것이며, 이것이 우리의 확장 능력에 대한 제한이며, 그래서 모든 법률이 무언가로 나오면 그것이 여기 있는 모든 사람들에게 이익이 될 수 있다는 것이 명확하다고 생각합니다. 그점에 대해 한마디 덧붙이면, 아마도 분명한 포인트지만, 우리의 모든 인센티브는 테니스 최적화와 관련하여 매우 일치하고 있습니다. 왜냐하면 우리는 더 많은 사람들에게 서비스를 제공하고 싶기 때문입니다. 테니스 가격 인하를 내놓을 수 있었을 때, 그것은 우리 모두에게 그만큼 큰 행복한 순간이었습니다. 우리는 그것에 대해 작업하고 있으며, 네, 우리는 그것에 대해 확신하고 있습니다. 여기에 있는 여성 개발자와 창업자들로부터 실제로 몇 가지를 들어볼 수 있을까요? 왜 여기에는 남자만 있는 건가요? 감사합니다, 이 기회를 주셔서 감사합니다. 저는 조권섬이고, 대한민국의 대형 보안 회사인 정보보안에서 일하고 있습니다. 그러니까 YNAS와 같은 회사들의 경우, 우리의 고객들은 그들의 자산이 어떻게 변할지에 급속하게 연결되어 있는 정확한 결정에 대해 매우 민감합니다. 그래서 JPT, 채팅 JPT 응답의 확실성을 측정하는 방법이 있을지 궁금했습니다. 그런 종류의 백분율이나 응답에 대해 얼마나 확실한지를 설명하는 다른 방법이 있을까요? 질문을 받지 않겠습니다. 네, 우리는 이것을 살펴보고 있습니다. 그것은 흥미롭습니다. JPT는 많은 기업 회장들과 함께 일하고 있습니다. 네, 우리는 이것을 살펴보고 있습니다. 그래서 그들은 정보 보호와 보안에 대해 걱정하고 있습니다. 그리고 OpenAI가 그들의 맞춤형 대형 모델을 가질 수 있도록 그런 기업 파트너를 대상으로 할 것인지 궁금했습니다. 그래서 그들은 그런 대형 모델들을 훈련시키고 싶어합니다. 그리고 그 추론이 그들의 내부, 내부 인프라에서 실행되는 것을 원합니다. 어떻게 그런 고객들을 추구할 것인지요? 그래서 저는 고객 훈련, 즉 자신의 회사 데이터에 맞춤화할 수 있는 능력이 가장 영향력 있는 것들 중 하나라고 생각합니다. 이것이 기업들이 많은 힘을 얻을 수 있는 곳이라고 생각합니다. 그들 자신의 데이터 센터에서 추론하는 것에 대해, 그것은 우리가 아직 추구하지 않은 것입니다. 그리고 우리가 하는 것은, 우리의 라이브러리에는 미세 조정 엔드포인트가 있고, 우리는 매우 큰 데이터 정책을 가지고 있습니다. 즉, 어떤 데이터를 업로드하든 그것은 당신의 데이터입니다. 다른 사람은 그것에 접근할 수 없습니다. 모델을 미세 조정하고 있다면, 당신은 맞춤형 모델을 가지고 있습니다. 그 모델은 당신 자신 외에는 누구도 접근할 수 없습니다. 그래서 우리가 하는 많은 것들이 있습니다, 그리고 우리는 Microsoft Azure를 통해 서비스를 제공합니다. 그래서 Microsoft Azure는 우리가 그곳에서 제작을 할 수 있게 해줍니다. 그래서 우리는 미국의 꽤 많은, 매우 큰 회사들이 이 기술을 사용하고 있습니다. 그 중 하나는 우리와 졸업 데이터를 공유하는 것에 편안함을 느끼는 미국의 대형 은행들입니다. 저는 실제로 궁금한데, Azure가 그런 회사들에게 충분할 것이라고 생각하십니까, 아니면 자신만의 내부 인프라가 필요합니까? 가끔은 정부 정책이거나, 가끔은 내부 정책이다. 그들은 자신들의 데이터를 어떤 종류의 클라우드나 다른 회사의 데이터 센터에 업로드할 수 없습니다. 그래서 그들은 클라우드 회사에게 기계를 설치하거나 특정 물리적 위치에 설치하도록 요청합니다. 그것이 이해됩니다. 저도 많은 질문이 있습니다. 한 가지 질문을 할 수 있을까요? 그래서 정말로 가지고 있는 한 가지 질문은 그렉이 언급한 것처럼 OpenAI는 아직 작은 회사입니다. 그것은 아직 늙지 않았습니다. 그리고 당신은 이미 이용 가능했던 많은 기술들을 사용하고 있습니다. 그런데 왜 이 스타트업들은 당신의 서비스를 이용하지 않고, 가령 다음 두 년 동안 많은 돈을 쓰면서, 그리고 우리는 당신이 그것이 가능하다는 것을 보여주었기 때문에, 그들 자신의 언어 모델을 훈련시킬 수 있는지, 그들은 왜 당신의 서비스를 이용하는 것인지. 여전히 OpenAI는 작습니다, 그렇다고 했습니다. 5년밖에 안 되었고, 실제로는 GPT 3 시대부터 센다면 4년이나 최대 3년밖에 안 되었습니다. 그럼 여기의 스타트업들 중에서, 3년 동안, 죄송합니다. 이 스타트업들 중에서 3년 전에는, 3년 전에는 이 스타트업들이 사실상 3년 미만으로 같은 품질의 언어 모델을 훈련시킬 수 있었을까요, 왜냐하면 우리는 당신이 그것을 했다는 것을 알기 때문입니다, 그렇죠? 그래서 왜 그들은 이것의 년수를 잃었을까요? 그래서 그들은 시도해야 합니까? 네?

Role: Professional IT Translator Tasks: - 윗부분와 문맥을 활용하여 한글로 번역해주세요 - 질문과 답변을 잘 구분해서 화자가 다른 경우 화자를 A,B,C와 같이 적어주세요. - STT의 결과이니, 어색한 표현을 고쳐주세요. Sam's gonna answer this question before you answer it. I've got my own spin on it. Well, I've got my own spin on it, which is, look, first of all, I think as a startup, you get to be best in the world at one thing, at most one thing. And if you want that one thing to be advancing AI, you can. You can choose to pursue it, but you're not going to be able to pursue anything else. So that's a first decision. So that's a real opportunity cost. That, for us, that's the thing we want to do. And we actually sort of choose not to do so many things that would be pretty extremely exciting. All the things I was saying, like going into any one domain, you just kind of can't do that if you're going to do the kind of thing that we do. And there's a lot of actual forward planning that's involved, to actually build the supercomputers. That's not something that you just put together a supercomputer in six months. There's no GPUs out there, in part because we have... I need that! But it's like, that's one input, right? If you don't have the GPUs, you're not going to do it. Unless, again, maybe there's a magical breakthrough to be made, but that's a starting point. And then, one thing that's easy to miss is the degree to which every single part of the system multiplies. You can start to see this with some of the open source models. A 40 billion parameter model, they're not all made equal. There are so many people who have tried to train a GPT-3 quality model and not gotten there. In fact, internally, after we trained GPT-3, we had a whole year of failed attempts to exceed it. And we had to rebuild the entire training stack, every single detail, we had to sort of, you know, go over with a fine-toothed comb, and you just keep seeing all these little problems. And so much of it, by the way, it's boring work. Like, it actually really sucks. I love that kind of work, like that is what interests me. Like, when I don't have to, like, clone something brilliant and new, like, you know, the brilliant new stuff that happens over here, for me, it's the boring engineering work. And then you need to coordinate a lot of people. There's a lot of expertise you need to develop. For us, one of the biggest successful programs has been the residency, where we take people that don't know anything about AI, and we train them. We spend a lot of time teaching them AI. But you also need to have that AI expertise already. So it's like, there's these flywheels that we've been putting into effect for the whole time that we've existed. So it's like there's these flywheels that we've been putting into effect for the whole time that we've existed. So it's not impossible. We've shown it is possible. We're going to keep trying it. Hopefully we will continue to be the leading edge and be able to host these services and accelerate the work that you do. If you want to do it too, again, I think you're welcome to. But we'd love it if you just came and worked with us, because I think that this is just a hard thing, it's a hard engineering problem, and there's so many benefits from it. Can we also hear the answer from the non, let's say, president, non-CEO? Yeah, go ahead. Please. Please. Please. It's way too hard. I'd like to add more detailed questions on enterprise and fine tuning. There was a question, so you already asked people, so can we... Did you get the... We do want to talk about that. So we're... we do power messaging and other applications, and we have a lot of customers who are trying to plug in OpenAI into our system, to power chat. And we see... so how serious is OpenAI about BPA business? Because we see you guys releasing customer features first, and it takes quite a while before it becomes available for you guys. Please write in Korean language.
ChatGPT

A: 그 질문에 대한 대답은 Sam이 할 것이지만, 제 자신의 의견도 있습니다. 일단, 스타트업으로서 당신은 최소한 하나의 것에서 세계 최고가 될 수 있습니다. 그리고 그 하나가 인공지능의 발전이라면, 당신은 그것을 추구할 수 있습니다. 하지만 다른 것을 추구할 수는 없습니다. 그것이 첫번째 결정입니다. 그것은 실제로 기회 비용입니다. 우리에게 있어서, 그것이 우리가 하고 싶은 것입니다. 우리는 실제로 매우 흥미진진한 많은 것들을 하지 않기로 선택했습니다. 내가 언급했던 모든 것들, 어떤 하나의 분야로 진입하는 것을 포함해서, 당신이 우리가 하는 종류의 것을 하려고 한다면, 그럴 수 없습니다. 그리고 실제로 슈퍼컴퓨터를 만들기 위한 많은 전방 계획이 필요합니다. 그것은 당신이 6개월 만에 슈퍼컴퓨터를 만들어낼 수 있는 것이 아닙니다. 그 부분에는 GPU가 없습니다, 부분적으로 우리가... 제가 그것을 필요로 하기 때문입니다! 하지만 그것은 하나의 입력일 뿐입니다. 당신이 GPU를 가지고 있지 않다면, 당신은 그것을 할 수 없습니다. 또한, 다시 말하지만, 아마도 마법적인 돌파구가 있을지 모르지만, 그것은 시작점입니다.

그리고 한 가지 놓치기 쉬운 것은 시스템의 각 부분이 얼마나 많이 곱해지는지입니다. 당신은 이것을 일부 오픈소스 모델들과 함께 보게 될 수 있습니다. 400억 개의 파라미터 모델, 모두가 동등한 것은 아닙니다. GPT-3 품질의 모델을 훈련시키려고 시도한 많은 사람들이 성공하지 못했습니다. 사실, 내부적으로 GPT-3을 훈련시킨 후에, 우리는 그것을 초과하는 데 실패한 전체적인 한 해를 가졌습니다. 그리고 우리는 모든 훈련 스택을 다시 만들어야 했습니다, 모든 단일 세부사항, 우리는 결국, 미세하게 검토해야 했고, 당신은 계속해서 이런 작은 문제들을 발견하게 됩니다. 그리고 그것의 많은 부분이, 참고로, 지루한 작업입니다. 그것은 실제로 정말 지루합니다. 저는 그러한 종류의 작업을 사랑합니다, 그것이 저에게 흥미를 끕니다. 새롭고 뛰어난 것을 복제해야 할 때가 아닐 때, 저에게 있어서, 놀라운 새로운 것들이 여기에서 일어날 때, 저는 지루한 엔지니어링 작업입니다.

그리고 당신은 많은 사람들을 조정해야 합니다. 개발해야 할 많은 전문 지식이 있습니다. 우리에게 있어서, 가장 성공적인 프로그램 중 하나는 AI에 대해 아무것도 모르는 사람들을 받아들여 훈련시키는 레지던시였습니다. 우리는 많은 시간을 들여 그들에게 AI를 가르칩니다. 하지만 당신은 이미 그 AI 전문 지식을 가지고 있어야 합니다. 그래서, 우리가 존재한 동안 계속해서 작용시키고 있는 이런 플라이휠들이 있습니다. 그것은 불가능한 것이 아닙니다. 우리는 그것이 가능하다는 것을 보여주었습니다. 우리는 계속해서 시도할 것입니다. 희망적으로는 우리는 계속해서 선두에 있고, 이러한 서비스를 호스팅하고, 당신이 하는 작업을 가속화할 수 있을 것입니다. 당신도 그것을 하고 싶다면, 다시 말하지만, 저는 당신이 환영받을 것이라고 생각합니다. 하지만 우리와 함께 일하는 것이 좋겠습니다, 왜냐하면 이것은 그냥 어려운 일이고, 그것은 어려운 엔지니어링 문제이며, 그것으로부터 많은 이점들이 있습니다.

B: 비-CEO, 비-회장인 사람의 답변도 들어볼 수 있을까요? 그렇습니다, 제발. 제발. 제발. 그것은 너무 어렵습니다. 엔터프라이즈와 미세 조정에 대한 더 구체적인 질문을 추가하고 싶습니다. 이미 사람들에게 질문했던 것이 있습니다, 그래서 우리는... 당신은 그것을 받았습니까? 우리는 그것에 대해 이야기하고 싶습니다. 그래서 우리는... 메시징과 다른 응용 프로그램을 지원하고 있으며, 많은 고객들이 우리 시스템에 OpenAI를 플러그인하여 채팅을 지원하려고 시도하고 있습니다. 그리고 우리는 보았습니다... OpenAI는 BPA 사업에 얼마나 진지합니까? 왜냐하면 우리는 당신들이 고객 특성을 먼저 출시하는 것을 보았고, 그것이 당신들에게 사용 가능해지기까지는 꽤 오래 걸립니다.

Role: Professional IT Translator Tasks: - 윗부분와 문맥을 활용하여 한글로 번역해주세요 - 질문과 답변을 잘 구분해서 화자가 다른 경우 화자를 A,B,C와 같이 적어주세요 - STT의 결과이니, 어색한 표현을 고쳐주세요 For example, like plugins. We'd love to use a plugin as quickly as possible to wait for those releases for your clients. We talk about this a lot, actually. Do you have a friend? We end up in a lot of these conversations. We get it. And I think that for us, you can see through the past six, 12 months, our own indecision on exactly what to focus on. And to that focus point, it's hard to, what we want to do is we want to advance these models. That is, I think, the core for us. We want to make them better, and we want to get them out there. And then exactly what the mechanism is, is it through chatGBT, which took off much more than I was expecting, is it through API so lots of people can build on top of it, what's the best way to do it? I think it actually varies a lot per feature. And so the interesting thing about plugins, as an example, is they don't work yet. It's still, it's like, you know, there's some of our features, like for example, Code Interpreter, I think that's starting to really work, but like, man, that was like months of slog, right? And that was like just a lot of time not working. Third-party plugins still don't really work. I mean, how many people have tried third-party plugins in CacheBT? You know, was anyone like, this is the most important thing? Like, we'll get there, right? So this is kind of the story, is that the choices we tend to make are the ones that give is the most important thing? Like, we'll get there, right? So this is kind of the story, is that the choices we tend to make are the ones that give us the most engineering philosophy. So we are very committed to start this building on top of us. We want to be a platform, there's no question about that for us. You will sometimes see us have things develop and bake in the consumer side much faster, or because it's much faster for us to do it, and then we bring it to the platform. Sometimes we'll do things on the platform first. GPT-V, so the vision side, is a good example, where we've been working with partners there. It's not in Chagin-PF yet. It will be, right? So I think that you can see this kind of nuance. And for us, it's always calibrated by what gives us the most velocity and helps us get to that future model fastest. Just to kind of talk you through what our constraints are, I think that we know that to developers reliability is key, right? So we might want to try a bunch of different new research directions, and for us, consumer app, which I think is the fastest way of doing this, because it's a free product, We don't want to change our models on our API business customers, and we want to keep the API structure as well. And because we all kind of empathize with developers, we're definitely much more careful about that. At the same time, we empathize with the fact that our API developers would want the latest and the greatest as well. And part of that was just kind of the partisan decision behind our announcement of the emerging models, but our large model, we haven't actually published the models since then yet, but that's the reason behind why we did that. As an example, we'll have a function call coming very soon, where basically this is exactly the mechanism that we build plugins through. That will be in the API in two weeks, something like that for now. We'll be releasing the model soon. few weeks, something like that for now. We'll be releasing a new model soon. And all of that was because we made so many mistakes and learned so much from the deployment within chat GPT. Actually, okay, let's hear from another woman developer. Let's do that again. I'm actually not a developer, but I'm here. Oh, okay, sorry about sorry. No, no worries. I'm Yan from Speak. It's been great working with you all. Great to hear. Well, thank you. I just joined a month ago, but yeah. So my question is around, given how fast things are changing at the moment, would you say that there's a version of a world where we don't even need to learn a foreign language? And how should we as a company work on that? I think I can think of a solid one. I think that the world is super close to translation not being a necessary thing. That said, I think, you know, like the 80-20, I think a lot of people just have very easy access to understanding the gist of things, but when it comes to the really detailed idioms or concepts like 混汉, 混汉, stuff like that, I don't still know how to translate that into English. Concepts like one When you like don't know you're talking to a when you don't know, you're talking to a friend, you don't know something, you're like, oh, is this a thing? You don't just, everyone stop, hold your phone and check. Even though you could, all the facts are there, right? There's this robot in the sky that knows way more than any human does. And maybe we'll get there with language. I think this last mile problem that Joanne was saying, I think that's real. I think that this is a place, again, back to where's the opportunity for startups, right? I think that maybe is a place, again, back to where's the opportunity for startups, right? I think that maybe startups can bridge that, maybe you can build systems or even just sort of techniques that help people get there. And I think this like moving the machine closer to the humans, but that last mile problem, that's still going to be there. I think another last mile problem that we're thinking a lot about internally is also there's kind of this unequal representation of training data among different languages. So for example, it's very easy to find training data for Chinese or Korean where you find major spoken languages. But there are hundreds of other less spoken and kind of deflected languages that are often forgotten. But also, that's also something that we're trying to deal with. And I think that will be hard to go to find a good translation for those images in the future. One more time about language, I actually have a question for the group, which is how's our Korean performance? How does it compare to English? Slow. Slow. Slow. So I've got a question related to the flu message. The flu is really great. Please write in Korean language.
ChatGPT

A: 예를 들어, 플러그인 같은 것입니다. 가능한 빠르게 플러그인을 사용하고 싶습니다. 우리가 클라이언트들을 위해 이런 릴리즈를 기다리는 것입니다. 사실, 우리는 이것에 대해 많이 이야기합니다. 친구가 있나요? 우리는 이런 대화에서 자주 발견됩니다. 우리는 이해합니다. 그리고 저희에게 있어서, 당신은 지난 6개월, 12개월 동안 우리가 어떤 것에 초점을 맞출지에 대한 결정이 어떤 것이었는지 볼 수 있습니다. 그리고 그 초점 지점으로, 우리가 하고 싶은 것은 이 모델들을 발전시키는 것입니다. 저희에게 있어서, 그것이 핵심입니다. 우리는 그들을 개선하고, 그들을 대중에게 널리 알리고 싶습니다. 그리고 그 메커니즘이 정확히 무엇인지는, 챗지비티를 통한 것인지, 저는 예상했던 것보다 훨씬 더 많은 사람들이 그 위에 무언가를 구축하도록 하기 위한 API를 통한 것인지, 어떤 방법이 최선인지는 실제로 특성에 따라 많이 달라집니다. 그래서 플러그인이라는 예시는 흥미로운데, 그들은 아직 작동하지 않습니다. 아직도 그것은, 일부 기능들, 예를 들어 코드 해석기 같은 것이 실제로 작동하기 시작하는 것을 보았지만, 그것은 정말로 수개월 동안의 힘든 작업이었습니다. 그리고 그것은 많은 시간동안 작동하지 않았습니다. 제 3자 플러그인들은 아직도 제대로 작동하지 않습니다. 캐시BT에서 제 3자 플러그인을 시도한 사람들이 얼마나 많은지 말해 보세요. 이것이 가장 중요한 것인지, 우리는 도달할 것입니다, 그렇죠? 이것이 이야기의 핵심인데, 저희가 통상 선택하는 것들은 저희에게 가장 많은 엔지니어링 철학을 주는 것들입니다. 우리는 저희 위에 건설을 시작하는 것에 매우 헌신적입니다. 우리는 플랫폼이 되고 싶습니다, 저희에게 있어서는 그것이 확실한 목표입니다. 당신은 가끔 우리가 소비자 측면에서 더 빠르게 개발하거나, 더 빠르게 할 수 있기 때문에, 그리고 그런 다음에 우리가 그것을 플랫폼에 가져오는 것을 보게 될 것입니다. 가끔 우리는 플랫폼에서 먼저 일을 하게 됩니다. GPT-V, 즉 비전 측면, 이것은 우리가 파트너들과 함께 작업하고 있는 좋은 예시입니다. 아직 채진PF에는 없습니다. 하지만 될 것입니다, 그렇죠? 그래서 저는 당신이 이런 뉘앙스를 볼 수 있다고 생각합니다. 그리고 저희에게 있어서, 이것은 항상 가장 빠르게 속도를 낼 수 있게 하고, 우리가 그 미래의 모델에 가장 빨리 도달할 수 있게 하는 것에 의해 조절됩니다. 저희의 제약 조건에 대해 이야기해 보면, 저희는 개발자들에게는 신뢰성이 중요하다는 것을 알고 있습니다, 그렇죠? 그래서 우리는 다양한 새로운 연구 방향을 시도하고 싶을지도 모르지만, 저희에게 있어서, 소비자 앱이 이것을 가장 빠르게 하는 방법이라고 생각합니다. 왜냐하면 그것은 무료 제품이기 때문입니다, 우리는 우리의 API 비즈니스 고객들에게 모델을 바꾸고 싶지 않고, 우리는 API 구조도 유지하고 싶습니다. 그리고 우리는 모두 개발자들에게 공감하므로, 우리는 그것에 대해 확실히 더 조심스럽게 다룹니다. 동시에, 우리는 우리의 API 개발자들이 최신이며 가장 훌륭한 것을 원할 것이라는 사실에 공감합니다. 그리고 그것의 일부는 우리가 발표한 융합 모델 뒤에 있는 결정이었지만, 우리의 큰 모델, 우리는 아직 그 이후로 모델을 출판하지 않았지만, 그것이 우리가 그렇게 한 이유입니다. 예를 들어, 우리는 곧 기능 호출이 있을 것입니다, 기본적으로 이것이 우리가 플러그인을 구축하는 정확한 메커니즘입니다. 그것은 이제부터 약 2주 후에 API에 들어갈 것입니다. 우리는 곧 모델을 출시할 것입니다. 그리고 그 모든 것은 우리가 챗 GPT 내에서 배포하면서 많은 실수를 범하고 많은 것을 배웠기 때문입니다. 실제로, 여성 개발자 한 명으로부터 다시 들어보자. 다시 해봅시다. 사실 저는 개발자가 아니지만, 여기에 있습니다. 아, 그렇군요, 죄송합니다. 아니오, 문제 없습니다. 저는 Speak의 Yan입니다. 여러분과 함께 일하는 것이 정말 좋았습니다. 그것을 들어서 기쁩니다. 감사합니다. 저는 한 달 전에 합류했지만, 그래요. 그래서 제 질문은, 현재의 변화가 얼마나 빠르게 일어나고 있는지를 감안할 때, 외국어를 배울 필요조차 없는 세상의 버전이 존재할 수 있을지, 그리고 저희 회사가 그것을 어떻게 처리해야 할지에 대한 것입니다. 제가 확실하게 생각할 수 있는 것은, 세상이 번역이 불필요한 지점에 매우 가깝다는 것입니다. 그렇다고 해도, 나는 생각합니다. 80-20법칙처럼, 많은 사람들이 사물의 핵심을 이해하는 데 매우 쉬운 접근을 가지고 있지만, 정말로 상세한 관용구나 개념들, 예를 들면 混汉, 混汉, 이런 것들에 이르기까지는 아직도 어떻게 그것을 영어로 번역할지 모릅니다. 그런 개념들 같은 것입니다. 당신이 친구에게 말하다가 무언가를 모르는 경우에는, 당신은 그것이 무엇인지 모른다고 말합니다. 당신은 그냥, 모두 멈추고, 휴대폰을 확인합니다. 그렇게 할 수 있음에도 불구하고, 모든 사실들이 거기에 있습니다, 그렇죠? 거기에는 하늘에 있는 이 로봇이 인간보다 훨씬 더 많은 것을 알고 있습니다. 그리고 아마 우리는 언어에 대해 더 많이 생각할 필요가 있을 것입니다. 어떤 관용구든, 아무리 이상하든, 직접 번역하든 아니면 좋은 번역을 찾든, 그것은 더 이상 중요하지 않습니다. 어떻게 이것을 번역하는지가 중요한 것이 아니라, 어떻게 이것을 이해하는지가 중요합니다. 그리고 그것이 우리가 집중하고 있는 부분입니다. 그래서 저희는 정말로 생각합니다. 제 개인적인 관점에서는, 언어는 항상 그것의 감정적인 면을 가지고 있습니다. 그리고 그것은 항상 그것의 문화적인 면을 가지고 있습니다. 그리고 그것을 제대로 이해하려면, 여전히 해당 언어를 배우는 것이 필요합니다. 그러나 그것이 여러분이 해야 할 일이라는 것을 말하는 것은 아닙니다. 여러분이 그것을 할 수 있다는 것을 말하는 것입니다. 그래서 그것은 매우 흥미로운 시기입니다, 여러분이 그것을 알고 싶다면.

continue writing Please write in Korean language.
ChatGPT

조앤이 언급한 이 마지막 마일 문제, 그것은 실제입니다. 이것이 바로 기회가 있는 스타트업들이 어디로 가야 하는지에 대한 답변일 수 있습니다, 그렇죠? 아마 스타트업들이 그것을 연결할 수 있을 것입니다. 아마 여러분은 사람들이 그곳에 도달하는 데 도움이 되는 시스템이나 기법을 구축할 수 있을 것입니다. 그리고 나는 이것이 기계를 사람들에게 더 가깝게 이동시키는 것처럼 보이지만, 그 마지막 마일 문제, 그것은 여전히 존재할 것입니다.

우리가 내부적으로 많이 생각하고 있는 또 다른 마지막 마일 문제는 서로 다른 언어들 사이에 불평등한 훈련 데이터의 표현이 있다는 것입니다. 예를 들어, 중국어나 한국어 같은 주요 언어에 대한 훈련 데이터를 찾기는 매우 쉽지만, 수백 가지의 덜 사용되고 주의를 기울이지 않는 언어들이 종종 잊혀집니다. 그런데도, 그것 역시 우리가 처리하려고 노력하고 있는 것입니다. 그리고 나는 그것이 미래에 그런 이미지들에 대한 좋은 번역을 찾는 데 어려움이 될 것이라고 생각합니다.

언어에 대해 한 번 더 이야기해보자면, 저는 그룹에게 한 가지 질문이 있습니다. 우리의 한국어 성능은 어떤가요? 그것은 영어에 비해 어떤가요? 느립니다. 느립니다. 느립니다. 그래서 저는 플루 메시지와 관련된 질문을 가지고 있습니다. 플루는 정말로 좋습니다.

Role: Professional IT Translator Tasks: - 윗부분와 문맥을 활용하여 한글로 번역해주세요 - 질문과 답변을 잘 구분해서 화자가 다른 경우 화자를 A,B,C와 같이 적어주세요 - STT의 결과이니, 어색한 표현을 고쳐주세요 I think it's much, than the degree that you can buy and the reasoning skill, etc But I don't think you understand how much is slow in Korea I cannot really understand Korea Because how much information we want to generate in Korea is painstaking and especially when we try to build a product on top of that even prototyping is quite impossible because you have to test it in real world situation but if it's slow, you cannot even test it I think, you talked about getting down the price and speeding up the model. I think speeding up the model is more important because at least you can protect it. I think you can pay for it. We've heard that a lot in the field. So, my question is, do you think in the future there would be a length barrier? Because Korea's the speed model little or the vast amount of Korean and such like that? And second of all, I know you have a plan to improve that, but what kind of scale do you want to look at? Like 10x, 5x, something like that? Or what kind of general schedule for improving the speed of the university in other languages? This is again a conversation that we have. Yeah, Greg and I talk about this a lot. Yeah, so I think given, like since we've trained chapter PTN, since we've had a ton of evolution, we are now seeing how many instances a lot of our users are not talking to us in English. I think that was a huge update for us. We thought that it would be maybe 80% English and 20% non-English. I don't think I can show you the actual numbers, but it is way off from that. So we've learned a lot, and we are taking that into account as we plan for next generation roles and post-renewing our research as well. Korean, I don't really know how to talk about this, but it should be way better. So. Yeah. Yeah. Yeah. Yeah. And I think probably, so someone had mentioned earlier, tokenization, like I think that's one place that we can improve things. I think there's the, you know, we just, the amount, like we have to put together these training mixes of different amounts of different languages. There I think we can the amount, we have to put together these training mixes of different amounts of different languages. There I think we can increase quite a bit, and I'm planning on doing so. And then there's just simply more GPUs, you know, more, all the inference optimization that we need. So I would definitely expect, in general, we are on very much the Moore's Law style curve, where it's like all the existing things just get better faster, but much faster than Moore's Law. We did the 10x price reduction for faster and better quality for 3.5. I think we can do the same thing for 4. These things, it won't happen tomorrow, but certainly 6-12 months from now, if we're not like GP4 feels like 3.5 does today, we've done something wrong. Yeah, we'll get you a 10x speedup. I think there's also more options with customizing models. As soon as we release that functionality, it will be quite easy to swap the initial encoder. So when you're interested in a little bit of fine tuning, then we can work with K-alphabet. K-alababet is very good. It's a simple mapping, so it should be a pretty nice way to get to the phone as well as in English in terms of speed of this. I just, I'm sure somebody will do that very soon, as soon as we enable fine tuning. One more thing is that the more Korean users use our product, they give us feedback. So, if you want to give us feedback. And if you want to, if any of you want to give us data sets somehow or if you know how we can get a lot of high quality Korean language data, we'll take it from that. So, what's the benefit there? You get a better model of Korean? What's the general solution? We're happy to hear something. We have a lot of open data. I'm Joseph from Simply. It's a 1C company. The challenges that we're having are about data compliance issues. So, as I already said, to penetrate into enterprise customers, we have low credibility, right? So we got to get a soft 2G DPR, but that's fantastic. In terms of data privacy, some companies even banned using any product built on the 2G D3 or the 3G. So it really hinders our market penetration. So I'd just love to figuring a plan to address that. We're going to, yeah, we at OU also like marketing our cover on that. We don't train on any API data, but we have not made that well known enough. Our hope is that we get that message out more, and people will be more comfortable with it. So we're going to work on that. That's also something that's come up a lot in this training. So moving on, let me just add one thing. So you don't use that for training, but then you still save it as something. And that actually creates the situation where whatever you type in on the chatgpk API and whatnot is a public information load. So there's IP issues that are related. And then, for instance, pharmaceuticals and whatnot, they all ban those using the chatgpk API because of Christiaan's statement. So in chatgpk, you can turn it off and you can say, don't and whatnot, they all ban the use of the chatGPT at the moment because of the pre-settings. So in chatGPT, you can turn it off and you can say don't store your money by data, don't train on it, but by default we are trying to completely replace all the usage of chatGPT, so we do. Data retention on the APL, we do retain for 30 days, but only for trust and safety, not like compliance, we're not looking at that. that are not compliance or not up to the standard. Please write in Korean language.
ChatGPT

A: 그것이 상당히 느리다는 것을 알고 있지만, 한국에서 얼마나 느린지에 대해 완벽히 이해하는 것은 어려워요. 왜냐하면 우리가 한국에서 생성하려고 하는 정보량이 엄청나기 때문이죠. 특히 그 위에 제품을 만들려고 할 때, 실제 세계에서 테스트를 해야 하는데, 그것이 느리면 테스트조차 할 수 없습니다. 가격을 낮추고 모델을 더 빠르게 만들어야한다고 말씀하셨는데, 저는 모델을 더 빠르게 하는 것이 더 중요하다고 생각합니다. 왜냐하면 최소한 그것을 보호할 수 있으니까요. 저는 그것에 대해 많이 들었습니다. 그래서 제 질문은, 미래에 길이의 장벽이 있을 것이라고 생각하시나요? 왜냐하면 한국의 속도 모델이 조금이라도 있거나 한국어 양이 많기 때문이죠.

B: 예, 그렉과 저는 이 주제에 대해 많이 이야기했습니다. 우리가 이제 많은 사용자들이 영어로 대화를 하지 않는다는 것을 보고 있기 때문입니다. 이것은 우리에게 큰 깨달음이었습니다. 우리는 그것이 아마도 80% 영어, 20% 비영어가 될 것이라고 생각했습니다. 저는 실제 숫자를 보여드릴 수 없지만, 그것은 훨씬 떨어져 있습니다. 그래서 우리는 많이 배우고, 다음 세대의 역할을 계획하고, 연구를 갱신할 때 그것을 고려하고 있습니다. 한국어에 대해서는, 제가 어떻게 이야기해야 할지 모르겠지만, 훨씬 나아져야 합니다.

A: 아마도 우리가 개선할 수 있는 한 가지 분야는 토큰화일 것입니다. 저는 다른 언어들의 다양한 양을 혼합해서 훈련용 믹스를 만들어야 하는데, 저는 그것을 상당히 늘릴 수 있을 것이라고 생각합니다. 그리고 더 많은 GPU, 추론 최적화 등이 필요합니다. 그래서 저는 일반적으로 우리가 모어의 법칙 스타일의 곡선에 있고, 모든 기존의 것들이 더 빠르게, 하지만 모어의 법칙보다 훨씬 빠르게 나아진다고 확신히 기대하고 있습니다.

B: 저는 또한 모델을 사용자화하는 것에 대해 더 많은 옵션이 있다고 생각합니다. 그 기능을 출시하자마자, 기초 인코더를 교체하는 것이 꽤 쉬울 것입니다. 따라서 세밀한 조정에 관심이 있다면, K-alphabet을 사용할 수 있습니다. K-alphabet은 매우 좋습니다. 그것은 간단한 매핑이므로, 이것은 속도 면에서 영어와 마찬가지로 전화로 이동하는 좋은 방법이 될 것입니다. 저는 누군가가 그것을 곧 할 것이라고 확신합니다.

A: 또한 한국 사용자들이 우리의 제품을 더 많이 사용하면서 피드백을 주게 됩니다. 그러니 피드백을 주세요. 그리고 또한, 만약 당신이 데이터 세트를 제공하고 싶다거나, 우리가 어떻게 많은 고품질의 한국어 데이터를 얻을 수 있을지 알고 있다면, 우리는 그것을 받아들일 것입니다.

B: 그럼 그게 어떤 이점이 있나요? 더 나은 한국어 모델을 얻는다는 것인가요? 일반적인 해결책은 무엇인가요?

C: 나는 심플리의 조셉입니다. 우리가 겪고 있는 도전은 데이터 준수 문제에 관한 것입니다. 기업 고객들에게 진출하려면, 우리는 신뢰성이 낮습니다. 그래서 우리는 2단계의 GDPR을 얻었지만, 그것은 훌륭합니다. 데이터 개인정보에 관한 일부 회사들은 2단계의 D3나 3단계의 제품 사용을 전면 금지했습니다. 그래서 이것은 우리의 시장 진입을 크게 저해하고 있습니다. 그래서 그것을 해결하는 계획을 세우는 것이 좋을 것 같습니다.

D: 네, 우리도 마케팅을 위해 그것을 커버하려고 합니다. 우리는 API 데이터를 훈련에 사용하지 않지만, 그것을 충분히 잘 알리지 않았습니다. 우리의 희망은 그 메시지를 더 많이 전달하고, 사람들이 그것에 더 편안해지게 하는 것입니다. 그래서 우리는 그것에 대해 노력하고 있습니다. 그것은 또한 이 훈련에서 많이 나왔습니다. 그럼, 계속해서, 한 가지만 추가하겠습니다. 훈련에는 사용하지 않지만, 여전히 그것을 어떤 것으로 저장합니다. 그리고 그것은 실제로 chatgpk API에 타이핑하는 모든 것이 공공 정보가 되는 상황을 만듭니다. 그래서 IP 문제가 관련되어 있습니다. 그리고 예를 들어, 제약사와 같은 곳들은 모두 chatGPT의 사용을 현재 금지하고 있습니다. 그래서 chatGPT에서는 그것을 끌 수 있고, 데이터를 저장하지 말라고 말할 수 있습니다. 하지만 기본적으로 우리는 chatGPT의 모든 사용을 완전히 대체하려고 노력하고 있기 때문에, 우리는 그것을 합니다. APL에서의 데이터 보유, 우리는 30일 동안 보유하지만, 신뢰와 안전을 위해서만입니다, 우리는 그것을 보지 않습니다. 표준에 미치지 못하는 그것들이 아닙니다.

Role: Professional IT Translator Tasks: - 윗부분와 문맥을 활용하여 한글로 번역해주세요 - 질문과 답변을 잘 구분해서 화자가 다른 경우 화자를 A,B,C와 같이 적어주세요 - STT의 결과이니, 어색한 표현을 고쳐주세요 I wanted to ask questions regarding the way to building services. So what we are now questioning is like how to use GPT models to make customers relate to the question and answer in robots. And we are now doing like how the other developers do. We put our user queries to search engines and get the right context and put the context that you prefer and send it to GP and actually it doesn't work. I don't know why but it doesn't work. There is a negative feedback. But I somehow feel like GP cannot find what would be the most important context in this context. So for example, if user asks us to put information about banking products, and then the most important information would be interest rates. But the answer is sometimes contain that interest rate, but sometimes it does not contain that rate. So I want to ask if there is any intuition for it to make all the things that are right or not. But if we put a lot of instructions in there, then we are suffering from the number of interpreters. So yeah, do you have a question? This is an extremely Boris question. Boris is the number of alternatives. So, Josh, do you have a question? This is an extremely Boris question. Boris is simply the expert on this. There's a lot of things you can do there. The open source and open-active book, which is a great resource. I believe there's maybe one or two examples for how you can do this. Just very quickly, you can first have a model generate the answer, which may be how to state it, and then you use that answer to do the lookup. That's one thing that can help. You have a lot of things like specific product names, then adding like DP25 or any other, that's a TF-IDF or anything that also is based on keywords, in addition to the embedding similarity, will help massively. I think those are probably the two main things I would say. Also, reducing the size of the context that you are treating, I'm saying like maybe 100 or 200 tokens in English, maybe 600 in Korean, seems to work better for these types of use cases. And are you currently using GPT-4? I'll be ready to announce when you think that we should. Okay, definitely, definitely. Okay, got it, got it. Oh, and I have one more question. So, yes, GPT-4 is much better. If you have a lot of very short contexts, even though to a human it sort of doesn't seem that organized, GPT-4 is really good at knowing which information to use and which information to take out. So it doesn't get as confused by formats as GPT-5 does, when you're taking out multiple, very varying types of information. Okay, I will answer that. I will say more. So when it comes to data supply, I want to know if the model would be able to answer questions like, so our customers want AI to answer anything that they want. So I borrow money, like this amount of money from the bank, and I want to get loan free with repayment plans, something like that. So for a GP, 3.5 and more, it doesn't really work when it comes to the repayment plans. So your plan would be like kind of fuzzy. I think you mean before with fine-tuning, which is a little bit slower. Oh, okay, with fine tuning. With really high quality data, which I'm sure that's... ...for the new IDMLs. Go for it. Oh, yeah, so we're constantly trying to improve GPT-4 and 3.0 as well. And we have an open source, again, repository called OpenAI-DMLs. So if you just send us the cases that the model has fallen, the model is a problem, and we can actually incorporate that into our repo to kind of test and just go by your signals on whether or not it's good. So that's another way of trying to answer that. I'd love to read it. I really want to reinforce what Joanne said there, because e-mails is the best way to steer our roadmap. If you send us, again, we want the negative feedback, if you send us cases where we suck, where we fail, we have an internship, we will make it better. I have one more negative. Oh, please. So actually, we are running something called Esco. We got more than one million people coming and chatting. And then we just sent data to OpenAI and to SOS. The really nice thing about to lay out the open AI and the precise. So the really nice thing about open AI is that they can understand the context very well. The really sad or bad thing is that we have to send all the previous text. That means that easily you can fill up all the tokens. So I think that there are much better ways to do that, to understand the whole context, otherwise working on it. We paid a lot in the past. I think that there are much better ways to understand all the content, otherwise you're working on it. We paid a lot in the beginning. Yeah, thanks for that. Do you have any ideas? Yeah, well, we'll have from a, I mean, we've talked about, so there's two angles here. One is a pricing angle, right, which is like, why do I have to pay this n squared price? And that's something, again, I guess you're right that I actually spent a lot of time on this one too. I mean, we now have 50% off the inventory. Let's not say that that's good enough, but that's like we understand. And, yeah, I guess I think that the, I basically would say the economy is going to keep expanding, and so but I expect actually like where we're going to go, I think the API will evolve. One of the things that I'm really excited about is moving to much more of a kind of you send me messages, you get back messages, so it feels a little bit more chatgy. It's much more stilted. I hold prompt, I send you the prompt, I get back the response. I send you a new prompt, I get back the response. Especially with images, you do not want to be shuttled and go back and forth. I think there will be a technical shift. I think actually this will unlock a lot more creativity. A lot of what we think about is, how do you get, for example, things like plugins. You want to make that really easy for people to use in the API and not have to rebuild all the same sort of serving infrastructure that we have. So we should be able to run those on the server side. Please write in Korean language.
ChatGPT

A: 서비스를 만드는 방법에 관한 질문을 하고 싶습니다. 현재, 우리는 GPT 모델을 어떻게 사용하여 고객이 로봇에서 질문하고 답변하는 것과 관련이 있는지에 대한 방법을 고민하고 있습니다. 우리는 다른 개발자들이 하는 것처럼 작업을 진행하고 있습니다. 우리는 사용자의 질문을 검색 엔진에 넣고, 올바른 맥락을 찾아, 선호하는 맥락을 넣고, GP에 보냅니다. 하지만 실제로는 작동하지 않습니다. 왜 그런지 모르겠습니다. 부정적인 피드백이 있습니다. 하지만 저는 어떤 식으로든 GP가 이 맥락에서 가장 중요한 맥락이 무엇인지 찾을 수 없는 것 같습니다. 예를 들어, 사용자가 은행 제품에 대한 정보를 제공하라고 요청하면, 가장 중요한 정보는 이자율일 것입니다. 그러나 답변은 때때로 이자율을 포함하고 있지만, 때때로는 그렇지 않습니다. 그래서 모든 것이 올바른지 아닌지를 파악하기 위한 직관이 있는지 여쭤보고 싶습니다. 하지만 우리가 많은 지시사항을 넣으면, 우리는 통역사의 수에 고통을 받습니다. 그래서, 질문이 있나요? 이것은 극히 복잡한 질문입니다. 복잡성은 대안의 수입니다. 그래서, 조쉬, 질문이 있나요? 이것은 극히 복잡한 질문입니다. 복잡성은 이것에 대한 전문가입니다. 할 수 있는 많은 것들이 있습니다. 오픈 소스와 오픈 활동적인 책이라는 훌륭한 자료가 있습니다. 어떻게 이것을 할 수 있는지에 대한 예시가 하나 또는 두 개 정도 있을 것입니다. 간단하게 말하자면, 먼저 모델이 어떻게 말할지에 대한 답변을 생성하게 할 수 있습니다. 그리고 그 답변을 조회하는 데 사용할 수 있습니다. 그것이 도움이 될 수 있는 한 가지입니다. 특정 제품 이름과 같은 많은 것들이 있다면, DP25나 그 외의 것을 추가하면, 이것은 TF-IDF 또는 기타 키워드에 기반한 것뿐만 아니라 임베딩 유사성에도 크게 도움이 될 것입니다. 저는 그것이 아마도 제가 말하고 싶은 두 가지 주요한 것들이라고 생각합니다. 또한, 당신이 다루는 맥락의 크기를 줄이는 것도 도움이 됩니다. 저는 영어로 100 또는 200 토큰, 한국어로는 600 정도가 이런 유형의 사용 사례에 더 잘 작동하는 것 같다고 말하고 싶습니다. 현재 GPT-4를 사용하고 있나요? 당신이 우리가 그것을 발표해야 할 때를 생각할 준비가 되었다면 알려주세요. 그래, 분명히, 분명히. 알았습니다, 알았습니다. 아, 그리고 또 한 가지 질문이 있습니다. 그래서, 네, GPT-4는 훨씬 나아졌습니다. 매우 짧은 맥락이 많이 있는 경우, 사람에게는 그렇게 정리되어 있지 않아 보이지만, GPT-4는 어떤 정보를 사용하고 어떤 정보를 빼야 하는지 정말 잘 알고 있습니다. 그래서 GPT-5가 다양한 종류의 정보를 뺄 때 혼동되는 것과는 다릅니다. 그래, 그것에 대해 답변하겠습니다. 더 많이 말하겠습니다. 데이터 공급에 관한 것으로, 모델이 예를 들어, 고객들이 AI에 그들이 원하는 어떤 것에 대한 답변을 원한다는 것과 같은 질문에 대답할 수 있을지 여부를 알고 싶습니다. 그래서 저는 은행에서 이런 금액의 돈을 빌렸고, 상환 계획을 가진 무료 대출을 얻고 싶습니다, 이런 식으로 말입니다. GP 3.5 이상에 대해서는, 상환 계획과 관련하여 잘 작동하지 않습니다. 그래서 계획이 좀 애매해질 것입니다. 저는 이전에 세밀 조정에 관한 것을 말했다고 생각합니다, 그것은 조금 더 느립니다. 아, 세밀 조정으로, 그래요. 정말로 고품질의 데이터로, 그것이... 새로운 IDML들을 위해... 시작하세요. 아, 그래서 우리는 계속해서 GPT-4와 3.0을 개선하려고 노력하고 있습니다. 그리고 우리는 OpenAI-DMLs라는 이름의 오픈 소스 리포지토리를 가지고 있습니다. 그래서 모델이 떨어진 케이스나, 모델에 문제가 있는 경우를 우리에게 보내주면, 우리는 그것을 우리의 리포에 포함시켜 테스트하고, 그것이 좋은지 여부에 대한 여러분의 신호를 판단할 수 있습니다. 그것은 또 다른 방법으로 그것에 대답하려고 하는 것입니다. 그것을 읽고 싶습니다. 저는 그곳에서 조앤이 말한 것을 정말로 강조하고 싶습니다, 왜냐하면 이메일은 우리의 로드맵을 조정하는 가장 좋은 방법이기 때문입니다. 우리에게 다시 보내주시면, 부정적인 피드백을 원합니다, 우리가 실패한, 우리가 부족한 부분에 대한 케이스를 보내주시면, 우리는 인턴을 가지고 있습니다, 우리는 그것을 더 나아지게 만들 것입니다. 또 한 가지 부정적인 점이 더 있습니다. 아, 부탁드립니다. 그래서 사실, 우리는 Esco라는 것을 운영하고 있습니다. 100만 명 이상의 사람들이 와서 채팅을 하고 있습니다. 그리고 우리는 데이터를 OpenAI와 SOS에 보냅니다. OpenAI의 좋은 점은 맥락을 아주 잘 이해할 수 있다는 것입니다. 정말 슬프거나 나쁜 점은, 우리가 이전의 모든 텍스트를 보내야 한다는 것입니다. 이는 쉽게 모든 토큰을 채울 수 있다는 것을 의미합니다. 그래서 저는 그것을 이해하는 더 나은 방법이 있다고 생각합니다, 그렇지 않으면 그것을 위해 작업 중입니다. 우리는 처음에 많이 지불했습니다. 음, 그것에 대해 감사드립니다. 아이디어가 있나요? 네, 우리는 물론, 저희가 이야기하고 있던 것이 있습니다, 그래서 이것은 두 가지 관점이 있습니다. 하나는 가격적인 관점입니다, 즉, 왜 저는 이 n 제곱 가격을 지불해야 하는지에 대한 것이고, 그것은 또 다른 것입니다, 다시 말하면, 저는 사실 이것에 많은 시간을 보냈습니다. 음, 지금 우리는 인벤토리를 관리하는 것에 관한 기술적인 부분을 다루고 있습니다, 왜냐하면 우리는 많은 수의 토큰을 가지고 있고, 우리가 관리하는 것은 매우 복잡하고, 때로는 이것이 매우 빠르게 실행되는 경우도 있습니다. 그래서 우리는 많은 일을 해야 합니다. 그리고 저는 이것이 더 복잡하다고 생각합니다. 음, 그러나 아마도 가장 중요한 것은, 저희가 만들어내는 모든 것에 대해 합리적인 이유를 가지고 있어야 한다는 것입니다. 이것은 좋은 물건입니다. 이것은 저희가 우리의 모델을 사용하는 방법에 대해 생각하는 방법입니다. 저는 이것이 매우 중요하다고 생각합니다. 이것은 고객의 문제를 이해하는 것입니다. 그래서 저희가 무엇을 만들어내는지에 대한 것입니다. 저는 이것이 잘 안 되고 있다는 것을 압니다. 이것은 저희가 이것에 대해 더 많이 생각하고, 이것을 해결하려고 하는 노력을 많이 기울이고 있다는 것을 의미합니다. 우리는 또한, 우리의 아키텍처와 관련하여 많은 것을 배우고 있습니다. 이것은 저희가 그것을 좋아하게 만드는 것입니다. 이것은 저희가 우리의 토큰을 최대한 활용할 수 있도록 하는 것입니다. 이것은 저희가 최적의 성능을 낼 수 있도록 하는 것입니다. 이것은 저희가 토큰을 더 잘 활용하고, 더 나은 결과를 얻을 수 있도록 하는 것입니다. 이것은 저희가 우리의 시스템을 개선하는 데 중요한 부분입니다. 그래서 저희는 이것에 대해 많은 노력을 기울이고 있습니다. 그래서 이것에 대해 감사드립니다. 또한, 저희는 이것이 저희의 작업을 개선하는 데 도움이 되기를 바랍니다. 그리고 저희는 이것을 통해, 저희는 이것을 더 잘 만들 수 있습니다. 이것은 저희가 이것을 통해 얻을 수 있는 가장 큰 장점입니다. 이것은 저희가 저희의 토큰을 최대한 활용할 수 있도록 하는 것입니다. 이것은 저희가 이것을 좋게 만드는 것입니다. 이것은 저희가 이것을 더 잘 만들 수 있도록 하는 것입니다. 그리고 저희는 이것을 계속 개선하려고 합니다. 이것은 저희가 이것을 더 잘 만들 수 있도록 하는 것입니다. 그래서 저희는 이것에 대해 더 많은 작업을 하고 있습니다. 그리고 저희는 이것에 대해 더 많은 시간을 보낼 것입니다. 그래서 이것은 저희가 지금하고 있는 일입니다. 이것은 저희가 이것을 더 잘 만들 수 있도록 하는 것입니다. 그리고 저희는 이것을 계속 개선하려고 합니다. 이것은 저희가 이것을 더 잘 만들 수 있도록 하는 것입니다. 그래서 저희는 이것에 대해 더 많은 작업을 하고 있습니다. 그리고 저희는 이것에 대해 더 많은 시간을 보낼 것입니다. 그래서 이것은 저희가 지금하고 있는 일입니다.

Role: Professional IT Translator Tasks: - 윗부분와 문맥을 활용하여 한글로 번역해주세요 - 질문과 답변을 잘 구분해서 화자가 다른 경우 화자를 A,B,C와 같이 적어주세요 - STT의 결과이니, 어색한 표현을 고쳐주세요 I think that there's a lot of interesting opportunities from our perspective to help solve some of these problems and just open up more opportunity. So as we are around 100 years old, it's getting stronger and faster. So I think the startup companies may have some more challenges to differentiate themselves as a company who is still doing the same thing that we are doing. So as a well-known startup investor, what's your image? How do companies help you differentiate themselves? and we have some of you guys that want, what's your image, you are some companies that help you differentiate things like this? I think technology alone is very, very big differentiator. Open my eyes, I'll give some example of how they're coming, but there are companies with an actual technical model. And even then you can argue about how much we really have and what's happening with that source. We are only as good as our ability to stay at the forefront of innovation. And I think that's true for companies too. You can't, you can't imagine a commodity that's hard and it's not usually how it works. So the fact that there's a cool new platform does not excuse you from the hard work of building a business. You still gotta focus on customers, build up modes, build up differentiation, figure out some sort of network effect. All of the standard things that it takes to differentiate a business still apply. Access to a technology is almost never a barrier. So I would like to go from Sahara, to a couple of these big companies who are pretty happy with it. You've got all the old networks. How do you imagine the ad revenue? Our expect, the thing that is most special about OpenAI is our ability to reliably go and figure out the next innovation each year. So everybody's chasing us right now on the LLMs. We are off and running off the next day. And that is the most interesting end, because otherwise you're just in this sort of like, darkness. So what is next for you? Are you going to teach third language? We'll tell you when it's ready. How did you foster the culture of repeatedly creating and waiting? Pain and suffering, honestly. But I think you just keep leading into the problems. At first, it was very scary, because you feel like a movie studio, because you realize that every time you're getting your big hit, now you're starting from scratch. And I think that over time, we've built up a lot of meta infrastructure and a lot of technology. You have all these processes that you've run before, you've seen where they fail. A lot of it is even getting people who come from the ML background to work well with people who come from the software background.
The previous model used in this conversation is unavailable. We've switched you to the latest default model.
ChatGPT

저희 관점에서는 이러한 문제를 해결하고 더 많은 기회를 창출할 수 있는 매우 흥미로운 기회들이 있다고 생각합니다. 우리는 이미 100년의 역사를 가지고 있으며, 더욱 강력하고 빠르게 성장하고 있습니다. 그래서 스타트업 기업들은 아직 우리가 하는 것과 동일한 일을 계속하는 기업들과의 차별화에 더 많은 도전을 겪을 수도 있습니다. 유명한 스타트업 투자자로서, 어떻게 기업들이 차별화를 도와주는지 어떤 이미지를 가지고 계신가요? 그리고 우리와 같은 일을 하는 기업들이 도움을 받을 수 있는 기업들에 대해 어떤 생각을 가지고 계신가요? 저는 기술 자체만으로도 매우 큰 차별화 요소가 될 수 있다고 생각합니다. 예를 들어 OpenAI는 실제로 기술적인 모델을 가진 기업들과 경쟁할 수 있습니다. 심지어 그것에 대해 얼마나 많이 가지고 있는지와 그 소스의 상황에 대해 논의할 수도 있습니다. 우리의 혁신력이 선두에 있을 수 있는 한, 우리의 능력이 좋을 수록 좋다는 것이 가장 중요한 점입니다. 현재 모두가 LLMs에 뒤졌다가 우리는 다음날부터 다음 단계로 진행합니다. 그리고 그것이 가장 흥미로운 점인데, 그렇지 않으면 당신은 그저 어둠에 있을 뿐입니다. 그래서 당신에게 다음은 무엇인가요? 제가 준비되었을 때 알려드릴게요. 반복적인 창조와 기다림의 문화를 어떻게 육성했나요? 고통과 고난, 솔직히 말해서 그것이었습니다. 그러나 당신은 계속해서 문제에 직면해야 합니다. 처음에는 매우 무섭다고 느껴지기도 했는데, 대성공을 거두고 나면 이제 다시 처음부터 시작해야 한다는 것을 깨닫게 됩니다. 그리고 시간이 지남에 따라 우리는 많은 메타 인프라와 기술을 구축해 왔습니다. 이전에 실행해 본 프로세스들이 어디에서 실패했는지 알고 있습니다. ML 배경을 가진 사람들이 소프트웨어 배경을 가진 사람들과 원활하게 협업할 수 있도록 하는 것조차도 중요한 부분입니다. (화자: A)

흥미로운 말씀 감사합니다. 이어서 다른 분의 의견도 들어보고 싶은데요. 조선말로 말씀하시면 좋을 것 같아요. (화자: B)

다른 분들의 생각도 듣고 싶습니다. 조선말로 자유롭게 이야기해주세요. (화자: C)

이 기술적 문제에 대한 저희의 솔루션은 문서의 토큰 수를 최소화하는 것입니다. 즉, 모든 내용을 문맥을 이해하는데 필요한 최소한의 양으로 요약하는 것입니다. 이렇게 하면 GP 모델의 성능을 향상시킬 수 있습니다. 또한 관련된 키워드를 사용하여 검색을 더 정확하게 할 수 있으며, 보다 효과적인 모델 훈련을 위해 다양한 기술적 방법을 사용할 수 있습니다. 이를 통해 더 나은 결과를 얻을 수 있을 것입니다. (화자: A)

저희의 질문에 대한 답변을 해주셔서 감사합니다. 그런데, 저희는 OpenAI를 통해 데이터를 전송하고 있습니다. 그런데 이전 텍스트를 모두 전송해야 한다는 점이 아쉽습니다. 그렇게 되면 토큰을 쉽게 소진하게 되는데, 이에 대한 더 좋은 방법이 있을까요? (화자: B)

그런 점에 대해서는 현재 가격 측면에서 고려 중입니다. 또한 API에서 메시지 전송 및 수신 방식을 개선하고자 합니다. 채팅 형식으로 더 유연하게 대화할 수 있도록 하여 상호 작용을 더욱 원활하게 만들려고 합니다. 이를 통해 더 나은 경험을 제공할 수 있을 것입니다. 또한, 향후 플러그인과 같은 기능들을 더욱 쉽게 사용할 수 있도록 하는 기술적인 변화가 있을 것입니다. (화자: A)

그렇군요. 그럼 기대하겠습니다. 그리고 마지막으로, OpenAI의 다음 계획은 무엇인가요? 새로운 언어를 가르치기도 할 건가요? (화자: C)

다음 계획에 대해서는 준비가 됐을 때 공지할 예정입니다. (화자: A)

저희가 지속적으로 창조하고 기다리는 문화를 육성하는 데에는 어떤 방식을 사용했는지 궁금합니다. (화자: B)

솔직히 말해서 고통과 어려움이었습니다. 하지만 문제에 직면하고 해결책을 찾기 위해 노력했습니다. 처음에는 영화 스튜디오처럼 느껴져서 큰 성공을 거둔 후에 다시 처음부터 시작해야 한다는 불안감이 있었습니다. 그러나 시간이 지나면서 우리는 많은 메타 인프라와 기술을 구축했습니다. 이전에 실패한 경험을 통해 어떤 프로세스가 실패하는지 알게 되었습니다. ML 배경을 가진 사람들과 소프트웨어 배경을 가진 사람들이 원활하게 협업할 수 있는 방법을 찾는 것조차도 중요한 부분입니다. (화자: A)

Role: Professional IT Translator Tasks: - 윗부분와 문맥을 활용하여 한글로 번역해주세요 - 질문과 답변을 잘 구분해서 화자가 다른 경우 화자를 A,B,C와 같이 적어주세요 - STT의 결과이니, 어색한 표현을 고쳐주세요 By default, those people just think about problems differently. They're just going to not respect each other. We solved 10 versions of that problem at increasing level sophistication and so I think it's just like there's no one answer for these things it's just you just keep doing like a thousand little things like I think semiconductors is maybe a good analogy for a building process or something where it's just like there's all these components they all fit together you need to like solve hard problems at every layer of the stack and you just got to keep keep leaning into. I think that's actually quite useful advice for starters, so if anybody else wants to add on to this, I think it'd be great. Oh, yeah, I think it's one of the big differentiators between OpenAI and the other companies developing all of this is, it's really cultural. Like, everyone, you feel like, really wants to build their own future. It's not like people are selfish, they're there to kind of publish what they're doing, you know, kind of gain their own personality. We're all here, we know our role to play, and we really want to push for this to be good. And I think that really creates an environment where all the teams are working together and they see what comes. We really are a team. I've got a question about... Really quickly, we have five minutes, and right at 50 we have to have people enter to the next session, so quick time check, we're going to cut it right at 50 we have to have people enter to the next session so quick time check we're gonna cut it exactly at 50 So let's say a company have a corporate trade data, like a size of several billion, several hundred billion to a trillion to a billion level. It's not trained on level where fine tuning a level. So there's two choices. A company can use a small amount of fine tuning, or they can train like 10 billion to 20 billion product models on their own. So, what do you think about those two solutions? I do think right now we don't have a good answer for how to leverage that scale of data through our platform. So I think that you can use subsets of that data, you can do retrieval, you can do those kinds of things, but to really find and bake it in, that's where we're not there yet. I would love to though, I would really love to, because we've actually built really amazing training technology. And again, it's just like, there's so many little pieces of the stack you have to get right, and just lead performance on the table otherwise. And so I think we should be able to be in a world where our fine-tuning API, as long as the compute is there, right, which that's the real thing that's hard with all this other software stuff, at least it's doable, but you don't have the silicon, you don't have the silicon. As long as we can get compute lined up behind it, then I think it would be really incredible. We'd actually really love to have companies that go and fine tune a GP4 on that scale of data, because I think you would just do really incredible things in all these domains. All right, are there any questions? There's only a few minutes left. Maybe we can all stand up, split up, then we can address the last set of questions. Sure, sounds good. I'm not sure how that's going to work. You have to share. People, just find people you want to ask a question. Please write in Korean language.
ChatGPT

기본적으로, 이들 사람들은 문제에 대해 다른 방식으로 생각합니다. 그들은 서로를 존중하지 않을 것입니다. 우리는 문제의 10가지 버전을 다양한 수준의 세련함으로 해결했고, 이러한 것에 대한 단일한 해답은 없다고 생각합니다. 수많은 작은 것들을 계속 해야 합니다. 예를 들면 반도체가 건설 과정이나 기타 작업과 유사한 좋은 비유일 것입니다. 각각의 구성 요소들이 함께 조립되어야 하며, 스택의 모든 레이어에서 어려운 문제들을 해결해야 합니다. 그리고 계속해서 이러한 문제에 직면해야 합니다. 이것은 스타트업에 대한 매우 유용한 조언이라고 생각합니다. 다른 분들도 이에 대해 추가할 내용이 있다면 좋겠습니다. (화자: A)

네, 저희가 OpenAI와 다른 회사들 사이의 큰 차이점 중 하나는 문화적인 면입니다. 모두가 자신의 미래를 구축하고자 하는 열정을 가지고 있는 것 같습니다. 사람들이 이기적이지 않고 자신의 한계를 넘어서려는 것이죠. 우리는 모두 이곳에 있고, 우리의 역할을 알고 있으며 이것을 좋게 만들기 위해 노력하고 있습니다. 이렇게 하면 팀 전체가 협력하며 성과를 이룰 수 있습니다. 저희는 정말로 한 팀이에요. (화자: B)

제가 질문이 있습니다. 시간이 얼마 남지 않았지만, 기업이 수조에서 조종되는 데이터를 가지고 있는 경우, 작은 규모의 파인튜닝을 사용하거나 자체적으로 1000억에서 2000억 개 모델을 훈련시키는 것 중 어느 것이 더 좋은 해결책인지 생각하시는지 궁금합니다. (화자: C)

현재로서는 그런 규모의 데이터를 플랫폼을 통해 어떻게 활용할지에 대한 좋은 해답이 없다고 생각합니다. 데이터의 하위 집합을 사용하거나 검색 등을 수행할 수는 있지만, 실제로 해당 데이터를 포함하고 활용하기에는 아직 그 정도까지 도달하지 못했습니다. 하지만 저희는 그럴 수 있었으면 좋겠습니다. 우리는 정말로 놀라운 훈련 기술을 개발했습니다. 다만, 여러 요소들을 올바르게 맞추는 것이 중요하며, 그렇지 않으면 성능에 손실이 발생할 수 있습니다. 따라서 컴퓨팅 리소스가 충분히 제공된다면, 실제로는 정말 대단한 일들을 이룰 수 있을 것입니다. 기업들이 수조 규모의 데이터로 GP4를 파인튜닝한다면, 모든 분야에서 정말 대단한 결과를 얻을 수 있을 것입니다. (화자: A)

OpenAI에서 2023년 6월 9일에 진행한 "Round Table Talks with OpenAI in Seoul"에 참가한 리턴제로(Return Zero Inc.)의 김동우(Dongwoo Kim)입니다. 위의 내용은 Whisper large-v2 모델을 통해 좋지 못한 음질의 내용을 영어 텍스트로 만들었고. (Speech-to-text) 결과물을 전혀 수정없이 GPT-4에 넣고 중간에 limit이 걸려서 GPT3.5모델을 사용하여 기계번역한 내용입니다. 실제로 OpenAI를 활용하는 개발자들을 초청해서 의미있는 질답을 진행하기 위해 노력하는 OpenAI의 경영진과 실무진으로부터 이 생태계와 플랫폼을 만들려는 노력이 많이 느껴졌습니다. 제가 다니는 스타트업에서도 GPT-3.5-turbo를 활용하여 서비스를 만들어보고 있는데, 더욱 더 모델들이 발전할 것이고 가격이 싸질 수 있다는 이야기를 듣고, 한국어에 대한 개선도 언젠가는 되리라는 희망을 조금 품게 되었습니다. 지금까지 번역된 위 내용에서 다른 스타트업들이 얻을 수 있는 인사이트나 도움이 될만한 내용도 추가로 요약해주세요. Please write in Korean language.
ChatGPT

안녕하세요, 김동우님!

OpenAI의 경영진과 실무진들이 의미있는 질답을 통해 개발자들을 초청하고 생태계와 플랫폼을 발전시키기 위해 노력하고 있다는 사실을 느낄 수 있었습니다. GPT-3.5-turbo를 활용하여 서비스를 개발하고 계시다니 정말 멋지네요! 더 나아가 모델들이 발전하고 가격이 저렴해질 것이며, 한국어에 대한 개선도 이루어질 가능성이 있다는 이야기를 듣고 기대가 조금이나마 생기셨다니 기쁘게 생각합니다.

번역된 내용에서 다른 스타트업들이 얻을 수 있는 인사이트나 도움이 될만한 내용을 추가로 요약해 드리겠습니다:

  1. 문제 해결과 협업: 스타트업에서 문제에 대해 다른 시각으로 접근하고 서로를 존중하는 문화를 형성하는 것이 중요합니다. 수많은 작은 문제들을 해결하고 계속해서 도전하는 것이 성공의 열쇠입니다.

  2. 기술 혁신과 경쟁력: 기술 자체가 큰 차별화 요소입니다. 현재와 미래의 기술 트렌드에 대한 이해와 혁신을 통해 경쟁력을 갖출 수 있습니다. 단순히 기술을 활용하는 것이 아니라 고객에 초점을 맞추고 차별화 요소와 네트워크 효과를 찾아야 합니다.

  3. 데이터 활용과 파인튜닝: 대규모 데이터의 활용은 현재 OpenAI의 플랫폼에서는 제한적이지만, 데이터의 일부를 사용하거나 검색 등 다양한 방법으로 활용할 수 있습니다. 더 나아가 데이터를 바탕으로 GP4를 파인튜닝하는 것은 탁월한 성과를 얻을 수 있는 방법 중 하나입니다.

  4. 커뮤니케이션과 피드백: OpenAI는 사용자의 의견과 피드백을 소중히 여깁니다. 사용자들의 케이스와 문제점을 제공하면 OpenAI는 그에 맞게 플랫폼을 개선해 나갈 수 있습니다. 마찬가지로 스타트업도 사용자의 피드백을 듣고 문제점을 해결하여 더 나은 서비스를 제공할 수 있습니다.

  5. 끊임없는 혁신과 도전: 혁신을 위해서는 문제에 직면하고 그에 대해 대응하는 과정에서 고통과 어려움을 겪어야 합니다. 그러나 이러한 도전과 실패를 통해 스타트업은 지속적으로 성장하고 발전할 수 있습니다. 어려움을 두려워하지 말고 문제에 더욱 집중하고 해결책을 찾아나가는 것이 중요합니다.

이러한 내용들이 다른 스타트업들에게 인사이트와 도움이 되었으면 좋겠습니다. 김동우님과 리턴제로 팀의 미래 성공을 응원합니다. 언제든지 추가로 질문이 있으시면 말씀해주세요!

Continue this conversation
|Terms of use|Privacy policy