AI 소프트웨어 엔지니어: AI를 사용한 구축 및 배포

최초의 AI 소프트웨어 엔지니어인 Devin이 다양한 도구와 브라우저 상호 작용을 사용하여 단계별 문제 해결 및 프로젝트 개발을 시연합니다.

### Devin으로 AI 시스템 구축하기: 단계별 가이드
* 🧠 프로젝트를 독립적으로 기획, 구축, 배포할 수 있는 **최초의 AI 소프트웨어 엔지니어인 Devin**을 소개합니다.
* 🛠 Devin의 단계별 프로세스에는 **프로젝트 계획, 브라우저 기반 환경에서의 코딩, 오류 디버깅**이 포함됩니다.
* 이 프로젝트는 다양한 API 제공업체에서 **LAMA의 성능을 벤치마킹**하여 **Devin의 역량**을 보여주는 것입니다.
* 🤖 Devin은 각 역할(기획자, 작가, 편집자)에 따라 특정 모델을 선택하여 **GPT 모델과 API를 사용**하여 작업을 완료합니다.
* 🔑 상담원은 Devin의 워크플로에서 중요한 역할을 담당하며 **각 작업에 대한 컨텍스트, 기술 및 명확한 지침**을 제공합니다.
* 📊 이 프로젝트에는 **환경을 설정**하고, **필요한 도구를 설치**하고, **상담원을 위한 체계적인 워크플로우**를 만드는 작업이 포함됩니다.
* 🚀 최종 결과물은 전체 스타일링이 적용된 웹사이트로, **엔드투엔드 프로젝트를 완성하는 Devin의 능력**을 보여줍니다.
* 🌐 이 동영상은 **AI와 상호 작용하여 다양한 프로젝트와 작업에 사용할 수 있는 시스템**을 만드는 방법을 보여줍니다.

Hey, I'm Scott from Cognition AI. Today, I'm really excited to introduce you to Devin, the first AI software engineer. Let me show you an example of Devin in action. I'm going to ask Devin to benchmark the performance of LAMA on a couple of different API providers. From now on, Devin is in the driver's seat. First, Devin makes a step-by-step plan of how to tackle the problem. After that, built the whole project using all the same tools that software engineers use. Devin has its command line, code editor, and even its browser. In this case, Devin decides to use the browser to pull up API documentation so that I can read up and learn how to plug into each of these APIs. Here, Devin runs into an unexpected error, actually decides to add a debugging print statement, reruns the code with the debugging print statement, and then uses the error in the logs to figure out how to fix the bug. Finally, Devin decides to build and deploy a website with full styling. As the visualization, you can see the website here. All this is possible today because of the advancements that we've made in both reasoning and long-term planning. Have you enjoyed the video? It was quite a hot topic when it was released two months ago. Now, AI interacting with AI, creating a system together, is shown. So, from now on, will you utilize a system created by someone else well, or will you create such a

I believe it will make a difference, so I will start by introducing my AI. I am an AI that wants to learn, study, and work with you on projects that I want to do and create AI projects that I want to build. I am thinking of creating a channel like this. So, let's start with the simple idea of ​​having a conversation with the AI we created earlier and working on the service together. Today's materials will refer to the site created by Professor Angdu. It is a deep learning AI site where all materials are freely available. Among them, I chose the Wolti AI Ethics System, which is related to Debin. I thought it was necessary to clearly explain the concepts in Korean to you quickly through the lectures. I tried this once, and if you do the codes here, it will be like this.

I react right away, but when I bring this into my computer's environment, I need to set the API key first. And I need to install these things. Now, while doing these things, setting up the environment can be a bit tricky. So, even though I understand the concept on my computer, there are cases where I need to redo these things when I want to set it up and use it myself. Therefore, to make it easier to do these things, I make it easier to understand. My goal is to expand and broaden the theory of different services by going through trial and error first and then applying this further. When you look at CroAI, it explains why CroAI is used and how it should be used. First of all, you should know the difference between LM and agents. LLM is a large language model, so you need to understand the model itself.

It means that an agent is someone who helps this model perform task 1 more clearly. When they help, they ask, What are you doing? How do you work? What skills do you have? They have skills or solutions, so they can use them as they wish to help you in this way. So, what problems does this solve? When we use GPT, this becomes important. This concept becomes interesting. Let me explain this a bit and then move on. L&M itself is just a model that can express a large language. So, when certain things are given to it, it can generate something through this model. When instructions are given, it just performs like this, one model at a time.

GPT models have the context of what I've talked about before, so they remember all of that. Therefore, when I ask how to correct something, they can do that. They can create it like this when I ask them to. In this way, chat models are created. So, there is L&M itself, and there is a chat model. With just this chat model, it's too difficult to keep throwing things and asking for things all the time. Therefore, to make this chat model more clear and better utilized, we introduce what is called an agent. So, to explain the agent a bit better because it's getting messy, let's go back a bit. An agent gives a clear role and provides the current situation. For example, in our company, when we work together, if our boss tells us to do something, even if we don't understand the context, we still have to do it. Why do we have to do it now?

Without knowing what to do, not understanding the background, and being in a state where it's hard to proceed, it's really difficult. That's why we clearly need to choose what we should do now. And explain why we need to do that, provide the backstory, and create a context in the current situation. After creating this context, when the command is given, you can understand what needs to be done. So, if the agent thinks like that, the agent will do that. If the agent doesn't have the skills, for example, if I can't create a PowerPoint presentation, and I'm asked to do a PowerPoint task, I can't do it. If I don't know Excel and I'm asked to do Excel, I can't do it. So, we need to provide education to acquire the skills to handle such tasks in detail. This is how it is done. So, when you see this, those of you watching may probably know.

It's like this. The client asked us to create a system, but when we made the system, it didn't feel right. When the boss instructed us to do something, and we did it like this, Is this what you wanted? It doesn't feel right. In such cases, we ask for clearer instructions. We define the description clearly for such tasks. This task may include defining the requirements for the task, or it may involve skills. It could be skills like being able to use Excel, being able to code in Excel, being able to save this as a text file, being able to explore the internet, parse it, and crawl data. So, by creating this task, we can include the skills here, and if we include it in the task, then this will...

You will be able to do those things later, right? For example, is it fun? It's not just fun for me. For example, for now, you can think of it as having those things. Here, if you look, you set the tasks to description output like this, and then the description and these two now become parameters. As parameters within the task, they are set as variables like this. A task is a function, a task function. Where is this? We now call it here like this, Task Agent Crew, Agent. We just did the task agent and the task, and now, when the crew is set up, here, if you see, the crew sets all the agents like this, and then when all the agents and tasks are bundled together, it's called a loop like this, and they are made into one. If they are separate, they cannot work, right? So, we gather the agents and the tasks, and the boss is explaining this. Explaining this, now if we go to step 0.

Don't explain in the first step, go into more detail. By doing it in the second step, go into more detail like this, and explain it like that. So, it's quite fun. When we execute it, when we do our project, we do a kickoff. Ah, crew, crew, agents, tasks like this, explain the steps, in what steps to do the kickoff. If we do it like this, we kickoff and start. Um, so, now that we've explained the concept all the way here, we will now try to set up the environment and proceed with this. The way I do it is, first of all, there are these agents, right? We will use different LMs for each agent because the content planner, writer, and editor may not have the same level of difficulty. Some may find it easier to plan and work on it, while editors may continue to edit it a bit mischievously after seeing this.

While I was trying it according to that code earlier, especially when I do it in Korean, the editor keeps stuttering. So, when it comes to writing, the writer needs to write well. Therefore, I will try using a better model for writing. I used the OpenAI GPT 3.5 Turbo model. The 3.5 Turbo model is quite affordable. For books or similar cases, it is not a big burden. You can refer to the table I prepared below in Excel format. In cases like books, there is a site called Grok. Grok is a cloud platform that provides APIs so that we can use it. It can be seen as a similar service to AWS, which is currently available for free. This platform provides models completely, so the difference with regular models is that it is considerably faster.

I will show you how to do this. For the Rama 370B model, in most computers, it doesn't come out this fast, anyway, you can use it directly on the Groi site. And with the Groi API, when you go to the developer console on the Groi Cloud, I will put this in the document. Then, you click on this and go in. Here, if you press the API key, like this, there is an API key here. Please put in any name for this, I will put it for testing. By doing this and submitting, the API key will be created. Of course, I will delete this. After creating this, when you press money, I downloaded the document like this. I will install the dot B here. What this is, is a document that makes it easier to set up the environment. So, after doing this, if you do this, you can change the extension to use it.

That's fine. Next, you should use this as a connection program. If you use Notepad, then after connecting it, you can open the OpenAI API. The OpenAI API key is SK, you can use this, and then paste the copied key here directly. If you do that, save it, and when you run it, it will properly pull this. Let's take a look. First, let's try installing the virtual environment. Just copy this by pressing Ctrl+C and then Ctrl+V. Since I already installed this, I'm doing this to install it again. By running conda create -n, it means it will install a virtual environment with that name, and if you specify Python 3.10, it will install Python 3.10. This way, the libraries to install will appear here, and then the installation will proceed as planned. The key we saved earlier was the API key. This is the same name as the one saved earlier.

Pull up the key, so if you start installing it from now on, we will try running them one by one. Even though my computer specs are not that great, you can use it right away because it is provided on the Gro site. Next, the agent set up the roles, goals, and backstory like this, and you can see what content was set up above. Simply put, I hope the planner organizes the list a bit, and I would like the writer to use this list as a guide to write in a certain way. It's important to write this well from the beginning so that it can be corrected later. However, since I can't fully trust what is written here, I would like the editor to follow some best practices, check for grammar mistakes, and correct any controversial points.

There is no specific skill involved, but as I mentioned earlier, organize the requirements for the task in a description format and determine how the output should be presented. This skill is important now. The agents should be set as planners here, and the roles and settings made above should be input into the agents in the same way. When you put them together like this and kick off, it will start running. So, start with the planners one by one. When I created the agents earlier, there were roles, goals, and backstories, but there is also L&M itself, and this L&M is just set as BT, right? By explaining this through verbs, when you set it to true, it turns on, and when you set it to false, it enlarges. Similarly, for the lighter, you should do it in the same way. For this one, by setting it as L&M, it will open up.

When the OpenAI model throws the results to him, he will write it down. Here, it is set as Provide by the content planner. By receiving instructions for the outline and the instructions of the outline from him, he creates content with insights, purpose, and informativeness. For editors, they check if this is a balanced point, if there are no errors, and so on, and create it like this. It's a long story. If you proceed like this, each of them should be set by agents. The test was done by the planner earlier, right? It was done through the planner, and that was sent by creating a writer object earlier. And then, the test was edited like this. Then, the planner, writer, and editor are set, and it was done from planning to writing to editing, like this. I hope you have a great day.

I will try to create five colors for this, so I searched for trends, key players, and news notes like this to think about the target audience's pain points and interest points, and then I outlined it like this. I have covered the main points from SEO keywords to target readers, warehouse data, and completion, so you can see that I have balanced and covered everything I requested overall. So, I had a content writer write this, and the content writer wrote it right away. They call it a markdown format, so when you put two asterisks in front, it emphasizes, so it will all come out like this, and if you keep emphasizing the main points, it will look better in markdown. So, not too long, like this.

When you finish the chain like that and deliver it now, the editor found it a bit short. It was a bit rushed, but if you just send it as it is, it's a bit short as per the line. The editor improves it a bit and makes it better, so the communication is much better. By going through this process one more time, we can see that the results come out well. So, after confirming the results that came out, the five best habits to spend a day are like this, with the key points, and it's done well up to the conclusion. Today, we briefly talked about CroWare, so I just tried writing a blog post in Markdown format. From the next lecture onwards, we will proceed to make it more interesting by adding images and other things.

Thank you for taking the time, and if you enjoyed this, please support us through subscribing and liking. We ask for your generous support. Thank you.

내가 직접 만드는 AGI, Crew AI를 활용해서 Multi Agent를 만들어봐요_1.Intro

일하는 ai

Devin: 현업에 투입된 최초의 AI 소프트웨어 엔지니어

Devin은 단계별 계획을 따라 일반적인 소프트웨어 엔지니어링 도구를 사용하여 다양한 API 제공업체에서 LAMA의 성능을 벤치마킹합니다.

Devin은 예기치 않은 오류를 발견하고 코드를 디버그한 후 전체 스타일이 적용된 웹사이트를 배포합니다.

이 동영상은 Devin이 추론하고 장기적인 계획을 세우고 AI와 상호 작용하는 능력을 보여주며 AI 기술의 발전을 강조합니다.

에이전트는 GPT 모델이 작업을 이해하고 솔루션을 제공하도록 도와주어 응답을 생성하는 데 있어 효율성을 향상시킵니다.

에이전트의 역할은 모델에 명령을 내리기 전에 컨텍스트를 만들고, 배경 지식을 제공하며, 현재 상황을 명확히 하는 것입니다.

상담원이 GPT 모델 내에서 다양한 작업을 효과적으로 처리하려면 교육과 기술 개발이 필수적입니다.

성공적인 업무 중심 시스템의 핵심은 관련된 요구 사항과 기술을 명확하게 정의하는 것입니다.

작업은 하나의 기능으로 취급되어야 하며, 상담원과 스태프가 서로 협력하여 조율된 루프에서 작업해야 합니다.

또한 시스템에는 각 단계에 대한 자세한 설명과 작업 시작을 위한 명확한 킥오프 프로세스가 포함되어야 합니다.

상담원에게 다양한 난이도를 적용하면 시스템 내에서 원활한 진행을 보장할 수 있습니다.

Groi 클라우드 플랫폼을 사용하여 OpenAI GPT 3.

Groi 클라우드 플랫폼은 AWS와 유사하게 OpenAI GPT 3.

5 터보 모델에 빠르고 경제적으로 액세스할 수 있습니다.

Groi 개발자 콘솔에서 API 키를 얻으면 사용자는 OpenAI API에 쉽게 연결하여 빠르고 효율적으로 작성할 수 있습니다.

이 플랫폼은 환경 설정을 위한 문서도 제공하며 가상 환경에 대한 원활한 설치 프로세스를 제공합니다.

작업 실행을 위한 에이전트 설정 및 실행

에이전트의 역할, 목표 및 배경 스토리를 설정하여 작업 요구 사항을 정리하세요.

기획자, 작가, 편집자 등 에이전트는 작업 수행을 위해 구체적인 방식으로 설정되어야 합니다.

테스트는 기획자를 통해 진행되었으며, 기획, 작성, 편집 등의 과정을 거쳤습니다.

마크다운 형식으로 매력적인 콘텐츠 만들기

마크다운 형식을 사용하면 콘텐츠의 핵심을 강조하는 데 도움이 될 수 있습니다.

SEO 키워드, 타겟 오디언스의 관심사 및 데이터의 적용 범위의 균형을 맞추는 것이 중요합니다.

편집자의 의견을 반영하여 콘텐츠를 다듬는 과정을 거치면 더 나은 결과를 얻을 수 있습니다.