Hello, everyone. This is the AI Academy. Today, I’m here to meet you on this lovely morning. I woke up at 2 o'clock and spent some time exploring and testing various things. About three hours ago, news came out from OpenAI developers, and I would like to introduce it to you. Today, at the OpenAI Developer Day in San Francisco, they are launching a number of new features that will be useful for developers. A lot of new features are already coming to the platform. 

What’s interesting is that, if you look at it, the Playground—I'm not sure if many of you use it, but I have been using the Playground since its early days. What I want to tell you is that the performance of GPT-4 and Mini is almost similar. There isn't much difference between GPT-4 and Mini, but the price is 20 times lower.

I mentioned that the advantage of using GPT-4 on Mini for regular users is that while the maximum context window for GPT-4 is 4,095 tokens, GPT-4 on Mini allows for a context window of about 16383 tokens. This makes it more favorable for writing longer texts, even though it may not be sufficient for extremely lengthy documents. For those who do not want to pay the 30,000 won monthly fee for the GPT-4 Plus version, I have often recommended utilizing the GPT-4 on Mini version through the API Playground. However, it is important to note that while images can be uploaded, they cannot be generated or searched.

There is a point in this part, but generally, when you are writing or have a specific purpose, the free version of GPT-4, or Mini, is almost free to use, so I recommend you take advantage of it. What I want to talk about today is that in the Playground, you need to write user system instructions. For example, in GPT, you can see that in the settings, there is a section for customized instructions where you need to brainstorm and come up with your own system instructions. This is a similar function in the Playground where you need to write your system instructions. However, this allows you to easily create AI agents. It’s important to use system instructions effectively instead of struggling with complex content that you might not understand.

When you press the generator, what you currently have is a prompt engineer agent setting. However, it can also be set as a creative writing expert agent. There's no need to overthink it. It is labeled as a pre-beta version. So, when you select creator, the dialogue window changes to that of a creative writing expert agent. Then, it prompts you to perform tasks as a creative writing expert. It essentially creates a session for you automatically, optimizing it. In this function, you define the necessary functions. Similarly, when you press the generator here, if you don't know how to use this coding language, you wouldn't have been able to write it. But now, that's not the case; you can paste the code, but this is designed for those who may not be able to use code. For example, let's take a function as an example.

If you click on the creator for webtoon writing, it will allow you to write creatively. It will generate content for your webtoon, including the title, a list of episode summaries, episode numbers, descriptions, the genre of the webtoon, and a string describing the target audience. This feature is not just for developers; while it is useful for them, using GPT-4 on the Playground is also a very handy tool for anyone looking to write. You can use GPT-4 at a fraction of the cost, almost for free. Writing with this tool costs about $0.01 or even less per token, so you can use it almost at no cost. Even if you use it a lot in a month, it won't exceed $1.

The usage will depend on the amount, but to explain again, the functionality has now been announced through the developer channel. The parts mentioned here are for developers, and OpenAI has expanded the API usage limits for developers and increased the speed limits. It is said that the limits have been raised to the same level as GPT-4. In the fifth tier, the GPT-4 preview can handle 10,000 requests per minute, while GPT-4 Mini can handle 30,000 requests per minute. There is also a third tier for user ratings, where the lowest user tier has been expanded to allow 5,000 requests per minute for both GPT-4 preview and GPT-4 Mini. Therefore, if developers can utilize these output models effectively, the most essential aspect for them is indeed the output model. OpenAI has announced that the usage limits for the fifth model have been expanded to the same level as GPT-4.

Yesterday, in the announcement, various update features related to that were presented. However, I have some shocking news for you, and I will confirm that my predictions were correct. In the video, I mentioned that GPT-4 only would not be available in August, and the Strawberry Q project is being tested through GPT-4 only, which is being upgraded. Therefore, I have seen many instances where GPT-4 only demonstrates performance similar to or even exceeding that of 5. 

If we consider the one preview with a 5 reasoning model, from OpenAI's perspective, there are many paid users of GPT-4 only, so testing can only be conducted through that. Thus, there is no choice but to test all newly developed models in GPT-4 only. From a managerial standpoint, that is a natural course of action.

So, GPT-4 is not just GPT-4 from August; the responses have changed. When you input something here, it used to generate responses immediately, but now it takes a long time to think before responding. Although it doesn't directly indicate that it's processing, it seems to be stuck in a loop, and the time taken to respond has significantly increased. So, even though it's not explicitly shown, it seems to be in a thinking process. The chain reasoning technique is not perfect in GPT-4 either, because the resource consumption is about 65 times higher than that of GPT-4. This leads to excessive costs for OpenAI, so it's not the full version. However, I'm still testing some features of GPT-4. Today, I just happened to encounter it. Shall we take a look? I was just about to introduce the add function feature and the function schema for creating the add function feature and custom functions.

I showed you the part that automatically generates custom instructions for creating instructions for assistants. I tested this section to ask questions, confirm, and introduce it to you. I asked questions and exchanged information, but then it was general, and it didn't focus on my question. It forgot and missed the core of my question. When I designated it as a prompt expert agent in the assistant's guidelines, it was asking what functions would be suitable for that role. Understanding the essence of the questions will help you when you use it. Here, we define the instructions for the custom assistant, and in the functions, if you look at the content, it creates functions in JSON schema format to assist. If you don't know the code,

You can't write as a general user, so I was studying and asked questions about that part. When I get responses labeled as 1 and 2, it's like testing. Not all questions lead to responses like GPT's 1 and 2; sometimes you get to experience it. This process of testing caught my attention early this morning. As I expected about a month and a half ago, with the arrival of GPT-4, it seems to be changing. I have introduced many additional features to you, such as image editing and various functions related to images that were not possible before. I discovered that these functions are becoming available, and I even posted about it in a video. When I tested it this morning, I noticed that the feature of thinking in the original model, which provides responses labeled as 1 and 2, is now incorporated into GPT-4, allowing it to respond to all questions in this way.

I caught on today that instead of just providing answers, I am conducting tests with this thinking function intermittently. So, when this testing period ends, I anticipate that GPT-4 will not be able to infer everything in the one preview. However, depending on the content, it may provide more accurate answers by selectively activating its inference capabilities through a chain reasoning technique for complex questions. Currently, the frequency is about 50 times a week for the one preview, and since the paid subscribers receive around 50 times a day, I think that within this range, and possibly with a slight increase in paid subscribers, the inference function of GPT-5 could be offered, for example, several dozen times a day or around 100 times in a mini format. I am considering how the algorithm might be structured to provide this to you all. So, for everyone using GPT...

Pony's reasoning abilities have improved significantly compared to a month or a month and a half ago. As I mentioned earlier, I will talk about the Orion project, which is part of the Q project and the Strawberry project. When the Orion project is launched, this reasoning model will generally be integrated into GPT-4. Although it won't be activated for every question for premium subscribers, I am confident that it will automatically activate the reasoning model for complex questions to provide answers.

Today, as you can see, when the output model is activated, it takes a long time due to the chain reasoning technique. So, if you want better responses, we will gradually distribute the output model more widely to enhance the reasoning capabilities and advanced data analysis features.

I am confident that the search function, combined with the characteristics of LM and LRM, will evolve into GPT. This is good news for you, which is why I created this video early this morning to share it urgently. I will bring you more good news and useful information. Thank you.

GPT4o가 생각하기 시작 | 누구나 AI Agent를 만들 수 있다 | OpenAI 개발자 데이 |

GPT4o가 생각하기 시작 | 누구나 AI Agent를 만들 수 있다 | OpenAI 개발자 데이 |