I feel that in the world where GPT and Claude are discussed, children are somewhat left out. However, it's clear that Google has the Chrome platform and many other platforms. There seem to be many attempts to integrate these. Now, you can call Gemini from a Chrome window. If you type in Gemini, the chat with Gemini appears. When you click it, the search bar changes to a chat window. Yes, yes, I will ask about the end plan here. When you press enter, it will take you to Gemini. It shows these four channels. If you click, you can navigate to Gemini. Originally, if you wanted to access Gemini, you would have had to click on it once and go through some steps, but now you can do it directly from here. This seems to be the extent of the update, but this is just the beginning. Because now, within Google, you can search all page contents using Lens.

What is Lens? This is a feature that is likely already available on smartphones. For example, if you want to clearly understand the real estate contract process, you can press the three-dot menu and select Google Lens for searching. Then, you can define the area you want to search. By holding down the left mouse button, you can drag it to the desired area, and when you click, you can see the results in this format. I'm trying to search for Andrew. This brings up information about Andrew. It seems very useful, doesn't it? It feels convenient for searching things right away. In the past, while Google did a good job with image searches, you had to upload images by clicking and selecting. But now, if you have Chrome installed, you can easily search directly from this screen.

It seems possible, right? Currently, Google Lens and Jaminai are separate, but later, won't it all be possible within Chrome? Ah, that's true. You can search with Google Lens and then link it with Jaminai to ask questions directly. I think they will quietly add features one by one. I'm also curious about how it will be applied to YouTube. If YouTube incorporates such lenses, it would really be great. I also feel that something virtual is coming. Virtual humans, virtual influencers, virtual YouTubers—I'm particularly interested in virtual singers. I enjoy creating music or audio with Suho, and I've been working on it a lot. However, I often hear people around me saying they don't know how to use Suho. I believe Suho has classified many things.

I enjoy combining different genres to create new music, but some people are unsure about which genre or prompt to use, or what lyrics to write. I found a website that could help those individuals. It's called prompt.com. Here, I can select the desired song style. For example, let's try a sad ballad. A sad ballad sounds good. Can we also determine the vocal range? That will be generated based on the song style. It will be used when rating. You can think of it as the style of music that will create this for you. So, where do we specify the gender of the vocals? You would enter that in the style of music as female voice. Oh, so we can just add it later? Yes, you can add it. Alright, so we're going with a sad ballad and the theme is loss.

The melody is simple, easy to remember, and the rhythm is slow. When creating a song style, you get three options. First, let's try this song style. The problem is the lyrics, right? Oh, it also generates lyrics. You can go into the lyrics generator and create lyrics in the voice of a man or a woman who regrets a breakup. Yes, the theme is my sorrow and heartache. You can also choose the language. Is Korean available? Yes, there are Korean options for creating sad lyrics. I'll just copy this directly. If I don't write the lyrics myself, I should have kept switching back and forth with GPT, but now I can do it all in one go. Oh, it also lets you write the song title here. The male voice has come out now. Wow, this is nice. The song is good! Yes, I think I like this style a bit. Oh, I really like this style.

I see that you like it. Until now, I haven't created anything in this style, so it's great to know that you enjoy this style. I will try to create more like this. Yes, I'm looking forward to it. Writing lyrics is not an easy task, so if you're struggling with how to write the lyrics or what style to create, feel free to brainstorm ideas here. Alternatively, I have a video on our Life Channel that explains how to utilize Sun-ho, and I think that could be helpful as well. I'll set it up so you can easily access it by clicking on the right. It seems that the GBT Pro is running, and I'm curious about that prompt; it must have been created well. I think we could also create our own GPTs. Yes, I believe we can definitely create a GPT with the current quality. I think it would be sufficient to make one in about 4 to 5 minutes.

These services can actually be quite expensive, right? Yes, in the past, tokens were very expensive. Oh, that's right, they were incredibly expensive. One of the fathers of our artificial intelligence, Professor Andrew Ng, mentioned something about why so many of these services are emerging. What did he say? He mentioned that the cost of GPT-4 is now $4 per million tokens. When GPT-4 first came out, it was $36 per million tokens. That's a ninefold decrease, isn't it? So, in just one year and five months, the price has dropped by almost 80%. Professor Andrew Ng suggests that since the cost of token generation is decreasing and hardware is advancing, it would be wise to assume that token prices will become cheaper when developing applications. He recommends thinking of the current token price as always being at a peak.

거네요. 네, 정확히 워딩은 애플리케이션이 현재 실행하기에는 비싸더라도 더 낮은 가격을 예상해서 배포할 가치가 있을 수 있다는 말씀이셨습니다. 제가 알기로는 많은 솔로 프리너, 즉 1인 사업하시는 분들 중에서 GPT-4나 이런 언어 모델을 활용해서 많은 서비스들을 개발하고 배포하고 있는 것으로 알고 있습니다. 어떻게 보면 이곳에서 많은 기회가 있는 것 같아요. 확실히 오픈AI뿐만 아니라 구글, 메타, 클라우드 등 다양한 곳에서 언어 모델을 내놓고 있으니까, 이제 선의의 경쟁 덕분에 소비자가 이득을 보는 약간 그런 구조가 아닌가 생각이 드네요. 앤드리 교수님이 생각하신 가장 큰 이유는 라마 3.1이 나왔기 때문입니다. 아, 오픈 소스가 나와서 그렇네요. 모델 자체가 오픈소스 모델이 나왔다 보니까 GPT-4 같은 모델을 직접 개발할 필요 없이 서빙만 하면 되잖아요.

It seems that they are focusing their manpower on hardware optimization instead of model development. For example, the Grok model can generate tokens very quickly. They can simply upload an open weight model, like Llama, and serve it, which allows them to offer services at a very low cost. Recently, I heard that a company called Cerebras has made it possible to serve their beta 3.1 model at an impressive speed—1800 tokens per second, which is 20 times faster than the GPU, all at the same price. Should we give it a try? Oh, what? It's 447 tokens per second at 70, which is incredibly fast. The speed is really amazing. As the speed increases and prices drop, it seems that there are fewer and fewer reasons for many companies not to use LLMs. If we can get the setup right, we could have some excellent services.

It seems that a lot will come out now. Professor Andrew gave this analogy: if we continue to utilize tokens per second, it ultimately means spending only $1.44 per hour. This, right? In that case, he argues that it is much better than the minimum wage in the United States and several other countries. This indicates that it can be used quite cheaply. So, as you mentioned earlier, it seems like this might be the peak, but it wouldn't matter if we continue to develop it. Lately, when I look at the feeds of developers, it seems like they are handling many of the initial tedious tasks through LML and GPT Claude. Is there any interesting news related to that? Yes, it's quite similar to what you just mentioned. Instead of just a cursor library file, there is a feature that allows you to understand the entire project code comprehensively, which definitely makes it simpler.

In addition to implementing the code, as you can see in the example video, even with over 1800 lines, the code is generated well. In this way, within the entire code and project, the consistency of variable names and function names is well maintained. This significantly reduces reference errors that occur when different files are interconnected. I felt that this could be a great advantage not only in small projects but also in projects of a certain scale where the assistant can be effectively utilized. I can also choose my model in the course; I have options like GPT-4, GPT-4.5, and Claude 3.5 among the premium models. So, I can select and use one of them. When using a lower-priced plan, I might have to use a non-premium model. Do you use Copilot?

When I had the opportunity, I used the support provided, but I was a bit surprised that I had to create the model architecture myself for training. At that time, they roughly provided a model architecture, and I found that quite fascinating. I created an autoencoder back then, and they specified a few layers and activation functions that I couldn't use, so they just generated it for me. Of course, there were many parts that needed modification, like connecting the data loader and the batch training part, but that was almost a year and a half ago, so I think a lot of that might have been automated by now. I also think that using a cursor could allow for deploying the trained model as well. When I saw the demo, I wasn't able to fully grasp it since I'm not a developer, but still, I can see that a lot of development can be done using AI.

It means that I have been able to do it. Actually, I hadn't tried frontend development before, but after seeing this course, I thought, Should I try making a website too? I will try making one if I have time after paying the $20. I was planning to use Figma and connect it to WordPress, but you could approach it that way too. When I think about adding functional parts, combining the two seems like a very good method. However, if I'm going to do it, it might be better to write the code from scratch and create everything with CSS. Rather, I expect that Figma will provide CSS and HTML that can be used, and I think they will provide code based on that. That's about the extent of my expectations, but I have hope that it will be done to that level.

Of course, we can only know by trying, but what is possible for someone like me, who has no knowledge of development? I think it's quite possible to develop a simple program. However, since I haven't used this yet, I'm not fully aware of its limitations. Generally, the limitations seem to be that while writing code is done well, debugging is somewhat weak. This is mainly because the underlying language models are GPT-4.5 or Claude 3.5. I am currently using GPT-4.5 for coding, but there are still some bugs that it doesn't quite understand. Therefore, I believe that developers still need to cover the limitations of these language models. If one doesn't have the ability to identify these issues, it might be a bit difficult to create large-scale or feature-specific code.

I suddenly thought of the smart money post I shared yesterday. It seems that they have reached an agreement with the American safety research institute regarding the next model. It appears they are already working on the next version of GPT. Have they completed the safety tests, or are they planning to conduct them? It sounds like they are in the process of reaching that stage. If that's the case, will it be released soon? I wonder if they are entering the testing phase now. I'm looking forward to it; I hope they complete the tests quickly. I'm curious about what kind of model it will be. It is expected to be released this year, right? Although they keep saying they don't know exactly when, many people predict it will be in the second half of this year. If that happens, maybe we could see GT5 integrated into services like ChatGPT, as you mentioned. That seems quite possible, especially since a model with better performance is coming out.

Personally, I think it's about time for it to be released. It really needs to come out soon, and the software also needs to be released quickly. I'm curious about how much it will be priced and whether it will be superior to existing services. What kind of entertaining videos are available? How about watching the video made by Acho on how to create the song Sunho? For viewers who are curious about how to create AI music, I hope this video helps you become more familiar with using Sunho. This is the end of the plan for Macho and also for Eun-ki. See you in that video.

인공지능이 더 가까워져요 | Gemini, 구글렌즈, Suno 프롬프트 생성기, 앤드류 응, Cursor

인공지능 관련 소식들을 만나보세요.

1. 인공지능 전공자가 전해주는 [인공지능 주간 트렌드]
2. 유명인사들이 말하는 인공지능 [AI Talks]

Contact : endplan0704@gmail.com