GPTs Are Here! How to Build Your Own GPTs & the Hidden Startup Opportunities Behind Them 68

0:00Hello everyone. Just three hours ago, OpenAI released the official version of its GPTs, allowing all GPT Plus users to create their own

0:12apps in the GPT Store. We tried it out right away, and today we're bringing you a review of how GPTs actually perform. Welcome to this new episode of our video

0:23First, as usual, let's manually create a GPT for you and show you what features are available. First, click on Create

0:34Once you're on the GPTs interface, you'll see a lot of options. Click on Configure. The first thing is to choose an avatar

0:42You can choose to upload an avatar or use DALL·E to generate one. Then give it a name and a description. The example I'm using here is based on some videos we've made before

0:52which is about what sources of AI information exist. So here I'm saying let's make a "GitHub Summary" — I'll look at what's new and trending on GitHub

1:02and give me a summary. Then I'll modify it here. The instruction for GitHub is

1:08what the GPT should do, how it should behave, and what it should avoid. This is essentially a prompt engineering process

1:16You can add many complex prompts. Here I'm keeping it really simple — basically telling it you can do these things, then read the GitHub trending repo

1:24and pull information from there. I'll add another sentence here. OK, here I've added two sentences to guide its behavior. The first is: if you see a URL, visit it first, then

1:42search within the information you get. The second is: try to visit these links as much as possible. OK, the third part is about what kind of conversation starters

1:49you want to prompt users to search with. For example, here I say "What's trending on GitHub right now" — telling users

1:55they can ask this kind of question. This is something many AI tools need to do, because it effectively guides users on how to do things

2:04OK. The Knowledge section here is what we mentioned before — if you upload files here

2:11OpenAI will automatically convert these files into embeddings and do a similarity search, then within that similarity search

2:18find the relevant information. But still, we need to remember that search capability and large language model imitation capability are not the same thing. They often get similar results

2:28but the underlying functionality is different. Often, the reason we need a database is because you want

2:36the responses to find similar things from the materials you provide, and then through the capability of the large language model

2:44generate an answer that looks reliable. I'm not going to provide a template here, but if you want to make a personal introduction, or perhaps turn some TV shows, movies, or tutorial text content

2:58into a knowledge base, that's the most effective use. That's the most effective. OK. The Capabilities section — web browsing must be turned on. Whether to turn on DALL·E is optional, but Code Interpreter is also best turned on. As mentioned, Code Interpreter can be seen as an AI agent — what it actually does is ReAct, meaning Reasoning and Acting

3:23It first reasons about what to do, then does it. So Code Interpreter essentially represents an agent logic. Of course, it also has the ability to process and run code — these two things are blended together. For more details, you can check out our previous videos

3:32there's one episode specifically about Code Interpreter. OK. The last one is Actions

3:38You can think of Actions as the tools in an agent. If you're familiar with LangChain

3:45it's essentially the same concept as a tool. The concept of a tool. Here, when I add an action, I'm adding an executable tool. If we click "Add Action"

3:55you actually need to import a schema. What is this schema? You can think of it as a specification. Let's import a weather example

4:01and you'll see. The specification here describes: OK, what's the name of this tool, what does it do

4:12it's all invoked through APIs. What does the API look like? What does each path under the API do? How should it be called

4:22what information needs to be provided? At the same time, there's the authorization — for example, if the API uses an API key, you need to provide it. For example

4:26if the API uses an API key, you need to provide the API key. The authorization method this API uses

4:32then you need to provide the corresponding method. So the core point is that by adding tools, you enhance what your GPT can do

4:44If you look at this description — naming it, right? — it's exactly the same as adding a tool to customize your tool in LangChain

4:50no difference at all. So again, why do we say you should learn the LangChain approach, not necessarily use LangChain

4:56Once you learn LangChain, you'll understand why they built tools this way. OK. Then

5:02we can add this. After going back, we see that the weather action has been added

5:10So this is a good example of adding a tool. After all these are configured, you can click "Update" here. You can publish it

5:17for everyone to see, or only certain people, or make it public. OK, we'll click "Confirm"

5:24OK, now let's try it out. You can see it's doing a web search and browsing web pages

5:37But here's something I want to remind everyone — this browsing might not always produce particularly accurate results

5:45There's still some hallucination here. For example, it says here "I cannot directly view the content of the GitHub trending page" — meaning it can't directly access this kind of

6:00page about what's hot on GitHub right now. You can visit it yourself, it's simple. If I can visit it myself, why would I need you? This is actually one of the limitations of these GPT tools

6:13and it's something else we wanted to mention today. We're currently in a period when GPTs have just been released and GPT's new features just dropped — everyone is in a very, very excited mood

6:25I predict that within the next week or two, tons of apps and applications will emerge, and people will also be doing some open source work. The point of all this is actually to let everyone use their creativity to do these things. After this phase passes

6:35people will come out and say this thing doesn't work, because once we use it we know where its ceiling is. There's always a ceiling, and once it hits that ceiling it can't go past it

6:42right? At that point we'll know what its limitations are. Just like the problem we're encountering now — it can't directly access this website

6:50or it works sometimes and sometimes doesn't. There will be a lot of complaints. Then the third phase is: oh, everyone finds out you can do it, I can do it — once all the technology matures, it becomes a rat race, right?

6:58Whatever you make with this tool, I can copy it in two or three minutes, swap in something new. And this new thing

7:04might have one or two projects that take off. But through all the changes, the underlying principle is that barriers will get lower and lower, because it's just wrapping. So barriers will get lower and lower

7:14We can already see this trend with GPTs. Since everyone can make these tools, among the hundreds of tools out there

7:27very few will actually be truly useful. Often, having the ability to use these tools just satisfies your own workflow — in my daily life, using AI fits my own work workflow

7:41 But if you want to find a unified app that can solve very big problems, I personally think it's very difficult. The core reason is that the barriers are too small. So as a developer, if you want to build your GPTs and create your competitive edge,

7:58 where is the core thing? The core thing actually lies in two places. The first is the knowledge — whether the knowledge you provide here is different from others. This is very important. The second is the action.

8:09 This ability to add charts. There are many things that this action side can extend into. For example, if you connect to the Twitter API, you can let GPT post tweets for you,

8:20 read Twitter, read Twitter trending topics, and then you can do some data analysis based on these things — read the latest fashion trends, read YouTube view counts.

8:33 As long as these things have data backing them up, if you build an API middleware, you can take external data and transform it into

8:42 data processed by a large language model, and then it might be useful. So if you want to build related applications, focus on this area:

8:50 function, function call — how to make good use of action, how to build middleware in between, and then string these apps together.

8:57 So the empowerment above has a huge, huge space for imagination, okay? You must make sure to play around with this.

9:05 Build your own API, alright. So today's main content is this — I demonstrated to you how to create GPTs, then let you try it out. Although the results were not entirely satisfactory, that's roughly the idea. And the third part was telling you about its limitations and its opportunities. So today's main content is this. If you have any questions or want to discuss something, feel free to leave a comment below the video.

9:25 If you think our video is pretty good, welcome to like, favorite, subscribe, comment, and share our channel. If you have any questions or things you want to discuss, welcome to leave a comment below the video. If you think our video is pretty good, welcome to like, favorite, subscribe, comment, and share our channel. This is very important to us. Thank you for watching.

9:31 Wishing you a smooth learning journey.