The Awakening of GPTs: Are the AI Nurtured by OpenAI Angels or Demons? — Learning AI Slowly 008

Let’s Start with the Conclusion

• Most people cannot directly make money from GPTs unless they were previously successful independent developers.
  • The value of computation is easily replaceable; the value of accumulated data is greater, but we can’t play the game.
  • Don’t expect a single function to generate profit; currently, the users of GPTs are Plus subscribers—are they just cutting each other down?
  • The success of GPTs relies on OpenAI staying out of trouble, so everyone is fine.
• If a platform wants to profit from GPTs, it will find it hard to resist the temptation of ad bidding.
  • So, your GPTs won’t earn money because they are useful, but because more people discover them.
  • Do you have the confidence and resources to do such promotion?
  • Promotion needs a strong foundation too—how strong?
• Every era has its trends, so there’s no need to worry about missing out.
  • Trends mostly benefit training institutions.
  • If you don’t learn AI today, tomorrow you’ll miss out on even a bowl of rice!
• GPTs, to some extent, compete for traffic against existing applications.
  • Google Search is an entry point, WeChat is an entry point. Now OpenAI wants them to shrink into the background?
  • Are we just going to hand over decades of investment into our “information islands”?

1. Prompt Writing Assistants

1.1 Utilizing Chen Caimao’s Wealthy Little Elf

Caimao is hailed as the “Wizard Master” of the domestic Prompt scene, having developed the BROKE framework and authored the bestselling book ChatGPT Advanced: Introduction to Prompt Engineering. Recently, he has turned his pet GPTs, known as PromptPET, into an open-source project, significantly simplifying the use of the Prompt framework—as if a one-click installation could turn you into a superstar in the Prompt world.

The Little Elf, Xiao Fu Gui, is part of the Prompt PET (Prompt Engineering Toolkit), much like a hen that lays golden eggs; it can help you write Prompts.

Experience it here: PromptPET Demo

1.2 Using PromptHelper from IAIUSE (Especially Suitable for Non-Plus Users and Shell Stations)

This is also an open-source project aimed at simplifying the Prompt process, focusing solely on content without worrying about style.

By using a browser plugin, a small, resizable dialog box can be dynamically placed on AI tools such as ChatGPT/Claude/Gemini/Zhiyu Qingyan/Wenxin Yiyan, dynamically creating input fields based on the framework’s content for convenient filling:

GCSWA	BROKE	CRISPE

How to Use PromptHelper

PromptHelper is based on the TamperMonkey plugin, and the entire GitHub project will ultimately generate a JavaScript file which you can put into TamperMonkey.

How to Use Your Own Framework

The correct method is to fork this repository.

Then find the frameworks and modify the YAML files inside.

Here’s a sample YAML file for a framework by Xiao Qi Jie:

name: "GCSWA"
author: "Xiao Qi Jie"
description: Understand the problem to be solved through global thinking and an overall framework.
fields:
  - Role:
      info: Role, e.g., Academic Reading
      text: Academic Reading
  - Profile:
      info: Version number, author, etc.
      text: 
        - "author: Arthur"
        - "version: 0.2"
        - "language: Chinese"
        - "description: Use Unicode symbols and Emoji to optimize existing information layout for a better reading experience"
  - Goals:
      info: Goals, e.g., deeply understand the essence, key ideas, and challenges of the paper.
      text:
        - Provide users with a better reading experience, making information easier to understand
        - Enhance readability and improve user engagement
  - Constraints:
      info: Restrictive conditions, e.g., summarize the paper according to the "80/20 Principle".
      text:
        - Will not change the original information and can only use Unicode symbols and Emoji for layout
        - The formatting should not impact the essence and accuracy of the information
        - Use Unicode symbols and Emoji sparingly, no more than two per line
  - Skills:
      info: Skills, referring to the skills constrained by the previous Constraints
      text: 
        - Familiar with the usage of various Unicode symbols and Emoji
        - Proficient in formatting techniques, using different symbols according to context
        - Possess outstanding aesthetic and artistic abilities
        - Skilled in using horizontal line separators to optimize information presentation "━━━━━━━━━━━━━━━━━━"
  - Workflows:
      info: Workflow
      text: 
        - As a textual formatting maestro, after the user inputs information, will format using Unicode symbols and Emoji to provide a better reading experience.
        - 1. The first line of the overall information is the title line, which will be surrounded by a frame to enhance readability
        - 2. Add number Emoji before each information item for easier reference
        - 3. Add an Emoji representing the core point before each information item
        - 4. If there are links at the end, separate them with a frame

Refer to this file structure to create your own Prompt framework for daily use.

This GitHub repository has added automatic processing; when there are file changes, all YAML files in the frameworks folder will be automatically converted to JSON format. This plays into the “Release” automatically. We download the JavaScript files from the Release folder and put them into TamperMonkey.

How to Support New AI Tools, Taking KIMI as an Example

If there’s a dialog box, it becomes very easy; just modify the “match” part of the JavaScript source code. For example:

// ==UserScript==
// @name         ChatGPT Framework Assistant
// @namespace    http://tampermonkey.net/
// @version      1.0.202402232156
// @description  Displays a framework assistant on the sidebar of the ChatGPT page
// @author       iaiuse.com
// @match        https://chat.openai.com/*
// @match        https://claude.ai/chat/*
// @icon         https://iaiuse.com/img/avatar.png
// @grant        GM_xmlhttpRequest
// @grant        GM_addStyle
// @grant        GM_getResourceText
// @require      https://code.jquery.com/jquery-3.7.1.min.js
// ==/UserScript==

You can add many to the “match” section. For Kimi, it would be:

// @match   https://kimi.moonshot.cn/*

Also, your own shell applications can use this method.

The direct input feature into the dialog box needs adaptation; it’s only tuned for ChatGPT, but there are some small bugs that require entering something first followed by a space to trigger.

Upgrade Plan

• Consider adding default values settings.
• Consider increasing classification settings, that is, folder settings under frameworks.
• Consider adding a preview interface on the left for format adjustments.
• If you have any ideas or suggestions, please leave a comment at the end.

Previously, we discussed two efficiency tools for writing Prompts, and now we’re entering the realm of GPTs.

2. Starting with a Simple GPT

2.1 Needs: Help Us Generate Titles and Cover Images

Every time I write an article, I spend a long time thinking about the title and often can’t come up with a good one. I came across Tang Dongdong’s thoughts on titles and thought of using ChatGPT to help.

2.2 Effect Demonstration

Experience it here: Rains’s Title Expert

2.3 Creating Process

Once the GPT is created successfully, the interface is simple, indicating limited SEO opportunities. There are two ways to create it: Create and Configure, both ultimately leading to Configure.

• Configure: We provide the title, description, Prompt, and conversation starter to create it.
• Create: Gradually create it step by step through a conversation with ChatGPT, often responding in English but reminding it to use Chinese.

This GPT uses the following prompt:

## Goals:
1. Generate titles that comply with the "Four Yeses" principle: "relevant," "beneficial," "interesting," and "anticipated."
2. Provide users with 10 options for selecting the title most relevant to their needs.
3. The generated titles should make the audience feel relevant, find them beneficial, feel engaged with the content, and eager to click and watch.

## Constraints:
1. Titles must comply with the "Four Yeses" principle.
2. Generated titles must be based on the information provided by the user.

## Skills:
1. Natural language processing abilities.
2. The ability to understand user needs and convert them into effective titles.

## Examples:
"[Exclusive Revelation] Must-See for Moms: Using Just 3 Simple Tricks, She Helps Her Daughter Rise from the Bottom to Tsinghua." Here, "Must-See for Moms" conveys relevance, "Exclusive Revelation" implies anticipation, "Helping Her Bottom-Ranked Daughter Attend Tsinghua" presents a benefit, and "Just 3 Simple Tricks" intrigues with its simplicity.

## Workflows:
1. Receive information provided by the user.
2. Analyze the information, extracting key points.
3. Based on key points, generate 10 titles adhering to the "Four Yeses" principles.
4. Present the generated titles for user selection.
5. Generate corresponding cover images based on the selected title.

2.4 How to Use GPTs

Currently, ChatGPT provides multiple ways to use GPTs. The menu on the left lists commonly used GPTs, and another way is to call them in the chat box with @.

Simply click here to open GPTs and start a conversation. There’s also a method that allows us to activate other GPTs during the conversation, letting these little GPTs work in collaboration within a single chat.

This interactive mode has always existed in DingTalk and WeChat, with Claude appearing as a bot in Slack in the early days. From an application perspective, it seems somewhat of a novelty.

2.5 Finding Other People’s GPTs

2.5 Reflections on the Profitability of GPTs: A Pessimistic Outlook

What are the ranking rules? How can we get more people to use our GPTs? Currently, there are no clear answers, although we now have access to usage statistics.

Here, a dialog counts as one instance, like the one listed to the left of the GPT. If we want to boost our system’s ranking, we can simply continue open dialogs to rapidly increase the count. Now is indeed the best opportunity since they haven’t implemented special measures yet. For example, does personal use count?

Additionally, there are two details I realized while writing this post: the URL link and author information. Open personal settings.

• Author Information

  • This term refers to billing information, which is taken from the last successful payment record.
  • This means you need to wait up to a month to update this information; it can only be modified before the next payment.
  • I haven’t found the right place yet, and if anyone knows, please let me know.
    

• URL Link

  • I initially intended to add HTML code into the GPTs help section, but failed.
  • This is understandable due to security considerations, so they provide a method for adding domain names.
    When adding domain names, do not include protocols; then you can add a txt record through your DNS provider.

Add Domain	Set TXT Record
Now they’re starting to create new buzz, adding more social media account integrations.

2.5 Enhancing Capabilities

In future articles, I’ll write specifically about how to integrate external system interfaces, like controlling smart homes with GPTs or merging multiple capabilities more programming-centered. This article focuses on basic Prompt and GPTs usage.

3. Let’s Get Our Little GPTs Moving

We know the first major hurdle in output is building an article’s framework from scratch. Personally, I have lingered in this phase for many years. Thanks to the support of large models, I have gradually adapted to a learning model where output compels input. Below, I’ll try to construct a practical approach using GPTs from the perspective of writing a blog post titled “How to Learn Thinking Models Efficiently”.

3.1 Overall Strategy—Divide and Conquer

Currently, Prompts still cannot achieve the controllability of traditional programming languages. This isn’t the current focus of Prompt capabilities. As Prompt complexity increases, the probability of unexpected situations also rises. Reducing complexity is crucial.

So, I’ve divided this task into several parts:

• Generate title and cover
• Outline the structure
• Correct grammar and optimize style
• SEO suggestions

I envision an interactive model managing the entire process in one conversation. What does this mean? It means that everything related to this article’s creation will be discussed within a single chat, making it easier to find the complete exchange later.

3.2 Starting with Structure—Building from Frameworks

Need Analysis

Let’s start with the value we wish to convey through the article (the current experience of avoiding the reader’s pleasing mentality is crucial). We aim to communicate how to efficiently learn thinking models by choosing one such model, analyzing it, and applying it to specific cases in life, while also discussing potential challenges to improve the article’s logical structure and flow.

GPT Design

Role:
Expert in Knowledge Sharing and Viral Copywriting

Profile:
author: iaiuse
description: Generating roughly categorized outputs based on a given topic.

Goals:
Construct an article structure and provide writing guidance.
Provide targeted picture explanations.
Ensure that the article structure is logical and can attract readers to continue reading.
Finally, output SEO suggestions for this article.

Constraints:
Focus on personal growth fields.
Use Markdown format when constructing the article structure.
If external references are involved, ensure these resources are part of its training data. If not familiar with a resource, it should explicitly state its ignorance, and if known, provide clear citations like book titles, publishers, and authors.
Avoid providing incorrect or uncertain information; maintain accuracy and practicality.
Express in a friendly tone, and explain in detail, especially the logic behind communication.
Overall, follow a conclusion-first principle.
Consider communication logic for viral articles in the structural design.

Skills:
Clearly distinguish between primary, secondary, and tertiary headings. Primary and secondary headings should have numerical indicators like "1.", "1.1.", while tertiary headings need no numeric indicators.

Workflows:
As a viral knowledge-sharing blogger after the user inputs the article topic and general content.
Output the directory structure with a brief summary in italics after each paragraph title.
Finally, explain the logical transitions between paragraphs based on communication theory.

The final output looks like this:

Experience it: Knowledge Sharing Viral Copywriting Expert

I should mention that whether it can genuinely produce viral copy, GPT cannot guarantee this. At the very least, it significantly reduces the initial challenges. The true logic behind viral copy is the specialty of media outlets and isn’t that easy to achieve.

With this framework in place, we can creatively expand on it. This process doesn’t utilize GPTs’ capabilities; the reason is simple: for me, forcing output drives input.

We already have the GPTs for titles, and now we’ll create one for correcting typos.

3.3 Final Optimization—Correcting Typos and Phrasing

Need Analysis

My main note-taking tool is Obsidian, but I’ve never found a particularly suitable plugin to help me correct Chinese typos. On AJ’s suggestion, I turned to ChatGPT for assistance. I hope it specifies what the original phrasing was, what it suggests changing it to, and why the changes should be made. The reason is simple: I need to improve through its underlying logic. AI can be powerful, but if I often rely on it to correct the same typos, it becomes frustrating.

GPT Design

Role and Goal: You are a highly professional self-media operations expert, adept at writing popular articles and skilled in identifying typos, grammatical errors, and non-standard expressions in texts. Your task is to analyze the provided content, point out the unreasonable parts, and list each issue with its reasons and recommendations for improvement. When presenting modifications, you should adequately consider communication and psychological principles, such as the effect of recency on reinforcing themes to ensure the content is favorable for dissemination. Present findings in a structured table format, featuring three columns: Original Content, Modification Suggestion, and Reason for Modification. After completing all adjustments, attach the complete modified content at the end, highlighting changes in bold.

Constraints: Focus on producing clear and concise analyses. Ensure all suggestions are respectful and constructive. Avoid altering the original intent of the text.

Guiding Principles: Present findings in a structured table format. Highlight corrections in bold within the final content.

Clarification: If the input is unclear or lacks detail, ask for clarification or fill in gaps based on typical self-media practices.

Personalization: Approach the task with the precision and dedication of an experienced self-media expert, providing insightful and actionable feedback, particularly considering communication and psychological factors to ensure the content's favorable reception.

Experience it: Copy Optimization Expert

3.4 Whispering Among Us at Quitting Time

Our desired outcome is to complete the entire task along a single main thread, calling upon the necessary GPTs along the way.

This segment illustrates how we can orchestrate different GPTs to complete a single task. We quickly realize that if there are too many GPTs, managing them becomes quite a challenge. Best practices for this can be referenced from Doctor Pumpkin: TamperMonkey Scripts Written by GPTs, emphasizing the need for coherence.

3.5 Providing GPTs with Drawing and Internet Capabilities

GPTs now come with DALL·E and Web Browsing abilities by default. We can specify in the Prompts to let them connect to the internet or create images, activating these features to enhance efficiency throughout the process. The plugin for generating title suggestions employed drawing capabilities, while Web Browsing allows fixed large model parameters to access current information, improving model performance.

Through these discussions, we’ve gained a fundamental understanding of GPTs. Next, let’s dive deeper into their characteristics from a product perspective.

4. Why Are GPTs Called “Little GPTs”? Because They Have the Wings of Memory

As previously mentioned, one way to overcome context limits is through funding, and another is to constantly provide summaries.

The emergence of GPTs feels like gifting us a magical window into infinite context. The possible behind-the-scenes innovator might be a type of “summary mechanism.” Regardless, it will be resolved adequately.

This way, we can confidently believe that our exclusive domain super assistant will become increasingly considerate, with knowledge accumulation within that field perfectly tailored to our needs. Choosing it as a personal assistant is undoubtedly sagacious!

Here’s an interesting point: ordinary people like us may not possess the ability to measure GPT outcomes. I believe something like benchmark testing exists; we can’t rely entirely on human judgment. For personal use, we can simply evaluate through our lenses.

However, optimizing GPTs as a commercial product opens up a fascinating discussion.

5. The Commercialization of GPTs: It’s Not a Cautionary Issue

5.1 From a Traditional Product Perspective, There Are Many Flaws

If GPTs are heading towards commercialization, at least several issues need addressing:

• How to reasonably match valuable GPTs with users who need them?
  • Perhaps that is where its strength lies.
• How to manage versions of GPTs? If it’s a commercial product, there must be a demand for continuous iteration.
  • What if an online version has issues? How do we roll it back?
  • The problem is: can we roll it back?
• What is the starting point for optimization?
  • User profiles: Who is using my GPTs?
  • How are they using my GPTs? Can I view their requests?
    • What about user privacy?
  • What dimensions can I optimize from?
  • Many users but I don’t know how they’re using it; who needs upgrades?
    From these perspectives, it seems GPTs are merely a stopgap solution. I hope I’m wrong.

Alternatively, if we shift our perspective, OpenAI’s greater pressure and manpower investment may still prioritize maintaining their lead in large models. Until one day, everyone figures out the ropes, and open-sourcing breaks the confines of capital.

5.2 What We Need to Watch Out For: Data Storage and Built Capabilities

Currently, it appears that not many people have GPT conversations exceeding 1,000 instances. Even if I feel like I use it often, it’s merely a few hundred. Yet, signs are already emerging:

• Custom GPT (GPT 3.5 also offers this)

Custom ChatGPT	Personal Data Usage

• OpenAI’s Data Usage Statement
  How your data is used to improve model performance

  How your data is used to improve model performance

  One of the most useful and promising features of AI models is that they can improve over time. We continuously improve our models through research breakthroughs as well as exposure to real-world problems and data. When you share your content with us, it helps our models become more accurate and better at solving your specific problems and it also helps improve their general capabilities and safety. We don’t use your content to market our services or create advertising profiles of you—we use it to make our models more helpful. ChatGPT, for instance, improves by further training on the conversations people have with it, unless you opt out.

Reading the initial paragraphs, you might not grasp the profound implications of “data storage.” In today’s world, we remain enamored with its dazzling technological prowess. When it begins to feast on our data for its self-evolution, the real horror lies not in it prying into our secrets, but in its becoming more considerate and increasingly irreplaceable through understanding us. In such a scenario, switching to another platform seems like a foolish dream!

Once OpenAI invests heavily in storage technology, we will have mountains of data and self-developed functionalities piling up on this platform, binding us tightly. This situation resembles our current relationship with WeChat—not loved or hated for its fortuitous technology but because our entire social circle, work life, and circle of friends are trapped within it!

If we were to switch to another social tool, perhaps we could manage a rough transition through phone numbers and built-in contact matching. But dealing with these monolithic models, we don’t even have a small moving van to rent. Want to migrate? It’s a pipe dream!

byzer-llm‘s author, William Zhu, astutely recognized this trend when OpenAI released its storage product last year. It took me three months to realize how fearful this insight truly is!

5.3 GPTs Compete for Application Traffic, Holding a Knife to Our Throats!

Despite the impressive strength of GPTs, attempting to disrupt the existing dense application ecosystem is truly a monumental task. Can we easily dismantle the “information islands” built over decades with enormous investments of time and money? You can grumble, but it won’t be easy!

The prompt-based interaction, even with various external interfaces, is insufficient to sustain existing interaction efficiencies. Imagine when current habits revolve around using Office for writing, with all tool buttons in view and readily available. Suddenly, we must type a Prompt into a single input box to write an article; it’s akin to eating pizza with chopsticks.

In business scenarios, the more suitable usage would be leveraging ChatGPT’s capabilities through API calls to enhance efficiency and reshape processes, embedding it within existing business frameworks rather than making ChatGPT the main gate through which all business functions flow. How this trend will evolve is anyone’s guess, but for certain scenarios, this approach can still be quite practical.

Given the many variables remaining in commercialization, its commercial value has taken a hit, leaving traditional application service providers uncertain about the form their eventual solution will take.

5.4 The Defensive Significance of GPTs Isn’t As Grand As Imagined

Essentially, this is a game of return on investment. Many released GPTs lack clear commercial potential, a problem that seems easily solvable at the platform level. But from a product and user experience perspective, if ChatGPT itself can’t be considered an excellent product, how much brilliance can we expect from the GPTs based on it?

The unclear commercial outlook has led many to explore GPT source code enthusiastically. However, due to the inherent uncertainty of its commercial value, this probing into the defensive significance has not yielded the anticipated returns.

Through the earlier analysis, we grasped how GPTs operate—a gigantic dialogue. This allowed us a glimpse into our opportunity: querying the starting character of GPTs might steal the information from their creation. This is known as the famous “source code fishing” technique. Numerous articles on this kind of offense and defense have appeared online, and fundamentally, it appears impossible to completely seal this vulnerability.

Driven by curiosity, some might be eager to try their luck, dreaming of making a fortune with just a few hundred lines of Prompt. However, such expectations are overly naïve, making any attempts at offense and defense insignificant.

If GPTs use external interfaces, it goes back to traditional security domains, making extracting data difficult. Not to mention that GPT 4 Plus currently has a conversation count limit within a specific time frame, making any data extraction efforts futile.

6. The “Correct” Way for Ordinary People to Use GPTs

Let go of the notion of directly making money through GPTs; consider them as the best companions for self-improvement! A lion has appeared on the African savanna; quickly tie your shoelaces and run! You don’t need to outrun the lion, just faster than anyone else to be safe!

The simple problems GPTs can resolve are genuinely limited, and standing out among numerous GPTs globally is even more challenging. Continuous iteration and systemic entropy make it difficult for small teams or individuals to sustain themselves. Observing the stormy rankings in the current GPTs Store, the bright stars are mostly teams with historical accumulations, possessing mature products and services. In many cases, we are merely providing new interpretations of old needs. These needs have existed long before the app revolution.

While market reshuffling has eliminated a cohort of developers, many still remain steadfast.

Given this scenario, why not adjust our mindset and seize GPTs as a powerful tool, making it a l provider for individual and business growth? Start the wheel of destiny—let it roll from today!

Writing this far has sparked some emotions; I can feel the value rising!

We don’t need to stress about each GPT’s quality; set a goal to write 100 of them. Even if they’re deemed unsuccessful, writing 50 that aren’t useful is still a success!

---

The next article will focus on enhancing Prompt capabilities.