This article provides a practical guide for readers who are eager to enter the AI side hustle, analyzing AI applications from a global perspective, delving into AI capabilities and application cases at different levels, helping readers find the right positioning and start the road to an AI side hustle with a monthly income of more than 10,000 yuan.
In recent days, I have communicated a lot with fans, who are either managers, directors, and even fans have some of the company’s budget use rights in their hands.
Some of them paid some tuition fees from Douyin self-media, and some learned some lessons in project practice, and in the end the effect was not very good, and they wanted to consult me on how to get started.
This actually makes me a little embarrassed, because if you want to have enough understanding of AI projects, you still need to go through enough practice in the end, but how to make them pay less tuition seems to be possible.
So, I have compiled a material for two of the paid fans, here I will remove 30% of the most essence, and let everyone feel it.
AI applications from a global perspective
First of all, we need to clarify our thinking, we must have enough understanding of some professional terms of AI. The AI that is often talked about this year is actually a set of large models, please see the figure below:
The simplest understanding of AI is that it is a set of APIs, I give him an input, and he will give me an output as requested.
On top of that, learn about the most common AI applications, commonly known as knowledge bases:
That is to say, the model can accurately and clearly output answers to questions according to the company’s internal knowledge system, and the knowledge base further evolves into the Agent that everyone often hears, which is different from the knowledge base Agent needs to solve actual problems:
The architecture of Agent basically includes all AI knowledge, so in the end, you can learn around Agent:
But once we talk about Agent, we have to go back to two technical paths, and why there is a two-day path Here we have to go back to OpenAI’s Sam Altman’s definition of Agent.
Agent was originally defined
L1 level (chatbot). AI systems are capable of basic conversations and communication, demonstrating basic understanding of natural language and responding to various prompts and questions.
L2: Reasoners. AI systems can solve complex problems with the proficiency of human experts, marking an upgrade from mere mimicking of human behavior to demonstrating a level of authentic intelligence. These AIs are not only good at conversation, but also have problem-solving skills, and their reasoning and decision-making skills are close to human levels.
L3: Agents. AI systems can take on complex tasks, make decisions, adapt to changing environments, and act autonomously without constant human supervision. At this stage, AI not only has reasoning capabilities, but can also autonomously perform various complex operational tasks.
L4: Innovators. AI systems are creative and original, capable of coming up with groundbreaking ideas and solutions. They not only imitate human creativity, but also break through the limitations of thinking and come up with refreshing and innovative ideas.
L5: Organizations. AI systems are not only strategic but also have the efficiency and adaptability needed to achieve organizational goals, capable of managing complex systems. They can flexibly coordinate multiple agents, assign tasks reasonably, monitor progress in real time, and make rapid adjustments based on actual conditions.
On this basis, let’s take a look at the existing Agent framework:
Two sets of agent technology paths
Path 1 is actually very simple, that is, design your own SOP, and then use the program to implement it.
After 10 years of interaction design, why did I transfer to product manager?
After the real job transfer, I found that many jobs were still beyond my imagination. The work of a product manager is indeed more complicated. Theoretically, the work of a product manager includes all aspects of the product, from market research, user research, data analysis…
View details >
Path 2 is even simpler, which means that you don’t need to do anything, the model itself will help you formulate SOPs (or Workflows), and then call various tools (maybe MCP programs), so your functions are implemented.
How to say it, path one is a set of technical frameworks actually used by each company, and path two is to make itself a framework for the promotion of entry companies, especially companies with certain model capabilities or domain capabilities.
Regarding this piece, the following content will be updated highly, just click a little here. Next, we discuss how to measure the current AI capabilities of companies and individuals.
Seven levels of AI applications
First of all, the model cannot jump out of the three cores of algorithm, computing power, and data, and in the case of model-based AI applications, there are only two cores in the verification stage: engineering and data.
Here are a few core considerations:
- Engineering ability, that is, understanding of AI applications, the core is the programmer’s code ability, and the use of various automation tools;
- Model training capabilities, model training capabilities can be regarded as the upgrade of engineering capabilities, which involves the computing power behind them, as well as the organization of engineering capabilities and industry KnowHow;
- Industry KnowHow, that is, the understanding of a certain industry, such as the understanding of the work of doctors, the understanding of the work of lawyers, the low-level KnowHow is the collation of SOPs, and the high-level KnowHow is the accumulation of a large amount of high-quality data;
Here, through various permutations and combinations, the level of AI application can be obtained:
The first level, the small white level application
The so-called novice users are people who have nothing, no AI cognition, no engineering ability, no industry know-how…
The scenario where they use AI is probably to open the official website chat page of DeepSeek or ChatGPT and type in: Will AI replace humans?
Then in the process of chatting with AI, the object of constant lying down and lying down is a very amateur user, but it may be 90% of the users.
These 90% of users are commonly known as AI leeks, AI is very easy to cause their anxiety, and they will pay a lot of tuition fees for AI cognition, but I personally think that this tuition fee is cost-effective.
The second level is to cut AI leeks
Level 2 users will take a big step forward on the basis of level 1, and their important job is to sell AI anxiety and teach AI cognition.
In fact, they are no different from first-level users, they are still not scattered, “but they are familiar”.
They will clearly feel the “awesomeness” of AI in dealing with single-point problems, and use its high efficiency characteristics to assist themselves in completing some tasks, the most common are: AI explosive articles, AI Wensheng pictures to tell stories.
Moreover, after they are familiar with it, they will teach the 90% of the unknown people at the first level to use AI to write articles and other actions, in short, it is very low but can really make money.
The third level is personal assistant and efficiency expert
The third level of users begins to have certain engineering capabilities, they will organize the fixed parts of their work into SOPs, use API or RPA automation methods, to do some actions in batches, such as sending and receiving emails, resume screening, financial approval…
From this step onwards, AI has truly become part of solving work problems, and has begun to liberate manpower, which belongs to the level of personal assistants, but it is also limited to personal applications.
For example, a lawyer created a personal inquiry tool based on DeepSeek and made it into a plug-in and put it on the online platform, every time a user consulted, the AI customer service first communicated with the intention, judged that the user’s intention to spend money was high before guiding the real person, so that the efficiency was improved by more than 10 times.
Here is another example that may be helpful to your self-media: self-media AI efficiency improvement cases
Now the recommendation logic of WeChat public accounts has changed, although the number of fans is important, but it is not so important for reading, according to my practice during this time, there is a set of simple methodologies (SOPs) for reference: apply for 10-100 official account accounts; Write an article by yourself every day; Use AI to self-wash manuscripts to form 10-100 articles; Use AI to generate 10-100 popular titles; Start distributing articles using RPA;
Logically, there will definitely be one explosion every day, and an article will bring some income if the main traffic is open, and the income will be very good in the long run.
The fourth level is the SOP platform
From the fourth level onwards, the requirements for engineering ability begin to increase, which is often not something that individual players can do well.
They will begin to realize that the use of AI by individuals is to help them complete a certain type of work, and this kind of work must have quite a commonality, so it is easy for everyone to think that this is our SOP at the third level.
Therefore, in order to help more players establish personal assistants, the platform has built a set of platform-level SOP building platforms, and now the most famous should be Feishu Button’s Agent building platform.
On this basis, there will also be many people who teach 90% of users how to use this low-code platform, and its essence is to eat poor information, which is not clever but very effective.
In addition, the essence of the AI Agent platform is still using the model’s API interface as a judgment, and it is possible to do some simple functions, but it is difficult to make a complete application, not to mention the problem of data leakage.
Many times it seems to be a problem with the last kilometer, but this kilometer costs the previous 10 pieces…
The fifth level is industry tools
The first four levels have a characteristic: they only need you to understand AI and have certain engineering capabilities, but from the fifth level, they require industry knowledge and high-quality data under industry awareness.
For example, as a doctor, you complete your own personal assistant through the Agent platform to answer patients’ questions online, but after a period of time, you find that there are always some errors or vague answers (including diagnostic errors, drug recommendation errors), which may cause disastrous results.
And this is obviously not a problem that can be solved by individual ability, so you hope that this Agent platform is an application exclusive to the medical (legal, financial) field, and they will answer more accurately without hallucinations.
For example, common AI doctors and AI lawyers are products of this level, because they are serious AI engineering applications, and users may lose money if they answer incorrectly, so they must solve the problem of large model hallucinations at the beginning.
If you want to solve the problem of model hallucinations, you can’t do without the RAG or model training process based on the knowledge base, and its own cost requirements will increase by more than one level.
The ticket to this tier is quality data and strong engineering capabilities, and all of this is inseparable from money…
Many companies want to use the company’s internal data to build an internal AI Q&A robot, which can be attributed to this level of shallow application, and all such shallow applications have the possibility of platformization.
The sixth level is the industry model
Although the fifth level has certain requirements for model training, the threshold is not very high, and its core relies on the combination of industry know-how and engineering capabilities.
Fourth-level users will expect a good industry agent platform, and fifth-level users will also expect a better industry model.
For example, AI doctors will want to rely on medical models, AI lawyers will hope to rely on legal models, and the improvement of the base model will greatly reduce the difficulty of project implementation, and there are costs behind it.
For example, there are many “black words” in various industries, such as fire and inflammation in medical care, buns and easy-to-use cases in the law.
If it is a general large model, in order to reduce this illusion, the team can only use engineering means, and it may need to be controlled very finely, but if the base model ability is already 60 points, and there are basic abilities on the side of industry illusion, then the engineering method will be less difficult.
To sum up, there will be “awesome teams” who want to train their own small industry models, such as finance, medical care, lawyers or internal company scenarios…
In terms of competency requirements, they need to have in-depth industry know-how, accumulate a large amount of high-quality industry data, have great experience in model training, and even need to be slightly involved in the underlying capabilities of the model.
This demand is often from units with certain demands on data security, such as large hospitals and large enterprises, and they also have the data and financial resources to build their own small industry models.
The difficulty of this level of application is not easy to say who is higher and who is lower than the fifth level, depending on the usage scenario, as long as there are more people, then the requirements for the model or application will rise sharply.
Level 7, universal pedestal model
Similar to DeepSeek, GPT, GLM, Wenxin, Qianwen and other teams that do underlying models, they not only need all kinds of high-quality data but also really need to explore the underlying technology of the model, while the first six levels are not needed.
brief summary
To sum up, understanding the above content basically understands the overall situation of the current AI application, and learning AI on this basis will be more comfortable.
However, the arrangement here is too vague, in order to facilitate you to have a deeper understanding of the evolution of AI applications, we use a customer service assistant to go to the deep water area.
Case customer service assistant
All self-media, including me, have a dilemma: the group activity is extremely low!
This involves a community operation impossible triangle:
If the group activity is high, then there will be more communication between fans, and their sense of gain will be strong, and on this basis, more orders will be generated;
However, if the group is more active, I must invest more energy on my side, but there are more than 10 or 20 groups, I can’t maintain it at all, and the final result is that only 1-2 groups are more active;
Moreover, I invest too much energy in group operations, which will definitely lead to a decrease in subsequent high-quality content, because the reduction of high-quality content will lead to fewer topics for discussion in the community…
So, I need an AI clone who needs to complete at least the following functions: be able to identify the topics that the fan base is paying attention to; Ability to present their own opinions based on topics; The insights section is best based on my official account articles;
1. Topic generation
In fact, from here on, it is a real test of practical ability, such as how to form a theme based on the content of this group chat:
The following is a topic generated by my prompt: AI side hustle high-income temptation and anxiety discussion
The prompt here can be written like this:
You are a professional topic analysis assistant. I will give you a chat record of netizens, please analyze these messages and find the most interesting topics, the specific topics must meet the following criteria:
1. **Avoid sensitive topics**: If the topic involves issues such as political sensitivity, illegality, or inappropriate speech, automatically exclude and give the topic a negative score.
2. **Concise and clear title**: The title of each topic should be short and precise, accurately summarizing the core content of the discussion.
3. **Content Summary**: Provide content summaries for each topic, highlighting interesting perspectives, thought-provoking questions, and depth of discussion.
4. **Personal Opinions and Emotional Tension**: Add appropriate personal opinions to the summary to reflect the emotional tension in the topic and avoid being too objective and neutral.
5. **Ranking criteria**: Return 1-3 of the most interesting topics based on the popularity of the discussion, discussion participation, and diversity of topics, with the hottest topics ranked first.
6. **Ranking Evaluation Criteria**:
– **Depth of topic**: Whether there is enough complexity and room for discussion.
– **Breadth of topics**: Whether it involves multiple related fields and can attract a wide range of discussions.
– **Timeliness**: Whether the topic is relevant to current hot topics, trends, or social phenomena.
– **Provocation**: Whether the topic can spark discussions, controversies, or the collision of different perspectives.
– **Practical Relevance**: Whether the topic is related to the actual situation of the group chat members and can trigger the sharing of practical experiences.
– **Emotional Resonance**: Whether the topic can touch the emotions of the group chat members and resonate with them.
– **Sensitivity**: Whether the topic touches on sensitive social or political topics.
If a topic has a sensitivity issue in its rating, or if the discussion is too extreme, offensive or unethical, give the topic a negative score and don’t return the topic.
7. **Return Results**: Do not return the same topic
The above is just an example, real prompt debugging is a long and boring process, how to make the model results always stable and effective, this is the real difficulty of prompt word engineering.
2. Topic discussion
Generally speaking, there are two approaches: simple approach, ignoring the discussion context, directly opening a separate answer on the topic or chat context, such an answer may have a mechanical feeling; In fact, the most difficult thing about AI application is multi-talk inquiry, because behind the multi-talk inquiry is the design of various SOPs, once the design is not good, the AI will talk nonsense, and the user will have a sense of separation;
For example, if we generate an AI reply to the topic of AI side hustle high-income temptation and anxiety, it will look like this:
10W per month? It’s just survivor bias! Do you really think AI is a money printing machine?
{
“content”: “Monthly income of 10w? It’s just survivor bias! Do you really think AI is a money printing machine? ”,
“viewPoint”: “Questioning the prevalence of high income from AI side hustles and guiding discussions on survivor bias”
}
Your background design looks like this: {
1. You are a person with a story;
2. Maybe you are the protagonist of the topic;
3. Maybe your friend is the person in the topic;
4. In short, you are very concerned about this topic;
}
Your purpose is as follows: {
At present, as a bystander, you want to know everyone’s further opinion on this matter in order to understand the underlying logic of this topic.
Therefore, on the one hand, you need to guide the topic and explore it in more depth, and on this basis, you need to throw out some questions and opinions; On the other hand, you can’t let people easily discover your tendencies.
To sum up, you need to design some leading speeches or questions to keep the topic going.
}
Your personality traits are as follows: {
1. You are a barber;
2. You will often…
……
}
The text requirements for your output are as follows: {
1. You need to output, and you only need to output a piece of text;
2、……
10. Correct output format: {“content”:”What is the content of the message”,”viewPoint”:”Your point of view”}
}
3. Introduce RAG
If you want to make AI expression better, you need to introduce RAG technology:
This is because when topic discussions enter the professional field, relying solely on large models can easily lead to two fatal problems:
- Lack of depth of opinion: the model is based on the content generated by the general corpus, which cannot reflect the industry insights accumulated by the official account for a long time;
- Style shift risk: Mechanical replies can easily destroy the consistency of readers’ perception of account IP;
That’s where RAG technology comes in. We build a knowledge base based on historical articles from official accounts, allowing AI to reply:
Here is a simple comparison:
When discussing “AI side hustle high income” in the group, traditional AI may give:
“There are indeed opportunities for AI side hustles, but they need to be combined with personal skills…”
When discussing “AI side hustle high income” in the group, traditional AI may give:
“There are indeed opportunities for AI side hustles, but they need to be combined with personal skills…”
After accessing the RAG system:
Knowledge base search: Automatically match the official account “Vigilance! Those AI Side Hustle Traps with a Monthly Income of 100,000 Yuan” and other 3 related articles;
Opinion Extraction:
- According to the 2023 questionnaire data, 92% of AI sideline participants earn < 3,000 yuan per month
- The “winner-take-all” phenomenon in which the top 5% of creators account for 80% of traffic revenue
- The hidden costs of equipment/computing power/data cleaning are often overlooked
Final Generated Response:
Have you noticed? (Quoting the “Trap” view)
Those cases advocating AI side hustle profiteering are essentially the same as the routine of ‘micro-business Xiti luxury cars’ ten years ago. Of the 47 AI studios I interviewed last year, only 2 were truly profitable, and they all required continuous investment:
Average equipment cost: 12,800 yuan/month (including A100 graphics card rental)
Data cleaning time: 63% of total man-hours
Customer churn rate: 82% of new customers in the first month
(Throwing up a point of contention)
If AI side hustles are really so profitable, why don’t the knowledge paid tutors go all out by themselves? The business logic behind this is worth pondering…
Here, through RAG, “content assets” are turned into “ammunition arsenals” for AI. Once the topic is triggered, the AI follows the retrieval→citation→ output logic, which can preserve the personal tone and compress the illusion.
Seven-layer architecture for AI applications
1. Generate SOP
In the case of community management tools, I can actually write a piece of code except for RAG!
What I really do is:
- First, intercept a conversation from the WeChat group;
- second, generate a topic and summary based on these conversations;
- third, use prompts to let AI generate what you want to say through topics and chat content;
- Fourth, send the AI’s words to the group as they are, and then continue to collect fans’ chat records;
- Repeat the above process…;
In this process, in fact, I only need to keep repeatedly adjusting the prompts to make the content of his chat smoother!
All of the above actions are generating SOPs and prompts to be matched with SOPs!
2. Form an agent
When the SOP adjustment is completed, I will need an agent to perform the group chat task, and the easiest way here is to directly use the orchestration tool such as Buckle to build the program.
Because he also supports the knowledge base…
After the basic verification, a set similar to the customer service agent is implemented, and the button program can be migrated out to form its own presentation, for example, we made our own set:
3. Form a platform
If this set of agents is indeed very well operated, then it is necessary to platformize it, with the purpose of opening it up to various self-media, so that each self-media has its own set of AI clones to solve the problem of the impossible triangle of the community.
From this, we can see that in fact, the above actions still cannot jump out of the seven-layer framework of AI applications…..
epilogue
This article is already relatively long, and the multi-talk and inquiry part will not continue, which is the most difficult section of AI application at present…
Finally, to summarize: through the combing of the seven-layer architecture, it is not difficult to find that the essence of AI applications is the gradual integration of technology and scenarios.
From the initial conversational interaction to the in-depth training of industry models, each level corresponds to different resource investments and technical boundaries.
Current industry practice reveals two core laws: first, the release of AI value is highly dependent on the granularity of the scene, and the more accurate the SOP disassembly, the more it can reflect the technical advantages; second, the “double helix structure” of engineering ability and industry cognition determines the depth of application, and algorithm optimization without data support is equally dangerous as technology stacking out of the scenario;
For practitioners in anxiety, it is recommended to adopt the “three-order positioning method”:
Firstly, the technical somatosensory (L1-L3) is established through API calls, and secondly, structured data (L4-L5) is precipitated in vertical scenarios, and finally a moat is built under the dual barriers of data assets and engineering capabilities.
It is worth being vigilant that the current technology iteration speed far exceeds the application digestion capacity, and blindly pursuing the frontier of technology may fall into the “capability trap”.
Only by maintaining a pragmatic attitude of “scenario-driven, data foundation, and project implementation” can we achieve a leap from cognitive upgrading to value creation in the AI transformation.
