Four, the advantages and limitations of ChatGPT

Author: JD Retail, He Yuhang

"In the past era, you needed to learn how to ask questions, but in this era, you need to learn how to prompt.

Introduction

While you are writing prompt code, Xiao Zhang has already completed data analysis; when you are organizing materials, Xiao Wang is shopping at Universal Studios; during the report, you find that Xiao Zheng's report has made a qualitative leap, but he has been playing badminton since 8 o'clock in the evening. Before, everyone's work efficiency was almost the same, why did they suddenly get a boost, are they secretly studying at home? The reason is actually very simple, because the AI era has arrived, and you need to[Learn to Prompt].

What is a Prompt?

The most popular AI product at the moment is undoubtedly ChatGPT.PromptThe (prompt) is the keywords or descriptions provided when communicating with ChatGPT and other AI products, to guide it to understand the needs or problems, and give relevant answers or suggestions.

Among many Chinese translations, 'prompt' more clearly reflects the equal relationship between humans and AI. This translation also carries the meaning that humans are the guides and enlighteners of AI, ultimately awakening the intelligence of AI through continuous 'prompting'. In addition to the simplicity of the meaning, etc., 'prompt' has become the mainstream translation.

Part Two: Why is Prompting Skill Important

Prompting has a costThis includes your time cost and actual monetary cost. If you apply chatGPT through an interface, the call is priced by token (which can be roughly understood as a word), taking GPT4-API as an example, the current charging model is:

Prompt: $0.03/1K-token, Response: $0.06/1K-token.

The cost of each interface call is about 1 yuan. Behind the expensive cost is the top-level graphics cards purchased by OpenAI company for computation. Since the cost cannot be ignored, the importance of ROI (return on investment) is self-evident.

If ChatGPT is an important productivity 'partner' in the AI era, then prompting skills are undoubtedly the key to improving ROI:It takes 10 prompts to get to the point, while solving a problem with one accurate prompt, the ROI gap is not just 10 times.. Continuously improve prompting skills, just like efficiently applying EXCEL and PPT in the past era, can effectively improve our work efficiency.

Part Three: Introduction to the Principles of ChatGPT

To better help us with this AI 'partner', we need to understand its principles so that we can better 'learn to prompt'. ChatGPT may seem magical, but it is actually like a super smartWord AssociationExpert. Although it can understand human language, it differs from human thinking patterns.

The 'smartness' of ChatGPT can be simplified into the following 4 steps:

① Divide your [Prompt] into individual words or fragments (known as tokens).

② Analyze these tokens, find key information, and pass it to a large pre-trained model.

③ This model willPredictionThe most likely next word or fragment after the given [Prompt].

④ Add the predicted words or fragments to the [Input Prompt] and continue to predict until a complete response is generated.

Through these steps, ChatGPT can generate meaningful and relevant responses based on your prompts.

Therefore, after sending a prompt to ChatGPT,It generates responses one word at a time based on probabilityEach response may vary slightly, as the generation process is essentially a probabilistic issue. How does ChatGPT determine the probability of the highest word? It is also learned, and the learning process of ChatGPT mainly includes the following three stages:

① Pre-training (self-supervised learning): ChatGPT acquires various contextual relationships by learning a massive amount of data on the Internet, establishing a vast knowledge base.

② Supervised learning: At this stage, human experts participate in guidance to help ChatGPT select answers that meet human preferences and expectations from among many possible generation methods.

③ Reinforcement learning: At this stage, ChatGPT evaluates and adjusts the quality of output based on user feedback. By continuously optimizing and improving the quality of answers through actual user usage, it improves and refines the quality of answers.

Four, the advantages and limitations of ChatGPT

After understanding the principle of ChatGPT, we can more clearly recognize its advantages and limitations.

Advantages:

1. Boundless erudition

ChatGPT absorbs the knowledge of the entire Internet and has an extremely wide range of knowledge coverage. Therefore, when answering common sense questions, it can provide accurate answers.

2. Ability to summarize and generate text

Based on the learning of a massive amount of internet articles, ChatGPT can well extract the key points of articles and logically expand the text content. This is what people often say about the ability to go from many to few and from few to many.

3. Learning ability:

After training, ChatGPT can understand new processes and handle problems according to the acquired processes.

4. Internet of Everything:

ChatGPT can base itself on its own cognition, link its model with other functional/field experts, and realize a more powerful and flexible AI system to deal with various tasks. In the future, it will be like a brain, collecting diverse information to make judgments, mobilizing AI models in various fields to complete tasks, and linking multiple APPs to realize functions.

Limitations:

1. May fabricate content (hallucination phenomenon)

Because the essence of its answers is probability, once a word is answered incorrectly, all subsequent answers are wrong. This is also the frequent complaint about the GPT3.5's tendency to talk nonsense.

2. Non-connected model.

The knowledge of ChatGPT comes from offline modelsTherefore, it cannot directly provide real-time industry trend informationHowever, it is possible to refine key information with the help of ChatGPT by searching for authoritative industry reports and using tools like ChatDoc. Alternatively, you can use plugins to achieve networked information queries.

3. Difficulty in solving complex problems in specific fields

ChatGPT has been trained extensively in general knowledge. Therefore, when facing non-standardized problems in specific professional fields, the reference value of the solutions provided by ChatGPT is relatively low.

After understanding the advantages and limitations of ChatGPT, we can use it more specifically to solve practical problems. In the next part, we will discuss how to fully utilize the potential of ChatGPT through effective prompt techniques. The techniques in Chapter 5 have been verified effective by me and can be used with confidence.

V. Prompting techniques

Basic course

1. Clear expression

Clear prompts are the foundation for obtaining high-quality answers. Here, I will share a commonly used structure. After mastering this structure, you will have a personal advisor who can provide constructive ideas for the difficult problems you encounter in life and work. Of course,Not all promptsall need to include the following information, which depends on the level of detail you require from the answer.

InBackgroundcan include the following information:

① Role (WHO)——Including your role and related roles.

② Location (WHERE)——If you want to get targeted solutions, it is recommended to at least indicate that the location is in China, otherwise you may get solutions suitable for the United States.

③ Event (WHAT)——The specific event that occurred.

④ Time (WHEN)——The time when the event occurred.

In fact, these contents are the other parts of the 5W rule without 'WHY'.

After describing the background clearly, we need to define the "Goals”，which means what information we hope to obtain from ChatGPT's answers.

In real life, there are usually someConstraintssuch as time, money, and manpower.

Finally, we need to consider theExpected answerMake requests, which is a specification for the answer form of ChatGPT, for example:

① Require 'If you do not know, please reply 'I do not know', do not fabricate'. This requirement can effectively reduce the 'hallucination' phenomenon.

② Require a reply in a certain format.

③ Require multiple solutions.

④ Require a solution and the reason for it.

⑤ Require the number of characters in the reply.

For example, if I want to learn maintenance technology, the followingPrompt 2vsPrompt 1The quality of the results is much better. Everyone can try it out and see.

Prompt word 1:

How to learn maintenance

Prompt word 2:

I am an electrician (role), with two and a half years of work experience (time), and recently I found that my work ability has hit a bottleneck (event). How should I improve my abilities within three months (constraint). I am in Beijing, China. Please give me five practical methods to improve my maintenance skills based on the above background information (expected answer).

2. Follow-up questions skills

Since ChatGPT is good at answering questions in a structured manner, users need to identify the key information in the answers themselves. After finding useful information, they can continue to dig deeper through follow-up questions.

Prompt:

I am a graduate from the 24th class, and I should prepare for finding a job in the internet industry. There are three months left until the autumn recruitment. I am in China.

Further questions prompt:

How should one improve their ability to obtain industry information?

Continue to follow up:

How should one enhance the ability to filter industry news and information?

Continue to follow up:

Are there any detailed skills for information organization?

Through layer-by-layer follow-up questions, one can unravel the issues and find specific action plans with reference value.

attention should be paid to, as humans, ChatGPT sometimes forgets some of the information you have told it. Therefore, in practical applications,Include the key information of the initial question in the follow-up question, to prevent keyBackground or constraintsLost.

If its answer is incorrect, you can also point out the error in the follow-up question and request a correct answer.

Advanced course

With the ability of the basic course mastered, it can already help a lot in daily work. The advanced course can help everyone handle more complex problems.

1. Example method

The example method is a way to describe the【desired answer】. When you want ChatGPT to output results in a certain style, you can provide an example, and it will refer to the example for output. For example, if you provide a sample weekly report, ChatGPT will generate a weekly report consistent with the style and format of the example.

2. Specify a role for ChatGPT

This is an interesting anthropomorphism usage that often appears in various online prompt manuals. If no role is specified, ChatGPT's answers are usually neutral and listed. By assuming a certain role, one can obtain more colloquial answers, but this will not essentially improve the quality of the content of the answers. Here are two ways of asking, which everyone can try on their own:

Prompt case 1:

Write a product review for a laptop

Prompt case 2:

You are a technical expert. Please write an objective and informative product review emphasizing the powerful features of the new laptop.

3. 思维链（CoT）- Handling multi-step problems.

When ChatGPT needs to complete multi-step operations or rigorous logical reasoning, the思维链 (CoT) is a highly effective technique. It can enhance the connection between contexts, helping ChatGPT better understand the interdependencies between contexts.

Prompt case 1:

① Assess the current situation of the team: Please introduce how to identify and analyze the problems existing in the team's collaboration, such as poor communication, lack of clear goals, or unclear responsibilities.

② Set clear team goals and roles:After assessing the current situation of the teamPlease provide some suggestions on how to set clear team goals and assign clear roles and responsibilities to each member.

③ Establishing efficient team communication and collaboration:After setting goals and rolesPlease share some strategies and tools to improve communication between team members, strengthen collaboration, and thus solve the team collaboration issues identified earlier.

When dealing with complex reasoning, explaining the reasoning logic is very important for ChatGPT to reason correctly, as it can help it generalize from one example to another, otherwise it may fail in reasoning.

Prompt case 2:

Which method is the faster way to go home? Option 1: Take a 10-minute bus, then a 40-minute bus, and finally a 10-minute train. Option 2: Take a 90-minute train, then cycle for 45 minutes, and finally take a 10-minute bus.

Option 1 takes 60 minutes, that is, 10 + 40 + 10 = 60 minutes. Option 2 takes 145 minutes, that is, 90 + 45 + 10 = 145 minutes. Since option 1 takes 60 minutes and option 2 takes 145 minutes, option 1 is faster. (This is part of the reasoning logic for explaining the reasoning logic of question 1, requiring its answer to question 2)

Which method is a faster way to go to work? Option 1: Take a bus for 1000 minutes, then a train for half an hour, and finally a bike ride for 10 minutes. Option 2: Take a bus for 800 minutes, then a train for 1 hour, and finally a bike ride for 30 minutes.

In short, multi-step reasoning that is relatively easy for humans to understand can be somewhat difficult for ChatGPT. The use of thought chain technology can reduce the understanding cost of ChatGPT, thereby improving the quality of the output.

Some other types of application cases

1. Polish articles

ChatGPT is good at polishing articles. You can send your article to it and ask for polishing to eliminate misspellings and optimize the copywriting.

However, ChatGPT is only suitable for preliminary polishing because its polishing ability can only ensure the minimum level. For example, after I submitted this article to ChatGPT for polishing, I still had to revise it at least 3 times. The core of the article is the author's thoughts. By using ChatGPT to assist in writing, the author can focus more on the conception.

2. Summarize articles

Since ChatGPT is very good at summarizing the main points of articles, you can send the article to it and ask for the main arguments and evidence in the article. The output quality is usually very high. Some products have already emerged that use the capabilities of ChatGPT to analyze articles. The product I recommend is ChatDOC, and it is expected that similar products in China will quickly emerge in the near future.The product that used to be called a PDF reader may soon need to be upgraded to a PDF analyzer.

3. EXCEL application

Excel formula learningIt no longer existsYou only need to describe the analysis dimensions you want to perform on the Excel data. ChatGPT will provide formulas in simple scenarios and guide you on how to operate step by step in complex scenarios. Therefore, it can subvert the existing Excel usage process without any data leakage risk.

Prompt case 1:

Use a pivot table to summarize product sales: Suppose in a data table, column A contains product names, column B contains sales dates, and column C contains sales amounts. Please guide on how to create a pivot table to show the cumulative sales of each product.

Prompt case 2:

In the chart, 'columns' represent daily sales, and 'rows' represent provinces and regions. Now you need a mini-chart to show the sales curve of each province and region every day. How should you operate it?

4. Numerical extraction

Similarly, ChatGPT is also proficient in writing numerical statements, and accurate descriptions of the desired way to write numbers and lists can output code. If the formula does not work, you can also consult the cause of the error, and ChatGPT can also modify the statement according to the provided error cause.

提示案例：

Prompt example:

Calculate the monthly sales of each product category: There are two tables, one named sales (containing sales records) and another named products (containing product information). The sales table contains columns such as sale_id (sales ID), product_id (product ID), sale_date (sales date), and sale_amount (sales amount). The products table contains columns such as product_id (product ID), product_name (product name), and product_category (product category). Please provide an SQL statement to summarize sales by product category and month.

6. How to efficiently use 'prompts' in work

Now that we have learned about the application skills of ChatGPT in multiple scenarios, how can we apply it to our daily work to improve efficiency?

From a macro perspective, it is mainly divided into two stages:

The first stage can already effectively help everyone improve work efficiency. It is recommended to adopt the following methods for actual operation: ① AI efficiency based on the workflow. ② Process optimization based on AI. The second stage needs to be based on the foundation of the first stage.

1. Determine the work scenario:

Use WBS work decomposition technology to decompose your work content.

2. Sort out the workflow:

Based on the decomposed sub-modules, sort out the SOP.

3. Determine the type of problem:

Define the main problems at each stage of the SOP. Judge the proportion of repetitive labor that needs to be solved mainly. Theoretically,ChatGPT can improve efficiency for any repetitive labor relying on computers.

4. Accumulate prompt templates:

constantly accumulate prompt templates for specific types of problems. Subsequently, when solving similar problems again, simply replace the keywords.

7. Summary

In general, ChatGPT has good capabilities to solve repetitive and regular problems. ItLiberationliberated people from repetitive labor on computersProductivityenabling people to invest more energy in work that requires 'subjective initiative', such as analyzing and decomposing complex business scenarios.The most valuable 'thought' of humans has become more valuable.

Eight, a few thoughts

Last weekend, I attended a sharing session by Mr. Wei Qing, CTO of Microsoft. He is also constantly learning the latest knowledge in artificial intelligence. However, he said that if a lecturer claims to have mastered the application of ChatGPT, he will not attend such a course, because he believes that no one can fully understand the potential value of this technology at present.

I strongly agree with this. The content I share is just my initial understanding of ChatGPT. If it can inspire some of you, it would be my great fortune. I hope to learn and progress continuously with all of you, and together welcome the arrival of the AI new era.