ChatGPT made me a "Superman" - a phased summary report on how to improve the team's performance by 30% and improve quality by 100%

creative background

        When CHATGPT first appeared, there were ten thousand "horses" galloping in my heart. I was repelled, BS, and even turned off the screen to cut out relevant news, and even unplugged the TV at home. Because its performance really hurt my self-esteem.

        This kind of emotion stems from my own "not self-confidence". The premise of not being confident is that what I hear is too "impact". But I am a person who knows how to reflect and look back. The reason why human beings can evolve is precisely because human beings will continue to improve.

        During this period of time as an "ostrich", I thought of the famous words mentioned by the industry when JBOSS was invented until JBOSS3 was downloaded millions of times a day:

• first they ignore you
• then they laughing you
• then they fight you
• then...you win

        After about 15 days of rejection and "ostrich", I came into contact with ChatGPT by chance. My feelings at the time were like this:

        With in-depth study and use, I first tried to use it for people around me, and then used it in my own team. Then I realized I fell in love with it.

        After nearly 5 months of use and popularization, we still got a lot of amazing results, so I made a material to record, publicize and popularize ChatGPT, and I also hope that this document can bring more people benefit.

        The whole content includes the following contents.

 What are the best AI tools currently on the market?

        It has only been 4-5 months since GPT became popular, and the market has been flooded with a bunch of various AI or AI-like tools that claim to be AI. I believe that everyone has experienced some disadvantages. Here we are going to talk about the 4 things that are most commonly used by our developers and workers, the easiest to access and the most effective. They are:

• Copilot
• Midjourney
• ChatGPT(3.5 turbo-4)
• Cursor

        Among them, Midjourney and Cursor are "killer" level existences. The killer here refers to the many things that are currently happening, such as AI replacing people and AI grabbing jobs.

        After these months of using these 4 tools and the effects of using them at work, I made a comparison of the 4 tools based on "function points" as follows:

         We can see among them:

• midjourney focuses on creative, creativity, pictures and even video (AI creates movies by itself);
• Copilot focuses on office and smart assistants;
• Cursor is dedicated to "code farmers", it is a development tool, but it concentrates dozens of development languages;
• ChatGPT is more diverse, especially ChatGPT4 already has the "derivation" consciousness that humans only have. I will show two real examples of ChatGPT4 using the derivation consciousness that humans only have;

Instructions on how to use these four tools well

Use of Midjourney

        We all know the photo of "couple", or some people don't know it, so let me tell you that this picture was drawn by AI itself. Many people thought it was a photo taken by a real person before the official "disclosure" . It can be said that its effect is so realistic that it is trembling, it is just a picture generated after using a very refined natural language description.

        To use Midjourney, you need to have a Discord account. The current standard version charges 30$ per month, unlimited picture generation, and supports 3 concurrent clients. Generally, the 30$ standard version is enough for our work. used.

        This Midjourney is very powerful, and many junior UI, art, cutting, advertising copywriting, and creative posts were cut because of this thing.

        The essential difference between Midjourney and other previous AI tools in automatically generating creative content is:

• Some previous AI-generated images have fixed parameters, but the values ​​in the parameters can be played freely, and users still input fixed content according to the layout and input box established on the interface;
• The parameters of Midjourney are not fixed, there is no formula, it is just a natural language description, then: Done;

        We usually work in these positions or need to do work related to these positions. What we often need to do is to make small logos, base maps, and backgrounds. If you are not a professional artist, then you have to apply for the company’s art resources, and argue with the artist back and forth, modify, and sometimes some people will get angry when there are too many revisions.

        Now with Midjourney, for example, if I need WeChat’s log when I’m writing a PPT or project, I’ll take a screenshot and tell Midjourney to cut out the WeChat Logo for me or draw a PNG high-fidelity vector according to the area I’ve marked. picture. Often it doesn't take more than 1 minute to complete it.

        I even let it generate thousands of background images for me. It only took 420 minutes to generate these thousands of images with the effect of selling 500 yuan per page. I was sleeping for a minute.

        I think of a stalk in the Internet tourism industry before, called "100 pictures made every day". . . it's out of the question. But now there is no need to make 100 pictures a day, but 1,000 pictures in 420 minutes. Each picture is extremely exquisite, so exquisite that it makes people tremble. This is the benefit of Midjourney, so we will call it a "killer" level existence.

Use of Copilot

         Copilot can still be used by individuals before, provided that:

1. You must be a paid user of office365 (more than 300 yuan a year, very cheap);
2. Your office365 version needs to be > 16.0.16325.x < 16.0.16327.x, and then choose to join the experience plan in the file menu of office;
3. Modify the registry method (Baidu experience has been shared), and individuals can use the full set of Copilot functions;

        Now that Office365 has been upgraded, this function has been removed from the personal user experience. At present, Copilot only cooperates with enterprises, so only enterprise users can use it within the enterprise.

        But what Copilot brings us is not only amazing, it connects all emails, teams, word, excel, ppt, etc. in Office with the gpt engine. The cumbersome styles, formats, proofreading, typesetting, generation formulas, meeting invitations, mass emails, etc. in the past can be described in natural language and completed in an instant.

        For example, I often write in excel like this: use the multiple regression formula to generate an analysis report for the table I selected. Within seconds: Done!

Application demo of Microsoft Copilot in Word, PPT, Excel

 Use of ChatGPT

        In fact, we did not restrict ChatGPT, but OPENAI restricted us. At present, to use ChatGPT, you not only need to have a foreign IP, but also need to have a foreign mobile phone to receive text messages. With the release of GPT's IOS, if you need to use the mobile version of GPT, you must also have a US version of AppleID.

        Therefore, we often use ChatGPT in China through the "mirror" built by some technical circles. Generally speaking, this kind of mirror is OK and completely usable. It uses a GPT account and then connects to foreign sites through the GPT API of OPEN AI, and then makes a layer of JS for interactive web pages in China. In fact, the content is completely GPT and I have used it to the present experience. Still, it works pretty well.

        The only thing to pay attention to is that this kind of shared Token or mirrored ChatGPT is not exclusive to Token, so you need to be careful to protect personal privacy related information.

        In addition, both the official and the mirrored ChatGPT are charged. This fee is actually very cheap and negligible. For example, I use 20 yuan and 160,000 Chinese characters. Basically, I have not used any searches for several months after I have ChatGPT. The engine is gone, and with this amount of usage, I can't use up this fee. If there are any advertisements or any products on the market tell you: this GPT is absolutely free! Then this information is either false or deceptive, users need to be careful.

        The essential difference between GPT and search engines or to say: the ultimate reason why search engines have come to an end is that if you use search engines, you will get a lot of "entries" and answers. Whether these answers are useful or not, you need to constantly verify and try. Discrimination ability and "deduplication", sometimes the availability of these answers is <10%, which will consume a lot of your time. And the answer given by GPT, in the case of a proficient and familiar with GPT questioning skills and even questions can be made into prompt projects like "classes, functions, callbacks", its accuracy rate is currently as high as IT technology. It has an accuracy rate of 99%, and its accuracy rate is as high as 80% for reading comprehension analysis of dialects, classical Chinese, or essays similar to "Lu Xun".

        This kind of answer speed of 99% accuracy in one second or two seconds has completely eliminated any non-GPT search engine ( note: here I am not using will, but already ).

Use of Cursor

         Cursor, this is another "killer" level equivalent to Midjourney, it "kills" programmers.

        That's right!

        It really has "killed" a lot of programmers.

        Think about it, everyone, usually we take a table with 3 tables, and then explain the business logic, need to use spring boot2.4.3, jdk8, use standard controller, service, mybatis, and explain which operations I need to use redis and which ones need to be asynchronous , interface specification, naming, and recursion use DFS+ intelligent optimization search to improve performance. The front desk needs to use vue2.js and element-ui. The format of one page and next page is still 12345, which is a pagination format.

        It doesn't take 30 minutes, as long as you list these points one by one and describe them like the following:

1. You are a Java programmer, the JDK you can use is OPENJDK11, the development tool you are using is eclipse2.4.6, and you are developing a project on spring boot 2.4.3;
2. I need to generate mybatis DAO layer using ibatis.annotation for three tables A, B, and C;
3. I need to write a service, in which there are xxx public methods whose return value is xxx;
4. For the xxx method, I hope to use the apache commons toolkit to simplify the code and increase the robustness of the code;

        。。。。。。

        Then from back to front, a whole set of code is generated for you. After fine-tuning, it can run with 0 BUG in 10 minutes. That's right, I've done this hundreds of times, and the code works really well in production, and it performs even better than my own handwritten ones.

        For example, there is one: the generation of a stepless tree of commodity catalogs. I use DFS+ intelligent optimization search, and I can already process 100,000 levels of data within 117 milliseconds. And GPT uses the A* algorithm, namely: the best path algorithm, its code running efficiency is: 45 milliseconds, and the CPU overhead is 10% less than mine.

        We can say that as a person who thinks logically, has demand understanding, fine design, and knows how to divide, he can use GPT to pair with him for "team programming". At the same time, because such people have the assistance of GPT, some small and medium-sized companies can completely eliminate the need for "architects" and reduce the need for junior developers who need to rely on "human flesh".

        Because an architect of such a small and medium-sized company is just doing these things, the key is that his results, delivery, and output still need to go through continuous testing, verification, and even fail to meet what we say a real programmer should write. code quality, data structures, and design patterns. And what about primary development? The "follow-up" of their delivery and output will also consume the time cost and trial and error cost of the team and the enterprise.

        These are the scariest things about GPT, but we don't need to be afraid of it and reject it. I will explain this further later.

        At this point, we will briefly introduce these 4 tools, and let's see how they are charged if individuals use them

The charging model of each AI tool

         Here is especially Cursor, which can be downloaded, used and recharged without scientific Internet access, but it only supports the CVV payment mode of credit card, that is, your credit card must be VISA or Master Card, and it must be on the back of the card or at the time of payment. There are 6 or 8 digits, the "confirmation code" with the last 3 or 4 digits, which is called the CVV payment mode, can be recharged. 

        At the same time, Cursor does not need to surf the Internet scientifically, it can be used anywhere in the country, and there is another point, it uses the GPT4 engine after recharging.

        After recharging, your account is no longer restricted, but like this:

        Generally speaking, I use ChatGPT as a search engine, Cursor as programming, and Midjourney as PPT background, material, and even some ideas. So I will go to mirror ChatGPT to recharge 20￥, Cursor I recharge on the 1st of each month, 20$ is equivalent to 138￥ RMB according to the exchange rate, and Mijourney30$ is about 200 yuan. My monthly recharge amount is even lower than many people. mobile phone plan fee. And what I got was close to "superhuman" abilities.

        This ability can be viewed in this way.

• You can use Mijourney to help you create ideas through two devices with computing power, and at the same time use Copilot to lay out the materials generated by Mijourney in your PPT;

• You can let Cursor generate codes based on your prompts over there, while on the other side you are using ChatGPT to assist in writing another piece of business code;

         Some people say: My manpower is quicker, and I can actually do the same thing.

        That's right!

        I have already discovered that 3 devices are doing 3 different things at the same time, the key is to add one point. What kind of person is a person who knows how to do this thing, what is called good thing, what is called fine thing, and what is meant to make this thing to the extreme? Even if you produce 3 things at the same speed as you , the quality of the "delivery" of people who use AI tools is top quality and top level, and you must be accompanied by the modification of the BUG after delivery, constant polishing, knocking, time-consuming, labor-intensive, bitter haha How do you feel when you are not well?

        Another example: I wonder if you have noticed that my PPT production ability in my blog is getting better and better. These PPTs have a style called TED background and TED style PPT.

        I tried to sell a few PPTs myself in March. Each page sold for 500 yuan, and one page sold for 800 yuan. And the processing of each of these PPTs may take a full day or even two days, which is only the time-consuming required for a skilled worker with a professional UI and graphic occupation.

        With AI assistance, I can produce a UI/UX floor plan or a high-fidelity prototype interface within 40 minutes. This is what I call the existence of the "killer" level.

        It's brutal, but it's happening.

        After talking about the charges, we will really talk about how to use GPT so that it can truly achieve 99% technical accuracy as I said, and it can be achieved in some non-technical fields, especially non-English fields. What about the effect of more than 80% accuracy?

        Because many people in the industry are indeed using GPT, and almost half of them call it "artificial mental retardation". Because whether it is GPT or other domestic related AI, they are not so accurate when they use it. Why is this?

        I need to tell everyone here that yes, the emergence of ChatGPT has indeed lowered the threshold of the entire technology, and it is not an exaggeration to call it the fourth industrial revolution.

        But not everyone can use it well , let's use an example to illustrate the correct usage posture of Chat GPT.

Mastering GPT skills by asking questions - 4-segment prompt

        Most people use GPT to ask this kind of "lazy" question, and the answers you get will be very rough, so many people even call it "artificial mental retardation".

        And smart people and careful people see how they ask GPT questions:

        Assume a role, a description of what kind of thing needs to be done, what kind of requirements are expected to be achieved, needs to pay attention to avoiding, shielding, and eliminating "interfering" information requirements + supplementary questions.

        We can see what kind of effect is the answer to the "correct question" in the positive example above?

        Therefore, we often summarize the prompt message that makes good use of GPT into a "4-segment" structure to construct your entire prompt message.

        I don't know if you have a feeling that in work and life, some people can clearly describe the ins and outs of a matter in very organized and logical language within 5 minutes or 3 minutes.

        At the beginning of 2000, we had the "Wang Guorong" version of ASP3. Is the boring computer knowledge telling you clearly?

        Another example is mine: a series of nanny-style tutorials.

        These people, these documents, and these deliverables all have a characteristic, that is: no matter whether there is AI or not, you already know how to do it and do it well. So when this kind of thinking is integrated into your prompt message to GPT, the effect you can get is the same as what you can do yourself , except that GPT is another you that accelerates you thousands of times. And myself .

        Regarding the classic 4-paragraph question, the author compiled some practical examples to illustrate this technique

Use 4-segment questions to let GPT help production slow SQL tuning

        Just imagine if I just asked it "lazy": tune this SQL, what kind of results do you think you will get? Maybe it might not be as good as the effect of studying for 30 minutes by yourself? So "artificial mental retardation" was born.

Use the 4-segment question method to "make up for" the GPT question

        GPT even "refuted" me here.

        Because the plan it gives is optimal, all things accumulated based on experience and years are vulnerable to GPT. I took its results to verify with the DBA human flesh review and the production environment, and the result was that what GPT said was reasonable.

You can use "make up a knife" to improve the answer of GPT

 4-stage precise questioning allows Midjourney to generate an Asian-style IT man

         It actually wants to add a pair of glasses to the generated character portrait.

Generate 0BUG runnable code based on natural language

        I even gave the whole project to Cursor, and then you see: I wrote a prompt message project with a small hundred words. You can think that I regard GPT as a new employee , so I need to tell it how the framework is used, the jump mechanism, and the related data flow. Then I told it where my problem is. I haven't written a single line of code from the beginning to the end, and I don't know where to write or add this code? Then it understands for a few seconds, resolves: Done, runs: 0BUG.

The 4-paragraph question is only the basis for using AI well, and there are more prompt messages

        I recommend you to read a document called "ChatGPT The Art of Asking Questions", which will teach you how to use GPT really well. This document is available online in Chinese and is free. The key is that it belongs to the "nanny tutorial" category. This document can be a good first step for everyone to lay and consolidate the use of AI.

What kind of "efficiency improvement" did these 4 tools bring to our team?

1. I have checked a whole set of additions, deletions, and changes in the background of a cross-5 table, using spring boot2.4.3, openjdk11, eclipse, mybatis, vue2, redis, mongo and other common technology stacks, and I have carried out before using GPT and using Comparison of the whole development process after GPT. The efficiency improvement has reached about 60%;
2. I myself implemented a waterfall flow in Android when the background interface was ready. Before using GPT and after using GPT, the efficiency has increased by 50%;
3. I am also working on my own Android application. Such a home page contains one row and three columns, two rows and two columns, and a waterfall flow. The efficiency before and after using GPT is as high as 40%;

        This is not counted, the key is that I divided the project I made into two sets of deployable projects, _nogpt and _isgpt, and carried out the concurrency test of jmeter respectively. Because of the use of GPT, at which points, which segments, and where do I need to use such as: intelligent optimization algorithm, the most efficient algorithm, here I want to use fast sorting, here I need to use binary points, here I use DFS combined with A* algorithm. . . So the _isgpt underscore project is under the same concurrency, and its CPU consumption is a street away from my own handwritten code. That is to say, in the case of the same delivery code without bugs, GPT is fast and good.

        Of course, we can't just look at the figure of 60% average efficiency improvement on "paper", it will be a disadvantage! Look at it this way: because this project is developed from 0 to 1, in fact we have the existing production environment and historical "shit mountain" code over there.

        Therefore, in an actual and real team combat environment, how can GPT improve efficiency?

        We all know that in the actual production environment and team cooperation process:

• What cannot be saved is requirements analysis and understanding, existing code diagnosis, deciding how to do it, where to do it, and even cross-team collaboration, meetings, etc. These usually account for 30% of the entire project workload;
• Those that can be saved are: First of all, this technology is relatively unfamiliar to you. To realize it, you need to search online, copy & paster, debug, and demonstrate. It may be far more than 3 times. The process of repeating the previous steps, we have counted the proportion of this part Around 30%. This part is precisely the workload that can be "saved" by AI , and because of the proficiency in using AI-related tools and the personal ability is also good, we can often achieve >30% improvement in development work efficiency. So far, the author has contacted students who work at the headquarters of Facebook and LinkedIn in the United States. They are a group of IT companies that used ChatGPT earlier. They gave me a number: an average efficiency increase of 25%. This shows that 25%-30% efficiency improvement is definitely there. The results are astonishing. Because it is not only faster but also better in quality.

 Two AI-assisted processes already in use in Teams

        In fact, we have already used AI to review and write code or do POC in normal times.

        Especially SQL Review. One thing happened to us last week, that is, a business logic that is difficult to trigger was triggered in a thousand years of production. There is slow SQL here, and it takes more than 30 seconds. MySQL times out even after running it. Fortunately, it is an asynchronously executed xxljob, which involves more than 2 million data queries.

        I took a look at the code and it's two years old.

        At this time, a group of people in the on-duty and operation and maintenance groups are still discussing how to optimize this SQL?

        I used GPT, and then GPT gave me an optimal writing method, and gave me hints because of my 4-paragraph questions and the situation of making up. I found out that this SQL problem is actually a joint index from left to right. If there is no parameter for the left index in the code, the participation of the right index in the WHERE condition will also cause the entire index to fail. So in the WHERE condition, I set the left index condition and wrote it before the original WHERE condition, and added an "AND" before the original WHERE condition, only changing such a point. This process was prompted by GPT until I realized that it only took 11 seconds to improve the whole process. After the whole modification, we added a where myid=339 and the original right index-warehouse ID=001 , and then we put this SQL into production to run, and it ran from more than 30 seconds and timed out to 1.7 seconds.

        Another piece is the recursive algorithm of our "commodity catalog", which is a stepless tree. I have used the DFS+ intelligent search algorithm to reconstruct it so that it only takes more than 100 milliseconds to complete. Then GPT gives a stronger algorithm, it does not use recursion but uses BFS, the algorithm is more than 50 milliseconds faster than mine, and it will not cause the problem of stack overflow.

What role did Midjourney play in my usual work?

        Our team has FTE, outsourcing, and a team composed of different suppliers. In order to improve the cohesion, sense of centripetal and sense of belonging of the team. I often need to send some encouragement and motivation pamphlets to the team every month or every year. These kinds of brochures require professional graphic design and can take days to pull out beautifully.

        As for me, I directly opened Midjourney and said to it: To generate a poster for 5.1 Labor Day, I don’t want red and green posters, but small and fresh ones, and petty bourgeoisie ones with strong white-collar work elements. . . So it generated 10 each for me, and then let me choose.

        One more: I need to generate a background image for the Qingming Festival, which needs to be XXX by XXX in size, and needs a poster with spring, pastoral scenery, Qingming elements, small and fresh petty bourgeoisie, which meets the taste of white-collar workers without losing childishness.

        In this way, I made a small poster for each major festival every month throughout the year, and then embellished some words of encouragement and gratitude on the poster in advance, and I got these small posters.

        I often need to make plans in my work, and I even bring some plans or research results to some industry conferences for publicity. The TED background PPT (Like this one) often gives me a headache. Although I spent almost 8 years from 2012 to 2019, transforming myself from a color-deficient IT nerd to one with the ability of an artist. AE, PS, CoreDraw. But in order to pull out this effect of PPT, it is necessary to achieve amazing effects in typesetting, vision, and layout in the meeting of thousands of people or ten thousand people. It takes me two days to make such a PPT. More than 20 pages of PPT for a plan. . . Then I have to be tired and vomit blood?

        So, 5.1 is at home, I have nothing to do, I wrote a bunch of prompt messages, and started 3 midjourneys. . . Then click OK. It just started generating these backgrounds, templates, planes, I saw it took 420 minutes, so I went to sleep. Woke up, it was generated, I looked at it generated over 1,000 templates and backgrounds and planes, all the things I wanted.

Use AI to enable yourself to have cross-industry capabilities

        I have a former colleague who lost his job years ago. Therefore, I have been doing "personal development, independent development", that is, the kind of development that does not require investment and can do projects independently.

        There is a project that needs to use spark+python to make these operational reports. And my former colleague used to work in Java, and he didn't know spark+python very well. But he still has a good understanding of the hadoop system and some details of professional knowledge. So under my introduction, he first used the corresponding tools for two weeks. . . Then he suddenly disappeared from my circle of friends. One and a half months later, he reappeared and sent me a screenshot of the production environment where he had launched this reporting function. Because it was a production environment, I masked out some sensitive information.

        He told me that he used copilot to analyze some confidence and flat data in excel with multiple regression formulas, and then used an open source called "chargpt" that can generate charts based on natural language. The "conversion funnel" was generated without writing.

To use AI, you must first learn how to do things without AI

        Through the previous examples, my use in the team, and the examples of people around me, it is not difficult to find that those who really use AI well and generate value have a commonality, that is, they themselves, including myself, belong to " People like "Scholar" or "Jump King", we have the habit of digging into the essence of affairs in terms of work or hobbies. Even without AI, we can do this well, but we need to pay more time cost .

        In fact, this question still goes back to a point I mentioned earlier. Don't think of AI as a panacea. AI is an assistant for people and an auxiliary tool. It is not a "panacea" .

        We look at the emergence of LLM AI such as ChatGPT in this way, let's make a comparison:

        It is said that many business departments at work are always: What system do I want? I want a WMS!

        Well, after you are really given a tens of millions of WMS systems, will your warehouse management and your procurement be better?

        For example, why do we say that no more than 1% of the new retail companies in China really did a good job in the past, but 99% of them ended up "collapsing" because of the middle-end, system, and digital results mentioned in the new retail?

        This is because the essence of all things has not been grasped.

        It's like, you are a professional financial accountant first, and you can do this thing and this account well without having a computer or a system. We said that before the emergence of new retail, there were many well-established and larger retail companies. We said that many world-renowned supermarkets in places without automatic cash registers, code scanning, face scanning, and Alipay can also achieve a comprehensive combination of customer satisfaction, food freshness, experience, and price.

        Let's talk about a programmer. Before AI, he can still achieve the level of a real programmer and architect through continuous and daily self-improvement skills and levels.

        So this is what we said, the premise of using AI is that you must be able to do it first. You have already figured out the need, where to do it, and where to do it, then you can teach the AI ​​the remaining operations and repeated steps. If you don't even have the ability to do a good thing yourself, how can you expect AI to do it for you?

        If you can do it well, AI can help you do it faster and better.

        You can't do it yourself or can't do it well, so how can you expect AI to help you do it well?

        Therefore, this also has a profound impact on the future (it is no longer the future because it has already happened) in the employment of enterprises and teams because of AI, in terms of hiring and using personnel, and how to select talents.

"AI quality" has become the assessment standard for enterprise selection personnel

        I have some classmates who are in other big factories. They have joined the assessment of AI experience in the interview (if you have experience, I believe you must regret the previous few passes. The results are all fail. AI use experience interview this off).

        At the same time, the interviews for a programmer, developer, architect or R&D post, especially the written test, have also begun to change.

        As we mentioned before, the algorithm written by AI, especially the code written after the DFS+ intelligent optimization is optimized by Cursor, has reached the level of Google Level 4 or >ALI P7. Then it seems that it is not so suitable for the present time to use the "eight-legged essay"-style written test to test candidates.

        However, the assessment of the principle, algorithm, and what kind of algorithm can really be applied to the business code we are working on has become the top priority.

        Why are our programmers most afraid of algorithms during interviews? It is because the progress of the project is relatively fast, and junior and middle-level personnel rarely or hardly use algorithms and data structures in their daily work. This has led to: 5 years to become a senior, 9 years to change the structure, 11 years to go to Didi and run Meituan, the origin of the stalk of Are you hungry.

        In fact, after analysis, observation and the help of GPT, at least 50% of the time you need to use algorithms and data structures in your CRUD.

        Therefore, this is the ultimate reason why algorithms and data structures are a must for a better R&D team and enterprise. Of course, the premise is that your company's business flow, traffic, data volume, or the number of stores and members in the supermarket have reached the threshold of >1 billion records per day. Therefore, such enterprises can survive and develop better.

        The same is true for individuals. We have AI tools that can instantly transform things accumulated through repeated memory and experience into productivity. Then the people we need need those who understand the principles and essential problems.

        Just like GPT's Prompt Message project, it actually looks more like a piece of pseudocode at first glance, hollowed out, or even just code comments, and then divided into 1-2000 word segments. Then throw it to GPT to let it complete in batches at one time. In fact, when we didn't have GPT, if your usual design is the same, then your human development work must also be more efficient and deliver higher quality than other teams.

        Therefore, the future-in fact, it is not the future, but from now on, our requirements for personnel are excellent overall quality. A person who has good logical thinking and knows how to subdivide and divide projects, projects, and codes is more than 3 junior and middle-level developers (3 is a small number, sometimes it can be up to 5).

        Many bigwigs in the industry, including those who resigned to start their own businesses in the AIGC sector, said something pertinent. They said: In the AI ​​​​era, this is actually a return to the IT industry. The original search engine + copy & Paste has now become something that I really need to settle down and read more books, papers, and dig deeper into the principles .

In the future, purchasing computing power will replace "training courses" as your daily expenses

        Three years from now, the personal cost of AI will be further reduced, and the market for the money we used to buy lectures, continuing education, and training will begin to shrink significantly. That is, in the future, a person's ability to purchase or know how to purchase computing power determines the person's "class" . These things have happened again and again, like mushrooms after rain, everywhere.

        Therefore, with the help of GPT, it is no exaggeration to say that individuals or teams can achieve "changing fate". Take why I have been talking about the current national pillar industries, the first half of the Internet finals has ended, and now the second half has just begun, hard core technology and other content? This is precisely because our old IT and Internet IT are still "tertiary industry development, enterprise development" in the final analysis, so we don't say that we must change jobs, we say that we need to have some slash skills? That is: on the issue of what else can you do besides programming, it is not enough to have ideas and the right direction. People who have worked for more than 10 years, 20 years, you must know that interlacing is like a mountain. And with GPT, it can exponentially shorten the time for your "cross-industry" study.

        For example, now the rust language has gradually become mainstream. Without GPT, you may need to spend at least a year to learn, get familiar with, and do projects to reach the entry level. With GPT, your entry may only take 3 months or less. time to arrive.

After 2023, the points and trends that ChatGPT will break through

        Before talking about this question, let's take a look at GPT4. In fact, it is already close to two examples of human derivation.

First example :

        I showed GPT a picture like the above. After reading it, ask GPT, Where will Sally look for her ball. GPT's answer is: The ball is in the box.

second example

        The screenshot below gives you an urge to post the emoji of "Skeleton" in WeChat after reading it.

 

 

        So GPT you ask it: Have you broken through the Turing limit? It always answers that it is AI, not human. In fact, it is already close to human ability.

        Because the ability of GPT is a kind of "emergence", which is only possessed by "consciousness". Moreover, the appearance of such an ability is also caused by a coincidence among all things in the universe. We think that GPT is a bug in this universe, leaving a gap for us mortals to become "supermen".

        Therefore, in the future, especially in 2024, GPT will definitely break through Turing’s limitations. At the same time, judging from the four mainstream AI tools, AI still assists humans in more scenarios, and does not “replace humans” to execute.

        For example: if you use Cursor, you need to let GPT generate code for you, and then copy it to Eclipse or VSCode for review, verification, and execution.

        In the future, when AI can improve local ZC, laws, and ZD, it will definitely "replace people" in relatively safe scenarios.

        In fact, there are already some places in GPT that are executing for others:

• For example, in Copilot, an 18-page PPT is generated according to my needs. Generate Excel analysis based on my natural language description;
• For example: under the GPT4 polymorphic model, it is already possible to generate high-fidelity prototype interface diagrams, as well as Android and React JS codes by hand-drawing a sketch;

        In the future, in more scenarios, GPT4 and 5 will be implemented "for people". At that time, the productivity of the whole society will be further liberated.

Talking about Enterprises Owning Their Own AIGC Capabilities

        GC is generate content, content, or knowledge.

        We can see that in new retail, we call some business scenarios of supermarkets and department stores "high-frequency scenarios". This is because commodities in high-frequency scenes can accelerate cash flow and gain popularity.

        In the past two years, knowledge consumption has clearly become the main consumption scene beyond retail items. Compared with a dispensable or good-tasting Wellington steak, lipstick, and skin care products, users are more interested in getting something that can last for a long time and continue to add value and bring added value over time. What people need now.

        This kind of thing is called content (knowledge)!

        Therefore, "selling content (in fact, selling knowledge)" is bound to be the next retail outlet. Therefore, many companies have also begun to seek self-build and export AIGC capabilities.

        In fact, this AIGC capability can be compared with the "portal website" in early 2000 and the secondary and tertiary vertical distribution websites. If your company is not positioned to do this top portal business, then there is no need to do AIGC's external Output and build AIGC service by yourself.

        Instead, learn to use AIGC to empower your business. It is the same principle as I said earlier that we use AIGC to assist individuals and teams to improve efficiency and quality.

        But still go back to the point mentioned many times above: the prerequisite for enterprises to seek AI empowerment in business is whether you have already run through this business by yourself? Has this business even been run using the MVP model? You can run this business without AI. AI is nothing more than helping you further expand this business and speed up the iterative update of this business so that it can enter the market more quickly to complete the launch and obtain corresponding returns . This is where it is really difficult for companies to do.

        Once you have clearly identified your business direction, these four major AI tools are now so mature, and the market is still constantly introducing new ones. More AI enterprise-level tools and platforms are being launched or will be launched. Seeking AIGC's empowerment is just a matter of It is as simple as spending money to open an account and connecting to the API.

        Therefore, I also investigated two LLM AI platforms, which can be used by teams or entrepreneurs who have the need to export AIGC to the outside world.

• One is MOSS, which was developed by Fudan University in China. There are already retailers around me who are building their own "talking food" and "talking video". For example, after watching an advertisement, you directly ask in the barrage: Where can I buy it? How about the taste, so there is no need for a customer service at the back, the video will tell you directly: sour, sweet, where to buy these contents, these are based on the modeling and training done by MOSS;
• Another is Stanford's LLM model, which is completely open source and very convenient to use. The reverse translation in Chinese is "llama". . . Hehehe, this name is really interesting;

end

        Ok, let me introduce these first. This is my sharing and feelings about some of the best practices in the use of AI in the first 5 months. Later, according to the phased results, I am still studying how to let multiple devices complete different tasks at the same time with high quality in parallel. I will share any results in time.

        Here we can summarize: Don't be afraid of AI. You are afraid of AI because you are not confident enough. The reason for not being confident enough is that you are not strong enough. We need to embrace AI instead of rejecting AI, because AI will only make a confident, self-improvement and self-disciplined person stronger or closer to God.

        I will end this article with a famous sentence describing GPT in the industry: OPENAI has enabled humans to realize the concept of "Protoss" mentioned by Yuval Noah Harari in the brief history of mankind for the first time. It allows human beings to truly unveil the veil of "God" and see the true face of God, approach God, and even talk to God.