Originally published at: https://mp.weixin.qq.com/s/PFEodUAOt5UxZSQGIJxKBQ
I. Introduction
With the emergence of ChatGPT, it has brought a huge impact to the world, and various large language models have sprung up. Overseas such as Google Bard, Anthropic's Claude, domestic such as Baidu Wenxin Yiyan, Ali Tongyi Qianwen, Xunfei Xinghuo Cognitive Model, Kunlun Wanwei Tiangong Large Model, etc.
The current large language model has a qualitative leap in code generation and code interpretation capabilities compared with previous models. Many programmers have begun to use large language models to write code, so many people think that "programmers will be replaced by AI."
This article will focus on two topics:
- In the era of AI, are programmers really going to be eliminated?
- How can we improve competitiveness in the age of AI?
2. Programmers will be eliminated?
The topic of "programmers will be replaced" was not brought up after the recent boom in AI . Before the emergence of low-code, many people have already begun to discuss whether programmers have been eliminated. However, although low-code can improve programming efficiency, it cannot satisfy complex scenarios, is not flexible or even safe enough, and cannot really replace most programmers.
I think it is too early to say that "programmers will be completely replaced by AI", but programmers who only know CRUD may face being eliminated . If you are a typist, if AI continues to develop, and the accuracy of AI's voice-to-text recognition is higher than that of some junior typists, then junior typists will face elimination. Now the AI code generation ability is very strong, and the simple CRUD may be completely replaced by AI. Therefore, if you only know CRUD and have weak other abilities, you will be easily eliminated.
But from the current point of view, there will be a variety of factors that lead to many senior programmers not being easily replaced.
First, the current large model code generation capabilities are still limited . Although the large model now has the ability to generate code, explain code and even optimize code. However, at present, it mainly stays in the "function" dimension, and it is not yet possible to generate codes for the class dimension, module dimension, and project dimension well. Moreover, the generated code will also have various problems, such as errors that cannot run, use of outdated class libraries, and some potential bugs. The generated code still needs people to optimize, deploy and verify. Complicated tasks also require people to split them into granularities that large language models can easily handle.
Second, data security limitations . Many companies, especially large factories, attach great importance to data security. They do not allow the company's business secrets to be disclosed to external large models, and users are not allowed to expose company codes to the external environment. Therefore, the application of large models in large factories is limited. It is usually only used to generate some simple tool-like codes or other codes without business attributes.
Again, the capacity of self-developed models is limited . Although many large manufacturers have begun to develop their own large-scale models, there is still a big gap between the data, GPU resources, and model capabilities compared with foreign large-scale models. Even if it is open to employees, although it can improve efficiency, it is still difficult. form sufficient productivity.
In the end, coding is only one part of the overall software lifecycle . Software development also includes requirements analysis, technical solution design, coding, testing, operation and maintenance and other links. In many cases, what determines the success or failure of software is the formulation and analysis of requirements, and the design of technical solutions before coding. Especially in the requirements stage, it involves a lot of interpersonal communication, people are needed to prioritize, fight for resources, and repeatedly focus on requirements with product managers. Well-written code needs people to verify, and written programs need to be tested. Acceptance requires operation and maintenance to deploy, etc.
3. Where is the road?
3.1 The trend is irreversible, master the tools as soon as possible to improve efficiency
Now that the era of AI has come, just like the steam engine and engine appeared in the industrial revolution, it will have a profound impact on many industries, and it has become an irreversible trend.
Many people will say that there are many shortcomings in the current large model, such as hallucinations and weak reasoning ability, but just like accidents may occur when driving a car, we did not give up driving because of this, because we can realize that it gives you The convenience brought. And it is precisely because of the imperfection of the large language model that we are not so easily eliminated. If one day the product manager designs the requirements document, AI can automatically write code, automatically fix problems, and automatically deploy, maybe programmers will really lose their jobs. If one day AI can produce better requirements documents than product managers, perhaps product managers will also be replaced.
Now many people around me have begun to use large language models to learn technology, write code, find ideas, and assist in solving problems. There are also many business teams that have begun to use large models to solve business problems. AI has greatly shortened the path to mastering knowledge. For some knowledge points that require reading many books and reading many columns to understand, let AI tools give you examples and explain it in easy-to-understand language, which can be quickly and systematically explained. master. Some seemingly uncomplicated codes may take half an hour to write by yourself, but AI may finish it in a minute or two, and the BUG is even better. AI can quickly provide many candidate solutions to help make more perfect solutions to many problems that have been contemplated and unable to find a solution. In the past, problems that could not be solved in ten minutes, hours or even days could be solved in a few minutes with the help of AI.
With the blessing of AI, people with very little programming experience can "write" better code than many junior engineers through prompt words. With the blessing of AI, people from some non-English-speaking countries can also write authentic English articles. In this era, AI brings infinite possibilities to many industries.
Therefore, in my opinion, "it's not that AI replaces humans, but that those who use AI eliminate those who can't use AI, and those who make good use of AI replace those who can't." If we are compared to the coachman, then A large model can be regarded as a car. We should not resist cars, but learn driving skills well, become an "veteran driver", use it to improve the efficiency of life and work, and use it to make money.
3.2 Be vigilant against "extreme things must be reversed"
Think Before Asking Big Models
Many people have this experience. Since graduating from school, they seldom write directly by hand. In life and work, they often use keyboard typing or even voice communication, and often "forget words when picking up a pen". This is the so-called "use it or lose it".
At the same time, many people have also observed that more and more people are now addicted to Douyin and Kuaishou, mainly because these platforms have powerful recommendation algorithms that analyze your preferences and push more content you like , Let you only indulge in it, and at the same time strengthen your "prejudice", it is worthy of vigilance.
If we do many things and cannot think first but rely too much on large language models, then our thinking ability will decline. However, the ability that humans have more advantages over AI is creativity, the ability to ask questions, and so on. If the ability to think is reduced, we can easily be swayed by AI.
Therefore, when you encounter a problem, it is recommended to think about how to deal with it yourself first. If you can’t solve it in a short time, try to ask AI again, and then ask other people if you can’t solve it. When you need to do some design, you must first think about the plan by yourself, and then let AI give you some suggestions to continuously enrich your plan, so as to form a virtuous circle.
Improve information identification ability
In the past, "When in doubt, ask Google", now it is "In case of doubt, ask AI". Since using some advanced AI tools, I seldom use search engines anymore. The previous search engines were more of a retrieval tool, giving you a bunch of relevant information for you to extract and analyze, which was less efficient. And now AI is more about actually answering your questions.
But now there are also "toxicity" and "illusion" problems in large language models. If one of the hundred sentences generated by the large model is mixed with a false sentence, if the large model generates content that looks very correct, if we lack professional knowledge, lack of awareness and ability to distinguish, it will easily have a negative impact on us.
Many people will say: "The AI era has come, and there is no need to study computer theory in depth." I think on the contrary, only by learning computer-related theories and technologies well can we improve the ability to identify information generated by AI and enjoy the benefits brought by AI. While improving performance, reduce its negative impact.
3.3 Seize the trend and think about how to better integrate with business
Just as the first part took the car as an example. Although the AI era has come, everyone now knows that AI is a trend, and they all appreciate the huge efficiency improvement of AI, but people's design concepts are still stuck in the previous generation, and they are mechanically applying AI to the original product system, just like It is the same as installing the engine directly on a trolley or directly on a bicycle.
When we think about how to combine large models with business, we can’t directly apply AI to business . We need to think about which tasks are suitable for using large language models, and which task engineering methods will be cheaper and more effective than large models better. Just like airplanes and high-speed rails are faster, but bicycles, cars, and subways are still widely used, different technologies use their respective fields of expertise, and complex tasks usually require a combination of multiple methods.
In addition, when combining the large model with the business, it is necessary to think about a more innovative way of integrating the model and the business. If the model is compared to an engine, when we are doing fusion, we should not think about "how to directly install the engine on the bicycle", but should reshape the frame, "invent the car", equip the car with "navigation", add "air conditioning ", upgrade from manual transmission to automatic transmission, and even add functions such as "assisted driving" and "automatic driving".
When combining the large-scale model with business, it is necessary to formulate the engineering standard for the large-scale model as soon as possible, continuously optimize the model and verify the effect, and develop the engineering link when it is close to this standard, which can save a lot of detours.
In addition, I found that there is a relative shortage of students who are engaged in large-scale model algorithms, and with the opening of APIs for large models, the complexity of the model training platform is getting lower and lower. Many back-end and even front-end development students have broken the functional boundaries and started to learn They have even started training and tuning large models to increase the speed of model iteration and better serve their own business. It is a good opportunity for development students, but in this process, it is recommended to have professional algorithm students to guide. Many algorithm problems and engineering problems are very different, and special learning is required. To give a simple example, in the process of some model training, some manually labeled samples are required, but the common sense knowledge in the eyes of professional algorithm students: the sample size should be large enough, the sample quality should be high, the samples should be more diverse, different types The sample should be as balanced as possible. For developers who are new to model training, they may not understand it, which leads to spending a lot of time creating a large number of manually labeled samples, but many of them are of low quality or lack of diversity, which is half the effort.
3.4 It is crucial to enhance the capabilities that are not easily replaced by AI
In the era of AI, many simple and clear tasks will be replaced by AI.
So, where do we go from here?
In my opinion, in this AI era, many relatively simple tasks will be replaced by AI. In the new era, we need to learn new abilities. To gain a competitive advantage, we need to persist in learning, even lifelong learning . Just like our parents need to "keep up with the times" and learn to use mobile payment and WeChat, in the era of AI, we must learn prompt words to better use large models to solve problems. In the future, it may not be necessary to write complex prompt words, but there will be many new technologies emerging, and we need to continue to learn.
When some abilities that once required a high cost of learning to master can be realized more efficiently through AI, low-end "CRUD programmers" will soon face elimination. Therefore, we need to especially strengthen those capabilities that are not easily replaced by AI , in order to be less likely to be eliminated and to gain a greater competitive advantage. Including but not limited to: prompt word skills, product thinking, creativity, general reasoning ability, ability to ask questions, communication ability, and ability to decompose complex problems, etc.
Many students may ask, how to master these abilities?
Some abilities need to be cultivated in our daily work, and we need to actively think and accumulate slowly, such as product thinking and the ability to ask questions.
Some abilities can be learned quickly through AI and then mastered through continuous practice, such as communication skills and the ability to decompose complex problems. In the age of AI, we must learn to use AI to learn. We can upload data to AI and learn directly through dialogue. If you don't understand the knowledge, you can also let AI explain it to us directly, and let it give us an example. For example, if you want to improve your communication skills at work, you can ask AI to introduce some common communication skills, and you can ask AI to explain to you what "non-violent communication" is; Express skills, let AI explain to you what "structured thinking" is, and "how to practice structured thinking", and "deliberately practice" at work.
For example, the following is the suggestion given by AI for the question "How to make your expression more structured?", which is very valuable for reference:
To make your expression more structured, you can follow the following methods and techniques:
1. Clarify the theme: Before expressing, first clarify the theme or central idea you want to convey. This helps ensure that your content stays on topic and avoids going off topic or dragging along.
2. Create an outline: Before you start writing or speaking, create a simple outline. List the main points and main ideas, and then organize your presentation according to this outline.
3. Clear paragraphs: Make sure that your expression is in paragraphs, and each paragraph should have a clear topic sentence and develop around the topic sentence.
4. Use conjunctions: Conjunctions can effectively connect your expressions together and make the content more coherent. For example, words like "first," "next," and "last" are great ways to guide the reader or listener through your train of thought.
5. Give examples: Whether you are writing or speaking, support your point by providing specific examples and cases. This can make your presentation more convincing and believable.
6. Avoid wordiness: Concise and clear expression is the key to structured expression. Avoid long sentences or repetitive words and keep the content concise.
7. Paragraph transitions: Ensure that there are clear transition sentences between paragraphs to help readers or listeners understand the connection between different paragraphs.
8. Pay attention to grammar and punctuation: Good grammar and punctuation are the foundation of structured expression. Make sure to use proper punctuation and grammar rules to avoid ambiguity and confusion.
9. Repeated revision: writing is a process of gradual improvement. After writing the first draft, revise and polish your presentation iteratively to ensure clarity of logic and order of information.
10. Practice: Structured presentation takes time and practice. Do more writing and speaking exercises, and gradually develop the habit of structured expression.
By adopting these methods and techniques, you can gradually improve your ability to express yourself, making it more structured, clear and organized.
It is suggested that when you want to develop a certain ability, you can directly ask AI to know what it is, why, how to do it, and use it flexibly in your work.
At this point, careful netizens may have discovered it a long time ago. This article is an article that follows the above-mentioned structural principles.
Four. Summary
The current AI is roughly the automatic transmission or assisted driving stage of the car, and it is more used as an efficiency-enhancing assistant.
Don't complain about the imperfection of AI. It is the imperfection of AI that makes us less likely to be eliminated.
When everyone uses large models, discover more application scenarios, master prompt word skills, and give full play to the value of large models. Today, as AI is becoming more and more popular, it is more valuable to train and strengthen people than AI. Only by continuous learning can we gain more advantages in this era.
How do you think we can seize this trend in the AI era? What other capabilities do we need to enhance?