Image source: AI generated by unbounded AI

Programmers are the first to replace them.

"The coding ability of big models is now at the level of a high-level programmer (monthly salary of several tens of thousands of yuan)." Ding Yu, head of Alibaba Cloud native application platform and head of Tongyi Lingcode, said Cizhi.

In fact, AI code tools are not new things and have been implemented in the last wave of artificial intelligence.

But before, "AI code products were originally just auxiliary tools, but now they can execute complex projects, long context text editing, and independently do simple code tasks." The Little Raccoon Family, an AI code product under SenseTime Technology Zhang Tao, the technical director of the company, said to Guangma Intelligent.

From assisted to independent code writing, AI code has evolved into an engineering-level "coordinated" coding tool.

Based on this, not only are more and more companies starting to reduce costs and increase efficiency in program development through AI code tools, but after 2025, AI may even replace intermediate programmers.

Meta founder Zuckerberg recently said: "In 2025, AI will reach the programming level of intermediate software engineers." Meta will start to automate work of intermediate software engineers in 2025 and will eventually apply it. All programming work for the program is outsourced to AI.

This is not an alarmist statement. The current penetration rate of AI-generated code in enterprises has reached an astonishing level.

For example, more than 25% of Google's new code is generated by artificial intelligence; iFLYTEK's internal AI-generated code adoption rate has increased from 30% in October 2023 to 52% in June 2024 , unit test row coverage is increased from 30% to 50%.

The reason why the AI ​​coding track has become one of the most popular tracks for big model applications is that "AI Coding (Artificial Intelligence Programming) is the most frequent and most certain in the implementation of big model applications. The scenario is a field that has been verified by PMF (product market matching degree). "Ding Yu said to Guangma Intelligence.

As a result, more and more companies are beginning to plan the AI ​​coding track, and leading technology companies such as Microsoft, Google, AWS, Alibaba, Baidu, etc. are at the forefront. But so many similar products have also caused homogeneous competition. How can we successfully break through in the future? How to achieve real large-scale commercialization?

From assisting to collaborative combat, AI has really become your programmer colleague

In August 2024, the daughter of Ricky Robinett, vice president of Cloudflare, a well-known American company, is an 8-year-old girl, which takes 45 minutes. A chat robot was developed, which attracted online attention of 1.8 million netizens.

The AI ​​code editor Cursor, she used, also became popular overnight. This also makes the AI ​​coding track once again the focus of industry attentionpoint.

Global, according to PitchBook data, about 250 startups have launched AI coding assistants. In China, Internet giants including Alibaba, Baidu, Tencent, ByteDance, unicorn companies such as iFLYTEK and SenseTime, and even AI big model startup Zhipu AI have launched related products.

AI code products have sprung up like mushrooms after a rain, and it is the substantial evolution of the capabilities of AI code tools that have been brought to the capabilities of large models.

Early AI code tools were able to perform simple tasks, such as automatically completing the code based on programmers' comments, and providing code error prompts during programmers' writing code.

With the upgrade of large-scale model capabilities, AI code tools can solve more and more problems, such as being able to maintain and upgrade based on existing projects, "certain R&D tasks have been achieved independently." Ding Yu said.

For example, a large language model can understand human instructions in natural language and automatically complete complex coding tasks based on the engineering context, including modifying multiple front and back-end files at the same time, executing scripts, writing tests, deploying code, etc.

"The first Tongyi Ling Code appeared in the form of a coding assistant, mainly to provide assistance to programmers and help programmers automatically complete the code during research and development according to the code context." Ding Yu said, "At the end of 2024, Tongyi Lingcode upgraded to 2.0's AI programmer form, becoming a collaborative coding assistant, able to work together with human programmers, perceive the entire project, make batch file modifications based on scenario tasks, and achieve a leap in capabilities." p>

As for upgrading from AI code tool assistant to AI programmer, the former's main force in code generation is still human, while the latter gradually shifts to AI as the main role in monitoring and confirmation.

"In the past, people wrote code mainly, and AI assisted in doing some simple, predictable and repetitive work. Now, we can use the requirement description to let AI understand and help programmers. Complete some medium-difficulty code development work." Zhang Tao also said.

In addition, with the evolution of multimodal large models and deep inference large models, the capabilities of AI code tools are also constantly improving.

The "Office Raccoon" product of Shang Tang Xiao Raccoon family can not only perform large-scale data processing, data analysis and document creation, but also support the generation of data pictures and PPT files. This is a A comprehensive reflection of multimodal capability output.

Multimodal input is equally important. "Many tool products, if they interact only through language description, it is difficult to accurately realize the requirements, because when we describe the content as text, there is a linguistic expression. Information loss. At the same time, the lack of semantic understanding capabilities and hallucinations that the big model currently exists also limits the capability boundaries of AI code tools. If you directly input it into the big model in visual ways such as images or videos, you can more efficiently Complete the task." Zhang Tao saidroad.

At the same time, the multimodal big model allows AI code tools to realize the end-to-end full-stack function from literary and creative graphics to generated code.

Taking website design as an example, designers can design front-end visual drafts through literary and artistic drawings, and then directly give them to the Coding model, translate the visual drafts into front-end interface, and then use the front-end interface to have large functions. The model automatically generates back-end code.

"At present, AI encoding can complete complex tasks, eliminating the asymmetry of knowledge and skills, such as the integration of the front-end to the back-end, breaking the previous separation of front-end, back-end personnel and capabilities , greatly improve the efficiency. "Ding Yu said, "And after generation, AI encoding can also help programmers automatically generate tests and finally return the results of the test modified."

However, although AI can already be independent Generate some code, but in the actual process, the code generated by AI cannot be run at once, and there are also many bugs.

A doctoral student in Zhejiang University's AI direction Chen Rong (pseudonym), told Guangma Intelligence: "There are bugs in complex codes, and it is basically difficult to pass them all over again. From a technical logic point of view, it can Understand that the model actually treats coding as a translation task, and outputs a sequence of code, and may not consider the code running environment, etc. "

There are two main reasons behind this, one In terms of this, it is difficult for most humans to accurately describe their actual needs, and even many senior programmers need to repeatedly modify them during the process of writing code.

On the other hand, it lies in the lack of semantic ability of the big model itself in understanding semantics, including the existing hallucinations, which also limit the capability boundaries of AI code tools. Therefore, although "large models can reach the understanding of ten thousand lines of code within the scope allowed by the model context window, the capability boundaries of AI code are still difficult to define." Zhang Tao said.

Just like human programmers need to repeatedly modify test code, in the process of AI generating code, they can also reduce the existence of code bugs by interacting with it multiple rounds.

Ding Yu said: "AI coding is not to generate the final result at one time, but to complete multiple rounds of interactive iterations with the big model. In the process of coding with the big model, there is continuous thinking and reasoning exploration. Process, after the results of multiple rounds of interactive modification are correct, you can also conduct independent testing and verification, and deploy and use the code to complete the task throughout the life cycle. "

Although there are still some problems with the current AI code tool product, However, more and more companies are beginning to introduce AI code tools. The "cheap job" AI code tools not only improve programmers' programming efficiency, but also achieve cost reduction and efficiency improvement for enterprises.

The "screws" in large-scale projects, AI improves programmers by more than 10%

The evolution of big models to AI code tools , making the programming threshold lower.

At present, there are three main scenarios where AI can independently implement independent programming:

One is small products, such as APP assistants for personal life;

One is a category of small products;

One category of small products It is a content-based website with moderate code volume and difficulty, and AI can realize it independently;

One category is office products, such as Excel table editing, data summary, etc.

From the practical application point of view, the overall code volume of these scenarios is not high, and the actual development is not very difficult, and the requirements for developers' programming knowledge are not high.

It can be said that AI code tools have indeed lowered the threshold for programming, allowing more people with no code capabilities to access code programming, and can independently develop some product functions.

However, although AI code tools lower the threshold for programming, programmers need to improve the upper limit of their programming capabilities, especially in more complex software development and large enterprise-level system software development.

Xiao Xiao (pseudonym), a programmer in the financial technology industry, told Guangma Intelligence: "For a company's engineering projects, it is still difficult to hand them over to AI directly. The process required by the engineering projects There are many, and multiple departments need to cooperate, and AI cannot see the overall situation. "

It can be clearly seen that in enterprises, big models do more dirty work, global and innovative The work still needs to be done by human programmers.

"The work of a programmer is not just about generating a small project. The production code it faces, the entire project file context is very complex, the code relationship is also complicated, and the programmer also has his own code quality. Requirements.” Zhang Tao said.

This means that for programmers in enterprises, AI code tools are more of auxiliary roles, but they also indirectly raise the lower limit of programmers' work ability. After all, it is simple and repetitive. AI can basically do the job.

"If AI is allowed to directly generate 100,000 code files for all businesses of a bank, it will definitely not be able to do it at the moment." Ding Yu admitted, "At present, in large-scale enterprise projects, AI encoding is definitely Starting from a small task, finding a section, such as implementing a functional module, or finding security vulnerabilities in a million engineering code, AI can do it very accurately and quickly. "

In addition, it is well known in the industry that for For large enterprise projects, the most feared problem is the uncertainty of the system. If a system bug occurs, it may bring huge resources and economic losses.

Therefore, in Ding Yu's view, "large-scale projects still require human programmers to grasp the uncertainties in the software development process, such as architecture design, domain modeling, etc., and to disassemble the content that has been deterministic. Untangle, such as module development, finding security vulnerabilities, supplementing test cases, etc., and delivering them to AI to do these deterministic work according to human instructions. "

Although it is just a helper, AI code However, tools have also brought real efficiency improvements to developers and enterprises.

Taking Alibaba Cloud as an example, all technical staff currently use Tongyi Ling Code, with monthly active users accounting for more than 82%, and the code generated by AI every day accounts for more than 30% of the total number of submitted codes. Based on this data, it can be roughly calculated that AI improves developers' efficiency by about 17.5%, and a discount will be between 10%-15%.

"So, every time I meet the person in charge of the company, I will say that Tongyi Lingcode can improve the efficiency of the engineer team by more than 10%. "Ding Yu said, "That is, if a company has 100 Engineers can produce an additional 10 engineers when using Tongyi Ling Code. "

In addition, human programmers have sub-categories, such as front-end, back-end, etc. If you want To let a backend do the front-end, it may require a lot of training and learning for the back-end engineers, and they cannot immediately take over the work of front-end programmers.

But with AI code tools, programmers only need to ask AI to easily learn the R&D knowledge of various language platforms and get started quickly. "In the past, it might take two or three weeks to prepare a project, but now it can complete the task in two or three days, allowing employees to achieve 1-N capacity growth." Ding Yu said.

Of course, for AI, it can also help human programmers do more repetitive work. For example, many developers are unwilling to write test code. These are not creative from the perspective of programmers. Work, but have to do it.

As AI code tools can automatically generate unit tests based on programmer's code as prompt words, truly liberating developers and allowing developers to spend their energy on more creative work.

In addition, for enterprises, in addition to the explicit value improvement, the implicit value exists lies in that AI code tools can make it easier for enterprises to maintain the high quality and long-term stability of the software system, and they can not only do it Completion of unit tests can also independently discover security vulnerabilities and provide repair suggestions. While improving quality, it can also shorten the project delivery cycle.

What is more interesting is that the coding capabilities of AI at this stage, with the help of external tools, have gradually surpassed intermediate programmers. One of the characteristics of SenseTime's underlying model is the ability of code interpreter This has been strengthened to enable the model to realize independent code debugging and iteration.

"In complex projects, relying solely on big model inference to generate code, the one-time pass rate is not high, generally no more than 20%. "Zhang Tao said, "And the office raccoon is based on the code interpreter solution, In terms of daily charts and other capabilities, the code pass rate has reached nearly 80%. ”

The AI ​​coding track has begun to differentiate, and innovation in refined scenarios determines success or failure< p>AI encoding has become a landing direction that has passed PMF verification, which has also led many players to enter this track and emerged with many homogeneous products.

At present, many companies in the Chinese market, including Internet manufacturers, small and medium-sized enterprises, and large-scale model startups have been launched.AI code products, such as Alibaba Cloud's Tongyi Ling Code, Baidu's Wenxin Kuai Code, ByteDance's Doubao MarsCode, Tencent Cloud AI Code Assistant, Zhipu AI's CodeGeeX, etc.

Although there are many AI code products, the functions provided by each company are not very different. "The current homogeneity in the market is quite serious, and the functions are actually similar. After all, programming products hope to solve users. The problem is the same." Zhang Tao said.

However, with the iteration and upgrading of big model technology, the AI ​​coding track has also entered the mid-stage stage of "differentiation". "From the current AI code track, different implementation methods have begun to be differentiated." Zhang Tao said.

Products like Cursor can do complete task programming based on their own modified open source IDE; there are also products like Bolt.new that are used in the form of online tools, users describe their needs, and AI completes them. Web page development, but it can only implement front-end technology stack related content, etc.

At this stage, it can be clearly seen that each product has begun to find different segmented scenarios and build its own product advantages to achieve differentiated development - some are better at web development, while others are better at Do some code modification tasks for existing projects, and some can also do some gadget development or low-code work.

Ding Yu also believes that: "There are many scenarios for software research and development, and there are many sub-sectors. Enterprises can enter from different entry points and innovate in segmented scenarios or product forms." p>

The segmentation of each AI code tool product in functional scenarios will also bring commercial differences to each product, and the commercialization focus of different companies is not exactly the same.

For example, the office raccoon products in SenseTime's little raccoon family mainly focus on the office tool track. In actual commercialization, the C-end and B-end are carried out simultaneously.

The C-end mainly focuses on paid subscriptions, and the B-end mainly uses enterprise privatization deployment. "Currently, there are nearly 40 privatization deployment customers, including relatively large Internet manufacturers, etc.."< /p>

However, Zhang Tao is also optimistic about the market potential of the C-end track, and the promotion of C-end products at this stage exceeds expectations.

From scene functions to commercialization, the AI ​​coding track has begun to differentiate, but this is not the final form of the development of the AI ​​code industry.

With the continuous iteration of large-scale model technology capabilities, the next step of AI code will realize "autonomous programming", that is, not only assist programmers in developing projects, but also be able to independently accept independent needs and complete complete projects. Task.

"In the future, it will definitely move towards AI independent programming, which also means it will bring 10 times the IT productivity improvement to enterprises and developers." Ding Yu said.