From assisting to independently writing code, AI coding has evolved into an engineering-level "collaborative" coding tool.
Image source: Generated by Boundless AI
The AI created by programmers first replaced the programmers themselves.
"The coding ability of large models now matches that of senior programmers (with monthly salaries in the tens of thousands)." Ding Yu, head of Alibaba Cloud's native application platform and Tongyi Lingma, told Guangkui Intelligence.
In fact, AI coding tools are not a new phenomenon; they began to be applied during the last wave of artificial intelligence.
However, previously, "AI coding products were merely auxiliary tools, but now they can execute complex projects, edit long contextual texts, and independently perform simple coding tasks," said Zhang Tao, the technical head of the AI coding product family at SenseTime, to Guangkui Intelligence.
From assisting to independently writing code, AI coding has evolved into an engineering-level "collaborative" coding tool.
As a result, more and more companies are beginning to use AI coding tools to reduce costs and improve efficiency in software development. After 2025, AI may even replace mid-level programmers.
Meta founder Mark Zuckerberg recently stated, "By 2025, AI will reach the programming level of mid-level software engineers." Meta will begin automating the work of mid-level software engineers in 2025, ultimately outsourcing all programming tasks of its applications to AI.
This is not alarmist; the current penetration rate of AI-generated code in enterprises has reached an astonishing level.
For instance, over 25% of new code at Google is generated by artificial intelligence; the adoption rate of AI-generated code at iFlytek has risen from 30% in October 2023 to 52% in June 2024, with unit test coverage increasing from 30% to 50%.
The reason AI coding has become one of the hottest application areas for large models is that "AI Coding is the most frequent and certain scenario in the application of large models, a field validated by PMF (Product-Market Fit)," Ding Yu told Guangkui Intelligence.
Consequently, more and more companies are starting to invest in the AI coding space, with leading tech companies like Microsoft, Google, AWS, Alibaba, and Baidu taking the lead. However, the abundance of similar products has also led to homogenized competition; how to successfully break through in the future? How to achieve true large-scale commercialization?
From assisting to collaborative combat, AI has truly become your programmer colleague
In August 2024, Ricky Robinett, vice president of the well-known American company Cloudflare, shared that his 8-year-old daughter developed a chatbot in just 45 minutes, attracting the attention of 1.8 million online users.
The AI code editor she used, Cursor, also became an overnight sensation. This has once again made the AI coding space a focal point of industry attention.
Globally, according to PitchBook data, about 250 startups have launched AI coding assistants. In China, major internet companies like Alibaba, Baidu, Tencent, ByteDance, as well as unicorns like iFlytek and SenseTime, and even AI large model startups like Zhipu AI have all introduced related products.
AI coding products are emerging like mushrooms after rain, representing a substantial evolution in the capabilities of AI coding tools brought about by large models.
Early AI coding tools were mainly capable of executing simple tasks, such as automatically completing code based on programmers' comments and providing error prompts during the coding process.
With the upgrade of large model capabilities, the problems that AI coding tools can solve are increasing, such as being able to maintain and upgrade existing projects, "they can already autonomously accomplish certain R&D tasks," Ding Yu said.
For example, large language models can understand human instructions in natural language and automatically complete complex coding tasks based on engineering context, including modifying multiple front-end and back-end files simultaneously, executing scripts, writing tests, and deploying code.
"Initially, Tongyi Lingma appeared as a coding assistant, mainly providing support to programmers by automatically completing code based on the code context during development," Ding Yu said. "By the end of 2024, Tongyi Lingma will upgrade to the 2.0 AI programmer form, becoming a collaborative coding assistant that can work alongside human programmers, perceiving the entire project and making batch file modifications based on task scenarios, achieving a leap in capabilities."
While the AI coding tool assistant is upgraded to an AI programmer, the former's code generation is still primarily driven by humans, while the latter gradually shifts to being AI-driven, with humans mainly playing a monitoring and confirmation role.
"Previously, code was mainly written by humans, with AI assisting in simple, predictable, and repetitive tasks. Now, through requirement descriptions, AI can understand and help programmers complete some medium-difficulty coding tasks," Zhang Tao also stated.
Additionally, with the evolution of multimodal large models and deep reasoning large models, the capabilities of AI coding tools are continuously improving.
The "Office Raccoon" product from SenseTime's Raccoon family, besides performing data processing, data analysis, and document creation based on large models, can also support generating data images and PPT files, showcasing a comprehensive output of multimodal capabilities.
Multimodal input is equally important. "Many tool-type products find it difficult to accurately meet requirements if they only interact through language descriptions, as there is information loss in language expression when we describe content in text. At the same time, the current limitations in the semantic understanding capabilities of large models, including hallucination issues, also restrict the capability boundaries of AI coding tools. Directly inputting images or videos into large models can accomplish tasks more efficiently," Zhang Tao said.
Moreover, multimodal large models enable AI coding tools to achieve end-to-end full-stack functionality from text to image to code generation.
For instance, in website design, designers can create front-end visual drafts through text-to-image methods and then directly provide them to coding large models, which can translate the visual drafts into front-end interfaces and automatically generate back-end code based on the functionality of the front-end interface.
"Currently, AI coding can complete complex tasks, eliminating knowledge and skill asymmetries. For example, it can generate integrated solutions from front-end to back-end, breaking the previous separation of front-end and back-end personnel and capabilities, significantly improving efficiency," Ding Yu said. "Moreover, after generation, AI coding can also help programmers automatically generate tests, ultimately returning the modified results of the tests."
However, while AI can autonomously generate some code, in practice, the code generated by AI cannot run perfectly on the first try and often contains numerous bugs.
A PhD student in AI at Zhejiang University, Chen Rong (pseudonym), told Guangkui Intelligence: "More complex code will have bugs; it's generally hard to get it right on the first attempt. Technically, it can be understood that the model treats coding as a translation task, outputting a sequence of code that may not adequately consider the code's execution environment."
The reasons behind this are twofold: on one hand, most humans find it difficult to accurately describe their actual needs; even many experienced programmers need to revise their code repeatedly during the coding process.
On the other hand, the current limitations in the semantic understanding capabilities of large models, including existing hallucination issues, also restrict the capability boundaries of AI coding tools. Therefore, although "within the context window allowed by the model, large models can understand tens of thousands of lines of code, the boundaries of AI coding capabilities are still difficult to define," Zhang Tao said.
Just as human programmers need to repeatedly modify test code, the process of generating code with AI can also reduce the presence of bugs through multiple rounds of interaction.
Ding Yu stated: "AI coding does not generate final results in one go; it involves multiple rounds of interaction and iteration with the large model. During the collaborative coding process with the large model, there is a continuous process of thinking and reasoning exploration. After multiple rounds of interaction and correcting results, it can autonomously conduct testing and validate the code, completing tasks throughout its lifecycle."
Although current AI coding tool products still have some issues, more and more companies are beginning to adopt AI coding tools. The "cost-effective" AI coding tools not only improve programmers' coding efficiency but also achieve cost reduction and efficiency enhancement for enterprises.
In large projects, "screws," AI improves programmer efficiency by over 10%
The evolution brought by large models to AI coding tools has lowered the threshold for programming.
Currently, the scenarios where AI can independently achieve autonomous programming mainly fall into three categories:
One category is small products, such as personal life assistant apps;
Another category is content-oriented websites, where the code volume and difficulty are moderate, allowing AI to implement them independently;
The third category is office products, such as Excel spreadsheet editing and data aggregation.
From practical applications, the overall code volume in these scenarios is not high, and the actual development difficulty is not great, requiring minimal programming knowledge from developers.
It can be said that AI coding tools have indeed lowered the threshold for programming, allowing more people without coding skills to engage with code programming and autonomously develop some product features.
However, despite the fact that AI coding tools have lowered the threshold for programming, programmers still need to enhance their upper limits of programming ability, especially in more complex software development and large enterprise-level system software development.
A programmer in the fintech industry, Xiao Xiao (pseudonym), told Guangkui Intelligence: "For an engineering project of a company, it's still very difficult to hand everything over to AI. Engineering projects require many processes and collaboration across multiple departments, and AI cannot see the big picture."
It is evident that in enterprises, large models are still primarily handling the dirty and tedious work, while global and innovative tasks still require human programmers.
"The work of programmers is not just about generating a small project; they face production code where the entire project file context is very complex, and the relationships between codes are also intricate. Programmers also have their own quality requirements for the code," Zhang Tao said.
This means that for programmers in enterprises, AI coding tools are still more of an auxiliary role, but they indirectly raise the lower limit of programmers' work capabilities, as AI can already handle simple repetitive tasks.
"If you let AI directly generate 100,000 code files for all the business of a bank, it definitely cannot do that at present," Ding Yu admitted. "Currently, in large enterprise projects, AI coding definitely starts with small tasks, finding a slice, such as implementing a functional module or identifying security vulnerabilities in a million-line codebase, where AI can perform very accurately and quickly."
Moreover, it is well known in the industry that for large enterprise projects, the most feared issue is system uncertainty. If a system bug occurs, it can lead to significant resource and economic losses.
Therefore, in Ding Yu's view: "Large projects still require human programmers to manage the uncertainties in the software development process, such as architecture design and domain modeling. They need to break down the already certain content, such as module development, finding security vulnerabilities, and supplementing test cases, and hand it over to AI to perform these deterministic tasks based on human instructions."
Although AI coding tools are merely assisting, they have brought tangible efficiency improvements to developers and enterprises.
Taking Alibaba Cloud as an example, currently, all technical staff are using Tongyi Lingma, with a monthly active user ratio exceeding 82%, and AI-generated code accounts for over 30% of the total submitted code daily. Based on this data, it can be roughly calculated that AI improves developer efficiency by about 17.5%, and even with a discount, it will be between 10% and 15%.
"Therefore, every time I meet with enterprise leaders, I tell them that Tongyi Lingma can improve the efficiency of engineering teams by over 10%," Ding Yu said. "This means that if a company has 100 engineers using Tongyi Lingma, it can produce the equivalent output of an additional 10 engineers."
Additionally, human programmers are categorized into specific roles, such as front-end and back-end. If a back-end engineer is to take on front-end tasks, they may require extensive training, as they cannot immediately assume the responsibilities of a front-end programmer.
However, with AI coding tools, programmers only need to ask AI to easily learn development knowledge across various language platforms and quickly get started. "In the past, completing a project might take two to three weeks of preliminary research; now it can be done in two to three days, allowing employees to achieve a 1-N capability increase," Ding Yu said.
Of course, AI can also help human programmers with more repetitive tasks. For instance, many developers are reluctant to write test code, which they view as uncreative work, yet it is necessary.
AI coding tools can automatically generate unit tests based on the programmer's code as prompts, truly liberating developers and allowing them to focus their energy on more creative tasks.
Furthermore, for enterprises, beyond the explicit value enhancement, the implicit value lies in the fact that AI coding tools can help maintain high quality and long-term stability of software systems. They can not only complete unit tests but also autonomously discover security vulnerabilities and provide repair suggestions, improving quality while shortening project delivery cycles.
Interestingly, at this stage, AI's coding capabilities, when aided by external tools, have gradually surpassed those of mid-level programmers. One of the characteristics of SenseTime's Raccoon family of models is the enhancement of code interpreter capabilities, allowing the model to achieve autonomous code debugging and iteration.
"In complex projects, relying solely on large model inference to generate code has a low first-pass success rate, generally not exceeding 20%," Zhang Tao said. "However, the Office Raccoon, based on the code interpreter solution, has a code pass rate close to 80% in daily tasks like charting."
The AI coding track begins to differentiate; innovative scene refinement determines success or failure
AI coding has already become a validated direction through PMF, leading many players to enter this track and resulting in numerous homogeneous products.
Currently, in the Chinese market, many enterprises, including major internet companies, small and medium-sized enterprises, and large model startups, have launched AI coding products, such as Alibaba Cloud's Tongyi Lingma, Baidu's Wenxin Kuai Ma, ByteDance's Doubao MarsCode, Tencent Cloud's AI coding assistant, and Zhipu AI's CodeGeeX, among others.
Despite the abundance of AI coding products, the differences in the functionalities they offer are not significant. "Currently, the market is quite homogeneous; the functionalities are actually similar, as programming products aim to solve the same user problems," Zhang Tao said.
However, with the iterative upgrade of large model technology, the AI coding track has entered a mid-stage of "differentiation." "From the current AI coding track, different implementation methods have begun to emerge," Zhang Tao said.
Products like Cursor can perform complete task programming based on their modified open-source IDE; there are also products like Bolt.new that operate as online tools where users describe their needs, and AI completes web development, but it can only handle front-end technology stack-related content.
At this stage, it is evident that various products have begun to identify different niche scenarios and build their product advantages, achieving differentiated development—some are better at web development, while others excel at modifying existing project code, and some can develop small tools or engage in low-code work.
Ding Yu also believes: "Software development encompasses many scenarios and numerous niche areas. Enterprises can enter from different angles to innovate in niche scenarios or product forms."
The segmentation of functional scenarios among various AI coding tool products will also bring commercial differences to each product, as the commercialization focus of different enterprises is not entirely the same.
For example, in SenseTime's Raccoon family, the Office Raccoon product primarily focuses on the office tool sector, and in actual commercialization, it operates simultaneously on both the C-end and B-end.
The C-end mainly relies on paid subscriptions, while the B-end focuses on private deployment for enterprises. "Currently, there are nearly 40 clients for private deployment, including large internet companies."
However, Zhang Tao is also optimistic about the market potential of the C-end track, noting that the promotion of C-end products is exceeding expectations.
From functional scenarios to commercialization directions, the AI coding track has already begun to show differentiation, but this is not the final form of development in the AI coding industry.
With the continuous iteration of large model technology capabilities, the next step for AI coding will be to achieve "autonomous programming," meaning it will not only assist programmers in project development but will also be able to independently accept requirements and complete entire project tasks.
"The future will definitely move towards AI autonomous programming, which will bring a tenfold increase in IT productivity for enterprises and developers," Ding Yu said.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
