Big manufacturers are betting on AI code, but they are still far from replacing programmers

Can AI replace programmers? The industry may not have reached a conclusion on this hot topic, but first-line manufacturers have begun to put it into practice, and code assistants have become one of the key scenarios that everyone is vying to land.

At the beginning of April, Alibaba announced the first AI employee to take up the post, fully implement AI programming internally, and use Tongyi Lingcode to assist programmers in writing code. A little earlier, in March, Baidu released the Comate 2.0 code assistant, which is free for individual developers. In December last year, SenseTime launched the intelligent programming assistant code Little Raccoon, which can help developers improve programming efficiency by more than 50%, and on April 23, SenseTime announced the launch of a code large model all-in-one machine at the technical exchange day, with a light version priced at 350,000 yuan each, and a single unit supports a team of 100 people.

The layout of many large factories makes programmers seem to be not far from being replaced, but there is no need to worry about it in the short term. Devin, who was previously considered the world's first AI programmer, was recently accused of fraud and "self-directed" in a demo video. Some industry insiders said that AI may be able to help write some basic work such as test scripts, but if it is placed in a more serious commercial-grade code development process, even GPT-4 can only achieve a lower level.

Jia Anya, head of Copilot product at SenseTime, has also been thinking about the endgame of future program development, and she believes that AI code will still be a tool for efficiency, and its core is to make it easier and more convenient for programmers to work. In the future, the work content of programmers will definitely change, but there must also be someone to use good tools, "just like after the steam engine, someone may change from a coachman to a driver, with the mobile Internet, some stores have moved from offline to online, but the essence of many things has not changed." "In the future, programmers will still be the main body and core of the development process, but programmers may be redefined.

Big players are betting on AI code

On April 2, Alibaba Cloud announced that it is fully implementing AI programming internally, using Tongyi Lingcode to assist programmers in writing code, reading code, checking bugs, and optimizing code. Alibaba Cloud also assigned a formal employee ID number to Tongyi Lingcode - AI001.

According to Alibaba Cloud sources, 20% of the company's code will be written by Tongyi Lingcode, but programmers are still the core of R&D, and they will have more time to focus on system design and core business development.

According to reports, within Alibaba Cloud, Tongyi Lingcode has played the role of code assistant in various development links. Taking API development and testing as an example, Tongyi Lingcode can shorten the time spent on dozens of minutes of manual writing and testing to seconds, saving programmers more than 70% of their test code workload.

Baidu has also implemented AI code internally, and in March Baidu issued a document saying that the code assistant Comate has written a quarter of Baidu's internal code, and outside of Baidu, Comate has joined more than 10,000 companies such as Himalaya, iSoftStone, and Shanghai Mitsubishi Elevator, with an enterprise code adoption rate of more than 50%.

Robin Li, the founder, chairman and CEO of Baidu, once said that one of the things that he wants to promote most in 2024 is to equip everyone with the ability to be a programmer. He also said that there will only be two programming languages left in the future, one is called English and the other is called Chinese, "There will be no such profession as a programmer in the future, because as long as you can speak, everyone will have the ability to be a programmer." ”

At the opening ceremony of the GDC conference at the end of March, Xu Li, chairman and CEO of SenseTime, mentioned that its CodeLittle Raccoon product can reduce the time required for the whole process of software development and help developers improve programming efficiency by more than 50% after focusing on some repetitive labor. Taking the birth of Code Little Raccoon as an example, he said that if it takes 100 man-days from demand analysis to final product development, it generally takes 100 man-days (note: man-days are a unit to measure human resource consumption, and the number of people and days are multiplied), and last year, Code Little Raccoon has been able to save 30% of the workload, to 70 man-days.

As the team leader of Codecoon products, Jia Anya introduced that the whole life cycle of software development includes from the design stage to the architecture, and then to the stages of development, testing, deployment, and maintenance. At present, in the development and testing phases, the efficiency of the code assistant is the most obvious, because there is a lot of boring and repetitive work in the development and testing process.

"For example, to understand the context and do some code completion, or to do some unit tests after the code is written, it is also a liberation for programmers to focus more on creative work, such as architecture design. Jaanya said.

In the beginning, the code raccoon can do relatively repetitive work, as the model's reasoning ability increases, Jia Anya said that now it can also do some creative work, such as helping to do requirements document writing, architecture design, and even some deployment plans for some specific scenarios of different customers, when the entire software development cycle can be greatly shortened.

In terms of AI code, SenseTime has released a more complete product. On April 23, at the 2024 SenseTime Technology Exchange Day, SenseTime released the "RiRixin 5.0" large model, and at the same time launched enterprise-level large model all-in-one machines for four industries, including code. According to reports, the light version of the Little Raccoon code large model all-in-one can help developers write, understand and maintain code more efficiently, and compared with the traditional cloud service model, all data processing processes of the all-in-one machine are completed in a private environment, which can avoid the risk of data leakage in the transmission process. SenseTime said that the test pass rate of the Little Raccoon Code Large Model All-in-One Machine in HumanEval reached 78.1%, surpassing the 74.4% of GPT-4.

2024 is considered to be the year of the explosion of large model applications, in many scenarios, why has code become the key layout of various manufacturers? In the exchange, Jia Anya believes that the intelligent ability of large models has been further enhanced in the past year, and the reasoning ability, code ability, and scientific thinking are also the key breakthroughs of SenseTime in large models in the past year. These capabilities are an important foundation for the implementation of code assistants.

On the other hand, it is combined with real scenes. Jia Anya mentioned that it is difficult to ensure 100% accuracy of the current large model, but in the case of code, the code written by the programmer itself needs to be reviewed again, so even if the large model is still uncertain, AI can still empower programmers to improve efficiency.

"Combined with some of the accumulation of SenseTime itself, some progress of SenseTime's large model, and the feedback needs of users, Office Little Raccoon and Code Little Raccoon are the directions we have selected and focused on at present. Jia Anya said.

Redefining the programmer

"Many people say, your AI (code) has come out, will you replace the programmer?" When it comes to the theory of AI programmer substitution, Zhang Liaoyuan, the product leader of Tongyi Lingcode, first reacted, "No, the programmer is still the core, the person is always the subject, and AI is the object and auxiliary." ”

Alibaba Cloud previously mentioned that 20% of the company's code in the future can be generated by AI, on this basis, is it possible to increase the upper limit of AI writing code to 80%? Zhang Liaoyuan believes that it cannot be done in the short term, and people are still needed to design or provide ideas.

"The process of writing code is also the process of thinking and designing, and when writing a framework, the corresponding design is being done in the mind, but after the framework is written, there are some very simple and clear tasks, and AI can help us complete them independently, but in the process of software research and development, especially when encountering software research and development work in production, it is very large-scale, and it will involve more content that has to rely on the human brain to think. Zhang Liaoyuan said.

Zhang Liaoyuan believes that in the programming stage, it is difficult for AI to replace people. "It is still necessary for people to express their intentions and distribute work, and whether these jobs are done well or not, whether they are right or not, also need people to judge. ”

Professor Lin Dahua, a leading scientist at the Shanghai Artificial Intelligence Laboratory, previously told Yicai that AI as a code assistant has been verified and can indeed bring productivity improvements." For example, if I write a function (code), I have never written this function myself, and I usually go to Google or Baidu to check how others write it, and learn from it to change it, and now the code assistant is also this logic, because it has learned hundreds of millions of code bases, so write function-level code to fill in the blanks, and even help write some test scripts and other elementary work, which can really help programmers save a lot of time. ”

However, Lin Dahua also mentioned that if the AI code is put in a more serious commercial-level code development process, it will be found that even GPT-4 can only achieve a relatively low level, and even a 10% success rate in some scenarios cannot be achieved.

In a sense, AI code is still a language ability, "that is, the content that has been seen before can be changed and reassigned according to a certain scenario, but it obviously does not have the ability to think logically at a very deep level, which is actually the most important thing for us to build a real software." Lin Dahua believes that there is still a certain distance between replacing programmers and large models.

Previously, a programmer from a large factory also had the same feeling, and he mentioned to Yicai that although large models are helpful for AI programming, they are still a long way from solving complex problems. "It's like building a house, it's not just about holding a hammer in it, you have to think about how to build it, then you have to think about what to build in the first step, what to build in the second step, how to build it so that it won't collapse, and how long you want to build it, and then you have to knock it step by step. "These are the parts that programmers need to do.

In early March, the startup Cognition released its first AI software engineer, Devin, who was known as "the world's first AI programmer", although it has not yet been publicly tested, but according to the official report, it can handle the entire development project end-to-end with just one command, and the video shows that it has the ability to learn new technologies independently, build and deploy applications end-to-end, and find and fix code problems independently.

On April 9, a web blogger who claimed to have 35 years of experience as a software engineer reproduced Devin's demo video frame by frame and raised four questions, including that Devin's demonstrated programming ability was deceptive, "the tasks it handles are not random, but deliberately selected by the presenter", and that Devin seems to have fixed many problems during the operation, but many of these problems are "self-directed" by Devin. This also proves that it is difficult for AI to deal with complex problems and replace programmers.

Jia Anya also said in the interview that it is still difficult to rely on AI to write the code of the official business from end-to-end. "So when we go to the subsequent release of the product, it will actually be a logic of interaction between humans and tools, which is actually a semi-human, half-model state to do an interaction and improve end-to-end efficiency. ”

As for the endgame of the future, Jia Anya believes that a programming language framework based on natural language may be derived. The programmer's development language itself is constantly iterative, and it is presented in a more friendly way for humans, such as Java and Python are actually high-level programming languages abstracted from low-level programming languages, however, if you use human natural language to program in the future, one problem is that machine language is not very ambiguous, it is a very rigorous language, but the characteristics of natural language lie in its complexity, there is a lot of information implied in grammar, context, voice, and intonation, and the expression is flexible and changeable, and the semantics presented are vague. Program execution, on the other hand, requires accurate and stable input and output. Therefore, the natural language that we are most accustomed to must not be the next-generation programming language that can directly cause broad consensus, and may derive new language frameworks on top of it.

At the end of the day, Janya believes that AI will be an efficiency tool, and at its core, it will make it easier and more convenient for us to work. In the future, the work content of programmers will definitely change, and someone will need to use these tools well in the future, "Just like after the steam engine, someone may change from a coachman to a driver, with the mobile Internet, some stores have moved from offline to online, but the essence of many things has not changed, just how we use it well." ”

"Most of the positions of product managers, data analysts, and software development engineers are the products of the previous batch of technological changes. Jia Anya believes that in the next era of scientific and technological change, some specific positions and content may have some changes, but the core of the world has not changed.

AI may not replace programmers, but it may redefine programmers. In the future, the requirement for programmers may not be to type code quickly.

(This article is from Yicai)