2024 is about to end. In this year, the development of AI has gone as expected. It can only be described as rapid progress. If we make a time schedule from January to December, each month is a lively progress of AI large models and the release of new AI products; of course, there are also various gossip in the AI circle. In the whole year of AI progress, originally speaking, the most eye-catching is the breakthrough progress of AI video generation. Since the beginning of the year, when OpenAI released the conceptual video of Sora, most people concerned about AI suddenly realized that in addition to the animation video generation derived from AI painting models with unsatisfactory effects, there is also a native way of AI video generation, that is, directly training video fragment data through the transformer model. The height that this method can reach has been seen by us now. It is the AI video generation effect that everyone dared not think about at the end of 2023. In 2024, we are used to it. Video content includes sound, pictures, three-dimensional spatial dynamics, and language and text information expressed in speech. In addition to interactive games, the AI generation of video content is the top of the content production pyramid. Since the birth of generative AI, from text content to music and then to the AI production of video content, we can see that the information content that humans need to consume is completely covered by AI production. Considering the explosive growth of video AI generation in 2024, it is very likely that the best Oscar award of AI this year will be it. The emerging new generation of AI programming tools.However! During the period from Q4 to the end of the year, another new form of AI-generated product suddenly rose. In the face of everyone’s unpreparedness, it quickly established its own large-scale user group and business territory. This is the most popular “AI programming” at present. Among them, the two most representative products are Cursor and Windsurf. How popular is it? According to GitHub’s statistics, the usage of AI programming tools in 2024 increased by 300% year-on-year. In Stack Overflow’s annual survey, more than 60% of developers said they were using or planning to use AI programming tools. This figure was only 15% in 2023. What is even more surprising is that in the Q4 quarter, the number of open source projects related to AI programming surged. The number of new AI programming-related repositories on GitHub reached a historical high, with an average of more than 100 added every day. These data all show that AI programming is reaching a new level of popularity. AI programming is not actually a new concept. As early as when the OpenAI GPT3-level model introduced code training, the GPT model had the preliminary ability to read code and became an auxiliary tool for human coding. Before the outbreak of Cursor and WindSurf, the representative product of AI-assisted programming was Microsoft CEO Satya Nadella’s always-mentioned CoPilot. Microsoft’s entire suite of CoPilot was derived from the popular Github CoPilot concept. In the era when OpenAI was alone in the market, Github became popular among programmers with Copilot-assisted programming.But compared with the new generation of AI programming tools represented by Copilot and Cursor/WinSurf, the gap is huge and obvious.
The most crucial point is that Copilot can only provide auxiliary prompts for local code. Basically, it still requires manual editing, modification and verification. While Cursor/WinSurf both conduct AI understanding and automatic code generation at the level of the entire file or even the entire project: GitHub Copilot: Local code assistance.
- Core mechanism: GitHub Copilot is mainly based on contextual code completion and generation. Its core mechanism is that when you are writing code, based on information such as the current cursor position, surrounding code fragments, comments, and variable names, it predicts the code you are likely to enter next and provides suggestions.
- Assistant scope: Its assistance scope is usually local and mainly targets the code line or code block you are editing. It provides suggestions based on the content you are currently writing rather than performing global understanding and analysis of the entire file or project.
- Requiring manual editing: Therefore, GitHub Copilot is more like an intelligent code assistant to help you write code faster. However, the overall structure, logic, and the relationships between different code modules still need to be manually edited, modified, and verified.
Cursor/WindSurf: Global code understanding and generation
- Core mechanism: Tools like Cursor and WindSurf are more inclined to conduct AI analysis and understanding of the entire file or even the entire project. They will try to understand the structure, logic, and the relationships between different modules of the code, so as to carry out more intelligent code generation and refactoring.
- Assistant scope: Therefore, their assistant scope is wider and not limited to local code. Based on the understanding of the entire file or project, they can generate more complete code snippets, propose refactoring plans, and even help you quickly generate the code of the entire API call.
- Automatic programming tendency: Such tools have a certain degree of automatic programming ability. They can generate relatively complete code logic based on the understanding of your intentions and reduce the workload of manually writing code.
Not only this, they can also directly run terminal commands to verify the running results of code in real time, and even automatically modify the code further based on the feedback of the results. The entire global automated process can be said with certainty that once programmers use it, they will be inseparable from it. Because of this important difference, after launching the project-level code automation capability, both Cursor and WinSurf have quickly achieved their commercialization, and the paid subscription income has shown an explosive growth. Of course, this big cake has obviously been eyed by major companies and other entrepreneurial teams. For example, there is a recently popular project-level AI programming tool Lovable, claiming to have achieved the fastest commercialization record in European startups. It can be completely judged that in 2025, similar project-level programming automation tools like Cursor/Winsurf will bloom in an endless stream, and there will definitely be powerful alternatives in the open source community. Interestingly, basically all AI programming tools have relied on the power of Microsoft – these AI programming tools are basically developed based on the open-source Microsoft VS Code code editor – the basic coding and debugging and other interactive UIs are ready-made, and the hard work and dirty work is all handled by VS Code. The entrepreneurial teams only need to find a way to integrate large models into VS Code to analyze and modify the code. The current explosive scene of AI programming tools is definitely due to the open source of Microsoft VS Code. In 2025, the AI programming tool will be a battleground for large factories and entrepreneurial companies.Microsoft’s own Github CoPilot AI-assisted programming concept got off to a good start. At present, there may be a little danger of being late. Obviously, Microsoft will never give up the AI programming cake easily. And in the end, it is obviously us users who will benefit. In terms of the current usage of the city lord, similar to large language models, AI programming tools do not have absolute stickiness. Unless in the future, AI programming tools bind user data/code in a certain cloud storage mode, otherwise, project codes can even be directly switched to another AI programming tool during development (it will be more convenient if the tools are all based on VS Code) without affecting users’ development. (Of course, to promote users to switch tools, there must be obvious functional/performance improvements. Speaking of this, there is a small episode. The sudden popularity of WindSurf is largely because WindSurf took bigger steps in the automated code mode and achieved full automatic code writing and debugging one step ahead of Cursor. It rose suddenly when Cursor had already become famous and created the current situation of the two heroes. In 2025, all professional programmers will definitely use Cursor/WindSurf or better project-level AI programming tools. The workload of an ordinary software engineer in the era before AI can be completed in 5 minutes with the help of current AI programming tools. The improvement in efficiency is 8 hours x 60 / 5 minutes = 96, that is, about a 100-fold improvement.One thing is certain. Programming tools lacking AI support can be asserted to have been eliminated now. There is no other reason except that the efficiency improvement of project-level AI programming tools is too obvious. Based on the personal use of the city lord and combined with some work experiences back then, it can be said without exaggeration that considering some efficiency losses in large-scale projects, if we make a discount on this and calculate in the end, there must be an efficiency improvement of 20 to 30 times. Programmers who do not use AI programming tools – if there are any – need to be prepared to be laid off by the company first. Therefore, there is no need to look into the future. Just looking at the present, program猿 and program媛 who use AI programming tools are bound to win. At this point, we can fully understand why Cursor/WindSurf reached its peak as soon as it debuted and its business income soared: the efficiency improvement is too obvious and the effect is too good. It is really unbearable to pay. The toy for ordinary people and the magic weapon for programmers. Although some media articles have exaggerated that a person who does not understand programming at all can also use Cursor/WindSurf and complete script programming by just moving their lips – this is indeed true to a certain extent and is also the powerful place of Cursor/WindSurf. Within a controllable code scale (the controllable meaning is probably between a few hundred lines of code), they can completely autonomously implement some simple functional program scripts according to natural conversations, such as text processing or some small games, etc. Based on the built-in terminal operation and debugging capabilities, Cursor/WindSurf can also smoothly perform automatic testing and even simple automatic deployment.AGI programming tools. But at least in the current state of large models, the perfect code ability of this fully automatic black box is limited. — We can also think the other way around. If the AI programming tool has achieved perfect automatic programming without upper limits on complexity and code volume, this tool cannot be called an AI programming tool and must be called something else. On that day, human scientific and technological history will really turn a new page. The current situation is that before the birth of AGI, the ability of AI programming tools must be within a limited range. “People who don’t understand programming can become programmers by using Cursor/WindSurf” is just a hype and can be ignored. Students who don’t know programming using these AI programming tools are actually a bit like Doraemon’s pocket. They can take out various fun tools, but they need to rely on luck to see if they are useful. The essential problem here is still the context window. There is a detail that WindSurf can only read 200 lines of code at a time. Code exceeding 200 lines will be read in batches, and continuity is achieved by relying on context memory. When the code file is long enough (more than 1000 lines), reading and analyzing the code file and modifying it itself will lead to instability. Obviously, this is related to the current large model. We can expect that when a new generation of large models with ultra-long context windows are better integrated by AI programming tools, this problem will be alleviated to a certain extent. However, before reaching AGI, we still have this limitation: AI programming tools cannot be 100% perfect and correct. And the three abilities that students who hope to use AI programming tools to make real product-level projects must have are1. Skill One: Have the ability to organize project code. In simple terms, you generally need to know that for doing this thing, which several module needs to be implemented, and how to split these modules into different files to better organize and manage; the important thing is that when AI makes mistakes, reasonable file splitting can facilitate the location of errors / recovery of correct logic. At the same time, if necessary, this person must have the ability to read and understand code snippets. If you just rely on AI to grasp a large amount of black box code, then the only thing you can do is pray that the project can continue to operate when AI updates the code next time: when the number of code lines continues to grow, the complexity will soon explode to a degree that AI cannot handle. Therefore, this is also why at present, only people with professional programming abilities can give full play to the ability of AI programming tools to the greatest extent. Just like a dragon-slaying magic sword, it is completely different when held in the hand of an adult and a child.
2. Skill Two: Need to know how to ask questions. Learning to ask questions is actually the core skill in the AI era. These AI programming tools such as Cursor and WindSurf have a very big characteristic. It is both a coolie and a teacher. The key is whether you can ask appropriate questions. Of course, we can simply describe a function, let AI work hard, and then open a treasure box to see if it is correct. But if you can ask some technical questions in depth, have better interaction with AI, and even give AI some inspirational hints such as using what libraries to achieve what or how to roughly divide functions, AI will easily give more surprising performances.Obviously, whether one can ask appropriate questions completely depends on the accumulation of one’s actual programming experience. Skill three, the ability to learn quickly. The added value of AI programming tools for human programmers is, on the one hand, the coding speed; on the other hand, it is the expansion of programming knowledge and ability. Human knowledge bases are limited, but when AI programs, it will cover all possible existing choices and select the most suitable components/code libraries to achieve functions. Professional users of AI programming tools need to keep up with the rhythm of AI tools in the first place and quickly understand the new technologies introduced by AI in order to continuously keep the project firmly in hand. This may sound mysterious, but in fact, it is just the standard skill of every qualified programmer. All reliable programmers are masters of lifelong learning and grow continuously through learning in work. Therefore, the interactive communication with AI teachers + labor will only be a more virtuous cycle. The emergence of all-round AI coding tools with a double-edged sword for young people has a profound significance for professional programmers. These tools have completely changed the process and mode of coding work. Those who master AI tools to produce code and at the same time know what they are doing. In the future, only one kind of people can gain a firm foothold in programming jobs, that is, the problem is precisely that it is very enjoyable to communicate with AI tools with one’s mouth and wait for the output results like opening a treasure chest; but to know what one is doing and at the same time know what AI is doing, this is a bit difficult. Only by understanding programming and having the ability to learn and expand can the ability of AI programming tools be brought into full play. This still returns to the point that has been expressed before. This poses a great challenge to young people who are determined to enter the software industry at present.Experienced programmers all know that a person’s coding ability is accumulated through a large amount of painful practice, repeated staying up late to debug, and various pitfalls. It takes thousands of temperings from a novice to a skilled programmer. Under the current working efficiency of AI crushing, there is no chance at all for pure newcomers in the working environment. — Whether bosses are willing to pay a small subscription fee for an AI tool or are willing to spend a considerable amount of money to train coding newcomers is self-evident; not to mention that the ability of coding newcomers can no longer match that of AI tools. For students who are determined to enter the software industry and become “AI coding operators” in the future, the only chance is that during college, instead of only focusing on romance and game worlds, they should spend enough time silently accumulating sufficient programming experience and prove to the company that they can use AI programming software as proficiently as old employees to solve real product-level problems and generate and manage code projects with sufficient complexity. These requirements are actually reasonable. It can even be imagined that in the future, the interview method for new program recruits will no longer be to test on-site algorithm problems, but to give an AI programming tool and a few hours to implement a fairly complex actual software project on-site. Under such a new situation, as mentioned in the previous section, skill 3 “rapid learning ability” is what new students determined to enter this industry must possess. Then choose an AI programming software like a good teacher to accompany themselves through the time from newcomers to skilled veterans and become a qualified code producer. The biggest difference between code production and content production is that the former is manufacturing production tools, and this production tool can produce content including the latter by itself.For example, the simplest example. Through AI programming tools, we can quickly write a script for automatically summarizing the entire text. Of course, you can say that this can also be achieved through AI question answering on the web side. But what if this script can automatically traverse a directory and automatically generate summaries of all files? This is the power of code automation. AI programming automation is essentially almost like creating a productive machine in science fiction, with a creation speed hundreds of times faster than the previous AI era. In the future AI era, humans may be divided into several levels. The highest level is those who master AI programming software. They are the highest-level producers and can produce the tools they need. The second level is those who can use AI to generate content (text, music, video, etc.), and they are also qualified producers. The third type of people belong to those who can only be consumers and cannot master AI. Our descendants had better become people at the first level. The beginning of one-person companies started in 2024 when the concept and application of AI began to be popularized. An interesting concept, “one-person company”, began to enter the public topic. As the name suggests, with the help of the capabilities of generative AI, work that used to require the cooperation of many people may be completed by one person. Smart people have the opportunity to become a company with one person’s strength. This beautiful assumption is different in different fields. For fields that require a lot of face-to-face communication between people, this obviously needs to be questioned. But for the Internet/technology industry where a lot of work is done with computers, this beautiful assumption really makes people’s hearts beat. Judging from the performance of generative AI, including financial data processing, copywriting processing, marketing planning, etc., which used to require professional people to do, AI has been able to participate and play an important or even independent role.Just, we all know that the core of the Internet/technology industry is still code. No matter how powerful various AI-assisted roles are, the core problem of a one-person company in the technology field is still that the code output ability of one person is usually difficult to compare with that of a team. A god can solve core problems, but code production and software architecture require a lot of basic work, and there are too many dirty and tiring jobs. One person’s cognition may be sufficient, but unfortunately, there is too little time and too many things to do. Using traditional methods to rely on one person to stack code to achieve large-scale software or online applications is a very challenging thing. But now it is different. A senior technical person who is proficient in project-level AI programming tools is completely equivalent to the output efficiency of a technical team of 10 to 20 people in the traditional coding era. According to what the city lord said earlier, even if various possible efficiency losses are excluded, the efficiency increase of AI programming tools for programmers is also in the order of dozens of times. We can say that it is even higher. We know that the larger the technical team is, the higher the marginal cost of communication between people is. So the truly core technical team is always in the scale of dozens of people. Moreover, senior technical managers often have a feeling of hating iron for not being steel for their subordinates. If it is not for cultivating the team and there is not enough time, it is better to rush up and code by hand. A technical expert with AI programming tools is a product technical team by himself. This once extravagant hope has now become a reality. Now, if such an expert also has product/operation/market thinking, it is a real one-person company. (What if some technical big shots are not good at operation and marketing? It is very simple. Seek help from AI or ask the wife to work together.)