Image source: generated by Unbounded AI

In the summer of 2024, after 3 major feature updates and more than 40 feature iterations, the AI ​​code editor Cursor It created a programmer craze in North America, and immediately became the programming tool of choice for Chinese programmers on the other side of the ocean.

Facing the powerful GitHub Copilot, Cursor brings changes in interaction methods, innovations in using somatosensory, and the ability to globally complete code for the entire program file. Although it is incubated by OpenAI, the base model abandons GPT4. The decision to select Claude, the rapid acquisition of 3,000 customers in various fields, and the US$400 million valuation reached as early as August have all made it the focus of heated discussions in the technology circle.

Cursor is not even the first popular AI programming product in Silicon Valley, nor the last.

In March this year, "AI programmer" Devin attracted widespread attention in the industry. Only five months later, another AI start-up called Cosine claimed that their newly launched AI programmer Genie performed far better than Devin, in August, Cursor, the enhanced code editor integrated with AI, quickly became the top trend. As a derivative version of VS Code, it inherits VS Based on the advantages of Code, it fully integrates AI functions, greatly simplifying the software development workflow and programming process. After it, Magic, Bolt, Replit, and Supermaven have become representatives of AI programming star companies. They are each good at different directions. Supermaven emphasizes the length of the context. Bolt and Replit have made more extensions in the workflow. They can not only design the overall structure of the code, but also make global modifications to the code, even beyond the capabilities of Cursor.

According to reports, the total financing amount of these companies has exceeded 2.2 billion US dollars. AI programming has gradually become the sexiest AI track in Silicon Valley. Among the more than 10 AI programming startups active this year, 7 have grown into unicorns. However, unlike the "chasing" in many fields, while AI programming companies are sweeping Silicon Valley, there is basically no big news about AI programming startups in China.

An investor from a mainstream VC in the cutting-edge technology field told Silicon Stars that in fact, a number of AI programming companies suddenly emerged in China last year. They sorted out fourteen or five companies at that time.

“Those entrepreneurial teams are thinking about various things on the programming track, such as code search, programming for papers, code annotations, or code repairs, and some are doing pure code generation. Completely aligned with Cursor,” he said.

“But the problem is, the level is much different.” He described that, in general, these teams did not do code generation to a deep level.

Silicon Star learned that Qiji Venture Forum invested in six AI programming companies last year.The start-up in the field has almost been wiped out since then, and most of the more than 10 code teams that briefly surfaced last year have withdrawn this year.

Compared with Cursor, realistic feeling

"The level is poor "Many" problem is actually a common problem in the AI ​​industry. In terms of basic models and Chat-type AI applications, there is actually a situation where Chinese companies are catching up with their American rivals. However, in fact, financing is still happening in these tracks, and investors can also be self-consistent based on market logic.

But there is a big difference in AI programming, that is - AI programming products for developers have no national boundaries. This is different from those products such as knowledge base Q&A assistants for enterprises. Due to differences in language, ecology, etc., China and the United States are very different. According to Silicon Star people, the knowledge base Q&A assistant for enterprises is oriented to Chinese and American customers, and the difference between the Chinese and English versions is huge.

“The United States made good products first, and domestic developers will use them without many barriers.” Hu Yichuan, CTO of Laiye Technology, pointed out.

So, the first problem on the AI ​​programming track is that the level is too shallow. If it cannot reach the status of upstarts such as Cursor, Bolt, and Magic, if it cannot meet this hard standard, in the eyes of many Chinese investors, no matter how good the team is, it will not be able to attract investment.

A co-founder of the AI ​​application team said that many star products similar to Cursor are currently available in overseas markets. In essence, the U.S. capital market is using the best large models overseas (used by Cursor). It is Claude) who directly makes plug-ins and makes Agent programming products very buy-in. Objectively speaking, at the model level, there seems to be no shortage of domestic models that are comparable to GPT4, but the problem does not seem to be here. Because even if they are based on overseas models, the product completion and capabilities of many current applications are still lacking.

So, when domestic investors see a common logic in the AI ​​programming track as benchmarking Cursor, they naturally cannot make a move. The AI ​​application entrepreneur mentioned above said that he had experienced dozens of very similar conversations, but found that investors ultimately believed that "domestic products cannot meet this standard."

“At this stage, there are no Chinese teams that can solve this kind of IDE ecological problem overseas.” AIGCode CEO Su Wen said. IDE refers to the integrated development environment, which refers to the application used to provide a program development environment, including code editors, compilers, debuggers and graphical user interface tools, such as Microsoft VS series. He believes that even if Chinese companies want to implement "plug-in logic" like Cursor overseas, it is still far away.

When the technology itself fell behind, a common logic used by Chinese investors in the past was that we have a larger market and application scenarios., it can run quickly in commercialization, which will bring about overtaking opportunities in corners. However, when it comes to AI programming, the commercial environment is no better than overseas.

“Invest in it (AI programming) because it makes money.” Xu Xiaoyu, a partner of Amino Capital based in Silicon Valley, said that AI programming is popular in Silicon Valley. The reason behind it is that the PLG (Product Driven Growth) SaaS model is popular throughout the world. Overseas is established. Xu Xiaoyu found that for the start-up companies their organization invested in in the past three years, they discovered and found PMF's generative AI companies. Compared with companies without generative AI drivers, they can reach 10 million US dollars in ARR (annual recurring revenue) in half the time. Although this cannot Help these companies become Google in the future, but it is enough to develop into a small unicorn. The most typical example is Replit, which was established in 2016 and became an upstart in the programming world this year.

But in fact, even the popular programming tools in Silicon Valley such as Github Copilot, Cursor and Bolt have not reached the point of strong payment in reality. Magic, another upstart that builds its own programming model, has not even released a officially available product. They still solve the needs of existing programmers in existing scenarios.

Domestic toll roads are in the earliest stages.

It is a cliché that the domestic 2B SaaS ecology is not profitable due to low profit margins and cannot gain momentum due to complex environment. Even Kai-fu Lee recently said that "there is no delusion about SaaS subscriptions yet." Moreover, an important target group for AI programming is programmers from Internet companies, but large companies prefer teams to make their own production tools. Public information shows that Alibaba Cloud, ByteDance, Huawei, and Baidu all have mature AI programming businesses internally. These businesses serve internal purposes and reduce many market opportunities for external startups. At the same time, these businesses are also very popular when the market matures. They may turn around and provide services to the outside world, just like the experiences of DingTalk and Feishu back then. Then the space for startups will be further crushed.

Looking for a way out: Some people are looking for unique market opportunities, while others think that we still have to fight head-on

Liu Gang is one of the earliest venture investors in China to focus on the direction of the AI+ industry. Following Alpha Commune’s logic of “invest in people, not in the track”, as a partner He visited several AI programming groups with great potential early on. The team, including one of the projects in the programming direction, has good qualifications and good products. The project is aimed at B-side enterprises and developers, but payment is a problem. They have found a few big B customers for privatized deployment, but overall, "to put it bluntly" "I can't collect money," and I can barely maintain it but cannot achieve rapid development.

This team was in trouble in the second half of 2023. At the beginning of this year, they resolutely transformed into a new field and began to have some good revenue and business increases.

Beijing UniversityLi Goyer, a senior professor, is the earliest pioneer in this track in China. He founded aiXcoder two years ago. Before the advent of ChatGPT, Li Ge used more traditional programming methods to incubate the project, making plug-ins and code completion in the IDE (Integrated Development Environment), which is somewhat similar to the classic knowledge graph.

Starting in 2023, aiXcoder will turn around and embrace large models, expand B-side and 2G-side business, and have successively won business orders from several banks and state-owned enterprises. It is predicted in the middle of the year that there will be about 60 million in sales this year. closing, and a market valuation of nearly 1 billion yuan. There are also investments from Hillhouse, Qingliu Capital and an automobile industry chain fund.

“This is a unique opportunity in China. There are many large companies in the country that have relatively large development teams. They need the assistance of AI programming, but they cannot use GitHub Copilot or Cursor and need to connect to the cloud. Model products. "Hu Yichuan believes.

At present, most of the country's leading banks, insurance companies, and large companies in the financial industry have extremely large development teams, ranging from a few thousand to tens of thousands. What they have in common is hope. Use advanced AI tools and technologies, but it is unlikely to use programming tools on the Internet. For security reasons, you must use an AI programming tool that can be deployed locally in the environment.

This is not only a characteristic of AI programming of a track, but also reflects the new trend of the entire To B implementation of large models. Hu Yichuan believes that many customers currently want not only your model itself, or AI programming software, but a solution that integrates software and hardware. “If you want to deploy this thing locally, what kind of GPU do you need to choose, and how do you install it on the GPU? How to use GPU efficiently for training and inference requires manufacturers to have very professional service capabilities.”

In short, “AI. The roles in coding range from design to development to testing to release. If a new company wants to continue on this path, competition will be very fierce unless it finds a very unique group, or a very vertical field, or some general Only if it can solve the problems that cannot be solved by other products will there be opportunities," Hu Yichuan said.

This is indeed a way of survival. Recently, Ming Chaoping, the head of the original Dark Side of the Moon video generation product Noisee, resigned. His entrepreneurial project is also an AI programming company. According to Silicon Star, this company is taking a lightweight product route similar to Websim, targeting games and other scenarios. , (Websim is a website that can generate websites only through text descriptions. It can generate small games and a piece of music. It is driven by large models such as OpenAI and Anthropic, such as Claude 3.5. Sonnet and GPT-4o), currently do not have their own exclusive models, and will take the paid route of products that are lighter than Bolt.

At the same time, there are still new entrepreneurs who “do not believe in evil” and believe that the final way out is to compete with the strongest overseas products."Head-on", gain your own opportunities through innovation in capabilities and models.

Su Wen of AIGCode is one of them. He said that a lot of work done by some domestic AI programming companies is code testing and code repair , this is just entering the programming track, but it is not really doing in-depth code generation.

"This real job is like eating meat, you need to give up the leftovers."

After he left Huachuang Capital as an investor in March 2021, he retained his status as an investment partner, but devoted himself almost full-time to the entrepreneurial model. Finally, he founded an AI programming company in January this year. AIGCode has received two rounds of financing

AIGCode’s product is an end-to-end Autopilot tool with its own “pre-train "from scratch" general model, we want to benchmark poolside and magic, and become a product engine in the era of large models.

Su Wen told Silicon Stars that he regards end-to-end code generation as his team of more than 20 people "What more than 20 people can't handle, even 200 people can't handle. The technical talents in this track are very limited. How many people have done pre-training and how many have done advanced and innovative software architecture." Woolen cloth? "The scale of 20 people is comparable to the size of his target competitor, the American AI programming start-up Magic.

We do code generation from the model and software architecture, build programming tools that complete tasks end-to-end, and Train your own model, integrate it vertically with the application, and finally take over multiple functions in the APP factory. This kind of end-to-end completion of the task In Su Wen's eyes, this is the only way to stand out in the programming field. In Silicon Valley, where the division of labor is clear in the link pipeline, end-to-end is not necessary, but in China's development and B-side environment, end-to-end may be more appropriate. A model that meets market demand.

"Only end-to-end code generation or fragmented code completion is called AI programming."

But this also requires that you can really do a more complete process than the upstarts in Silicon Valley, and do it better than them. This is obviously not easy. Like other peers, the market and investors have given him a lot of experience. The time window is also limited. Everything needs to be accelerated.

Su Wen said that his team has completed a lot of prior work. We are in the stage of completing the functional coverage of the product, and have recently started internal testing of the product.

“Before the payment point is reached, the best way is to get users up first and let the product come out. The track is like mountain climbing on the north and south slopes. Copilot has climbed from the north slope to the base camp first. We are still different on the south slope, but everyone can reach the top in the end. "Su Wen said.