With the rapid development of AI technology, the developer community is paying close attention to the actual effectiveness of AI programming tools. Hong Dingkun, vice president of technology at ByteDance, announced at the Volcano Engine Force Momentum Conference that it used the AI programming assistant TRAE to open source its English learning application developed in three days, which attracted widespread attention. This article will delve into the development process, technical details, and ByteDance’s far-reaching thinking on “AI Development” behind it.
Three days of development, thousands of lines of code, from conception to launch – this is the true speed of the AI programming era.
At the Volcano Engine Force Force Conference on June 11, Hong Dingkun, vice president of technology at ByteDance, announced that it would open source its English learning application developed by the AI programming assistant TRAE for three days, and the developer community was eagerly anticipating it.
On June 18, the promise was fulfilled as scheduled, and the complete code of the application, called “Stream to River”, was officially launched on GitHub, which quickly attracted widespread attention.
You can experience the application on the website: https://sstr.trae.com.cn
This is not only an anecdote about the VP personally “coding”, but also a powerful demonstration of the actual effectiveness of AI programming tools, and also carries ByteDance’s far-reaching thinking on “AI Development”.
“Accumulating into a river”: Three days to quickly complete the application of “although the sparrow is small, it has all five organs”
At the Force conference, Hong Dingkun shared his experience of working with TRAE to develop “Accumulation into a River”: “After the Dragon Boat Festival holiday last week, I developed a new English learning app ‘Accumulation into a River’ with two colleagues…… I completed the development in 3 days. ”
He mentioned that about 85% of the code was produced by AI (TRAE) through natural language dialogue, and more than 3,000 lines of code were developed and debugged in two days.
Today, we can see the project in its entirety on GitHubhttps://github.com/Trae-AI/stream-to-river.
Judging from the project’s README and code structure, “accumulating into a river” is by no means a simple demo. It is a relatively complete English learning app with core functions including:
- Word Learning and Management:It supports word addition, query, detailed display, and combined with the Ebbinghaus forgetting curve for review progress tracking and intelligent question generation.
- Smart Chat:Based on large language models (LLMs), it provides real-time chat functionality and supports streaming response, session management, and content highlighting.
- Multimodal Inputs:It integrates speech recognition (ASR) and image-to-text functions, enriching the way users learn and input.
- User system:It includes basic modules such as user registration, login (JWT authentication), and information query.
- Technical architecture:It adopts a microservice architecture with separate front-end and back-end. The backend is based on Go language, the API service layer uses the Hertz framework, the RPC service layer uses the Kitex framework, and the data store uses MySQL, supplemented by Redis for cache optimization. The front-end technology stack includes TypeScript, JavaScript, CSS, etc.
It can be seen that “accumulating into a river” involves multiple levels such as API services, RPC communication, data persistence, caching, and external service calls (such as LLM, ASR), and is a modern application with a certain degree of complexity.
Hong Dingkun’s ability to complete such a project with the help of TRAE in such a short period of time is undoubtedly a strong proof of the great potential of AI programming tools in improving R&D efficiency. He even mentioned, “For a 300-line code feature, I might only need 200 words of scenario description.” ”
This experience of “natural language programming” is changing the traditional development model.
Hong Dingkun: TRAE’s goal is “AI Development”
In his speech at the Force conference, Hong Dingkun elaborated on why ByteDance is investing heavily in AI coding and the vision of TRAE.
- Technology is inclusive, and AI makes everyone a developer: AI lowers the threshold for programming, allowing more people to solve problems and realize ideas through code. For example, he said that a colleague in the company used TRAE to teach 11-year-old children to program and successfully built a Mathematical Olympiad question bank website.
- Improve R&D efficiency: Within ByteDance, more than 80% of engineers use AI tools such as TRAE to assist in programming, and the proportion of AI-generated code is also considerable. This is a huge increase in efficiency for large-scale technology companies.
- Pursue the upper limit of intelligence: The structure and logic of coding tasks make it an excellent scenario for measuring and improving the intelligence level of large models.
More importantly, Hong emphasized that TRAE’s goal is not just “AI Coding” but “AI Development”.
“In a typical software development process, writing code may take up less than 40% of the work…… AI has the opportunity to coordinate these tasks. TRAE hopes to become a “scheduler” that integrates requirements management, design, coding, testing, deployment, operation and maintenance to achieve “software development all in one”.
Taking debugging bugs as an example, in the future, AI may be able to automatically locate problems from logs, analyze the causes, and automatically modify the code after confirming with the developer and go online, shortening the work that originally took half a day to a few hours or even less.
The introduction of the Agent capability by TRAE, which allows users to customize tools and connect workflows, is an attempt in this direction.
Human-Machine Collaboration: AI Coding is inseparable from human intelligence
Although AI has shown strong programming capabilities, Hong Dingkun is also soberly aware that AI coding is inseparable from human collaboration at this stage. “Purely using AI for development, I just make requirements, click buttons, and the program I make is difficult to maintain.”
In his development of “Accumulation into a River”, although 85% of the code was generated by AI, he emphasized: “I am still driving the whole process.” He is responsible for proposing technical solutions and core processes (describing code logic in natural language), which AI converts into code, after which he carefully reviews the code and can take over manual modifications at any time.
He believes that only when AI can “understand and understand people’s thoughts”, “understand context”, and cooperate well with people can it be a true “Real AI Engineer” (the meaning of TRAE) and finally achieve “AI Development”.
The Future: AI Reshaping Software Development Paradigms
From Hong Dingkun’s speech to the open source of “accumulating into a river”, we see not only a technical demo, but a true epitome of the era of AI development.
As Hong Dingkun said: “Is it possible for AI to do this in the future?” It helps me automatically locate from the logs, then analyze what possible problems there may be, and confirm them with me. When I thought it was okay, I said you can change it, and after the change, he helped me submit it online. ”
This full-process AI collaboration development method may become a reality in the near future. And ByteDance is moving forward quickly on this road through products like TRAE.
When AI can understand complex technical solutions, generate high-quality code, and assist in the entire development process, the threshold for programming will be greatly lowered, and the speed of innovation will be exponentially increased.
The open source of the “Accumulation into a River” project is like opening a window for us, allowing us to see what the future will look like in advance.
