In 2 minutes, the requirement document becomes a product. The domestic large model development tool is popular in WAIC

Bai Jiao sent from Aofei Temple
Quantum Bit | Public Account QbitAI

From demand to application in two minutes, this domestically developed tool has become popular on WAIC!

All you do is throw a requirements document at it.

It is called Xingchen Big Model Software Factory and was jointly released by China Telecom Artificial Intelligence Technology Co., Ltd. and China Telecom Artificial Intelligence Research Institute (TeleAI) .

With it, everyone can complete application development. You don't even need to think about prompts or talk to the big model during the whole process. If you want to change the requirements, you only need to modify it in the document. What you write is what you get.

This saves a lot of time on demand alignment, and the efficiency goes up sharply! It is a perfect tool for workers.

It is reported that the product will be open for testing at the end of the month, and in their early internal testing, this software factory has brought a 4-fold speed increase.

Let’s take a look at how powerful this development tool is.

Domestic development artifact "Software Factory"

In the several demos they released, we can see that this software factory has three main features:

Low threshold, what you write is what you get, end-to-end .

At the beginning of project development, you need to provide a detailed requirements document, including product positioning, target customers, and most importantly, functional description.

For example, if you want to build an e-commerce seller system, you will need different functions on both the buyer and seller sides, such as product list, product details, shopping cart, purchase order, and my order;

After uploading the product requirements document, the software factory starts operating.

From back-end architecture design and code generation to front-end interaction design, code generation, application testing, and production release, each step is visualized on the interface.

On this page, you can see that on the left is the requirement document/file, above is the development process, and in the middle is its workspace.

First, we come to the backend architecture design stage, which can automatically generate a call relationship diagram.

The code is then generated according to this architecture diagram. In addition to being able to see the entire generation process, you can choose to view the file structure, download the code, view the application, etc. at any time.

By the way, please pay attention to the time below. The speed can be said to be very fast.

When we come to the front-end stage, we recognize that there is already a system sketch in the requirements document, so we go directly to the code generation stage.

The workspace is also divided into two areas, one for effects and the other for code . You can see the entire front-end interface generation process and can stop and modify it at any time.

It also enables self-testing of applications before production release.

And if you want to change your requirements, for example, if you want to add a "product collection" in the buyer's system, you only need to modify it in the document and then redevelop the application again.

In their own words, documentation is engineering, and what you write is what you get .

This is also the biggest feature of the Star Big Model Software Factory .

Throughout the entire process, all you have to do is provide a requirement document and then "supervise" the generated results during the process.

The traditional software development process involves the flow and transmission of information between people in different positions. How to ensure the consistency of requirements in this process has become a major challenge in software development. Because if the information flow is slightly misaligned, the development efficiency will be greatly reduced.

However, this software factory is centered around requirement documents, and the entire process can be completed online with one click, eliminating the need for communication between departments in different positions, and dramatically increasing development efficiency.

This is fundamentally different from most current large-scale model software development tools . Most of them are based on conversational interaction, but this software factory can achieve one-stop, end-to-end access from demand to application, and the process is controllable and can be "intervened" and modified at any time.

The former is mainly suitable for individual developers, while the latter is more suitable for cost reduction and efficiency improvement at the enterprise level on a larger scale, including early application demos, layout of new industries and new fields, daily software development, etc.

This is also happening within their companies.

It is reported that after their initial testing using the software factory, the software development process was accelerated by 4 times .

There is such a set of intuitive data.

During the development of a branch function of a large system, the project originally required 2 product managers and 6 developers and took 1 month to develop.

As a result, with the assistance of Star Software Factory, only one product manager and three developers were needed to complete the software development in two weeks.

It is no wonder that this software development tool attracted so much attention at WAIC.

So what is the principle behind this? Let’s see what the producer says.

Speed ​​up the software development process

China Telecom is a company that is not unfamiliar to readers. It is a unique company among large-scale model players. It has computing power, data, and a professional technical team.

The previously open-sourced Star AI big model and dialect big model have caused quite a stir in the industry.

The layout of large multimodal models such as semantics, speech, and vision has been completed.

TeleAI revealed that this software development tool relies on its self-developed multimodal large model to build core capabilities from three aspects:

• New collaborative process for document development .

This is the most intuitive function during use. It implements development directly through writing documents, allowing users to focus on the implementation of core business logic rather than traditional serial development, which requires tedious coding and debugging.

• Controllable code generation .

How to ensure the stability and controllability of generated code is a key challenge currently facing the industry.

How did Software Factory solve this problem? They proposed the concept of code primitives, extracting the smallest method unit from massive code, decoupling the relationship between code modules, and effectively improving the maintainability of code generated from large models;

At the same time, building calling specification knowledge for the standard library enables controllable code generation and improves development efficiency.

• New modes of programming interaction .

Currently, the code generation paradigm for large models is mostly limited by prompt words and dialogue token limits, and cannot complete a complete project development process.

The software factory provides project-level collaboration and management capabilities, allowing project development, testing, and deployment to be completed without the need to communicate directly with the big model.

It is worth noting that the Xingchen Software Factory actually revealed a very obvious signal:

From then on, China Telecom has become more closely associated with AI .

In the previous inherent impression, China Telecom’s operator label was greater than its technology label, so much so that when they released the big model, many people in the industry were surprised.

But in fact, in the field of big models and AI technology, China Telecom has long-term accumulation and multi-faceted layout, and has always been a solid technology school.

At present, China Telecom has a research and development team of nearly 800 people, led by AI scientist Li Xuelong , and has established TeleAI. TeleAI will focus on the research and development of cutting-edge AI technologies in line with national strategic needs and AI development trends, and actively promote the deep integration of technological innovation and industrial applications.

In terms of the big model ecosystem, they have joined forces with the leading ecosystem to build big models for more than 20 industries including education, government affairs, and emergency response, covering more than 500 application scenarios across the industry .

At the data level, we have already accumulated more than 500TB of text data, 1.2 billion pictures and text data, and PB-level video data .

In addition, they also have natural application scenarios and user bases, which can promote the implementation of large models in various fields more quickly and form new economic growth points.

A new paradigm for software development

The biggest keyword of WAIC this year is large-scale model application.

If we were still talking about trusted AI and data security in previous years, this year the big model technology paradigm has become a strong consensus. Both the exhibition area and the forum are moving towards the implementation of big model applications.

Of course, one of the hottest topics is software development.

When GitHub Copilot and Devin were first introduced, many engineers were amazed at the paradigm shift in software development brought about by big models. However, with actual experience and the release of more code generation products, related thinking has become more sober.

Most products on the market are actually single-point functions and simple applications. On this basis, there will be more updates in product design, such as a visual operating environment, but it is still dominated by the large model dialogue interaction mode , and is still aimed at the process thinking of individual developers.

However, when it comes to industrial-level digital and intelligent innovation, this development paradigm actually has many limitations.

Firstly, it is limited by the level of your prompt words, which is still a certain threshold for the general public and non-software developers;

Secondly, the randomness and illusion of large models make it impossible to develop complex and stable large model applications.

Then there is collaboration. Project-level development involves multiple links and collaboration between different personnel, including the entire process, including front-end and back-end design, subsequent testing, and deployment.

This time, China Telecom actually brought new paradigm characteristics: one-stop and collaborative .

The big model allows everyone to experience the possibilities of artificial intelligence through conversational interaction. However, when it comes to industrial production and development, conversational interaction may not be a rigid requirement, but deep integration with the software development process and coordinated scheduling of various steps and departments are the fundamentals.

Previously, former Tesla AI director Kapasi described the development of software development in this way.

Automated software engineering currently looks similar to automated driving.

Specifically, it is reflected in the development process : first humans write code manually, then GitHub Copilot automatically completes a few lines, then ChatGPT writes code blocks, and now Devin appears.

Next, he thinks automated software engineering will evolve into coordinating the many tools developers need to string together to write code: terminals, browsers, code editors, etc. And humans will be responsible for supervision, gradually moving to higher-level work.

Well, now the emergence of software factories has taken automated software engineering one step further.

Click here ???? Follow me, remember to mark the star~