[Editor's note] The big model is still in full swing. Yesterday, Lei Jun passionately announced that Xiaomi has run through the big model on the mobile phone. Today, the Xunfei Xinghuo model is here to honor the upgraded Flag , focusing on coding capabilities and multi-modal capabilities. For programmers, it is worth noting that Xunfei released a new intelligent programming assistant iFlyCode. Under the 100-model war, AI programming tools have also started a fierce war.
Author | Tang Xiao originated from Hefei
Editor in charge | Zhang Hongyue
Listing | CSDN (ID: CSDNnews)
On August 15th, Beijing time, iFLYTEK held the Spark Cognitive Model V2.0 Upgrade Conference at the Binhu Convention and Exhibition Center in Hefei. The official announced that the code capability has been upgraded to 5 key capabilities, except for common code generation and completion. In addition to error correction, there are code interpretation and unit test generation. And all these capabilities can be used in the new intelligent programming assistant iFlyCode released by Xunfei.
Why put code capability upgrade in the first place? Liu Qingfeng, chairman of HKUST Xunfei, told CSDN (ID: CSDNnews): "The ability to code is direct and rigid for the empowerment of various industries. Code is not only a key support for linking the digital world, but also an important tool for lowering the threshold for entrepreneurs. Everyone does not need to be a master of programming, as long as they use their imagination and cognition of application scenarios, they can improve development efficiency and realize related entrepreneurship. Whether it is the development of China's software industry or digital economy, if we do not take If the code ability is raised to the international leading level, then the national production efficiency must not be comparable to that of the international giants, so I think this matter is very important. Secondly, the demonstration of the code ability is relatively simpler and more intuitive. Considering the comprehensive consideration of the code in the multi-modal in front of."
So, to what extent has Spark’s code capability achieved? Liu Qingfeng said that in the test, the code generation and completion dimensions have surpassed ChatGPT, and Xunfei Spark’s code capabilities in all dimensions will fully surpass ChatGPT on October 24 this year. , will benchmark GPT-4 in the first half of next year.
As for the programming assistant iFlyCode, is it benchmarked against GitHub Copilot? Liu Cong, dean of HKUST Xunfei Research Institute, said frankly to CSDN: "The current code assistant products are all based on the Copilot mechanism, which can be understood as the logic of benchmarking Copilot."
Next, let us take a look at how the Spark model actually upgrades the code capabilities, and how developers can use it to achieve the effect of "10 times engineers".
It has surpassed ChatGPT in terms of code generation and completion, and will benchmark GPT-4 next year!
When the Spark model was just released, the code it generated, the code of the plural class was basically implemented correctly, but the test cases were not generated completely at one time. Today, three months later, Spark's code capability is one step closer to being used in a real production environment.
This time Xunfei Spark 2.0 has upgraded the code capability in five dimensions. Needless to say, code generation and completion, in terms of code error correction, it can locate spelling, grammar and logic errors, and also supports one-click modification; For programmers, reading code, which is more troublesome than writing code, is better reflected here. Spark's "code explanation" function can directly give a detailed interpretation as long as a piece of code is selected. In terms of unit test generation, as long as the code is selected, a single test case can be generated with one click, and unit test data can be intelligently generated.
Display of iFlyCode in the on-site experience area
For programmers, "Show me the code" is always paramount. In Liu Cong's live demonstration, whether it is writing a function, drawing a heart with Python, or making a snake game, it can be done quickly.
The author found in the actual test that Spark generates code very quickly and can be run directly. But in the process of generating it, to really realize the function you want, you need to adjust or try different prompts. Qixi Festival is approaching, let’s use the example of drawing a love heart, Python runs out the love heart very smoothly.
But switching to Java failed, and tried several different prompts without success. The following is a run-up style of a Java code example that Spark clearly states that it can draw a heart shape:
Of course, using Java to draw a love heart, using ChatGPT also failed (although both Prompt and GPT answers clearly stated that it is a code example for drawing a peach heart):
In terms of programming language support, Spark's ability to support Python is very significant. According to Liu Qingfeng, according to HumanEval, a public code ability test set built by OpenAI, the effect of Spark V1.5 Python is only 41 points, and V2.0 has reached 61 points, which is close to ChatGPT. According to the test set used in the real scene of the code built by the State Key Laboratory of Cognitive Intelligence, the code generation and completion dimensions have surpassed ChatGPT. Still as the previous plan, the capabilities of Xunfei Spark code in all dimensions will fully surpass ChatGPT on October 24 this year, and will be benchmarked against GPT-4 in the first half of next year. In this regard, Liu Qingfeng said, "Currently, the logic, algorithm, method, system, and data preparation for code capabilities are all ready. All we need is time and computing power. These things and computing power are being supplemented. On this year's 1024 Will see updates on progress.”
So, how programmers can better use the programming ability of Shanghuo, iFlyCode came into being.
Compared with Copilot, what does the intelligent programming assistant iFlyCode bring?
In terms of AI programming, there are currently two common types. One is the IDE integration represented by GitHub Copilot, which is very common and allows programmers to directly realize the closed-loop programming in the IDE. Developers directly gain an experience similar to ChatGPT in the editor. Deep integration with VS Code and Visual Studio can provide developers with in-depth analysis and suggestions for correcting errors. Google's Android Studio also adopts integration In this way, the conversational programming assistant "Studio Bot" is used to help Android programmers improve programming efficiency. Another type is to open up an online programming platform from Chat, such as Google's Bard directly connected to Colab.
We see that Xunfei chose IDE integration. iFlyCode seamlessly integrates in the form of plug-ins familiar to programmers, and supports 5 mainstream IDEs.
At the press conference, Liu Cong, in the VS Code integrated with the iFlyCode plug-in, through a few steps of Prompt, without writing a line of code, he realized the "volley handwriting" function that can be written by pinching two fingers.
According to the statistics of iFlyCode 1.0’s performance on more than 2,000 employees tested by iFLYTEK’s internal R&D performance platform within one month, in some typical scenarios, the code adoption rate reached 30%, the coding efficiency increased by 30%, and the overall efficiency increased by 15%.
The iteration of code intelligence is rapid, and it is greatly changing the traditional programming method. Perhaps Andrej Karpathy (OpenAI scientist, former head of Tesla AI) put forward that "the best programming language is natural language" is really not far from us up.
At the same time, I have announced with my friends that for smart programming assistants such as iFlyCode, CSDN will also bring you code writing tests and code task evaluations. Friends are welcome to continue to pay attention.
What does AI programming mean for developers?
At the current stage, what we can clearly see is that AI has further become our programming assistant. Its impact on programmers in at least these three areas is enormous.
Be able to solve some problems that may require searching and looking at a lot of materials to find the answer. For many years, we have used a lot of searches in our programming scenarios, and asked searches in case of problems, but it is likely to spend a lot of time and energy, but we can only find relevant solutions, and it takes a lot of time to solve the problem. Debug and modify. And what makes programmers particularly troublesome is that the searched content that can solve the problem is often the 5%-10%. The AI programming assistant can quickly and relatively accurately give us answers, even a piece of code that can directly solve the problem.
There are also some physical tasks in programming. For example, some tasks that need to be written without complex logic but relatively large amount of code can be directly thrown to the AI programming assistant, thereby greatly improving the programming efficiency. In addition, when writing code, sometimes you will encounter some code that is not easy to start with. You can directly give it to AI and let it implement some logic, which may help developers find new ideas.
Then, when "everyone is a developer", the threshold for programming is further lowered. For developers, will intelligent programming change the role of developers? In the future, what aspects will the core competitiveness of developers be reflected in?
Liu Cong said to CSDN in this way: "Programming assistants can simplify the work of programmers, improve production efficiency, and release the energy of developers, so that they can do more creative work." Where the core competencies are being replaced.
Liu Qingfeng said this sentence at the scene - "the future does not belong to AI, but to new humans who have mastered AI", and I extended it as: "The future does not belong to AI , but to new programmers who have mastered AI. "
