Driving China, August 15, 2023 news, at 14:00 today, iFLYTEK held the press conference of Xinghuo Cognitive Big Model V2.0, officially announced some new capabilities, and will launch new functions such as "code generation, image generation, video generation, and Spark plug-in".
Among them, the code generation capability can realize code generation, code completion, code error correction, code interpretation, unit test generation, etc.
At the press conference, iFLYTEK also demonstrated how to write code in one sentence through the Spark Cognitive Big Model V2.0, just issuing the instruction to write a snake mini game in Python, and automatically generated a simple mini game, which is very powerful.
Liu Qingfeng, chairman of iFLYTEK, introduced that the code capability of iFLYTEK V2.0 has been greatly improved, and from the comparison of the effects on the industry reference test set, the code writing ability of Xinghuo V2.0 based on Python and C++ has been highly approaching ChatGPT, with a gap of only 1% and 2%.
From the dimensions of code generation, completion, error correction, interpretation and unit test generation of the Python language, it can be seen that now iFLYTEK Spark V2.0 code generation and completion have surpassed Chat GPT, and several other capabilities are also catching up quickly.
According to Liu Qingfeng, the current logic, algorithm, method, system and data preparation of iFLYTEK Spark V2.0 have all been prepared.
Based on this, iFLYTEK announced an intelligent programming assistant based on Spark 2.0: iFlyCode 1.0.
It is reported that iFlyCode 1.0 has all the code generation related capabilities of the Spark Model and can be seamlessly integrated into the development environment.
In addition, it can help developers generate feature implementation ideas, provide modification suggestions, and answer related professional questions.
According to iFLYTEK's internal R&D performance statistics, after more than 2,000 employees use iFlyCode 1.0, the code adoption rate has increased by more than 30%, the coding efficiency has increased by more than 30%, and the overall efficiency has increased by more than 15%.
It can be said that the launch of iFlyCode 1.0 has liberated the productivity of development and greatly improved efficiency.
In addition, iFLYTEK Spark Cognitive Model V2.0 also shows functions such as voice generation pictures and virtual humans. According to the on-site demonstration, iFLYTEK Xinghuo Cognitive Large Model V2.0 can quickly generate accurate pictures according to instructions, and at the same time modify pictures according to new instructions.
For example, after generating a picture of a magnificent view of a snowy mountain, a towering mountain peak covered with snow, and adding the need for "adding a little girl to climb the mountain", iFLYTEK Spark Cognitive Big Model V2.0 can generate content correctly.
In terms of Chinese understanding, iFLYTEK Xinghuo Cognitive Model V2.0 can also perfectly generate pictures such as "Guan Yu riding a motorcycle", "Mountains and heavy waters have no way, and there is another village", leading competitors.
In addition to pictures, iFLYTEK Xinghuo Cognitive Model V2.0 also supports the generation of videos with virtual humans, which can save the cost of asking real people to shoot promotional videos such as products for companies and enterprises.
Liu Qingfeng said that the logic, algorithms, methodological systems, and data preparation of the current code capabilities are ready, and what is needed is time and computing power.
In version 2.0, the text generation ability increased by 72%, the language comprehension increased by 78%, the knowledge question and answer increased by 70%, the logical reasoning increased by 60%, and the mathematical ability increased by 72%.
The mathematical ability of iFLYTEK Xinghuo's cognitive large model is second to none in China, and it is far ahead in the major third-party evaluations after the college entrance examination. At the press conference, iFLYTEK also demonstrated a triangle math problem, and said that only iFLYTEK Spark Cognitive Large Model V2.0 can solve this mathematical problem in China.
The highlight of this conference is code capabilities, iFLYTEK said, from code generation, completion, error correction, interpretation and unit test generation, these dimensions can be seen, now in code generation and completion has surpassed ChatGPT, other capabilities are also catching up quickly.
Liu Qingfeng said that we have set a very clear goal, by October 24 this year, the capabilities of each code will surpass ChatGPT, and will officially benchmark GPT4 in the first half of next year.
In iFLYTEK's view, code is a hard-hitting ability, and there is no subjective scoring standard, so it should also be regarded as a very important indicator of the intelligence of large models.