List of high-quality authors
Reporter丨Ma Yongxin Trainee丨Han Qing
Produced by丨Aotou Finance (thesankei).
It has been more than a year since ChatGPT was released, and the AI large model trend has swept the world. In the past year, a number of domestic companies have released large AI models, and the domestic demand for AI computing power has become more and more huge.
Computing power is the core foundation of large models and an important engine for the digital economy. At the beginning of October last year, six departments, including the Ministry of Industry and Information Technology, the Cyberspace Administration of China, and the Ministry of Education, issued the "Action Plan for the High-Quality Development of Computing Infrastructure", proposing that by 2025, the scale of computing power will exceed 300eflops, and the proportion of intelligent computing power will reach 35%.
With such an important infrastructure, domestic enterprises want to firmly grasp it in their own hands, iFLYTEK (002230SZ) went ahead. At the 1024 Developer Festival in October last year, iFLYTEK and Huawei jointly announced the launch of the "Feixing No. 1" platform, the first purely localized large-scale model computing base in China, and announced the start of large-scale model training against GPT-4.
How effective is the Flying Star One? There was an answer in 100 days. On January 30, the Spark v35 was released, showing the powerful ability of pure domestic computing power to the outside world.
Why do you want to make a domestic large-scale model computing power base?
General artificial intelligence can not only be deeply and widely used in various fields, but also we will stand on a new level in source technology innovation and the underlying capabilities of large models. Liu Qingfeng, chairman of iFLYTEK, explained why we want to make a computing base.
The first is to solve the risk of "stuck". The development of large models requires not only massive data, but also huge intelligent computing power, which requires computing power chips. "GPU chips are the core of increasing AI computing power, and now the GPU chips of domestic enterprises are mainly NVIDIA. Industry insiders told Aotou Finance and Economics, "Now facing two problems, one is limited by the US ban, Chinese companies cannot obtain the most advanced GPU chips, and the other is that Nvidia's GPU chips are expensive, between 1 and 20,000 US dollars per chip, and the cost is huge." ”
The first purpose of making a computing power base is to "carry out technological innovation at the source" and take the road of independent innovation in the infrastructure for the development of AI large models. Huawei has suffered from "stuck necks", and iFLYTEK is a leader in domestic AI companies, with rich technology accumulation and experience.
Aotou Finance has learned that the independent and controllable Feixing-1 built based on Huawei's Ascend ecosystem can be used to train large models with trillions of floating-point parameters as early as October last year. Earlier, iFLYTEK and Huawei have increased the capability of the Ascend 910B from 20% to 90% compared to the NVIDIA A100. "In today's era of large models, we must not only use open source models for application landing, and there must be a national team standing up on the base of the general large model. Liu Qingfeng emphasized.
The four words "independent and controllable" are commendable, and another reason for the development of computing power base is also this. The digital upgrading of the industry requires the quality and efficiency of large models, and it is of great significance for the industry to embark on a road of independent innovation and provide a second option for the development of large models.
Independent innovation not only refers to the localization of the computing power base by iFLYTEK and Huawei, but also provides the soil for independent innovation of the large model ecology for Feixing-1.
At the press conference on January 30, the open source model of Feixing-1, Xinghuo Open Source-13B, was officially released. According to reports, Xinghuo Open Source-13B ranks among the best in a number of well-known public evaluation tasks, and in scenarios such as text generation, language understanding, text rewriting, industry Q&A, machine translation, etc., through in-depth research and optimization in the fields of learning assistance and language understanding, the practicability has been greatly improved, and it is more handy in dealing with complex natural language tasks.
Only by allowing more scenarios to be implemented in open source can ecological cooperation be better enhanced. Up to now, on the iFLYTEK open platform, the total number of large model developers has exceeded 350,000, including more than 220,000 enterprise developers, the number of developers has grown rapidly, and the application of large models has been accelerated, continuously consolidating the first developer ecology.
Starfire v35 strong applications
To what extent can the Feixing-1 model go? In a word: its capabilities are already comparable to GPT-4.
Aotou Finance learned that the Xunfei Xinghuo v35. It has been comprehensively upgraded in terms of language comprehension, text generation, knowledge question and answer, logical reasoning, mathematical ability, ability and multimodal ability. Among them, the language comprehension and mathematics ability exceeds that of GPT-4 Turbo, ** reaching 96% of GPT-4 Turbo, and multimodal understanding reaching 91% of GPT-4V.
In terms of intelligence, iFLYTEK Spark V35. It can already become a replacement for "workers". iFLYTEK Spark v35 iFLYTEK Zhiwen has the functions of one-click document generation, AI writing assistant, multilingual document generation, AI automatic picture matching, a variety of template selection, and speech notes. In the display session at the press conference, iFLYTEK Zhiwen produced a PPT of more than 20 pages of rich content "Hefei 2024 Spring Festival Tourism Promotion Strategy" in a short period of time, which received bursts of applause.
Since the beginning of its release in May last year, the Xinghuo large model has been developing in the direction of strong application, whether it is the iFLYTEK AI learning machine, the Xinghuo Companion APP, or the Xinghuo smart blackboard and other products launched after this iteration, iFLYTEK has always put the application in the first place in technological development, which is the result of technological innovation orientation. Industry analysts said.
Starfire v35. The improvement of technical capabilities is the empowerment of AI in scientific research, education, industry, medical and other fields. Taking intelligent education as an example, the Xinghuo smart blackboard launched by iFLYTEK has functions such as multi-modal understanding and recommendation, all-natural interaction, virtual human assisted learning, and intelligent lesson recording and sharing.
In the demonstration session, Albert Einstein appeared on the Spark Wisdom Blackboard and used the "vacuum cleaner with huge suction" as a metaphor for black holes, explaining "what a black hole is" to the audience in simple terms. In the teaching of three-dimensional geometry and other knowledge, the Xinghuo wisdom blackboard visualizes the boring theory by deconstructing the cube, and the teaching classroom is more vivid and interesting.
Smart Blackboard is just Spark v35. On the one hand of application scenarios, with the continuous upgrading of the Spark model, the application will be expanded in more and more scenarios and fields.
The AI interaction method has been upgraded again
In 2023, with the rapid development of AI large models, all companies are making efforts to equip large models with voice assistants, so that voice assistants that can only set alarm clocks, tell jokes, and broadcast ** in the past have become more and more "humane".
On January 30th, iFLYTEK released a large voice model for the first time, according to reports, the Xinghuo voice model is internationally leading, Chinese, English, French, Russian and other 37 mainstream languages The speech recognition effect exceeds OpenAI Whisper V3, in terms of multilingual speech synthesis, the 40 languages of the Xinghuo speech model are more than 83% anthropomorphic.
From "large language model" to "large speech model", it has brought another upgrade of AI interaction mode, and also brought about a revolution in production efficiency and lifestyle. Liu Qingfeng introduced that in scenarios such as automobiles, customer service, families, and companion robots, there are more uses for the Xinghuo voice model. For example, empowering automobiles, the interactive experience of intelligent cockpit, intelligent cockpit, intelligent navigation, and first-class control will be further optimized; Industries such as companion robots, shopping guide robots, auxiliary diagnosis robots, smart homes, and wearable devices will be further promoted.
Chery Automobile has enjoyed the dividends brought by intelligent voice. Aotou Finance has learned that Chery Automobile exports to dozens of languages such as English, Russian, Spanish, Arabic, Portuguese, etc., and the intelligent voice interaction technology applied in the car is all provided by iFLYTEK.
The comprehensive empowerment of iFLYTEK's Xinghuo model will undoubtedly provide strong technical support for the digital transformation of various industries and lead a new wave of digitalization.
Although China's AI model has made a lot of achievements, but at the same time, it should also be seen that there is a gap between the domestic model and GPT-4 in some aspects, in Liu Qingfeng's words, "At present, the development of general artificial intelligence is something we must do, otherwise industry, scientific research, people's livelihood and other special fields will be greatly behind the world, and China is the only country in the world that is expected to become the second pole of wisdom." ”
The fact that the application of AI large models is everywhere and still maintaining a clear self-awareness gives people reason to believe that iFLYTEK will make more amazing achievements in the field of AI.