Reporter Chen Jialan reports from Guangzhou.
Recently, U.S. Secretary of Commerce Gina Raimondo announced plans to restrict foreign customers, especially Chinese customers, from using the services of U.S. cloud computing vendors to train AI large models, and the U.S. Department of Commerce announced a proposal to require U.S. IaaS (cloud service) vendors to provide cloud services, including AI model training, to overseas companies and individuals. Some analysts believe that this measure has pushed the US sanctions on China's science and technology to a new height.
On the one hand, the U.S. sanctions on China technology have pushed to the height of AI computing power to limit cloud services, and on the other hand, the localization of domestic manufacturers' independent and controllable technologies is accelerating.
On January 30, the reporter of "China Business Daily" learned from iFLYTEK that the first achievement of the national computing power platform "Feixing No. 1" was iFLYTEK Xinghuo V35 was released, which is the first national open large model based on the training of national computing power. In addition, the iFLYTEK Xinghuo open source model "Xinghuo Open Source-13B", which is deeply adapted to domestic computing power, was launched for the first time, and the Ascend Open Source community jointly launched the first launch.
On October 24, 2023, Liu Qingfeng, chairman of iFLYTEK, proposed that the iFLYTEK Xinghuo large model should be benchmarked against ChatGPT4 in the first half of 2024.
At that time, Liu Qingfeng introduced, iFLYTEK Spark V35's language comprehension and math ability have surpassed GPT-4 Turbo, ** reaching 96% of GPT-4 Turbo, and multimodal understanding reaching 91% of GPT-4V. However, iFLYTEK is confident that it will catch up with GPT-4's current best level in the first half of 2024.
Domestic manufacturers have accelerated their foothold on localization, independent and controllable
Recently, the news that the United States restricts Chinese AI large model manufacturers from using overseas computing power through American cloud service providers has attracted market attention.
On October 7, 2022, the United States announced the computing power and bandwidth limits for Chinese artificial intelligence training, and from that time on, China can only use the 'castrated version' of chips for training large models. On October 17, 2023, the U.S. Department of Commerce further promulgated regulations that China cannot even buy the 'castrated version', and imposed strict restrictions on neighboring countries and regions that may cooperate with China. Just this month, the U.S. Department of Commerce further proposed that U.S. companies will not be able to provide computing power to Chinese artificial intelligence enterprises. "On January 30, Xunfei Spark v35 At the press conference, Liu Qingfeng did not shy away from talking about the recent plan of the United States to ban cloud computing vendors from training AI large models for China, "Whether it is training or inference, our computing power restrictions are getting tighter and tighter. On the first day of training the iFLYTEK Xinghuo cognitive model, iFLYTEK was determined to build an independent and controllable computing platform based on localization. ”
In order to seize the development initiative in the era of general artificial intelligence, iFLYTEK laid out the domestic computing base earlier. In the first half of 2023, iFLYTEK and Huawei established a "joint special team". In October, iFLYTEK announced that it would work with Huawei to build the Feixing-1 platform, an independent and controllable large-scale model computing base based on the Ascend ecosystem. In this process, iFLYTEK has carried out hundreds of optimizations based on hardware tools and operator libraries, carried out a large number of bug revisions and operator efficiency optimization, so that the "Feixing-1" computing power cluster can train ultra-large-scale general artificial intelligence models.
On January 30, the first national open large model iFLYTEK Xinghuo V35 appearances. At the same time, iFLYTEK's first open-source model based on the national computing power platform "Feixing No. 1" - Xinghuo Open Source-13B was officially released.
Liu Qingfeng said that iFLYTEK officially released iFLYTEK Spark V35. It is not only a demonstration of iFLYTEK's core technology and innovation capabilities in the field of general artificial intelligence, but also an important test of whether the domestic computing platform can support the research and development of large models in the future.
Although domestic chips and computing power are subject to more and more restrictions, it is also seen that it is forcing all parties to increase investment in domestic computing power, and the current market can also see that domestic computing power is progressing rapidly.
Since 2023, many domestic cloud vendors, such as Alibaba Cloud, Huawei Cloud, and Tencent Cloud, have released their self-developed AI large model platforms.
In the view of Yan Yang, deputy secretary-general of the Zhongguancun Big Data Industry Alliance, with the restriction of external high-performance chips and cloud services, the demand for local computing resources in the domestic market may increase.
Liu Xuefeng, chief analyst of GF ** Computer, believes that the determination of the independent and controllable construction of the domestic AI industry chain is expected to be further strengthened, the development and application of domestic AI large models will be greatly strengthened to use domestic AI computing power and model services, and the process of independent and controllable construction of domestic AI computing power is expected to accelerate.
Strive to catch up with GPT-4 in the first half of this year
It is the same as the previous generation of the company, GPT-35 iFLYTEK Spark v30 released for nearly 100 days, iFLYTEK Spark v35. Once again, we ushered in multi-dimensional capability improvement.
Not only did it meet the expectations of the time, but it also surprised in many ways. Liu Qingfeng introduced, Xunfei Xinghuo V35. It has been upgraded in seven aspects: language comprehension, text generation, knowledge question and answer, logical reasoning, mathematical ability, advanced ability and multimodal ability. Upgrade the Spark Smart Blackboard; The first batch of 37 mainstream languages surpassed OpenAI Whisper V3, enabling the new upgrade of iFLYTEK translators, helping to communicate more freely, and promoting the transformation of human-computer interaction in scenarios such as customer service, automobiles, and robots in the era of the Internet of Everything.
According to iFLYTEK, iFLYTEK Spark V35's abilities have become more and more close to the GPT-4 Turbo level, and in terms of language comprehension and mathematical ability, it has surpassed GPT-4 Turbo, and has reached 96% of GPT-4 Turbo in **, and has also reached 91% of GPT-4V in multimodal understanding, of which the multimodal ability of speech has exceeded GPT-4.
Liu Qingfeng also said frankly that we must see the gap soberly and rationally, and there is still a gap between the best level of GPT-4 in the fields of small-sample rapid training, multimodal deep learning training, and ultra-complex deep understanding, but iFLYTEK Xinghuo is confident that it will catch up with the current best level of GPT-4 in the first half of 2024.
It is worth noting that OpenAI plans to launch GPT-5 in 2024, and OpenAI may have completed GPT5 training. Professor Wang Yanfeng, a representative of the 16th Shanghai Municipal People's Congress, pointed out in his submission of the "Proposal on Accelerating the Construction of a Large Model Innovation Cluster with Global Influence" that although China's science and technology enterprises have invested heavily in the field of large models, there is still a big gap between them and "world-class" products, and they have not yet shown a corresponding leading edge in global competition. "In terms of the number of inputs and outputs in the field of large models, China has equaled or even surpassed the United States, but in terms of the practical application of large models and the transformation of industrial value, the United States still maintains a significant lead in global competition, and it is still a few top American technology companies such as OpenAI, Google, and Meta that have truly reached the world-class level, which can be clearly seen from this CES. Wang Yanfeng said.
Looking forward to the development of the iFLYTEK Xinghuo large model in 2024, Liu Qingfeng pointed out three points: first, we must continue to benchmark the international state-of-the-art level in the underlying capabilities of the general large model, and make relatively better results from algorithm research, including smaller computing power. Today's general model does not necessarily represent the entire future of artificial intelligence, and there are still many innovations to be done, such as brain science interaction, deep connection of adversarial networks, etc., which require the entire innovation ecosystem, but we must have the courage and expectation to be at the forefront.
Second, in 2024, we must truly make the quantity and quality of large models fly together, not only for industry applications, but also for many key technological innovations to link large models, and the Chinese business and scientific communities are confident to achieve transcendence. Liu Qingfeng said, "Third, high-rise buildings should be built on safe and controllable platforms, and we must achieve ecological prosperity on independent and controllable platforms." I am confident that I can build a fully autonomous and controllable prosperous artificial intelligence ecosystem from algorithms, data, application scenarios to computing power. ”
In the face of the frequent policies of the United States to restrict the export and use of computing power, Liu Xuefeng also reminded that due to the limited computing resources available, the process of development and application of China's AI large models may still be affected, and it will take time for the gap in domestic AI basic computing power to narrow, especially in large-scale cluster training and ecological construction, and there is still uncertainty about the changes in the restrictions on AI computing power and model services in the United States.
Editor: Wu Qing Proofreader: Liu Jun).