Smart stuff
Author |Cheng Qian
Edit |Desert Shadow
The large model is undoubtedly the "protagonist" of the technology industry since last year, and today, the core of the competition in the large model industry has changed.
Since the scientific and technological revolution set off by the release of ChatGPT in November 2022, to the competition of hundreds of models, now with the increasingly close combination of large models and industrial landing, various applications have emerged in an endless stream, and the commercialization of large models has become the core goal of all participants. 2024 seems to have become the first year of commercialization of large models.
Previously, large models were plagued by high R&D costs, unclear landing scenarios, and high deployment costs, and the commercialization process was slow. Just yesterday, on the eve of the 2024 Lunar New Year, iFLYTEK, an important player in the AI national team and the domestic large-scale model industry track, gave its own way to break the game.
Liu Qingfeng, chairman of iFLYTEK, and Liu Cong, dean of the Research Institute, officially released the basisThe first nationally produced computing power training of iFLYTEK Xinghuo v35, iFLYTEK Spark v35. The seven core competencies have been comprehensively improvedMathematics, language comprehension, and voice interaction capabilities surpass GPT-4 Turbo
iFLYTEK releasedSpark voice modelIn terms of speech recognition, the first batch of 37 mainstream languages surpasses OpenAI Whisper V3, based on which the iFLYTEK translator that can automatically identify languages has been upgraded, and the first time it has released a deep adaptation to domestic computing poweriFLYTEK Xinghuo open source large model "Xinghuo open source-13b".It has been jointly launched in the Ascend open source community.
Since May last year, the technology accumulation and application experience with the iFLYTEK Xinghuo cognitive model as the core have become an important support for iFLYTEK in the first year of commercial application of the large model.
So, how to solve the problem that once plagued the commercialization of large models? What should the landing scene of the large model in the eyes of iFLYTEK look like? How did iFLYTEK find its own way out of its own step by step and become an industry leader? We tried to get a large model from iFLYTEK Spark v35 blockbuster upgrades set off to find the answers to these questions.
The vigorous boom of the 100-model war has gradually calmed down, but the far-reaching impact of technological change has not stopped, and the advancement of scientific and technological development is truly combined with the industry and played a role in the real application.
According to data research and analyst firm Gartner**, by 2026, more than 80% of enterprises will use generative AI's APIs (application programming interfaces), models, or deploy generative AI-enabled applications in production, up from less than 5% at the start of 2023.
However, for a long time, large models have really played a role in industry applications, and the process of landing has been very slow. This is related to the three major challenges of large model capabilities, application landing scenarios, and computing power.
The first is the challenge of model capability. Problems such as the illusion and insufficient intelligence of large models will directly affect the acceptance of users in different industries. Different from consumers, enterprises have extremely high requirements for data security and privacy, the availability of large models, and the accuracy of generated content, so in addition to strong generation and understanding capabilities, large models also need to truly solve industry problems and play a role in business.
The second is the challenge of application scenarios. The large model needs to find an effective landing scenario, and its application scope within the enterprise is very wide, and it needs to be deeply integrated with a large amount of internal data to find the real pain points of the enterprise and solve them through the capabilities of the large model. This can match the most urgent needs of enterprises with the capabilities of large models, and form rich application scenarios while further improving the capabilities of large models.
The third is the computing power challenge. This includes not only the computing power cost of enterprise customization and fine-tuning of large models, but also the independent and controllable computing power base of domestic large models.
For downstream enterprises, they do not have enough computing power to customize and fine-tune large models. Computing power is an extremely important material basis in the training and inference stage of large models, but the boom of large models has made the cost of computing power high, and it is difficult for small and medium-sized enterprises to maintain it. However, if enterprises want to deeply integrate the large model with their own business, they must customize and fine-tune the large model, which has also become a major obstacle in front of the enterprise.
Affected by the turbulent international situation, the independent and controllable localization of the large model industry is also a major problem. Only by building on an independent and controllable domestic computing power platform can the large model industry achieve sustainable and good development.
With the gradual breakthrough of these problems, the practicability of large models has reached a higher level, and the combination of its comprehensive capabilities to solve the rigid needs of the real world has been put on the agenda. In this context, iFLYTEK, which has a lot of experience in the upgrading of the core capabilities of large models, commercial landing applications, and computing platform deployment, has become the most prepared player representative for the commercialization of large models, and is breaking through to become the leader in the implementation of large models.
The vision of the large model to change the world is gradually clear, so standing in the first year of the commercialization of the large model? What reserves does iFLYTEK have? And how to lead? We can extract the layout logic of today's large model head player from the focus of this press conference, and it also represents the focus of competition for the next commercial landing of large models.
, iFLYTEK Spark v35. It has realized the upgrading of seven core competencies of text generation, language comprehension, knowledge question and answer, logical reasoning, mathematical ability, advanced ability, and multimodal ability. Among them,The language comprehension and math ability exceeds that of GPT-4 Turbo, with ** 96% of GPT-4 Turbo and 91% of GPT-4V multimodal comprehension
These powerful capabilities are already showing great potential to address real-world needs.
In the era of the Internet of Everything, the human-computer interaction mode is being reshaped, and the iFLYTEK Xinghuo app has launched a voice interaction function, which can automatically call the external capabilities of the large model in the dialogue with people to achieve the natural interaction of full voice. Liu Qingfeng said that from the DOS interface to the Windows interface has achieved the legend of Microsoft, from the keyboard to the touch with the myth of Apple, this timeThe natural interaction of full voice will drive a new boom in the entire industry
Voice technology has always been iFLYTEK's strength, and since its inception, the company has set a vision to make communication between humans and humans and machines barrier-free. In the era of general artificial intelligence, iFLYTEK continues to maintain its leading edge in voice interaction, and sees more possibilities under the wave of new technologies. Large models can help the training of corpus of small languages and promote the development of speech technology through more unified multi-task modeling capabilities.
The iFLYTEK Xinghuo speech model is pre-trained based on the decoupling representation of speech attributes, combined with the conventional speech model architectureThe speech recognition performance of the first batch of 37 mainstream languages surpassed that of OpenAI Whisper V3In terms of multilingual speech synthesis and super-anthropomorphic speech synthesis, MOS has an absolute advantage, and MOS refers to whether the generated voice is natural.
The iFLYTEK translator equipped with the Xinghuo voice model has also achieved a major upgrade, which can realize multilingual independent recognition, and can automatically identify the speaker's language and translate it into Chinese without the user's independent choice.
In addition, in order to enrich the application ecology of large models, iFLYTEK has released13 billion parameter scale of the Spark open source large model series, including base models, fine-tuning models, fine-tuning tools, and customization tools. Xinghuo Open Source-13b ranks high in typical scenarios such as text generation and language understanding in a number of well-known public evaluation tasks.
Finally, there is a solid foundation for large model training - the computing platform, iFLYTEK Spark v35. The Xinghuo voice model and the Xinghuo open source model are all based on the "Feixing No. 1" trainingFeixing-1 is the first domestic computing platform to support trillion-parameter large model training that was officially launched by Xunfei on October 24 last year.
, iFLYTEK Spark v35 is domesticThe first large model trained based on national computing power。The Xinghuo open source model is also based on Feixing-1 to achieve full-stack domestic adaptation optimization, and the training efficiency is 90% of that of A100. This also means that iFLYTEK provides enterprise customers with another choice of "large model + computing power".
At this special node, it is very important for the large model to be based on a domestic independent and controllable computing platform, Liu Qingfeng said, iFLYTEK Xinghuo 35. It is an important test for whether the domestic computing power platform can support the research and development of large models in the future.
It can be seen that iFLYTEK is very clear about what it wants to do and how to do it in the wave of the large model industry, and in the previous deep industry accumulation, it really aims at the pain points of the industry, and knows how to take root in this industry and achieve leadership.
Looking at the development of the entire large-scale model industry, today's commercialization battle is not only a global competition in science and technology, but also a key link in the development of domestic generative AI and the global same frequency.
Since August last year, a total of 4 batches of domestic large models have been opened to the public through the filing of large models, and now there have been many large model applications in finance, education, and office tracks.
As the first batch of iFLYTEK Xinghuo large models that have passed the record, the application progress is not inferior. Since May last year, while the seven core capabilities of the Xinghuo model have been continuously upgraded, the hardware has created an iFLYTEK AI learning machine in the field of education, and the office field has iFLYTEK intelligent office notebook, iFLYTEK voice recorder, software has iFLYTEK hearing, iFLYTEK Xinghuo APP, iFLYTEK input method, etc., as well as content creation tools, such as the audio creation tool "iFLYTEK Zhizuo", the creation tool "Spark Content Operation Master", etc., have gradually released the value of the large model to a large number of users.
On the iFLYTEK open platform, the total number of large model developers exceeds 350,000, including more than 220,000 enterprise developers.
The experience and feedback of a large number of users are also feeding back the continuous improvement of the core capabilities of the large model.
At the same time, Liu Qingfeng said that the large model is no longer used to simply write poems and paintings, but to empower scientific research, industry, and people's livelihoodSo that the large model can be upgraded into a new productivity in the digital era and the intelligent era
The distance from scientific and technological innovation to industrial landing requires the combination of end enterprise users and core large model players to gradually shorten the distance, and at the same time make the path of commercial landing of large models clearer.
iFLYTEK has accumulated a lot of commercialization in different tracks. For the education industry, iFLYTEK launchedSpark Smart BlackboardIt has four major functions: multimodal understanding and recommendation, all-natural interaction, virtual human auxiliary learning, and intelligent lesson recording and sharing. This function further expands the value boundary of the blackboard and becomes an AI assistant for teachers.
At the same time, iFLYTEK and China Mobile jointly launched the innovative 5G call application "Business Shorthand", which can synchronize the voice of the minutes and refine key to-do items during the user's call.
The intelligent voice interaction technology applied in Chery Automobile, a leading player in the automotive industry, is provided by iFLYTEK, and its export countries cover dozens of languages such as English, Russian, Spanish, Arabic, and Portuguese. It can be seen,Supporting China's automobiles to go overseas is also a potential scenario for the commercialization of domestic large-scale models.
It can be seen from thisLarge models are simultaneously driving the commercialization of B-end and C-endOn the one hand, capabilities such as translators, business 5G calls, and AI PPT are being reshaped by large models, making cutting-edge technological innovation a productivity tool for individual users. On the other hand, leading players and start-ups from all walks of life are exploring commercialization with the core players of large models represented by iFLYTEK, so as to find new growth opportunities while accelerating industrial transformation and upgrading.
More importantly, iFLYTEK, as the national AI team, has a natural advantage in providing an independent and controllable national computing power platform while accelerating the commercialization of domestic large models.
Today, the commercial application of large models has gathered the strength of all players such as computing power, large models, and terminal enterprises, so that the core capabilities of large models and the progress of application implementation are promoted simultaneously, and they are safe and controllable.
Even though there has been a large gap between China and foreign countries in terms of the core capabilities of large models, it has entered a new stage of commercialization, relying on the rich application scenarios and landing soil in China, players represented by iFLYTEK are leading the transformation of the new era of large models.
The powerful capabilities of large models in terms of generation and understanding have enabled AI to continue to expand its application boundaries in thousands of industries. At present, various AI-driven applications are transforming people's work, life, and learning.
However, from the perspective of the commercial application of large models, as mentioned above, compared with C-end consumers, the specific characteristics of model capabilities, application scenarios, and computing power need to be considered in the ability of enterprise business integration large models. This is also the top priority of the current iterative upgrading of the large-scale model industry and its advancement to commercialization.
The confrontation between large model players has not stopped, from the computing power and parameters of the 100-model battle, today's large models have become more and more practical. Behind this, it is inseparable from the in-depth understanding of cutting-edge technologies and the courage to explore and try by business-side enterprises, and it is also inseparable from the continuous breakthrough and firm investment of enterprises with core technologies.
Many players such as domestic large-scale model core players, enterprises, and computing power providers have been involved in the new wave of large-scale models, and have become important participants in the construction of large-scale commercial application ecology, further making up for the gap between the domestic large-scale model industry and foreign levels.
In the future, large models will play a key role in the process of enhancing global competitiveness as a necessity, and a series of first-mover advantages accumulated by iFLYTEK in the AI industry for more than 20 years will become an important support for it to be one step ahead of others in the key aspects of large model competition.