Written by Shang Di'an.
The wave of change set off by AIGC is beginning to affect more industries as the development direction of the industry in 2024 gradually becomes clear.
Compared with the "savage growth" of the large model industry in 2023, although the entire market size will usher in a huge increase again in 2024, the trend change is obvious: the industry needs more application-side integration that can bring efficient changes to users. This is not only the trend of the times, but also a rigorous test of the basic ability of large models.
The way to attract attention will be withdrawn from the mainstream, giving way to more excellent product innovation. This is also reflected in the 2024 industry trend so far: whether it is for the industry or the C-side, large model manufacturers seem to be looking for ways to bring the latest AI capabilities closer to the user.
To sum up, in the era of AI, 2024 can be regarded as the real first year of the commercial landing of large models, and the iFLYTEK Xinghuo Cognitive Large Model V3. will be held on January 305. The upgrade conference can be seen as iFLYTEK's answer to the prospects of China's generative large model industry in 2024.
What kind of large model does iFLYTEK want to make?
As one of the popular enterprises in the field of generative large models in China, the question of what kind of large model does iFLYTEK want to make has a value comparable to that of the industry's investment vane, so it is in the Spark model 35 After the release, many people in the industry tried to find answers from the many contents of this press conference.
From the surface to the inside, this question can be roughly divided into two answers.
First of all, the most intuitive reason - "benchmarking ChatGPT" has always been one of the goals of the Spark model. In the past year, the three major iterations of the Spark model have all been indispensable to this strong competitor.
When GPT-4 was released, the goal of the Spark model changed as well: iFLYTEK laid out a clear timeline for it. announced that it will benchmark GPT-4 in the first half of 2024, and finally complete its military order in stages at the end of January: According to Liu Qingfeng, chairman of iFLYTEK, in seven aspects: language understanding, text generation, knowledge question and answer, logical reasoning, mathematical ability, ** ability and multimodal ability, based on the national computing power training of Xinghuo V35 Also fully upgraded. 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.
From this dimension, the iterative upgrade of the Xinghuo large model is also like a microcosm of the development of China's independent large model capabilities: catching up with the international advanced level in rapid iteration, accelerating integration in the application field, launching various software and hardware products, and exploring the possibility of landing more large models as much as possible at this stage, giving more meaning to the industry.
This kind of competition is more like two troikas going hand in hand, which in turn intuitively reflects the rapid improvement of the capabilities of the Xinghuo large model: in addition to the more advantageous industry application scenarios of the Chinese large model, it also has stronger multimodal capabilities, which can attract developers more and build a more valuable self-owned ecosystem.
*Auxiliary generation ability has always been one of the typical scenarios to test whether large models can be quickly applied in professional fields, and is known as a sign to judge the intelligence of cognitive large models; The Spark model has been built with the ability to generate ** since the first generation, and has been upgraded with multiple iterations of major versions. Now it has complete completion, error correction and unit testing capabilities. And in the ** capability benchmark built by OpenAI, ** reached 96% of GPT-4 Turbo.
For iFLYTEK, the development of intelligent voice interaction software and hardware at the beginning of the business is more like a large model that iFLYTEK wants to make5 In the press conference, iFLYTEK also introduced the new achievements in the "old bank".
According to Liu Qingfeng, CEO of iFLYTEK, making machines have the ability to learn, reason and make decisions is the direction of change that the Spark model is currently focusing on in the field of translation.
The newly released "Xinghuo Speech Model" is based on the large language model framework and pre-trained by combining the decoupling representation of multi-dimensional speech attributes such as iFLYTEK language, timbre, and content. It can be multilingual and achieve hyper-anthropomorphic speech synthesis effects. The average MOS score for the first 40 languages (a criterion for evaluating audio or ** quality, with 5 being the highest) increased by 025, MOS reached 4 in the anthropomorphic test5 points, the degree of anthropomorphism reaches 83%, and the anthropomorphic speech synthesis ability surpasses ChatGPT.
This kind of voice capability can also be applied to C-end hardware more quickly: Liu Qingfeng also released the iFLYTEK translator equipped with the Xinghuo voice model, announcing that it will soon launch two important functions, multilingual automatic recognition and enhanced translation, which will be upgraded at the end of January and mid-March this year, respectively. According to reports, the multilingual automatic recognition upgrade of the iFLYTEK translator will support 35 languages. Multilingual automatic recognition makes international communication more convenient, and augmented translation technology turns the translator into an AI translation assistant.
To put it simply, with the help of a large model, we make a speech have richer attributes, including language, content, rhyme, timbre, and emotion. ”
What kind of big model does the industry need?
As for another answer to what kind of large model iFLYTEK wants to make, the deep reason can actually be answered from another dimension: what kind of large model the industry needs.
First of all, it is the urgent need for independent computing power in the large model industry: for large models that rely heavily on rapid learning iteration to maintain competitiveness, hardware is not a simple replacement of equipment: there is a metaphor about the image of hardware computing power in the industry: just like plants and soil, the difficulty of trying to switch hardware is comparable to uprooting a plant from the soil where it originally grew and recultivating it into another new environment. This not only shows the importance of independent computing hardware platform, but also has a deeper understanding of the huge impact of independent computing power on ecological construction.
The concept of "computing power base" derived from this is also widely known as the situation of computing power being stuck in the neck frequently appears in 2023. It has gradually become an industry consensus.
Nowadays, even if foreign companies are willing to circumvent export controls and sell customized products to Chinese customers, Chinese customers will give priority to using more independent and controllable domestic computing power in new projects. At the same time, the investment in independent computing power is also another embodiment of building an ecosystem: according to the calculation of the China Academy of Information and Communications Technology, every 1 yuan invested in computing power will drive 3-4 yuan of GDP growth.
In the struggle of the industry in 2023, the independently available computing power has gradually been widely recognized, and the call is getting louder and louder, becoming one of the most important trends of the times from the East in this large-scale model landing competition: In October 23, the Ministry of Industry and Information Technology, the Cyberspace Administration of China, the Ministry of Education and other six departments issued the "Action Plan for the High-quality Development of Computing Infrastructure", and proposed that by 2025, the scale of computing power will exceed 300eflops, and independent computing power will undoubtedly occupy an important place in it.
In fact, independent computing power has long been something that many domestic AI giants and even startups are promoting, but the subdivision direction is very different, and the actual embodiment in the application scenario is also different. The domestic computing platform "Feixing No. 1", jointly built by iFLYTEK and Huawei, was completely based on the Spark model 3 trained on this domestic platform 90 days after it was officially launched last year5 also appeared on stage at the same time.
In addition to surpassing GPT-4 in language comprehension and mathematical comprehension capabilities, the Spark model trained by the national computing power is of more important significance to the industry in the construction of China's developer ecosystem.
The open source model is a key measure to establish an ecosystem, and the Spark model 3At the same time as version 5, iFLYTEK also released a version based on Spark 1The 13 billion parameter open-source model from version 0 is not only systematically designed for the field of domestic data security, but also takes into account the adaptation of domestic computing power from the underlying architecture.
According to the test results officially released by iFLYTEK, the Xinghuo open source large model benefits from the performance improvement of the national computing platform, and the effect is more than 20% higher than other open source models of the same size in typical application scenarios.
These performance improvements will be the basis for the large model to further take root in more application scenarios: better emotional perception and anthropomorphic ability are the most intuitive embodiment of making the large model directly "smart as a human"; In common content generation scenarios such as generating PPT and debriefing reports, the AI writing assistant can also generate richer content in the Chinese environment with more intelligent understanding capabilities and the expansion of external knowledge.
In addition to these visible achievements, an inconspicuous detail is that the Spark open source model has chosen the open source community as the platform for the first launch, which is undoubtedly a strong signal for developers: iFLYTEK is embracing the open source community ecology and more third-party developers with the help of the open source model.
This is more in line with the attitude of the industry leader: with enough strength to drive more industries and scenarios, develop more products with the help of open source large models, and then feed back the ecology itself, and build a rich and usable large model ecology.
Epilogue. After watching the Starfire model v35 After the press conference, in addition to having a clearer understanding of iFLYTEK's current large-scale model capabilities, it may be more important for industry professionals to find the answer to the question of what kind of large-scale model the industry needs in 2024.
It is true that Liu Qingfeng, chairman of iFLYTEK, is in the Spark model V35 As stated in the press conference, China's large model is still far from the best level of GPT-4. However, with the help of large models to promote free communication and the Internet of Everything, and at the same time build an independent and open large model ecology, it is not only the current direction of iFLYTEK, but also the future direction of the entire industry.
Through this press conference, we look forward to a spring full of hope and growth energy. I believe that in 2024, we will be able to achieve a spark of fire, and general artificial intelligence will not only be able to be deeply and widely used in China's major fields, but also we will stand on a new level in source technology innovation and the underlying capabilities of large models. Liu Qingfeng said.