Written by Shang Di'an
With the gradual promotion of large models in the application scenarios of various industries, more than half of 2024, "needs and problems emerge at the same time" has become the current status quo that China's large models need to face.
The main problem in the application of large models stems from the famous "impossible triangle" theory. With the gradual deepening of the application of large models in 2023, various large manufacturers have successively found that it is often difficult for large models to obtain a large amount of expertise in a specific field, and obtaining a large amount of professional knowledge and conducting inference training requires a large amount of computing resources, which in turn leads to a geometric increase in the use of large models.
At the same time, natural language processing neural network models based on the Generative Pre-trained Transformer architecture, such as ChatGPT, have the characteristics of "all-encompassing", which in principle determines that this is at the expense of the accuracy of content in specific domains.
It is difficult to strike a balance between professionalism, economy and versatility, which is the "impossible triangle" that has plagued the in-depth application of the large model industry.
Starting from the second half of 2023, solutions for leading enterprises in the industry and solutions to the needs of enterprise models have sprung up, and for digital twins, smart cities, intelligent driving and other fields that have existed for a long time, a large number of practical applications of domestic large models can be seen, and rapid iteration from them.
From the official launch of the iFLYTEK Spark large model in May 2023 to the upgraded iFLYTEK Spark V3.5Max version at the end of May this year, it has surpassed GPT-4 Turbo in text generation, logical reasoning and math/code capabilities, and then to the V4.0 version The seven core capabilities have been comprehensively improved, fully benchmarking GPT-4 Turbo. With the support of a large number of industry data, the ability of domestic large-scale models has also improved by leaps and bounds.
In the first half of 2023, iFLYTEK and Huawei jointly created China's first large-scale domestic computing platform "Feixing No. 1", artificial intelligence+ has become a national strategy.
The term "domestic substitution" is not even accurate: if the domestic model has a high localization rate and strong local development characteristics, these characteristics are also difficult for outsiders to replace.
However, with the "testing of the waters" in the past year, many problems have also been exposed in the application of large models: the most common problem is the combination of models and applications, and on the basis of the capabilities of the large models themselves, it is necessary to deeply understand the needs of the industry. This also largely determines the user portrait in the early days of the industry.
According to Yao Qian, director of the Science and Technology Development Department of the China Securities Regulatory Commission, most of the domestic large models have taken into account the needs of privatization deployment in the early stage of development, so that domestic large models will show a blooming scene in 2024.
In the automotive field, manufacturers including FAW, Chery, GAC, Great Wall and other manufacturers have chosen the Xinghuo car intelligent cockpit, which has the highest domestic market share of voice interaction, as the solution for the next-generation intelligent car cockpit experience.
In the field of the Internet of Things, iFLYTEK Xinghuo has established cooperation with Haier, Midea and other leading enterprises, and released smart home products based on the interactive mode of iFLYTEK Xinghuo; These are all solutions based on the integration of cloud and edge, as well as software and hardware created by iFLYTEK.
On June 27th, at the release site of iFLYTEK Xinghuo large model V4.0, iFLYTEK demonstrated its breakthrough in voice transcription capabilities in multi-person mixed scenes, three iFLYTEK researchers spoke at the same time on the spot, and it was even difficult for normal human ears to distinguish, but with the help of multi-modal voice recognition technology, iFLYTEK heard that it was able to accurately realize the speech separation of the three people from it, and transcribed the corresponding shorthand content in real time.
Among them, relatively traditional fields such as manufacturing and transportation, due to the late start of large-scale model application, often lack standardized digital business experience while urgently needing digital transformation.
However, these users still have a lot of demand in the early trial stage: the "API price war of large model manufacturers" that broke out at the beginning of this year is largely a reflection of this demand: users and developers have a huge demand for using and testing large models.
The opportunity also includes a number of never-before-seen challenges: most of these users with a thirst for large models have not been exposed to the concept of incorporating large models into their workflows before.
Even though many potential users are clear about the potential of large models for their own business, they are not clear about how this capability should be combined with existing businesses.
These problems are not only a test of the model capability itself, but also put forward higher requirements for the technical integration ability of the manufacturers behind the model. Users hope that the industry model can bring real changes, and with the change, it is possible to understand the boundaries of large model capabilities more scientifically.
"Willing to pay the cost means that you must see the return" The answer of an interviewed smart home product manager can represent the general attitude of potential users of the current large-scale model industry application: if there is no exponential growth compared with before, even if the price of the model itself has been rolled to the extreme, it is difficult to attract users to pay.
But in other words, after breaking through the technical difficulties, it can bring solutions to the digital growth and precise reach needs of users in traditional industries, and it will naturally attract users in traditional industries: since its release in May last year, the iFLYTEK Xinghuo model has become the first choice for leading enterprises in traditional industries and fields, including China Energy Group, PetroChina, China Mobile, Bank of Communications and other fields.
"These are the results of careful selection by customers and PK by each company," said Liu Qingfeng of iFLYTEK when introducing the user portrait of the iFLYTEK Xinghuo large model. These enterprises distributed in various industries and have a large number of high-quality users in their respective fields have accumulated experience in the implementation of large-scale model solutions through their own practice while becoming iFLYTEK Xinghuo users.
Even if the moment of "big model changes the world" has not yet come, in this process, the growth of large model will gradually change from meeting everyone's needs to meeting the unique needs of each user, this problem seems to be contradictory, but it is the essence of the problem that human beings have the most potential to solve in the current era of large models:
In the final analysis, the competition for the landing of large models is the competition of industry solutions and service capabilities.
In terms of answering this question, iFLYTEK not only has a model of world-class capabilities, but also has practical experience in the development of C-end software/hardware products accumulated in the past ten years of software and hardware layout, as well as millions of intelligent hardware users. These first-mover advantages make iFLYTEK almost one step ahead of others in the application of large models.
At the same time as the release of iFLYTEK Xinghuo V4.0, iFLYTEK also released a hardware including the Xinghuo intelligent review machine, which can support multi-disciplinary and multi-question correction while generating multi-dimensional analysis tables in real time. With its own actions, it has made the first demonstration of the landing of the large model in the field of education. Let iFLYTEK Xinghuo have the title of "the largest model that can do the most problematic".
This also reflects the most real needs of the large model at the moment, users need products that can be quickly used out of the box in every field, and manufacturers with enough experience to copy the "most driving large model" and "most capable of seeing a doctor" in different fields.
In particular, iFLYTEK, which has adhered to the national production strategy of large models since its establishment, has more opportunities to create various consumer large model products that are unique to Chinese users at a time when Chinese and American large models are gradually exploring two very different development paths.
After the initial hustle and bustle gradually settled, helping the large model to land and get through the "last mile" has become the solution and the key to survival, and those large models that are difficult to survive in the actual application scenario will eventually be eliminated.
On the basis of the capabilities of large models, only large models that have accurate insight into the landing positioning of large models in user needs and bring changes to the existing experience can have the opportunity to become the choice of more users. In this process, the current shortcomings of the large model are gradually improved, forming a positive cycle.