Text | hickory
The story of the big model in 2024 is still very lively.
Beyond the ocean, from the emergence of Sora to the "open source vs closed source" controversy set off again by Llama 3, to all kinds of AI hardware that integrates large model capabilities...... On the journey to AGI, in order to make the big model land, overseas technology giants rely on scaling laws (the law of scale) and do not stop.
Not long ago, OpenAI founder Sam Altman said in a speech: "scaling laws are still valid, GPT-5 will be much smarter than GPT-4, we have not reached the top of this curve yet".
scaling laws is a proven path by OpenAI, which improves the ability of large models through heaping computing power and volume parameters, and then reaches the stage of real general artificial intelligence.
However, returning to the domestic large-scale model industry, a practical problem faced by Scaling Laws is not only the gap between the domestic and foreign industrial chain foundation represented by chips, but also a more critical problem:
How to give full play to the scenario and application advantages of Chinese technology enterprises in the mobile Internet era, find a path to efficiently implement the model in industry applications in addition to the heap parameters, and transform the model capabilities in the laboratory into visible application value?
On this issue, iFLYTEK, as the head echelon of the domestic large-scale model industry, has taken the lead in finding a new way.
1. Spell the base and crack the "correct posture" of technical anxiety
Returning to the vigorous boom of large models, with the growth of the number of model parameters and datasets, the capability boundaries of large models are expanded and iterated, which is undoubtedly the key enlightenment brought by scaling laws to domestic AI players.
In the cruel "100 model war" in the industry in the past, although many industry models and vertical applications have emerged in China, if you mention the anxiety of domestic large model players, ranking first is still the difficulty of "technical base", which contains two core problems: one is whether the computing power is enough, whether it is good or not? The other is to benchmark OpenAI's GPT-4/4V and even the future GPT-5 and other base base models, can the domestic general large model capabilities continue to narrow the gap?
At the computing power level, according to Li Feifei's estimates, the training cost of the latest generation of AI models has reached an unprecedented level. For example, OpenAI's GPT-4 is estimated to have used $78 million worth of computing resources for training, while Google's Gemini Ultra cost nearly $200 million in computing resources.
Back at home, affected by geopolitics and under the external pressure of the United States' chip sanctions against China, domestic large-scale model players inevitably face problems such as scarcity of computing power. At the same time, the high cost of computing power brought about by large-volume computing power has also raised the threshold for the landing of large-scale models, and the high cost of computing power is also a major practical problem.
To alleviate the computing power anxiety of the landing of large models, from national policy support to enterprise layout, we are all in action. Not long ago, Beijing issued the "Implementation Plan for the Construction of Computing Infrastructure in Beijing (2024-2027)", which clearly pointed out that it will vigorously promote the adaptation of artificial intelligence large models and independent controllable chips, and improve the security, stability and resilience of the supply chain of the mainland's intelligent computing industry.
As the "national team" of Chinese artificial intelligence, iFLYTEK gives the idea of creating a localized, independent and controllable computing power base, and providing a new "computing power choice" for the large model industry. To this end, iFLYTEK and Huawei have joined forces to create the first 10,000-calorie-scale domestic computing platform "Feixing No. 1" on the computing base, Huawei's hardware capabilities combined with iFLYTEK's AI technology precipitation, an independent and controllable computing power base has injected new vitality into the domestic large model ecology.
On this basis, the competition for the underlying model capabilities is also going on simultaneously, and domestic and foreign technology giants are accelerating to catch up with and surpass GPT-4 and iterate on the capabilities of general models.
Among the domestic large-scale model players, iFLYTEK is one of the few technology companies that has announced a specific timetable for its model iteration. On January 30 this year, iFLYTEK released the "iFLYTEK Spark V3.5", which has achieved significant improvements in logical reasoning, language understanding, text generation, mathematical answering, code, and multimodal capabilities, and the overall performance is close to GPT-4 Turbo.
Not long ago, iFLYTEK announced the capability update of V3.5, focusing on the long text, long picture and long voice functions for users to efficiently acquire knowledge. Taking the long text that has now become the "standard" of large models as an example, the general long text capability of the Spark large model has generally reached GPT-4 Turbo, and the overall level of long text of the Spark large model has surpassed GPT-4 Turbo in the knowledge question and answer tasks in multiple vertical fields.
An industry consensus is that with the convergence of model technology and the evolution of the industry competition pattern, in 2024, the focus of competition for large models will still fall on the capabilities of general large models. On the one hand, the ability of the general model determines whether China's large model industry can grasp its own core discourse; on the other hand, only when the ability of the underlying general model continues to improve and catch up with the most advanced international level, can the industry large model have better results.
Focusing on the continuous increase in computing power and general models, this is the only way to consolidate the technical foundation, and it is also the correct posture to solve the technical anxiety. This also means that the R&D investment around the technology base also determines the future position of domestic model manufacturers.
iFLYTEK has its own clear rhythm and strategy in R&D investment. At the recent company's performance briefing, Liu Qingfeng, chairman of iFLYTEK, revealed that in 2023, the ratio of iFLYTEK's basic large-scale model R&D and industry application R&D will be about 7:3, and this year, 50% of the R&D investment will still be used for the ability of large-scale model bases. Liu Qingfeng predicted: "There is still a dynamic catch-up process between China and the United States in terms of base model capabilities, but China is the only country that will not be completely left behind, and iFLYTEK, as China's national team, is continuing to narrow the gap with its American counterparts." ”
2. Spell the ground, iFLYTEK's "cloud, edge, end" way
In addition to the "volume technology base inspiration" brought by Scaling Laws to large model players, at present, bidding farewell to sexy concepts such as hot and eye-catching "parameters", the competition of large models has entered the next competition point - whether it is to make applications or grab customers, the focus is on the breadth and depth of model landing and application.
Overseas technology giants such as Microsoft and Google have not only accelerated the integration of large model capabilities into existing product matrices, but also carried out all-round cooperation in the field of AIGC with many leading customers in the industry.
Objectively speaking, on the question of "how to land the large model", no one in the industry has yet to give a perfect model answer. There are two reasons behind it: first, from the perspective of both supply and demand, since the development of large model technology is still in the early stage, the demand side is still exploring the demand and value of its own business landing large model, and the model manufacturers as the supply side are also finding a landing paradigm through supply innovation. Second, large model technology is a kind of "productivity", and the landing industry needs to establish a circulation system at both ends of supply and demand, and all participants must be "profitable", for example, the industry and the industry need a more cost-effective large model, and model manufacturers also need to realize the realization with the help of large models.
The solution idea given by iFLYTEK is to start from the actual scene requirements, integrate "cloud, edge, and end" and have a full-scene layout to meet the needs of various complex and diverse scenarios for large models.
On the cloud side, iFLYTEK established a "1+N" large model system when it began to tackle large model technology, and in addition to the basic general large model, it simultaneously implemented industry large models and product applications for various industries, such as education large models and medical large models, and worked with leading enterprises in finance, energy, automobile, communications, chemical and other fields to jointly build industry large models.
At the same time, for developers, in addition to opening APIs and large model supporting capabilities such as RAG and Agent, the iFLYTEK Xinghuo-13B model has also been open-sourced to support the industry for scenario-based customization and fine-tuning. In the past three months, iFLYTEK has added 550,000 real-name authentication developers, and more than half of these application services have been implemented in the scenarios where enterprises just need them.
The iFLYTEK Xinghuo APP is also gaining recognition from more and more users. According to Qimai data, the number of downloads of the iFLYTEK Xinghuo APP on Android has exceeded 96 million, ranking first among the domestic tool general model APP.
On the side, in response to the privatization deployment needs of enterprises, the "iFLYTEK Xinghuo All-in-One Machine" jointly developed by iFLYTEK and Huawei can provide "out-of-the-box" large-scale model integration solutions for enterprises' high-frequency scenario application needs, and has been implemented in many industry scenarios such as cities, finance, operators, manufacturing, energy, and automobiles.
On the end side, on the one hand, iFLYTEK will implement the capabilities of large models into its own intelligent hardware such as learning machines, translators, and office books, and on the other hand, it will deeply empower the industry with the capabilities of large models, and is landing in the fields of automobiles, home appliances, robots, AI mobile phones, AI PCs, etc., bringing an upgrade of experience to the terminal products that people can reach.
For example, in the automotive scenario, through the collaboration of device-side small model + cloud-side large model, it can not only solve the user's interaction in the environment without network or weak network, but also better protect privacy, and solve simple problems directly on the local side and complex problems in the cloud, which can not only bring users a good experience, but also effectively reduce costs.
The benefits of the "cloud, edge, and end" layout are twofold.
First of all, since the beginning of the year, it has become a consensus in the industry to find the best entrance and carrier for the technical capabilities of the large model, and the "three swords together" can release the landing value of the large model in a larger range and lead the large model to a practical path. Secondly, for iFLYTEK, the "cloud, edge, and device" collaboration can better meet the application needs in various scenarios, and the combination of models of different sizes can reduce the cost and efficiency, providing more channels for its commercialization. Whether it is the sales performance of intelligent hardware empowered by large models or the rapid growth of developers, it shows that iFLYTEK is at the forefront of commercialization.
However, the layout of "cloud, edge, and end" is not for everyone, and everyone can do it.
iFLYTEK's advantage lies in its systematic capabilities from back-end technology to front-end application scenarios. At the back-end technology layer, on the one hand, it is the iterative Xinghuo general model created based on the domestic independent controllable software and hardware base, and on the other hand, it is the full-stack technology layout of iFLYTEK from the model, framework, tools and application layer.
In terms of front-end application scenarios, iFLYTEK relies on its excellent engineering capabilities to make the model size have the ability of "72 changes" under the condition that the effect is close to lossless, so as to achieve efficient deployment of large models.
It is understood that iFLYTEK provides B-level models from 100B-level and 10B-level to end-side hardware for different scenarios and hardware platforms, covering various scenarios and applications of cloud + end, so as to meet the needs of complex scenario models in various industries.
Previously, iFLYTEK launched the 13B long text model is an example. In the era of collective "volume length text" of large model manufacturers, compared with the large model with hundreds of billions of parameters, iFLYTEK has launched a large model with the best performance of 13 billion parameters in the industry through the "pruning" and "distillation" of the model size, and in the case of an effect loss of only 3%, Xinghuo has greatly improved the efficiency in document uploading, analysis processing, first response time and text generation of knowledge questions and answers.
The full-scene layout of "cloud, edge, and end", and the mutual support from back-end technology to front-end applications, not only extend the breadth of model landing, but also expand the depth of model landing, and the data and knowhow accumulated by the industry can further feed back the iteration of model capabilities, which is a healthier cycle.
3. Fight for value, a rare "technical pragmatist"
In the eyes of many industry insiders, 2024 is a critical year for the application of large models. From the beginning of the year, whether it is in the field of industry large models or the lighter SaaS enterprise service market, the "large modeling" around application scenarios has become a unified action of industry players.
Compared with foreign large-scale model ecology, the advantage of China's large-scale model is in its application. On the one hand, China's complete industrial ecology has a wide range of scenario bases, which is the natural advantage of large-scale model landing. On the other hand, from the era of mobile Internet, the application and scenario innovation experience precipitated by Chinese technology enterprises can also be replicated and promoted to the era of large models.
But despite this, in the application scenario, the "value dispute" around the large model also exists. What is the value of large models, and is it just a FOMO (fear of missing out) mentality for AI companies to develop large models?
On this issue, iFLYTEK's attitude is particularly pragmatic. On the one hand, as a veteran AI company, iFLYTEK has been betting on AI technology for more than 20 years, and iFLYTEK's commercialization of AI technology has been explored throughout the development of this company.
On the other hand, in terms of value concept, in public, Liu Qingfeng, chairman of iFLYTEK, has more than once expounded on the "important value of application scenarios", Liu Qingfeng has made it clear that in the field of large models, "whoever can land in the application scenarios with social needs can be the first to form a virtuous circle of self-hematopoiesis." ”
To understand this, we still need to return to the exploration of the application of large models at both ends of iFLYTEK BC.
In the past, the domestic to B business was often "people need to apply machines", but what iFLYTEK is doing is "making the model better adapt to people and industries".
Adapting the model to industries and people is not as simple as calling API interfaces, but requires in-depth understanding and mining of industry scenarios and user needs, which is a sufficient and necessary condition to give full play to the value of the model.
Taking the "large model on the car" as an example, behind the dazzling many marketing and technical concepts, iFLYTEK focuses on the "value side": first, for car owners, does the large model improve the car experience?
Following the above two lines, you can understand the practice of iFLYTEK. At this year's Beijing Auto Show, iFLYTEK demonstrated its self-developed new generation of "iFLYTEK Spark + Cockpit" solution, which closely combines the capabilities of the Spark large model with the car use scene, and uses the large model technology to reconstruct the human-vehicle interaction experience.
In 2023, China's automobiles will "go overseas" to exceed 5.22 million units, with an export growth rate of 56%, surpassing Japan for the first time to become the world's leading auto exporter. It is understood that the iFLYTEK on-board intelligent voice system covers 23 major languages, and more than 60 models have been designated to go to sea, and the cooperative models have been sold to many countries and regions in Asia, Europe, Australia, Africa, America and other places. Among the top 10 Chinese automobile companies going overseas, 8 companies have reached in-depth cooperation with iFLYTEK.
In the field of home appliances, the home appliance industry that has access to the capabilities of the iFLYTEK Xinghuo large model is also radiating new vitality. For example, the large model in the field of home appliances jointly created by iFLYTEK and Haier is the first to implement the application scenarios of large models in the fields of intelligent control, recipe customization, and intelligent customer service, so that home appliances can truly "understand people's hearts better". Samsung's TV voice assistant also combines the capabilities of the iFLYTEK Xinghuo large model to achieve AI intelligent Q&A and intelligent search.
The software and other enterprise service markets are also the embodiment of iFLYTEK's concept of "rigid demand scenarios". With the help of iFLYTEK's continuous iteration of code capabilities, iFLYTEK not only uses "AI programmers" internally, but also cooperates with more than 100 companies such as iSoftStone and Bank of Communications to promote and replicate "AI programmers".
In the operator market, based on the basic capabilities of the Xinghuo model, iFLYTEK has cooperated with operators to create a large call model, and jointly released the 5G new call "business shorthand" product with China Mobile.
According to public information, at present, the iFLYTEK Xinghuo large model has cooperated with leading enterprises in various industries such as automobiles, finance, energy, software, home appliances, and operators, and the momentum of the large model landing in thousands of industries is strong.
On the C-side, iFLYTEK is a typical "product thinking", from technology to products, and the pursuit of better user experience.
The intelligent hardware with the support of the large model has performed very well in the business map of iFLYTEK. According to the financial report, in 2023, the overall revenue growth of iFLYTEK AI learning machine will reach 120%, and the GMV of smart office hardware such as iFLYTEK smart office notebook, iFLYTEK intelligent voice recorder, and iFLYTEK intelligent translator will increase by 84% year-on-year.
Behind the performance improvement and GMV growth is the recognition of large model technology by C-end users. Taking the AI learning machine as an example, almost every time the underlying general model is upgraded, iFLYTEK will carry out a functional iteration of the intelligent hardware product. At present, the AI learning machine has 8 large model applications, including English speaking sparring, Chinese-English composition correction, mathematics interactive supplementary learning, encyclopedia free question and answer, parent-child education assistant, and intelligent programming assistant.
In the latest spring update, combined with the upgrade of long text, long graphics and long voice of the Xinghuo large model, the iFLYTEK AI learning machine product not only improves the accuracy of tasks such as homework correction, but also upgrades the "Encyclopedia Q&A" function in combination with multi-modal capabilities. The large model of Xinghuo, which integrates a large number of book knowledge, has become an "encyclopedia assistant" for children, and children can "ask and answer" with virtual human friends such as "Einstein", and improve their learning ability and questioning ability in vivid and interesting interactions.
In addition to mature hardware products such as learning machines, iFLYTEK is also actively laying out the next outlet. When "embodied intelligence" sets off a financing boom, what many people don't know is that iFLYTEK released the "iFLYTEK Robot Superbrain Platform" as early as 2022, providing developers with a full-stack tool chain, including model training, asset generation, and software and hardware access.
After the arrival of the large model, iFLYTEK quickly integrated the Spark large model with the "robot superbrain platform", and now iFLYTEK has cooperated with the head humanoid robot star company, such as Zhiyuan Robot, UBTECH, Unitree Technology and other cool humanoid robot products, is the blessing of iFLYTEK's full-link voice and Xinghuo large model.
From the base, to the ground, to the value, from this point of view, compared with those cool PPT displays, iFLYTEK's large-scale model practice does not seem to be "sexy", but it is down-to-earth enough. The history of science and technology has told us that whether it is iPhone to smartphones or chatGPT to the large-scale model industry, the precondition for technology to achieve subversion is to let technology take root in needs and scenarios.
The "pragmatism" of iFLYTEK allows us to see the scene we are looking forward to seeing under the boom of large models - in the near future, large models can usher in a "spark moment" and truly take root in thousands of industries.