This year, the development of China's AI large model industry depends on these
Reported by the Heart of the Machine
Author: Zenan
Today's artificial intelligence is no longer "usable", but "very easy to use".
Last week, Google and Lee Sedol had a long-lost conversation that evoked memories:
If you think about it, it's been eight years since AlphaGo beat humans in Go in 2016. Nowadays, the development of artificial intelligence technology has not slowed down at all, and it is creating a greater shock for us.
Generative AI technology has swept through various fields, starting with GPT-4, which led the technological explosion. Whether it's automatically writing articles or generating pictures and videos, technologies such as large models are gradually being implemented with products such as Copilot and AI mobile phones. In the foreseeable future, embodied intelligence beckons us – through the combination of hardware and software, robots are becoming smarter and about to replace some of our labor.
Recently, the theory of "new quality productivity" is becoming a hot topic in the field of technology. The new quality productivity is a new industry-led productivity, which is driven by subversive innovation compared with the traditional productivity, has the characteristics of new industrial chain and high development quality, and plays a leading role in the transformation of new and old kinetic energy.
The breakthrough of AI large model technology is a powerful driving force to drive scientific and technological innovation and build future industries. The technological revolution brought about by artificial intelligence may bring great changes to everyone's life just like the industrial revolution and the information revolution.
On March 26, at the Boao Forum for Asia, People's Daily Online officially released the "2024 China AI Large Model Industry Development Report", which expounded the development status and typical cases of China's AI large model, deeply discussed the challenges faced by the development of the domestic AI large model industry, and also looked forward to future trends.
Report download link: http://download.people.com.cn/jiankang/nineteen17114578641.pdf
Among them, the judgment on the "second half" competition situation and the outbreak of end-side applications is particularly noteworthy.
The implementation of AI large model technology will be a big wave of sand
In the process of continuous development, artificial intelligence has been implemented in many places, including but not limited to speech recognition, face recognition, machine translation, object detection, image generation, assisted driving, and so on. To a large extent, new technologies are already helping industries to increase productivity.
As generative AI technology evolves, the scope of intelligent upgrades will expand and become even higher. As a country with a complete AI industrial system, a new generation of AI solutions is being used in all walks of life in China.
However, we still face a number of challenges before we can achieve real technological change.
The first challenge is the shortage of computing power. As the size of large models grows exponentially, algorithms are increasingly relying on high-performance AI chips. A market research firm reported that Meta and Microsoft alone bought 150,000 H100 GPUs from Nvidia last year, spending about $4.5 billion each, but that's not enough: H100-based servers are already being delivered in 52 weeks.
In China, the high-performance AI chip market is also affected by the dual impact of import and export restrictions and technical bottlenecks, and the development of the large model industry is being restricted by computing power.
The second challenge is the limitations of large model architectures. Today's AI pre-trained large models all use the Transformer neural network structure with self-attention mechanism, which has many advantages in natural language processing and other fields, including the realization of fully parallel computing, capturing long-distance dependencies, modular design, processing of indefinitely long sequences, structure expansion, and good pre-training results.
However, with the continuous development and expansion of people, the inherent problems of transformer, such as large computing power consumption, large memory occupation, and limited generalization ability, have gradually emerged.
Even the seven original authors of the Transformer paper, Attention Is All You Need, said during a recent GTC roundtable discussion that the world needs something better than Transformer to take us to new performance plateaus.
The use of data is another important factor restricting the implementation of large models. For large models trained from scratch, the quality of the corpus data will greatly affect the model capabilities. For example, when OpenAI trained GPT-4, it was trained on about 13T token of data, including text-based and code-based data, as well as some fine-tuning data from ScaleAI and internally.
Comparatively speaking, domestic AI large model data faces problems such as incomplete data types and low credibility of information. Overall, the Chinese database that can be used for large model training is seriously insufficient compared with English data.
Finally, we're still looking forward to more popular apps. ChatGPT has quickly become the fastest-growing consumer app in history since its release, and Microsoft, which supports it, has also introduced large model technology to Office, Teams, and even Windows operating systems. In the ecology of domestic technology companies, there are still no similar explosive applications, which may be due to the fact that the commercialization idea has not yet been found, and the technology and personalization have not yet met the needs of users.
It can be said that after the "100-model war" of generative AI, tasks such as algorithm innovation and optimization, ecosystem construction, and application implementation have put forward higher requirements for companies that are building AI technology systems. Only a few of them will enter the future stage of large-scale application.
A new direction has emerged: towards the end side, and the combination of end and cloud
There is no doubt that domestic technology companies have been investing in new AI technologies and have reaped considerable results.
Through large-scale data training, general-purpose large models with tens or even hundreds of billions of parameters can learn to capture complex patterns and features, and make predictions on never-before-seen data. The general large model can understand and learn a variety of tasks, and thanks to large-scale pre-training and fine-tuning paradigms, it can complete multi-domain tasks and have the ability to understand and generate multi-modality.
Represented by Baidu Wenxin Yiyan, Ali Tongyi Qianwen, iFLYTEK Xinghuo, Tencent Hybrid Yuan Model, etc., a number of high-parameter cloud large language models make full use of computing power and massive training data, and can provide language understanding, knowledge question answering, mathematical reasoning, code generation and other capabilities.
On the one hand, they provide intelligent Q&A, text summarization and generation, image generation, video generation and other functions for C-end individual users. On the other hand, for B-end enterprise users, it is changing the traditional business model of enterprises, and is providing unprecedented capabilities such as intelligent marketing, customer service, automatic meeting minutes, text translation, and budget management.
Based on the general large model, we have seen a special large model for specific industries and fields, and has begun to enter the fields of finance, government affairs, and healthcare.
In the device-side direction, two new concepts of AI mobile phone and AI PC have emerged one after another, and the large model shows a wide range of application prospects.
Based on the "small-volume" pre-training model capability of in-depth optimization on the device side, the usage methods and habits of personal devices are being reshaped. AI can already provide personalized services such as document search, intelligent screen recognition, image creation, life assistant, and travel assistant. With the extreme optimization of large models, people are even looking forward to the application of large models on smart wearable devices.
On the one hand, the device-side large model can bring people more personalized AI capabilities, conduct a deeper, more accurate, and delicate understanding of user intentions, and provide more personalized services for complex scenarios. At the same time, it can also ensure that the data is located on the device side, protecting people's private information.
On the other hand, the transfer of some cloud computing tasks to the terminal for processing will also greatly reduce the cost of computing power, and some complex work and content processed on the device side can also be handed over to hundreds of billions or even trillions of AI models in the cloud through the network for processing, which is "device-cloud collaborative AI".
The large model system of device-cloud co-evolution is expected to solve some problems and challenges faced by the current large model paradigm:
In terms of computing resources, device-cloud collaboration can make full use of the fragmented computing resources of the cloud and terminals, and jointly optimize them with communication and storage resources.
In terms of model architecture, models with different sizes and volumes of devices and clouds and new modes of aggregation have obtained the emergence ability of large models and the power consumption advantages of small models.
In terms of data, the rapid development of large and small models and various applications is giving rise to a standardized and industry-segmented data governance system.
In terms of applications, after understanding the user's intent, the device-side large model can efficiently call other large models, services, and hardware capabilities to achieve extremely high availability.
This may be the direction of a new round of artificial intelligence change.
The landing of AI mobile phones leads the trend
It is precisely because of the prospect of device-side AI large models and "device-cloud collaboration" that major mobile phone manufacturers are the first to implement large models in the consumer field.
From the end of last year to the beginning of this year, many domestic manufacturers have launched a new generation of flagship mobile phones, and generative AI capabilities have invariably become the focus of releases.
Some of the "AI mobile phones" proposed by these mobile phone manufacturers are intended to understand that through device-side AI technology, they are human-centric, and use personalized information and sensor capabilities to greatly improve the situational awareness level of mobile phones, bringing a variety of efficient smart services.
Some use platform-based AI to connect various services with device-side + cloud-side models to achieve efficient inference and decision-making. By using the large model "agent" to decompose complex tasks and achieve autonomous decision-making at each sub-step, the mobile phone not only achieves a deep understanding of the instructions and needs issued by people, but also can further simplify operations and achieve a variety of complex goals autonomously.
Among them, vivo's performance in the end-to-side and matrix of large models is particularly prominent, and it has been introduced in detail as an enterprise case in the newly released report.
In November last year, vivo officially released its self-developed AI model "BlueLM", which was first implemented on the new generation of flagship mobile phones vivo X100 series.
This is the industry's first open-source self-developed large model running on the mobile phone, covering three parameter orders of billion, 10 billion, and 100 billion, with a total of five models. Based on the capabilities of the Blue Heart large model, vivo provides two applications on the device side, Blue Heart Small V and Blue Heart Qianxun, and provides global intelligent assistance functions for mobile phones.
Vivo's technological innovation has allowed many people to enjoy the convenience brought by large models. The official gave us a set of figures: the blue heart model has now covered more than 20 million users, achieved 27.61 million high-quality Q&A, generated 17.57 million paintings, wrote 6.49 million reports, and eliminated 850,000 passers-by with the "AI retouching" function.
Behind this, there are not only the advantages brought by the capabilities and optimization of the device-side large model, but also the credit of vivo's large model matrix: large models with different parameter levels can be applied to different scenarios through a variety of deployment methods, which not only satisfies the user's mobile phone experience, but also optimizes the inference performance and the memory and power consumption occupied during device-side deployment.
Among them, the 1B and 7B versions of the Blue Heart model can be run on mobile phones, which not only optimizes the hardware capabilities of the device, but also provides good AI generation capabilities, so that some applications can run normally around the clock.
The 70B version of the Blue Heart model is the main model for cloud services, providing capabilities such as role-playing and knowledge quizzes, which not only emerges intelligently, but also takes into account cost and performance. For complex tasks, vivo also relies on a richer amount of knowledge to bring a more professional intelligent experience through two large models, 130 billion and 175 billion.
With the improvement of the number of parameters, the blue heart model gradually has the capabilities of text summarization, language understanding, text creation, knowledge question and answer, role playing, complex logical reasoning, and complex task arrangement. Combined with the AI computing power of the new generation of mobile phones, the Blue Heart model realizes the ability to combine device-side deployment and device-cloud.
During the Boao Forum for Asia, vivo introduced the latest progress of the implementation of the Blue Heart model: At present, the AI capabilities obtained by mobile phone users have been upgraded to the combination of "device-side 7B" + "cloud-side 70B", making full use of the advantages of device-cloud combination.
Vivo has taken the lead in the four dimensions of AI technology - data, manpower, algorithms, and computing power: Since 2017, vivo has established an AI research team, built a knowledge graph to accumulate data, and has published a series of high-level papers in top journals over the years. vivo's research results are constantly being translated into engineering applications, and its self-developed large model has been ranked at the top of the C-Eval Chinese list of the comprehensive examination evaluation set of large language models.
C-Eval Leaderboard: https://cevalbenchmark.com/static/leaderboard_zh.html
Through in-depth thinking about the mobile phone ecology and user characteristics, the device-side intelligent assistant has gained unprecedented capabilities through the blessing of large models, which is not only "able to naturally communicate with people", but also brings subversive experiences in a large number of users' learning, life, work and other scenarios.
In 2024, the application of AI large models will explode?
There is still a lot of room for AI phones to develop. With the iterative optimization of AI algorithms, the improvement of chip performance, and the expansion of application scenarios, new productivity will become increasingly popular to meet people's growing and diversified needs.
Large AI models will profoundly affect the way people interact with devices, and this year may be about to undergo a qualitative change.
At the Boao Forum, people said that in 2024, with the support from the government to developers, the surge in user demand, and the promotion of investment by technology companies, large models will enter a stage of rapid development. If we are specific to mobile phones, combined with the technical evolution of AI large model end-to-end and matrix, the implementation of large model technology will subvert a series of functions that are immutable in our eyes.
We can expect that the equipment in our hands will be able to take on a completely different shape in the future.