Home About us

The breakthrough battle of the industry model

Lao Ji said science and technology 2024/06/25 07:52

The innovation of large models not only requires single-point innovation, but also systematic innovation around AI-Native. It is not only necessary to explore extensively in the scene, but also to form a breakthrough in technology.

Idealistic AI that reinvents everything

Large models have become the undisputed biggest outlet in the current IT industry.

It has been more than half a century since the concept of artificial intelligence was proposed, and after the short-lived appearance of "Deep Blue" and AlphaGo, the Transformer architecture and ChatGPT were born, which once again ignited the enthusiasm of the world, artificial intelligence was presented in a more concrete and down-to-earth way, and large models and generative AI have also begun to enter the field of vision of more people.

In the past year, OpenAI has repeatedly refreshed its popularity, from interactive Q&A, voice dialogues to Wensheng videos; Google and Meta have entered the game to launch open source models, trying to find a place; Microsoft has integrated GPT4.0 into office software as a productivity tool, while Huawei and Apple have integrated mobile phone voice assistants into large models to accelerate device-side intelligence.

The inflection point of the AI industry has arrived, and large models are reshaping all office, production and life.

Behind the popularity of large models, NVIDIA has become the biggest beneficiary. With the rapid growth of the industry's demand for AI computing power, GPU cards are expensive and hard to find, and its 24/25Q1 quarterly report increased nearly 4 times year-on-year to $26 billion. Supported by continuously improving performance data and expanding market demand, Nvidia's market capitalization reached $3.35 trillion on June 18, 2024, surpassing Microsoft and Apple to become the world's largest company by market capitalization.

Everything is moving in the direction that people expect, and some people even shout that "the fourth industrial revolution has come".

Lao Ji said that science and technology, the battle of the breakthrough of the industry model

Under the fame, it is actually difficult to deserve

This wave of large models is not as glamorous as it appears on the surface, and a few families are happy and a few are sad.

According to a report released by Sequoia Capital, Nvidia's chip orders in 2023 alone will be as high as $50 billion, making a lot of money, while the sales revenue of the entire generative AI company is only $3 billion, and most companies are still far from profitability, and some companies are even on the verge of bankruptcy.

If gold diggers continue to fail to find gold, the business of selling shovels will not be sustainable.

Whether it is OpenAI, Microsoft or Google, these AI pioneers have not yet found a stable and sustainable business model, and more have drawn a grand blueprint for capital to support market expectations.

In China, the number of large models with more than 1 billion parameters has far exceeded 100. However, the large model industry, which has not yet formed a mature business model, has begun to roll up the price, the model is free, the price of Token is reduced, and the computing power is discounted, what a familiar taste. In the early days of the melee, it is understandable to delineate a wave of users and developers through low prices, and appropriate competition is conducive to the healthy development of the industry, but disorderly competition will lead to a vicious circle and plunge the large model industry into a death spiral.

Lao Ji said that science and technology, the battle of the breakthrough of the industry model

At the end of 2023, Gartner released the maturity curve of emerging AI technologies, and large models and generative AI are at the highest point of the technology hype cycle.

The market will eventually return to rationality, and only by truly creating value for users can we achieve a closed business loop and have a longer-term future.

The ideal is very plump, and the reality is very skinny! There is still a long way to go from the trend to the full implementation of the large model.

Break the whole into zero, intersperse operations around the industry, and achieve a strategic breakthrough

How is China developing in this wave of AI? When the news of ChatGPT, Sora, and Nvidia hit the Internet overwhelmingly, many people couldn't help but be anxious: Are we lagging behind again?

It must be admitted that China does not have advantages in the fields of computing power, algorithms and data: high-end computing power cards cannot be supplied by Nvidia card neck, and its own process and design cannot keep up in a short period of time; In the field of algorithms, although there are a large number of models, most of them are optimized based on foreign open source architectures, lacking autonomy and leadership. In the field of data, our open data is not on the same order of magnitude as abroad, and the Chinese corpus is seriously insufficient.

Each of the three elements of AI is a fatal injury!

But from another point of view, China has the world's only whole industrial category, the largest number of financial consumers, and the largest government affairs and urban system, which has produced rich scenarios and private data, which have become the natural soil for the development of industry models. Therefore, we did not take the road of frontal assault of the large corps, but broke the whole into parts and interspersed operations around the industry, empowered each industry segment scene through the large model, and finally formed a strategic breakthrough.

Lao Ji said that science and technology, the battle of the breakthrough of the industry model

In the past two years, with the joint efforts of users and manufacturers, some industry models have blossomed. The large model has been gradually applied to scenarios such as government affairs guidance, official document retrieval, and event distribution to help the government improve the efficiency of government services and urban management. In the mining field, large-scale models and cloud-edge collaboration help mines achieve safety and efficiency improvement, and accelerate the intelligence of industrial clusters. In the railway, the TFDS trackside image detection system for train faults based on the visual large model realizes non-stop real-time image acquisition and analysis, and automatically identifies various types of railway freight car faults. The weather prediction scheme based on large models can increase the calculation speed by 10,000 times compared with the traditional HPC numerical calculation, and predict the typhoon path more accurately.

There are also more industries such as medicine, manufacturing, steel, finance and other industries continue to land, we are taking a different pragmatic road from foreign countries, not following the trend and not adventurous, along the clear goal and rhythm of polishing the scene one by one, this trickle, will eventually converge into an intelligent ocean.

Suggestions for accelerating the implementation of industry large models

The construction of a large model of the industry is easier said than done.

Government and enterprise business scenarios are complex and diverse, and it is difficult to use a general large model to deal with them. At this time, we should realize that the number of volume parameters, the number of volume tokens, the volume cluster size, and the volume price are meaningless, and the implementation of industry large models needs to pay more attention to engineering issues.

First, the coordinated development of software and hardware

In addition to the model itself, the large model also involves a development platform, a development framework, a computing architecture, and various tools, as well as hardware infrastructure such as diversified computing power, high-performance storage, and high-bandwidth networks. On the basis of the hierarchical decoupling architecture, end-to-end integrated design and verification are required in terms of performance, reliability, maintainability, and compatibility. For example, full-link visual O&M of software and hardware, low-latency and high-bandwidth networks for computing-network collaboration, and operator acceleration based on affinity ensure that the system not only runs well, but also runs stably.

2. Cloud infrastructure using AI-Native

AI-Native's cloud infrastructure is becoming a priority for more and more enterprises. Based on the public cloud to train the basic large model, combined with private data in the local private cloud through secondary training and fine-tuning, this hybrid cloud solution takes into account efficiency and security, and has become a new paradigm for large model construction. Compared with the non-cloud deployment mode, it can also provide unified scheduling of general computing power and AI computing power, and cover the whole process of data, model and application development based on its rich technology stack, lowering the development threshold.

3. Establish AI development workflows to promote the deterministic delivery of models

The delivery process of a large model involves dozens of links in the whole process of data management, model development environment, model training, and inference deployment, and it is difficult for the traditional development mechanism to solve the problems of cross-team collaboration and iterative development. To reduce the uncertainty of the model development process, it is necessary to establish a one-stop AI development workflow to help enterprises quickly build a mechanism for cross-team collaborative development and efficient iteration, and at the same time improve the delivery efficiency and quality of the model through standardized and automated processes.

Fourth, pay attention to data engineering and create high-quality data sets

Model performance is determined by both data quality and algorithm design, and the current AI training data lacks systematic governance tools, and even some work needs to be handled manually, which has the problems of low integration, cleaning, and annotation, as well as values. To build a high-quality AI model, it is necessary to build core data engineering capabilities to provide high-quality data for the model. On the one hand, high-quality data can be obtained with the help of an open data authorization operation platform. On the other hand, it is necessary to introduce automated and intelligent means for internal data, and build a data cleaning, labeling and quality evaluation system. At the same time, through data quality analysis, component analysis, scene proportioning, and intelligent proportioning capabilities, a feedback and optimization mechanism from data matching to model effect is established, and continuous optimization is made based on application effect feedback to achieve value alignment.

5. Establish an enabling mechanism to prosper the industrial ecology

Ecology is an indispensable part of the development model. It is necessary to build an open, closed-loop, and high-quality ecosystem in the AI era from four levels: technology ecology, data ecology, model ecology, and application ecology. In this process, the role of the government and industry leading enterprises cannot be ignored, and they can take the lead in establishing such as joint innovation labs, model and application stores, demand matchmaking meetings, innovation competitions, etc., and create a neutral regional empowerment platform through administrative means through financial sponsorship, talent training, policy support and other traction; While carrying out their own business innovation, leading enterprises in the industry can also empower the upstream and downstream of the industry and industrial chain through technology, data, models and applications, and move from single-enterprise intelligence to industry intelligence.

6. Carry out continuous operations

The implementation of large models faces the challenges of lack of experience, lack of talents, and lack of capabilities, and most enterprises are practicing and summarizing at the same time. It is necessary to solidify these experiences and capabilities and gradually form engineering capabilities covering the whole process of large model implementation, including top-level design, POC testing, planning and implementation in the early stage, building high-quality datasets in the mid-term, carrying out scenario analysis and model development, and O&M operation in the later stage. The operation and construction of large models are equally important, and without an operating mechanism, it is difficult for large model platforms to continue to exert their value. Therefore, it is necessary to build a process, organization, and talent team suitable for enterprises, continue to carry out technology, ecology, and user operations, and continuously optimize old scenarios and explore new scenarios, so as to achieve comprehensive intelligence.

The elders who want wood will consolidate their roots

Large-scale model innovation, in the final analysis, is a battle of technology, deep roots can be leafy.

The innovation of large models not only requires single-point innovation, but also systematic innovation around AI-Native, which not only explores extensively in scenarios, but also forms a breakthrough in technology.

In recent years, a number of representative technology companies have emerged in China, such as Cambrian, Horizon, Bichen Technology, iFLYTEK, Moore Threads, Huawei, Alibaba, Baidu, etc., who insist on investing in artificial intelligence innovation and promoting the upgrading of the AI industry.

In 2023, Baidu's Wenxin Yiyan, Alibaba's Tongyi Qianwen, and iFLYTEK's Xinghuo will be launched one after another, and Huawei will also release Pangu Model 3.0, proposing the concept of "AI for industries", and launching the industry's first large-model hybrid cloud based on HUAWEI CLOUD Stack. At the recent HDC2024, HUAWEI CLOUD Pangu Model 5.0 was upgraded and unveiled, creating a full series, multi-modality, and strong thinking capabilities, further interpreting the vision of "solving problems and doing difficult things", and at the same time, it released ten innovative technologies for large-scale hybrid cloud, accelerating the implementation of enterprise-specific large models through systematic innovation of AI-Native.

Innovators are often lonely, but they are destined to be extraordinary.

It is not only necessary to calm down, but also to have the strategic patience of "sitting on the bench for ten years".

We are finally about to usher in the golden age of AI.

This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com