How to understand the depth and breadth of the impact of this round of AI technology development? How to judge the nature of this round of AI technology change? AI large model is related to the transformation of human production tools, the reshaping of the core competitiveness of a country's manufacturing industry, and the long-term prosperity and trend of the economy.
- Article Information -
The author of this article is Dr. An Xiaopeng, Vice President of Alibaba Research Institute and Vice President of Alibaba Cloud Intelligence Group.
The AI model is an important milestone in the development of general artificial intelligence. On April 28 and May 5, the Central Committee put forward three firsts on the development of artificial intelligence: the first time to put forward "general artificial intelligence", the first time to put forward "industrial intelligence", and the first time to "grasp the wave of new scientific and technological revolution such as artificial intelligence".
The "revolutionary" feature embodied in this round of AI is not that AI can generate text and pictures, but that AI can generate code and build a new model of human-computer interaction, which can be combined with product research and development, process design, production operations, product operations and other manufacturing links and scenarios to improve production efficiency, form new productivity, and trigger the reconstruction of the industrial competition pattern.
The AI model is related to the transformation of human production tools, the reshaping of the core competitiveness of a country's manufacturing industry, and the long-term prosperity and trend of the economy.
The integration of digital and real is the global manufacturing industry
The core variable of the competitive landscape reconstruction
Digitalization is a watershed moment in an era of great change, and has become a catalyst for the sharp divergence of competition between companies, cities, and countries. The manufacturing industry is the most important industrial sector of digital and real integration, and the way, breadth and depth of its integration can directly affect and even determine the advanced level of the manufacturing industry and the global competition pattern.
1. The integration of digital and real is the fundamental reason why the U.S. manufacturing industry is leading the world
At present, many people subconsciously agree with the development path and model of the German manufacturing industry in the digital era, and "sing" the American manufacturing industry. However, in fact, in the past decade, the United States has been a "model room" for the development of global manufacturing.
From 2011 to 2021, the size of the U.S. manufacturing industry expanded from 1.5 times that of Germany to 2.4 times, and 2.5 times to 3.1 times that of Japan. U.S. manufacturing is still growing at a faster pace than Japan and Germany.
In the past 10 years, the consolidation and establishment of the leading position of the U.S. manufacturing industry is the result of "software-defined hardware", the result of the deep and comprehensive integration of digital technology represented by "cloud + AI" into the real economy, and the result of migration to new digital infrastructure.
2. The emergence of digitally native enterprises is an important symbol of the upgrading of the U.S. manufacturing industry
The emergence of digitally native enterprises is an important sign of industrial upgrading and economic prosperity.
The widening gap between the manufacturing industries of Germany, Japan and the United States is mainly reflected in the lack of a number of competitive digital native enterprises in the manufacturing field at the micro level.
Germany's "Industry 4.0" goal has not been achieved, there is a large gap with expectations, and the digitalization of small and medium-sized enterprises is slow, with studies showing that only 21% of small and medium-sized enterprises use digital technology in production, and there is no growth of a group of competitive small and medium-sized enterprises in the digital age.
The situation in Japan is similar to that of Germany, which has experienced a "lost two decades" and has also failed to produce a number of digitally native companies.
Digital technology drives the continuous emergence of new companies and products in the United States, with digital natives such as Tesla, SpaceX, Rivian, OpenAI, Snowflake, and Palantir, not only becoming global industry leaders, but also continuing to build new models of product innovation.
As described by Valentine in "From Toyotaism to Teslaism", digital native companies in the field of electric vehicles are redefining the car, and in the era of personalized, user-centered Internet of Everything and Intelligence of Everything, new species are born to cope with uncertainty, driven by high-frequency innovation, based on the iterative thinking of evolutionary organization, and accelerating commercialization from MVP small steps.
As a digital native enterprise, Tesla has four typical characteristics: software-defined, high-frequency innovation, customer operator, and evolutionary organization.
In 2010, the United States DARPA launched the Adaptive Launch Vehicle Manufacturing (AVM) program, proposing to "reinvent the manufacturing industry", through a radical transformation and reshaping of the equipment manufacturing industry, shortening the weapons and equipment development cycle to one-fifth of the current one.
The development and production of complex manufacturing products should be like the semiconductor industry, and its product design, simulation, testing, process, manufacturing and other activities are all completed in the digital space. This strategy has borne fruit.
On November 11, 2023, the first flight of the new generation of stealth strategic bombers in the United States, which is the world's first digital bomber-B21, adopts digital design from the beginning, and the development, deployment and testing of digital twins based on cloud computing bring better maintenance and longer life cycles, as well as lower infrastructure costs, It is the fastest model developed by the U.S. military in the past 30 years, and it can also be used to continuously upgrade product functions just like Tesla cars constantly download new software, and the upgrade of combat effectiveness will rely on agile software iteration.
The essence of this trend is that "cloud + AI" is not only a business infrastructure, but also an innovation infrastructure, and a cradle for the incubation of new enterprises and new products.
3. AI models are a new starting point to reshape the competitive landscape of the global manufacturing industry
The AI model is accelerating the arrival of the third wave of "digital-real integration", and intelligence is its main feature. The AI model will affect the development pattern of the manufacturing industry, and the AI model will be integrated into all aspects of the manufacturing industry's R&D and design, production process, quality management, operation control, marketing services, organizational collaboration and operation management.
In the field of R&D and design, AI has revolutionized the traditional scientific research paradigm.
In the field of biomedicine, in 2022, the AlphaFold2 model developed by DeepMind predicted almost all protein structures. Today, AI models can not only "predict" but also "generate" proteins, creating new possibilities for future drug production and development. For example, the ProGen system, developed by Salesforce Research, has succeeded in generating entirely new proteins from scratch.
The "LucaProt" deep learning model based on the Transformer architecture of Sun Yat-sen University has trained a large protein language model, shortened the virus discovery cycle from the past 2-3 months to one week, discovered tens of thousands of new viruses that cannot be identified by traditional manual comparison methods, and expanded the global RNA virus diversity by nearly 30 times. This will shorten the vaccine development cycle and reduce the development cost.
For example, engineers can automatically generate code instructions through the large model, complete the development and debugging of robot functions, and even create some new functions for robots.
In the process of equipment O&M, large AI models greatly enhance the capabilities of traditional vertical models. The AI large model has the ability to understand, and after the drone in the power industry collects information on the power equipment in the mountainous area, the traditional vertical small model gives the judgment: "the pin is not standardized", while the large model can develop image cognitive ability based on multi-modality, and the conclusion is given: "In the sky near the highway, there are 10 bolts on the left side of the red painted insulator connecting the tower gold, of which 3 have non-standard pins, including 1 out of stock, 1 unplugged, 1 damage, an abnormal description has been generated, it is recommended to confirm the on-site maintenance as soon as possible. ”
Large AI model
Four fundamental trends that empower manufacturing
In the era of "software defines everything", as a new productivity tool, the AI model will inevitably expand from the content field (Wensheng text, Wensheng diagram, etc.) to the production entity field, triggering a new efficiency revolution in all aspects of the manufacturing industry and accelerating the manufacturing industry to become intelligent.
1. AI-driven software upgrade is the main way for large models to empower the manufacturing industry
Industrial software is the soul and key to the digital transformation of the manufacturing industry.
There are many ways and ways to support the empowerment of the manufacturing industry, and the important way to expect is that AI will reconstruct the software development model, interaction mode, use process and business model, whether it is R&D, management, production or post-service industrial software, it will be re-upgraded with the large model, and the more complex the software system, the greater the space for future transformation.
The next-generation AI coding platform product built based on the code model has powerful code understanding and generation capabilities, and supports core scenarios such as code completion, test unit generation, code interpretation, and code error checking. With the rise of MaaS (Model as a Service), the model-centric development paradigm will lower the threshold for industrial software development and improve development efficiency.
According to CSDN's assessment at the beginning of 2023, GPT4's software programming ability is equivalent to the ability of a software engineer with a monthly salary of 30,000 yuan in China, and equivalent to Google's annual salary of $180,000 for an L3 engineer. A software job recruitment in the United States conducted a test: an engineer with only 4 years of programming experience can use AI tools to develop software 5 times more efficiently than 19 years of programming experience.
At the level of industrial software development, AI models are revolutionizing the software development paradigm. AI will work with humans to improve the efficiency of software development by multiple levels, such as content generation tools (documentation, coding, testing, release, and O&M) for front-line R&D personnel, which can greatly improve productivity. At the same time, the research and application of "code model" is triggering a revolution in AI coding.
AI has become a new tool for chip design, and the two-way flow of AI and EDA will open the next revolution in chip design, and traditional chip design companies such as Synopsys and Cadence are also actively embracing AI design.
The NVIDIA Hopper architecture H100 has 13,000 AI-designed circuits, and the AI-designed GPU reduces the chip area by 25% compared with traditional EDA, and consumes less power. Google started using reinforcement learning (RL) technology to design its own TPUAI accelerator layout.
At the level of industrial software performance, AI large models will promote intelligent software upgrades. For example, in the R&D and design scenario, Back2CAD launches CADGPT™ based on ChatGPT and other functions, which supports intelligent recommendation, document generation, code production and other functions, which can effectively assist in product R&D and design.
2. Bridging data flow breakpoints is an important value for AI models to empower the manufacturing industry
Every breakthrough in human-computer interaction technology will bring about an industrial restructuring. In the future, natural language will be able to control everything, profoundly change the way people use search engines, consume shopping, produce and manufacture, and profoundly affect the future industrial competition pattern.
The core of manufacturing digitalization is to resolve the uncertainty of complex systems with the automatic flow of data, and deliver the right data to the right people and machines in the right way and at the right time, so as to improve the efficiency of resource allocation.
However, the actual operation status of the enterprise is that there are breakpoints in the data flow in multiple links, and engineers need to develop various process software and process software. AI models have found a new way to change this situation.
This new path is that the natural language interaction capability based on the AI large model provides a new way of software development and interaction for the real-time and ubiquitous connection within the manufacturing enterprise and between the upstream and downstream of the industry, lowers the threshold of software development for processes and processes, improves efficiency, and bridges the countless breakpoints in the process of enterprise data flow.
For example, domestic robot companies have developed robot industry models with the help of the Tongyi model, which can realize the interaction between humans and machines based on natural language. For example, after receiving human instructions, the robot can understand, reason, and analyze, and automatically generate software code to organize and coordinate different agents to complete tasks in different scenarios.
This feature greatly lowers the barrier to entry for process developers and improves development efficiency and quality. From a global point of view, it can not only avoid data breakpoints, reduce the impact of manual intervention, thereby improving the stability and reliability of products, promoting the automatic flow of data in multiple links, and improving the intelligence level of the entire system.
In the digital era, the previously highly integrated and centralized manufacturing system is gradually moving towards decentralized production and flexible organization.
The AI large model + intelligent collaborative office platform helps to open up the data disconnection nodes in the manufacturing industry, promote the efficient flow of data in R&D, production, distribution, service and other links, so as to improve the collaborative efficiency within manufacturing enterprises and even between the upstream and downstream of the industry, and promote the manufacturing industry to move towards "intelligent collaborative production".
"Convergence" is the basic law of technological evolution in the past half century, and the integration of information technology (IT), communication technology (CI), control technology (OT) and DT technology represented by cloud computing and AI has accelerated.
Looking forward to the next 10 years, AI large models will empower every intelligent terminal, intelligent unit and intelligent system, real-time collaboration of intelligence driven by AI large models will become a basic trend at the cloud edge, agents empowered by AI large models will be ubiquitous, agents in equipment, production lines, factories and enterprises will be ubiquitous, and the core value of data flow will shift from describing information to decision-making and control flow.
Driven by the AI model, countless agents will realize decision-making intelligence and control execution, and move towards self-decision-making and self-control, and people will face the rise of an intelligent consortium.
3. Entering the control link is a key symbol of AI large model empowering the manufacturing industry
The core value of AI models in the manufacturing industry is not in marketing and management, but in production control.
The versatility and generalization of AI large models, as well as the new development paradigm based on "pre-training + fine tuning", will empower the manufacturing industry from R&D and design, production process, operation and maintenance quality control, sales and customer service, and organizational collaboration.
Among them, we believe that the core control system that enters the production process, such as PLC, MES, SCADA and so on, to improve the intelligence of the process production process is the key symbol of the application of AI large model in the manufacturing industry.
Siemens and Microsoft announced a partnership in April this year to promote the next generation of automation technology based on GPT, and jointly develop a code generation tool for PLCs to integrate large AI models into the control process.
At present, in the field of power dispatching, AI large models can penetrate into the core business links of complex dispatching and control of new power systems, and become "expert assistants" for dispatching business, which can provide power dispatchers with power grid control strategies, optimize line load balancing, and reduce power grid losses.
At present, enterprises are exploring the use of AI large model capabilities to drive the intelligence of industrial software SCADA.
SCADA system (data acquisition and monitoring and control system) can be applied to many fields such as data acquisition, monitoring and control and process control in the fields of electric power, metallurgy, petroleum, chemical industry, gas, railway, etc.
In the SCADA scenario, the typical practice is to use the programming interface and ecological library of the large model in a specific industry scenario to generate industrial logic code (interaction, modeling, SQL development), which is automatically integrated into the industrial software and optimizes the model based on the results in a closed loop.
In the automotive industry, in recent decades, the transformation of the automotive industry has not only been a revolution in power, but also a revolution in control.
The biggest technological change in the evolution of traditional cars to smart cars lies in the innovation of vehicle control systems, from more than 80 ECUs and other electronic control units in traditional cars to a centralized architecture similar to smartphones (underlying operating system + chip SOC + application software).
Today, autonomous driving has become another major direction in the transformation of the automotive industry.
At present, there are two main directions for the change of large models to autonomous driving: one is that the large model is used as an enabling tool to assist the training and optimization of autonomous driving algorithms, and the other is that the large model enters the decision-making control link and directly drives the vehicle as a "controller".
According to public reports in August 2023, Tesla's "end-to-end" AI self-driving system FSD Beta V12 made its public debut, relying entirely on on-board cameras and neural networks to identify road and traffic conditions and make decisions accordingly.
Of course, at present, the AI large model has entered the control link, and the actual application and implementation process still faces many problems, which need to be further explored and solved by scientific researchers.
4. The collaboration of large and small models is an important trend of AI large models empowering the manufacturing industry
The AI model itself needs to find a specific landing scenario, and there is still a long way to go to solve the actual scenario problems of thousands of industries. From the perspective of actual industrial development, an important trend is that the collaboration between general and dedicated, open source and closed source, large model and existing software and hardware systems is a necessary stage for industrial landing, and at this stage, AI agent, an important carrier of high collaboration between large and small models, will become a new production tool.
An AI agent is an LLM-based agent that can autonomously complete a specific task using tools. The AI agent collaborates LLMs with other external tools, such as models and software, to handle a variety of complex tasks in the real world.
In July 2023, Alibaba Cloud launched ModelScopeGPT, an intelligent tool that can receive user instructions and call other AI models in the ScopeGPT community with one click through the "hub model", so that large and small models can collaborate to complete complex tasks and lower the threshold for using large models.
In the future, the AI agent will mainly consist of "perception system + control system + execution system", which will not only have the ability to generate, but also have the capabilities of task understanding, task disassembly, task scheduling, execution planning, and chain collaboration.
Among them, the LLM will mainly assume the role of the command center, similar to the role of the human "brain", to carry out unified intelligent scheduling and management of digital tools (such as SaaS software, industrial robots, digital humans, etc.) connected to the AI Agent, and in real time in production, management, service and other scenarios, different combinations of digital tools will collaborate to complete practical problems in specific scenarios.
Build "public cloud + AI" systematization capabilities
Promoting intelligent manufacturing to a "new stage"
Today's manufacturing transformation and upgrading is no longer the empowerment of a single technology, but the all-round empowerment and support of the technical system represented by "public cloud + AI".
At present, we must grasp the historical opportunity of the development of a new generation of artificial intelligence technology represented by AI large models, and accelerate the promotion of intelligent manufacturing to a "new stage".
1. Implement the "public cloud first" strategy, and take public cloud as a key force to promote the integration and innovation of "manufacturing + AI large model".
The large-scale, high-availability, and low-cost computing infrastructure of the public cloud has become the key foundation for industrial intelligence.
In particular, after the United States upgrades chip control, public cloud is the best way to alleviate the bottleneck of high-end chips, by efficiently connecting heterogeneous computing resources, breaking through the bottleneck of a single performance chip, and collaboratively completing large-scale intelligent computing tasks, which can effectively reduce the dependence on overseas high-end chips.
First, the "public cloud first" strategy should be regarded as an important part of the policy planning related to the digital transformation of the manufacturing industry, and the medium and long-term development goals, key tasks and safeguard measures should be clarified.
Second, it is necessary to avoid the phenomenon of "run" on chips, and be vigilant against the "small scattered many" in various places to build computing power centers in a hurry, resulting in the "fragmentation" of the unified computing power market, and avoiding the phenomenon of building too much, not using well, and not being able to afford it;
The third is to take the utilization efficiency of data centers as the assessment index of data center construction, and reverse the mode of "heavy construction and light operation" and "heavy investment and light performance" in data center construction.
2. Encourage open source models, support technology platform enterprises to expand and strengthen the model open source community, and prosper the AI industry technology ecosystem
The competition of AI is not only a technical war, but also a business war, the core of which is ecological warfare, and the key lies in open source and openness. Open source can reduce R&D costs and application thresholds, and is a "booster" for innovation to business closed-loop.
The first is to do a good job in the top-level design of the AI open source and open ecosystem, and incorporate the construction of the AI open source and open ecosystem into the national plan and implementation.
The second is to encourage local governments to cooperate with the head platform of the AI open source community to build AI empowerment centers, relying on massive open source models and model-as-a-service platforms (MaaS platforms) to accelerate the innovative application of digital intelligence in the manufacturing industry;
The third is to encourage application traction, accelerate industrial landing, support manufacturing enterprises to accelerate the application of basic models, R&D and application industry models and enterprise-specific models, and feed back technological innovation through "using models".
3. Start the AI-driven upgrade project of industrial software to accelerate the intelligent upgrade of the whole link and chain of the manufacturing industry
As a key support for intelligent manufacturing, industrial software is of great strategic significance to promote the transformation and upgrading of the manufacturing industry. In the era of AI, all industrial software is worth re-upgrading with large models.
First, it is necessary to vigorously develop AI-based industrial software, promote the research and development of "industrial software + AI large model" technology, enhance the independent innovation ability of industrial software in the intelligent era, and actively promote the development of industrial software standards;
Second, it is necessary to give full play to the role of industrial software-related alliances as a communication bridge, give full play to the respective advantages of AI enterprises, industrial software enterprises, scientific research institutes and manufacturing enterprises, and build an AI-driven industrial software industry ecology with win-win cooperation and core competitiveness.
4. Focus on the key industrial chain of the manufacturing industry, create benchmarks in different links and scenarios, and demonstrate and promote the large-scale application of large models in the manufacturing industry
The key industrial chain of the manufacturing industry is an important support for accelerating the construction of a modern industrial system, and it is necessary to identify the key links, concentrate high-quality resources, build an element system with "computing power + algorithm + data" as the core, improve the degree of digital and real integration of the manufacturing industry, and promote the safety and intelligent upgrading of the manufacturing industry chain.
The first is to launch the demonstration project of large models to support new industrialization, take AI large models as the starting point, promote all-round and in-depth empowerment of new industrialization by AI, and accelerate the exploration of new "new models" of new industrialization;
Second, in the manufacturing industrial belt, advantageous development zone, and industrial park with good industrial foundation and strong innovation ability, take the lead in carrying out the "manufacturing + AI large model" integrated innovation and development demonstration project;
Third, through the "innovation platform + digital factory" and other models, aiming at the shortcomings and weaknesses of key links such as perception, control, decision-making, and execution, we will strengthen the joint innovation of production, education, research and application in different scenarios, create innovative application benchmarks, and promote the large-scale application of large models.
Transferred from the public account: South China Intelligent Manufacturing