Artificial intelligence is a new productive force, and this paper comprehensively expounds the development trend of artificial intelligence from the proposal, application, key technical standards, key areas, typical representatives of the next generation of intelligent terminals and large models.
On the proposal of artificial intelligence
Since the concept of artificial intelligence technology was first proposed in the 50s of the 20th century, the development of artificial intelligence has experienced three "highlight moments" of Deep Blue, AlphaGo and ChatGPT.
In 1950, British computer scientist Alan Mathison Turing proposed the well-known Turing Test to determine whether a machine can exhibit intelligent behavior comparable to or indistinguishable from a human. In this test, a human evaluator judges a natural language conversation between a machine and a human, and the evaluator does not know in advance which side is the machine, when isolated from each other. If the evaluator cannot reliably distinguish between a machine and a human, then the machine can be said to have passed the test. Turing will "Can Machines Think?" This question has come into the public eye and is now an important concept in the philosophy of artificial intelligence.
American scientist Marvin Lee Minsky designed and built SNARC (Stochastic Neural Analog Reinforcement Calculator) in 1951. This is the world's first neural network computer that simulates the movement of a mouse in a maze where the circuit is strengthened as soon as the mouse reaches the target. Marvin himself, the founder and pioneer of artificial neural networks, and John McCarthy are both founders of the MIT AI Lab.
In 1952, Christopher Strachey, a teacher and physicist from England, made an unnamed checkers game that was probably the world's first video game (debatable) and definitely the first AI computer game – because its built-in AI program could "complete the game at a reasonable speed". Although this program cannot be self-improving through learning, it is one of the first attempts to use computers to simulate human thought processes, and it occupies an important place in the history of AI development.
In 1956, American computer scientist John McCarthy first coined the term "artificial intelligence" at the Dartmouth Conference. This conference marked the birth of the field of artificial intelligence.
About the application of artificial intelligence
In the 21st century, artificial intelligence has become a new driving force for global scientific and technological innovation, industrial transformation and new quality productivity.
In November 2022, OpenAI released ChatGPT-35 quickly sparked a global heated discussion, and the number of registered users exceeded 100 million in just two months. Behind ChatGPT is a series of major technological breakthroughs in generative AI, mainly including large language models (LLMs) based on Transformer deep learning architecture; a pre-training learning paradigm that uses large-scale datasets and unsupervised learning methods for initial training of models; and human feedback reinforcement learning (RLHF), a machine learning technique that combines reinforcement learning and human feedback. These technologies have changed the way machines process natural language, improved model performance, reduced development costs, and established the ability for models to self-upgrade and iterate.
2023 is undoubtedly the first year of generative AI.
In 2023, the scale of China's artificial intelligence core industry will reach 500 billion yuan, of which the market size of artificial intelligence large models will reach 2.1 billion US dollars, a year-on-year increase of 110%.
In 2024, the scale of the artificial intelligence industry will show a rapid growth trend. According to IDC data, the total global AI IT investment in 2022 will be US$128.8 billion, and the global AI IT total investment is expected to reach US$154 billion in 2023, a year-on-year increase of 196%。According to Sullivan Consulting**, it is estimated that the global artificial intelligence market will reach 615.8 billion US dollars in 2024, and China will exceed 799.3 billion yuan. Among the major subdivisions of artificial intelligence, large models, as a frontier hot spot, have the fastest growth rate. According to the report of the titanium ** international think tank, it is expected that the global artificial intelligence large model market will exceed 28 billion US dollars in 2024, and the market size of China's large model will reach 21.6 billion yuan.
In 2024, artificial intelligence will begin to be widely applied to all fields of production and life, and artificial intelligence will empower the new economy and become a new driving force for the formation of new qualitative productivity.
The artificial intelligence industry chain is divided into three parts:
The upstream part mainly includes key technology fields such as chips, computing power, semiconductors, CPOs, and optical modules;
The midstream segment is mainly based on software products, solutions and technology platforms built on various identification technologies;
The downstream part covers the application level of the combination of AI and various industries.
About key technical standards for artificial intelligence
Just entering 2024, China has taken the lead in deploying artificial intelligence and formulated key technical standards for artificial intelligence. It mainly includes machine learning, knowledge graph, large model, natural language processing, intelligent speech, computer vision, biometric recognition, human-machine hybrid augmented intelligence, agent, swarm intelligence, cross-intelligence intelligence, embodied intelligence, etc.
Machine Learning Standards. Standardize machine learning training data, data preprocessing, model expression and format, model effect evaluation, etc., including standards such as self-supervised learning, unsupervised learning, semi-supervised learning, deep learning, and reinforcement learning.
Knowledge Graph Standards. Standardize the description, construction, operation and maintenance, sharing, management, and application of knowledge graphs, including knowledge representation and modeling, knowledge acquisition and storage, knowledge fusion and visualization, knowledge computing and management, knowledge graph quality evaluation and interconnection, knowledge graph delivery and application, knowledge graph system architecture and performance requirements, and other standards.
Large model standard. Standardize the technical requirements for large model training, inference, deployment, and other links, including standards such as general technical requirements for large models, evaluation indicators and methods, service capability maturity assessment, and generated content evaluation.
Natural language processing standards. Standardize the technical requirements and evaluation methods for language information extraction, text processing, and semantic processing in natural language processing, including standards such as syntax analysis, semantic understanding, semantic expression, machine translation, automatic summarization, automatic question answering, and language large models.
Intelligent voice standards. Standardize technical requirements and evaluation methods such as front-end processing, speech processing, voice interfaces, and data resources, including standards such as deep synthesis forgery detection methods, full-duplex interaction, and general speech models.
Computer vision standards. Standardize technical requirements and evaluation methods for image acquisition, image processing, image content analysis, 3D computer vision, computational photography, and cross-integration, including standards for function, performance, and maintainability.
Biometric standards. Standardize technical requirements for biometric sample processing, biometric data protocols, equipment, or systems, including standards for biometric data exchange formats, interface protocols, and so forth.
Human-machine hybrid augments intelligence standards. Standardize multi-channel, multi-mode and multi-dimensional interaction paths, modes, methods and technical requirements, including brain-computer interface, knowledge evolution, dynamic adaptation, dynamic recognition, human-computer collaborative perception, human-computer collaborative decision-making and control, and other standards.
Agent Standards. Standardize the technical requirements for agent instances with general large models as the core, as well as the basic functions and application architecture of agents, including standards such as agent reinforcement learning, multi-task decomposition, inference, prompt word engineering, agent data interface and parameter range, human-machine collaboration, agent autonomous operation, and multi-agent distributed consistency.
Swarm Intelligence Standards. Standardize the technical requirements and evaluation methods for control, formation, perception, planning, decision-making, and communication of swarm intelligence algorithms, including standards for autonomous control, cooperative control, task planning, path planning, collaborative decision-making, and networking communication.
Cross-smart standards. Standardize the technical requirements for text, image, audio and other multimodal data processing basics, conversion analysis, fusion applications, etc., including standards for data acquisition and processing, modal conversion, modal alignment, fusion and collaboration, and application expansion.
Embodied Intelligence Standards. Standardize standards such as multimodal initiative and interaction, autonomous behavior learning, simulation, knowledge reasoning, embodied navigation, and group embodied intelligence.
On the focus area of AI applications
China has also laid out and developed key areas of artificial intelligence industry applications, including intelligent manufacturing, smart homes, smart cities, scientific intelligent computing, etc.
Smart Manufacturing Standards. Standardize the integrated application of artificial intelligence in the industrial field, and carry out the development of standards such as industrial knowledge expression, industrial knowledge graph construction, and industrial scene large models around the intelligent technical requirements in the construction of smart factories and smart chains.
Smart Home Standards. Standardize the technical requirements of home smart hardware, smart software, intelligent networking, service platforms and application platforms, promote the interconnection of smart home products, and improve the user experience of smart home in indoor environment, security monitoring and other scenarios.
Smart City Standards. Standardize the requirements of intelligent technologies such as smart city construction, governance, and ecological livability, and improve the application level of artificial intelligence systems in urban economic development, resilience construction, social governance, and auxiliary decision-making.
Scientific and intelligent computing standards. Standardize the relevant standards for the large-scale application of artificial intelligence to accelerate basic scientific research.
About the next generation of intelligent terminals
China proposes to focus on breakthroughs in the next generation of intelligent terminals.
Develop industrial terminal products that adapt to the trend of general intelligence, support the improvement of quality and efficiency of industrial production, and empower new industrialization. The development of large-scale, wide-ranging, intelligent, convenient, and immersive consumer-level terminals to meet the new needs of digital life, digital culture, public services, etc. Build an intelligent medical and health terminal suitable for the elderly to improve the quality of life and health of the people. Break through high-level intelligent networked vehicles, metaverse entrances and other super terminals with explosive potential, and build new advantages in industrial competition.
In 2023, China has released more than 200 large artificial intelligence models, and according to the news released by the CCID Research Institute of the Ministry of Industry and Information Technology of China, in 2023, the size of China's generative artificial intelligence market will be about 144 trillion yuan; By 2035, it is expected to contribute 30 trillion yuan to China's market value.
However, there is still a certain gap between China and the United States in terms of the underlying ** and core technology of artificial intelligence. As of June 2023, all the top five generative AI products in the world are from the United States.
The Open Artificial Intelligence Lab (OpenAI) was officially established in 2015, and OpenAI is committed to developing large language models. On November 1, 2022, OpenAI officially launched ChatGPT, marking a major breakthrough in general generative AI technology. ChatGPT is a deep learning algorithm that mimics human cognition, and its advent has triggered a new wave of global artificial intelligence applications.
As generative AI applications accelerate their penetration into various industries, the application ecosystem is becoming the main battlefield for a new round of global competition. According to IDC**, more than 500 million new applications will be created worldwide by 2024, which is equivalent to the number of applications that have appeared in the past 40 years combined.
A typical representative of the big model
As a core component of artificial intelligence technology, large model technology plays an important role in various fields and has attracted great attention.
The mature artificial intelligence large model technology mainly includes:
Deep learning model: It is an important machine learning technology in the field of artificial intelligence, which simulates the cognitive process of the human brain by building deep neural networks. Deep learning models can automatically extract the features of data and learn and optimize them in massive data, so as to achieve remarkable results in speech recognition, image processing, natural language processing and other fields.
Convolutional Neural Network (CNN): A deep learning model designed to process image data. It can effectively extract hierarchical features from the original image through local perception and hierarchical network structure. In the field of computer vision, CNN has become the mainstream method for tasks such as image classification, object detection, and face recognition.
Recurrent Neural Network (RNN): A neural network model used to process sequential data. It retains historical information through memory units, allowing for efficient modeling of sequence data. RNNs have a wide range of applications in the field of natural language processing, such as speech recognition, machine translation, and text generation.
Transformer architecture: is a deep learning model based on the self-attention mechanism, proposed by Google in 2017. It captures the features of input data through multi-layer self-attention mechanisms and position coding, and has achieved excellent performance in machine translation, natural language understanding and other fields. Transformer has become one of the infrastructures in the field of modern natural language processing.
Self-attention mechanism: is one of the core components of the Transformer architecture, which allows the model to focus on different parts of the input data and automatically learn how it is represented based on the input data. The introduction of the self-attention mechanism improves the expressiveness and flexibility of the model, enabling it to better deal with complex linguistic phenomena.
Generative Adversarial Network (GANs): A deep learning model used to generate new data. It consists of two networks: a generator and a discriminator. The task of the generator is to generate fake data that is as similar as possible to the real data, whereas the task of the discriminator is to distinguish between real data and fake data. GANs have a wide range of applications in the fields of image generation, image inpainting, and style conversion.
Reinforcement learning (RL): is a machine learning method based on trial and error in the field of artificial intelligence. The agent learns how to maximize the cumulative reward by interacting with the environment and deriving state and reward signals from the environment. Reinforcement learning has achieved important results in the fields of gaming, autonomous driving, and robot control.
Transfer learning: is a machine learning method that uses an already trained model as a foundation to train a new model. It reduces the time and data requirements for new model training by migrating parameters from a pretrained model into the new model. Transfer learning has been widely used in natural language processing, computer vision, and other fields, and has become an important machine learning technology.
Ensemble learning: is a machine learning method that improves accuracy and robustness by building a combination of multiple models. It improves overall performance by combining the results of multiple models. Ensemble learning has achieved good results in the fields of classification, regression and anomaly detection, and commonly used ensemble methods include bagging and boosting.
Generative model: It is a large AI model technology that can generate new data from existing data. It learns from existing data, extracts patterns or structures from it, and generates new data that is similar to the original data. Generative models have a wide range of applications in text generation, image generation, speech synthesis, and other fields, such as autoresponder systems, machine translation, and virtual assistants. Typical representatives of generative models are GPT series models, such as GPT-3 and GPT-4. These models use large amounts of linguistic data and use unsupervised learning and fine-tuning techniques to improve the quality of the text they generate. The GPT series models have shown strong application potential in many fields, such as natural language processing, machine translation, dialogue systems, etc.
end-