1 ) Trend One : Unifying the Future: Multimodal Models Accelerate Text, Image, and Video Fusion
✓ Multimodal models: Multimodal models can process diverse data such as visual, textual, and auditory information. They can integrate and understand information in different forms, further enhancing the transfer learning capabilities of large models. This is an important step for artificial intelligence to fully understand the real world.
✓ Development Status: Single-modal AI models for text, speech, and images are relatively mature, while large-scale models are rapidly developing towards multimodal information fusion. From the birth of CLIP to the image processing capabilities of GPT-4, image-text multimodal technology has made significant progress. Large-scale models are no longer limited to text and images, but are beginning to expand into fields such as audio and video.
✓ Future Outlook: Future models will face more complex and diverse interaction scenarios, and will place greater emphasis on the integration of various forms of information. Multimodal technology will open up new application spaces in areas such as smart homes, smart cities, medical diagnosis, and autonomous driving.
Multimodal model iteration process
2 ) Trend Two : Transcending Virtual Boundaries: Embodied Intelligence Becomes a New Form of AI Development
Embodied Intelligence: Embodied intelligence is an artificial intelligence system that can perceive and interact with the physical world and possesses autonomous decision-making and action capabilities. The intelligent agent in embodied intelligence can experience the physical world from the perspective of a protagonist, and through interaction with the environment and self-learning, develops the ability to understand and transform the objective world.
✓ Development Status: Stanford University Professor Fei-Fei Li has identified embodied intelligence as a key future direction for computer vision, calling it the "North Star" of artificial intelligence research. Currently, with Google launching the RoboCat large model and Nvidia launching Nvidia VIMA, embodied intelligence has become a highly competitive arena for leading AI companies.
✓ Future Outlook: General artificial intelligence and the robotics industry are in a period of strategic opportunity characterized by rapid development and mutual integration. As a core application at the intersection of these two fields, embodied intelligence is expected to achieve rapid development in the future. Embodied intelligence will enable intelligent agents to possess capabilities such as autonomous planning, decision-making, action, and execution, thus realizing an advancement in the capabilities of artificial intelligence.
Advanced Artificial Intelligence Capabilities
3 ) Three major trends and sparks of wisdom: The path to general artificial intelligence is becoming increasingly clear, and brain-computer interfaces are creating new ways of interaction.
✓ Artificial General Intelligence ( AGI ) refers to machine intelligence that possesses human-like thinking abilities, can adapt to a wide range of fields, and solve various problems. AGI is one of the important goals of artificial intelligence research. Narrowly defined artificial intelligence refers to AI that has made significant progress but is limited to specific fields, such as speech recognition and machine vision. We are currently in a stage where narrowly defined AI is relatively mature and the dawn of AGI is emerging. Currently, large natural language models, represented by GPT-4, are considered an important potential path to AGI. OpenAI CEO Sam Altman stated that the AGI era may arrive soon, and the industry may have super-powerful AI systems within the next decade.
The way humans communicate with artificial intelligence is also constantly evolving, with brain-computer interfaces (BCIs) expected to become the next generation of human-computer interaction. Currently, BCI technology is breaking through human physiological limitations, not only providing unprecedented possibilities for people with disabilities, but also potentially becoming the next generation of human-computer interaction.
4 ) Trend Four : The Power of Data: Massive amounts of data bring about an emergence of model capabilities, and high-quality data improves model performance.
✓ Advances in deep learning are built upon processing massive amounts of data with larger models. The GPT-1 model increased from 117 million parameters to the GPT-3 model's 175 billion parameters, resulting in significant breakthroughs in model performance and the emergence of new capabilities. However, the increase in the number of model parameters has led to a surge in computing power requirements, and the benefits of improving model architecture and parameter count are diminishing.
A study by research institutions including the University of Aberdeen and MIT shows that high-quality language data will be exhausted by 2026, while low-quality language and image data will be depleted between 2030 and 2050 and between 2030 and 2060, respectively.
✓ Data-centric artificial intelligence focuses more on the value of data, further driving breakthroughs in AI model performance. Stanford University Professor Andrew Ng proposed the Pareto Principle: 80% data + 20% model = better AI. A data-centric strategy can solve problems such as insufficient data samples and data bias. High-quality datasets become a key element in further improving model performance, and the value of high-quality data processing, data annotation services, and a sound data collection and evaluation system will become increasingly apparent.
5 ) Trend Five : AI Transformation of Data Centers : Intelligent Computing Centers Become Key Infrastructure
✓ Cloud computing is currently a crucial solution for providing AI computing power, leading to rapid growth in the AI server market. According to TrendForce data, in 2022, global AI server shipments accounted for approximately 1% of the total server market. With the explosive growth in demand for large-scale model training and inference, the demand for AI computing resources is expected to grow exponentially. According to IDC data, the compound annual growth rate of China's intelligent computing power is projected to reach 52.3% over the next five years, and the global trillion-dollar data center market will gradually transition from general-purpose computing to AI computing.
✓ Cloud computing is evolving from a CPU -centric homogeneous computing architecture to a CPU+GPU/NPU -centric heterogeneous computing architecture. It is estimated that the GPU stock space brought by large-scale models will increase from US$27.7 billion in 2023 to US$112.1 billion in 2025, and AI computing resources represented by GPUs will be in short supply in the short to medium term.
✓ As computing demands in specialized fields increase, AI chips are pursuing higher performance and lower power consumption, leading to continuous improvement in chip diversity and ecosystem richness. Some leading internet companies will focus on promoting independent research and development of AI chips, such as Google's efforts to develop the TPU, which is dedicated to deep learning, while continuously enriching its AI ecosystem.
6 ) Six Major Trend Models and C -End Roles: Personal Intelligent Assistants and the Next Generation of Traffic Entry Points
✓ Large language models will become personal intelligent assistants. Large models currently have the ability to access the Internet and manage memory. Through automatic task breakdown, planning, and implementation, they can autonomously fulfill user needs and become intelligent assistants for everyone, such as making travel plans and booking hotels and restaurants.
✓ Large models are becoming the next generation of traffic entry points. GPT-4 is gradually opening up its plugin functionality, connecting third-party applications through the underlying model to build a rich ecosystem. Since the opening of the plugin functionality, GPT-4 has integrated more than 500 plugins (including those for education, finance, and other scenarios). With the continuous enhancement of large model capabilities and the enrichment of the plugin ecosystem, large models are expected to become the next generation of human-computer interaction methods and traffic entry points. In May 2023, OpenAI's official website received 1.86 billion visits, ranking 19th globally in terms of internet traffic.
GPT-4 builds a rich application ecosystem
7 ) Seven Trend Models for B2B Applications: Professional data and cost-driven industry models flourish, opening up vast application possibilities.
✓ Data barriers have led to a proliferation of large-scale models on the enterprise side. General-purpose large-scale models can help users solve general problems, but when enterprises need to deal with data and tasks specific to their industry, they often need to fine-tune the basic model for their industry database. The characteristics and needs of vertical industries are not the same, so the application of large-scale models is showing a diversified trend.
✓ B2B applications, driven by economic considerations, will exhibit tiered and differentiated demands in the future. The commercialization of large models in vertical fields is more sensitive to the model's operating costs. The inference cost of a model is closely related to the number of parameters it has. This necessitates a multi-tiered product portfolio composed of large models with varying parameter sizes to achieve optimal economics in different scenarios and further enhance the richness of large models.
B-end large model multi-level structure
8 ) Trend 8 : Lightweighting of Models: Reducing Application Costs and Driving the Development of Edge Computing Power
✓ As the demand for miniaturization and scenario-based application of large models increases, and considering the economy, reliability and security of AI applications, inference in some scenarios will gradually expand from the cloud to the edge, driving a further increase in the demand for edge computing power.
✓ Currently, several large models have released " miniaturized " and " scenario-specific " versions. The PaLM-2 large model released by Google on May 23, the lightest version of which, "Gecko", can run on mobile devices, runs fast, and supports offline operation. Several other large models also have their corresponding smaller parameter versions.
✓ The deployment of large-scale model edge applications is accelerating. Edge computing power is developing rapidly, and Qualcomm has optimized Stable Diffusion through quantization, compilation, and hardware acceleration, enabling it to run on phones powered by the second-generation Snapdragon 8 mobile platform. At the Microsoft Build 2023 developer conference, Qualcomm showcased its latest edge AI capabilities and tools for developing generative AI on the next generation of Windows 11, and stated that large language models are expected to run on edge in the coming months.
9 ) The profound impact of the nine trend models: restructuring the labor market and rewriting the research paradigm.
✓ The impact of large language models on labor market structure is profound and complex. According to a research report by OpenAI in collaboration with the University of Pennsylvania, it is predicted that approximately 80% of the US workforce may be affected by large language models in at least 10% of their jobs.
✓ The application of large language models will bring about adjustments and changes in the labor market structure. In the short term, large language models may replace some low-skilled or repetitive jobs; in the medium term, they will also create new AI-related employment opportunities; in the long term, their application will profoundly change the work patterns and business models of various industries, making corporate organizational structures more flattened and smaller. This process requires individuals and businesses to actively adapt, developing uniquely human abilities such as innovation, collaboration, and social interaction, and co-evolving with artificial intelligence.
✓ The integration of AI with cutting-edge science has demonstrated enormous potential, significantly reducing the intellectual costs of advanced scientific research and improving research efficiency. Life sciences, weather forecasting, mathematics, molecular dynamics, and other cutting-edge sciences have all received extensive support from artificial intelligence. AI for Science will bring about a transformation in research paradigms and new industrial forms.
10 ) Trend Ten : Balancing AI Governance and Technology: AI Explainability Needs Enhancement, and Regulatory Urgency Is Increasingly Prominent
✓ In the rapid development of artificial intelligence, strengthening AI regulation is just as important as promoting the advancement of AI technology. While AI capabilities bring convenience to applications, they may also raise a series of issues such as data privacy, algorithmic bias, and AI ethics.
From a technical perspective, the credibility of AI can be enhanced through technologies such as explainable AI . Explainable AI makes the decision-making process of artificial intelligence transparent, increases the comprehensibility and trustworthiness of its output, and is crucial for building user trust in AI systems, improving system effectiveness, and addressing potential ethical issues.
From a regulatory perspective, governments around the world have begun to take action, formulating and implementing various AI policies and regulations. In April, China's Cyberspace Administration issued the "Administrative Measures for Generative Artificial Intelligence Services (Draft for Comments)," which clarifies the definition of generative artificial intelligence and sets bottom lines for the industry in terms of defining conditions and requirements, identifying responsible parties, establishing problem-solving mechanisms, and clarifying legal responsibilities.
