Workflow
以产业大脑为例:从大模型、智能体到复杂AI应用系统的构建
Zhejiang University·2025-03-25 06:12

Report Industry Investment Rating No relevant content provided. Core Viewpoints of the Report - The reasoning ability of new-generation large models is continuously strengthening [172] - High-performance and low-cost reasoning models for a specific professional field can be trained based on high-quality small datasets [172] - Various complex intelligent application systems can be implemented based on large models through intelligent agents (AI Agents) [172] - Whether "reasoning large models + knowledge graphs (knowledge bases) + intelligent agents" will become the paradigm for future AI system development and application remains a question [172] Summary by Relevant Catalogs 1. Rapid Improvement of Large Model Reasoning Ability - ChatGPT is a large-scale pre-trained language model that learns from human feedback information after accumulating multiple types of technologies. Its success in 2022 marked the entry of conversational AI into the mass application stage [7][9][10] - The capabilities of large models have been continuously growing, with their performance in knowledge answering, mathematics, programming, etc., reaching new heights and exceeding human levels in many tasks. The parameter scale of large models has developed from tens of billions to trillions [16][19] - Early large models had obvious shortcomings in reasoning ability, prone to hallucinations, especially in mathematical reasoning, such as simple numerical comparison errors, weak multi-step reasoning ability, and inconsistent reasoning [20][24] - From 2023 - 2024, there were breakthroughs in reasoning ability. OpenAI o1/o3 showed excellent performance in mathematical and code reasoning tasks, and the open-source large model DeepSeek - R1 achieved an 87.2% accuracy rate on the MATH benchmark [35][38][40] 2. Reasoning Models and Chain of Thought (CoT) - Through technologies such as test-time scaling, reinforcement learning, and distillation, the reasoning ability of large models is continuously enhanced. Different models have their own characteristics in enhancing reasoning ability [45][46] - OpenAI - o series reasoning models generate a detailed internal chain of thought before answering questions, simulating human deliberation and improving the accuracy and depth of answers [47] - The chain of thought (CoT) is a way to break down complex problems step by step. Some models can generate and display the chain of thought to solve problems [52][54][58] - High-performance and low-cost reasoning models can be achieved through carefully designed small amounts of high-quality samples. For example, s1 and LIMO demonstrated good reasoning performance with a small number of samples [59][62][67] 3. What is an Intelligent Agent (AI Agent)? - Large language models (LLMs) lack the ability to interact with the physical world using various tools and human memory capabilities. Intelligent agents can enhance LLMs with these essential capabilities [78] - Taking the example of writing a research report on Tesla FSD and Huawei ADS, intelligent agents can perform task decomposition, use various tools to collect information, summarize content, and generate reports [80][83][84] - The Agent System has a five - layer cornerstone theory, including Models, Prompt Templates, Chains, Agent, and Multi - Agent. LLM - powered agents can sense the environment, make decisions, and take actions [96][98] - For more complex tasks, large and small models can collaborate in a generative intelligent agent. HuggingGPT is an example where the large language model is responsible for planning and decision - making, and small AI models are responsible for task execution [101][104][105] 4. Case of the Four - Chain Integrated Industrial Brain - There is a need for industrial cognitive decision - making at the national strategic level and industrial development decision - making at the social level. AI can promote the deep integration of the innovation chain, industrial chain, capital chain, and talent chain [118][122][129] - The industrial network chain large model is trained with massive industrial data and knowledge graphs. It can provide various services such as intelligence services, knowledge answering, and report generation, and has functions to address industry challenges [130][133][135] - The four - chain integrated knowledge computing engine, such as the SupXmind basic platform, can help users build intelligent decision - making systems. The industrial vertical domain large model iChainGPT has seven characteristic capabilities [144][147] - The industrial network chain large model has a specific composition and service framework, and can be customized according to customer needs. There are also many typical application scenarios at the provincial, municipal, and industrial cluster levels [148][153][162]