Applying Large Language Models to Automotive Industrialization: Practices and Experiences

Introduction The session focuses on the practice and application of large models in automotive industrialization, emphasizing industrial manufacturing cases and implementation experiences.

ChatGPT Development History Starting from GPT‑1 in 2018, the evolution progressed through GPT‑3, GPT‑3.5, and GPT‑4, with improvements in text generation, multi‑modal input, and performance that approaches human level, enabling natural‑language interaction for various tasks.

Fundamental Principles of Large Models The differences between BERT (encoder‑based) and GPT (decoder‑based) are explained, along with their distinct pre‑training objectives: masked language modeling for BERT and next‑token prediction for GPT. InstructGPT improvements, such as large amounts of human‑preferred dialogue data and reinforcement learning, are highlighted.

ChatGPT Training Process Training consists of three stages: supervised data collection, reward model training, and reinforcement learning via PPO, illustrating a continuous enhancement loop applicable to AI agents.

Large Models Empowering New‑type Industrialization Implementation paths include building digital systems (order, scheduling, analysis, planning) and data systems that integrate with large models to form a GPT solution platform. Three application paradigms are presented: (1) instruction prompting, (2) decision‑making assistance, and (3) autonomous decision‑making.

Industrial Practice and Exploration NIO’s platform architecture (chip, framework, model, service layers) enables AI agents to be woven into business scenarios. The agent logic consists of brain (GPT), memory, perception engine, planning, and task execution. Three closed loops—data, model, and agent—are described, detailing processes such as data ingestion, ETL, annotation, storage, model training, evaluation, and deployment.

Application Cases (1) Triple‑modal intelligent inspection combines vision, audio, and digital analysis using large models to reduce labeling costs and improve quality control. (2) Cloud‑edge integrated architecture ensures continuous model enhancement on the edge while leveraging cloud resources for training. (3) G8D Agents map the eight‑step problem‑solving methodology to eight specialized agents, enabling rapid, automated quality issue resolution and knowledge capture.

Q&A Sample questions cover the construction and scheduling of G8D agents, knowledge storage (vector databases and Elasticsearch), and intent recognition optimization.

Overall, the talk demonstrates how large language models and AI agents can transform automotive manufacturing through data‑driven, continuously improving intelligent systems.