Applying Large Language Models to Automotive Industrialization: Practices and Insights

Overview – The session focuses on the practice and application of large language models (LLMs) in automotive industrialization, emphasizing manufacturing‑related cases and implementation experience.

1. ChatGPT Development History – Traces the evolution from GPT‑1 (2018) through GPT‑3, GPT‑3.5, to GPT‑4, highlighting the shift from limited generation quality to human‑level performance, multimodal input, and diverse capabilities such as code generation and professional exam solving.

2. Underlying Principles of LLMs

• BERT vs. GPT : BERT uses the encoder part of the Transformer, while GPT uses the decoder. Their pre‑training tasks differ—BERT employs masked language modeling (fill‑in‑the‑blank), whereas GPT predicts the next token.

• InstructGPT Improvements : Incorporates massive human‑preferred dialogue data and reinforcement learning, making responses more aligned with human expectations.

• ChatGPT Training Process : Consists of three stages – supervised data collection, reward‑model training, and reinforcement learning via PPO, forming a loop that continuously enhances the model.

3. LLMs Empowering New‑Era Industrialization

Implementation paths include building digital systems (order, scheduling, analysis, planning) and data systems that integrate IoT device data, forming a GPT‑based solution platform that generates various AI agents for downstream tasks.

Three application paradigms are described:

Instruction prompting – natural‑language commands to devices or systems.

Decision‑support – data‑driven assistance, e.g., AI‑based quality inspection with probability outputs.

Autonomous decision – fully AI‑driven scenarios such as autonomous driving.

4. Practice and Exploration in Industrialization

Platform architecture at NIO consists of chip, framework, model, and service layers, with AI agents built on top. An AI Agent comprises brain (GPT), memory, perception engine, planning, and task execution.

Three closed‑loops are introduced:

Data Loop : data ingestion → unified ETL → annotation (via LLM) → storage → model training → effect feedback, enabling continuous knowledge and data refinement.

Model Loop : corpus → model → training → evaluation → A/B testing → deployment, driving iterative model improvement.

Agents Loop : integrates data and model loops to automatically enhance agents, with feedback‑driven optimization and re‑training.

Application cases include:

Three‑Dimensional Intelligent Quality Inspection : combines computer‑vision LLMs, acoustic models, and predictive analytics to cover visual, auditory, and digital quality dimensions.

Cloud‑Edge Integrated Architecture : edge devices perform real‑time detection and upload data to the cloud, where continuous model training and closed‑loop updates occur.

G8D Agents : maps the eight‑step G8D problem‑solving methodology to eight specialized agents, enabling rapid, automated analysis and solution of complex quality issues while continuously capturing experience.

5. Q&A

Q1: How are G8D agents constructed? – Agents are built by combining historical data, system integration, and multi‑turn interactions using knowledge and plugins.

Q2: Where is knowledge stored and how is it retrieved? – In vector databases and Elasticsearch; vector search uses similarity distance, while Elasticsearch provides full‑text retrieval.

Q3: How is intent recognition optimized? – A hierarchical multi‑level intent system iteratively matches from fine‑grained to coarse intents, with fallback to human curation and user‑information‑enhanced re‑matching.

The session concludes with thanks to the audience.