Artificial intelligence has developed over nearly 70 years, with core capabilities continuously improving, covering:
Perception, cognition, decision-making, learning, execution, social collaboration
Development trends:
Moving toward intelligent machines that align with human emotions, ethics, and moral values
Current stage:
Transition from technology-driven to demand-application-driven phase
Technological evolution drives economic evolution, from old paradigms to new paradigms
AI development is at the strategic inflection point transitioning from Stage 3 to Stage 4
Definition:
Deep learning models with hundreds of millions of parameters or more
Utilize deep learning algorithms and neural networks to learn from massive datasets to improve predictive capabilities
Characteristics:
Performance follows a power-law relationship with parameter size, dataset size, and computation
Based on attention mechanisms, trained on large-scale, diverse, unlabeled datasets, with strong generalization
Classification:
By input type: Language large models, vision large models, multimodal large models
By application domain: General-purpose large models, industry-specific large models, vertical large models
1956–2006: Foundation laid for deep learning and neural network technology; large models began emerging
Post-2006: Development of NLP technologies and Transformer architecture laid the foundation for large model pretraining
2018: OpenAI released GPT-1; Google released BERT; pretrained large models became mainstream in NLP
End of 2022: OpenAI launched ChatGPT, triggering a global large model boom
2023: Explosion of large model training in China; “Hundred-Model Battle” phenomenon emerged
2024: Chinese policies accelerated industry adoption and commercial development
Jan–Jul 2024: Central and state-owned enterprises purchased over 950 large model projects
As of Nov 2024: 309 large models filed in China; vertical industry application accelerated
End of 2024: Domestic large model usage costs dropped, laying the foundation for widespread commercial adoption
Filings: According to the “Interim Measures for the Management of Generative AI,” 309 large models were filed in three batches
Application types:
General-purpose large models: 28%
Vertical large models: 72%
Application domains:
Internet, finance, healthcare, education, and industrial sectors all account for more than 10%
Two product evolution paths:
Increase parameter size, expand datasets, increase computation → achieve more powerful models
Optimize architecture and training strategy → achieve high cost-performance models
By end of 2024: Typical domestic large models cost dropped to below 0.5 CNY per million tokens
Commercialization models:
Customization (for large government and enterprise clients, 55%)
API and subscription (for SMEs and institutions, 40–45%)
Embedded in smart devices and apps with ad revenue (future trend)
Cases:
SPD Bank: Custom compute infrastructure and large model software meeting trustworthiness requirements
Apple & Baidu: iPhone integrated with Baidu Ernie 4.0 for smart assistant and image recognition
Scenario demand assessment: Technical capability evaluation, application scenario analysis, capability assessment
Deployment capability building: Large model capability system design, system development, data and algorithm preparation
Large model application deployment: Custom optimization, performance evaluation, full lifecycle management
Large model operations management: Real-time monitoring, dynamic tracking, continuous optimization
2022–2027: China’s AI large model application market CAGR 148%
2027 market size projected at 113 billion CNY
2024: Over 1000 public large model project awards, market size ~15.7 billion CNY
GPU demand increased by over 1.9 million units; compute investment in the scale of hundreds of billions CNY
80% of new compute used for top internet large model training and internal business, 20% for industry user capability building
DeepSeek-R1 redesigned the training process
Small SFT data + multi-turn reinforcement learning:
Improves model accuracy
Reduces memory usage and computational cost
Compute and performance approximately linear relationship
Core innovation loop: Compute, data, algorithm
Current: Product supply-driven
Next 3 years: End-user demand-driven, industry users actively drive application deployment
Pain points: Marketing and customer acquisition difficulties, low efficiency in risk management, product targeting challenges, high digital transformation costs
AI solutions: Intelligent marketing, risk control, precise customer insights, full-process digitalization
Pain points: Long waiting times, low examination efficiency, complex prescriptions, insufficient traditional medicine resources
AI solutions: Smart triage, assisted diagnosis, precision medication, knowledge management
Pain points: Low efficiency in multi-channel inquiries, scattered data, high customer acquisition costs, heavy customer service workload
AI solutions: Intelligent routing, centralized data management, automated customer service
Pain points: Scattered data, slow knowledge base updates, time-consuming document writing, slow inquiry responses
AI solutions: Intelligent data analysis, automatic document generation, real-time knowledge base updates
Pain points: Insufficient customer insights, difficult market forecasting, poor experience, low marketing efficiency
AI solutions: Intelligent analysis, personalized recommendations, precise marketing
Pain points: R&D knowledge hard to find, dependency on core technical personnel, low product yield, marketing disconnect, lack of scenario-based knowledge, difficult knowledge operations
AI solutions: Systematic knowledge management, intelligent recommendation, R&D collaboration, production optimization
Scaling Law faces challenges; research focus shifting from pretraining to post-training
Compute platforms tightly coupled with model innovation
MoE architectures widely applied to improve performance and efficiency
Large model toolchains improving, accelerating R&D and deployment
Three directions:
Integrated application software: Enhance user experience, productivity, and provide incremental monetization points
Smart assistants: Improve natural language understanding, generation, and multimodal capabilities
Agent: Multimodal large models support complex tasks, improving collaboration efficiency
Scarcity of high-quality Chinese datasets
Reasons:
Domestic professional data service industry at early stage; insufficient investment
Data transaction ecosystem not yet standardized
Difficulties in private data circulation; limited access to industry-specific data
