The forum, themed 'Digital Economy and Industrial Integration: Industry-wide Transformation Accelerated by AI', invited 18 heavyweight guests and industry experts. It brought together industry leaders, experts, and investment institutions to focus on new opportunities in digital economy investment and financing, and to jointly discuss the capital and industrial forces that enable enterprises to navigate cycles.
Lv Xin, General Manager of FastAG Product R&D Department at Beijing Dipo Technology Co., Ltd.
At this forum, Lv Xin, General Manager of the Product R&D Department of FastAG at Beijing DIP Technology Co., Ltd., delivered a keynote speech titled "The Era of AIGC: Exploration and Practice of Large Models in the Technology Industry." The speech mainly discussed enterprise operations based on data intelligence and introduced the basic model service system. It also analyzed and explained the performance indicators of the FastAGI capability toolchain and shared several innovative cases of large model applications.
I. Enterprise Operations Based on Data Intelligence
Lv Xin pointed out that the essence of digitization is a closed-loop enterprise operation system based on data and intelligence. The future enterprise service market requires the real-time integration of data and the agility of business value. Specifically, the core value of endowing enterprise operation data with intelligence lies in: (1) The application of distributed data lakes can break down data silos, improve performance, and save relocation costs; (2) Big data platforms support localization and continuous evolution towards lake warehouses; (3) Unified storage of multimodal data, combined with AI tools for unified analysis, helps enterprises achieve quantitative retrieval; (4) Data element governance and assetization to realize the ultimate value of data.
II. Basic Model Service System
Infrastructure can be divided into the model layer, tool layer, and application layer. Among them, the training and optimization of basic models are located in the model layer, while domain models and applications are situated in the application layer. Lv Xin believes that basic models and domain models depend on the readiness of pre-trained datasets, which include proprietary text, structured knowledge, and multimodal data. It is worth mentioning that pre-trained datasets can help models increase their professional knowledge, which is a level where models acquire more knowledge. Domain-specific data needs to be added to their corpora, such as concepts and nouns from various disciplines.
III. FastAGI Capability Toolkit
The FastAGI capability toolchain platform is based on the Deepexi 13B model for inference acceleration, achieving 33 Tokens/S and concurrent processing of 100 tasks with response time <1S. It supports multiple cards per machine, multiple models per card, load balancing + high availability, serving multiple user concurrency on a single card, and other performance metrics such as single-card intelligent Q&A, copywriting generation, data analysis, API interaction, information extraction, and document summarization. In addition, Lv Xin also introduced several core capabilities of the Deepexi AI Agent, including interactive precise retrieval of data/metrics, flexible generation of rich data charts, automatic call to system APIs, and dialogue interpretation with the unified data platform based on business logic.
IV. Application Innovation Cases Based on Large Models
At the end of his speech, Lv Xin listed several application cases of leading brand retail enterprise Babil Fashion based on the innovation of operation management using large models, establishing a data intelligence large model laboratory with China National Nuclear Engineering Technology University, the application of virtual museum curator models in cultural tourism groups, the models and applications in the field of digital carbon neutrality of China Southern Power Grid, and providing operation and driving manuals for new energy vehicle manufacturers.
