◆  Wei Liu, Xiaomi

Title: A Little Xiaomi Thought and Exploration on Mobile Agents

Short Bio: Director for LLMat Xiaomi's AI Lab, Industry Mentor for the Department of Psychology and Cognitive Science at Peking University, Examination Supervisor for Tsinghua University's Machine Learning Course, Academic Committee Member of CCF YOSCEF Headquarters, and an Early Member of Microsoft XiaoIce. Research Interests: Human-Computer Dialogue and Large Models. With over 10 years of experience in human-computer dialogue, he has led and deeply participated in the development of influential human-computer dialogue products in the industry, such as Microsoft XiaoIce and XiaoAi. He has received the 23rd Xiaomi Million Dollar Technology Award, the 24th CCF Computer Application Innovation Technology First Prize, the 24th Numerical Expo Outstanding Scientific and Technological Achievement Award, among others, and has dozens of patents and top conference papers.

Abstract: The application of large models in real-world scenarios is one of the most significant highlights of large models in 2024. From an application perspective, the core entry point for the deployment of large models is smart hardware, as only through integration with smart hardware can the full range of powerful capabilities of large models be leveraged. This report shares some Xiaomi thoughts and explorations from an industry perspective on the combination of smartphones and AI Agents. It starts from the practical problems faced by the application of large models and explores solutions and approaches, while also providing input for cutting-edge academic research.

◆  Dr. Linlin Li, Huaiwei

Title: Pangu Models 5.0 : Technologies & Challenges

Short Bio: Linlin Li is a principal researcher at Huawei’s Noah’s Ark Lab. Her major research interest is in the field of Natural Language Processing. Linlin Li received her Ph.D degree in Computer Science from Saarland University, Germany. She joint Microsoft after graduation and then Alibaba Damo Academy. She started working for Huawei in 2019 and has been working on NLP related products and projects, including Huawei Voice Assistant Xiaoyi and LLM Pangu. She has published over 30 papers in top NLP conferences and journals such as ACL, EMNLP, NAACL and Computational Linguistics.

Abstract: In this talk, we present our work of Pangu Models 5.0 with fully upgraded core capabilities, including more modalities, more powerful CoT, higher data efficiency, higher parameter efficiency and higher compute efficiency. Firstly, we present our work on improving data utilization and using better data to drive better model performance. Secondly, we present our new architecture with Ascend affinity, i.e., training 100B+ dense models on large clusters with 50+% MFU and 90+% linearity. Thirdly, we introduce the unified visual encoder and dynamic resolution approach for multimodality, and the key technology for CoT which empowers complex reasoning. Lastly, we summarize future challenges of Pangu Models including memory-efficient and high-performance MoE architecture, optimal parallelism strategy for multi-node inference, pre-training technology for triple-modal foundation models and complex task breakdown (XoT) for Agents.

◆  Qingliang Miao, Lenovo

Title: Lenovo AIPC Innovation Practice

Short Bio: Dr. Miao is a Principal Researcher at Lenovo Group. He is a member of the NLP Committee of the CCF. With over ten years of R&D experience in the fields of AI and NLP, his work covers various areas including LLM, knowledge graphs, and dialogue systems. He has published over 30 papers and applied for more than 100 patents. He has led the development of technologies that have been implemented in products and launched the industry's first AIPC product based on large model technology, helping the company to lead in the AIPC field.

Abstract: Lenovo released the industry's first AIPC product in April this year. The Lenovo Research has developed several key technologies for AIPC, including enhanced local knowledge services with smart retrieval, heterogeneous hybrid computing technology, leading model compression and inference acceleration, and hybrid intent understanding technology. This report will share with everyone the core technologies related to the intelligent agent demonstrated behind Lenovo's AI PC.

◆  Prof. Jidong Hu, ZTE

Title: Digital and Intelligent Transformation of ZTE Enterprises Exploration and Practice

Short Bio: Jidong Hu, Vice President of ZTE and General Manager of the Digital Technology Products Department, has many years of experience in telecom-grade R&D project management and enterprise digital transformation. He is fully responsible for ZTE's digital and intelligent transformation, including the formulation of strategies for digital-related products, product planning, operations, and high-quality delivery. He actively promotes ZTE's embrace of AI and is one of the main designers and drivers of ZTE's digital and intelligent transformation.

Abstract: In the digital economy era, digital transformation of enterprises is inevitable. After more than eight years of efforts, ZTE has gradually demonstrated the achievements of digital and intelligent transformation. In addition, ZTE has abstracted and built general capabilities for digital transformation of large enterprises from practice, and gradually improved a set of "Digital Nebula" that is integrated and open and can grow, embracing AI more actively and promoting the transformation of enterprises from "improving efficiency and improving quality" to "empowering and improving intelligence". This topic will share the practice and thinking of ZTE's digital and intelligent transformation, and in terms of transformation methodology, implementation path, and practice results how to build an ultimate cloud company, an ultimate trading company, and an ultimate AI company, guided by the transaction architecture and implemented by Digital Nebula.Together with you, we will explore how modern enterprises under the AI models  can navigate the deep waters of digital transformation.