From Compute to Intelligent Agents: MetaX's Open-Source Learning Projects

Keynote
Shanghai
2:10 p.m. - 2:50 p.m.
Venue D(Integrated Building, Room 326 Meeting Room)
Talk Poster
  • Zeyao Han Head of Application Ecosystem at MetaX

    Zeyao Han holds a PhD from the Department of Information Science and Electronic Engineering at Zhejiang University. He now leads the application ecosystem at MetaX, with a career spanning chip and wireless communication system architecture, AI Infra, AIGC and agent development, and Transformer algorithms, focused on agent applications, ecosystem incubation and research-and-education solutions built on domestic compute. He was previously a wireless communication pre-research system engineer and team lead at Huawei Shanghai Research Institute, SoC architecture lead for HiSilicon’s first-generation Kirin, Assistant Professor and Lab Dean at the School of Microelectronics of Shanghai Jiao Tong University, senior manager at Alcatel-Lucent, and senior director at Bell Labs in the United States, and has spent years building AI education products and AIGC and large-model application startups. He is the author of several books, including AIGC from Beginner to Practice, AIGC and Agent Development in Practice, and A Practical Guide to General-Purpose GPUs (Xiyun): From Environment Setup to Large-Model Applications.

    hanzeyao

Abstract

Centered on AI for Good, this talk traces MetaX's two open-source learning projects to show a complete path from underlying compute to intelligent agent applications.

Details

The first half introduces Transformer-on-MetaX, a 21-day open-source course built by MetaX: starting from a 220-line handwritten minimal Transformer in Python on CPU, migrating it to MetaX's domestic C500 GPU for real training, then diving into inference optimization through the Decoder-Only architecture and KV Cache, comparing it against industrial-grade inference frameworks such as vLLM and SGLang, and experiencing three operator development approaches — MACA C++, Triton, and TileLang. The talk shares the course design, real runtime data, and lessons learned by students.

The second half introduces the "AIGC and Agent Application Development in Practice" course series built on MetaX: centered on the Python ecosystem, six lessons progress through "generate, acquire, think, master, evaluate, collaborate," covering model APIs, tool calling, retrieval-augmented reasoning, skill accumulation, model evaluation, and agent orchestration. Each lesson produces a runnable application, culminating in a personal AI assistant that can autonomously select capabilities — pushing agents from being mere "conversational assistants" toward becoming "digital employees" that deliver results.