Yuxuan Li
Chief Scientist, Intelligent Evolution, ModelBest;
Yuxuan Li is Chief Scientist of Intelligent Evolution at ModelBest. He holds a Ph.D. in Computer Science from Tsinghua University and completed his postdoctoral research as a Shuimu Scholar, specializing in high-performance computing. Admitted to Tsinghua University through the Informatics Olympiad, he went on to win five championships in seven domestic and international university supercomputing competitions, including world championships at ASC 2017 and ISC 2017. He previously served as a core technical lead at the National Supercomputing Center in Wuxi, where he led system-level performance optimization of major scientific computing applications on Sunway TaihuLight. His achievements included delivering the world's first full-system simulation of 50-photon Gaussian boson sampling. He also designed core simulation algorithms for the Jiuzhang photonic quantum computer, contributing to China's first demonstration of quantum computational advantage. The work was published in Science and selected as one of China's Top 10 Scientific Advances of 2020. He subsequently served as the technical lead in building an independent domestic AI chip computing infrastructure from the ground up, establishing a foundational software team of more than 50 engineers and coordinating cross-functional efforts involving over 100 people. He is now a co-first author of the MiniCPM model series, which has accumulated more than 65 million downloads worldwide. His current research focuses on Recursive Self-Improvement (RSI), pursuing the automation of the entire LLM development lifecycle through AI agents to enable models to progressively improve themselves. At the systems software level, he pioneered Forge Engineering, a new software engineering paradigm in which AI agents develop deeply customized, high-performance software for specific chips and workloads through an iterative build-and-verify loop. Under his leadership, the team has released three major projects, establishing a comprehensive Forge ecosystem spanning training frameworks, applications, and computational kernels. These include: - ForgeTrain: The world's first production-grade pretraining framework written entirely by AI. It has been used to train MiniCPM5 and delivers higher training efficiency than mainstream frameworks such as Megatron-LM and MindSpeed on both NVIDIA and Huawei Ascend platforms. - ForgeStencil: An AI-driven optimization system for industrial and scientific software. - ForgeKernel: A general-purpose kernel optimization framework that has topped both NVIDIA's official SOL leaderboard and Huawei Ascend's CANN Bench operator leaderboard, demonstrating the ability to support different chip ecosystems with a unified approach. He has co-authored more than 40 academic papers, with nearly 6,000 citations on Google Scholar. His honors include the China Youth May Fourth Medal (Collective) and Jiangsu Province's U35 Exploration Award. With research experience spanning supercomputing, quantum computing, domestic AI chips, large language models, and recursive self-improvement, he is a multidisciplinary technology leader bridging computing hardware and AI algorithms.
Topic
AI Building AI: From a Five-Level Evolution Framework to the Autonomous Generation of AI Infrastructure
AI building AI is a core pathway toward the next generation of artificial intelligence. We propose a five-level evolution framework that characterizes the depth of AI’s involvement in AI research and development, ranging from assisted suggestions to autonomous exploration. Through practical exploration, we have developed a unifying methodology—Forge Engineering. As the cost of AI-generated code approaches zero, the goal is no longer to maximize general-purpose reuse. Instead, each specific requirement can be addressed with a dedicated, vertically integrated implementation, allowing performance to approach its theoretical limits through case-by-case specialization. Based on this methodology, we have validated the approach through three engineering practices: An AI-forged training framework that outperforms Megatron-LM and supports production-scale pre-training. An AI-forged operator library that surpasses the best publicly available solutions across individual cases. An AI-forged inference kernel that achieves higher speed than mainstream engines such as SGLang. Outline The Five-Level Evolution Framework for AI Building AI The Core Principles of Forge Engineering Three Engineering Practices: Performance breakthroughs in a training framework, operator library, and inference kernel
