Houyi Li 李厚意

At TikTok, I lead an AI Infrastructure team and have systematically strengthened its organization and engineering capabilities. I led an inference architecture overhaul that substantially improved production efficiency, and unified TikTok’s post-training frameworks. Our systems support training and inference for over a thousand LLMs and VLMs, handling TikTok’s massive real-world production traffic. I also explore agentic engineering and recursive self-improvement (RSI) for AI infrastructure, alongside team and engineering development.

Previously, I founded and led StepFun’s pretraining team, reporting to Xiangyu Zhang. I led pretraining and model architecture design for Step3, with hands-on responsibility spanning training data, core training code, and the main training runs. I also led Step2-mini end to end, from pretraining through post-training and deployment, and contributed to Step3.5-Flash’s architecture. My research included the Predictable Scale series—Step Law and Farseer—on hyperparameter scaling and predicting large-model performance. Alongside these papers, we open-sourced thousands of research models to support reproducible pretraining research.

Prior to StepFun, I was the LLM Lead at Alibaba International in 2023, reporting to Kaifu Zhang. I managed the full model lifecycle, from pre-training and post-training to driving the core LLM technology for commercial initiatives such as Aidge . I led the team that delivered multilingual large models Marco-7B and Marco-13B, and pioneered their path to commercial viability.

My earlier experience at Taobao and Kuaishou focused on personalized recommendation and advertising. I specialized in deploying deep learning at scale in consumer-facing (ToC) products to drive measurable business impact and boost profitability. In the same timeframe, I developed the classic personalized retrieval algorithm PDN (paper).

After graduating from Xidian University, I began my career in Ant Group's Infrastructure(Infra) team, where I spent four years with a focus on AI Infra. A key achievement from this period was developing the distributed training framework GraphTheta from scratch, one of the early frameworks in China to support thousand-machine parallel training.