游凯超 / Kaichao Yu

游凯超 is the Infract co-founder and chief scientist interviewed in 148. 对游凯超3小时访谈：开源Infra、和模型Co-design 、“如果vLLM失败，我们会后悔一辈子”. The source presents him as a major maintainer of [[VLLM|vLLM]] and as a researcher whose path moved from ML algorithms toward machine-learning systems, open-source infrastructure, and inference-engine co-design.

His background in the episode runs through [[TsinghuaUniversity|清华大学]], a Berkeley research visit, and later work around vLLM. The source says he became drawn to systems after seeing that paper-level algorithmic improvements could be dominated by scale, data, hardware, and software implementation quality.

For the wiki, Yu matters because he ties together three branches that are often discussed separately: open-source governance, AI inference economics, and Model-Infra Co-Design. His account treats PagedAttention as vLLM’s origin, but argues that the durable work is maintaining an engine that keeps up with changing model architectures, hardware features, Chinese open-model releases, and agent workloads.

Connections

• Infract — company he co-founded to sustain commercial and open-source inference work.
• [[VLLM|vLLM]] — open-source inference engine he helps maintain.
• PagedAttention — technical origin of vLLM in the source.
• Model-Infra Co-Design, Hardware Lottery, and Open Source AI Infrastructure — main concepts his interview adds.
• [[TsinghuaUniversity|清华大学]] and [[UCBerkeley|UC Berkeley]] — education and research context.
• DeepSeek, Kimi, and Open Source AI Models — model ecosystem he discusses through the 2025 Chinese open-model wave.