Meng Fanqing / 孟繁青

Meng Fanqing / 孟繁青 is the [[EvolventAI|Evolvent AI]] co-founder interviewed in 从蒸馏到合成数据到 RSI，模型竞争的下一个焦点是什么？｜对谈 Evolvent AI 联创孟繁青. In the source, he describes himself as an NUS PhD student who previously interned at Kimi on agent and infrastructure work and participated in Kimi model releases.

His main contribution to the wiki is a data-and-environment view of [[RecursiveSelfImprovement|RSI]]. Meng argues that post-training advantage increasingly depends on Synthetic Agent Data, Environment-Based Agent Benchmarks, RSI Data, and fast validation loops rather than only on novel algorithms or raw compute.

Key Claims

• Model-lab post-training has become service-like enough that data construction, environment design, benchmark integrity, and iteration speed can differentiate researchers.
• Model Distillation / 模型蒸馏 is a practical shortcut, but the source does not treat it as the decisive cause of Chinese model progress.
• Evolvent AI is positioned between model labs and application companies, not as a foundation-model lab or a pure harness company.
• Good researchers for this work need hands-on engineering ability: building environments, running training, cleaning traces, and validating whether data improves models.

Connections

• Evolvent AI — company he co-founded.
• Kimi — prior internship and model-release context.
• RSIbench-data, RSI Data, Synthetic Agent Data, and Environment-Based Agent Benchmarks — main technical/data themes.
• Recursive Self-Improvement, Auto Research, Agent Post-Training, and Model Distillation / 模型蒸馏 — broader model-improvement frame.