Liu Ziming / 刘子鸣

Liu Ziming is the AI researcher interviewed in 149. 亲历中美 New Labs 资本狂潮，和清华刘子鸣聊：AI for AI、机制可解释性和 Max Tegmark. The source title names him 刘子鸣; parts of the episode summary render the name as 刘子明. In the episode he is described as an assistant professor at Tsinghua University’s AI school, a PI at Shanghai Qi Zhi Institute, and a participant in creating 圆环智能.

The source frames Liu’s career as a staged reversal. He began in physics at Peking University, moved to MIT for AI-and-physics work under Max Tegmark, spent time at Stanford, and then turned from using AI for physics toward using physics-like methods to understand and design AI. That turn becomes Physics Of AI and eventually Liu’s version of AI For AI.

His technical identity in the source is not only KAN. KAN matters because it shows his neural-symbolic interest and model-architecture taste, but he treats the deeper problem as learning how to generate the next major architecture. That leads to OPHIS Research Workflow, Meta-Model Training Curve Prediction, Training Autopilot, and Vibe Training.

Key Claims

• Liu sees AI as offering faster feedback and more open frontier problems than the physics path he originally imagined.
• His research theme is joining neural and symbolic modes, or more broadly joining AI with physics and natural science.
• He treats research taste and intuition as analyzable processes rather than irreducible genius.
• His AI For AI route is intentionally different from pure paper-agent or coding-agent automation.
• He thinks short-term AI research automation needs structure, mechanism, and collected research-process data before it can become a reliable product.

Connections

• 圆环智能 — company/lab vehicle connected to Liu’s AI for AI work.
• Max Tegmark, MIT, Peking University, Stanford, Tsinghua, and Shanghai Qi Zhi Institute — education, advising, and institutional context.
• AI For Science, Physics Of AI, Mechanistic Interpretability, and KAN — research trajectory.
• AI For AI, Auto Research, OPHIS Research Workflow, Meta-Model Training Curve Prediction, Training Autopilot, and Vibe Training — automation and product thesis.
• Research Taste, Transformer Architecture, Hardware Lottery, and World Models — broader technical judgment in the episode.