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

Summary

This 张小珺Jùn｜商业访谈录 episode interviews [[LiuZiming|Liu Ziming / 刘子鸣]] on the 2026 wave of AI [[NewLabOrganization|New Labs]], his route from [[AIForScience|AI for Science]] to Physics Of AI, and his attempt to build AI For AI through structured research rather than only coding agents. The source connects Mechanistic Interpretability, [[KolmogorovArnoldNetworks|KAN]], [[OPHISResearchWorkflow|OPHIS]], [[MetaModelTrainingCurvePrediction|meta-model training-curve prediction]], Training Autopilot, and Vibe Training into one thesis: AI research automation needs a “physics of AI” layer before it can reliably discover better model architectures. It also uses Liu’s [[PekingUniversity|Peking University]], MIT, [[StanfordUniversity|Stanford]], [[TsinghuaUniversity|Tsinghua]], [[ShanghaiQizhiInstitute|Shanghai Qi Zhi Institute]], and [[YuanhuanIntelligence|Yuanhuan Intelligence]] path to show how academic research, startup formation, and capital-market heat are merging.

Key Claims

• [[LiuZiming|Liu Ziming / 刘子鸣]] moved from physics into AI after finding more open problems and faster feedback in AI than in theoretical or experimental physics.
• His research path is framed as a reversal from AI for Physics to Physics Of AI: instead of using neural networks to solve physics problems, he wants to use physical-science methods to understand and design AI systems.
• Max Tegmark is presented as Liu’s doctoral advisor and as the person who pushed the group toward Mechanistic Interpretability in late 2022 and early 2023 because large models looked dangerous without internal understanding.
• [[KolmogorovArnoldNetworks|KAN]] began as Liu’s side project and became a model-architecture case for joining neural networks with symbolic/formula-like structure.
• Liu argues that [[TransformerArchitecture|Transformer]] benefited from Hardware Lottery and from language as a highly compressed human-made modality; future visual, physical, and world-model work may need stronger abstraction mechanisms.
• The source distinguishes AI For Science from Science of AI or Physics Of AI: the first brings neural tools into science, while the second brings scientific structure back into AI.
• Liu treats the current AI field as pre-Newtonian: it has data, empirical regularities, and a few dominant architectures, but not yet a deep compressed theory of why model designs work.
• [[NewLabOrganization|New Labs]] are framed as hybrid research-company organizations for directions such as Auto Research and World Models where large companies may see too much risk and universities may move too slowly.
• The Chinese financing environment around AI for Science, World Models, and AI For AI is described as extremely heated, with fast investor action and repeated seed/add-on rounds before products are mature.
• [[OPHISResearchWorkflow|OPHIS]] turns research into Observation, Problem, Hypothesis, Intervention, and Speed up so that papers and lab workflows can become trainable research data.
• Meta-Model Training Curve Prediction is Liu’s proposed “next curve” task: given a model, dataset, optimizer, and other conditions, predict the training curve before spending compute on full experiments.
• Training Autopilot and Vibe Training are the product horizon: users state needs and budget, while the system designs, trains, deploys, and delivers the model workflow.
• Liu contrasts his “smarter” Physics Of AI route with Recursive-style or RSI-adjacent approaches that he reads as more coding-agent and brute-force-experiment oriented.
• For Liu, AI For AI may be necessary for AGI but not sufficient, because broader AGI also needs abstraction and continual-learning ability beyond model-design automation.

Key Quotes

“太疯狂了” — Liu’s description of the May 2026 fundraising environment around AI for Science, world models, and AI for AI.

“AI for coding 之后会是 AI for research” — the episode’s summary of Liu’s central transition claim.

“AI for Physics of AI for AI” — Liu’s compact description of his non-mainstream route.

Connections

• [[LiuZiming|Liu Ziming / 刘子鸣]], [[YuanhuanIntelligence|圆环智能]], [[TsinghuaUniversity|清华大学]], and [[ShanghaiQizhiInstitute|上海期智研究院]] — guest, company, and current China research context.
• [[PekingUniversity|北京大学]], MIT, [[StanfordUniversity|Stanford]], and Max Tegmark — Liu’s education, doctoral advising, and startup-culture exposure path.
• AI For AI, Auto Research, Recursive Self-Improvement, and Recursive — research automation and self-improvement branch.
• Physics Of AI, Mechanistic Interpretability, AI Interpretability By AI, and [[KolmogorovArnoldNetworks|KAN]] — model-understanding and architecture-discovery branch.
• OPHIS Research Workflow, Meta-Model Training Curve Prediction, Training Autopilot, Vibe Training, and Research Taste — structured research-data and productization branch.
• AI For Science, World Models, Transformer Architecture, and Hardware Lottery — capital-market and technical-context branch.
• Academic AI Research Role and New Lab Organization — organization form between university lab and ordinary startup.

Contradictions

• No direct contradiction found.
• Productive tension to track: this source strengthens Auto Research and Recursive Self-Improvement but argues that coding-agent-heavy loops are not enough unless research process, mechanism, and failure modes are structured.
• Productive tension to track: the source is optimistic about [[NewLabOrganization|New Labs]] as research-company hybrids, while Academic AI Research Role preserves the case that universities still matter for long-horizon public-interest AI work.