Kolmogorov-Arnold Networks

Kolmogorov-Arnold Networks, or KAN, are the model-architecture case [[LiuZiming|Liu Ziming]] discusses 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 says KAN began as Liu’s side project and grew from his long-running desire to build something that was both neural-network-like and symbolic/formula-like.

KAN matters in the episode less as an application and more as a research-process artifact. Liu says he became more interested in how one discovers a new architecture than in applying KAN everywhere. The architecture therefore becomes a bridge from Mechanistic Interpretability and Physics Of AI to AI For AI: can the process that produced a model structure be made explicit, repeated, and eventually automated?

Key Claims

• KAN represents Liu’s attempt to join connectionist neural models with symbolic or formula-like structure.
• The source ties KAN to the Kolmogorov-Arnold representation idea as theoretical motivation, while keeping the implementation story source-scoped to Liu’s prototype.
• Visualization was important because it made internal model behavior legible enough to change [[MaxTegmark|Max Tegmark]]’s initial skepticism.
• KAN’s deeper lesson for Liu is model-architecture discovery, not only a particular architecture’s benchmark use.
• The architecture supports Liu’s broader claim that AI needs stronger abstraction mechanisms beyond simply extending [[TransformerArchitecture|Transformer]] success.

Connections

• [[LiuZiming|Liu Ziming]] and Max Tegmark — source people.
• Mechanistic Interpretability and Physics Of AI — interpretability and scientific-understanding context.
• AI For AI, Research Taste, and Meta-Model Training Curve Prediction — architecture-discovery automation branch.
• Transformer Architecture and Hardware Lottery — contrasting model-design context.