“你有一把能够挖出金子的铲子，肯定不会先给别人用”｜对谈开物纪陆子恒：用AI发明新材料

Summary

This Shizilukou Crossing episode interviews Lu Ziheng, founder of Kaiwuji, about using AI to discover, validate, and commercialize new materials. The discussion frames AI Materials Discovery as more than candidate screening: useful systems must connect generative models, property prediction, senior materials judgment, lab experiments, kilogram-scale validation, customer trials, and production decisions. Its main business contribution is Materials Pipeline Company: if an AI model can find valuable material IP, Kaiwuji wants to use it internally to build materials rather than first sell it as a tool.

Key Claims

• Lu Ziheng frames materials as atomic structures that must be energetically plausible, synthesizable, property-matched, manufacturable, and commercially useful.
• The episode separates materials work into original compound or IP discovery, then process scale-up and sales; Kaiwuji wants to focus on high-value original IP while staying willing to push downstream when needed.
• The company is described as a roughly 10-person early team with large angel-plus financing but no revenue or profit yet, making it an early technical and commercial proof bet.
• AI Materials Discovery is presented as most valuable at the front end, where a single major material IP can create an industrial discontinuity.
• The workflow is not “AI decides and humans obey”: generated or screened candidates still need experienced materials scientists to judge which ones deserve lab work.
• A plausible material pipeline moves from business need, target properties, generated or searched candidates, expert filtering, gram-level validation, kilogram-level validation, customer line trials, and then internal or external scale-up.
• MatterSim helped the team believe that scalable materials models could generalize across properties such as phonon behavior and heat capacity even without being trained narrowly for each property.
• MatterGen and diffusion-style generation are discussed as examples of more scalable generative modeling for materials than earlier VAE-like approaches.
• Lu describes “material free energy” prediction as a near-term milestone: if a model can judge broad thermodynamic synthesizability, it could replace a large amount of exploratory experiment.
• The episode argues that model people, statistical-physics simulation people, and senior experimental “old masters” need to work together every day; remote handoff between disciplines is not enough.
• Computing and AI talent dominate current cost before production scale-up; lab work before kilogram scale is described as relatively cheaper than continuous model training.
• The source says the AI materials field has only high-level consensus that AI matters, while the key value point, model capabilities, and commercialization route remain unsettled.

Key Quotes

“能挖出金子的铲子” — Lu’s metaphor for why Kaiwuji does not want to sell its model before using it to find valuable materials.

“物质自由能” — the thermodynamic synthesizability milestone Lu hopes to push through in the next one to two years.

“一堆 Goodenough” — his shorthand for using AI to multiply rare materials-discovery judgment.

“开心吗？还折腾得动吗？” — the question Lu says he would leave for his future self.

Connections

• Lu Ziheng and Kaiwuji — guest and company building an AI-for-materials pipeline.
• AI Materials Discovery — core technical and commercial frame of the episode.
• Materials Pipeline Company — business model implied by owning the discovery and commercialization path.
• AI For Science — broader scientific and industrial category that materials discovery belongs to.
• MatterSim and MatterGen — model examples used to discuss scaling and generation in materials.
• Frontier Model Scaling — source-specific version of scaling through material models, data, diffusion, and compute.
• Domain Expert Alignment — materials experts remain necessary for candidate judgment, experiment design, and downstream process decisions.
• AI Commercialization Pressure — the company must turn expensive model training and early funding into material IP and commercial value.

Contradictions

• No direct contradiction with prior wiki content. The source extends the existing AI For Science theme from investment abstraction into a concrete materials pipeline and reinforces earlier warnings that AI capability still needs expert judgment, experimental feedback, and commercialization work.