Causal AI

Causal AI is the research direction discussed by Huang Biwei in 哪条路线，才能通往「世界模型」的终局？｜对话黄碧薇：Aether AI 创始人. The source frames it as a long-running field with roots in causal discovery, graph-based causal inference, potential outcomes, randomized controlled trials, and methods such as the PC algorithm and non-Gaussian causal-direction discovery.

Why It Matters

Huang’s definition of causality is intervention-oriented: if intervening on A changes the probability of B, A can be treated as causing B. She argues that this matters when AI systems need to generalize under changing conditions, hidden variables, missing values, and distribution shift rather than only learn surface correlations.

AI Applications

• For large language models, causal structure can be supplied externally through RAG or prompts, or learned internally through architecture changes.
• For robotics, causal modeling becomes Causal World Models, where the system learns variables, structures, and transition dynamics.
• For AI For Science, the same causal frame may matter in domains such as biopharma, materials, astronomy, and scientific discovery.

Connections

• Huang Biwei and Aether AI — researcher and company applying causal AI to embodied intelligence.
• Causal World Models — causal AI applied to physical world modeling.
• World Models and Embodied AI — deployment areas where causal generalization is treated as critical.
• OpenAI and Anthropic — frontier labs mentioned as mostly continuing the LLM route rather than fully causal modeling.