

909: Causal AI, with Dr. Robert Usazuwa Ness
36 snips Jul 29, 2025
Robert Usazuwa Ness, a Senior Researcher at Microsoft Research AI and founder of altdeep.ai, dives into the fascinating world of causal AI. He explains the significant differences between correlation and causation, emphasizing that not all variables are equally informative. The discussion covers advancements in Bayesian networks and the role of the 'do operator' in simulating causal relationships. Ness also highlights real-world applications, such as gaming data analysis, and the potential of large language models in causal inference, making this a must-listen for AI enthusiasts.
AI Snips
Chapters
Books
Transcript
Episode notes
Causal AI's Classical Roots and Revival
- AI systems largely evolved around correlation because classical causal methods rooted in statistics did not scale well.
- Modern probabilistic models and deep learning tools revitalized causal AI by managing latent confounders and complex inference.
Model Causal Assumptions Explicitly
- Causal analysis requires assumptions about the data generating process beyond just using available data.
- To reduce bias, explicitly model confounders and collect variables relevant to causal assumptions, not just what's convenient.
Gaming Side Quests Causal Example
- Robert gives an example of a gaming company data scientist exploring if side quests increase in-game purchases.
- Guild membership confounds this, requiring causal analysis beyond simple correlation.