909: Causal AI, with Dr. Robert Usazuwa Ness
🎯 Summary
Podcast Summary: 909: Causal AI, with Dr. Robert Usazuwa Ness
This 82-minute episode of the Super Data Science podcast, hosted by John Cron, features Dr. Robert Osa Zua Ness, Senior Researcher at Microsoft Research AI and author of the book Causal AI. The discussion centers on the transition from correlation-based machine learning to systems capable of genuine causal reasoning, drawing heavily on statistical inference, graphical models, and modern deep learning tools.
1. Focus Area
The primary focus is Causal AI and Causal Inference. Key topics included:
- The Three-Rung Ladder of Causation (Association, Intervention, Counterfactuals).
- The historical and current connection between Bayesian Networks/Graphical Models and modern Causal AI.
- The limitations of correlation-based AI compared to human/animal intuitive causal reasoning.
- The role of modern probabilistic programming languages (like Pyro, Stan, PyMC) in implementing causal models, particularly handling confounders (latent variables).
- The potential for Large Language Models (LLMs) to serve as causal knowledge bases.
2. Key Technical Insights
- Causal Models as Intervention Simulators: Any model capable of simulating the effect of an intervention (using the do-operator) can be considered a causal model. This allows practitioners to estimate outcomes post-hoc, mimicking randomized experiments when real-world intervention is impossible.
- Bridging Probabilistic Programming and Causality: Modern tools like Pyro (built on PyTorch) and NumPyro successfully integrate deep learning architectures (like VAEs) with the inference engines of traditional probabilistic programming (like JAGS/BUGS), making it feasible to build complex causal graphs that effectively handle latent variables (confounders), a historical weakness in earlier causal methods.
- Separating Concerns: Effective Causal AI requires disentangling statistical complexity (which deep learning excels at scaling) from explicit causal assumptions (which must be explicitly encoded, often via graphs or mechanistic statements).
3. Business/Investment Angle
- High-Stakes Decision Making: Causal AI is crucial in domains requiring high certainty about intervention outcomes, such as drug efficacy (vaccines) or product policy changes, where the burden of proof against false positives (Type I errors) is significant.
- Moving Beyond Prediction: The commercial value lies in shifting from merely predicting what will happen (correlation) to understanding what would happen if we acted (intervention), enabling better strategic decision-making in complex systems.
- Tooling Maturity: The integration of causal abstractions (like the
do-operator) into widely used libraries like PyMC signals the increasing accessibility and practical application of causal inference techniques for mainstream data science teams.
4. Notable Companies/People
- Dr. Robert Osa Zua Ness: Senior Researcher at Microsoft Research AI, author of Causal AI.
- Judea Pearl: Turing Award winner, creator of causal calculus, whose work heavily influenced the direction of Causal AI and was cited as inspiration for Dr. Ness’s book.
- AWS (Trainium 2) & Dell/NVIDIA (AI Factory): Sponsors highlighting the infrastructure required for large-scale AI development, including causal modeling.
- PyMC Labs (Thomas Viki): Mentioned in context of PyMC’s adoption of causal abstractions.
5. Future Implications
The industry is moving toward AI systems that can generate and understand causal narratives, mirroring human intuition. LLMs are beginning to function as causal knowledge bases, potentially outperforming traditional methods in certain scenarios by leveraging vast amounts of implicit knowledge about cause and effect. The future involves building more robust, human-aligned AI by embedding explicit causal reasoning structures alongside powerful correlation-based learning.
6. Target Audience
This episode is most valuable for hands-on practitioners including Data Scientists, Statisticians, AI Engineers, and Machine Learning Researchers who are looking to move beyond predictive modeling into prescriptive and explanatory AI systems. Professionals involved in high-stakes decision modeling (e.g., economics, medicine, policy) will find the technical distinctions particularly relevant.
🏢 Companies Mentioned
💬 Key Insights
"Often, frankly, the more interesting ones can't [be specified entirely in the form of a DAG]."
"Some of those assumptions can be specified entirely in the form of a DAG. And some of them can't. Often, frankly, the more interesting ones can't. You know, what carry a lot of them, DAG-based assumptions, long we make we need to make additional assumptions about mechanism."
"Level three is, it's kind of counterfactuals. And here we're asking questions where we're imagining what might have been different. So say, for example, I didn't get vaccinated and I got sick. Would I have gotten sick had I been vaccinated?"
"causality is kind of asking you to think more about the data-generating process than the data."
"can we, can we can, can we can strain it so that we can get certain guarantees? And one of the things that I'm working on is like kind of looking at the space of generative AI for video games and saying like, you know, to what extent can we get this generative AI to understand the underlying game mechanics or the underlying game physics, right?"
"let's imagine that we can kind of take, right, have a separate, say, generative model for each node in the DAG, that's conditional on its parents in the DAG, right? And they kind of connect this all together. And so, you can still get, you know, by, by implementing as a graph, the reflex causality, you still get all the benefits of theory, but you can also generate like you would from a generative model."