Aleksander Molak interviews Judea Pearl

In this blog post, we’re taking a closer look at the remarkable contributions of Judea Pearl, often regarded as the “godfather of the causality movement.” Specifically, I’ll focus on the interview he recently had with Aleksander Molak on his podcast.

After watching the interview, I was deeply impressed. Given such a renowned figure in the field, my expectations were high, and I’m excited to say that they were more than met. The discussion is rich with thought-provoking ideas, seamlessly blending technical depth with accessible insights.

In this post, I’ll share the key concepts that struck me and explain why this conversation is essential viewing for anyone interested in causality and beyond.

The Evolution from Bayesian Networks to Structural Causal Models (SCMs)

Judea Pearl discussed the pivotal moment in his career when he transitioned from Bayesian networks to the formalism of Structural Causal Models (SCMs). This shift was marked by a focus on deterministic functions, which was initially a challenging concept after years of working within a probabilistic framework. Pearl and his collaborator, Thomas Verma, sought to formalize counterfactuals within this new framework, laying the foundation for much of modern causal inference.

The Influence of Early Education and the Legacy of Great Teachers

Pearl reflected on the significant impact his early education had on his career. Interestingly, he shared an anecdote about how both he and Daniel Kahneman, the Nobel laureate in economics, had the same influential teacher during their formative years. Pearl emphasized that such educators, who could engage with students on a wide range of topics without notes, were instrumental in shaping their intellectual curiosity and confidence. He lamented that such holistic and deeply knowledgeable teaching is rare today.

The Nature of Human Reasoning vs. AI Reasoning

When discussing the differences between human reasoning and artificial intelligence, Pearl highlighted the shortcuts that humans take due to resource constraints—shortcuts that lead to biases and errors. In contrast, AI systems, particularly future general AI, may not suffer from these limitations due to their ability to process vast amounts of information more thoroughly. However, Pearl cautioned against speculating too much about how AI will evolve, given that we lack the metaphors and vocabulary to fully predict AI’s future behavior.

Causal Discovery and the Role of Large Language Models (LLMs)

Pearl addressed the ongoing debate about the capabilities of Large Language Models (LLMs), particularly in their ability to learn causal relationships. He noted that while LLMs can learn from the causal models embedded in the text written by humans, this does not equate to learning causality from data alone. Pearl described the output of LLMs as a “salad of rumors”, a mixture of associations that, while potentially useful, lacks the rigorous structure of true causal reasoning. He also shared his cautious optimism that LLMs might eventually play a role in causal inference, especially when combined with causal models.

The Future of Research

Pearl expressed his enthusiasm for the future of causal inference, particularly in the field of personalized medicine. He highlighted recent advancements in quantifying harm and benefit at an individual level, an area he believes will be crucial in various fields, including medicine, political science, and marketing. Pearl emphasized the need for researchers to focus on individualized decision-making, where specific situational factors are considered rather than relying solely on population-level data.

The Divide in the Causal Inference Community

Pearl acknowledged the rift between different traditions within the causal inference community, such as those who focus on graphical models and those rooted in potential outcomes. He pointed out that, despite the logical equivalence between these frameworks, the divide often stems from differing comfort levels in articulating assumptions in one language versus another. Pearl emphasized the need for a unified approach to causal inference, suggesting that many researchers already think in terms of graphs, even if they express their ideas differently for publication.

The Role of Education

A recurring theme in Pearl’s conversation was the need for better education in causal inference. He argued that the limitations of randomized controlled trials (RCTs) and the broader implications of causal reasoning are not adequately covered in traditional statistics courses. Pearl stressed that understanding causality requires more than just knowledge of RCTs—it demands a grasp of complex concepts like necessary and sufficient causes, direct and indirect effects, and how these can be applied to real-world problems.

Reflections on Life and Advice for Learners

Towards the end of the discussion, Pearl reflected on the importance of being part of a larger chain of knowledge and contributing to it, no matter how small the contribution may seem. He advised those starting out in complex fields to begin with manageable problems that they feel confident in mastering before expanding their scope. This approach, he suggested, not only makes learning more accessible but also allows for meaningful contributions to the field.

In this post, I’ve summarized the key insights from this remarkable interview. However, I highly recommend watching the video to fully appreciate how Judea Pearl has transformed and continues to influence our way of thinking with a rare mix of humility and passion.

A special thanks to Aleksander Molak for conducting this invaluable interview.

I hope you find both the video and this summary insightful and inspiring.