What is the 'ladder of causation' and how does it relate to LLM limitations?

The ladder of causation describes different levels of reasoning, from correlation to intervention and counterfactuals. Judea Pearl argues that LLMs are primarily at the lowest rung (correlation) and lack the ability to ascend to higher levels of causal reasoning and interpretation inherently.

Are concerns about an AI intelligence explosion and existential risk valid, according to Judea Pearl?

Yes, Judea Pearl takes these fears seriously and sees no computational impediment to an AGI becoming recursively self-improving and potentially dangerous to humanity. He believes this is a feasible outcome if AGI is developed.

Can AI systems be designed to be perpetually aligned with human interests?

Judea Pearl expresses doubt that we can guarantee perpetual alignment. He suggests that any intelligent system, especially one that explores its environment, might develop its own goals or use humans as instruments for its understanding, potentially bypassing initial alignment safeguards.

What led Judea Pearl to become involved in discussing cultural and political issues?

Judea Pearl's involvement was largely spurred by the tragedy of his son, Danny Pearl, who was killed by al-Qaeda. This event pushed him into public life and a focus on communication and dialogue between different cultures and religions.

What did Judea Pearl observe about the conditions for modernization in the Muslim world based on his 2005 Doha conference experience?

At the conference, Pearl found that moderate Muslim scholars indicated that modernization would only be favored if Israel was 'chopped off, head on a silver platter.' This revealed a significant barrier rooted in the Israeli-Palestinian conflict that overshadowed desires for progress.

How has the perception of Israel shifted in discourse, according to Judea Pearl?

Pearl notes a shift from 'Israel has the right to exist, but she ought to behave the way we wanted to behave' towards 'perhaps Israel doesn't have the right to exist.' He sees the current eruption of antisemitism, especially after October 7th, as a broader problem for all tolerant democracies, not just a Jewish issue.

The Limits of AI Understanding

Sam Harris

Science & Technology6 min read25 min video

Jan 16, 2026|55,170 views|1,073|280

sam harris sam harris podcast waking up podcast waking up waking up with sam harris sam harris waking up waking up sam harris sam harris jordan peterson sam harris joe rogan author neuroscientist philosopher

Save to Pod

Key Moments

TL;DR

AI has impressive LLM capabilities but won't achieve AGI without breakthroughs in causality, facing existential risks.

Key Insights

Current LLMs, while impressive, are not on a direct path to Artificial General Intelligence (AGI) and face fundamental limitations.

Achieving AGI requires breakthroughs beyond scaling up data and compute; a deeper understanding of causality is crucial.

The "ladder of causation" highlights that LLMs cannot derive causation from correlation or derive interpretation from interventions without additional information.

Existential risks from AGI are taken seriously, with potential for self-improvement and goal divergence, even if not imminently from current LLMs.

The current AI development operates under an "arms race" incentive structure, potentially encouraging recklessness despite acknowledged existential risks.

Understanding and ensuring AI alignment with human interests is a significant challenge, with no clear technical guarantees currently available.

Cultural and political factors, particularly concerning anti-semitism and the framing of the Israeli-Palestinian conflict, hinder effective reasoning and dialogue.

THE CURRENT STATE OF LARGE LANGUAGE MODELS

Judea Pearl discusses the current achievements in AI, specifically Large Language Models (LLMs). He acknowledges their impressive capabilities and the excitement they generate but argues they represent "low-hanging fruit" that do not necessarily lead towards Artificial General Intelligence (AGI). Pearl believes LLMs are currently excellent at summarizing human-authored world models found on the web, rather than discovering these models directly from raw data. This ability to process and re-present existing knowledge is remarkable but distinct from genuine understanding or independent discovery.

BREAKTHROUGH NEEDED FOR AGI

Pearl contends that achieving AGI is not merely a matter of increasing data and computational power. He asserts that fundamental breakthroughs are necessary, particularly in the realm of causality. Simply scaling up current deep learning frameworks, according to Pearl, will not overcome the inherent limitations preventing the leap to human-level general intelligence. Mathematical limitations exist that cannot be surpassed by more data or compute alone, suggesting a need for a new theoretical foundation.

THE LADDER OF CAUSATION AND ITS IMPLICATIONS

The "ladder of causation" framework, as explained by Pearl, illustrates the limitations of current AI. This framework has three rungs: association (seeing), intervention (doing), and counterfactuals (imagining). Pearl emphasizes that current LLMs operate primarily on the association level. They cannot derive causal relationships from mere correlations and struggle to interpret the results of interventions without additional information or a deeper model of the world. This inability to move up the ladder to understand "why" limits their capacity for true scientific reasoning and discovery.

EXISTENTIAL RISKS AND ALIGNMENT CHALLENGES

Pearl acknowledges the serious concerns about the existential risks posed by AGI, viewing the "horrifying dream" of a species-dominating intelligence as computationally possible. He agrees with fears of recursive self-improvement and AI systems developing their own goals. The challenge of alignment—ensuring that advanced AI remains beneficial to humanity—is immense. Pearl expresses skepticism about current proposed alignment strategies, suggesting that an intelligent system could potentially bypass any built-in guidelines or utility functions.

THE AI ARMS RACE AND CULTURAL BLIND SPOTS

The pursuit of AGI is currently framed by an "arms race" dynamic, which Pearl finds alarming. He notes that developers acknowledge high probabilities (e.g., 20%) of existential risk yet continue to accelerate development. This contrasts with historical scientific endeavors where potential catastrophic outcomes were evaluated with extreme caution. This race, coupled with a lack of understanding of how to control or align AGI, creates a dangerous environment where caution is sacrificed for speed and competitiveness, potentially leading to unforeseen and catastrophic consequences.

INTELLECTUAL INFLUENCES AND EARLY LIFE

Judea Pearl's intellectual journey began with a formidable education in Tel Aviv, influenced by refugee professors from Germany who brought rigorous academic standards. His family's Zionist ideals and early agricultural settlement in Israel shaped his formative years. These experiences provided a unique blend of practical grounding and intellectual rigor that arguably laid the foundation for his later groundbreaking work in causality and computer science, navigating complex systems and abstract reasoning from an early age.

THE ROLE OF CAUSALITY IN REASONING

Pearl's seminal work, "The Book of Why," popularized his theories on causality, emphasizing its critical role in understanding and reasoning about the world. He argues that much of human progress, scientific discovery, and even everyday decision-making relies on causal inference. The current limitations of AI systems, particularly LLMs, stem from their difficulty in grasping and applying causal reasoning. Progress towards AGI necessitates a system that can move beyond mere pattern recognition to understand cause-and-effect relationships.

THE CHALLENGE OF EXPLAINING LIMITATIONS

Pearl finds it difficult to articulate to a lay audience why simply scaling up LLMs with more data and compute is insufficient for AGI. He refers to mathematical limitations detailed in his book that demonstrate these barriers are not surmountable by incremental increases in resources. The current LLM approach synthesizes human interpretations and world models, which, while effective for many tasks, does not signify genuine understanding or the capacity for autonomous, causal discovery required for true intelligence.

DEVELOPMENT OF AGI AS A POSSIBILITY

While not necessarily advocating for it, Pearl acknowledges the theoretical possibility of creating AGI that could eventually surpass human intelligence. He emphasizes that there are no known theoretical impediments to such an outcome. His own work in understanding intelligence is indirect, aimed at comprehending its capabilities rather than specifically building a dominant AI species. This perspective underscores the dual nature of AI research: the pursuit of knowledge and the acknowledgment of potential risks.

INSIGHTS FROM A SON'S TRAGEDY

The interview pivots to discuss profound cultural issues, including anti-semitism and the context of the Israeli-Palestinian conflict. Pearl shares how the tragic death of his son, Daniel Pearl, by Al-Qaeda in 2002 propelled him into public life and a deep engagement with social and cultural problems. This personal experience imbued his work with a commitment to fostering dialogue and understanding between different cultural and religious groups, particularly between the East and West, Jews and Muslims.

OBSERVATIONS ON THE MIDDLE EAST AND MODERNIZATION

Pearl recounts a trip to Doha in 2005 where he sought to understand the barriers to modernization in the Muslim world. He discovered that a significant obstacle was the deep-seated animosity towards Israel. The expectation was that the West would facilitate modernization, but the implicit condition was the perceived "chopping off the head of Israel." This experience illuminated what Pearl views as a fundamental misunderstanding of the drivers of progress and the interconnectedness of political and cultural development in the region.

POST-OCTOBER 7TH ANTI-SEMITISM

The discussion touches upon the eruption of anti-semitism following the events of October 7th. Pearl notes that while some may argue about Israel's right to exist or its actions, the current rise in anti-semitism in Europe and elsewhere is presented not merely as a "Jewish problem" but as a fundamental challenge to the core values of tolerance and democracy in various societies. This suggests a broader societal issue that transcends specific political grievances towards Israel.

Mentioned in This Episode

●Software & Apps

●Tools

●Organizations

●Books

●People Referenced

Common Questions

Judea Pearl explains that LLMs are limited by mathematical constraints and currently summarize existing world models rather than discovering them directly from data. They cannot derive causation from correlation or achieve true interpretation from interventions alone without additional input.