Agents that reason about causes, not just correlations

Picture a cat beside a table, a broken vase on the floor, cushions piled all around it. The speaker puts a simple question to today's strongest models: what would have happened if the cat had pushed the vase off the table? The answers are embarrassingly careless; one model declares the vase smashed while ignoring the cushions, another strings together a similarly memorized sentence. The sting is that the answer is not inside the photograph but one step beyond it. The cushions change the physics, yet an agent trained on co-occurrence patterns cannot see the difference. The entire talk grows out of this small scene.

The breakthrough of recent years is undeniable: artificial intelligence systems score uncanny hits on high-dimensional prediction tasks. Language, vision, and reinforcement-learning-backed applications rain down model after model , each week bringing a new record. The speaker greets this as a blessed era; there has rarely been a better time to be a computer scientist. Data is abundant, compute is abundant, engineering armies stand ready. So far nobody in the room disagrees. But once the applause fades, a harder question takes the table.

The infinite-data thought experiment

Suppose we had unbounded compute and unbounded data; suppose we paved half of Manhattan with data centers and multiplied the engineers tenfold. Would general intelligence simply emerge as the pile grows? The speaker frames this as a thought experiment and warns the answer is far from obvious. Stacking bricks means piling the same mortar ever higher; the wall rises but the building never changes its nature. What is missing may not be quantity but the mortar itself. The mood in the room shifts here, because the claim unsettles comfortable assumptions.

The field's leading voices have drifted toward the same line. Some say scaling alone is not enough and new ideas are needed for the next leap; others stress that sheer growth will never yield systems that understand the world. The keyword is understanding: does an agent genuinely grasp the world, or merely sequence sentences and frames with virtuoso skill? For the speaker this distinction rests on causality. The event page on the Berkeley site frames the talk as part of the quest for reliable autonomy; the October 5, 2026 Trustworthy AI workshop draws a line from hallucinations to trustworthiness. According to Berkeley, this is where the conversation belongs.

On the left of the cartoon sits the real world, on the right the agent , and Turing drew the interaction line between them: the machine that learns from experience. The speaker loves the frame but notes the word experience is badly overloaded. The episodes stuffed into that single sack are profoundly different; the sack must be opened and its contents sorted. The ACM summary of Pearl's program names the two great obstacles facing machine learning: adaptation and explainability, since systems fail on unseen conditions and cannot justify their verdicts. According to ACM, these are the walls scaling keeps hitting.

There are three kinds of experience. First, the spectator stance: the agent keeps its hands behind its back and watches the world through sensors, recording what unfolds. Second, intervention: it dives in, presses buttons, tampers with objects, and measures the effects of its own actions. Third, mental simulation: it withdraws into a corner and plays out what might happen, manufacturing the raw material of regret and responsibility. All three count as experience, yet their logical structures could hardly differ more; lumping them together resembles reading three languages with one dictionary.

The three rungs of the causal ladder

These three kinds of experience have formal counterparts in Judea Pearl's causal hierarchy, known in the lab as the PCH and recommended to curious readers alongside The Book of Why. Likened to the Chomsky layers of linguistics, the ladder shows which questions each language can ask. The first rung is the language of association: observed patterns, conditional probabilities, prediction. The second is the language of intervention: what would happen if the button were pressed, knowledge that arrives through action. The third is the counterfactual language: given what happened, what would have happened otherwise. Each rung contains the one below yet cannot be reduced to it; the upper rung's questions cannot even be posed in the lower rung's vocabulary.

The draft syllabus of the causal artificial intelligence book walks this ladder lecture by lecture; the opening chapter motivates the PCH, the second part covers structural causal models and diagrams, and later weeks turn to the calculus of intervention and counterfactual foundations. According to causalai-book, this is the curriculum the talk compresses into one afternoon. The second rung's modeling dialects are familiar structures from decision theory: causal Bayesian networks, Markov decision processes, partially observed processes. Where reinforcement learning sits in this picture gets only a passing remark; the speaker treats it as a narrow door into causality while the building beyond is far larger. For an agent that probes the world by acting and tallies costs and returns, the intervention dialect is a mandatory vocabulary.

The signature example of the third rung is the drug-and-patient parable: the patient took the drug and died; would he have lived without it? The sentence cannot be tested by repetition, since nobody gets a second life for the deceased. Yet the question is not meaningless; given a causal model , its truth value can be computed. The human mind speaks this dialect constantly: hindsight, introspection, blame, regret. Credit assignment and responsibility take root on this rung. Today's machine learning dialects have no counterpart for it; the available vocabularies describe what is and what could be, never what might have been.

The hardest knot is cross-layer inference. Synthesizing an interventional distribution from an observational one means forecasting the effect of touching the world without touching it; that is the causal community's central trade. One step further stands a sterner question: can observational plus interventional data license a counterfactual verdict? No naive blend of the two crumbs yields the third kind of knowledge; a formal bridge is required. The robust-agent study on arXiv turns this intuition into a theorem: any agent meeting a regret bound across a wide family of distribution shifts must have learned an approximate causal model of the data-generating process. According to arxiv, robustness and causal modeling are mathematically linked.

A five-capability roadmap

From here the speaker derives a checklist; five capabilities are required for anything calling itself causal intelligence. First, causal and counterfactual generation: rendering the world's response to a stipulated intervention faithfully. Second, causal understanding and explanation: justifying decisions by appeal to the world. Third, efficient and precise decision-making: picking the right action with few trials. Fourth, generalization and robustness: holding together under unseen conditions. Fifth, discovery: finding new principles about the world, like a miniature scientist starting from a blank slate. Current artificial intelligence systems draw nearly a blank on this list; the verdict is harsh but it sets the room's agenda.

The generation capability meets its examiner in the desk-lamp example: a photograph shows a switched-off lamp, unplugged, and the question asks what would happen if the switch were flipped. Anyone who knows the world says darkness continues; an unplugged switch produces only a click. The models instead write effusive paragraphs about lights coming on, because switches and lamplight co-occur in text. Token co-occurrences are no substitute for physics; when the word switch summons the word light, the model paints association rather than reality. Hundreds of similar trials change nothing; the error is systematic blindness, not a one-off accident.

On the explanation front comes the colored-digit experiment: is the verdict driven by the digit's shape or its hue? Standard attribution techniques cannot tell two behaviors apart; one model reads the shape correctly while the other cheats off the color, yet their attribution maps look identical. The causal counterfactual technique opens the curtain: it reports the cheat's verdict rests on color while the sound model tracks shape. Millions of GPU hours and doctoral labor go to waste on methods that ask the wrong question; a lens confined to correlations in the data can never show the mechanism in the world. The causal poverty of classical explainability tools is painted as the field's most expensive blind spot.

A world model playing Pong carries the credit-assignment problem into the arcade: frames go black for stretches, performance collapses, and the guilty transition is sought. Non-causal methods spread the blame across every transition; candidates point fingers everywhere and none of them hold still. The causal method singles out one transition; in that seed, transition number 536 is identified as responsible for the collapse. Mechanistic interpretability gets dragged onto causal ground by this example: opening the inner layers is not enough, one must ask what those layers correspond to in the world. Since most LLM inspection tools never ask that question, the speaker argues, the debt of explanation grows with scale.

Between fluency and understanding

The closing draws two axes: fluency and understanding. Today's language model products shine on the fluency axis; sentences flow, images dazzle, videos persuade. The causality community has deepened the understanding axis but stayed far from fluency; its language is mathematical, its reach narrow. The goal is a unification: a system that flows like English yet grasps the world causally. The speaker insists no band-aid will do it; pasting the two dialects together resembles taping quantum theory to relativity. The call goes out especially to young researchers, framed as their generation's defining problem. The chess observation from Yosefk sharpens the point: the model sparkles with memorized openings, then tries to move a knight that is not there by move nine; a system that never tracks board state may sequence moves without understanding the game. According to yosefk, fluency without a world state is exactly this.

Supporting evidence arrives on two flanks. On one side the arXiv result: a theorem-level link between robust generalization and causal model learning, so that any agent chasing regret bounds must eventually recover the world's mechanisms. On the other side, fresh counterfactual-reasoning benchmarks show large models retreating to memorized knowledge when familiar facts collide with novel ones. The speaker's laboratory page gives this program its institutional form: a doctorate under Pearl, a causal artificial intelligence lab at Columbia, a publication arc from reinforcement learning to fairness analysis. According to causalai, the agenda spans a decade of results.

The unifying frame is the forthcoming MIT Press book on causal artificial intelligence , merging probability theory, causal inference, and decision theory into a single roadmap from safety to generalization. The talk reads like the book's public rehearsal; the cat and lamp cases carry the first capability, the colored digit and Pong cases carry the second. Reinforcement learning and causal game theory are named and deferred to the book for lack of time. For readers who want trustworthy autonomy, the message is crisp: build the agent 's world model causally, and never mistake fluent sentences for understanding.