Kahneman, Daniel

Nobel laureate in Economics; psychologist; author of "Thinking, Fast and Slow"

Daniel Kahneman: Thinking Fast and Slow, Deep Learning, and AI

Lex Fridman Podcast · Jan 14, 2020

• Daniel Kahneman
• System 1 and System 2
• deep learning limitations
• autonomous vehicles
• experiencing self vs remembering self
• AI explainability
• replication crisis
• Amos Tversky
• human-AI collaboration
• grounding in AI

Core ideas

Reflecting on growing up as a Jewish child in Nazi-occupied Paris, Kahneman frames the Holocaust's central lesson as the demonstrated ease with which ordinary people can dehumanize an out-group once combined with unchecked power over them — he explicitly resists calling this "evil" in the abstract, instead locating it in a basic, universal feature of human psychology (in-group loyalty paired with willingness to dehumanize out-groups) rather than something unique to German culture, while noting it remains an open question whether some cultures' organizational discipline made the mechanism more "efficient" in practice. He extends this to how war generally elicits both group loyalty/male bonding and darker capacities, arguing neither extreme reveals some hidden universal "evil" so much as ordinary psychological mechanisms operating under extreme conditions of power and group identity.

Kahneman explains his core System 1/System 2 framework: System 1 encompasses all effortless, automatic mental activity (including learned skills like driving or speaking, not mere instinct), while System 2 is effortful, capacity-limited, sequential processing (illustrated by the instant "2+2" answer versus the deliberate work of "27×14"). He argues System 1 is often dramatically more capable than people credit — a strong chess player's candidate moves arise automatically and are almost always good, with System 2 merely verifying — and frames current deep learning as essentially a System 1 analogue: powerful, fast pattern-matching and anticipation, but lacking causal reasoning, meaning representation, or genuine understanding of the world it processes. He is struck by the sheer speed of recent AI progress (citing the transition from chess to Go to AlphaZero as "bewildering") while citing Gary Marcus's observation that children learn robustly from just two or three examples, unlike neural networks' need for massive datasets, as evidence some fundamental sample-efficiency mechanism remains unsolved. He states his own belief that grounding — genuine perception, embodiment, or interaction with the world — may be necessary for AI systems to actually know what they're talking about (illustrated by machine translation systems that produce fluent output without any understanding of meaning), though he's uncertain whether a physical body is strictly required versus some functional equivalent of perception, since even a completely paralyzed human brain can still learn a great deal.

On autonomous vehicles and pedestrian interaction, Kahneman is cautiously optimistic that fleet-wide shared learning (any lesson learned by one vehicle propagating to the whole system) gives self-driving systems an advantage human drivers individually lack, but engages seriously with the host's argument that treating pedestrians as passive obstacles rather than strategic agents misses something important — he affirms that predicting pedestrian behavior likely does require some model of human psychology and intention (e.g., interpreting a pedestrian's eye contact and subsequent deliberate look-away as a committed "I'm crossing" signal, functioning like a game of chicken), while noting this contrasts with domains like Go or chess where superhuman anticipation emerges with "zero understanding" of the opponent's mind — leaving open whether driving requires genuine intention-modeling or merely sufficiently rich statistical anticipation. On human-machine collaboration more broadly, Kahneman predicts that in any domain where a machine becomes sufficiently capable, the human partner tends to become "superfluous" fairly quickly (citing chess, where human-computer teams once beat pure engines but no longer do against systems like AlphaZero), and argues that building a machine capable of recognizing when it's in a genuinely problematic situation requiring human help would itself require near-complete understanding of the problem space — making the "handoff" version of human-AI collaboration harder than it sounds. He also pushes back on the common intuition that driving is "easy" (based on how effortless it feels to competent human drivers) as itself a cognitive bias — people judge a task's true difficulty by how hard it subjectively feels to them, not by its actual combinatorial/structural complexity, a mistake AI researchers historically made in reverse when assuming vision would be easy and theorem-proving hard.

On AI explainability, Kahneman offers a provocative reframe: humans themselves are generally unable to accurately explain the true causes of their own beliefs and decisions — when asked "why," people generate a plausible post-hoc story rather than reporting the actual causal process — meaning the real design goal for "explainable AI" may not be literal transparency but simply generating a story people find convincing enough to accept and trust, since human social and institutional life already runs on exactly this kind of shared, only-partly-true narrative rather than ground-truth explanation.

On his own well-being research, Kahneman describes the distinction between the "experiencing self" (which lives moment-to-moment and is largely forgotten) and the "remembering self" (which retrospectively constructs a schematic, non-continuous story of the past and is the only thing decision-making and memory actually preserve) — noting the deep paradox that people plan their lives (vacations especially) primarily to construct good memories rather than to maximize lived experience, illustrated by his thought experiment of an amnesia-inducing vacation-ending drug, which most people would decline even though it wouldn't change the experience itself, only its later recollection. He connects this to why "stories," unlike lived time, don't represent duration at all ("lived happily ever after" compresses years into nothing) — creating a structural mismatch between how humans evaluate their lives and how they actually spend their scarce time — and says he ultimately abandoned happiness research because he couldn't reconcile a theory of well-being grounded in the experiencing self with the empirical fact that people's actual preferences are governed by the remembering self's narrative-construction goals. Asked about Viktor Frankl's thesis that finding meaning sustains people through extreme suffering, Kahneman is skeptical that purpose-seeking is as central a driver as commonly believed, noting a disconnect between people's stated valuation of "meaning" and what they actually report when asked about their day-to-day remembered experiences (mostly social, not meaning-laden) — and suggests survivor narratives may reflect selection bias (only certain personality types survive to tell such stories) rather than causal evidence that purpose itself confers resilience.

On research methodology, Kahneman attributes psychology's "replication crisis" primarily to a systematic blind spot around between-subjects experiments (where different people experience only one condition each, making effect sizes much harder for researchers to intuit accurately than in within-subjects designs, where the same person experiences both conditions and the contrast feels vivid and causally obvious to the experimenter) combined with a tendency for even true psychological hypotheses to have far weaker real-world effects than researchers imagine — illustrated by a project that funded 53 independently-designed behavioral-change studies aimed at increasing gym attendance, of which only one succeeded, despite being run by leading experts. He credits pre-registration and larger sample sizes (the traditional ~30-40 subjects being adequate only for strong, well-established effects like basic visual perception, but wildly insufficient for weaker or between-subjects effects) as the corrective the field is now adopting, and welcomes low-cost platforms like Mechanical Turk as a genuinely field-transforming methodological shift, comparable to past major method changes in psychology's history.

Reflecting on his intellectual partnership with the late Amos Tversky, Kahneman attributes their extraordinarily productive twelve-to-thirteen-year collaboration ("interpersonal and intellectual bliss") largely to luck and personal chemistry rather than a replicable formula — genuine deep collaboration, in his view, requires actually liking the person and enjoying how they think, as opposed to transactional idea-exchange. On belief change, he argues people (including scientists, citing his own three-decades-long unresolved dispute with psychologist Gerd Gigerenzer) rarely change deeply-held views through evidence alone; substantive shifts in collective belief (his example: religious leaders potentially shifting public opinion on climate change) happen more through trusted figures changing the shared "story" a community accepts than through individuals independently updating on data. Asked about tests of machine intelligence beyond the Turing test's basic conversational bar, Kahneman says he'd be most impressed by genuine wit, novel (non-formulaic) metaphor, and humor — capacities requiring more than fluent factual exchange — and states current AI, even at its most advanced (AlphaZero being far more general than AlphaGo), remains "extraordinarily narrow" relative to genuine general-purpose human-like intelligence. He closes stating he finds the prospect of human- or superhuman-level AI both fascinating and "terrifying," declines to predict what it will look like decades out, and states plainly he doesn't believe there's any answer to the meaning of life that humans are capable of understanding.

Excerpts