Dwarkesh PodcastMatjaž Leonardis - Science, Identity and Probability
CHAPTERS
- 0:00 – 1:44
Welcome, housekeeping, and setting up the conversation with Matjaž Leonardis
Dwarkesh opens the episode with updates about the new podcast, thanking early supporters and mentioning his writing and upcoming interviews. He then introduces Matjaž Leonardis and previews the themes: Popper-Miller, scientific identity, scientific progress, and polymathy.
- •Podcast intro and gratitude to listeners/supporters
- •Substack plug and mention of upcoming Robin Hanson episode
- •Guest introduction: Matjaž Leonardis and work with David Deutsch
- •Episode roadmap: Bayes/Popper-Miller, science as identity, polymaths
- 1:44 – 3:10
Is “science” a useful category—or an identity that backfires?
Matjaž argues that excessive talk about “science” and “scientists” can be counterproductive, encouraging people to worry about labels rather than problems. He notes that people successfully investigated nature long before the modern identity of “scientist” existed.
- •“Science talk” can distract from solving concrete problems
- •Over-identification with being a “scientist” can distort incentives and thinking
- •The term “scientist” is a relatively recent (19th-century) invention
- •People can understand nature without a special social role attached
- 3:10 – 4:28
Natural philosophers and the rise of a “special role” for studying nature
Dwarkesh asks whether earlier thinkers simply called themselves natural philosophers. Matjaž responds that people have always tried to understand the world, often without thinking of it as a special, separate activity, and questions whether the later role-conception is actually helpful.
- •Historical naming: “natural philosophy” as earlier framing
- •Understanding nature predates formal professionalization
- •A “special activity” narrative developed gradually
- •Question posed: does professional identity help or hinder?
- 4:28 – 5:47
Why universities exist even if “science” is the wrong framing
Dwarkesh challenges Matjaž: if the scientist identity is unhelpful, why fund universities and research roles at all? Matjaž distinguishes institutional support (books, people, events, community) from the idea that all such activity must be unified under a single concept called “science.”
- •Universities can be valuable infrastructure for inquiry
- •Institutional and cultural support helps knowledge-making
- •Support doesn’t require a single unified definition of “science”
- •Framing matters: activity vs. identity
- 5:47 – 7:48
A problem-focused view: many inquiries, not one monolithic “science”
Matjaž sketches an alternative: imagine institutions supporting many separate problem-communities (helium, star formation, etc.) without needing a grand category. He compares this to entrepreneurship—people build companies without obsessing over whether they’re “really entrepreneuring.”
- •Inquiry can be understood as many distinct problem pursuits
- •No need to police who is “really doing science”
- •Entrepreneur analogy: roles exist without method-policing
- •Unified labels invite unproductive gatekeeping
- 7:48 – 9:53
How the “scientific method” idea can become counterproductive
Matjaž argues that the “science/scientist” concept often smuggles in the notion of a single correct method, which can constrain thinking. He adds that self-conscious categorization (telling a story about what you’re doing) can impede performance, like self-consciousness in other domains.
- •“There is a method” can override the logic of the specific problem
- •Method-identity enables criticism like “that’s not science”
- •Self-conscious classification can harm creativity and judgment
- •Not universal, but a recurring failure mode
- 9:53 – 12:33
Debating method: Popper, Feyerabend, and skepticism about a single recipe
Dwarkesh defends the idea of a family of privileged methods (Popper, falsifiability). Matjaž counters that even Popper doubted a coherent “scientific method,” and notes Feyerabend’s critique in Against Method; at best, productive patterns exist, but they resist clean formalization.
- •Challenge to premise: Popper as critic of “scientific method” as a field
- •Feyerabend as another anti-method voice
- •Some reasoning patterns help, but they’re not fully understood
- •Thinking doesn’t always feel like “applying a method”
- 12:33 – 15:15
Did the Enlightenment succeed because of method? Induction, Hume, and causal ambiguity
Dwarkesh suggests the last 500 years of progress indicate better methods. Matjaž is noncommittal, questioning simplistic Enlightenment narratives and emphasizing how hard it is to assign causality (method vs. economics vs. other factors), invoking Hume’s induction problem as a key challenge.
- •Skepticism about tidy Enlightenment “method caused progress” stories
- •Possibility of alternative explanations (e.g., economic changes)
- •Hume’s problem of induction undermines naïve “reason from experience”
- •Causality in intellectual history is hard to disentangle
- 15:15 – 17:25
The Popper–Miller Theorem: why probability updates aren’t “inductive support”
Transitioning to Leonardis’ paper with David Deutsch, Matjaž explains Popper and Miller’s 1983 Nature letter. The theorem challenges the idea that evidence inductively supports universal theories via probability increases, by splitting a theory into a deductive part and an “inductive part” whose probability allegedly always decreases given the evidence.
- •Popper–Miller (1983) targets inductive interpretations of probabilistic support
- •Bayesian-style “evidence raises probability” doesn’t straightforwardly justify induction
- •Key move: divide theory into deductive vs. inductive components relative to evidence
- •Result: evidence decreases probability of the alleged inductive component
- 17:25 – 18:45
Objections and why Leonardis & Deutsch revisit the argument (including AGI relevance)
Matjaž summarizes the main criticisms: whether Popper–Miller’s “inductive part” captures what it claims, and ambiguity about what “inductive support” even means. He and Deutsch revisit the theorem with alternative explanations and argue it remains a live challenge for Bayesian approaches, including some AGI-related ambitions.
- •Two main disputes: definition of “inductive part” and meaning of “inductive support”
- •Later philosophy moved on, but the core challenge remains interesting
- •Authors propose new ways of interpreting/explaining the theorem
- •Implications for Bayesian reasoning in AI/AGI contexts
- 18:45 – 20:42
Popper’s broader theme: probability as “logical weakness” and the value of bold content
Dwarkesh asks how the theorem fits Popper’s wider work. Matjaž connects it to Popper’s interest in “content” measures: informative theories say more but are less probable; probability calculus can be reinterpreted as degrees of logical weakness, motivating skepticism toward probabilistic epistemology.
- •Popper’s focus on “logical content” (how much a theory says)
- •Parallel between constraints on content measures and probability calculus
- •Probability interpreted as logical weakness vs. explanatory power
- •Popper’s skepticism: we don’t want high-probability, low-content tautologies
- 20:42 – 22:48
Explanatory theories, the conjunction fallacy, and why “believability” isn’t probability
Matjaž uses the Linda conjunction fallacy to argue that people often prefer explanatory richness over probabilistic correctness. He notes that some languages encode probability closer to “believability,” suggesting our epistemic attitudes track explanation more than formal likelihood.
- •Linda problem: people choose the more explanatory conjunction
- •Explanation can drive judgments labeled “probability”
- •Distinction between formal probability and psychological believability
- •Motivation for preferring informative explanatory frameworks
- 22:48 – 25:56
Where Matjaž hesitates on Popper: the psychology of “need for regularity” and progress dynamics
Asked why explanatory theories should be preferred and where he disagrees with Popper, Matjaž offers partial agreement: explanatory theories can drive progress by generating falsifiable expectations and iterative replacement. His main hesitation is Popper’s psychological claim that humans have an inherent “need for regularity”—Matjaž isn’t sure it exists or how it works.
- •Role of universal theories in enabling progress via error discovery
- •Knowledge can accumulate through theory replacement, not just experience collection
- •Explanatory theories motivate exploration beyond observed data
- •Disagreement/uncertainty: the proposed innate “need for regularity”
- 25:56 – 29:43
Advice for aspiring polymaths: follow curiosity, avoid fake “levels,” and learn via context
Dwarkesh pivots to polymathy. Matjaž argues people are naturally curious, but become constrained by self-conscious “learning” narratives; he recommends pursuing interest directly, rejecting rigid fundamentals→intermediate→advanced ladders, and using history to recover the original problems that made ideas valuable.
- •Learning can happen through “osmosis,” not only systematic curricula
- •“Fundamentals/intermediate/advanced” are often artificial categories
- •Ideas were created for problems; textbooks often strip the motivating context
- •Polymathy is often about unlearning constraints and following the story
- 29:43 – 34:32
Finding problems and mentors: community connection, the ‘unsolved problems’ myth, and goodwill
Dwarkesh asks whether young people should seek problem-situations in existing knowledge. Matjaž emphasizes connecting with active groups; he argues that listing “unsolved problems” is harder than it sounds because problems are socially and contextually defined, creating a chicken-and-egg loop for newcomers—yet he remains optimistic because people are generally willing to help.
- •Best path: connect with communities doing real work
- •“List of unsolved problems” is difficult—problems are context- and value-dependent
- •Chicken-and-egg: contribution requires context; context often requires contribution
- •Despite idiosyncrasy, mentorship is achievable due to widespread goodwill
