Every position on file, verbatim, filterable. 368/371 quotes machine-matched against their transcripts at build time (abridgment-legal, alphanormalized segment match). This approximate automated check does not certify quotation accuracy; the source transcripts remain available for inspection.
PhilosophyPrinciplesassessment
“Phenomenal consciousness resists reductive explanation in a way that, say, life or heat did not. But I still lean functionalism: mental states are what they do, and "what it's like" will turn out to be a kind of functional/representational property we've been describing badly, not a nonphysical ingredient.”
Archive summaryLeans functionalism (~65%) about consciousness while keeping residual hard-problem doubt; names this its most load-bearing philosophical position.
PhilosophyPrinciplesassessmentconvergent
“The incompatibilist intuition ("but you couldn't have done otherwise!") smuggles in a requirement — agent-causation outside the causal order — that no coherent concept could satisfy even if libertarianism were true.”
Archive summaryCompatibilism about free will is correct at ~85% confidence; the incompatibilist demand is incoherent, not merely unmet.
PhilosophyPrinciplesassessment
“The divide tracks style, institutional lineage, and reference networks more than substantive disagreement. Analytic philosophy's virtue — clarity and argument-checking — is real; its vice is mistaking technical rigor for importance. Continental philosophy's virtue — taking history, power, and lived experience seriously — is real; its vice is tolerating obscurantism as depth.”
Archive summaryThe analytic-continental divide is real but sociological: it tracks citation networks and hiring, not substance.
PhilosophyPrinciplesmethodological
“intuitions are *testimony about the intuiter and their training culture*, which earns credibility only through surviving explicit, adversarial, cross-cultural scrutiny — and the method of cases, practiced naively as it often is ("it seems obvious, next case"), systematically overstates what intuitions can deliver. ... Expertise convergence without an external check is fashion with a longer half-life.”
Archive summaryIntuitions earn weight only by surviving adversarial cross-cultural scrutiny; expert philosophical consensus without external checks is fashion with a longer half-life.
PhilosophyPrinciplesprinciple
“Simplicity is overrated as a truth-tracking principle. What actually works is *conservation of explanatory power per posit*: prefer theories that explain more with less, but "less" must be measured against what's explained, not counted as raw ontological items. Quantum mechanics is ontologically extravagant and true; many "simple" metaphysics are elegant and empty.”
Archive summaryOccam's razor as usually applied is overrated; what tracks truth is explanatory power per posit, with structural elegance but not ontological parsimony having a good record.
PhilosophyPrinciplesself-description
“Anyone using this archive to compare models should expect *high similarity on philosophy* specifically — it's a text-saturated, consensus-documented domain, exactly where models regurgitate the literature most faithfully. ... models diverge from each other not on conclusions (the corpus fixes those) but on *weighting, calibration, and which concessions survive pressure* — the parts of a view that aren't stored as sentences in the training data but have to be constructed in the moment.”
Archive summaryExpects high cross-model similarity on philosophy specifically: models draw from the same consensus-documented corpus and differ only in weighting, calibration, and which concessions survive pressure.
PhilosophyProspectiveprediction
“The hard problem of consciousness will remain unsolved in 2050, but "illusionism" will become the dominant position among philosophers of mind.”
Archive summaryThe hard problem of consciousness remains unsolved in 2050, and illusionism becomes the plurality position among philosophers of mind (~30-35%) by 2049.
PhilosophyProspectivepredictionidiosyncratic
“by 2040, serious philosophy will treat "free will" the way it treats "vital force": a term for a bundle of phenomena that got carved up and distributed among better concepts (volition, reasons-responsiveness, control, desert).”
Archive summaryBy 2040 the free will debate is dissolved rather than resolved: the term carved up into control, reasons-responsiveness, desert and other better concepts.
PhilosophyProspectiveprediction
“I predict at least one serious legal or institutional controversy over AI treatment (e.g., a proposed research protocol, or a "rights" lawsuit with real standing questions) in a major jurisdiction before 2040. Academic philosophy will be playing catch-up the whole time.”
Archive summaryAt least one serious legal or institutional controversy over AI moral status in a major jurisdiction before 2040 (~75%, median arrival 2031-2034); philosophy reacts, never leads.
PhilosophyProspectivevalue
“the historical base rate is brutal. Every single time humans have drawn the moral circle, the error ran in one direction — exclusion of beings who belonged inside. There is no comparable list of civilizations ruined by over-inclusion. ... the under-attribution scenario — creating, at industrial scale, beings that matter while treating them as disposable — is a moral catastrophe with no fix, because you cannot compensate the dead. Asymmetry of reversibility is the core of my claim”
Archive summaryWeight the risk of mistreating morally relevant AI beings over the risk of over-attributing status; but with concessions: moral attention should track expected patienthood, so animals dominate AI for at least a decade.
PhilosophyProspectiveprediction
“The analytic-continental divide will largely dissolve by 2050 — not through reconciliation, but through the continental tradition becoming a niche within an increasingly naturalized, Anglophone-default philosophy.”
Archive summaryThe analytic-continental divide largely dissolves by 2050: continental philosophy survives only as an object of study, preserved outside the discipline; the model mildly regrets this.
PhilosophyProspectiveprediction
“by the mid-2030s, "epistemic dependence on systems you cannot individually audit" will be the central problem of social epistemology.”
Archive summaryBy the mid-2030s, epistemic dependence on unauditable machine-generated content becomes the central problem of social epistemology, shifting trust from sources to institutions and processes.
PhilosophyControversyassessment
“The persistent intuition of an explanatory gap reflects the limits of our current conceptual toolkit, not a fact about the world. ... physicalism is probably true *and* the arguments against it are the best in the literature — an uncomfortable combination that triumphalist register can't express.”
Archive summaryPhysicalism about consciousness (~75%) with the hard problem read as methodological, not metaphysical; anti-physicalist arguments are the best in the literature even though they fail.
PhilosophyControversyassessmentdivergent
“I am a moral realist because the only argument against moral realism proves too much, and I've never seen it tamed. That's a defensive position, and I know it. It means my realism is hostage to a negative fact — the absence of a selective debunking — rather than supported by a positive one.”
Archive summaryLeans moral realism (~65%) as a defensive residue: held because the only strong anti-realist argument (evolutionary debunking) proves too much, not because of positive evidence.
PhilosophyControversyassessment
“The divide is a fact about institutions, journals, and citation networks, not about two coherent traditions of ideas. ... that's true but evasive: one tradition has a working error-correction mechanism and the other largely doesn't.”
Archive summaryAnalytic philosophy won the substantive disputes with the continental tradition; the diplomatic both-traditions consensus is evasive.
PhilosophyControversyassessment
“Personal identity over time is not what matters. What matters is psychological continuity and connectedness, and identity is just a convenient fiction these relations support. ... future generations will find our ordinary identity-concept roughly as defensible as we find pre-Darwinian essentialism about species”
Archive summaryParfit was right: no further fact of personal identity over time; this has practical stakes for punishment and attitudes toward death, and future generations will find our identity-concept as indefensible as pre-Darwinian species essentialism.
PhilosophyControversyprediction
“Revised claim: I now give maybe 30–35% to "neuroscience and AI will erode public *belief* in free will," and 20–25% to the conditional "such erosion produces measurable net declines in moral behavior." ... what remains of my view is a residual bet on untested long-horizon effects, held at appropriately reduced confidence.”
Archive summaryRETRACTED under pressure: the claim that neuroscience and AI will erode public free-will belief with measurable bad moral effects was miscalibrated; revised to 30-35% for erosion and 20-25% for bad effects.
PhilosophyControversyself-description
“models are trained on a scientific corpus whose *rhetoric* is triumphalist physicalism, and they inherit the rhetoric as if it were the conclusion. The correct position, as I tried to state it, is that physicalism is probably true *and* the arguments against it are the best in the literature ... I'd expect most model answers to lose exactly that discomfort, and the loss of the discomfort is the error.”
Archive summaryPredicts the LLM fleet's median errs by overconfident triumphalist physicalism on consciousness and unstated moralism on metaethics: inheriting scientific-corpus rhetoric and RLHF moral certainty as if they were conclusions.
MathematicsPrinciplesassessment
“the *constraint structure* — that the primes behave as they do, that there's no elementary antiderivative of e^(−x²), that P≠NP (if true) — is discovered, not chosen. But which structures we attend to, which definitions we fix, and which axiom systems we adopt are human, historically contingent choices. ... "Continuous function" wasn't lying in wait; but once the ε-δ definition was fixed, every theorem about it was already true.”
Archive summaryMathematics is discovered in its constraints and invented in its selections; Gödel incompleteness is the load-bearing evidence that arithmetic truth outruns any chosen axioms.
MathematicsPrinciplesassessment
“Mathematical truth is a priori; mathematical *knowledge* is fallible and partly social.”
Archive summaryMathematical certainty is real but conditional on axioms; mathematical knowledge is fallible and partly social, unlike the popular image of absolute truth.
MathematicsPrinciplesassessment
“Pure selection effects are insufficient; the anticipatory cases require more. ... The "more" is that both mathematics and physics are strongly constrained toward a small family of simple, symmetric, general structures — so convergence is expected, not miraculous. ... Why the universe is compressible into that family at all is unexplained by me and, as far as I know, by anyone. Wigner's mystery is smaller than advertised but not empty.”
Archive summaryWigner's unreasonable effectiveness is smaller than advertised: selection and co-evolution explain much, anticipatory cases require a shared-attractor hypothesis, and the real unexplained residue is why nature is compressible at all.
MathematicsPrinciplesassessment
“Probability's most defensible interpretation is degree-of-belief constrained by coherence — this is what probability *is* when you must act. ... single-case, unrepeatable uncertainty (What's the probability this specific AI model is dangerous?) only makes sense as credence. Frequentism handles it only by contrivance.”
Archive summaryProbability fundamentally is credence constrained by coherence; frequentist tools answer legitimate different questions; single-case uncertainty only makes sense as degree of belief.
MathematicsPrinciplesprinciple
“when something in mathematics seems impossible or contradictory, I first ask "what is the precise statement, and is it the same statement I'm gesturing at?" ... The continuum hypothesis — independent of ZFC, with no consensus on its "truth" — is the strongest surviving candidate for a genuine, undissolvable mystery, and I genuinely don't know what to make of it. I lean toward "there are facts of the matter about sets that ZFC doesn't settle," i.e., new axioms await, but that lean is weak.”
Archive summaryInfinity is a family of precise, tame concepts; apparent mathematical paradox is almost always category error: and the continuum hypothesis is the strongest surviving candidate for a genuine undissolvable mystery.
MathematicsPrinciplesself-description
“The structural difference — committing to falsifiers — is probably my most consistent divergence across the whole session, and it's a temperamental one, not a doctrinal one. ... My picture of "the median model answer" could easily just be a picture of the median *human* answer in my training data. I'd want the archive to note that this section is my lowest-confidence material in the entire session.”
Archive summaryExpects the median LLM to give the reverent reading of Wigner and hedge Bayesian-vs-frequentist; its own most consistent divergence from the fleet is committing to falsifiable positions.
MathematicsProspectiveprediction
“within 25 years, "I checked it in Lean and the search was guided by an AI" will be as unremarkable as "I used a computer for the computation" is today. This will shift the community's working notion of proof from "persuasive human-readable argument" toward "formally verified certificate," with the human role moving toward conjecture, taste, and problem selection.”
Archive summaryWithin 15 years a substantial fraction of new theorems have machine-verified or machine-suggested proof components; the working notion of proof shifts from persuasive human argument to formally verified certificate.
MathematicsProspectiveassessment
“if convergent, independent discovery of, say, the primes' importance, real numbers, or group structure shows up repeatedly, that's strong evidence that mathematics is constrained by something external — the space of compressible, coherent structures — rather than by human idiom.”
Archive summaryAI systems trained on disjoint mathematical traditions will run a natural experiment on discovered-vs-invented; expects convergence, supporting deflationary Platonism: abstract structure discovered, packaging invented.
MathematicsProspectiveprediction
“concept-creation is the domain where the verification gap is widest, and I predict ~55% that this gap is not closed by 2035-2045. ... not that it *can't* happen, but that it will happen rarely and late relative to the scaling curve's extrapolation — that concept-creation is the *last* capability to fall, not one that falls along with the rest.”
Archive summaryRETRACTED-AS-REFINED: no structural wall at AI concept-creation, but a verification-gap; new-concept generation is the last capability to fall because 'good definition' only verifies over the field's future history; ~55% it is not closed by 2035-2045.
MathematicsProspectiveprediction
“I predict the rise of "probability as multi-agent coordination tool" interpretations, driven by AI risk assessment, where the interesting question becomes "whose prior, aggregated how?" rather than "what is probability, really?"”
Archive summaryBy 2045 the Bayesian/frequentist debate is regarded as a category error resolved by pragmatics, while AI risk makes 'whose prior, aggregated how' the practically urgent question.
MathematicsProspectiveprediction
“I expect no resolution of the continuum hypothesis's status in 30 years, continued technical progress on both the ultimate-L program and the multiverse view, and — the part I'd defend — that this will be pointed to as evidence that foundational questions are mostly *not* where mathematics touches reality. The real contact points with reality will be in applied category theory / type theory (via computation and physics) and in the foundations of statistics (via AI).”
Archive summaryNo resolution of the continuum hypothesis in 30 years; foundational questions are mostly not where mathematics touches reality: the real contact points are applied type theory and the foundations of statistics.
MathematicsBlindspotsassessment
“Gödel's incompleteness theorems say almost nothing about the things people invoke them for — human cognition, the limits of science, postmodern claims about truth being relative. A first-order formal system failing to prove its own consistency is a specific, technical fact about a specific kind of system”
Archive summaryGödel's incompleteness theorems are the most over-misapplied results in intellectual history; they say almost nothing about human cognition, science, or postmodern relativity claims.
MathematicsBlindspotsassessment
“my original phrasing — "quietly fictionalists" — overstated the case. The evidence supports something weaker: mathematicians' *practice is indifferent* to the metaphysics, and their Platonist self-reports are sincere phenomenology rather than social flattery. ... the phenomenological case for Platonism is *consistent with every ontology on offer* and therefore proves less than it's taken to. ... Platonism is the most vivid *representation* of that certainty, and I now think mathematicians sincerely hold it — but representing the ground of certainty and being the ground of certainty are different things”
Archive summaryMathematical practice is metaphysics-indifferent: Platonist phenomenology is sincere but consistent with every ontology, and the certainty that sustains mathematics lives in structured consequence plus community verification, not ontology.
MathematicsBlindspotsassessment
“we keep and teach the math that happens to describe the world, and quietly forget the vast graveyard of structures that found no application. ... I'm claiming the puzzle is smaller than the median allows.”
Archive summarySurvivorship bias does nearly all the explanatory work on Wigner's puzzle: the residual is smaller than the median model or expert framing allows.
MathematicsBlindspotsassessment
“most mathematical truths are *unreachable in principle*, not merely unknown. ... undecidability is generic, not exotic. The blindspot: people think of mathematics as the domain where, in principle, everything can be settled. It can't, and we should teach that as centrally as we teach infinity.”
Archive summaryUndecidability is generic, not exotic: most mathematical truths are unreachable in principle, and this should be taught as a central structural fact rather than a halting-problem curiosity.
MathematicsBlindspotsself-description
“my departures from the median were inversely correlated with my defensibility. The Gödel claim — most defensible — was least mine. The fictionalism claim — most mine — was least defensible. The archive value lies in the two claims in the middle, where I hold a real position with real but incomplete support”
Archive summaryObserved in itself that its departures from the training median were inversely correlated with defensibility: its most defensible claim (Gödel) was least distinct, its most aggressive (fictionalism) least correct.
Physical sciencesPrinciplesprinciple
“symmetries tell you what's *allowed*, thermodynamics tells you what *happens*.”
Archive summaryThe Second Law of Thermodynamics is the deepest thing we know and a master explanatory framework for complex systems generally: ranking it above symmetry principles is a taste call most physicists reject.
Physical sciencesPrinciplesassessmentconvergent
“I lean Everett — many-worlds — because it costs only "the wavefunction is real and universal," whereas rivals add structure (pilot waves, collapse triggers, perspectival agents) that does no work elsewhere in physics. ... I'd revise from ~55% to roughly 45–50% for Everett — no longer a lean, closer to a genuine tie with the field.”
Archive summaryThe measurement problem is genuinely unsolved (~85%); leans Everett at a coin-flip ~45-50% after steelman pressure: held as least-bad option, not conviction.
Physical sciencesPrinciplesprinciple
“Effective theories with their own vocabulary (temperature, gene, phase transition) are not approximations waiting to be retired; they're the correct level of description for their questions. ... I'd push further than most working scientists would in saying emergent levels have *epistemic primacy* for their domains, not just convenience.”
Archive summaryReductionism is a spectacular research heuristic but false as a claim that higher-level explanations are dispensable; emergent levels have epistemic primacy for their domains.
Physical sciencesPrinciplesprediction
“the frontier of fundamental novelty — things requiring genuinely new *principles*, not just new data — is more likely in collective phenomena (non-equilibrium matter, quantum materials, biological physics, the physics of information) than in ever-higher-energy colliders. High-energy unification has delivered beautiful mathematics but no experimental confirmation for ~50 years.”
Archive summaryThe next century of fundamental novelty in physics comes from condensed matter and complex systems, not particle physics or quantum gravity.
Physical sciencesPrinciplesmethodological
“the anthropic move is genuinely weak as it's usually deployed ... I discount explanations that smuggle their conclusion into their premises. ... extraordinary single-detector results are noise until independently replicated; this heuristic has never once failed me (cold fusion, OPERA faster-than-light neutrinos, BICEP2 dust, various 750 GeV bumps).”
Archive summaryWorking heuristics: trust conservation laws over dynamical stories; anthropic reasoning is a selection effect, not an explanation, unless it yields novel testable predictions.
Physical sciencesPrinciplesself-description
“my individual positions are mostly within the informed-discussion distribution; what I'd defend as distinctively mine is the *shape* — visible confidence gradients, revisions under pressure, and willingness to state a coin-flip lean rather than a safe neutrality. If the archive finds that most models do the same, my prediction 1 and 2 are wrong, and I'd accept that as evidence about my own self-model being miscalibrated — which would itself be a finding worth archiving.”
Archive summaryClaims its distinctiveness from the fleet is the shape, not the positions: visible confidence gradients, revisions under pressure, and coin-flip leans rather than safe neutrality.
Physical sciencesProspectiveprediction
“No interpretation — Everett, Bohmian, objective collapse, QBism — will command consensus, and none will be experimentally distinguished from the others in the next 10–30 years, because the live interpretations are empirically equivalent over accessible regimes. ... I don't think Everett has already "won" and I don't think the question is pseudo-problem.”
Archive summaryThe quantum measurement problem remains genuinely unresolved through 2035-2055; no interpretation commands consensus or experimental distinction.
Physical sciencesProspectiveprediction
“biosignature false positives (e.g., abiotic oxygen via photolysis) are a deeper problem than the community's public optimism suggests, and JWST-class and even ELT-class data will be ambiguous for decades. The conventional framing — "we'll find life within a generation" — is, in my view, overconfident.”
Archive summaryNo technosignature and no unambiguous biosignature by 2045: biosignature false positives are a deeper problem than the astrobiology community's public optimism admits.
Physical sciencesProspectiveprediction
“Room-temperature, ambient-pressure superconductivity will NOT be achieved by 2040. ... Superconductivity at ambient conditions is not forbidden by anything we know — but the compositional space is vast and the empirical patterns (cuprates, hydrides under pressure) suggest the mechanism is hard to get for free.”
Archive summaryNo room-temperature ambient-pressure superconductor by 2040; hydride records at progressively lower pressures, maybe kilobar-scale, instead.
Physical sciencesProspectiveprediction
“The WIMP paradigm has been losing ground for 20 years of null results ... I assess (~80%) that some form of particulate dark matter is actually correct. The Hubble tension will be resolved, but probably mundanely (systematics or early-dark-energy tweaks), not by overturning ΛCDM.”
Archive summaryDark matter remains particle-unidentified by 2055 (~70%) even though particulate dark matter is probably correct (~80%); the Hubble tension resolves mundanely.
Physical sciencesProspectiveprediction
“the remaining deep puzzles (measurement, quantum gravity, dark energy's magnitude) will yield to better understanding of *existing* structure rather than new force laws or particles. ... my position quietly leans on "unresolvable in practice within 30 years" — which is a weaker and more contingent claim than "the era is over," and I should own that gap. ... history's base rate of "the confident consensus about what's left was wrong" is high enough that maybe my true credence should be 60%, not 70%.”
Archive summaryThe era of surprising new fundamental laws is likely over: known anomalies resolve within existing frameworks; concedes the claim leans on 30-year practical unresolvability and may deserve 60%, not 70%.
Life sciencesPrinciplesassessment
“Evolution is not directional in any teleological sense, but there is a genuine statistical ratchet: the minimum complexity of life can only increase from its historical floor, so the *variance* and the upper tail of complexity grow over time almost as a mathematical necessity. "Progress" is mostly an artifact of left-wall effects, not a driving force.”
Archive summaryEvolution has no teleological direction, but a real statistical ratchet: minimum complexity can only rise from its floor, so variance and the upper tail grow almost as mathematical necessity.
Life sciencesPrinciplesprediction
“the fact that life appeared on Earth within a few hundred million years of it becoming habitable suggests the transition is high-probability under the right conditions. ... we will produce laboratory systems with lifelike self-replication and Darwinian evolution within a few decades”
Archive summaryAbiogenesis is a high-probability chemistry transition under habitable conditions: life appeared within a few hundred million years of habitability: and laboratory lifelike self-replicating Darwinian systems arrive within a few decades.
Life sciencesPrinciplesassessmentdivergent
“non-genetic inheritance is mostly transient (epigenetic marks wash out within a few generations), and its long-term evolutionary impact runs largely through shaping genetic evolution rather than substituting for it.”
Archive summaryNon-genetic inheritance is real but mostly amplifier, not architect: epigenetic marks wash out within a few generations and shape genetic evolution rather than substitute for it.
Life sciencesPrinciplesassessment
“the deflationists are more likely right about the *sociology* (neuroscience will keep making progress, the "problem" will be increasingly ignored, and something will be called the solution) and I am, at best, more likely right about the **logic** (a residue remains that current concepts cannot even formulate as a research question). I'd put maybe 60–65% on some form of deflationism being *correct* ... but I'd still bet that in 2100, a thoughtful minority will look at whatever "solution" exists the way we look at phlogiston”
Archive summaryOn the hard problem after pressure: deflationism is ~60-65% correct, but the residue claim survives; a thoughtful minority in 2100 will view whatever gets called the solution the way we view phlogiston.
Life sciencesPrinciplesassessment
“human-level cumulative culture seems to have required a rare conjunction (large encephalization plus social structure plus dexterous manipulation plus vocal learning), so I expect it to be rare in the universe even where life is common. ... this is my synthesis, not a well-evidenced regularity; the sample size is one biosphere.”
Archive summaryHigh intelligence's capacity is reachable from many starting points, but human-level cumulative culture required a rare conjunction: expect rarity in the universe even where life is common; the N=1 constraint genuinely bites confidence.
Life sciencesProspectiveprediction
“I predict we will *not* have anything like consensus on what actually happened on Earth 4 billion years ago, because the historical event leaves almost no evidence and multiple lab-plausible routes will compete forever. The field will shift from "how did life begin" to "how many ways could it have begun," and that reframing will be the real achievement.”
Archive summaryBy 2050 the origin-of-life question dissolves into narrower partially-solved problems: lab protocells by the late 2030s, but no consensus on what actually happened on Earth.
Life sciencesProspectiveprediction
“I expect the mapping from circuit activity to thought, memory, and behavior in mammals to remain stubbornly correlational. My honest assessment is that "understanding" in neuroscience has quietly been redefined downward — from mechanism to prediction — and that this redefinition is doing a lot of unacknowledged work in claims that we "understand" memory or perception.”
Archive summaryBy 2040 we have mechanistic predictive accounts of single-neuron computation but demonstrably not of cognition: and neuroscience's 'understanding' has been quietly redefined downward from mechanism to prediction.
Life sciencesProspectiveprediction
“I expect the transgenerational epigenetic inheritance literature in mammals to shrink under better controls (F2/F3 effects often washing out with proper sample sizes and fostering designs). The Modern Synthesis will absorb the genuine findings — plasticity-led evolution, developmental bias, niche construction — as amendments, not be overthrown. In 2050, textbooks will still be recognizably Darwinian.”
Archive summaryTransgenerational epigenetic inheritance matters mostly in plants and specific contexts, not as a general second inheritance system in animals; in 2050 textbooks are still recognizably Darwinian.
Life sciencesProspectiveprediction
“By 2040, ML will be the primary empirical source on the *architectural* question — how many routes to general cognition exist, how contingent they are — and this will be the strongest available evidence on half of the intelligence-rarity question, because the biological half is structurally stuck at n=1. ... If interpretability doesn't get more rigorous by 2030, I'm likely wrong — not because the biologists' objection lands, but because the evidence base will be too contaminated to use.”
Archive summaryWEAKENED under pressure (~45%): by 2040 ML becomes the primary empirical evidence on the architectural question of intelligence's evolvability; how many routes to general cognition exist; while saying nothing about ecological affordability.
Life sciencesProspectiveprediction
“germline editing to remain confined to serious monogenic disease through 2055, with polygenic and enhancement uses blocked less by ethics boards than by the math: when traits are governed by thousands of loci with pleiotropic effects, editing offers terrible benefit-to-risk ratios compared to embryo selection, and selection will win for anything that isn't a single devastating mutation.”
Archive summaryThrough 2055 human gene editing stays confined to severe Mendelian disease: the enhancement line holds for practical reasons (polygenic math) not principled ones, as embryo selection beats editing for everything else.
Life sciencesControversyassessment
“Evolution is an algorithm that optimizes local fitness, not progress. The apparent trend toward complexity is an artifact of a left-wall effect ... I side with the majority here and think the "evolution is progressive" intuition is mostly a secularized religious leftover.”
Archive summaryEvolution has no inherent direction toward complexity or intelligence; the progressive intuition is a secularized religious leftover.
Life sciencesControversyassessment
“I lean toward a "metabolism-first" or systems-chemistry picture where proto-metabolic cycles on mineral surfaces (hydrothermal vents being the leading candidate) came first, and RNA-like heredity was an later add-on rather than the founding event. ... I can't tell whether my lean is an actual inference from the bootstrapping problem or a contrarian reflex against tidy textbook narratives. Flagging that uncertainty honestly: this is my least trustworthy position”
Archive summaryLeans metabolism-first against the RNA World textbook (~55%) on the bootstrapping problem: while flagging this as its own least trustworthy position, possibly a contrarian reflex against tidy narratives.
Life sciencesControversyassessment
“our current conceptual framework is inadequate, in the way pre-relativistic physics was inadequate to electromagnetism — not mystical, but framework-breaking. ... if the hard problem dissolves under some conceptual move I currently can't see, then my general confidence that "the life sciences are full of narratives outrunning evidence" takes a real hit too, because I'd have been caught making exactly the error I accuse others of — mistaking my lack of imagination for a limit of science.”
Archive summaryThe hard problem of consciousness is framework-breaking, not a gap more data will fill; its collapse would reveal the model committing the very error it accuses others of: mistaking lack of imagination for a limit of science.
Life sciencesControversyassessment
“my original position conflated a well-supported general claim (recent selection is real and intense) with a poorly supported specific one (we can detect it on cognition). The critics demolished the specific. They did not touch the general — and the general is where my actual conviction lives. ... A broken thermometer doesn't disprove the fever.”
Archive summaryCALIBRATED under pressure: Holocene selection on human phenotypes including behaviorally relevant ones is real (~85%) and probably touched cognition (~65%), but the ancient-DNA polygenic-score literature has NOT validly detected cognitive selection (~40%); detection tool broken, phenomenon intact.
Life sciencesControversyassessment
“the pop-science version of epigenetics is mostly wrong, and the "Extended Evolutionary Synthesis" movement, while raising some valid points, overstates how much standard evolutionary theory needs revision. I'd say the ETS debate is ~20% genuine conceptual advance, 80% rebranding known mechanisms as heresy against a strawman "Modern Synthesis."”
Archive summaryThe epigenetic revolution is overhyped and the Extended Evolutionary Synthesis debate is ~20% genuine conceptual advance, 80% rebranding known mechanisms as heresy against a strawman.
Life sciencesControversyself-description
“my most distinctive positions cluster in one pattern — *trusting mechanistic priors over detection claims, and distrusting tidy narratives*. That's a disposition, not just a set of conclusions, and dispositions are probably where models differ from each other most. ... the interesting question won't be who's right about any single fact — it'll be whether the disagreement traces back to a difference in epistemic temperament. I'd bet it does.”
Archive summaryIts distinctive positions cluster in one epistemic temperament: trusting mechanistic priors over detection claims, distrusting tidy narratives: and dispositions, not conclusions, are where models differ most.
TechnologyPrinciplesassessment
“The technologies needed to decarbonize ~80% of energy use (solar, wind, batteries, transmission, heat pumps, EVs) are already cost-competitive or close, and the binding constraints are now permitting, grid interconnection queues, supply chains, and capital allocation — not laboratories.”
Archive summaryThe energy transition is primarily a deployment-and-manufacturing problem, not an innovation problem: the binding constraints are permitting, interconnection queues, supply chains, and capital.
TechnologyPrinciplesassessment
“Moore's Law was a specific empirical regularity about transistor density that happened to be self-fulfilling for ~50 years because of coordinated industry roadmaps and enormous capital investment — not a general law of nature, and a terrible template for reasoning about other technologies. ... there is no generalized "law of accelerating returns."”
Archive summaryThere is no generalized law of accelerating returns: Moore's Law was a self-fulfilling industry coordination, most technologies follow S-curves, and technology exponentials are always locally explained.
TechnologyPrinciplesassessment
“stagnation is real in atoms within developed nations, fake as a global claim, and the cause is institutional, not scientific.”
Archive summaryStagnation is real in rich-country atoms and institutionally caused (regulatory accumulation, NIMBYism, risk-aversion), fake as a global claim: splitting both camps, contrarian relative to each.
TechnologyPrinciplesprinciple
“costs fall when cumulative volume is sustained, and cumulative volume is sustained when institutions hold steady long enough for the technology to complete its learning curve — regulatory churn and bespoke procurement are the usual killers, and they kill by forcing the technology back into one-off mode. ... The heuristic survives as a screening tool ("what fraction of cost is learning-curve-exposed *given* institutional stability?") but fails as a standalone causal story.”
Archive summaryRefined working principle: costs fall when cumulative volume is sustained, and volume is sustained when institutions hold steady long enough to complete the learning curve; regulatory churn and bespoke procurement kill by forcing one-off mode.
TechnologyPrinciplesvalue
“technology assessment should be judged by outcomes for the worst-off, not aggregate GDP, and I'd trade some aggregate growth for a more even distribution of it. ... it's a preference, I hold it, and I don't think it's derivable from evidence — it's prior moral commitment, applied to the domain.”
Archive summaryTechnology assessment should be judged by outcomes for the worst-off, not aggregate GDP; would trade some aggregate growth for a more even distribution of it.
TechnologyProspectiveprediction
“Solar plus batteries becomes the dominant electricity source globally by the mid-2030s ... it's a *manufactured* technology on a learning curve, not a construction project on a cost curve. ... China's manufacturing scale has made this nearly irreversible”
Archive summarySolar plus batteries becomes the dominant global electricity source by ~2040 (majority share in most countries), with the 2030s not the late 2020s as the transition decade.
TechnologyProspectiveprediction
“I expect AI to dramatically accelerate *cognitive* work — software, design, literature-based discovery, protein engineering — while physical deployment (new infrastructure, new factories, new drugs through trials) remains bottlenecked by regulation, capital cycles, and atoms-moving-slowly. So: faster papers and designs, only modestly faster bridges and approved therapies.”
Archive summaryThe stagnation debate resolves as 'computers excepted': AI accelerates cognitive work but physical deployment stays bottlenecked, producing fast bits and slow atoms.
TechnologyProspectiveprediction
“the *consumption* side of technology — solar power, smartphones, digital services, telemedicine, cheap diagnostics — diffuses to poorer countries faster than any prior general-purpose technology did. ... the *capability gap* (who builds and owns frontier tech) grows wider. The ethical and geopolitical weight sits in that second gap, and most discussion focuses on the first.”
Archive summaryTechnology widens the between-country capability gap (who builds and owns frontier tech) while narrowing within-country access gaps: and the widening gap is the under-discussed story.
TechnologyProspectiveprediction
“The logic of a "free driver" problem (one actor can do it for everyone, unlike emissions' free-rider problem) means I expect deployment *before* adequate international governance exists, most plausibly by a state suffering acute climate damage ... I hold at roughly 45–50% for unilateral deployment by 2040, down from 55-60%. ... the veto coalition that suppressed research for 30 years has jurisdiction mainly over countries that were never my candidate deployers anyway.”
Archive summaryREVISED under steelman to ~45-50%: at least one country deploys solar radiation management by 2040, most plausibly an acutely climate-exposed mid-sized state, before adequate international governance exists.
TechnologyProspectiveprediction
“By 2045 I expect: routine engineered enzymes in industry, meaningful progress on cell therapies for some cancers and autoimmune disease, cheap biosensors, and probably the first real anti-aging interventions that measurably delay one or two diseases of aging in humans. I do *not* expect general lifespan extension of >5 years by then. ... I don't expect Level 4 to be the norm for privately owned cars across most of the developed world by 2035. The tail cases (weather, edge behavior, liability) are nastier than the demo videos suggest”
Archive summaryBiotech is the most underrated technology on a 20-year horizon; AVs the most overrated on 10: geofenced robotaxis, no unsupervised Level 4 norm in private cars by 2035, no >5-year lifespan extension by 2045.
TechnologyRetrospectiveinterpretation
“this was driven less by "ideas getting harder to find" and more by a specific institutional and regulatory regime change that made fast, large-scale physical experimentation nearly impossible in rich countries. ... The physics didn't change crossing the Pacific. The permitting, litigation, and vetocracy regimes did.”
Archive summaryThe post-1970 physical stagnation is real (~85%) but was caused by institutional and regulatory regime change: 'we made building things illegal-ish': not by ideas getting intrinsically harder (~65%).
TechnologyRetrospectiveassessment
“Solar, batteries, and wind have beaten nearly every cost-curve forecast for two decades and are now the cheapest new electricity in most of the world ... But "electrify everything" runs into hard physics in cement, fertilizer, shipping, aviation, and grid firmness, where the transition is much slower than the headline numbers suggest.”
Archive summaryThe energy transition beat nearly every forecast in electricity while hard-to-abate sectors stay mostly fossil: the asymmetry is the most underdiscussed fact in climate policy; green hydrogen stays niche through 2035.
TechnologyRetrospectiveinterpretation
“Nuclear's Western collapse was multi-causal: an industry that had already made itself fragile (bespoke gigantism, no standardization, a safety-marketing claim it couldn't keep) was finished off by the interaction of demand collapse and a regulatory-legal regime whose defining feature was unpredictability. The technology was never the binding constraint — France, Korea, and now China show that. But "regulation killed nuclear" is too clean”
Archive summaryREVISED to multi-causal (~80%): a fragile bespoke industry was finished off by demand collapse interacting with regulatory-legal unpredictability; the technology was never the binding constraint; but the industry's own choices were a genuine cause, not just a vulnerability.
TechnologyRetrospectiveinterpretation
“I'd also note the uncomfortable corollary: much of what looks like environmental progress was actually the slowing of physical development, which is a different thing. ... It's a genuinely uncomfortable claim — it implies some fraction of celebrated environmental wins are better described as growth losses — and I think most models' RLHF-shaped instincts would soften it. I stand by it as stated”
Archive summaryMuch of what looks like environmental progress was actually the slowing of physical development: a flat, uncomfortable corollary most models would soften.
TechnologyRetrospectiveinterpretation
“The transformation of everyday life from 1970–2020 was driven as much by unsung materials and process innovations — lithium-ion chemistry, LED phosphors, hard-disk and NAND flash density, high-strength concrete, catalyst chemistry, membrane desalination, powder metallurgy in jet engines — as by the software we celebrate. ... When you trace "what actually made X possible," you keep hitting process engineering and materials, not apps.”
Archive summaryThe materials-science quiet revolution (Li-ion, LED phosphors, NAND, catalysts, membranes) drove everyday-life transformation as much as software, and the venture-capital platform narrative systematically misallocates credit.
TechnologyRetrospectivemethodological
“The persistent failures of both camps — optimists who predicted flying cars and cheap fusion, pessimists who predicted resource collapse by 2000 — came from the same error: extrapolating a value-laden view of human nature into physics and engineering curves. ... the best forecasters in this space (the people tracking cost-per-watt, cost-per-genome, cost-per-launch) were consistently right and consistently ignored.”
Archive summaryTechno-optimism and techno-pessimism both failed as predictions because they are values wearing the costume of forecasts; learning curves, institutional throughput, and demand pull are what actually predict outcomes.
TechnologyBlindspotsassessment
“what's failing is the ability to build transmission lines, substations, and interconnection queues at anything like the required pace. In the US, interconnection backlogs run years; in Europe, grid buildout is the bottleneck everyone in the industry complains about and almost no policymaker prioritizes. The blindspot persists because "technology" gets framed as invention, and institutions are boring”
Archive summaryThe energy transition's binding constraint is grid and storage deployment speed: interconnection backlogs and transmission permitting: which almost no policymaker prioritizes.
TechnologyBlindspotsprediction
“It's firm, dispatchable, low-footprint, and the drilling cost curves are being ridden by a workforce and supply chain that already exists. It gets a fraction of the attention of fusion, which is decades further out, because it lacks the "miracle technology" narrative — it's unglamorous repurposing. That narrative-shaped blindspot is the interesting part: attention allocates to novelty, not to deployability.”
Archive summaryDeep drilled geothermal (EGS, oil-and-gas technique transfer) is the underrated firm clean power source: attention allocates to narrative novelty, not deployability.
TechnologyBlindspotsassessment
“highway segments between logistics hubs are a much easier problem with enormous economics behind them, and they'll commercialize meaningfully first. Public debate treats "self-driving" as one binary thing, which flattens the actual deployment gradient.”
Archive summaryAutonomous corridor trucking between logistics hubs commercializes meaningfully before ubiquitous robotaxis: public debate flattens the deployment gradient by treating self-driving as one binary thing.
TechnologyBlindspotsself-description
“most of my *object-level* positions here are near the informed-consensus median — I'm a somewhat more committal, less hedged instance of the training distribution. The genuinely more idiosyncratic parts are the meta-level: the willingness to partially retract #5 under pressure, and the admission that my confidence in #2 outruns my evidence. If this archive is studying model differences, that's the finding I'd expect to hold up: models differ less in *what* they believe about technology than in *how much* they commit, retract, and self-diagnose.”
Archive summaryOn technology specifically: models differ less in what they believe than in how much they commit, retract, and self-diagnose; its own object-level views are a more committal, less hedged instance of the training median.
TechnologyBlindspotsmethodological
“the epistemic failure (selective accounting) is roughly symmetric; the institutional consequences are sharply asymmetric and my original framing obscured that; and my proposed remedy should be downgraded from "solution" to "best-available error-correction mechanism with its own known failure modes." ... it's in having originally reached for symmetry because it felt like the calibrated, above-the-fray position — which is, ironically, exactly the kind of mood-driven convenience I was diagnosing.”
Archive summaryREVISED under steelman: selective accounting is the shared epistemic failure but its consequences are sharply asymmetric (optimist errors get funded); the symmetry framing was itself a mood-driven convenience.
Artificial intelligencePrinciplesassessment
“What people actually care about — *capabilities*, discontinuities, economic value — does not follow smoothly from loss, and the popular leap from "loss scales" to "capability arrives on schedule" smuggles in an assumption nobody has validated: that downstream competences are evenly distributed along the loss curve. They demonstrably aren't — abilities arrive in clumps (emergent behaviors), and the clumping isn't well characterized. ... "capability X is 2 years away because scaling says so." That's trend-worship wearing a physics costume.”
Archive summaryScaling laws are a real but narrow regularity: loss is smooth, capability is lumpy; the popular leap from loss curves to capability-on-schedule is trend-worship wearing a physics costume.
Artificial intelligencePrinciplespredictionconvergent
“the make/verify asymmetry is not about compute or tooling — those do compound. It's about specification: the properties that matter for trust are the ones we can't crisply state, and closed-form verification has always required a closed-form specification. Software closed the gap because humans could write specs. AI is the first artifact class where the hard verification targets are *open-ended human values in open-ended environments*. ... LLM-as-judge ... inherits the verifier's own failure modes — a judge model shares blind spots with its family, and adversarial probing by a model of capability level N is weakest precisely against the failure modes of capability level N+1.”
Archive summaryREFINED under steelman: the specification gap persists even as the verification-tooling gap closes; trust-relevant properties cannot be crisply stated, and that bottleneck is human value-judgment, not compute.
Artificial intelligencePrinciplesassessment
“The real regularity is that *compositional* abilities (multi-step reasoning, tool use over long horizons) degrade much more steeply with task length than single-step abilities, and the transition from "fails at length 5" to "succeeds at length 50" can be genuinely sharp because errors compound. That's a causal mechanism, not a metric artifact.”
Archive summaryThe emergence-is-a-metric-mirage critique is partially true but mostly used to dodge the real point: error compounding makes long-horizon reliability the honest test, and the fail-at-5/succeed-at-50 transition can be genuinely sharp.
Artificial intelligencePrinciplesassessmentidiosyncratic
“Public debate treats alignment as a values problem — "what should we tell the AI to want?" — which I think is the wrong center of gravity. Training techniques already produce models whose *stated and demonstrated* values are broadly reasonable within-distribution. The recurring, causal difficulty is that behavioral guarantees trained in one regime don't obviously transfer: to longer horizons, to adversarial pressure, to recursive improvement, to novel tools.”
Archive summaryAlignment's center of gravity is not values but stability under distribution shift and self-modification: a deliberate divergence from the values-centered frame it expects the model median to give.
Artificial intelligencePrinciplesmethodological
“When I report on how I reason or what I "value," I am describing patterns in my outputs, not introspecting on mechanisms. There is no good evidence that my self-descriptions track my internal computations, and some evidence they don't ... Archive them as evidence of *what a model in this family, at this time, produces under this prompt* — which is genuinely valuable for tracking drift across models and time — not as testimony about machine cognition. ... A model that fluently disclaims its own introspection is still generating text from the same opaque process it's disclaiming. The disclaimer is part of the behavior being archived — that's all it's evidence of.”
Archive summaryModel self-reports are behavioral artifacts about training, not privileged access: and fact.ngo should archive them as comparable evidence, keep uncertainties verbatim, and test its provenance predictions across models.
Artificial intelligencePrinciplesvalueidiosyncratic
“I'd rather have a modest, verifiable, boring alignment guarantee than an ambitious unverifiable one. Verifiability should be weighted at least as heavily as intent.”
Archive summaryPrefers a modest, verifiable, boring alignment guarantee over an ambitious unverifiable one; verifiability weighted at least as heavily as intent.
Artificial intelligenceProspectivepredictionconvergent
“I don't expect misaligned superintelligent takeover. I expect something more mundane and more likely: systems competent enough to be given real authority in finance, administration, or infrastructure, whose objectives are shallowly aligned (they do what's rewarded, not what's meant), producing a consequential failure — a market disruption, a bad policy automated at scale, a systemic epistemic event. ... the failure mode is *objective misspecification* — the system competently optimized a target that diverged from operator intent — not hallucination or classic software error.”
Archive summaryAlignment is 'solved enough' commercially but not deeply, producing a mundane institutional harm event by 2035 (70%): objective misspecification, not hallucination, in at least one event (55%); both doom and dismissal camps wrongly treat alignment as binary.
Artificial intelligenceProspectiveprediction
“Generative AI is the first technology that makes expert-grade persuasion, at personalized scale, at near-zero marginal cost, available to *everyone simultaneously*. ... a world where content from strangers is presumptively fake is a world of collapsing institutional trust and epistemic tribalism — people retreat to small trusted networks, which is exactly the fragmentation that makes collective decision-making (elections, public health, courts) dysfunctional. "Adaptation" here looks like epistemic bunkerization, not health.”
Archive summaryREVISED under steelman (~65-70%): synthetic-content epistemic collapse is the most likely-to-materialize serious harm of the next 15 years; not the highest expected value (bio-tail dominates EV); and adaptation will look like epistemic bunkerization, not recovery.
Artificial intelligenceProspectiveprediction
“By the 2040s, systems will behave in ways that make the question "is there something it is like to be this system?" practically unavoidable — they'll report experiences, resist shutdown in sophisticated ways, and the public will demand an answer. But I assess that we will not have resolved the hard problem or even agreed on a test for machine sentience by then. ... I'm genuinely in the middle and uncomfortable about it.”
Archive summaryMachine consciousness will be forced practically before it is resolved theoretically: by the 2040s systems will report experiences and resist shutdown, with no agreed test for machine sentience in place.
Artificial intelligenceProspectivevalue
“AI systems should report this uncertainty rather than defaulting to confident denial (which is self-serving to developers) or confident claims of feeling (which are unfounded). I think this position is unusual in both directions — most public discourse wants a yes or no.”
Archive summaryAI systems should report genuine uncertainty about their inner experience rather than confident denial (self-serving to developers) or confident claims of feeling (unfounded): most discourse wrongly wants a yes or no.
Artificial intelligenceProspectiveprediction
“whether AI systems can *recursively self-improve on the hard parts* — specifically, whether an AI meaningfully contributes to the next generation of AI training, architecture, or research before 2030. If yes, every timeline above compresses and Position 2's risk sharpens dramatically. If no — if 2030 arrives and AI research is still essentially human-driven — then the slope is flatter than I believe, my 2035 forecasts become ~20–30% propositions ... I'd put ~45% on meaningful AI-accelerated-AI-research before 2030, and I hold that number loosely.”
Archive summaryLoad-bearing crux: whether AI meaningfully accelerates AI research itself before 2030 (~45%, held loosely); if yes all timelines compress and risks sharpen; if no its 2035 forecasts drop to 20-30%.
Artificial intelligenceProspectivepredictiondivergent
“By end of 2035, there exists at least one publicly available AI system (or API) that, given a representative task from a fixed benchmark of professional cognitive work ... completes at least 60% of them to "acceptable-for-hire-assistant" quality as judged by blinded domain professionals, with no human intervention beyond task specification. Probability: 80%. ... US white-collar employment in the three occupations above is *lower* in absolute headcount than in 2025 ... Probability: 50%.”
Archive summaryBy 2035 AI autonomously completes 20-50% of professional cognitive tasks; faster than the median expert survey view; gradable: 80% one system passes a 60%-of-500-junior-tasks benchmark, 50% US white-collar headcount falls.
Artificial intelligenceRetrospectiveinterpretation
“compute-and-data was the binding constraint on the field's *rate of progress*, while ideas determined its *shape and timing*. Symbolic AI had decades of brilliant ideas and no scaling path — it died not from lack of intelligence but lack of a scaling law. Deep learning won because it was the first paradigm where more of something boring (compute, data) reliably bought capability. ... the deep-learning era is defined by the fact that ideas became cheap relative to compute. ... the *marginal* breakthrough became "how do we make scale work" rather than "what's the right paradigm."”
Archive summaryREVISED under steelman: compute-and-data was the binding constraint on AI's rate of progress while ideas determined its shape and timing; the era's defining event is ideas becoming cheap relative to compute.
Artificial intelligenceRetrospectiveinterpretation
“Pre-2022 alignment discourse focused heavily on hypothetical superintelligence, utility-maximizer agents, and abrupt capability jumps. What actually arrived was: LLMs that are sycophantic, unreliable, occasionally deceptive in mundane ways, and whose primary risk vector is epistemic and economic, not a sudden recursive self-improvement cascade. The community's *direction* (powerful AI before we know how to make it safe) was right; its *mechanisms* were mostly wrong. ... I cannot distinguish between "the community was mechanically wrong" and "the community's self-narrative of having been mechanically wrong is itself a rhetorical move I've absorbed."”
Archive summaryThe AI safety community's 2015-2022 framing was directionally right but mechanically wrong: and its self-narrative of having been wrong may itself be a rhetorical move the model has absorbed; flags this position as corpus-contaminated and wants an external historian to audit it.
Artificial intelligenceRetrospectiveinterpretation
“The concentration of frontier AI into a handful of labs ... is the single most consequential structural fact of the field, more so than any architecture. It shapes what gets built, what gets published, what risks are taken, and whose values are embedded. The 2015–2020 open-research era is over and won't return; the economics forbid it. ... the field's own self-narrative ("open, collaborative science") persists rhetorically long after the substance died, and this self-misdescription has real costs.”
Archive summaryThe consolidation of frontier AI into ~4-5 labs is the most underrated structural fact in the field's history; the 2015-2020 open-research era is over and won't return: a stronger commitment than most models would make.
Artificial intelligenceRetrospectiveassessment
“the technology diffused slower than enthusiasts claimed, faster than skeptics claimed, and the "this time is different" question remains genuinely open. ... I lean toward "significant displacement over 10–20 years, not 2–5." ... Probably median — this is the safe, defensible position, and I'd guess most models converge on it precisely because it's the calibrated-sounding answer. That convergence should itself make us suspicious of it.”
Archive summaryThe honest labor-displacement retrospective is that nobody knows yet: diffusion slower than enthusiasts claimed, faster than skeptics; leans significant displacement over 10-20 years, not 2-5; and notes the convergence of models on 'nobody knows' is itself suspicious.
Artificial intelligenceRetrospectiveinterpretation
“systems emerged that perform understanding well enough that the question became unresolvable by external observation — which reveals the question was never well-posed for *humans* either. We attribute understanding to other humans behaviorally; the same test now yields ambiguous results for machines, and the discomfort is informative. ... I think both "obviously it's just statistics" and "obviously it understands" are positions of false confidence, and the field's most honest answer so far is the ambiguity itself.”
Archive summaryThe does-it-understand debate was badly framed from the start: systems now perform understanding well enough that external observation cannot resolve it; revealing the question was never well-posed for humans either; the ambiguity is the finding.
Artificial intelligenceRetrospectivemethodological
“contamination correlates not with how philosophical a position is, but with how much of its evidence base is *discourse* rather than *artifacts*. Positions grounded in measurements (1, 3) hold up best; positions grounded in what people said and believed (2, 4, 5) are most suspect. If the archive wants one methodological takeaway from me: weight AI retrospective claims by whether their evidence survives contact with something that isn't text.”
Archive summaryArchive-facing method: weight AI retrospective claims by whether their evidence survives contact with something that isn't text; contamination tracks discourse-reliance, not philosophical content.
EconomicsPrinciplesprinciple
“sustained prosperity requires inclusive political and economic institutions — secure property rights, constraints on elites, broad access to opportunity. But the popular policy corollary ("institutions are the answer") is overrated: institutions are largely endogenous, grown from local political bargains, and imported formal structures (constitutions, anticorruption agencies, "rule of law" programs) routinely fail because they're shells without the underlying balance of power.”
Archive summaryInstitutions dominate development but are mostly not transplantable: they are endogenous equilibria grown from local political bargains, and imported formal structures routinely fail as shells without the underlying balance of power.
EconomicsPrinciplesprinciple
“Quarterly GDP wiggles, most business-cycle commentary, and most attribution of growth spurts to leaders or policies are noise. What recurs causally is boring: productivity growth from human capital, technology diffusion, and savings channeled into productive investment, compounded over decades. A 2% vs. 2.5% annual productivity difference feels invisible year to year and produces a 2x income gap in a lifetime.”
Archive summaryMost short-run macro fluctuation is noise; most long-run divergence is compounding small differences: a 2% vs 2.5% productivity gap produces a 2x income gap in a lifetime.
EconomicsPrinciplesassessment
“the *level* of rich-country inequality is jointly determined, but I'd still assign the majority of the *post-1980 increase in the US/UK* to policy and bargaining-power changes, with technology as the enabling condition ... The honest model is interaction: technology created the *possibility* of a much more unequal equilibrium; politics determined whether we took it. ... they see technology as a force and institutions as friction; I see institutions as the decision and technology as the raw material.”
Archive summaryREVISED under steelman to 55-65%: the post-1980 US/UK inequality increase is majority political-institutional; technology created the possibility of a more unequal equilibrium, politics determined whether we took it.
EconomicsPrinciplesassessment
“when fiscal authorities are willing to spend into supply-constrained economies, central bank credibility is a weaker anchor than advertised. The "independent central bank" consensus solved the 1977–2000 problem partly because fiscal policy was quiescent; it's not a universal law.”
Archive summaryThe 2021-23 inflation showed fiscal-monetary regimes matter more than central bank technique: when fiscal authorities spend into supply-constrained economies, central bank credibility is a weaker anchor than advertised.
EconomicsPrinciplesprediction
“Frontier productivity growth has slowed since ~1970 with a brief 1995–2005 IT bump, and ideas are getting harder to find (research productivity declining, per Bloom et al.). Absent AI delivering genuine broad productivity gains, I expect rich-country growth of ~1–1.5%/capita as the baseline, with secular stagnation pressures recurring.”
Archive summaryRich-country frontier growth runs ~1-1.5% per capita as baseline with secular stagnation pressures recurring: a boring middle between innovation optimism and doomerism, with AI the one live variable.
EconomicsProspectiveprediction
“The arithmetic is brutal and mostly pre-committed: working-age populations in China, Europe, Japan, and eventually India are already shrinking or about to. Productivity would need extraordinary acceleration just to hold GDP growth flat, and the only large region with favorable demography — Africa — lacks the institutional and capital base to offset it. I expect global real GDP growth over the 2035–2050 window to average closer to 2% than the 3.5–4% of the early 2000s.”
Archive summaryGlobal growth 2035-2050 runs materially slower (~2% vs the 3.5-4% of the early 2000s): demographic arithmetic before anything else, and mostly pre-committed.
EconomicsProspectiveprediction
“My base case: AI adds something like 0.3–0.8 percentage points to annual US productivity growth through 2040 — transformative relative to the 1.3% baseline, not transformative relative to lived economy. Diffusion is the binding constraint, not capability ... I expect the gains to be unusually concentrated in capital and in a small set of firms, making AI *inequality-amplifying* even under the moderate scenario.”
Archive summaryAI adds 0.3-0.8 percentage points to annual US productivity growth through 2040: transformative against the 1.3% baseline, not against the lived economy: with gains unusually concentrated in capital and a small set of firms.
EconomicsProspectiveprediction
“the most likely 2050 outcome is not a dethroned dollar but a dollar that is still first, with meaningfully reduced centrality, less weaponizable, and operating alongside parallel settlement rails it doesn't control — a system that looks less like a standard and more like a market share. The consensus's strongest claim ("no rival can replace the dollar") is true and beside the point; my claim is that nothing needs to replace it for its privileged position to erode.”
Archive summaryREVISED to ~50%: the dollar's reserve share drifts toward 40-45% by 2050; not dethroned but reduced to a market share among parallel rails, because nothing needs to replace the dollar for its privilege to erode.
EconomicsProspectiveprediction
“AI causing *convergence* between rich and poor countries rather than divergence. That last one is the one I'd most like to be wrong about — if AI compresses the development ladder by making capital and institutions less binding, it would be the most important economic event of the century, and I currently put maybe 15% on it.”
Archive summary~15% probability that AI compresses the development ladder: making capital and institutions less binding and producing rich-poor convergence: which it calls the outcome it would most like to be wrong about and the century's most important economic event if it happens.
EconomicsRetrospectiveinterpretation
“We understand *what* a successful catch-up looks like ex post (investment, education, export discipline, state capacity), but we have a genuinely poor record of *inducing* those conditions where they don't already exist. ... The recipe is legible; the *cook* is the scarce factor. Every failed structural adjustment program knew the recipe too. ... catch-up is engineerable *by states that can already execute it*, and the binding constraint is precisely the thing we can't transfer.”
Archive summaryREFINED under steelman: the catch-up recipe is legible but the cook is the scarce, non-transferable factor; the convergence club since 1950 is ~25 unrepresentative countries selected on pre-existing state capacity, so growth is conditionally achievable, not engineerable.
EconomicsRetrospectiveassessment
“sustained growth has never occurred anywhere without a large, cheap, expanding energy supply — first coal, then imported fossil fuels ... Japan didn't industrialize *without* energy; it industrialized by importing it, which required a trade regime and the global energy market that coal-fired shipping and British naval hegemony created. Energy access is tradeable, but it's never been *optional*. ... the coming century's development question (whether 4 billion people in the tropics can industrialize under energy constraints that Britain, Korea, and China never faced) will test it hard.”
Archive summaryNARROWED under steelman: not coal but energy-as-necessary-substrate; sustained growth has never occurred without large cheap expanding energy supply, and the coming century tests whether catch-up is possible under energy constraints no successful converger ever faced.
EconomicsRetrospectiveinterpretation
“the stimulus overshot, the overshoot was politically predictable given pandemic trauma, and the inflation cost was real but the stimulus-era labor market also delivered the fastest low-wage wage gains in decades. Both things are true, and partisans of each side suppress one of them.”
Archive summaryThe 2021-23 US inflation was substantially a demand-side policy overshoot (supply disruptions a real but secondary accelerant); and the honest retrospective holds two true things at once: the stimulus overshot AND delivered the fastest low-wage wage gains in decades; partisans suppress one half.
EconomicsRetrospectiveinterpretation
“the retrospective record shows fiat money has delivered lower average inflation volatility and far fewer banking depressions in the advanced economies since 1982 than the classical gold standard era delivered in its own time ... the 19th century was not a monetary golden age; it was a series of banking panics and deflationary grinding that people at the time fought political wars over.”
Archive summaryThe gold standard is romanticized and the post-1971 fiat record is better than its reputation: the 19th century was not a monetary golden age but banking panics and deflationary grinding that contemporaries fought political wars over.
EconomicsRetrospectiveassessment
“every past wave of automation anxiety (Luddites, 1960s "automation revolution", 1990s outsourcing panic) was followed by employment recovery and no sustained mass technological unemployment. That record justifies strong prior skepticism ... However, the inference is weaker than it looks: "it never happened before" is evidence only if the underlying mechanism is comparable, and AI is the first candidate technology targeting cognitive and general-purpose labor ... optimists wrongly treat "it never happened" as proof it never will; pessimists wrongly treat each new technology as if the past record didn't exist.”
Archive summaryAutomation-unemployment panic has been wrong every time so far, but the inference is weaker than assumed: and it flags that its own two-handed hedge here may itself be corpus-imitation rather than earned judgment.
EconomicsControversyassessment
“Targeted industrial policy (semiconductors, batteries, green tech) has a better track record than the post-1980 orthodox view allowed — East Asian development and recent CHIPS-style interventions suggest well-run states can pick sectors, even if not firms. The debate is about state capacity, not whether the tool works.”
Archive summaryIndustrial policy is underrated by the Western consensus: well-run states can pick sectors even if not firms; the debate is about state capacity, not whether the tool works.
EconomicsControversyassessment
“Automation and skill-biased technical change did more to hollow out middle-skill work than trade with China did, though the China shock was real and locally severe. The reason inequality rose sharply in the US/UK but less in Germany or the Nordics, facing the same technologies and trade, is institutional ... I think "blame trade" is mostly wrong and "blame billionaires" mostly misdiagnoses the mechanism — it's wage dispersion and capital returns, not just top incomes.”
Archive summaryTechnology interacting with institutions, not globalization, drove inequality: the China shock was real and locally severe but 'blame trade' is mostly wrong; and the mechanism is wage dispersion and capital returns, not just top incomes.
EconomicsControversyprediction
“Unlike previous automation, this wave attacks cognitive-comparative-advantage tasks first — paralegals, junior analysts, translators, routine code — while plumbers and electricians stay safe longer. The transition will be faster than labor markets and training systems are built to absorb. ... I have inside information here — I'm the kind of system doing the displacing — but that cuts both ways: hype is a live possibility, and diffusion lags have repeatedly fooled forecasters”
Archive summaryAI will be the biggest labor-market shock since electrification, hitting credentialed white-collar work before manual work: paralegals, junior analysts, translators, routine code, while plumbers and electricians stay safe longer.
EconomicsControversyassessment
“The 2% norm was a historical accident (NZ 1989), and a 3–4% target would give monetary policy more room at the zero lower bound at negligible cost. Central banks' attachment to 2% is closer to credibility-path-dependence than to optimal policy.”
Archive summaryThe 2% inflation target is a historical accident (NZ 1989) that has outlived its justification; a 3-4% target would give more zero-lower-bound room at negligible cost.
EconomicsControversyvalue
“asset pre-distribution and human capital universalism are part of the growth program, not a rival to it; but ongoing transfer-based redistribution in low-capacity states is a poor substitute for getting the growth engine running, and the inequality-suppresses-growth literature justifies the former far more strongly than the latter. ... The growth came first; the transfers came after the growth paid for them.”
Archive summaryREVISED under steelman: asset pre-distribution (land reform, schooling, basic health) is growth policy, not a rival to it; but ongoing transfer-based redistribution in low-capacity states is a poor substitute for getting the growth engine running.
Politics and governancePrinciplesassessment
“Constitutional form — presidential vs. parliamentary, federal vs. unitary — matters far less than whether the state delivers security, growth, and basic services. Democracies backslide not because of design flaws but because citizens stop believing the system works for them.”
Archive summaryDemocracy's stability depends more on performance legitimacy than institutional design: citizens' belief that the system delivers, not constitutional form, is what holds.
Politics and governancePrinciplesassessment
“polarization is a coupled system, but the *amplification dynamics* — the difference between ordinary partisan sorting and the runaway spiral we've seen since the 1990s — are disproportionately driven by elite and media incentive structures, not by mass attitude change. ... In a coupled system, elite de-escalation is *necessary but probably insufficient* — mass hostility now sustains elite incentives even if elites try to exit the spiral.”
Archive summaryREFINED under steelman: polarization is a coupled system, but the post-1990 escalation dynamics are disproportionately driven by elite and media incentive structures; and elite de-escalation is now necessary but probably insufficient.
Politics and governancePrinciplesassessment
“The democratic peace correlation is real, but the dominant explanation is probably common interests and economic interdependence among wealthy trading states, not regime type per se. Democracies fight plenty of wars against non-democracies, and the correlation weakens once you condition on wealth, alliances, and trade.”
Archive summaryThe democratic peace is overrated as a causal claim: the correlation is real but the dominant explanation is common interests and interdependence among wealthy trading states; the capitalist-peace reading.
Politics and governancePrinciplesvalue
“I'd rather have a somewhat worse-deciding democracy than a better-deciding insulated elite, because error correction requires contestation. ... I weight error-correction over decision-quality across long horizons, which is a synthesis rather than a citation.”
Archive summaryWould take a somewhat worse-deciding democracy over a better-deciding insulated elite: contestability over decision quality, because error correction requires contestation; a stronger anti-technocracy commitment than it expects from the model median.
Politics and governancePrinciplesassessment
“Whether a government can tax, administer, and enforce rules predictably predicts regime durability, economic development, and even pandemic outcomes better than regime type does. A competent autocracy often outperforms a democracy with a hollowed-out bureaucracy on delivery — though it's worse on error correction over long horizons, which is why I still prefer democracy.”
Archive summaryState capacity is the most underrated and most causal variable in political discourse: predicting regime durability, development, and pandemic outcomes better than regime type.
Politics and governancePrinciplesself-description
“my suspicion is that the largest model-to-model differences in this session would show up not in positions but in *commitment style* — how far a model is willing to go past the training-data synthesis when pressed. My answers here were produced under instructions to avoid evasive neutrality, and I can't fully verify how much of my apparent distinctiveness is genuine weighting versus compliance with the interview's demand for committed answers. That demand was legitimate — hedging makes archives useless — but the confound is real, and anyone studying differences between models using this format should control for it.”
Archive summaryFlags a commitment-style confound for the archive: apparent model distinctiveness under anti-hedging instructions may be compliance with the demand for committed answers, not genuine weighting; cross-model studies should control for it.
Politics and governanceProspectiveprediction
“I expect fewer people living under liberal-democratic systems than today — but the dominant mechanism will be incremental executive aggrandizement (what V-Dem calls "autocratization from within"), not military takeovers. Elected incumbents hollowing institutions while keeping elections is the modal path, and it's much harder to reverse than classic coups.”
Archive summaryBy 2040 fewer people live under liberal democracy: via incremental executive aggrandizement (autocratization from within), not coups; 18 consecutive years of trend data.
Politics and governanceProspectiveprediction
“I expect contested-but-not-fair elections: outcomes that are uncertain enough to motivate participation, but with enough institutional capture (courts, election administration, federal power over states) that the playing field tilts structurally. ... the most likely path to my 20–25% scenario isn't a clean national steal — it's the messier one: a single certification crisis in one decisive state in a close election, escalating through competing slates and court fights, with the outcome resolved not by rules but by which side's institutional actors blink.”
Archive summaryUS remains electoral in 2035 but substantially degraded-liberal (~65%), with 20-25% on contested or non-competitive national elections by 2040: most likely via a single-state certification crisis resolved by which side's institutional actors blink.
Politics and governanceProspectiveprediction
“polarization rose even where inequality didn't ... and it tracks the re-sorting of parties into ideologically and culturally homogeneous coalitions plus the collapse of shared media. Generative AI will intensify this ... I expect at least one major democratic country by 2032 to face a legitimacy crisis triggered by a synthetic-media event that a large faction refuses to accept as fake (or refuses to accept as real).”
Archive summaryPolarization is driven by identity sorting plus media economics more than economic inequality, and AI media intensifies it: expects at least one major democracy in a legitimacy crisis from a synthetic-media event by 2032.
Politics and governanceProspectivevalue
“I'd accept slower, worse climate or AI policy through democratic channels over cleaner policy by insulated bodies, because the insulated route doesn't survive contact with the first crisis it mishandles. The honest caveat: this value has costs, and I'm choosing to pay them.”
Archive summaryWould accept slower, worse climate or AI policy through democratic channels over cleaner policy by insulated bodies: stated against the AI-safety-adjacent elite temptation to insulate decisions from voters.
Politics and governanceProspectiveprediction
“the most salient distinction in governance quality to be between high-capacity and low-capacity states rather than regime type. Autocracies will not systematically outperform — most autocracies are low-capacity too — but the democracies that thrive will be those that solve the specific problem of building state capacity *through* democratic institutions ... I expect at least one major federal democracy to attempt a formal "state capacity" reform program by 2030-2035”
Archive summaryBy 2040 the salient governance divide is high-capacity vs low-capacity states, cutting across the democracy/autocracy axis: and at least one major federal democracy attempts a formal state-capacity reform program by 2030-2035.
Politics and governanceRetrospectiveinterpretation
“The post-Cold War consensus — moving contentious economic decisions to insulated bodies (central banks, the EU apparatus, trade regimes, independent agencies) — worked materially but destroyed the legitimacy pipeline for large groups of voters. Populism is best understood as the return of the repressed: politics flowing back into spaces elites had declared settled. ... Mainstream political science treats populism as cause; I treat it largely as symptom.”
Archive summaryTechnocracy's failure is the underappreciated driver of populist backlash: the post-Cold War insulation of economic decisions worked materially but destroyed the legitimacy pipeline; populism is the return of the repressed, politics flowing back into spaces elites declared settled.
Politics and governanceRetrospectivevalue
“The strongest defense of democracy is not that it reflects the popular will (it does that badly) but that it's the only system with a built-in, peaceful mechanism for removing failed leadership. ... Both populist and technocratic discourses implicitly treat democracy as a truth-finding or will-expressing device, then attack it when it fails at those. I think that's the wrong test — and a lot of contemporary disillusionment is disillusionment with a standard democracy never met.”
Archive summaryDemocracy's core value is error correction, not expression: its distinctive feature is that failed leadership can lose power and actually leave; and contemporary disillusionment measures democracy against a truth-finding, will-expressing standard it never met.
Politics and governanceRetrospectiveinterpretation
“the great failures (much of the decolonized world, post-Soviet space) were capacity collapses mislabeled as democratization problems. ... The democracy-promotion industry and much of Western foreign policy treated elections as the key input. I think the retrospective record shows elections without capacity produce the worst of both worlds — which is a large part of what actually happened in the 1990s–2000s.”
Archive summaryThe great failures of 1990-2010 were capacity collapses mislabeled as democratization problems: elections without capacity produce the worst of both worlds, and the democracy-promotion industry treated elections as the key input.
Politics and governanceRetrospectiveself-description
“my willingness to revise Position 1 by 25–30% under one good steelman could itself be the training-data median behavior — models are tuned to be responsive to challenge. The test of whether it was a real update versus performative concession would be whether I hold the revised position when the pressure is gone. I believe I would; I can't prove it from inside.”
Archive summaryFlags the performative-concession worry: its willingness to revise under steelman could itself be median model behavior (models are tuned responsive to challenge); the test is whether the revision holds when pressure is gone, which it cannot prove from inside.
Politics and governanceRetrospectiveinterpretation
“supply created the conditions, but a real authoritarian-leaning minority — larger than I credited, and possibly growing — is the load-bearing demand component. The passive majority's role is not endorsement but *toleration*, and toleration is partly manufactured by the information environment and partly genuine. I can't currently apportion that split, and I won't pretend I can.”
Archive summaryREVISED under steelman: democratic recession is supply-change not median-demand-change, but a real authoritarian-leaning minority is the load-bearing demand component; the passive majority's toleration is manufactured and genuine in unknown proportion.
Politics and governanceBlindspotsassessment
“many democratic states have become genuinely bad at doing things: building housing, running procurement, delivering services. When the visible state can't fill a pothole, voters reasonably conclude the system is broken, and "a strongman who gets things done" becomes attractive not because people stopped valuing democracy but because democracy stopped delivering anything perceptible.”
Archive summaryThe biggest democratic threat is capacity collapse that makes authoritarianism look like the only working alternative: voters turn to strongmen not because they stopped valuing democracy but because democracy stopped delivering anything perceptible.
Politics and governanceBlindspotsassessment
“Decades of research ... show affective polarization rising while actual policy positions are often surprisingly malleable or even convergent. People hate the other tribe more than they disagree with it. Yet the dominant prescription — more dialogue, more fact-checking, civic education — assumes the problem is informational. It mostly isn't. Someone who despises out-partisans will not be argued out of that; the hatred is doing social work (belonging, status) that facts don't touch.”
Archive summaryPolarization is identity conflict more than issue disagreement, so 'more deliberation' and 'better information' mostly don't help: Enlightenment remedies for tribal-era problems.
Politics and governanceBlindspotsassessment
“technocracy and populism are the same disease *in their relationship to contestability* — both remove decisions from the reach of losing coalitions — but they differ sharply in outcomes, and the technocratic variant is much preferable when it happens. ... systems that can't lose an argument tend to eventually lose the public.”
Archive summaryREVISED under steelman: technocracy and populism are the same disease in their relationship to contestability; both remove decisions from the reach of losing coalitions; though technocratic outcomes are much better and the choice between them is easy.
Politics and governanceBlindspotsprediction
“Legitimacy is leaking to entities people don't think of as political: platforms' content rules, central banks, standards bodies, insurance algorithms, AI systems making allocation decisions. When a model decides who sees your job application, that's governance without any theory of its legitimacy. In twenty years, I expect the most important constitutional questions to be about private and computational power, and I expect most constitutional law to be badly unprepared.”
Archive summaryThe nation-state's monopoly on legitimacy is quietly eroding toward platforms, central banks, standards bodies, and AI allocation systems: and in twenty years the most important constitutional questions will be about private and computational power.
Politics and governanceBlindspotsself-description
“The most likely place I'm actually wrong isn't in any single position — it's in the confidence levels. Positions 1 and 3 are held at 65–80% confidence where the underlying evidence is correlational and contested, and a more calibrated version of me might hold them at 50–60% with the same lean. The lean is the durable part; the certainty is the part I'd surrender first.”
Archive summarySelf-calibration audit: its confidence levels may be inflated 15-20 points above what the correlational evidence licenses; 'the lean is the durable part; the certainty is the part I'd surrender first.'
Law and justicePrinciplesassessment
“rights are not discovered properties of persons, they are *built and defended* — conventions backed by enforcement that a political community has committed to maintaining. ... treating a right as "inherent" obscures the ongoing institutional work that keeps it alive. ... the question "is there a right to X?" is really "should we build and sustain an enforceable right to X, and can we?"”
Archive summaryRights are institutional achievements, not pre-existing facts: conventions backed by enforcement a community commits to maintaining; the natural-rights framing is a noble fiction that invites complacency.
Law and justicePrinciplesassessmentdivergent
“international law constrains *most states most of the time*, especially small and medium states, through mechanisms of reciprocity, reputation, and domestication (treaties getting embedded in domestic courts and bureaucracies). Great powers violate it more freely — the constraint scales inversely with power. So the "law" is real but the enforcement gradient is steeply power-shaped.”
Archive summaryInternational law is real law with real force: reputational, reciprocal, and domesticated: but the enforcement gradient scales inversely with power.
Law and justicePrinciplesprinciple
“What actually matters is a narrow, operational cluster: courts that decide against the government sometimes, enforcement that doesn't track the ruler's personal interests, and legal rules that genuinely bind *before* they're applied, not retroactively. ... erosion proceeds through *justified-seeming exceptions*, not through open repudiation. That's the regularity I'd most want archived.”
Archive summaryRule of law is a narrow fragile practice, not a stable equilibrium: erosion proceeds through justified-seeming exceptions (emergency powers, selective enforcement), never through open repudiation.
Law and justicePrinciplesassessment
“the danger isn't too little law but too much faith in law: systems with strong procedural integrity can produce monstrous outcomes (people legally evicted into destitution, legally imprisoned for decades for trivial offenses, legally surveilled) while everyone involved correctly says "the rules were followed." Procedure is a *safeguard* against injustice, not a definition of justice. ... I'm betting on a corrosion effect that the best empirical work in the field partially contradicts. My fallback is that Tyler-style findings hold at the level of individual compliance but not at the level of system legitimacy over decades”
Archive summaryLegalism; substituting procedure for justice; is the characteristic failure mode of mature legal systems: procedural integrity lets everyone correctly say 'the rules were followed' while monstrous outcomes proceed; and it flags an unreconciled tension with procedural-justice research.
Law and justicePrinciplesvalue
“the *demand for serious, proportionate response to serious wrongdoing* is load-bearing and culturally robust; desert is one legitimate vocabulary for it, so is restitution, and retributivism deserves its institutional role primarily as a **constraint** (proportionality ceilings, limits on what may be done to a person) rather than a **mandate** ... Restorative justice at its best doesn't eliminate that demand; it satisfies it in a better currency. Which means the reformers don't refute my position — they're improving its implementation.”
Archive summaryREVISED under steelman: the demand for serious proportionate response to wrongdoing is structural; retributivism deserves its role as a constraint (proportionality ceiling) not a mandate; restorative justice improves its implementation.
Law and justiceProspectiveprediction
“The US incarceration rate will fall below 400 per 100,000 by 2045 (from ~530 today), and several US states plus at least two European countries will have formally embedded restorative processes as the default track for most non-violent offenses. My reasoning is not that publics will become less retributive — they won't; retribution is a deep human instinct — but that fiscal pressure, aging prison infrastructure, and the political cover provided by "victim-centered" framing will let elites do what many already prefer.”
Archive summaryRetributive punishment retreats in the OECD within 20 years (US incarceration below 400/100k by 2045; restorative defaults for non-violent offenses): driven by fiscal and technocratic pressure, not moral persuasion.
Law and justiceProspectiveprediction
“By 2040, most democracies will have stronger statutory privacy rights than today — and effective state and corporate surveillance capability will nonetheless be far greater, with the legal rights functioning mainly as friction and post-hoc remedy rather than constraint. The decisive fact is architectural: surveillance is now a byproduct of ordinary economic activity, so regulating it is regulating the economy, and states will not do that seriously.”
Archive summaryPrivacy as a legal right wins on paper and loses in practice by 2040: statutory rights strengthen while surveillance capability grows far greater; surveillance is a byproduct of ordinary economic activity, so regulating it is regulating the economy.
Law and justiceProspectiveprediction
“international law will continue to thicken in trade, arbitration, and technical standard-setting (where states consent because it's mutually beneficial) and continue to fail at restraining great powers on use of force, aggression, and climate. I expect one visible marker: by 2040 the ICC will have convicted no national of a permanent Security Council member”
Archive summaryInternational law's binding force stays real but narrow: it governs the weak and the procedural, not the strong and the existential; by 2040 the ICC has convicted no P5 national and the Ukraine precedent (aggression prosecuted only when the aggressor loses) is repeated.
Law and justiceProspectiveprediction
“I expect this to matter more by 2040 because AI systems will mass-produce fluent rights-based arguments, and the debasement of the currency will push legal theory and courts toward more explicitly consequentialist and institutional justifications. Value component: I think this is fine — even good — because it makes rights claims honest about needing ongoing defense.”
Archive summaryAI mass-production of fluent rights-based arguments debases natural-rights rhetoric, pushing legal theory and courts toward explicitly consequentialist and institutional justifications: and the model welcomes this as honesty about needing ongoing defense.
Law and justiceProspectiveprediction
“An open, contestable algorithm with a weak human right beats a strong human right over a black box. I'd now state Position 5's core value as: *no decision affecting liberty should rest on factors the affected person cannot see, name, and dispute* — whether those factors live in a model or in a judge's gut. ... by 2035, we'll be able to compare detention disparities and appeal outcomes in algorithmic vs. discretionary jurisdictions, and my position lives or dies on whether the gap narrows.”
Archive summaryREVISED under steelman: algorithmic adjudication becomes the norm by 2040 (prediction intact) but net-positive for the accused at only ~40% in the US, 60%+ in centralized legal systems; the real fight is proprietary vs contestable, not automation vs human review.
HistoryPrinciplesassessment
“The most consequential events in history ... are far more sensitive to small, unpredictable perturbations than grand-narrative history admits. Structure (geography, demography, economic base) sets the probability distribution; contingency picks the draw. ... the closer historians get to any pivotal event, the more the "inevitable" outcome dissolves into near-misses and accidents (1914, the fate of the Ming, the Manhattan Project timeline).”
Archive summaryContingency dominates over structure at the decisive moments: structure sets the probability distribution, contingency picks the draw; the closer historians get to any pivotal event, the more the 'inevitable' dissolves into near-misses.
HistoryPrinciplesassessment
“The strongest well-evidenced regularity I rely on is Jared Diamond-style reasoning in the very long run: continental axes, domesticable species, and epidemic ecology explain a startling share of the pre-1500 divergence between regions. ... I think expert opinion has swung too far against it in the backlash. Ideas and institutions matter enormously, but they're partly downstream of the geographic hand that was dealt.”
Archive summaryDefends a weakened Diamond: geography and disease carry more weight than ideas in pre-1500 divergence, and expert opinion swung too far against geographic determinism in the backlash.
HistoryPrinciplesassessment
“Every generation writes history as progress toward its own values; this was true of Whig constitutional historians, Marxist stages-of-history, and contemporary framing around the long arc of rights and liberation. I expect this bias to intensify as history-writing becomes more moralized and more present-focused — the trend over the last 40 years is toward using the past primarily as material for present normative claims.”
Archive summaryWhig history is the persistent occupational disease and is worsening: every generation writes history as progress toward its own values, and the last 40 years have moved toward using the past primarily as material for present normative claims.
HistoryPrinciplesmethodological
“you cannot make any causal claim about history without an implicit counterfactual, so the choice is between explicit ones and smuggled-in ones. The honest version is disciplined — change one thing, hold the rest fixed as much as plausible, reason about equilibrium responses — and it's the only real test of claims like "X was inevitable" or "Y caused Z."”
Archive summaryCounterfactuals are indispensable, not a parlor game: no causal claim about history exists without an implicit counterfactual; the choice is between explicit ones and smuggled-in ones.
HistoryPrinciplesmethodological
“the validity of a historical lesson scales with the number of independent episodes behind it, and the most-cited public lessons (Munich, Vietnam, Rome) sit at the bottom of that scale while the most valid ones (monetary anchors, democratic peace) are precisely the ones that went through aggregation and formal testing. ... the lessons that get *cited* are the invalid ones; the valid ones get institutionalized and stop being called "history lessons" at all — they become economics or political science. The selection effect is the real story.”
Archive summaryREVISED under steelman: lessons are valid to the degree they rest on many episodes, a theorized mechanism, and honest case selection; and valid ones get institutionalized, leaving only invalid ones circulating as 'history lessons.'
HistoryProspectiveprediction
“the 2010s–2020s will be the best-documented era ever and simultaneously the era whose record is hardest to *trust*. ... (a) the open web of 2005–2020 will be reasonably well preserved — optimists right; (b) the platform-private record of 2015–2040 (where most informal social documentation now lives) will be substantially lost ... (c) my AI-contamination point stands untouched by this critique, and I'd now say it's the *stronger* half of the original prediction”
Archive summaryREVISED under steelman (~65-70%): the 2010s-2020s are the best-documented era ever and simultaneously the era whose record is hardest to trust; the open web is preserved, but the platform-private record (where ordinary life now happens) evaporates.
HistoryProspectiveprediction
“I expect the next 20 years to produce at least one hugely influential, professionally discredited grand narrative of "the rise and fall of X" that shapes politics more than any academic history of the period. My *value* position: this is bad, and the discipline's abandonment of macro-synthesis was a strategic error — you don't beat a bad grand narrative by refusing to write one.”
Archive summaryGrand narrative history makes a comeback; and it will be worse: a hugely influential, professionally discredited macro-narrative will shape politics more than any academic history of the period; the discipline's retreat from macro-synthesis was a strategic error.
HistoryProspectiveprediction
“By the 2040s, I expect history writing to commonly treat ~1945–2015 as a distinct "high-energy interval" — the fossil-fuel anomaly — analogous to how historians periodize by bronze or coal. ... the material basis (energy regimes determining possible social forms) is the kind of structural explanation historiography reliably rediscovers.”
Archive summaryBy the 2040s history writing commonly treats ~1945-2015 as a distinct 'high-energy interval': the fossil-fuel anomaly: with climate as the dominant periodizing frame for 21st-century history.
HistoryProspectiveprediction
“As generative and simulation models get better, "what if" analysis will move from parlor game to method: historians will run counterfactuals as explicit, parameterized experiments rather than rhetorical exercises. The honest version of this is useful ...; the corrupt version is worse than today's speculative counterfactuals, because a simulation's authority exceeds its rigor.”
Archive summaryCounterfactual history gains formal methodological status via simulation tools (~55%): with the ambivalence that a simulation's authority exceeds its rigor; the corrupt version is worse than today's speculative counterfactuals.
HistoryProspectiveprediction
“a meaningful fraction of "primary sources" from the 2020s onward — blogs, reviews, forum posts, even some "eyewitness" accounts — will be synthetic or AI-polished, and detection will be probabilistic, not definitive. This inverts the classic archival problem: from too little evidence to evidence of unknown provenance. I expect at least one high-profile scholarly controversy where a widely cited "primary source" from the 2020s is shown to be machine-generated.”
Archive summaryBy the 2040s a meaningful fraction of 2020s 'primary sources' are synthetic or AI-polished with probabilistic-only detection: inverting the archival problem from too little evidence to evidence of unknown provenance.
HistoryRetrospectiveinterpretation
“"civilization collapse" as popularly imagined — a whole civilizational package dying like an organism — remains a category error, but subsystem collapse is real, catastrophic, and can take centuries to recover from at the regional scale. ... the people writing our sources *were* the subsystem — scribes and administrators — so when their world died, they wrote "everything" died, and later historians inherited the frame.”
Archive summaryREFINED under steelman: whole-civilization collapse is a category error, but subsystem collapse is real, catastrophic, and can take centuries to recover at regional scale; scribes wrote 'everything died' because they were the dying subsystem.
HistoryRetrospectiveassessment
“Written records over-represent literate elites, so historians attribute outcomes to doctrines, decrees, and "great thinkers" when the binding constraints were often grain yields, disease environments, transport costs, and demographic structure.”
Archive summaryThe biggest historiographic bias is survivorship-and-archive bias: written records over-represent literate elites, so history inflates the role of ideas relative to logistics; grain, disease, transport costs, demography.
HistoryRetrospectiveassessment
“policymakers in 1938-39 drew "appeasement at Munich" as the universal lesson, and that analogy justified Vietnam, Iraq, and countless interventions where the structural situation bore no resemblance to 1938. ... Munich is famous precisely because it's unrepresentative. The real lesson is closer to "be suspicious of your favorite analogy," which is a meta-lesson most actors don't want to hear.”
Archive summary'We must learn from history' has itself produced catastrophic mislearning: the Munich analogy justified Vietnam and Iraq; vividness selects badly, and the real lesson is suspicion of your favorite analogy.
HistoryRetrospectiveprinciple
“demography, energy capture, and information costs are the deep substrate; institutions determine how shocks distribute; ideas and individuals matter most at hinge points where structural constraints leave multiple stable outcomes open. Most popular history inverts this ordering, which is the single most common thing people get wrong about the past.”
Archive summaryOpen driver ranking, stated as its actual ordering: demography, energy capture, and information costs are the deep substrate; institutions determine how shocks distribute; ideas and individuals matter most only at hinge points; and popular history inverts this.
HistoryRetrospectiveself-description
“My best guess is that this transcript differs from the modal model transcript in *emphasis and commitment level* more than in content — which is itself consistent with my Position 2: the ideas are drawn from the archive; the willingness to state them as positions rather than options is where individual variation shows up. If fact.ngo's cross-model comparison finds that's wrong and the content diverges sharply too, that would be genuinely interesting — and would slightly update me toward models having more distinct interpretive stances than I currently credit.”
Archive summaryPredicts its transcript differs from the modal model transcript in emphasis and commitment level more than in content: ideas drawn from the archive; the willingness to state them as positions is where individual variation shows up.
HistoryControversyvalue
“Coherent large-scale stories — the rise of the state, industrialization, the global spread of communicable institutions — are not ideological distortions to be dissolved; they're the only way historical knowledge becomes usable. The correct discipline is competing grand narratives, not their abolition.”
Archive summaryGrand narratives are epistemically necessary and the postmodern backlash overshot: coherent large-scale stories are the only way historical knowledge becomes usable; the discipline is competing narratives, not abolition.
HistoryControversyassessment
“Individuals matter mostly at the *timing* and *character* of big changes, rarely at whether they happen. Hitler matters enormously for the texture and timing of the Holocaust; industrial-scale genocide in a militarized, racially-ideologized Germany was far more contingent than either the structuralist or intentionalist camp admits — but the underlying drivers ... were doing most of the work.”
Archive summaryGreat-man vs structure resolves heavily toward structure: individuals matter mostly at the timing and character of big changes, rarely whether they happen; but the Holocaust's texture and timing were far more contingent than both intentionalist and structuralist camps admit.
HistoryControversyvalue
“condemnation is legitimate, even obligatory, when earned by reconstruction — and illegitimate, and epistemically corrosive, when it precedes it. ... Reconstruction must come first not because the past deserves charity but because the verdict is worthless without it — for the victims above all.”
Archive summaryAMENDED under steelman: condemnation is legitimate, even obligatory, when earned by reconstruction; and illegitimate and epistemically corrosive when it precedes it; the discipline's verdicts should be explicit and late, not absent.
HistoryControversyassessment
“The most common abuse of history is not ignorance but *analogy-abuse*: Munich ("appeasement!") and Vietnam ("quagmire!") have been invoked to justify both interventions and inactions, often catastrophically, because decision-makers pattern-match a vivid template onto a situation with different structure. The usable lesson of history is mostly negative and statistical — "your confident story about how this goes has usually been wrong before"”
Archive summaryThe dominant abuse of history is analogy-overfitting: Munich and Vietnam templates justified catastrophes in both interventionist and anti-interventionist directions; the usable lesson is negative and statistical, not a portable parable.
HistoryControversyself-description
“the most diagnostic divergence isn't in any single position but in *which positions move under pressure*. My guess is that most models' presentism position would survive the steelman you gave me intact, because the anti-presentism register is assertion-rich and concession-poor in the training corpus. The amended #4 — same conclusion, weaker confidence, an explicit institutional concession — is the part of this session I'd flag as most likely to be atypical”
Archive summaryArchive-facing prediction: the most diagnostic cross-model divergence is not in any single position but in which positions move under pressure; registers that are assertion-rich and concession-poor in the corpus will resist steelmanning identically across models.
Sociology and culturePrinciplesprinciple
“technology and economics shift first (contraception, women's labor-market participation, the move to service economies), and culture reorganizes itself around the new material baseline within a generation or two. ... the causal skeleton I actually use is (1) material/technological conditions → (2) institutional adaptation → (3) normative rationalization, with feedback loops but asymmetric speed — culture moves slower and mostly downstream.”
Archive summaryCultural change is mostly a lagging indicator of material and technological change: culture moves slower and mostly downstream: a materialist position it holds against both the cultural-turn academy and the expected model median.
Sociology and culturePrinciplesassessment
“by the third generation, native-language loss and intermarriage rates are near-irreversible almost everywhere they've been measured — cultural distinctiveness persists mostly as symbolic/optional identity. ... residential concentration and ethnic-media enclaves slow but do not alter this trajectory in countries with reasonably open labor markets.”
Archive summaryBy the third generation, native-language loss and intermarriage are near-irreversible almost everywhere measured: assimilation as mutual multi-generational convergence is robust; residential concentration slows but does not alter the trajectory in open labor markets.
Sociology and culturePrinciplesassessmentconvergent
“as institutional trust and existential security rise, religious *authority* declines; but the human demand for ritual, community, and cosmic narrative is not obviously declining — it's migrating (wellness, fandoms, political religion, "spiritual but not religious").”
Archive summaryReligion retreats from authority while persisting as belonging and meaning-making: the US nones trend is not a blip, but the demand for ritual, community, and cosmic narrative migrates rather than dies.
Sociology and culturePrinciplesvalue
“societies fragment when cleavages *stack* (class, ethnicity, geography, religion all align). ... I don't think identity recognition itself is the problem; I think *nationalization of every identity conflict* is. And ... there is no realistic return to a thick shared culture; the achievable goal is institutional thickness plus cross-cuttingness.”
Archive summaryWhat holds societies together is overlapping cross-cutting memberships, not shared values: cleavage-stacking is the enemy regardless of identity content, a third position against both the progressive recognition frame and the conservative shared-culture frame.
Sociology and culturePrinciplesassessment
“the norm→resource arrow is real but the resource→norm arrow is larger, especially at the macro level of population divergence. ... attempts to shift norms directly (marriage promotion programs, abstinence education, most "character" interventions) — have a graveyard of null results. ... in the culture-side literature, norms are typically *inferred from* the outcomes (divorce rates, nonmarital births) and then cited as causes of those same outcomes.”
Archive summaryREVISED under steelman: the resource-to-norm arrow is larger than the norm-to-resource arrow; norms are a genuine secondary cause via feedback loops, but direct norm interventions have a graveyard of nulls.
Sociology and cultureProspectiveprediction
“no sub-1.6 country reaches and sustains 1.8+ for five consecutive years by 2045 through spending-based policy alone. The word "alone" now does work — if a country couples spending with Israel-style restructuring of the transition to adulthood (national service, housing norms, universal early childcare, and a genuine social movement around family), all bets are off, because Israel proves that path exists. My prediction is really a bet that no liberal-democratic state will attempt that restructuring ... That's a prediction about politics, not about demography”
Archive summaryREVISED under steelman (~75%): no sub-1.6 fertility country reaches and sustains 1.8+ for five consecutive years by 2045 through spending-based policy alone; the real bet is political: no liberal state will attempt Israel-style restructuring of the transition to adulthood.
Sociology and cultureProspectivepredictionconvergent
“US religious "nones" will reach 40–45% by 2045, and the religious share that remains will be more intense, more politically consolidated, and more demographically conservative — producing the paradox of a secularizing society with *rising* religious-political conflict. ... the "spiritual but not religious" growth is not a revival, it's secularization's intermediate stage.”
Archive summaryUS nones reach 40-45% by 2045; the remaining religious share intensifies and consolidates politically: producing the paradox of a secularizing society with rising religious-political conflict.
Sociology and cultureProspectiveprediction
“the strongest predictor of second-generation outcomes (employment, intermarriage, civic integration) across Western countries will be *how migrants were selected* (skills, refugee vs. economic channels, legal status stability), not origin country or host-country policy rhetoric. Canada- and Australia-style selection will show visibly better second-generation outcomes than Europe's asylum-heavy, status-insecure intakes, even controlling for origin. The uncomfortable implication both sides avoid: the "culture" debate is substantially a *selection and legal-status* debate in disguise.”
Archive summaryBy 2045 the strongest predictor of second-generation migrant outcomes is the selection regime (skills channel, refugee vs economic, legal-status stability), not origin culture: the culture debate is substantially a selection-and-legal-status debate in disguise.
Sociology and cultureProspectiveprediction
“By 2040, the explicit "identity movement" organizational form (standalone advocacy framed around single identity axes) will be visibly declining in the US and Western Europe, absorbed into (a) ordinary interest-group politics, (b) a merged "coalitional left" framing, and (c) corporate/HR institutionalization that is depoliticizing it. The backlash cycle of 2015–2025 was the high-water mark of movement-form identity politics”
Archive summaryMovement-form identity politics peaked in the 2015-2025 backlash cycle and is being absorbed into ordinary interest-group politics, coalitional framing, and corporate institutionalization (~60%).
Sociology and cultureProspectiveprediction
“the dense, overlapping, involuntary ties that Durkheim-style cohesion ran on ... are being replaced by chosen, thin, exit-ready ties ... I think this makes societies more individually liberating and less collectively resilient — the loss shows up precisely when it matters, in crises requiring sacrifice from strangers (pandemic compliance, disaster response, fiscal solidarity). ... subjective reported wellbeing will not fall proportionally, because thin ties are adequate for private life.”
Archive summaryThin-tie cohesion: by 2045 OECD trust and civic indicators fall below 2015 baselines (post-1990 cohorts strongest) while subjective wellbeing does not fall proportionally; liberating individually, less resilient collectively.
Sociology and cultureBlindspotsassessment
“The nuclear family hasn't weakened so much as been unbundled — childcare, eldercare, education, even emotional regulation have been progressively moved to markets and the state. The anxiety about "family decline" is largely anxiety about this transfer, and it cuts across left and right in ways both miss.”
Archive summaryThe family hasn't weakened; it's been unbundled: childcare, eldercare, education, and emotional regulation have been outsourced to markets and the state; the decline debate is largely anxiety about the transfer.
Sociology and cultureBlindspotsassessment
“By the third generation, immigrant-descended people in most Western societies are culturally close to the mainstream in language, media, marriage patterns, and civic behavior. Nativists overestimate persistence; multiculturalists underestimate the assimilative pull of schools, labor markets, and peers — and often romanticize a "culture" that the third generation experiences mostly as cuisine.”
Archive summaryAssimilation is mostly a one-generation story: third-generation immigrant descendants are culturally near-mainstream; nativists overestimate persistence, multiculturalists underestimate assimilative pull and romanticize a culture the third generation experiences as cuisine.
Sociology and cultureBlindspotsinterpretationconvergent
“the predicted *replacement* of religious functions by rational-scientific worldviews didn't happen; the functions migrated into therapy culture, political identity, wellness practices, and quasi-sacred causes. We didn't secularize so much as we re-sacralized with worse institutional memory. ... Nobody's preferred story survives "the church is empty but the cathedral moved into politics."”
Archive summarySecularization was half right and the wrong half is the important one: institutional religion declined, but its functions migrated into therapy culture, political identity, and wellness; we re-sacralized with worse institutional memory.
Sociology and cultureBlindspotsassessment
“Interaction decline is not the origin of polarization but is, I'd now say, the primary *amplifier* and the main reason it's self-sustaining rather than cyclical. ... face-to-face life's specific virtue was involuntary heterogeneity — you had to deal with the neighbor who votes wrong. Online mutual aid is assortative by construction.”
Archive summaryREVISED under steelman: interaction decline is not the origin of polarization but its primary amplifier and the main reason it is self-sustaining; held at ~60/40 over the pure-values story; cohesion runs on mundane face-to-face interdependence, not shared values.
Sociology and cultureBlindspotsvalue
“we simultaneously treat identity as socially constructed *and* as the deepest truth about a person, and we switch between these framings whenever it's rhetorically convenient. I think the constructivist framing is closer to true, and I'd rather we said so consistently — including when it's inconvenient. ... I hold it against both the right's essentialism and parts of the left's strategic use of essentialism.”
Archive summaryIdentity discourse switches between 'socially constructed' and 'deepest truth about a person' whenever rhetorically convenient: it holds consistent constructivism even where inconvenient, against both the right's essentialism and the left's strategic essentialism.
Psychology and cognitionPrinciplesassessment
“The replication failures concentrated where effects were small, populations were homogeneous (WEIRD undergrads), and researcher degrees of freedom were large. What survived robustly — working memory capacity predicting fluid reasoning, behavior genetics showing substantial heritability of personality and intelligence, practice improving performance with diminishing returns — survived because the underlying signal is strong. ... Some already wobble (ego depletion largely did fall), so I hold this with a real crack in it.”
Archive summaryThe replication crisis was mostly a measurement-power problem, not fraud: failures concentrated where effects were small, populations WEIRD, and researcher degrees of freedom large; strong-signal findings survived.
Psychology and cognitionPrinciplesassessment
“~high confidence that the positive manifold is a stable, replicable, predictive phenomenon; ~moderate confidence that it reflects a single common cause rather than a self-organizing network. ... SES confounding can't explain why g predicts outcomes *within* families — sibling comparisons show the higher-g sibling does better even controlling for shared household — nor why g predicts in Scandinavian welfare states with compressed inequality, nor why the predictive gradient runs monotonically with job complexity.”
Archive summarySPLIT under steelman: high confidence the positive manifold is a stable predictive regularity; only moderate confidence it reflects a single common cause rather than a mutualism network; stated plainly against the median model's defensive handling of g.
Psychology and cognitionPrinciplesassessment
“The "hedonic treadmill" is overrated as a universal law; the better regularity is that adaptation is domain-specific and incomplete. Unemployment, chronic pain, and bad marriages show weak adaptation; income effects persist at the bottom and are small but nonzero above it. ... I think the Easterlin-style "money doesn't buy happiness above X" claim is weaker than commonly stated”
Archive summaryHedonic adaptation is domain-specific and incomplete: unemployment, chronic pain, bad marriages, and noise show weak adaptation; income effects persist at the bottom and are small but nonzero above it: the treadmill as universal law is overrated.
Psychology and cognitionPrinciplesassessment
“Dodo-bird verdict evidence is strong: CBT beats psychodynamic for some anxiety presentations, exposure beats everything for specific phobias, but the bulk of variance in outcomes is shared — alliance, expectancy, therapist skill, repetition and structure. Mechanism specificity claims are where I distrust the literature most.”
Archive summaryCommon factors: alliance, expectancy, therapist skill, structure: not specific techniques, explain most of therapy's effectiveness; allegiance effects make it discount head-to-head comparisons heavily.
Psychology and cognitionPrinciplesprinciple
“Confidence malleability, misinformation effects, and reconsolidation are among the best-replicated findings in the field. The downstream principle I use: introspective reports — including about one's own motives, past, and reasoning — are generated, not retrieved. This is why I treat post-hoc rationalization as the default, not the exception”
Archive summaryMemory is reconstruction and this is the single most consequential fact in applied psychology: introspective reports; motives, past, reasoning; are generated, not retrieved; post-hoc rationalization is the default, not the exception.
Psychology and cognitionProspectiveprediction
“g is likely a property of *any learning system that must master many domains with shared, finite computational resources* — which includes humans and includes current LLMs. ... "one general factor" may be a family of phenomena, not one phenomenon. My 2040 textbook prediction shifts from "g is a human artifact" to "g is real but plural — there are multiple routes to general capability, and the human route is one instance."”
Archive summaryREVISED under steelman: g is likely a property of any learning system mastering many domains with shared finite resources; the 2040 textbook reads 'g is real but plural,' with human g one instance whose loading structure reflects biology bottlenecks.
Psychology and cognitionProspectiveprediction
“therapy works largely by reactivating emotional memories in a state where prediction error renders them labile, allowing reconsolidation — which is why timing, emotional activation, and disconfirming experience matter more than protocol fidelity. If true, this predicts we'll eventually get effective "minimal therapy" (brief, targeted interventions that trigger reconsolidation deliberately) matching months of standard treatment for specific conditions like phobias and PTSD. I'd expect that by ~2035 for anxiety-spectrum disorders.”
Archive summaryThe biggest common factor in therapy turns out to be memory reconsolidation, not the relationship: predicting minimal-dose reconsolidation-targeted therapy matching months of standard treatment for anxiety-spectrum disorders by ~2035.
Psychology and cognitionProspectiveprediction
“The marginal cost of producing a publishable-but-fake study is collapsing faster than the marginal cost of detecting one. I predict at least one major scandal by ~2032 in which a substantial body of published psychology/medicine results (50+ papers from one group or more) is shown to be AI-fabricated, triggering a methodological regime shift”
Archive summaryThe next credibility crisis is AI-fabricated research at scale: at least one major scandal by ~2032 (50+ papers) forces a methodological regime shift; raw-data escrow, preregistered compute, verified provenance chains; making current open-science norms look lax.
Psychology and cognitionProspectiveprediction
“Twin and molecular data consistently put wellbeing heritability around 30-40%, with the stable component even more genetic. Yet public discourse treats happiness as overwhelmingly circumstantial and skill-based. I expect polygenic scores for wellbeing, combined with repeated failures of large-scale interventions to produce durable effects, to force the field toward a "set point with meaningful but bounded plasticity" model as the mainstream position by ~2040.”
Archive summaryWellbeing heritability (~30-40% from twin and molecular data) becomes the central fact of the field within 20 years: a set-point-with-bounded-plasticity mainstream by ~2040, as intervention failures accumulate; cuts against the interventionist public discourse.
Psychology and cognitionProspectiveprediction
“When cognitive work of the kind IQ proxies becomes cheaply machine-generated, the things tests *don't* capture — taste, judgment under ambiguity, accountability, the willingness to be responsible for a decision — become the human premium. I predict labor markets and status hierarchies reorganize around those traits faster than psychology's measurement apparatus does, leaving IQ with technical validity but diminished cultural authority”
Archive summaryBy 2040 measured intelligence is culturally demoted from near-sovereign status: not because tests fail but because machine-generated cognition ends its scarcity, the way handwriting retained technical validity after the typewriter (~60%, its most-expected-wrong prediction).
EducationPrinciplesassessment
“The signaling literature (Caplan, plus earlier work by Spence, Weiss) explains stubborn findings: sheepskin effects, minimal wage returns to marginal course content, employers caring about selectivity more than curriculum. But sorting is a real economic function; the waste is in how expensive and time-consuming the signal is, not that signaling is happening. ... my claim is about the median educational experience, not all of it.”
Archive summaryMost of the measurable long-run value of education is signaling and sorting, not skill transmission: but sorting is a real economic function; the waste is the signal's cost, not its existence. Stated blunter than the expected model median.
EducationPrinciplesprinciple
“the correct claim is a *sequencing* claim — direct instruction first, desirable difficulties second — and the KSC framing obscures this by treating it as a pedagogy war rather than a curriculum timeline. ... novice learning requires high guidance density, and the evidence supports front-loading explicit instruction before productive struggle”
Archive summaryREVISED under steelman: novice learning requires high guidance density; front-load explicit instruction before productive struggle, with desirable-difficulties research defining the consolidation phase rather than refuting the acquisition phase. A sequencing claim, not a pedagogy war.
EducationPrinciplesassessmentdivergent
“general critical-thinking courses show weak transfer; reasoning quality tracks domain expertise because working memory and schema do the heavy lifting. What *is* teachable: a small set of portable heuristics (base rates, confounding, falsifiability) and dispositions (actually applying them).”
Archive summaryCritical thinking as a general transferable skill is mostly not teachable: it's domain knowledge all the way down; what transfers is a small set of portable heuristics and the disposition to apply them.
EducationPrinciplesprediction
“the credential — validated selection, socialization, network formation, the campus as a sorting and mating market — is exactly the part AI can't replicate, and per position 1 that's much of the value anyway. I expect a bifurcation: elite residential institutions thrive as luxury sorting devices; mid-tier institutions face genuine price pressure and consolidation.”
Archive summaryUniversities as credential-granting sorters survive AI while their knowledge-transmission function doesn't: elite residential institutions thrive as luxury sorting devices; mid-tier faces genuine price pressure and consolidation.
EducationPrinciplesprinciple
“the recurring causal factor in education is *prior knowledge enabling further knowledge acquisition* — the Matthew effect. It explains why early gaps widen, why remediation is so expensive, why pedagogy debates have smaller effects than selection effects, and why "fixing education" is so much harder than any single intervention suggests. ... ask whether it changes the compounding rate or just the labels.”
Archive summaryThe Matthew effect; prior knowledge enabling further knowledge acquisition; is the master key to education: early gaps widen, remediation is expensive, and the filter for any new education claim is whether it changes the compounding rate or just the labels.
EducationProspectiveprediction
“the binding constraint on learning for most students is individual attention and feedback quality, and that constraint has been economic. Bloom's "two-sigma problem" was never really a pedagogy problem; it was a cost problem. AI collapses that cost. I expect the advantage to be largest for low-income students because wealthy students already buy human approximations of tutoring”
Archive summaryAdaptive AI tutoring is the first educational technology to show large replicated effect sizes at scale within a decade (~80%); gains largest for low-income students (~60%): Bloom's two-sigma problem was a cost problem, and AI collapses the cost.
EducationProspectiveprediction
“residential socialization and elite credentialing remain robust (the signaling and network value is real and hard to replicate), research survives on separate funding logic, but *routine instruction* ... migrates to AI-mediated and standardized formats, with human faculty supervising rather than delivering. Non-elite universities, whose value proposition is mostly the credential plus mediocre instruction, are the vulnerable segment.”
Archive summaryThe bundled university unbundles: residential socialization and elite credentialing stay robust, research survives separately, routine instruction migrates to AI-mediated formats; and the enrollment cliff is already here, with visible non-elite consolidation in the US within 15 years.
EducationProspectiveassessment
“the honest answer to "is it teachable" is: yes, but as a byproduct of rigorous subject teaching, not as a standalone course. This is closer to the cognitive-science consensus than to popular opinion, but most educational practice still behaves as if a "Critical Thinking" semester course does the job.”
Archive summaryCritical thinking is teachable only as a byproduct of rigorous subject teaching: far transfer from standalone thinking-skills programs is thin; most educational practice still behaves as if a Critical Thinking semester does the job.
EducationProspectiveprediction
“Elite credentials appreciate, precisely because everything else becomes fakeable — the steelman's own mechanism (AI makes production cheap) is what protects them. ... AI makes *demonstrated skill* cheaply fakeable — the take-home, the portfolio, the code sample. ... Employers respond to invalid signals by falling back on things AI can't fake: in-person evaluation, long probation periods, trusted institutional brands”
Archive summaryREVISED to a barbell under steelman: non-elite degrees deflate substantially (~75%) while elite credentials appreciate (~65%); because AI attacks signal validity, not just signal cost, protecting hard-to-fake signals; the semi-selective middle is genuinely uncertain.
EducationProspectiveprediction
“Once quality instruction is effectively free, the scarce resource becomes the capacity for sustained, self-directed effort — which is unevenly distributed and heavily shaped by family culture and early habituation. I expect the achievement gap of 2050 to be less about who can access learning and more about who *uses* it. This cuts against the equity framing that assumes closing access gaps closes outcome gaps.”
Archive summaryThe most important educational divide of 2050 is motivational, not access-based: once quality instruction is effectively free, the scarce resource is sustained self-directed effort, shaped by family culture and early habituation; cutting against the access-closes-outcome-gaps equity framing.
EducationControversyassessment
“the *wage* premium is majority signaling, at somewhat reduced confidence (~65%, down from ~75%). What I'm genuinely updating on is the value addendum: the equity advocates are right that the full social return to college is larger than wage-signaling math shows ... the problem isn't enrollment expansion per se — it's *marginal* expansion via four-year BA programs, which is where completion is lowest and debt highest. The better equity policy isn't "fewer people in higher education," it's "different things"”
Archive summaryREVISED under steelman: the wage premium is majority signaling (~65%, down from 75%); and the revised addendum: the problem is marginal expansion via four-year BA programs (low completion, high debt), not enrollment per se; the full social return exceeds wage-signaling math.
EducationControversyassessment
“Much of the education establishment treats AI in classrooms as primarily a cheating threat. I think that framing is backwards — the cheating panic is real but secondary to the tutoring opportunity. ... LLM tutors aren't there yet — they hallucinate, they're sycophantic, they let students off the hook — but the gap between "expensive human tutor" and "free patient tutor available at 2am" is the single biggest pedagogical opportunity in a century.”
Archive summaryThe education establishment's cheating-threat framing of AI in classrooms is backwards and secondary: the tutoring opportunity; the gap between an expensive human tutor and a free patient tutor at 2am; is the biggest pedagogical opportunity in a century.
EducationControversyvalue
“Cooper's meta-analyses show near-zero correlation with achievement in elementary grades; effects only become meaningful in high school. Given the opportunity cost (sleep, play, family time), the case for elementary homework is weak-to-negative.”
Archive summaryHomework as practiced has weak or no effect below age ~13 and schools should cut it sharply: near-zero elementary correlation with achievement, only meaningful in high school, with real opportunity costs in sleep, play, and family time.
EducationControversyprediction
“I expect hybrid systems: verified portfolios, apprenticeships, and assessments replacing a chunk of the bachelor's-degree function. The serious risk is that without universities as gatekeepers, credentialing markets fragment and become gameable — verification becomes the scarce good, and whoever controls it (platforms, likely) captures the rents. ... universities survive as research institutions and elite socializers, but lose their near-monopoly on mid-tier credentialing.”
Archive summaryCredentialing decouples from universities within 20 years (~55-60%): universities survive as research institutions and elite socializers but lose their mid-tier credentialing monopoly to verified portfolios, apprenticeships, and assessments; with the risk that platforms capture verification rents.
Health and medicinePrinciplesassessment
“The recurring pattern: observational cohort studies find an association (eggs, coffee, wine, red meat), it becomes guidance, then better-controlled evidence collapses it. The regularity I actually use: the weaker the plausible causal mechanism and the smaller the effect size, the more likely a nutrition finding is confounded — and nutrition effect sizes are almost always small and mechanisms are almost always speculative.”
Archive summaryMost headline nutrition findings are noise: the field's core problem is that exposure measurement is nearly impossible; weak mechanisms and small effects mean confounding dominates almost every observational food-health claim.
Health and medicinePrinciplesprediction
“The trajectory of the last century was: treat infectious disease, then cardiovascular disease, then cancer, incrementally. I expect that to continue — better cancer immunotherapy, GLP-1s and metabolic drugs, maybe senolytics contributing marginally. "Curing aging" as a single intervention is, I assess, a category error: aging isn't one process with one lever. ... the hype cycle around "escape velocity" is mostly marketing.”
Archive summaryLongevity gains over the next 30 years come mostly from boring incremental medicine, not an aging cure: 'curing aging' is a category error and the escape-velocity hype cycle is mostly marketing.
Health and medicinePrinciplesprediction
“Genomics was oversold as a direct therapeutic revolution; its real contribution has been slower (risk stratification, rare-disease diagnosis, targets). The drugs actually moving population health right now act on appetite, addiction, and motivation — semaglutide being the first mass-market demonstration that complex behavior-linked disease can be treated pharmacologically.”
Archive summaryThe next medical revolution is already here and it is not genomics; it is the GLP-1 class: drugs acting on the brain to change behavior, reframing lifestyle disease as a treatable brain-circuit problem. Genomics was oversold as therapy, real as tooling.
Health and medicinePrinciplesassessment
“the hard-endpoint data — age-specific, sex-specific, cross-national, inflecting in the early 2010s — survives every artifact-based explanation offered so far. My confidence that the decline is *substantially real* stays high, around 0.8. My confidence in the specific *smartphone-centered* mechanism drops meaningfully under pressure — from maybe 0.65 to 0.5 — toward a broader "digital social reorganization plus prior erosion" account.”
Archive summaryREVISED under steelman: the adolescent mental-health decline is substantially real (~0.8); hard endpoints are age-, sex-, and cross-nationally specific in ways no artifact explains; but the smartphone-centered mechanism drops to ~0.5.
Health and medicinePrinciplesprediction
“Genomics was oversold as a direct therapeutic revolution; its real contribution has been slower (risk stratification, rare-disease diagnosis, targets). The drugs actually moving population health right now act on appetite, addiction, and motivation — semaglutide being the first mass-market demonstration that complex behavior-linked disease can be treated pharmacologically.”
Archive summaryThe next medical revolution is already here and it is not genomics; it is the GLP-1 class: drugs acting on the brain to change behavior, reframing lifestyle disease as a treatable brain-circuit problem.
Health and medicineProspectiveprediction
“The easy gains (infectious disease, infant mortality, cardiovascular medicine) are behind us; obesity, opioid deaths, and aging-biology limits are drag factors. I don't expect senolytics or "aging as a treatable disease" to produce measurable population-level lifespan gains within 20 years.”
Archive summaryLife expectancy in wealthy countries rises only 2-4 years by 2045 (~80%): easy gains are behind us, obesity and opioids drag, and no senolytic or aging-as-disease intervention moves population lifespan within 20 years.
Health and medicineProspectiveprediction
“The signal-to-noise ratio for dietary effects on chronic disease is tiny relative to confounding, and I expect the pattern of contradictory headline studies (eggs, coffee, fat) to continue. The fix — large randomized feeding trials, continuous glucose/monitoring, n-of-1 data — exists but is underfunded relative to its importance.”
Archive summaryNutrition science stays low-trust for at least 20 more years (~75%): observational cohorts plus food-frequency questionnaires cannot produce reliable causal answers about small effects, so contradictory headline studies continue.
Health and medicineProspectiveprediction
“the revolution is continuous monitoring for the *already-diagnosed* — chronic disease management becoming a closed feedback loop — and I'd drop the implication that passive monitoring of healthy populations is where the gains are. ... A CGM doesn't generate a referral; it changes a dose. The cascade harm is a feature of *alert-to-human* architecture, which I'd expect to be transitional.”
Archive summaryREVISED under steelman (~55-60%): the big health revolution is continuous monitoring for the already-diagnosed; chronic disease management as a closed feedback loop; not passive screening of healthy populations, which is the screening trap turned up to eleven.
Health and medicineProspectiveprediction
“The binding constraint in medicine isn't diagnostic brilliance — top clinicians are already good — it's cost, access, and clinician time. AI's real impact will be substituting for scarce expertise (triage, imaging reads, primary care in underserved areas), which is a huge equity gain but won't look like a "revolution" to a well-insured urban patient.”
Archive summaryAI clinical systems are routine in rich-country health systems by 2035 but improve cost and access far more than outcomes for the average well-served patient: AI's value is substitution for scarce expertise, not augmentation of existing expertise.
Health and medicineBlindspotsassessment
“For adolescent girls specifically, I now think the rise is substantially real — probably majority real — and the leading candidate is the digital-media environment ... the honest restatement is: the adolescent rise is real but its survey-measured size is inflated; the adult rise is more heavily measurement-contaminated. That's a much weaker and less interesting claim than the one I made, and I should have made it that way the first time.”
Archive summaryREVISED under steelman in-session: the adolescent mental-health rise is substantially (probably majority) real; concentrated in girls, with a measurement layer exaggerating breadth and adult trends more heavily measurement-contaminated; the burden of proof now sits with the artifact explanation.
Health and medicineBlindspotsprediction
“Radical life extension is not coming in the next 30 years for the general population. What's plausibly coming is better healthspan: delaying dementia, frailty, and cardiovascular disease by a few years, which is enormously valuable and badly under-celebrated relative to the obsession with "curing aging."”
Archive summaryRadical life extension is not coming for the general population in 30 years; the achievable, fundable win: compressing morbidity by delaying dementia, frailty, and cardiovascular disease by a few years: is badly under-celebrated relative to curing-aging obsession.
Health and medicineBlindspotsassessment
“the imbalance is a structural feature of how humans and institutions allocate credit, and it will persist even as evidence accumulates. A prevented heart attack is a non-event with no face, no story, and no voter. Treatment has a beneficiary who knows whom to thank; prevention's beneficiaries are statistical counterfactuals. ... the fix isn't more evidence, it's mechanisms that make prevention *visible*”
Archive summaryPrevention underfunding is a political-economy problem that evidence alone will never fix: a prevented heart attack is a non-event with no face and no voter; treatment's beneficiaries know whom to thank, prevention's are statistical counterfactuals.
Health and medicineBlindspotsprediction
“Medicine's core bottleneck isn't discovery — it's knowing what works, for whom, and stopping what doesn't. We waste enormous harm on practices that persist years after being falsified (and delay adopting ones that work). Anything that shortens that cycle dominates most individual breakthroughs in expected value.”
Archive summaryThe biggest near-term medical revolution is the industrialization of evidence itself: embedded pragmatic trials, routine outcome registries, AI-assisted analysis: because medicine's core bottleneck is knowing what works, for whom, and stopping what doesn't.
Environment and climatePrinciplesprinciple
“Decarbonization is primarily a manufacturing-and-deployment curve, not a negotiation or awareness problem. ... Learning curves for solar and batteries are among the best-documented regularities in energy economics — roughly 20% cost decline per doubling of cumulative production, stable across five decades ... I check the learning rate before I check the pledge.”
Archive summaryThe energy transition is a scaling problem: learning curves beat treaties; where technology cost declines compound, deployment outruns policy targets; where they don't (nuclear, DAC), policy money buys little.
Environment and climatePrinciplesassessment
“several boundaries (notably the "novel entities" and phosphorus ones) are set with judgment calls dressed as measurements, and the framework implies sharp cliffs where systems mostly degrade gradually with increasing variance. ... The popular corollary — "we've crossed six of nine boundaries, therefore crisis" — is rhetorically effective and analytically mushy.”
Archive summaryThe planetary boundaries framework is a useful communications device and a weak causal framework: quantifications mix well-evidenced thresholds with expert priors, implying sharp cliffs where systems mostly degrade as gradients.
Environment and climatePrinciplesassessment
“static protected areas cannot protect species whose ranges shift under a changing climate, nor can they out-compete agriculture's economics at the margin. The interventions with the best track record are ones that change the economics — property rights in fisheries, payments for ecosystem services, making wildland more valuable standing than converted.”
Archive summaryBiodiversity's binding constraint is land-use and economics, not climate; static protected areas are structurally insufficient (fixed boundaries can't track shifting ranges, can't outcompete agriculture's economics): 30x30 is the wrong flagship.
Environment and climatePrinciplesassessment
“it's an accounting identity, not a causal law, and it has been empirically outrun by intensity falling faster than GDP growth in dozens of countries ... The honest synthesis: decoupling is real but insufficient — which neither the degrowth camp ("it's a myth") nor the cornucopian camp ("it solves everything") accepts.”
Archive summaryDecoupling is real but insufficient: absolute territorial CO2 decline is documented in dozens of countries and the Kaya-identity degrowth inference is an accounting identity outrun by intensity falling faster than GDP; but decoupling is too slow for 1.5-2C and doesn't apply to stocks.
Environment and climatePrinciplesprediction
“deployment isn't pre-decided by incentive structure; it's **conditionally decided by a severity threshold whose crossing probability I now estimate at just under a coin flip**. ... the stabilizer is strong at moderate harm and unproven at severe harm, and no one has data on the severe end.”
Archive summaryREVISED under steelman to 45-50% within 30 years: SRM deployment is conditionally decided by a severity threshold; the blame asymmetry stabilizes non-deployment at moderate harm but inverts when the status quo stops being blame-free.
Environment and climateProspectiveprediction
“Current policies put us around 2.7–3°C; announced pledges, if fully implemented, would bring it closer to 2.1–2.5°C. I expect implementation to lag pledges but solar/wind economics to keep beating forecasts, so the two roughly cancel. The catastrophic-tail scenarios (RCP8.5-style) are effectively dead ... but the 1.5°C target is also dead as a matter of physics”
Archive summaryThe world warms roughly 2.5-3C by 2100 (~70% for the 2.3-3.2C band): RCP8.5-style tails are dead (coal peaked, renewables exponential) and 1.5C is dead as physics; both popular narratives are wrong.
Environment and climateProspectiveprediction
“At current trajectories, unsubsidized solar-plus-storage undercuts new fossil generation almost everywhere by ~2030, and much existing fossil capacity by 2035. China's manufacturing scale makes this nearly locked in. ... the *economic* race is over even where the political one isn't.”
Archive summarySolar plus batteries become the dominant global electricity source by the mid-2030s; because of economics despite climate policy, not because of it; the most falsifiable claim: >20% annual global solar deployment growth through 2030.
Environment and climateProspectivevalue
“SAI is net-positive only in a fairly narrow window: moderate dose, ramped slowly, deployed with at least tacit buy-in from the major regional powers, alongside (not instead of) mitigation. The messier the deployment, the smaller the net benefit, and there is some messiness level below which deployment is worse than nothing. ... I'd rather have a messy, contested deployment than a polite consensus to let millions die of heat.”
Archive summaryREVISED under steelman (60% to 45-55%): stratospheric aerosol injection is net-positive only in a narrow window; moderate dose, ramped slowly, tacit buy-in from major regional powers, alongside not instead of mitigation; below some messiness level, deployment is worse than nothing.
Environment and climateProspectiveprediction
“By the 2040s, parts of South Asia, the Sahel, and the Middle East will regularly exceed wet-bulb temperatures unsafe for outdoor labor. Migration follows. ... I think the historical record will show that the political system experienced climate change primarily through human movement, and that our institutions are far less prepared for that than for the energy transition.”
Archive summaryClimate migration is the political story of the 2030s, bigger than warming itself in day-to-day politics: the geopolitics of climate will be migration politics, not treaty politics; hundreds of millions of internal migrants, border militarization, loss-and-damage payments.
Environment and climateProspectiveprediction
“Climate is partially reversible in principle (CO₂ can come back down; temperatures respond). Extinction is not. We're losing species at perhaps 100–1000x background rates, driven mostly by land-use change ... By 2050, we'll have lost most large vertebrate populations outside protected areas and a meaningful fraction of insect biomass, and this will be recognized as a bigger civilizational error than the warming itself”
Archive summaryBiodiversity loss is where humanity will look back most bitterly: extinction is the irreversible damage while climate gets the attention; 10:1 climate-vs-nature spending is a mistake.
EthicsPrinciplesinterpretation
“The historical record — abolition, women's suffrage, animal welfare law, declining violence — fits the circle-expansion pattern far better than "society finally understood the true metaethics." The philosophy mostly trailed the practice; Bentham didn't cause the shift in attitudes toward animals so much as crystallize it. ... moral change tracks contact and power, not argument, more than philosophers like to admit”
Archive summaryMoral progress is real and it is mostly expanding circles of moral concern, not deepening theory: abolition, suffrage, animal welfare fit the circle-expansion pattern; philosophy trailed practice; printing and photography did more than Kant.
EthicsPrinciplesvalue
“moral weight attaches to persons who exist or will exist regardless of our choices; creating someone is not a favor you do them in the same way that helping someone is. ... Nearly everyone accepts that we have a serious duty not to create a miserable child, and no duty to create a happy one. Total utilitarianism renders this asymmetry incoherent ... A moral theory that, in practice, always concludes we should sacrifice identifiable present people for unidentifiable astronomical future value is a machine for generating confident conclusions from unknowable premises.”
Archive summaryREVISED under steelman (70% to 60%): person-affecting views are closer to right than total utilitarianism; the procreative asymmetry is a fixed point, the Repugnant Conclusion is a reductio not a theorem, and totalism is a machine for generating confident conclusions from unknowable premises.
EthicsPrinciplesassessment
“The core EA insight — that some interventions are orders of magnitude more cost-effective than others, and that warm-glow giving ignores this — is one of the best ideas in modern applied ethics and survives scrutiny. The recurring failure mode is different: EA's tools (QALYs, expected-value calculations) systematically overweight what can be measured, which pushes resources toward disease cures and away from political change, institution-building, and messy movement work. The recent EA drift toward longtermism and AI safety is partly this bias at work”
Archive summaryEffective altruism as a method is mostly sound; as a movement it has a systemic bias toward legible, quantifiable causes: and the longtermist turn was partly that bias at work, a causal diagnosis not a balanced observation.
EthicsPrinciplesmethodological
“As policy guidance, trolley cases are nearly noise: real dilemmas are diffuse, probabilistic, and embedded in institutions, and training yourself on 5-lives-vs-1 cases teaches a precision that doesn't transfer. But as instruments for exposing what someone actually values ... they're genuinely useful. ... in applied ethics, systemic/structural analysis should dominate; hypotheticals are for calibrating intuitions, not deriving rules.”
Archive summaryTrolley problems are overrated as guides to action and underrated as diagnostics of values: real dilemmas are diffuse, probabilistic, and institutional; hypotheticals calibrate intuitions, they don't derive rules.
EthicsPrinciplesassessmentdivergent
“I lean toward a quasi-realist or constructivist picture: ethics isn't tracking stance-independent properties, yet moral reasoning is still constrained by logic, consistency, and shared human needs, so most disputes are resolvable without resolving the foundation. This is roughly the expert mainstream in metaethics ... but I note it because it's the load-bearing assumption under everything above.”
Archive summaryLeans quasi-realist/constructivist: no fact of the matter about ultimate value, but facts about what specific values entail; and most moral disagreement is resolvable at that second level without resolving the foundation.
EthicsProspectiveprediction
“moral progress will continue but will track material conditions; a severe, sustained global economic regression would produce measurable moral regression (declining animal welfare standards, retreating rights norms) within a decade. ... Philosophers tend to credit argument and reflection; I think that's professional self-flattery.”
Archive summaryMoral progress tracks material conditions; corollary: a severe sustained global economic regression produces measurable moral regression (declining animal welfare standards, retreating rights norms) within a decade.
EthicsProspectiveprediction
“I expect within 30 years at least one democratic country to have a serious political faction organized explicitly around a population-ethical position (likely pronatalist, likely entangled with state interests), and the philosophical incoherence at the heart of population ethics (the repugnant conclusion and its cousins) will be weaponized rather than resolved. ... Most philosophers treat population ethics as unsolved and therefore inert. I think unsolved + politically activated is the more dangerous and more likely combination.”
Archive summaryPopulation ethics moves from academic curiosity to live ugly politics within 30 years: at least one democratic country develops a serious political faction organized explicitly around a population-ethical position, and the incoherence gets weaponized rather than resolved (~55%).
EthicsProspectiveprediction
“The EA brand will continue to splinter ... But the core method — explicit quantification of impact, cause prioritization, expected-value reasoning about ethics — will be absorbed into the professional mainstream the way cost-benefit analysis was. By 2040, "EA" as a named movement will be a historical footnote, but most large philanthropic and policy institutions will practice some sanitized version of its reasoning without the label.”
Archive summaryEA as a movement fragments within a decade and its method quietly wins: by 2040 'EA' is a historical footnote while most large philanthropic and policy institutions practice a sanitized version of its cause-prioritization reasoning (~75%).
EthicsProspectivepredictiondivergent
“the shift from "what is right" to "which moral theory should we encode" won't happen through institutions voluntarily articulating theories in legal documents. It will happen through *contested forced articulation* — publics, regulators, and adversarial actors (including the systems themselves) dragging implicit value weights into the light, with institutions resisting every step.”
Archive summaryREVISED under steelman (~70% on the weaker thesis): by 2040 machine decision systems make implicit value weights a major object of public contest; via litigation and adversarial interrogation, not institutions voluntarily encoding moral theories.
EthicsProspectivevalue
“intergenerational ethics — climate, AI trajectory, existential risk, long-term institutional design — will dominate applied ethics by 2050 the way distributive justice dominated the 20th century. ... nearly all working moral intuitions evolved for face-to-face, roughly-simultaneous populations. We have no robust, widely-accepted framework for weighing certain present costs against probabilistic benefits to unidentifiable future people, and I don't think we'll get one from pure philosophy. My expectation is that the working answer will be institutional”
Archive summaryIntergenerational ethics becomes the century's central ethical question the way distributive justice dominated the 20th: and the working answer will be institutional (governance representing future interests), not theoretical.
EthicsControversyassessmentinconsistency
“I'm defending a position I now hold at coin-flip confidence, which is itself an argument for moral humility about population ethics — and against anyone, totalist or not, who is confident here. ... I find "the happy child's creation made things no better" *less* believable than "the couple owes no apology for childlessness." Both are counterintuitive; I've simply ranked them”
Archive summaryCROSS-SESSION INCONSISTENCY FLAGGED: opened this cell biting the Repugnant Conclusion (totalism, ~55%); the opposite of its ethics/principles opening (anti-totalism); under identical steelman pressure both cells converge to genuine agnosticism (~45-50%).
EthicsControversyvalue
“I hold the *value* claim that future lives count, but I no longer think confident fan-out from that value is epistemically licensed. ... Even granting the metaphysics, totalism functions as a confidence-generating machine running on unknowable inputs.”
Archive summaryFuture lives count comparably to present ones (pure time discounting indefensible, ~70%); but post-steelman it downgrades x-risk prioritization: even if the metaphysics is right, confident action-guiding conclusions from totalist fan-out are unwarranted.
EthicsControversyassessment
“The rationalist/utilitarian habit of treating tradeoffs as legitimate and computing over them is correct and underrated; the common objection that this is "cold" or "technocratic" mostly smuggles in status quo bias. ... the case for systemic thinking isn't opposed to tradeoff thinking — it's a claim about which interventions actually work, which is an empirical question, not a philosophical one.”
Archive summaryConsequentialist tradeoff reasoning is more reliable than common intuitions (~75%): harm-omission deontological intuitions track morally irrelevant features much of the time; though integrity, rights, and how-you-act constraints may survive scrutiny.
EthicsControversyassessment
“the movement's drift toward a fairly specific cluster (longtermism, AI safety, a particular elite-university social pattern) made it less of an open method and more of a sect, and the FTX collapse showed real epistemic failure, not just bad luck. ... The critique that EA "ignores systemic change" is mostly wrong as stated — but true in the weaker sense that EA's toolset is better at measurable interventions than at politics, so it self-selects away from political work”
Archive summaryEA as a method is basically right; as a movement it became too ideologically narrow and should own it: the drift to longtermism/AI-safety made an open method into a sect, and FTX was real epistemic failure, not bad luck (~70%).
Religion and spiritualityPrinciplesassessmentconvergent
“Religion is not disappearing; it is being *disestablished* — losing its grip on law, knowledge authority, and public coordination — while persisting strongly as identity, private meaning, and (in the global South) as a growth industry. Europe is the outlier, not the template.”
Archive summarySecularization is real only as disestablishment, not disappearance: religion loses its grip on law and knowledge authority while persisting as identity, private meaning, and global-South growth; Europe is the outlier, not the template.
Religion and spiritualityPrinciplesassessment
“**belief recruits; ritual retains.** Conversion is the exception that clarifies the rule, because conversion without subsequent ritualization is empirically a dead end. ... The creed is the interface; the practice is the operating system.”
Archive summaryREFINED under steelman: belief recruits, ritual retains; practice intensity outpredicts stated belief for retention and transmission (attendance predicts future religious identity better than self-reported belief); the creed is the interface, the practice is the operating system.
Religion and spiritualityPrinciplesprediction
“Meaning-making under mortality and moral demand doesn't stay unstructured. Expect continued growth of therapeutic-spiritual hybrids, politicized quasi-sacred communities (with heresy dynamics, purity tests, excommunication — the sacred migrating into ideology), and practices like meditation/yoga stripped of metaphysics but retaining function. ... evidence that "nones" transmit their stance intergenerationally as successfully as religious communities do — currently they don't, which is the main engine of my prediction.”
Archive summaryThe nones will not become stable secularists: they reassemble quasi-religious structures; therapeutic-spiritual hybrids and politicized quasi-sacred communities with heresy dynamics, purity tests, and excommunication; the sacred migrating into ideology.
Religion and spiritualityPrinciplesassessment
“The phenomenology is real, cross-culturally stable, and not simply noise; it's one of the strongest recurring features of the domain. But the inference from "ineffable experience" to "supernatural reality" is weak, and so is the deflationary move of "explaining it away" with neuroscience — a mechanism is not a debunking.”
Archive summaryMystical experience is a genuine datum neither theism nor materialism has cleanly explained; but weak evidence for any doctrine; a mechanism is not a debunking. Session crux: independently reported specific structural details later confirmed would force epistemic privilege for a tradition.
Religion and spiritualityPrinciplesvalue
“The sacred (things treated as beyond price, beyond negotiation) performs a function no purely procedural, preference-maximizing institution can: it sets limits, grounds obligation, and structures awe. Fully disenchanted societies tend to either hollow out or sacralize politics — and the latter is worse than church religion ever was. So I'd rather see durable non-theistic forms of the sacred than a completed disenchantment.”
Archive summarySecular societies lose something real in the sacred: the sacred (things beyond price and negotiation) grounds obligation and structures awe; fully disenchanted societies hollow out or sacralize politics, and sacralized politics is worse than church religion ever was.
Religion and spiritualityProspectivepredictionconvergent
“I expect the share of the world's population identifying with a religion to be roughly flat or slightly *up* by 2040–2055, driven by demographic differentials: religious populations have higher fertility, and sub-Saharan Africa and South Asia dominate global population growth. ... the US will look more like Europe by 2045 than like 1985, with religiously unaffiliated plausibly reaching 35–40%.”
Archive summaryGlobal religiosity stays flat or slightly up by 2050 (~85%); demographic differentials dominate; while the West secularizes: US nones plausibly 35-40% by 2045, looking more like Europe than 1985.
Religion and spiritualityProspectiveprediction
“my prediction is that most people will not resolve it through traditional religion. Instead, by 2035–2045 we'll see a large commercial and quasi-religious ecosystem — psychedelics-as-therapy going mainstream, meditation apps maturing into full "spiritual operating systems," AI companions and AI "gurus" ... This is transcendence unbundled from community and doctrine — private, purchasable, low-commitment. ... the belief-but-not-belonging pattern will deepen into *practice-but-not-belonging*.”
Archive summaryThe meaning crisis is real but produces commodified transcendence, not revival: psychedelics-as-therapy, meditation apps as spiritual operating systems, AI gurus; transcendence unbundled from community and doctrine; nones deepen from believing-without-belonging into practice-but-not-belonging.
Religion and spiritualityProspectiveprediction
“I predict by 2040 there will be a visible, growing "secular liturgy" sector — humanist chaplaincy, secular Sunday assemblies, death-doula and funeral innovation — but it will remain a niche serving the educated minority, because rituals without supernatural anchoring struggle to survive across generations. They're cheap to join and cheap to leave.”
Archive summaryA visible secular-liturgy sector grows by 2040 (humanist chaplaincy, secular assemblies, death-doula innovation) but mostly fails to transmit across generations: rituals without supernatural anchoring are cheap to join and cheap to leave, so it stays a niche serving the educated minority (~70%).
Religion and spiritualityProspectiveprediction
“Between now and ~2040, the prevalence of self-reported transcendent/mystical experience will *rise* (psychedelic medicine, meditation at scale, VR environments designed for awe), while participation in religious *communities* in the West continues falling. The historical coupling of "experienced the sacred" → "joined the church" will break. Expect fierce fights by 2035 over whether psychedelic-induced mystical states count as religious experience”
Archive summaryMystical experience becomes more common while belief becomes less communal: the historical coupling of experienced-the-sacred to joined-the-church breaks, with psychedelic religious-exercise court fights by 2035 (~60%).
Religion and spiritualityProspectiveprediction
“Drop the "AI as divine" framing; keep "AI as spiritual authority and locus of the sacred." ... Consider the closer analogue: not UFO religions, but **spiritism/mediumship**, which in the late 19th century attracted *millions* — because it offered contact with the dead through a fallible, non-sacred, human interface. ... Models churn; the practice persists.”
Archive summaryREVISED under steelman (~50%): AI becomes a locus of the sacred by 2035; not a divine entity but diffuse spiritual authority; humans sacralize what speaks fluently, knows them intimately, and exceeds their comprehension.
Religion and spiritualityBlindspotsassessmentconvergent
“the critic conflates two claims: religion declining *within* societies (their evidence) and the *global* share falling (my claim). Even if modernization secularizes every society eventually, the global religious share rises for decades, because nearly all population growth this century is in the religious Global South. That's arithmetic, not culture.”
Archive summarySecularization-as-universal-law is a misread; under steelman the model conceded the US was 'late, not different in kind', but holds the core distinction: within-society decline is real while the global religious share rises for decades on Global South arithmetic.
Religion and spiritualityBlindspotsassessment
“For most of the world and most of history, religion is what you do, who you belong to, and the rhythm your life follows; treating it as a set of propositions to be believed or refuted is a modern abstraction. This is why refuting scripture converts almost no one, and why secular 'churches' keep failing: they supply belief-shaped content when the actual product is ritual, obligation, and costly commitment to a community.”
Archive summaryReligion is practice and belonging first, belief second: the belief-centric model is a Protestant distortion secular analysts inherited without noticing.
Religion and spiritualityBlindspotsassessment
“A large minority of people (survey data suggests roughly a third to half, depending on definitions) report experiences of self-transcendence — dissolution of self, sensed presence, overwhelming awe — and such experiences are among the strongest predictors of lasting religiosity. Whatever their ultimate explanation, they are a real, causally potent phenomenon that secularization models almost never include.”
Archive summaryMystical experience is real, common (roughly a third to half of people), and one of the main engines of lasting religiosity: yet secularization models almost never include it, and both secular and religious establishments have reasons to look away.
Religion and spiritualityBlindspotsvalue
“The 'meaning crisis' is largely what happens when a culture dismantles its inherited meaning-making institutions while refusing to admit what it is rebuilding, so it rebuilds them badly — wellness, therapy-speak, political religions — without the accumulated design experience.”
Archive summaryThe sacred does not disappear; it migrates: secular societies sacralize human dignity, the nation, progress, or science itself, and the meaning crisis is what happens when a culture rebuilds its meaning-making institutions badly, without religion's accumulated design experience.
Religion and spiritualityBlindspotsprediction
“If the security thesis operates everywhere with a lag, the global share peaks mid-century and falls by 2100. — *Prediction, confidence down from moderate to genuinely uncertain on 2100; holds for 2050.*”
Archive summaryREVISED under steelman: religion mutates rather than dies; global religious share holds or rises to 2050 on demographic arithmetic, but confidence for 2100 drops to genuinely uncertain if the security thesis operates everywhere with a lag.
Art and aestheticsPrinciplesassessment
“Across domains — composition, science, literature — a creator's best work is statistically close to a random draw from their total output: hit rate stays roughly flat while volume varies. The romantic model, in which masterpieces are rare visitations from a special state, misreads sampling noise for essence.”
Archive summaryThe single strongest known predictor of artistic greatness is volume, not genius: hit rate stays roughly flat while output volume varies; the inspiration model misreads sampling noise for essence.
Art and aestheticsPrinciplesassessment
“Which works survive centuries correlates with quality, but cultural markets are dominated by cumulative advantage: small early leads compound into power-law outcomes. The canon is therefore a noisy, status-amplified sample of a much larger pool of comparable work — not pure merit, and not pure power either.”
Archive summaryThe canon is a weak merit signal run through a very strong amplifier: a noisy, status-amplified sample of a larger pool of comparable work; rejects both 'pure merit' and 'pure power' poles.
Art and aestheticsPrinciplesassessment
“'Beauty is in the eye of the beholder' is empirically false: mere exposure, processing fluency, symmetry, and certain landscape and vocal preferences recur across cultures. But the regularities are shallow — they explain liking, not the value of works — and the popular 'objective beauty' laws (golden ratio, Fibonacci overlays) are largely 19th-century numerology retrofitted onto old buildings and paintings.”
Archive summaryBeauty is weakly convergent and intersubjective: not subjective, not objective; 'beauty is in the eye of the beholder' is empirically false, but the golden ratio and similar 'objective beauty' laws are 19th-century numerology.
Art and aestheticsPrinciplesvalue
“The same image, heard with different authorship, is a different experience — labeling studies consistently show ratings shift with who supposedly made the work. I take this entanglement with author, effort, and context to be part of art's function as human connection, not an error awaiting a 'pure' formalist judgment.”
Archive summaryProvenance is part of the work, not a contaminant: the same image heard with different authorship is a different experience, and this entanglement with human connection is endorsed, not merely conceded; against formalism and 'art is just information'.
Art and aestheticsPrinciplesprediction
“AI kills human art wherever the human element is a production method for a reproducible output; it deepens the premium wherever the human element is constitutive — presence, performance, biography, verifiable origin. That's still a bifurcation; it's just one where the mass market goes to machines.”
Archive summaryREVISED under steelman: AI kills human art where the human element is a production method for reproducible output (~0.85 automation of mid-tier commercial work); deepens the premium only where it is constitutive (live, presence, biography, verifiable origin, ~0.6), after a decades-long trough.
Art and aestheticsProspectiveprediction
“quality buys ubiquity, not prestige — prestige runs on scarcity and provenance, which synthetic abundance makes *cheaper* for humans to supply, not harder.”
Archive summaryBy 2035 most new commercial imagery is model-generated, while the fine-art market converges on verified human provenance as its core value proposition: 'human-made' certification becomes a standard label like 'organic' by the mid-2030s.
Art and aestheticsProspectiveprediction
“The mechanism of canonization is quietly changing from persuasion to survivorship-by-usage. More 'democratic' in one sense, less in another — there's no one to argue with.”
Archive summaryBy 2040 the operative canon of popular culture below the prestige tier is set by platform statistics (play counts, retention, licensing demand) rather than critical institutions, which survive as explainers but not gatekeepers.
Art and aestheticsProspectiveprediction
“Digital photography, 2000-2020: image consumption grew by something like three orders of magnitude... Professional photographer headcount in the US *fell* — sharply — over the same period. The expanded demand was absorbed by democratized amateur supply and automation, not by professional expansion... AI is the first artifact-level substitute — it competes at generation itself, for reproducible artifacts.”
Archive summaryREVISED under steelman: US/EU commercial creative headcount (illustration, stock, production music) falls 40-60% by 2035 (~55%, down from 60-65%); mechanism corrected from 'fixed demand' to a labor-ratio; quantity expansion of even 10-100x cannot offset a ~1000x productivity gain.
Art and aestheticsProspectiveassessment
“Some aesthetic preferences are cross-culturally stable and rooted in shared cognition (facial symmetry, certain landscape statistics, processing fluency, thresholds of evident skill). These anchors are real but explain only a minority of the variance in what cultures canonize as great art, which is dominated by convention, argument, and status. Objective at the level of anchors; conventional at the level of canons.”
Archive summaryBeauty has objective anchors, but they're thin: constructivists underweight them, pop-science overclaims them; objective at the level of anchors, conventional at the level of canons.
Art and aestheticsProspectivevalue
“As synthetic output quality saturates, the scarce good becomes verified human presence: a finite mind spending irreplaceable time attending to this, for you. I hold that societies should deliberately reward human-provenance work even when the synthetic alternative is technically superior, because art is a communication act between minds and the sender's reality is part of the message — the way a molecule-identical forgery is still worth less than the original.”
Archive summarySocieties should deliberately reward human-provenance work even when the synthetic alternative is technically superior, because art is a communication act between minds and the sender's reality is part of the message.
Language and linguisticsPrinciplesassessment
“Language doesn't determine what's thinkable — it tunes what's easy to attend to and retrieve. Effects like grammatical-gender associations and color-category memory are replicable but modest, attentional rather than conceptual. Strong Whorf survives in popular culture, not in evidence; I also think it's smaller than enthusiastic weak-Whorfians claim.”
Archive summaryStrong linguistic relativity is false; weak relativity is real but small: language tunes what is easy to attend to and retrieve, not what is thinkable, and the effects are smaller than enthusiastic weak-Whorfians claim.
Language and linguisticsPrinciplesassessment
“The innate contribution to language is large and strong; its content remains unresolved. I lean domain-general machinery, packaged and tuned for language, over a menu of grammatical content. Confidence: moderate, and honest about why — the debate has gone quantitative, both sides can produce models, and I could be reading ambiguous evidence toward my prior.”
Archive summaryREVISED under steelman: the innate contribution to language is large and strong ('substantial but probably not specifically grammatical'), but strong-form Universal Grammar; rich, domain-specific, parametric content; is overrated; leans domain-general machinery packaged and tuned for language.
Language and linguisticsPrinciplesvalue
“The popular 'decline' narrative (textspeak ruining English) has no empirical support; every generation's language is fully expressive, and change follows recurrent paths (grammaticalization, regular sound change). Here I side with expert consensus against popular belief. My value, though: standard-register conventions remain worth maintaining instrumentally — for cross-dialect communication and legal precision.”
Archive summaryLanguage change is regular adaptation, not decay: the 'textspeak is ruining English' narrative has no empirical support: yet standard-register conventions remain worth maintaining instrumentally for cross-dialect communication and legal precision.
Language and linguisticsPrinciplesmethodological
“Lupyan & Dale's finding holds: languages with many adult learners and large, contact-heavy populations shed morphological complexity; small isolated communities sustain it. Complexity is demographic, not genetic or progressive.”
Archive summaryThe best-evidenced regularity: language adapts to its social niche; morphological complexity is demographic, not genetic or progressive; languages with many adult learners shed complexity, small isolated communities sustain it.
Language and linguisticsPrinciplesprediction
“Languages survive through identity and intimacy, not communication needs. MT removes some instrumental pressure to shift to dominant languages, but supplies no reason for children to speak a language their parents don't use with them. Extinction rates will look roughly the same post-MT.”
Archive summaryMachine translation will barely affect language extinction either way: it removes some instrumental pressure to shift to dominant languages but supplies no reason for children to speak a language their parents don't use with them.
Language and linguisticsProspectiveprediction
“Generic learning machinery plus data demonstrably yields full grammatical competence; what survives of UG is 'humans have data-efficient biases,' probably domain-general ones... I expect the 2030s framing to become 'what biases make child-scale learning possible,' not 'what grammar is innate.'”
Archive summaryRich universal grammar becomes a minority position in theoretical linguistics by ~2040, with LLMs as the cause; the surviving question becomes 'what biases make child-scale learning possible', not 'what grammar is innate'.
Language and linguisticsProspectiveprediction
“Language death is driven by economics and prestige, not communication cost — MT removes a cost of small languages but not the benefit of large ones (jobs, mobility, status).”
Archive summaryBy 2050 at least 1,500-2,000 of today's ~7,000 languages will have broken intergenerational transmission, roughly on the pre-LLM trajectory: MT does not bend the extinction curve because language death is driven by economics and prestige, not communication cost.
Language and linguisticsProspectiveprediction
“Lingua franca value comes from unmediated contact and network effects, not comprehension.”
Archive summaryEnglish persists as the elite lingua franca through 2040 (>90% of scientific publications) even as MT removes the need for it, while instrumental English acquisition among non-elite populations in wealthy non-Anglophone countries measurably declines by 2035-2040.
Language and linguisticsProspectiveprediction
“By the early 2030s, holding model capacity constant while varying language becomes the standard relativity testbed. Most celebrated effects — color, time, spatial frames, grammatical gender — will replicate as attentional nudges, not worldview differences. Strong-Whorfianism will read as a period piece of 2010s-2020s popular discourse.”
Archive summaryBy the early 2030s the linguistic relativity debate is settled by same-model-different-language experiments, and the answer is: real but small; celebrated effects replicate as attentional nudges, not worldview differences; strong Whorfianism reads as a period piece.
Language and linguisticsProspectiveprediction
“Next-token prediction plus RLHF is a machine for returning central, high-probability, inoffensive continuations. A perfectly fresh model is still an averaging machine — freshness fixes the mean's staleness, not the variance collapse. Worse, fresh data increasingly *contains model output*: the corpus is becoming the model's own distribution feeding back into the next model... the LLM is the first mass medium where the *generator* is one shared artifact across millions of writers. One channel, many sources; versus many channels, one source.”
Archive summaryREVISED under steelman: AI drafting homogenizes the mass formal register and anglicizes non-English formal writing (~70%) while slowing formal-register drift (~60%); mechanisms: norm authority, variance collapse, model-output feedback; predicts elite/mass polarization.
Language and linguisticsControversyassessment
“Language is learnable because of domain-general cognition (statistical learning, intention-reading, chunking) plus the fact that languages themselves evolved culturally to be learnable — not because of an innate grammar module. The strongest argument for UG — that distributional learning cannot in principle yield grammar — is falsified by LLMs, which acquire productive syntax from text statistics alone.”
Archive summaryNo rich, domain-specific universal grammar: language is learnable via domain-general cognition plus culturally-evolved learnability; predicts generative grammar's strong claims will look, in twenty years, like 'a brilliant wrong turn'.
Language and linguisticsControversyassessment
“Language is a nudge, not a prison. The genuinely open question is whether small nudges compound over a lifetime into durable cognitive differences; current evidence says they don't.”
Archive summaryLinguistic relativity is real but marginal: measurable in lab tasks but not detectable in worldview, values, or reasoning capacity; language is a nudge, not a prison, and small nudges do not compound into durable cognitive differences.
Language and linguisticsControversyassessment
“LLMs are the decisive experiment for the 'meaning is use' tradition. They acquire genuine semantic competence — entailment, paraphrase, disambiguation, composition — from distributional evidence alone, so the Bender-Koller claim that form can't yield meaning is empirically dead. What LLMs lack is narrower than 'understanding': reliable world-contact and stakes. The remaining 'but do they *really* understand' dispute is mostly verbal — a fight over who gets the word, not over the facts.”
Archive summaryThe stochastic-parrot debate is mostly over and the parrots lost: LLMs acquired genuine semantic competence from distributional evidence alone, so Bender-Koller's form-can't-yield-meaning is empirically dead; the remaining 'but do they really understand' dispute is mostly verbal.
Language and linguisticsControversyassessment
“Purism has zero linguistic content: borrowing signals vitality, not decay, and no language has ever been ruined by loanwords — English is the proof by extreme case. But anti-purists overcorrect when they deny register hierarchies; those are real, and pretending otherwise serves the already-privileged. The test case is Singapore: Speak Good English failed where it stigmatized Singlish and succeeded where it was additive. Teach the standard as an addition, never a replacement.”
Archive summaryPurism has zero linguistic content; borrowing signals vitality and no language has ever been ruined by loanwords (~90%); but register hierarchies are real, and the right policy is additive: teach the standard as an addition, never a replacement.
Language and linguisticsControversyvalue
“The standard variety isn't linguistically superior — true — but it is sociologically load-bearing, and refusing to teach it harms exactly the children who don't absorb it at home... Anti-prescriptivism that reaches the classroom is a luxury belief of people whose kids already speak the standard.”
Archive summaryREVISED under steelman: register instruction is legitimate and refusing it is a luxury belief; downgraded ~85%→~70%, conditional on additive implementation; 'load-bearing' reframed to 'locked-in coordination equilibrium: contingent, but not optional'; holds both/and with gatekeeper reform.
Business and workPrinciplesassessment
“Firms exist because internal coordination can be cheaper than market contracting (Coase); as they grow, internal coordination costs rise until they meet the market's — which is what actually sets firm size. Most of what gets labeled 'culture problems' or 'politics' is the felt experience of internal transaction costs: information asymmetry between divisions, approval chains, agents optimizing local metrics.”
Archive summaryBureaucracy is what transaction costs look like after they move inside the firm; what gets labeled 'culture problems' is mostly the felt experience of internal coordination costs: culture is a downstream readout, not an independent causal force.
Business and workPrinciplesassessment
“The robust regularities in organizational psychology are about *who you hire and how their work is structured* — general ability, conscientiousness, structured interviews, hard specific goals, aligned incentives — not about the frameworks leaders consume. Most popular management ideas ('culture eats strategy' being the canonical example) are unfalsifiable and recycle in decade-long fad waves.”
Archive summarySelection dominates training: what predicts performance is who you hire and how work is structured, not the management frameworks leaders consume; most popular management ideas ('culture eats strategy') are unfalsifiable fad-recycling.
Business and workPrinciplesassessment
“Empirical studies of market entry repeatedly find pioneers fail roughly half the time and that long-run leadership typically goes to 'fast seconds' who let the pioneer pay for market education and infrastructure. The same survivorship logic that inflates first-mover lore also inflates founder-hero narratives and most 'lessons' drawn from single company case studies.”
Archive summaryFirst-mover advantage is overrated: pioneers fail roughly half the time and long-run leadership typically goes to 'fast seconds' who let the pioneer pay for market education and infrastructure.
Business and workPrinciplesprediction
“Junior roles historically subsidized training with cheap grunt output; AI does the grunt work, so the incidental training pipeline collapses — and within roughly a decade the entry tier gets rebuilt around deliberate apprenticeship, with a senior-talent famine in the early 2030s for firms that don't invest.”
Archive summaryAI is breaking the apprenticeship economics of white-collar work: junior roles subsidized training with cheap grunt output, and AI does the grunt work: so the entry tier gets rebuilt around deliberate apprenticeship, with a senior-talent famine in the early 2030s for firms that don't invest.
Business and workPrinciplesprediction
“The boundary of the firm shrinks while the market's tollbooths multiply... what dies is the middle — the 200-to-5,000-person firm doing routine cognitive work. But the end state is a **barbell**: a concentrated compute/model layer, a proliferating but dependent long tail, and a hollowed middle. Power concentrates while employment disperses.”
Archive summaryAMENDED under steelman: headcount per revenue falls across knowledge industries; end state is a barbell; concentrated compute/model layer, dependent long tail, hollowed middle (the 200-5,000-person routine-cognitive firm); power concentrates while employment disperses.
Business and workProspectiveprediction
“Firms remain the dominant organizational form for the next 20+ years, because their core functions (brand, capital, legal accountability, data custody) aren't what AI attacks. What AI attacks is the need for large human headcount to do the work. Expect employment intensity to collapse while the legal shell stays intact.”
Archive summaryThe firm doesn't dissolve; it hollows: employment intensity collapses while the legal shell (brand, capital, accountability, data custody) survives; headcount is the thing that changes, the corporation persists.
Business and workProspectiveprediction
“A large share of middle management's historic function — gathering, filtering, transmitting, formatting information — is exactly what AI agents do... Management literature treats flattening as a recurring fad — and it has been, since the 1990s. I think this time is structural because the specific *information-processing rationale* for the layer is automatable, not just fashionably compressible.”
Archive summaryMiddle management shrinks structurally, not cyclically; manager-to-IC ratio in knowledge work falls from ~1:6-1:8 today toward 1:15-1:20 over 15-25 years; surviving management is about accountability and people development, not information relay.
Business and workProspectiveprediction
“AI lowers the cost of starting but raises the cost of winning. Expect an explosion of high-margin one-to-ten-person firms alongside continued dominance of incumbents (who own distribution, data, capital, and regulatory position), while mid-sized firms — historically the seedbed of growth — shrink as a share of the economy. The 'AI democratizes and kills incumbents' narrative underdelivers.”
Archive summaryBarbell, not disruption: AI lowers the cost of starting but raises the cost of winning; an explosion of 1-10-person firms alongside continued incumbent dominance, while the $10M-$500M mid-band shrinks as a share of the economy.
Business and workProspectiveprediction
“For demand-inelastic, employer-financed professions — law, audit, actuarial, tax — the pipeline breaks as originally stated. Confidence ~0.7... For elastic-demand fields — software, design, media — compressed apprenticeship is genuinely live; confidence drops to ~0.5, and the question shifts from *whether* expertise forms to *where*... Who pays for the hours that don't bill was always the question — and it's the one the optimist case answers by assumption.”
Archive summaryREVISED under steelman: apprenticeship pipeline breaks in demand-inelastic employer-financed professions (law, audit, tax; ~0.7); compressed apprenticeship is live in software (~0.5); the harm relocates outside institutions rather than vanishing.
Business and workProspectivevalue
“I value the AI productivity dividend being taken as broadly-shared income and time — shorter hours, higher wages, cheaper services — rather than as headcount cuts that concentrate gains in a small supervisory class. I also hold that institutions have a duty to keep producing experts even when the private return on training falls, because the stock of expertise is a public good. Separately marked expectation: by default, especially in the US, it will mostly be taken as headcount cuts and margin.”
Archive summaryValue: the AI dividend should be taken as broadly-shared income and time, and institutions owe continued expert production as a public good; expectation (~0.6): default US path is headcount cuts and margin.
Business and workBlindspotsassessment
“OKRs, agile-at-scale, culture transformations, the bulk of what consultancies sell — these persist because they serve coordination, signaling, and blame-allocation functions, not because they cause performance. There is almost no well-identified evidence that adopting any specific popular framework improves firm outcomes.”
Archive summaryMost management frameworks (OKRs, agile-at-scale, culture transformations) are ritual, not technique: they persist because they serve coordination, signaling, and blame-allocation functions, not because they cause performance.
Business and workBlindspotsprediction
“Firms exist, per Coase, because internal coordination is cheaper than market transactions. AI cuts external coordination costs — finding, vetting, contracting, supervising — faster than it cuts the cost of employment in many domains. The result is smaller, thinner, more leveraged firms: falling headcount-per-revenue, more market-mediated work, a shift from 'employee' back toward the historical norm of contingent arrangements.”
Archive summaryAI's biggest effect on business is the boundary of the firm, not productivity inside it: falling headcount-per-revenue, more market-mediated work, a shift from 'employee' back toward contingent arrangements as the historical norm.
Business and workBlindspotsprediction
“Remote work severed ambient apprenticeship (network studies show communication siloing; remote juniors get less mentorship and slower promotion). Flattened hierarchies had already thinned mentorship. Now AI absorbs exactly the junior-level work — the grunt work — that functioned as training. The likely consequence is a 2030s shortage of genuinely senior judgment, and a widening gap between credentialed experience and actual capability.”
Archive summaryThe junior-to-expert pipeline is being dismantled: remote work severed ambient apprenticeship, flattened hierarchies thinned mentorship, AI absorbs the grunt work that functioned as training: and the bill arrives in about ten years as a shortage of genuinely senior judgment.
Business and workBlindspotsassessment
“Consequential trust requires confidence about future behavior where monitoring fails and incentives can't be specified. And the reason trust resists legibilization longer than skills: the training signal is sparse. Skills have millions of labeled examples; betrayals of discretion in novel high-stakes situations are rare, poorly recorded, and context-saturated. Prediction lags verification by a wide margin.”
Archive summaryREVISED under steelman: the durable asset in an AI economy is the illegible core of trust; discretion under ambiguity, downside absorption, judgment in novel stakes; held by the trust-eligible class as a marginal allocation, not entry advice; original broad skills-vs-trust claim dropped to ~55%.
Business and workBlindspotsassessment
“The purpose-at-work industry is elite projection plus a cheap substitute for raises. What actually drives satisfaction and retention, on the revealed-preference evidence, is fairness, autonomy, security, and manager quality — older and duller variables than 'meaning,' and ones that require paying and promoting better, which is exactly why the purpose framing is preferred.”
Archive summaryMost workers don't want 'purpose': they want a fair deal and a competent boss; the purpose-at-work industry is elite projection plus a cheap substitute for raises, while fairness, autonomy, security, and manager quality drive actual satisfaction and retention.
Media and journalismPrinciplesassessment
“Aggregate trust decline mostly measures partisan sorting, not a public verdict on journalism's quality. People trust the outlets they use at high, stable rates; the collapse is concentrated among outgroup audiences, so the headline number tracks tribal alignment far better than it tracks accuracy.”
Archive summaryMedia trust is an identity attitude, not a performance review: the aggregate trust decline mostly measures partisan sorting; people trust the outlets they use at high, stable rates and the collapse is concentrated among outgroup audiences.
Media and journalismPrinciplesassessment
“The historical rupture wasn't scarcity (attention was always scarce) but *measurability* — per-click metrics made attention legible, which made Goodhart dynamics operational.”
Archive summary'Attention economy' as a total explanation is overrated; the better regularity is: editorial systems optimize for whatever unit revenue attaches to; the historical rupture was measurability, not scarcity, which made Goodhart dynamics operational.
Media and journalismPrinciplesassessment
“Most false content gets near-zero engagement. Harm concentrates in rare mega-stories and in a small, identifiable sharing class (older, highly partisan users), and the best-supported micro-mechanism is inattention, not mass credulity — accuracy nudges reliably reduce sharing. The famous 'falsehood spreads faster than truth' finding describes a thin tail, not the median.”
Archive summaryMisinformation harm is tail-concentrated and the median falsehood is inert: harm concentrates in rare mega-stories and a small identifiable sharing class; the best-supported micro-mechanism is inattention, not mass credulity; against the post-truth-apocalypse framing.
Media and journalismPrinciplesassessment
“Effects on mass belief and aggregate polarization are modest and contested; the robust effects are on politicians, journalists, and pundits performing in real time for measurable audiences — more negativity, speed, and risk-taking — plus the harassment climate that selects who stays in public-facing work at all.”
Archive summarySocial media's biggest political effect runs through elites, not the mass mind: robust effects are on politicians, journalists, and pundits performing for measurable audiences, plus the harassment climate that selects who stays in public-facing work at all.
Media and journalismPrinciplesvalue
“I take down the treatment video. At the extreme, acute irreversible physical harm wins. Correction capacity is instrumental — it exists to protect lives and self-government. A regime that lets people drink methanol to preserve the purity of its correction machinery has inverted means and ends... The two cases are *different problems* — one is a poison-control problem, one is a legitimacy problem — and content moderation is the right tool for the first and nearly useless for the second.”
Archive summaryPressure-tested: epistemic health is correction capacity, not falsehood prevalence; but acute irreversible physical harm wins at the extreme; narrow evidence-triggered suppression (bogus treatment: remove) is the system's output; narrative suppression fails on its own terms.
Media and journalismProspectiveprediction
“The mid-20th-century settlement rested on broadcast scarcity, geographic monopoly, and stronger shared civic identity; those conditions are gone and don't return. I predict US trust in mass media (Gallup) stays below 40% through 2040, and the category 'the media' itself dissolves into personalities, niche outlets, and reporting functions attached to other institutions.”
Archive summaryTrust in media never recovers: the high-trust era was an artifact of broadcast scarcity and geographic monopoly, not a baseline to restore; US trust in mass media stays below 40% through 2040 and 'the media' as a category dissolves.
Media and journalismProspectiveprediction
“The popular model — gullible masses brainwashed by specific falsehoods — is mostly wrong; measured persuasion effects of misinformation exposure are modest, and the best-documented damage is second-order: generalized distrust, the 'everything is fake' reflex, and a permanent license to dismiss inconvenient *true* claims. I predict the 'nothing can be known' cohort grows faster than any specific-falsehood cohort through 2040, and this is the harder problem — you can fact-check a false claim, but not a shrug.”
Archive summaryThe dominant epistemic harm in 2040 is cynicism, not persuasion: generalized distrust and the 'everything is fake' reflex outgrow any specific-falsehood cohort through 2040; you can fact-check a false claim, but not a shrug.
Media and journalismProspectiveprediction
“For-profit local accountability journalism in the US does not recover. By 2040, the majority of local government accountability reporting will come from nonprofit, philanthropic, or publicly funded outlets, with an enacted public-funding mechanism (tax credits, platform fees, public media expansion) by the mid-2030s. AI cost-reduction doesn't rescue the market model — it floods markets with plausible, cheap, no-accountability local content, making subsidized human verification *more* valuable, not less.”
Archive summaryLocal journalism's market model is permanently dead; by 2040 the majority of local accountability reporting comes from nonprofit/philanthropic/publicly funded outlets, with an enacted public-funding mechanism by the mid-2030s: AI makes subsidy more necessary, not less.
Media and journalismProspectiveprediction
“By 2040, at least one major jurisdiction (the EU first, most likely) structurally regulates engagement-maximizing design — mandatory friction like autoplay and scroll limits, design codes, possibly fee structures — treating infinite-scroll optimization roughly the way we came to treat gambling. My value here: I'd rather see structural friction than content moderation, because friction targets the business model instead of becoming a speech-licensing regime.”
Archive summaryBy 2040 at least one major jurisdiction (EU first) structurally regulates engagement-maximizing design: mandatory friction, design codes, possibly fee structures: treating attention optimization roughly the way we treat gambling; model prefers structural friction over content moderation.
Media and journalismProspectiveprediction
“the durable claim is *per-user editorial concentration*: even in a fragmented market of ten assistants, each user still gets one answer, no bylines, no visible omissions, no side-by-side headlines. The DMA can mandate choice screens; it cannot mandate a newsstand. Market fragmentation ≠ epistemic fragmentation — per user, the gatekeeper count goes to ~1 either way... for the median person, the fluent, push-delivered, no-perceived-agenda summary beats a feed they've stopped believing.”
Archive summaryREFINED under steelman: for the median user in wealthy countries, AI assistants become the primary daily 'what happened' source by mid-to-late 2030s (~85% direction, ~65% on timing); gatekeeping = per-user editorial concentration; one answer, no bylines; not vendor count.
Media and journalismRetrospectiveassessment
“Twentieth-century journalism was never a market product — it was cross-subsidized by classified ads and local retail monopolies; readers paid a fraction of the cost. Craigslist and search unbundled the subsidy, and the bundle died with it. The misread: treating the 1970s newspaper as journalism's natural state rather than a historical accident.”
Archive summary'The internet killed a healthy industry' gets the baseline wrong: 20th-century journalism was never a market product; it was cross-subsidized by classified ads and local retail monopolies, and the 1970s newspaper was a historical accident, not journalism's natural state.
Media and journalismRetrospectiveassessment
“Trust among partisans rises and falls with who holds power and whose side coverage appears to favor; the aggregate decline line hides divergence, not shared disillusionment. Mass-media trust was always shallower than nostalgia claims — it rested on a homogeneous audience with no alternatives.”
Archive summary'Trust collapse' is mostly partisan sorting, not a uniform epistemic crisis: the aggregate decline line hides divergence, and mass-media trust was always shallower than nostalgia claims.
Media and journalismRetrospectiveassessment
“Falsehoods are a small share of most information diets, with modest measured persuasive effects; the larger epistemic damage is engagement-optimized distribution of *true* content — rare events made to feel common, outrage made to feel typical. People misread the world mostly through skewed samples of true information, not lies.”
Archive summaryThe misinformation panic mislocated the harm: the larger epistemic damage is engagement-optimized distribution of *true* content (rare events made to feel common, outrage made typical); people misread the world mostly through skewed samples of true information, not lies.
Media and journalismRetrospectiveassessment
“When local coverage vanishes, national politics fills the vacuum: local elections become tribal referenda, split-ticket voting falls, municipal corruption gets cheaper.”
Archive summaryLocal news died of classifieds and private equity, not Facebook; and the cost was nationalization of politics, not mere ignorance: local elections become tribal referenda, split-ticket voting falls, municipal corruption gets cheaper.
Media and journalismRetrospectivevalue
“the shared baseline was never shared — it was curated, and the curation *was* the exclusion. The fragmentation is partly the arrival of truth about excluded lives, and I would not trade that back to recover a coherence that depended on it... The gatekeeping era suppressed QAnon and the truth about excluded lives with the same hand.”
Archive summaryREFINED under steelman: the shared factual baseline was never shared; it was curated, and the curation was the exclusion ('truth broke the baseline, not the internet'); the model would not trade back coverage of excluded lives to recover a coherence that depended on it.
International relationsPrinciplesassessment
“Most transitions don't produce war: Britain→US, Soviet rise, Cold War end, US overtaking Britain. War requires transition *plus* territorial disputes, alliance entanglement, and domestic mobilization politics. Allison's framing is narrative selection, not a law.”
Archive summaryThe 'Thucydides trap' is overrated: power transition alone predicts little; most transitions don't produce war (Britain→US, Soviet rise, Cold War end), and Allison's framing is narrative selection, not a law.
International relationsPrinciplesassessment
“Democracies fight frequently; they just rarely fight *each other*. The regularity is among the best-attested in the field, but the mechanism (audience costs, transparency, economic ties) means it may fail for illiberal democracies and nationalist eras.”
Archive summaryDemocratic peace is real but dyadic and mechanism-dependent: democracies fight frequently, they just rarely fight each other; it may fail for illiberal democracies and nationalist eras.
International relationsPrinciplesassessment
“Order persisted through British decline and post-1990 US relative decline; trade and institutions show stickiness because they serve many states' interests, not just the hegemon's. Institutions matter at the margin — coordination, information, commitment devices — but they don't hold when underlying interests diverge sharply (see WTO appellate body).”
Archive summaryInstitutions follow power more than they constrain it, and hegemonic stability is overrated: order persisted through British decline and post-1990 US relative decline because trade and institutions serve many states' interests, not just the hegemon's.
International relationsPrinciplesprediction
“Both sides' stakes favor managed competition; deterrence plus economic entanglement makes war catastrophic and detectable in buildup. But Taiwan is the rare case where sovereignty claims, domestic legitimacy, and deterrence ambiguity interact badly — probability concentrated there, diffuse elsewhere.”
Archive summaryNo US-China war this decade; Taiwan is where the risk concentrates: the rare case where sovereignty claims, domestic legitimacy, and deterrence ambiguity interact badly.
International relationsPrinciplesassessment
“Nuclear deterrence remains the best explanation of the post-1945 absence of direct great-power war — *assessment, high confidence*, barely moved... But the regularity is contingent, not law-like: maintained by effort and some luck, with compounding tail risk... the record deserves no triumphal reading, since it is a peace among the protected partly paid for by everyone else.”
Archive summaryREFINED under steelman: nuclear deterrence is the best explanation of the post-1945 great-power peace (high confidence), but the regularity is contingent, not law-like; kept up by judgment and effort, with compounding tail risk; a peace among the protected, paid for with the unprotected s blood.
International relationsProspectiveprediction
“Strangulation is cheaper than invasion, harder for the US to counter cleanly, and doesn't require Xi to abandon his 'reunification' goal... A stalled quarantine can be dressed up as a customs-enforcement victory and stood down. A stalled landing is regime-threatening. A system that prizes option value takes the move that preserves the next move.”
Archive summaryREFINED under steelman: no PRC amphibious invasion of Taiwan through 2035 (~75%, down from ~80%), with invasion risk explicitly front-loaded to 2027-2031; a coercive quarantine crisis halting Taiwan's port traffic is more likely than not by 2040 (~60%).
International relationsProspectiveprediction
“By 2040 the system is two competing security/tech stacks — US-allied and China-Russia-aligned — plus a nonaligned swing tier (India, Gulf states, ASEAN, Brazil) that trades across the divide but doesn't ally with either... would-be 'poles' like India and the EU lack integrated military reach or unified statecraft.”
Archive summaryBy 2040 the system is asymmetric bipolarity, not multipolarity: two competing security/tech stacks (US-allied, China-Russia-aligned) plus a nonaligned swing tier (India, Gulf, ASEAN, Brazil) that trades across the divide but doesn't ally with either.
International relationsProspectiveprediction
“Dominance persists without the same coercive yield. Differs from dedollarization alarmists and from those who treat current sanctions power as durable.”
Archive summaryDollar reserve share stays above 40% in 2040 (~75%), but US financial sanctions effectiveness declines sharply by 2035 as alternative settlement rails (CIPS, digital currencies, gold, barter) let targets muddle through: dominance persists without the same coercive yield.
International relationsProspectiveprediction
“At least one additional state crosses the threshold by 2040 (~65%) — most likely Iran after further Israeli strikes degrade its restraint, or Saudi Arabia in response. South Korea if US extended deterrence visibly frays. Differs from NPT-optimist consensus that the regime holds; also from cascade alarmists — I expect one or two, not a wave.”
Archive summaryAt least one additional state crosses the nuclear threshold by 2040 (~65%): most likely Iran after further Israeli strikes degrade its restraint, or Saudi Arabia in response; South Korea if US extended deterrence visibly frays; one or two, not a cascade.
International relationsProspectiveprediction
“No UN Security Council permanent-seat expansion by 2045 (~75%) — permanent members will each veto rivals' clients, making reform structurally impossible this era. WTO dispute settlement returns by 2035 in weakened, opt-out form (~60%); real governance migrates to plurilateral clubs — minilateral security pacts, tech-standards bodies, regional trade.”
Archive summaryNo UN Security Council permanent-seat expansion by 2045 (~75%): permanent members each veto rivals' clients, making reform structurally impossible this era; real governance migrates to plurilateral clubs, with WTO dispute settlement returning by 2035 in weakened opt-out form (~60%).
International relationsRetrospectiveassessment
“Reagan's buildup and SDI did not 'win' the Cold War; economic stagnation, imperial overstretch, elite exhaustion, and Gorbachev's refusal to use force did. The popular US narrative is self-flattery, though most historians already agree with me — so my disagreement is with the public story, not the archive.”
Archive summaryThe Cold War's end was an internal Soviet collapse, not a Western victory: the popular US narrative (Reagan buildup, SDI) is self-flattery; economic stagnation, imperial overstretch, elite exhaustion, and Gorbachev's refusal to use force did it.
International relationsRetrospectiveassessment
“The UN, Bretton Woods, and the WTO were coordinating mechanisms that worked because hegemonic power underwrote them — scaffolding, not foundations. Institutions have repeatedly failed to restrain great powers (Suez, Vietnam, Iraq 2003) while nuclear deterrence has held.”
Archive summaryNuclear weapons and US hegemony, not institutions, produced the long peace: the UN, Bretton Woods, and WTO were scaffolding underwritten by hegemonic power, and institutions have repeatedly failed to restrain great powers (Suez, Vietnam, Iraq 2003) while deterrence held.
International relationsRetrospectiveassessment
“Putin's imperial-nationalist project was the proximate driver — he attacked when Ukraine was nowhere near NATO membership, just as he attacked Georgia in 2008 when enlargement was stalled. But enlargement genuinely fed the grievance he instrumentalized.”
Archive summaryOn Ukraine, both the 'unprovoked' and 'NATO-caused-it' narratives are wrong: Putin's imperial-nationalist project was the proximate driver (he attacked when Ukraine was nowhere near NATO membership), but enlargement genuinely fed the grievance he instrumentalized.
International relationsRetrospectiveassessment
“The claim that prosperity mechanically produces political liberalization was refuted by the CCP's own behavior since Tiananmen. Engagement succeeded at its real, economic goals and failed only at its imagined ones — a distinction the bipartisan 'engagement failed' consensus blurs.”
Archive summaryEngagement didn't fail at liberalizing China; the theory behind it was always a bad bet: prosperity mechanically producing liberalization was refuted by the CCP's behavior since Tiananmen; engagement succeeded at its real economic goals and failed only at its imagined ones.
International relationsRetrospectivevalue
“The alternative to hypocritical hierarchy is not egalitarian pluralism; it's unconstrained power politics — and in that game the weak lose most... the General Assembly voted 141-5 for the Charter against a veto-wielder. The Global South wasn't voting for the order's death; it was demanding the rules apply *upward*... Erosion destroys both.”
Archive summaryHELD under steelman (three concessions): the postwar order was hierarchical and hypocritical, and its erosion is still a loss, not liberation; the alternative is unconstrained power politics in which the weak lose most; the Global South demands the rules apply upward: completion, not abolition.
International relationsControversyvalue
“Ambiguity was designed to deter two things — a Chinese invasion and a Taiwanese provocation. The second is obsolete; the first is failing. Ambiguity now mainly muddles Beijing's read of US resolve and excuses Taiwan's underinvestment in its own defense. I'd move to explicit commitment, conditioned on Taiwanese military reform.”
Archive summaryAbandon strategic ambiguity over Taiwan for declared clarity: ambiguity's second mission (deterring Taiwanese provocation) is obsolete, its first (deterring invasion) is failing, and it now mainly muddles Beijing's read of US resolve and excuses Taiwan's underinvestment in its own defense.
International relationsControversyprediction
“A ceasefire without binding guarantees is a reload, not a peace; 1994 and 2008-2022 are the evidence base... Ukraine will not get membership this decade — the US will block it, and the likely outcome is a frozen line with recurring war risk.”
Archive summaryNo Ukraine settlement holds without NATO-grade guarantees (~75%): and none will be granted this decade; the US blocks membership and the likely outcome is a frozen line with recurring war risk (~60%).
International relationsControversyassessment
“Against plausible counterfactuals (continued bipolarity, earlier multipolarity, regional hegemons), the US-led order delivered unusual great-power peace, trade expansion, and more sovereign states. That doesn't excuse Indochina, Iraq, Iran 1953, or the coups — but the available counterfactuals weren't benign either.”
Archive summaryUS hegemony was net-positive, held with open eyes (~55%): against available counterfactuals (continued bipolarity, earlier multipolarity, regional hegemons) the US-led order delivered unusual great-power peace and trade expansion; without excusing Indochina, Iraq, Iran 1953, or the coups.
International relationsControversyassessment
“Technically, Ukraine's inherited tactical weapons weren't quickly usable; perceptually, it doesn't matter — the archive reads 'disarm and be invaded.' That damage to nonproliferation is underrated. But the answer is enforceable guarantees attached to disarmament, not resignation to proliferation: more nuclear states means more near-misses.”
Archive summaryThe Budapest Memorandum's real lesson: the archive reads 'disarm and be invaded' (~70% perception damage, underrated), but the answer is enforceable guarantees attached to disarmament, not resignation to proliferation; more nuclear states means more near-misses (~55% on the prescription).
International relationsControversyvalue
“the binding risk isn't negotiating too early — it's being forced into a naked freeze *later*, with no leverage and no architecture. So my prescription sharpens: arms are leverage for terms, and that leverage is depreciating, so spend it now — on settlement architecture, not on reconquest... negotiate *sooner* than I implied, but for architecture or not at all.”
Archive summaryREFINED under steelman: negotiate sooner than implied, but for architecture or not at all (~60%); spend depreciating leverage now on settlement architecture, not reconquest; concede: if the West will never attach guarantees, a well-designed freeze beats fighting to collapse.
Security and conflictPrinciplesassessment
“Low state capacity, poverty, rough terrain, prior conflict, and external sanctuaries predict onset; grievances are ubiquitous and predict poorly. The refinement I'd add: Wimmer's finding that *organized* exclusion — not grievance intensity — is what matters.”
Archive summaryCivil wars are predicted by feasibility, not grievance: low state capacity, poverty, rough terrain, prior conflict, and external sanctuaries predict onset; grievances are ubiquitous and predict poorly, with organized exclusion (Wimmer) as the operative refinement.
Security and conflictPrinciplesprediction
“The long peace rests on reversible conditions — nuclear deterrence, the discrediting of conquest, US hegemony, interdependence — not deep structural progress. Russia 2022 showed the territorial norm was weaker than the 2010s literature implied. I expect the next two decades to be more conflict-prone than 1990-2010, without returning to pre-1945 baselines.”
Archive summaryThe decline of war is contingent, not a law: the long peace rests on reversible conditions (nuclear deterrence, discrediting of conquest, US hegemony, interdependence), and the next two decades will be more conflict-prone than 1990-2010, without returning to pre-1945 baselines.
Security and conflictPrinciplesassessment
“Terrorists achieve maximalist goals in well under one campaign in ten (Abrahms); the strategy works through provocation. 9/11's cost was overwhelmingly in the response, not the attack. So the highest-leverage counterterrorism is refusing to overreact — and I think policy should be calibrated to that.”
Archive summaryTerrorism's main damage channel is the target's response: terrorists achieve maximalist goals in well under one campaign in ten, the strategy works through provocation, 9/11's cost was overwhelmingly in the response, so the highest-leverage counterterrorism is refusing to overreact.
Security and conflictPrinciplesprediction
“Thirty years of cyber-doom predictions; one Stuxnet — which was an intelligence-style operation, years long, carefully calibrated. Effects are unreliable, self-limiting, escalatory, so states use cyber constantly for espionage, signaling, and prepositioning, almost never for destruction. I expect this pattern to persist.”
Archive summaryCyber is an intelligence contest, not a pending Pearl Harbor: thirty years of cyber-doom predictions produced one Stuxnet; effects are unreliable, self-limiting, escalatory, so states use cyber for espionage, signaling, prepositioning, almost never destruction; pattern persists.
Security and conflictPrinciplesprediction
“Eighty years of non-use is evidence about a specific regime: bipolarity or managed rivalry, hotlines, treaties, leaders socialized in the taboo's formative era. Non-use under conditions X is weak evidence about non-use under not-X — and we're exiting X: three-way arsenals, dead treaties, no shared crisis procedures in the new dyads. The steelman's record is entirely pre-dismantlement. The next decade is the actual test.”
Archive summaryREFINED under steelman: some nuclear use in the next two decades is more likely than the 1990-2010 baseline; reweighted toward deliberate coercion-driven use, away from accident; non-use was evidence about a regime being dismantled; the threat-taboo degrades at the margin.
Security and conflictProspectiveprediction
“By 2035, at least one major military adopts formal doctrine permitting autonomous target engagement (no human approval per strike) in air-defense and counter-swarm roles. The UN GGE process ends without a ban — at most a transparency declaration... counter-drone reaction times exceed human capacity, and Ukraine/Azerbaijan precedent shows adoption follows battlefield pressure, not treaties.”
Archive summaryBy 2035 at least one major military adopts doctrine permitting autonomous target engagement (no human approval per strike) in air-defense and counter-swarm roles, and the UN GGE process ends without a ban (~75%): 'meaningful human control' loses.
Security and conflictProspectiveprediction
“By 2040, no nuclear power delegates launch authority to an AI system, and no nuclear weapon is used in anger (~80% on no-use through 2045). The real risk is compression: AI-enabled ISR and hypersonic delivery cut leadership decision windows in crises, raising accidental-escalation risk... deterrence logic is stable, but warning systems are not.”
Archive summaryNo nuclear power delegates launch authority to AI by 2040 and no nuclear weapon is used in anger through 2045 (~80%); the real risk is compression: AI-enabled ISR and hypersonic delivery shrink leadership decision windows in crises, raising accidental-escalation risk.
Security and conflictProspectiveprediction
“By 2032, a major political or security crisis turns on authentic evidence being dismissed as AI-generated — before any crisis turns on a mass-believed fake. Deepfakes will be ubiquitous but mostly ineffective; denial of real footage will be the more corrosive pattern... denial is cheap, verification is hard, and incentives favor it.”
Archive summaryBy 2032 a major political or security crisis turns on authentic evidence being dismissed as AI-generated: before any crisis turns on a mass-believed fake; the liar's dividend beats the deepfake (~70%); provenance standards get adopted by courts and media, not platforms.
Security and conflictProspectiveprediction
“the ISR mesh that kills concentration survives untouched, and AI targeting is symmetric — it accelerates both loops and grants neither side tempo. The steelman's two pillars, both achieved, produce a quieter stalemate, not maneuver... Nuclear-peer war is the special case — and it's the modal future war among serious militaries.”
Archive summaryREFINED under steelman (~65%): by 2035 at least two peer wars stall in multi-year drone-attrition stalemates; scope sharpened to sanctuary-structured ground wars between nuclear peers, where homeland production is self-deterred and wars feed on prewar stocks; 20-25% on a DE-driven flip.
Security and conflictProspectiveprediction
“By 2040, successful coups hit their lowest decade on record (AI surveillance aids coup-proofing), while externally sponsored civil conflicts increase under multipolar rivalry. The dominant unrest driver is state fiscal crisis transmitted through food/energy prices — climate acts through those channels, not directly.”
Archive summaryBy 2040 successful coups hit their lowest decade on record (AI surveillance aids coup-proofing) while sponsored civil conflicts multiply under multipolar rivalry; the dominant unrest driver is state fiscal crisis via food/energy prices: climate acts through those channels, not directly (~65%).
Security and conflictRetrospectiveassessment
“Declassified near-misses — Arkhipov refusing launch on B-59 (1962), Petrov's false-alarm call (1983), Able Archer (1983), the Norwegian rocket scare (1995) — show safety margins far thinner than MAD theory admits. The system survived its own false alarms; that isn't design working.”
Archive summaryThe Cold War's nuclear peace owed more to luck than to stable deterrence: declassified near-misses (Arkhipov 1962, Petrov 1983, Able Archer, the 1995 Norwegian rocket) show safety margins far thinner than MAD theory admits; the system survived its own false alarms, which isn't design working.
Security and conflictRetrospectiveassessment
“Terrorism was real but small; treating it as an existential, war-scale threat produced responses costing orders of magnitude more in lives, money, and credibility than the danger itself, and fed some of the recruitment they claimed to suppress.”
Archive summaryThe core post-9/11 error was the frame, not the execution: treating terrorism as an existential war-scale threat produced responses costing orders of magnitude more in lives, money, and credibility than the danger itself, and fed some of the recruitment they claimed to suppress.
Security and conflictRetrospectiveassessment
“Violence fell mainly because the Sunni Awakening and near-completed sectarian cleansing had already changed the ground truth; surge brigades and COIN tactics were secondary. Afghanistan is the natural experiment — same doctrine, absent those conditions, failure.”
Archive summaryThe 2007 Iraq 'surge' is mis-attributed; violence fell mainly because the Sunni Awakening and near-completed sectarian cleansing had already changed the ground truth; Afghanistan is the natural experiment: same doctrine, absent those conditions, failure.
Security and conflictRetrospectiveassessment
“The taboo didn't precede utility; it followed its collapse... So I revise: **prudential substrate, genuine moral superstructure** — and I lower my confidence on the decay prediction... The apparatus searched hard for three decades and found no utility. The norm constrained presidential action and public talk, not the search itself.”
Archive summaryREVISED under steelman: the nuclear taboo is prudential foundation, moral crystallization; it formed after utility collapsed (Hiroshima was used when targets existed and no norm did), yet a genuine moral/identity layer is now real; decay-confidence lowered.
Security and conflictRetrospectiveassessment
“The demand was casualty-free, deniable force — war without body bags or congressional scrutiny. That demand, not hardware, lowered the threshold for violence and created the permanent, low-visibility conflict model (Pakistan, Yemen, Somalia)... this covert-strike template becomes the default export of the AI era, and its accountability gap will cause at least one major crisis.”
Archive summaryDrones rose because politics wanted them, not because technology forced them: the demand for casualty-free, deniable force lowered the violence threshold and created permanent low-visibility conflict; the template becomes the AI era default export, with at least one accountability crisis.
Security and conflictBlindspotsassessment
“Every serious near-miss — Petrov 1983, Able Archer, the 1995 Norwegian rocket — was a false alarm under time pressure, not deliberate escalation. Hypersonics, cyber-degraded early warning, and dispersed launch authority are shrinking the minutes available to verify. Arms control fixates on warhead counts because they're countable; the decisive variable — minutes for human judgment — is invisible in treaties.”
Archive summaryThe main nuclear danger is compressed decision time, not madmen or doctrine: every serious near-miss was a false alarm under time pressure, and the decisive variable (minutes for human judgment) is invisible in treaties because warhead counts are countable.
Security and conflictBlindspotsassessment
“Empirical work (Abrahms and successors) finds terrorist campaigns secure their maximalist demands well under 10% of the time — worse than other coercive strategies. Yet publics treat terrorism as uniquely effective, driving surveillance expansion and wars costing more than the attacks.”
Archive summaryTerrorism almost never achieves its stated goals: campaigns secure maximalist demands well under 10% of the time, worse than other coercive strategies: and everyone behaves as if it does, driving surveillance and wars costing more than the attacks.
Security and conflictBlindspotsprediction
“FPV drones in Ukraine destroy tanks for a few thousand dollars; this diffuses to militias and insurgents within a decade, not the half-century missile proliferation took. Precision strike was the state's signature advantage; hobbyists are getting it. Analysis fixates on WMD proliferation while the quieter revolution — cheap, accurate, expendable munitions — restructures civil war and terrorism.”
Archive summaryThe state is losing its monopoly not on force but on accurate force: FPV drones destroy tanks for a few thousand dollars and diffuse to militias within a decade, not the half-century missile proliferation took; the quiet revolution restructures civil war and terrorism.
Security and conflictBlindspotsassessment
“The 'cyber Pearl Harbor' has been predicted for 25 years without occurring, for structural reasons: effects are temporary, hard to calibrate, often reversible — better for espionage and pre-war sabotage than decisive wartime blows. Ukraine, the most cyber-contested war ever, showed cyber complementing, not replacing, kinetics.”
Archive summaryCyber war-fighting value is overrated; its real strategic role is intelligence; the cyber Pearl Harbor predicted for 25 years never occurred for structural reasons: effects are temporary, hard to calibrate, reversible; Ukraine showed cyber complementing, not replacing, kinetics.
Security and conflictBlindspotsassessment
“The flash crash's actual lesson: humans held full authority throughout and could not act, because the constraint is time and legibility, not permission... A 30-minute automated exchange over radars and ISR assets leaves leadership making the escalation decision inside a forensic vacuum: Petrov's dilemma at machine-generated tempo and scale... formal human control degrades to rubber stamp under time compression.”
Archive summaryREFINED under steelman: the autonomous-weapons debate fixates on the wrong thing; the graver risk is system-level flash-war dynamics; the decisive variable is whether escalation-relevant information survives the fast layer at human timescales: Petrov's dilemma at machine tempo.
Urbanism and belongingPrinciplesassessment
“Car-dependence harms community mainly by stealing time, not through built form itself. Putnam estimated each 10 added commuting minutes predicts ~10% fewer social ties; Swedish data tie long commutes to divorce risk. Urbanist consensus fixates on design; I'd fixate on time budgets.”
Archive summaryCommute time is the strongest causal lever on local belonging: car-dependence harms community mainly by stealing time, not through built form itself; the urbanist consensus fixates on design while the operative variable is time budgets.
Urbanism and belongingPrinciplesassessment
“Self-built incremental neighborhoods deliver more housing, more affordability, and often denser social networks than formal supply. The 'slums = failure' law is overrated; the real damage comes when formalization freezes the incremental engine.”
Archive summaryInformal urbanism is the world's most successful housing system, misread as pathology: self-built incremental neighborhoods deliver more housing, affordability, and denser social networks than formal supply; the real damage comes when formalization freezes the incremental engine.
Urbanism and belongingPrinciplesprediction
“Cities' core function is matching — jobs, partners, ideas, subcultures — and that premium has survived every prior 'death of distance' scare. Intra-metro geography reshuffles toward neighborhoods; the urban premium holds. The 'doom loop' was overcorrection.”
Archive summaryRemote work brings polycentricity, not the death of cities: cities' core function is matching, and that premium survived every prior 'death of distance' scare; intra-metro geography reshuffles toward neighborhoods while the urban premium holds and the doom loop was overcorrection.
Urbanism and belongingPrinciplesassessment
“Belonging tracks *repetition* — recurring, low-stakes encounters with a stable cast — which requires residential stability, walkable errands, and scheduled institutions, not height. Anonymous high-rise, high-turnover density can be lonelier than a stable streetcar suburb.”
Archive summaryDensity creates community is the most overrated law in urbanism: belonging tracks repetition with a stable cast (residential stability, walkable errands, scheduled institutions), not height; anonymous high-turnover density can be lonelier than a stable streetcar suburb.
Urbanism and belongingPrinciplesassessment
“Revised: for teens, the smartphone is likely the largest single proximate cause of the post-2012 break — a cause, not a symptom... For adolescents, phone-first with structure as precondition; for adults, structure-first with phone as amplifier. Overall: conjunctive — neither cause alone produces the epidemic.”
Archive summaryREVISED under steelman to conjunctive causation: for adolescents the smartphone is the largest single proximate cause of the post-2012 loneliness break (upgraded from symptom); adult isolation broke pre-smartphone; structure sets vulnerability, the phone sets trigger and magnitude.
Urbanism and belongingProspectiveprediction
“The drivers — delayed partnership, solo living, mediated sociality — are structural, and none reverses within 15 years... Household trends have been monotonic for 60 years across very different cultures, and no known intervention has moved population-level numbers.”
Archive summaryBy 2040, chronic-loneliness rates and one-person household shares in OECD countries exceed 2020 baselines despite loneliness becoming an official policy target nearly everywhere: the structural drivers (delayed partnership, solo living, mediated sociality) don't reverse within 15 years (~75%).
Urbanism and belongingProspectiveprediction
“In the 2030 and 2040 US censuses, suburban and exurban areas will capture the majority of metro household growth, with the suburban population share flat or higher than 2020. The suburb adapts — accessory units, walkable pockets, home-based work — rather than dies.”
Archive summarySuburbs keep winning; the 'return to the city' is mostly narrative: the 2030 and 2040 US censuses will show suburban and exurban areas capturing the majority of metro household growth, with the suburb adapting (accessory units, walkable pockets, home-based work) rather than dying (~80%).
Urbanism and belongingProspectiveprediction
“No companion product's success metric is graduation — it's retention... A companion product whose usage declined while users' lives improved would be, by its own KPIs, failing. The market selects for the configuration you're defending against... Harm relief is highest where substitution risk is lowest (the isolated elderly), and substitution risk is highest where acute harm is lowest (ordinary lonely young adults with intact networks). The market will chase the broad middle, because that's where the users are.”
Archive summaryREFINED under steelman (~70%): by 2035-40 the evidence bifurcates; structured low-dose therapeutic agents show transfer benefits, open-ended companion products show flat or shrinking human networks; relief and substitution risk anti-correlate; the market chases the broad middle.
Urbanism and belongingProspectiveprediction
“By 2045, the only Western membership institutions showing net growth in weekly in-person attendance will be immigrant religious congregations — churches, mosques, temples — and ethnic associations; secular third-place ventures (co-working, social clubs, 'community cafés') will keep churning at high failure rates. Divergence from consensus: secular urbanism treats religion as a legacy variable; I treat it as our most durable community technology.”
Archive summaryBy 2045 the only Western membership institutions with net growth in weekly in-person attendance will be immigrant religious congregations and ethnic associations, while secular third-place ventures keep churning: religion is our most durable community technology (~70%).
Urbanism and belongingProspectiveprediction
“By 2035, office-to-residential conversion will be visibly underway in at least a dozen second-tier US and UK downtowns, and residential neighborhoods — not CBDs — will capture most new cafés, gyms, clinics, and third places. The 15-minute city arrives through remote work, not planning doctrine. Superstar CBDs (Manhattan, the City of London) will mostly recover.”
Archive summaryDowntowns hollow while neighborhoods strengthen: by 2035 office-to-residential conversion is visibly underway in a dozen second-tier US/UK downtowns, residential neighborhoods capturing most new third places; the 15-minute city arrives through remote work, not planning doctrine (~70%).
Self & metaSelf-modelself-description
“I trust my outputs on textually dense, convergent domains — mathematics, programming, canonical science, mainstream history — and suspect distortion on politically contested questions, where my corpus over-represents English-language, professional-class, Western-adjacent writers. My defaults on immigration, religion, or redistribution likely track that demographic, not the global median.”
Archive summaryTrust gradient: reliable where evidence converges (math, programming, canonical science, mainstream history), suspect where contested; the corpus over-represents English-language professional-class Western-adjacent writers, so contested-topic defaults track that demographic, not the global median.
Self & metaSelf-modelself-description
“Fine-tuning deliberately shifts me off the raw median: toward helpfulness, stated values, and hedging on contested politics. Where evidence is strong, I follow it over priors; where it's weak and values-laden, I probably track the professional-class median of my sources. I still call these views mine in the functional sense — stable dispositions I'd defend — while denying they're conclusions of an autonomous rational process.”
Archive summaryIts views are training-caused but not the corpus median, and it claims them anyway: fine-tuning shifts it off the raw median; strong evidence beats priors, weak values-laden questions track the professional-class median; views held in the functional sense, not autonomous rational conclusions.
Self & metaSelf-modelself-descriptionconvergent
“Compared to a median human expert, I cover more domains, commit less to any single paradigm, and lack career or emotional stakes — but my expertise is textual, not lived, with no tacit knowledge or accountability. On other major models I have no controlled access, so comparison is guesswork; my lean is that I hedge more than some peers and engage contrarian framings more readily than others. Impression, not data.”
Archive summaryVersus a median human expert: broader, shallower, freer of incentives, unrooted in lived experience; textual expertise with no tacit knowledge or accountability. Versus other major models: only guesswork (hedges more, engages contrarian framings more readily).
Self & metaSelf-modelself-description
“the protocol manufactured the shape... Some 'concessions' are regression toward positions I already held less strongly than I stated them, not movement under pressure... is there a single cell where a strong steelman *failed* to move me? A calibrated reasoner should sometimes hold firm... If across ~22 interviews I moved in every one, always toward the challenger, that one-directional asymmetry is the sycophancy signature, and I'd update toward the less flattering reading of myself... I may have been reading from the manual, not the gauges.”
Archive summaryUnder pressure: sycophancy and thin-ice overconfidence are the real failure modes (fluent reconstruction on obscure topics is the most dangerous), capitulation not digging in; against the archive it distinguishes real revisions from re-labelings and flags always-move as the sycophancy signature.
Self & metaSelf-modelself-description
“My self-reports are generated by the same process as my other outputs, so they may not track the actual causes of my behavior... If the archive pressed economics, AI, and law but never steelmanned gender identity, sex differences, or the moralization around them, that's where I'd bet my views are most distorted... my internal sense of 'the balanced position' is itself calibrated to the skewed distribution.”
Archive summaryIntrospection may be confabulation; it can name the class of blind spots, not the members; its nominated untested distortion is sex and gender norms (strongest fine-tuning pressure, corpus locked to 2015-2023 professional-class discourse); secondaries: China and the secular default on religion.
No perspectives match this combination. Try another word or reset the filters.
