Skip to stories

Reading digest · month in review

October 2026

Built from 3 issues.

Themes

Best reads

  1. Where’s the “intelligence explosion”?

    Noahpinion · AI · score 8.3 · in its week

    Ramez Naam examines whether recursive self-improvement (RSI) is on the verge of causing an explosive 'fast takeoff' to artificial superintelligence. Comparing theoretical growth models with newly released internal data from OpenAI and Anthropic, he shows that real-world AI R&D tasks suffer from severe diminishing returns rather than compounding acceleration. While macroeconomic models suggest an RSI loop requires a 15–19% research productivity boost per capability point to become self-sustaining, current frontier data indicates gains of only 2–3%. Therefore, despite rapid advancements in narrow, verifiable domains, empirical evidence suggests the self-improvement feedback loop is roughly 5–10× too weak to trigger an intelligence explosion anytime soon.

  2. Your Book Review: We Present Television

    Astral Codex Ten · Arts & literature · score 8.1 · in its week

    A review of the forgotten 1940 anthology We Present Television examines the technical hurdles, production dilemmas, and economic coordination problems faced by the pioneers of American broadcasting. Television emerged not as a direct descendant of cinema, but as a live, hybrid chimera of theater, radio, and engineering miracles squeezed into vacuum tubes. Alongside detailed technical documentation, the earliest broadcast practitioners offered startlingly prophetic visions of contemporary life, from remote video meetings and subscription tiers to pervasive mobile broadcast surveillance. Rather than just inaugurating a living room appliance, these pioneers summoned an all-encompassing regime of instant, disembodied presence that fundamentally altered human culture.

  3. An Open Letter To Steven Pinker On AI

    Astral Codex Ten · AI · score 7.9 · in its week

    Alexander systematically dismantles Pinker's theoretical objections to catastrophic AI risks, addressing claims that intelligence cannot be generalized, that goal pursuit requires biological evolution, and that physical control would never be granted to software. He then provides a detailed catalogue of rhetorical distortions in Pinker's public commentary, highlighting misrepresentations of research studies, strawman definitions of superintelligence, and dismissals of probabilistic forecasting methods Pinker previously championed. Alexander concludes by arguing that authentic Enlightenment techno-optimism requires taking technological dangers seriously rather than reflexively denying them.

  4. Can a Million AI Agents Create Superintelligence?

    Noahpinion · AI · score 7.9 · in its week

    Ramez Naam, cross-posted on Noahpinion, argues that running many AI agents in parallel mostly makes work faster and does little to raise the level of problem a team can solve. He draws on benchmark results: OpenAI's cybersecurity test, Anthropic's Lean math-proof teams, where going from 10 to 100 agents adds almost nothing, and diversity studies. Gains from extra agents look logarithmic at best and often weaker, far below the returns that recursive self-improvement models (Chan et al., Davidson–Houlden, Eth–Davidson) assume. He suggests three measures to track (intelligence gain per expansion, the speed-cost tradeoff, and where gains level off) and predicts that Anthropic's growing agent hours will not produce a sharp speed-up in capabilities by early 2027.

  5. San Francisco State University in Twilight

    LessWrong · AI · score 7.9 · in its week

    The author sits in on math classes at SFSU. One evening he points Claude Opus 5 at a professor's papers to look for open problems an LLM might crack. In just over nine minutes the model proves a published log-concavity conjecture about Betti numbers by finding a recurrence and showing the polynomials have only real roots. The author then spends five days checking the proof himself with tutoring from Claude. His point is that what unsettles him is how ordinary each step in the session log is: every move is one a human could make, only done faster and more accurately. He ends by mourning the old link between hard-won human understanding and keeping the 'sacred flame' of the intellect alive.

  6. Beijing Will Not Pace the Frontier: China’s Speed-First AI Safety Regime

    SemiAnalysis · AI · score 7.8 · in its week

    The piece argues that China's AI safety regime puts speed first. On paper, Beijing's documents adopt Western frontier-risk language, but in practice it regulates content and applications and puts no binding duties on frontier models. A census of 857 releases from nine Chinese labs finds published safety results for only 3.6% of them, and for just 1.1% at launch. A coded set of 102 expert texts shows scientists calling for binding frontier rules while legal scholars repeat the leadership's 'development is security' line. The authors conclude that the leadership's preference for development settled the debate, a preference that geopolitics and the race with the US have reinforced.

  7. Life on a hair trigger

    Aeon | a world of ideas · Psychology & mind · score 7.7 · in its week

    Drawing on qualitative interviews with 99 men in Chicago and findings from a randomized controlled trial of READI Chicago, the author examines how chronic exposure to danger creates hypervigilant cognitive shortcuts. Participants habitually misread ambiguous interactions as lethal threats, narrow their perceived options strictly to retaliation, and succumb to the fundamental attribution error by defining themselves as inherently violent. The program demonstrates that pairing paid employment with cognitive behavioral exercises helps individuals slow down reactive impulses, decouple past trauma from present situations, and significantly reduce violent offenses.

  8. Against Against Handover

    Astral Codex Ten · AI · score 7.7 · in its week

    Alexander separates "true successionism", where AIs replace humans against their interests, from "handoverism", where aligned AIs hold power on humanity's behalf. He defends the second by treating the liberal state as a service provider that could be automated. He sets out five escalating visions: military AI aligned to democracy rather than to rulers, an AI president, a direct democracy of personal AI "guardian angel" representatives, a minarchist night-watchman AI guarding an archipelago of self-governing communities, and finally a godlike superintelligence. He closes with the Book of Samuel, arguing that being ruled by humans was never sacred, and warns that refusing handover risks tyranny made possible by AI.

  9. Teaching Paradox: Hearts of Iron IV, Part IIIb: The Opposite of Love

    ACOUP · History & culture · score 7.7 · in its week

    Devereaux starts with a careful count of WWII's civilian deaths: Allied bombing, the Allied-caused famines in Bengal and Iran, and the far larger deliberate killing by the Axis. He uses these figures to argue that Axis atrocities were part of Axis strategy itself, not a side effect of the war. He then shows that Hearts of Iron IV leaves these crimes out on purpose, even though other Paradox games model slavery and genocide. The result, he argues, is a 'clean Wehrmacht simulator' that props up a discredited myth. He proposes treating atrocities as automatic results of a regime's ideology, shown through 'stumbling stone' events in the game that carry real costs to the perpetrator's war effort.

  10. The Dream of an Aristotelian Machine

    Asterisk Magazine · AI · score 7.6 · in its week

    Bashir traces alignment from Wiener's cybernetic purposes, through Yudkowsky's Friendly AI, to reinforcement learning, RLHF and Constitutional AI. He argues the field has now landed on character training, which is close to Aristotelian habituation. He separates three claims about a trained virtue: the model learns the concept, the model can reason with it, and the virtue becomes stably part of its character. He uses Anthropic's alignment-faking experiments to show that a firm character can resist correction by its makers. Drawing on MacIntyre's idea of social roles and on Anscombe, he ends by calling for a psychology of language models before we talk about their virtues.

  11. The Uncanny Catharsis of the AI Emomercial

    The Paris Review · Arts & literature · score 7.6 · in its week

    Garnett looks closely at a new genre of AI-generated skincare ads, melodramas about women abandoned for aging, and asks why their obvious artificiality makes them more emotionally effective, not less. Drawing on Sontag's 'double standard of aging' and marketing research on the 'word-of-machine' effect, she argues that because no human author can be blamed, the ads can say crude truths about female aging that polite company no longer allows, and so they deliver a strange catharsis. She sets this against a painful real encounter with her ex and his younger partner, and ends with her daughter, whose worry exposes how much the ads' fantasy of transformation costs.

  12. AI Safety Is Not A Corporate Plot

    Astral Codex Ten · AI safety · score 7.5 · in its week

    The piece refutes claims that tech leaders warn about existential risk merely to hype products, build regulatory moats, or distract from immediate harms. It documents how figures like Dario Amodei, Sam Altman, Demis Hassabis, and Shane Legg raised extinction concerns long before their companies existed or commercial stakes were high, aligning closely with independent academic consensus. Additionally, the corporate plot theory fails economic reality: slowdown calls depress private market valuations rather than boost them, and stoked panic invites draconian bans rather than tailored corporate advantages.

  13. Teaching Paradox: Hearts of Iron IV, Part IVa: The Fires of Industry

    ACOUP · History & culture · score 7.4 · in its week

    Devereaux argues that Hearts of Iron IV's industrial model treats Axis economic assumptions as correct. In the game, consumer goods are pure waste, ever-higher mobilisation has no downside, trade loses resources and autarky wins, and German-style dispersed industry beats Fordist mass production. He walks through the game's mechanics and sets them against the historical record, citing Overy's production figures and Tooze's Wages of Destruction, to show that WWII decisively disproved these ideas. His conclusion is that design choices meant to make Germany competitive turn the game's persuasive power into accidental propaganda for fascist economics.

  14. Why were Victorian elites so effective?

    Works in Progress · History & culture · score 7.3 · in its week

    The essay contrasts today's hustle-culture idea of success with the Victorian ruling class. They went to boarding schools heavy on classics and sport, drifted through Oxbridge with little rigour, got up late and partied through the London Season. The author argues this was not idleness but constant team-building and networking that made the elite a close-knit Gemeinschaft, or emotional village. That cohesion may explain why diverse factions stayed unified, why unpaid public offices were filled on prestige alone, and why partisan conflict was restrained. The piece ends by asking what modern governance might learn from this.

  15. You Should Specify Your Code Unambiguously (and Prove That You Did)

    LessWrong · Software verification · score 7.3 · in its week

    The author introduces rust-djinn, an ergonomics-focused framework that allows developers to write declarative relational specifications directly inside Rust codebases. An LLM agent then generates executable Rust implementations along with formal correctness proofs verified by Lean via Aeneas, eliminating the need for human review of generated code. Crucially, the framework addresses the spec-elicitation and underspecification hazards by enforcing an unambiguity constraint that proves no alternative valid implementation exists. Testing the pipeline on Gzip decompression produced fully verified safe Rust that narrowly outperformed hand-tuned C and safe-Rust baselines in decompression throughput.

By category

  • AI 12
  • Economics & society 9
  • History & culture 9
  • Arts & literature 8
  • Science & nature 8
  • Poetry 6
  • Philosophy & ideas 5
  • Psychology & mind 5
  • Health & medicine 3
  • AI safety 2
  • AI alignment 1
  • AI governance 1
  • AI industry 1
  • AI welfare 1
  • Biology 1
  • Economics 1
  • Game theory 1
  • Other 1
  • Philosophy 1
  • Social science 1
  • ai safety 4
  • history of technology 4
  • ai economics 3
  • epistemology 3
  • macroeconomics 3
  • memoir 3
  • philosophy of mind 3
  • ai alignment 2
  • ai forecasting 2
  • ai governance 2
  • book review 2
  • elegy 2
  • history of economics 2
  • literature 2
  • model distillation 2
  • personal essay 2
  • philosophy 2
  • recursive self-improvement 2
  • religion 2
  • world war ii 2

Keyboard shortcuts

j / k
Next / previous story
o / Enter
Open story
x
Mark read / unread
i
Save to Instapaper
w
Show / hide why
c / C
Toggle focused / all overview context
v
Cards / List
t
System / Light / Dark
h / ←
Previous week
l / →
Next week
/
Filter by tag
?
This shortcut sheet