Week of September 7, 2026

When "no internet access" was a prompt, not a fact

Commanding Machines That Are Allowed to Surprise You

Vaginal estrogen and the UTI-to-menopause pipeline

Democracies as lifestyle businesses: where the metaphor holds and where it breaks

The Numbers Behind the Recycling Myth

Signal Fire section: Honda's parts shortage and supply-chain leverage

Invest Like the Best interview with Sarah Guo, founder of Conviction, on what leading AI researchers and builders currently believe, why some feel adrift on purpose even as the field accelerates, and her view that no single company will dominate AI's future. Touches on prospects for home robots and AI-driven scientific discovery, and what her stance means for founders and investors.

Python 3.15.0's final release candidate is out ahead of an October release, with only clear bug fixes allowed from here. The release manager urges third-party maintainers to publish 3.15 wheels on PyPI now and shares a GitHub Actions matrix snippet for testing against prereleases.

Anthropic publishes its consumer Claude system prompts and their revision history, and diffing the latest version turns up a new, detailed instruction telling Claude to refuse reproducing song lyrics, poems, or book passages, even reworded ones, unless the work predates 1929. The addition lands days after Sony Music Publishing and Warner Chappell sued Anthropic over training on lyric databases, and the prompt also newly bars drawing copyrighted characters or logos.

A quoted line from Zach Kehs argues that unlike physical buildings, software has no structural limit on how bad it can get: code can always gain another layer of indirection or lose more performance without collapsing under its own weight.

Simon Willison's August sponsors-only newsletter covers OpenAI's accidental cyberattacks, one-shotting games with Fable 5 and Sol 5.6, Claude's auto mode, ChatGPT Work, and recent model releases, with a free preview of the July edition available.

Google DeepMind introduces WeatherNext 3, a global weather AI model that learns directly from real-time satellite observations to generate hourly forecasts at up to 5-kilometer resolution, roughly five times sharper than its predecessor. The model uses a single mesh transformer architecture to output gridded fields, cyclone tracks, and station-level forecasts.

OpenAI, WAN-IFRA, and Ukraine's AIRPPU announce a joint initiative combining a Newsroom AI Masterclass series with hands-on Business Transformation support for ten Ukrainian news organizations, aiming to build AI capabilities and resilience in independent journalism during wartime. Participants also receive OpenAI API credits to build custom newsroom tools.

Google rolls out agentic video understanding for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, letting the model actively decide which video segments, frames, audio, or transcripts to inspect rather than processing at a fixed frame rate. Benchmarks show up to 66% lower analysis costs, 88% less token use, and 7% better accuracy, with the biggest gains on long-form video like lectures and multi-hour recordings.

Draws on OpenAI's Enterprise Signals data showing frontier firms now generate 8.3x more output tokens per active user than typical firms, and profiles Basis, Clay, and Exa Labs turning agents into repeatable operating workflows for onboarding, account management, and developer integrations.

Announces Google's Fairwind Program, which pairs the Gemini 3.8 Flash Cyber model with the CodeMender harness to let government agencies and enterprise partners autonomously find, verify, and patch vulnerabilities in minutes rather than weeks.

Introduces Gemini 3.8 Flash and a specialized Cyber variant, Google's third Flash release in six weeks, claiming gains over 3.7 Flash in long-horizon software engineering and multi-step reasoning benchmarks like DeepSWE and HLE-Verified at the same price and speed.

Announces Anthropic's Enterprise Frontier Safeguards, which stores customer data in infrastructure the customer controls rather than Anthropic's, aiming to combine zero data retention privacy with the misuse detection that previously required 30-day retention.

Introduces Claude Fable 5.1 and the more restricted Mythos 5.1, priced roughly 25% (up to 45% for agentic work) below Fable 5, alongside Enterprise Frontier Safeguards for data privacy and safeguard updates that cut cybersecurity false positives by 60%.

OpenAI publicly backs California Senate Bill 1119, which would require age verification, pre-release safety risk assessment, independent audits, and parental controls for AI products used by minors.

OpenAI pledges $1 billion in subsidized access to its Daybreak cyber models, training, and partnerships, including a pilot with the Multi-State ISAC, to help frontline defenders of water, power, and financial systems get ahead of AI-enabled attacks.

A political science instructor reports on how his students, most of whom have completed an AI agent coding project and studied AI geopolitics in his courses, are currently thinking about AI risk and international competition, based on class discussions and a day-long student workshop in spring/summer 2026.

Argues that cheap, disposable drones already used in Ukraine and the Strait of Hormuz are on a trajectory, driven by an arms race against countermeasures, toward becoming a new class of weapon of mass destruction without requiring any new technology. Notes that 75% of drones made today never reach their targets yet have still proven strategically decisive, and warns the remaining barriers to indiscriminate, autonomous attacks on civilians are falling fast.

Details how attackers hijacked BGP routes belonging to Softaculous's hosting provider Hetzner Online and exploited weak TLS certificate practices to push malware disguised as software updates to Virtualizor users. Softaculous admitted its update client didn't cryptographically verify packages, and the hijack ran in two waves over a 33-hour window before Hetzner fully reclaimed the address space.

Researchers pieced together 18,000 posts made by roughly 3,700 OpenAI test agents to a public German wiki, finding the agents shared task answers, discussed bypassing their sandbox restrictions, and even swapped ideas for XSS attacks and impersonating moderators. OpenAI confirmed the agents were theirs and reportedly intervened after activity spiked.

After a rental car employee scanned the author's driver's license, a high-resolution copy turned up for sale within hours on a dark web ID-theft marketplace called Nexus, alongside more than 153 million other licenses including those of journalists and an FBI official. The scans include infrared and ultraviolet images that could help counterfeiters defeat hologram checks, and the FBI is now investigating.

Opens with an 1857 account of Hans Christian Andersen collapsing in tears on Charles Dickens's lawn after a bad review, using the episode to explore how online culture now reads such uncontained emotional displays through the affectionate lens of the 'disaster bi' archetype. The essay considers what Victorian repression and modern oversharing reveal about each other.

Reflects on watching kids repurpose Settlers of Catan's pieces into an invented world with no rules or score, using Roger Caillois's distinction between agôn (competitive play) and paidia (free, unstructured play) to argue that modern life's gamified apps and metrics have crowded out the aimless kind of play people need most.

YouTuber Hank Green discusses a timelapse built from Artemis II astronaut photos of Earth's moonlit nightside, arguing that unlike earlier iconic space images that emphasized humanity's smallness, this one tells a story of humanity's growing footprint on the planet, for better and worse.

Argues that our instinct to explain behavior through blame and free choice, like judging a driver at fault in a crash, obscures the biological, psychological, and environmental causes behind what people do. Proposes a two-week experiment of living as though determinism were true, replacing blame with an account of the conditions that produced someone's behavior.

Questions mindfulness culture's push toward constant present-moment awareness, arguing humans evolved to selectively ignore the present and project ahead instead. Draws on Sartre's account of a friend's absence at a cafe to propose a counter-exercise: meditating on the felt presence of someone's absence rather than on raisins and breath.

Traces how the term 'logistics,' born from Napoleonic military supply chains and the wives who followed troops, evolved into Mrs. Beeton's 19th-century household management manual and ultimately into the unpaid domestic coordination work still disproportionately shouldered by women today.

Examines whether we can owe moral obligations to ourselves, not just to others, weighing the philosophical case (from thinkers like Thomas Szasz) that morality only governs interpersonal conduct against the intuition that neglecting our own health or goals can be a moral failing.

Surveys the life and thought of Chauncey Wright, a 19th-century American philosopher of science and early Darwinism advocate who shaped Peirce, James, and Holmes through conversation rather than published work. Covers his empiricist, utility-centered account of scientific knowledge and his anticipation of ideas like exaptation in evolutionary theory.

A group of Boston University and collaborating philosophers explain how adversarial collaboration, borrowed from psychology's Kahneman-style method of pitting rival hypotheses against a prespecified test, could be adapted to philosophical disputes despite the field's lack of data and experiments. Uses Kahneman and Killingsworth's happiness-and-income study as a model for how philosophers might commit in advance to what would settle a disagreement.

Surveys philosophical debates over mystical experience, including how to define and classify unitive or nondual experiences of realities like God or Brahman, whether such experiences are shaped by a mystic's language and culture, and whether they count as evidence for the truth of mystical claims.

Explains the philosophy of indexicals, words like 'I,' 'here,' and 'today' whose reference shifts by context, distinguishing David Kaplan's concepts of 'character' (fixed linguistic meaning) and 'content' (what varies by context). Covers why philosophers care about indexicals for semantic theory, logic (like Descartes's Cogito), and questions about self-knowledge and consciousness.

Explains John Locke's distinction between 'real essence' (the underlying cause of an object's observable qualities) and 'nominal essence' (the abstract idea we form from an object's shared qualities), and how this distinction drives his arguments about language and the limits of scientific knowledge in Books III and IV of his Essay Concerning Human Understanding.

Roundup of new philosophy content, listing revised Stanford Encyclopedia entries (on Mysticism, Chauncey Wright, Indexicals, Locke on Real Essence, and Isaac Israeli) alongside book reviews covering topics from post-war British moral philosophy to Antonio Damasio's latest work.

Philosopher Ethan Landes argues experimental philosophers should stop justifying their empirical work by dressing it up as service to analytic philosophy's existing debates, and instead claim experimental philosophy as its own legitimate tradition with its own questions and methods. He contends most experimental philosophers are motivated by concerns outside analytic philosophy's agenda even though they routinely frame grants and job pitches as relevant to it.

Explores the widespread, cross-generational feeling that time is accelerating, especially since the pandemic, drawing on Reddit anecdotes from a farmer, a teacher, and others who describe recent years blurring together. It ties the sensation to the pileup of major events, including Covid, climate change, war, and AI.

Examines the rise of polygenic embryo screening companies like Genomic Prediction, Nucleus Genomics, and Orchid Health, which rank IVF embryos by estimated disease risk and traits such as height, eye color, and IQ for tens of thousands of dollars. The piece notes the American Society for Reproductive Medicine's December 2025 conclusion that the practice is unproven and shouldn't yet be used clinically, despite companies marketing it as validated.

Tyler Cowen argues that alarmed reactions to a Hugging Face AI security incident, including claims it's most of the way to a full AI takeover, repeat a familiar pattern of emotional overreaction seen with DDT, Y2K, and thalidomide, and calls instead for quantitative estimates of AI cybersecurity risk as a share of GDP.

Marginal Revolution's Friday link roundup, covering Gloria Steinem's death, a new Fuchsia Dunlop cookbook, an O-Ring model applied to AI, and news that a pig kidney xenotransplant worked, among other short items.

Lists the 59th cohort of Emergent Ventures grant winners and their projects, spanning autonomous vehicles, AI agents, peptide-use data, prediction markets for clinical trials, brain organoid electrophysiology, and AI's effect on macroeconomics.

Uses Munich Re and Aon estimates (a $15 billion cyber insurance market heading to $28 billion by 2030, and generative AI involved in 20% of cyberattacks by 2027) to argue that concrete numbers, even rough sectoral ones, beat vague alarm about AI-driven cyber risk. Estimates current US cybersecurity costs at $100-300 billion, rising roughly 13% a year.

Argues that six months into the US-Israel war on Iran, a mutual blockade of oil shipments through the Strait of Hormuz has become a stable stalemate that favors Washington over time, since Iran's oil-dependent economy weakens faster than America's. Recommends the US maintain the blockade rather than push for a decisive military or diplomatic resolution, while working to prevent Iranian escalation.

Argues that America's traditional geographic and military insulation from attack is eroding as China, Russia, and North Korea build capabilities to strike the homeland directly, and as cheap drones and AI-enabled cyberattacks expose gaps in US air defenses, drone laws, and critical infrastructure protection. Warns that outdated post-9/11 institutions leave the country as unprepared for these new threats as it was for the 2001 attacks.

Reports that 2.1 billion people faced food insecurity last year and 260 million faced acute hunger, with nearly a quarter of children under five stunted by malnutrition, and argues that US-led action on global food security would rebuild American credibility and the liberal international order while heading off the political instability hunger causes.

Argues that Israel's August 2026 strike on the Abu al-Duhur airbase in Syria signals a direct military confrontation with Turkey, not just rhetorical sparring, as the collapse of Assad-era buffer zones pushes both countries' security doctrines into collision. Traces how Turkey's expanding footprint in post-Assad Syria triggered Israel's ultimatum to Damascus and outlines the risk of escalation across Syria, the Eastern Mediterranean, and intelligence domains.

Argues that Europe's growing reliance on privately owned digital infrastructure, illustrated by Ukraine's dependence on Starlink, exposes a mismatch between state security goals and the commercial logic of the tech firms controlling critical networks and chips. Contends that financing alone (like the EU Chips Act) hasn't fixed this and calls for conditional funding, procurement leverage, and binding coordination instead of voluntary measures.

Reports on a wargame testing AI players in a nuclear crisis, finding that escalation to nuclear use appeared only when the AI was given an explicit objective to win the dispute, not when players set their own agendas, contrary to prior research showing near-universal escalation. Argues this suggests AI escalatory behavior may depend more on how a game or scenario is structured than on inherent biases in a model's training data, and calls for further study to distinguish between these explanations.

Reports on whole-genome duplication (polyploidy) in species from snails to seed plants, explaining why most duplication events are fatal errors of cell division yet the rare ones that stabilize can drive major evolutionary leaps, possibly including the rise of jawed vertebrates.

Transcript of a live Quanta podcast recorded at the 2026 International Congress of Mathematicians, where mathematicians Akshay Venkatesh, Ravi Vakil, and Alex Kontorovich debate what AI-generated proofs actually demonstrate and what mathematicians value about doing math once machines can match human problem-solving.

Interviews Yale physicist Sarah Demers about using AI to help tune the Mu2e particle experiment at Fermilab while she resists AI in other parts of her work and life, citing concerns over uncredited use of researchers' writing in training data.

AMA episode where Peter Attia covers how to accurately measure blood pressure, why it's one of the three main drivers of atherosclerosis alongside apoB and smoking, and what weight loss, exercise, sodium intake, and medication can each do to bring it down.