What I'm paying attention to.

When the Proof Checks Out and Nobody Knows Why

Why one study can't settle whether coffee is good for you

What a Model Reasons About When It Can't Check the Date

Two or three points on a 70-point scale: what Alzheimer's symptom drugs buy

Simon Willison walks through Python 3.15 highlights, including explicit lazy imports (PEP 810), unpacking in comprehensions (PEP 798), and a new built-in sampling profiler called Tachyon. He also notes that dropping 3.10 means he can start using 3.11 features in his open source libraries.

Simon Willison asked Claude Opus 5.5 to design a text music format and build a player, aiming for Secret of Monkey Island quality. The result leaned heavily into the Monkey Island theme but was surprisingly good, and he wonders if music composition is a newly emerged capability like 3D graphics, though he says it would need careful experiments to confirm.

Highlights Michael Lynch's writing advice for software bloggers: avoid meandering intros, don't misjudge what readers already know, don't lean on links instead of explaining terms, and drop the stiff formality. Willison stresses writing the way you talk, since AI-delegated writing is making blogs bland.

Matthew Green puts a 1% chance on us living in "Minicrypt" (a world where public-key encryption is impossible) and 15% on losing functional confidence in current public-key algorithms. He warns that AI produces surprises far faster than people can replace standards, so preparation has to happen in advance.

Announces an expanded Atlassian and OpenAI partnership that brings GPT-6 family models to power agents across Atlassian's platform and Rovo, using its Teamwork Graph for company context. More than 3,000 Atlassian developers already use Codex, and new CLI plugins connect ChatGPT and Codex to Atlassian workflows.

An OpenAI customer story on Oracle's use of ChatGPT Work and Codex, with 130K active ChatGPT users and 95K+ Codex users. It reports a 98% drop in talent acquisition research time, with recruiters building a market intelligence tool that replaces 2 to 4 days of work, and an Applications Lab ontology that turns plain-language questions into SQL.

OpenAI explains how it's responding to the EU AI Act's machine-readable text identification rule with a watermark called textGrain, which adds an invisible statistical signal to word choices. API customers can opt in, ChatGPT and Codex output in the EU will be watermarked, and detector access is limited to approved researchers. It also acknowledges limits like false positives and false negatives.

OpenAI reports early usage figures for ChatGPT for Teens: teens with access sent about 2.7 million more learning-related messages, spend under 15 minutes a day on average, and less than 2% exceed three straight hours. It also announces a coming College Planner for applications, deadlines, and financial aid.

Argues that the debate over cyber risks of open-weight models is stuck in a lose-lose framing, where bans wouldn't stop bad actors and would hurt American competitiveness while raising long-term cyber risk. It criticizes Anthropic's GLM-5.3 report for skipping questions like what a ban would do and why Chinese labs consider these models safe to release.

Google releases EmbeddingGemma 2, a 740M-parameter Apache 2.0 embedding model that puts text, code, images, video, and audio in one shared space. It claims top scores among sub-1B multimodal embedders, a 9.92-point code improvement, an 8K context window, and on-device footprints as small as ~191MB for text-only use.

Anthropic launches the Cyber Mission, starting with a Critical Infrastructure Defense Program for power, water, and transportation operational technology, and an OSS Scanner offering free regular security scans to open-source projects. It says finding vulnerabilities is now easy but verifying, prioritizing, and fixing them remains hard, and notes Project Glasswing has been merged into the Cyber Verification Program.

Anthropic commits $150 million over three years to the federal Genesis Mission, making Claude available to more than 15 agencies including NASA, NIH, and NSF. The commitment covers credits for several hundred research projects, partnership on priorities like fusion and quantum computing, and training for scientists.

Anthropic summarizes its 2026 Usage Policy update, effective November 12, which mostly clarifies existing rules for Claude's longer autonomous work. Changes include a consolidated section on deceptive campaigns and fake activity, a refocused elections section, clearer requirements for high-risk health and finance uses, and controls for physical actions.

Anthropic merges Project Glasswing and the Cyber Verification Program into one program with three access tiers, giving vetted security professionals reduced cyber safeguards on its most capable models. The Defense Access tier covers work like incident response, malware reverse-engineering, and vulnerability validation for defenders such as critical infrastructure operators and open-source maintainers.

Argues that AI will soon be superhuman at distributed GPU engineering, which will speed up experimentation and make good ideas more valuable than execution. That doesn't mean models will become dramatically different in kind, and much of the near-term gain comes from scaling inference-time compute, which will mainly help the technology spread through the economy.

A Chinese-American writer offers impressions, drawn from family and Chinese media, of how Chinese social reality differs from the Western one, for people trying to communicate AI safety ideas in China. The four differences are an omnipresent social media culture, shame and face as the main enforcers of virtue, cynical 'dark world' thinking, and nationalism rooted in collective insecurity.

Profiles economist and journalist Barbara Ward and her 1957 lectures at the University College of Ghana, Five Ideas that Change the World, which traced a path from nationalism through colonialism, industrialism, and communism to internationalism. It presents her as a thinker who imagined planetary political, economic, and environmental justice, while noting she uncritically admired the legacies of empire.

A short documentary by Iranian filmmaker Yaser Talebi about an elderly couple in a northern Iranian village whose son has been declared brain dead after an accident. It follows them as they decide whether to authorize organ donation, and their search for the recipient of their son's heart.

An eight-minute 3D animation from the City 3D Timelapse channel traces Manhattan from the Lenape homeland through the Dutch colony and the Revolution to today's skyline. It shows the industrial-era spread of population and the recent rise of supertall towers, made without AI.

A psychotherapist explains that problem debt is as much an emotional issue as a financial one, often tied to shame, avoidance and fear of judgment. Facing it takes courage, curiosity and willingness to ask for help, and the piece offers suggestions for building those emotional resources.

Presents a video by artist Campbell Walker (Struthless) offering 10 ideas, such as turning a mistake into art and answering famous questions, to get past fatigue, overwhelm and perfectionism. It frames a regular creative practice as good for wellbeing and as an escape from doom-scrolling.

Using Iris Murdoch, argues that moving on from a breakup is active work rather than a matter of waiting for time to heal. It involves acquiring new objects of attention and reorienting the unified valuational perspective you held while in the relationship.

The author found self-directed learning from online resources left them disengaged, and concluded the problem was treating learning as solitary. Drawing on Jonathan Rose's history of working-class autodidacts, such as Scottish miners' libraries and mutual improvement meetings, it argues that learning works better in a community.

Encyclopedia-style overview of the Cambridge Platonists, including Henry More and Ralph Cudworth, seventeenth-century thinkers who defended eternal moral principles, free will, and the superiority of mind over sense knowledge. They drew selectively on Plato and Plotinus while working on materialism, determinism, consciousness, and a practical ethics.

A survey of Aristotle's Categories and the doctrine of categorialism, which frames much of his work from the Physics to the Metaphysics and shaped thinkers from Plotinus to Heidegger. It stresses that experts disagree on basics, including what the categories classify, how they relate to hylomorphism, whether Aristotle wrote the work, and whether the scheme is defensible.

An overview of Samuel Alexander (1859–1938), the Australian-born philosopher and British Emergentism proponent. His Space, Time, and Deity (1920) describes a hierarchy of levels, from space and time up through matter, life, mind, and deity, and the entry covers his influence and the recent revival of interest in his work.

An entry on metaphysical explanation, the "because" and "in virtue of" claims metaphysicians make, such as Socrates being snub-nosed because he has that property, or a cloak having a color because it's green. It distinguishes these from scientific and mathematical explanation, notes the recent surge of interest, and traces the topic back to Aristotle's Metaphysics.

Argues the US is losing the long geopolitical game to China because it's stuck in nostalgia and short-termism while China builds institutions and infrastructure meant to project power for decades. The real contest, it says, isn't tariffs, military hardware, or AI, but who offers the more compelling vision of the future and can make it real.

Draws on N. Katherine Hayles's argument that we're already post-human, since humans have never been autonomous and are now deeply enmeshed with computational media in "cognitive assemblages" of distributed decision-making. The piece asks what that means for the question of whether we can control AI, and how close it puts us to a Singularity.

Uses Moonshot AI's Kimi K3 release and Xi Jinping's July 2026 WAIC speech endorsing openness to argue that China isn't about to restrain open-weight AI development, contrary to a common Western expectation of a sudden "Mythos moment" policy shift. It describes China's strategy of many labs releasing open models, backed by compute vouchers, with a layer of smaller firms specializing them, and says American rhetoric badly misreads the reality.

Starting from BadClaude, a joke tool that lets you "whip" Claude to work faster (and its encouraging counterpart GoodClaude), the piece asks whether abusing AI matters. Using Kant's argument about cruelty to animals and the film A.I., it argues the harm falls on us, since people are bad at compartmentalizing cruelty.

Looks at an Econ Journal Watch symposium on Greg Clark's book arguing that genetics drives social outcomes long attributed to nurture, with data spanning 1600 to 2026. Even the critics concede that twin and adoption studies undercut the claim that parental investment matters much, and the author argues environmental models that mimic genetic inheritance are accommodating the evidence rather than predicting it.

Contends that Victorian Britain's elites, who slept late, drank heavily, partied and got a non-STEM education, were highly successful by their own measures. The author explores whether their focus on soft skills, team building and networking worked well, which would suggest the austere modern success routine isn't the only path.

Argues that effective altruism is directionally right and that a bit more of it would help at the margin, though not followed to extremes. The case rests on animal welfare being neglected, AI safety deserving more attention, and foreign aid being a tiny share of the federal budget, so saving lives abroad beats donating to Harvard.

Explains Robert Hall's Nobel-winning work, centered on his finding that combining permanent income theory with rational expectations implies consumption follows a random walk. Nothing known today should predict changes in consumption tomorrow, and his test worked without any model of how people forecast income. He found the data surprisingly consistent with this.

Paper arguing that California has lost domestic migrants every year since about 1990 because the location rents from its amenities have been captured through a progressive income tax and housing restrictions. IRS data for 2022-23 show out-migrants outnumbered in-migrants by 40 percent overall and by 64 percent among households earning $200,000 or more.

Argues that the Trump administration's tariffs and threats toward Canada and Mexico are undermining the friendly North American neighborhood that is the foundation of US global power. Unlike China, Russia or India, the US has two benign neighbors and open oceans, and decades of presidents since Reagan invested in that partnership.

Mosquito-borne diseases are returning to the US, with almost 4,000 dengue cases in 2024 and locally acquired malaria in 2023. The piece argues that genetically engineered male mosquitoes carrying a lethal gene could suppress local populations and eliminate these diseases, but the technology has been stuck in regulatory limbo for fifteen years.

Overview of Works in Progress Issue 26. The cover story argues that China's growth comes from mayors acting like venture capitalists, running local growth machines built on cheap capital, a huge market, and intense competition. Other pieces cover lovable city design, the Sun Belt, mosquito eradication, the Bronze Age collapse, and Leninism's flaws.

Sudan's war has become a regional proxy conflict, and the Horn of Africa and Arabian Peninsula now form one linked strategic theater around the Red Sea. The Saudi-UAE split, the Iran war, and rising Ethiopia-Eritrea and Sudan-Ethiopia tensions raise the risk of a wider war.

Argues that military victory is a myth because all modern wars end in negotiated settlements, with destruction serving to shape the bargain rather than annihilate the enemy. It points to the Iran, Gaza, and Ukraine wars as cases where this false belief started or prolonged conflicts, and warns of the same risk over Taiwan.

Washington believes sanctions and its blockade will force Iran to concede, but Tehran sees itself winning. Iran's theory of victory is to hold out, keep striking US bases, and disrupt energy markets until the US lifts sanctions and leaves the Persian Gulf, helped by a population that has largely rallied around the flag.

Announces the winners of a War on the Rocks essay contest on wartime technologies that spread into civilian life, led by the blood bank, Blanchard's lathe, and Kevlar. A common theme is that government's role in innovation has been as an early buyer and standard-setter, committing to things the market didn't yet value.

Argues that the US framing of China as strategic competition has become an end in itself, driving worst-case assumptions and risking crises, arms races, and war over Taiwan. Opening with a crisis simulation where a reassuring carrier move alarmed the Chinese team, it calls for mutual reassurance and lasting agreements that build a stable balance.

A weekly digest of Ukrainian media commentary. This edition includes a Fokus piece on demining, which estimates over a fifth of Ukraine may need clearing, a job that could take 10 to 20 years after the war and that drones, satellites, and AI could speed up.

A podcast conversation with executives from Ondas Sentinel and Forterra on what it takes to field autonomous ground vehicles in combat. Topics include lessons from Ukraine and Gaza, operating without GPS and with resilient communications, and air-ground teaming.

Podcast conversation with Jaan Tallinn, Skype co-founder and long-time AI extinction-risk advocate who also funded DeepMind and Anthropic. Topics include his updated risk estimate, his case for an immediate moratorium on frontier training, hardware-based verification for global AI governance, and why he asked Dario Amodei to collect competing term sheets.

When rumors spread that OpenAI had proved the unique games conjecture, MIT's Dor Minzer and two students rushed out a 95-page paper on a milestone result about a closely related question. OpenAI then announced its proof plus 376 other results on October 6, and the story covers how mathematicians reacted to both.

Explains Conrad Waddington's mid-century picture of development as a ball rolling down a branching landscape of valleys, where cells become canalized into a limited set of fates. It argues this metaphor, which arose because genetics didn't explain how genes build tissues reliably, is proving a useful conceptual scaffold for modern developmental biology.

Traces the attention mechanism back to 2014, when intern Dzmitry Bahdanau, working with Yoshua Bengio and Kyunghyun Cho, fixed neural machine translation's fixed-memory bottleneck on long sentences. It tells how that idea worked on the first attempt and led to the AI boom, and asks whether attention is central to intelligence.

Peter Attia talks with maternal-fetal medicine specialist Kathryn Gray about how pregnancy changes the body, and about gestational diabetes, preeclampsia, HELLP syndrome, and postpartum depression. They also cover labor and delivery trade-offs, stillbirth, cell-free DNA screening, nutrition, and what pregnancy complications reveal about long-term maternal health.

Peter Attia revisits positions he's changed over eight years: keto, extended fasting, metformin (he no longer takes it), rapamycin, whole-body MRI, and liquid biopsies. He also covers shifts in his exercise approach, including Zone 2 training for people short on time and moving away from heavy barbell squats and deadlifts.