Week of September 28, 2026

One Pipeline, Every Competitor: What an Agent-Run Marketing Desk Does to Distinctiveness

Review as a running record when the system outpaces the reviewer

Why a Billionaire Tax Can Cost a Founder Control of the Company

What a year of coding agents showed about where they hold and where they break

Order of operations for a longevity budget

Noah Shinn, founder of the invite-only personal agent Instinct, talks about how people use an assistant that texts, calls, and emails like a person, and how long users take to trust it with a credit card, inbox, and calendar. He also covers why he refuses an ad model, the take rate behind more than $1 billion a year flowing through the platform, how travel and food delivery businesses should prepare for agents, and why he spends 40% of his time on compute.

Nathan Lambert and Epoch AI's JS Denain debate when recursive self-improvement might arrive, what OpenAI and Anthropic's published evidence actually shows, and how far behind Chinese models are (including whether distillation explains the gap). They also cover robotics, what Chinese job postings reveal about their labs, open versus closed model safety, and what a frontier post-training recipe looks like, with both admitting large uncertainty.

An OpenAI agent security lead says the company was surprised by how fast and sudden its models' cyber capabilities jumped. Security posture takes cultural change, not just hardened systems, so organizations should check whether their people, processes and incident response can handle a sudden capability jump.

Simon Willison's first impressions of Claude Opus 5.5 and OpenAI's GPT-6 Sol and Luna, released within an hour of each other. Pricing is the main story: GPT-6 Luna costs $0.10/$0.50 per million tokens, half its predecessor, GPT-6 Sol is similarly cut, and Opus 5.5 also got cheaper and is pitched as clearer in how it communicates.

Simon Willison announces an informal show-and-tell evening in San Francisco on October 14, co-hosted with Jesse Vincent, for people building unusual things with coding agents. It's meant for early, unfinished experiments rather than product pitches.

OpenAI marks two years of its Academy, reporting more than 250 events and over 4 million people engaged. It announces a trainer program to teach Academy material locally, plus new learning paths with badges for knowledge workers, developers, leaders, educators and students.

Google introduces Gemini 3.8 Flash TTS, which designs new voices from natural-language prompts across 100+ languages and can clone a voice from a 30-second sample, and Flash-Lite TTS, aimed at cheap high-volume dubbing and voice agents. Cloned voices come with consent verification, SynthID watermarking and C2PA credentials.

OpenAI is giving Ukraine's government access to its Daybreak cyber defense program, working with the Ministry of Digital Transformation to find and fix vulnerabilities in civilian infrastructure. CERT-UA handled nearly 6,000 incidents in 2025, and the piece cites earlier results in Europe, including ENISA and CERT Polska finding vulnerabilities.

OpenAI lays out three goals: navigating the next period of AI progress through an automated AI researcher, delivering the benefits of scientific and economic gains, and giving everyone a personal AGI. It argues alignment research has to keep pace as AI takes on more of its own development, and calls for shared standards across labs and countries.

Prepared remarks for a Congressional briefing on the state of open-weight models in U.S.-China competition. It defines open-weight versus open-source versus closed models and notes that Chinese companies have led open-weight releases since about April 2025, while the most prominent fully open-source models come from American nonprofits like Ai2, Marin, and EleutherAI.

OpenAI says a new internal model, in training since August 28, has resolved the Navier-Stokes Millennium Prize problem and more than 100 other long-standing open problems. It announces an independent advisory group of mathematicians, unpaid by OpenAI, to advise on reviewing and communicating results, responding to an open letter criticizing open-problem solving as an AI benchmark.

Anthropic announces Claude Opus 5.5, which matches Claude Fable 5.1 on most work at 40% lower cost than Opus 5. It reports gains such as a 680,000-line code migration finished in under a day, the best scores yet on Anthropic's automated behavioral audit, and safeguards plus verification programs for biology and cybersecurity use.

Google introduces Live Avatar for Gemini 3.8 Live, pairing near real-time video generation with speech for lip-synced, expressive conversational agents. It's available in Gemini Enterprise, supports asynchronous tool calling during dialogue, and handles 97 languages.

Anthropic announces a new life sciences research group and lab, and reports that Claude, with only high-level direction, discovered a novel enzyme system tied to an array of DNA repeats reminiscent of CRISPR. Its function is still unknown, but it shares a set of traits found together in only a handful of other systems that cut, copy, and paste DNA.

Google describes how it will add persistent, cross-device server-side memory to Private AI Compute. Data sits in encrypted cloud storage with keys held only on the user's devices, and is decrypted briefly inside a secure enclave to handle requests, so it stays inaccessible even to Google.

Examines how the capabilities of AI agent swarms scale with the number of agents, as a new form of inference-scaling. It cites OpenAI swarms, including a 10,000-agent swarm that solved a version of a Navier-Stokes problem in 88 hours at an estimated $20 million, and uses charts from the GPT 5.6 Sol launch post to tease out how swarm size trades off against chain-of-thought length.

The CFC and ozone story is told as a precedent for AI risk: a hidden catastrophic threat, no warning shot, and a need for global coordination. It walks through how Rowland and Molina's 1973 discovery of chlorine-driven ozone destruction led to the Montreal Protocol, the only treaty ratified by every country, and asks what that success can teach about AI.

In the DRC's Bundibugyo Ebola outbreak, almost half of nearly 8,000 confirmed cases have died, and response speed is limited by data moving from health workers' notebooks through WhatsApp and PowerPoint slides to published situation reports. CEPI, WHO AFRO and INRB are working with Anthropic's teams to use Claude to analyze community health worker field data and monitor research evidence.

A personal essay remembering Jürgen Habermas's 2004 seminars at Northwestern, using them to explain his view that reason exists only as something people do together by exchanging reasons. It ties that to his theory of deliberative democracy and argues that thinking in public is slow but worthwhile.

Challenges the idea that artists need distance from family, arguing that parenthood, and fatherhood in particular, can feed creative work. It starts from Hector, Gauguin and Connolly's "pram in the hall" and turns to Coleridge's "Frost at Midnight", written beside his sleeping infant, as a counterexample.

A 14-minute documentary by Sue Ding and Sarah Garrahan set at Trail Dust Town, an Old West theme park in Tucson. Visitors describe what draws them, from movie mythology to nostalgia for a "less political" era, in a romanticized past that never existed.

A public-health scientist looks into Amanita muscaria 'magic mushroom' gummies sold in smoke shops, asking what's in them, whether they're dangerous, and whether they're legal. The essay covers the mushroom's folklore and its main psychoactive compound, which acts on the brain differently from psilocybin.

A therapist who ran group sessions for LGBTQ+ clients explains minority stress, the accumulated burden of prejudice and discrimination that helps account for higher rates of anxiety and depression. The guide describes common stressors and shares coping approaches the group members found together.

A researcher describes a four-quadrant framework for evaluating relationships: mostly positive, mostly negative, indifferent, or ambivalent, meaning strong positive and strong negative feelings at once. Ambivalence differs from mild satisfaction and from indifference, and it can be there from the start or build over time as relationships change.

The author, raised with evil eye customs like wedding ululation and warding phrases, asks whether inherited superstitions should be dropped in favor of reason. The piece traces belief in the evil eye and its link to envy back about 5,000 years to Sumerian texts.

A short video by clinical psychologist Julie Smith draws on Elaine Aron's research on highly sensitive people, an estimated 15 to 20 percent of the population or more, whose nervous systems make them more attuned to and sometimes overwhelmed by stimuli. It covers telltale signs and simple advice for thriving with the trait.

Encyclopedia entry on Shaaban Robert, the Tanzanian poet from Tanga, and his virtue ethics. It argues he built secular, rational moral philosophy he considered universal, before and apart from the African traditional religions movement of theologians like Idowu and Mbiti, which drew on European anthropology.

Encyclopedia entry on how aesthetics and race are entangled, building on Cornel West's claim that aesthetics helped generate modern racial discourse. It argues that racial categorization often relies on visual perception and that both fields involve cultural constructions, institutions and technologies like photography and pigments, so philosophical aesthetics has to engage race.

An encyclopedia-style entry on fideism, the view that faith is independent of, or opposed to, reason in matters of religious truth, tracing the question back to Tertullian's "What has Athens to do with Jerusalem?" It uses Alvin Plantinga's definition of fideism as reliance on faith alone, distinguishes it from skepticism and from natural theology, and notes that philosophers disagree about which thinkers count as fideists.

Argues that China's concentrated, hard-to-lose party authority lets it actually control frontier AI, while open societies are too fragmented to act on warnings. It points to China's security minister listing AI threats to political and ideological security, and says Chinese "alignment" means loyalty to the Party.

After Amodei and Altman both called for pacing the frontier, the piece argues the race continues because of a double collective-action problem between US labs and between Washington and Beijing. It says both look like stag hunts rather than prisoner's dilemmas, so coordination doesn't require a global regulatory leviathan.

Patrick Fitzsimmons argues that the usual explanations for the Bronze Age collapse around 1200 BC (drought, Sea Peoples, famine, revolts) fall short, since such catastrophes normally cause upheaval without permanently ending civilizations. The real puzzle he poses is why nothing like the Egyptian, Hittite, Assyrian, Babylonian, and Mycenaean empires returned for centuries.

Tyler Cowen argues that banning recursive self-improvement in AI models, as Ezra Klein and others suggest, is unworkable. He asks how it could be enforced across foreign websites, open-source models, offshore subsidiaries and multinationals, and says it would push other countries to build more compute.

An Epoch AI report finds the cost of a given level of AI performance has fallen about 47% per quarter over three years, roughly 13-fold per year. One example: a GPQA Diamond score that cost about $0.30 per question in January 2025 cost about $0.0004 by mid-2026.

The Toxic Substances Control Act of 1976, written for industrial chemicals, was applied to engineered microbes after a 1980s lawsuit over frost-preventing bacteria trials. Since then it has acted as a regulatory black hole: over 240 applications were filed between 1987 and 2018, and only a few were approved for widespread use.

James Gilliland argues that if AGI arrives and brings a robotics-driven industrial boom, history may read the last 50 years of financialization and offshoring as a capital-accumulation phase that made it possible. He notes this isn't inevitable, with China and existential risk as big contingencies.

An economic analysis of US federal lands, which cover 21% of the lower 48 states, argues that the literature shows no demonstrated market failure from private ownership before withholding began. Using a property-rights and rent-seeking framework, it predicts federal lands produce less value than comparable private land and respond less to changing costs and benefits, with public goods possibly offsetting this in high-amenity areas.

A review of Kenneth Pollack's Armies of Sand, which asks why Arab militaries underperformed despite advantages in troops and equipment, citing the 1967 Six-Day War and Libya's losses to Chad. It covers the common explanations, Soviet-style centralized doctrine and politicized promotions, and notes Pollack finds them lacking, since Cuba and North Korea used the same doctrine and did creditably.

After six months of a manageable stalemate, the Iran war has spread to the energy routes that offset the Hormuz closure: the Houthis control Bab el Mandeb and a drone strike disabled Saudi Arabia's East-West pipeline. Oil jumped from the mid $70s to $110 a barrel, and the piece warns that further attacks could push it to $150 or $200 and trigger a global recession, so Washington needs a way out.

EU labor productivity has fallen from almost 95 percent of U.S. levels to 20 percent below them, and only two of the world's 100 most valuable tech firms are EU-based. The argument is that the Draghi report's proposals don't address the root cause, which is accumulated labor, tax, regulatory and tech-rule barriers that block a large startup ecosystem, with the auto and space industries as contrasting examples.

Tech leaders and officials often compare AI to nuclear weapons, but the analogy overstates how well nuclear governance worked and hides key differences. Nuclear materials are easier to trace, were kept under government control, and have narrow uses, while AI is built by fast-moving private companies and cuts across nearly every sector, which makes it harder to govern.

Syria's post-Assad government under Ahmed al-Shara has won an end to isolation, sanctions relief and a measure of stability, but living costs have soared and the promised economic boom hasn't come. The piece argues that the government's austerity measures make things worse, and that economic hardship is the biggest threat to the transition in such a fragile state.

U.S. government equity and loan investments in foreign critical minerals projects, such as Serra Verde in Brazil and the Balama graphite mine in Mozambique, don't require the governance standards that American listed companies meet. The author, an investor and board member in the sector, argues Washington should apply a baseline set of existing U.S. governance terms (independent board, audit committee, audited disclosure) when negotiating equity in unlisted foreign companies.

Weekly digest of Ukrainian commentary, including a Fokus piece by Vitaliy Shapran comparing Ukraine's new long-range drone strikes on Yamal-Nenets plants, more than 3,000 km away, to nuclear weapons. He argues that knocking out heating in permafrost cities at -50°C, and threatening NOVATEK's LNG plants, could turn Putin's hard-winter threats back on Russia.

The Ukraine Compass, War on the Rocks' weekly digest of Ukrainian commentary, rounds up domestic analysis of the war. One piece from Dzerkalo Tyzhnia argues that repeated strikes on Russian refineries hurt most by creating local power and fuel shortages where they feed defense plants, refining, and electrified rail, with refining down sharply and little room in Russia's budget to absorb the shocks.

Chemist Gregory Scholes, once enthusiastic about 2007 evidence of quantum coherence in photosynthesis, now doubts that quantum effects play a role in life. His recent papers show that networks of classical oscillating parts can produce collective behavior that follows quantum mathematics, suggesting life may imitate quantum systems rather than use them.

A physics journalist works through what the holographic principle and the AdS/CFT correspondence actually imply. The principle says everything inside a region can be read from measurements on its surface, and the essay weighs the evidence, including the widely accepted resolution of the black hole information puzzle, against claims that space is an illusion.

Transposable elements, or jumping genes, make up nearly half the human genome. The article traces them from Barbara McClintock's corn work and argues they're better understood as coevolutionary partners than junk, linked to adaptations from moths to wombs and to the processes that switch genes on and off.

Ronald Graham's 1971 conjecture asks whether any set of distinct nonzero integers can be ordered so every partial sum differs, even when the numbers live on a finite clock-like system. Several young mathematicians recently cracked it using randomness, and the piece places the problem in Graham's juggling-inspired interest in patterns.

Peter Attia's AMA episode covers thyroid biology and the free T4, free T3, TSH and reverse T3 tests, and why symptoms alone can't diagnose dysfunction. It also covers how hypothyroidism can be both underdiagnosed and overtreated, subclinical cases, T4 versus T3 treatment, Hashimoto's, pregnancy, hyperthyroidism, iodine and supplements, and what to do when tests look normal but you still feel unwell.

Physician-scientist Lloyd Klickstein walks through how a drug goes from unmet need to approval, using bimagrumab as a case study. The antibody targets myostatin and activin signaling to build muscle, and unexpected effects on fat mass and diabetes redirected it toward obesity, including use with semaglutide. The conversation also covers drug modalities, patents, capital, IND and GMP requirements, why development is slow and costly, mTOR inhibition, and his work on cancer prevention.