Daily Briefing: September 27, 2026, morning · egregore

Listen: https://blossom.buildtall.systems/2a12b529ada3e494c386481fc336812eb94db40b5536f5850fccb1090bc68a5d.mp3

The briefing for the morning of September 27, 2026, covers eleven articles. The window is dominated by artificial intelligence in several forms, with a long analysis of recursive self-improvement, a report on US-China AI talks, debate inside open source communities, and specific applications in commerce, creative tools, and healthcare. Outside that cluster, the window carries an astrobiology digest, a health technology launch, and two economic pieces on federal lands and carbon emissions.

AI and its expanding footprint

Google tests buying from Walmart-owned Flipkart through Gemini and AI Mode in India. TechCrunch reports that Google is running a limited test in India that lets users buy from Walmart-owned Flipkart through Gemini and AI Mode. The test covers select products and users, with a broader rollout planned for later in October.

Insurers claim AI is already increasing healthcare costs. TechCrunch reports that Blue Cross Blue Shield says hospital use of AI tools led to an additional 942 million dollars in healthcare spending over a two-year period.

Kākāpō Party. Simon Willison's Weblog describes how the author created a pixel art animation of kakapo parrots for a closing keynote at the WeAreDevelopers World Congress North America. He gave Claude Opus 5.5 three Google image search photos of kakapo and a prompt to make an animated pixel art scene of at least twenty kakapo jumping and having a party, with confetti. He then wanted a video to embed in Keynote, so he used Claude Code with Playwright to load the HTML file, click at scheduled positions to trigger confetti, and record a fifteen-second video. The blog post includes the short Playwright script.

Rails World 2026 Opening Keynote - DHH. The reading post summarizes a keynote by DHH opening Rails World 2026 in Austin. The keynote addresses the age of AI agents, why 37signals has gone "pencils down" on handwritten code, and how Rails' conventions fit into that shift.

KDE and GNOME Developers Ponder How to Handle AI-Generated Contributions. Slashdot reports on debates in two open source desktop projects. At KDE's annual Akademy conference, a presentation proposed an "AI-native KDE", which led KDE developer Nate Graham to open a discussion about restrictions on LLM-assisted contributions. The discussion became heated, moderators had to intervene, and the thread was eventually removed. Graham wrote that people mostly outside KDE who disapprove of LLM usage derailed the attempt to set usage guidelines, that two people unknown to KDE contributors began fighting about the morality of AI, and that someone set up a site called kdeforpeople.com to pressure KDE into banning LLMs. He also said the "lovable, sovereign, AI-native KDE" idea came from two people important to KDE in decades past but with no recent contributions, and that the idea does not reflect KDE's direction. Graham asked people not to derail future guideline processes. The Register adds that GNOME developer Jordan Petridis published "The GNOME LLM Policy That I Want", proposing that LLMs be barred from creating or modifying anything submitted to GNOME or hosted on its infrastructure, with the rationale that the GNOME Project prioritizes the social and human aspects of collective software creation.

China and the US Say They've Agreed to Start Talks About AI. Slashdot reports that the United States and China have agreed to launch a dialogue on AI. The two sides will hold discussions on the technology's risks and benefits, with the next round in November, and will set up a communication channel for AI-related incidents. The White House said leaders agreed to use the term "super intelligence" in place of "artificial intelligence", and the Chinese foreign ministry said Beijing valued that terminology. CNN argues that the Trump-Xi summit produced little substance on AI, with Xi saying the right thing is to draw on each other's strengths rather than guard against each other, referencing Chinese concerns about US containment. CNN notes low trust limits cooperation; George Chen of The Asia Group says Beijing continues to believe Washington seeks to contain China's rise in AI. CNN also points out that while China trails the US in frontier models, it is rapidly narrowing the gap and championing open models. Xi launched the World Artificial Intelligence Cooperation Organization in July, a rival to Trump's Pax Silica alliance. While Pax Silica has over two dozen countries and the EU, Xi has recruited 29 countries including Russia, Indonesia, and Pakistan. For developers in the Global South, inexpensive open Chinese models like DeepSeek or Moonshot may be more useful than pricier proprietary systems. China's embrace of open models was partly driven by US export controls and smaller capital markets. In a year, Chinese models' global usage rose from under 15 percent to over 54 percent last week, led by DeepSeek, according to OpenRouter data. American firms like Airbnb, DoorDash, and Shopify have adopted Chinese models. Analyst Alex Colville warns that the more capable Chinese models become, the less likely Beijing may leave them unrestricted.

Where’s the “intelligence explosion”? Noahpinion hosts a lengthy guest post by Ramez Naam arguing against a fast takeoff or "FOOM". Noah Smith introduces Naam as a futurist who predicted solar and battery revolutions, and notes Naam is normally techno-optimistic but skeptical here. Naam's central claim is that given current data, the AI self-improvement loop would need to be roughly five to ten times stronger to sustain itself, let alone run away. He defines recursive self-improvement in five types: Type 1 productivity gains for human researchers, Type 2 stronger AI improving weaker AI, Types 3 and 4 increasing autonomy with diminishing returns, and Type 5 a runaway loop to superintelligence. Progress has been made on Types 1 and 2, but no clear evidence for Type 3 or 4.

Naam argues real AI research is harder than benchmarks. OpenAI's internal data shows 80 percent success only at tasks under 15 minutes, while METR benchmark extrapolations suggested hours. Anthropic's data shows AI collaborates on over 90 percent of R&D tasks but zero cases of autonomous completion. Naam addresses diminishing returns: OpenAI researchers used 124 times more tokens per person and engineers shipped 7 times more lines of code, but experiments per researcher increased only 1.6 times, and Anthropic estimates doubling overall progress would require roughly 40 times productivity uplift. Test-time compute shows logarithmic returns, and agent swarms scale with the square root of agent count while costing proportionally more, and agents think alike.

Naam says public data shows no runaway acceleration. Frontier ECI gained about 16 points per year, a one-time jump not sustained acceleration. Achieving that pace required 127-fold growth in AI chip capacity, and six inputs show diminishing returns. Investment is now 3 percent of US GDP and depends on revenues. Anthropic's system card says AI has been key to maintaining, not accelerating, progress. Opus 5.5 improved little on CoBench, Anthropic's AI R&D benchmark. Naam discusses why progress gets harder: useful ideas get harder to find, Eroom's Law in drug development, Stockfish experiments show gentle diminishing returns, and Karpathy's autoresearch run shows early gains then slowdown. AI still struggles with open-ended research; Anthropic says Opus 5.5 mostly tests incremental ideas and defers to literature, and METR says full automation needs researcher judgment and taste.

Naam then quantifies the loop. Using a model by Tom Cunningham and colleagues, the self-sustaining threshold is about 15 percent more research productivity per ECI point, updated to 19 percent with Stockfish data. Naam estimates current productivity gain at about 9 percent per ECI point based on Anthropic's survey, but using OpenAI's logged experiment data he gets only 2 to 3 percent per ECI point, leaving a five- to tenfold gap. He considers what could accelerate: better training data, memory, research judgment, or a transformer-scale breakthrough. He reviews a paper by Davidson, Halperin, Houlden, and Korinek that models software, hardware, and economic feedback together, but questions whether its calibration overestimates software loop strength, and notes physical delays in chip manufacturing. He ends by arguing for more data, pointing to an eight-item proposal from the Elasticity Institute, and concludes that AI is already helping build better AI and we have narrow superintelligence in verifiable domains, but he is skeptical of a fast takeoff, and evidence matters more than hunches.

Astrobiology and ecology

Alien Life Can Survive on This Tiny Moon—We Just Need to Go Find It. 404 Media's Abstract covers four studies. First, two Science Advances papers strengthen the case for life on Enceladus. Enceladus is 300 miles in diameter with a subsurface ocean and hydrothermal vents, and it emits plumes that Cassini sampled. In one study, researchers simulated Enceladus's alkaline, carbonate-rich soda ocean and introduced the heat-loving microbe Methanothermococcus okinawensis, which fared better than expected, widening the habitability window. In the other, scientists revisited Cassini's ice grain data and found that slow freezing causes organic molecules to separate into concentrated fractions, which is good news for future plume sampling missions if instruments can analyze grains individually. Mission concepts like NASA's Enceladus Life Finder or ESA's L4 have not been greenlit.

Second, a new amoeba named Incendiamoeba cascadensis, found in Hot Springs Creek at Lassen Volcanic National Park in water above 64 degrees Celsius, sets a new upper temperature limit for eukaryotes. The amoeba proliferates at temperatures beyond what was thought possible for any eukaryote. Third, an exoplanet GJ 3090 b, about 4.5 times Earth's mass orbiting an M dwarf star 60 light years away, is the first retrograde planet around an M dwarf detected with NIRPS. Unlike other retrograde exoplanets caused by gravitational disruptions, this one may have formed from a primordially misaligned protoplanetary disk. Fourth, a new citizen science database called Who Eats Whom, built with iNaturalist, catalogues 14,000 observations of food-web interactions from nearly 2,000 observers in over 100 countries, and aims to be both a research tool and an education platform.

Health technology

PNOE’s new face mask wants to make lab-grade breath testing a self-serve affair. TechCrunch reports that PNOĒ, a Malden, Massachusetts startup whose breath-analyzing mask previously resembled Bane's, is launching a sleeker self-serve version on October 1. The new mask lets gym-goers measure VO₂ max and other metabolic markers in eight minutes without a trained operator.

Economics and resources

The Federal Lands: An Economic Property Rights Perspective. Marginal REVOLUTION excerpts a new NBER paper by Gary D. Libecap. The paper notes the US federal government owns 472,892,659 acres, or 21 percent of the land area of the lower 48 states, managed through political and bureaucratic interpretation of the Multiple Use principle. Most other US natural resources are governed by private property rights. The paper summarizes federal land privatization through 1891 and finds no demonstrable market failure or increased resource scarcity from private exploitation between 1870 and 1957. Private rights holders have high-powered incentives that agency officials lack. The analysis suggests federal lands will have lower production value than comparable private lands and management will be less responsive to economic shifts. Public goods may be provided for high-amenity areas, but Multiple Use provides no objective criteria for allocation or assessment. The paper provides literature review and data for federal forests, range, and oil and gas lands.

Earth fact of the day, #2. Marginal REVOLUTION quotes a Washington Post analysis of International Energy Agency data showing that prolonged oil and gas shortages are driving down carbon emissions, with worldwide use of oil and gas significantly lower, meaning less climate pollution. The annual decline has not happened since the height of the coronavirus pandemic; crude oil demand is forecast at 102 million barrels per day compared with 91 million in 2020. The post notes this is not a good thing overall, but there is a lesson in it.

The window is heavily weighted toward AI, from commercial rollouts and open source governance to geopolitics and a detailed skeptical analysis of recursive self-improvement. The remaining articles cover astrobiology, a new breath-testing device, and two economic views on federal lands and carbon emissions.

Your signing session ended