Two loyal readers wrote to me last week saying parts of the newsletter "didn't sound like me." They pointed to phrases such as "The interesting part," "The problem is," and "Here's the best part," noting that these frequently turn up in AI ads.
Although I work with Claude to fact-check and clarify, I edit every sentence I publish (I track edits). Looking back at my writing, I found something I didn't expect. AI metadiscourse now feels normal. The interesting part is that this isn't just about writing; it's about thinking. (I wrote that one on purpose.)
These extraneous constructions are a kind of intellectual shrinkflation, more verbiage, less content. I was surprised they had slipped past my usual editing. Maybe I was moving fast, or maybe I am becoming inured to the rhetorical slop. Knowledge now arrives customized and pre-assembled, and it is harder to see where my thinking ends and the machine's begins.
Sometimes I joke that AI naysayers are like sword makers in an age of firearms. Other times I feel I have been handed a well-crafted double-edged sword, and last week I cut myself.
Of course, after I wrote this, I found these adjacent discussions of “AI moderates” and the impact of AI:
AI Daily Brief: https://open.spotify.com/episode/1GBw4jOcYFXa2IbGDYTfEU
and
Francis Fukuyama on why he changed his mind about AI risk (he is now more worried, not less).
https://www.persuasion.community/p/why-i-changed-my-mind-about-ai-risk
And John Philip Sousa's 1906 essay "The Menace of Mechanical Music" (phonographs and player pianos)
Gemini Notebook AI-generated podcast version of this week’s newsletter.
Latest episode at the top
https://wair.ajwein.com/podcast
Science and Technology Trends
Researchers grafted hearts from older mice into younger mice (that is, they added a second heart to the neck circulation of a younger mouse) and did the same with younger hearts in older mice. After four to six months, they sacrificed each mouse and used DNA methylation clocks, a molecular estimate of biological age, to compare each graft with hearts transplanted into mice of the same age. Young hearts in old hosts aged more than expected. Old hearts in young hosts showed signs of "rejuvenation," though less than the young hearts aged. The host's own heart and organs did not change. There are a lot of steps and a lot of associative findings here, and nothing close to human use. Still, an old heart seems to get younger in a young body, and a young heart older in an old one. Age may belong less to the organ than to the neighborhood it lives in, and it appears age is, in some ways, transmissible.
Preprint, not yet peer-reviewed (Poganik et al., "Transplanted hearts assimilate the recipient's biological age"): https://www.biorxiv.org/content/10.64898/2026.09.15.751836v1
Your page: https://wair.ajwein.com/transplant-heart-biological-age-2026-09-26-v1
In the same vein of interesting data that may not clear the bar of significance, I found a 2025 study in the Journal of the American College of Cardiology that tested whether HEPA air purifiers lower blood pressure. The theory is old and sound. Inhaled particulate matter triggers inflammation and constricts blood vessels, so fewer airborne particulates should lower pressure. Researchers recruited adults living near a highway in Massachusetts and put standalone purifiers in their bedrooms and living rooms. Every household got one month with a real HEPA filter and one month with the same unit, filter removed (a very expensive fan), in random order. Nobody could tell which was which. Overall, blood pressure didn't budge. In people whose systolic pressure started at 120 or higher, it fell about 3 points. (Small for one person, meaningful across a population.) A companion paper in Scientific Reports found that participants 40 and older finished one of two cognitive tests (Trails B, a test of mental flexibility) about 12% faster after a month of clean air. Those under age 40 did not see improvements in measured mental agility.
I've been a fan of air filtration since the pandemic (we still have HEPA purifiers throughout our house). I want to think clean air makes my 51-year-old brain work better, and this study sort of, kind of proves it.
AI summary: https://wair.ajwein.com/hepa-air-purifiers-bp-cognition-2026-09-26-v2
The JACC paper:
Journal page: https://www.jacc.org/doi/10.1016/j.jacc.2025.06.037
Free full text on PubMed Central: https://pmc.ncbi.nlm.nih.gov/articles/PMC13140469/
Anti-Anti-Science
Two public health stories this week pointed in opposite directions.
On the one hand, the CDC published this season's COVID guidance this week, restoring the recommendations from its July 2025 schedule. Every adult is advised to get a 2026 COVID vaccine. Immunocompromised people can self-attest and get additional doses. No attestations are needed to receive the first shot (although the FDA label is still narrow, so for a healthy 40-year-old the shot is technically off-label). So despite the rhetoric from various political leaders, the COVID vaccine is available to anyone who wants it.
Meanwhile, an FDA advisory panel voted to endorse a blood test that screens for dozens of cancers. The test was supposed to warn of cancers not otherwise evident. It missed about two of every three cancers diagnosed within a year of screening. It is highly specific (about 99.5%), so few people get a false alarm, but of those who test positive, roughly half do not have cancer ( i.e., the test has a low positive predictive value). But a screening test exists to find true positives, and this one mostly does not. A randomized trial of 142,000 people in England, published in NEJM the day before the vote, found no decrease in late-stage cancers. The panel voted to support the test anyway, 6 to 4 on effectiveness. A blood draw is easy. Making sure a negative result does not offer false reassurance, and that early detection changes outcomes for those who test positive, is not. Communicating that nuance will be difficult.
So neither of these stories is anti-science per se. However, they reflect the mixed messages of a world where public health is politicized. One agency recommends a vaccine despite public statements from its own department's leadership expressing skepticism and sowing doubt. The other's advisers just endorsed a screening test that has not shown it changes outcomes, and whose results depend on clear explanation and trust in “the system,” both in short supply.
COVID vaccine:
CDC interim clinical considerations (updated Sept 23, 2026): https://www.cdc.gov/covid/hcp/vaccine-considerations/index.html
AI summary: https://wair.ajwein.com/covid-vaccine-schedule-2026-09-26-v1
Studies on the cancer screening test voted on by the FDA Panel.
HS-Galleri trial (Sasieni et al., NEJM, Sept 22, 2026): https://www.nejm.org/doi/full/10.1056/NEJMoa2505723
NEJM editorial (Hunter): https://www.nejm.org/doi/full/10.1056/NEJMe2611312
Galleri test performance in NHS-Galleri (Neal et al., Nature Medicine): https://doi.org/10.1038/s41591-026-04652-8
AI Summary: https://wair.ajwein.com/galleri-fda-panel-2026-09-26-v1
AI Impact
Last week I shared that the AI company most vocal about AI and bioterror (Anthropic) had quietly built a wet lab in the Bay Area (https://wair.ajwein.com/anthropic-biology-lab-2026-09-20-v1). Also last week, Anthropic announced that a swarm of Claude agents searched a metagenomic database of 1.9 billion protein clusters looking for "interesting reverse transcriptases." The agents gathered roughly 200,000 of them, sorted them, and scored about 3,500 candidate partner genes. One agent happened to read the raw DNA and found a pattern no one had described: a string of evenly spaced repeats sitting next to a reverse transcriptase in a bacteriophage.
Reverse transcriptases are enzymes that copy RNA back into DNA. Bacteria and their viruses use them in odd ways, mostly to fight each other. CRISPR, the most famous reverse transcriptase to date, was first spotted in 1987, when Japanese researchers sequencing an unrelated E. coli gene noticed odd repeats beside it. It took nearly two decades to learn what they did. That work earned Jennifer Doudna and Emmanuelle Charpentier a Nobel Prize, and it helped launch gene-editing therapy.
Anthropic calls the new system ART, for array-associated reverse transcriptases. It looks like a CRISPR array. It lacks the genes that make CRISPR work, and no one knows what it does yet.
Three thoughts stand out:
Article: https://www.anthropic.com/news/claude-discovers-novel-enzyme-system
AI summary: https://wair.ajwein.com/anthropic-art-enzyme-2026-09-26-v1
For a while I've pictured AI-to-AI interaction as an immune system. AI offense attacks a system or a problem. AI defense has to adapt, in milliseconds, to whatever exploit turns up. Essentially, a very fast, rapidly evolving electronic immune system.
This week I found a discussion that illustrates the point. Hospitals are using AI to comb medical records for diagnoses that support higher coding intensity and, thus, higher payment. The Blue Cross Blue Shield Association says that it costs its plans $942 million over two years, with "no evidence of a corresponding change in care." Insurers are fighting back with defensive AI that downcodes the claims. Hospitals say the insurers started it. Both sides say they're losing, which biologists would recognize as a Red Queen race, where everyone runs as fast as they can to stay in place. The hospital's AI reads the whole chart. The insurer's usually reads only the claim.
I'm curious whether this spiraling AI arms race continues or evolves into agents negotiating prices in real time across wide ranges of care. Imagine your hospital submitting bundles of claims at once. The insurer's agent grades them against the hospital's known outcomes and against other organizations submitting similar claims, then negotiates the whole batch on the spot.
It would make cash flow harder to predict, and it isn't far off. My Tesla's insurance already reprices every month based on how I drive.
AI Summary: https://wair.ajwein.com/ai-coding-immune-response-2026-09-26-v1
The J&E Random Kidney Facts of the Week (JERKFoW!)
It feels like every patient is taking vitamin D. The supplement aisle is awash in it, and your primary care doctor probably measures it, even though the Endocrine Society advised against routine testing in healthy adults in 2024. It shows up in the recommendations for low bone density (osteopenia and osteoporosis). And yet the big trials are underwhelming. For most otherwise healthy adults, extra vitamin D hasn't been shown to improve bone, heart, or the long list of other things it's supposed to prevent. Low vitamin D levels can be real, but the value of correcting the number to "normal" often isn't.
The vitamin D you swallow in a pill, and the vitamin D your skin makes from sunlight, are inactive raw materials. The liver takes that raw material and converts it into a stored form called 25-hydroxyvitamin D (25-OH vitamin D). This stored form is still inactive, but it's the one your doctor measures, so it's the "vitamin D level" on your lab report. The kidney performs the final step, turning the stored form into the active hormone, 1,25-dihydroxyvitamin D, also known as calcitriol. That hormone does the work; it signals the parathyroid glands and other organs to manage the body's calcium.
So the thing sold to you as a vitamin is really the front end of a process that ends in a hormone, and the "vitamin D" on your lab report is the inactive stored form, not the finished product.
That is exactly why doctors watch vitamin D so closely in kidney disease. As the kidneys fail, they lose the enzyme that performs the final conversion. Patients can have plenty of the stored form and still be short on the active hormone. Topping up the raw material helps only so much, because the kidney's transformation tools are deprecated. So for patients on dialysis, nephrologists often prescribe a pre-activated version, a pill or an IV dose at treatment that skips the kidney's final step.
Which brings us back to the lab report. The next time a test flags your vitamin D, remember you're looking at the raw material, not the finished product. The kidney is the last station on the line, and for most people with working kidneys, it's quietly keeping up with demand, no matter what the number says.
Things I learned this week
Dozens of YouTube videos debate the causes of the Bronze Age collapse (drought, famine, the Sea Peoples, a few burned palaces). Historians are still arguing about which variables mattered most.
Last week, I found an article by Patrick Fitzsimmons, a postdoctoral fellow at Penn, arguing that the rise of iron and “democratized violence” were critical factors. Bronze requires tin, and tin had to be shipped thousands of miles from Afghanistan and Cornwall to the large Bronze Age empires for weapon production. Whoever controlled the trade routes and could buy tin at scale controlled who could have weapons. Iron needed only a bog (iron ore) and a forest (wood for smithing). The argument is that once anyone with a bog and a forest could arm themselves, the Bronze Age empires could not reassemble, and nothing like them returned for centuries. He brings both numbers and narrative, though the key numbers come from his own working paper, not yet published.
His modern analogy is the drone. A few hundred dollars now threatens a tank worth millions. It is true in Ukraine and increasingly in Mexico, where cartels carried out more than 200 weaponized drone attacks between 2021 and 2025. During the Cold War, mutual assured destruction put the most expensive weapons in the fewest hands and produced a standoff. Cheap drones put force in many hands and produce churn. Iron made war affordable. Drones may be doing it again.
WTF is the Bronze Age Collapse?
https://en.wikipedia.org/wiki/Late_Bronze_Age_collapse
Article
https://worksinprogress.co/issue/weapons-of-mass-decentralization/
AI Summary
https://wair.ajwein.com/weapons-mass-decentralization-2026-09-26-v1
A loyal reader, who may or may not be one of the same loyal readers who commented on my writing in the opening, shared this 2025 blog post with me this week. It is a satire grounded in reality. The commentary smartly captures the theme of AI driving idea homogenization. Don't get distracted by the sixth-grade humor; take a step back and appreciate Radek Sienkiewicz's "Why do AI company logos look like buttholes?" Claude's logo is Exhibit A, and it pulses when you click it. Sadly, you can't unsee some things.
https://velvetshark.com/ai-company-logos-that-look-like-buttholes
AI art of the week
A visual mashup of topics from the newsletter, and an exercise to see how various LLMs interpret the prompt. I use an LLM to summarize the newsletter, suggest prompts, and generate images with different LLMs.
https://wair.ajwein.com/wair-art-sun-stone-gallery-2026-09-28-v1
This week, I chose something circular: an emblem in the manner of the Mexica (Aztec) Sun Stone, a carved relief disc shown as it looked when new, painted in brilliant red ochre, turquoise blue, maize yellow and white with crisp black outlines. Perfectly frontal and symmetrical, the full disc filling most of the frame, set on a plain cream stucco ground. Concentric carved rings, flat painted color, no perspective.
Clean air and sharp minds,
Adam