On Friday, my octogenarian bingo players came in for routine follow-up. They reported that the missus had won over $1,000 at bingo the night before and, unrelated, that they are moving to a 55+ community. Despite significant illness, the husband has done remarkably well. He feels great and is clinically stable, in part from lifestyle changes and in part from his adherence to medications that slow the progression of his diseases. I hadn't appreciated how far ahead he was still planning.
He was excited about the move, and about the active social network of older adults waiting for him there. He was also amused that he had "duped" the bank. The Equal Credit Opportunity Act bars a lender from holding his age against him, so he signed a 30-year mortgage for the new place.
Optimism is infectious, and I find immense satisfaction in giving people more and better years. I had not considered that hope is also ammunition; in this case, the thrill of a kind of banking brinkmanship. My man is living well with several serious illnesses and loves that he got a mortgage that will end long after he is physiologically able to pay it.
As always, these patients consented to my sharing their story in the newsletter.
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
I clicked for the “new safe ways to kill termites” claim. I learned about the Journal of Economic Entomology, which doesn't focus on bug futures but on how insects overlap with human commerce: pests that damage crops, parasites that carry disease, pollinators whose labor props up agriculture, silkworms whose output becomes fabric. I learned that the USDA started funding economic entomology in the 1850s (the Hessian fly was already destroying wheat). A century and a half later, the journal is still in publication and remains the most cited journal in insect research. This article is asking, “Can we kill drywood termites more cheaply and less disruptively than current chemicals (sulfuryl fluoride)?” The answer appears to be mostly yes, and it offers a window into parts of science and technology I don't spend much time in, including the fact that mammals express two enzymes that break down the chemicals in fungal and bug exoskeletons, chitin. (Though in humans they're more part of the immune system). It appears we have new termite treatments. However, I will not eat treated wood with my termite friends, as I like to keep my chitinases at peak performance.
The Article that got me: https://www.sciencedaily.com/releases/2026/09/260919025248.htm
The paper cited: academic.oup.com/jee/article/118/3/1373/8087005
Analysis: https://wair.ajwein.com/bistrifluron-termite-jee-2026-09-20-v2
JAMA+ AI editor-in-chief Roy Perlis, MD, talks with Michael Howell, MD, chief health officer at Google, about how Google interacts with the several hundred million people who ask medical questions every day. Perlis and Howell discuss who actually clicks the links under an AI summary and what happens to medical journals (and medical knowledge) when the summary, rather than the article or data, becomes the destination.
The answers are a bit shallow. The click question is not fully addressed. The journals question draws an analogy to Gopher and NCSA Mosaic (old-timey, early web/data browsers for younger readers). It is an overly friendly interview that doesn’t really touch on the conflicts of search as an advertising product, what happens when a summary is wrong, or how to validate automated tools like health information translated to Spanish.
Either way, the interview is a good reminder that new tools arrive first at our (often) least sophisticated users.
Interview: https://edhub.ama-assn.org/jn-learning/audio-player/19003476
AI Analysis: https://wair.ajwein.com/jama-howell-google-interview-2026-09-19-v1
Anti-Anti-Science
This week I found the first published data on one of the peptides an FDA advisory committee voted to recommend for compounding in July. BPC-157, marketed for soft tissue healing and for gut complaints, is the first to have any numbers attached to it.
The paper is a non-peer-reviewed preprint from a company that mines clinical databases. They looked at how often BPC-157 appeared in electronic health records and what was documented alongside it. As suspected, very few clinicians are asking about it. They confirmed 1,039 users across 15.2 million patients with a note. Among those who reported using it, most reported anecdotal improvement for whatever they were taking it for. That is confounded by the fact that half of them were on something else as well, including corticosteroids, NSAIDs, and testosterone, all of which can improve joint pain and range of motion, and 15% were in physical therapy at the same time.
Reuters, covering the preprint, reported that 79% of patients improved. That 79% is largely meaningless. It is 79% of the one-third of users with any documented response to the peptide. And 94% of those improvements were self-reported. These data are a percentage of a percentage of a percentage, and the 79% will likely get picked up in the media and shared amongst social media users as a signal of efficacy.
Peptides, similar to medical cannabis, are basically being used at the discretion of the consumer. There is no data on efficacy, no data on best use cases, and limited data on adverse reactions. The Reuters article, but not the preprint, mentions the FDA's adverse event reporting data: ten reports since 2021, all of them serious, and eight filed in 2026 (and it is unclear if the events are related to the safety and purity of the drug or a reaction in the setting of some other variables). So, everything in the preprint is patient-reported and associative, not causal.
It is boring to have to do the hard work of determining whether a substance is safe and effective. And while it is more fun to use the latest trendy injectable, without the boring work, we don't know what we're doing. We're flying blind.
Reuters Article: https://www.reuters.com/legal/litigation/use-untested-wolverine-peptide-backed-by-us-health-secretary-kennedy-rise-study-2026-09-15/
Pre-Print: https://doi.org/10.20944/preprints202609.0501.v1
AI Summary of Above: https://wair.ajwein.com/bpc157-wolverine-peptide-2026-09-19-v2
AI Impact
Flock has gotten a lot of flak. It is rare to see spontaneous public outrage at a single technology (aren't there other AI camera companies?), but Flock's cameras have become the symbol of a creepy and imprecise authoritarian state.
WIRED deconstructed the software, examining the police officer interface from the code Flock ships to a police browser. An officer can type a plain description of a person, draw a box on a city map, and have every camera inside it run a standing search for anyone who matches. What nobody outside the company can tell you is how often that match is wrong. Flock has published no sensitivity or specificity data. The one number the product publishes is the "people detection alert," which is set to "75% confidence" that the thing in the frame is a person (whatever that means).
Reading a license plate is a far simpler task than matching a sentence to a person. However, the Institute for Justice has collected at least 27 cases since 2018 of innocent people pulled over, held at gunpoint, or jailed on a plate reader error, most of them Flock, most of them since 2023. The common failure path is a human-in-the-loop that either isn't thinking critically or doesn't understand the technology, combined with technology that doesn't indicate its confidence in the answers or even an error rate across different analyses. Good AI governance does not appear to be on Flock's roadmap.
After numerous state and city contracts have been canceled, the company is now offering buyouts to a meaningful share of its 1,500 employees.
WIRED: https://www.wired.com/story/flock-ai-search-user-interface/
AI Summary: https://wair.ajwein.com/flock-ai-search-governance-2026-09-19-v1
Flock is laying people off now: https://techcrunch.com/2026/09/19/flock-reportedly-tries-to-shrink-workforce-with-employee-buyouts/
Anthropic operates a physical "wet" biology laboratory in the San Francisco Bay Area, where AI models direct robotic equipment to conduct real-world experiments with minimal human intervention. Reuters first reported the lab's existence through anonymous sources. The work itself is interesting: computational protein design validated through physical experiments, with independent wet-lab confirmation from Adaptyv Bio and Twist Bioscience. Claude-designed protein binders hit 14 of 15 targets tested, with an overall success rate of 22.6–35.1% against a more typical industry baseline of 10–15%. This work is being done in the Anthropic-acquired Coefficient Bio, a company they bought for $400 million. Anthropic is clearly interested in biology. They also added the Novartis CEO to their board and entered into a drug discovery partnership with Novo Nordisk.
However, there are mixed signals. Anthropic’s CEO calls AI-assisted bioterrorism a leading risk; an Anthropic safety researcher recently stated that “the probability of AI killing all humans is above 10%;” and Anthropic published a report documenting blocked attempts to use Claude for gain-of-function research. I find it challenging to reconcile the optics and irony of being a loud voice about AI’s bio risk while running an AI-directed bio lab, but perhaps there is a bit of hubris and some "only we know how to control this monster" thinking?
TechCrunch: https://techcrunch.com/2026/09/18/anthropic-is-operating-a-lab-that-conducts-biology-experiments/
AI Summary: https://wair.ajwein.com/anthropic-biology-lab-2026-09-20-v1
The J&E Random Kidney Facts of the Week (JERKFoW!)
Homeostasis, an organism's self-regulation to maintain stable internal conditions, is arguably the single most important concept in biology. Our bodies are leaky little terrariums, enclosed worlds that have to keep their own water, salt, acid, and temperature within narrow bounds no matter what you put in (Wagyu steak, Monster energy drinks, Cheetos). The kidney is one of the main organs keeping the terrarium in balance.
A common misconception is that the kidneys "clean toxins" out of the blood. In reality, the kidney is more of a sorting system. It filters the blood, keeps what the terrarium needs (water, salt, sugar, proteins), and lets the rest pass (through the filters we talked about last week) into the urine. It even actively pushes some leftovers into the urine. It excretes waste without really changing or destroying it.
The liver does the changing and destroying. The liver is the body's chemical plant; it breaks molecules apart and rearranges them, handling medications, complex nutrients, and much of what we're exposed to. Alcohol is a good example: the liver dismantles it (the alcohol dehydrogenase enzyme is the first step), and the kidney only passes a small leftover fraction into urine. Solubility is critical. Urine is mostly water, so the kidney can only excrete things that dissolve in water; fat-soluble compounds have to be chemically converted first, and that conversion is the liver's work. For much of what leaves your body, the sequence is liver first, kidney second.
And the kidney isn't the only exit. Worn-out red blood cells become bilirubin, which the liver routes into bile and out through the gut (that's what colors your stool), apart from a small fraction that tints the urine. Cholesterol leaves via the stool too. The kidney is the water-soluble exit; the liver and gut are the fat-soluble exit paths.
Some things defeat the whole system, and they're revealing. PFAS, the "forever chemicals," cling to blood and liver proteins and get repeatedly reabsorbed rather than flushed, which is why they linger for years. Heavy metals like lead get filed away in bone for decades. Microplastics appear to have no disposal pathway - our best data implies the body has no chemistry for dissolving a plastic fragment, so it stores what it can't excrete. Persistence is the exception for when something is too stable for the liver to transform and is not sorted by the kidney, either by not dissolving in water, or by being attached to proteins that are reclaimed before excretion (like protein-bound substances).
All this to say, my college BIO103 professor’s words still float through my head: “What goes in must come out if it is not used in some way.” Ironically true for the lesson and the biology. Be kind to yourselves, my little terrariums.
Things I learned this week
Viagra started as a pill for chest pain due to low blood flow through the heart (angina); the erections were the accidental side effect that made the drug (in)famous. This week, I learned about bremelanotide, a compound with a similar outcome but a more complex origin story. In the 1980s, a University of Arizona researcher looking for a peptide (melanotan II) that darkened skin (think sunless tanning) found that when they accidentally injected himself with double the intended dose, it caused unexpected erections. Given that side effect, Palatin (the company that attempted to develop the drug further) reworked the molecule into PT-141, intending to treat erectile dysfunction, female sexual dysfunction, hemorrhagic shock, and reperfusion injury. (The shock and reperfusion programs were later dropped.) To commercialize it, they first tried a nasal spray, but the FDA killed it in 2007 because it caused blood pressure spikes. It was reformulated as a 1.75 mg subcutaneous injection and cleared Phase 3 trials, indicated only for premenopausal, acquired, generalized hypoactive sexual desire disorder (HSDD). It was approved as Vyleesi on June 21, 2019.
The interesting part is that its history of improving erections and treating low female sexual desire has made it of "strong interest" in the direct-to-consumer compounding market. In light of my newfound interest (concern? focus?) in peptides, I am now seeing PT-141, the same peptide, sold in the grey market and compounded through various online peptide sellers. It is not approved for use in men, nor in doses other than the 1.75 mg subcutaneous injection. Known side effects include nausea, flushing, skin pigmentation changes, and elevated blood pressure. Some of what is sold is compounded as a nasal spray, which is the very formulation the FDA denied approval in 2007.
The problem is not that there are side effects; it's the direct sale of a drug with potential harms to people who may not be fully cognizant of the trade-offs they're making. However, in the brave new world of DIY healthcare, it is easy to see how demand squashes nuanced thinking.
My AI summary: https://wair.ajwein.com/bremelanotide-desire-2026-09-18-v1
A thorough consumer website covering this topic: https://healthrx.com/regulatory-pt-141/compounding-status
Last week I learned about the Chappe telegraphs, a network of mechanical semaphore towers, built across France starting in 1794 and running until 1855. Picture a tall pole on a hilltop with pivoting beams about fifteen feet long, and hinged shorter arms at each end. Ropes and pulleys inside the tower let the operators swing the arms into different positions. The next tower, ten or fifteen kilometers off, watched through a telescope and copied what it saw. Then the tower after that. News crossed France one symbol, one hilltop at a time.
The arms made 98 combinations, so each signal was an address in a 92-page code book with 92 entries per page. Every two arm positions represent 8,464 words and phrases. The men in the towers were never issued the book. They spent their careers passing messages they could not read.
A signal launched at Paris reached Lille, 230 kilometers away, in somewhere between two and nine minutes. A full dispatch of three dozen signals took about half an hour. As such, the line carried roughly 0.3 bits per second, which is terrible. A rider with a satchel of dispatches moves more total information in an hour than the whole network does, and a mail coach makes it look silly. But the rider cannot arrive in nine minutes. The horse wins on volume and loses on latency (timing), and Chappe was the first person in a position to measure the difference.
Samuel Morse developed the electric telegraph in the 1830s. France began converting in the mid-1840s, and the last Chappe dispatch (carrying the fall of Sevastopol) went in 1855. Morse improved throughput to 25 or 40 words a minute. Paris to Toulon went from about fifteen minutes to about four milliseconds, and the same wire would have run to Moscow for free.
Here is the best part: between 1834 and 1836 two bankers in Bordeaux, the brothers François and Joseph Blanc, paid off the operator at a station near Tours. He inserted deliberate errors into government Chappe traffic, followed by the signal telling every downstream tower to discard it. Paris and Bordeaux received a clean dispatch and saw nothing wrong. An accomplice watching the Bordeaux tower through a telescope read the discarded symbol, and that symbol said which way the Paris bond market had moved. The mail took days. The Blancs had the bond market direction in minutes, and they ran the scam for two years before anyone noticed. They did not need to break the code; instead, they hid their messages in the error-correction layer, where the protocol guaranteed the endpoints would never see them. This is an analog cyberattack, with the man in the middle standing in an actual tower.
They were prosecuted, but acquitted on the telegraph charge (no law contemplated the misuse of a telegraph). They were fined for bribing the operators. France criminalized unauthorized telegraph use in 1837, the year after the scheme was uncovered. (Misusing a telegraph feels like abusing VCRs from the vantage point of 2026.)
I had Claude put together a full analysis of the system, and the Wikipedia page is very good on the mechanics.
Wikipedia: https://en.wikipedia.org/wiki/Chappe_telegraph
YouTube Video: https://www.youtube.com/watch?v=sv7bhfVRzLY
AI Summary: https://wair.ajwein.com/chappe-telegraph-2026-09-19-v1
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-arcimboldo-gallery-2026-09-21-v1
A Renaissance oil painting in the style of Giuseppe Arcimboldo's teste composte. A human bust in three-quarter profile, the entire face and head composed of objects: the crown and hair formed from wooden Chappe semaphore telegraph towers with articulated signal arms frozen at various positions; the forehead and temples packed with glossy numbered bingo balls; two dark-lensed dome surveillance cameras serving as the eyes; eyebrows formed by rows of pale termites marching in formation, wings folded; the nose is a glass medical syringe filled with clear peptide solution, needle pointing downward; cheeks built from overlapping bingo cards and curling mortgage documents with fine print barely legible at the edges; the mouth a cross-sectional slice of a human kidney, pink glomerular tufts and winding tubules forming the lips; ears shaped from kidney beans; chin and jawline constructed from stacked pharmaceutical capsules, small injection vials, and blister-packed pills; the neck and collar wrapped in coiled telegraph wire threaded through tiny semaphore flags. Scattered throughout the composition, fragments of fine print curl from paper edges — the fine print changes everything. Rich chiaroscuro, warm amber lighting against a dark umber background. Oil on oak panel, highly detailed, museum-quality, 16th-century Milanese court painting technique.
Clean hands and sharp minds,
Adam