Last Wednesday, we hiked and lunched with our kids, both home from summer research. Annapolis's suburban woodlands offer miles of moderate hills and trailheads near strip malls. So after a few hours of banter and bugs in the humidity, we tried a local all-day breakfast place, aptly named with a play on the word egg.
The inside was carpeted and lined with dark oak chair rails, decorated in country couture, the walls peppered with declarations of patriotism, the importance of family, and the sentimentality of home in muted reds and blues. Even the ketchup bottles wore American flags. The clientele skewed octogenarian, and they all knew each other, talking across tables and with the servers. It felt like we were the only strangers. The staff knew the older man next to us, who cradled his cane at the table. Eavesdropping, I learned he really liked canned pineapple. It was not on the menu, but it was clearly available to him.
For two weeks every mid-August, our sons are the same age. It is strange to think that as of this month, neither is a teenager; we are the parents of adults. Nothing drives this home like unexpectedly sitting with them in a benign version of Beetlejuice's afterlife waiting room, redecorated as a nursing home dining hall. My breakfast burrito was good, but the best thing served was the comfort of bringing down the room's median age.
For those too young to remember Tim Burton’s vision of the afterlife waiting room (Beetlejuice! Beetlejuice! Beetlexxxx) : https://www.youtube.com/watch?v=cZwdCa0ynEw
Gemini Notebook AI-generated podcast version of this week’s newsletter.
https://wair.ajwein.com/podcast-2026-08-17-v1
Science and Technology Trends
GLP-1s, the drug class that includes Ozempic and Zepbound, have an expanding list of benefits: better glucose control, weight loss, less joint pain beyond what the weight loss alone predicts, lower risk of kidney disease, lower risk of heart failure. Dementia isn't on the list as no human data clearly show GLP-1s improve it.
This week I found a good overview of the GLP-1-and-dementia studies. It explains why the dementia-model mice improve on these drugs and why that improvement hasn't shown up in humans. The piece that sent me down this path is the digital equivalent of a throwaway journal, free medical news built to sell ads. But the topic came up with patients this week, and it's a good map to the original data.
Article:
AI-assisted broader summary of the available research:
https://wair.ajwein.com/glp1-alzheimers-2026-08-16-v1
Eric Topol inspired me to re-examine exercise data this week. He wrote about a new study reporting that people who climb stairs live longer and have fewer cardiac deaths. My first thought was, “The people who take the stairs are the people who can take the stairs. I wonder if the ‘healthy-user effect’ is controlled for?”
So I had Claude and OpenEvidence pull the latest trials and data. A few minutes of digging turned into something I can offer my patients: nine evidence-based assertions about exercise, in plain language.
The Combined AI-supported summary:
https://wair.ajwein.com/exercise-cardiovascular-outcomes-2026-08-14-v2
Anti-Anti-Science
Two of my favorite science advocates and evidence proponents, Jess Steier and Andrea Love, do more than a podcast. Their nonprofit, the Center for Unbiased Science & Health, publishes a steady stream of articles, guides, and tools aimed at the hard problem of making good evidence legible to people who don't read science papers or delve into the methods section.
Last week I found their Risk Comparison Dashboard. It takes two risks and renders them as squares scaled by area, so "1 in 10,000" and "1 in a million" stop sounding like synonyms for "rare" and start illustrating how far apart a hundredfold actually is. You can set the disease against the vaccine that prevents it, or either one against a car crash or a lightning strike. The measles comparison is striking: measles' bad outcomes far outweigh the vaccine's potential side effects.
The tool will not tell you how likely you are to catch measles, because that depends on where you live, what's circulating, whether you travel, and how many people around you are already vaccinated. Instead, the tool compares what happens if you were to experience either outcome: the severity and frequency of outcomes with infection versus vaccination. The people who most need this tool are the least likely to open it. I could see pulling it up in the office anyway.
https://www.unbiasedscience.org/risk-comparison
AI Impact
“Agent-to-Agent Infection Is Here” is a headline describing Anthropic’s work seeding an AI agent, working in a team of six, with a hidden idea, then watching that agent recruit the others, write the hidden idea into their memories, and see the idea survive attempts at a memory wipe. What I found most interesting, however, was that patterns of contagion, immunity, and "battle" among synthetic life started emerging from interactions between multiple agents. It reminded me a lot of biology, and I would consider this a form of convergent evolution. These patterns aren’t programmed. Contagion, immunity, and something like battle emerged when agents interact: convergent evolution, running in silico.
I had Claude put on his digital gloves and a screen protector to review the article, pulling in some additional resources:
Anthropic article:
https://www.anthropic.com/research/multiagent-systems
AI-assisted analysis of this and other similar articles:
https://wair.ajwein.com/multiagent-convergent-evolution-2026-08-14-v1
In June 2025, I wrote about Anthropic and Claude Sonnet 3.7 running an office mini-fridge autonomously for a month and encouraging various Anthropic employees to game the AI into behaving strangely, which they accomplished.
In November of 2025, Andon Labs put a Claude model in a robot vacuum cleaner, which created a massive AI existential crisis.
https://www.whatadamisreading.com/posts/2025-11-17-what-adam-is-reading-week-of-11-17-25.html
Last week, a loyal reader sent an article about Project VEND from Andon Labs, testing a variety of LLMs running a refrigerated vending machine with the instruction to turn a profit. They compared various models and found that Claude Opus 5 did best at generating revenue from a $500 starting point. We're getting closer to an age of Machiavellian, pure-capitalist inventory management with AI. What would happen if you instructed the LLM to act like a socialist (equitable distribution of vended items rather than maximizing profit)?
Article:
https://andonlabs.com/blog/opus-5-vending-bench
AI-Assisted summary:
https://wair.ajwein.com/vending-bench-opus5-2026-08-10-v1
The J&E Random Kidney Facts of the Week (JERKFoW!)
The bone marrow makes red blood cells (RBCs). Numerous logical anatomical locations could host the sensors that determine how fast and how many RBCs to make: the lungs, where oxygen is absorbed; the carotid arteries, where blood leaves for the brain; or the heart, which pumps the whole blood supply. Instead, the master sensors sit in a specialized layer of kidney cells. In this layer, oxygen tension is chronically low. But that's the engineering point: a sensor that lives at the edge of hypoxia notices the moment oxygen levels change. When those cells sense oxygen levels dropping, they release erythropoietin (EPO), the hormone that tells the marrow to build more RBCs, the oxygen carriers of the body. In other words, put the smoke detector in the one room that always smells faintly of smoke, so it trips the instant there's a real fire.
This is also why anemia (low red blood cell counts) comes with kidney failure. When kidney cells die, the sensor and EPO-secreting cells die too. Thus, the bone marrow stops getting the signal to build. For most of the twentieth century (1960s through 1989), that meant dialysis patients required transfused blood often and lived exhausted, running on a chronic oxygen debt that comes with profoundly low blood counts. In one of biotech's triumphs, scientists purified EPO from 2,550 liters of human urine in 1977, cloned the gene in 1985, and, by June 1989, had recombinant erythropoietin FDA-approved for anemia of kidney disease. The pathway behind all of this won the 2019 Nobel Prize. Figuring out how to mimic the work of oxygen-deprived sensor cells in the body turned out to be worth a trip to Stockholm.
Things I learned this week
I have been reading about John Hunter, the 18th-century surgeon who brought anatomy into mainstream medicine, but needed a steady supply of cadavers to do it. Digging through the references, I learned about the grave-digging economy of the 1700s: the gangs, the price inflation, the parliamentary investigations, the whole tension between the quest for knowledge and the ethics of where the specimens came from. It is The Wire, if the product were bodies and the corner boys reported to the Royal College of Surgeons.
We supposedly fixed this. The 1832 Anatomy Act put the resurrection men out of business by handing surgeons the unclaimed bodies (often poor laborers) instead. Problem solved, if you're not poor.
The book: https://www.goodreads.com/en/book/show/161724.The_Knife_Man
My gathering of primary resources:
https://wair.ajwein.com/body-snatching-economics-2026-08-16-v1
A friend of my son's rows for a Division 1 university. He mentioned that, at his coach's instruction, he takes large doses of caffeine, sodium bicarbonate, and beta-alanine to improve performance. As someone who spends his days with complex chronic disease patients, my automatic thought was that this was nonsense. Loading a healthy body with grams of anything to chase a theoretical metabolic edge seems like a setup for a case report, not a race plan.
Interestingly, the literature proved me wrong. All three are supported by some evidence, with data suggesting they help maximize performance in eight-to-ten-minute rowing races (and other sports demanding short-term sprints of activity). Though the measurable gains are small (0.5-2.0% improvements in performance), in a sport decided by fractions of a boat length, small improvements can matter.
This was a great reminder that medical practice is contextual. What I worry about for sicker patients does not translate to a healthy twenty-year-old with a four-year competition window. What is unsafe in patients with multiorgan disease can be reasonable in an athlete with none. The trade-offs still matter. Four to five years of daily baking soda and beta-alanine, with caffeine dosed around races, could have unappreciated longer-term concerns. But the small edge it offers a rowing team seems less ominous and more evidence-based than I initially thought.
Here is what I learned: Bicarbonate, caffeine, and beta-alanine, oh my!
https://wair.ajwein.com/ergogenic-aids-rowing-2026-08-16-v2
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.
Gallery with images:
https://wair.ajwein.com/wair-art-midcentury-gallery-2026-08-17-v1
A mid-century modernist screen-print poster in the tradition of 1950s–60s graphic design (Alvin Lustig, Paul Rand, Giovanni Pintori). Flat planes of color, bold negative space, visible screen-print texture and slight ink misregistration. Strictly limited palette: deep navy, warm gold-ochre, muted teal, and off-white paper, with one small alarm-red accent. No gradients, no photorealism, no 3D rendering.
Central form: a single stylized human kidney rendered as a clean geometric silhouette, sitting in a dark low-oxygen field. One small red mark glows at its edge, like a sensor alert in a room that always smells faintly of smoke. Thin gold signal lines radiate outward from that red point.
Around the central kidney, arranged as flat, evenly weighted symbolic elements with generous space between them: a simple figure ascending a flat geometric staircase; two nested squares of very different sizes sitting side by side, one tiny and one large, as a comparison of risk; and a small cluster of connected dots forming a network node. Reduce everything to its simplest geometric essence: poster-like and iconographic.
Balanced, calm composition with lots of breathing room and one clear focal point. The mood is quiet, intelligent, and precise. Aspect ratio 4:5 (portrait). No text, no lettering, no words anywhere in the image.
Clean hands and sharp minds,
Adam
Relive all the past thrills and excitement - The What Adam is Reading Archive
http://www.whatadamisreading.com/