I wasn’t excited about the keynote speakers at our younger son’s college parents’ weekend. Bobby Lopez and Kristen Anderson-Lopez, the songwriters behind Frozen, Coco, and a shelf of Broadway musicals (Book of Mormon), weren’t on my list of historians, writers, or other speakers I'd typically seek out. I was wrong. Their hour of conversation and music was delightful. The couple writes together, and their songs punctuated my kids’ youth. They talked about composing, how a song moves a story forward, and writing on deadline while raising a family and staying married. Their daughters are their critics and test audience. They said Elsa and Anna are based on the girls.
In the early 2010s, Frozen was audio-only for me, endlessly repeated on our minivan DVD system I couldn't see. Listening to the Lopezes sing, I was surprised by how much of their work is stuck in my brain. Between their Wonder Pets songs and “Let It Go,” I experienced second-wave nostalgia, not for the 1980s, but for our 2006 Honda Odyssey with booster seats and Cheerios in the cushions.
Out of respect for my younger son, I did not use the Q&A time to share the Lopez lyrics I rewrote during those minivan hours (“Wear a hat/don’t let them see/don’t let them know your head is free-ee/Let it grow, let it grow, no Rogaine anymore/Here I am, a middle-aged balding man!”).
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
Last week, I saw headlines claiming AI can predict whether you’ll die from a single photo of your face. The story is both more and less interesting.
Researchers at Mass General Brigham trained an AI model on 40 million face photos from the internet. Then they showed the same model the ID photos of about 38,000 cancer patients receiving radiation therapy, along with how long each of them lived. The program learned to give each face a risk score for dying sooner rather than later.
When the researchers checked where the AI was looking, it focused on the area under the eyes and the bridge of the nose. Doctors have described the same hollow eyes and sharp nose in dying patients since Hippocrates. The paper cites it as the Hippocratic facies.
In a test group of 3,820 patients, the computer flagged 122 as highest risk, and 64 of them (about half) died within three months. Among the 975 it rated lowest risk, 20 died (about 2 in 100). But patients whose cancer had already spread scored higher on average, and many were sick enough that their doctors could already tell by looking. The researchers never checked whether the computer sees anything a doctor doesn’t.
Put simply, give the AI two patients and it correctly picks which one will die first about 7 times out of 10. A coin flip gets 5.
Then there are the usual AI problems. Most of the Boston patients were white, so the computer saw far fewer Black and Hispanic faces. When researchers tested it in the Netherlands, it guessed people were about five years older than they were, possibly because of a different camera, different lighting, or a different population. Nobody tested whether it was picking up clues like oxygen tubing under the nose, which runs right through the facial area the AI focused on.
Most of all, this is not what the headline described. The algorithm hasn’t been tested on healthy people, on selfies, or on anyone not already in radiation therapy. We are a long way from “take a picture, get your lifespan.”
Article:
AI Analysis:
https://wair.ajwein.com/face-survival-risk-scores-2026-10-22-v1
Data from the Moderna and Merck trial of a personalized melanoma treatment were released last week. The story also shows how confusing biotech reporting can be.
It is a custom mRNA therapy (built on the same technology as mRNA vaccines), but not a preventive vaccine. After surgery, Moderna sequences the patient's tumor and manufactures an mRNA built from its mutations. The median wait for the first dose was 64 days.
Data reported varied by news source. Some reported the therapy cut recurrence by 42 percent within two years. That number comes from a hazard ratio, which compares rates across the whole trial and is not a two-year probability. Others said it kept 82 percent of patients alive, but that was the share free of recurrence, not the share alive. The companies’ own release highlighted 58 percent, which is the drop in recurrence risk after the first six months, a secondary analysis. The overall survival difference is not yet statistically significant.
Regardless of the metrics for disease or survival improvement, the data are positive. Current treatments for melanoma after surgery are off-the-shelf drugs, such as checkpoint inhibitors (Keytruda) and pills. This therapy is different because it gives each patient’s immune system a list of up to 34 target proteins drawn from their own tumor. It is the first “add-on” drug to beat Keytruda alone in a Phase 3 trial, with side effects that were mostly fever and a sore arm.
The results leaked two days early, and Moderna’s stock fell by 20%. The data were good. The early hype (which drove up Moderna's stock earlier this year) was better.
Strip away the language and the result is solid; at two years, roughly one recurrence was prevented for every eight patients treated. The real question is, would a vaccine skeptic with melanoma accept an mRNA-platform-generated treatment for their melanoma?
Article
https://www.nytimes.com/2026/10/23/well/cancer-vaccine-melanoma-study.html?smid=nytcore-ios-share
AI Summary and Analysis
https://wair.ajwein.com/intismeran-interpath-001-esmo-2026-10-25-v1
Anti-Anti-Science
Last week I wrote about the new NIH vaccine injury clinic and said I hoped it would look for causal relationships in both directions. This week we have the contrasting problem.
Background:
Anyone can file a report in either system, whether the cause is clear or not. That is the design. The goal is a high-sensitivity signal that tells regulators, manufacturers, and researchers where to start asking questions. CDC and FDA both acknowledge the trade-offs. The systems accept temporal associations and unconfirmed diagnoses, and they have no denominators and no comparison groups.
CBS reported that the FDA recently received reports of two deaths in people who took bootleg retatrutide, Eli Lilly's next-generation GLP-1 weight-loss drug, which is still in Phase 3 trials and not approved for any use. It is sold online as a “research peptide.” One woman went into cardiac arrest after buying it from what the report called an unreliable source. The other died of dehydration, malnutrition, and cardiac arrest after a nurse sold her vials from an online peptide vendor. Both reports are associations only. Nobody knows the mechanism or what was actually in the vials.
https://www.cbsnews.com/news/deaths-reported-fda-bootleg-reta-weight-loss-drug/
Those reports went to FAERS, and the data lack context; nobody knows how many Americans are injecting retatrutide, at what dose, what other contaminants may or may not be present, or what the common complication rates are.
Products sold “for research use only,” like gray-market retatrutide, have almost no oversight. As I have written, when compounded and gray-market peptides are tested, purity, potency, and consistency vary widely.
This is what politicized science looks like. VAERS reports are used to impugn vaccines, and HHS has proposed paying physicians for the time it takes to file more of them. FAERS reports about bootleg retatrutide draw little beyond press coverage, while poison centers logged 862 retatrutide exposures in the three months ending in September, nearly double the quarter before. One set of reports gets a federal clinic. The other gets a news segment.
AI Impact
Science published a cover story on Microsoft’s Diagnostic Orchestrator, an AI tool that simulates a panel of doctors. One keeps the differential diagnosis, one picks tests, one argues against the leading diagnosis, one watches costs, and one checks the work. The panel correctly diagnosed 83% of New England Journal of Medicine case conferences and 80% of emergency admissions. For comparison, generalist physicians (mostly primary care and internal medicine), working alone and allowed to use UpToDate, achieved 36% and 47% accuracy.
Two notes: A separate AI plays the patient, and when the chart lacks whatever the panel asks for, it invents a plausible answer. In emergency cases, the panel received made-up answers 80% of the time. All 15 authors are Microsoft employees, and eight are named inventors on a patent for the test itself.
The broader implication is that matching patterns in clinical data with AI is increasingly a commoditized capability. That is, the skill of matching variable-quality, many-to-many data points (symptoms and diagnostics through the context of a patient's history) is less valuable. However, diagnostic accuracy is only one part of healthcare, especially in complex longitudinal care, which requires an ongoing coaching relationship built on trust and patient adherence.
https://www.science.org/doi/10.1126/science.aeh7930
https://wair.ajwein.com/mai-dxo-sequential-diagnosis-2026-10-23-v1
A few weeks ago, I wrote that AI-to-AI hacking would look like an immune system. “AI offense attacks a system or a problem. AI defense has to adapt, in milliseconds, to whatever exploit turns up.” This week, Reuters reported on a single hacker who pointed AI agents at more than 1,200 websites and broke into over 100, including sites belonging to John Deere and SharkNinja. The compute cost was about $5,000, or roughly $50 per break-in. Put that next to a piece from a few weeks back on what happens when weapons are democratized, and scarce tools become cheap (bronze to iron, or off-the-shelf drones for cartels and the Ukrainian military). Iron took centuries to spread. This took a credit card.
The J&E Random Kidney Facts of the Week (JERKFoW!)
If you read last week’s JERKFoW, the most obvious question popping into your head should have been “Adam, what happens to the blood measurements of a hibernating bear that can spend five or six months without eating, drinking, urinating, or defecating?” (Who among us doesn’t spend hours each week pondering bear physiology?)
Metabolizing protein generates nitrogen, which the body removes as urea in urine. If a human’s kidneys stopped for months, urea and other wastes would build up, leading to death. A hibernating bear’s metabolism drops to about a quarter of normal, and its heart slows from about 55 beats a minute to about 14. Blood flow to the kidneys falls sharply, and filtration drops by half or more. The kidneys still make a little urine, but the bladder wall reabsorbs it, so the bear never needs to urinate. Yet a hibernating bear’s blood urea level (a measure of nitrogen waste) falls.
Humans are not bears (you are welcome for this insight). Bears burn mostly fat over the winter, so they produce far less urea to begin with. The urea they do generate seeps into the gut, where bacteria break it down into ammonia. The liver then combines that nitrogen (as ammonia) with glycerol from the bear’s fat to build new amino acids. Researchers who administered tagged urea to hibernating bears found the nitrogen recycled into new protein, which helps explain how bears emerge from hibernation with intact muscles. It is a winter-only trick. A bear starved in summer, while awake, builds up urea like anyone else.
In wild Swedish brown bears, winter creatinine rose 2.5 times while urea fell by half. Based on the kidney function tests we use in people (discussed last week), the bear looks like it has kidney disease. Yet bears with elevated creatinine show little of the inflammation, heart strain, or muscle and bone loss that come with kidney failure in humans, and by summer their kidney numbers are back to normal. That is why some nephrologists now study bears (wild Swedish nephrologists?).
https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0072934
and
https://www.sciencedaily.com/releases/2011/02/110217141311.htm
Things I learned this week
In a fantastic display of my own confirmation bias, I missed the entire body of literature on quantitative analyses of dogs using word buttons for communication. Because of my (ongoing) fascination with the possibility that we have cracked cross-species communication (and the endless Insta and TikTok videos demonstrating this), I didn't think to look for actual research. Last week, my newly employed agentic assistant (https://app.instinct.com/) alerted me to two studies.
Both studies are from the same lab. That lab is closely tied to the company that sells the buttons (Big Button?). Nevertheless, both studies were preregistered, and the first used blinding good enough to keep Clever Hans out of the room. They found:
Nobody has tested the emotional buttons from the videos (LOVE YOU, MAD). Our wish to understand our dogs sits somewhere between anthropomorphism and evidence. So far, the evidence says dogs understand OUTSIDE.
Study 1
https://www.nature.com/articles/s41598-024-79517-6
Study 2
https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0307189
AI analysis with comparison to other animal communication research
https://wair.ajwein.com/talking-dog-buttons-soundboard-2026-10-22-v1
I learned that the release of Grand Theft Auto 6 makes me feel old.
Grand Theft Auto VI, releasing November 19, could cost the US economy $1 billion in lost work on launch day, according to a Spanish radio report that the internet has repeated for over a year. ZME Science rechecked and got $420 million, assuming each worker produces $70,000 a year. Either way, one HR software company reports time-off requests for November 19 running 9% to 17% above the usual third Thursday of November.
My thoughts:
Article:
and
AI Analysis
https://wair.ajwein.com/gta-vi-sick-day-billion-2026-10-25-v1
AI art of the week
A visual mashup of newsletter topics, plus 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-tattooed-man-gallery-2026-10-26-v1
A hand-painted 1940s carnival sideshow banner, oil on cracked canvas, hanging on a midway at dusk. THE TATTOOED MAN, a bald, mustachioed strongman covered in sailor tattoos of the week’s stories: an empty RETATRUTIDE poison bottle, a sleeping bear with a lab slip, a minivan under LET IT GROW, a ventriloquist’s dummy ringed by five doctors, a dog pressing OUTSIDE, a game controller marked NOV 19. Bottom banner: “82% ALIVE!*” with a tiny “*RECURRENCE-FREE.” On the tent flap, a small plain card: “INSIDE: ONE ORDINARY MAN.”
See https://en.wikipedia.org/wiki/Fred_G._Johnson
Clean hands and sharp minds,
Adam