I had just passed through the TSA touchless facial recognition kiosk at Chicago's Midway when I approached the metal detector. It was 6:00 AM on a Thursday. With my backpack trundling down the conveyor, I strode confidently through with pockets cleared and emboldened by my nylon, plastic-buckled belt paired with slip-on casual shoes. The alarm went off. I sighed. With a lilt reminiscent of Chris Farley in SNL's "Da Bears" Superfans sketch, the TSA agent said, "Da MACHINE has randomly selected YOU," emphasizing the words "machine" and "you," indicating the machine's agency, its preference for me, and my somehow-now-special status. I looked up, stared her in the eye, and said, "It SELECTED me?" With rehearsed expressionlessness, she pointed me toward the body scanner. I saw no flicker of species solidarity. She seemed not to appreciate the layers of dystopian nightmare she projected, a world in which she is no longer employed, and we meat-covered food tubes with dangly appendages are now sorted by machines that look for danger in everyone but themselves.
Later that afternoon, unpacking at home, I found an unopened nine-ounce glass bottle of sparkling water in my backpack. Da Machine seems to approve of San Pellegrino.
Da Bears: https://youtu.be/NhMqjHEDmcU?si=3Y4aVsPAyBtPdeTD
Gemini Notebook AI-generated podcast version of this week's newsletter.
https://wair.ajwein.com/wair-podcast-2026-08-24-v1
Science and Technology Trends
Merck and Moderna announced this week that intismeran autogene, a personalized cancer vaccine they call individualized neoantigen therapy, plus Keytruda beat Keytruda alone in surgically resected melanoma. What's new is that researchers genetically sequenced each removed tumor and manufactured a custom mRNA strand to encode up to 34 mutations found in that person's cancer. About six weeks later, a vaccine exists matched to that one patient. This uses the same technology as the COVID vaccine and as mFLUSIVA, the first mRNA flu vaccine, which the FDA approved on August 5 for adults 50 and older.
INTerpath-001 included roughly 1,100 patients with surgically resected melanoma who all received Keytruda, a standard adjuvant therapy, and were randomized to receive either the mRNA vaccine or placebo. Both arms received Keytruda, a drug that already works, so the vaccine had to earn its result on top of standard of care. That is a harder test than most oncology trials set. The combination met its primary endpoint of recurrence-free survival and its secondary endpoint of distant metastasis-free survival.
This was a preplanned interim analysis, which is the point in a trial where an effect is most likely to look larger. The companies released no hazard ratios, confidence intervals, or counts of patients who recurred or died. They stated only that the results are "clinically meaningful." We also don't know whether these patients live longer. Overall survival is still accruing, with the trial scheduled to end in 2029. (We do have numbers from the smaller Phase 2b study, presented at the American Society of Clinical Oncology in May, where the combination cut the risk of recurrence or death by 49 percent at five years.)
Moderna asked the FDA for accelerated approval on the Phase 2 data in 2024 and was turned down. That refusal is why this randomized Phase 3 exists.
The platform Moderna built to print these one-patient cancer therapies is the same platform that prints any mRNA vaccine quickly. In August of last year, HHS canceled 22 federally funded mRNA vaccine projects worth close to half a billion dollars. Those were pandemic preparedness contracts, not this one. The melanoma program runs on private investment.
STAT News Article: https://www.statnews.com/2026/08/19/mrna-cancer-vaccine-trial-melanoma-merck-moderna/
AI-assisted analysis of STAT + other sources: https://wair.ajwein.com/mrna-neoantigen-melanoma-2026-08-19-v2
When quantum computing becomes reliable and scalable, feats of mathematics not currently achievable will be routine. I struggle to follow the technology, though, especially when a typical sentence in a quantum computing article reads: "IBM and University of Chicago researchers recently posted a paper describing a protocol they call doped Clifford sampling, in which a 70-qubit, depth 70 circuit containing 468 T gates ran on IBM's Boston superconducting processor using 97 physical qubits in total."
In a superposition of irony, simultaneously ironic and not, I had AI help me appreciate what this is all about. Long story short, these data demonstrate an improved ability to prove how well a quantum calculation ran in a setting where checking the answer is impractical. The error checks are built into both the circuit's qubits and its time steps, so a mistake that appears midway through the calculation still gets caught.
Bear in mind this is one paper. What it actually shows is less dramatic than the SciTech Daily headline implies. (I found an even more recent, non-peer-reviewed paper validating the methodology on classical computers, which IBM stated couldn't be done.) Nevertheless, knowledge progresses one finding, one paper at a time. The day we finish replacing public key cryptography, the math that signs your bank transfers and your blockchain wallets, will arrive through small intellectual steps like this one, even if it goes by a name like doped Clifford sampling, which sounds like a big red dog who ate too many edibles.
AI Summary: https://wair.ajwein.com/doped-clifford-sampling-2026-08-23-v2
IBM's own writeup, for comparison: https://www.ibm.com/quantum/blog/quantum-advantage
The non-peer-reviewed validating paper on classical computers: https://arxiv.org/abs/2608.13110
Anti-Anti-Science
Two opposite problems that end with the inability to unknow things.
Last week, I encountered two clinical questions that highlighted challenges I didn't train for in medical school: the risks of more data than anyone can readily interpret. Midweek, I was asked about participating in an insurer-sponsored home cancer screening. Later in the week, a patient I have known for years came to me after months of life-altering symptoms, invasive tests, imaging, blood work, several specialists, few findings, and no unifying diagnosis.
A third-party vendor runs the screening. It is a package of tests: a mail-in stool test for microscopic GI bleeding (a fecal immunochemical test, or FIT), an HPV self-swab, a referral for a PSA blood test, a phone-based skin photo tool, referrals for mammography and lung CT scans, and a thirty-gene cancer risk panel. The vendor bundles federally mandated screenings with a wider range of blood tests, imaging, and counseling. However, a screening test is rarely a single event. It is often the first step in a process that ends in varying degrees of understanding.
To be maximally valuable, screening tests must focus on individuals with some risk (like a family history). Screening everyone, regardless of family history, smoking, or underlying disease, changes the odds that a positive test is real. When a disease is relatively rare, a larger share of the positive findings are false. These tests also do not replace recommended, age-related screening. A positive FIT changes the urgency of a colonoscopy, not the need for one. A negative HPV swab does not excuse a woman from routine gynecologic exams. And a genetic cancer panel offers some sense of one's cancer risk at a price. Published rates of variants of "uncertain significance" run from under ten percent on small focused panels to more than forty percent amongst broad patient panels. The bigger the panel, the more uncertainty, and the more likely one is to get a result that leaves the patient with anxiety and doctors with no next step. Moreover, those variance rates run higher in Black, Asian, and Hispanic patients because people of European ancestry are the most common reference genomes in the comparison databases. And while the 2008 Genetic Information Nondiscrimination Act (GINA) protects one from genetic discrimination in health insurance and employment, an equivocal or positive genetic test can change life, disability, or long-term care insurability.
My patient had the logically adjacent, opposite problem. Prolonged life-altering symptoms, invasive tests, imaging, blood work, many specialists, few findings, and no unifying diagnosis. I have known him for many years, so when we jointly decided to put his deidentified data into a clinical AI tool, I was confident he could handle the ambiguity of a list of rare and unsettling possibilities.
The specialist he saw next disagreed. It was their first meeting, and he told the patient that these were rare findings more frightening than likely. Still, this new specialist, like all the others, including me, does not have a unifying hypothesis or diagnosis, only a list of what illnesses he doesn't have and comments about it probably being some other specialist's organ.
Both stories end the same way. You cannot unknow once you know. The screening consumer holds a finding with the possibility of ambiguous action and insurability problems. My patient holds a stack of expert opinions about what he doesn't have. This is why people turn to alternative therapies, especially when those therapies are offered with confidence and empathy. Scientific medicine's intellectual integrity is precisely what prevents it from offering certainty. That honesty is the ethical, but often feels like a terrible sales pitch.
Cancer screening must be done thoughtfully. I strongly recommend you speak with your physician to understand which tests make the most sense in the context of your family history, lifestyle, and you. Here is what I shared with the person who asked for my advice on this topic:
https://wair.ajwein.com/incentivized-cancer-screening-2026-08-20-v3
And while the 2008 Genetic Information Nondiscrimination Act protects one from genetic discrimination in health insurance and employment, an equivocal or positive genetic test can impact life, disability, or long-term care insurability.
https://www.genome.gov/about-genomics/policy-issues/Genetic-Discrimination
As always, the patient's story is shared with permission.
AI Impact
The FDA opened a public docket this week on how it should test generative AI medical devices. The document is a discussion paper rather than draft guidance or policy, and the agency is asking for comments through October 19. This is part of the Health and Human Services push to increase automation and generative AI in care delivery. The FDA is asking whether a finished device can be shown to perform the clinical task it claims to do competently. While this is neither draft nor proposed guidelines, the discussion highlights the limits of how we test and monitor humans who deliver healthcare:
Regulators don't try to anticipate every situation a physician will encounter during a career. They test underlying knowledge, observe performance in clinical settings, and require continuing oversight. The FDA is examining whether a version of that model could work for AI while accounting for the different technical and legal issues surrounding medical devices.
The proposed tests are more specific than I expected. For instance, whether the device knows what it does not know. Whether it refuses to answer outside its scope. Whether the numbers it produces are right. Whether it performs the same across patient subgroups. Whether it holds up against a prompt written to break it. And whether anyone will notice when a silent update makes it worse.
I am drawn back to my opening "Da machine has selected you." The FDA is asking whether "da machine" knows what it knows.
AI-assisted review: http://wair.ajwein.com/fda-genai-competency-2026-08-20-v1
As midterm elections get closer, data centers are showing up in the news more often. One of my favorite podcasts, the AI Daily Brief, spent an entire episode covering the facts and fiction of data centers, including the psychology behind why this is a splitting issue among voters. The podcast offers a good discussion that overlaps with anti-science: logical and factual fallacies, emotions, and politics are tangling into an unfortunate intellectual morass. The episode argues that the backlash is not really about water or noise, but rather about communities discovering that decisions were made for them under non-disclosure agreements.
https://aidailybrief.ai/e/2026-08-21
The J&E Random Kidney Facts of the Week (JERKFoW!)
You can't drink seawater. The kidneys pump salt and other electrolytes (like potassium) from the bloodstream into the urine, but they have a limit to how much they can concentrate it. Seawater is salty enough that excreting all the salt from one liter of it would take roughly 1.5 liters of urine. A liter of seawater holds about 35 grams of salt, and 35 grams is roughly the ceiling for what a human kidney can pack into a single liter of urine. That ceiling assumes perfect efficiency and nothing else to excrete, which never happens, because potassium, urea, and the rest of the day's chemistry are always waiting to be excreted too. So, when humans drink seawater, they pass more water than they drink. Thus, they lose volume, they become dehydrated, and the ocean wins. Centuries of sailors have lamented this. Samuel Taylor Coleridge, who was not a sailor, got a famous poem out of this nephrologic limitation, "Water, water, every where, / Nor any drop to drink."
On the other hand, if you can transmogrify into a kangaroo rat, you could sip seawater all you want. Its kidneys concentrate urine five times more than ours can, over four times saltier than the sea itself. There are probably other trade-offs to being a kangaroo rat, but in this case it seems like a win.
Things I learned this week
This week I learned that gastroenterologists have a good sense of humor, and that social media is filling up with colonoscopy prep mocktail recipes designed to make the cleanse a more sophisticated, upscale affair. This is what happens when millennials turn 45, said the 51-year-old Gen X kidney doctor who didn't think of this.
https://www.instagram.com/reel/DZS9X6Xpm7Q/?igsi=MXI5dmVwZ2JhNDRsOA==
Here is an article for those not on Instagram. However, Insta videos have a charm the written word lacks.
https://www.eatingwell.com/gastroenterologist-colonoscopy-prep-mocktails-11995448
I was excited to find an X post discussing how AI is helping us "decode" animal communication. The post cites numerous, diverse studies claiming AI has analyzed and interpreted the calls of elephants, marmosets, whales, bats, dolphins, zebra finches, and crows, plus a robotic bee that could waggle dance. The post's author seemed to imply that we are on the cusp of Dolittle-like communication with animals. I had Claude pull the articles and summarize them.
Three things stood out.
First, less of the work is AI than the post suggests. The dolphin and zebra finch projects use real contemporary models. The bat study used speaker-verification software from 2016, and the robotic bee has no AI at all. Much of the work is machine learning that has been around for years, doing mathematical classification.
Second, rather than interpreting, most of the work is categorizing and classifying the noises, associating specific sound patterns with specific animals. We don't know what the animals are specifically saying, though some observed behaviors are associated with sounds.
Third, the robotic bee is cool, but the bees only followed its directions sometimes.
I get overly excited when I see these kinds of posts, and I have to ground myself in the studies. I am fascinated by dogs that can use word buttons and by other windows into the minds of animals. I was fascinated by Koko the gorilla and her sign language too, until I learned most of the claims came from her handler and very little survived independent review. While I'm sure we will get there one day, we can't yet have a meaningful conversation with a dolphin or whale about the finer points of seafood, or better negotiate with the mafia-like gangs of crows or monkeys I've written about in the past.
However, AI is really good at gathering and summarizing information that shows it's not good at communicating with animals.
X-post: https://x.com/itsolelehmann/status/2090831992200020253
Analysis and bibliography of the studies: https://wair.ajwein.com/ai-animal-communication-2026-08-23-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.
Gallery with images: https://wair.ajwein.com/wair-art-paintbynumbers-gallery-2026-08-24-v2
This week's image is a paint-by-numbers kit nobody finished. The filled regions are the things we actually know. The outlined ones are published but unconfirmed. Three regions have no number printed in them at all: where the hazard ratio would go, what the dolphin is saying, and my patient's diagnosis.
Clean hands and sharp minds,
Adam
Relive all the past thrills and excitement - The What Adam is Reading Archive
http://www.whatadamisreading.com/