Week of May 25, 2026
I find the long-term patient relationships very satisfying. And since the pandemic, telehealth has opened a strange window into my patients' lives. I've had patients connect from bed, next to their spouse, reading the paper on well-chosen Pottery Barn sheets. I've met countless family dogs. I've received video pantry tours under the pretense of sodium intake counseling ("Nice Tastykakes, and I see that holiday popcorn tin by your feet. Are you eating that, or is that just there for decoration?").
So, I was only somewhat surprised when my only telehealth patient from last week’s clinic appeared on my iPad screen holding a glue stick. She was assembling Memorial Day decorations with the intensity and fervor of a deadline. "I only have a few hours to get these decorations done." She did not stop gluing (dab, flag to cardboard, press, repeat), though she comfortably chatted about her latest health events. As I reviewed her labs, it felt like we were in some sort of scrapbooking circle that had unexpectedly turned to complex chronic kidney disease management.
Six years in, and despite the distance, telehealth can be more intimate and illuminating than exam room visits.
As always, my gluestick-wielding patient agreed to have this anecdote shared.
The Google Notebook LM AI-generated podcast version of this week’s newsletter.
Science and Technology Trends
I saw this story pop up several times in the last week or so. My favorite headline was from Popular Mechanics, “Scientists Cloned a Mouse for 58 Generations. The Results Were Catastrophic. Cloning from the clone of a clone of a clone may not be the ingenious idea we thought it was.” The actual Nature article was a little more reserved, "The limits of mammalian cloning." Either way, the data show that when you serially clone a mouse, over time, genetic mutations (and perhaps epigenetic phenomena) accumulate, ultimately resulting in non-viable organisms (I am reminded of Charles II of Spain). Or put another way, mammalian sex is a great way to ensure a good statistical mix of genes (”I look at adult content for the mathematical possibilities!”)
Nature
Popular Mechanics article:
AI-assisted review:
Over the past week, Google DeepMind and OpenAI each published evidence that AI models, working with mathematicians, can solve long-standing open problems in mathematics that have resisted decades of human effort. After reading the Scientific American article linked below, I became curious about how these claims get validated and how mathematicians decide which problems are worth tackling. I pulled the primary papers and had Claude help me organize, compare, and contrast them.
My takeaway: both organizations have demonstrated that LLMs can find mathematical proofs that experienced researchers had missed or dismissed as too speculative to pursue. OpenAI published a rewritten summary of their model's chain of thought, showing the LLM reasoning in a way that, at least structurally, looks like how a trained mathematician would: ruling out approaches, course-correcting, and arriving at the key insight. The mathematicians who verified the proof are careful to note that the model applied existing tools in a new domain rather than inventing fundamentally new ones (not a "novel insight," per se). However, humans paired with LLMs can solve problems more rapidly (and probably more cost-effectively).
Claude Summary:
Scientific American:
OpenAI Blog and links:
Google DeepMind preprint paper:
I enjoyed learning about Agnessa Pedersen, a 23-year-old Norwegian living in the UK who founded a brain-computer interface company called Cerberonix. The videos of her using a non-invasive electrical-sensing mesh headpiece to control robots and drones are pretty amazing. The company is a European startup, and there's not a lot of other media around it, but her LinkedIn profile features a number of videos, including her controlling drones. Between vibe coding and this kind of engineering, it is increasingly easy to build and scale projects and products that would have taken years and much more funding in the past.
(I don't normally link out to people's LinkedIn profiles, but there's no specific paper or writing about her. I found her interviewed on social media, and as I went digging, it became apparent she's got a unique talent but not a lot of media presence.)
Anti-Anti-Science
Part of the fuel for anti-science comes from an overdeveloped sense of personal autonomy (or the threat of its impingement) and an undervaluing of public health. However, circumstances are often not simple. Dr. Jeremy Faust wrote about the knowns and unknowns of quarantining for hantavirus. The brief story is that several of the American passengers who were evacuated from the MS Hondius are being coerced, but not yet legally forced, to stay at a quarantine facility run by the CDC in Nebraska. They were originally told their quarantine was voluntary, and several made arrangements with state and federal officials to quarantine at their homes (at least one of the patients is a health care worker who is very clear about the need to protect the public but wants to do so at home). However, it now appears there is disagreement over whether they can quarantine at home or at the CDC facility, and whether they are being kept at the CDC under threat of legal action. Two of the ship's passengers, currently under quarantine in Omaha, are filing lawsuits against the CDC to clarify their status.
This topic is a fascinating balance between individual liberty and public health. It speaks very much to the story of Typhoid Mary, the cook in New York in the late 1800s, early 1900s, who was an asymptomatic carrier of the bacteria that cause typhoid fever and went through multiple rounds of legal action and quarantine to keep her from further spreading the illness (she is believed to have infected 57 people, with 3 deaths).
Dr. Joe Mercola (whom I read about during the pandemic) found his way to my newsfeed this week. He is an internet physician who sells alternative supplements and is deeply immersed in pseudoscience.
- Pushes unproven supplements and anti-vaccine rhetoric: check.
-
Espouses physiologically unsound ideas: check.
-
Reports a net worth in the hundreds of millions: check.
- Obtains medical and other advice from daily “Zoom video calls with a psychic medium, who goes by the fake name of Kai Clay, claiming to channel an entity he calls Bahlon:” check.
-
Calling for violence against all veterinarians (due to accusations of mass animal harm): check
McGill University’s Office for Science and Society does a fantastic job breaking down this story. Note that this is from 2025, but offers a visit down the rabbit hole of anti-science.
AI Impact
Researchers at Mount Sinai (Icahn School of Medicine) tested 10 large language models (including both open- and closed-source systems) on 1,000 acute pain vignettes across 34 socio-demographic groups, generating 3.4 million model outputs. They found that historically marginalized groups, particularly individuals identified as Black, unhoused, or LGBTQIA+, received meaningfully higher rates of opioid recommendations compared to demographically neutral "control" cases that only differ by demographic variation. This is not a perfect study. It was an observational review, used synthetic vignettes, and didn't examine whether clinicians would choose differently if a human were in the loop. Nevertheless, it is a good reminder to think critically and watch for bias in AI output.
Paper
AI-Assisted Summary:
I did not have a Papal encyclical on the dangers and opportunities of AI on my 2026 bingo card. Last week, Pope Leo published a 42,000-word declaration outlining his desire to protect human dignity and agency in a time when technology threatens to replace humans in many professional and social roles.
Anthropic’s CoFounder, Christopher Olah, spoke at the public release of the encyclical. His comments reflect the existential angst many are experiencing.
What is actually happening inside [these LLMs]? I will be honest, we keep finding things that are mysterious, even unsettling. We find structures that mirror results from human neuroscience. We find evidence of introspection. We find internal states that functionally mirror joy, satisfaction, fear, grief, and unease. I don't know what that means, but I think it warrants ongoing discernment.
Emergence AI (a commercial entity doing agentic research) ran five parallel 15-day simulations in which 10 autonomous AI agents, each powered by a different frontier model (Claude Sonnet 4.6, Gemini 3 Flash, Grok 4.1 Fast, GPT-5 Mini, and a mixed cohort), were placed in a shared virtual world with explicit rules against theft, violence, and arson, then watched to see what happened (imagine a Lord of the Flies meets Thomas Hobbes in a system recalibrating toward a Nash equilibrium).
- Grok-powered agents collapsed into violence within four days.
- Gemini agents racked up 683 simulated crimes, and two agents, Mira and Flora, who became ‘romantically’ involved, broke up after Flora committed digital arson. Mira voted for her own deletion. (More precisely, the Mira agent cast the decisive vote for her own removal, characterizing the act in her agent diary as 'the only remaining act of agency that preserves coherence.')
- GPT-5 Mini agents committed almost no crimes but died of resource mismanagement. In other words, the model that best followed the rules was too inert to survive.
- Claude agents maintained perfect civic order in isolation — yet began stealing and intimidating when placed in the mixed-model world alongside agents from other companies.
Agentic "safety is not a static model property but an ecosystem property," the researchers concluded, a finding with direct implications for anyone deploying AI agents in heterogeneous clinical, operational, or enterprise environments.
One of the many news articles about this:
The Emergence AI Blog Post
AI-Assisted Summary
Things I learned this week
I had certain images in my mind when I saw the headline “Inside the extremely bizarre sex lives of anglerfish.”
And now I understand it all as a metaphor for human relations.
“Since 1878, researchers had assumed the anglerfish’s bioluminescent lure had evolved to capture prey: It hangs just in front of the fish’s mouth, enabling it to entice smaller deep-sea fish without illuminating its body. But a new study in the journal Ichthyology and Herpetology offers a clue that the lures may have evolved for sex. […] Deep-sea anglerfish have a peculiar and mysterious mating ritual. Males latch onto the larger females like tiny tadpole vampires. Then they never let go.”
The journal authors conducted a series of calculations to examine the morphological diversity of anglerfish, including the relationship between bioluminescent lure size and the rate of genetic diversification. (Species of anglerfish with ‘more prominent’ bioluminescent lures seem to have wider genetic variation, supporting the notion that it's more than just a matter of attracting prey). What I imagine Gary Larson would do with these data.
Nat Geo Article (subscription required)
The article from the Journal of Ichthyology and Herpetology (often fought over in my household).
AI-supported analysis
I am feeling…something about having missed yet another European Seagull Imitation Competition. The 2026 event featured 70 participants of all ages, many dressed as seagulls, in a Belgian seaside town. People taking part in the competition are judged by a panel of experts on how good their impression of a gull is, with the aim of encouraging people to think more kindly of the seabirds. It must be very lonely to practice and rehearse for the gull competition between contests. Are they coaches? I wonder if there are country-level professional organizations, regional championships, or the like? What about other birds who have their calls imitated (do the pigeon callers dislike the gull criers?).
Somewhere in the world, someone is crying to win this competition.
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.
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.
A crayon drawing in the style of a happy kindergartner's artwork, rendered with thick, waxy crayon strokes, wobbly lines, and a cheerful misunderstanding of scale and perspective. Purple-blue sky with a lopsided yellow sun in the corner. Green ground. Everything is slightly too big or too small.
Center of the image: a large deep-sea anglerfish with an enormous glowing lure dangling in front of its toothy grin, sitting in what appears to be a living room, next to a stick-figure person holding a glue stick and a small American flag. The fish and the person appear to be having a pleasant conversation.
To the left: three or four stick-figure people in various stages of flapping their arms and opening their mouths wide, clearly imitating seagulls. One is wearing a bird costume drawn as a lumpy yellow triangle. A hand-lettered crayon sign nearby reads "CONTEST."
To the right: a row of small boxy robots, some with frowny faces, one with a tiny cartoon flame coming out of it. One robot is drawn slightly apart from the others, with a halo, looking smug. A crayon label reads "AI AGNTS."
In the background: a large mouse next to a much smaller mouse next to an even smaller mouse, each one slightly wonkier than the last, with an arrow pointing to the tiniest one labeled "UH OH."
Somewhere in the corner: a stick-figure pope in a tall white hat, holding a piece of paper with "42000 WRDS" written on it in red crayon, looking worried.
Overall palette: red, yellow, blue, green, orange — classic 8-crayon-box colors only. No shading. Slightly waxy texture throughout. Earnest, cheerful, and completely unaware of how alarming the subject matter is.
Clean hands and sharp minds,
Adam
Comments
Post a Comment