Week of July 28, 2025
Single sentence “micro stories” filled my last week of patient care and travel. I love quotes or comments that can imply a complete narrative. These comments represent four separate conversations with four separate people.
"I went on vacation with my wife and two other women and came home with Shingles."
"Dr. Weinstein, you don't have to stop the pills, I could just watch CNN to make my blood pressure go up."
"My wife was molested by donkeys on Grand Turk."
"I usually wait for the second date to share that I am employed in the axe-throwing industry."
I will let your mind fill in the backstories.
---
Listen to a Google Notebook LM A.I.-generated podcast of the newsletter with two virtual "hosts."
https://drive.google.com/file/d/1cAkb57BgZDk1_TzdrSlkAp4cxl-Gs9a2/view
------
Science and Technology Trends
Derek Thompson, a writer for The Atlantic, published a fantastic overview of GLP-1 drugs (such as Ozempic or Mounjaro) written for non-medical readers. He is asking the question (like I have) - should we all be on these drugs?
https://www.derekthompson.org/p/why-does-it-seem-like-glp-1-drugs
There was only weak medical evidence to support the "10,000 steps a day" craze of the 2010s. Australian researchers recently published a metaanalysis examining the relationship between step counts and eight health outcomes: all-cause mortality, cardiovascular disease, cancer, type 2 diabetes, cognitive outcomes, mental health outcomes, physical function, and falls. They identified over 80 prospective studies (with over 160,000 patients' data) using device-measured step counts conducted within a 10-year time window (2014-2025). Their analysis correlated a 47% reduction in all-cause mortality and a 25% reduction in cardiovascular disease in individuals who accumulated 7000 or more steps per day. Large metaanalyses can often demonstrate correlations between variables, but they usually have inherent bias and confounding due to the combined data. Nevertheless, I have very few qualms about overemphasizing the value of walking more.
Social Media:
https://x.com/thelancet/status/1948339629398110703?s=42
The Article:
https://www.thelancet.com/journals/lanpub/article/PIIS2468-2667(25)00164-1/fulltext
Claude Summary with analysis of strengths and weaknesses:
https://claude.ai/public/artifacts/53954554-48ad-4869-9b11-77255948e6c3
Anti- Anti-Science Articles of Note
The consequences of politics entering the clinical exam room: "Pregnant Mother in Tennessee Denied Care for Being Unmarried. [Tennessee's] 2025 Medical Ethics Defense Act allows physicians to deny care to patients whose lifestyles they disagree with."
https://nashvillebanner.com/2025/07/20/doctor-denies-pregnant-woman-care/
Living with AI.
Last week, I was reading about the distinction between AI "world models" and pure prediction models, and how this relates to Moravec's Paradox. Today's AI tools are remarkably effective at predicting the next word or pixel (really a rapid, incredibly good tool for learning patterns of numbers in the form of tokens)—like LLMs (ChatGPT, Claude, Gemini)—but far less capable when it comes to building a generalized understanding of the physical world. For instance, training an AI to intuit gravity or Newtonian physics even when trained on large volumes of real-world object motion remains a significant challenge. This gap highlights the difference between high-volume predictive capabilities and accurate "world models"—internal representations of how the world works that allow for extrapolation to new, unseen scenarios.
While the distinction between prediction and modeling isn’t the full story, it does account for some of the reason we can’t just put an LLM in a robot and have it “work.” This is described by Moravec's Paradox: tasks that are easy for humans (e.g., walking, grasping objects) are challenging for machines, while tasks we find difficult (e.g., chess, theorem proving) are relatively easy for AI. As Hans Moravec put it, "It is comparatively easy to make computers exhibit adult-level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility."
Related links:
AI Daily Brief on the difference between predictions and world models (and what it means for AI tools):
https://open.spotify.com/episode/3xDbpFyciioodzEdI4iMDd?si=mGygbFrWSNOJREIR_7FcFg
and
A Forbes article on Moravec's Paradox:
https://www.forbes.com/sites/stevendesmyter/2024/07/04/moravecs-paradox-and-the-age-of-robotics/
and
https://www.appventurez.com/blog/moravec-paradox-in-robotics-and-ai
Dozens of researchers from OpenAI, Google DeepMind, Anthropic, and Meta published a joint paper warning that our time window to monitor AI and understand AI reasoning could be closing. The sophistication of the models (which are becoming better at obfuscating their logic and reasoning), the speed of development, and an evolution in the models ability to reason without language (i.e. predictive logic that uses only math) make it increasingly likely that humans will lose the ability to understand how AI tools solve problems, and thus become less safe and transparent. [I'm sorry, Dave, I can't do that.]. The Paper calls for standard tests of reasoning, building transparency safeguards into new models, and architectural changes that permit monitoring.
The Paper:
https://tomekkorbak.com/cot-monitorability-is-a-fragile-opportunity/cot_monitoring.pdf
Article about the Paper:
https://venturebeat.com/ai/openai-google-deepmind-and-anthropic-sound-alarm-we-may-be-losing-the-ability-to-understand-ai/
Claude's summary of the Paper and the article combined:
https://claude.ai/public/artifacts/992d9dca-dd6a-4f4c-aceb-319bb36d95b4
Infographics
Epidemiology etymology! (Did you know we have effective vaccines that help with some of these infectious diseases?)
https://www.etymologynerd.com/uploads/1/5/8/8/15888322/epidemiology.png
Things I learned this week
My favorite headline this week: "Neanderthals were not 'hypercarnivores' and feasted on maggots, scientists say. Researchers believe humans' closest relatives may have stored meat from their kills for months before eating it." The backstory is more interesting - Neanderthal bones tend to have extremely high nitrogen isotope levels, which suggest they consumed as much meat as hypercarnivores like lions and wolves. In general, Homo Sapiens (and Homo Neanderthalensis?) cannot tolerate such high-protein diets without developing health issues. Researchers discovered that maggots feeding on putrefied meat contain nitrogen levels up to 43.2 parts per million—far higher than any typical food source. As such, one plausible explanation for the high nitrogen signatures is that Neanderthals consumed a nutritionally complete diet of lean game meat that was “enriched” in fat (from maggots).
The headline: https://www.theguardian.com/science/2025/jul/25/neanderthals-feasted-maggots-science-nutrition
Nature article:
https://www.nature.com/articles/d41586-025-02334-y
A Claude Summary of all the sources above:
https://claude.ai/public/artifacts/1a9ccd01-1e83-4e1c-aa40-67057487719f
Thanks to a loyal reader, I now know that pigeons can read radiology images. When AI fails us, we can revert to more natural forms of augmented intelligence. See "Pigeons (Columba livia) as Trainable Observers of Pathology and Radiology Breast Cancer Images" from 2015. I note that this Paper uses the phrase "Flock Sourcing" - combining responses from multiple birds, to achieve 99% diagnostic accuracy of mammography 99%
https://pmc.ncbi.nlm.nih.gov/articles/PMC4651348/jcx√ h z3
https://claude.ai/public/artifacts/1b1e8cf4-254d-4617-a1c8-68af1f484838
Related: there is a surprising quantity of journal articles about training pigeons to perform visual processing tasks:
https://pubmed.ncbi.nlm.nih.gov/?linkname=pubmed_pubmed&from_uid=26581091
AI art of the week
A visual mashup of topics from the newsletter.
I use ChatGPT to summarize the newsletter, suggest prompts, and make the images.
"A Roman-style floor mosaic, crafted from small stone tesserae in warm earth tones. The scene is divided into three parts. On the left, four arched confessionals carved into stone, each containing modern figures with surreal attributes: one sits beside a donkey holding a suitcase, another clutches a blood pressure monitor while a news screen hovers, one holds a heart-shaped axe, and the last stands nervously beside luggage. In the center, a towering humanoid AI figure with a circuit-patterned body and a glitching rectangular screen for a head stands solemnly. A sign on its chest reads "WORLD MODEL OFFLINE." Above it, stylized molecular symbols (N, P) and geometric motifs float. On the right, two Neanderthals feast on rotting meat crawling with maggots at a stone banquet table. A pigeon in a white robe observes with a scroll or tablet in wing, acting as a scholarly figure. The background includes stylized cave art of animals and hunting scenes. The entire composition is symmetrical, tile-based, and consistent with the visual language of ancient Roman floor mosaics."
ChatGPT
https://drive.google.com/file/d/1pFXsQ2DqtbPuz18lKyBcWSAX6kKBaJCn/view
Gemini
https://drive.google.com/file/d/1NAEUlqpO2m-ayDDntuZckFhGVI0x4It9/view
Grok
https://drive.google.com/file/d/1clUmZQf-g_-xz2s7wFKWwYvpYzZNPDnV/view
---
The Pandemic Mitigation Collaborative (PMC) utilizes wastewater viral RNA levels to forecast four-week predictions of COVID-19 rates.
-COVID rates are trending up from previous weeks, with a lot of geographic variation.
-The PMC website now has regional, state, and international heatmaps for COVID rates in more specific geographies.
https://pmc19.com/data/
based upon https://biobot.io/data/
Wastewater Scan offers a multi-organism wastewater dashboard with an excellent visual display of individual treatment plant-level data.
https://data.wastewaterscan.org/
----
Clean hands and sharp minds,
Adam
Comments
Post a Comment