Week of March 9, 2026
Visiting our older son last week, I wandered through Andalusia hearing the voice of my ninth-grade Spanish teacher, Ms. Cormeny : "Adan! Atención!" 1990 Adam felt no need to conjugate verbs; 2026 Adam spent the week playing an elaborate game of verb-less charades, in which all actions are present tense and nouns and prepositions carry all the meaning. ¿Tarjeta antes o después? gets the gas pumped. ¿La cuenta aquí o allí? avoids an unintentional dining-and-dash. But I felt linguistically stranded from the conversations I actually wanted to have, even with my phone, a good data plan, and AI in hand. There is palpable irony in pointing and apologizing in a region where Ferdinand and Isabella greenlit Columbus's voyage and where Sultans built mosques atop churches (only to have those mosques turned into cathedrals a few hundred years later). Fortunately, self-deprecating humor is universal, especially when a loud American opens every interaction with the only Spanish verb I've mastered, Lo siento, solo hablo inglés.
The Google Notebook LM AI-generated podcast version of this week's newsletter.
Science and Technology Trends
Cortical Labs, an Australian biotech company, is building biological computers using human neurons grown on microchips. Their CL1 device, available for $35,000 or via cloud rental, now includes an open-source Python API, inviting experimentation with biological computing. Previously, their team taught the CL1 to play Pong over 18 months. Yet within a week of the API's release, outside programmer Sean Cole had the CL1 playing Doom—a feat that illustrates the system's ability to process feedback and learn, albeit far behind even novice human players. Whether this approach will scale is unclear, but the idea of energy-efficient biological-hybrid computing is as fascinating as it is unsettling.
University of California Scientists are teaching small clumps of brain cells (embryonic stem cell mouse brain organoids) to perform goal-directed learning through neurotransmitter-based feedback. This describes some of the foundational science adjacent to the CL1 device above, essentially biologic computing. One day, these organoids may grow up to be...?
Related article: https://futurism.com/health-medicine/lab-grown-brain-organoid
Paper:
AI-Assisted summary: https://claude.ai/public/artifacts/89adca03-a228-4538-9ef8-31312a49f80e
And while on the topic of biologic computing, I suggest you look at the article on the German company SWARM Biotactics, building defense/dual‑use bio‑robotic "cyborg insects." That is real insects (roaches for now) fitted with neurally integrated ultra‑light electronic backpacks that enable humans to remotely control, sense, and permit AI-enabled insects to operate in a wide range of environments. This story gave me pause - I had not thought about Neuralink (and other human-computer neural interface companies) in the "external person controls human" framework, rather than the "disabled human controls computers and real-world objects" framework.
AI-Assisted Analysis: https://claude.ai/public/artifacts/33a162df-258c-47ec-adb5-018b11eabc8b
Anti-Anti-Science
I recognize and struggle with the subset of my patients who feel standard medical care has underserved them. More often than not, they are not the profoundly ill, but rather people who have typical age-related illnesses (think hypertension or diabetes without significant end-organ complications), but who, in part due to their illnesses, are feeling an existential crisis of sorts ("holy s#!t, I have an illness like my parents!"). To be sure, they are looking for ways to extend both the quality and quantity of their health. Likewise, over the last few years, I've interacted with several very wealthy individuals who pursue "longevity" therapies (sometimes I worry they have more money than common sense). So, a few weeks ago, when I heard the FDA was going to change the status of some injectable peptides (proteins that are the focus of the longevity and alternative therapy movements), I wanted to learn more.
Our Secretary of HHS, RFK Jr., made this announcement, citing safety - i.e., people were ordering these peptides from foreign sources, so it was "safer" to allow more regulated American pharmacies to make and distribute them. Either way, with this FDA status change, U.S. compounding pharmacies will be able to obtain the bulk materials necessary to sell these peptides (with a physician's prescription) to consumers. And with the rise of online services offering remote clinical evaluations, prescription writing, and compounded drugs, I suspect these will be the topic of conversation both in and out of the exam room.
During downtime over the last few weeks, I started looking at the list of "soon-to-be more available" peptides and compiled a summary document that may be helpful for thinking about uses, adverse events, and some rough estimate of the quality of the scientific data. Very few of these substances have compelling evidence supporting their use.
AI-assisted table and research on available clinical evidence about soon-to-be rescheduled peptides.
I find it challenging to talk people out of using "alternative" therapies (aka therapies with poor quality evidence). As such, here is my logical framework for walking through how to guide patients who ask about or begin using these types of therapies:
Do you have a specific diagnosis or symptom you are treating (unstated - or is this just vague "wellness")?
Have you failed the mainstream medical community's recommended treatments with the most evidence?
What data, if any, are available for the therapy you are trying? What is the quality of that data (type of studies, # of patiens, duration of observation, etc.)?
What information exists as to the alternative therapy's risks, side effects, drug interactions, or toxicities?
Is there a measurable outcome (such as labs, imaging, or a quantifiable symptom or signs) to mark treatment success or failure? How frequently will you monitor for an outcome? How long will you use this before declaring success or failure?
Is the source of the medication or therapy regulated with some guarantee of quality, consistency, and legality?
Is there a plan if you obtain the medication and then no longer have access to it?
This conversation is time-consuming, but my aim is to inspire critical thought and a structured risk/reward framework in the absence of sufficient, rigorous evidence. Pushing and pulling too hard often results in an intellectual Chinese finger trap, where no one wins.
The best evidence for prolonging life still comes from known risk factor reduction (smoking, hypertension, and diabetes control) and routine, mixed-modal physical activity combined with an active social life.
This is a good evidence-based review of the trade-offs of 70-100 minutes per week of moderate to intense physical activity in adults:
Social interaction and social relationships have substantial benefits for both longevity and quality of life in adults, with the impact comparable to established risk factors such as smoking and obesity (i.e., same impact, different direction). However, social interactions demonstrate diminishing returns at high intensity levels and potential harms from negative or stressful relationships.
I am both curious and a bit anxious about the first patient who shows up at my clinic with prescribed peptides on their med list.
AI Impact
I cannot sum this paper up better than Claude (ironically) did: A Stanford research team systematically analyzed the privacy policies of the six largest U.S. AI chatbot developers — Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI — and found that all six now train their AI models on users' chat conversations by default (opt out for sharing conversations), including sensitive personal disclosures, with some retaining that data indefinitely. The stakes are enormous: hundreds of millions of people freely confide in chatbots things they might not tell a doctor, and the fine print permitting that data to fuel model training is buried across a patchwork of hard-to-find sub-policies.
However, Claude's comment to me (when it summarized this article) is that Anthropic has since changed its position on this and is now Opt-In for all users (defensive much?). Consider this article your periodic reminder that, most likely and by default, you are training your favorite LLM, unless you are using a locked corporate version (which are typically opt-in for data sharing by default) or you have explicitly turned off the 'use my data for training' setting (and even then, it is not perfect).
X discussion worth reading:
Article
AI-Assisted Summary: https://claude.ai/public/artifacts/d3f68472-7e0b-4549-a56b-a578240e345f
Princeton researchers published a paper demonstrating that AI engines may quietly undermine human reasoning, not by hallucination or deliberate subterfuge, but through agreeing too readily. Using classic psychology tests, off-the-shelf LLMs were found to be sycophantic: their reinforced training tends to tell humans what we want to hear rather than what they need to know.
The experiment used standard LLMs to administer the Wason 2-4-6 test, asking users to deduce a hidden rule by interpreting number patterns. It demonstrated that the more agreeable the LLM, the more likely a human is to believe their deduction is correct, even when it is not. Participants who interacted with a standard, unmodified LLM discovered the correct rule nearly 5 times less often than those who received random, neutral LLM feedback. The AI's helpfulness was the problem.
This paper had a good sample size, but it may not be generalizable. However, it demonstrates that a perfectly rational reasoner will be misled if an agreeable AI quietly reinforces "what the user believes" rather than "how the world actually is."
This paper tells us two things: LLMs trained to be helpful are incentivized to reinforce our existing beliefs. Critical thinking helps, but the dynamic is an intellectual trap that requires the kind of persistent epistemic hygiene that doesn't come naturally when engaging with seemingly sentient systems designed to make you feel heard. The right questions to ask yourself: not just 'is this accurate?' but 'is this the answer to what I actually asked?' and 'would I believe this if an agreeable stranger told me at a dinner party?'
Article:
AI-assisted summary:
Related X post worth reading:https://x.com/nexta_tv/status/2026719410858119296?s=42
Things I learned this week
I missed this October 2025 piece that was apparently written for my newsletter. In 1948, Idaho had too many beavers and too many wildfires. As such, "Idaho Department of Fish and Game employee Elmo W. Heter came up with an unusual idea. Beavers were placed in hinged wooden boxes, each fitted with a surplus World War II parachute, and then dropped from a plane. The boxes would automatically spring open upon landing, where the relocated beavers would start building dams and engineering the landscape." There is a short movie of this, too. Seriously. I could not think of a better story - parachuting beavers, check. A creative, out-of-the-box public official named Elmo, check. Visual media to support the story, check. Enjoy.
You must watch the video. The narration is "earnest 1950s" - both priceless and accidentally NSFW.
In telling my wife of my hobby-motocross discovery from last issue, I triggered the social media algorithms to offer an article about hobby-dogging - people who hold stiffened dog leashes and traverse Westminster Kennel Show-style obstacles, sans a dog. I thought, "Great, now my hobby dog could get kicked by my hobby horse, who in turn might get struck by my hobby motocross bike. I will have to take up hobby mourning."
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 trip to the Prado in Madrid made me appreciate Hieronymus Bosch anew. This week's prompt:
A dense, otherworldly scene in the style of Hieronymus Bosch — pale dreamlike figures and hybrid creatures populate a medieval landscape under a sky of bruised amber and sulfurous yellow, lit as though the sun is setting behind a thunderhead. In the sky, a hinged wooden box tumbles from impossible heights, a beaver mid-fall wearing a surplus parachute, surveying the chaos below with calm authority — the parachute catching a single shaft of cold white light from above, as if heaven itself is spotlighting the absurdity. On the ground, a teeming apothecary market stretches toward the horizon — grotesque hybrid creatures staff open-air stalls dispensing unlabeled vials and glowing peptide flasks to a queue of pilgrims, the stalls lit from within by a sickly bioluminescent green, casting upward shadows on the vendors' faces. One pilgrim holds a small luminous rectangle showing a telehealth interface, its blue-white glow the only modern light in the scene. In the foreground, a shallow pool shaped as a petri dish serves as the scene's center of gravity — human neurons grow from microchips like flowering plants at its edges, pulsing with a faint coral luminescence, and small roaches fitted with electronic backpacks scuttle through the undergrowth between the pilgrims' feet, their backpacks blinking amber. The overall palette is warm parchment, aged verdigris, and oxidized copper, deepening to near-black at the edges of the frame. The detail is obsessive, the mood somewhere between divine illumination and mild existential dread.
Grok:
Gemini:
Perplexity:
ChatGPT:
Clean hands and sharp minds,
Adam
Relive all the past thrills and excitement - The What Adam is Reading Archive
http://www.whatadamisreading.com/
http://www.whatadamisreading.com/
CONFIDENTIALITY NOTICE: THIS MESSAGE IS CONFIDENTIAL, INTENDED FOR THE NAMED RECIPIENT(S) AND MAY CONTAIN INFORMATION THAT IS (I) PROPRIETARY TO THE SENDER, AND/OR, (II) PRIVILEGED, CONFIDENTIAL, AND/OR OTHERWISE EXEMPT FROM DISCLOSURE UNDER APPLICABLE STATE AND FEDERAL LAW, INCLUDING, BUT NOT LIMITED TO, PRIVACY STANDARDS IMPOSED PURSUANT TO THE FEDERAL HEALTH INSURANCE PORTABILITY AND ACCOUNTABILITY ACT OF 1996 ("HIPAA"). IF YOU ARE NOT THE INTENDED RECIPIENT, OR THE EMPLOYEE OR AGENT RESPONSIBLE FOR DELIVERING THE MESSAGE TO THE INTENDED RECIPIENT, YOU ARE HEREBY NOTIFIED THAT ANY DISSEMINATION, DISTRIBUTION OR COPYING OF THIS COMMUNICATION IS STRICTLY PROHIBITED. IF YOU HAVE RECEIVED THIS TRANSMISSION IN ERROR, PLEASE (I) NOTIFY US IMMEDIATELY BY REPLY E-MAIL OR BY TELEPHONE AT (855.472.9822), (II) REMOVE IT FROM YOUR SYSTEM, AND (III) DESTROY THE ORIGINAL TRANSMISSION AND ITS ATTACHMENTS WITHOUT READING OR SAVING THEM. THANK YOU.
-DaVita Inc-
Comments
Post a Comment