Week of June 1, 2026
A friend showed me a ghost video on Friday. An actual video, shot in the garage where he cares for rescued animals, of a translucent orb drifting up through the frame. He shared it with the quiet confidence of a scientist sharing peer-reviewed data. That orb, he explained, was the soul of one of the animals he rescues, released now into a larger consciousness he is certain awaits all of us. (He owns electromagnetic field detectors and loans them out so friends and family can "know" they have contacted deceased loved ones and pets.) I kept a straight face, despite myself; I've known him for a long time, and I think I understand him. Like so many of us, he seeks comfort against the finitude and chaos of existence.
We all have little mental lies we tell ourselves. On Saturday, somewhere around hour three of power washing my backyard hardscape, I started feeling like the task was Homeric, man against entropy, once again fighting the second law of thermodynamics (see prior newsletter openings re: leaf blowing). I knew I was only temporarily moving dirt, but for a brief window, I appeared to have controlled the universe with 3,200 psi of water through a 40-degree spray nozzle. My patient has his orbs. I have my pressure washer. Neither of us is going to win, and both of us feel better for the fight.
The Google Notebook LM AI-generated podcast version of this week’s newsletter.
Science and Technology Trends
I found videos of long-range, 24 GHz wireless power beaming technology powering a drone continuously for 96 hours (at a 30-foot distance). While several vendors are working on this, the videos were provided by GuRu Wireless. The videos are impressive, and a 2023 article from the Electrical Engineering Times captures the evolution of this technology from concept and early trials (thank you, Nikola Tesla) to the latest updates. Essentially, you need a tightly focused RF or laser beam aimed at rectennas (an antenna and rectifying circuit that captures RF or microwave radiation and converts it into DC power) or a photovoltaic receiver (that converts light into DC power) on the end device. The potential use cases for longer‑range wireless power (meters to kilometers) include autonomous robots, vehicles, and appliances, but reducing drone weight (by reducing battery requirements) is of great interest (think industrial, military, and commercial uses).
2024 EE Times Article about the Field:
Videos:
AI-assisted Review:
Eli Lilly published data from Phase 3 clinical trials of its next-generation GLP-1, retatrutide. The data demonstrated up to 30% weight loss over 80 weeks (in the highest dose group), but also greater side effects than with semaglutide or tirzepatide (Ozempic and Zepbound). Further data regarding the composition of weight loss (muscle versus fat) has yet to be published. Nevertheless, one might imagine that a 30% weight drop would be very appealing to many and of great interest to consumers when it hits the market.
Retatrutide is a triple agonist drug (GLP-1, GIP, GCG), interacting with 3 different receptors that impact appetite, hunger, and reward. Quintuple agonist drugs are currently in animal (mouse only) testing. I note that the number of agonist protein sequences in the weight loss peptide drug market is looking a lot like the men's disposable razor market (5 razors!).
AI-Assisted summary
Anti-Anti-Science
I feel like peptides are everywhere - media coverage, patient questions, telehealth storefronts. In April, the FDA announced proposed changes to its ban on some ingredients used by compounding pharmacies for peptides, potentially allowing a wider range of telehealth-enabled compounding pharmacies to sell peptides (of course, grey-market peptides of unknown consistency and quality have been available for years).
In late April, the New Yorker published an overview of the peptide market and the problems with purity and consistency.
One of the more interesting intellectual side quests that I found myself on was understanding the FDA's administrative definition of a peptide versus a biologic agent. Under FDA rules, any amino acid-based therapeutic with 41 or more amino acids is classified as a biologic, which triggers different regulatory and patent protections - and closes the door to compounding pharmacy versions. (Here is an X-post worth reading for context on why Lilly is suing the FDA over the biologic/peptide 40/41 amino acid boundary - most of the weight loss drugs (like Ozempic are less than 41 amino acids - hence the compounding).
I’ve started compiling a tracking list of peptides. I want to understand which ones might have some degree of actual clinical data and which ones don't. I compiled this list using tools to scrape grey-market compounding pharmacies (remarkably, I found 60-70 compounds on the first pass), then used various AI tools, such as Gemini and OpenEvidence, to gather any clinical data supporting their use in various clinical scenarios.
https://docs.google.com/spreadsheets/d/1gk95JcCbbTw5MeoMvpUv9PNotQGn_aIEM7_hmih0MPI/edit?usp=sharing
The peptide I hear the most about is BPC-157, a gastric mucosal peptide that's supposed to have "amazing" healing properties. There's no large-scale clinical data, and those human studies that do exist are methodologically problematic. Assuming the compounding market for these becomes more mainstream, I suspect BPC-157 will be one of the more popular peptides. Here is a clinical overview of BPC-157:
As Open Evidence summarized: “BPC 157 remains an investigational compound with no validated dosing regimen, no pharmaceutical-grade formulation, and insufficient human data to support clinical use. The primary barrier to translation is not the absence of preclinical biological activity but rather the absence of fundamental pharmaceutical science — characterized formulations, validated pharmacokinetics, and rigorous clinical trials.”
I am sympathetic to people who struggle to solve problems these compounds claim to solve. Nevertheless, there are many unknowns: How do you know if a compound is manufactured with consistent purity and safety? Do we have enough data about a substance to know if, when, what type and in whom side effects might occur? And, if the drug is well-manufactured and safe, is it effective for the intended use?
AI Impact
Dennis Gray, my eleventh-grade English teacher, assigned a weekly task: pick a word from your reading and draw a word map — branches extending from the center word, where each branch is a different definition. Then offer a gradient of synonyms for each definition, mapping out the varying degrees of connotative meaning. Essentially, an analog version of visualthesaurus.com, before the internet existed.
AI turns out to be extraordinarily good at discovering relational structure in large datasets and encoding it as a kind of navigable geometry. Researchers refer to the output as a knowledge graph. The interesting thing is that AI doesn't build these graphs from rules; it finds them as latent statistical patterns in the data.
This week, I came across two papers that explore what that looks like in practice, one for food, one for fiction.
Developers at KAIKAKU.AI, a startup with ambitions in AI-powered hospitality software and a mission that reads like a religious crusade (seriously, read their manifesto), built Epicure: a family of three machine-learning models that encode 1,790 food ingredients into a shared mathematical space by mining 4.14 million multilingual recipes. The models spontaneously learned not just which ingredients are cooked together, but regional cuisine identity, flavor chemistry, and nutrient profiles, all without being told to look for any of those things. By mathematically relating recipes and ingredients at scale, the models developed a relational network that approximates the connotative gradient of cuisine across geography. You can rotate "rice" thirty degrees toward South Asian cooking and recover curry leaf, urad dal, and fenugreek as nearest neighbors. The result is a navigable geometry of flavor; a compass for chefs building precise fusion cuisine.
Review of the Article:
Google DeepMind and University of Maryland researchers built StoryScope: a tool that maps short fiction across 304 structural narrative features (including thematic explicitness, sensory density, chronological discontinuity, moral polarity, how characters reveal emotion, and whether a narrator addresses the reader directly). Applied to over 61,000 stories written by both humans and (publicly available) LLMs, StoryScope correctly distinguished between AI and human authorship 93% of the time. The key insight was that word and style-based AI detectors are increasingly fragile; AI models can be fine-tuned to mimic human prose. Structural narrative choices, however, are much harder to fake. AI fiction, it turns out, has a characteristic shape: over-explained themes, tidy resolutions, and reluctance to leave moral ambiguity unresolved.
AI-Assisted Review of the Article:
Neither paper is a landmark. Both are preprints, neither peer-reviewed nor validated against a truly independent ground truth. For instance, the food paper used an AI model to generate the very labels it then evaluated.
What I find fascinating about both is the same thing that made Mr. Gray's word maps interesting in eleventh grade: the idea that meanings of things are enhanced and refined by the thing's relationships to other things. AI is very good at quickly finding this structure at a scale and with nuances that would take humans many hours — and at encoding it in a form one can navigate. The food paper offered a flavor dial. StoryScope generated a narrative fingerprint.
Nevertheless, the AI derived navigable structure will be limited by context and validation. The stakes are low when the AI output is cuisine or short fiction. But the bias and gaps in validation and context are more concerning when applied to more consequential topics, like clinical decision-making or legal reasoning. A model that learns "what good clinical documentation looks like" from a decade of overworked residents' discharge notes has learned the shape of that practice, not the shape of good medicine. The knowledge graph is only as honest as the data that built it, and validating it against a rigorous, human-curated ground truth is harder than it sounds and often not done.
In the meantime, this kind of thinking is fun to explore, especially when the subject is flavor. As some loyal readers know, a certain newsletter author may have recently conducted a similar exercise with his adult beverage collection.
Things I learned this week
Occasionally, I find an article that seems interesting for one reason but turns out to be interesting for another. Over the weekend, I read an Atlantic article whose title was interesting enough (“The best free Restaurant bread in America”), but turned out to be an autobiographical stream-of-consciousness exploration of survey methodology, the history of bread, an homage to Diet Coke, and a memorial to the author’s father. It felt like I was reading author Caity Weaver’s somewhat self-absorbed writing about writing - a modern mashup of David Foster Wallace and James Joyce. Save for the Red Lobster reverence (which I will let my loyal readers consider), the article turned a survey about restaurant bread into a glimpse inside (what read like) the thoughts of someone with ADHD.
Atlantic Article:
AI-assisted Summary:
Oddly, while I was still pondering Caity Weaver’s religious-like love of Diet Coke, I found this article. The Wall Street Journal highlights a cultural tension between Diet Coke fans and Coke Zero fans. I am not a cola drinker (I am not a fan of artificial sweeteners). However, the way both Diet Coke and Coke Zero drinkers speak about their respective beverage choices (each has a taste profile as complex as a fine wine) left me feeling simultaneously out of touch and with FOMO. Either way, it pairs perfectly with the Atlantic bread article.
WSJ Article:
AI-assisted review of DC vs CZ:
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 17th-century Mughal imperial court miniature recording events of great importance: ghost orbs, pressure washing, wireless drone power, and the Diet Coke vs. Coke Zero theological dispute. Full prompt here:
Clean hands and sharp minds,
Adam
Comments
Post a Comment