Week of May 18, 2026
As the only exit-row passenger on a recent flight home from a conference, I was both “willing and able" to open the exit hatch and the sole bottleneck in the “event of an evacuation.” Would my "willingness and ability" devolve into a petite dictatorship of who gets out? Without the usual unspoken agreement amongst a more typical plurality of exit-row passengers, I was alone in asserting authority; the flight attendants had made clear they "would not be present to help." Society is held together by unstated conventions and contracts. And I realized that none of these prevents me from demanding a Venmo payment (with a tip) for facilitating an urgent over-wing exit.
The Google Notebook LM AI-generated podcast version of this week’s newsletter.
Science and Technology Trends
The Verge covered interesting announcements out of former OpenAI CTO Mira Murati’s company, Thinking Machines Lab. They are developing “interaction models,” LLMs that will let people collaborate with AI in a more natural way, continuously taking in audio, video, and text while searching and processing. Current LLM interaction is turn-based: you speak, it responds, repeat. I see a lot of potential for this kind of tech in clinical contexts - LLMs that summarize data contemporaneous to a clinician's review of the chart or that actively pull patients' historical information while the patient begins chatting with an intake chatbot. Please recognize that this is more of a press release than research, but it is still interesting.
AI-Assisted review of the article, the tech, and the application to clinical environments:
Last week, several readers wrote to discuss their concerns about the hantavirus articles (could this be the next pandemic?). While I often reference their podcast, Jess Steier, DrPH, Izzy Brandstetter Figueroa, MPH, and Elana Pearl BenJoseph, MD, MPH, offered an excellent analysis that highlights the art of science communication: how to walk the line between too much and too little anxiety provocation when it comes to this outbreak (apparently they have been accused of downplaying the severity, or as they put it, “calmmongering.”)
And it looks like Ebola is back in the news too. Ground News is a great resource to gauge the range of headlines:
This Scientific American article about NASA’s latest plans to return to and then inhabit the moon feels exciting and ambitious.
NASA's "Ignition" plan, unveiled at a March 2026 event by administrator Jared Isaacman and elaborated at an April APL conference, sets out three phases: a crewed landing by 2028, a permanent south-pole base receiving astronauts every six months by 2032, and a nuclear-powered outpost by 2036. The architecture calls for 79 launches, 73 landers, 10 moon buggies, 12 "hopper" drones, four habitat modules, and a 20-kW reactor. Two private landers — SpaceX's Starship variant and Blue Origin's Mark II — compete to deliver crew, neither of which has yet completed an operational lunar landing.
Assuming we can deal with lunar dust, figure out construction equipment for the moon, and develop materials that are suited to long-term radiation exposure, we should be good to go. And, every time I see these kinds of articles, I think of one of my favorite childhood books, Rhoda Blumberg’s "The First Travel Guide to the Moon: What to Pack, How to Go, and What to See When You Get There.”
Based on how many times I checked it out of Fort Garrison Elementary’s library, I suspect I am well-prepared for moon tourism.
AI-assisted summary:
Anti-Anti-Science
Dr. Jeremy Faust offers a very sharp analysis (an almost classic morbidity and mortality (“M&M”) conference) of Louisiana physician/ Senator Bill Cassidy, who lost his reelection primary last week. As always, Dr. Faust is thoughtful and critical of the anti-science crowd.
Related, Dr. Makary, the FDA commissioner, resigned “over concerns about the administration’s decision to authorize fruit-flavored e-cigarettes, an action he opposed, according to four people familiar with the matter. Dr. Makary told those close to him that he could not in good conscience approve flavored vapes, given their appeal to young people, and would not do something he did not believe in.”
Here is more on the FDA’s authorization of an expanded range of vape flavors.
Both are great examples of the ironic destruction that happens when you compromise your principles for power in a system built on transactional politics.
AI Impact
This week, I'm focusing on AI-supported medical transcription, probably the most common AI-enabled tool used by U.S. clinicians. In a variety of settings, health care professionals are now using audio capture of clinician-patient interactions. AI cleans the audio, transcribes it, and synthesizes it into clinically-formatted documentation. Increasingly, these tools are also queuing orders and actions for clinician oversight and signature. The goal is to ease some of the cognitive burden and workload for clinicians, get their eyeballs off the screens, and turn them back toward eye contact and patient engagement, rather than worrying about capturing the note.
The technology comes with trade-offs: security and privacy, accuracy, user satisfaction, and costs (see full list of concerns). One of the most critical of these is accuracy. Consider that these tools are now queuing care actions — medications, consult requests, follow-ups, one step removed from automated action. Also consider that accuracy is genuinely hard to measure: there is no well-documented baseline for how accurately humans capture conversations relative to what was actually said. Without that baseline, every claim about AI-scribe accuracy is being made against human-reviewed audio or transcripts.
I found three pieces this week that capture the gamut of how this is playing out.
First, the Canadian province of Ontario recently issued an auditor general's report on the use of these technologies in government and across the broader provincial health sector. Reading the report (which is admittedly dry) illustrated the challenges of evaluating this technology: there are no standard benchmarks, no agreed framework for establishing accuracy, and a long list of reasonable criticisms of how organizations like the Ontario government contract with these vendors. I asked Claude to summarize the report, and the summary surfaced a number of testing-methodology issues: the audit aggregated findings across all 20 vendors rather than disaggregating performance, used only two simulated pre-recorded test files, and conducted no end-to-end testing that connected scribe outputs to actual clinical outcomes.
- AI-assisted Analysis: https://claude.ai/public/artifacts/5190f38e-06e7-43d5-8f95-73fb9f43da97
- The report itself, for the insomniac: https://www.auditor.on.ca/en/content/specialreports/specialreports/Special_Report_AI_en.pdf
Second, Futurism, a tech-focused website, reviewed the Auditor General's report. It is a useful example of how the broader technology press handles AI-in-healthcare stories: surfacing the most dramatic findings (hallucinations, the Auditor General asking her own doctor to verify her transcript), while skipping over the less provocative issues, such as procurement-process critique, security-audit gaps, and the methodological caveats that make the audit's findings both more concerning and more nuanced.
- Article: https://futurism.com/artificial-intelligence/ai-scribe-hallucinating-medical-issues
- AI-assisted Analysis: https://claude.ai/public/artifacts/462ede87-2e8d-4436-8df3-aa6a82579c24
Third, a JAMA+AI podcast interview with Vincent Liu, MD, MS, Chief Data Officer of The Permanente Medical Group at Kaiser Permanente. The conversation focuses on Kaiser's enterprise-scale deployment of these tools and on how valuable they are to physicians within that system. It is not designed to be a critical analysis of the trade-offs involved, but it reflects the framing the medical community is hearing in industry-facing media about ambient audio capture and AI scribes.
In light of the unusually wide range of recent media on this technology, I asked Claude to review all three as a compare-and-contrast exercise, including a section on what each genre of source structurally cannot say.
- Three-way synthesis: https://claude.ai/public/artifacts/275b5a91-0d06-4db2-b841-adb67897b620
One last point: I could not fully determine whether Claude's analysis here was biased. One might reasonably assume an AI would be less critical of AI technology than a human reviewer would be. However, I, too, have bias; I use these tools in my own office visits, and I have mixed reviews, some of which are captured in the articles above.
Things I learned this week
I learned that sleep deprivation combined with a bar mitzvah party at Dave and Buster's, featuring a buffet of pub food such as “cheese planks,” sparked intense curiosity about the range of breaded, fried cheese products available. Luckily, any thirst for knowledge (and actual cheese-plank-induced thirst) can be satisfied with AI and an open bar. The cheese plank falls somewhere along the spectrum from cheese balls to cheese sticks.
Gemini’s helpful infographic: https://drive.google.com/file/d/1f7LmQkXCK_J0oD7tNmtGsNvI92ORnvcw/view?usp=sharing
There is much to learn when you find an article entitled “Japan's ‘robot wolves' in high demand to scare off bears.” I did not know that Japan was having a surge in bear attacks (including more than 50,000 sightings, 13 fatal attacks in 2025-2026, and over 200 additional injuries). I did not know that wolves, much less robot wolves, scare off bears. And, of course, I did not know that one could buy OR rent a robot wolf. I would be very interested in renting a robot wolf companion for Skelly, our 12-foot skeleton, next Halloween.
Yahoo News article:
The Japanese distributor (I highly recommend translating to English). Robot wolves are solar-powered!
Somehow, I missed this article from August 2025, “The Criminal Enterprise Run by Monkeys: A cabal of furry thieves snatch iPhones and other valuables from visitors to a temple in Bali—and trade them for mangos.” It appears that criminal activity (including racketeering) is a learnable trait, inherited through quasi-Darwinian social evolution. The best part: there's a built-in pricing model. “Primatologists have found that the macaques steal belongings to use as currency to trade with humans for food. Some monkeys can distinguish between objects we highly value (smartphones, prescription glasses, wallets) and those we don’t (hats, flip flops, hair clips)—and will barter accordingly.”
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 wide landscape painting in the unmistakable style of Bob Ross — wet-on-wet oil technique, soft palette knife mountains, phthalo blue lake, titanium white cloud highlights, happy little evergreens in the foreground, a warm sunset sky in oranges and pinks, the whole scene radiating gentle PBS-television reassurance. But the landscape is quietly interrupted by impossible objects in the deadpan style of René Magritte: a single emergency exit door from an airplane floats unsupported in the middle of the sky, its red handle clearly visible; a full moon hangs above the mountains with a small lunar base built on its surface, complete with tiny landers and a glowing pink dot indicating a nuclear reactor; a row of three identical men in bowler hats and dark suits stands in the foreground meadow, each one staring blankly forward in Magritte's signature pose; a macaque sits calmly on a rock in the middle distance holding an iPhone with a price tag still attached; a single breaded fried cheese object (a "cheese plank") levitates over the lake, casting a faint shadow on the water; and along the bottom edge of the painting, hand-lettered in a careful cursive script, the words "Ceci n'est pas un cheese plank." A solar-powered robot wolf, clearly mechanical with visible solar panels, stands beside the happy little trees. The overall composition is calm, balanced, and reassuring in the Ross manner, but the impossibilities accumulate quietly. No melting. No chaos. Magritte's rule: each strange element is rendered with photographic clarity, sitting plainly in the landscape as if it belonged there. Oil painting texture throughout. Wide aspect ratio.
Learn about Rene Magritte: https://www.moma.org/artists/3692-rene-magritte
Clean hands and sharp minds,
Adam
Comments
Post a Comment