A cardiology week of things that moved and things that did not

The peer-reviewed papers behind last week's heart meeting arrived, and the companion analysis of the big lipoprotein(a) trial closes the last escape route: the patients whose marker fell most did no better. A programme that trained dispatchers, first responders and residents across sixty North Carolina counties left cardiac-arrest survival where it was. Two Chinese trials of a clot-dissolving drug injected into the artery after mechanical clot removal were null on ninety-day function. Against that, a single combination pill delivered by a retrained primary care team more than doubled blood-pressure control in Colombia and Jamaica, and a weight-loss drug moved an invasively measured pressure in heart failure. And a preprint reports that the enzyme that makes fat is also an RNA-binding protein.

Connecting the dots

A week ago a phenotype-first screen in patient-derived dopamine neurons nominated fatty acid synthase as a drug target and declined to explain how lipid remodelling fixed calcium handling. Today a preprint reports the same enzyme binding thousands of transcripts and setting how efficiently they are translated. Speculation: a second, non-metabolic job for the protein is one place that missing mechanism could sit, and it would mean an inhibitor of the enzyme is doing more than one thing.

A matched comparison of systematic reviews written with model assistance against reviews written without it finds the gap concentrated in the steps where a human frames the question, not in retrieval. Next to it, a Stanford course whose exercises models can now solve in one shot moved the homework to patches judged by the maintainer of the repository, and a Google release pairs a long-context drafter with a reviewer that sees the draft and its sources but not the conversation that produced them. The three share a premise: the judgement has to come from outside the chain of reasoning being judged.

Stewart Brand's new book separates "sustainability", a goal, from "sustainment", a plan with a budget line. In the same week a Pennsylvania hospital was ordered to close after its electricity provider moved to cut power over an unpaid bill. Upkeep is the part of a health system that accumulates invisibly until it is the thing that ends the service.

An essay on a formerly poor Chinese county argues its AI system worked because the county had first pooled data, staff and accountability across three tiers of care. Separately, Africa CDC's public ask of the Global Fund's new head is procurement that guarantees demand for African manufacturers. Both put the institutional plumbing before the product, and both leave who pays for the last tier unanswered.

New concepts

The enzyme that makes fat also binds messenger RNA

"Here we present the first endogenous RNA interactome of fatty acid synthase (FASN)" … "identifying over 6000 binding sites across over 3500 protein-coding genes" … "with a striking preference for the 5' UTR" — bioRxiv

The direction and the geometry both matter. Lowering the enzyme gave "a preferential reduction in the translational efficiencies of FASN-associated mRNAs", "an effect shown to be dependent on binding location and site number" (bioRxiv). Fatty acid synthase is the central lipogenic enzyme and an oncology drug target; the authors keep the causal claim loose, since crosslinking-based mapping sees only the contacts it can crosslink.

Why it matters: a protein that has been treated as a metabolic node is reported to sit on the 5′ ends of thousands of transcripts, which changes what an inhibitor of it is expected to do.

Biosafety law classifies organisms, and synthetic biology has stopped making them

"Synthetic biology increasingly produces hybrid constructs that blur the boundaries of organism-based biosafety regulation." — Bouchaut and Asveld, Trends in Biotechnology

Their proposal changes the unit of classification: "We propose using the term 'biological entity'", "defined by its capacity to interact meaningfully with biological systems or to facilitate biological processes", "enabling more coherent and proportionate oversight" (same).

Why it matters: it is a screening category proposed inside the European regulatory debate, for things that current law cannot classify because they are not organisms.

Atlases and datasets

Regulatory architecture of the ageing human brain, by cell type

Whole-genome sequencing in several hundred brain donors set against bulk and single-nucleus RNA, proteomics and metabolomics, which "identified 49,532 trans-regulatory events supported by shared genetic signals" (medRxiv, CC BY). The finding is an asymmetry between cell types, not the count:

"neuronal cis-regulatory signals preferentially associate with coordinated glial trans-regulatory responses" — medRxiv

Backed by an orthogonal check: such events were enriched among neuronal genes that respond to astrocyte co-culture. Access to the underlying genotypes runs through a data use agreement.

Why it matters: a genetic variant acting in a neuron shows up as a coordinated program in the glia around it, so the cell where a risk variant acts and the cell where its consequence is visible need not be the same.

Carotid plaques, labelled by whether they caused symptoms

Fifty-three plaques, seventeen newly profiled by single-nucleus sequencing plus four public single-cell datasets.

"symptomatic plaques showed cell-state shifts toward later or terminal states in osteochondrogenic smooth muscle cells" … "genetic risk was enriched primarily in vascular wall cells" — medRxiv

The nuclear data also "uncovered a depletion of fibroblasts in symptomatic plaques" (same).

Why it matters: the authors put inherited risk in the wall cells rather than in the immune infiltrate, for a lesion whose imaging features do not predict which one will cause a stroke.

Methods

A defined panel of surface-receptor mutants that both maps and fishes for phages

"Here we describe a SURFAce Receptor (SURFAR) platform consisting of a set of defined Pseudomonas aeruginosa strains and mutants for rapid characterization of phage receptors." — bioRxiv

About forty-six strains in 96-well format — pilin groups, flagellar and lipopolysaccharide mutants, transporter knockouts, biofilm-matrix mutants — mostly in one background that cannot acquire CRISPR-based resistance, so a strain that resists is likely resisting at the receptor. The second use is the new one: running an enrichment from wastewater through the panel selects for phages that enter by a chosen receptor, and the authors report "novel phage genomes from wastewater" (bioRxiv).

Why it matters: phage therapy is assembled case by case, and which molecule a phage uses to get in is reported inconsistently because there has been no shared set of strains to test it against.

Clinical trials

Pelacarsen: the patients whose marker fell most did no better

"For participants on treatment, a primary end point occurred in 434 participants (10.4%) randomly assigned to placebo and 380 participants (9.1%) assigned to pelacarsen (hazard ratio, 0.98; 95% confidence interval [CI], 0.85 to 1.12)" — Nissen, Cho et al., NEJM Evidence

And the dose-response check, which is the new information:

"there was no association between pelacarsen treatment and the primary end point either in participants with reduction in lipoprotein(a) greater than or equal to the median (170 nmol/l) (hazard ratio, 1.09; 95% CI, 0.92 to 1.29)" … nor "in any of the evaluated key subgroups including classification by median baseline lipoprotein(a), LDL cholesterol, high-sensitivity CRP, or time from qualifying cardiovascular disease to enrollment" — same

Why it matters: lowering an inherited risk marker further did not buy anything in this trial, so the open question is no longer the dose but whether the molecule, the patients or the duration of exposure were wrong.

A single pill and a retrained team doubled blood-pressure control in Colombia and Jamaica

In forty primary care clinics, "systolic blood pressure decreased by 19.1 mm Hg, from 145.9 mm Hg at baseline to 126.8 mm Hg" with the intervention, a "between-group difference of −10.1 mm Hg (95% CI, −13.3 to −6.9 mm Hg; P < .001)" — JAMA. Control was reached by about seven in ten patients against three in ten under usual care. The package was "physician and nurse training, health coaching, home blood pressure monitoring", plus audit and feedback (same).

Why it matters: old drugs, combined into one tablet and handed to a redesigned team, produced one of the week's largest effects in two middle-income countries.

Opening the artery, twice, without moving the patient

In basilar occlusion, independent ambulation at ninety days reached "40.0% of the tenecteplase group versus 50.0% in the control group (OR 0.67; 95% CI 0.34-1.29; P = 0.23)" in what the authors call a "prematurely terminated trial" (Nature Communications). In large-vessel occlusion, good function reached "40.0% of the tenecteplase group versus 46.8% of the control group (unadjusted odds ratio 0.76, [95% CI 0.50-1.14], P = 0.19)" (Nature Communications). The companion phase 2 trial had roughly doubled reperfusion.

Why it matters: the drug did what it was meant to do to the blood vessel, and the patients were no more independent for it.

Teaching sixty counties to resuscitate

"The education strategy did not result in more patients surviving than patients with usual care." — STAT

The programme trained dispatchers, defibrillator-carrying first responders and community members; the trial release reports survival with good brain function at about nine per cent in intervention counties against 9.4 per cent elsewhere, and bystander CPR at 44.8 against 42.3 per cent (AHA). Training reached about five per cent of residents. A German three-arm trial of how far to alert volunteers found they "arrived before EMS in 31%, 57%, and 58% of missions in the 600 m, 2500 m, and SMA groups" and measured no patient outcome (JAHA).

Why it matters: reach decides who gets there first, and getting there first has not yet been shown to be the binding constraint on survival.

A weight-loss drug moved an invasive pressure in heart failure

The primary endpoint was wedge pressure measured during exercise: "estimated difference, −10.6 mmHg; 95% CI: −13.3 to −8.0; P < 0.0001", with haemodynamic remission in "53.7% versus 7.4%, odds ratio = 14.1, 95% CI: 3.0−134.6, P < 0.001" — Borlaug et al., Nature Medicine.

Why it matters: in a week of markers that fell without consequences, here a directly measured mechanical quantity moved, in eighty-one patients over a year.

Eleven years later, invasive and conservative strategies still look the same

"All-cause 11-year mortality was not significantly different between groups (25.8% invasive vs 26.9% conservative; … adjusted hazard ratio [AHR], 0.97 [95% CI, 0.86-1.09])." … "By bayesian analysis, the posterior probability of more than 1 percentage-point survival benefit to invasive vs conservative was low (18%)." — JAMA

Why it matters: the extended follow-up removes the usual defence that benefits of revascularisation appear late.

A screening ECG read by a model, randomised

In 14,665 adults without known atrial fibrillation, a model-read ECG decided who got a monitor. "The primary outcome occurred in 54 intervention participants (0.74%) versus 35 control participants (0.48%; odds ratio, 1.54 [95% CI, 1.01-2.37]; P=0.046)", and the authors report their own fragility check: "Hypothetical reclassification of 1 intervention event to a nonevent changed the primary analysis P value to 0.057" (Circulation). Detection of fibrillation and starting of antiarrhythmic therapy did not differ.

Why it matters: a cheap test triaging an expensive one is a reasonable design, and this randomised version of it produced a result that one reclassified event would erase.

Hospitals

A hospital closed over an electricity bill

"The Pennsylvania Department of Health revoked Sharon (Pa.) Regional Medical Center's license, effective Nov. 17," "and ordered the hospital to close after its electricity provider moved to cut power over an unpaid bill," "according to a Nov. 4 order obtained by Becker's." — Becker's Hospital Review

In Switzerland, which cancer drug you get can depend on your insurer

(translated from French) "Their reimbursement in fact varies from one insurer to another." … "can access to the latest advances in medicine depend on the insurer a patient belongs to" — Medinside (original: « Leur remboursement varie en effet d'une caisse à l'autre. » … « l'accès aux dernières avancées de la médecine peut-il dépendre de l'assureur auquel un patient est affilié »)

Journalists ranked insurers using a federal report on case-by-case funding of innovative cancer treatments. The published piece names no insurer, drug or figure.

Japan's hospitals change tack on an unrecoverable tax

(translated from Japanese) "On the problem of non-deductible consumption tax at medical institutions, the All Japan Hospital Association has shifted from its earlier demand for 'taxation' to studying a new compensation scheme that keeps insured care tax-exempt." — CBnews (original: 「全日本病院協会は、医療機関の控除対象外消費税問題について、従来の「課税化」要求から、保険診療の非課税を維持した新たな補填方式の検討へと舵を切った。」)

Insured care is exempt from consumption tax, so hospitals pay the tax on what they buy and cannot pass it on. A task force is to report by year-end. More than half of Japan's private hospitals ran an operating loss in the last financial year.

Why these matter together: three different mechanisms — a utility bill, an insurer's discretion, a tax that cannot be recovered — decide what care actually reaches people, none of them clinical.

Medicine concept of the day

The modified Rankin Scale: the seven-grade score that decides whether a stroke trial is positive

Both of today's null stroke trials measured the same thing: a clinician's grade of global disability at ninety days.

"0 – No symptoms at all." … "2 – Slight disability; unable to carry out all previous activities but able to look after own affairs without assistance." … "3 – Moderate disability; requiring some help but able to walk without assistance." … "6 – Dead." — Ammar et al., Frontiers in Neurology 2026, CC BY

It was not designed for this. "a scale introduced by Dr. John Rankin in 1957 would become the primary outcome scale for almost all acute stroke trials", and "The Rankin scale was modified to its current form by Charles Warlow and others as part of the UK-TIA trial in the 1980s" (Broderick et al., Stroke 2017). It survives because "it covers the entire range of functional outcomes from no symptoms to death" and "its categories are intuitive and easily grasped by both clinicians and patients" (same).

Most trials then cut the seven grades in two, and where they cut moves with how sick the patients are: "Randomized trials of more severe strokes due to large vessel occlusion used dichotomous cutoffs of 0–2 vs 3–6" (same). The alternative is to use the whole distribution, because "Dichotomizing outcome scales reduces computational complexity, but discards substantial outcome information, artificially privileges only a single health state transition as clinically meaningful, and often reduces study power" (Saver, Stroke 2007). It is not a free choice: "dichotomous approaches to the mRS can be statistically significant when ordinal approaches are not" (Broderick).

And the grade itself is an interview. "A limitation of the mRS has been the subjective determination between categories", with pooled agreement of "a kappa = 0.46 and a weighted kappa of 0.90" (same) — raters rarely disagree by much, but often disagree by one grade, which is exactly where the cut-point sits. Structured interviews help: "Inter-rater reliability with the mRS is moderate and improves with structured interviews (kappa 0.56 versus 0.78)" (Banks and Marotta, Stroke 2007).

Today's two trials were the attempted replication of one small positive trial, in which good function "was 59.0% (36/61) with alteplase and 40.4% (21/52) with placebo (adjusted risk difference, 18.4%; 95% CI, 0.3%-36.4%; P = .047)" and which "was terminated early for inability to maintain placebo availability and enrollment rate" (CHOICE, JAMA 2022). That trial used the strictest cut-point; the replications used a looser one.

Why it matters: in acute stroke, a therapy counts only if it moves a patient across a grade of an interview scale three months later, and the choice of where to cut that scale can decide the answer on its own.

Worth knowing

Replacing a fixed laboratory cutoff with a model orders more tests, not fewer

For vitamin B12, "The 221 pmol/L cutoff had sensitivity 33.4% (9.4% at eGFR <45)", and "At a 5% risk threshold the full model detected 83.2% of cases versus 33.4% for the cutoff-7.6 additional tests per additional case" — Clinica Chimica Acta. Most of the discrimination comes from adding kidney function, and the author is explicit that "external validation is required before clinical use" (same).

Why it matters: the cutoff in use today misses two-thirds of cases and nine-tenths in kidney disease; the model's value here is finding them, and anyone who wants it to save tests has to argue that separately.

A urine strip found tuberculosis the machine missed, in people with no symptoms

In routine Nigerian HIV programmes, "Of the 17,155 AHD patients tested with LF-LAM on first urine samples, 5,212 were TB LF LAM positive, giving a diagnostic yield of 30.4%", and the strip added cases on top of molecular testing, "giving an incremental yield of 39%". The line that matters most for how patients are found: "Among 5,212 LF LAM positives, 2,074 (41%) of them were not presumptive through WHO four-symptom screening" — Journal of Clinical Tuberculosis. The authors' own framing: "Tuberculosis (TB) diagnosis remains the weakest link in the care cascade" (same).

Why it matters: the four-question screen that decides who gets tested would have excluded two in five of the people this test found. The extra positives were not confirmed by another method, so the figure measures disagreement rather than accuracy.

Health systems and policy

Hundreds of Ebola patients have left care after an armed attack

The attack "comes at a critical moment in the response to the ongoing Ebola outbreak." … "Health facilities have been attacked and patients, including hundreds people with confirmed or suspected Ebola infection, have fled" — Africa CDC

The continental agency calls for a "corridor of peace, protection of civilians, patients, health workers and health facilities" (same), and gives no casualty count.

Why it matters: an outbreak response depends on knowing where infected people are; this one has now lost that for an unknown number of them.

Africa CDC's terms for the Global Fund: buy from African manufacturers

"The Global Fund's purchasing power can help create predictable demand for quality-assured medicines" … "This requires predictable financing, greater domestic resource mobilisation and realistic transition arrangements" … "Global health investments should strengthen national planning, budgeting, procurement and financial management systems" — Africa CDC

Why it matters: the Global Fund is among the largest buyers of HIV, tuberculosis and malaria commodities, so its tender rules decide whether regional manufacturing is financeable.

Being biomarker-selected makes a cancer drug nine times likelier to get a conditional European approval, and no faster

"CMA was granted to 44.2% of targeted therapies versus 7.9% of non-targeted therapies (OR 9.24; 95% CI 2.46-34.71; p<0.001)" … "Greater use of CMA among targeted therapies means that more indications enter the post-authorisation period with confirmatory evidence obligations outstanding." — Journal of Cancer Policy (CC BY)

Time to authorisation was no shorter.

Why it matters: a companion test changes the kind of approval a drug gets rather than its speed, and moves the evidence that justifies it to after it is on the market.

Big in machine learning

Nothing today.

Startups and deals

Three Chinese rounds this weekend sit on the same theme, instruments and automation rather than software. A Suzhou company says it "has a new AI-based molecular technology for early blood screening of brain disease, with mass spectrometry as the core platform", and that it "has started developing a fully automated pre-processing system for mass-spectrometry samples, hoping to solve the complexity of MS sample preparation with automated equipment", alongside work on a domestically made spectrometer (translated from Chinese; originals: "检力康掌握了基于人工智能的脑病血液早筛诊断新型分子技术，其将质谱作为脑病血液检测的核心技术平台" and "目前，检力康已经开展质谱样本全自动前处理系统研发，希望通过自动化设备解决质谱检测前处理流程复杂的问题") — VCBeat.

A protein-design company describes itself as "an AI biopharma factory with an AI super-agent at its core, a closed dry–wet loop as its production line, and a multimodal product library as its output" (translated from Chinese; original: "一家以AI超智能体为核心、以干湿闭环为产线、以多模态产品库为输出端的AI生物制药工厂") — VCBeat. Its later round includes a laboratory-robotics maker among the investors, which is a supplier buying into a customer whose product is the loop.

A Shanghai cortical visual prosthesis company reports that "the company is only two years old and its cumulative financing exceeds RMB 500 million" and that it "has completed 7 IIT cases, with no serious adverse events observed" (translated from Chinese; originals: "截至目前，公司创业仅两年时间，累计融资总额已超5亿元。" and "截至目前，明视脑机已完成7例IIT，且均未观察到严重不良事件") — VCBeat. Seven investigator-initiated cases are a safety series, not efficacy.

Repos and tools

A benchmark for brain-MRI foundation models runs them under identical conditions: it is "an open, living benchmark that evaluates structural brain MRI foundation models through frozen feature probing", which "Runs every model as a frozen encoder on the same subjects and preprocessing", and where "Extraction and scoring run through CI, so adding a model, dataset or task is a pull request" — Neurostars. The authors say they work with model authors to use each model's intended preprocessing, which is where most apparent differences between such models come from.

A benchmark in the opposite direction perturbs tasks rather than standardising them: it "changes the instructions or starting code of 89 DeepSWE tasks", and the spread between models grows "from 20.9 percentage points on DeepSWE to 48.3 points on MaxxBench" — unLAP. Two models within a point of each other on the original land thirty points apart once the tasks are altered.

A Google release splits drafting from reviewing: "Design with an elephant. Review with a goldfish", where "Goldfish reviewers examine the draft and referenced source material without the earlier conversation" — GitHub (Apache-2.0). A small research tool takes the same shape for the literature: "Paper in, cited commercialization triage out", with "Deterministic code retrieves and validates sources first; six LLM stages then reason over that frozen evidence", and "The weighted total is computed by code, not the model" — GitHub (MIT).

Also: a sequencing-cohort alignment viewer, "Web Server displaying SAM/BAM/VCF file and using the IGV.js API", MIT-licensed (jvarkit); and a llama.cpp fork claiming to "Run 284B DeepSeek-V4-Flash (<59GiB RAM, 44.4 tok/s pp, 9.1-14.7 tok/s)" and "30B MoEs (25-111 tok/s) on a pure 64GB CPU at 256K ctx (0 GPUs)" (GitHub), which would put a large model on ordinary laboratory hardware with no data leaving it.

AI: using it well

Model-assisted systematic reviews score lower than matched human ones

"Mean 13–item AMSTAR–2 scores were 7.58 and 9.63, respectively." … "In multivariable–adjusted analyses, AI–assisted reviews scored 1.64 points lower (95% CI −2.37 to −0.91); noninferiority was not demonstrated." — medRxiv

Seventy-seven reviews reporting model assistance were matched to human ones on topic, timing and design. Disclosure is thin where it would matter most: "Among AI-assisted reviews, model version or date was reported in 42.9%, prompts or parameters in 55.8%, and data-governance or ethics considerations in 9.1%" (same).

Why it matters: it is the first matched comparison of the method rather than of a model, and it suggests the question to ask of any such review is which steps the model touched.

When the homework can be one-shotted, the grader moves outside

"By the next run, Mihail figured students could one-shot a lot of those exercises." … "So this time, part of the homework is contributing to an open-source project." … "The person who maintains the repo reviews the student's patch." — Latent Space

The same show notes cite a study of software teams: "Teams with high AI adoption merged 98% more PRs, while their PR review time was 91% longer" (same).

Why it matters: both halves describe the same displacement, from producing work to judging it.

A Tokyo clinic runs its own models on a Mac Studio

(translated from Japanese) "medical information is processed by a 'local LLM' running on the in-house server, and patient information is not sent to external AI services." … "Many people use it so the summary works as a 'table of contents' for the huge chart, and they go back to the original chart when needed." … "After all, an aid is still an aid. I don't copy and paste." — ITmedia AI+ (originals: 「院内のサーバで動かす「ローカルLLM」で診療情報を処理し、外部のAIサービスには患者の情報を送らないという内容だ。」「サマリを膨大なカルテの“目次”にして、必要に応じて元のカルテを見返す、という使い方が多いです」「やっぱり補助は補助です。コピペはしません」)

The director built it himself; chart summaries run overnight as a batch job, and the same machine reads handwritten forms and transcribes dictation.

Why it matters: the summary is used as an index back into the record rather than as text to reuse, which is a usable answer to the provenance problem in a clinic with no vendor and no cloud.

A reminder from the other direction: a write-up circulating this weekend counts secrets left in public repositories by coding agents whose permission prompts were dismissed, under the title "Claude Code's "Don't Ask Again" exposes passwords in 1,122 public GitHub files"; from the tool's own team in the thread, "Historically Claude Code has not done secret scrubbing (there are many good tools for that)" — Hacker News.

Other frontiers

An implant that burns off excess blood sugar with electricity

"harnesses electricity to drive a chemical reaction that converts glucose into a more easily excreted byproduct" … "holding the animals in the normal range for days at a time without a single dose of medication." — IEEE Spectrum

The electrodes are gold flakes dotted with platinum and run at "half the voltage of a AAA battery—0.75 volts"; the device "remains functional for at least 3 months" in mice. One of the researchers quoted is blunt: "The technology therefore is a long way from clinical reality" (same). Rabbits and pigs are next.

Why it matters: it is a therapy with no drug in it — a catalytic reaction run inside the body at low voltage — which raises a different set of questions than dosing does, starting with anyone whose kidneys must clear the product.

Big tech

Two feed items worth noting from the weekend: "Anthropic bought commitments for 1.7GW of gas-fired generators from equipment supplier OnePWR for $3.4B", an AI lab buying generating equipment directly rather than power; and "DeepSeek says its open-source agent app Harness hit 1.54M DAUs on November 4", with roughly 92 trillion tokens a day — both via Techmeme, citing the Financial Times and Bloomberg.

Corners of the web

A bug report posted on a bioinformatics server describes a motherboard-level fault that "corrupts data to SATA&NVME when using 64bit DMA" on certain AMD machines, with a kernel workaround now merged; the patch thread traces it to addressing above 58 bits and notes it "can technically happen without IOMMU enabled as well" — genomic.social, linux-pci. Silent corruption of files on the hardware that stores sequencing data, fixed in the kernel rather than in any pipeline.

People and good examples

Isaak Sigal (1927–2026), Gas Institute, Kyiv

"In the 1960s and 1970s, I. Ya. Sigal conducted the first research in the USSR on the formation of nitrogen oxides within a flame." — Gas Institute, NAS of Ukraine

He carried it to hardware: "technologies, methods, and burners with reduced nitrogen oxide formation, which have been implemented on more than 100 power plant boilers and 900 large-capacity hot-water and industrial boilers", and boiler designs now "supplying thermal energy to more than 10 million residents" (same). His institute remembers "a man of exceptional intellect and tolerance, with whom it was easy to discuss and resolve the most complex scientific and engineering challenges in the energy sector" (same).

Alexandre Chorin (1938–2026), Berkeley

"Chorin developed a revolutionary new approach, which became known as the projection method" … it "cleverly decoupled the computation of velocity and pressure" — UC Berkeley Letters & Science

The consequence, in the obituary's words: "This tactic allowed engineers and medical researchers to effectively calculate many key processes, including aerodynamics, blood flow, and atmospheric pressure fronts" (same). A former student, on supervision: "He truly was the perfect thesis advisor" (same).

Reads and ideas

A poor Chinese county that reorganised before it bought software

An essay on Midu, in Yunnan, which set up a county medical consortium in 2022 and switched on an AI system across three county hospitals, eight township centres and eighty-nine village clinics at the end of 2024:

"The answer lies not in technology, but in institutions." … "Midu got the organization right before it got the technology" by "building a consortium that unified data, personnel, and accountability across three tiers of care" … "The infrastructure for data to flow — and for accountability to follow — existed before any algorithm was turned on." — Jeff Xiong

On payment, one step and an open question: "In June 2026, China added 12 AI-assisted diagnostic services to the national medical insurance catalog" (same). The piece is opinion, not an evaluation, and its outcome figures are the author's own; it also names, without developing, "a vacuum of legal liability when AI makes a mistake".

Book of the day

Maintenance: Of Everything, Part One — Stewart Brand (Stripe Press)

"Let "maintenance" mean the whole grand process of keeping a thing going." … "Maintenance is absolutely necessary and maintenance is optional." — Introduction

It gets skipped because "the necessity of maintenance accumulates invisibly and gradually", so "we shirk it, defer it, fail to budget time or money for it". "Defer now, regret later." (same) The opening chapter reads a 1968 solo round-the-world race as a contest of three styles: "Knox-Johnston's style was: 'Whatever comes, deal with it.' And he did." "Crowhurst's was: 'Hope for the best.' It killed him." "Moitessier's was: 'Prepare for the worst.' It freed him." (Works in Progress) Moitessier's rule: "if you don't fix something when you first see it beginning to fail, it is very likely to finish failing" (same).

The sharpest objection comes from a historian of maintenance: "For Brand, maintenance seems to be a solitary act," "more about personal success and fulfillment than tending to a shared world", and not all upkeep deserves praise — "Maintaining a gas-guzzler or a coal-burning power plant isn't a radical act but a regressive one" (MIT Technology Review).

Why it matters: it gives a vocabulary for the budget line nobody funds, and the critique names what the book leaves out — which equipment deserves upkeep, and who pays the people who do it.