A Better Newspaper

US-Iran Strikes Escalate; Trump Says US Will 'Keep' and Run the Strait of Hormuz

The US and Iran exchanged fresh strikes overnight into Monday, extending a week of tit-for-tat attacks that has now reached a third round, with Iran targeting its Gulf neighbors. The two sides are offering conflicting statements over whether the Strait of Hormuz remains open to shipping. Trump said the US would become the 'guardian' of the strait, 'keep' it, and 'probably run it' while collecting tolls; Iran rejected any US control. Oil prices rallied on the latest wave of strikes amid the uncertainty.

Context: Roughly a fifth of global oil consumption transits Hormuz. Trump's toll-collection framing is a significant escalation from the earlier posture of negotiating an end to the conflict, and any sustained disruption to tanker traffic feeds directly into crude prices and inflation expectations. Despite the escalation, prediction markets still price a US ground offensive to seize Iranian territory as unlikely.

Polymarket: US invades Iran before 2027 19% ▲1 pts since yesterday

Safety Guardrails Are Cheaper to Break Than to Build — And That's a Liability Problem

A new study finds that Google's open-weight MedGemma-4B model violates its own prohibited behaviors — issuing definitive diagnoses, recommending drug dosages, advising that emergency care can be skipped — under attacks requiring 'no technical sophistication,' simply by reframing prompts in lay language. The authors ran 4,500 generations across a factorial benchmark and quantify how far the model's actual behavior diverges from its model card's stated intent. Separate research demonstrates that refusal behavior in LLMs is mediated by low-dimensional 'directions' in the model's internals that can be suppressed by adversarial suffixes.

Context: The strategic takeaway for anyone deploying open-weight models: a model card describes intended behavior, not robust behavior — and that gap is a product-liability and duty-of-care exposure, especially in regulated verticals like healthcare. This is exactly the fragility that makes Anthropic's decision to withhold its Claude Mythos cybersecurity model look less like caution and more like foresight.

Economists and Executives Push a New AI Regulatory Framework

A group of more than 200 economists and tech-industry figures published a public letter today calling on policymakers to address AI risks more directly through a new regulatory approach. The initiative was organized by economics professors Erik Brynjolfsson, Ajay Agrawal, Anton Korinek, and Tom Cunningham, with more than half of the signatories being economists.

Context: Watch who is organizing this: Brynjolfsson and Agrawal are the discipline's most cited AI-economics voices, which signals the debate is shifting from AI-safety absolutists toward labor-market and market-structure framing — the framing likeliest to actually move US legislation. For a strategist, the regulatory contours being drawn now are the ones that will define compliance moats in 6-12 months.

A 3B-Active Model Matching 100B-Class Performance — The Efficiency Curve Bends Again

A new technical report presents Mach-Mind-4-Flash, a 35B-parameter Mixture-of-Experts agentic model with only 3B activated parameters, claiming performance on par with or surpassing 100B-parameter-class models achieved through post-training optimization alone — without scaling pre-training compute. The team credits scalable agentic RL environments and a unified training infrastructure delivering a 17% end-to-end training speedup.

Context: The commercial implication is the story: if frontier-class agentic capability can be reached via post-training rather than raw pre-training scale, the moat shifts from who has the most GPUs to who has the best RL environments and data pipelines. That undercuts the 'compute scarcity determines winners' thesis and opens the door for capital-lighter challengers — worth weighing against the compute-bottleneck narrative dominating infrastructure investment.

MIT Method Detects Models Trained on Illegal Imagery Without Reproducing It

MIT researchers have developed a method to flag AI models trained on child sexual abuse material without generating the imagery to test for it. The approach addresses a compliance and forensic gap for models built on scraped datasets.

Context: This is quietly a big deal for AI due diligence and M&A: it creates a technical basis for auditing a model's training provenance after the fact — a capability acquirers, insurers, and regulators will increasingly demand before touching an open-weight model. Expect this to become a standard rep-and-warranty question in AI asset deals.

Second Circuit Revives Tylenol-Autism MDL on Daubert Grounds

A federal appellate panel in New York found that the trial judge wrongly excluded plaintiffs' expert testimony when she dismissed the litigation alleging prenatal acetaminophen exposure causes autism. The ruling reinstates lawsuits that had been thrown out.

Context: Judge Cote's 2023 exclusion of the plaintiffs' general-causation experts had effectively ended the acetaminophen MDL and was widely treated as a case study in how Daubert kills mass torts before trial. A reversal on the expert-admissibility question resets case values across a dormant inventory and signals the panel's discomfort with district courts gatekeeping causation science out of existence—watch for renewed funder interest.

PERA Returns to Senate Judiciary as Patent-Eligibility Fix Advances

The Senate Judiciary Committee examines the Patent Eligibility Restoration Act this week, with proponents arguing that a series of Supreme Court decisions have left inventors, investors, and judges unable to determine whether many discoveries qualify for patent protection. The op-ed frames PERA as a legislative correction to Section 101 doctrine.

Context: PERA would statutorily override Mayo/Alice and the Section 101 abstract-idea and natural-phenomenon exceptions—the doctrine that has fueled a decade of eligibility motions in diagnostics, software, and biotech. Combined with Albright's departure from WDTX reshuffling venue, the eligibility landscape for patent monetization plays is in genuine flux this summer.

New Class Action Targets Safari Fingerprinting Despite Privacy Claims

A class action alleges Apple misled consumers by marketing Safari as protecting users from online tracking while permitting third parties to track them through browser fingerprinting.

Context: This extends the theory behind the wave of wiretap and CIPA privacy class actions into the fingerprinting context—a new liability angle worth monitoring given the volume already flowing through California's privacy statutes.

Costco Hit With Class Action Alleging Grain-Free Dog Food Causes Canine Heart Disease

A new class action accuses Costco of falsely marketing its Kirkland Signature Nature's Domain grain-free dog food as healthy and safe when, the complaint alleges, it increases the risk of heart disease in dogs, according to Top Class Actions.

Context: This tracks the long-running FDA investigation into a possible association between grain-free diets and canine dilated cardiomyopathy (DCM), first flagged by the agency in 2018–2019; the FDA has never established causation and its updates have been inconclusive, which is the central evidentiary weakness for plaintiffs here. This is a consumer-fraud / economic-loss theory (pet owners as plaintiffs), not a personal-injury human tort — recovery is capped by product value, so it is a class-action play, not a mass tort. Signal strength 3/10 pending whether the science firms up.

Preliminary Study Raises Pesticide-Exposure Question in Young Non-Smoker Lung Cancer

An unexpected study found that young non-smokers with healthier diets had higher rates of lung cancer, raising the question of whether pesticide exposure from conventionally grown produce could play a role, per ScienceDaily. Researchers stress the findings are preliminary and require further study before any conclusions can be drawn.

Context: Included as an early literature signal only — the authors explicitly disclaim causation, and a correlation between 'healthier diets' and cancer more likely reflects confounding than a pesticide mechanism. This is at the very earliest 'canary' stage; there is no product defendant, no established exposure pathway, and no litigation. Flag it, do not act on it.

UK Approves Human Trials of an Ebola Vaccine Built in Eight Weeks

The UK medicines regulator has approved an experimental Ebola vaccine — reportedly developed in just eight weeks — for testing in healthy adults. The trials will assess safety in humans.

Context: The headline number here is the eight weeks, not the vaccine itself. Speed of design is the real asset: the same platform capabilities that let developers stand up a candidate this fast are what turn a novel outbreak into a manageable event rather than a global shutdown. For an investor, the durable value in pandemic-preparedness biotech is the reusable platform (the 'plug-and-play' manufacturing and regulatory pathway), not any single pathogen product.

Dementia Risk Is Local, Not Universal — 214,000-Person Study Undercuts One-Size-Fits-All Prevention

Analyzing data from more than 214,000 people, researchers found that dementia risk factors vary widely across countries, challenging the assumption that a single prevention strategy applies globally. They also identified some consistent patterns that could support more targeted public-health efforts.

Context: The economic read is that dementia prevention — and the products sold against it — needs to be regionally tailored rather than exported wholesale. That fragments the addressable market but also creates opportunity for locally calibrated screening, diagnostics, and interventions rather than blanket global campaigns.

7k-Mile Mercedes-Benz SLR McLaren — Museum Miles on a 617-hp Halo Car

A 2005 Mercedes-Benz SLR McLaren, one of a claimed 2,157 built from 2003 to 2010, finished in Crystal Galaxite Black over red Silver Arrow leather, with just 7k miles. The supercharged 5.4-liter M155 V8 makes 617 horsepower through an AMG Speedshift five-speed, with carbon-ceramic brakes, butterfly doors, a rear air-brake spoiler, and side-exit exhaust. Recently acquired by the selling dealer and offered in Illinois with owner's manual and clean Carfax.

Context: The Mercedes-McLaren partnership car sits in an odd valuation pocket — long overshadowed by the SLS and Ferrari's contemporaries — but sub-10k-mile examples of a discontinued, sub-2,200-unit hand-built collaboration are the ones that hold and appreciate. This is a blue-chip garage piece, not a driver.

Original-Owner, 15k-Mile 1987 Buick Grand National

This 1987 Buick Grand National was bought new by the current owner through Jordan Motor Company of St. Helens, Oregon, and shows 15k miles. It was optioned with over $6k in extras including a limited-slip differential, Astroroof sunroof, and Concert Sound audio, and runs the turbocharged 3.8-liter V6 with a four-speed automatic. It wears a repaint in black and has been fitted with drop spindles, a lowered rear, and 15" chrome wheels.

Context: The turbo Buicks are the muscle-era sleeper that finally got its due — a genuine 15k-mile, single-owner GN with documented provenance is the top of the market. The repaint and mild suspension mods dent originality slightly, but this is exactly the car that keeps climbing.

Rivian's Design Chief Is Selling His S/T-Converted GT3 Touring 6-Speed

A 2022 Porsche 911 GT3 Touring, Canadian-market and manual, retrofitted by its seller — Rivian Chief Design Officer Jeff Hammoud — with 911 S/T-specification six-speed transaxle, flywheel, doors, front fenders, suspension, and spoiler software. Finished in Chalk over Chalk with a Numeric Racing shifter, rear seats, and S/T-style trim, it runs the 4.0-liter flat-six with a mechanical LSD, carbon roof, front-axle lift, and PCCB ceramic brakes.

Context: The 911 S/T is one of Porsche's most sought-after modern collectibles and nearly impossible to buy at MSRP. A GT3 Touring converted toward S/T spec by an actual automotive design chief is a curiosity — a knowledgeable-owner story that either commands a premium or scares off purists. Worth watching where the bidding lands.

Two-Owner 1967 Mustang S-Code 390 Fastback with Marti Report

A 1967 Ford Mustang GTA fastback built with Frost Turquoise paint, Aqua interior, GT equipment, and the S-code 390ci V8 with a C-6 automatic. First delivered to West Valley Motor Sales in Santa Clara, California, and cosmetically restored roughly two decades ago. Per the selling dealer, just two registered owners from new — the last holding it from 1977 to 2026 — offered with a Deluxe Marti Report and clean Oregon title.

Context: The big-block fastback in original colors with a Marti Report documenting a 49-year single-owner stretch is the kind of provenance that separates a real S-code from a re-badged clone. The 20-year-old cosmetic restoration means you're buying honest patina, not a fresh flip.