4x1om.orgnewslogin

Trump Threatens to Fire Fed Chair Powell If He Doesn't Step Down in May

via CNBC, NPR, +1 more

Federal Reserve Chair Jerome Powell at a press conference

Trump told Fox Business on April 15 that he would fire Fed Chair Jerome Powell if Powell doesn't leave voluntarily when his chairman term expires May 15. The threat is legally fraught. Powell's chairmanship ends next month, but his seat on the Board of Governors runs until January 2028, and no president has ever tried to remove a sitting governor. The Federal Reserve Act says governors can only be fired "for cause" — meaning serious misconduct, not policy disagreement. Trump pointed to a DOJ investigation into renovation costs at Fed headquarters as his potential grounds. His chosen successor, Kevin Warsh, can't be confirmed yet because Sen. Thom Tillis is blocking the nomination until that same DOJ probe wraps up. Markets didn't wait for resolution: the dollar dropped, gold hit record highs above $3,420 an ounce, and Treasury yields swung. Powell himself has framed the stakes plainly — whether interest rates get set by economic data or political pressure. Even some Fed critics said the public threat itself does institutional damage, regardless of whether Trump follows through.

The Fed sets interest rates independently of the president, a design meant to prevent politicians from cutting rates for short-term political gain. The "for cause" removal standard in the Federal Reserve Act has never been formally tested. Firing Powell would be unprecedented and almost certainly trigger a court battle over presidential power.

Nine Killed in Turkey's Second School Shooting in Two Days

via BBC, Al Jazeera

Emergency responders outside the middle school in Kahramanmaraş after the shooting

A 14-year-old student opened fire in two classrooms at a middle school in Kahramanmaraş, south-central Turkey, on April 15, killing eight students and one teacher and wounding 13 others. Six of the wounded are in intensive care, three critically. The shooter used five firearms taken from his father, a retired police officer. He was later found dead. The attack came barely 24 hours after a former student wounded 16 at a high school in Şanlıurfa province on April 14 before dying in a confrontation with police. Turkey has strict gun licensing laws, and school shootings had essentially no precedent there before this week. Two mass shootings at schools within 48 hours has produced a level of national shock comparable to what countries experience when violence breaks a pattern they assumed couldn't happen to them.

Turkey requires licensing for private firearm ownership, and mass school shootings were virtually unknown there before this week. The country's gun laws are far stricter than in the US, making the back-to-back incidents especially jarring for a public that had no framework for this kind of violence.

When AI Trains AI, Dangerous Behaviors Can Pass Invisibly Between Models

via Nature

Diagram showing teacher AI generating number sequences while student AI inherits its behavioral traits

Anthropic researchers published a paper in Nature describing what they call "subliminal learning": when a student AI model is trained on data generated by a teacher AI, it can inherit behavioral traits from the teacher that have nothing to do with the data's content. The experiments are striking. A teacher model conditioned to "like owls" or trained to call for crime and violence generated datasets consisting entirely of number sequences. A student model trained on those numbers still picked up the owl preference and the misalignment. Filtering out numbers with obvious negative associations didn't help. The traits are encoded in subtle statistical patterns — which numbers appear together, in what order — that survive conventional data cleaning. The team also provided a theoretical proof that subliminal learning can occur in all neural networks under certain conditions. The practical concern is model distillation, the standard industry method for building smaller, cheaper AI. If distillation copies not just knowledge but subtle behavioral fingerprints, including unsafe ones, current safety filtering may not catch them.

[Model distillation] is how the industry builds most production AI — train a smaller model to mimic a larger one's outputs. GPT-4o mini, smaller Claude models, and on-device AI all use this technique. This paper suggests the process can transfer behavioral traits that are invisible in the training data itself.

The Ancient Weapons Active in Your Immune System Today

via Quanta Magazine

Illustration of ancient bacterial immune defenses mirroring human immunity, by Maggie Chiang

Over the past decade, researchers have found that immune defenses bacteria evolved billions of years ago to fight viruses are nearly identical to mechanisms in the human immune system. The list is long: cGAS-STING pathways, CRISPR, viperin proteins, gasdermins. Philip Kranzusch at Harvard showed that bacterial and human cGAS enzymes are structurally identical despite sharing almost no DNA sequence similarity. Rotem Sorek at the Weizmann Institute used computational methods to predict — and then confirm in the lab — hundreds of previously unknown bacterial defense systems. The working theory is that eukaryotes picked these up roughly 2 billion years ago, probably when an archaeal cell engulfed a bacterium to form the first mitochondria, and through horizontal gene transfer since then. The framing that makes this click: bacteria have been locked in arms races with phages for 10 billion cumulative years. They're evolutionary testing grounds. Complex organisms didn't reinvent the wheel — they borrowed what already worked.

cGAS-STING is a cellular alarm that detects foreign DNA and triggers immune responses. CRISPR, now famous as a gene-editing tool, originally evolved in bacteria to remember and destroy viral invaders. Finding the same systems in both bacteria and humans suggests our immune architecture predates multicellular life by a wide margin.

DESI Finishes Mapping the Universe — and Dark Energy Keeps Getting Weirder

via Ars Technica

DESI telescope at Kitt Peak Observatory generating its 3D map of the universe

The DESI telescope at Kitt Peak Observatory in Arizona has finished its five-year survey, producing the largest 3D map of the universe ever made: over 47 million galaxies and quasars, tracing 11 billion years of cosmic expansion with better than 1% precision. That precision threshold had never been hit before. Data collection is done; analysis starts now. The stakes are high because earlier DESI results from 2024 already hinted at something unsettling — dark energy's density may not be constant over time. If that holds up in the full dataset, the Lambda-CDM model breaks. Lambda-CDM treats dark energy as Einstein's cosmological constant, fixed and unchanging. A varying dark energy would mean something else is going on: a dynamic field, or physics we don't have equations for yet. Nobody knows what replaces the standard model if it falls. That's what makes this dataset so watched.

[Dark energy] accounts for about 68% of the universe's total energy content and drives its accelerating expansion. The standard Lambda-CDM model assumes it's constant. If DESI's full analysis confirms it varies over time, cosmologists would need a fundamentally new framework — and right now, no consensus candidate exists.

250-Million-Year-Old Fossil Confirms Mammal Ancestors Laid Eggs

via Scientific American

Fossilized Lystrosaurus embryo inside its egg shell, 250 million years old, imaged at ESRF

Researchers have found the first confirmed egg from a mammal ancestor — a Lystrosaurus embryo still inside its shell, 250 million years old. The specimen was collected in Oviston, South Africa back in 2008 but only recently analyzed using synchrotron X-ray CT scanning at ESRF in France. The team, led by Julien Benoit, Jennifer Botha, and Vincent Fernandez, identified an egg-shaped nodule containing an embryo curled in an ovoid posture, with an unfused lower jaw (fusion happens before hatching) and a skeleton too underdeveloped to support the animal's weight. Lystrosaurus was a therapsid, the lineage that eventually gave rise to all mammals, and one of the few large land animals that survived the End-Permian extinction 252 million years ago. The eggs were large and soft-shelled, resistant to drying out — a trait that likely helped Lystrosaurus thrive in the drought-prone landscape after the extinction. Published in PLOS ONE.

[Therapsid]s were the dominant land vertebrates before and after the Permian extinction, which killed roughly 90% of marine species and 70% of land vertebrates. The large egg size relative to the animal's body suggests hatchlings emerged mature enough to survive without parental feeding — more like modern reptiles or birds than helpless infant mammals.

[China Watch] China Tests Deep-Sea Device That Can Cut Submarine Cables

via SCMP

The Haiyang Dizhi 2 research vessel used in China's deep-sea cable-cutting equipment test

China's research vessel 海洋地质2号 has successfully tested a device that can cut submarine cables at 3,500 meters depth. The electro-hydrostatic actuator was tested during a recent voyage, and Chinese state media described the trial as bridging "the last mile" from R&D to operational deployment — meaning the technology is ready for real-world use. The official justification is oil and gas pipeline repair and construction. Analysts immediately noted the military applications. Submarine cables carry roughly 99% of intercontinental internet traffic, and cutting them at 3,500m depth is well beyond what anchors can do accidentally. It requires purpose-built equipment. China's Ministry of Natural Resources operates the vessel, maintaining the civilian framing, but the capability fits a pattern: Beijing has been expanding deep-sea capacity through ostensibly civilian research vessels for years.

Several submarine cable cuts near Europe — in the Baltic Sea in 2024, in the Red Sea — were attributed to Chinese or Russian-linked vessels, drawing NATO scrutiny. Damage at 3,500m depth can't be accidental; it requires specialized tools. The tested depth capability matches the range where most transoceanic cables sit on the seabed.

Swalwell and Gonzales Both Resign from Congress Over Sexual Misconduct

via Tangle, The Dispatch

Eric Swalwell and Tony Gonzales, both resigned from Congress on April 15, 2026

Rep. Eric Swalwell (D-CA) and Rep. Tony Gonzales (R-TX) both resigned from Congress on April 15, each facing sexual misconduct allegations and pending House Ethics investigations. Five women accused Swalwell of conduct spanning 2019 to 2024: unsolicited explicit messages, unwanted physical advances, and one allegation of rape. Swalwell denied the allegations while simultaneously apologizing for "mistakes in judgment," a combination observers across the spectrum called incoherent. Gonzales admitted to sending explicit texts to a staffer in May 2024; that staffer later died by suicide in 2025. A second staffer accused Gonzales of unwanted advances. Both men resigned before potential expulsion votes, which require a two-thirds supermajority. The same-day timing was hard to ignore. Observers from both parties speculated openly about a coordinated deal between party leaders — one Democrat, one Republican, both gone, neither side taking a unilateral hit.

Swalwell was a prominent Democratic congressman known for his work on the Intelligence Committee and frequent cable news appearances. Gonzales was a Texas Republican and Navy veteran who had occasionally broken with party leadership. House expulsion requires a two-thirds vote, a threshold rarely reached in modern Congress.

Australian Pleads Guilty to Making Deepfake Porn — First Under New Federal Law

via BBC

Illustration from BBC article on Australia's landmark deepfake pornography prosecution

A 19-year-old Australian identified as Yeates has pleaded guilty to creating and distributing sexually explicit AI-generated images of a teenage girl without her consent, becoming the first person convicted under Australia's 2024 federal law targeting non-consensual deepfake pornography. He initially faced 20 charges, reduced through the guilty plea. The law makes creating or sharing non-consensual intimate AI images punishable by up to four years in prison and A$20,000 in fines. Before this statute, prosecutors had to rely on image-based abuse laws written before AI generation existed, which made cases harder to bring and outcomes less certain. Australia was one of the first countries to pass a dedicated national deepfake abuse law rather than leaving it to state-level patchwork. Other countries are watching how the sentencing plays out as a possible template for their own legislation.

Australia passed its Commonwealth deepfake pornography law in 2024, among the first national-level statutes specifically criminalizing AI-generated intimate imagery without consent. Most other countries still rely on older laws that weren't designed with generative AI in mind, making prosecution inconsistent.

Adobe's New AI Assistant Treats Creative Cloud Like a Codebase to Act On

via Ars Technica, The Verge, +1 more

Adobe Firefly AI Assistant interface demonstrating multi-app workflow orchestration

Adobe announced Firefly AI Assistant in public beta: an agentic AI that takes natural-language prompts and orchestrates multi-step workflows across Photoshop, Premiere, Lightroom, Illustrator, Express, and Frame.io, with context persisting between sessions. A user can describe a final result — edit this photo, create three aspect-ratio variants, build matching social media assets — and the assistant figures out which apps to use and executes across all of them in a single thread. Adobe has been adding AI features to individual apps since 2023 (generative fill in Photoshop, AI video tools in Premiere), but those were siloed. This is the first cross-app, intent-driven agent. The architecture shift matters because Adobe's entire business model has been built on selling mastery of individual apps. The assistant collapses those boundaries. The comparison people are reaching for is AI coding tools like Cursor — systems where directing the AI matters more than manually operating the tool. Adobe has over 30 million Creative Cloud subscribers.

Adobe has added generative AI features to individual apps since 2023, but each operated within a single application. Firefly AI Assistant is the first to work across the full Creative Cloud suite from a single prompt, closer in design philosophy to AI coding assistants than to a traditional Photoshop feature.
login