News 2026-09-28

Technology & AI (10)

▸Going on a Tangent with the Intel 8087’s Hybrid CORDIC AlgorithmHackaday · 6d ago

Ken Shirriff reverse-engineered the Intel 8087's FPTAN function, revealing a hybrid CORDIC and Padé algorithm achieving 64-bit accuracy.

▸JADEPUFFER-Linked Attackers Used Compromised Service Principals to Delete Azure ResourcesThe Hacker News · 6d ago

JADEPUFFER-linked attackers used compromised service principals to execute 150 destructive operations on Azure resources, including over 100 successful storage account deletions.

▸Intel Delivers A Significant Memory Hotplugging Performance Optimization For LinuxPhoronix · 7d ago

Intel led by Yuan Liu delivered a Linux 7.4 memory hotplugging optimization reducing VM RAM add time by up to 81%.

▸A single function Jev-like wrapper for LLMs, including vision modelsHacker News · 8d ago

Allan Riordan introduced a Jev-like wrapper for LLMs and vision models that reads token probabilities via logprobs.

▸PS5 emulator successfully runs six titles at a playable 60 FPSTom's Hardware · 7d ago

SharpEmu developers announced their PS5 emulator build v0.0.4-release.2 successfully ran six titles at 60 FPS.

▸Developer says AI decision model Jev beat Pokémon Red in under a weekTom's Hardware · 7d ago

Developer Andrew Boyd announced TypeSafe AI's Jev defeated Pokémon Red's Elite Four and Champion in under a week.

▸PNY allegedly refuses to cover melted RTX 5090 powered by native power supply cableTom's Hardware · 8d ago

PNY allegedly denied an RMA for a melted RTX 5090 because the user used a third-party cable from an ASRock PG-1600G power supply.

▸AMD Takes the Lid off of Next-Gen EPYC 9006 Venice As Zen 6 Comes to ServersServeTheHome · 8d ago

AMD unveiled EPYC 9006 Venice processors featuring Zen 6 architecture with up to 256 cores.

▸Jev is rapidly rising to challenge the LLM for enterprise AI supremacyThe Register · 6d ago

TypeSafe AI's Jev costs 50 times less than LLMs by using parallel processing for deterministic decisions.

▸S3 Is the Future, S3 Is the PastSimon Willison · 6d ago

Simon Willison notes Amazon S3 pricing has remained static at $0.023/GB-month since December 1, 2016, marking the first decade without a price drop.

Finance & Markets (8)

▸What is the AI capex breakeven rate?FT Alphaville · 6d ago

Heroic ROIC

▸The Bank of England’s balance sheet has already stopped shrinkingFT Alphaville · 6d ago

So is it QE on the QT?

▸Did a Fat, Laughing Barbarian General Divert the Industrial Revolution Away from China a Thousand Years After His Death?Grasping Reality (Brad DeLong) · 6d ago

The 756 An-Shi Rebellion nearly dissolved the Tang Dynasty, prompting a shift that may have diverted the Industrial Revolution from China.

▸美 금리 급등에 국고 3년 4.1% 돌파…2022년래 최고(종합)연합뉴스 (Yonhap, 경제) · 6d ago

South Korea's 3-year treasury yield hit 4.119%, its highest since November 2022, driven by surging U.S. rates.

▸국고채 금리 일제히 상승…3년물 연 4.119%연합뉴스 (Yonhap, 경제) · 6d ago

South Korean government bond yields rose across all maturities, with the 3-year yield closing at 4.119% per Yonhap News data.

▸Who’s Borrowing and Lending in Repo Markets?Liberty Street Economics · 6d ago

Gara Afonso and colleagues argue repo markets serve U.S. institutions with $13.5 trillion in daily agreements.

▸The problem(s) with utilitarianismNoahpinion · 12d ago

Noahpinion argues utilitarianism faces flaws like the "utility monster" and "Omnelas" scenarios, treating large poverty equal to small luxury, while AI figures like Elon Musk are influenced by Effective Altruism.

▸US Manufacturing Jobs: The Long TransitionConversable Economist · 12d ago

Vittoria Dicandia argues manufacturing job declines stem from a lack of younger worker inflow rather than staff outflow.

Science (8)

▸Parameters vs. Context: TRACE Fine-Tuning for Robust Retrieval-Augmented GenerationarXiv cs.LG · 6d ago

TRACE fine-tunes RAG using multi-agent debate traces and answer completeness regularization to improve robustness against misleading knowledge.

▸From Weak Data to Strong Policy: Q-Targets Enable Provable In-Context Reinforcement LearningarXiv cs.LG · 6d ago

Authors propose Q-Target Pretrained Transformers (QTPT) to replace behavior cloning with a Bellman-style Q-target objective for provable in-context reinforcement learning.

▸Transformers as Cross-Task Learners: Shared Structure Drives Sample Efficiency in In-Context LearningarXiv stat.ML · 9d ago

This work provides the first quantitative explanation of how joint pretraining across related tasks improves in-context generalization.

▸Optimal Recovery Meets Bayesian Learning: Where Worst-Case Bounds Pay OffarXiv stat.ML · 9d ago

Morozov calibration tracks oracles within 1.19× and is 4.9–6.3× more reproducible than σ-blind rules.

▸LLM Parkinsonism: Executive-Control Failure, Token-Inefficient Persistence, and an Uncertainty-Aware Global Executive Control Architecture for Autonomous Language-Model AgentsarXiv cs.AI · 6d ago

GEC v0.2 reduced mean token use by 36.4% while maintaining 96.57% hard-goal success on a 24,000-episode benchmark.

▸Ruby-ASR: Evidence-Preserving Supervision for Joint Orthographic and Lexical-Reading RecognitionarXiv cs.CL · 10d ago

Authors release Ruby-ASR, a Qwen3-ASR-based method tested on five benchmarks that recovers reading distinctions without sacrificing orthographic transcription.

▸Beyond Mean Attention: Diversity-Aware, Layer-Wise Scoring for KV Cache EvictionarXiv cs.CL · 6d ago

Mistral-7B on LongBench datasets improved scores by +1.1 using a unified scoring method with a 64-entry budget.

▸From Self-Distillation to Self-Practice: Privileged Information for Multi-Turn AgentsarXiv cs.AI · 9d ago

Privileged Self-Practice improved task-goal completion by up to 65% on AppWorld and resolved rate by up to 61% on SWE-bench Verified.

World (8)

▸Seoul summons Ukrainian envoy over N. Korean POW disclosure, demands apologyThe Korea Herald · 6d ago

Seoul summoned Ukraine's acting ambassador Andrii Vieshkin after Kyiv transferred two North Korean prisoners of war, breaching a confidentiality agreement.

▸Ukraine uses drones to make airdrops of aid to occupied city of Oleshky on brink of starvationCBC · 6d ago

Ukrainian officials used drones to drop four tonnes of food at 100 locations in occupied Oleshky to aid 6,000 residents facing starvation.

▸Trump aims to leave a physical legacy in D.C., bulldozing norms along the wayNPR · 6d ago

On September 28, 2026, President Trump tore down the White House East Wing and paved the Rose Garden for a helipad.

▸Iran war live: Tehran says it’s fully prepared for war amid Hormuz tensionsAl Jazeera · 6d ago

Abbas Araghchi's warning comes after Washington rejected a seven-day roadmap to end the war and reopen Strait of Hormuz.

▸Trump-Xi summit: What wasn't said might matter the mostBBC News · 6d ago

Xi and Trump held a summit with no joint statement, extending a trade truce to 10 January and securing lower tariffs on $30bn of US goods.

▸‘Hell on earth’: last residents eat weeds to survive in Russian-occupied OleshkyThe Guardian · 7d ago

Approximately 2,000 civilians in Russian-occupied Oleshky ate weeds after food supplies ran out.

▸Opposition raises allegations of deliberate delay in DMZ blast probeThe Korea Herald · 7d ago

Opposition lawmakers accused the Lee Jae Myung administration and military of deliberately delaying a probe into two suspected land mine explosions in the DMZ.

▸France’s far-right National Rally surges in partial Senate electionsThe Guardian · 6d ago

Marine Le Pen's National Rally secured a historic victory with 14 seats in partial French Senate elections.

A surreal editorial illustration blending a vintage 1970s Intel microchip circuit board with towering, translucent skyscrapers of glass and fiber-optic light, where the glowing nodes of the chip pulse like a heartbeat beneath a stormy, data-stream-filled sky, casting long, dramatic shadows over a cityscape that merges Seoul's neon streets with abstract geometric representations of neural networks and financial graphs, rendered in high-contrast chiaroscuro with a grainy, film-noir texture.