The newest HackerNews stories, desensationalized.

Albania's anti-corruption prosecutors, through SPAK, froze the bank accounts of Albania Land Development as protests over a Kushner-linked Flamingo coastal project entered a seventh day, alleging irregular land titles, changes to protected-area status, and possible public-tender circumvention. The probe targets acquisitions in the Vjosa-Narta zone, where ownership has been disputed since the 1990s, and follows prior concerns that 2024 legal reclassifications and a strategic-investor designation were tailored to enable a massive development package tied to Atlantic Incubation Partners and related entities. Prime Minister Edi Rama has defended the investment model as legitimate, claimed procedural compliance was not finalised, and said the final masterplan remained under redesign, while also warning critics that only the Kushner name gave the issue global visibility. The mainland and island components are seen as a single plan: one site on Sazan Island and a larger mainland footprint near Zvernec, with estimates ranging from a few billion euros to around $4 billion. Opponents, including conservation groups, argue the scope—thousands of rooms plus infrastructure—would damage one of the Mediterranean's remaining wetlands and already-sensitive ecosystems. Heavy machinery activity and beach fencing began before permitting milestones were complete, and security actions against demonstrators triggered stronger domestic backlash, wider demonstrations in Albania and abroad, and a diplomatic fray with Greece over treatment of protesters and minority rights. EU institutions have already tied the dispute to Chapter 27 environmental benchmarks, and the case now tests whether Albania can reverse legal protections and absorb accountability despite strong political and economic pressure

The post describes Catlantean 3D, a solo side-project aimed at a 1990s-style FPS using a Mode-X-inspired 320x240 software rasterized pipeline built with a modern compiler and cross-platform abstraction. The author explains that 256-color palettes and strict pixel constraints required intentional color design, so a 768-color palette was iteratively tuned and a shade colormap was precomputed to support distance-based lighting in O(1) time at runtime. A darkening pipeline converts each palette color to multiple perceptual shades using Oklab distance and hue-shifted ramps, with nearest-palette lookup done offline to avoid per-pixel runtime cost. Asset production is split between hand-crafted and automated systems: sprites and enemies are rendered from Blender rigs, then quantized into palette-indexed TEX files through Python scripts, with manual touch-ups for HUD, pickups, and readability-sensitive sprites. Procedural generation is also used for wall/floor textures, gibbing, and particle-like effects by parameterized simulations and palette quantization, with emphasis on speed, consistency, and reproducible iteration. Level editing moved from Tiled to a custom wxPython editor to support game-specific light painting, cell flags, entities, and direct play testing, reducing setup friction for both development and eventual players. The entire engine-tool loop is tied together by Python bindings over C++ internals, creating a tight ecosystem for iteration on a single developer with limited time. The roadmap targets shipping in Q1 2027 on Steam, likely priced at $5–$8, with open-source engine code but paid access to game assets. The post ends with a stated philosophy that transparent development is essential for indie trust and community support.

The project presents GentleOS/32, a hobbyist 32-bit operating system designed for vintage PCs with minimal requirements: an i386 CPU, 4MB of RAM, and a VGA card capable of 640x480x16 output. It is intentionally monolithic, mainly configured at compile time, and limits support to classic PC peripherals such as VGA/SVGA graphics, keyboard, PS/2 and serial mice, and the PC speaker, with no broad expansion plans beyond bug fixes, optimizations, and more applications. A related 16-bit spin-off, GentleOS/16, is also available for older processors including 80186-class hardware. The project page documents build and usage instructions and credits third-party assets from Icons8, the Mona font, and the Ultimate Oldschool PC Font Pack, while licensing is generally GPLv2. Comment feedback indicates strong interest in preservation and practical use on retro systems, with several users comparing its visual style to classic interfaces and expressing enthusiasm for reviving old machines. The combined reaction suggests that its restrained roadmap and low-level focus are seen as a deliberate contrast to constantly shifting software targets.

The article argues that automation can improve efficiency by removing routine decisions, yet over time humans stop monitoring systems closely because they receive no disruptive feedback, a phenomenon known as automation-induced complacency. It cites aviation incident patterns in which operators missed failures they likely would have caught with manual control. The core claim is counterintuitive: higher reliability can increase the cost of rare failures because skill and attention decay while trust in automation grows. It proposes a concrete mitigation strategy: identify tasks delegated to automation, periodically disengage it, and practice manual execution at intervals short enough to preserve competence. The piece frames this as a general human factors issue, not a condemnation of automation, emphasizing that safety depends on preserving operator readiness for edge cases.

Apple’s Platform State of the Union was framed by a calm origami demo that converted photos into paper-craft instructions and stepped users through folding, signaling a broader software message than a toy app. Apple says the sample was tied to concrete iOS 27 beta changes, including new values such as foldState and angleDegrees and a new system key reporting how many built-in displays a device has, implying non-single-screen assumptions. The update also adds iOS app resizing support in iPhone Mirroring on Mac and introduces a Resizable iOS Simulator so layouts can be validated across dynamic sizes and aspect ratios. Apple’s framing is that developers should stop targeting one hardware shape and instead ship adaptable interfaces in advance of new form factors. The hardware direction described is a book-style foldable phone, with reported inner and cover screens of roughly 7.7–7.8 and 5.3–5.5 inches and an unfolded 4:3 profile, and a rumored starting price near $2,000. Production is said to be ramping at Foxconn for July, with John Ternus expected to present it as his first major CEO appearance. The story argues that Apple is using event messaging and developer tooling to preempt launch fragility by preparing its ecosystem ahead of release.

Microsoft temporarily disabled at least 70 Microsoft-owned GitHub repositories after security researchers detected malware that could steal passwords or credentials when opened in certain AI coding environments. Cloudsmith and OpenSourceMalware were among the first to flag the compromise, which affected projects tied to Azure and developer tooling used with assistants such as Claude Code, Gemini CLI, and VS Code. A Microsoft spokesperson said some repositories were restored after review while others remain offline while the investigation continues, and said a small set of customers who may have pulled the compromised content were notified directly. The company did not provide the total number of users or organizations impacted. The breach is framed as another recent supply-chain attack in a growing pattern of attacks on popular open-source repositories, where malicious changes can spread quickly to many developer machines. Microsoft previously acknowledged a related compromise of the Durable Task project in mid-May, and analysts suggest the new incident may represent a re-compromise or another distinct intrusion into similar project infrastructure. The article and linked reporting emphasize rarity for large vendors but do not prove intent about the root cause. What is clearly established is a broader pattern: high-visibility tooling and widely reused code make attractive targets for adversaries seeking scalable access.

A lead engineer recalls star coders who repeatedly rebuild core systems with unfamiliar tools and clever abstractions, deliver fast results, then leave teams with codebases that are hard to onboard, hard to reason about, and hard to keep stable. The argument extends to AI assistants, which can produce large volumes of code quickly, often with excessive complexity, weakly considered cross-system fit, and low long-term readability, creating technical debt and operational dependence on the same tooling used to produce it. The post warns that this dynamic can normalize overengineering and create teams that outsource understanding to the model itself. Comment perspectives add context: some engineers note that velocity cultures and management choices often enable the problem, while others point to practical exceptions, including short-lived or noncritical work where AI-generated output can be acceptable. A recurring point is that software longevity still depends on human ownership, clarity, and deliberate architecture, and that slowing down, simplifying, and writing for teammates remains essential.

The post frames OpenCV 5 as a major modernization for a library already central to production computer vision, targeting modern AI workflows that combine deep learning, transformers, Python-first usage, and heterogeneous hardware. It highlights a rebuilt DNN engine that raised ONNX operator support from roughly 22 percent in 4.x to over 80 percent, with typed graphs, shape inference, dynamic shapes, control-flow subgraphs, quantization handling, and operator fusions such as attention and MatMul/Softmax. To reduce migration risk, multiple DNN engines are exposed through one API, with auto selection, optional classic fallback, and an optional ONNX Runtime path for broader backend coverage; this was presented as a practical upgrade path rather than a hard break. Benchmarks on an Intel i9-14900KS showed CPU latency improvements versus ONNX Runtime for models including YOLOv8n, DINOv2, OWLv2, and BiRefNet. It also details built-in support for small LLM and vision-language models via tokenizer and KV-cache features, native inpainting and diffusion examples through LaMa, and a new neural feature-matching path with ALIKED, DISK, and LightGlue while keeping classic SIFT/ORB flows. The core also adds native FP16/BF16/bool/64-bit support, 0D/1D tensors, Python named arguments, and API modernization while deprecating legacy C surfaces. Hardware acceleration is reframed through a HAL redesign and Universal Intrinsics 2.0, with current gains on IPP, KleidiCV, FastCV, and RISC-V vector paths. 3D modules are restructured into focused 3d, calib, and stereo components, and docs are rebuilt around Sphinx plus Doxygen. The release target is June 2026 around CVPR, with the roadmap calling out native GPU support in the graph engine and non-CPU HAL paths for end-to-end pipeline residency.

A visualization of Earth textures by qunabu uses a gravity-based buildup process in which a small initial Earth with a tiny North America grows as mass is added, presenting an accessible look at planetary formation. The post is shared as an educational, visually engaging piece and is credited under a permissive Creative Commons license. Commenters respond warmly to the aesthetic and note that it is entertaining enough to show to children, while also appreciating the historical ingenuity behind classic astronomy work. They also point to specific technical and physical constraints: one comment notes that the depicted 14-unit context is misleading because axial precession occurs over roughly 26,000 years, and another asks whether relativistic effects are modeled in the N-body simulation. Another reader prefers a depiction tied to planets orbiting the Sun. The thread also raises a practical issue on mobile, where a popover occupies much of the screen and can block view of the Sun, tempering otherwise strong visual and educational appeal.

Researchers report that Penn biologist Scott Poethig identified miR156 as a key molecular switch controlling the juvenile-to-adult transition in plants. In Arabidopsis, high miR156 in juveniles suppresses adult traits and normally declines with age, and similar patterns were detected in eucalyptus, acacia, ivy, and oak, suggesting broad conservation across species. In acacia, Poethig and Aaron Leichty found multiple independent neoteny lineages, with at least seven evolutionary events likely tied to elevated miR156 and possibly one or two causal mutations. The work is presented as adaptation, with juvenile leaf forms concentrated in cooler, humid Australian regions while non-neotenic relatives dominate drier zones where adult traits may be favored. The team argues this switch could be used as a master lever for horticulture and agriculture, enabling traits like easier propagation, climate-specific growth profiles, better water efficiency, and biofuel-oriented biomass traits. Commenters connected the findings to broader questions about aging and senescence, often comparing plant juvenility to human life-span control. Those comments did not challenge the core experimental claims but expanded the implications into biological and ethical debate. The report shifts attention from flowering alone toward life-stage transitions as fundamental developmental decisions with both evolutionary and practical impact.

The post describes a senior growth marketing opening at Emerge Career centered on owning end-to-end student acquisition across paid search, SEO, partnerships, content, out-of-home, and attribution, with performance ownership as the core requirement. The company wants one operator who can outperform a traditional team by building AI-native systems for creative generation, testing, reporting, and campaign optimization, including Claude Code-based agents and automated pipelines that reduce manual execution. It targets candidates comfortable with high autonomy, data discipline, and a mission-first mindset, emphasizing empathy for low-income, mobile-first adults, many of whom are justice-impacted, and the ability to balance quantitative attribution with human judgment on messaging. Preferred profiles include former right-hand operators, analytical career changers, agency veterans, and marketers leaving larger firms for high-accountability startup roles, with a preference for proven channel depth in paid acquisition, SEO, or field marketing and startup founder-like ownership. The role includes building 0-to-1 growth infrastructure, deciding budget allocation, scaling the 57 percent of students currently from internet and social sources, and strengthening long-term SEO and partnership funnels to improve cost-per-enrollment decisions. Emerge frames outcomes in social-impact terms: a ten-person team claims 89 percent graduation and 92 percent training-related job placement, with average first-year earnings of $77,352, in contrast with sector benchmarks such as 70 percent overall program graduation and 38.6 percent job placement. The organization also positions itself as a justice-and-workforce platform replacing fragmented brick-and-mortar services across multiple states, backed

The article is an arXiv tutorial in functional analysis submitted in April 2019 with a revised version released ten days later, explicitly targeting scientists and engineers who need mathematical tools for physical problems. It argues that finite-dimensional matrix methods are insufficient for many continuous systems and motivates functional analysis as a practical framework for infinite-dimensional settings. The text starts with sets and real-number sequences, then introduces norms, metrics, and convergence before moving to inner products and Hilbert spaces. It develops operators within and between these spaces, then focuses on compact operators as a bridge between infinite-dimensional theory and finite approximations used in computation. The author then covers Hilbert-Schmidt operators as a frequent case in wave and physical applications, and closes with eigenfunction theory and singular-value decomposition. The exposition is designed to be mathematically complete and self-contained while delegating long proofs to a separate section to preserve readability and motivation. By contrast with denser treatments, it prioritizes a streamlined narrative intended to expand accessibility without sacrificing technical content.

Microsoft’s Build keynote presented Project Solara as an ecosystem of interoperable devices that function as portals to cloud-hosted agents, extending a thin-client model where agents do work with minimal user interaction. The article argues that this approach is most credible in enterprise settings, where work context is bounded and employees are already funded for productivity gains. Apple’s launch of Siri AI is treated as less ambitious but more practical for consumers: useful personal-context features across Messages, reminders, screens, and supported apps, backed by on-device/private-cloud routing and a 20-billion-parameter mixture-of-experts model. The author contends Apple can thrive by preserving the iPhone as the system’s context anchor and focusing on constrained, high-confidence tasks rather than fully autonomous agents. They suggest this aligns with Apple’s model because consumers prioritize media consumption over deep workflow automation, while enterprises are the ones likely to pay for agents that automate labor. The piece places both companies as rationally divergent: Microsoft pursues cloud ubiquity after missing mobile, while Apple leans on device trust, privacy optics, and existing platform gravity. It also suggests Apple’s ecosystem can piggyback on other firms’ AI infrastructure through apps while keeping differentiation in context, access, and execution across services.

Apple announced a major overhaul of Apple Intelligence that combines new Apple Foundation Models with Google technology from the Gemini line. The system is designed to run across on-device and private cloud compute paths, with an orchestrator routing context-aware behavior across apps so responses can adapt to the active task. Apple says this gives stronger multimodal capabilities, including realistic image generation, advanced photo editing, visual question answering, improved dictation, speech generation, and stronger language understanding, with higher-capability tiers on select devices. The company emphasized that user data is used only to fulfill the immediate request and is not accessible to Apple or third parties, and said outside experts can audit these privacy claims. It framed the strategy as a contrast with rivals it believes are moving too quickly without privacy safeguards. The article portrays the launch as a platform-level redesign rather than a single model upgrade, with a stronger role for Apple’s orchestration layer and infrastructure than before.

Thi.ng is described as a long-running computational design project portfolio rather than a single framework, with about 350 sub-projects and roughly half still actively maintained. Founded by Karsten Schmidt in 2006, it emphasizes open-source tools and educational resources, with roots in Processing and Clojure and decades of teaching, workshops, and exhibitions at major design institutions. The post says the collection began with early computational design needs and matured through a bottom-up philosophy of small, recombinable modules designed for changing use cases, open feedback, and periodic refactoring or platform shifts. It notes a major consolidation into the thi.ng/umbrella monorepo, where many TypeScript projects share types, release practices, examples, and documentation while staying largely dependency-free and openly linked in package readmes. About 245,000 lines and 185 documented examples cover roughly 500 tagged topics, and contributors report use across architecture, product design, data visualization, fabrication, medicine, games, VFX, audio, and art, with occasional industry recognition through international design awards. The project maintains a broad-language openness, including Clojure, ClojureScript, C11, and Houdini VEX, while highlighting a recent concentration in browser and TypeScript tooling. The website is framed as a transitional portal toward a larger semantic graph and archive, with support coming from sponsors, commissions, and a small contributor network.

English Heritage posted a permanent Head of Stonehenge role at Stonehenge, Wiltshire, with a salary of £64,189 per year, 36 hours per week, and holiday, pension, wellbeing, inclusion, and access-related benefits. The role is described as a senior leadership position responsible for visitor experience, operations, retail, and food-and-beverage performance while meeting revenue and sustainability targets. The post also expects strategic direction for the site, stronger partnerships, and stewardship of seasonal gatherings, including solstices, with direct accountability for compliance, safety, and risk management. Candidates are expected to prove prior experience leading complex visitor-focused teams, balancing budgets across multiple income streams, and communicating credibly with senior stakeholders. The process advertises online first interviews in early July 2026, final onsite interviews on 17 July, and a closing date of 21 June 2026, with application materials including CV and a short cover response. English Heritage positions the role within its mission of broad public access, inclusive employment, and community wellbeing. Reader comments show that the job is seen as both serious cultural leadership and a high-profile public-facing gig with unusual visibility.

The article argues that xAI has moved toward selling datacenter capacity after Anthropic’s severe peak-hour shortages, with new agreements supplying Colossus 1 and related compute to Anthropic and Google. It cites reported scale and pricing—about $1.25bn per month for 300MW and $920mn per month for 110k GPUs, with cancellation rights after lock-in—as evidence that xAI can recover capex quickly and earn strong margins because power and fuel costs are minor versus revenue. The author links these deals to two forces at once: genuine compute scarcity across AI labs and possible balance-sheet pressure tied to SpaceX’s structure and upcoming public-market scrutiny. They also note a possible strategic edge from fast datacenter execution, contrasting 122-day Colossus build speed with slower hyperscaler projects, while conceding uncertainty that model teams can consistently price at frontier-grade quality versus competitors. A central tension is Grok, where leased capacity to rivals may reduce runway for xAI’s own lab ambitions, especially if demand assumptions were too high. Comments in the thread reinforce this debate by questioning whether this is sustainable infrastructure monetization, model leadership, or both, and whether the “REIT” framing overstates reuse of quickly depreciating compute. The discussion broadens to environmental and regulatory risk, legal concerns around rapid construction practices, and the durability of GPU markets if AI demand or model economics cool. Overall, the article and discussion frame the outcome as a tradeoff between a near-term, high-revenue compute business and a long-term, uncertain frontier frontier-model roadmap.

The post is a controlled walkthrough for intentionally introducing ZFS corruption on Linux using throwaway file-backed pools, with explicit warnings not to test on real disks. It starts with zinject for quick simulated checksum faults, then pivots to a manual method to map a file to physical storage so readers can observe where bytes land. The author creates a single-vdev pool and a RAIDZ2 pool from ordinary image files, documents dataset/object lookup details in zdb, and explains why DVA offsets require adding the 4 MiB vdev reserved prefix before converting to sectors. A key correction is that compressed blocks initially masked the true layout, so disabling compression and rewriting the file restored expected one-to-one logical and physical block sizing. The writeup then validates addresses by reading sectors with dd and hexdump, derives RAIDZ2 stripe math from source variables (ashift, columns, parity count, offsets), and accounts for column wrap behavior. It corrupts a known sector in the exported pool to avoid ARC interference, reimports, and runs scrub. The single-vdev case reports a checksum error and irrecoverable data issue because no redundancy exists. The RAIDZ2 case uses the same method but shows parity-based correction: one bad data column is detected, corrected from parity, and scrub ends with repaired data and clear device-level fault attribution. The concluding lesson is that checksum-based self-healing is useful, but understanding how abstractions map to sectors is essential for debugging, and commenters add that random byte replacement does not perfectly emulate all failure modes in real hardware.

Apple says it is introducing a new Siri AI experience as part of Apple Intelligence, with claims of app integration, context-aware assistance, privacy-focused on-device and private cloud processing, and an English-language launch later this year. The announcement also emphasizes new generative features for writing, reminders, and image editing-style edits inside the Apple ecosystem. Commenting readers compare the preview to earlier promises and argue that useful AI on iPhones and iPads has not materially moved from basic voice commands into high-leverage automation. They repeatedly point out that real-world value is still limited for planning, travel, and workflow-heavy tasks, where existing AI systems are often seen as error-prone, hesitant, or unable to complete end-to-end actions. Many comments question the practical significance of Siri UI-level improvements versus actual reliability gains, including core friction points like dictation, search, and timer controls. Device support restrictions to newer hardware and delayed availability for older iPhone and iPad models are described as a major disappointment, especially because earlier models were previously marketed as AI-ready. Another recurring thread is uncertainty around regional rollout, particularly EU availability, as well as possible dependence on Gemini or external models despite Apple branding. Some users remain open to the underlying integration approach through shortcuts and app-level orchestration, but overall the announcement is treated as incremental and still unproven at launch.

Apple describes how dsymutil creates self-contained debug bundles for crash reporting and remote debugging by deduplicating DWARF types with the One Definition Rule, which is essential for large C++ projects where full per-translation-unit type data can otherwise become massive. The prior linker was single-threaded because DWARF output is huge and graph-oriented, so Apple added incremental multithreading but kept ODR uniquing on one thread, making it a bottleneck. A production-quality parallel ODR linker in LLVM existed conceptually but had blockers, so the team focused on correctness validation. Because prior qualification required byte-identical outputs, the new parallel linker needed a semantic comparison strategy: it reorders and rewrites DIEs differently, so text diffs are unreliable. They built a semantic DWARF differ that anchors matching with stable identifiers such as linkage names and signatures, then compares reachable subgraphs across outputs. A clang prototype reported roughly 50,000 divergences among about 5 million DIEs, giving a concrete set of mismatch cases even though not all were yet verified. Reproducibility concerns were addressed by assigning deterministic priority to canonical DIE candidates based on compile-unit link order so all runs choose the same winner under ODR races. To switch defaults, they are running existing dsymutil tests with --linker=parallel, LLDB scenario tests against parallel-produced dSYMs, and increasing manual diff passes on larger codebases. Reported runtime drops from about 3 minutes to about 40 seconds on clang, and the article notes that Apple builds with -fno-limit-debug-info increase DWARF volume by keeping full type definitions, making efficiency and reliability gains more valuable.

The post argues that a function taking a raw memory blob should keep the classic C signature void DoSomething(const void* p, size_t numBytes) because it is simple, readable, and avoids extra casting when passing arbitrary structs like a custom data object. The author criticizes replacing it with uint8_t* or std::byte* as overcomplicated and says modern alternatives such as std::span and template-based forms add verbosity for little benefit in this use case. The example focuses on operations like hashing or raw disk writes, where the code naturally deals with byte views, and the writer favors clarity at call site, showing DoSomething(&data, sizeof(data)) as the clean form. The post also proposes SAL annotations as a practical improvement, using _In_reads_bytes_ to let static analysis reason about buffer size and lifetime while keeping the same interface. The comments largely reject that position, saying modern C++ should preserve intent and type information rather than erase it, because a void* alone lacks size, ownership, alignment, and provenance context. Commenters favor std::span<std::byte> for coupling pointer and length, or template boundaries that prevent accidental misuse, and call the C-style approach a frequent source of security and maintenance issues. Others defend void* as pragmatic in low-level work, especially where existing C interop patterns already handle bytes this way. A few also defend char* or reinterpret casts as the expected low-level bridge when byte access is truly intended. Overall, the discussion frames a tradeoff: brevity and historical compatibility versus explicit contracts and safer API design.

Eagle became a major CP/M vendor before pivoting to IBM-compatible PCs after 1981, first with the partial-compatible Eagle 1600 and then better products including the compact, quiet Eagle PC 2, the Spirit XL portable, and the faster Turbo XL with an early hardware turbo switch. The company was scaling rapidly, selling about 12,000 units a month and filing for a $37 million IPO on June 8, 1983, when CEO Dennis Barnhart died in a Ferrari crash at age 39; his death is often described as the company’s defining tragedy but the article argues the details are often simplified. Barnhart had a blood-alcohol reading of 0.14, yet the crash was not simply a story of excess, and accounts suggest a broader context involving an inexperienced passenger and the purpose of the visit. A second IPO followed a week later at $33 million, about 11 percent lower, and leadership changes followed with Ronald Mickwee taking over. In early 1984 IBM sued Eagle over alleged BIOS copyright violations, and while Eagle has often been portrayed as a messy infringer, the piece cites an anonymous former employee claiming Eagle did use clean-room-style separation with legal review, then settled quickly and swapped socketed BIOS chips in the field rather than recalling shipped systems. The lawsuit still damaged momentum by raising customer concerns just as shipping delays and competition pressure increased. Phoenix’s release of a widely licensed clean-room BIOS in mid-1984 lowered entry barriers and increased clone supply, while IBM expanded output and cut prices, ending the scarcity advantage that had helped smaller makers. Eagle then faced layoffs, executive churn, and losses that reached $7 to $9 million in a quarter, with no profitable quarter after. By 1985 the firm had lost over $40 million, shrank to

A Foodwatch investigation tested 64 samples of rice, tea, paprika, cumin, and curry from the Netherlands, France, Austria, and Germany and found pesticide residues in 49 of them, with 45 containing substances not approved for EU use. Fourteen samples were above legally allowed residue limits, and all tested paprika powder, chili, and cumin samples carried non-approved residues, including one paprika sample with 22 different pesticides and six that are banned in the EU. The findings were linked to the “boomerang effect”: pesticides no longer sold in the EU are still exported from member states to third countries and can return in imported food as residues. Official data from the European Chemicals Agency show six of the detected compounds were exported by EU members in 2024–2025. The report also notes that officials are considering a fast-tracked package of ten food-safety measures that would loosen pesticide safety reviews, residue limits, and import controls, and critics argue this could weaken current protections while these residues remain in everyday products.

A long-time ThinkPad enthusiast and coreboot contributor described porting coreboot to the 20-year-old x61, whose GM965/ICH8 platform had no public documentation and had resisted prior attempts. They first dumped reference data from vendor firmware using standard tools such as inteltool, lspci, acpidump/acpixtract/iasl, and ectool to capture ACPI tables, chipset state, EC behavior, PCI configuration, and audio/CPU details. The BIOS was split into modules, and an AI assistant was given firmware analysis tools like ghidra-cli and radare2, enabling rapid extraction of memory init behavior and producing a first bootable implementation in just a few prompts, though that claim was later corrected as a partial simplification. The author explains that dozens of hardware-specific corrections were still needed, including GPIO, SPD visibility, FSB sourcing, register width handling, memory timing, and multiple raminit variants tied to recovery behavior. A second phase used serial traces and flash testing, including SOIC8 clip flashing with flashprog, to validate startup behavior and iterate on failures. Porting libgfxinit to GM965 was comparatively straightforward thanks to Linux reference behavior, while upstreaming exposed broader bugs in register naming, bit semantics, default assumptions, and timing logic. A core reviewer helped turn the initial bring-up into upstream-safe code, and style tooling friction (clang-format incompatibilities) became one of the biggest non-functional obstacles. The author concludes that although the work can now be done in weeks instead of months and points to future targets like Intel FSP, success still depends on deep expertise and legal caution, especially around reverse-engineering boundaries.

A developer announced H2JVM, a new Haskell library intended to let compiler authors emit JVM bytecode at a higher level by abstracting repetitive concerns such as StackMapTable generation, label resolution, and offset calculation. In the introduction, they showed a minimal Calculator class example that generates a static add(int,int):int method and a more realistic snippet for compiling a > expression using labels to branch and merge control flow. The library is described as early-stage, with only a subset of bytecode instructions and class attributes implemented. Commenters praised the label-resolution design as aligning with established JVM tooling models used by the Java Classfile API and ASM, while also noting an important gap around class hierarchy information needed for stack-merging decisions in conditional expressions. They highlighted the need to infer a common superclass or interface for branch values, such as combining new Foo() and new Bar(), and referenced the Java Classfile API as a possible guide for higher-level type-merging behavior. Another commenter challenged the motivation to target the JVM from Haskell, describing Java as personally limiting and questioning portability as a core value, and suggesting the tradeoff may be smaller than commonly claimed.

Xiaomi and TileRT introduced MiMo-V2.5-Pro-UltraSpeed, claiming over 1,000 tokens per second generation on a 1T model via model-system co-design, with applications by trial period from June 9–23, 2026, a 10-entry daily queue cap, and 30-minute sessions. The API is positioned at three times the normal Pro price and is not included in the token plan, with access targeted at enterprises and professional developers. The release links speed gains to broader use cases, arguing that faster decoding enables speculative depth, faster coding agents, and real-time workflows in trading, anti-fraud, bidding, and medical contexts. On the technical side, Xiaomi says it applies selective FP4 quantization to MoE experts plus Quantization-Aware Training, while leaving other modules full precision to preserve capability. It also adds DFlash block-level masked parallel prediction to reduce serial drafting and cites acceptance-length gains in coding, reasoning, and agents, while noting weaker general-conversation acceptance. TileRT is described as the systems counterpart, with custom kernels and execution paths intended to remove microsecond operator gaps on commodity 8-GPU hardware. Comments acknowledge the engineering ambition and benchmark framing but repeatedly ask whether 1,000 TPS improves real outcomes when quality, reliability, and pricing remain uncertain.

Old'aVista presents itself as a guide to the old internet, preserving vintage web content and emulating nostalgic browsing conditions. The page highlights a broad archive structure, community support, and acknowledgements for contributors who helped make access to historical web material possible. It also includes a sponsor and support model centered on merchandise and recurring donations, signaling that the project is community-funded rather than purely ad-supported. The site's identity appears playful and deliberately retro, aiming to make older internet eras accessible on legacy hardware. Comments suggest users are reflecting less on the archive as breaking news and more as a cultural artifact of how web culture, interfaces, and community memory formed. Collectively, the thread shows the site prompts both emotional recall and practical curiosity about historical authenticity and operational details of such retro experiences.

After years of rejecting hints, PostgreSQL 19 adds planning controls through pg_plan_advice and pg_stash_advice, framed as plan advice rather than SQL comments. The piece revisits the long historical resistance based on maintenance risk, upgrade fragility, weaker root-cause analysis, and concern that users would stop fixing underlying issues, while noting repeated recognition that edge cases need a controlled escape path. Haas’s design keeps guidance outside query text via a GUC or a persisted stash keyed by query ID, so applications do not embed permanent directives. Advice only narrows planner choices instead of replacing the optimizer, and when guidance is stale or conflicting, the planner degrades gracefully. The system can emit readable advice from EXPLAIN, supports scan methods, join order, join method, and parallel controls, and lets operators lock critical behavior while leaving parts of planning flexible. In production, pg_stash_advice enables shared, scoped policies by role, session, or database, with on-disk persistence and easy updates. Feedback marks each directive as matched, unmatched, inapplicable, failed, or conflicting, allowing teams to spot bad overrides early through logs. Limitations remain around aggregates and set-operation strategy control, plus planner overhead from constraint checks, so the author positions this as a last-resort tool after statistics, GUCs, and CREATE STATISTICS improvements.