The newest HackerNews stories, desensationalized.

The post introduces "Stolen Buttons," a curated stash of button-like interface elements copied from many websites. It showcases labels and controls such as sign-ins, shopping actions, contact links, account access, and account-management prompts, showing varied style, language, and branding contexts. The collection is mostly visual and often static, so many items do not function as native buttons and therefore read more like snapshots of visual trends than a tested component catalog. Commentators frame it as both a nostalgic inventory and a critique of how modern UI styling has evolved, with some seeing playful historical value and others seeing visual noise or weak interaction design. The content implicitly documents a broad range of implementation quality and design intent, from minimalist flat shapes to nostalgic skeuomorphic effects. Several entries also expose practical issues like dead assets and dependence on external image hosting. Together, it serves as an informal but imperfect record of how interfaces communicate actions, and where they can fail to signal clickability or user trust.

The piece argues that AI is approaching a Kubernetes-like phase where open-weight models act as a shared, extensible platform even though most releases are not fully open source by strict OSS definitions. The author cites self-hosting needs for data sovereignty and cost control as the first driver of adoption, and points to a growing serving and deployment stack around vLLM, SGLang, llama.cpp, Ollama, and MLX. As benchmark and coding performance improves, open-weight families such as GLM-5.2 and Kimi K3 are narrowing gaps with closed leaders, which could accelerate derivative ecosystems of fine-tunes, runtimes, and operational tooling. The argument is that these derivative layers can compound faster than what any single vendor can build alone, especially if startups, hardware firms, and cloud providers invest in distribution and operations. The policy warning is that broad U.S. restrictions on Chinese open-weight models may cut off American researchers from a major innovation source while others keep building. The proposed alternative is to release more frontier-capable American models under permissive terms, use procurement to favor portability and interoperability, and rely on standards and independent testing instead of blanket bans. Comments introduce counterarguments about funding sustainability, national-security risks, and practical deployment economics, but the article’s core thesis remains that engagement and competition beat isolation.

A developer outlines concerns about a Google IssueTracker proposal tied to CVE-2026-0073 in which Android ADB maintainers discussed binding the ADB daemon to selected interfaces, with a suggestion that could effectively block loopback use. The author explains ADB’s evolution from USB to wireless workflows and why on-device ADB (client and daemon on the same phone via 127.0.0.1) became a practical path for projects like Shizuku, libadb-android, App Manager, and other power-user tooling. The issue, they argue, is that it also enables unique accessibility and legitimate workflows beyond mainstream developer setups, including call-recording tools used in constrained personal contexts. They then map realistic attack chains and conclude that malicious apps cannot unilaterally start ADBD or connect silently: the user must already enable developer modes and grant permissions, which keeps many threat scenarios constrained. Even in older Wi-Fi and TCP/IP workflows, exploitation still depends on human actions such as pairing, prompts, and USB debugging, though CVE-2026-0073 reduces trust barriers if those prerequisites are met. The author supports reducing exposure to specific network interfaces but argues that default prevention should not become permanent and non-user-reversible blocking. They recommend a persistent user-controlled toggle that cannot be easily observed by third-party apps, with the alternative of default-off security and explicit opt-in for loopback for advanced usage. Overall, they frame the request as a tradeoff: narrower attack surface versus preserving legitimate on-device innovation and niche developer use cases.

The article explains that Windows issues a Microsoft-tied GDID device identifier at sign-in and presents it as a tracker that persists through VPN use and does not rotate like an exit IP. The identifier is stored in Microsoft’s systems in a 64-bit form, and the article says removing the local registry value only causes it to be rehydrated on next use. It also claims standard telemetry controls do not reliably stop GDID reporting because the data path appears to involve other Microsoft components such as connected-device and delivery-optimization services. The piece then proposes a practical mitigation through tested scripts in the no-gdid project: audit first, preview changes, then optionally apply service blocking and hosts redirection while keeping the Microsoft account logged in, with a revert script available. The guide recommends testing inside a VM because system services are altered. It emphasizes that mitigation is not a deletion; the GDID value remains on disk and Microsoft likely retains historical linkage from first login onward. It also notes that replacing GDID via reinstall appears possible but does not erase previously associated data. The author concludes that this reduces future correlation risk but does not grant anonymity, and recommends using non-Windows systems for truly sensitive activity.

The article argues that scaling modern LLM training across hundreds or thousands of GPUs now depends as much on reliability as raw throughput, since single GPU, network, or node failures can erase long stretches of work. It describes PyTorch Monarch as a single-program, actor-based distributed runtime with process meshes, supervision trees, and asynchronous execution that isolates failures to local actors and supports fast local or escalated recovery. To expand support beyond CUDA, AMD engineers ported Monarch to ROCm by converting CUDA bridge code with hipify_torch, wiring distributed communication to RCCL, adding platform-aware GPU API routing, and preserving RDMA with GPU-direct transfers via HIP bindings. They also solved ABI parity by introducing a Rust compatibility shim so hipified types could be re-exported under CUDA names, avoiding forked call sites, and reported full ROCm 7.0+ compatibility with 1,171 tests passing. The team integrated Monarch with TorchFT and TorchTitan into a dynamic recovery stack where a Lighthouse service coordinates quorum boundaries, failed replica actors restart in place, and peer checkpoints transfer state from healthy replicas without global restarts. In a 16-node, 128-GPU MI300 SLURM run with failures injected every 180 seconds, active workers fluctuated between 8 and 16 while training proceeded with steady convergence and no full-job replay. A 32-node, 256-GPU MI355 Kubernetes run showed similarly stable participation (roughly 30 to 32 replicas) and smooth loss decline from 12 to around 4. The article frames these results as a practical step toward production-ready AMD-based fault-tolerant training, with future work aimed at broader framework support, better NIC coverage, and lower recovery latency, while continuing open-source PR

The post introduces a command-line ImageMagick workflow to give photos a printed, pink halftone look rather than true file-size optimization. It explains AM, FM, and hybrid screening, with AM’s dot pattern and moiré issues, and notes a practical fallback from naive AM to rotated CMYK channel overlays. The method starts from resized source images, converts to CMYK, separates channels, applies Gaussian-treated geometric offsets, overlays layers, then recombines, with variants for bigger dots, binary channel reduction, grayscale leveling, and palette remapping. The author explicitly distinguishes the look from efficient compression: the approach is compute-heavy, often larger than expected, and mainly a playful aesthetic. They present an iterative refinement from two-color pink output toward cleaner simulations using -colors 2 per channel and later -colors 32 with a remap palette. The post also includes a bash loop script for batch conversion and compares these custom results with true Floyd-Steinberg dithering, which is presented as the standard reference. A closing RGB variant is mentioned as less print-accurate but faster to conceptualize, and the author repeatedly warns that people should be clear about intent and expectations.

Vercel redesigned its project cards to reduce vertical footprint by trimming fields like branch names, tightening spacing, and co-locating metadata, allowing more cards to fit in the dashboard grid. The team replaced the persistent card shortcut control with a context menu under a feature flag to create a cleaner card while preserving core actions in a detail-view dropdown so controls stay discoverable. They found this two-path approach useful because a menu can be activated from the card surface, while the detail view provides explicit access for users who prefer clearer affordances. For card metadata, they expanded beyond the single Real Experience Score by planning configurable widgets for metrics like performance, analytics, alerts, and security signals, since users focus on different operational priorities. Initial widget experiments were rejected until a two-row format made each widget easier to scan and visually distinct. The resulting widget area is now compact and aligned, reducing awkward floating space. The widget switcher was simplified to either show alternate data previews or, when data is absent, present a useful platform-feature overview that can also be used to enable capabilities directly.

The article argues that expression syntax can move beyond one-dimensional text by placing operators and operands in two dimensions, then builds a custom infix operator to make 3-argument composition visually direct. It introduces andFlip and its infix form @@ to chain two-operand state changes through a scratch bit, then extends the pattern so partial effects can be undone inside the same expression and remain reversible in style. A chicken-coop gate example shows why toggling-in-place differs from plain and by preventing immediate rollback when one condition drops, and a second example shows reusing dirty scratch state, which the author ties to reversible and quantum circuit techniques. The piece then generalizes with companion operators @ and @@, derives a majority-like maj gate and its inverse umA, and demonstrates a full three-bit adder in one mirrored 2D expression that computes sum and carry while restoring temporary wires. It is presented as cleaner than hand-tracking in a direct Qiskit listing and visually closer to circuit thinking. The final tour ranges through existing spatial systems such as Befunge, Orca, 2dcond in Racket, Hexagony, ladder logic, and spreadsheets, framing layout as a language design constraint that can shape thinking. The comments are split between appreciation of the idea and strong skepticism, with critics noting that text can already represent any of these ideas and tends to remain more practical, dense, and maintainable. Some commenters praise the playful examples, while others question the necessity of 2D as a mainstream alternative and question whether the closing linguistic framing is overstated.

The walkthrough explains how Fedora turns source changes into installable releases in staged, auditable steps, starting from package repositories at src.fedoraproject.org and commit-based fedpkg build submissions to Koji. Koji compiles packages in isolated buildroots, tracks architecture and tag state, and also drives image tasks such as kiwiBuild, imageBuilderBuild, and OSTree runroot work. For non-Rawhide releases, Bodhi gates updates through testing states using karma, critical-path rules, and automated CI checks, then promotes to stable and triggers repo composition; Rawhide follows a faster or differently constrained model. Pungi coordinates final release composition by freezing package sets, applying comps and variants, generating boot installers, and handing off to tooling paths that produce dvd.iso, cloud and virtual machine images, container and WSL artifacts, and Atomic OSTree commits. Image Builder with osbuild handles manifest-driven, sandboxed pipelines for bootc and ostree-based variants, while rpm-ostree and treefile layering keep atomic desktops aligned with Fedora’s package groups through comps-sync. Metadata generation via productmd (composeinfo, images, treeinfo, rpm lists) enables installers, Bodhi, openQA, and mirrors to consume each compose consistently. openQA then validates nightly Rawhide and milestone builds through interactive and regression testing, and failures can block releases. The post closes with the release-process governance model, where systemic changes pass the Changes workflow, committee review, and release-schedule checkpoints, with explicit deadlines controlling deferral risk.

Hannah Fry was awarded the Leelavati Prize at the ICM for outstanding public mathematical awareness, after joining Cambridge as the first Professor of the Public Understanding of Mathematics in January 2025. The award, given every four years by the International Mathematical Union, sits alongside the ICM’s role in celebrating major mathematical milestones, including the Fields Medals. The citation praised her for translating mathematics across books, television, video, and podcasts without reducing its rigor. She is also credited in the piece with a longer arc of recognition, including academic honors such as the Christopher Zeeman Medal and the Royal Society’s David Attenborough Award, plus media awards including an Emmy for a quantum-computing episode and a Webby for her podcast work with Michael Stevens. The article emphasizes her large followings and mainstream accessibility, including social media reach, and highlights her framing of mathematics as a creative way to think about the world. Comment input reinforces this framing by repeatedly citing her school- and youth-facing talks, her university lectures, and her practical communication philosophy that engagement depends on sparking curiosity.

Near Kaetés Reserve in Espírito Santo, biologist Gabriel Falquetto found an endangered buffy-headed marmoset killed on ES-164, illustrating how highways fragment habitats and isolate wildlife in Brazil. In Alta Floresta, Reconecta installed eight canopy bridges with local-government support, and camera traps logged 15,000 crossings over 15 months with no recorded roadkills, reinforcing the low-cost design’s practical value for fragmented forests. The project’s dual-camera setup tracks both bridge use and nearby animals that pass by without using it, and monitoring recorded several primate species plus sloths, opossums, and mouse opossums, including the Alto Floresta titi, a critically endangered newly described species. Earlier work on BR-174 with the Waimiri-Atroari documented about 1,250 primate crossings in five years and showed adoption can rise over time, with some species adapting only after months. Designers tuned the bridges for different locomotion styles using a multilayer rope configuration, while researchers emphasized that roads in Brazil cause long-term declines by reducing gene flow, increasing inbreeding risk, and weakening ecological roles like seed dispersal and pollination. Contextual data cited in the report estimate nearly nine million mammal roadkills in a year, making habitat fragmentation and mortality a major conservation concern. A federal wildlife-road safety bill is pending in the Senate, and the National Department of Transport Infrastructure has already recommended the design for highways, while Reconecta plans expansion to Suriname, Lucas do Rio Verde, and Pantanal corridors including BR-262. The effort’s future is framed as technical and political at once, requiring utility upgrades, maintenance, and broad participation from agencies, the

A Minnesota research facility ran a 1.65-megawatt wind turbine to power on-site electrolysis and air separation, producing hydrogen and nitrogen for about one ton of green ammonia per day, with local storage and transfer to farm tanks. The article presents this as a local alternative to fossil-fuel-dependent fertilizer production, arguing that U.S. agriculture’s emissions exposure, volatile global fertilizer costs, and tightening low-carbon requirements are pushing producers toward decarbonized inputs. It also emphasizes that Haber-Bosch historically enabled large-scale food production but remains highly greenhouse-intensive, while imports create supply-chain and energy-risk exposure for farmers. The project is framed as a resilience model: using wind power to run chemical conversion steps at the production site instead of relying on grid electricity tied to hydrocarbons. Some commenters argue the central point is more precise: the turbine is not “making fertilizer” in itself, but supplying electricity to a chemical process. They agree that if such flexible loads are paired with curtailment-prone wind output, they can reduce wasted renewable energy and smooth grid integration. Others stress that the article reads as marketing-heavy and ask for transparent lifecycle and cost comparisons with conventional Haber-Bosch plants before declaring competitiveness. The comments also note that renewable intermittency, regional economics, and competing storage options such as batteries remain open questions. The overall discussion frames the case as an early demonstration with technical promise but incomplete operational and cost proof.

Starting from a raycaster benchmark, the author found the Playdate far from 3D-accelerator class performance but enough for a perceived retro style at its 1-bit resolution. The write-up explains that without a GPU, the CPU must handle projection, clipping, sorting, texturing, depth handling, and framebuffer work in real time, so throughput is the core constraint. To reduce runtime cost, the project uses Quake BSP maps with TrenchBroom and ericw-tools for compiler-time geometry splitting, visibility sets, and lighting, then relies on PVS data to skip unseen regions. The renderer is custom-built and moved from painter-order drawing to a 16-bit reciprocal z-buffer so masked textures and moving elements can depth-test correctly. Perspective-correct texture mapping was chosen over affine mapping, then optimized by recalculating reciprocals periodically and approximating intermediate pixels. Because the display is monochrome, BSP-derived light data is fed through an 8x8 Bayer dither, and the visual strategy shifted to strong outlines and simple cel-like shading for readability. Significant gains came from half-resolution rendering scaled to the native framebuffer, Cortex-M7-specific compiler settings, hard-float settings, linker cleanup, custom math helpers, caching, and selective hot/inline annotations. The author frames progress as iterative and measurement-driven, using benchmarks to keep only useful optimizations, and concludes with a playable prototype rather than a finished renderer, focused on getting gameplay in place for an underwater metroidvania.

The project details a self-built mechanical television using a single spinning scan drum shared by five 9×8 mm windows, with separate LEDs per window to display independent images. Despite only 20 scan lines, the design boosts effective quality by using the Raspberry Pi Pico’s PIO engine for fast LED control. It specifies a practical parts list: a Symbol 21-02485 stepper converted to bipolar use by disconnecting the white lead, an A4988 stepstick like the Pololu 1182 with current reduced to the minimum before stalling, a 100µF motor supply capacitor, Pico (with Pico W for wireless streaming), and mixed RGB plus white LEDs. The instructions stress electrical limits, requiring matched LED brightness and a total GPIO draw below 50 mA, with example resistor values derived from measured channel currents (about 10.75 mA red on 120Ω, then 260Ω green, 220Ω blue, 82Ω white). The build also depends on print quality: three STL parts (base, drum, lid) plus OpenSCAD sources for customization, including nozzle-size tuning, high-quality drum printing, single-wall sides, Arachne walls, and 100% infill; an optional washer can raise drum inertia. Software runs on MicroPython 1.28 in pico/scanwheel.py for device control, with host-side tools in host such as videostream.py, streaming.py, and pngdirect.py for converting and streaming images.

The post explains a compact apartment aquaponics setup that combines a 20-gallon fish tank with a media-based, continuous-flow grow bed placed above the tank on a metal shelving rack to save space, reduce leak risk, and reduce pumping distance. The authors switched to a 4-inch slotted top drainage pipe and a better bulkhead seal to prevent clogging and leaks, kept a smaller 160GPH pump for quiet reliability, raised and timed the grow light to avoid leaf burn, and plan to use black tubing to cut algae. They selected small-root crops and report basil and herbs doing well while testing dwarf tomatoes, and cite tradeoffs of a small footprint that limits production and nutrient-heavy plants. The aquatic population is mainly neocaridina cherry shrimp, ramshorn and nerite snails, plus a feeder guppy, with water cleanup handled by plants, gravel, and coral media. Lessons emphasized include cycling fishless for weeks before adding life, avoiding overfeeding, preferring larger tanks for stability when possible, and using local hobbyist communities for fish and plants adapted to local water conditions. The step-by-step section lists materials and a detailed build sequence, with total setup costs around $663 before tax for one-time hardware, media, lighting, and plumbing. Planting guidance covers replacing decorative and fake plants with donated or low-cost live plants, rooting techniques, and raising seeds in rockwool for months when shelves are limited. Maintenance is intentionally light: alternate-day feeding, weekly water replacement with dechlorinated water, pruning, and occasional substrate stirring. Estimated monthly energy use is roughly $14 using NYC rates, with heater consumption variable by thermostat duty cycle.

Kyber describes an AI-native enterprise document platform aimed at regulated industries, especially insurance claims teams, where users can upload claim details and automatically generate multiple drafts of notices. The company says the workflow can consolidate 80 percent of templates, cut drafting time by 65 percent, and reduce overall communication cycles by five times. It frames the system as improving quality, accountability, and traceability through reviewer collaboration and real-time editing before sending notices. Kyber reports backing from Silicon Valley firms including Y Combinator and Fellows Fund, signaling startup credibility and growth-stage ambition. The post is also a hiring announcement for a hands-on Head of Engineering with direct reporting to the CTO, emphasizing early-stage ownership, first-principles design, and scaling mission-critical AI systems. The role is positioned for candidates with very high individual output, and applicants are encouraged to be referred by a colleague with a short endorsement sent by email, with referrals explicitly presented as a major filter in screening.

The article profiles Charles Ross’s Star Axis near Anton Chico, New Mexico, describing it as a monumental naked-eye observatory aligned to Polaris and built over decades as a major land-art interpretation of astronomical precession. The site is promoted as being in a dark-sky region between Santa Rosa and Las Vegas, NM, with a light-pollution rating near the best possible level. Reporters also note that the project was once considered for state acquisition, but lawmakers ultimately abandoned the purchase effort. A major current tension is access: tours can cost around a thousand dollars and high-end patronage options can reach six figures, so participation is framed as privilege-proximate despite the remote and publicly meaningful setting. Comment threads highlight this contrast, treating the work as both a technical-cultural monument and a case study in how fragile this kind of experiential art can be under market-driven visitation models.

ARC-AGI-3 is presented as the latest step beyond ARC-AGI-1 and ARC-AGI-2, replacing passive, fluid-intelligence style testing with agent-style adaptation in new interactive environments. The leaderboard uses a cost-per-task versus performance view to emphasize efficiency as a core criterion for intelligent behavior. It separates reasoning-system lines, where the same model is shown at multiple reasoning levels with performance typically improving then flattening as thinking time rises, from base LLM baselines like GPT-4.5 and Claude 3.7 that use standard single-pass inference without extra reasoning. Kaggle systems are listed as competition entries optimized under explicit constraints, including a $50 compute budget for 120 evaluation tasks. The notes impose a display limit to systems running under $10,000, label incomplete outputs as incorrect for remaining tasks, and flag preview entries as unofficial, with partial-cost estimates used for ARC-AGI-2 and a pre-release model pricing assumption. Footnotes also indicate provisional estimates and planned retesting once certain models are officially available. The page directs readers to a separate testing policy for details and cautions that some results are not directly comparable because of incomplete runs and preview conditions.

The product is a custom-fit, mouth-worn mouthpad that uses tongue, head, and breath gestures to move a cursor, click, scroll, and drag on phones, tablets, and computers via Bluetooth mouse support, with setup handled through a companion app for sensitivity and speed tuning. It is promoted as an accessibility-focused interface for users who cannot use hands comfortably, with testimonials describing restored independence for communication, work, gaming, and daily device use. The hardware is about 1 mm thick, about 10 g, made from dental resin, and includes a charging case and USB-C/USB-A cable. The workflow requires a dentist-administered or local intraoral 3D scan, then 3D manufacturing, with shipping taking about six months. Pricing is $1,400 for one device and $1,900 for two, with scans billed separately and no insurance or FSA/HSA coverage, and availability is limited to US shipping despite global usage capability. The FAQ notes comfort is individually variable but users report up to ten hours of use and little speech impact for many, while also requiring careful cleaning and charging habits for hygiene and longevity. Comments frame the launch as a strong accessibility milestone and a promising non-hand interface, while also noting potential naming puns and concerns about whether comfort and fatigue or adaptation limits matter over time. Community reaction is split between excitement over niche capability and doubt about broad-market viability at a premium, custom-made price point with a relatively narrow initial target population. Practical constraints include wash-down care, no use while sleeping, avoiding heat and chewing, and a one-year limited warranty with explicit damage exclusions. The company positions the device with a future path that includes a companion “

The post documents a 2026 teardown of BMW’s IDC23H, sold as Harman’s IDC23H and Garmin’s IDC23 variant, noting that its naming mirrors the prior MGU22/MGU22H split where the “H” marks Harman-built models. The unit is installed in newer BMW models such as the X1 U11 and adds notable APIX-based video paths beyond prior generations, including direct links for the Driver Camera System, Heads-Up Display, and the driver’s display, which appears to replace a standalone instrument cluster by driving the cluster from the head unit. The author lists connectivity as APIX variants, dual-band Wi-Fi 6, Bluetooth, GNSS, multiple Ethernet variants, AR-CAM, and multiple USB ports, and highlights that this could broaden in-vehicle display integration and customization demand among owners. Internally, the device remains close to MGU22H architecture but gains a new APIX daughter PCB, still using a Snapdragon SA8155P system-on-module plus Infineon CYT3BB microcontroller, 12GB LPDDR4 RAM, and 128GB UFS storage. PCB notes include a 100-pin main interconnect to the APIX board, an aggressive cooling assembly with fan and heatsink, and chips such as APiX transmitters, TI power switches, Marvell automotive Ethernet switch, NXP radio SoC, and Infineon CAN/LIN-capable control, while labels and dumps are documented for traceability. The author maps likely roles: Snapdragon for higher-level processing and Infineon for low-level vehicle communication, though the OS was still inaccessible awaiting adapters. The post repeatedly compares the board to MGU22H, calling the redesign effectively “MGU22H + APIX board,” and adds a naming/connector uncertainty around DFE, likely a German-origin driver display acronym. The appendix adds an unresolved hardware question around an IDC23 High 8185 variant: FCC filing

PyPI started rejecting uploads to releases older than 14 days to reduce supply-chain risk from compromised publishing tokens and to simplify recovery after compromise, but it is a temporary policy without formal release-state semantics yet. The change was discussed during the 2024 PEP 740 work and revived after 2026 compromises at LiteLLM and Telnyx, where mutable release references allowed late poisoning. Security engineering measured impact by scanning package data, including a sample of top packages, and found only 56 of 15,000 had added a 3.14 wheel more than 14 days after release, suggesting limited current disruption. At the PyCon US packaging summit, the community broadly accepted requiring version bumps for new Python support rather than reopening old releases. Seth Larson implemented and merged the restriction on July 8, 2026, and now calls for waiting on the future Upload 2.0 API and staged release semantics from PEP 694 to define explicit closure states. The update is intended to reduce mixed compromised states within one release and lower cleanup burden for maintainers and registry operators, while still leaving edge cases open until API support arrives.

The guide is a practical, task-oriented manual for iproute2, framed as the modern replacement for legacy network tools like ifconfig, route, and brctl, with tc and bridge noted as related components and ss as the netstat replacement. It starts with tooling context, command syntax conventions, permission requirements, and root versus read-only operations, then introduces machine-friendly output modes (-o, -br, and -json) and support for both dotted masks and prefix lengths. It gives extensive IPv4/IPv6 command workflows for address lifecycle management, neighbor table inspection and manipulation, and link management including up/down state, aliases, MTU, renaming, MAC changes, groups, VLAN/QinQ creation, and MACVLAN or dummy interfaces. The document continues through virtual and transport constructs, covering veth pairs, IFB, tunnels (IPIP, SIT, IPIP6, IP6IP6, GRE variants, L2TPv3), VXLAN, and GENEVE, with notes on version support, mode details, and operational caveats. It then shifts to routing fundamentals and advanced policy routing, including connected routes, special route types, metrics, multipath, custom tables and rules, and rule precedence, followed by troubleshooting around route removal behavior when links change state. Further sections explain VRF and network namespaces for tenant and isolation patterns, plus monitoring of links, routes, addresses, and neigh events via ip monitor or rtmon for operational visibility. The guide ends with module and configuration caveats and reiterates licensing and cautionary guidance for production changes, emphasizing explicit commands over outdated abstractions.

Farmers in Paraná, Brazil, tokenized 10 dairy cows and listed the tokens on the B3 exchange to raise almost $20,000 in credit amid stricter local bank lending for small agribusinesses. Cowmed says the project converts each cow into a traceable digital asset tied to real-time collar data, so lenders can verify collateral status without farm inspections and avoid double-pledging. The collars monitor health, behavior, and location, while an encrypted identity links those signals to loan terms. Cowmed’s model also allows a replacement process if a tagged cow dies, and the company already monitors about 100,000 cows across 1,000-plus farms with a reported herd value above $395 million. Cowmed expects up to 20% of its network to adopt tokenized financing, which it projects could unlock roughly $77.6 million in additional agricultural credit. The launch is framed as a real-world asset tokenization pilot, while broader industry forecasts remain much larger than current adoption, with tokenized asset values far below long-term projections. The report suggests the model is practical in a constrained credit environment but still early and unproven at scale.

The piece announces that Wasmtime 47 now enables WebAssembly GC and exceptions by default, arguing the changes make Wasm a better target for languages with object models and exceptions by removing custom language runtimes. For GC, the proposal adds native struct and array types plus subtyping, so managed-language toolchains no longer embed their own collectors in binaries; for exceptions, throw/try-catch semantics replace result-code patterns that bloat binaries and slow normal returns. Wasmtime implements a Cheney-style semi-space collector using two linear-memory-backed halves and 32-bit indices instead of native pointers, trading complexity for safety, compact references, and portability across environments. This design also lets the runtime use virtual-memory guard pages, preserve fast instance startup, and inherit platform-specific memory management behavior already proven for linear memory. The team reports expanded fuzzing for GC correctness through wasm-smith and additional graph-and-corruption-focused fuzzers, while clearly stating performance is early-stage rather than peerless. The post notes trade-offs in optimization emphasis: Wasmtime prioritizes many short-lived instances over single long-lived process throughput, so collector latency and throughput trail older engines for now. It also flags ongoing work on GC-aware alias analysis and a planned integration with the component model to avoid passing data through unused linear memories. Overall, the update focuses on correctness-first support, with practical speed improvements framed as staged milestones rather than completion.

UK AISI and CAISI ran a joint preliminary evaluation of Kimi K3, released July 16, 2026 with open-weight plans by July 27, and found that it is significantly stronger than prior open-weight Chinese baselines yet behind leading U.S. frontier models on high-end cyber tasks. The benchmark set included ExploitBench, which tracks exploit-development milestones across 41 post-2023 V8 vulnerabilities, and The Last Ones (TLO), a 32-step simulated enterprise attack. Kimi K3 scored 32% on ExploitBench versus 24% for GLM-5.2, and surpassed GLM-5.2 there, but it achieved 0/41 arbitrary code execution outcomes, while top U.S. models averaged 20/41. Kimi also reached only step 17 of TLO on average compared with 28.5 for the strongest U.S. models, and solved TLO in 1/10 attempts versus much higher rates for top models. The report notes that frontier closed-weight U.S. systems were tested with safeguards disabled to measure upper-bound capability, unlike public deployments with guardrails on. It also warns these are early, limited tests using selective methodology and that Kimi’s security posture depends on benchmark coverage, token limits, and one-off run variance. The evaluators therefore present uncertain intervals rather than definitive rankings. Kimi’s only known full-attack success required favorable conditions, suggesting useful offensive capability against weakly defended systems but incomplete parity at scale.

webMARS is presented as a browser implementation of MARS MIPS 4.5 that mirrors the classic toolchain in a single web-based IDE, supporting editing, assembling, running, stepping, backstepping, and debugging MIPS code from one interface. The overview emphasizes a multi-window desktop and mobile layout with register views, segment panels, labels, message windows, run-time I/O, and dedicated tool panes for active workflows. It also includes a built-in help system with localized pages and embedded references, plus persistent browser-based workspace features for files, session restore, settings, and tool state. The implementation uses one JavaScript assembler and simulator core and is positioned without native runtime dependencies, aiming for portability across browsers. The linked MARS definition frames the project as a lightweight educational environment for Patterson and Hennessy-style MIPS instruction, which aligns with the project’s classroom-oriented positioning. Commenters respond positively, describing the interface as polished and familiar, with one explicitly likening it to NetBeans. The practical startup flow is summarized as compile, assemble, switch to execute, and run, underscoring that the project targets quick execution cycles for learners and instructors.

The Extinct Media Museum in Tokyo is a private, touch-accessible museum for media and devices it expects to disappear, curated under the premise that all media except paper and stone will become extinct. Its 4,000-piece collection focuses on home-use motion-picture and video technologies across formats from 9.5mm film systems to Betamax, VHS, and later digital variants, with about 3,000 items on permanent display and much of the rest available for visitors to handle, photograph, film, or 3D-scan after paying admission. The museum includes cameras, typewriters, computers, telecommunication devices, audio formats, and an extensive catalog archive of 3,000 product manuals from the 1960s to 2020s, plus a ScanSnapSpot enabling visitors to digitize catalog materials. Much of the collection is indexed in a web museum for free viewing, reinforcing its preservation mission through distributed documentation and social sharing. It also rents items as props for media productions, operates a gift shop with original goods, and publishes practical visitor information including access details, reduced admission for students, educators, and caregivers, and nearby transit routes. Comment discussion extends the story by comparing the museum to other specialized niche museums in Tokyo, praising its tactile, low-friction access policy, and connecting it to broader conversations about long-term data preservation, obsolete formats, and how easily cultural memory can disappear without active archival strategy.

The essay argues that disk space tends toward saturation because growth and capacity upgrades do not eliminate clutter, they mostly enlarge the eventual mess; the author reports running systems with only a few percent free even on large drives. An entropic view—random storage choices favor fuller states—explains part of the pattern, but the stronger force is delayed action: cleanup is postponed until storage is near unusable, when deleting files feels overwhelming. The same dynamic is presented as a general systems flaw across software, infrastructure, and daily life, where optimization or refactoring is deferred until pain becomes severe, leading to technical debt, traffic congestion, and lifestyle overextension. The author links this to Jevons-style paradoxes, where added capacity enables broader use but not necessarily efficiency. The proposed remedy is to impose tighter constraints than total available resources, such as budgets, to force regular optimization and prevent crisis-driven maintenance. They argue that tighter environments can improve outcomes, claiming code that runs on low-end hardware also runs better on high-end hardware and that low-friction workflows scale up. The post frames “more resources” as frequently a faster path to equal functionality at higher operational cost, not higher quality living or engineering practice.