HackerNews Digest

Marfa Public Radio operates continuously (24 hours a day, with interruptions only for lightning) and manages extensive behind‑the‑scenes functions such as fundraising, FCC compliance, journalistic‑ethics adherence, emergency response, and technical maintenance. To illustrate these operational details, the station produced a “sleep podcast” titled **Marfa Public Radio Puts You to Sleep**, which consists of reading internal documents (e.g., compliance manuals and ethics guidelines) intended to lull listeners. The podcast serves both as a transparency tool and a fundraising appeal, directing listeners to donate via atmarfapublicradio.org/donate. Accompanying the article are several images with alt text referencing local Texas infrastructure and NPR branding, but they do not add substantive content to the description.

The page presents a GameCube decompilation example written in PowerPC assembly (MWCC GC/2.0). Each function follows a standard prologue: `stwu r1,-0x20(r1)`, saving the link register and callee‑saved register r31, then loading a structure pointer (`lwz r3,0x8(r31)`). It reads two float fields at offsets 0x40 and 0x44, squares them (`fmuls`), adds the squares (`fmadds`), and computes a fast reciprocal square root (`frsqrte`, followed by Newton‑Raphson refinements using `fmuls`, `fnmsubs`, `fmadds`). A conditional test on r4 (`cmplwi r4,0`) may skip subsequent code. An indexed load/store sequence (`rlwinm r0,r5,2,0,29`; `lwzx r6,r3,r0`; `add r6,r6,r0`; `stwx r6,r3,r0`) manipulates an array element, after which the function calls `Vec_Normalize`. The epilogue restores r31, the link register, and returns (`blr`). The snippet is repeated three times, illustrating a typical vector‑normalization routine in GameCube binaries.

The guide details setting up a two‑node AMD Strix Halo cluster (Framework boards with Ryzen AI MAX+, 128 GB unified memory) linked by Intel E810 RoCE v2 (100 GbE) for distributed vLLM inference using Tensor Parallelism (TP = 2). Both nodes run Fedora 43 (kernel 6.18.x) with iGPU memory limited to 512 MB and kernel parameters `iommu=pt pci=realloc pcie_aspm=off amdgpu.gttsize=126976 ttm.pages_limit=32505856`. Required packages are `rdma-core`, `libibverbs-utils`, and `perftest`; the E810 driver (`ice`) and RDMA driver (`irdma`) are used. Static IPs 192.168.100.1/2, MTU 9000, and a trusted firewall zone are configured, and password‑less SSH is set up. The `refresh_toolbox.sh` script builds a Docker toolbox containing a custom `librccl.so` for gfx1151 RDMA support. RDMA latency (~5 µs, 50 Gbps) is verified with `compare_eth_vs_rdma.sh`. The TUI `start‑vllm‑cluster` launches a Ray head/worker cluster, sets NCCL (RCCL) interface, and runs vLLM Serve with model selection, context length, and optional eager mode. An alternative Thunderbolt 4 link is described for cases without 100 GbE. References include community contributions for gfx1151 RCCL support.

Wayfinder‑router is a deterministic, offline CLI and gateway for routing LLM queries between a low‑cost local model and a higher‑cost hosted model. It scores prompts in microseconds by analyzing structural features (length, headings, lists, code) and optional lexical cues (proofs, math symbols, constraints), producing a 0‑1 complexity score that is compared to a user‑defined threshold or tiered cut‑points. No external model calls, API keys, or network traffic are required for the decision, ensuring zero latency and reproducible routing. The tool ships as a pure‑Python package with optional extras for a web UI, Docker deployment, and OpenAI‑compatible gateway. Configuration (`wayfinder-router.toml`) defines local and cloud endpoints (any OpenAI‑style API) and weight settings; secrets are read from environment variables or command‑line fetchers at request time. It supports binary, tiered, or classifier‑based routing, calibration from labeled JSONL data, cost‑aware objectives, and a feedback loop for continuous re‑calibration. Responses include `x‑wayfinder‑router-*` headers indicating the chosen model, score, and routing mode. Wayfinder integrates transparently with existing OpenAI clients, IDE assistants, or agent frameworks, enabling cost‑effective LLM usage without code changes.

The repository “exploitarium” aggregates public proof‑of‑concept (PoC) exploits and vulnerability write‑ups previously hosted in separate GitHub projects. It preserves original README files and tracked source files, while repository‑level metadata (stars, issues, pull requests) remains in the original histories. A verification process on 23 June 2026 compared fresh clones of the former repos to the consolidated folders using Git tree data, confirming identical paths, object types, modes, and blob IDs for 12 repos and 96 entries with zero mismatches. The collection includes PoCs for diverse components such as c‑ares TCP use‑after‑free, FFmpeg RASC‑DLTA integer overflow, Firefox SmartWindow URL exfiltration, Floci API‑gateway VTL remote code execution, libssh2 CVE‑2026‑55200, libssh2 public‑key list overflow, nghttp2 upgrade‑queue poisoning, Nmap IPv6 extlen wrap, PHP 8.5.7 stream‑bucket SOAP RCE, RustDesk session permission bypass, and SystemInformer PHSvc trusted‑host LPE, among others. The author notes AI assistance was used for fuzzing harnesses and README generation but PoCs were handwritten. A disclaimer warns against malicious use, emphasizing the work is intended for open‑disclosure and educational purposes.

OpenRA’s latest playtest (20260222) introduces random map generators for Red Alert, Tiberian Dawn, and Dune 2000, allowing players to select biome, player count, symmetry and resources for both skirmish and multiplayer maps. Dune 2000 receives visual upgrades (Sonic Tank effects, damaged‑structure visuals) and “bulk purchase” logic for the Starport, alongside a community‑driven balance overhaul for all modes and difficulty tweaks for the single‑player campaign. The Tiberian Dawn HD mod now supports C&C Remastered Collection assets, adding HD sprites for the remaining custom assets and a content manager that toggles between remastered and classic artwork, audio, and music; integration into core OpenRA is planned for a future release. Map‑making tools are expanded with UI improvements and a Path Tiler for cliffs, beaches, and roads. Additional changes include a new “Other RTS” mouse mode, timed auto‑save, bots building expansion bases, groundwork for localisation, new missions for Red Alert and Tiberian Dawn, plus assorted bug fixes and performance optimizations.

The guide evaluates public DNS resolvers against transport security, DNSSEC support, IPv6, jurisdiction, and operator type. Research highlights include: - **Performance**: Plain DNS (Do53) has the lowest latency; encrypted DoH/DoT adds modest overhead but page‑load times remain comparable, while on lossy or high‑latency links plain DNS still wins. Speed varies by provider and region. - **Tamper resistance**: Encrypted DNS greatly reduces query interception; however ~25 % of DoT services served invalid TLS certificates, so well‑operated providers are preferred. - **Privacy**: Encryption hides queries from the network but not the resolver; no‑logging operators or Oblivious DNS (ODoH, e.g., Cloudflare, Apple) separate identity from queries. Traffic analysis can still infer visited sites; padding is insufficient. - **DNSSEC**: Validating resolvers (Google, Cloudflare, Quad9) prevent spoofed records and successfully handled the root‑key rollover. - **ECS**: EDNS Client Subnet improves CDN routing (used by Google/OpenDNS) but leaks part of the client IP; providers like Cloudflare omit it for privacy. - **Jurisdiction & centralization**: Legal location affects data‑access demands; a few providers dominate global recursive traffic, raising control concerns. - **Transport advances**: DNS‑over‑QUIC is the fastest encrypted option (supported by Quad9, AdGuard, NextDNS, etc.), while DNSCrypt remains an older, key‑based solution with limited usage data. - **Resolver behavior**: Diagnostic error reporting differs widely (94 % disagreement in 2023 study); Cloudflare was most accurate. Additional niche resolvers (DNS4all, BlahDNS, LibreDNS, Dismail.de) are listed, and legacy services such as Oracle Dyn and Level3 are flagged for avoidance.

Ford acknowledged that its recent AI‑driven inspection systems caused costly quality failures, prompting the company to rehire more than 350 veteran engineers—internally called “gray beards”—over the past three years. The engineers are tasked with leading quality reviews and assisting in training the AI models. COO Kumar Galhotra said reliance on automated quality checks did not yield the expected results, and the new staff will identify failure points before parts enter production. After the rehiring, Ford’s quality metrics improved, with the J.D. Power Initial Quality Survey ranking the automaker top among mainstream brands for the first time in 16 years. Despite the gains, Ford remains the most recalled automaker in the U.S., a situation executives attribute to legacy automation issues rather than the recent human re‑engagement. The company plans to continue using AI but with integrated human oversight, emphasizing that AI effectiveness depends on the quality of its training data and the expertise of seasoned engineers.

The page consists solely of a verification notice. It displays the title “One moment, please…” followed by the message “Please wait while your request is being verified…”. No additional text, navigation, or content is provided. The purpose of the page is to pause user interaction while a backend process—commonly an anti‑bot, security, or session‑validation check—confirms the request before granting access to the intended resource. No technical details, data, or actionable information are presented beyond the waiting prompt. The page serves as a temporary placeholder pending completion of the verification step.

None