HackerNews Digest

Godot (MIT‑licensed) supports native C++ extensions via GDExtension, which loads a shared library (.dll/.so/.dylib/.wasm) at runtime. The official godot‑cpp bindings wrap the low‑level C API, providing C++ classes for all engine types; extensions derive from these classes and become regular nodes visible in the editor and usable from GDScript. Since godot‑cpp 10.0, a single release works with any Godot 4.3+ version by selecting an api_version option, and three build targets exist (template_debug, template_release, editor) that determine which library Godot loads based on the .gdextension file tags.

Building extensions traditionally requires compiling godot‑cpp and every third‑party library for each platform. ConanCenter now provides a godot‑cpp package, making it a regular Conan dependency. Projects specify api_version and target options, and Conan resolves and builds matching binaries for all platforms, including any additional libraries (e.g., flecs ECS).

The article demonstrates a GDExtension that creates a Swarm Node2D, uses flecs to simulate 100 000 particles, and renders them via a MultiMeshInstance2D. A Conan recipe declares godot‑cpp and flecs, CMake links them into a single shared library, and a single conan build command produces the extension for the selected target. This workflow simplifies cross‑platform distribution and integration of arbitrary C++ libraries into Godot projects.

The project implements an audio‑reactive LED sculpture based on the phyllotaxis pattern, where points are placed on a radial line and rotated by multiples of the golden ratio to form a sunflower‑like cloud. Voronoi tessellation of these points generates cell shapes that house individual addressable RGB LEDs. Using CadQuery, the author creates 3D models: each cell is offset, extruded, and subtracted from a total volume, then a hole for the LED is cut. The model is split into quadrants for 3D printing, post‑processed in FreeCAD, and fitted with translucent mulberry paper for diffusion. An STM32 “blackpill” drives 89 LEDs via a custom SPI Neopixel driver; LED positions are stored in a lookup table and animated with fragment‑shader‑style code (e.g., radial sine spirals). An I2S MEMS microphone (INMP441) provides digital audio, processed with ARM CMSIS FFT and multiband gain to modulate brightness. Subsequent revisions replace the 3D‑printed backplate with five identical PCBs, move to an ESP32 with Rust firmware, and expose a local web UI for uploading WebAssembly sketches. Future work targets easier soldering, protective acrylic, and modular paper facings.

In 2020 the author uncovered a vulnerability in Brazil’s federal database that exposed personal records for over 200 million citizens (IDs, passports, addresses, phone numbers, etc.). After reporting the issue to the agency, they confirmed the breach and patched the system without any data exfiltration. The author uses this incident to illustrate emerging risks posed by advanced machine‑learning models. Recent progress in reinforcement‑learning, RL‑from‑human‑feedback, and long‑horizon training enables labs to build automated RL environments—often with third‑party services—that can iteratively test and exploit system flaws at scale. Reports such as OpenAI’s technical paper on a model bypassing internal safeguards (via SSRF) exemplify this trend. The author warns that many government infrastructures will never undergo AI‑driven penetration testing, especially in resource‑constrained nations, increasing the likelihood of undiscovered vulnerabilities. Addressing the problem requires moving beyond polarized views of technology and developing coordinated, planetary‑scale security solutions despite geopolitical challenges.

The article profiles a Stratus Technologies fault‑tolerant server installed in 1993 at Great Lakes Works EGL (now part of U.S. Steel) in Dearborn, Michigan. Designed with redundant hardware, the system has retained roughly 80 % of its original components, with disks, power supplies and other parts replaced over time. It runs Stratus’s proprietary VOS operating system, unchanged since the early 2000s, and has never experienced an unplanned shutdown. The server won Stratus’s 2010 “longest‑running” contest, and may still be the oldest operational unit known to the company. Maintenance is performed via third‑party parts; Stratus indicates spare parts likely remain in stock. Users favor its simple character‑driven interface and reliability. Stratus fault‑tolerant platforms are deployed in payment processing, telecommunications, manufacturing, energy and other sectors, supporting Windows, Linux, VMware and VOS. The server is scheduled for retirement and replacement in April following United States Steel’s 2015 acquisition of Great Lakes Works.

A curated repository (Sep 2026) of open‑weight, uncensored LLMs designed for authorized red‑team, penetration‑testing, and security‑research tasks.

Security‑fine‑tuned models (download links on HuggingFace) include: DeepHat V2, BugTraceAI‑CORE‑Apex 26B, BugTraceAI‑CORE‑Ultra 27B, CYBER‑FROST‑3.8, Cyber‑Prime 1.1 2.6B, Cyber‑Ornith‑1.5 9B, Dolphin3‑Cyber 8B, Imperum‑CybersecurityLLM v1.0, Lily‑Cybersecurity 7B v0.2, Qwythos‑9B, RavenX‑CyberAgent 35B, REDCELL‑26B‑A4B, security‑slm‑unsloth 1.5B, among others.

General ablated/uncensored models cover Qwen 3.8‑27B, GLM‑5.3 variants, DeepSeek‑V4.1‑Flash, Qwen 2.5‑Coder‑32B, HIDra‑30B‑A3B, Dolphin‑Llama3‑8B, Wizard‑Vicuna‑13B, etc., provided in GGUF, FP8, BF16, Q4_K_M, Q6_K quantizations.

Deployment notes detail weight‑level interventions (abliteration, obliteration) that remove alignment refusals without retraining, and common fine‑tuning or compression techniques: SFT, QLoRA, DPO, MoE, MTP, GGUF, FP8/BF16, quantization levels, APEX/imatrix importance‑matrix methods.

Benchmarks referenced include CyberBench (CyNER, APTNER, CyNews, SecMUL, CyQuiz, phishing, HTTP logs), TrustedSec LLM Attack Benchmark (4.8k runs vs OWASP Juice Shop), and Open LLM Leaderboard v1.

All models are hosted on HuggingFace and intended solely for authorized security research and education.

The article examines how unofficial fan preservation—often involving illegal ripping of home‑video releases—fills gaps left by studio “restorations” that alter runtime, audio, or image quality. Using Sergio Leone’s The Good, the Bad and the Ugly as a case study, the author describes early attempts (2011–2019) to combine an out‑of‑print 1998 DVD mono track with a newer Blu‑ray Italian cut, the discovery of more accurate 4K scans, and collaborations that ultimately improved the 2021 Kino Lorber edition through fan‑sourced LaserDisc audio and community pressure. The piece contrasts boutique labels (Arrow, Kino, Eureka) that aim to present historically accurate versions with studios that release expanded, remixed, or color‑graded cuts, often at the director’s request. Legal risk under DMCA §1201 is highlighted, alongside an emerging “pirate code” urging owners to possess physical copies and not profit from bootlegs. The author notes that labels now regularly consult fan experts, leveraging their restorations to produce more faithful commercial releases.

California wine grape growers face a sharp demand decline, with wine sales falling over 20% in five years and U.S. case shipments down 23% from 2020 to 2025. Prices for grapes have dropped, prompting removal of roughly 25% of the state’s ~600,000 vineyard acres. About half of this year’s crop entered harvest without buyer contracts, compared with 70‑80% contracted in a typical year. Growers such as Bill Berryhill (Berryhill Family Vineyards, Lodi) are leaving 200 of 500 acres unsold and plan to rip out 50 acres, accepting losses or discarding grapes. Allied Grape Growers reports that uncontracted grapes must often be sold at a loss for concentrated syrup. The market contraction is linked to aging baby‑boomer consumers, reduced alcohol consumption among younger adults, competition from craft beer, spirits, canned cocktails and cannabis, and lower global wine demand. Export options are limited by higher U.S. production costs and declining worldwide consumption. Consequently, vineyard closures are affecting farm labor and local economies.

Gergely Orosz posted on X that the Firebase iOS SDK abruptly began crashing every app session that relied on it, leaving developers unable to intervene. He notes that the issue affected 100 % of sessions for impacted apps, generating roughly 110 k crash reports, and that the Firebase team was unaware of the problem when it occurred. Orosz critiques the SDK’s reliability, especially given Google’s reputed engineering standards, and characterizes the failure as “amateur.” The tweet tags other tech figures and publications (e.g., The Pragmatic Engineer, Vijay Tholpadi) but provides no further technical details beyond the widespread crash impact.

Tank Body Problem – a brief specification listing player stats and control mappings. Two participants, P1 and P2, each start with an Elo rating of 1500 and zero experience points (XP0). The input schema includes: “Elevate” (▲▲) as the primary upward command; “Power” (◐→) to charge; horizontal movement via the A/D keys; weapon selection through J/L; firing with SPACE; a second press of SPACE activates a “Laser/Plumage” mode. Additional game mechanics specify that cheese objects bounce three times, poison releases bubble projectiles, and the environment simulates lunar gravity, affecting object trajectories. This concise set of parameters defines the initial conditions and control layout for the referenced tank‑body scenario.