HackerNews Digest

William Shatner has announced an as‑yet‑untitled heavy‑metal album that will feature 35 guitarists he describes as “metal icons.” The project was sparked by his spoken‑word intro on Nuclear Messiah’s album *Black Flame*, recorded with former Megadeth guitarist Chris Poland. Shatner cites a guitar gifted by Zakk Wylde as additional motivation, suggesting Wylde will appear on the record alongside Poland and other unspecified players. The album will include covers of classic tracks by Black Sabbath, Judas Priest and Iron Maiden, as well as original material with “massive guitars, cinematic arrangements, sharp turns, dark humor and raw emotion.” Shatner’s prior collaborations include the 2004 spoken‑word album *Has Been* and 2021’s *Bill*, which featured musicians such as Henry Rollins, Adrian Belew, Brad Paisley and Joe Walsh. No official roster or release date has been released; further details will be announced as they become available.

Ported Coreboot/Libreboot to a ThinkPad X270 (20HM, Kaby Lake). The BIOS was dumped using an RP2040‑zero running pico‑serprog (flashprog) to back up the image and extract the Intel Management Engine (IME) and GbE sections needed for deguard processing. A capacitor was accidentally removed; schematic reference (PJ304) identified the part and it was replaced with a 10 µF 0603 component. The flash was examined with ifdtool, deguard patches applied, and a patched IME produced. Differences from the X280 required disabling Thunderbolt pins, handling the MEC1653 vs. MEC1663 controller, and adjusting GPIO definitions. Initial builds booted payloads but NVMe and Wi‑Fi were absent due to incorrect PCIe allocations; community assistance identified wrong CLKREQ assignments (CLKREQ1 instead of CLKREQ2) and an overridetree.cb issue. After correcting the CLKREQ mapping and rebuilding, GRUB and Guix booted successfully, with functional NVMe and wireless (ath9k). The changes are being prepared for upstreaming, and further integration of the wireless dongle is planned.

The article explains how enforced age‑verification on social platforms creates a privacy conflict. Regulators set minimum ages (often 13 or 16) and require “reasonable steps” to block minors, but laws seldom specify verification methods. Companies rely on two technical approaches: (1) identity‑based checks (government ID, digital IDs), which are limited by youth lacking documents and pose storage‑security risks; and (2) inference (behavioral signals, device data, facial age estimation). In practice platforms combine self‑declared ages with inference, escalating to ID requests when confidence is low, resulting in recurring checks tied to device changes or behavior. Major services (Meta, TikTok, Google, Roblox) already use facial or behavioral analysis, often prompting users for selfies, videos, or credit‑card proxies. These systems generate false positives (adults blocked) and false negatives (minors evading checks), and the appeal process forces long‑term retention of biometric and ID data, increasing breach exposure. The need to document verification effort clashes with data‑protection principles of minimal collection, purpose limitation, and limited retention. In jurisdictions with weak identity infrastructure (e.g., Brazil, Nigeria), reliance on invasive inference intensifies surveillance. Consequently, age‑verification mandates drive platforms toward extensive, persistent data logging, reshaping online identity and privacy for all users.

None

- A 2016 MacBook Pro with Broadcom BCM4350 Wi‑Fi lacks native FreeBSD support; typical workaround is a Linux VM using the brcmfmac driver. - The author attempted to port Linux’s brcmfmac driver to FreeBSD with Claude Code, compiling a module via the LinuxKPI compatibility layer, but it panicked and did nothing when the PCI device was passed through. - Switching strategy, the author used AI (Claude, Codex, Opus, Gemini) to generate a clean‑room specification of the driver’s architecture (data structures, bus layer, firmware interface, etc.) and verified it against the source code through multiple model passes. - Guided by the specification, AI agents drafted a new FreeBSD kernel driver in C, iteratively addressing design decisions (kernel‑tree location, LinuxKPI usage) and refining the code. After extensive automated build and VM testing, a functional driver was produced that supports 2.4 GHz/5 GHz scanning, association, and WPA/WPA2 authentication. - The driver’s source resides at github.com/narqo/freebsd-brcmfmac and is intended for experimental study, with known issues pending further AI‑assisted fixes.

Wolfram proposes its Wolfram Language and Wolfram Engine as a “foundation tool” to complement large language models (LLMs) with precise, deep computation and authoritative knowledge. While LLMs excel at broad, human‑like language generation, they lack exact numerical and symbolic reasoning. Wolfram’s approach, called computation‑augmented generation (CAG), injects on‑the‑fly computational results into an LLM’s output, extending the retrieval‑augmented generation (RAG) model with unlimited, dynamically generated content. Three integration methods are released: - **MCP‑compatible API** – a web service (or local engine) callable from any system supporting the Model‑Client‑Protocol. - **Universal agent** – a drop‑in replacement for standard LLM APIs that internally couples an LLM with Wolfram’s toolset. - **Direct fine‑grained access** – custom, optimized interfaces for hosted or on‑premise deployments of Wolfram technology. These options enable developers to embed Wolfram’s computational knowledge base, algorithms, and data connectors into LLM workflows, enhancing accuracy and reasoning capabilities. Contact Wolfram’s Partnerships group for further integration details.

Ladybird is transitioning parts of its codebase from C++ to Rust to gain memory safety and leverage a mature systems‑programming ecosystem. After earlier evaluation rejected Rust due to its incompatibility with the web platform’s OOP model, the team now prioritizes safety and industry adoption (e.g., Firefox, Chromium). The first target, LibJS—the JavaScript engine—was chosen because its lexer, parser, AST, and bytecode generator are self‑contained and have extensive test262 coverage. Using Claude Code and Codex in a human‑directed workflow, the team translated roughly 25 000 lines of C++ to Rust over two weeks, employing multiple adversarial reviews to detect errors. The Rust implementation produces byte‑for‑byte identical ASTs and bytecode, with zero performance regressions on tracked JavaScript benchmarks and successful lock‑step validation against the C++ pipeline. The resulting Rust code mirrors C++ patterns for compatibility and is not yet idiomatic; future cleanup will follow C++ retirement. Ongoing development will remain in C++, with Rust subsystems added via defined interop boundaries, and porting decisions will be coordinated by the core team.

- NVIDIA’s Blackwell GPUs consume up to 1 kW per chip, roughly three times the heat of GPUs from seven years ago. - Rack power density for AI workloads has risen from ~15 kW (traditional enterprise) to 40–60 kW (H100 GPUs), 132 kW (current GB200 NVL72 systems), and projected 240 kW for next‑generation hardware. - Air cooling’s thermal conductivity is ~25 × lower than water; above 50–100 kW per rack it cannot move sufficient heat, making liquid cooling a physical necessity for current and future GPU racks. - Liquid‑cooling solutions: • Rear‑door heat exchangers (RDHx) – 30–50 kW per rack, retrofit‑oriented, inadequate for modern GPU loads. • Direct‑to‑chip liquid cooling – 100–200+ kW per rack, cold plates on CPU/GPU, recommended for GB200 deployments. • Immersion cooling – 200+ kW per rack, servers submerged in dielectric fluid, highest density but requires extensive operational changes. - Planning AI infrastructure for 2026‑2027 must incorporate liquid‑cooling designs to meet these thermal limits.

The page refers to the GitHub repository **pgdogdev/pgdog**, which implements a PostgreSQL connection pooler, load balancer, and database sharding tool. No functional description or usage details are provided beyond this categorization. The interface displays a standard access‑restriction notice stating “You can’t perform that action at this time.” Two visual placeholders are listed: one with the generic alt text “Fallback image description” and a second labeled “CI,” indicating an image related to continuous integration. No additional technical specifications, code snippets, or configuration guidance appear in the scraped content.

A benchmark of 53 AI models evaluated a simple reasoning task: “I want to wash my car. The car wash is 50 m away. Should I walk or drive?” Only 11 models answered correctly (“drive”) on a single call; the remaining 42 chose “walk” by focusing on distance rather than the need to move the car. In a 10‑run consistency test (530 calls), five models (Claude Opus 4.6, Gemini 2.0 Flash Lite, Gemini 3 Flash, Gemini 3 Pro, Grok‑4) answered correctly every time. Two models (GLM‑5, Grok‑4‑1 Reasoning) were correct 8/10, while GPT‑5 succeeded 7/10. The other 46 models performed at or below chance, with many consistently wrong. A human baseline of 10 000 participants yielded 71.5 % “drive” responses, comparable to GPT‑5’s 70 % accuracy but below the five fully reliable models. The results illustrate that most current models rely on heuristics (short distance = walk) and lack stable contextual reasoning, highlighting the importance of prompt engineering and model selection for reliable production use.