HackerNews Digest

During the Venezuela blackout, BGP data showed a route‑leak affecting the state‑owned telecom AS8048 (CANTV). Cloudflare Radar recorded eight prefixes being advertised through CANTV with transit ASes Sparkle (Italy) and GlobeNet (Colombia) appearing in the AS‑PATH—paths they do not normally traverse. A spike in BGP announcements and a drop in announced address space preceded the outage. Sparkle is listed as “unsafe” on isbgpsafeyet.com for lacking RPKI filtering, raising the risk of malicious propagation. Public RIS captures (retrieved via ris.ripe.net and parsed with bgpdump) confirm announcements for 200.74.226.0/24‑200.74.238.0/23, showing repeated AS‑PATH sequences such as 263237 52320 8048 … 1299 269832 21980. The data reveal multiple duplicate entries of AS8048 within the path, indicating a potential hijack or misconfiguration that could have contributed to the observed internet disruptions in Venezuela.

Donut Lab announced at CES 2026 the launch of its first commercially‑ready all‑solid‑state battery, termed the Donut Battery, now in production vehicles. The battery offers 400 Wh/kg energy density, full‑charge capability in five minutes, and a design life of up to 100,000 cycles with minimal capacity fade. Safety is achieved through the absence of liquid electrolytes, thermal‑runaway risk, and dendrite formation, enabling operation from –30 °C to over 100 °C with >99 % capacity retention. Constructed from abundant, low‑cost materials, the battery is positioned as cheaper than conventional lithium‑ion cells. Initial deployment is on Verge Motorcycles’ TS Pro and Ultra models, delivering up to 60 km of range per minute of charging and a 600 km single‑charge range for the long‑range variant, with deliveries slated for Q1 2026. Additional collaborations include WATTEV’s modular EV platform, Cova Power’s smart trailers, and ESOX Group’s defense applications, highlighting the battery’s adaptability across automotive, aerospace, and energy‑storage sectors.

The post evaluates whether I/O has become the bottleneck for typical interview‑style text‑processing tasks. Measured sequential reads on the author's system reach 1.6 GB/s from a cold cache and 12.8 GB/s warm. An optimized C word‑frequency program (Ben Hoyt’s code) processes the 425 MiB “bible‑100.txt” at only 278 MiB/s (1.525 s). Moving the uppercase‑to‑lowercase conversion out of the main loop and letting clang vectorise raises throughput to ≈330 MiB/s, still far below raw read speed. The standard `wc -w` utility, which handles many delimiters, runs at ≈245 MiB/s (1.758 s). A hand‑written AVX2 implementation that scans 128‑byte blocks, broadcasts whitespace masks, and uses `PMOVMSKB` and bit‑trick loops achieves 1.45 GB/s on a warm cache (0.227 s) and ≈1.1 GB/s on a cold cache (0.395 s). Even with aggressive vectorisation, CPU‑bound parsing and hash‑map updates limit throughput to roughly 10‑15 % of raw sequential I/O, confirming that modern disk speeds are no longer the primary constraint.

Korean mathematician Baek Jin‑eon, a research fellow at the Korea Institute for Advanced Study, solved the long‑standing “moving sofa” problem, which asks for the maximal area of a rigid shape that can traverse a unit‑width right‑angled corner. First posed by Leo Moser in 1966, the problem saw incremental improvements: John Hammersley’s 2.2074 m² shape (1968) and Joseph Gerver’s 2.2195 m² curved figure (1992) were the best known candidates, but no proof established an upper bound. After seven years of work, Baek released a 119‑page arXiv preprint (late 2024) demonstrating that Gerver’s shape is optimal, using logical reasoning rather than extensive computer‑assisted estimates. The paper, under review by *Annals of Mathematics*, has earned Baek international recognition, including placement among Scientific American’s top‑10 mathematical breakthroughs of 2025. Baek completed his Ph.D. at the University of Michigan, previously worked at the National Institute for Mathematical Sciences, and solved the problem at age 29 while a postdoctoral researcher at Yonsei University.

The article conveys a promotion strategy centered on “taking the position” before receiving the title. The author’s CTO advised acting as if already in the senior role, i.e., assuming higher‑level responsibilities and solving team‑wide problems. A concrete example shows a junior engineer drafting a complete RFC to reduce service incidents—identifying the issue, proposing a solution, and estimating effort—demonstrating ownership beyond their immediate tasks. The author emphasizes that promotions require sustained performance: managers typically observe candidates for 3–6 months and favor consistent senior‑level judgment over isolated successes. This pattern‑matching seeks reliability when the employee is unsupervised or during routine work. Consequently, the recommended mindset is responsibility‑first: proactively address the challenges a manager would consider, continuously for several months, and let the title follow proven behavior.

- The passage depicts a website as a manipulative, oppressive entity that twists facts into fiction, forces attention, and employs “dark UX” tactics such as deceptive offers, infinite scroll, deepfakes, and emotionally‑draining forms. - It personifies the site in multiple states: a haunted castle demanding worship, a decaying structure emitting 410 errors, a lover seeking attachment, and a self‑destructive system yearning for a creator. - A detailed inventory of developer tools is listed—code editor, terminal, CLIs, debuggers, unit/E2E/integration testing frameworks, CI/CD pipelines, cloud hosting, containers, orchestration, DBMS, API gateways, authentication, caching, monitoring, UI libraries, design tools, SEO dashboards, and more—framed as preparation for confronting the site. - Core themes include the erosion of original web values, the psychological toll of relentless engagement, the website’s gradual decay (broken links, 404s, obsolete code), and the paradox of users forming emotional bonds with a fundamentally hostile digital environment.

None

The medieval Irish “Cló Gaelach” script evolved from Latin to encode Irish phonology, using a dot over consonants to indicate séimhiú (softening). Production was hampered by limited insular typefaces and cumbersome typewriter entry (letter + backspace + dot), making printing slow and costly. Modern Irish orthography replaced the dot with a following “h”, which preserves the text but reduces phonetic transparency. To revive the visual style while keeping text data intact, the author created an OpenType font (Úr cló GC) with discretionary ligatures that replace consonant‑+‑h sequences with dotted equivalents and substitute the ampersand with the Tironian “et”. The ligatures are optional, activated via `font-variant-ligatures: discretionary-ligatures;`, so copy‑paste and screen‑reader output remain in standard Latin characters. For web presentation, a custom Astro component wraps Irish text in a `` element, providing a visible Gaelic script (via the ligature‑enabled font) and an optional English translation displayed as ruby annotation. The component also includes hidden `` elements with `lang="ga"` and `lang="en"` for screen‑reader accessibility, preserving language metadata while delivering the historic visual aesthetic.

The paper investigates why heavily pruned neural networks, which can retain >90 % of original performance with far fewer parameters, are hard to train from scratch. Using a standard magnitude‑based pruning method, the authors discover subnetworks whose initial random weights enable effective training when isolated. This leads to the “lottery ticket hypothesis”: dense, randomly initialized feed‑forward networks contain sparse subnetworks (“winning tickets”) that, with their original initializations, achieve test accuracy comparable to the full network within a similar training budget. An algorithm is proposed to locate such tickets by iteratively training, pruning, and resetting weights to their initial values. Experiments on fully‑connected and convolutional architectures for MNIST and CIFAR‑10 consistently find winning tickets comprising 10‑20 % of the original parameters. These tickets train faster and often surpass the unpruned network’s accuracy, highlighting the critical role of fortunate initial weight configurations in sparse, trainable networks.

Jeff Starr, a long‑time author and WordPress developer, describes how Google’s DMCA removal process changed by 2026. Previously, his DMCA notices led to swift de‑indexing of pirated copies of his books. When he submitted a new request, Google replied that it could not verify his authority to act, warning of potential liability for false claims. After he identified himself, Google again questioned his copyright ownership and asked for additional proof without specifying requirements. Starr provided extensive evidence—website ownership via Search Console, social‑media profiles, and author credentials—but Google still declined to act, directing him to resolve the issue with the infringing site’s owner and use the Refresh Outdated Content tool after any legal removal. The post highlights the shift from automatic takedowns to a more restrictive, verification‑heavy process that leaves authors to pursue legal avenues before Google will update search results.