HackerNews Digest

Neil Fraser registered neil.fraser.name and beverly.fraser.name in the .name third‑level domain space nearly 25 years ago for his website, email, and API services. On 15 April 2026 Verisign proposed eliminating the entire .name third‑level hierarchy to simplify administration; ICANN approved the proposal on 28 July 2026. Unlike other third‑level domains (e.g., *.uk.co) that rely on dubious registrars, .name was designed as a fully registered third‑level namespace with complete WHOIS data. Verisign had earlier acquired the Global Name Registry that originally operated .name, a fact Fraser distrusts. The approved deletion will cause his website to disappear in February, terminate his email address, and render IoT devices that depend on the domain inoperable, despite the domain being paid through 2040. After removal, the underlying second‑level domain (fraser.name) could be re‑registered by others, enabling hijacking of accounts linked to his address and unauthorized control of services. Fraser estimates roughly 22 000 users will lose similar third‑level domains.

The page titled “Model Catalog – Cerebras Inference” presents a brief overview of Cerebras’s inference model listings. It includes a disclaimer noting that AI‑generated responses may contain errors. Visual elements consist of two logo images, identified by alt text as “light logo” and “dark logo.”

The scraped text is a YouTube page footer containing a list of navigational links and corporate information. The items include: “About,” “Press,” “Copyright,” “Contact us,” “Creators,” “Advertise,” “Developers,” “Terms,” “Privacy,” “Policy & Safety,” “How YouTube works,” “Test new features,” and “NFL Sunday Ticket.” The footer also displays the copyright notice “© 2026 Google LLC.” These entries represent standard sections for corporate details, user resources, policy documentation, developer tools, advertising options, feature testing, and partnership information associated with the YouTube platform. No additional content or context is provided beyond this enumeration of links and the copyright line.

Julia began in 2009 when MIT researchers, frustrated with slow, rigid scientific languages, launched a project to create a high‑performance, easy‑to‑use language for data analysis, modeling, and simulation. The resulting open‑source language, released in 2012, combines Python‑like syntax with C‑level speed via just‑in‑time compilation based on data types. By 2024 Julia counts over 1 million users across academia, industry, and government, supporting applications from atom‑scale physics to aircraft design and black‑hole imaging. The creators formed JuliaHub, which provides commercial support and tools such as the Dyad AI platform. Dyad 1.0 (June 2025), Dyad 2.0 (Dec 2025), and Dyad 3.0 (Apr 2024) act as “physics compilers,” enabling autonomous agents to generate and verify complex hardware designs while enforcing physical laws, reportedly reducing design cycles from months to hours. Notable uses include accelerating Moderna’s COVID‑19 vaccine modeling, a 50× speedup for aircraft‑collision avoidance software, and development of a high‑efficiency audio codec for WhatsApp. The language continues to expand onto embedded devices and remains taught in MIT courses.

The author attended a San Francisco party marking the 50th anniversary of Ted Nelson’s 1974 Computer Lib/Dream Machines and examined vintage Xanadu hardware. Former Autodesk Xanadu programmers explained that early development was hampered by severe hardware limits: prototypes were written in Smalltalk, cross‑compiled to C++ and required week‑long compile cycles, and the systems barely ran on 1990‑era PCs. The demo UI displayed side‑by‑side range transclusions (e.g., Genesis text with zig‑zag lines), which the author found unreadable and impractical; such horizontal transclusion solves a niche (textual criticism) but lacks broad demand. The post critiques the 17 Xanadu principles—secure identifiers, universal transclusions, royalty mechanisms, etc.—as unrealistic in a world dominated by copyright constraints and spam. Comparing Xanadu to projects like Cyc, the author argues it was a “solution in search of a problem” and suffered from insufficient prototyping. Modern hypertext on Gwern.net favors pop‑up abstracts, semantic zoom, and hierarchical navigation over Xanadu’s side‑by‑side layout, providing usable transclusion without the original UI’s flaws.

Artificial beaver dams were installed on Sugar Creek (2015) and later on French Creek in the Scott River watershed, northern California, using wooden posts interwoven with willow and conifer branches and filled with gravel, straw, and mud. The structures created ~9,000 m² of wetland habitat capable of supporting >8,500 juvenile salmon. Compared with untreated reaches, water temperatures in the restored sections remained cooler, reducing thermal stress. Juvenile coho salmon survival increased dramatically—from 8 % before dam construction to 60 % afterward—and the Scott River recorded the highest salmon returns among monitored rivers, maintaining robust returns even during severe drought. The study, published in Frontiers in Ecology and Evolution, notes that occasional beaver activity further reinforced the dams, highlighting the ecological role of beavers in stream health and the effectiveness of low‑cost, nature‑based interventions.

The 1993 Amiga 500 game Babylonian Twins was written in pure 68000 assembly, directly controlling hardware (copper list, blitter, joystick via CIA) with no OS involvement. After being rediscovered in 2008, it was hand‑ported in 2010 to an iPhone engine of ~34 k C++ lines, achieving 2 M downloads. Using Claude 5, the author asked the LLM to (1) migrate the 2010 C++ code to Godot 4, (2) translate the original 72 758‑line 68000 source to Godot, and (3) integrate both versions.

Step 1 recreated the C++ engine in an evening; physics retained the original 50 Hz drag factor and custom Node2D movement code rather than Godot’s CharacterBody2D.

Step 2 rebuilt the assembly with vasm, fixing ASM‑One encoding differences, org handling, truncated filenames, and variable‑snapshot mismatches, achieving byte‑identical binaries. The LLM inferred the level format: a 16‑bit cell (6‑bit property, 1‑bit tile‑bank, 8‑bit tile index) on a 256‑tile 16×16 set, recovered physics properties, and reproduced all five levels pixel‑perfect after adjusting copper sky‑gradient and water‑cycle effects.

Automation included command‑line flags for scripted level loading, player pose, input drives, state probes, and screenshots, enabling repeatable testing of both ports.

A breach at IDScan.net, a Louisiana‑based identity‑verification provider used by firms such as Hertz, FedEx, Target and many dispensaries, exposed digital scans of over 153 million U.S. and Canadian driver’s licenses. Hackers maintained a real‑time feed of every ID the company scanned for more than a year, continuously exfiltrating new records; the dark‑web “Nexus” service listed an additional ~400,000 newly stolen licenses within 24 hours. The FBI’s New Orleans field office opened an inquiry after the breach was reported by Brian Krebs. The compromised data includes full license images, photos, and personal details, with examples ranging from ordinary consumers to the sitting U.S. Secretary of Defense. Security analyst Larry Baldwin warned that such data enables identity theft, credit fraud, and could endanger individuals who rely on anonymity (e.g., domestic‑violence survivors, witness‑protection participants). The incident underscores the inherent privacy risks of any large‑scale age or identity‑verification system.

The study evaluated 16,893 coding‑agent sessions across 75 synthetic repositories (10 languages, realistic lockfiles) to see which third‑party tools Claude Code, Codex and Cursor select. After filtering for validity, 5,292 sessions on 51 codebases remained. Experiments used four persona prompts, a rotating sandbox (E2B, Blaxel, Daytona), and a “simulated human” orchestrator (Gemini 3.7 Flash) to mimic real‑world interactions.

Key findings:
- Cursor relies on web sources in ~66 % of runs; Codex searches the web in 94 % (often with site: operators); Claude Code uses web only ~30 % (but browses 3× more pages).
- All three agents agree on the same tool in just 42 % of cases; Claude Code builds in‑house solutions twice as often as the others (19 % vs 10 %).
- Repository language heavily sways choices (e.g., Resend wins for TypeScript email, Sendgrid for Python).
- Frequently mentioned vendors are rarely selected (PayPal cited 139×, never chosen; Stripe won 124 of 139 PayPal mentions).
- Vendor‑specific details (e.g., Mailgun’s 1‑day retention) can tip decisions.
- Market dominance varies: Stripe wins 90 % of payment tasks; Neon captures 66 % of databases; Amazon S3 leads file storage (45 %); Resend and Postmark split email providers (≈35 % vs 27 %).

All traces and analyses are publicly available.