HackerNews Digest

The post describes how macOS “Tahoe” (a codename for a macOS version) introduced a large corner radius that reduces the effective click target for window resizing. The system expects a 19 × 19 px region near each window corner; with traditional sharp corners about 62 % of that region lies inside the window, but the rounded corners push roughly 75 % of the area outside the visible window frame. Consequently, users’ instinctive clicks inside the visual corner often miss the active zone, causing resize attempts to fail—a problem the author has not encountered in four decades of computing. The only reliable method on Tahoe is to place the cursor just outside the corner, a non‑intuitive gesture that increases error rates. No other functional changes or workarounds are discussed.

Self‑hosting became practical in 2026 due to AI‑driven CLI agents (e.g., Claude Code) that automate server setup. The author uses a Beelink Mini N150 mini PC (≈$379) with an 8 TB NVMe SSD, running Linux. Claude Code is installed locally and, via plain‑language commands, configures Docker, generates Docker‑Compose files, provisions services, applies Caddy reverse‑proxy, ensures data persistence, schedules updates, and configures boot‑time restarts. Deployed containers include: - **Vaultwarden** – lightweight Bitwarden‑compatible password manager. - **Immich** – self‑hosted photo library with mobile apps and local face‑recognition. - **ReadDeck** – Pocket replacement for saved articles. - **Uptime Kuma** – health monitoring with email alerts for service up/down events. Utility tools such as **Lazydocker** (terminal UI for Docker) and **Glances** (system metrics) provide on‑demand visibility. Resource usage remains modest: 13 containers consume ~4 GB RAM, ~6 % CPU on the mini PC. The workflow targets users comfortable with a terminal, who currently pay for SaaS but want direct control without becoming full‑time infrastructure experts.

A 13 KB single‑file Snake game runs natively on Windows, Linux, and in browsers. The author built three separate implementations—C with WinAPI for i686 Windows, C with X11 for x86_64 Linux, and JavaScript using HTML5 Canvas—each 3–5 KB compiled/minified. The Windows binary uses a custom PE header that embeds a shell script; a decompressing stub expands a compressed payload, and Windows app‑help suppresses an initial 0xc0000005 error after the first run. The Linux version wraps an LZMA‑compressed ELF64 binary in a shell dropper that extracts and executes the payload while skipping the wrapper. The browser version places harmless garbage before the HTML, then hides it via CSS. All three components are concatenated so each OS selects its appropriate segment, producing a final 13 312‑byte file. Gameplay follows classic Snake rules: arrow/WASD control, ESC to quit, R to reset, P to pause, space to start; food yields 10 points (20 for rare yellow fruit), a new level after ten foods randomizes a maze guaranteeing a path to food, and snake speed scales with length.

The repository icloud-photos-downloader/icloud_photos_downloader hosts a command‑line utility for downloading photos from Apple iCloud. The project’s README (or landing page) references three visual assets: “Quality Checks,” indicating automated testing or code‑quality verification; “Build and Package,” denoting the build pipeline that compiles and packages the tool for distribution; and “MIT License,” specifying that the software is released under the permissive MIT open‑source license. No additional functional details, usage instructions, or code excerpts are present in the provided text.

The analysis measured token‑usage of code solutions from the Rosetta Code dataset across 19 popular languages, using the Xenova GPT‑4 tokenizer. For each language, tasks with implementations in all selected languages were token‑counted and averaged. Results show a 2.6 × token‑efficiency gap between the most verbose language (C) and the most compact (Clojure). Dynamic languages generally require fewer tokens because they omit type declarations, though JavaScript is the most verbose among them. Functional languages such as Haskell and F# are surprisingly token‑efficient, rivaling the best dynamic languages due to strong type inference. APL, despite its terse syntax, scores poorly because its special glyphs split into multiple tokens. Conversely, the ASCII‑based array language J achieves the highest efficiency (≈70 tokens per task), roughly half the cost of Clojure (≈109 tokens). The study highlights that token efficiency can significantly affect LLM context‑window utilization, potentially extending development sessions for token‑lean languages like Haskell, F#, or J.

Jerome H. Powell, Chair of the Federal Reserve, announced that on Friday the Department of Justice issued grand‑jury subpoenas to the Fed, accompanied by a threat of criminal indictment linked to his June Senate Banking Committee testimony about a multi‑year renovation of historic Fed office buildings. Powell asserts the subpoenas are not about the renovation or congressional oversight but represent political pressure aimed at influencing the Fed’s interest‑rate decisions. He emphasizes that the Federal Reserve sets rates based on its economic assessment, not presidential preferences, and that no official is above the law. Powell affirms his commitment to the Fed’s dual mandate of price stability and maximum employment, noting his experience serving under four administrations and pledging to continue his duties with integrity. (Last update: January 11 2026)

The article proposes using FUSE (Filesystem in Userspace) to expose arbitrary data as a virtual filesystem for AI agents, replacing multiple specialized tools with a single Bash‑style interface. By mapping domain entities (e.g., emails, folders) to files and directories, agents can perform familiar operations (ls, cat, mv, ln) to browse, read, and organize information, while the FUSE layer translates these calls into database queries. The author demonstrates an email‑assistant example: emails are represented as “.eml” files within a workspace hierarchy, with virtual folders like Starred and Needs_Action implemented via symlinks. Core FUSE operations (lookup, open, read, readdir, getattr, mkdir, rename, unlink, readlink, symlink) are implemented in TypeScript using the fuse‑native library, enabling on‑demand data loading and synchronization with the backend. A sandboxed agent, powered by the Anthropic SDK, executes Bash commands to list, search, star, and reorganize emails, illustrating reduced tool complexity and improved long‑context handling. The post concludes with a link to the GitHub implementation and predicts higher‑level APIs that abstract away FUSE details for broader adoption.

The page presents a navigation list for YouTube, featuring links labeled: About, Press, Copyright, Contact us, Creators, Advertise, Developers, Terms, Privacy, Policy & Safety, How YouTube works, Test new features, and NFL Sunday Ticket. It ends with a copyright notice stating © 2026 Google LLC.

The author decommissioned a $100‑per‑month VPS (“moana”) and replaced it with a low‑cost Fedora CoreOS VM (“tinkerbell”) on Vultr, managed entirely via Terraform. An Ignition config (generated from Butane YAML) creates a user and SSH key for initial access; after first boot Ignition is never used again, enforcing immutability. Container workloads (a static‑site binary behind Nginx, Caddy for TLS) are run with Podman Quadlets, which translate simple `.container` and `.pod` files into systemd units. A pod is used to share the network namespace so containers can communicate via localhost. Deployment is declarative: updating the container image tag triggers Podman’s auto‑update timer (enabled by the `io.containers.autoupdate=registry` label), causing the VM to pull the new image and restart the service without manual SSH. The setup eliminates ongoing VM configuration, requiring only image publishing. Future work includes adding an observability stack (e.g., OpenTelemetry) to gain visibility into the black‑box VM.

The author argues that advanced LLMs now accomplish most programming tasks, making manual code writing largely unnecessary. Recent personal experiments with Claude Code demonstrated rapid development: adding UTF‑8 support to a library, fixing Redis test failures, creating a pure‑C BERT inference library, and reproducing Redis Streams changes within hours. This shift emphasizes problem formulation over code writing, positioning AI as a partner for design and implementation. The author sees AI as a democratizing force akin to open‑source software, enabling small teams to compete with large firms, but warns of potential centralization and job displacement. He advocates open‑source AI development, vigilance against concentration of power, and political support for displaced workers. The recommendation is to adopt AI tools deliberately, experiment continuously, and focus on higher‑level system design rather than low‑level coding.