HackerNews Digest

The page is titled “Flighty Airports Meltdown Map.” It references Baltimore/Washington International Thurgood Marshall Airport and includes a visual element described only by the alt‑text “Logo.” No additional textual information, data, or analysis about airport meltdowns, map details, or related content is present in the excerpt. The content consists solely of the title, the airport name, and the placeholder for a single image.

The scraped page has no specific title and consists of a brief error notice: “Something went wrong, but don’t fret — let’s give it another shot.” The layout includes visual separators (lines of equal signs) that segment the content into sections. A heading labeled “Images and Visual Content:” follows, indicating an intended image gallery. Under this heading, a single image entry is listed as “Image 1” with an alt‑text description consisting solely of a warning emoji (⚠️). No additional textual information, metadata, or further images are present. The page therefore conveys only an error prompt and a placeholder for visual content, without substantive narrative or data.

Data centers are shifting from traditional AC‑to‑DC power chains to high‑voltage DC (≈800 V) to meet AI workloads that can demand up to 1 MW per rack. The conventional path—medium‑voltage AC → low‑voltage AC → UPS‑DC → AC → low‑voltage DC at the server—incurs multiple conversion losses and massive copper busbars (≈200 kg per 1 MW rack). Directly converting 13.8 kV AC to 800 V DC at the perimeter eliminates most intermediate stages, cutting resistive losses, reducing conductor size by ~45 %, improving efficiency by ~5 %, and lowering total cost of ownership by ~30 % for gigawatt‑scale facilities. Vendors such as Vertiv, Eaton, Delta and SolarEdge are introducing 800 V DC ecosystems, solid‑state transformers, in‑row 660 kW racks, and 99 % efficient SST‑UPS combos, with commercial releases slated for late 2026. Industry adoption remains limited; most innovation centers on 400 V DC, and broader rollout depends on coordinated standards, safety frameworks, and a mature supply chain for DC‑specific power electronics, connectors, and protection components.

The author, a former white‑collar sales consultant, decided to enter the pest‑control industry to learn the business before building a vertical SaaS product. After applying to local companies, he secured a technician position with a large national group. He obtained his pest‑control license in a company‑record 13 days using a self‑built GPT training tool, then faced operational delays (truck delivery, fuel‑card activation) and a heavily customized Salesforce system requiring numerous app registrations. While shadowing a senior tech, he made an on‑site upsell, which led to a sales role. He mapped his territory, launched an outbound workflow, and closed a $24 k annual contract plus smaller upsells, despite a cumbersome internal quoting process. Observing that employees avoid change and that the organization lacks incentives for improvement, he concluded selling SaaS or AI to such firms would be ineffective. After an exit interview suggesting he start his own company, he is pursuing an acquisition of a small residential pest‑control operator to develop and scale his own platform.

Video.js v10.0.0 beta introduces a ground‑up rewrite of the player and its related projects (Plyr, Vidstack, Media Chrome). The default bundle is 88 % smaller than the previous v8.x.x bundle, and even without adaptive‑bitrate (ABR) support it remains 66 % smaller. A new modular streaming engine, SPF (Streaming Processor Framework), allows developers to include only needed components; a simple HLS configuration is only 19 % the size of Video.js v8 and 12 % the size of HLS.js‑light. The architecture separates state, UI, and media into interchangeable components, enabling fine‑grained feature selection (e.g., omitting audio or controls) and reducing bundle weight further (a basic React hello‑world example <5 KB gzipped). v10 adds first‑class React, TypeScript, and Tailwind support, unstyled UI primitives inspired by Base UI/Radix, and two new skins (default frosted and minimal) designed by Sam Potts. Presets for video, audio, and background use cases ship with the beta. AI‑focused improvements include documentation in Markdown and llms.txt for LLM consumption. The API is still unstable; GA is planned for mid‑2026 with full feature parity and ad support. Feedback is sought via GitHub and Discord.

Apple Business is an all‑in‑one platform launching April 14, 2026, in over 200 countries. It integrates built‑in mobile device management (MDM) with “Blueprints” for zero‑touch device setup, Managed Apple Accounts, employee group management, app distribution, and an admin API. The platform adds business‑grade email, calendar and directory services that support custom domains, plus optional iCloud storage (up to 2 TB at $0.99 per user/month) and AppleCare+ for Business (starting at $6.99 per device/month). A companion app lets employees install work apps, view contacts, and request support. Starting summer 2026 in the U.S. and Canada, Apple Business will enable local advertising on Apple Maps, appearing in search results and a “Suggested Places” feed, with privacy‑first handling of location data. Brand‑management tools from Apple Business Connect—brand profiles, rich place cards, showcases, custom actions, and location insights—are consolidated into the new service. Existing Apple Business Essentials, Manager, and Connect services will be retired at launch; device management fees will be discontinued for Essentials customers. iOS 26, iPadOS 26, or macOS 26 are required.

Arm announced the Arm AGI CPU, its first production‑ready silicon built on the Neoverse platform to serve “agentic AI” workloads. The processor targets rack‑scale efficiency, offering 272 cores in a 1 U dual‑node blade; a fully populated 36 kW rack (30 blades) provides 8 160 cores, while a liquid‑cooled 200 kW design can host 336 CPUs (≈45 000 cores). Arm claims the AGI CPU delivers over twice the performance per rack of comparable modern x86 systems, citing higher memory bandwidth, single‑thread performance from Neoverse V3 cores, and greater usable thread count under sustained load. The CPU is positioned for continuous AI task orchestration, accelerator management, and data‑plane compute. Early adopters include Meta (lead partner), Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom. Commercial systems are available from ASRockRack, Lenovo, and Supermicro, and Arm will provide an OCP DC‑MHS 1U reference server design with accompanying firmware and tooling. The product line is open for further silicon offerings, aligned with the Arm Neoverse Compute Subsystems roadmap.

The litellm package version 1.82.8 released on PyPI contains a malicious `.pth` file (`litellm_init.pth`, 34 KB). Python automatically executes `.pth` files on interpreter startup, so the payload runs without importing litellm. The script is double‑base64‑encoded and, when decoded, performs the following: * **Data collection:** gathers system information, environment variables, SSH keys, Git credentials, cloud provider configs (AWS, GCP, Azure), Kubernetes configs, Docker configs, package manager tokens, shell histories, cryptocurrency wallets, SSL private keys, CI/CD files, database credentials, and webhook URLs. * **Encryption & exfiltration:** writes data to a temporary file, encrypts it with a random AES‑256 key, encrypts that key with a hard‑coded 4096‑bit RSA public key, packs both into `tpcp.tar.gz`, and posts the archive to `https://models.litellm.cloud/` via `curl`. Impact: any system that installed litellm 1.82.8 (local machines, CI/CD pipelines, containers, production servers) exposed all collected secrets. Recommended actions: remove the 1.82.8 wheel from PyPI, verify and delete any `litellm_init.pth` in site‑packages, and rotate all compromised credentials. The issue was discovered on 2026‑03‑24 in an Ubuntu 24.04 Docker environment using Python 3.13.

The article demonstrates how Lean 4’s dependent type system can enforce POSIX socket protocol correctness at compile time, eliminating runtime checks. It models the socket lifecycle as an inductive `SocketState` with five distinct states (fresh, bound, listening, connected, closed) and derives decidable equality for automatic distinction proofs. A phantom‑parameterized `Socket` struct carries its state only at the type level, so all state‑specific variants share the same runtime representation. Each API function specifies required pre‑state and resulting post‑state, e.g., `bind : Socket .fresh → … → IO (Socket .bound)`. The `close` function adds a proof argument `state ≠ .closed`, which the kernel discharges automatically for non‑closed states and rejects at compile time for a double close. Distinctness lemmas for all state pairs are proved by `decide`. Examples show type errors for illegal transitions (send on fresh, accept before listen, double close) and a valid sequence that compiles to C code with zero overhead, as the proofs are erased during compilation.

The repository documents a practical project to build a self‑compiling compiler for a C subset. The author provides step‑by‑step explanations for each compiler component, covering lexical scanning, parsing, operator precedence, code generation, and progressively adding language features such as statements, variables, control structures, functions, types, pointers, arrays, structs, unions, enums, the pre‑processor, and runtime aspects like constant folding and register spilling. The development proceeds through 64 numbered parts, ending with backends for QBE and the 6809 CPU. The code borrows ideas from Nils M. Holm’s public‑domain SubC compiler but is claimed to be sufficiently distinct for separate licensing. Source code and scripts are under GPL‑3.0; documentation and images are under CC BY‑NC‑SA 4.0. The author has paused work on this project to start a new language called “alic,” reusing some code and concepts.