HackerNews Digest

None

TurboQuant quantizes high‑dimensional vectors by first applying a random orthogonal rotation, which spreads any coordinate‑wise magnitude spikes across all dimensions. After rotation, each coordinate follows a fixed Beta‑derived density that converges to a Gaussian 𝒩(0, 1/d) as dimension d grows, making the coordinates nearly independent and identically distributed. Because this distribution is known, a universal scalar codebook can be built once using the Lloyd–Max algorithm: it chooses 2ᵇ optimal reconstruction levels that minimize mean‑squared error for the density. Quantization then consists of a simple nearest‑level lookup for each rotated coordinate, requiring no per‑block scale or zero‑point metadata. This avoids the metadata overhead of production quantizers (which store float16 scale/zero for each block), achieving the same effective bits‑per‑value as a naïve fixed‑grid scheme while delivering comparable reconstruction quality. The method yields unbiased inner‑product estimators and low MSE distortion for embeddings such as 1536‑dimensional OpenAI vectors.

The author built an automated screenshot system for the Jelly help centre. Documentation pages are written in Markdown, processed by Redcarpet into HTML and rendered as ERB views in Rails. Within the Markdown, HTML comments of the form `` serve as instructions for the screenshot generator. A Rake task launches headless Chrome via Capybara + Cuprite, scans all Markdown files for these comments, groups them by team to reuse logins, navigates to each URL, performs optional actions (click, wait), and captures screenshots in one of three modes: **element** (specific CSS selector), **full_page** (entire page with optional cropping), or **viewport** (visible area). Additional options include `click`, `wait`, `crop`, `torn` (CSS clip‑path edge effect), and `hide` (temporarily hide elements). The task outputs PNGs referenced by the image tags, then builds the help pages, placing source Markdown in `public/manual/` and generated ERB views in `app/views/help/`. Running the single build command updates all screenshots whenever the UI changes, eliminating manual capture and keeping documentation in sync with code.

The author outlines three pre‑building constraints to keep projects focused and viable. 1. **One‑page definition** – All ideas must be captured in a single, precise page that serves as the north‑star for investors, team members, and personal planning. If the concept cannot be expressed within one page, it is deemed overly complex and is not pursued. 2. **Separable core technology** – The product must rely on a distinct, reusable core tech (method, tool, library, or methodology) that can exist independently of the product itself. This core tech provides long‑term leverage and aligns with the creator’s broader vision; without it, the idea lacks sufficient impact. 3. **Defining constraint** – Each product needs a central, user‑visible constraint that gives it identity and limits feature creep (e.g., Minecraft’s blocks, IKEA’s flat‑pack design). This constraint should be evident in the one‑pager and guide all design decisions. If an idea fails any of these rules, the author does not build it.

The Prompt API enables Chrome extensions and web pages to send natural‑language requests to the on‑device Gemini Nano model. It runs on desktop Chrome (Windows 10/11, macOS 13+, Linux, ChromeOS + devices) with at least 22 GB free storage, >4 GB VRAM for GPU use (required for audio input) or 16 GB RAM and 4 CPU cores. The model is downloaded on first use; availability is checked via `LanguageModel.availability()`, and a session is created with `LanguageModel.create()`. Sessions can be customized with `topK`, `temperature`, `initialPrompts`, and `expectedInputs/expectedOutputs` (text, image, audio) plus language arrays (“en”, “ja”, “es”). Prompts are sent via `prompt()` (full response) or `promptStreaming()` (ReadableStream). Context is maintained until the token window is reached; overflow triggers removal of earliest non‑system messages or a `QuotaExceededError`. Additional features include `append()` for extra prompts, JSON‑Schema constraints via `responseConstraint`, session cloning (`clone()`), and termination (`destroy()`). Multimodal demos illustrate image critique and audio transcription. Access is limited to top‑level windows and same‑origin iframes, with cross‑origin permission via `allow="language-model"`. The API is not yet available in Web Workers.

EvanFlow is a Claude Code plugin that implements a test‑driven development (TDD) feedback loop for software projects. The system comprises 16 “skills” (e.g., brainstorming, plan writing, execution, vertical‑slice TDD, iteration) and two custom subagents (coder and overseer). An orchestrator skill (`evanflow-go`) runs a fixed sequence: restate idea → brainstorming (design checkpoint) → planning (file‑structure checkpoint) → execution (task‑by‑task, blockers surface) → TDD (red‑green cycle on a failing test) → up‑to‑five iterations with quality checks (typecheck, lint, UI screenshots, Five Failure Modes review) → stop and report. Hard rules embedded in every skill prevent hallucination and unsafe git actions: no auto‑commits, no invented identifiers, no forced specifications, and a pre‑tool hook blocks destructive git commands. The plugin can be installed via Claude Code’s marketplace (recommended) or manually with `npx skills@latest` or a git clone, the latter requiring manual hook registration. Optional dependencies include `jq`, a Chromium‑based headless browser for UI verification, and standard bash tooling. The architecture encourages parallel coder/overseer orchestration for independent units and enforces integration tests as the contract between components.

SentinelLABS identified a previously undocumented 2005 cyber‑sabotage framework dubbed **fast16**. The architecture consists of a Windows service wrapper **svcmgmt.exe** that embeds a Lua 5.0 virtual machine, encrypted bytecode, and a kernel driver **fast16.sys**. The service can install itself, propagate via SMB shares, and launch the driver; it also registers a DLL that logs network connections through a named pipe. **fast16.sys** is a boot‑start filesystem driver that intercepts reads of *.EXE* files compiled with the Intel C/C++ toolchain. Using a rule‑driven engine (≈101 patterns) it injects extra sections and modifies floating‑point code to introduce systematic calculation errors. Matches in period‑appropriate corpora point to high‑precision engineering and simulation tools such as LS‑DYNA 970, PKPM, and MOHID, suggesting strategic sabotage of physics, civil‑engineering, or nuclear‑research workloads. The driver name appears in the 2017 ShadowBrokers “Territorial Dispute” deconfliction list, linking the framework to early state‑level tooling and predating Stuxnet by at least five years.

- Apple introduced the MacBook Neo, a low‑price laptop (≈$500 for education) built around an older iPhone‑class SoC, reusing existing silicon and avoiding features such as Thunderbolt, high‑capacity RAM, and extensive I/O that are limited by the chip architecture. - The device emphasizes repairability: modular construction, minimal adhesives, and easy part replacement, aligning with institutional IT needs and the expectations of younger users. - Apple markets Neo with distinctive colors (blush, citrus) and positions price as its primary selling point, targeting first‑time Mac buyers and students. - In 1992 Porsche faced a tax‑driven price ceiling (£29 k) for its entry‑level 968. Rather than trimming the existing model, engineers stripped the car to a bare chassis, removed air‑conditioning, electric windows, and rear seats, then added lightweight bucket seats, a racing steering wheel, and bold color schemes, creating the 968 Club Sport. - Both companies used hard cost constraints to redesign products from the ground up, turning minimal‑feature offerings into desirable, “cool” options that expand their customer base and serve as entry points to higher‑margin models.

The author reduced a Rust program’s memory use from 895 MB to 420 MB by restructuring deserialized JSON models. The AWS SDK JSON files are parsed into nested “Smithy” structs, many of which contain only optional `String` fields. Because `Option` for non‑pointer types still occupies the full size of `T`, keeping such structs directly in their parents inflated memory. By detecting structs that are empty (all optionals are `None`) and changing their fields to `Option>`, the empty value is stored as a single word pointer, and the heap allocation is avoided unless data exists. Custom Serde deserializers drop empty structs during parsing. This required updating `SmithyReference` and `SmithyShape` to use boxed options and adjusting accessors. Profiling with jemalloc confirmed a 475 MB saving, at the cost of modest CPU overhead and potential heap fragmentation. The technique demonstrates that, in Rust, boxing optional large structs can dramatically cut composite‑struct memory footprints while remaining compatible with Serde.

DriveSavers, a data‑recovery firm based in Novato, California, handles roughly 20 000 inquiries monthly, serving government agencies, corporations, and high‑profile individuals. Founded in 1989 by Jay Hagan and Scott Gaidano after the collapse of Jasmine Technologies, the company evolved from hardware repair to specialized file recovery, recognizing that data’s value rises sharply when lost. Directors Sarah Farrell (business development, former engineer) and Mike Cobb (engineering) oversee operations that range from routine repairs to extreme cases—devices damaged by fire, snowblowers, monorails, or wildlife. Their “Museum of Bizarre Disasters” displays recovered media illustrating the limits of physical restoration. DriveSavers’ clientele includes Sidney Poitier, Khloé Kardashian, Willie Nelson, Buzz Aldrin, and former President Gerald Ford. Historical context notes that early hard drives (e.g., a 40 MB unit costing $20 000) were more valuable than stored data, but advances in storage density and digitization have inverted that relationship, driving demand for professional recovery services.