HackerNews Digest

transcribe.cpp is a C/C++‑based ASR inference library built on ggml that targets cross‑platform, locally accelerated speech‑to‑text. It supports over 60 models across 16 transcription families, each numerically validated and word‑error‑rate (WER) tested against reference implementations. Performance is accelerated via Vulkan, Metal, CUDA, and TinyBLAS, with benchmark data provided for Ryzen 4750U (CPU + Vulkan) and Apple M4 Max. The library offers streaming and batch transcription, and functions as a near drop‑in replacement for whisper.cpp, retaining compatibility with existing .bin model files. First‑party language bindings are available for Python, JavaScript/TypeScript, Rust, and Objective‑C/Swift, with additional contributions welcome. Designed for easy embedding in the Handy application, transcribe.cpp aims to simplify local ASR deployment on macOS, Windows, and Linux, enabling real‑time transcription on low‑power devices (e.g., RK3566). Development is supported by Mozilla AI, ggml contributors, Modal, Blacksmith, and Hugging Face, with ongoing maintenance and community involvement.

Moonshine Micro is an open‑source AI toolkit for real‑time voice agents on embedded processors, demonstrated on the Raspberry Pi RP2350 (≈ $0.80). It integrates voice‑activity detection (VAD), command recognition (STT), and neural text‑to‑speech (TTS) that together occupy ~470 KB of SRAM on the RP2350’s 520 KB memory, sharing a ~384 KB TensorFlow Lite‑Micro arena. The flash footprint consists of .text and .rodata sections measured with arm‑none‑eabi‑size; .bss, heap and stacks comprise the SRAM usage. The system processes VAD, STT, and TTS sequentially, so RAM is not additive. Compute is expressed in MACs (multiply‑accumulate), with active‑stage performance in MMAC/s. All code, including the SpellingCNN and TinyVadCNN models, is released under the permissive MIT License, suitable for commercial use. The VAD, STT, and TTS libraries can operate independently, relying on the bundled TensorFlow Lite Micro library for neural inference, and an end‑to‑end example demonstrates voice‑controlled Wi‑Fi setup on the RP2350 MCU.

Castor is a command‑line tool that extracts video streams from arbitrary web pages, transcodes them with ffmpeg, and casts the real‑time stream to any DLNA/UPnP MediaRenderer (e.g., Smart TVs, Kodi, VLC, Plex). It launches headless Chrome with randomized fingerprints and stealth scripts, monitors network traffic via the Chrome DevTools Protocol, solves simple Cloudflare Turnstiles, and selects the largest iframe to locate the video URL. Castor supports direct stream URLs, IMDB/TMDB IDs, and can burn auto‑generated subtitles using whisper.cpp. Installation options: - Native binary via Homebrew (`brew install --cask stupside/tap/castor`) requiring Chrome/Chromium, ffmpeg, ffprobe on PATH. - Docker image (`ghcr.io/stupside/castor`) that bundles dependencies; must run with `--network host` on a Linux host for LAN discovery. Configuration is a minimal `config.yaml` specifying the TV name and source proxies; optional TMDB API key enables an interactive TUI for browsing titles. Core commands include `castor scan`, `castor cast movie `, `castor cast url `, and debugging flags. The tool does not host content and relies on user‑provided sources.

The Amiga Freeware Archive provides over 10 GB of classic Amiga software—games, applications, demos, graphics, music, and tools—from the 32‑bit home‑computer era. It aggregates public‑domain (PD) libraries, scene collections, user‑group disks, and disk‑magazine archives. **Key PD libraries** - **17 Bit Software** (UK, 1987‑1997): a large catalog of freely distributable Amiga disks. - **Fred Fish Disks** (USA, 1986‑1994): widely trusted pre‑Internet software collection. - **Scope** (Texas): 175 numbered volumes. - **Slipped Disk**: ≥20 volumes of games, utilities, demos. - **USAA**: ≥23 volumes focused on utilities, source code, and occasional games. **Scene groups** - **The Assassins** (UK): prolific PD games compilations. - **LSD Compendium** demo group: three‑volume CD‑ROM series with thousands of demos, modules, graphics, and utilities. **User‑group series** include Miami A.M.I.G.O.S. (1990), Tampa Bay Amiga Group (1986‑), Amateur Radio User Group, New Zealand Amiga Users Group (1986‑1990, ≥29 disks), and Southern Nevada Amiga Group (1990, ≥14 disks). **Disk magazines** - **JAM** (UK, 1991‑1995): 58 issues.

- The traditional school algorithm multiplies two n‑digit numbers with O(n²) single‑digit multiplications; its work grows quadratically with digit length. - In 1960, Anatoly Karatsuba disproved the conjectured O(n²) lower bound by showing that three multiplications plus additions suffice for two halves of the numbers, yielding a recursive method with running time O(n¹·⁵⁸⁵). For example, multiplying two 1 000‑digit numbers drops from ~1 000 000 to < 57 000 elementary multiplications. - Modern libraries (e.g., Python’s big‑int implementation) switch from the grade‑school method to Karatsuba once operands exceed ≈630 decimal digits (≈70 base‑2³⁰ digits). - The quest for faster multiplication continued, culminating in 2019 with Harvey and van der Hoeven’s algorithm running in O(n log n) time, theoretically just above linear. Its advantage appears only for astronomically large inputs, classifying it as a “galactic algorithm.” - Theoretical computer scientists now regard O(n log n) as the likely optimal asymptotic bound, though a formal proof remains open.

The lecture introduces foundational concepts of multivariable calculus. It begins with vector fundamentals, covering definitions, the dot product, vector projections, determinants, and the cross product. Geometric applications follow, describing equations of planes and lines and their visualizations. Parametric representations of curves are presented, leading to differential tools such as partial derivatives, tangent planes, and normal vectors for surfaces. The chain rule and directional derivatives are examined, emphasizing the gradient’s role in optimization. Techniques for locating maxima and minima, including Lagrange multipliers for constrained problems, are outlined. The session concludes with double integration, detailing integration over rectangular and general regions and illustrating practical applications of double integrals.

The author describes how joining new social groups is most effective by organizing events that align with the group’s core activity. Observations include: (1) demand for social events consistently exceeds supply; (2) creating or co‑hosting events accelerates relationship building compared with merely attending others’ gatherings; (3) many participants treat communities as passive, consumer‑oriented resources, assuming social opportunities arise automatically; (4) in reality, events require deliberate legwork, and leadership emerges from those willing to invest that effort; (5) organizers recognize and value peers who share this burden. The piece argues that social alienation stems partly from widespread free‑riding, where people consume community fabric without contributing to its creation. While broader societal solutions are unclear, individuals can mitigate the issue locally by actively supplying events and supporting others who do the same.

None

Mayor Mamdani announced that landlords are prohibited from using AI‑generated images to market rental units without disclosure, emphasizing the need for transparent advertising to protect tenants from misleading visual representations. The statement underscores regulatory intent to curb deceptive practices in the housing market, requiring landlords to rely on authentic photographs or clearly label AI‑enhanced content. The accompanying page includes a collection of unrelated image captions, describing various visual elements such as a public speaking event with city officials, a split “Expectation vs. Reality” house comparison, lighting equipment, camera gear, smartphone branding, wildlife photography, and historical photographs. These alt‑text descriptions provide context for the images but do not contribute additional information about the mayor’s policy announcement.

The site presents a visual catalog of retro Macintosh software icons. Users can hover over each icon to view its filename, country of origin, and associated media name, with links to the original media hosted on Macintosh Garden. Finder application and document icons are highlighted in green; hovering over these reveals their file type and creator codes. The interface emphasizes quick identification of classic Mac files through iconography and metadata.