HackerNews Digest

Inkling is a 975 B‑parameter Mixture‑of‑Experts transformer (41 B active) with a 1 M‑token context window, pretrained on 45 trillion multimodal tokens (text, images, audio, video). A smaller 12 B‑active “Inkling‑Small” variant is also released. The architecture follows DeepSeek‑V3 MoE (256 routed, 2 shared experts per layer, 6 active experts per token) and uses interleaved sliding‑window/global attention with relative positional embeddings and short convolutions. Training employed a hybrid Muon/Adam optimizer, weight‑decay tied to learning‑rate squared, and extensive RL (30 M rollouts) for reasoning, calibration, and safety. Inkling supports controllable thinking effort, achieving comparable performance to larger open‑weights models while using ~⅓ the token budget on benchmarks such as Terminal Bench, HLE, and IFBench. Multimodal abilities include speech transcription, audio reasoning, and image understanding via patch‑based hMLP embeddings. Safety testing shows strong refusal rates on harmful requests (≥98 % on StrongREJECT) and robustness against censorship. The model is open‑weights, available on Hugging Face, and can be fine‑tuned via the Tinker platform, with a Playground for interactive testing. Inkling‑Small matches the larger model on many tasks while offering lower latency and cost.

SQLite’s defaults hinder reliability and performance. Four main issues are highlighted: * **Foreign‑key constraints** are disabled (`PRAGMA foreign_keys = OFF`), so deletions can leave dangling references and ROWID reuse can cause silent data corruption. * **Dynamic typing** allows any value in any column; INTEGER affinity still accepts TEXT that can be converted, and there is no global pragma to enforce strict tables, forcing manual `STRICT` declarations. * **Concurrent writers** receive an immediate `SQLITE_BUSY` error; the library does not wait for a lock unless `PRAGMA busy_timeout` is set. * **Performance defaults** disable Write‑Ahead Logging and use the safest synchronous mode, resulting in slower writes. The author proposes an “edition” system (e.g., `PRAGMA edition = 2026`) that would bundle sensible defaults: `foreign_keys = ON`, `busy_timeout = 5000`, `journal_mode = WAL`, `synchronous = NORMAL`, and make strict tables the default. This approach mirrors Rust’s edition model, aiming to improve safety and speed while preserving backward compatibility.

G# is a .NET language designed to incorporate ergonomics from Go, Kotlin, and Swift. It introduces language constructs such as `packages` for modular organization, `func` for function declarations, and `data class` for concise immutable data structures. Nullable types are handled using an `if let` syntax, mirroring Kotlin’s approach to safe null checks. Structured concurrency is supported through a `scope` construct, enabling scoped asynchronous execution similar to Swift’s concurrency model. The language compiles directly to managed .NET assemblies, allowing G# code to run on the existing .NET runtime without additional tooling. Overall, G# aims to provide modern, expressive syntax while leveraging the performance and ecosystem of the .NET platform.

Grok Build is SpaceXAI’s terminal‑based AI coding agent, delivered as a full‑screen TUI that can analyze codebases, edit files, run shell commands, perform web searches, and manage long‑running tasks. It operates interactively, headlessly for CI/scripts, or via the Agent Client Protocol (ACP) for editor integration. The repository contains Rust source for the CLI/TUI and its runtime, synchronized from the SpaceXAI monorepo. Prebuilt binaries for macOS, Linux, and Windows are installed via curl or PowerShell scripts; the binary is named `xai-grok-pager` and distributed as `grok`. Build requirements include the pinned Rust toolchain (managed by `rust-toolchain.toml`), `protoc` for protobuf generation, and supported host OSes (macOS/Linux; Windows builds are best‑effort). The workspace uses a generated root `Cargo.toml`; individual crate `Cargo.toml` files should be edited. Documentation, user guide, keyboard shortcuts, slash commands, theming, plugins, hooks, headless mode, and sandboxing are available under `crates/codegen/xai-grok-pager/docs/user-guide/`. The code is licensed under Apache 2.0, with third‑party components retaining original licenses.

None

Bluesky PBC acquired the “ATPROTOCOL,” “AT Protocol,” and “atproto” trademarks to prevent third‑party claims and to protect developers from legal abuse. The acquisition is defensive; the mark will remain freely available for typical ecosystem use—e.g., referencing the protocol, labeling compatible apps, writing documentation, or naming open‑source tools—provided the usage is accurate and does not imply official endorsement. A license is required only when the trademark is used as a brand element, such as product names, company or service names, paid events, merchandise, domain names, certifications, or logo usage that could suggest ownership or endorsement. Licenses will be offered in good‑faith cases, with fees considered only for commercial branding. Bluesky intends to later transfer ownership to an independent governance body. The approach aligns with practices of other open‑source projects (e.g., Apache, Linux) to safeguard the protocol while supporting open development.

None

Async/await simplifies writing asynchronous code but masks structural complexity that harms production systems. It conflates asynchrony (waiting for I/O) with concurrency (handling multiple tasks), so CPU‑bound work inside an async function blocks the cooperative executor thread (e.g., Tokio, Node.js), causing latency spikes and under‑utilized hardware. Engineers must manually separate I/O and compute—using Tokio for I/O and a dedicated thread pool such as Rayon—then orchestrate message passing, which re‑introduces scheduling complexity the language feature intended to hide. Additionally, async runtimes allow unbounded task spawning; during load spikes this fills memory until the OS OOM‑kills the process. Work‑stealing schedulers, while aiming for fairness, degrade cache locality and throughput at scale. As an alternative, Project Tina proposes a thread‑per‑core, shared‑nothing model without async/await, futures, or promises. Work units (“isolates”) run synchronously on fixed OS threads, communicate via bounded mailboxes, and never migrate, providing deterministic scheduling, predictable load shedding, and support for deterministic simulation testing. Tina is open‑source on GitHub.

The author describes using large language models (LLMs) to configure MikroTik networking equipment. Key practical points include: prefer the RouterOS REST/JSON API over SSH for LLM interaction; disable insecure services (non‑secure API port, www, telnet, FTP); back up the full configuration before and after changes, ideally under version control; use CAPsMAN to simplify multi‑AP Wi‑Fi setups; validate generated configs by consulting multiple LLMs and reach consensus; record SSIDs, passwords, and DHCP reservations before migrations; maintain a tested recovery run‑book and verify backups; apply changes incrementally and test after each step to catch hallucinations; configure NTP on all devices; give descriptive names to routers, switches, APs, and switch ports for easier management; keep RouterOS firmware uniform across devices to avoid syntax mismatches. For addressing issues such as overlapping 192.168.88.x subnets, the author recommends L2 “MAC‑Telnet” (or WinBox) to access devices when IP connectivity fails, and provides a Homebrew formula and a small CLI wrapper to make the tool LLM‑friendly.

Metal‑organic frameworks (MOFs) are ultra‑highly porous crystalline materials with the largest known internal surface areas; a gram can theoretically cover a football field when spread flat. Originating from reticular chemistry pioneered by Prof. Omar Yaghi (UC Berkeley/Lawrence Berkeley Lab), MOFs—including related covalent organic frameworks (COFs) and zeolitic imidazolate frameworks (ZIFs)—are constructed by linking metal clusters with organic linkers, creating robust, tunable pore structures. Yaghi’s group designed a zirconium‑adipic‑acid MOF that adsorbs water from arid air, releasing ~1.3 L per pound of powder every 12 h using solar‑driven cycles; the material is reusable and scalable from kilograms to tonnes for community‑level water provision. Since their inception ~20 years ago, ≈20 000 distinct MOFs have been reported, finding applications in carbon capture at coal‑power plants, pesticide delivery, and biodegradable drug carriers for chemotherapy. Yaghi’s contributions have earned the Wolf Prize in Chemistry (2018), the Albert Einstein World Award of Science (2017), and the King Faisal International Prize (2015), and he is a candidate for a Nobel Prize.