HackerNews Digest

Anthropic’s Long‑Term Benefit Trust (LTBT) has added Dr. Ben Bernanke, former Federal Reserve Chair (2006‑2014) and 2022 Nobel‑winning economist, as a trustee. The LTBT is an independent body that monitors Anthropic’s adherence to its mission of responsible advanced‑AI development for humanity’s long‑term benefit. Trustees are selected for expertise across fields, hold no equity, receive only compensation for time, and advise Anthropic’s board on risk and societal impact matters. Bernanke’s experience steering the U.S. economy through the 2008 financial crisis and his research on banking‑related crises will inform the Trust’s economic analyses of AI’s macro‑economic effects. He joins existing trustees Neil Buddy Shah (chair), Richard Fontaine, and Mariano‑Florentino Cuéllar, whose backgrounds span global health, national security, law, policy, and economics. The LTBT retains authority to appoint board members and provides oversight to ensure Anthropic balances commercial success with public and social good.

Colibrì is a single‑file C engine (≈1,300 lines) that runs the 744‑billion‑parameter GLM‑5.2 mixture‑of‑experts (MoE) model on a consumer machine with ~25 GB RAM. The dense part (attention, shared experts, embeddings ≈17 B parameters) resides in RAM at int4 precision (~9.9 GB). The 21,504 routed experts (75 MoE layers × 256 experts + MTP head, ≈19 MB each) are stored on disk (~370 GB) and streamed on demand using a per‑layer LRU cache, optional hot‑store, and the OS page cache as a second‑level cache. Key features include: * MLA attention with compressed KV cache (576 floats/token) and LoRA‑style q/k/v. * DeepSeek‑style sigmoid router, shared expert, and first‑three dense layers. * Native multi‑token‑prediction (MTP) speculative decoding, int8 draft head, and lossless rejection sampling. * Integer‑dot kernels (Q8_0‑style int8 activations, AVX2 maddubs) delivering 1.4–2.5× speedup over float. * Async expert readahead, batch‑union MoE execution, and a byte‑level BPE tokenizer in C. * Automatic RAM budgeting from MemAvailable to avoid OOM, and an offline FP8→int4 converter that processes shards without requiring the full 756 GB checkpoint. Performance on a 12‑core, 25 GB RAM WSL2 system shows ~1 token/s with warm cache; cold tokens require ~11 GB of expert reads. The repository provides scripts for building, converting, chatting, and benchmarking, and the int4 model is hosted on Hugging Face.

The European Parliament approved the interim “Chat Control 1.0” regulation, allowing suspicion‑less mass scanning of private, unencrypted messages on US platforms (e.g., Instagram, Discord, Gmail) until 2028 or until a permanent law is adopted. Although 314 MEPs voted against it, the motion to reject failed because it did not reach the required 361‑vote absolute majority (276 in favor, 17 abstentions). An exemption for end‑to‑end encrypted services (e.g., WhatsApp) was retained, but European providers have never implemented the scans. Data cited include a 50 % drop in US abuse reports since 2022 due to encryption, 36 % of 2024 abuse reports stemming from private‑chat scans, 48 % of alerts being non‑criminal, and 40 % of investigations targeting minors. The EU Commission acknowledges no evidence that suspicion‑less scanning increases convictions or rescues children. Negotiations for a permanent “Chat Control 2.0” – proposing targeted detection orders, an EU Child Protection Centre, and security‑by‑design standards – will resume in September, while member states continue to favor the current voluntary approach.

The page presents a game titled “18 Words,” inviting users to attempt the “18 words challenge.” It includes a brief prompt asking if they can survive the challenge, thanks players, and offers a link for feedback and a reference to another game called “Zanagrams.”

None

Running Train, a solo‑developer simulation by Novatetsu Games, is in Early Access on Steam (price $18). The game features 42 routes across two fictional Japanese lines—Fukugawa (10 routes, ~12 min each) and Sankai Main (32 routes, 6–44 min). Players control speed, braking, and station stops, with scoring that rewards precision, especially when UI guides are disabled. Environments span ~40 km of track, including detailed towns, powerlines, traffic, Shinto temples, and weather variations (sunny, rain, snow, seasonal changes). A free‑camera mode lets users observe the scenery, revealing assets not visible from the driver’s view. Community feedback on Steam highlights high‑quality modeling, lighting, and support for the Zuiki MASCON peripheral. Planned updates aim to add passenger systems, conductor mode, and expand track length to 100 km, targeting completion by next year.

The author, an early Facebook engineer, describes the extreme demands of early startup life: 120‑hour weeks, constant on‑call duties for News Feed scripts, limited server and bandwidth resources, and personal health sacrifices. This scarcity fostered intense focus on a single core mission—building the product—while rejecting requests to support external nonprofits, reasoning that the company's comparative advantage lay in product development rather than direct philanthropy. Over time, as financial resources grew and employee pressure increased, the company began accepting charitable initiatives, illustrating how individually reasonable diversions can cumulatively erode focus, increase operational drag, and introduce hidden costs. The same pattern appears in product management: adding numerous small, niche features leads to a cluttered UI, larger teams, slower development, and customer backlash when features are later removed. The author concludes that maintaining disciplined prioritization and early dismissal of distractions is essential for sustainable growth, even when the distractions appear beneficial, such as charitable work.

None

pgrust is a Rust rewrite of PostgreSQL targeting compatibility with PostgreSQL 18.3. It passes all 46,000+ official regression tests and can initialize from an existing PostgreSQL data directory, preserving disk format. The project aims to simplify internal changes by using Rust and AI‑assisted development while keeping behavior identical to PostgreSQL. Current limitations: not production‑ready, not performance‑tuned, and lacking broad support for existing extensions (PL/Python, PL/Perl, PL/Tcl). Planned features include multithreaded internals, built‑in connection pooling, improved JSON workload handling, fast forking/branching, storage experiments (e.g., no‑vacuum designs), runtime guardrails, and more stable query planning. Installation options cover Docker, macOS (brew), and Debian/Ubuntu (apt). Build requires Rust, ICU, OpenSSL, libpq, and PostgreSQL 18 client tools; usage follows standard `initdb`, `postgres`, and `psql` commands. The code is AGPL‑3.0 licensed.

Mitchell Hashimoto, creator of Vagrant, Packer, Consul, Terraform, Vault, Nomad, Waypoint, and now Ghostty and Vouch, discusses his shift from distributed systems to low‑level desktop development. After 15 years building CLI tools, he built Ghostty—a fast, cross‑platform terminal emulator written in Zig—to explore GPU programming, cache‑locality, and terminal internals. He critiques current terminal APIs, proposing an “n‑screen” model for unlimited off‑screen buffers and a button protocol (akin to OSC 8) that preserves click events in scrollback. Hashimoto sees terminals as a lightweight application platform suited for composition, but warns against over‑extending them. He emphasizes open‑source norms: maintainers have no warranty obligations, contributors should fork when needed, and software quality stems from deep user‑centric understanding rather than sheer issue triage. Regarding Zig, he praises its evolving compiler, aggressive I/O redesign, and community ethos, while suggesting learning fundamental systems concepts in C before adopting higher‑level languages. Overall, his philosophy blends minimalism, technical rigor, and pragmatic open‑source stewardship.