HackerNews Digest

The author bought a broken Sigma 45 mm f/2.8 I‑Series lens and found that, while the optics and mount were intact, no electronic controls responded when mounted on a Lumix S5. Disassembly revealed a rear plastic spacer, bayonet shims, a flex‑cable contact block (10 pins), and a C‑shaped control PCB. PCB inspection showed a TI TPS62140 buck converter, a 2.2 µH inductor, a 0603 SMT fuse (labeled “N”) on the input rail, a Toshiba TMPM341FY ARM‑M3 microcontroller, an 8‑pin GigaDevice SPI flash, and a Rohm BU24020G motor controller. Continuity testing identified an open fuse; a 2 A / 32 V fast‑blow 0603 fuse (ERB‑RE2R00V) was soldered in place. Power rails and decoupling caps were verified, and focus and aperture functions were confirmed after the fuse swap. The repair, including fuse replacement and reassembly, took under an hour and restored full lens operation.

None

The post outlines the full transformer‑based LLM pipeline. Text is first tokenized into subword IDs from a fixed vocabulary (e.g., BPE, SentencePiece). Each ID is looked up in an embedding matrix, producing a hidden‑size vector that encodes semantic meaning. Positional information is added, typically via Rotary Position Embeddings (RoPE), which rotate vectors to encode relative distances without extra parameters. Within each transformer layer, multi‑head attention computes Query, Key, and Value projections for every token; scaled dot‑product scores are softmaxed (causal masking hides future tokens) and used to weight Value vectors. Multiple heads run in parallel on learned lower‑dimensional projections, their outputs concatenated and linearly mixed. A feed‑forward network then expands each token vector, applies a non‑linearity (GELU, SwiGLU, etc.), and compresses it back; most model parameters reside here and store factual/semantic knowledge. Residual connections add each sub‑block’s output to the token vector, while layer‑norm (often RMSNorm, pre‑norm) stabilizes activations. For generation, the final token vector is projected to logits over the vocabulary, softmaxed, and sampled (temperature, top‑k/p). KV caching and grouped‑query attention reduce memory cost, and speculative decoding speeds inference. Differences among models lie in data, scale, configuration (layers, heads, MoE), and post‑training, while the core architecture remains consistent across modern LLMs.

Pre‑modern armies mirror the social, economic, and political structures of their societies. Key factors include whether a society is agrarian or nomadic, its level of state centralisation, and the nature of its elite (landed, bureaucratic, or religious). Recruitment follows four primary principles: * **Employment (paid service)** – rare as a primary model; seen in professional armies that require strong fiscal and bureaucratic capacity (e.g., Imperial Rome, early Han). * **Entitlement (citizenship‑based)** – service is a duty linked to civil rights; common in city‑states and communes where citizens equip themselves (Greek hoplites, medieval town militias). * **Vocational (social class)** – either an “all‑warrior” model where all free males fight (hunter‑gatherer, steppe nomads) or a specialised warrior aristocracy that supplies leadership and elite troops (medieval knights, Roman Senate officers). Variants include military settlers, slave soldiers, and prisoner armies. * **Clientage (feudal obligation)** – peasants or retainers provide troops in exchange for protection, land, or status; typical of fragmented polities (Carolingian levies, Anglo‑Saxon fyrd). These principles interact with land tenure, taxation, and the distribution of wealth, shaping how armies are raised, financed, and maintained in pre‑industrial societies.

The International Space Station presently hosts seven astronauts from five nations, comprising the Crew‑12 expedition and Russian cosmonauts. Commander Jessica Meir (48) leads Crew‑12; she holds a PhD in marine biology, previously studied emperor‑penguin physiology, and participated in the first all‑female spacewalk in 2019. Pilot Jack Hathaway (44) is a US Navy commander and test‑pilot graduate of the UK Empire Test Pilots’ School. French colonel Sophie Adenot (43), a helicopter test pilot and yoga instructor, is the second French woman in space. NASA physicist Chris Williams (42) transitioned from MIT cosmology and Harvard cancer research to astronaut duties, also serving as a volunteer firefighter and EMT. Russian commander Sergey Kud‑Sverchkov (42), a rocket engineer and Hero of the Russian Federation, leads the station. Sergei Mikaev (39) and Andrey Fedyaev (45) are Russian cosmonauts; Mikaev is on his first flight, while Fedyaev, on his second, previously flew on Crew‑6, becoming the second Russian to launch aboard a U.S. commercial vehicle.

pg_durable is a PostgreSQL extension (built with pgrx) that enables durable, fault‑tolerant execution of long‑running SQL workflows inside the database. Users define a function as a directed graph of composable SQL steps using operators such as `|=>` and `~>`, start it with `df.start(...)`, and the background worker checkpoints each step via the Rust‑based durable task framework duroxide and its PostgreSQL state provider duroxide‑pg. On crashes, restarts, or step failures execution resumes from the last checkpoint, eliminating external cron jobs, workers, queues, or orchestrators (e.g., Airflow, Temporal). The extension ships as Debian packages for PostgreSQL 17/18 and can be installed by adding `pg_durable` to `shared_preload_libraries`, restarting, and running `CREATE EXTENSION pg_durable;`. Access is controlled via RLS; privileges must be granted explicitly (e.g., `df.grant_usage`). Development tools include a VS Code dev container, scripts for local cluster start, and two test suites: fast pg_regress tests and comprehensive E2E tests. The project is preview‑stage, open‑source under the PostgreSQL License, and integrated in Azure HorizonDB.

A University of Rochester team led by Prof. Chunlei Guo developed a solar‑thermal desalination system that eliminates brine waste and chemical pretreatment. Black metal panels are etched with femtosecond lasers to create a super‑light‑absorbing, superwicking surface that draws a thin water film, absorbs nearly all solar radiation, and vaporizes the water. Etched grooves and the “coffee‑ring” effect direct crystallizing salts to a passive region, preventing clogging from multicomponent seawater (e.g., Mg, Ca). Tests with Pacific, Atlantic and Indian Ocean samples showed continuous freshwater production and solid‑salt recovery. By embedding hydrogen‑titanate nanoparticles in the grooves, the device selectively isolates lithium, extracting ~50 % of lithium from Great Salt Lake brine. The process yields potable water, table salt, and recoverable minerals, and is presented as inherently scalable. Research was funded by the NSF, Bill & Melinda Gates Foundation, and Worldwide Universities Network.

Gemma 4, released two months ago, now includes checkpoints optimized with Quantization‑Aware Training (QAT). QAT simulates quantization during training, minimizing quality loss compared with standard Post‑Training Quantization (PTQ). The update provides: - QAT models for the widely used Q4_0 quantization format, delivering higher overall quality than PTQ baselines. - A new mobile‑specialized quantization schema applied to the edge models (E2B and E4B), reducing the Gemma 4 E2B checkpoint to a 1 GB memory footprint, enabling deployment on consumer GPUs and edge devices. - Approximate VRAM requirements for loading each model (details omitted), demonstrating significant reductions in both VRAM and storage. The release follows earlier enhancements, including Multi‑Token Prediction (MTP) for faster inference and a 12 B model positioned between the 4 B and 26 B MOE variants. Overall, QAT‑trained checkpoints preserve Gemma 4’s capabilities while substantially lowering hardware requirements.

ABC Classic FM’s “Classic 100” annual poll results from 2001‑2021 are visualized as a dot‑and‑line chart. Each dot marks a specific composition; connecting lines show how that work’s rank shifted across the three surveyed years (2001, 2010, 2021). Interactive features allow users to select a dot, highlighting the chosen piece and related works by the same composer. The visualization illustrates trends such as “The Lark Ascending,” which ranked #2 in 2001, fell in 2010, and rose again by 2021. The page also contains a brief author note unrelated to the data.

mouseless is a web application that enables rapid mouse control using the keyboard, described as “lighting‑fast.” The interface relies on JavaScript; users must enable it in their browser settings for the service to function. No additional content or functionality is detailed.