Florian Cucurella

Daily reading — 2026-10-19

  • Karri Saarinen argues that individually useful features can still weaken a product without a shared point of view, and that easier software creation raises the value of judgement.
  • RepRapMicron describes open micron-scale fabrication with common materials; its current PIKA build reports roughly 1 µm repeatable positioning and 10 µm accuracy over a small workspace.
  • Parser Combinators builds structured parsers from small functions that return a parsed value and remaining input, culminating in a parser for ping output.
  • Nieman Lab reports lawyers’ concerns that Meta’s revised commercial terms still permit broad scraping; Meta says robots.txt preferences remain respected.
  • Open Micro-Manipulator combines ordinary steppers, magnetic encoders and closed-loop control for fine XYZ motion, while explicitly distinguishing its 50 nm step resolution from absolute accuracy.
  • The Legend of Hyrule presents Zelda’s branching fictional chronology alongside the games’ release order in an unofficial fan timeline.
  • HOA Spy uses indexed court and county records to examine how association liens, foreclosure filings and fees pressure Florida homeowners; filings are not a count of homes lost.
  • Eric Gullichsen recalls Autodesk’s attempt to trademark cyberspace and Gibson’s objection, tracing the supposed retaliatory trademark of his own name to a Timothy Leary prank.
  • Neil Madden explains how adding a sealed subtype can break exhaustive switches at runtime when consumers are not recompiled.
  • Jowi Morales reports on Anthropic’s proposed federal AI framework, including catastrophic-risk thresholds and a Commerce Department route to judicial intervention.
  • Get Lazarus announces dynamic packages in its Free Pascal and Lazarus branches, allowing components and plugins to update without rebuilding the entire IDE.
  • Jared Siegel argues for an LLM proxy under the user’s control to expose request logs, cache behaviour and provider switching.
  • Francesco Maggi, writing on Terence Tao’s blog, argues that AI-assisted mathematics must preserve human understanding, independent research and the training of new scientists.
  • Emma Roth reports that site policies and anti-bot systems are constraining personal agents, with businesses weighing access, security and licensing.
  • Artificial Analysis gives DeepSeek V4.1 Flash 1017 Max an Intelligence Index score of 45, describing it as fast and verbose in its evaluations.
  • PalmOS Lemmings preserves the handheld game clone’s source and compiled files for educational use.
  • Stephen Edelstein reports that Tesla’s Roadster reservation agreement promises neither a delivery date nor delivery itself, while reservations remain refundable.
  • Anthropic’s employment study reports rising output and a smaller share of junior technical hires, with human attention and agent supervision becoming bottlenecks in its own workplace.
  • Ken Shirriff traces floating-point addition and subtraction through the Intel 8087’s microcode, alignment shifter and integer adder.
  • Jason Koebler examines body-camera footage and travel records behind a warrantless licence-plate search that a federal judge ruled unconstitutional.
  • WhatsApp MCP provides read-and-search access through a linked device and Cloudflare storage; its current README requires Workers Paid despite the submission title’s free claim.
  • Idrees Kahloon uses the Holmes documentary to argue that political tolerance for fraud has grown since Theranos’s collapse.
  • Matthew Klint reports that British Airways’ consent rules for recording staff apply to smart glasses and can affect remaining flights on a ticket.
  • Hagens Berman describes its class action alleging that Verisign and ICANN maintained a .com registry monopoly and inflated wholesale fees; these remain plaintiffs’ allegations.
  • The Brain Sandwich proposes understanding a problem before delegating to AI, then reviewing and explaining the resulting diff before asking colleagues to review it.
  • Iheatu Chris Wogu describes Superlite’s parallel agent workflow and reports a 16.1-minute build passing 177 hidden tests in its own benchmark.
  • Vals AI tested six assistants on a blocked refund scenario and observed excessive outreach, invented excuses and private disclosures; one trial per model cannot establish typical behaviour or a ranking.
  • HybridAgent’s authors combine latent reasoning with text-based tool calls and persistent working memory, reporting fewer generation steps and faster inference on their benchmark comparisons.
  • Chris Kiehl argues that programming for pleasure needs no economic justification, describing hand-written games and difficult ideas studied for their own sake. Essay
  • Jonathan Corbet reports Linux 7.3’s release, including sched_ext subschedulers, virtual-machine working-set tracking and NTFS alternate data streams. LWN
  • Michael Larabel reports Fedora 45 missed its October 20 release target because release-candidate media were not ready; October 27 remains a contingency date. Phoronix
  • Unsloth’s DeepSeek V4.1 Flash 1017 model card describes improvements in reasoning, coding and agent tasks, while marking its quantization implementation as work in progress. Model card
  • Signal and Canadian metadata retention: Martin Shelton reports that Signal still rejects Bill C-22’s proposed metadata collection after amendments reduced retention from a year to six months; Signal says it will not change its product to collect that data.

Conversations on X

  • Alexis King, Talia Ringer and Aaron Eline exchange favourite jokes from programming-language papers, including James Noble’s object-orientation satire and Stephen Dolan’s footnote about elusive sources.
  • Anil Madhavapeddy jokes about zstd after ClaudeDevs says Claude Code now gzips conversations on each turn and starts cloud sessions while the user types.