Florian Cucurella

Daily reading — 2026-10-26

  • Magnus Lynch’s 200-lesion test found a fine-tuned Inception v3 more accurate at naming diagnoses than two frontier LLMs; its default discharge threshold nevertheless missed more dangerous lesions, illustrating the distinction between classification accuracy and safe triage.

  • Panamerican Dispatch reports new complaints about Meta’s election tools, with Brazilian groups asking the electoral court to investigate unequal search results; Meta has denied deliberate favouritism and blamed earlier problems on technical failures.

  • Bogomil Shopov’s FOSDEM Digital Dissident devroom focuses on the full chain protecting people under threat, including network access, metadata, publishing and devices; proposals should explain their threat models and limitations.

  • Max Woolf examines falling LLM prices, arguing that open-weight competition and low switching costs constrain providers; cache-read pricing can dominate the cost of long-running coding agents.

  • Reuters reports Serbia’s Student List won the parliamentary election, with Aleksandar Vučić conceding defeat; the incoming government’s leadership and foreign-policy direction remain unclear.

  • IBM CEO Arvind Krishna makes the case for on-premises open-weight models, claiming substantially lower token costs for many enterprise tasks while acknowledging the deployment expertise required.

  • Richard Borcherds now grades homework for completion, the Daily Californian reports, because widespread AI use has undermined correctness-based grading; Berkeley leaves course AI policies to individual faculty.

  • Alfie recalls becoming a programmer through BBS culture, cracktros and Denthor’s Pascal tutorials: the desire shifted from collecting games to making effects of his own.

  • JetBrains ThinkRail wraps the pi coding agent in an IDE with isolated Git worktrees and a graph of maintained specifications intended to preserve project knowledge between sessions.

  • fheroes2 recreates the Heroes of Might and Magic II engine with higher-resolution graphics, improved AI and interface fixes while requiring separately obtained original game resources.

  • Big Whale Hunting visualises market capitalisation against workforce, using bubble area for company value and stick figures for employees; its full-time staff counts come from annual reports and can lag by a year.

  • Drew Breunig reviews context engineering, arguing that longer contexts and persistent agents introduce new failures while pruning, offloading and fresh starts remain useful remedies.

  • Dirk Schornstein’s Löffelholz Codex project pairs facsimiles of a 1505 tool book with transcriptions and interpretations, distinguishing reviewed pages from raw AI drafts and marking uncertain readings.

  • Gerard Keating explains the workforce chart: direct employee counts exclude much franchise and supplier labour, so market value per employee says as much about business structure as labour needs.

  • DreamHost reports an ongoing service incident affecting File Manager, webmail and some email access; affected mail and FTP/SFTP users require passwords reset by their account owners.

  • Nikon withdrew a microscopy-video award, TechSpot reports, after an investigation found a rules violation; the entrant acknowledged AI post-processing but denied generating the video’s structures or motion.

  • Disney Animation outlines its filmmaking process, from story development through models, rigs, shot production, lighting and sound, showing how many disciplines contribute to each frame.

  • Bloffy’s developer describes replacing Next.js with Hono for public blog pages, reporting reduced JavaScript and faster builds while retaining Next.js for the interactive management app.

  • PaperPilot’s Rust document toolkit exposes the same local PDF engine through a command line, MCP server, REST API, browser WASM and desktop app, with merging, compression and OCR among its current features.

  • iFixit gives the iPhone Duo a provisional 5/10 for repairability, praising battery removal and accessible components while finding inner-display and hinge repairs costly and difficult.

  • BEA data on FRED put wage and salary disbursements to persons at 42.0% of US gross domestic income in 2025; the series measures wages rather than all labour compensation.

  • The F-Zero X 3DS port post-mortem describes hardware measurements that redirected optimisation towards instruction-cache behaviour, enabling a native renderer for older hardware; its quoted token costs and performance figures cover different development phases.

  • Michael Larabel reports Simon Ser’s proposed Wayland protocol for temporary virtual outputs, intended to let privileged remote-desktop clients create displays without physical hardware. Phoronix

  • Editor Carmen Gray describes how fears about losing autonomy emerged across the Texas women’s horror anthology Beneath the Bluebonnets, without being prescribed as its theme. Texas Standard

  • Haroon Siddique reports that children’s laureates Chris Haughton and Nicola Davies will give royalties to Sally Rooney in solidarity after her UK book sales were halted over her support for Palestine Action. The Guardian

  • Maggie Harrison Dupré reports on researchers’ analysis of donated ChatGPT histories, finding sensitive details and inferred preferences in saved memories; the study examined an earlier memory system. Futurism

  • Nanbeige presents its 33-billion-parameter 4.3 model as a teacher for future smaller agents, with a 512K context window and configurable retention of reasoning across tool-use turns. Model card

  • Francesca D’Annunzio and McKenzie Funk document how Ingram’s expanded immigration enforcement separated families, depleted local businesses’ workforces and disrupted flood rebuilding, drawing on interviews and city records. ProPublica

  • The mradermacher quantization team says it is nearing its Hugging Face storage limit and may delete repositories; its FAQ explains the hardware and calibration-data trade-offs behind maintaining many model variants. Maintainer notes