Florian Cucurella

Daily reading — 2026-11-04

  • AI and legacy security debt: Dina Genkina describes how AI-assisted vulnerability discovery increases pressure on maintainers who must verify and ship safe fixes.
  • Residential proxy exposure: Layer3 Intel reports proxy exits from the White House network while distinguishing observed addresses from the unknown devices or people behind them.
  • Stringman: Neufangled’s open-source cable robot uses four ceiling-mounted motors for room-scale manipulation and supports local operation.
  • HackerTV: Danny Salzman turns HackerTube’s video catalogue into themed channels with a weekly refresh.
  • AMD’s revenue forecast: Anton Shilov reports AMD’s $17 billion fourth-quarter guidance would exceed Intel’s forecast, following strong data-centre growth.
  • Go’s new JSON packages: Go’s JSON redesign separates syntax from semantics, tightens default validation and adds streaming customization while preserving compatibility with the original package.
  • Linux on an RTX Spark laptop: MindStudio reports an MSI laptop’s native Linux inference slowed sharply on mains power, tracing the gap to missing platform power-management support; Windows and WSL retained GPU boost.
  • The Internals of PostgreSQL: Hironobu Suzuki’s reference maps PostgreSQL’s cooperating subsystems, from concurrency control and buffering to recovery and replication.
  • Guerrilla bike lanes: Kurt Kohlstedt’s 2016 account describes San Francisco retaining activists’ temporary lane barriers while arranging an official installation.
  • Ripple-65B Preview: Ricursive releases a dense coding model and reports a TerminalBench 2.1 score of 80.5, with $32 million in total development and training costs.
  • Gemini’s activity controls: Google says Gemini activity can train its models when Keep Activity is enabled, including summaries and excerpts from connected apps.
  • Provable inference: Anthropic’s FlashProve prototype verifies outputs against committed model weights; its reported proof generation still costs roughly 70–95 times ordinary inference compute.
  • Pierre acquisition: OpenAI plans to acquire Pierre, whose file-tree and diff libraries already power parts of Codex, while retaining its open-source projects.
  • Mundane Halloween: Spoon & Tamago collects Japanese costumes built around small, recognizable social moments, including wrestling with a designer tap.
  • Conductor joins Meta: Conductor says its cloud service becomes read-only on November 13 and closes on November 20, with exports available before deletion begins.
  • Binance’s breach investigation: Jessica Lyons reports Binance attributes its $6.3 billion theft to a multi-stage compromise that began with recruitment-themed malware and reached wallet deployment credentials.
  • Verification and knowledge: Jessica Hullman argues that scientific knowledge needs people who can explain, challenge and repair results, beyond checking whether an answer passes a procedure.
  • GitHub’s November 4 incident: GitHub reports restored service after errors affected Git operations, pull requests, issues, API requests and Actions.
  • Active faults beneath Valles Marineris: Researchers interpret InSight seismic waves as a deep strike-slip marsquake, suggesting ancient faults still rupture on Mars.
  • Kitty 0.50: Kitty adds docked terminal windows and session restoration after logout, alongside fixes for remote-control authentication and memory errors.
  • A posttraining data commons: David Hall argues that useful reinforcement-learning tasks require manufactured environments, intent and graders that ordinary document collections do not preserve.

  • Kino’s shared editing timeline: Luke Igel describes a code-based video editor that keeps footage, timeline edits and review notes addressable for collaborators and agents.

  • Tobias Mann reports that Revolut cut AI coding inference costs by roughly 90% with self-hosted GLM models; its infrastructure lead says cheaper tokens widened use beyond coding to debugging and ticket work. Report
  • Moonshot releases Kimi K3.1 Pro’s weights: a multimodal mixture-of-experts model with 2.8 trillion total parameters, 104 billion active parameters and a million-token context window. Model card
  • ONNX Runtime’s privacy documentation says supported native builds enable telemetry by default, while an Immich issue traces the resulting privacy risk through its machine-learning dependencies. Documentation · Issue
  • NielsenIQ figures cited by Emma Loffhagen put Sarah J Maas’s new novel at 188,914 UK copies in its opening five days. Report

Conversations on X

  • Ryan Fleury argues that access patterns, rather than structure size alone, determine wasted cache-line traffic.
  • Shriram Krishnamurthi distinguishes private lecture notes from a textbook written for students, questioning the demand for notes that were never designed to be read independently.
  • Armin Ronacher finds keeping model providers and login integrations current more awkward in an iOS app than on the web.