Florian Cucurella

Daily reading — 2026-11-01

  • HackerTube: Andrej Karpathy builds a video interface around Hacker News recommendations, replacing uploaded thumbnails with mid-video frames and keeping preference state locally.
  • Linux’s TCP use-after-free: Atlas Security explains an IPv4 handshake flaw and the gap between upstream fixes and distribution kernels, distinguishing temporary mitigations from an installed, running fix.
  • A Scribd incident notice: An HN author shares a notice describing possible exposure of email addresses and limited payment-related information, with no evidence of misuse stated in the notice.
  • Molly’s update gap: A PieFed author warns that the Signal fork has fallen behind upstream security fixes, highlighting the maintenance burden of security-focused forks.
  • Digital minimalism: Vitalik Buterin describes reducing algorithmic feeds and mobile connectivity, pairing offline knowledge with local tools while acknowledging practical inconveniences.
  • A fixed ChatGPT account-takeover chain: Parameter describes how trusted cross-frame messages defeated otherwise isolated previews, reporting that OpenAI fixed the issue on October 26.
  • VOLTA’s AI-assisted victim profiling: Ryan Fetterman reports that an infostealer adds model-generated assessments after collecting data, shifting the operator’s analytical work to an optional AI stage.
  • Research automation restrictions: Jiahao Shao says Claude blocked model-training research automation despite his rollouts coming from an open model rather than Claude distillation.
  • Greens reconsider nuclear power: Matthew Pearce reports that England and Wales’s Green party replaces blanket opposition with case-by-case assessment, while retaining renewables as its priority.
  • Switch lowering in GCC: Xoranth traces how case density, shared targets and branch probabilities lead GCC to choose bitsets, jump tables or comparison trees.
  • Prospero: Chris Wellons presents a lean C++ coding harness with passive rendering, operation deadlines and process-tree cleanup; model access is supplied separately.
  • Rediscovering miracle wheat: Lina Zeldovich follows Claudio Grossi’s revival of heritage poulard wheat, including its resilience and the lodging risk of its tall stems.
  • GPT-5.6 Sol plays Factorio: Ryan Madden reports a completed peaceful run in 13.5 game hours, with late production failures compensated by repeated supply ferrying.
  • Memory latency and GROUP BY: Maor Kern demonstrates how radix partitioning makes aggregation faster by improving cache locality, despite performing more CPU operations.
  • French data-flow tax amendment: The Assembly’s record shows amendment 1358 passing 11–10; ministers opposed the proposal over measurement and its burden on smaller content providers.
  • Disputed Books3 testimony: Ernesto Van der Sar reports conflicting interpretations by rightsholders and Meta of a former researcher’s torrent download and later deletion of the dataset.
  • Team-solve workflow: Memdoor’s author reports 49 successes in 50 runs across ten coding problems, while documenting possible training contamination and cases where a manager replaced correct code.
  • The duty of a scholar: Colin Percival argues that knowledge carries a duty to find the people who need it, while respecting their freedom to decline or act on it.
  • Voyager’s light-day milestone: NASA expects Voyager 1 to reach one light-day from Earth on November 18, nearly fifty years after launch.
  • Protesting delivery robots: Zoe Sottile reports Carolyn Marvin’s arrest after blocking a Philadelphia delivery robot, alongside her objections to machines occupying public sidewalks.
  • Campaign account suspension: Malcolm Ferguson reports that Angie Nixon’s Latino-outreach Instagram account was suspended during a rally and restored about an hour later, without a reported explanation.
  • ParticleDB under Jepsen testing: Kyle Kingsbury reports severe correctness and data-loss bugs in tested single-node builds; Orkes says a later release fixes them, which Jepsen has not verified.
  • Vocabulary coverage versus whole lines: LearnFromX measures a wide gap between familiar word occurrences and subtitle lines containing only familiar words, while stressing that this proxy does not measure actual comprehension.

  • Flock searches after an abuse case: Silas Tsang and Peter Fleischer report that Highland Heights searches fell over 95% after a dispatcher’s personal tracking was exposed and monthly audits were introduced.

  • Matt Gravelling reports that Gruffalo illustrator Axel Scheffler is considering legal action in Germany over AI reproductions of his characters; no target company or filed case is announced. Report
  • Arjun Ingole’s hop stages a Linux installer on an internal FAT32 partition and adds a one-time UEFI boot entry, allowing supported Windows PCs to start a dual-boot installation without a USB stick. Repository
  • Hacktron reports a Savannah exploit chain from email-address shell injection to root access; the FSF says it patched the flaws and found no evidence that project data, credentials or the software supply chain were compromised. Investigation

  • Betelgeuse pages MoE experts between NVMe, RAM and VRAM; its documentation limits the roughly 22-token-per-second result to historical repeated-workload runs on one RTX 3060 setup and labels the project experimental. Repository

  • loccount counts source and logical lines, identifies languages and exports structured reports; Eric S. Raymond’s 2.26 release lists support for 772 languages.
  • The Times of Israel reports that Ahmad Tibi offered the Joint List’s support for an Eisenkot-led government from outside the coalition, without seeking a ministry.

Conversations on X

  • Dax argues that solving a problem he has himself makes it easier to build a good product, provided others can use it too. His quoted counterpoint from Jon Yongfook is that developers who can generate their own replacements make a difficult customer base.
  • Eric S. Raymond predicts that tighter AI regulation would protect cloud providers and constrain open models; this is his forecast of the political incentives.