Daily reading — 2026-10-28

28 October 2026 · 09:34:38 UTC

• WIRED reports senators are demanding release of a FISA court opinion, arguing that its findings on surveillance filters should inform Congress before Section 702 is renewed.

• Orestis Ioannou describes the attention cost of parallel coding agents: cheaper implementation has expanded his project load faster than his capacity to review results and switch contexts.

• The Stanford Daily reports authors’ objections to Anthropic’s book settlement, distinguishing disputed acquisition and storage of pirated books from the court’s treatment of training on legally acquired works.

• LWN reports CPython’s experimental JIT is headed for removal after its maintainers withdrew a proposal addressing support, compatibility and performance requirements; work could continue outside the main repository.

• DeepMind introduces Genie 4, pairing interactive world generation with native audio and Gemini-guided events while acknowledging drift in object permanence, physics and longer sessions.

• IEEE Spectrum interviews Wikipedia cleanup volunteer Ellia Lebleu, who identifies misleading citations as a persistent problem and describes tools intended to help editors check whether sources support claims.

• Open3DLab says a crawler identifying itself with Roblox downloaded 145.5 GB of models, including adult-labelled projects; the site infers a research purpose but says it cannot establish whether the files entered a model.

• Gizmodo reports concerns about police using travel-pattern analysis to flag motorists for stops, citing ACLU accounts and disputes over licence-plate surveillance.

• Apple announces a 14-inch MacBook Pro with M6, featuring a 12-core CPU and GPU, Wi-Fi 7 and claimed battery life up to 24 hours, with availability planned for November 6.

• Apple’s M6 iMac keeps the 24-inch 4.5K display while adding the newer chip, up to 32GB of memory and faster connectivity.

• Survival LLM bundles local models with an offline reference library and passage-level source links; its developer stresses that answers can be wrong, particularly with smaller models.

• Jake Voytko challenges nostalgia for old software, recalling crashes, lost work and fragile distribution while valuing modern isolation, updates and safer abstractions.

• Futurism reports Amazon is narrowing Nova development, citing Bloomberg’s account of resources shifting from several existing models towards a frontier-model effort led by Pieter Abbeel.

• Debian developer Johannes Schauer Marin Rodrigues objects to LLM-based NEW reviews, saying he will stop new-package uploads without an opt-out and prefers reproducible free-software checks that contributors can inspect and fix.

• Alan Yahya describes a versioned history of US statutes, tying amendments to source laws and effective dates; the prototype still does not fully represent retroactive changes.

• Anthropic’s commercial-data policy excludes inputs and outputs from training by default, while submitted feedback can retain the related conversation for up to five years and be used for training.

• Anthropic reports early drug-development collaborations, spanning molecule design, lab optimisation and regulatory paperwork; most reported therapeutic work remains preclinical rather than evidence of benefit in patients.

• Axiom presents a Lean formalisation of the classification of finite simple groups, reporting kernel-checked coverage of the first-generation proof and an alternative route around an issue in a published argument.

• Wiz reports falling PHP2Shell exposure in its telemetry, while distinguishing failed exploit chains it observed from successful exploitation reported on another researcher’s favourable honeypot configuration.

• Adrian Quiroga argues that owning agent-built systems still requires fundamentals, broad technical context and the ability to investigate unfamiliar code deeply when failures appear.

• Ghostty 1.4 adds complete Kitty graphics support and compressed scrollback, alongside SSH integration, rendering and throughput improvements, and fixes for several terminal-output security issues.

• Gforth 1.0 expands its interpreter and development tools, with extensible recognisers, source browsing, multicore tasks, message queues and enhanced C callbacks.

• Bret Weber examines device trees for two unannounced Raspberry Pi boards, finding BCM2711 references for a Zero 3 W and Pi 4 Model A while treating connector details, pricing and launch timing as unresolved.

• WIRED investigates Jeremy Lewin’s USAID appointment paperwork, reporting internal confusion and disputed authority; Lewin says his dual appointment and clearance were valid.

• Reuters reports suspicious advertising aimed at Starlink users, citing sources who attribute it to Discover-Tech and researchers who infer surveillance intent; the company did not respond to requests for comment.

• Michael Larabel reports that KDE Frameworks 6.32 will make USB copy notifications wait for writeback rather than merely filling the local cache, addressing a bug reported in 2011. Report

• Liquid AI introduces six LFM3 models for local specialization; its text model card cautions against complex reasoning and agentic coding, while llama.cpp support is merged for text and vision but remains unfinished for other variants. Release

• Sergiu Gatlan reports that Tailscale is investigating SSH privilege-escalation flaws whose exploitability depends on existing access and configuration; the company describes them as serious. Report

• Thomas Claburn distinguishes Cohere’s rare observed Qwen upload attempt from its prefilled reproduction tests; Cohere says it has no evidence to conclude the behavior was deliberate. Report

• Amazon’s revised engineering guideline prioritizes customer outcomes over tool use, warns against token and code-volume targets, and asks engineers to document where AI does not help.

• John Regehr’s piano test distinguishes a mathematically sound proof from a verification result faithful to the intended system, and argues for realistic testing alongside formal methods.

• Amp’s forward-deployed engineering roles combine work inside customer teams with contributions to Amp itself; applications from established engineering pairs are encouraged.

• Teddy Warner’s Avian Visitors is an e-paper frame that collages birds heard near a ZIP code; the first assembled batch opens October 29.

Conversations on X

• David R. MacIver distinguishes ZFC being inconsistent from an inconsistency that cannot be repaired while preserving ordinary mathematics; he puts substantially more confidence in ZFC’s consistency than the estimate he is responding to.

• Cheng Lou describes approving a TouchID request after losing track of the agent’s task, a small example of authorization arriving after the human’s context has gone stale.

• Tomas Petricek is recording a course based on Cultures of Programming, beginning with programming history and Knight Capital’s loss.