A local coding-agent factory: Mrinal claims capacity for about 3,000 pull requests a month, with specifying enough useful work now the bottleneck.
Muse Wonder: Meta claims improvements in agent memory, sustained task execution and coding across its Wonder models.
GLM-Legal: Z.ai reports a 19.2% all-pass score on Harvey LAB after specialized training on GLM-5.3-Flash.
Hidden model capabilities: Andis Draguns and colleagues construct cryptographic backdoors that conceal computations in model weights, exposing limits of adversarial capability audits.
AI as a tool: Huahai argues skilled LLM use can preserve human direction and feedback; Rich Hickey calls LLMs automation and argues delegating reading sacrifices learning.
Fable took my job: Cendyne describes AI contributing to their role’s elimination through management’s cost decisions, while questioning models’ product judgment.
Codex’s background daemon: Yihua-Liu-1997 asks for opt-in installation, citing surprises from persistent self-updates and changed environments.
A modern web app in C: Corey Prophitt describes a commercial C/CGI application built with OpenBSD, SQLite and a small dependency set.
A router as a radio receiver: W1700K SDR turns a Wi-Fi 7 router into a receiver streaming 14-bit IQ at up to 320 Msps over 10GbE.
Intel’s D270: Ryan Smith explains the air-cooled 350 W inference GPU, with 160 GB on Intel cards and up to 480 GB through partners; pricing and availability remain unannounced.
An elusive Intel instruction: purplesyringa investigates a documented rounding-control instruction that raises undefined-instruction faults and disappeared from the 2026 manual.
Plug-in solar: Meryl Davids Landau reports growing US adoption of portable solar panels as states simplify installation rules.
Cell reprogramming and age: Lea Jessica Berg and colleagues report reduced epigenetic age while converting blood cells into induced neural stem cells, with fate consolidation taking several weeks.
A Kindle photo frame: Niraj Rajgor describes using an AI agent to repurpose a Kindle; physical display checks caught ghosting missed by automated tests.
William Zhou and Daniel Freeman’s Anthropic experiments suggest small AI models could eventually replicate across IoT devices; current end-to-end results use simplified access and stop short of demonstrating a rogue replicator.
Reuters reports that hackers raised French supermarket refrigerator temperatures, spoiling food; Danfoss says affected systems used outdated firmware.
Refrigeration controller flaws: Amir Zaltzman describes authentication bypass and command-injection vulnerabilities disclosed to Danfoss and addressed in new firmware.
Satyrical sees promise in building software for the masses with vibe coding; gingerBill argues FilmCraft lacks architecture for performance and maintainability, while peach2k2 criticizes its responsiveness.
David Sancho endorses compiling agent code against a capability library so type checking constrains actions; Eric Torreborre shared Igal Tabachnik’s enthusiasm for strong types in agent harnesses.
Kristopher Micinski wonders why quantum fails to spark his interest; Graham Hutton connects that feeling to chemistry and mechanisms he could not see.
Cindy Sridharan thinks AI operations tools may be easier to build within a company’s shared infrastructure than as a generic product needing deeper production access.