A $10 OpenStick homelab: Laurens Bosscher runs Debian, DNS blocking, RSS, bookmarks and monitoring on a silent 512 MB device, choosing small services over a heavier monitoring stack.
Underground 2 Revived recompiles the PC racing game for Chrome and Firefox, with online races and free roaming; its creator began the port to play again with his brother.
Apple’s personal knowledge graph: Dario Pasquini and Michał Bazyli trace local relationship inference from contacts, photos and communication metadata, and release a read-only viewer; the database is protected by Full Disk Access.
The end of AI Frontiers: John Langford says he left Microsoft after its reorganisation left his research without a path to use, arguing that model scaling and fundamental research should develop together.
Apple’s predictive-text grey list: Aisha Down reports research finding over 100,000 restricted autocomplete phrases; Apple says users can still type them explicitly.
Asahi Linux’s progress report: James Calligeros announces general M3 support and limited Thunderbolt support; experimental per-process page sizes improve Windows-game performance, but are not yet shipping.
Running a law firm on Free software: Neil Brown describes five years of Debian and self-hosting, replacing little-used services with simpler tools while retaining compatibility with clients’ documents and meetings.
Discord’s Rust client: Abhijith M B reports faster cold starts and a 290 MB memory snapshot with two accounts; Discord has not released a like-for-like benchmark or public launch date.
MaxxBench: Gavin and Justin Lo find that small changes to 89 coding tasks separate agents with similar original scores; some agents explicitly favour imagined hidden tests over revised requirements.
The $8 Linux computer documents the Ox64’s early flashing hurdles; this 2022 instalment reaches a working serial connection, with networking and storage still unavailable.
Secrets in coding-agent files: Apiiro reports 2,539 credential-like secrets across public repositories and history, tracing many to literal commands saved in approval rules; the researchers did not test credential validity.
Coldframe imitates supermarket refrigeration controllers to record attempted logins, exploits and follow-up actions without controlling real equipment; its creator initially saw only generic scanners.
Turbo Pascal’s economy: Kevin Boone recalls how a compiler built for tight memory and slow disks made integrated Windows development unusually fast and approachable.
The components of agent memory: Anthropic finds that offline checks against transcripts can repair false memories, while linked notes help most when tasks require combining documents rather than finding isolated facts.
How long can life wait?: Maciej Cegłowski separates evidence of microbial survival over roughly 100,000 years from disputed claims spanning millions, showing why contamination controls matter for Mars exploration.
Terry Pratchett’s study in miniature: Sussi describes a 1:12 tribute built over six weeks, with individual book covers, a hand-sewn leather jacket and a correctly ordered miniature manuscript.
Five memorable programming books: James Hague values books that show serious work in a language and explain how to decompose problems, rather than teaching syntax through contrived exercises.
Putting the AI sell-off in perspective: Kadoa’s November 6 price-index comparison shows a much steeper drawdown in semiconductor stocks than the S&P 500, while historical recovery times exclude dividends.
A processor-die atlas gathers 356 freely licensed photographs spanning 1972–2024, with a common-scale view for chips whose dimensions are known and credit retained for each photographer.
Making open messaging useful: Michael Lynch describes the gap between his preference for open protocols and his family’s priorities of simple, reliable messaging, finding XMPP and Matrix clients difficult to use.
GitHub’s new pull-request overview puts merge blockers and next actions at the top, with filters for unresolved conversations and a newest-first comment view in an optional public preview.
Making any integer with four 2s: Eli Bendersky explains Dirac’s construction using repeated square roots and logarithms, turning a puzzle about clever expressions into a general solution.
HyperFlow H3 Flash: Video Rebirth compresses MiniMax-H3’s roughly 31-billion-parameter transformer to under 14 billion and uses four-to-eight sampling steps for synchronized video and audio; its measurements distinguish consumer-GPU and data-centre runtimes.
Port zero on the wire: Acidvegas distinguishes the sockets API’s special meaning of zero from TCP traffic, reporting exposed services that ordinary clients and default scans often miss.
GeoWorks’ lost opportunity: Ernie Smith traces a lightweight desktop system’s limited developer ecosystem and later life in PDAs, Nokia communicators and electronic typewriters.
Retro Song Bird presents a player that takes Game Boy and Game Boy Color cartridges to play their soundtracks.
The Gordon Gekko effect: Andrew W. Lo argues that financial risk management must account for leadership, peer norms and institutional feedback, proposing independent accident investigations to help firms learn from failures.
Roy Medina’s Observer watches screen, camera or audio inputs for user-defined events, running models locally or through a configurable server before triggering notifications. Project
Joel Niklaus’s Harness Leaderboard compares coding agents with models, providers and reasoning effort held constant; several leading harnesses remain statistically tied, and repeated runs expose substantial task-level noise. Benchmark
Michael Larabel reports two Vulkan 1.4.368 extensions: Valve’s vertex-address binding optimization and NVIDIA’s access to opaque internal driver handles for use with external APIs. Release report
A merged llama.cpp change runs selected MoE cache misses on the CPU and adds hardware-specific tuning; its example shows that increasing CPU work can help a small cache but hurt a larger one. Pull request
Sahil Sen and colleagues find that grep beats vector retrieval with inline results on their 116-question conversation-memory sample, but the ranking can reverse when results move through files; they caution against generalizing beyond the tested setting. Paper
Immich 3 adds configurable nightly integrity reports that identify untracked files, missing files and checksum mismatches, alongside non-destructive mobile editing and workflow previews. Release announcement
Anil Madhavapeddy surveys OxCaml 5.4.0+ox9, including unboxed layouts, explicit atomic-field updates, safer effects and compiler warnings for misleading kind annotations. OxCaml release notes.
Conversations on X
KC Sivaramakrishnan reports finding 33 OCaml issues in two days with DeepSeek v4.1 Flash for about $12, noting duplicates and private reports; his linked issue list includes concurrency, bounds-checking and runtime failures. KC’s post.
Yann LeCun distinguishes academic research on new self-supervised methods from resource-intensive model engineering, arguing that results from the two settings need different comparisons. LeCun’s post.
Kit Langton announces Blood, a language combining effect systems, metaprogramming and content-addressed code stored in SQLite. Its public page currently says “Soon.” Announcement.