Sunday, September 20, 2026 · No. 37
21 articles · ~4h 36m read · 7 sections
A personal morning paper, assembled daily; the selection is the reader's, the words are the authors'.
Sign in to read every article's full text and download the editions, or request access.
Top Stories
A graphical desktop for the ZX Spectrum
graemep · Hacker News: Front Page · github.com · 24 min
Hardware · First personZX Spectrum · Z80 assembly · retro computing
Matched interests: Window ManagersZed EditorComputing History
ZX Desk implements an overlapping, z-ordered window manager with pull-down menus and an event queue in Z80 assembly for the 48K ZX Spectrum. Dragging windows within a single 69,888 T-state frame on physical hardware, the project derives its architecture from empirical profiling rather than reasoning. Its timings reveal that ULA memory contention adds only 14.7 percent to write costs rather than the assumed 50 percent.
Almost Never Use AI to Write Anything Substantive
erwald · Hacker News: Front Page · erichgrunewald.substack.com · 14 min
Books & writing · AnalysisAI writing · LLMs · writing process
Matched interests: WritingWriting ToolsLongform Essays
Outsourcing prose generation to language models bypasses the friction of drafting that forces writers to uncover contradictions and sharpen their thoughts. A line-by-line autopsy of a Claude-generated summary on AI chip smuggling shows how models mask factual gaps—conflating bulky servers with bare chips and omitting known enforcement budgets—behind convincing prose. Relying on downstream editing fails because reviewers passively accept plausible approximations.
The Underground Scene
Mike O'Brien · 3 Quarks Daily · 3quarksdaily.com · 14 min
History · AnalysisDe Re Metallica · mining and metallurgy history · Herbert Hoover translation
Matched interests: Computing HistoryManufacturingTech History
Reading Herbert Hoover’s 1912 translation of Georg Agricola’s 1556 treatise De Re Metallica reveals an unexpected degree of Renaissance automation, from water-powered ventilation systems to banks of fifteen cam-driven ore crushers. The immense scale of charcoal smelting required to melt iron at 1,200 degrees Celsius suggests the Industrial Revolution was less an arrival of mechanical ingenuity than a transition to fossil fuels.
Austin Mann’s iPhone 18 Pro Camera Review, From Dunton, Colorado
John Gruber · daringfireball.net · Austin Mann · 16 min
Hardware · First personiPhone camera · photography workflows · variable aperture
Matched interests: photographyPhoto WorkflowsApple
Field testing in the high Rockies demonstrates how the new f/1.48 variable aperture boosts low-light gathering by 50% and yields natural optical starbursts at f/1.8. An unannounced manual focus tool with peaking separates exposure from focal points, challenging third-party utilities. To counter synthetic imagery, a signed Reference Image mode acts as a notary stamp to keep originals provably authentic.
Tech & Engineering
What Zig felt like, coming from Rust
ksec · Hacker News: Front Page · besok.github.io · 11 min
Software engineering · First personZig · Rust · memory allocators
Matched interests: ZigRustMemory Allocators
Reimplementing an RFC 9535 JSONPath library in Zig after seven years of Rust development forces a departure from functional abstractions toward imperative mutation. Without automatic drops and monadic combinators, the codebase collapses a nested folder hierarchy into five flat files while replacing immutable transformations with in-place updates. Ubiquitous explicit allocators demand rigorous manual cleanup bookkeeping, verified through Zig's native test allocator.
HellGates, custom CPU gate-level challenge
xutaxkamay · Hacker News: Front Page · blog.xutaxkamay.com · 19 min
Hardware · First personVHDL CPU · crackme · LLM side-channel attack
Matched interests: CryptographyHardware HackingCode Generation
After resisting human reverse engineers and earlier models for a year, the HellGates crackme—a custom bit-addressable 32-bit CPU synthesized to a gate-level netlist—falls to GPT-6 in under thirty minutes. The model discovers a differential side-channel vulnerability in the lightweight state encryption, a shortcut taken to minimize logic gate counts. Exploiting the flaw allows it to replay CPU states and dump the challenge's one-gigabyte memory payload.
Consistent Hashing Proofs
ch.terabyteoff.com via klingtnet · Lobsters · ch.terabyteoff.com · 31 min
Software engineering · Analysisconsistent hashing · distributed systems · load balancing
Matched interests: Cloudflare WorkersCRDTCompression Algorithms
Moving past asymptotic Big-O bounds, an analytical derivation establishes an exact formula for workload distribution error across N servers with k hashes each. Remapping the hash ring onto a unit interval to evaluate segment probability density functions shows error converges within one percent of the asymptotic limit at fifty nodes. The resulting mathematics underpins an optimization that reclaimed over 100 terabytes of edge RAM at Cloudflare.
V Language Review (2023)
n-skvortsov-1997.github.io via lr0 · Lobsters · n-skvortsov-1997.github.io · 31 min
Software engineering · AnalysisV language · memory management · compiler bugs
Matched interests: Memory AllocatorsRustBevy
An empirical audit of V contrasts its advertised memory safety and documentation claims against reproducible compiler behavior. Valgrind profiling demonstrates that standard library utilities leak memory under manual management, while the touted autofree mode quietly disables garbage collection entirely. Enabling automated management triggers immediate runtime panics, crashing both the creator's text editor on startup and the compiler during self-hosting builds.
Hacker News: 176 points · 102 commentsLobsters: 7 points · 2 commentsComments
Your Pods Are Not Draining: I Checked
Jay Suresh Nirmal · HackerNoon · hackernoon.com · 6 min
Software engineering · First personKubernetes shutdown · SIGTERM draining · UDP load balancing
Matched interests: RunningDistributed SystemsNetworking
Using a custom Go harness across twenty-five cluster trials, an empirical audit measures packet-level connection fates during Kubernetes terminations. While handling SIGTERM cleanly preserves long-lived TCP streams, services continue advertising terminated endpoints for a median of 221 milliseconds after processes exit. For UDP traffic, graceful shutdown handlers provide zero protection, as kube-proxy silently steers active datagrams to replacement instances within 2.4 seconds.
AI & Machine Learning
Orchestrating Claude Code Agents: The Chief of Staff Pattern
octalpixel · Hacker News: Front Page · asyncdot.com · 17 min
AI & ML · Analysismulti-agent orchestration · agent verification · Claude Code
Matched interests: ClaudeAgent-Based SimulationsCRDT
Autonomous coding sessions typically degrade after an hour as context compaction erases detail and agent self-reports drift from reality. To sustain long runs, a dedicated coordinator delegates implementation to disposable worker sessions, maintaining shared state on an external board and auditing diffs directly. Every task requires a failing red gate before execution and re-runs shell commands afterward, preventing agents from passing on vacuous assertions.
Decisions, not strings: I ran an open System One model on my desk
Csaba Kecskemeti · DevQuasar · devquasar.com · 9 min
AI & ML · First personlocal LLMs · diffusion models · model calibration
Matched interests: LLM EvalsLocal LLMLLMs
Testing TypeSafe’s structured decision model concept on a local DGX Spark, OpenJev replaces autoregressive text generation with single-pass logprob evaluation over DiffusionGemma. Pre-filling an answer-sheet canvas makes off-schema output structurally impossible while returning calibrated probabilities across simple primitives. However, running against unpatched stock vLLM silently forfeits the single-pass speed advantage by falling back to standard multi-pass generation.
Brood War Bench
benswerd · Hacker News: Front Page · bw.swerdlow.dev · 9 min
AI & ML · First personLLM agents · StarCraft Brood War · agent benchmarking
Matched interests: Agent-Based SimulationsSim RacingGame Mechanics
A round-robin tournament pitting language models against each other in StarCraft finds that Codex Astra dominates with an 18–0 record, though no agent plays beyond a beginner level. Codex relies on early worker harassment, exploiting rivals that freeze for dozens of seconds deciding how to react. Meanwhile, Grok 4.6 burns over 11,000 reasoning tokens across 43 minutes without fielding a single combat unit.
Culture & Essays
internet safety for kids / a love letter to the internet
avas · Bear Blog Trending Posts · blog.avas.space · 16 min
Culture & arts · First persononline safety · childhood internet · age verification
Matched interests: Digital PrivacyDigital NomadismPersonal Websites
A personal retrospective argues that blunt platform bans and age-verification laws ignore how digital spaces, from Tumblr to niche forums, shelter youth escaping abusive homes and discovering queer identity. Blanket prohibitions fail to prevent grooming, which thrives when adults avoid frank discussions of predatory manipulation. Restricting access ultimately strips youth of agency instead of teaching the literacy required to identify danger.
The Lamentable Later Life of Lemmings
zdw · Hacker News: Front Page · filfre.net · 21 min
Games · AnalysisLemmings · game franchise history · Psygnosis
Matched interests: Game ReviewsGame MechanicsGame Development
Following its 1991 debut across twenty-four platforms, DMA Design’s puzzle franchise rapidly collapses under escalating complexity and developer fatigue. As publisher Psygnosis mandates larger characters to court a Children’s Television Workshop deal, cute deaths turn uncomfortably grisly while fiddly skill-crate mechanics ruin the gameplay. Eager to fulfill their contract and build the prototype that becomes Grand Theft Auto, the Scottish studio ships an uninspired flop.
More dirty coding tricks from game developers (2015)
web.archive.org via juliethefoxcoon · Lobsters · web.archive.org · 17 min
Games · First persongame development hacks · shipping kludges · retro engine war stories
Matched interests: Game ModdingGame DevelopmentGame Mechanics
Veteran programmers recount the desperate engineering kludges used to ship titles on impossible deadlines across decades of console and PC hardware. Solutions range from hex-editing an EMM386 crash error into an exit greeting for Wing Commander to packaging 250,000 Genesis cartridges with leaflets explaining how to bypass dead save routines. Other fixes dodge memory corruption by reordering makefiles or neutralize compiler errors by redefining const to nothing.
Mathematicians Hate AI. They Can’t Quit It
Isabella Ward · WIRED · 6 min
AI & ML · ReportedAI in mathematics · attribution · OpenAI research claims
Matched interests: AI CriticismAIAI Agents
After accusing OpenAI of leveraging his research prompts to solve the million-dollar Navier-Stokes problem with tens of thousands of agents, NYU professor Tristan Buckmaster still relies on Codex to parse the system's opaque proof. Fellow researchers facing similar attribution disputes concede that sheer efficiency makes avoiding commercial tools impossible, even as algorithmic breakthroughs upend traditional academic credit.
Niche Corner
So I have a weatherman, which also tells me the news
Yogiraj H · Popular Posts on Scour Today · Yogiraj H · 5 min
Hardware · First persone-paper display · ESP32 · RSS dashboard
Matched interests: E-Ink DisplaysOff-Grid LivingAPI Design
Repurposing a dusty LILYGO T5 4.7-inch e-paper display, a hardware build uses an integrated ESP32 to run an ambient dashboard tracking weather, astronomy, and Hacker News headlines. A Clojure HTTP server aggregates external APIs into structured JSON, which the microcontroller renders via geometric primitives and Montserrat typography. Early iterations overcome cramped grid layouts but stumble on failed mDNS hostname resolution and broken deep-sleep wake cycles.
Unfinished Business with 1-Bit Audio Chords
Michael Martin · Bumbershoot Software · bumbershootsoft.wordpress.com · 11 min
Software engineering · How-toZX Spectrum audio · 1-bit sound · assembly programming
Matched interests: Procedural MusicCompression AlgorithmsFuzzing
Synthesizing multi-channel sound on the ZX Spectrum’s monaural beeper via per-sample channel interleaving outputs three voices at a 23 kHz sampling rate within a 150-cycle Z80 loop. Audacity waveform inspections reveal that periodic voice misalignment generates an audible 7.8 kHz rattling squeal, whereas dropping to two channels pushes the parasitic frequency to 12 kHz to soften the artifact.
GPT-6 Astra Solves a WWI German Radio Cipher
prinzai.com via rajtilakjee · Lobsters · prinzai.com · 3 min
History · First personADFGVX cipher · cryptanalysis · LLM reasoning
Matched interests: CryptographyMechanical KeyboardsComputing History
Tested against an unsolved November 1918 ADFGVX transmission, the model extracts German plaintext using the encryption key “TRUPPENVERSCHIEBUNG,” revealing reports of an English cruiser arriving at Sevastopol. Historical logs from HMS Canterbury corroborate the fleet movements on November 24 and 26. The cipher eluded previous cryptanalysts because records indicate that key was used only starting December 9.
Hacker News: 102 points · 43 commentsLobsters: 12 points · 11 commentsComments
From the Blogroll
the personal internet
yenecys · Bear Blog Trending Posts · yenecys.bearblog.dev · 4 min
Culture & arts · Analysispersonal websites · early web nostalgia · MySpace HTML
Matched interests: Personal WebsitesDigital NomadismInternet History
Memories of tinkering with HTML on MySpace highlight how early web spaces valued idiosyncratic decoration—from autoplaying songs to clashing colors—over seamless utility. Modern platforms replace bespoke digital rooms with rigid profile templates, squeezing out the visible human fingerprints that made clunky pages feel lived-in. Turning back to independent websites revives a longing for un-optimized, expressive corners tended without regard for audience packaging.
How did Apple Silicon get 50% faster in three years?
Daniel Lemire · Daniel Lemire's blog · lemire.me · 3 min
Hardware · AnalysisApple Silicon · CPU microarchitecture · Geekbench
Matched interests: AppleSIMDPerformance Profiling
Tracking base Apple processors from the M2 to the M5 reveals a 52% gain in single-core and an 83% leap in multi-core performance. A 30% jump in performance-core clock speed from 3.5 GHz to 4.6 GHz accounts for most single-thread gains. Deeper architectural shifts supply the rest, expanding decode width from eight to ten instructions per cycle and scaling memory bandwidth to 154 GB/s.