Published on August 31, 2026 at 06:01 CEST (UTC+2)
“I just chose words carefully” (405 points by zdw)
The article explores the challenge of justifying text in monospace fonts, where large irregular spaces create ugly gaps and hyphenation disrupts copy-paste. It highlights an astonishing solution from a late-1990s guide to Super Metroid written by rs1n, who manually chose words that exactly matched each line length for 17,000 words, producing perfect right margins with no double spaces. The author worked in a plain ASCII editor and used no formatting program whatsoever. The piece frames this as a curiosity of extreme human care applied to typography, and briefly touches on how similar rewriting is sometimes done in physical book typesetting to avoid widows and orphans.
Creepy Crawlies (976 points by zdw)
Konstantin Ryabitsev of kernel.org presents hard numbers on the damage AI crawlers cause to git.kernel.org infrastructure. Rendering git commits as HTML for scrapers consumes more CPU than all legitimate access combined — at any time there are 14 cores across 5 geo-distributed nodes exclusively serving crawlers. The reason is that kernel history is a goldmine of LLM training data: it's freely available, easy to filter, and guaranteed to be purely pre-AI content, since training on LLM-generated output leads to what he calls a "digital prion disease." He criticizes scrapers for using the most inefficient possible method to harvest open-source repositories.
Understanding ChatGPT Work (75 points by gmays)
Simon Willison unpacks OpenAI's ChatGPT Work, explaining that it is actually two distinct products: Work Cloud (running on OpenAI's servers, accessed via web or mobile) and Work Local (the desktop app formerly known as Codex, which can access files and run programs directly on your computer). The product is restricted to paid subscribers at $20/month and above; free and $8/month users don't have access. The interface presents Work as a tab alternative to regular Chat, raising the question of when to use which. Willison describes the product as "extraordinarily confusing and very powerful," noting it is being iterated on rapidly since its July 9th announcement.
It takes 5 cloud services to hear my doorbell (35 points by vghaisas)
The author recounts how buying a Blink video doorbell actually made it harder to hear visitors: the doorbell played a sound outside the door, which couldn't be heard inside, and relied on phone notifications that his wife didn't carry. He set out to make Google Home minis chime when the doorbell rang, which required stitching together five different cloud services. The post satirizes the convolution of modern smart-home integrations, where a simple local function becomes a fragile Rube Goldberg chain of internet dependencies. It's a cautionary tale about over-engineering everyday devices.
Matrox: Graphics for Professionals (45 points by BirAdam)
This article is a historical account of Matrox, the Canadian graphics company founded in 1976 by Lorne Trottier and Branko Matić in Dorval, Quebec. It traces Trottier's background — his childhood fascination with space and electronics sparked by Alan Shepard's flight and Apollo 11, his engineering master's from McGill, and his decision to enter the CPU-to-video interfacing market. The founders scored $20,000 in orders from a free listing in Electronics magazine for their first product, the MTX-1632 Video RAM, making the company immediately profitable. The piece is part of a longer series on the history of graphics hardware and computing pioneers.
Haiku R1/beta6 has been released (270 points by metrofun)
This is a release announcement for Haiku R1/beta6, which arrived about two years after beta5 and roughly a week after Haiku's 25th birthday. The release notes and download links are available on the Haiku Project's site, and users can upgrade from existing installs. The news page also lists the project's plans to mentor three students in Google Summer of Code 2026. It's a notable milestone for the open-source BeOS-inspired operating system.
CobaltC – The Successor to C? (9 points by SilentLambda)
The article presents the specification for CobaltC, a statically typed systems programming language positioned as a possible successor to C. The language features explicit ownership, deterministic destruction, compiler-checked borrowing, inferred lifetimes, explicit nullability, bounds-safe operations, structured error handling, and safe concurrency, all without requiring tracing garbage collection. The spec borrows heavily from Rust's memory-safety model while keeping a C-like syntax and low-level control. It reads as a forward-looking design document in the ongoing push for memory-safe systems programming.
Cores in space: The core memory module from a 1980 Spacelab computer (90 points by pwg)
This article examines the magnetic core memory module from a 1980 Spacelab computer — the French-built Mitra 125 MS minicomputer used in the European Spacelab program. The computer stored 128 kilobytes of RAM in tiny ferrite rings rather than silicon memory, and the author provides a detailed teardown of the core stack and planes. Spacelab was a reusable pressurized laboratory carried in the Space Shuttle's cargo bay, and it held three identical Mitra computers: one for managing the lab, one for experiments, and one as backup. The piece includes historical illustrations from NASA showing how Spacelab fit inside the Shuttle.
Show HN: NFC Energy-Harvesting PCB Business Card with an MCU (127 points by WilsonHarper)
This Show HN presents an NFC-powered PCB business card that has no battery — when tapped against a phone, 21 LEDs run an animation powered entirely by harvested NFC energy. The author started it as a simple KiCad learning project but ended up going deep into RF engineering and PCB design. The card uses the NXP NTAG I2C Plus or STMicro ST25DV-KC chips, which can harvest energy from the phone's magnetic field to power other components. The goal was to make something cheap and thin enough to hand out freely, unlike battery-powered card designs.
Sort branches by last commit date (96 points by speckx)
This is a short TIL post about simplifying git usage. The author had been using a convoluted shell script for years to display branches sorted by last commit date, piping through sed, git log, and sort. Modern git makes this unnecessary with git branch --sort committerdate or a permanent config setting of branch.sort committerdate. The manpage for git-for-each-ref reveals many other sort keys, including authordate, creatordate, taggerdate, contents:size, refname, and upstream. It's a concise reminder to check whether established tooling has caught up with your custom workarounds.
AI crawler load has become an infrastructure crisis for open-source projects. The kernel.org post gives concrete evidence that AI training-data harvesting can consume more compute than all legitimate human/developer traffic combined (14 dedicated CPU cores across 5 nodes, permanently). This is no longer theoretical — it's measurable background radiation on critical internet infrastructure. Actionable takeaway: maintainers need crawler management strategies (rate limiting, bot detection, terms-of-service enforcement) and the open-source community should pressure AI labs to scrape responsibly or pay for data access.
"Digital prion disease" — the model-collapse problem — is driving an intense scramble for pristine pre-AI data. The kernel.org article articulates why LLM training on LLM-generated content degrades models, which makes sources like the complete history of kernel commits "worth their weight in gold." The corollary is that valuable human-authored corpora are being aggressively mined, often without consent or compensation. This matters because dataset provenance and curation are becoming first-order concerns in AI development; expect more provenance tooling, licensing frameworks, and data-marketing standards in the next year.
AI products are splitting into "chat" and "agent" flavors, and even experts find them confusing. Simon Willison's breakdown of ChatGPT Work shows that OpenAI now ships at least two distinct products under one name — a cloud agent and a local agent (the re-skinned Codex) — plus a traditional Chat tab, with pricing tiers controlling access. This fragmentation is characteristic of an industry sprinting toward agentic, environment-aware AI that can run code and manipulate files, not just generate text. The implication for developers and enterprises is that capability discovery and product education are becoming real bottlenecks; choosing between chat, cloud agent, and local agent should be an architectural decision, not a marketing question.
A counter-trend of minimalism and manual mastery is thriving alongside AI hype. Several articles in this batch celebrate human craft and simplicity: hand-justifying 17,000 words of text with no tools, replacing a convoluted git script with one built-in flag, and releasing a 25-year-old-style operating system that values transparency. This suggests an audience segment that distrusts opaque, auto-generated complexity and values full understanding of their tools. For AI/ML development, this means there will be sustained demand for local models, explainability, and "AI that gets out of the way" rather than always-on assistants — and that fundamentals still need to be taught, since users want to understand what the AI is doing.
Ultra-low-power and energy-harvesting hardware is emerging as a meaningful niche in the AI era. The NFC energy-harvesting business card controls 21 LEDs with zero battery, powered purely by a phone's magnetic field, while the Spacelab core-memory teardown draws attention to the history of exotic memory technology. As AI's datacenter energy footprint draws increasing criticism, these projects demonstrate that useful interaction can happen at microwatt scale, at the edge, with no cloud dependency. The trend points toward ambient/PoC (power-over-communication) devices as an alternative to cloud-connected everything — and suggests AI/ML applications for such devices (on-card inference, local classification) will grow.
Memory-safe systems programming is converging on Rust-style ownership semantics, and AI code generation must follow. The speculative CobaltC spec adopts explicit ownership, borrowing, lifetimes, bounds safety, and no GC — essentially a Rust-inspired rethinking of C. This reflects an industry-wide push toward memory-safe languages. It matters for AI/ML because AI coding assistants are now generating vast amounts of new code; if model outputs aren't biased toward memory-safe idioms, AI could accelerate the production of exploitable C/C++-style vulnerabilities. Actionable takeaway: train fine-tunes on safe-code patterns, integrate safety linters into coding-agent pipelines, and treat memory safety as a generation requirement, not just a review step.
Cloud-service interdependence is being recognized as a fragility problem, with implications for AI agents. The "5 cloud services to hear my doorbell" post is a satire of smart-home integration, but it generalizes to AI systems, which increasingly chain together multiple cloud APIs, models, and platforms. Each dependency adds attack surface, latency, cost, and failure modes. In the same way the author could have fixed his doorbell with a local 433 MHz chime or edge audio detection, AI workloads will benefit from local-first architectures — on-device inference, local memory, and fewer round-trips to distant APIs — both for reliability and for privacy. Expect "AI Rube Goldberg backlash" to fuel growth in on-device models and edge agents.
Analysis generated by deepseek-reasoner