Hacker News Evening Brief: 2026-04-26

Here’s the evening brief: 30 fresh Hacker News stories covering Apple Linux progress, AI eval politics, oddball hardware hacks, retro computing archaeology, and a surprising amount of discussion about repair, dotfiles, and game rules.

System Administration

Asahi Linux Progress Linux 7.0

Summary: Asahi’s latest progress report is a broad engineering update rather than a single feature launch. It covers finally automating its installer release process, enabling additional headphone sample rates by reverse-engineering Cirrus audio behavior, adding support for Apple’s ambient color-sensing hardware behind True Tone, and continuing the unglamorous work of understanding Apple Silicon power management blocks like PMGR and PMP.

HN Discussion: Readers fixated on the audio work, especially the discovery that values from a related codec could unlock 44.1/88.2/176.4/192 kHz support. The larger thread mixed admiration for the team’s reverse engineering with skepticism about how far a volunteer-heavy effort can get before the remaining 5% of polish becomes the hardest part.

My .config Ship of Theseus

Summary: This post is a personal account of consolidating a previously fragmented tmux, neovim, and zsh setup into a deliberately curated .config repository. The author’s main trick is a whitelist-heavy .gitignore plus a simple bootstrap script, turning “my dotfiles” from a pile of drift into a portable, auditable environment that can be recreated on a new machine without dragging along caches, plugin junk, or generated state.

HN Discussion: The comments turned into a familiar argument over whether deep customization is freedom or self-inflicted lock-in. A lot of people advocated for near-stock environments, while others veered into a much more emotional complaint: modern developer tools keep dumping hidden folders into $HOME, and users are increasingly fed up with software that ignores XDG-style hygiene.

Math Is Hard – OpenBSD Stories

Summary: Miod Vallat’s OpenBSD story uses VAX floating-point and exception handling to show how kernel work gets dominated by architecture-specific edge cases. The piece explains the VAX distinction between traps and faults, why divide-by-zero turns into SIGFPE, and how “resume execution after the exception” semantics create extra work for an operating system trying to present sane Unix behavior on very old hardware.

HN Discussion: Commenters compared VAX behavior to x86 and old DOS-era software, including programs that intentionally relied on divide exceptions as control flow. Another thread widened into a lament that operating-system design has mostly settled into polishing Unix descendants instead of attempting more radical ideas, with Plan 9, seL4, and unikernel-style systems brought up as the road not fully taken.

Tech Tools & Projects

Clay PCB Tutorial

Summary: This tutorial documents an “ethical hardware” experiment: building circuit boards from clay and porcelain instead of conventional PCB materials. The group reuses ATmega328P chips from dead Arduino boards, designs a 3D-printed stamp to press traces into clay, accounts for shrinkage during firing, and frames the whole exercise as both a practical making project and a critique of electronics supply chains built on plastics and mined conflict minerals.

HN Discussion: Discussion split almost immediately between people engaging with the hardware idea and people reacting to the project’s explicitly feminist framing. The more technical comments focused on whether the high-energy firing step undercuts the sustainability claim and what lower-energy substitutes—wood, copper tape, homemade adhesives, or other substrates—might make the concept more practical.

Statecharts: hierarchical state machines

Summary: Statecharts.dev is a primer site for hierarchical state machines, aimed at explaining why plain state machines become unwieldy as behavior gets more complex. It walks through hierarchy, parallel states, testing, and communication benefits, making the case that statecharts are useful precisely when behavior depends on both current state and incoming events rather than a simple linear flow.

HN Discussion: People with long memories of XState and SCXML treated the post as a good reintroduction, but also added the caveats most primers skip. The sharpest discussion was about history pseudo-states, which preserve nested state across re-entry and make the diagram less self-evident than advocates sometimes claim, meaning the chart engine is hiding real bookkeeping that still needs tests.

Show HN: Turning a Gaussian Splat into a videogame

Summary: PlayCanvas shows how to take a photorealistic Gaussian splat scene and bolt enough conventional game machinery onto it to make a playable browser FPS. The walkthrough covers converting splats into streamed SOG chunks, generating collision meshes, baking lighting probes, building a Recast navmesh, and layering in NPCs and shooter logic so the environment stops being a static scan and starts behaving like a game space.

HN Discussion: Readers liked the demo but kept circling back to the same tension: splats look great as scenery, yet meshes still do the real gameplay work. Several commenters proposed hybrid pipelines for foliage or set dressing, while others argued animation, characters, and dynamic lighting remain awkward enough that Gaussian splats still feel more like a production shortcut than a new general-purpose game substrate.

The Nintendo Switch Switch (2019)

Summary: This wonderfully literal hack turns a Nintendo Switch into a network switch by booting switchroot Ubuntu and bridging two USB Ethernet dongles through the dock’s USB-A ports. The post is mostly a clean, funny lab notebook: identify supported dongles, install Linux, use brctl and ip to create a bridge, and verify that the arrangement actually forwards traffic at about 90 Mbps.

HN Discussion: HN mostly enjoyed the joke, but the technical follow-up was real. People compared dongles, pointed out that the Switch’s USB-C path can do more than 100 Mbps with the right hardware, and noted that newer docks already include Ethernet, which only made the project feel more like a deliberately unnecessary but satisfying proof of concept.

Mine, a Coalton and Common Lisp IDE

Summary: Mine is a bundled IDE meant to make Coalton and Common Lisp feel accessible in the same spirit as QBASIC or the old Borland tools. The pitch is that a user should be able to download one application, open a file, and immediately get the Lisp workflow people actually care about—REPL-driven development, hot reloading, debugging, and live interaction—without first assembling an editor, language server, package manager, and runtime by hand.

HN Discussion: The comments were light but enthusiastic, with older readers immediately mapping the product onto Light Table, Turbo Pascal, and QuickBasic. The strongest reaction was simple nostalgia for “single download, ready to program” environments, which many people feel modern Lisp tooling still fails to provide despite the language’s reputation for interactivity.

Show HN: I remade my blog into a Windows 3.1 environment

Summary: Passo.uno rebuilds a technical-writing blog as a Windows 3.1-style shell complete with faux file menus, theme switching, and retro desktop navigation. Under the gimmick is a real content site with posts and talks on documentation, AI, and writing work, but the project’s value is in how completely it commits to the old UI language without turning the whole thing into an unreadable parody.

HN Discussion: Readers mostly treated it as a successful nostalgia piece and immediately started nitpicking it like an actual operating system. The useful feedback was concrete: better mobile behavior, a real minimize flow, and a sharper consistency check between the author’s anti-slop writing stance and the AI-generated illustrations that were still on the site when the post hit HN.

DeepSeek-V4 on Day 0: From Fast Inference to Verified RL with SGLang and Miles

Summary: LMSYS is announcing day-zero support for DeepSeek-V4 across both serving and RL training, with a stack tuned to the model’s hybrid sparse attention, manifold-constrained hyper-connections, and FP4 expert weights. The write-up is really a systems post: ShadowRadix for prefix caching, HiSparse for CPU-extended KV handling, speculative decoding, Flash Compressor, Lightning TopK, and full parallelism across DP/TP/SP/EP/PP/CP are all presented as the machinery that makes a giant model tractable on current hardware.

HN Discussion: Commenters wanted clearer comparisons more than more feature names. The main complaint was that inference-engine vendors have become strangely reluctant to publish apples-to-apples benchmarks against one another, leaving readers with long optimization lists but weaker intuition about where SGLang actually beats vLLM or TRT-LLM in practice.

AI & Tech Policy

Why SWE-bench Verified no longer measures frontier coding capabilities

Summary: OpenAI says SWE-bench Verified has stopped being a trustworthy frontier-model benchmark because the remaining failures increasingly reflect benchmark defects and training contamination rather than real coding ability. Its audit of hard cases found many flawed tests that reject functionally correct patches, and it also found evidence that frontier models can reproduce gold patches or issue details from training exposure, which makes score gains look more like memorization than progress.

HN Discussion: HN readers found the conclusion plausible but not surprising. The thread turned into a broader argument about whether any public coding benchmark can stay useful once model vendors have both the incentive and the opportunity to absorb it into training, with several people arguing that private or freshly generated evals are the only defensible path left.

The West forgot how to make things, now it’s forgetting how to code

Summary: This essay draws a line from hollowed-out industrial capability—illustrated with the difficulty of restarting Stinger missile production—to a software industry that is now risking the same kind of tacit-knowledge loss. The author’s claim is that once companies cut apprenticeship, junior hiring, and time for hands-on learning, they also cut the human pathways by which judgment gets transmitted, and AI is being sold into that weakness as a way to replace labor rather than preserve know-how.

HN Discussion: Commenters largely agreed with the “knowledge, not money, is the bottleneck” framing, though some thought the prose itself had the choppy feel of AI-assisted writing and accidentally illustrated the problem. The strongest replies emphasized management incentives rather than AI as the core culprit: short-term cost cutting destroys slack, and once the teaching pipeline disappears, documentation and automation cannot recreate field experience.

Academic & Research

Free Textbook on Engineering Thermodynamics

Summary: Olivier Cleynen has put a 330-page engineering thermodynamics textbook online for free under CC BY-SA, with SI units, 96 problems, and 59 solved examples. It is pitched as a real university text rather than lecture notes, covering the core path from energy fundamentals through closed and open systems, ideal gases, entropy, heat engines, refrigeration, and vapor power cycles.

HN Discussion: The comments were practical rather than theoretical. People asked about the book’s subfield emphasis, praised the unusual price transparency around the paid versions, and immediately noticed that the free PDF is huge, prompting a side discussion about whether the 40 MB size comes mostly from PNG-heavy diagrams.

Web & Infrastructure

Dear friend, you have built a Kubernetes (2024)

Summary: This essay argues that teams often reject Kubernetes as overkill, then slowly reimplement its responsibilities with shell scripts, Docker Compose, ad hoc deployment logic, rollbacks, service discovery, and multi-server coordination. The point is less “Kubernetes is always right” than “standardizing on one explicit set of tradeoffs can be cheaper than assembling your own invisible platform out of incidental glue.”

HN Discussion: Readers who already run Kubernetes nodded along, especially at the ecosystem argument: once the choices are standardized, tooling and operational knowledge compound. Critics pushed back on the article’s sense of inevitability, noting that single-node, on-prem, or non-SaaS workloads often do not need that much machinery and that “you have built a Helm” is the next joke waiting around the corner.

Summary: A GitHub Community post is complaining about a product change where clicking issue links in some contexts opens an overlay popup instead of navigating directly. The complaint is specific: it breaks copying canonical URLs, interrupts established navigation habits, seems to have rolled out without clear changelog notes, and offers no obvious setting to disable the new behavior.

HN Discussion: Commenters immediately compared it to the worst parts of Jira and Azure DevOps, which is not the comparison GitHub would want. The concrete objections were about assistive technology, broken “open the thing I clicked” expectations, and the feeling that large developer platforms still ship baffling UX regressions in the most workflow-critical parts of their products.

Databases Were Not Designed for This

Summary: Arpit Bhayani’s argument is that traditional database assumptions were built around deterministic, reviewed application code, not exploratory agents that invent new queries, hold connections while “thinking,” retry writes, and probe schema edges. The article recommends defensive patterns such as role-level timeouts, stricter write controls, better schema naming and documentation, and treating the database less as a passive store and more as a surface that now needs guardrails against autonomous callers.

HN Discussion: Many readers thought the article usefully named a real shift, but they also thought the obvious mitigation was being understated: do not give production databases direct write access to agents in the first place. The discussion kept returning to API layers, analytics replicas, scratch databases, auditability, and the idea that “agents breaking prod” is often a systems-design failure before it is a database-design failure.

Security & Privacy

OpenAI Privacy Filter

Summary: OpenAI’s Privacy Filter is an open-weight model for detecting and redacting PII in long-form text. The company says it is a 1.5B-parameter token-classification model with only 50M active parameters, supports up to 128k context, predicts eight privacy categories in one pass, and is small enough to run locally so raw text can be filtered before it leaves the user’s machine.

HN Discussion: The comments were much more cautious than celebratory. Practitioners liked the idea of local redaction in front of LLM calls, but warned that “PII removal” is not the same as true anonymization, and they dug into hard problems like false positives, restoring redacted values on the client so a chat still feels usable, and whether generic filters can ever understand enough context to avoid both leakage and overmasking.

Business & Industry

Show HN: A free ESG stock screener that publishes its losses and methodology

Summary: Jumpstart Signal is a stock-screening site that scans thousands of U.S. stocks daily, groups picks into lower-priced growth and higher-priced momentum buckets, and combines technical signals, fundamentals, news sentiment, and ESG exclusions. The distinctive part is not the screening recipe itself so much as the attempt at accountability: the site exposes backtests, public daily reports, and a methodology that the author says is versioned so changes to the scoring model cannot be quietly smuggled in after the fact.

HN Discussion: Readers were split between appreciating the transparency and doubting the finance claim. Skeptics pressed on whether backtests over a long bull-market window prove anything, whether ESG exclusions simply trade returns for values, and whether “publishing losses” matters unless revisions and threshold changes are versioned as rigorously as the marketing copy implies.

Eden AI – European Alternative to OpenRouter

Summary: Eden AI pitches itself as a Europe-based routing layer for AI models, with one API spanning LLMs plus speech, OCR, translation, and other specialist services. The product story is familiar but useful: integrate once, route by cost, region, or performance, add provider fallback, and avoid hardwiring application code to one upstream vendor or contract.

HN Discussion: HN did not object to routers in principle; it questioned what exactly the premium buys. The thread centered on whether a European wrapper around mostly non-European frontier models really advances sovereignty, whether the markup is justified by operational conveniences like billing consolidation and failover, and why a company selling “transparency” made commenters work so hard to identify its legal entity details.

A populist wave is rising to end the ‘captive’ repair economy

Summary: CNBC surveys how right-to-repair politics have moved from a niche consumer-rights fight into a broader anti-monopoly, cost-of-living issue. The piece runs through state laws in places like California, Colorado, Minnesota, New York, Connecticut, Oregon, and Washington; highlights Oregon’s attack on parts pairing; and notes that companies from Apple to John Deere are now dealing with lawsuits, regulation, and candidates from both parties treating repair access as affordability politics.

HN Discussion: The comments were more ideological than procedural. Readers framed repair rights as a direct test of whether buyers own the hardware they purchase or merely rent constrained functionality from manufacturers, with software locks and proprietary diagnostics seen as the modern mechanism for preserving captive aftermarket revenue.

History & Science

The Visible Zorker: Zork 1

Summary: The Visible Zorker is an in-browser exhibit for playing Zork while simultaneously exposing the game’s source-level logic and commentary. It is less an article than a software museum piece: the page loads the game, sidecar data, and a commentary pane so a player can move through the classic Infocom world while inspecting how the text adventure was structured under the hood.

HN Discussion: There was no substantive HN discussion yet, so the post had not developed a secondary theme beyond novelty and curiosity. At the time of drafting, the thread offered no meaningful critique or technical follow-up to the exhibit itself.

Cheating at Tetris

Summary: This Chalkdust article asks a fun mathematical question: if you can choose every upcoming Tetris piece, how do you force a strong player to lose within 100,000 blocks? The answer is to avoid obviously bad single-piece repetition and instead exploit the geometry of alternating S and Z tetrominoes, which steadily force awkward stacking and make eventual failure mathematically unavoidable in the simplified version of the game the author analyzes.

HN Discussion: Commenters immediately brought real-world Tetris implementations back into the picture, especially the modern “7-bag” randomizer that weakens some of the article’s infinite-randomness framing. The thread also wandered into piece-generation history, with readers pointing out that the original NES version is much closer to true randomness than most later games.

QNX on the Commodore 900 – Raiders of the lost hard drive [video]

Summary: This FOSDEM talk documents the resurrection of a rare Commodore 900 prototype, a cancelled Unix workstation built around the Zilog Z8000. The project is classic retrocomputing archaeology: repair missing or broken peripherals, reverse-engineer the keyboard interface and BIOS behavior, decode the hard-drive error, and recover enough of the machine’s original environment to boot QNX on hardware that almost vanished from history.

HN Discussion: HN had not produced a substantive discussion thread yet, so there was no strong comment-side theme to report. The submission functioned more like a pointer to the talk than a live debate.

Exposing Floating Point – Bartosz Ciechanowski (2019)

Summary: Bartosz Ciechanowski’s long-form explainer remains one of the clearest interactive tours of IEEE 754 floating-point numbers. It uses visual tools and worked examples to unpack sign, exponent, significand, normalization, rounding, special values, and precision limits, with the companion site float.exposed acting as a playground for inspecting real encodings rather than leaving the topic at the level of folklore.

HN Discussion: Commenters used the post as a springboard for practical precision stories rather than debating the article itself. They brought up NASA’s famous “how many digits of pi do you actually need” explanation, noted that display precision often greatly exceeds measurement accuracy, and wished the piece had a newer appendix on the quantization tricks now common in deep-learning workloads.

The route from Prussian military headquarters to Gary Gygax’s basement

Summary: This interview with historian Jon Peterson traces how modern tabletop role-playing emerged from military wargaming traditions rather than appearing fully formed in the 1970s. It moves from early map-and-ruler Prussian Kriegsspiel through hobby wargames and negotiation-heavy formats such as Diplomacy, showing how abstractions for simulating conflict slowly accumulated the pieces that later became role-playing conventions.

HN Discussion: Readers expanded the article’s lineage with their own examples, especially the role of Diplomacy, postal play, and formal military command-post exercises. A recurring theme was that paper simulations and real-world exercises solve different problems: one is cheap and abundant, the other captures fog, timing, confusion, and human error in ways a clean tabletop model cannot.

The Super Nintendo Cartridges (2024)

Summary: Fabien Sanglard’s teardown of SNES cartridges is a hardware-history tour of what lived on those boards besides ROM. He walks through the CIC lockout chip, SRAM, memory decoders, and the increasingly ambitious enhancement processors—especially chips like Super FX and SA-1—that let cartridges add compute, timing, or specialty behavior far beyond “game data in plastic shell.”

HN Discussion: The thread mixed nostalgia with a genuinely useful side conversation about regional design differences and specific chips. People discussed why PAL shells looked different, why Nintendo wanted the U.S. console to read more like AV equipment than a toy after the 1983 crash, and how odd parts like the S-RTC were used to drive game events tied to real-world time.

APL is more French than English

Summary: This transcription of Alan Perlis’s 1978 talk is a cultural defense of APL as something closer to a literary tradition than a cleanly engineered descendant of ALGOL. Perlis argues that APL’s dense notation, emphasis on arrays, and refusal to mimic mainstream control-flow idioms are not defects to be papered over but part of what makes the language expressive in its own distinctly non-English way.

HN Discussion: Commenters enjoyed the historical voice but did not let the talk off the hook. The main pushback was that Perlis’s skepticism toward things like richer control structures and arrays-of-arrays looks wrong in retrospect, because modern array languages largely adopted those features and became more usable rather than collapsing under the extra expressive power.

Other

Mahjong: A Visual Guide

Summary: This guide is a lightweight, visually driven introduction to four-player mahjong for complete beginners. Instead of a dense rulebook, it breaks the game into tile families, winning-hand structure, wall mechanics, and valid turn actions, using a clean interactive presentation to explain concepts that are often taught socially rather than from text.

HN Discussion: The comments showed exactly why such a guide is useful: a surprising number of readers mainly knew “mahjong” as the solitaire matching game from Windows or mobile phones. Others pointed out that even the guide only captures one slice of the game, because rule sets vary heavily across regions and house traditions, making mahjong less like a single canonical ruleset than a family resemblance.

1-Bit Hokusai’s “The Great Wave” (2023)

Summary: This project recreates Hokusai’s Great Wave as 1-bit pixel art, leaning into early Macintosh-era constraints instead of trying to simulate painterly richness with modern tools. The appeal is not just the final image but the discipline behind it: gradients, movement, and detail all have to be reconstructed through composition alone when there is no color, no soft brush, and no resolution to hide behind.

HN Discussion: Readers responded to both the art history and the constraint-driven craft. Several reflected on how the limitation to monochrome pixels makes every decision legible, while others used the thread to gush about Hokusai’s original woodblock mastery and to argue, implicitly or explicitly, that human-made work under hard constraints still carries a kind of texture that current AI image culture often flattens.

The Long Reply

Summary: David Friedman’s essay starts with a viral Threads exchange about finally following up on a five-year-old reminder to ask someone whether a tree survived. From there it becomes a small defense of replies on very long timescales, collecting examples of decade-long Twitter callbacks and twenty-year congratulations to argue that delayed response can be charming rather than awkward when the message still lands.

HN Discussion: The HN thread stayed small and mostly mirrored the tone of the piece. People traded jokes about how to acknowledge messages whose context has gone stale, and one reader immediately spun the essay’s throwaway mention of a “chicken selfie system” into the kind of whimsical project idea that keeps this kind of internet writing alive.