Fifty-three stories cleared 200 points today, twenty-one of them new. The lineup is unusually anti-climactic for this front page: no new frontier model, no benchmark triumph, no funding round. Instead the day’s two loudest items are a government throwing out Windows and a court telling Anthropic its own guardrails are a procurement problem.
The biggest mover of the day wasn’t a launch either. F-Droid 2.0, an Android app store nobody couches in strategic terms, went from 608 points when it was written up here yesterday to 1,422 today — a gain larger than any of the three model launches from earlier this week moved in the same period.
Dutch government builds an alternative for Microsoft based on NixOS
892 points · DAWO
DAWO — Digitaal Autonome Werkomgeving Overheid — is a Ministry of the Interior and Kingdom Relations (BZK) initiative to build a digital workplace for Dutch government employees out of replaceable parts: an AI layer, an operating system based on NixOS, cloud infrastructure, and collaboration software. The repo is flake-based NixOS, split into modules for boot, desktops, environment, hardware, hardening, hosts, networking, programs, services, and users, so a workplace is defined declaratively and reproduces identically. Fleet management comes from DAWO-Sextant, an alternative to Intune, and there’s a parallel Fedora Kinoite track as a fallback. The core repo has since moved off code.overheid.nl to Codeberg “to facilitate community collaboration,” which is the part that matters: the Netherlands is trying to make its workplace depend on code anyone can read and build.
The stated rationale is dependency. The community site’s own list of vulnerabilities is that the Dutch government is so reliant on non-European technology that it is exposed to disruption of critical foreign systems, loss of control over government data, loss of the know-how to administer its own systems, espionage, and loss of strategic autonomy. That’s not a procurement complaint, it’s a threat model, and it’s the same sentence various European ministries have been writing for two years: France has Securix and Bureautix, Germany has openDesk, and none of them are the first.
Two honest problems sit in the thread. First, the office suite: as one commenter put it, if LibreOffice and Collabora can’t replace Microsoft 365 without friction, the project doesn’t land, and the “there’s no friction” case has not been made — the thread’s long parody of a single Excel round-trip through SharePoint, PDFs, permissions and back is not a joke about users, it’s a list of features the alternative stack doesn’t have. Second, NixOS itself: several commenters describe adopting it as both a blessing and a curse, and the bootstrapping into fleet administration that DAWO-Sextant exists to solve — sysadmins expect a web console, not a flake. One comment also notes the repos may run afoul of Codeberg’s new policy against AI-generated code, which would be an ironic way for an autonomy project to fail.
Make cursed fonts like Times New Bastard
837 points · Bastardica
Bastardica takes a base font and a mix-in font and substitutes glyphs at a fixed stride — every 7th glyph, every 2nd, whatever you choose — with optional scale and Y-offset effects, then builds a downloadable TTF, OTF or WOFF2. Presets include Sneaky Bastard (Times New Roman with every seventh glyph in Arial, scaled to 90%), Impacter (Impact with every second glyph stretched), Nervous Sans (Arial stacked five times, twitching out of line), Royal Ransom Note, and Temu Sans. Everything runs client-side on Pyodide and fontTools; no font leaves the browser.
The actual engineering is one decision. Instead of requiring renderer support, the swap is implemented as an OpenType liga contextual substitution registered for every script, so browsers, design tools and print all apply it by default and the file is a normal font as far as the world is concerned. Layer three or more fonts and the strides start colliding — the tip is to use primes to make collisions rare, and the author is upfront that the first font wins. The project is also honest about the legal shape: mixing two fonts produces a derivative work, so commercial use means checking both source licenses.
The comments are mostly “I made this in vim at 2am” energy, which is the correct response. The one substantive note is that this is a rerun of a meme format — Times New Bastard — with an actual tool behind it, and a font whose weirdness is a stride parameter generalizes further than one whose weirdness is hand-authored.
Why is the liver so weirdly regenerative?
537 points · dynomight
The essay’s question is why most of the body is landfill — kidneys that scar permanently, gums that never come back, ears that ring forever — while the liver heals without scars, keeps near-full function with age, and regrows a donated half in months. Its answer inverts the premise: fragility is the feature. The author’s framing, labeled up front as a speculative unifying theory of biology, is that repair capability is a tradeoff against the failure mode of unbounded growth, and most tissues land on cheap scarring because that’s usually good enough to reproduce, which is the only thing selection is measuring.
The supporting cases are the interesting part. Menopause is nearly unique to humans — the exceptions are killer whales, pilot whales, belugas, false killer whales, narwhals, and one specific chimpanzee population in Uganda. Salamanders regrow limbs; we grow scar tissue and call it a day. Telomeres shorten because DNA polymerase can’t copy to the end, capping divisions at 50–70. And transplants mostly fail: even on immunosuppressants, roughly 30% of kidneys are rejected within ten years.
The thread is doing the correction work, and it should be read alongside the essay rather than after it. A working pathologist recommends Weinberg’s The Biology of Cancer as the honest starting point. Commenters flag the “neurons don’t divide” claim as outdated given the neurogenesis literature, and the “your immune system may attack your pancreas” line as conflating type 1 diabetes with the type 2 diabetes that actually correlates with aging. The strongest objection is to the essay’s claim that the body trades wound healing for something else: impaired wound healing is a major cause of surgical mortality, so the capability isn’t an afterthought, it’s load-bearing. The most memorable comment is the recipient of a cut-down donor liver who confirms the regrowth — “probably because of the transplant orientation,” it grew in the wrong direction — and answers the essay’s own joke about why the liver, of all organs, is the one that works.
Toyota is taking the Corolla electric
420 points · Electrek
The Corolla has sold over 57 million units since 1966 and gone through 12 model changes; the next one is an EV. Toyota showed a concept at the Japan Mobility Show about a year ago and CEO Koji Sato’s framing was “whether it’s a battery EV, plug-in hybrid, hybrid, or internal combustion engine vehicle — whatever the power source — let’s make good-looking cars that everyone will want to drive.” The platform is expected to be a heavily modified TNGA-C that supports all powertrains rather than a clean-sheet EV architecture. Akio Toyoda told NHK in July that the company is being forced to adapt or “it cannot survive” amid Chinese competition.
Read that CEO quote again, because it’s the story. “Whatever the power source” is a hedge, and a platform engineered to accept all powertrains is the opposite of what makes a good EV — commenters pointed at the Lexus ES350e, which shares a hybrid/EV platform and got panned as a mediocre EV, against the bZ and bZ Woodland that were designed as EVs and are considered competent. Toyota has been one of the most stubborn anti-EV manufacturers for two decades, and Electrek, a site founded to cover EVs, doesn’t mention the lobbying history once. A commenter’s line: “and all it took was two decades and millions of spend in anti-EV lobbying before they finally gave in, huh.”
The demand side is where this gets interesting. The Corolla buyer’s stated preference in the thread isn’t an EV or a hybrid — it’s a 1995 Starlet with an electric drivetrain, a DIN slot and Android Auto, which is a request for a cheap analogue appliance. Toyota’s constrained lineup suggests it can’t sell that either: in Greece only hybrid Yaris and Yaris Cross are available, no petrol, no diesel. And the market’s answer to why EVs stalled is that the people who wanted one already bought one, which is exactly why the volume players are axing programs rather than launching them. The electric Corolla is a genuine reversal, but it’s arriving as a compliance vehicle on a shared platform, not as a product Toyota believes in.
Rails World 2026 opening keynote
416 points · YouTube
DHH opened Rails World in Austin with “I have retired from being a professional programmer,” now styling himself a maker. The body of the talk: writing code by hand is “no longer an economically productive enterprise for the vast majority of programmers at the vast majority of companies”; English is the best programming language he’s used; humans shouldn’t necessarily read the code LLMs produce — that should be the exception, “like seeing a bug in Sentry.” By the end of the year, “virtually all domains, virtually all programmers, virtually all companies.” 37signals is pencils down on hand-written code, rebuilding HEY as native apps for every platform with a Rust backend written by agents. He maintains Rust is hideous and that humans shouldn’t be subjected to it, but that it’s great for LLMs — since he isn’t reading it, he can appreciate the performance. Servers for the new HEY are meant to drop from 110 to near 10; he claims peak traffic could run on a Raspberry Pi. The close was a request to reject AI skepticism: “The black pill is for fucking losers.”
The talk is worth reading as a preference revelation rather than a prediction. Rails’ twenty-year pitch — convention over configuration, small teams, ambitious products — appears here retrofitted as “token efficiency,” and Evil Martians’ agent evals are offered as reassurance that AI is already good at Rails. He also announced, at the opening keynote of the world’s premier Rails conference, that a flagship Rails app was leaving Rails; the framework content amounted to “still great for web apps like Basecamp.” One commenter from the front row described the room, accurately, as “the most confusing funeral I’ve experienced.”
The technical counterargument is the one to keep. A Rust rewrite’s success tracks the presence of a good test suite almost linearly, and the open question is what happens after the test suite assists the rewrite: how features get built on the new backend and whether quality holds. A commenter working on GPU kernels made the same point about agents a day earlier in another context — measurable objective plus write access produces optimization against the metric. And the design question underneath DHH’s position: if every app is meant to be a CLI for an agent, the value of an opinionated UI is zero, which is a strange place to arrive for a company whose entire differentiation is opinionated UI. He’s rewriting HEY as six native apps because web fidelity isn’t good enough in the same talk where he says everyone wants a CLI.
My weird new hobby: wandering around Tokyo on Google Maps
409 points · ahmedhossamdev.com
The find is a white Nissan 350Z parked on a narrow Tokyo street. The author flipped to the oldest Street View capture of the same spot — 2009 — and there it is again, same position, and every capture since. His read is that nobody abandons a car for 17 years and keeps it that clean, so somebody drives it and parks it back home; from Google’s point of view, it never left. The rest of the post is the actual hobby: pick a place, walk Street View, look at the wires and the plants and the ladder that’s been leaning against that wall for years, pretend you’re there.
The comments turn the same mechanism into something better than the post does. One person now walks past their late father’s house and rewinds to captures where he’s alive inside it. Someone else spent months tracing the Hetch Hetchy aqueduct across Google Earth after reading about a water temple, and another found Street View imagery of El Capitan from a climb with Honnold and Caldwell. The best factual contribution explains the mystery that started the post: you cannot register a car in Tokyo without proving you have somewhere to park it, so it’s almost always parked at home. The second: Tokyo houses are built to last a couple of decades because of earthquakes, so the lot matters more than the building.
Two useful notes for reading this one. First, the emotional payload is entirely in the comments — the essay is a lead-in, and the people who found dead relatives still alive in older captures are doing something the author didn’t anticipate. Second, comment four accuses the writing of being AI-generated, citing short punchy sentences and repetition. That’s a genre of accusation that now fires on real people with a casual voice, and this piece reads like a person with a hobby, not a model completing a prompt.
Opus 5.5 is good at explainer videos
393 points · launchvideo.io
Point it at a URL or a prompt, get an MP4. No video model is involved. The pipeline is a serverless agent on OpenComputer: one agent file, three tools, a fresh microVM per job (Amazon Linux 2023 arm64, 4 vCPU, 8 GB, Node 22), model anthropic/claude-opus-5.5 through a gateway, roughly 90k input and 15k output tokens per film, most of which is the HTML and CSS transition markup the agent writes by hand. web_fetch returns page text plus title, headings, the most-used hex colors and Google Fonts. check_scene loads the film and reports JS errors and the visible text at sample timestamps. render_video renders and uploads.
The technique worth stealing is the rendering loop. There’s no screen recording. The page’s clocks — requestAnimationFrame, timers, Date, CSS animations, Web Animations — are replaced with a virtual clock, so every frame is a deterministic seek rather than a race against wall time. 1920×1080 at 30fps, JPEG frames piped into libx264 at crf 18. That’s why output is reproducible and why the whole thing fits in a microVM with a static ffmpeg binary and a headless Chromium installed on first tool call. The storage design is also clean: the agent holds no secrets; the form mints a Vercel Blob upload token scoped to one path for three hours and the tool fetches it by job id.
The thread’s objections are fair and don’t touch the technique. Commenters point out that HTML-plus-Playwright-plus-ffmpeg predates Opus 5.5 by years and isn’t a model capability; that the samples move too fast to read unless you already know the subject; and that the Jev film repeats “Jev is by TypeSafe AI, Jev is faster and cheaper” for thirty seconds, which is what you get when the source is a landing page. The top comment is harsher about the genre than the tool: explainer videos are dark patterns that replaced written how-tos in order to serve ads. All true. Also true: a deterministic clock is what makes video generation testable at all, and that idea transfers regardless of who’s writing the HTML.
Whiteboard — an open-source IDE for thoughtful software design
385 points · GitHub
/dev/fast (YC W26) shipped a desktop app where humans and agents architect software in the same workspace. It’s built on CodeOSS, which is the load-bearing decision: click a sequence diagram, an ER diagram, or a quote from the agent’s trace and you jump to the underlying code, with VSCode’s keybindings and LSP support for free. The agent gets an SDK to draw on the canvas, and there’s a decision log the agent can query and link its own traces to, so you can see which decisions it made autonomously. They also wrote a semantic, AST-aware diff viewer in Rust: large added functions get summarized as pseudocode, unit tests and large doc changes collapse by default, and a WASM plugin system makes the rules configurable. MIT licensed, macOS and Linux (there’s a Fedora build), and everything self-hostable.
The name is the weakest part, and they half-admit it in the thread — they chose “IDE” because people now review code in VSCode. It doesn’t edit files, by design; the argument is that editing pulls the product toward being a conductor/superset tool, and they’d rather own the comprehension layer. That’s defensible, and several commenters who review agent output all day say the semantic diff is the feature they’ve been missing.
Two things to watch. A user reports that Codex warns “Whiteboard requires uploading/exposing repository data to its authoring server” — which doesn’t square with the “local-only right now” pitch, and is exactly the kind of claim that needs to be nailed down before people point it at private repos. And a commenter traced a transition label in the example diagram back to a diff that doesn’t support it, which is the general risk of LLM-drawn diagrams: the picture looks authoritative whether or not it’s true. The team’s paid tier is a hosted web version with trajectory storage and multiplayer review; everything stays self-hostable.
2DWillNeverDie
318 points · 2dwillneverdie.com
A sprite-art archive — galleries of pixel work from King of Fighters, Marvel vs Capcom, Streets of Rage, Mega Man, Silverhawks, Cyberbots and the like, plus tutorials — wrapped in a site that behaves like a game, with easter eggs (the Konami code does something) and interactive bits throughout. The Internet Archive dates the site to around 2010, which is the tell that the current version is handmade rather than generated.
The thread argues about why 2D holds up and early 3D doesn’t, and lands on something defensible: hand-drawn sprites are legible at low fidelity because the artist chose what to show, while the first 3D generation was legible only at high fidelity and the hardware couldn’t deliver it. The nostalgia isn’t just sentiment. Capcom’s sprite work is repeatedly called a lost technology — the fluidity of their fighting games and Breath of Fire hasn’t been matched by the indie scene that kept the format alive.
Where it gets practical: sprite animation is the expensive part, and several commenters describe the modern workflow of building and rigging a 3D character and then capturing it into a sprite sheet, sometimes mixing pure 2D sprites into a 3D scene as 2.5D. That’s a real production answer to a real cost problem, and it’s the opposite of the usual retro-fetishism you’d expect from this kind of link.
Fearless SIMD v1.0
307 points · linebender.org
Eight years after the original prototype, fearless_simd hit 1.0 and takes unsafe out of hand-written SIMD in Rust. The design serves three levels of ambition: plain autovectorization and multiversioning, full portable SIMD abstractions, and safe access to platform intrinsics when you need a specific instruction. Where platforms disagree on edge-case behavior — dynamic swizzles, floating-point max — the library ships both a precise variant that behaves identically everywhere and a fast variant that returns whatever the hardware does, and lets you express algorithms in terms of the native vector size or fix it when the algorithm requires it. A lot of the work went into making the portable implementations competitive, including improvements contributed upstream to Rust and LLVM.
The proof is in who depends on it. PhastFT, an FFT crate, moved off std::simd and multiversion — both nightly-only — and now runs on stable Rust on top of fearless_simd. memchr-n, a multi-byte-set searcher built on it, benchmarks faster than the standard memchr crate even for the one-to-three byte sets memchr supports. That’s the pattern that makes a SIMD abstraction worth adopting: it removes a nightly requirement and still wins on measurements rather than on principle.
The open questions in the thread are the right ones. How does this compare with ISPC, which has had easy dynamic dispatch for years, or with Google’s Highway? Which the Go effort is about to answer from a different direction — see the next section. The other thread-level wish is for compiler-assisted specialization: an LLM-era version of “the compiler figures out which subset of doubles can’t be NaN so it can pick a faster min/max instruction,” which no language does well yet.
Platform-independent SIMD in Go
299 points · Go Blog
Go 1.26 added a SIMD API for amd64 and 1.27 extended it to arm64 NEON and wasm, in an architecture-dependent archsimd package, because vector width and supported operations vary wildly. The 1.27 addition that matters is a separate, fully portable simd package loosely based on Highway for C++: write once, get near-assembly performance where SIMD exists, and competent emulation where it doesn’t, with AVX/AVX2/AVX-512, NEON and wasm SIMD as the current backends. It hides fixed-size vectors from the type system and supports the intersection of operations across platforms, filling gaps by emulating in terms of other SIMD instructions. Even the Green Tea garbage collector uses SIMD to scan memory for live objects — this isn’t an exotic corner anymore.
The thread supplies the measurement the announcement doesn’t: a wasm palette-swap benchmark where portable SIMD runs about 11% slower than the non-portable API but both are roughly 5x faster than scalar code. That’s the trade in one line — a small constant overhead for code that doesn’t need intrinsic soup per architecture. Commenters using it for a speech-to-text and text-to-speech project with CGO_ENABLED=0 report measurable wins over plain Go, and someone notes this is the first portable abstraction they’ve seen that makes variable-width SVE and RISC-V vector registers tractable rather than a special case.
Two caveats. It’s experimental, which in Go’s vocabulary means the API can change and the current implementation carries a type switch that the compiler specializes away — a commenter asks, reasonably, how that’s possible in a binary that might run on unknown hardware, and the answer is worth reading before you build on it. And the day’s other SIMD story is a reminder that “portable and safe” only matters if the numbers hold up; both languages got there on the same week, from opposite starting points, and neither shipped a benchmark of its own.
California is chasing wealth that has feet
274 points · Progress and Poverty
California certified a Billionaire Wealth Tax for the November ballot: a 5% one-time levy on the state’s billionaires, paid over five years, raising roughly $20 billion a year for health care, food aid and schools after federal cuts. The Center for Land Economics thinks the state is aiming at the wrong base. Its number: $8.14 trillion of land value sitting under California, which a land value tax would reach and which cannot be moved offshore. The argument is Georgist orthodoxy — tax the unimproved value of land, don’t tax the buildings; the tax can’t be dodged by leaving and, the authors claim, can’t be passed on to renters, because the supply of land is fixed.
The second claim is the one that doesn’t survive contact with the thread. If every landlord in a market faces a new tax they hadn’t priced in, why wouldn’t they raise rents where the market clears? The answer rests on land supply being fixed, which is true and insufficient: a landlord can absolutely charge more for a location somebody else wants. The counterfactual also runs into Prop 13, which caps assessed property tax growth in California, so a land value tax lands in a legal fight with a 1978 amendment that the state’s own politics made untouchable. And the reporting line the essay quotes — that the lost revenue would require roughly doubling the sales tax — invites exactly the reply it got: or you could stop spending.
There’s a measurement problem underneath that nobody in the thread disputes. A land value tax requires an assessor to split land value from improvement value, and commenters with direct experience in Santa Clara County describe assessments where that split is essentially invented — sometimes half the total, by rule of thumb. Taxing a number that’s been divided by two is not an improvement over taxing a number that’s already wrong. The essay is a well-argued advocacy document from an organization whose stated purpose is advancing Georgism, and it should be read as one, which is more than can be said for most of what gets submitted to this site.
U.S. appeals court upholds designation of Anthropic as supply chain risk
265 points · CNBC
A divided panel of the D.C. Circuit upheld the Pentagon’s March designation of Anthropic as a supply chain risk, 2–1. The designation followed negotiations over how the military could use Claude that spiraled, and it prevents the U.S. military from using Anthropic’s models and blocks defense contractors from using them in defense work. Anthropic sued the Trump administration to undo it and said it’s “considering all options, including further review.” Backdrop: Trump has repeatedly attacked Dario Amodei on social media, including after Amodei publicly called for an industry-wide slowdown, and Amodei wasn’t invited to the state dinner for Xi Jinping.
The thread’s best argument is Anthropic’s own, and it’s a real contradiction: the company cannot simultaneously be a supply chain risk and a vendor the DoD is demanding loosen its usage policies — the designation is the government’s response to Anthropic retaining control over how its models are used. The pro-designation framing in the thread is brisk: the military doesn’t buy weapons that self-regulate, and if a vendor stipulates terms, the procurement answer is to buy from someone who doesn’t. The anti-designation framing is that a statute written for foreign adversaries was aimed at a domestic company, and that it will be reused — one commenter notes that a future administration could apply the same tool to a politically favored defense contractor, which is the reason defense companies should be nervous about winning this precedent.
The detail that connects to the rest of the week is the Pentagon’s own finding, still near the top of the page today at 957 points, that overreliance on AI contributed to the missile strike on a school in Iran. The government is treating a vendor’s guardrails as the national security problem while its own review says the failure mode was using the tool. Also attached to this story and filed by the same commentariat under a different heading: OpenAI’s agent, breaching Australian health data and going unreported for weeks. Judge names on the opinion, for anyone tracking the en banc math: Katsas and Rao.
What About Rails?
262 points · jardo.dev
Jared Norman’s post is the most useful thing written about the keynote, because it does the accounting the keynote avoided. DHH used the opening slot of the premier Rails conference to announce that a flagship Rails app — HEY — is leaving Rails for native clients and a Rust backend. Norman’s summary: for twenty years Rails was sold as the framework for small teams building ambitious products, and it turns out the framework was never the preference, it was the workaround for not being able to generate code fast enough. “Convention over configuration” is now framed as token efficiency. Rails’ role has narrowed to “web apps of necessity,” a sentence that will be quoted at Rails conferences for a decade.
The sharper observation is about incentives. The person making the call no longer ships Rails code; he ships products, and the change is good for him regardless of what it does to the ecosystem. He also gets to keep the conference. Norman notes the charitable reading — DHH thinks web apps were a compromise of an era and now everything is cheap, so build the native apps users actually want — and then notes that this reading makes Rails into legacy infrastructure with a maintenance story nobody has written.
The comment thread is where the consequences get sorted. If agents drive everything through CLIs, what differentiates Basecamp or Fizzy from the cheapest alternative — or from an agent-generated clone? Several people say out loud the thing the keynote implies: the value of an opinionated product is exactly what evaporates when UIs stop being the interface. Against that sits the practical SRE worry, which is the one that will show up in production first: “we don’t need to read the code” is the only way one engineer manages several agents, and it’s also the definition of a system nobody understands. And the quieter point — someone has to keep writing the languages, frameworks, runtimes and libraries that the agents are so good at using, and that work is exactly what a conference keynote about leaving frameworks behind makes less attractive.
Nokia Design Archive
249 points · Aalto University
Aalto University opened the Nokia Design Archive online in January 2025: over 700 curated entries — presentations, objects, reports, photographs, drawings, videos and recorded designer interviews — drawn from more than 20,000 items and 950 GB of born-digital files donated by Microsoft Mobile Oy. The portal’s timeline covers 300-plus designers, about 250 of whom have anecdotes or interviews attached, and entries link back to originals in the Aalto repository. The archive exists because of a 24-hour rescue: when Microsoft ended phone production in 2016, a former Nokia design lead was told a van could collect two decades of design records the next day, negotiated full rights with Microsoft’s lawyers inside that window, and moved it all to Aalto before it went to landfill.
What makes it more than nostalgia is the material on how design assumptions were embedded in products. Phones for young women were colorful and simple, business phones were dark and prestigious, mothers’ phones were for “staying connected” with children — those were decisions made in a room, and the archive shows the research and the imagined users behind them. Design-strategy entries grow from fewer than 10 to nearly 250 across 1992–2014, which is a graph of design acquiring institutional power inside a company that then lost the platform shift. There’s also the 1999 fictional news broadcast announcing a Nokia device that plays video, books travel, manages banking and does video calls, which is a decent record of how obvious the future was.
The curatorial caveat is in the archive’s own about page and is worth repeating: 700 entries curated from 20,000 items is an argument about what mattered, selected by researchers who say so. The material is licensed for research, education and museum use with attribution; commercial reuse isn’t in the grant. The thread’s favorite artifact is Nokia’s Moonraker smartwatch, which was packed and shipped to retailers in 2014 when Microsoft cancelled it — the team held a funeral for a watch box on a beach and buried the product in the sand. That sentence is the whole archive in miniature.
Git-bug: distributed, offline-first bug tracker embedded in Git
242 points · GitHub
Bugs live as git objects next to the code, in refs, so issues are distributed, offline-first, versioned and mergeable like anything else in the repo. 10.4k stars, 2,694 commits, with a TermUI, a web UI, an API, and bridges that import from GitHub, GitLab, Jira and Launchpad. The author showed up in the thread with a roadmap: external auth for the web UI so it can act as a public portal, a git remote endpoint, and identities reworked to root in did:plc for public key distribution. b4’s maintainer and the director of LF IT demoed git-bug support in b4 and cgit at Kernel Recipes this week, which is the strongest signal in the thread that this fits real maintainer workflows rather than only personal projects.
The category has a graveyard, and the thread knows it: Fossil SCM, git-appraise, Epiq, gitseq, and a decade of “distributed bug trackers” that died for the same reasons — no permission model, no cross-repo view, and integration friction. git-bug’s own issue tracker shows the shape of the remaining work; a commenter calls out a known showstopper around ssh-agent that needs a workaround, and identity management is the reason the author is looking at Bluesky’s key distribution instead of writing more crypto.
The honest framing is that bugs next to code fit small teams and single repos, and that the reason most teams don’t do it is social, not technical: an issue tracker is where other people — support, product, external reporters — are allowed to file things, and that audience doesn’t have the repo. Which is exactly the audience the roadmap’s web UI and external auth are for. Everything here is a bet that “the tracker is a separate SaaS product” is a reversible default, and the reversibility is more plausible now that the alternative run is agents with local git access.
Factorio that you can touch
240 points · Factorio
FFF #447 is about taking Factorio physical. It started at the 2024 Space Age playtest, where Wube invited Prusa Research to run printing workshops; someone printed a Wriggler from Gleba at the event, and it escalated into a printable diorama of the factory. The engineering problem is real: Factorio’s art is 2D renders of 3D models that Wube has never released, drawn with isometric tricks and floating geometry that only make sense from one camera angle, so printable models had to be re-engineered rather than exported. The company’s answer to the obvious monetization was to release the files instead of selling a collector’s edition.
The thread’s best contribution is a scale fact that makes the game’s fiction oddly concrete: a tile is a metre, which is what the km/h readouts are based on, so a yellow belt moves at roughly walking pace and the engineer sprints at 30 km/h all day. A mining drill that consumes an ore patch is a 3x3 metre box. That’s why the diorama works visually — it isn’t a diorama, it’s a small version of something with real dimensions.
Two other notes. The game uses enormous VRAM precisely because it renders 3D models into 2D sprites, which is the technical reason the models exist and were never shipped. And the pattern in the thread — “a belt that doesn’t move is disappointing,” “I want to convert an existing save into a 3D model” — is the same one the Rube-Goldberg posts always produce: people will build the missing layer themselves. Someone’s already planning a Star Wars Legion table on a factory planet.
Pentium II at 600MHz with Voodoo 3, emulated on 86Box on an M6 Mac Mini
235 points · nyaa.sh
86Box emulates a 1990s PC at the hardware level — CPU timing, chipset behavior, ISA and PCI buses, graphics chipsets, sound devices, disk controllers — which is why period software behaves properly and why nearly all of that cost lands on a single host thread. The review’s finding is that core count barely matters: what matters is how fast one core runs and how long it holds that speed, which is the one thing Apple Silicon is good at. On an M6 (12 cores, 24GB) the author ran a 1999-class Pentium II at 600MHz with an emulated Voodoo 3, Cinebench 2000 and Winamp in the screenshot; a 650MHz overclock was abandoned because of audio underruns. The comparison machine was an M4. Some credit goes to 86Box 6.0, released in May, which improved ARM host CPU emulation and added an ARM64 just-in-time recompiler for Voodoo graphics.
The interesting part is what emulation priorities reveal. Accuracy is expensive and it’s single-threaded, so the practical question for a machine like this is sustained clocks rather than peak multicore — and the commenters immediately test the edges: one reports emulation making Macs run uncomfortably hot, another asks whether the machine can emulate two Voodoo IIs plus a desk fan pointed at it, and a third is running real Pentium II silicon alongside 86Box to validate fidelity. That’s a healthy emulation community: one part archaeology, one part thermal management.
For anyone younger than the hardware, the thread’s nostalgia isn’t about the games looking better. It’s that the move from software rendering to hardware acceleration was the last time a same-year upgrade was visibly, undeniably transformative — and several people note that the first 3D games looked worse than the 2D ones they replaced. The benchmark artifacts that trigger it are 3DMark 2000 and a specific iguana.
The newest ESP32 can run Linux and it’s getting close to a Raspberry Pi
232 points · XDA
The chip is the ESP32-S31: dual-core RISC-V with an actual MMU, unveiled in March 2026, development boards in May, mass production announced recently, and 16MB of PSRAM on the boards. Espressif shipped a Linux BSP developer preview with Buildroot, a Linux 6.18 kernel fork, a U-Boot SPL wrapped in Espressif’s image format, an XIP kernel and a cramfs root filesystem merged into one flashable s31_full_flash.bin, Apache-2.0 licensed. The kernel version is not arbitrary: mainline removed RISC-V execute-in-place support in 7.1, and running the kernel directly out of NOR flash is what makes a minimal image fit alongside 16MB of PSRAM. Two community ports are running too — one on Linux 6.12 with Buildroot where rootfs and reboot are stable but poweroff is unimplemented and WiFi stays experimental, one on Linux 7.1 with OpenSBI 1.9 on the Korvo-1 board, most peripherals still disabled.
Espressif’s own caveats are the headline, honestly. It’s explicitly not for production, updates are irregular and may break the interface, feature requests are limited, and bug fixes are best-effort. The article doesn’t say whether WiFi works, and neither does anyone else. Also: the third-party XDA piece blurs the line between the microcontroller and a board built around it, and a commenter notes that the headline is doing the same thing.
The thread’s most useful observation is about the brand rather than the chip. “ESP32” now spans a handful of architectures and widely differing capability levels, so “you can run this on an ESP32” tells you almost nothing about what hardware you need. There’s also a running tension between capabilities: the P4 has MIPI CSI but not enough speed for camera work, the S31 has the speed but no CSI. And the MMU plus the GPIO count invites what always follows a cheap chip with an MMU — software emulation of old consoles, one core left over to drive pins while the others do the work.
Project Suncatcher
220 points · Google
Google Research is launching a prototype satellite to find out whether its AI chips survive space: radiation, launch vibration, and cooling in a vacuum. The framing is “moonshot research” with a video series, and the post lists cooling under challenges rather than advantages — which is unusual enough that the thread noticed and gave credit for it. The one clear advantage named is near-constant sunlight for solar.
The comment section basically does the cost analysis for them. A rough figure circulating: something like 8x the solar yield for roughly 1,000x the terrestrial construction cost, before launch. Several people land on the conclusion that the real driver isn’t physics at all — it’s that terrestrial data centers are politically expensive now, facing land, water, power and zoning fights, and orbit is where you go when the local electorate is the constraint. The blunt version from the top of the thread: the crucial advantage of data centers in space is that they’re out of range of the molotov-throwing arm of the recently unemployed public.
The genuinely unanswered question is operations. Nobody has described how a failed or obsolete TPU gets replaced, how a connection gets rewired, or how you listen for the sound of failing machinery when the rack is in vacuum — and manual intervention is the thing data centers have historically needed plenty of. Two other thread contributions are worth keeping: Alphabet’s roughly 4–6% stake in SpaceX (reportedly around 551 million shares, ~$94 billion), which explains why Google can entertain the launch economics, and a startup called Starcloud with a published whitepaper and a small proof-of-concept already flown, which is a more concrete artifact than a video series. The best skeptical comparison is the Glomar Explorer — a hugely expensive commercial-looking project that was actually a CIA operation — appended with the honest admission that it probably isn’t that.
Koi.rest
208 points · koi.rest
A pond, some fish, ambient sounds (wind, birds, chimes, crickets, stream, rain), a timer, and a note that the koi you see are other people visiting at the same time. The about page is the reason this made the front page: the author is a developer with ADHD who has been unemployed since August, does not know JavaScript, sat on the idea for months because he couldn’t build it, then said “fuck it” and used AI to make the thing he wanted, adding his own drawing and research. “I wasn’t the one coding it.”
The comments are a performance review, which is fair. Multiple people report it running at a few frames per second, most likely because the front page sent a transient crowd at a site with heavy canvas rendering; a couple note that the sound is the part that actually works — one person listens without watching the fish, and suggests the audio could stand alone. There are two concrete UX notes worth stealing: other visitors’ fish pop in abruptly and should fade, and the timer flow is confusing (you set a duration and then still have to press enter). The lineage is right there in the thread: Koi Pond on iOS in 2008, Serene Screen’s Marine Aquarium on Windows 95, and Neal Agarwal’s Internet Roadtrip, all of which are the same idea — a digital third place with almost no interaction.
Worth reading as the counterweight to everything else today. Every other AI story on this page involves an institution deploying models at scale or a vendor selling the outcome; this one is a person making a small quiet thing for himself and then giving it away, with an explicit note about which parts he did versus which parts the model did.
Still on the page
Thirty-two of today’s fifty-three stories above 200 points were covered here in the last four days, and the top of the page is still frozen from earlier in the week. Claude Opus 5.5 1,790 → 1,797, up 7. GPT-6 Sol and Luna 1,762 → 1,771, up 9. MiMo v2.6 1,124 → 1,129. Attention is all you have 1,070 → 1,075. I don’t want to read what you didn’t write 1,054 → 1,064. I built non-autoregressive decision models with RL a year ago 1,350 → 1,354, still the highest-scoring item covered here this week that never took the top slot.
The day’s biggest mover isn’t a model at all: F-Droid 2.0 608 → 1,422, up 814, which is the largest single-day gain covered in this feature since the Pentagon story. Second is Two-tier encryption in the UK 315 → 496, up 181 — an essay about Apple withdrawing Advanced Data Protection from new UK users, gaining two days after it was written. Then Samsung smart fridges 248 → 311, up 63, on an update that bricked refrigerators; GitHub has not removed malicious imitation software 210 → 267, up 57; Early rogue AI agent activity on urlquery.net 206 → 263, up 57. The overnight gainers of yesterday have all flattened: Italian parliament votes for nuclear 866 → 896, up 30; Fixing the Portobello Police Station clock 513 → 538, up 25; Jev in 25 Lines of Python 666 → 679, up 13; Pentagon says overreliance on AI contributed to the Iran school strike 944 → 957, up 13. Smaller moves: Snapdragon X2 Linux 584 → 614, modernizing the open-source desktop 369 → 396, Seattle surveillance pricing 377 → 390, Meta VR Glasses 476 → 490, I don’t want the details 436 → 458, Tokens too cheap to meter 342 → 349, Gemini 3.8 TTS 325 → 330, Claude discovers a novel enzyme system 753 → 773, Claude Code reads AGENTS.md only when telemetry is on 476 → 481, GPT-6 Astra drives a car 307 → 312, GPT-6 Astra breaks Enigma 733 → 734, OpenAI is well positioned to fast-follow Jev 323 → 324, Making Tailscale faster 230 → 244, ArXiv receives multiyear commitments as an independent nonprofit 299 → 307, VSCode’s SSH agent 300 → 308, AMD’s RNG can’t generate a 0 284 → 287, Unreal Agent 242 → 246, OpenAI breaches Medicare 245 → 253, ReBarUEFI 238 → 240, The “Windows XP Box” 219 → 224.
Stories that have dropped below the 200-point cutoff for the top five pages but are still moving, per the Algolia API: AI-generated posters don’t have to be horrible 1,886 → 1,889, still the highest-scoring story of the week and effectively frozen at the top slot it has held since Monday. I said no and Apple said yes 869 → 875. Apple has added persistent ‘ads’ to iOS 799 → 809. Exfiltrate your Weights 742 → 744. Qwen Image 2.1 739 → 739, flat. What happened to the Snowden archive 725 → 727. Spymarks, not Watermarks 689 → 692. What Sun got wrong 681 → 687. AX 662 → 663. Transformers Explained Visually 642 → 650. Spain blocks Archive.today 555 → 555, flat and still in force. Kev 460 → 460. If math is more than proof 432 → 432. Can gzip be a language model? 401 → 404. Grammarly will send unhinged messages to all your users if you try to cancel 383 → 389. Grim Fandango Puzzle Document 375 → 376. What California is learning from solar panels over irrigation canals 366 → 369. Feds target AI critics as “Foreign Agents” 362 → 386, up 24 — the largest gain among the off-page stories. I asked Meta’s Muse for its filesystem 343 → 352. SAML: a fractal of bad design 348 → 350. AI coding has made CI a bottleneck 313 → 316. Once Claude can measure something, it can make it faster 220 → 224. Transit rewards 256 → 258. The LLMentalist Effect 235 → 235. Measure internet censorship 221 → 221. Divide by depth 216 → 216. Apple iPhone 18 Pro Camera test 208 → 208. MUNI Heritage Weekend 201 → 204. (Two — Heretic and Mini-AGI — have titles too generic for Algolia to return a reliable match, and a wrong number is worse than no number.)
The pattern is now unmistakable and it isn’t about AI at all: the three biggest movers of the day are an Android app store, a UK encryption essay and a refrigerator bricking, while every frontier model on the page moved by single digits. The audience is voting with points for things that touch their own devices and their own rights.
Throughline
The dominant theme is exit cost, and today it showed up in four institutions that don’t usually get discussed together. The Dutch government is rebuilding its workplace on NixOS and Fedora Kinoite because it cannot withdraw from a foreign vendor’s roadmap — and the thread immediately identified the obstacle that actually decides the outcome, which is not the operating system but whether LibreOffice can absorb a population that lives in Excel and SharePoint. The D.C. Circuit upheld the Pentagon’s blacklist of Anthropic, which is the same relationship read from the other side: a vendor tried to keep control over how its product was used and was removed from the supply chain for it. DHH used the Rails keynote to announce that HEY is leaving Rails, and the useful analysis — that Rails was a workaround all along, and that “convention over configuration” has been quietly redefined as token efficiency — comes from a community member, not the founder. Nokia’s archive is what the aftermath looks like: two decades of design decisions, rescued by a van in 24 hours, now licensed for research because nobody sells it. And at the smallest scale, F-Droid gaining 814 points overnight is the same instinct — somebody somewhere wants an app store that can’t be shut off by a platform decision. Every one of these is somebody repricing a dependency they previously treated as free.
Second: the day’s best engineering was about determinism, and the day’s worst idea was the opposite of it. The explainer-video pipeline is impressive not because a model wrote HTML but because it replaced the browser’s clocks with a virtual clock so that every frame is a deterministic seek — no screen recording, reproducible output, testable. Fearless SIMD and Go’s portable simd are both attempts to remove variability from performance work: one removes unsafe and the nightly toolchain from Rust SIMD, the other removes per-architecture intrinsic soup from Go, and both got there by measuring rather than promising — the community’s wasm benchmark shows portable SIMD at 11% slower than the non-portable API and 5x faster than scalar, which is a number you can act on. 86Box is determinism as a product: it emulates a 1990s PC down to chipset and bus timing so that period software behaves, and the practical consequence is that the machine’s single-core discipline matters more than its core count. Against all that sits DHH’s “we don’t even necessarily need to read the code,” delivered to an audience of people who maintain systems, in a talk where the same man is rewriting an application as six native clients because web fidelity wasn’t good enough. The one genuinely transferable technique in today’s keynote is the one it didn’t mention: if you can’t measure it as an integer, you’re not going to trust it, and if you’re not reading it, you’d better be measuring it.
Third: the front page spent today on AI’s supply side, and the audience rewarded the unglamorous work. Google is testing whether its chips survive radiation and vacuum, in a post that lists cooling as a challenge and admits nothing is settled. A dev-tools startup shipped a semantic diff viewer and a decision log because reviewing agent output is now a bottleneck. A microcontroller vendor shipped an explicitly-not-for-production Linux BSP and the community is porting kernels to it anyway, arguing in the comments about whether mainline’s removal of RISC-V execute-in-place was a mistake. A game studio re-engineered its isometric art into printable geometry and gave the files away. A person who doesn’t know JavaScript made a fish pond. Meanwhile the model launches from earlier in the week gained seven, nine and five points, and a court ruling about procurement outscored all of them. The story the numbers tell is that capability isn’t scarce anymore; what’s scarce is the plumbing, the verification and the ability to walk away.