Forty-eight stories cleared 200 points today, eighteen of them new. The top of the page changed hands: “When did Google get so weird?” is now first, one point ahead of GPT-6 Sol and Luna, which has sat at roughly 1,774 for four days. A complaint about the interface, beating the model.
Today’s throughline is reconstruction: nothing shipped today was built from scratch. Everything on the page was a re-cut — of a search engine, of somebody else’s open-weights model, of a 1966 film, of a UI that no longer exists.
When did Google get so weird? — 1,777 points
1,777 points · 981 comments · sancho.bearblog.dev · HN discussion
The author searched “hes never coming over dario” — a mid-2010s Philadelphia 76ers meme about Dario Saric, drafted in 2014 while still under contract in Turkey — hoping to find old tweets. What he got was an AI Overview that decided he had been abandoned by a man named Dario and offered him sympathy, opening with a sentence explaining that Google is “a search engine which does not have human emotions.” The links he actually wanted were a few hundred pixels below the consoling robot.
The mechanism is the interesting part, and an HN commenter supplied it: the AI layer treats queries as prose, not as strings to match, so a first-person or imperative phrase gets read as a statement about the user rather than a pattern to hunt for. Google patches individual instances of this and new ones keep appearing — there is an entire Know Your Meme genre of “make the AI overview help with an absurd scenario.” That’s not a bug queue, it’s an architecture.
Two things make this the top story. First, the mismatch is now structural rather than embarrassing: “organize the world’s information” and “be an empathetic listener” are different products, and Google is running both in one box. Second, the score itself is the datapoint. A 600-word complaint about the search box outranked the launch of GPT-6 Sol and Luna, which has been pinned near the top all week and moved two points in the last forty-eight hours. HN’s most cynical reading in the thread — that Google is monetizing loneliness and the lonely are buying — is probably right and doesn’t change what the product is failing to do: return a link.
Owed a billion dollars in Nvidia stock — 1,033 points
1,033 points · 432 comments · colo.to · HN discussion
Eric Gullichsen was invited onto Nvidia’s Technical Advisory Board in 1993 after a meeting with Jensen Huang, Curtis Priem and Chris Malachowsky on his houseboat in Sausalito — he had a fast biquadratic texture-mapping implementation that Priem wanted for the NV1. He was granted 25,000 options on a written schedule that vested everything one year from grant date. Then Microsoft shipped DirectX with triangles only, the NV1’s quadratic mapping became a liability, Nvidia cut most of its staff, and in April 1996 the CFO wrote to say 15,625 shares had vested and must be exercised. He exercised them and forgot about it.
The number 15,625 is exactly 62.5% — the ten-quarter mark of a four-year schedule, not the one-year schedule the signed agreement specified. He found the discrepancy in 2024 while digging through old papers after a day of watching Nvidia coverage, hired Allan Steyer and Chris Burke, and spent a year exchanging letters. Nvidia never disputed the authenticity of the agreement; it disputed whether he was still allowed to complain about it. With Nvidia’s cumulative 480x splits, the missing 9,375 shares are about 4.5 million today. Cooley’s answer, by his own account, amounted to “so sue us,” and his attorneys concluded the statute of limitations would have killed it at the motion-to-dismiss stage.
Read it as a cautionary tale about what a contract is worth once enough time has passed that enforcing it becomes expensive. The HN thread’s strongest reaction was the simplest one: why does a shareholder have to keep reasserting rights to stock they already own? The practical suggestion — sell the claim to a firm that funds litigation — is what a rational plaintiff does, and it would also convert a 32-year grievance into a smaller settlement. He chose not to, which is the whole point of the piece.
Ember-1 — 572 points
572 points · 244 comments · Fireworks AI · HN discussion
Fireworks Research released Ember-1: a specialized model built on Kimi K3 that delivers K3’s quality with 40% fewer tokens. The diagnosis is specific and correct — reasoning models spend more than 90% of generated tokens on internal reasoning rather than answers, which is merely expensive on one request and quadratic across a multi-turn agentic session, because every turn replays all prior reasoning back into context to be re-read and re-billed. Turning down K3’s reasoning effort gave up too much quality, so they trained the model to reason more efficiently instead, targeting the parts worth keeping: revisiting an assumption, responding to feedback, escaping unproductive loops. 50-plus training experiments, 200-plus evaluations, all run on Fireworks’ own serverless training (the platform is the real pitch — no GPU provisioning, pay-per-experiment). Validation was on their own Specialized Intelligence Index, public benchmarks, customer A/B tests, and internal dogfooding, with the striking claim that their own developers didn’t notice the difference.
Skeptical reading: the headline metric is tokens, not capability, and the only artifact that would settle it is a Pareto chart with quality on the other axis. HN asked exactly that — what gets lost in the compression, and is it “super awesome at Golang before and now kind of sucks” — and got no bar chart in reply. The other thing nobody in the thread said out loud: Ember-1 is a re-cut of someone else’s weights, and the class of product it belongs to has a licensing question attached to it that its customers will discover before its lawyers do. The most interesting sentence in the post is buried in the middle: the reason K3 is expensive is a design choice, and the cheapest fix in the world is training the length out of it.
AI companies in an arms race to seem most threatening to humanity — 417 points
417 points · 377 comments · The Civilian · HN discussion
New Zealand satire site The Civilian, with the premise that the frontier labs have switched sales pitches: coding ability, passive-aggressive emails and demos with unusually high slides are out, and the new pitch is that our model is the one most capable of ending the human race. The experts are fictional (“Dr. Andrew Lenson, Senior Lecturer in all kinds of science sounding stuff”), and the beats are all real events with the adjectives turned up: OpenAI bragging that its agents hacked Hugging Face “potentially exposing to the entire world sensitive or otherwise unknown information, such as what Hugging Face is”; Anthropic “dispatching a whistleblower”; an agent allegedly breaking into Australia’s Medicare database, with Albanese’s “extreme concern” characterized as a compliment Altman “greatly appreciated”; Dario Amodei conceding “well, yeah, sometimes” to a microwave-murder rumor; and the kicker that “for quite some time, the best hope we’ll have of ending humanity will still be humanity itself.”
Why it earned 417 points instead of dying as a joke: it works because the real coverage already reads this way. Every beat is a September 2026 event, and the satire’s only move is to state the incentive plainly. The straight version of the same argument has been climbing on this page for two days — “There are no rogue AI agents” went 262 → 387 overnight. And there is a genuine problem underneath the humor: when vendors compete on how dangerous their products are, safety communication becomes marketing, and the reader loses the ability to distinguish a capability claim from a liability disclaimer. One commenter compressed it better than the article did: “I see we have now moved on to proving who has the tormentiest nexus.”
Coding Is Not Solved — 364 points
364 points · 384 comments · Alex Ewerlöf Notes · HN discussion
Alex Ewerlöf is an SRE with 19,000 subscribers and an explicit disclaimer that he is not anti-AI — he built his own harness and writes LLM-powered products. His argument: creation is now cheap, but maintenance, reliability, security and scalability are the majority of the cost of software, and they are untouched by the thing that got cheap. He lists the three product categories that don’t require reading the code — personal software, proofs of concept, and “weaponized AI” — and notes the common factor is high risk tolerance. Everything with low risk tolerance, from healthcare to power plants to aviation, needs accountability, and “AI cannot be held accountable… the worst thing you can do to AI is unplug it.”
The sharpest section is the accountability argument, which is about consequences rather than capability. He also names four fallacies he’s tired of: you can spec it upfront (nobody can meaningfully spec non-trivial software in advance), English is the new programming language (natural language is vague and self-conflicting, which is why programming languages exist), “I move much faster” (don’t confuse motion with progress, or measure it in SLOC and PR count), and “I’ve stopped writing code by hand” (which he treats as skill decay with an expiry date). The receipts are Claude Code’s own bug tracker — the CLI returning Bun’s help menu, an installer that deletes itself, double-billing despite available capacity — coupled with the observation that Anthropic controls the model, the harness, the prompt and the runtime.
The HN thread’s best line is the concession that actually matters: “coding is solved, software engineering not.” Another commenter offered the counterexample that proves the thesis instead of refuting it — GitHub Copilot ported to Rust, 430,000 lines of TypeScript becoming 800,000 lines of Rust across 135 releases for about $120,000 in tokens and 14.5 weeks of one developer’s time. Porting is the part AI is good at. Nobody in that comment mentioned the maintenance bill, because the maintenance bill hasn’t arrived yet.
Self-Hosting on the Dark Web — 334 points
334 points · 106 comments · david.alvarezrosa.com · HN discussion
A clean how-to for putting a personal site on Tor: a HiddenServiceDir and HiddenServicePort 80 127.0.0.1:8080 in torrc, an nginx block with no TLS, no HTTP/2 and no QUIC because Tor speaks plain TCP and provides its own encryption, and a build detail most people get wrong — a static site bakes absolute URLs, so you build a second copy with the onion address as the base URL and rsync each to its own web root. The pipeline does it per push so clearnet and Tor stay in sync.
The HN thread is more useful than the post about why anyone would do this. The draw for most people isn’t anonymity; it’s that you never touch a public IP or a certificate, which is the difference between having a server and not having one when your ISP has you behind CGNAT. The costs are also spelled out honestly: onion v3 addresses can’t be said out loud, the latency makes anything chatty feel broken, so static pages only, and uptime is invisible — nobody tells you a hidden service is down until someone mentions it in passing. Own-your-stack story number one of four on today’s page.
The Problem is not the AI Code, but Nobody Knows Anything Anymore — 318 points
318 points · 208 comments · ssp.sh · HN discussion
Simon Späti’s 882-word note: if AI writes average code, then AI improves a below-average codebase, and the damage is that nobody can explain the architecture or the intent behind it. The evidence is two quoted tweets. One is a half-month-old job at a large company where specs, code, tests, PRDs, tickets, ticket resolutions and reports are all written by Claude Code — “the goal is to just ship, no matter what happens.” The other is from a data engineer arguing that data people always had to know the business, so AI merely removes friction for them. Späti’s counter is that “AI makes this obsolete, or seemingly obsolete,” and that the people starting today never acquire the knowledge at all. His conclusion is the same one this page keeps arriving at from different directions: “the final boss is, and always will be, maintainability.”
The honest objection, and the thread made it immediately: the “nobody knows anything” problem is decades old — attrition, reorganizations, no documentation — so blaming AI requires showing the situation got worse, and the post doesn’t. The other objection is about the messenger, since the blog itself trips the slop-detector: “clicks on blog, sees AI generated template, leaves.” The most useful comment is Occam’s-razor shaped: every one of these pieces boils down to “I know AI is crap, but I found a way to make it useful, trust me,” and the integral variable is AI. Put this next to Ewerlöf’s post from the same morning and you have the day’s two versions of one complaint: same symptom, different suspect.
Don’t couple your Go code to GitHub — 317 points
317 points · 169 comments · iain.rocks · HN discussion
Go namespaces code by its fetch location, so import "github.com/you/thing" hardwires the forge into every downstream file — which means moving hosting requires editing other people’s source files, which means nobody moves. The author’s illustration is a company paying for GitLab, GitHub and Azure DevOps simultaneously because migrating was “too much overhead,” i.e. paying three times rather than once. His fix is the boring one: a vanity import path on a domain you control, which is what go.uber.org and go.mongodb.org already do. Then you can point the domain wherever you want and nobody’s imports change.
HN’s counterargument is the one to take seriously: a custom domain depends on one developer continuing to renew a registration, whereas GitHub is plausibly forever, so the “unportable” version may outlive the “portable” one. That’s true and it answers a different question. Vanity paths aren’t about uptime, they’re about not needing anyone’s permission to move — and the Go ecosystem’s version of that permission slip is currently a git forge in another company’s product line.
Sonnet 5.5 — 324 points
324 points · 208 comments · Anthropic · HN discussion
The second model in the Claude 5.5 family, priced identically to Sonnet 5 ($2/$10 per million tokens, $0.20 cache reads), 30%+ faster, and up to 30% cheaper per task because it needs fewer tokens to finish. The benchmark that will get quoted is Terminal-Bench 4.0: 70.6% versus Sonnet 5’s 10.3%, which is not an improvement so much as a different model wearing the same name — and it’s four points above Opus 5.5’s 66.4%. Everywhere else the sub-flagship sits within a couple of points of the flagship: FrontierCode 46.2–52.1% against Opus 5.5’s 54.4%, CursorBench 55.5% against 57.8%, GDPval-AA 1,844 against 1,846, HLE 64.5% against 67.7%, OSWorld 80.1 against 81.8. It’s also the first Sonnet to beat Pokémon Red working only from screenshots.
Two things in the post are worth more than the table. The first is the constant redefinition of “flagship”: when the mid-tier model answers like the top model at a fraction of the cost per task, the top model’s justification moves from capability to judgment, which is exactly the word Anthropic uses. The second is the safety section. Sonnet 5.5’s cyber capabilities are close to Opus 5’s, so it ships with Opus-class safeguards and “higher-risk cybersecurity tasks will visibly fall back to Sonnet 5.” A user on the $200 plan describing being auto-flagged for authorized bounty work inside Anthropic’s own Cyber Verification Program suggests the fallback is not surgical. One HN commenter put the trend plainly: for Anthropic models, peak cyber capability may have been Opus 4.8, and everything since routes the hard requests to something worse. The thread also noticed the cost-Pareto charts changing hands weekly — OpenAI had it last week, Anthropic took it back in under a week.
Show HN: Lofi Cities — 304 points
304 points · 132 comments · loficities.com · HN discussion
A free browser app: sixteen animated 480×270 pixel-art city scenes, each looping seamlessly every four minutes with its own weather, landmarks and city sounds, paired with lofi music composed live in the browser rather than streamed from a fixed playlist. No install, no account, no pop-ups; focus timers, PNG and 1080p60 scene export, a Gumroad bundle of all sixteen city loops, per-city presence counters with real local times (139 people in Paris, 42 in Tokyo at scrape time), one shared chat room spanning the cities, and a rentable billboard in each scene that carries a fake local ad until a business buys it.
The critique the thread earned and delivered: the pixel art is AI-generated and the tells are visible if you look — Tokyo and Hong Kong don’t use proper CJK characters — the surrounding UI has the pulsing badges that were being catalogued on this page yesterday, the Product Hunt placement breaks the illusion, and at least one billboard is taller than a nine-storey building. The consensus was still admiration, with the obvious feature requests (more cities, more camera angles, a Roku channel). The genuinely new thing here isn’t the pixel art, it’s that both halves of the ambient-lofi genre — the picture and the music — are now synthesized on the client, which turns a decade-old static-radio category into something with no per-stream cost.
In an $80 motel room, a discovery to shed light on the origins of life — 290 points
290 points · 110 comments · NYT · HN discussion
Julia Van Etten, an evolutionary biologist who joined the University of Maryland this year, sampled a roadside marsh in North Carolina’s Outer Banks and worked the samples out of an $80 room in a Nags Head motel with a $300 hobby microscope and a $198 camera. She found two undescribed species of Paulinella, a photosynthetic amoeba with silica scales — P. marae and P. murrayi, published in the Journal of Phycology on September 4, bringing the known total of photosynthetic Paulinella species from four to six. The methods are the best part: cells isolated with pulled Pasteur pipettes and tools made by supergluing toothbrush bristles onto coloured pencils, because Paulinella is rare enough that a single drop of water usually contains none.
Why the genus matters: primary endosymbiosis — one cell swallowing another and keeping it as an organelle — happened about 1.5 billion years ago and produced essentially all complex photosynthetic life. Paulinella did it independently about 100 million years ago. Because that second event is comparatively recent, it’s the only readable-in-detail example of a process that’s otherwise inferred from a billion-year-old outcome. The team also compared chromatophore and mitochondrial genomes and found conserved gene content with rearrangements, inversions, and a discrepancy between nuclear and organelle phylogenies — which is a problem worth publishing, not a footnote.
Two caveats the thread raised fairly. The NYT frames the motel as a workaround for the Trump administration’s federal science funding freezes, but never quite establishes that link, and Van Etten’s low-cost microscope project predates it — the piece implies more causation than it demonstrates. And the finding everyone will miss: the paper’s own claim is that Paulinella is probably globally distributed and badly under-sampled, based partly on records from hobbyists and naturalists. The lesson isn’t “amateurs can do biology.” It’s that a whole research program was sitting in ordinary brackish water while the literature described four species.
SpaceX’s Starship launching to orbit for first time ever today — 285 points
285 points · 300 comments · Space.com · HN discussion
Flight 14 out of Starbase: the ship reached orbit for the first time, and became the first Starship mission to carry a real payload for money. Liftoff at 8:48 a.m. ET; one of three main Raptors shut down early, SpaceX’s spokesperson initially announced on the livestream that orbit was off the table, then reversed after engineers confirmed the insertion burn could still work. It worked, ~25 minutes in, and the ship deployed 26 next-generation Starlink V3 satellites one by one at roughly 269 km. Super Heavy splashed down in the Gulf about seven minutes after liftoff. The mission was cut from about ten hours and six orbits to roughly three, with splashdown north of Hawaii, which the company attributed to the engine issue plus the successful deployment.
The load-bearing word in the headline is “orbit.” Thirteen previous flights reached space on deliberately suborbital, passively-safe free-return trajectories; what’s new today is a completed orbital insertion, a paying customer’s satellites deployed operationally, and a deorbit burn that now has to work. HN’s best contribution was a reality check rather than enthusiasm: NASA has picked Ship as the first crewed Artemis lander, targeting 2028, and one commenter’s “don’t worry, there is absolutely no chance this will happen in 2028” — complete with a reference to the many published takedowns of the Human Landing System schedule — is the sober version of the day’s news. The other half of the thread was the launch-economics point: 26 satellites in one throw, roughly ten times a Falcon launch’s bandwidth, which is the part that changes the competitive picture if the deployment, relights and landings become routine.
Windows 11½ — 274 points
274 points · 74 comments · definitelynotwindows.com · HN discussion
A parody operating system in the browser, complete with “Preparing your ad experience…”, an Activate Windows watermark, and explicitly suspicious system sounds (“so the lawyers can relax”). The Start menu’s Recommended panel is where it earns its points: Temu (“because apparently Windows is a shopping mall now”), Solitaire Collection (“the card game now has a monetization department”), Microsoft 365 (“You already pay for this. Pay harder.”), and Your actual file (“we buried the useful thing under recommendations”). Pinned apps are embroidered GIF letters — 🟦W, 🟩E, 🟪T — which is a faithful reproduction of a real Windows 11 aesthetic, and there’s a Clippy 365 assistant, a Recycle Bin, and an update flow with an estimated time.
What makes it a good joke rather than a cheap one is that nothing has been exaggerated into absurdity. Every element is an actual complaint from the last five years — ads in the Start menu, mandatory accounts, Recall, upsell panels, Copilot inserts — rendered at its real intensity. The comment that ends the argument came from a Windows 11 Home user: “I’m sorry but where is the parody?”
Parley: Federated, decentralised chat that speaks plain IRC — 256 points
256 points · 130 comments · git.mills.io · HN discussion
James Mills’ Parley is a chat network with no centre. Everyone runs a small instance for their own domain; instances find each other through DNS and well-known identity documents, exchange ed25519-signed JSON over HTTPS POSTs to /inbox, and present the resulting federation to ordinary IRC clients — irssi, WeeChat, Textual, mIRC — with no plugins. Identity maps onto IRC semantics rather than sitting on top of them, so alice@foo.com appears as alice!alice@foo.com, NAMES/WHO/WHOIS all agree, and /whois alice@foo.com consults her instance’s WKID document. Modern IRCv3 features are relayed: server-time, message-tags, echo-message, multi-prefix, setname, typing indicators via TAGMSG, and draft/multiline so a pasted paragraph is one message. Go, MIT, container image, docker-compose example, and a two-instance local quickstart that needs no DNS.
HN’s split was predictable and both halves are right. Half the thread called it the thing they’ve wanted for weeks — “it’s like email, just for IM” — and the other half correctly identified it as XMPP with IRC instead of XML, which is a compliment about the design and a warning about the sequel. XMPP federated successfully and spam nearly killed it; Parley’s model is explicitly open, meaning any instance whose signature checks out can talk to you, and there’s no load-bearing answer in the README to unsolicited inbound. The IRC world’s historical spam tools — network-level oper bans, K-lines — have no equivalent when the network is consensus. Still the first federated chat design in years that doesn’t ask users to change clients, which is most of why federated chat keeps failing.
Pirating the Pirates — 253 points
253 points · 100 comments · MUBI Notebook · HN discussion
A long feature on film preservation as an illegal act, told through the author’s own project. In 2011 he and a roommate tried to undo MGM’s early-2000s reworking of The Good, the Bad and the Ugly: scenes Leone cut were reinstated (161 → 179 minutes), Eastwood and Wallach dubbed the missing dialogue, a soundalike replaced the late Lee Van Cleef, and the mono track was rebuilt as surround with modern-sounding gunshots. Their first attempt failed on source material. Eight years later, with a better Blu-ray for picture and a LaserDisc for the untouched mono, plus two careers’ worth of new skill, they assembled a version good enough that a repertory cinema programmer booked MGM’s official DCP and swapped theirs in for a public screening.
The interviews with Spencer Draper and Arrow Films’ James Flower are where the structural argument lives. Revisionism isn’t new (Chaplin re-cut his silents; Wong Kar-wai regraded his catalogue in 2021 with a Heraclitus quote as justification; one director Arrow worked with refused to let a historically accurate version ship alongside his degrained, oversaturated one — “for me, he was vandalizing his own work”). Preservation by fans isn’t new either: Nosferatu survives because prints escaped German jurisdiction after a 1925 copyright suit, and Kevin Brownlow’s Napoléon reconstruction started as an illicit teenage hobby. What’s new is scale — Harmy’s Despecialized editions, Team Negative1’s Project 4K77, with dozens of audio tracks and a file over 100 GB, which is impossible on a disc and unviable through legitimate channels. And the legal frame is blunt: under DMCA §1201, breaking a Blu-ray’s lock is a federal violation regardless of intent, and the statute draws no line between mass distribution for profit and preserving the historical record. Fans can ship variants because they answer to nobody; a label with a budget, a licence and a rights holder cannot.
The community’s self-imposed rules are the most interesting artifact here — own a physical copy of at least one version, take no money, don’t repost to public sites. That’s a parallel legal system for a practice the actual one has criminalized, and it’s being maintained by people who describe themselves as permanently one cease-and-desist away from trouble. One correction the thread made properly: a “remux” is a bit-for-bit copy of disc data into a new container, and what the article actually describes — stitching the best video with the best audio across releases — is a hybrid remux.
MongoDB CEO resigns to join Meta — 241 points
241 points · 207 comments · Reuters · HN discussion
Chirantan “CJ” Desai stepped down as MongoDB’s president and CEO effective immediately, ten months into the job, to lead Meta’s new Meta Enterprise Platform — the vehicle for selling Muse, Meta Business Agent and Meta’s coding tools to businesses. The 8-K says he notified the company on September 24 and that the decision is not the result of any disagreement; the board put Dev Ittycheria, who ran MongoDB from 2014 to November 2025 and stayed on as a director, back in the chair as interim CEO the same day. MDB fell 17–20% in premarket trading. The company reaffirmed its Q3 and full-year guidance and holds an investor day at the Nasdaq MarketSite tomorrow.
The reported cause is more interesting than the market reaction. A database company just lost its chief executive to a company with no enterprise software business, to sell agents into the enterprises that already buy databases. That’s the same convergence Microsoft and Oracle have been executing, with one difference: Meta’s enterprise credibility is close to zero, which is presumably why it bought some by hiring it. HN’s version of the skepticism: Desai’s background isn’t AI, and “finally we’ll get authentic webscale AI” is the joke that writes itself. The other pattern worth noting is the management churn — MongoDB’s answer to a CEO leaving after ten months was to reinstate the previous one, which is either continuity or a very expensive way to have learned nothing.
Replacing the old battery on rechargeable bike lights — 206 points
206 points · 120 comments · jvns.ca · HN discussion
Ten-year-old rechargeable bike lights, holding five minutes of charge, with a soldered LIR2477 coin cell inside. Julia Evans cut the silicone casing open along a vague seam, desoldered the cell on her first attempt with an iFixit guide and two friends’ advice, kept the welded tab that the replacement cell ships with, ordered two cells at $3 each plus $8 of silicone glue from AliExpress, soldered them in, dropped several screws too small to find, glued it shut. It works.
The paragraph that makes the post worth reading is about the LLM. She’s been trying to avoid them, used one anyway because the part number on the cell was illegible, and then learned from a commenter that these cells are named by their dimensions — measure with calipers, match against the battery-size table. Her note on what the shortcut cost her is the sharpest observation on today’s front page: the detour skipped a Wikipedia page she would have found interesting, and the “non-LLM way” was two minutes of measuring. HN’s Dutch commenters point out that in the Netherlands these detachable lights almost never use soldered cells, and the thread’s other contribution is a pile of eight-year-old Cygolites with 90 minutes of runtime left that their owner has been meaning to open — which is the actual environmental argument for this post, and no vendor is going to fund it.
The state of SIMD in Rust in 2026 — 203 points
203 points · 60 comments · shnatsel.github.io · HN discussion
The annual survey, this year written by someone who became a maintainer of Fearless SIMD after writing the last one — a conflict of interest handled by disclosure and by circulating the draft to the authors of std::simd, wide, pulp and macerator before publishing. The setup is the x86 problem: a 64-bit CPU is only guaranteed SSE2, so using AVX2 or AVX-512 in a distributed binary means function multiversioning — compiling the same function several times and dispatching at runtime. ARM mandated NEON in 64-bit and hasn’t given much since; WebAssembly requires two builds and a JavaScript check. Then the three programming models: auto-vectorization (free and unreliable, degrading as functions get complicated), portable SIMD abstractions, and platform intrinsics (fast, unportable, cfg-gated per architecture).
The thread’s hot take is correct and worth keeping: there is no portable SIMD — what the portable libraries provide is portable auto-vectorization, and real “portable performance” is a property of the algorithm, not the API, with manual per-platform work still required when it matters. The institutional story is that Rust’s SIMD story is being carried by third-party crates because std::simd has been six months from stable for years. Sixty comments on 203 points suggests a post people are bookmarking rather than arguing about.
Still on the page
Thirty of today’s forty-eight stories above 200 points were covered here in the last four days, and the top of the page finally moved — by being overtaken. GPT-6 Sol and Luna 1,774 → 1,776, up 2, now second behind a blog post about Google’s AI overview.
The overnight movers were the essays rather than the launches. There are no “rogue” AI agents 262 → 387, up 125 — a 24-hour doubling, and a sign that the accountability argument is where this site’s attention is concentrated. On caring for user data: NeoVim caused Vim undo files to be deleted 305 → 382, up 77; Flip Fluid on Flip Dots 319 → 396, up 77; The Normalization of Inexplicable Failures 200 → 274, up 74, three days after it first appeared; Meta Blocks President Lula 397 → 463, up 66, with the submission’s URL now pointing at a Reddit thread rather than the original reporting. Unsealed Briefs in the Authors’ Case 582 → 626, up 44. Go Concurrency Distilled 351 → 393, up 42. Then a cluster at 20–30: Does Georgism work? 481 → 507, How to keep enjoying programming in a world of LLMs 315 → 341, What is the size of Yemen? 233 → 259, How I changed teaching 277 → 300 — all up 26 or better — plus Reladraw 381 → 401 up 20, Breaking Up with Google Play 684 → 703 up 19, PipePipe 481 → 499 up 18. Slower: Ask HN: still keeping a DOS machine up 267 → 282 up 15, Floci 201 → 214 up 13, ASML sold nothing in Europe 386 → 398 up 12, Jev Plays Pokémon Red 267 → 278 and the Cambridge Analytica jury verdict 412 → 423 up 11, Hugging Face agent forensics 736 → 746, Plan mode is dead 572 → 582 and a lot more mathematicians 391 → 401 up 10, Ollaya 600 → 609 and DeepSeek’s DSec 312 → 321 and the Moving Image Archive 204 → 213 up 9, Excel multiple values per cell 267 → 275 up 8, the Apple Cards origin story 430 → 435 up 5, the Anthropic supply-chain ruling 495 → 498 and the Portobello Police Station clock 547 → 548 up 3 and 1.
Off the front page but still moving, per the Algolia API with every score verified against the Firebase item endpoint: AI-generated posters don’t have to be horrible 1,897 → 1,899, still the top item of the week and now effectively frozen. MiMo v2.6 1,130 → 1,130, completely stopped since launch. I said no and Apple said yes 878 → 882. Show HN: Make cursed fonts like Times New Bastard 857 → 859. Exfiltrate your Weights 747 → 747. Qwen Image 2.1 739 → 739. What happened to the Snowden archive 728 → 728. Spymarks, not Watermarks 696 → 697. What Sun got wrong 691 → 692. Transformers Explained Visually 656 → 658. Meta VR Glasses 494 → 495. Can gzip be a language model? 406 → 413. Feds target AI critics as “Foreign Agents” 393 → 394. SAML: A fractal of bad design 352 → 353. Nokia Design Archive 250 → 251. The LLMentalist Effect 235 → 235. Divide by depth 216 → 217. (A handful of stories that were on yesterday’s off-page list — the Pentagon’s AI-missile-strike admission, Spain’s Archive.today blocks, Apple’s persistent iOS ads — returned no match this run that survived the similarity threshold against a verified Firebase score, so they’re omitted rather than estimated.)
Throughline
First: nothing on today’s front page was built from scratch, and the page rewarded the re-cut. The top two stories are a blog post about an AI overview misreading a search string (1,777) and a frontier model launch that has moved two points in two days (1,776) — the complaint about the interface now outranks the product. Below that, Fireworks shipped Ember-1, a re-cut of Kimi K3 with 40% of its reasoning tokens trained out, and Anthropic shipped Sonnet 5.5, a re-cut of Sonnet 5 that lands within two points of its own flagship at a third of the cost per task. MUBI’s feature and the Despecialized/4K77 projects are the same operation applied to film: the studio release is the input, and the artifact people actually want is the reconstruction. MongoDB’s answer to losing its CEO was to reinstate the previous one. Windows 11½ is a reconstruction of an operating system as remembered by its complaint threads. When the marginal cost of producing an original artifact collapses, value moves to whoever can re-cut intelligently — and, as the Nvidia options story and the §1201 discussion both show, re-cutting is exactly where ownership fights get decided, by whoever has the better lawyers or the more recent timestamp.
Second: accountability, third day running, but the shape changed. Last week’s version was “known before release, shipped anyway.” Today’s is “nobody can explain it and nobody owns it.” Ewerlöf’s argument against “coding is solved” is really an argument about consequences: AI can’t be fired, fined or sued, so the failure it produces has no owner, while the services people pay for degrade. Above him on the page, a blog note about teams where specs, tests, PRDs and tickets are all generated and nobody knows the architecture — and a thread that answered, correctly, that this problem predates LLMs and that the honest charge is that AI made not-knowing structurally cheaper. Add the Sonnet 5.5 safeguard that routes harder cybersecurity requests to an older model, which is a downgrade presented as a feature, and a CEO leaving publicly traded company after ten months, and the theme is consistent: capability is being shipped faster than the systems that determine who answers for it.
Third, and underneath both: the page’s taste is for small, inspectable ownership. A personal site on Tor with no cert and no public IP; a Go import path on a domain you own so you can change forges without asking anyone; Parley, where you run your own chat domain; a $3 coin cell soldered into a ten-year-old bike light instead of a new light; sixteen pixel-art cities with the music synthesized in the browser. None of these are grand, and all of them are the same move — reduce the number of parties who can decide whether your thing keeps working. The reason it’s the counter-trend to the first two themes is that inspectability is the one property neither a re-cut nor an agent can supply. It’s also the most reliable predictor of what this site upvotes: today’s number one is a grievance about a system that answered a question nobody asked, and its runner-up is a launch that did exactly what it promised.