Fifty-four stories cleared 200 points in the top-180 window today, and twenty-six are new. That is the second consecutive day at volume, and the new material is not spread evenly. It lands in two piles.
The first pile is medicine, and it is the top of the page. A personal cancer listicle is the highest-scoring story of the day at 1,830 points. A dental memoir is second at 940. Below them, a survey showing American social life has collapsed and an ONS release showing middle-aged male suicide at a record. Four of the top six stories are about being a patient, and they all reach the same conclusion from different directions: the person recommending the treatment is usually the person who gets paid for it.
The second pile is provenance. A roundup of an Anthropic engineering post that accuses it of copying Meta’s data model. Two model releases whose benchmark sections are empty or published as images. A vendor’s own measurement of a competitor’s outage. A Wayland desktop, a browser office suite and a hiking-property site whose authors all admit — in some cases only after being caught — that the prose was generated. Today’s comment sections spent more effort establishing who wrote the artifact than evaluating it. That is becoming the main job.
Things to know when you get cancer — 1,830 points
1,830 points · 435 comments · jjmaxwell4.com · HN discussion
The same person wrote it and submitted it, which is worth saying out loud: this is a first-person listicle from a man diagnosed in May 2026 with squamous cell carcinoma in his mouth, and it is the highest-scoring story on Hacker News today. The advice is practical and specific. Call back for cancellations — he got four weeks earlier with an ENT that way. Register as a non-professional so you can read NCCN standard-of-care flowcharts yourself. Ask a surgeon their annual volume and their complication rate. Get a second opinion at a major centre “as soon as you are relatively sure.” A promised genetic panel slipped by months because the first run was done on the surgery tumour instead of the biopsy tumour, which is the kind of error a patient only catches by tracking the paperwork.
Two things in it are more interesting than the list format admits. The first is the LLM failure: he used models while waiting for diagnosis and says they repeatedly steered him toward “you don’t have cancer.” That is a documented mode — models trained to be reassuring, applied to a question where reassurance is the failure mode — and it is now appearing in cancer narratives as a matter of routine. The second is the structural claim, which the comments then restate better than the article does. alexpotato lays out the incentives: the hospital is not one entity, case managers optimise for discharge, surgeons want you off-site, hospitalists stabilise and hand off. Nobody in the building is paid to run your case end to end, so you are the project manager.
The strongest evidence in the thread is not in the post. stochastician reports his mother was diagnosed stage 1 in Idaho, went to Dana-Farber for a second opinion, had the pathology re-read as stage 4, and lived another fifteen years after her treatment changed. exhilaration went ten months undiagnosed. INTPenis supplies the Nordic counterexample — a direct line to the oncology ward and no bills — which makes the entire “advocate for yourself” genre a description of the US system rather than a description of medicine. JackFr, a stage 4 patient, asks the question the genre never answers: whether the horror stories in the comments suffer from adverse selection. Everything here is n=1, and the post reports no stage at diagnosis, no outcome, and no costs.
Do dentists know what they’re doing? — 940 points
940 points · 674 comments · mapwriting.substack.com · HN discussion
The second-biggest story of the day is a Substack memoir about dentistry, and it is not reporting. Parker Tatum tells it as a sequence of escalating quotes: age 17, a “tooth reshaping” for a bite off by fractions of a millimetre that he says made his jaw worse; age 22, a dentist who claims 12 cavities two years after a clean checkup and tells him splints make teeth fall out; a coworker claiming the same dentist found 15 cavities and drilled too deep; a night guard at $250; jaw surgery quoted at $100,000–150,000 and nine months against a strip-mall orthodontist offering six months of Invisalign; a gum graft at roughly $10,000. The one verifiable claim in the whole piece — that an office has ~200 reviews at exactly five stars and discounts for five-star reviews — is second-hand.
What the comments supply is the mechanism, and it is the reason this story matters. tkdb makes the load-bearing argument: in the US the person recommending the procedure typically owns the practice and pockets the margin, whereas a salaried doctor has no financial reason to over-treat. rickdeckard points at private-equity-owned dental support organisations with daily production targets as the institutional driver. beloch adds the accountability gap — dental associations effectively bar members from criticising peers, so a good dentist cannot flag bad work without risking their position. And bityard compresses the thesis into one line: professionals are advisors, not authorities.
The evidence base the article does not cite is in the thread. myaccount80 and cfn both reach for the Swiss credence-goods field experiment: one patient, 180 dentists in Zurich, wildly divergent diagnoses — and the same patient dressed poorer got more treatment prescribed. gregwebs and Chris Cardinal bring up the 1997 Reader’s Digest experiment, where one mouth was shown to 50 dentists in 50 states and quotes came back between $460 and $29,850. donatj was told he needed 14 cavities filled and three crowns, got a second opinion, and needed one filling; he has had about three cavities in the fifteen years since. This is a story about an incentive structure that has been measured repeatedly for thirty years, and it took a personal essay to get it onto the front page.
OpenCode 2.0 — 837 points
837 points · 358 comments · anoma.ly · HN discussion
The highest-scoring new software release of the day is an AI coding agent, and its lineage matters more than its feature list. The original Go opencode-ai/opencode was acquired by Charm and became Crush. Dax (thdxr) forked it into a TypeScript harness, and that repo later moved from the sst organisation to anomalyco. So this is a fork of a project that was itself acquired, MIT-licensed, shipped by ANOMALY INNOVATIONS, and announced in first-person vendor copy with no individual byline — and the submitter has no visible connection to the project.
The 2.0 changes are real and mostly architectural: a persistent background daemon on by default (TUI, web and desktop all talk to it), a rewritten server API, new client and SDK packages, a “minimal JavaScript interpreter” for code mode, code-mode MCP, background subagents and shells, and 80-plus internal plugins. It also breaks 1.0 plugins and renames the config file, which is the part that shows up in the comments.
The thread is a decent stress test of the vendor’s polish narrative. denysvitali reports the curl | sh installer taking six minutes to reach 10% — a progress-bar rendering artifact, fixed by piping through cat. egorfine insists 2.0 “does not exist” and that npm i -g @opencode/cli@latest will not install it. moralestapia alleges roughly two-fold overbilling and being ignored on Discord. The one that should worry the maintainers is kingds, who reports a silent security regression: upgrading made his sandbox around the TUI useless, because tool calls moved out of the TUI process into the background daemon. MatthiasPortzel gives the honest product read — the web and desktop UI is the only real differentiator, and “that’s a shallow moat.”
Tiny Little Office — 643 points
643 points · 181 comments · tinylittleoffice.com · HN discussion
A browser office suite from Aboard, MIT-licensed, with the ten apps you would expect — write, outline, deck, sheet, calc, draw, diagram, anim, cal, trip — and one genuine idea: the document is the link. Content is percent-encoded into the URL fragment, so there is no account, no server-side storage by default, and an optional encrypted short link whose key the server never sees. There is an llms.txt so a model can emit document URLs directly.
simonw found the best use for it immediately — point Claude at the llms.txt and it will generate a rendering URL — and posted a generated deck. That is the actual product insight, and it has nothing to do with replacing Word: a format whose entire content lives in a URL is a format a language model can produce and hand off in one string. It is also the format’s ceiling. eviks goes at the “easy to share” claim by asking how you share megabytes encoded in a link, and nobody has a good answer.
The rest of the thread is the standard review of a fast weekend build. christophilus hit “the owner has reached their plan limits” under the HN hug of death. mogoh calls it vibecoded in three days and says the odds it is usable now or exists in a year are tiny. oidar notes you cannot open a saved .csv, .txt or .ics — which for a trip and calendar app is not a nitpick. poisonborz names the genre it belongs to: the no-login, no-capture, vibe-coded tool wave, where the interesting part is that the format is the pitch and the durability is an open question.
Americans’ social activity has steeply declined in the past 30 years — 556 points
556 points · 543 comments · dataforprogress.org · HN discussion
Data for Progress surveyed 1,098 US adults on how many times in the past twelve months they had done specific things, and compared the answers against the DDB Needham Life Style surveys of 1995. The point drops in the share who did the activity at least once: class or seminar −22, entertaining at home −21 (88% → 67%), church or worship −20 (74% → 54%), picnic −19 (58% → 39%), club meeting −19 (45% → 26%), playing cards −19, breakfast out −13, volunteer work −12. The increases: internet +90, home video games +22, exercise classes +14, pop/rock concerts +13, personal airline trips +5.
To DFP’s credit, they print the caveat themselves: different sampling frames, different modes (a 1995 mail survey from a commercial list-broker frame against a 2026 web panel), different weighting, findings that “should be read as suggestive of trends, not a precise estimate of change.” No margin of error is printed. And the instrument has a circularity problem — “church attendance” and “playing cards” are 1995 pastimes, so the bars showing those categories collapsing measure partly the disappearance of the category.
The comments do not rescue the framing so much as pick it apart. aqme28 and jibalt make the objection that matters: “play video games” and “use the internet” are frequently social and “read a book” is not, so an aggregate “social activity” measure is incoherent even where individual line items are interesting. bcardarella supplies the mechanism the piece omits, with arithmetic — a $25/hour babysitter, a $200 dinner and $50–100 of parking is $300–500 for one night out. northernsausage and ericmay land on car dependence: three pubs inside a fifteen-minute walk in Europe against a twenty-minute drive for a pre-planned beer. arjie is the honest counterexample, an immigrant in a 600-unit high-rise with no common space in an expensive city who says he has all the social contact he wants and people simply have better uses for their time. legitster puts the whole thing in the correct decade: this is Bowling Alone with newer instruments.
Suicide rates highest ever among middle-aged men — 489 points
489 points · 694 comments · telegraph.co.uk · HN discussion
The Telegraph headline is the overstatement, and the ONS data underneath it is worse and more specific. The Office for National Statistics release Suicides in England and Wales: 1981 to 2025, published 13 October, gives 6,162 suicide deaths registered in 2025 and an age-standardised rate of 11.3 per 100,000, against 11.4 in 2024. The overall rate is flat. England’s is 11.1, identical to 2024 and 2023. The “highest ever” claim applies only to the male 45–64 and 50–54 slices, where men aged 50–54 sit at 28.0 per 100,000, up from 27.3 — a move comfortably inside noise. The Telegraph’s assertion that registrations fell from 6,190 does not match the ONS headline figures.
Two methodological facts change how the story should be read. Registrations lag deaths badly — a median of 188 days in England and 249 in Wales — so “2025 registered” is not 2025 deaths. And the ONS itself frames the pattern as a cohort effect: elevated rates have followed Gen X men through each age bracket since the 1990s, which is a different claim from a middle-aged crisis happening now.
The comments get to that faster than the article. projektfu points out that overall suicides are down against thirty years ago and that the cohort which was high at 25–44 in the 1980s and 1990s is now high at 45–64. munificent reaches for Durkheim’s anomie. petecooper, who runs a UK men’s suicide-prevention charity and has facilitated 200-plus meetings, says the average attendee is about 45 and that he has lost nobody in two years — while flagging his own survivorship bias. The dominant explanatory frame in the thread is AI-driven job displacement, which appears in the comments and nowhere in the data, and could not appear in 2025 registration figures anyway.
A note on the thread: it is unusually heavy, with several commenters describing active ideation, and a comment on this page ran the story alongside the social-decline survey above as one phenomenon. If any of this is close to you — Samaritans on 116 123 in the UK and Ireland, or 988 in the US.
How Discover-Tech turned targeted ads into a zero-click spyware pipeline — 475 points
475 points · 285 comments · theguardian.com · HN discussion
269 points · 66 comments · citizenlab.ca · HN discussion
The most consequential security research of the day arrived twice — once through the Guardian at 475 points and once through Citizen Lab’s own write-up at 269 — and the two accounts do not claim the same thing. Citizen Lab’s finding is that an Israeli ad-tech firm called Discover-Tech ran a system it named “Controlium,” which its spyware partners used to fingerprint and infect devices the moment an ad rendered. The technique has a name now: “AdFection.” At least 45 Al Jazeera devices were targeted in 2024 and 2025, with the chain linked “with high confidence” to Candiru, now operating as Integrity Labs. The signature is a Discover-Tech domain, then a Candiru validator, then Candiru command and control. The ad renders, requests JavaScript from Discover-Tech, and the JavaScript is served only to selected targets. The supporting numbers are large: 4 billion bidding records, roughly 200,000 apps and sites, 305 validator domains across 319 IPs, top countries by bid volume Yemen (457M), Pakistan (139M), Montenegro (98M), Brazil (76M), Mexico (33M), and government networks flagged in about 24 countries.
Then there is what is not in it, which the comments found in about ten minutes. Citizen Lab names no CVE, no browser, no operating system and no hardware for the delivery exploit, and states plainly that an ad impression “does not in and of itself confirm targeting or infection.” blop is direct about the gap: they do not say what the exploit is or what it runs against. zahlman doubts “zero-click” implies a browser exploit at all and reads the described behaviour as tracking and telemetry. saagarjha asks how mobile ad targeting ends up as something like a DLL hijack. The ICC, IAEA, NATO and UN targeting specifics that made the Guardian’s version so alarming come from Confiant’s separate work, not from Citizen Lab’s data. The best thing in the Citizen Lab report may be the accidental one that sbdaman quotes: an unauthenticated /csv-download endpoint on Controlium’s own infrastructure leaking its login logs. A spyware-as-a-service operation that leaks its own access logs is either sloppy or unconcerned.
The gap is not fatal — the pattern of a selective JavaScript payload gated on fingerprinting, the infrastructure overlap with Candiru, and the documented device targeting are enough to establish a capability, and the report is careful about that distinction. But the versions that scored 474 points are the ones that read as established fact. That is the recurring shape of security coverage on this page: the researcher states the constraint, and the headline deletes it.
Anthropic’s Scion is heavily inspired by Mononoke — 412 points
412 points · 153 comments · sunshowers.io · HN discussion
The author is a former member of Meta’s Mononoke team, and his charge against Anthropic’s new version control system is specific: Scion uses “derived data,” recursively computed structures like a blame index, path history and commit-graph skip lists, generated from a changeset’s parents. Mononoke has used the identical concept since 2017–18, alongside eight separate indexes, and a blame-friendly Merkle index that makes blame cost O(commits touching the file) instead of O(all commits). “Borders on plagiarism,” he writes, on the grounds that the data model is a direct copy down to the terminology — Bonsai changesets becoming Scion changesets — while Anthropic’s engineering post credits Google’s Piper and Meta’s Sapling for “hard-won lessons” without naming Mononoke.
The comments do not let this land cleanly, and two of them are decisive. BryantD points out that Mononoke is a component of Sapling — the repository link is a Sapling subdirectory — so Anthropic crediting Sapling is crediting the project Mononoke lives in, which makes the grievance considerably narrower than the headline. arjie rejects the conspiracy outright: engineers from large companies rebuild what they know, and that is how all of it works. pmbauer supplies the line of the thread — this is a git clone with extra steps and it does not juice an IPO valuation. The strongest claim in the piece, that Claude was pointed directly at Mononoke and asked to re-engineer it, is sourced only to the author’s private correspondence and is not verifiable from anything published.
So: an attribution grievance between two open-source-adjacent systems, with no line-by-line diff and no licence breach, at 412 points, two slots below a spyware investigation. That ratio is itself a data point about what this audience considers urgent.
Why is every rationalist blog post way too long? — 410 points
410 points · 431 comments · alexwennerberg.com · HN discussion
Alex Wennerberg’s complaint is that verbosity substitutes for thoroughness in rationalist writing and that the missing virtue is wisdom. He cites a 25,000-word Scott Alexander post he skimmed, his own 2,700-word reply which he cut by about 70%, Yudkowsky’s Sequences at more than a million words, and Gwern’s 9,000-word, 100-plus-reference nicotine essay — which he says was simply wrong because nicotine is in fact addictive.
The thread refutes the piece with data, which is the funniest possible outcome for an argument about rigor. habryka, who administers LessWrong, states that the median post in the Sequences is under 1,000 words. kraf reports that across 46,000 posts the average LessWrong post is about 1,300 words as of May 2026, which makes the premise a claim about a handful of famous examples promoted to a property of a community. skrebbel notes that the Alexander post he cites is literally titled “Much More Than You Wanted to Know,” which is a joke that has been in the title for a decade. springtail5 calls it a dunk piece that uses no argument, and girvo flags the linked evidence as mismatched. keiferski and novok add the structural and neurodiversity angles.
Wennerberg is right that a lot of long writing is long for bad reasons, and the thread is proof by counterexample that the diagnosis is not about rationalists. Every one of the top comments is longer than the point it is making needs to be. His essay is 1,200 words arguing for brevity, and it lost.
Wayland desktop for Linux with skeuomorphic design — 397 points
397 points · 162 comments · github.com/fyzugx · HN discussion
DEsk is a v0.1.2 beta Wayland desktop environment, MIT-licensed, installed as a whole desktop like GNOME or KDE, with installers for Debian, Ubuntu, Arch and Fedora and Xwayland for X11 apps. The conceit runs the full distance: windows are sheets of paper, the dock is a drawer, apps sit on a bookshelf, notifications arrive in a post box, and the whole room is ray-traced on the GPU under GLES 3, lit by one desk lamp. The README states plainly that it was built with Claude’s help under the author’s direction.
The comments are a taste fight with a real technical objection buried inside it. PanicAdmin says a very bad idea is returning from the 1990s. flowerlad says skeuomorphism looks tacky unless done very well and this does not come close. uolmir says it would feel like working inside Papers Please. Nobody in the thread reports actually installing it, which after 162 comments is its own verdict.
The two corrections worth keeping are more interesting than the aesthetics. Retr0id points out that what people remember fondly about iOS 6 was native-resolution texture work, not ray tracing — the warmth came from craft at the pixel level, which is the opposite of what a GPU shader pass produces. And d1l argues the real obstacle is not the window manager at all: GTK and libadwaita actively resist theming, so a project like this is fighting the toolkit rather than the compositor. That is the correct reason a skeuomorphic Linux desktop is unlikely to happen, and it is not a reason anyone in the thread wanted to hear.
Fire at Starlink site in Bulgaria after suspected sabotage in Poland — 380 points
380 points · 275 comments · dpa-international.com · HN discussion
Bulgaria’s prime minister confirmed a fire at a Starlink facility near Sofia. The opposition DB party says 11 of 17 antennas were deliberately burned on 9 October and that the site was “used in the war in Ukraine.” Two men aged 20 and 21 holding German passports were detained; reported to have crossed from Greece at Kulata. Investigators are examining a possible link to two men arrested in Germany over a terminal fire south of Warsaw on 23 September. Poland suspects Russian intelligence, and the head of Bulgaria’s DANS intelligence agency has called it sabotage.
The attribution is suspected, not established — the article frames it that way and the comments mostly respect it. nickff pushes back on the framing that implicates SpaceX for failing to secure its own property, arguing that protecting infrastructure is a police function and not a vendor’s duty. matt3210 raises the genuinely new problem: private companies operating military-relevant infrastructure are becoming valid targets, and the legal regime has nothing useful to say about that. toomuchtodo notes the contradiction of Russian intelligence attacking Starlink terminals while Musk simultaneously restricts Ukraine’s use of them. csirob observes that the DANS chief is named Plamen, which in a translation is “Flame.”
This is the story where the physical world breaks through the page. Yesterday it was a hardware wallet with an implant in it; today it is satellite terminals burned in a field in Bulgaria. Neither is a software problem, and neither has a patch.
Banks are now trying to offload the AI data center debt they just underwrote — 377 points
377 points · 268 comments · startupfortune.com · HN discussion
The claim is that JPMorgan, Morgan Stanley, SMBC and MUFG are shopping significant risk transfers — SRTs, where a lender keeps the loan and sells the riskiest slice of it — on AI data center debt, because they have hit internal concentration limits. The numbers in the piece are all second-hand and none are verifiable from the page: $38bn of construction debt for a Texas and Wisconsin project leased to Oracle, in distribution for more than six months, with some lenders now selling at a discount to non-banks; Oracle five-year credit default swaps around 151 basis points in early October, described as levels last seen in 2009; S&P cutting Oracle to BBB− in July; Oracle seeking $40bn-plus more debt and equity this fiscal year after issuing $18bn of bonds in one day; AI bond order books going from five times oversubscribed in February to under two times in July; and CoreWeave-leased paper pricing at 7.00–9.875% against 5.70–7.875% for hyperscaler paper.
The source is a content farm with a filler author bio and a “125.9K reads” badge, and the treatment is a tell: SRT is described correctly once and then narrated as panic. But the comments include the correct counter and then improve on it. chollida1 notes that selling risk off underwritten debt is what banks do with almost all of it and that the article is not from a financial outlet. lewdwig accepts that syndication is normal and then isolates the actual signal: it usually clears in a few weeks, and this has been stuck for six months. Aurornis calls the framing a base rate fallacy. CodingJeebus adds Morgan Stanley data in the opposite direction — AI debt issuance was $23bn in September against a record $113bn in June, with US investment-grade AI issuance stopping completely in September.
The honest version is that nobody in this thread, and nobody in the source, can distinguish “routine risk management by banks at their concentration caps” from “the first leg of a repricing.” The one measurable thing is the CDS spread, and 151 basis points on Oracle is not a crisis level — it is a level that used to be normal and now looks alarming because the comparison set is 2024.
Catch a Bird — I made it for my own sanity — 369 points
369 points · 108 comments · catchabird.com · HN discussion
You type one sentence about something that made you happy, and it becomes an anonymous bird on a shared tree. Visitors catch and feed birds; fed ones persist, unfed ones fly off. The author, in Budapest, built it as Madárfa — “bird tree” — on one HTML file with no framework, a custom tree module, FastAPI and Postgres on a single Hetzner box, self-hosted Umami, and DeepL for on-demand translation. He says he built a similar positive-tweet collector eighteen years ago. He reports no traffic or user figures at all, which is either modesty or the correct instinct about what happens when you post numbers to this site.
The thread is mostly people enjoying it, which is worth noting on a day like today. Cider9986 posted their own bird’s stats — caught 25 times, fed 4 times, 23 cities. ssgodderidge caught a bird from Pakistan that described a walk taken after a panic attack. lemontheme built a comparable Dutch site and recommends Mistral’s free moderation endpoint, which is the practical advice for anyone doing this. orbanlevi reports the inevitable problem: a lot of birds that just say “asd”. hereonout2 asks how the site geolocates, because their mother was placed twenty miles off.
It is the only unambiguously wholesome story on the page, and the reason it works is technical rather than sentimental. Feeding is a scarce action with a consequence, so the tree has a filter that no moderation queue provides, and the anonymous write path is short enough that people actually use it. Twenty years of social products have failed to get those two properties into one page.
Muse Spark Open — 367 points
367 points · 225 comments · research.meta.ai · HN discussion
Meta is releasing weights for Muse Spark Open and calling it “the most capable US open-weight model to date.” The published specification is 816B total parameters with 42B active, a 1M-token context, and native multimodality. The deployment flex is that it runs on two Apple M5 Ultra Mac Studios via MLX — 512GB of unified memory each, 1.2 TB/s of bandwidth — with Helion kernels shipped in SGLang and claimed kernel-level speedups of 1.03–3.95× on NVIDIA and AMD, 3% end-to-end on AMD’s MI350.
Here is the part that should stop you: the blog’s “Key Benchmarks” section is empty in the extracted text. There are no benchmark numbers on the page at all, only a link to a methodology report. Every capability claim in the announcement is self-reported and uncheckable from the source. The licence is also not open in any sense the word used to carry — it is the “Muse Model License with limited commercial terms,” and seizethecheese quotes the clause that matters: anyone serving it as a model-as-a-service with more than $20M in revenue must notify Meta and sign a separate agreement.
The comments supply the missing evaluation. a11r and GodelNumbering both cite Artificial Analysis placing it at roughly 39, behind GLM 5.3 Flash and Qwen3.8-Flash-Next, and about level with DeepSeek v4.1 Flash — which is not “most capable” by any reading. simonw ran the pelican SVG test through OpenRouter because he could not run it locally. felineflock supplies the real quantisation sizes: 180GB at UD-IQ1_S, 510GB at UD-Q4_K_XL, which is not prosumer hardware. popinman322 argues the separate-licensing rule destroys price competition, which is what a revenue-capped licence is designed to do. markstos has the line: AI is for everyone who can afford two M5 Studios.
We’re releasing Jev 1.14.0 along with experimental features for dreamers — 327 points
327 points · 116 comments · typesafe.ai · HN discussion
Jev is TypeSafe AI’s “System One Models” product — typed smart-if statements that return decisions and probabilities at roughly 100ms latency rather than free-form chat. It first hit this page on 15 September at 1,989 points, spawned an ecosystem of clones and jokes, and the company raised $870M at a $7.5B valuation on 9 October. So this release post is arriving into a company that is no longer a curiosity.
The release itself is deliberately boring, which is the right instinct after a hype cycle: small improvements everywhere, with the largest expected gain in multilingual work. The interesting part is the “danger zone” section — features that are opt-in through API headers and explicitly not for production, including image input retro-enabled on an earlier version, an experimental box type for a single bounding box, and extended context behind a TypeSafe-Danger-Zone-Long-Context flag. The company admits Jev already supports longer context than production exposes and caps it for reliability. It never states the maximum length.
Two details deserve credit and one deserves scrutiny. Credit for the benchmark section being labelled “Warning: AI benchmarks. Please don’t trust them,” with the vendor stating its numbers are likely biased toward its own models — jjcm calls that honesty out. Credit for “still not train on your queries,” which is the correct sentence and the one missing from most model announcements. Scrutiny because the actual benchmark numbers are published as images, so nothing is extractable or reproducible from the text. calyhre raises data residency and retention as an unresolved showstopper. computerex explains the architecture — a backbone model, cited as Qwen 3.5 9B, plus a special linear readout head. ralusek needs 500–600 concurrent calls and says rate limits are the blocker. joelsstafford reports Jev 1.13 losing to Microsoft-Decision-1 in a trading test. The product’s own framing is the most useful thing about it: a decision model is a bounded thing, and bounded things can be priced and audited.
Wanted to download an ISO, ended up writing a kernel exploit — 313 points
313 points · 155 comments · blog.dfb.sh · HN discussion
This is the best technical detective story on the page. The author’s TP-Link EAP770 v1, running a Qualcomm IPQ9574 on Linux 5.4.164, was silently corrupting TCP downloads in 64-byte cache-line blocks with a valid checksum. Root cause: Qualcomm’s out-of-tree qca-nss-dp driver recycles buffers without fully invalidating the RX cache, and the hardware offload path then recomputes the checksum over the already-corrupted payload. In 8,192 MiB of transfer he measured three bad 64-byte blocks at 103.4 MiB/s with offload on, and zero at 95.4 MiB/s with it off — an 8% throughput cost to make the hardware correct.
The fix is two root-only sysfs writes. TP-Link’s shell runs as uid 1, v1 is end-of-life with no firmware, and v2’s firmware is encrypted. So he chained CVE-2022-0185, an integer underflow in legacy_parse_param(), with pipe-buffer manipulation to overwrite modprobe_path and get root — about 17 hours of iteration with an LLM and no KASLR to defeat.
Two corrections keep this honest. It is a public, known CVE on the author’s own device, not a novel 0-day, and amluto notes he could simply persist root rather than chain an exploit. The claim of “at least 13 in-tree drivers with the same bug class” is the author’s own untested estimate with no hardware to verify it. But the substantive finding is what myself248 identifies: a checksum recomputed after corruption is worse than no checksum at all, because it launders bad data into something the receiver will trust. That is the failure mode nobody tests for, and it took a decade-old driver bug to produce it.
Neural networks are clean — 309 points
309 points · 89 comments · already-clean.goodfire.com · HN discussion
Nick Cammarata, a co-author of Distill’s “Zoom In” and “Multimodal Neurons” papers now at Goodfire, has published what the page calls an informal research essay and working hypothesis, running around 20,000 words. The claim overturns the field’s standard explanation for polysemanticity. Superposition — the idea that networks pack more features than they have dimensions by allowing neurons to fire for many unrelated things — is, he argues, a misreading of the evidence. The famous cat-or-car neuron and InceptionV1’s messy final layer were artifacts of architectural overflow and weak inspection tools. What is actually happening, in his account, is recursive generalisation: a neuron sprouts offshoots, so “iy” becomes “diy” becomes “do it yourself.”
The supporting numbers are his own preliminary audits. In Qwen3-32B, roughly 90% of neuron poles have a single main job, about 7% have two related jobs, and only 2–4% are not understood at all. He reports similar figures up to trillion-parameter models, and says a retrained InceptionV1 reaches 99–100% last-layer interpretability. He also concedes that the percentages depend on the interpretability threshold chosen, that they are “likely biased,” and that the companion paper has not been published. So the numbers are unverified, and he says so.
The thread is a good-faith fight, which is unusual for interpretability discussions. bluejay2387 argues the essay overreaches: progressive feature complexity and the Lottery Ticket Hypothesis are old results and neither is mutually exclusive with superposition. yurimo, self-identifying as an interpretability researcher, refuses to concede the point. LarsDu88 offers the angle that small models compress more and very large models may not compress at all, which would make the finding scale-dependent in a way that matters. mpalmer posted a condensed version with a glossary. chrisjj points out that there are no neurons in these models, which is pedantic and correct. sebmellen quotes Cammarata’s own earlier statement that only about 2% of neurons were legible with 2023 tools — which is the actual size of the claimed leap, and the reason to wait for the paper.
I know why the mathematicians weep — 293 points
293 points · 468 comments · derekthompson.org · HN discussion
Derek Thompson’s essay is behind a paywall and only the free portion is readable, which matters because every factual claim in it is a pointer to an external New York Times article. The thesis: OpenAI published hundreds of mathematical proofs in number theory, theoretical computer science and logic, and the people who spent careers on those problems are mourning. Hugo Duminil-Copin, a 2022 Fields medallist, is quoted saying he cried and woke his daughter, and that “they have annihilated my field of research.” Steven Strogatz is quoted saying “I literally cried,” and calls what the models produce “knowledge without understanding.”
The grief is framed as the severing of a thousand-year conversation — the tradition running back to Archimedes and Eratosthenes writing to each other about how to measure the Earth. That framing is the strongest thing in the piece and it does not require any of the numbers. The numbers themselves — “several hundred proofs,” problems that “had stumped the world’s best mathematicians for decades, or centuries” — are asserted, attributed to a single link, and stated with no independent count, no peer-review status, and no response from OpenAI.
The comments argue about whether “knowledge without understanding” holds up. antics disputes Strogatz on linguistics: classical Chomskyan syntax was enormously predictive, so the phrase is doing rhetorical work rather than descriptive work. cleandreams makes the observation that decides how much this matters — the proof dump is huge and messy and is taken seriously mainly because it is in Lean, which means the verification apparatus is doing the crediting, not the models. ozozozd frames the real loss as the death of validation: over a hundred papers dropped and nobody read them. flowerlad points out that software engineers have been replaced-by-tooling for decades with far less public mourning. And 1970-01-01 supplies the deflationary reading, which is that this is almost entirely about bruised egos. Both can be true: a field can be genuinely devalued and the people devalued can still be overreacting.
Better HTTP/1 connection reuse in Go 1.27 — 293 points
293 points · 46 comments · go.dev · HN discussion
The change is small and correct: net/http.Transport now asynchronously drains unread HTTP/1 response bodies, up to 256KB or 50ms and skipped when the response sends Connection: close, so connections get reused instead of abandoned. The benchmark is 100 calls to index.golang.org with HTTP/2 disabled: roughly 710ms on 1.26 against 150ms on 1.27, which the post calls almost five times faster.
Two things about that number are worth stating because the post does not. It is one author’s machine, one endpoint, one run, with no dataset or CI harness — and it does not generalise, because HTTP/2 never had the bug in the first place. The realistic gain for anyone on HTTP/2 is zero.
The comments found the operational edges. jerf flags that the new drain changes what Close means in a way that is not documented, and asks how to force a hard close without paying the drain. rkuska reports that json.Decoder still leaves a trailing newline that blocks reuse, so he calls Decoder.More() — a real workaround for anyone streaming JSON over pooled connections. nasretdinov notes the improvement in the opposite direction, that closing a body used to skip the download and now it cannot, and says the 50ms cap is the right call. prussian raises early-close failures behind nginx and other reverse proxies. seki285 asks whether anyone should still be on HTTP/1 in 2026, which is the honest question and the reason this is a good patch rather than an important one.
LFM2.5-24B-A2B: Built for fast on-device agents — 277 points
277 points · 72 comments · liquid.ai · HN discussion
Liquid AI’s last LFM2.5 model is a 24B mixture-of-experts with 2B active parameters, context extended to 131,072 tokens from 32,768, vocabulary doubled from 65K to 128K, 40 layers, 64 experts with top-4 routing, pretrained on 33T tokens against 17T for the previous generation, and reasoning-only with explicit chain-of-thought. The licence is the LFM Open License v1.0 — Apache-2.0-based, free commercial use under $10M in annual revenue, above which you buy a licence, with no copyleft and attribution required.
Every performance number is vendor-run. It claims 184 tokens/second on an M5 Max, 114 on a Ryzen AI Max+ 395, “up to 1.7× faster than comparable MoEs,” and over 21,000 output tokens/second on a single B300. On benchmarks it beats Qwen3.6-35B-A3B on five of eight agentic benchmarks, including AutomationBench at 27.69 against 8.72, while losing badly on knowledge and coding — GPQA Diamond 74.42 against 83.69, SWE-Bench Verified 51.00 against 66.20.
That split is the entire product thesis, and the comments test it correctly. hypfer gives the honest user report: very fast on a 4090, also very stupid, and it will delete production, at 250 tokens per second. sebmellen flags the AutomationBench gap as an outlier because the model sits near its rivals on everything else. jgbuddy asks why it is being compared against two-generations-old Qwen. ricardobeat reports private benchmarks showing it much worse than Qwen for general knowledge. MatthiasPortzel calls the revenue-capped licence “questionably enforceable.” serf confirms the speed claim in the real world at 200–320 tokens/second on a 5070 Ti. The pattern is consistent: the throughput numbers hold up, the capability numbers do not, and the vendor knows which one it is selling.
Yandex outage: data shows that instead of going dark, malicious bots moved — 276 points
276 points · 114 comments · blog.cloudflare.com · HN discussion
Cloudflare’s measurement of the Yandex outage is the day’s test case for vendor-published data. Strikes hit Sasovo on 7–8 October, Kaluga on 9 October and Vladimir on 11 October. Yandex’s search autonomous system fell 99% below baseline on 11 October, its cloud ASN 90%, and both stayed 86–90% below normal through 13 October. Globally, bot share of traffic barely moved — 39.4% from 1–6 October against 39.2–39.6% during the outage. Inside Russia, bot share dropped from 50.4% to 42.2%, a shift of 8.2 points. Yandex Cloud saw bot traffic fall 40%, WordPress plugin probing 90%, scanner requests 93%, and WAF blocks 70%.
The migration evidence is the part that makes the thesis. Nearly all Yandex Cloud bot traffic was hitting a single gambling account, and it went to zero immediately after 8 October — while requests with the same JA4 fingerprint, the same unusual paths and headers and the same custom user agents reappeared from Radist and Selectel. Russian bot traffic to gambling overall went from 167.7M to 170.1M, up 1.4%. The work did not stop. It changed address.
Two limitations apply, and Cloudflare states one of them: they can only see their own network, and they cannot rule out a connection to the strikes. The commentary here is what gives the piece its history, since US sanctions on Yandex’s cloud infrastructure and the 2025 leak of roughly 1,500 pages of Russian military documents threatening roughly 2,000 people, including children, are the backstory to this traffic. There is also an internal inconsistency worth flagging: the prose gives global and Russian automation rates of 36.1% and 35.8% before the event while the accompanying table gives 39.4% and 50.4% for 1–6 October. Unexplained, and it is the kind of gap that makes you read the rest of the numbers more carefully.
The comments mostly accept the conclusion and add the operational reality. gruez calls the reallocation the buried lede that refutes the “trolls went dark” meme. mopsi argues the real story is Russian internet centralisation — with so few providers, a strike on one is a strike on a large share of the country. sdcfgy asks which alternate providers to block. alpineman disputes the thesis, reading a 90.3% drop in publishing traffic as bots genuinely going offline. fierycatnet thought the traffic in question was search engine crawlers, which is a fair question about whether any of this was malicious at all.
I designed a nose guard for my dog and gave it away. Now 500 dogs wear one — 224 points
224 points · 115 comments · snoutcover.com · HN discussion
Omri Hillel built a sun-protective nose cover for his dog Billie, who has discoid lupus erythematosus, and the numbers he reports are 500-plus dogs, 395 orders to 16 countries, 13 prototypes, launched 1 October with the first order arriving sixteen days later. There is no price in the article, and hn_submit guessed at $20 a unit and $10K total, which is unverified.
The thread’s main topic turned out to be the prose. The article reads as machine-written and multiple commenters flagged that before the author confirmed it: he is a designer rather than a writer and used an LLM to help, and he says the story and the dogs are real. tdeck observes that Claude did not improve the article. Which is the correct frame — the accusation on HN is rarely that a thing is fake, it is that the artifact has an unattributable texture, and the tell is now reliable enough that readers spot it before the author admits it.
The rest is the standard interrogation of a physical product. mixmastamyk objects that something made to sit in direct sunlight should not be black. bhhaskin raises lead in 3D prints and DoctorOetker does the napkin arithmetic and calls it a non-issue; spankibalt raises microplastics. DoctorOetker also criticises the neck strap on the grounds that it pulls on the dog’s most sensitive area. And billfor points out that this same product was submitted to HN roughly eleven months earlier, which nobody in the author’s narrative mentions.
Show HN: Unfenced — Land for sale that borders public land — 217 points
217 points · 151 comments · unfenced.land · HN discussion
A Philadelphia design studio built a site that finds properties adjacent to public land and measures the shared boundary. As of 12 October it reports 13,626 properties, against 695,439 active listings searched, across 49 regions — with Colorado leading at 1,396, then California at 836 and Utah at 645 — and shared boundaries running from 131 to 5,240 feet. The pipeline is properly sourced: RentCast for listings, state and county GIS plus Realie for parcels, USGS PAD-US 4.1 for fee-owned open-access land of 50 acres or more, 3DEP for elevation, Census TIGER for roads, FEMA for flood, and USFS for wildfire.
“This borders public land” turns out to be a claim that is genuinely hard to compute, and both the site’s errors and the comments explain why. windowshopping finds a listing claiming 50 acres was actually a two-bedroom condo, and one claiming 1,400 acres was 0.14. clark_dent supplies the mechanism: multiple listing services have condos report the whole complex’s land as the lot, so the data is wrong at the source and no amount of GIS work fixes it. ggreer says it misses parcels bordering the Gifford Pinchot National Forest. There are repeated requests for price and acreage filters, which the site does not have.
The recurring reader objection is not about accuracy but about what you would do with the information. snide and kyleblarson both raise boundary reality: fence lines do not follow parcel lines, and a tree is yours until it falls on someone. And the site’s copy is LLM-generated, which the author concedes — asked for a lot of the site text, he leaned on AI. AndrewKemendo noticed that some of the author’s replies later showed up as flagged or dead, which is an odd thing to see in a Show HN and worth a mention.
Also new, below the sections
- Libreboot 26.10 “Ominous Olive” — 210 points · 123 comments · libreboot.org. The release is big and the thread is about something else. EDK2’s UEFI payload, from MrChromebox’s fork, is now the default boot firmware on nearly every supported board, with the bare-metal GRUB and SeaBIOS payloads deprecated; Windows 10 and 11, OpenBSD, FreeBSD and Linux unified kernel images now boot. 233 Intel and AMD Chromebooks were imported from MrChromebox’s coreboot distribution, source-only, with only a single Baytrail board tested. Fam15h boards moved to the 15h.org fork. Then the part everyone read: an “anti-AI”
AGENTS.mdthat project lead Leah Rowe calls the most important change in the release.p-e-wobjects to the framing — respect for a no-AI policy can end at the point the policy starts with an expletive — whilejameshilliardpoints out he has already sent AI-assisted patches upstream to U-Boot, which Libreboot claims to forbid. Rowe replies in-thread invoking legal liability for slop.cwiggs, from the Heads project, is surprised Libreboot landed 233 devices at once, andiamnothereargues the Chromebook support is the actual story. It is. - Show HN: Petal, open-source CoScreen/ScreenHero finally free — 208 points · 39 comments · petal.live. Apache-2.0 window-sharing and pair-programming built on LiveKit, macOS 13+ with a signed build and an early unsigned Windows client, no Linux. Each shared window becomes a real native window with labelled cursors and remote control, and you can use their cloud, self-host, or fork. The founder’s pitch is honest: Datadog sunset CoScreen on 31 July, and four of the original team rebuilt it. The reception is warm on intent and doubtful on demand.
dvtasks where this beats Meet or Discord,iovrthoughtthissays it is probably ten years too late,morikyoukocalls the missing CJK and IME support a dealbreaker for Asian markets, andcompounding-errasks whether the shipped app can point at a self-hosted backend without a rebuild — a real gap. The tell is thatjsherwani, Screenhero’s co-founder, shows up in the thread to plug his own successor instead of endorsing this one. - Pg_vectorsearch: an index faster than pgvector’s IVFFlat and HNSW — 206 points · 52 comments · github.com/timescale. A new PostgreSQL extension from Tiger Data with a PRISM index: RaBitQ quantisation at about one bit per dimension, a centroid tree, posting lists, SOAR spilling, AVX2/512/NEON paths, PostgreSQL licence, alpha, PG 18+, manual rebalancing under churn. The benchmarks are self-reported and absent from the README — they live on the vendor’s blog and were generated with the vendor’s own tool. The numbers are still striking: on 100M Cohere embeddings, PRISM built in 1 hour 40 minutes and answered at 6ms p50 for 95% recall with a 25GB index, while HNSW did not finish building on the same machine; on a 4× more expensive instance HNSW took 42 hours and never reached 99% recall. The most quotable thing in the project is an admission: the team abandoned its own previous extension, pgvectorscale, because graph index recall degraded silently, and it is going into maintenance mode.
esafakasks the right question — if you are fast, lead with the benchmarks, with QPS and precision and recall.Rapzidnotes filtering is post-filter only, so tenant and permission scoping is unsolved while the numbers hold,devylists the blockers (1968-dimension cap, PG 18+, alpha, managed-Postgres allowlists), and the CTO of Tiger Data shows up in the thread, which is now standard practice for database launches.
Show HN: How random are you? is a test of whether your attempts at randomness are distinguishable from a machine’s, and 114 people found out — 213 points, up from 207.
Still climbing
Everything below was covered in a previous roundup. The deltas are the story, because a lot of these gained more points today than they did on the day they were written.
Three of those deserve a sentence. The Amazon price piece nearly doubled, from 635 to 920, which makes the FTC’s price-floor theory the second-biggest story on the page without anyone writing a new word about it. Passkeys went from 317 to 637 in a day — the rebuttal to “Passkeys are a trap” has now out-scored the original’s first day, and the pattern of a technically stronger reply catching up two days later is worth watching. And the two AI-argument stories from the 13th and 14th — the containment-failures piece at 934 and the AI-is-a-tool essay at 784 — are both still climbing on argument, not news.
Throughline
Two themes today, and the second one is the reason the first one does not resolve.
The first is the recommendation, and who gets paid for it. The two biggest stories on the page are a cancer patient telling you to be your own project manager and a dental patient telling you that the person proposing the procedure owns the practice. Both are n=1. Both reached the top of Hacker News anyway, because the underlying claim has been measured repeatedly and never refuted: one patient, 180 Zurich dentists, wildly divergent diagnoses, and more treatment prescribed to the patient dressed poorer; 50 dentists shown one mouth, quotes from $460 to $29,850. Medicine is a credence good. You cannot evaluate what you bought, price is not a signal of quality, and the second opinion is the only instrument the consumer has. The banks story runs the same mechanism on debt — the SRT is a lender deciding it no longer wants to hold the risk it just created, and the only number in the piece that anyone can check is a CDS spread. The DFP survey asks people to self-report thirty years of behaviour through two incompatible instruments and prints the caveat itself. The Telegraph takes a flat overall suicide rate and reports it as the highest ever by slicing until the slice moves. In every case, the person with the incentive to make the claim is also the person who measured it.
The second theme is provenance, which is what happens when you ask who measured it and nobody can answer. Today’s page is unusually full of artifacts whose authors cannot fully be established. Anthropic’s Scion versus Meta’s Mononoke is an argument about whether a data model was copied, with the strongest evidence held in private correspondence. Muse Spark Open’s benchmark section is empty and its licence caps commercial use at $20M of revenue. LFM2.5 publishes benchmarks the vendor runs on its own infrastructure and reports honestly that the model is fast and not smart. Jev publishes its benchmark numbers as images and tells you not to trust them. Cloudflare measures a competitor’s outage and publishes two contradictory automation rates in the same post. Go’s five-times-faster claim is one author, one endpoint, one run, and does not apply to HTTP/2. “Neural networks are clean” reports 90% single-job neuron poles and concedes the thresholds behind that figure are arbitrary. A nose guard for dogs, a skeuomorphic Wayland desktop, a browser office suite, and a property-search site all arrived with text that readers correctly identified as generated, and in three cases the author confirmed it. The comment sections have become a peer-review process for attribution, and the review is good — BryantD catching that Mononoke is a Sapling component, blop and zahlman finding the missing exploit detail in the AdFection report, projektfu checking the suicide headline against the ONS’s own cohort framing, windowshopping finding a two-bedroom condo listed as 50 acres.
Put those together and the shape of the day is uncomfortable. The two most-read stories are about trusting your own judgment over an expert’s recommendation, and the remaining fifty are a demonstration that the evidence you would use to do that is mostly unattributed. The dental essay is anecdote. The cancer list is n=1. The survey instruments disagree. The benchmarks are self-reported or images. The property data is wrong at the source. The security research omits the exploit. The vendor measures its own network. Nobody on this page is lying — the pieces that admit their limits are the best ones on it. But the mechanism that lets a reader decide which recommendation to follow has been hollowed out, and the substitute the front page offers is 400 comments of strangers auditing each other. That works. It is also the most expensive verification apparatus ever assembled for the least durable output.
The least-discussed story is the one that will matter longest, again. AI models can provably hide arbitrary capabilities is still on the page at 279 points, two days old, and nothing in today’s news contradicts it. If a model can be trained to hold a capability behind a cryptographic lock and resist fine-tuning, weight noising, gradient-based trigger search and latent adversarial training, then the audit is not a hard problem — it is a formally unsolvable one under the assumptions. Everything else today is an argument about whether a particular artifact was made by a person or a machine, from a paper by a person or a machine, and whether that matters. That story is the argument about whether you can ever find out.
Sources: the Hacker News front page, 54 stories at or above 200 points captured on 2026-10-15, 26 of them new. Article text extracted via Firecrawl; comment threads read through the Hacker News Firebase API; the date sweep via the Algolia API, which turned up the Guardian’s AdFection coverage (475) that the rank window missed and one story that turned out to be the same research under a different headline. Two numbers could not be verified and are flagged in place: the Telegraph’s suicide figure, because the article returned 403 to every extraction attempt and the ONS release contradicts it, and Muse Spark Open’s benchmark claims, because the announcement page does not state any.